Use 32-bit big write location (OK for up to 512 TB OSDs)

This is required to:
1) later inline the last "big_write" entry into the object to slightly reduce memory usage
2) eliminate an ugly hack where entry type is determined by its size
3) make the storage scheme extensible i.e. when adding new entry types

.gitea/workflows/test.yml

@@ -414,6 +414,24 @@ jobs:
           echo ""
         done
 
+  test_degraded:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_degraded.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_rm:
     runs-on: ubuntu-latest
     needs: build
@@ -810,6 +828,60 @@ jobs:
           echo ""
         done
 
+  test_reweight_half:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_reweight_half.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
+  test_snapshot_pool2:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_snapshot_pool2.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
+  test_snapshot_read_bitmap:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_snapshot_read_bitmap.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_heal_csum_32k_dmj:
     runs-on: ubuntu-latest
     needs: build
@@ -954,24 +1026,6 @@ jobs:
           echo ""
         done
 
-  test_snapshot_pool2:
-    runs-on: ubuntu-latest
-    needs: build
-    container: ${{env.TEST_IMAGE}}:${{github.sha}}
-    steps:
-    - name: Run test
-      id: test
-      timeout-minutes: 3
-      run: /root/vitastor/tests/test_snapshot_pool2.sh
-    - name: Print logs
-      if: always() && steps.test.outcome == 'failure'
-      run: |
-        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
-          echo "-------- $i --------"
-          cat $i
-          echo ""
-        done
-
   test_osd_tags:
     runs-on: ubuntu-latest
     needs: build

CMakeLists.txt

@@ -2,6 +2,19 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VITASTOR_VERSION "2.3.0")
+set(VITASTOR_VERSION "2.4.3")
 
+include(CTest)
+
+add_custom_target(build_tests)
+add_custom_target(test
+	COMMAND
+	echo leak:tcmalloc > ${CMAKE_CURRENT_BINARY_DIR}/lsan-suppress.txt &&
+	env LSAN_OPTIONS=suppressions=${CMAKE_CURRENT_BINARY_DIR}/lsan-suppress.txt ${CMAKE_CTEST_COMMAND}
+)
+# make -j16 -C ../../build test_heap && ../../build/src/test/test_heap
+# make -j16 -C ../../build test_heap && rm -f $(find ../../build -name '*.gcda') && ctest -V -T test -T coverage -R heap --test-dir ../../build && (cd ../../build; gcovr -f ../src --html --html-nested -o coverage/index.html; cd ../src/test)
+# make -j16 -C ../../build test_blockstore && rm -f $(find ../../build -name '*.gcda') && ctest -V -T test -T coverage -R blockstore --test-dir ../../build && (cd ../../build; gcovr -f ../src --html --html-nested -o coverage/index.html; cd ../src/test)
+# kcov --include-path=../../../src ../../kcov ./test_blockstore
+add_dependencies(test build_tests)
 add_subdirectory(src)

csi/Dockerfile

@@ -36,7 +36,7 @@ RUN (echo deb http://vitastor.io/debian bookworm main > /etc/apt/sources.list.d/
     ((echo 'Package: *'; echo 'Pin: origin "vitastor.io"'; echo 'Pin-Priority: 1000') > /etc/apt/preferences.d/vitastor.pref) && \
     wget -q -O /etc/apt/trusted.gpg.d/vitastor.gpg https://vitastor.io/debian/pubkey.gpg && \
     apt-get update && \
-    apt-get install -y vitastor-client && \
+    apt-get install -y vitastor-client ibverbs-providers && \
     wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-utils_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
     wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-block-extra_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
     dpkg -x qemu-utils*.deb tmp1 && \

csi/Dockerfile.local

@@ -0,0 +1,49 @@
+# Compile stage
+FROM golang:bookworm AS build
+
+ADD go.sum go.mod /app/
+RUN cd /app; CGO_ENABLED=1 GOOS=linux GOARCH=amd64 go mod download -x
+ADD . /app
+RUN perl -i -e '$/ = undef; while(<>) { s/\n\s*(\{\s*\n)/$1\n/g; s/\}(\s*\n\s*)else\b/$1} else/g; print; }' `find /app -name '*.go'` && \
+    cd /app && \
+    CGO_ENABLED=1 GOOS=linux GOARCH=amd64 go build -o vitastor-csi
+
+# Final stage
+FROM debian:bookworm
+
+LABEL maintainers="Vitaliy Filippov <vitalif@yourcmc.ru>"
+LABEL description="Vitastor CSI Driver"
+
+ENV NODE_ID=""
+ENV CSI_ENDPOINT=""
+
+RUN apt-get update && \
+    apt-get install -y wget && \
+    (echo "APT::Install-Recommends false;" > /etc/apt/apt.conf) && \
+    apt-get update && \
+    apt-get install -y e2fsprogs xfsprogs kmod iproute2 \
+        # NFS mount dependencies
+        nfs-common netbase \
+        # dependencies of qemu-storage-daemon
+        libnuma1 liburing2 libglib2.0-0 libfuse3-3 libaio1 libzstd1 libnettle8 \
+        libgmp10 libhogweed6 libp11-kit0 libidn2-0 libunistring2 libtasn1-6 libpcre2-8-0 libffi8 && \
+    apt-get clean && \
+    (echo options nbd nbds_max=128 > /etc/modprobe.d/nbd.conf)
+
+COPY --from=build /app/vitastor-csi /bin/
+
+ADD deb /deb
+
+RUN apt-get update && \
+    apt-get -y install /deb/vitastor-client_*.deb && \
+    wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-utils_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
+    wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-block-extra_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
+    dpkg -x qemu-utils*.deb tmp1 && \
+    dpkg -x qemu-block-extra*.deb tmp1 && \
+    cp -a tmp1/usr/bin/qemu-storage-daemon /usr/bin/ && \
+    mkdir -p /usr/lib/x86_64-linux-gnu/qemu && \
+    cp -a tmp1/usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so /usr/lib/x86_64-linux-gnu/qemu/ && \
+    rm -rf tmp1 *.deb && \
+    apt-get clean
+
+ENTRYPOINT ["/bin/vitastor-csi"]

csi/Makefile

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v2.3.0
+VITASTOR_VERSION ?= v2.4.3
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v2.3.0
+          image: vitalif/vitastor-csi:v2.4.3
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/007-csi-provisioner.yaml

@@ -121,7 +121,7 @@ spec:
             privileged: true
             capabilities:
               add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v2.3.0
+          image: vitalif/vitastor-csi:v2.4.3
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "2.3.0"
+    vitastorCSIDriverVersion = "2.4.3"
 )
 
 // Config struct fills the parameters of request or user input

csi/src/nodeserver.go

@@ -33,7 +33,7 @@ import (
 type NodeServer struct
 {
     *Driver
-    useVduse        bool
+    method          MountMethod
     stateDir        string
     nfsStageDir     string
     mounter         mount.Interface
@@ -81,16 +81,23 @@ func NewNodeServer(driver *Driver) *NodeServer
     }
     ns := &NodeServer{
         Driver:      driver,
-        useVduse:    checkVduseSupport(),
+        method:      selectMountMethod(),
         stateDir:    stateDir,
         nfsStageDir: nfsStageDir,
         mounter:     mount.New(""),
         volumeLocks: make(map[string]bool),
     }
     ns.cond = sync.NewCond(&ns.mu)
-    if (ns.useVduse)
+    if (ns.method == MOUNT_VDUSE)
     {
         ns.restoreVduseDaemons()
+    }
+    else if (ns.method == MOUNT_UBLK)
+    {
+        ns.restoreUblkDaemons()
+    }
+    if (ns.method == MOUNT_VDUSE || ns.method == MOUNT_UBLK)
+    {
         dur, err := time.ParseDuration(os.Getenv("RESTART_INTERVAL"))
         if (err != nil)
         {
@@ -136,7 +143,14 @@ func (ns *NodeServer) restarter()
     for
     {
         <-ticker.C
-        ns.restoreVduseDaemons()
+        if (ns.method == MOUNT_VDUSE)
+        {
+            ns.restoreVduseDaemons()
+        }
+        else if (ns.method == MOUNT_UBLK)
+        {
+            ns.restoreUblkDaemons()
+        }
     }
 }
 
@@ -231,6 +245,78 @@ func (ns *NodeServer) checkVduseState(stateFile string, devs map[string]interfac
     }
 }
 
+func (ns *NodeServer) restoreUblkDaemons()
+{
+    pattern := ns.stateDir+"vitastor-ublk-*.json"
+    stateFiles, err := filepath.Glob(pattern)
+    if (err != nil)
+    {
+        klog.Errorf("failed to list %s: %v", pattern, err)
+    }
+    if (len(stateFiles) == 0)
+    {
+        return
+    }
+    for _, stateFile := range stateFiles
+    {
+        deviceNum := stateFile[len(ns.stateDir) + len("vitastor-ublk-") :]
+        deviceNum = deviceNum[0:len(deviceNum)-5]
+        ns.checkUblkState(deviceNum)
+    }
+}
+
+func (ns *NodeServer) checkUblkState(deviceNum string)
+{
+    // Check if the ublk daemon is still active
+
+    // Read state file
+    stateFile := ns.stateDir + "vitastor-ublk-" + deviceNum + ".json"
+    stateJSON, err := os.ReadFile(stateFile)
+    if (err != nil)
+    {
+        klog.Warningf("error reading state file %v: %v", stateFile, err)
+        return
+    }
+    var state DeviceState
+    err = json.Unmarshal(stateJSON, &state)
+    if (err != nil)
+    {
+        klog.Warningf("state file %v contains invalid JSON (error %v): %v", stateFile, err, string(stateJSON))
+        return
+    }
+
+    // Lock volume
+    ns.lockVolume(state.ConfigPath+":block:"+state.Image)
+    defer ns.unlockVolume(state.ConfigPath+":block:"+state.Image)
+
+    // Recheck state file after locking
+    _, err = os.ReadFile(stateFile)
+    if (err != nil)
+    {
+        klog.Warningf("state file %v disappeared, skipping volume", stateFile)
+        return
+    }
+
+    // Check if the vitastor-ublk process is still active
+    pidFile := ns.stateDir + "vitastor-ublk-" + deviceNum + ".pid"
+    exists := false
+    proc, err := findByPidFile(pidFile)
+    if (err == nil)
+    {
+        exists = proc.Signal(syscall.Signal(0)) == nil
+    }
+    if (!exists)
+    {
+        // Restart daemon
+        klog.Warningf("recovering UBLK device /dev/ublkb%v for volume %v", deviceNum, state.Image)
+        _, err = mapUblk(ns.stateDir, state.Image, state.ConfigPath, state.Readonly, "/dev/ublkb"+deviceNum)
+        if (err != nil)
+        {
+            klog.Warningf("failed to recover ublk device for volume %v: %v", state.Image, err)
+        }
+    }
+}
+
 func (ns *NodeServer) restoreNfsDaemons()
 {
     pattern := ns.stateDir+"vitastor-nfs-*.json"
@@ -417,14 +503,18 @@ func (ns *NodeServer) NodeStageVolume(ctx context.Context, req *csi.NodeStageVol
     }
 
     var devicePath, vdpaId string
-    if (!ns.useVduse)
+    if (ns.method == MOUNT_UBLK)
     {
-        devicePath, err = mapNbd(volName, ctxVars, false)
+        devicePath, err = mapUblk(ns.stateDir, volName, ctxVars["configPath"], false, "")
     }
-    else
+    else if (ns.method == MOUNT_VDUSE)
     {
         devicePath, vdpaId, err = mapVduse(ns.stateDir, volName, ctxVars, false)
     }
+    else /* if (ns.method == MOUNT_NBD) */
+    {
+        devicePath, err = mapNbd(volName, ctxVars, false)
+    }
     if (err != nil)
     {
         return nil, err
@@ -439,7 +529,8 @@ func (ns *NodeServer) NodeStageVolume(ctx context.Context, req *csi.NodeStageVol
     else
     {
         // Check existing format
-        existingFormat, err := diskMounter.GetDiskFormat(devicePath)
+        var existingFormat string
+        existingFormat, err = diskMounter.GetDiskFormat(devicePath)
         if (err != nil)
         {
             klog.Errorf("failed to get disk format for path %s, error: %v", err)
@@ -495,10 +586,6 @@ func (ns *NodeServer) NodeStageVolume(ctx context.Context, req *csi.NodeStageVol
                 case "xfs":
                     _, err = systemCombined("xfs_growfs", devicePath)
             }
-            if (err != nil)
-            {
-                goto unmap
-            }
         }
     }
     if (err != nil)
@@ -512,14 +599,18 @@ func (ns *NodeServer) NodeStageVolume(ctx context.Context, req *csi.NodeStageVol
     return &csi.NodeStageVolumeResponse{}, nil
 
 unmap:
-    if (!ns.useVduse || len(devicePath) >= 8 && devicePath[0:8] == "/dev/nbd")
+    if (ns.method == MOUNT_UBLK)
     {
-        unmapNbd(devicePath)
+        unmapUblk(ns.stateDir, devicePath)
     }
-    else
+    else if (ns.method == MOUNT_VDUSE)
     {
         unmapVduseById(ns.stateDir, vdpaId)
     }
+    else /* if (ns.method == MOUNT_NBD) */
+    {
+        unmapNbd(devicePath)
+    }
     return nil, err
 }
 
@@ -545,7 +636,7 @@ func (ns *NodeServer) NodeUnstageVolume(ctx context.Context, req *csi.NodeUnstag
     defer ns.unlockVolume(ctxVars["configPath"]+":block:"+volName)
 
     targetPath := req.GetStagingTargetPath()
-    devicePath, _, err := mount.GetDeviceNameFromMount(ns.mounter, targetPath)
+    devicePath, err := GetDeviceNameFromMount(targetPath)
     if (err != nil)
     {
         if (os.IsNotExist(err))
@@ -582,14 +673,18 @@ func (ns *NodeServer) NodeUnstageVolume(ctx context.Context, req *csi.NodeUnstag
     // unmap device
     if (len(refList) == 0)
     {
-        if (!ns.useVduse)
+        if (ns.method == MOUNT_UBLK)
         {
-            unmapNbd(devicePath)
+            unmapUblk(ns.stateDir, devicePath)
         }
-        else
+        else if (ns.method == MOUNT_VDUSE)
         {
             unmapVduse(ns.stateDir, devicePath)
         }
+        else /* if (ns.method == MOUNT_NBD) */
+        {
+            unmapNbd(devicePath)
+        }
     }
 
     return &csi.NodeUnstageVolumeResponse{}, nil
@@ -897,7 +992,7 @@ func (ns *NodeServer) NodeUnpublishVolume(ctx context.Context, req *csi.NodeUnpu
     }
 
     targetPath := req.GetTargetPath()
-    devicePath, _, err := mount.GetDeviceNameFromMount(ns.mounter, targetPath)
+    devicePath, err := GetDeviceNameFromMount(targetPath)
     if (err != nil)
     {
         if (os.IsNotExist(err))

csi/src/utils.go

@@ -16,10 +16,20 @@ import (
     "syscall"
 
     "k8s.io/klog"
+    "k8s.io/utils/mount"
+
     "google.golang.org/grpc/codes"
     "google.golang.org/grpc/status"
 )
 
+type MountMethod int
+
+const (
+    MOUNT_NBD   MountMethod = 0
+    MOUNT_VDUSE MountMethod = 1
+    MOUNT_UBLK  MountMethod = 2
+)
+
 func Contains(list []string, s string) bool
 {
     for i := 0; i < len(list); i++
@@ -32,29 +42,26 @@ func Contains(list []string, s string) bool
     return false
 }
 
-func checkVduseSupport() bool
+func selectMountMethod() MountMethod
 {
+    // Check UBLK support (ublk_drv kernel module)
+    if (checkModule("ublk_drv"))
+    {
+        klog.Infof("UBLK support enabled successfully")
+        return MOUNT_UBLK
+    }
+    klog.Errorf(
+        "Your host apparently has no UBLK support. UBLK support disabled."+
+        " For UBLK you need at least Linux 6.0 and the ublk_drv kernel module.",
+    )
     // Check VDUSE support (vdpa, vduse, virtio-vdpa kernel modules)
     vduse := true
     for _, mod := range []string{"vdpa", "vduse", "virtio-vdpa"}
     {
-        _, err := os.Stat("/sys/module/"+mod)
-        if (err != nil)
+        if (!checkModule(mod))
         {
-            if (!errors.Is(err, os.ErrNotExist))
-            {
-                klog.Errorf("failed to check /sys/module/%s: %v", mod, err)
-            }
-            c := exec.Command("/sbin/modprobe", mod)
-            c.Stdout = os.Stderr
-            c.Stderr = os.Stderr
-            err := c.Run()
-            if (err != nil)
-            {
-                klog.Errorf("/sbin/modprobe %s failed: %v", mod, err)
-                vduse = false
-                break
-            }
+            vduse = false
+            break
         }
     }
     // Check that vdpa tool functions
@@ -69,18 +76,38 @@ func checkVduseSupport() bool
             vduse = false
         }
     }
-    if (!vduse)
-    {
-        klog.Errorf(
-            "Your host apparently has no VDUSE support. VDUSE support disabled, NBD will be used to map devices."+
-            " For VDUSE you need at least Linux 5.15 and the following kernel modules: vdpa, virtio-vdpa, vduse.",
-        )
-    }
-    else
+    if (vduse)
     {
         klog.Infof("VDUSE support enabled successfully")
+        return MOUNT_VDUSE
     }
-    return vduse
+    klog.Errorf(
+        "Your host apparently has no VDUSE support. VDUSE support disabled, NBD will be used to map devices."+
+        " For VDUSE you need at least Linux 5.15 and the following kernel modules: vdpa, virtio-vdpa, vduse.",
+    )
+    return MOUNT_NBD
+}
+
+func checkModule(mod string) bool
+{
+    _, err := os.Stat("/sys/module/"+mod)
+    if (err != nil)
+    {
+        if (!errors.Is(err, os.ErrNotExist))
+        {
+            klog.Errorf("failed to check /sys/module/%s: %v", mod, err)
+        }
+        c := exec.Command("/sbin/modprobe", mod)
+        c.Stdout = os.Stderr
+        c.Stderr = os.Stderr
+        err := c.Run()
+        if (err != nil)
+        {
+            klog.Errorf("/sbin/modprobe %s failed: %v", mod, err)
+            return false
+        }
+    }
+    return true
 }
 
 func mapNbd(volName string, ctxVars map[string]string, readonly bool) (string, error)
@@ -217,6 +244,7 @@ func mapVduse(stateDir string, volName string, ctxVars map[string]string, readon
                     stateJSON, _ := json.Marshal(&DeviceState{
                         ConfigPath: ctxVars["configPath"],
                         VdpaId:     vdpaId,
+
                         Image:      volName,
                         Blockdev:   blockdev,
                         Readonly:   readonly,
@@ -309,6 +337,117 @@ func unmapVduseById(stateDir, vdpaId string)
     }
 }
 
+func mapUblk(stateDir string, volName string, configPath string, readonly bool, recoverDev string) (string, error)
+{
+    pidFile := ""
+    if (recoverDev != "")
+    {
+        if (len(recoverDev) < 10 || recoverDev[0:10] != "/dev/ublkb")
+        {
+            return "", fmt.Errorf("recover: %s does not start with /dev/ublkb", recoverDev)
+        }
+        pidFile = stateDir + "vitastor-ublk-" + recoverDev[10:] + ".pid"
+    }
+    else
+    {
+        pidFd, err := os.CreateTemp(stateDir, "vitastor-tmp-*.pid")
+        if (err != nil)
+        {
+            return "", err
+        }
+        pidFile = pidFd.Name()
+        pidFd.Close()
+    }
+    // Map device via vitastor-ublk
+    args := []string{
+        "map", "--image", volName, "--pidfile", pidFile,
+    }
+    if (configPath != "")
+    {
+        args = append(args, "--config_path", configPath)
+    }
+    if (readonly)
+    {
+        args = append(args, "--readonly")
+    }
+    if (recoverDev != "")
+    {
+        args = append(args, "--recover", recoverDev)
+    }
+    stdout, stderr, err := system("/usr/bin/vitastor-ublk", args...)
+    if (err != nil)
+    {
+        return "", err
+    }
+    devicePath := strings.TrimSpace(string(stdout))
+    if (devicePath == "")
+    {
+        return "", fmt.Errorf("vitastor-ublk did not return the name of the device. output: %s", stderr)
+    }
+    if (len(devicePath) >= 10 && devicePath[0:10] == "/dev/ublkb")
+    {
+        // Generate state file
+        devNum := devicePath[10:]
+        pidNew := stateDir + "vitastor-ublk-" + devNum + ".pid"
+        if (pidFile != pidNew)
+        {
+            err := os.Rename(pidFile, pidNew)
+            if (err != nil)
+            {
+                klog.Errorf("Failed to rename PID file %s to %s: %v", pidFile, pidNew, err)
+            }
+            else
+            {
+                pidFile = pidNew
+            }
+        }
+        stateFile := stateDir + "vitastor-ublk-" + devNum + ".json"
+        stateJSON, _ := json.Marshal(&DeviceState{
+            ConfigPath: configPath,
+            Image:      volName,
+            Readonly:   readonly,
+            PidFile:    pidFile,
+        })
+        err = os.WriteFile(stateFile, stateJSON, 0600)
+        if (err == nil)
+        {
+            klog.Infof("Attached volume %s via UBLK as %s", volName, devicePath)
+            return devicePath, nil
+        }
+        os.Remove(stateFile)
+    }
+    killErr := killByPidFile(pidFile)
+    if (killErr != nil)
+    {
+        klog.Errorf("Failed to kill started vitastor-ublk: %v", killErr)
+    }
+    os.Remove(pidFile)
+    return "", err
+}
+
+func unmapUblk(stateDir, devicePath string)
+{
+    if (len(devicePath) < 10 || devicePath[0:10] != "/dev/ublkb")
+    {
+        klog.Errorf("%s does not start with /dev/ublkb", devicePath)
+        return
+    }
+    unmapOut, unmapErr := exec.Command("/usr/bin/vitastor-ublk", "unmap", devicePath).CombinedOutput()
+    if (unmapErr != nil)
+    {
+        klog.Errorf("failed to unmap UBLK device %s: %s, error: %v", devicePath, unmapOut, unmapErr)
+    }
+    for _, ext := range []string{"json", "pid"}
+    {
+        fn := stateDir + "vitastor-ublk-" + devicePath[10:] + "." + ext
+        err := os.Remove(fn)
+        if (err != nil)
+        {
+            klog.Errorf("failed to remove %s: %v", fn, err)
+        }
+    }
+}
+
 func system(program string, args ...string) ([]byte, []byte, error)
 {
     klog.Infof("Running "+program+" "+strings.Join(args, " "))
@@ -340,3 +479,43 @@ func systemCombined(program string, args ...string) ([]byte, error)
     }
     return out.Bytes(), nil
 }
+
+func GetDeviceNameFromMount(mountPath string) (string, error)
+{
+    // Use /proc/self/mountinfo to correctly parse bind mounts for block device files
+    mps, err := mount.ParseMountInfo("/proc/self/mountinfo")
+    if (err != nil)
+    {
+        return "", err
+    }
+
+    slTarget, err := filepath.EvalSymlinks(mountPath)
+    if (err != nil)
+    {
+        slTarget = mountPath
+    }
+
+    device := ""
+    for _, mp := range mps
+    {
+        if (mp.MountPoint == slTarget)
+        {
+            device = mp.Source
+            if (device[0] != '/' && mp.Root != "/")
+            {
+                // Handle {Source=udev Root=/vdb MountPoint=/var/lib/kubelet/tralaleylo/tralala}
+                for _, other := range mps
+                {
+                    if (other.Root == "/" && other.Source == mp.Source)
+                    {
+                        device = other.MountPoint + mp.Root
+                        break
+                    }
+                }
+            }
+            break
+        }
+    }
+
+    return device, nil
+}

debian/build-vitastor-bookworm.sh

@@ -1,4 +1,4 @@
 #!/bin/bash
 
 docker build --build-arg DISTRO=debian --build-arg REL=bookworm -t vitastor-buildenv:bookworm -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=bookworm -v `dirname $0`/../:/root/vitastor vitastor-buildenv:bookworm /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=bookworm -v `dirname $0`/../:/root/vitastor vitastor-buildenv:bookworm /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-bullseye.sh

@@ -1,4 +1,4 @@
 #!/bin/bash
 
 docker build --build-arg DISTRO=debian --build-arg REL=bullseye -t vitastor-buildenv:bullseye -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=bullseye -v `dirname $0`/../:/root/vitastor vitastor-buildenv:bullseye /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=bullseye -v `dirname $0`/../:/root/vitastor vitastor-buildenv:bullseye /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-buster.sh

@@ -1,4 +1,4 @@
 #!/bin/bash
 
 docker build --build-arg DISTRO=debian --build-arg REL=buster -t vitastor-buildenv:buster -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=buster -v `dirname $0`/../:/root/vitastor vitastor-buildenv:buster /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=buster -v `dirname $0`/../:/root/vitastor vitastor-buildenv:buster /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-trixie.sh

@@ -1,4 +1,4 @@
 #!/bin/bash
 
 docker build --build-arg DISTRO=debian --build-arg REL=trixie -t vitastor-buildenv:trixie -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=trixie -v `dirname $0`/../:/root/vitastor vitastor-buildenv:trixie /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=trixie -v `dirname $0`/../:/root/vitastor vitastor-buildenv:trixie /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-ubuntu-jammy.sh

@@ -2,4 +2,4 @@
 # Ubuntu 22.04 Jammy Jellyfish
 
 docker build --build-arg DISTRO=ubuntu --build-arg REL=jammy -t vitastor-buildenv:jammy -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=jammy -v `dirname $0`/../:/root/vitastor vitastor-buildenv:jammy /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=jammy -v `dirname $0`/../:/root/vitastor vitastor-buildenv:jammy /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-ubuntu-noble.sh

@@ -2,4 +2,4 @@
 # 24.04 Noble Numbat
 
 docker build --build-arg DISTRO=ubuntu --build-arg REL=noble -t vitastor-buildenv:noble -f vitastor-buildenv.Dockerfile .
-docker run -i --rm -e REL=noble -v `dirname $0`/../:/root/vitastor vitastor-buildenv:noble /root/vitastor/debian/vitastor-build.sh
+docker run -it --rm -e REL=noble -v `dirname $0`/../:/root/vitastor vitastor-buildenv:noble /root/vitastor/debian/vitastor-build.sh

debian/build-vitastor-ubuntu-questing.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+# 25.10 Questing quokka
+
+docker build --build-arg DISTRO=ubuntu --build-arg REL=questing -t vitastor-buildenv:questing -f vitastor-buildenv.Dockerfile .
+docker run -it --rm -e REL=questing -v `dirname $0`/../:/root/vitastor vitastor-buildenv:questing /root/vitastor/debian/vitastor-build.sh

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (2.3.0-1) unstable; urgency=medium
+vitastor (2.4.3-1) unstable; urgency=medium
 
   * Bugfixes
 

debian/control

@@ -4,7 +4,7 @@ Priority: optional
 Maintainer: Vitaliy Filippov <vitalif@yourcmc.ru>
 Build-Depends: debhelper, g++ (>= 8), libstdc++6 (>= 8),
   linux-libc-dev, libgoogle-perftools-dev, libjerasure-dev, libgf-complete-dev,
-  libibverbs-dev, libisal-dev, cmake, pkg-config, libnl-3-dev, libnl-genl-3-dev,
+  libibverbs-dev, librdmacm-dev, libisal-dev, cmake, pkg-config, libnl-3-dev, libnl-genl-3-dev,
   node-bindings <!nocheck>, node-gyp, node-nan
 Standards-Version: 4.5.0
 Homepage: https://vitastor.io/

debian/vitastor-build.sh

@@ -44,7 +44,12 @@ curl -s https://git.yourcmc.ru/vitalif/antietcd/archive/master.tar.gz | tar -zx
 curl -s https://git.yourcmc.ru/vitalif/tinyraft/archive/master.tar.gz | tar -zx
 
 cd /root/vitastor/packages/vitastor-$REL
-tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_$VER.orig.tar.xz vitastor-$VER
+if [[ "$REL" = "trixie" && -e ../vitastor-bookworm/vitastor_$VER.orig.tar.xz ]]; then
+    # Fucking shit, archives differ between bookworm (xz 5.4.1) and trixie (xz 5.8.1)
+    cp ../vitastor-bookworm/vitastor_$VER.orig.tar.xz .
+else
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_$VER.orig.tar.xz vitastor-$VER
+fi
 cd vitastor-$VER
 DEBEMAIL="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v "$FULLVER""$REL" "Rebuild for $REL"
 DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa

docker/Dockerfile

@@ -3,7 +3,7 @@
 FROM debian:bookworm
 
 ADD etc/apt /etc/apt/
-RUN apt-get update && apt-get -y install vitastor udev systemd qemu-system-x86 qemu-system-common qemu-block-extra qemu-utils jq nfs-common && apt-get clean
+RUN apt-get update && apt-get -y install vitastor ibverbs-providers udev systemd qemu-system-x86 qemu-system-common qemu-block-extra qemu-utils jq nfs-common && apt-get clean
 ADD sleep.sh /usr/bin/
 ADD install.sh /usr/bin/
 ADD scripts /opt/scripts/

docker/Makefile

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v2.3.0
+VITASTOR_VERSION ?= v2.4.3
 
 all: build push
 

docker/etc/vitastor/docker.conf

@@ -4,7 +4,7 @@
 #
 
 # Desired Vitastor version
-VITASTOR_VERSION=v2.3.0
+VITASTOR_VERSION=v2.4.3
 
 # Additional arguments for all containers
 # For example, you may want to specify a custom logging driver here

docs/config/osd.en.md

@@ -491,7 +491,7 @@ Can be used to slow down scrubbing if it affects user load too much.
 ## scrub_list_limit
 
 - Type: integer
-- Default: 1000
+- Default: 262144
 - Can be changed online: yes
 
 Number of objects to list in one listing operation during scrub.

docs/config/osd.ru.md

@@ -514,7 +514,7 @@ fsync небезопасным даже с режимом "directsync".
 ## scrub_list_limit
 
 - Тип: целое число
-- Значение по умолчанию: 1000
+- Значение по умолчанию: 262144
 - Можно менять на лету: да
 
 Размер загружаемых за одну операцию списков объектов в процессе фоновой

docs/config/src/osd.yml

@@ -566,7 +566,7 @@
     сильно влияет на пользовательскую нагрузку.
 - name: scrub_list_limit
   type: int
-  default: 1000
+  default: 262144
   online: true
   info: |
     Number of objects to list in one listing operation during scrub.

docs/installation/docker.en.md

@@ -26,9 +26,9 @@ at Vitastor Kubernetes operator: https://github.com/Antilles7227/vitastor-operat
 The instruction is very simple.
 
 1. Download a Docker image of the desired version: \
-   `docker pull vitalif/vitastor:v2.3.0`
+   `docker pull vitalif/vitastor:v2.4.3`
 2. Install scripts to the host system: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v2.3.0 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v2.4.3 install.sh`
 3. Reload udev rules: \
    `udevadm control --reload-rules`
 

docs/installation/docker.ru.md

@@ -25,9 +25,9 @@ Vitastor можно установить в Docker/Podman. При этом etcd,
 Инструкция по установке максимально простая.
 
 1. Скачайте Docker-образ желаемой версии: \
-   `docker pull vitalif/vitastor:v2.3.0`
+   `docker pull vitalif/vitastor:v2.4.3`
 2. Установите скрипты в хост-систему командой: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v2.3.0 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v2.4.3 install.sh`
 3. Перезагрузите правила udev: \
    `udevadm control --reload-rules`
 

docs/installation/openstack.en.md

@@ -30,7 +30,7 @@ volume_backend_name = vitastor-testcluster
 image_volume_cache_enabled = True
 volume_clear = none
 vitastor_etcd_address = 192.168.7.2:2379
-vitastor_etcd_prefix =
+vitastor_etcd_prefix = /vitastor
 vitastor_config_path = /etc/vitastor/vitastor.conf
 vitastor_pool_id = 1
 image_upload_use_cinder_backend = True

docs/installation/openstack.ru.md

@@ -29,7 +29,7 @@ volume_backend_name = vitastor-testcluster
 image_volume_cache_enabled = True
 volume_clear = none
 vitastor_etcd_address = 192.168.7.2:2379
-vitastor_etcd_prefix =
+vitastor_etcd_prefix = /vitastor
 vitastor_config_path = /etc/vitastor/vitastor.conf
 vitastor_pool_id = 1
 image_upload_use_cinder_backend = True

docs/usage/cli.en.md

@@ -100,12 +100,14 @@ List images (only matching `<glob>` pattern(s) if passed).
 Options:
 
 ```
+--exact         Do not match glob patterns as names, select only exact name matches.
 -p|--pool POOL  Filter images by pool ID or name
 -l|--long       Also report allocated size and I/O statistics
 --del           Also include delete operation statistics
 --sort FIELD    Sort by specified field (name, size, used_size, <read|write|delete>_<iops|bps|lat|queue>)
 -r|--reverse    Sort in descending order
 -n|--count N    Only list first N items
+--tree          Show image snapshot/clone tree
 ```
 
 Example output:

docs/usage/cli.ru.md

@@ -102,12 +102,14 @@ kaveri    2/1     32   0 B      10 G    0 B        100%    0%
 Опции:
 
 ```
+--exact         Не применять ФС-шаблоны к именам, выводить только точные совпадения
 -p|--pool POOL  Фильтровать образы по пулу (ID или имени)
 -l|--long       Также выводить статистику занятого места и ввода-вывода
 --del           Также выводить статистику операций удаления
 --sort FIELD    Сортировать по заданному полю (name, size, used_size, <read|write|delete>_<iops|bps|lat|queue>)
 -r|--reverse    Сортировать в обратном порядке
 -n|--count N    Показывать только первые N записей
+--tree          Вывести снапшоты и клоны в виде дерева
 ```
 
 Пример вывода:

docs/usage/disk.en.md

@@ -73,6 +73,8 @@ Options (automatic mode):
 --max_other 10%
   Use disks for OSD data even if they already have non-Vitastor partitions,
   but only if these take up no more than this percent of disk space.
+--dry-run
+  Check and print new OSD count for each disk but do not actually create them.
 ```
 
 Options (single-device mode):

docs/usage/disk.ru.md

@@ -74,6 +74,8 @@ vitastor-disk - инструмент командной строки для уп
 --max_other 10%
   Использовать диски под данные OSD, даже если на них уже есть не-Vitastor-овые
   разделы, но только в случае, если они занимают не более данного процента диска.
+--dry-run
+  Проверить и вывести число новых OSD для каждого диска, но не создавать их.
 ```
 
 Опции для режима одного OSD:

mon/osd_tree.js

@@ -15,7 +15,7 @@ function get_osd_tree(global_config, state)
         const stat = state.osd.stats[osd_num];
         const osd_cfg = state.config.osd[osd_num];
         let reweight = osd_cfg == null ? 1 : Number(osd_cfg.reweight);
-        if (isNaN(reweight) || reweight < 0 || reweight > 0)
+        if (isNaN(reweight) || reweight < 0 || reweight > 1)
             reweight = 1;
         if (stat && stat.size && reweight && (state.osd.state[osd_num] || Number(stat.time) >= down_time ||
             osd_cfg && osd_cfg.noout))

mon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor-mon",
-  "version": "2.3.0",
+  "version": "2.4.3",
   "description": "Vitastor SDS monitor service",
   "main": "mon-main.js",
   "scripts": {

mon/scripts/make-etcd

@@ -52,15 +52,16 @@ async function run()
         process.exit(1);
     }
     const etcds = (config.etcd_address instanceof Array ? config.etcd_address : (''+config.etcd_address).split(/,/))
-        .map(s => (''+s).replace(/^https?:\/\/\[?|\]?(:\d+)?(\/.*)?$/g, '').toLowerCase());
+        .map(s => (''+s).replace(/^https?:\/\/|(:\d+)?(\/.*)?$/g, '').replace(/^\[(.*)\]$/, '$1').toLowerCase());
     const num = select_local_etcd(etcds);
     if (num < 0)
     {
         console.log('No matching IPs in etcd_address from '+config_path);
         process.exit(0);
     }
+    const etcd_url = 'http://' + (etcds[num].indexOf(':') >= 0 ? '['+etcds[num]+']' : etcds[num]);
     const etcd_name = 'etcd'+etcds[num].replace(/[^0-9a-z_]/ig, '_');
-    const etcd_cluster = etcds.map(e => `etcd${e.replace(/[^0-9a-z_]/ig, '_')}=http://${e}:2380`).join(',');
+    const etcd_cluster = etcds.map(e => `etcd${e.replace(/[^0-9a-z_]/ig, '_')}=http://${e.indexOf(':') >= 0 ? '['+e+']' : e}:2380`).join(',');
     if (in_docker)
     {
         let etcd_conf = fs.readFileSync("/etc/vitastor/etcd.conf", { encoding: 'utf-8' });
@@ -83,8 +84,8 @@ Wants=network-online.target local-fs.target time-sync.target
 Restart=always
 Environment=GOGC=50
 ExecStart=etcd --name ${etcd_name} --data-dir /var/lib/etcd/vitastor \\
-    --snapshot-count 10000 --advertise-client-urls http://${etcds[num]}:2379 --listen-client-urls http://${etcds[num]}:2379 \\
-    --initial-advertise-peer-urls http://${etcds[num]}:2380 --listen-peer-urls http://${etcds[num]}:2380 \\
+    --snapshot-count 10000 --advertise-client-urls ${etcd_url}:2379 --listen-client-urls ${etcd_url}:2379 \\
+    --initial-advertise-peer-urls ${etcd_url}:2380 --listen-peer-urls ${etcd_url}:2380 \\
     --initial-cluster-token vitastor-etcd-1 --initial-cluster ${etcd_cluster} \\
     --initial-cluster-state new --max-txn-ops=100000 --max-request-bytes=104857600 \\
     --auto-compaction-retention=10 --auto-compaction-mode=revision

mon/stats.js

@@ -276,6 +276,10 @@ function sum_inode_stats(state, prev_stats)
         }
         for (const pool_id in osd_diff.inode_stats)
         {
+            if (!inode_stats[pool_id])
+            {
+                continue;
+            }
             for (const inode_num in prev_stats.osd_diff[osd].inode_stats[pool_id])
             {
                 inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();

node-binding/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor",
-  "version": "2.3.0",
+  "version": "2.4.3",
   "description": "Low-level native bindings to Vitastor client library",
   "main": "index.js",
   "keywords": [

patches/VitastorPlugin.pm

@@ -499,4 +499,55 @@ sub rename_volume
     return "${storeid}:${base_name}${target_volname}";
 }
 
+sub _monkey_patch_qemu_blockdev_options
+{
+    my ($cfg, $volid, $machine_version, $options) = @_;
+    my ($storeid, $volname) = PVE::Storage::parse_volume_id($volid);
+
+    my $scfg = PVE::Storage::storage_config($cfg, $storeid);
+
+    my $plugin = PVE::Storage::Plugin->lookup($scfg->{type});
+
+    my ($vtype) = $plugin->parse_volname($volname);
+    die "cannot use volume of type '$vtype' as a QEMU blockdevice\n"
+        if $vtype ne 'images' && $vtype ne 'iso' && $vtype ne 'import';
+
+    return $plugin->qemu_blockdev_options($scfg, $storeid, $volname, $machine_version, $options);
+}
+
+sub qemu_blockdev_options
+{
+    my ($class, $scfg, $storeid, $volname, $machine_version, $options) = @_;
+    my $prefix = defined $scfg->{vitastor_prefix} ? $scfg->{vitastor_prefix} : 'pve/';
+    my ($vtype, $name, $vmid) = $class->parse_volname($volname);
+    $name .= '@'.$options->{'snapshot-name'} if $options->{'snapshot-name'};
+    if ($scfg->{vitastor_nbd})
+    {
+        my $mapped = run_cli($scfg, [ 'ls' ], binary => '/usr/bin/vitastor-nbd');
+        my ($kerneldev) = grep { $mapped->{$_}->{image} eq $prefix.$name } keys %$mapped;
+        die "Image not mapped via NBD" if !$kerneldev;
+        return { driver => 'host_device', filename => $kerneldev };
+    }
+    my $blockdev = {
+        driver => 'vitastor',
+        image => $prefix.$name,
+    };
+    if ($scfg->{vitastor_config_path})
+    {
+        $blockdev->{'config-path'} = $scfg->{vitastor_config_path};
+    }
+    if ($scfg->{vitastor_etcd_address})
+    {
+        # FIXME This is the only exception: etcd_address -> etcd_host for qemu
+        $blockdev->{'etcd-host'} = $scfg->{vitastor_etcd_address};
+    }
+    if ($scfg->{vitastor_etcd_prefix})
+    {
+        $blockdev->{'etcd-prefix'} = $scfg->{vitastor_etcd_prefix};
+    }
+    return $blockdev;
+}
+
+*PVE::Storage::qemu_blockdev_options = *_monkey_patch_qemu_blockdev_options;
+
 1;

patches/cinder-vitastor.py

@@ -50,7 +50,7 @@ from cinder.volume import configuration
 from cinder.volume import driver
 from cinder.volume import volume_utils
 
-VITASTOR_VERSION = '2.3.0'
+VITASTOR_VERSION = '2.4.3'
 
 LOG = logging.getLogger(__name__)
 

patches/pve-qemu-10.1-vitastor.patch

@@ -0,0 +1,172 @@
+Index: pve-qemu-kvm-10.1.2/block/meson.build
+===================================================================
+--- pve-qemu-kvm-10.1.2.orig/block/meson.build
++++ pve-qemu-kvm-10.1.2/block/meson.build
+@@ -126,6 +126,7 @@ foreach m : [
+   [libnfs, 'nfs', files('nfs.c')],
+   [libssh, 'ssh', files('ssh.c')],
+   [rbd, 'rbd', files('rbd.c')],
++  [vitastor, 'vitastor', files('vitastor.c')],
+ ]
+   if m[0].found()
+     module_ss = ss.source_set()
+Index: pve-qemu-kvm-10.1.2/meson.build
+===================================================================
+--- pve-qemu-kvm-10.1.2.orig/meson.build
++++ pve-qemu-kvm-10.1.2/meson.build
+@@ -1653,6 +1653,26 @@ if not get_option('rbd').auto() or have_
+   endif
+ endif
+ 
++vitastor = not_found
++if not get_option('vitastor').auto() or have_block
++  libvitastor_client = cc.find_library('vitastor_client', has_headers: ['vitastor_c.h'],
++    required: get_option('vitastor'))
++  if libvitastor_client.found()
++    if cc.links('''
++      #include <vitastor_c.h>
++      int main(void) {
++        vitastor_c_create_qemu(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
++        return 0;
++      }''', dependencies: libvitastor_client)
++      vitastor = declare_dependency(dependencies: libvitastor_client)
++    elif get_option('vitastor').enabled()
++      error('could not link libvitastor_client')
++    else
++      warning('could not link libvitastor_client, disabling')
++    endif
++  endif
++endif
++
+ glusterfs = not_found
+ glusterfs_ftruncate_has_stat = false
+ glusterfs_iocb_has_stat = false
+@@ -2552,6 +2572,7 @@ endif
+ config_host_data.set('CONFIG_OPENGL', opengl.found())
+ config_host_data.set('CONFIG_PLUGIN', get_option('plugins'))
+ config_host_data.set('CONFIG_RBD', rbd.found())
++config_host_data.set('CONFIG_VITASTOR', vitastor.found())
+ config_host_data.set('CONFIG_RDMA', rdma.found())
+ config_host_data.set('CONFIG_RELOCATABLE', get_option('relocatable'))
+ config_host_data.set('CONFIG_SAFESTACK', get_option('safe_stack'))
+@@ -4984,6 +5005,7 @@ summary_info += {'fdt support':       fd
+ summary_info += {'libcap-ng support': libcap_ng}
+ summary_info += {'bpf support':       libbpf}
+ summary_info += {'rbd support':       rbd}
++summary_info += {'vitastor support':  vitastor}
+ summary_info += {'smartcard support': cacard}
+ summary_info += {'U2F support':       u2f}
+ summary_info += {'libusb':            libusb}
+Index: pve-qemu-kvm-10.1.2/meson_options.txt
+===================================================================
+--- pve-qemu-kvm-10.1.2.orig/meson_options.txt
++++ pve-qemu-kvm-10.1.2/meson_options.txt
+@@ -202,6 +202,8 @@ option('pvg', type: 'feature', value: 'a
+        description: 'macOS paravirtualized graphics support')
+ option('rbd', type : 'feature', value : 'auto',
+        description: 'Ceph block device driver')
++option('vitastor', type : 'feature', value : 'auto',
++       description: 'Vitastor block device driver')
+ option('opengl', type : 'feature', value : 'auto',
+        description: 'OpenGL support')
+ option('rdma', type : 'feature', value : 'auto',
+Index: pve-qemu-kvm-10.1.2/qapi/block-core.json
+===================================================================
+--- pve-qemu-kvm-10.1.2.orig/qapi/block-core.json
++++ pve-qemu-kvm-10.1.2/qapi/block-core.json
+@@ -3647,7 +3647,7 @@
+             'raw', 'rbd',
+             { 'name': 'replication', 'if': 'CONFIG_REPLICATION' },
+             'pbs',
+-            'ssh', 'throttle', 'vdi', 'vhdx',
++            'ssh', 'throttle', 'vdi', 'vhdx', 'vitastor',
+             { 'name': 'virtio-blk-vfio-pci', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-user', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-vdpa', 'if': 'CONFIG_BLKIO' },
+@@ -4773,6 +4773,28 @@
+             '*server': ['InetSocketAddressBase'] } }
+ 
+ ##
++# @BlockdevOptionsVitastor:
++#
++# Driver specific block device options for vitastor
++#
++# @image:       Image name
++# @inode:       Inode number
++# @pool:        Pool ID
++# @size:        Desired image size in bytes
++# @config-path: Path to Vitastor configuration
++# @etcd-host:   etcd connection address(es)
++# @etcd-prefix: etcd key/value prefix
++##
++{ 'struct': 'BlockdevOptionsVitastor',
++  'data': { '*inode': 'uint64',
++            '*pool': 'uint64',
++            '*size': 'uint64',
++            '*image': 'str',
++            '*config-path': 'str',
++            '*etcd-host': 'str',
++            '*etcd-prefix': 'str' } }
++
++##
+ # @ReplicationMode:
+ #
+ # An enumeration of replication modes.
+@@ -5242,6 +5264,7 @@
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+       'vhdx':       'BlockdevOptionsGenericFormat',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'virtio-blk-vfio-pci':
+                     { 'type': 'BlockdevOptionsVirtioBlkVfioPci',
+                       'if': 'CONFIG_BLKIO' },
+@@ -5722,6 +5745,20 @@
+             '*encrypt' :        'RbdEncryptionCreateOptions' } }
+ 
+ ##
++# @BlockdevCreateOptionsVitastor:
++#
++# Driver specific image creation options for Vitastor.
++#
++# @location: Where to store the new image file.  This location cannot
++#     point to a snapshot.
++#
++# @size: Size of the virtual disk in bytes
++##
++{ 'struct': 'BlockdevCreateOptionsVitastor',
++  'data': { 'location':         'BlockdevOptionsVitastor',
++            'size':             'size' } }
++
++##
+ # @BlockdevVmdkSubformat:
+ #
+ # Subformat options for VMDK images
+@@ -5943,6 +5980,7 @@
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'vdi':            'BlockdevCreateOptionsVdi',
+       'vhdx':           'BlockdevCreateOptionsVhdx',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'vmdk':           'BlockdevCreateOptionsVmdk',
+       'vpc':            'BlockdevCreateOptionsVpc'
+   } }
+Index: pve-qemu-kvm-10.1.2/scripts/meson-buildoptions.sh
+===================================================================
+--- pve-qemu-kvm-10.1.2.orig/scripts/meson-buildoptions.sh
++++ pve-qemu-kvm-10.1.2/scripts/meson-buildoptions.sh
+@@ -175,6 +175,7 @@ meson_options_help() {
+   printf "%s\n" '  qga-vss         build QGA VSS support (broken with MinGW)'
+   printf "%s\n" '  qpl             Query Processing Library support'
+   printf "%s\n" '  rbd             Ceph block device driver'
++  printf "%s\n" '  vitastor        Vitastor block device driver'
+   printf "%s\n" '  rdma            Enable RDMA-based migration'
+   printf "%s\n" '  replication     replication support'
+   printf "%s\n" '  rust            Rust support'
+@@ -459,6 +460,8 @@ _meson_option_parse() {
+     --disable-qpl) printf "%s" -Dqpl=disabled ;;
+     --enable-rbd) printf "%s" -Drbd=enabled ;;
+     --disable-rbd) printf "%s" -Drbd=disabled ;;
++    --enable-vitastor) printf "%s" -Dvitastor=enabled ;;
++    --disable-vitastor) printf "%s" -Dvitastor=disabled ;;
+     --enable-rdma) printf "%s" -Drdma=enabled ;;
+     --disable-rdma) printf "%s" -Drdma=disabled ;;
+     --enable-relocatable) printf "%s" -Drelocatable=true ;;

patches/qemu-10.1-vitastor.patch

@@ -0,0 +1,29 @@
+diff --git a/src/client/qemu_driver.c b/src/client/qemu_driver.c
+index d8356dab..5f4cd50d 100644
+--- a/src/client/qemu_driver.c
++++ b/src/client/qemu_driver.c
+@@ -974,14 +974,21 @@ static void vitastor_co_read_bitmap_cb(void *opaque, long retval, uint8_t *bitma
+ #endif
+ }
+ 
+-static int coroutine_fn vitastor_co_block_status(
+-    BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
+-    int64_t *pnum, int64_t *map, BlockDriverState **file)
++static int coroutine_fn vitastor_co_block_status(BlockDriverState *bs,
++#if QEMU_VERSION_MAJOR > 10 || QEMU_VERSION_MAJOR == 10 && QEMU_VERSION_MINOR >= 1
++    unsigned int mode,
++#else
++    bool want_zero,
++#endif
++    int64_t offset, int64_t bytes, int64_t *pnum, int64_t *map, BlockDriverState **file)
+ {
+     // Allocated => return BDRV_BLOCK_DATA|BDRV_BLOCK_OFFSET_VALID
+     // Not allocated => return 0
+     // Error => return -errno
+     // Set pnum to length of the extent, `*map` = `offset`, `*file` = `bs`
++#if QEMU_VERSION_MAJOR > 10 || QEMU_VERSION_MAJOR == 10 && QEMU_VERSION_MINOR >= 1
++    int want_zero = (mode == BDRV_WANT_PRECISE);
++#endif
+     VitastorRPC task;
+     VitastorClient *client = bs->opaque;
+     uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        2.3.0
+Version:        2.4.3
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-2.3.0.el7.tar.gz
+Source0:        vitastor-2.4.3.el7.tar.gz
 
 BuildRequires:  gperftools-devel
 BuildRequires:  devtoolset-9-gcc-c++

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        2.3.0
+Version:        2.4.3
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-2.3.0.el8.tar.gz
+Source0:        vitastor-2.4.3.el8.tar.gz
 
 BuildRequires:  gperftools-devel
 BuildRequires:  gcc-toolset-9-gcc-c++

rpm/vitastor-el9.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        2.3.0
+Version:        2.4.3
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-2.3.0.el9.tar.gz
+Source0:        vitastor-2.4.3.el9.tar.gz
 
 BuildRequires:  gperftools-devel
 BuildRequires:  gcc-c++

src/CMakeLists.txt

@@ -19,8 +19,9 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	endif()
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
+set(ENABLE_COVERAGE false CACHE BOOL "Enable code coverage")
 
-add_definitions(-DVITASTOR_VERSION="2.3.0")
+add_definitions(-DVITASTOR_VERSION="2.4.3")
 add_definitions(-D_GNU_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64 -Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -fno-omit-frame-pointer -fvisibility=hidden -I ${CMAKE_SOURCE_DIR}/src)
 add_link_options(-fno-omit-frame-pointer)
 if (${WITH_ASAN})
@@ -31,6 +32,11 @@ set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fvisibility-inlines-hid
 set(CMAKE_CXX_FLAGS_MINSIZEREL "${CMAKE_CXX_FLAGS_MINSIZEREL} -fvisibility-inlines-hidden")
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -fvisibility-inlines-hidden")
 
+if (${ENABLE_COVERAGE})
+	add_definitions(-coverage)
+	add_link_options(-coverage)
+endif()
+
 set(CMAKE_BUILD_TYPE RelWithDebInfo)
 string(REGEX REPLACE "([\\/\\-]O)[^ \t\r\n]*" "\\13" CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
 string(REGEX REPLACE "([\\/\\-]O)[^ \t\r\n]*" "\\13" CMAKE_CXX_FLAGS_MINSIZEREL "${CMAKE_CXX_FLAGS_MINSIZEREL}")
@@ -78,14 +84,6 @@ else()
 	set(LIBURING_LIBRARIES uring)
 endif (${WITH_SYSTEM_LIBURING})
 
-add_custom_target(build_tests)
-add_custom_target(test
-	COMMAND
-	echo leak:tcmalloc > ${CMAKE_CURRENT_BINARY_DIR}/lsan-suppress.txt &&
-	env LSAN_OPTIONS=suppressions=${CMAKE_CURRENT_BINARY_DIR}/lsan-suppress.txt ${CMAKE_CTEST_COMMAND}
-)
-add_dependencies(test build_tests)
-
 include_directories(
 	../
 	${CMAKE_SOURCE_DIR}/src/blockstore
@@ -117,7 +115,7 @@ install_symlink(vitastor-disk ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_BINDIR}/vi
 install_symlink(vitastor-cli ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_BINDIR}/vitastor-rm)
 install_symlink(vitastor-cli ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_BINDIR}/vita)
 install(
-	TARGETS vitastor_blk vitastor_client vitastor_kv
+	TARGETS vitastor_client vitastor_kv
 	LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
 	PUBLIC_HEADER DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
 )

src/blockstore/CMakeLists.txt

@@ -2,15 +2,18 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-# libvitastor_blk.so
-add_library(vitastor_blk SHARED
-	../util/allocator.cpp blockstore.cpp blockstore_impl.cpp blockstore_disk.cpp blockstore_init.cpp blockstore_open.cpp blockstore_journal.cpp blockstore_read.cpp
-	blockstore_write.cpp blockstore_sync.cpp blockstore_stable.cpp blockstore_rollback.cpp blockstore_flush.cpp ../util/crc32c.c ../util/ringloop.cpp
+# libvitastor_blk.a
+add_library(vitastor_blk STATIC
+	../util/allocator.cpp ../util/crc32c.c ../util/ringloop.cpp
+	multilist.cpp blockstore_heap.cpp blockstore_disk.cpp
+	blockstore.cpp blockstore_impl.cpp blockstore_init.cpp blockstore_open.cpp
+	blockstore_flush.cpp blockstore_read.cpp blockstore_stable.cpp blockstore_sync.cpp blockstore_write.cpp
+	v1/flush.cpp v1/impl.cpp v1/init.cpp v1/journal.cpp v1/open.cpp v1/read.cpp v1/rollback.cpp v1/stable.cpp v1/sync.cpp v1/write.cpp
 )
+target_compile_options(vitastor_blk PUBLIC -fPIC)
 target_link_libraries(vitastor_blk
 	${LIBURING_LIBRARIES}
 	${ISAL_LIBRARIES}
-	tcmalloc_minimal
 	# for timerfd_manager
 	vitastor_common
 )

src/blockstore/blockstore.cpp

@@ -1,89 +1,16 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
+#include "str_util.h"
+
 #include "blockstore_impl.h"
+#include "v1/impl.h"
 
-blockstore_t::blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd)
+blockstore_i* blockstore_i::create(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd)
 {
-    impl = new blockstore_impl_t(config, ringloop, tfd);
-}
-
-blockstore_t::~blockstore_t()
-{
-    delete impl;
-}
-
-void blockstore_t::parse_config(blockstore_config_t & config)
-{
-    impl->parse_config(config, false);
-}
-
-void blockstore_t::loop()
-{
-    impl->loop();
-}
-
-bool blockstore_t::is_started()
-{
-    return impl->is_started();
-}
-
-bool blockstore_t::is_stalled()
-{
-    return impl->is_stalled();
-}
-
-bool blockstore_t::is_safe_to_stop()
-{
-    return impl->is_safe_to_stop();
-}
-
-void blockstore_t::enqueue_op(blockstore_op_t *op)
-{
-    impl->enqueue_op(op);
-}
-
-int blockstore_t::read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version)
-{
-    return impl->read_bitmap(oid, target_version, bitmap, result_version);
-}
-
-std::map<uint64_t, uint64_t> & blockstore_t::get_inode_space_stats()
-{
-    return impl->inode_space_stats;
-}
-
-void blockstore_t::dump_diagnostics()
-{
-    return impl->dump_diagnostics();
-}
-
-uint32_t blockstore_t::get_block_size()
-{
-    return impl->get_block_size();
-}
-
-uint64_t blockstore_t::get_block_count()
-{
-    return impl->get_block_count();
-}
-
-uint64_t blockstore_t::get_free_block_count()
-{
-    return impl->get_free_block_count();
-}
-
-uint64_t blockstore_t::get_journal_size()
-{
-    return impl->get_journal_size();
-}
-
-uint32_t blockstore_t::get_bitmap_granularity()
-{
-    return impl->get_bitmap_granularity();
-}
-
-void blockstore_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
-{
-    impl->set_no_inode_stats(pool_ids);
+    auto meta_format = stoull_full(config["meta_format"]);
+    if (meta_format == BLOCKSTORE_META_FORMAT_HEAP)
+        return new blockstore_impl_t(config, ringloop, tfd);
+    else
+        return new v1::blockstore_impl_t(config, ringloop, tfd);
 }

src/blockstore/blockstore.h

@@ -17,22 +17,14 @@
 #include "ringloop.h"
 #include "timerfd_manager.h"
 
-// Memory alignment for direct I/O (usually 512 bytes)
-#ifndef DIRECT_IO_ALIGNMENT
-#define DIRECT_IO_ALIGNMENT 512
-#endif
-
-// Memory allocation alignment (page size is usually optimal)
-#ifndef MEM_ALIGNMENT
-#define MEM_ALIGNMENT 4096
-#endif
-
 // Default block size is 128 KB, current allowed range is 4K - 128M
 #define DEFAULT_DATA_BLOCK_ORDER 17
 #define MIN_DATA_BLOCK_SIZE 4*1024
 #define MAX_DATA_BLOCK_SIZE 128*1024*1024
 #define DEFAULT_BITMAP_GRANULARITY 4096
 
+#define MIN_JOURNAL_SIZE 1024*1024
+
 #define BS_OP_MIN 1
 #define BS_OP_READ 1
 #define BS_OP_WRITE 2
@@ -46,8 +38,18 @@
 
 #define BS_OP_PRIVATE_DATA_SIZE 256
 
+#define IMMEDIATE_NONE 0
+#define IMMEDIATE_SMALL 1
+#define IMMEDIATE_ALL 2
+
 /*
 
+All operations may be submitted in any order, because reads only see completed writes,
+syncs only sync completed writes and writes don't depend on each other.
+
+The only restriction is that the external code MUST NOT submit multiple writes for one
+object in parallel. This is a natural restriction because `version` numbers are used though.
+
 Blockstore opcode documentation:
 
 ## BS_OP_READ / BS_OP_WRITE / BS_OP_WRITE_STABLE
@@ -162,8 +164,8 @@ struct __attribute__ ((visibility("default"))) blockstore_op_t
             uint32_t list_stable_limit;
         };
     };
-    void *buf = NULL;
-    void *bitmap = NULL;
+    uint8_t *buf = NULL;
+    uint8_t *bitmap = NULL;
     int retval = 0;
 
     uint8_t private_data[BS_OP_PRIVATE_DATA_SIZE];
@@ -171,53 +173,54 @@ struct __attribute__ ((visibility("default"))) blockstore_op_t
 
 typedef std::map<std::string, std::string> blockstore_config_t;
 
-class blockstore_impl_t;
-
-class __attribute__((visibility("default"))) blockstore_t
+class __attribute__((visibility("default"))) blockstore_i
 {
-    blockstore_impl_t *impl;
 public:
-    blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd);
-    ~blockstore_t();
+    static blockstore_i* create(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd);
+
+    virtual ~blockstore_i() = default;
 
     // Update configuration
-    void parse_config(blockstore_config_t & config);
+    virtual void parse_config(blockstore_config_t & config) = 0;
 
     // Event loop
-    void loop();
+    virtual void loop() = 0;
 
     // Returns true when blockstore is ready to process operations
     // (Although you're free to enqueue them before that)
-    bool is_started();
+    virtual bool is_started() = 0;
 
     // Returns true when blockstore is stalled
-    bool is_stalled();
+    virtual bool is_stalled() = 0;
 
     // Returns true when it's safe to destroy the instance. If destroying the instance
     // requires to purge some queues, starts that process. Should be called in the event
     // loop until it returns true.
-    bool is_safe_to_stop();
+    virtual bool is_safe_to_stop() = 0;
 
     // Submission
-    void enqueue_op(blockstore_op_t *op);
+    virtual void enqueue_op(blockstore_op_t *op) = 0;
 
     // Simplified synchronous operation: get object bitmap & current version
-    int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL);
+    virtual int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL) = 0;
 
     // Get per-inode space usage statistics
-    std::map<uint64_t, uint64_t> & get_inode_space_stats();
+    virtual const std::map<uint64_t, uint64_t> & get_inode_space_stats() = 0;
 
     // Set per-pool no_inode_stats
-    void set_no_inode_stats(const std::vector<uint64_t> & pool_ids);
+    virtual void set_no_inode_stats(const std::vector<uint64_t> & pool_ids) = 0;
 
     // Print diagnostics to stdout
-    void dump_diagnostics();
+    virtual void dump_diagnostics() = 0;
 
-    uint32_t get_block_size();
-    uint64_t get_block_count();
-    uint64_t get_free_block_count();
+    // Get diagnostic string for an operation
+    virtual std::string get_op_diag(blockstore_op_t *op) = 0;
 
-    uint64_t get_journal_size();
+    virtual uint32_t get_block_size() = 0;
+    virtual uint64_t get_block_count() = 0;
+    virtual uint64_t get_free_block_count() = 0;
 
-    uint32_t get_bitmap_granularity();
+    virtual uint64_t get_journal_size() = 0;
+
+    virtual uint32_t get_bitmap_granularity() = 0;
 };

src/blockstore/blockstore_disk.cpp

@@ -2,11 +2,15 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include <sys/file.h>
+#include <sys/ioctl.h>
+#include <unistd.h>
 
 #include <stdexcept>
 
-#include "blockstore_impl.h"
+#include "blockstore.h"
+#include "ondisk_formats.h"
 #include "blockstore_disk.h"
+#include "blockstore_heap.h"
 #include "str_util.h"
 #include "allocator.h"
 
@@ -44,8 +48,11 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     disk_alignment = parse_size(config["disk_alignment"]);
     journal_block_size = parse_size(config["journal_block_size"]);
     meta_block_size = parse_size(config["meta_block_size"]);
+    meta_block_target_free_space = parse_size(config["meta_block_target_free_space"]);
     bitmap_granularity = parse_size(config["bitmap_granularity"]);
     meta_format = stoull_full(config["meta_format"]);
+    atomic_write_size = (config.find("atomic_write_size") != config.end()
+        ? parse_size(config["atomic_write_size"]) : 4096);
     if (config.find("data_io") == config.end() &&
         config.find("meta_io") == config.end() &&
         config.find("journal_io") == config.end())
@@ -90,12 +97,28 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     if (!min_discard_size)
         min_discard_size = 1024*1024;
     discard_granularity = parse_size(config["discard_granularity"]);
+    inmemory_meta = config["inmemory_metadata"] != "false" && config["inmemory_metadata"] != "0" &&
+        config["inmemory_metadata"] != "no";
+    inmemory_journal = config["inmemory_journal"] != "false" && config["inmemory_journal"] != "0" &&
+        config["inmemory_journal"] != "no";
+    disable_data_fsync = config["disable_data_fsync"] == "true" || config["disable_data_fsync"] == "1" || config["disable_data_fsync"] == "yes";
+    disable_meta_fsync = config["disable_meta_fsync"] == "true" || config["disable_meta_fsync"] == "1" || config["disable_meta_fsync"] == "yes";
+    disable_journal_fsync = config["disable_journal_fsync"] == "true" || config["disable_journal_fsync"] == "1" || config["disable_journal_fsync"] == "yes";
+    if (mock_mode)
+    {
+        data_device_size = parse_size(config["data_device_size"]);
+        data_device_sect = parse_size(config["data_device_sect"]);
+        meta_device_size = parse_size(config["meta_device_size"]);
+        meta_device_sect = parse_size(config["meta_device_sect"]);
+        journal_device_size = parse_size(config["journal_device_size"]);
+        journal_device_sect = parse_size(config["journal_device_sect"]);
+    }
     // Validate
     if (!data_block_size)
     {
         data_block_size = (1 << DEFAULT_DATA_BLOCK_ORDER);
     }
-    if ((block_order = is_power_of_two(data_block_size)) >= 64 || data_block_size < MIN_DATA_BLOCK_SIZE || data_block_size >= MAX_DATA_BLOCK_SIZE)
+    if (is_power_of_two(data_block_size) >= 64 || data_block_size < MIN_DATA_BLOCK_SIZE || data_block_size >= MAX_DATA_BLOCK_SIZE)
     {
         throw std::runtime_error("Bad block size");
     }
@@ -131,6 +154,14 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     {
         throw std::runtime_error("meta_block_size must not exceed "+std::to_string(MAX_DATA_BLOCK_SIZE));
     }
+    if (!meta_block_target_free_space)
+    {
+        meta_block_target_free_space = 800;
+    }
+    if (meta_block_target_free_space >= meta_block_size)
+    {
+        throw std::runtime_error("meta_block_target_free_space must not exceed "+std::to_string(meta_block_size));
+    }
     if (data_offset % disk_alignment)
     {
         throw std::runtime_error("data_offset must be a multiple of disk_alignment = "+std::to_string(disk_alignment));
@@ -179,17 +210,29 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     {
         throw std::runtime_error("journal_offset must be a multiple of journal_block_size = "+std::to_string(journal_block_size));
     }
+    if (!meta_format)
+    {
+        meta_format = BLOCKSTORE_META_FORMAT_HEAP;
+    }
+    if (meta_device == data_device)
+    {
+        disable_meta_fsync = disable_data_fsync;
+    }
+    if (journal_device == meta_device)
+    {
+        disable_journal_fsync = disable_meta_fsync;
+    }
 }
 
-void blockstore_disk_t::calc_lengths(bool skip_meta_check)
+void blockstore_disk_t::calc_lengths()
 {
     // data
     data_len = data_device_size - data_offset;
-    if (data_fd == meta_fd && data_offset < meta_offset)
+    if (data_device == meta_device && data_offset < meta_offset)
     {
         data_len = meta_offset - data_offset;
     }
-    if (data_fd == journal_fd && data_offset < journal_offset)
+    if (data_device == journal_device && data_offset < journal_offset)
     {
         data_len = data_len < journal_offset-data_offset
             ? data_len : journal_offset-data_offset;
@@ -204,23 +247,23 @@ void blockstore_disk_t::calc_lengths(bool skip_meta_check)
         data_len = cfg_data_size;
     }
     // meta
-    uint64_t meta_area_size = (meta_fd == data_fd ? data_device_size : meta_device_size) - meta_offset;
-    if (meta_fd == data_fd && meta_offset <= data_offset)
+    meta_area_size = (meta_device == data_device ? data_device_size : meta_device_size) - meta_offset;
+    if (meta_device == data_device && meta_offset <= data_offset)
     {
         meta_area_size = data_offset - meta_offset;
     }
-    if (meta_fd == journal_fd && meta_offset <= journal_offset)
+    if (meta_device == journal_device && meta_offset <= journal_offset)
     {
         meta_area_size = meta_area_size < journal_offset-meta_offset
             ? meta_area_size : journal_offset-meta_offset;
     }
     // journal
-    journal_len = (journal_fd == data_fd ? data_device_size : (journal_fd == meta_fd ? meta_device_size : journal_device_size)) - journal_offset;
-    if (journal_fd == data_fd && journal_offset <= data_offset)
+    journal_len = (journal_device == data_device ? data_device_size : (journal_device == meta_device ? meta_device_size : journal_device_size)) - journal_offset;
+    if (journal_device == data_device && journal_offset <= data_offset)
     {
         journal_len = data_offset - journal_offset;
     }
-    if (journal_fd == meta_fd && journal_offset <= meta_offset)
+    if (journal_device == meta_device && journal_offset <= meta_offset)
     {
         journal_len = journal_len < meta_offset-journal_offset
             ? journal_len : meta_offset-journal_offset;
@@ -230,37 +273,37 @@ void blockstore_disk_t::calc_lengths(bool skip_meta_check)
     clean_entry_bitmap_size = data_block_size / bitmap_granularity / 8;
     clean_dyn_size = clean_entry_bitmap_size*2 + (csum_block_size
         ? data_block_size/csum_block_size*(data_csum_type & 0xFF) : 0);
-    clean_entry_size = sizeof(clean_disk_entry) + clean_dyn_size + 4 /*entry_csum*/;
-    meta_len = (1 + (block_count - 1 + meta_block_size / clean_entry_size) / (meta_block_size / clean_entry_size)) * meta_block_size;
-    bool new_doesnt_fit = (!meta_format && !skip_meta_check && meta_area_size < meta_len && !data_csum_type);
-    if (meta_format == BLOCKSTORE_META_FORMAT_V1 || new_doesnt_fit)
+    if (meta_format == BLOCKSTORE_META_FORMAT_HEAP)
     {
-        uint64_t clean_entry_v0_size = sizeof(clean_disk_entry) + 2*clean_entry_bitmap_size;
-        uint64_t meta_v0_len = (1 + (block_count - 1 + meta_block_size / clean_entry_v0_size)
-            / (meta_block_size / clean_entry_v0_size)) * meta_block_size;
-        if (meta_format == BLOCKSTORE_META_FORMAT_V1 || meta_area_size >= meta_v0_len)
-        {
-            // Old metadata fits.
-            if (new_doesnt_fit)
-            {
-                printf("Warning: Using old metadata format without checksums because the new format"
-                    " doesn't fit into provided area (%ju bytes required, %ju bytes available)\n", meta_len, meta_area_size);
-            }
-            clean_entry_size = clean_entry_v0_size;
-            meta_len = meta_v0_len;
-            meta_format = BLOCKSTORE_META_FORMAT_V1;
-        }
-        else
-            meta_format = BLOCKSTORE_META_FORMAT_V2;
+        uint32_t entries_per_block = ((meta_block_size-meta_block_target_free_space) /
+            (sizeof(heap_object_t) + sizeof(heap_write_t) + clean_dyn_size));
+        min_meta_len = (block_count+entries_per_block-1) / entries_per_block * meta_block_size;
+    }
+    else if (meta_format == BLOCKSTORE_META_FORMAT_V1)
+    {
+        clean_entry_size = 24 /*sizeof(clean_disk_entry)*/ + 2*clean_entry_bitmap_size;
+        min_meta_len = (1 + (block_count - 1 + meta_block_size / clean_entry_size)
+            / (meta_block_size / clean_entry_size)) * meta_block_size;
+    }
+    else if (meta_format == BLOCKSTORE_META_FORMAT_V2)
+    {
+        clean_entry_size = 24 /*sizeof(clean_disk_entry)*/ + clean_dyn_size + 4 /*entry_csum*/;
+        min_meta_len = (1 + (block_count - 1 + meta_block_size / clean_entry_size) / (meta_block_size / clean_entry_size)) * meta_block_size;
     }
     else
-        meta_format = BLOCKSTORE_META_FORMAT_V2;
-    if (!skip_meta_check && meta_area_size < meta_len)
     {
-        throw std::runtime_error("Metadata area is too small, need at least "+std::to_string(meta_len)+" bytes, have only "+std::to_string(meta_area_size)+" bytes");
+        throw std::runtime_error("meta_format = "+std::to_string(meta_format)+" is not supported");
+    }
+}
+
+void blockstore_disk_t::check_lengths()
+{
+    if (meta_area_size < min_meta_len)
+    {
+        throw std::runtime_error("Metadata area is too small, need at least "+std::to_string(min_meta_len)+" bytes, have only "+std::to_string(meta_area_size)+" bytes");
     }
     // requested journal size
-    if (!skip_meta_check && cfg_journal_size > journal_len)
+    if (cfg_journal_size > journal_len)
     {
         throw std::runtime_error("Requested journal_size is too large");
     }
@@ -321,12 +364,19 @@ static int bs_openmode(const std::string & mode)
 
 void blockstore_disk_t::open_data()
 {
-    data_fd = open(data_device.c_str(), bs_openmode(data_io) | O_RDWR);
+    if (data_fd >= 0)
+    {
+        throw std::runtime_error("data device is already opened");
+    }
+    data_fd = mock_mode ? MOCK_DATA_FD : open(data_device.c_str(), bs_openmode(data_io) | O_RDWR);
     if (data_fd == -1)
     {
         throw std::runtime_error("Failed to open data device "+data_device+": "+std::string(strerror(errno)));
     }
-    check_size(data_fd, &data_device_size, &data_device_sect, "data device");
+    if (!mock_mode)
+    {
+        check_size(data_fd, &data_device_size, &data_device_sect, "data device");
+    }
     if (disk_alignment % data_device_sect)
     {
         throw std::runtime_error(
@@ -338,7 +388,7 @@ void blockstore_disk_t::open_data()
     {
         throw std::runtime_error("data_offset exceeds device size = "+std::to_string(data_device_size));
     }
-    if (!disable_flock && flock(data_fd, LOCK_EX|LOCK_NB) != 0)
+    if (!mock_mode && !disable_flock && flock(data_fd, LOCK_EX|LOCK_NB) != 0)
     {
         throw std::runtime_error(std::string("Failed to lock data device: ") + strerror(errno));
     }
@@ -346,19 +396,26 @@ void blockstore_disk_t::open_data()
 
 void blockstore_disk_t::open_meta()
 {
+    if (meta_fd >= 0)
+    {
+        throw std::runtime_error("metadata device is already opened");
+    }
     if (meta_device != data_device || meta_io != data_io)
     {
-        meta_fd = open(meta_device.c_str(), bs_openmode(meta_io) | O_RDWR);
+        meta_fd = mock_mode ? MOCK_META_FD : open(meta_device.c_str(), bs_openmode(meta_io) | O_RDWR);
         if (meta_fd == -1)
         {
             throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
         }
-        check_size(meta_fd, &meta_device_size, &meta_device_sect, "metadata device");
+        if (!mock_mode)
+        {
+            check_size(meta_fd, &meta_device_size, &meta_device_sect, "metadata device");
+        }
         if (meta_offset >= meta_device_size)
         {
             throw std::runtime_error("meta_offset exceeds device size = "+std::to_string(meta_device_size));
         }
-        if (!disable_flock && meta_device != data_device && flock(meta_fd, LOCK_EX|LOCK_NB) != 0)
+        if (!mock_mode && !disable_flock && meta_device != data_device && flock(meta_fd, LOCK_EX|LOCK_NB) != 0)
         {
             throw std::runtime_error(std::string("Failed to lock metadata device: ") + strerror(errno));
         }
@@ -384,15 +441,26 @@ void blockstore_disk_t::open_meta()
 
 void blockstore_disk_t::open_journal()
 {
+    if (journal_fd >= 0)
+    {
+        throw std::runtime_error("journal device is already opened");
+    }
     if (journal_device != meta_device || journal_io != meta_io)
     {
-        journal_fd = open(journal_device.c_str(), bs_openmode(journal_io) | O_RDWR);
+        journal_fd = mock_mode ? MOCK_JOURNAL_FD : open(journal_device.c_str(), bs_openmode(journal_io) | O_RDWR);
         if (journal_fd == -1)
         {
             throw std::runtime_error("Failed to open journal device "+journal_device+": "+std::string(strerror(errno)));
         }
-        check_size(journal_fd, &journal_device_size, &journal_device_sect, "journal device");
-        if (!disable_flock && journal_device != meta_device && flock(journal_fd, LOCK_EX|LOCK_NB) != 0)
+        if (!mock_mode)
+        {
+            check_size(journal_fd, &journal_device_size, &journal_device_sect, "journal device");
+        }
+        if (journal_offset >= journal_device_size)
+        {
+            throw std::runtime_error("journal_offset exceeds device size = "+std::to_string(journal_device_size));
+        }
+        if (!mock_mode && !disable_flock && journal_device != meta_device && flock(journal_fd, LOCK_EX|LOCK_NB) != 0)
         {
             throw std::runtime_error(std::string("Failed to lock journal device: ") + strerror(errno));
         }
@@ -418,25 +486,32 @@ void blockstore_disk_t::open_journal()
 
 void blockstore_disk_t::close_all()
 {
-    if (data_fd >= 0)
-        close(data_fd);
-    if (meta_fd >= 0 && meta_fd != data_fd)
-        close(meta_fd);
-    if (journal_fd >= 0 && journal_fd != meta_fd)
-        close(journal_fd);
+    if (!mock_mode)
+    {
+        if (data_fd >= 0)
+            close(data_fd);
+        if (meta_fd >= 0 && meta_fd != data_fd)
+            close(meta_fd);
+        if (journal_fd >= 0 && journal_fd != meta_fd)
+            close(journal_fd);
+    }
     data_fd = meta_fd = journal_fd = -1;
 }
 
 // Sadly DISCARD only works through ioctl(), but it seems to always block the device queue,
 // so it's not a big deal that we can only run it synchronously.
-int blockstore_disk_t::trim_data(allocator_t *alloc)
+int blockstore_disk_t::trim_data(std::function<bool(uint64_t)> is_free)
 {
+    if (mock_mode)
+    {
+        return -EINVAL;
+    }
     int r = 0;
     uint64_t j = 0, i = 0;
     uint64_t discarded = 0;
     for (; i <= block_count; i++)
     {
-        if (i >= block_count || alloc->get(i))
+        if (i >= block_count || is_free(i))
         {
             if (i > j && (i-j)*data_block_size >= min_discard_size)
             {

src/blockstore/blockstore_disk.h

@@ -8,10 +8,19 @@
 #include <string>
 #include <map>
 
+// Memory alignment for direct I/O (usually 512 bytes)
+#ifndef DIRECT_IO_ALIGNMENT
+#define DIRECT_IO_ALIGNMENT 512
+#endif
+
 #define BLOCKSTORE_CSUM_NONE 0
 // Lower byte of checksum type is its length
 #define BLOCKSTORE_CSUM_CRC32C 0x104
 
+#define MOCK_DATA_FD 1000
+#define MOCK_META_FD 1001
+#define MOCK_JOURNAL_FD 1002
+
 class allocator_t;
 
 struct blockstore_disk_t
@@ -22,11 +31,15 @@ struct blockstore_disk_t
     // Required write alignment and journal/metadata/data areas' location alignment
     uint32_t disk_alignment = 4096;
     // Journal block size - minimum_io_size of the journal device is the best choice
-    uint64_t journal_block_size = 4096;
+    uint32_t journal_block_size = 4096;
     // Metadata block size - minimum_io_size of the metadata device is the best choice
-    uint64_t meta_block_size = 4096;
+    uint32_t meta_block_size = 4096;
+    // Atomic write size of the data block device
+    uint32_t atomic_write_size = 4096;
+    // Target free space in metadata blocks
+    uint32_t meta_block_target_free_space = 800;
     // Sparse write tracking granularity. 4 KB is a good choice. Must be a multiple of disk_alignment
-    uint64_t bitmap_granularity = 4096;
+    uint32_t bitmap_granularity = 4096;
     // Data checksum type, BLOCKSTORE_CSUM_NONE or BLOCKSTORE_CSUM_CRC32C
     uint32_t data_csum_type = BLOCKSTORE_CSUM_NONE;
     // Checksum block size, must be a multiple of bitmap_granularity
@@ -36,27 +49,36 @@ struct blockstore_disk_t
     // I/O modes for data, metadata and journal: direct or "" = O_DIRECT, cached = O_SYNC, directsync = O_DIRECT|O_SYNC
     // O_SYNC without O_DIRECT = use Linux page cache for reads and writes
     std::string data_io, meta_io, journal_io;
+    // It is safe to disable fsync() if drive write cache is writethrough
+    bool disable_data_fsync = false, disable_meta_fsync = false, disable_journal_fsync = false;
+    // Keep journal (buffered data) in memory?
+    bool inmemory_meta = true;
+    // Keep metadata in memory?
+    bool inmemory_journal = true;
     // Data discard granularity and minimum size (for the sake of performance)
     bool discard_on_start = false;
     uint64_t min_discard_size = 1024*1024;
     uint64_t discard_granularity = 0;
 
     int meta_fd = -1, data_fd = -1, journal_fd = -1;
-    uint64_t meta_offset, meta_device_sect, meta_device_size, meta_len, meta_format = 0;
+    uint64_t meta_offset, meta_device_sect, meta_device_size, meta_area_size, min_meta_len, meta_format = 0;
     uint64_t data_offset, data_device_sect, data_device_size, data_len;
     uint64_t journal_offset, journal_device_sect, journal_device_size, journal_len;
 
-    uint32_t block_order = 0;
     uint64_t block_count = 0;
-    uint32_t clean_entry_bitmap_size = 0, clean_entry_size = 0, clean_dyn_size = 0;
+    uint32_t clean_entry_bitmap_size = 0;
+    uint32_t clean_entry_size = 0, clean_dyn_size = 0; // for meta_v1/2
+
+    bool mock_mode = false;
 
     void parse_config(std::map<std::string, std::string> & config);
     void open_data();
     void open_meta();
     void open_journal();
-    void calc_lengths(bool skip_meta_check = false);
+    void calc_lengths();
+    void check_lengths();
     void close_all();
-    int trim_data(allocator_t *alloc);
+    int trim_data(std::function<bool(uint64_t)> is_free);
 
     inline uint64_t dirty_dyn_size(uint64_t offset, uint64_t len)
     {

src/blockstore/blockstore_flush.h

@@ -1,22 +1,12 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#define COPY_BUF_JOURNAL 1
-#define COPY_BUF_DATA 2
-#define COPY_BUF_ZERO 4
-#define COPY_BUF_CSUM_FILL 8
-#define COPY_BUF_COALESCED 16
-#define COPY_BUF_META_BLOCK 32
-#define COPY_BUF_JOURNALED_BIG 64
-
 struct copy_buffer_t
 {
-    int copy_flags;
-    uint64_t offset, len, disk_offset;
-    uint64_t journal_sector; // only for reads: sector+1 if used and !journal.inmemory, otherwise 0
-    void *buf;
-    uint8_t *csum_buf;
-    int *dyn_data;
+    uint32_t copy_flags;
+    uint64_t offset, len, disk_loc, disk_offset, disk_len;
+    uint8_t *buf;
+    uint64_t wr_lsn;
 };
 
 struct meta_sector_t
@@ -27,13 +17,6 @@ struct meta_sector_t
     int usage_count;
 };
 
-struct flusher_sync_t
-{
-    bool fsync_meta;
-    int ready_count;
-    int state;
-};
-
 struct flusher_meta_write_t
 {
     uint64_t sector, pos;
@@ -49,94 +32,75 @@ class journal_flusher_co
 {
     blockstore_impl_t *bs;
     journal_flusher_t *flusher;
-    int wait_state, wait_count, wait_journal_count;
+    int co_id;
+    int wait_state, wait_count;
     struct io_uring_sqe *sqe;
     struct ring_data_t *data;
 
-    std::list<flusher_sync_t>::iterator cur_sync;
+    std::function<void(ring_data_t*)> simple_callback_r, simple_callback_w;
 
-    obj_ver_id cur;
-    std::map<obj_ver_id, dirty_entry>::iterator dirty_it, dirty_start, dirty_end;
-    std::map<object_id, uint64_t>::iterator repeat_it;
-    std::function<void(ring_data_t*)> simple_callback_r, simple_callback_rj, simple_callback_w;
+    object_id cur_oid;
+    uint64_t copy_id;
+    uint64_t compact_lsn;
+    uint64_t cur_version;
+    heap_object_t *cur_obj;
+    heap_write_t *begin_wr, *end_wr;
+    uint32_t modified_block;
+    bool should_repeat;
 
-    bool try_trim = false;
-    bool skip_copy, has_delete, has_writes;
-    std::vector<copy_buffer_t> v;
-    std::vector<copy_buffer_t>::iterator it;
-    int i;
-    bool fill_incomplete, cleared_incomplete;
-    int read_to_fill_incomplete;
+    std::vector<copy_buffer_t> read_vec;
+    uint32_t overwrite_start, overwrite_end;
+    uint32_t big_start, big_end;
+    int i, res;
+    bool read_to_fill_incomplete;
     int copy_count;
-    uint64_t clean_loc, clean_ver, old_clean_loc, old_clean_ver;
+    uint64_t clean_loc;
     flusher_meta_write_t meta_old, meta_new;
-    bool clean_init_bitmap;
-    uint64_t clean_bitmap_offset, clean_bitmap_len;
-    uint8_t *clean_init_dyn_ptr;
-    uint8_t *new_clean_bitmap;
-
-    uint64_t new_trim_pos;
+    bool do_repeat = false;
 
     friend class journal_flusher_t;
-    void scan_dirty();
-    bool read_dirty(int wait_base);
-    bool modify_meta_do_reads(int wait_base);
-    bool wait_meta_reads(int wait_base);
-    bool modify_meta_read(uint64_t meta_loc, flusher_meta_write_t &wr, int wait_base);
-    bool clear_incomplete_csum_block_bits(int wait_base);
-    void calc_block_checksums(uint32_t *new_data_csums, bool skip_overwrites);
-    void update_metadata_entry();
-    bool write_meta_block(flusher_meta_write_t & meta_block, int wait_base);
-    void update_clean_db();
-    void free_data_blocks();
-    bool fsync_batch(bool fsync_meta, int wait_base);
-    bool trim_journal(int wait_base);
+
+    void iterate_checksum_holes(std::function<void(int & pos, uint32_t hole_start, uint32_t hole_end)> cb);
+    void fill_partial_checksum_blocks();
     void free_buffers();
+    int check_and_punch_checksums();
+    bool calc_block_checksums();
+    bool write_meta_block(int wait_base);
+    bool read_buffered(int wait_base);
+    bool fsync_meta(int wait_base);
+    int fsync_buffer(int wait_base);
+    bool trim_lsn(int wait_base);
 public:
     journal_flusher_co();
+    ~journal_flusher_co();
     bool loop();
 };
 
 // Journal flusher itself
 class journal_flusher_t
 {
-    int trim_wanted = 0;
-    bool dequeuing;
-    int min_flusher_count, max_flusher_count, cur_flusher_count, target_flusher_count;
-    int flusher_start_threshold;
+    int force_start = 0;
+    int min_flusher_count = 0, max_flusher_count = 0, cur_flusher_count = 0, target_flusher_count = 0;
     journal_flusher_co *co;
     blockstore_impl_t *bs;
     friend class journal_flusher_co;
 
-    int journal_trim_counter;
-    bool trimming;
-    void* journal_superblock;
+    int advance_lsn_counter = 0;
+    uint64_t compact_counter = 0;
 
-    int active_flushers;
-    int syncing_flushers;
-    std::list<flusher_sync_t> syncs;
-    std::map<object_id, uint64_t> sync_to_repeat;
-
-    std::map<uint64_t, meta_sector_t> meta_sectors;
-    std::deque<object_id> flush_queue;
-    std::unordered_map<object_id, uint64_t> flush_versions;
-    std::unordered_set<uint64_t> inflight_meta_sectors;
-
-    bool try_find_older(std::map<obj_ver_id, dirty_entry>::iterator & dirty_end, obj_ver_id & cur);
-    bool try_find_other(std::map<obj_ver_id, dirty_entry>::iterator & dirty_end, obj_ver_id & cur);
+    int active_flushers = 0;
+    int wanting_meta_fsync = 0;
+    bool fsyncing_meta = false;
+    int syncing_buffer = 0;
 
 public:
     journal_flusher_t(blockstore_impl_t *bs);
     ~journal_flusher_t();
     void loop();
-    bool is_trim_wanted() { return trim_wanted; }
+    int get_syncing_buffer();
+    uint64_t get_compact_counter();
     bool is_active();
-    void mark_trim_possible();
     void request_trim();
     void release_trim();
-    void enqueue_flush(obj_ver_id oid);
-    void unshift_flush(obj_ver_id oid, bool force);
-    void remove_flush(object_id oid);
     void dump_diagnostics();
-    bool is_mutated(uint64_t clean_loc);
 };

src/blockstore/blockstore_heap.h

@@ -0,0 +1,376 @@
+// Metadata storage version 3 ("heap")
+// Copyright (c) Vitaliy Filippov, 2025+
+// License: VNPL-1.1 (see README.md for details)
+
+#pragma once
+
+#include <map>
+#include <unordered_map>
+#include <set>
+#include <deque>
+#include <vector>
+
+#include "../client/object_id.h"
+#include "../util/robin_hood.h"
+#include "blockstore_disk.h"
+#include "multilist.h"
+
+struct pool_shard_settings_t
+{
+    uint32_t pg_count;
+    uint32_t pg_stripe_size;
+};
+
+#define BS_HEAP_TYPE 7
+#define BS_HEAP_OBJECT 1
+#define BS_HEAP_SMALL_WRITE 2
+#define BS_HEAP_BIG_WRITE 3
+#define BS_HEAP_TOMBSTONE 4
+#define BS_HEAP_INTENT_WRITE 5
+#define BS_HEAP_STABLE 8
+
+class blockstore_heap_t;
+
+struct __attribute__((__packed__)) heap_small_write_t
+{
+    uint16_t size;
+    int16_t next_pos;
+    uint8_t flags;
+    uint64_t lsn;
+    uint64_t version;
+    uint64_t location;
+    uint32_t offset;
+    uint32_t len;
+};
+
+struct __attribute__((__packed__)) heap_big_write_t
+{
+    uint16_t size;
+    int16_t next_pos;
+    uint8_t flags;
+    uint64_t lsn;
+    uint64_t version;
+    uint32_t block_num;
+};
+
+struct __attribute__((__packed__)) heap_tombstone_t
+{
+    uint16_t size;
+    int16_t next_pos;
+    uint8_t flags;
+    uint64_t lsn;
+    uint64_t version;
+};
+
+struct __attribute__((__packed__)) heap_write_t
+{
+    // size should have top bit cleared
+    uint16_t size = 0;
+    int16_t next_pos = 0;
+    uint8_t entry_type = 0; // BS_HEAP_*
+    uint64_t lsn = 0;
+    uint64_t version = 0;
+
+    // uint8_t[] external_bitmap
+    // uint8_t[] internal_bitmap
+    // uint32_t[] checksums
+
+    heap_write_t *next();
+    inline uint8_t type() const { return (entry_type & BS_HEAP_TYPE); }
+    inline heap_small_write_t& small() { return *(heap_small_write_t*)this; }
+    inline heap_big_write_t& big() { return *(heap_big_write_t*)this; }
+    uint32_t get_size(blockstore_heap_t *heap);
+    uint32_t get_csum_size(blockstore_heap_t *heap);
+    bool needs_recheck(blockstore_heap_t *heap);
+    bool needs_compact(blockstore_heap_t *heap);
+    bool is_compacted(uint64_t compacted_lsn);
+    bool can_be_collapsed(blockstore_heap_t *heap);
+    bool is_allowed_before_compacted(uint64_t compacted_lsn, bool is_last_entry);
+    uint8_t *get_ext_bitmap(blockstore_heap_t *heap);
+    uint8_t *get_int_bitmap(blockstore_heap_t *heap);
+    uint8_t *get_checksums(blockstore_heap_t *heap);
+    uint32_t *get_checksum(blockstore_heap_t *heap);
+    uint64_t big_location(blockstore_heap_t *heap);
+    void set_big_location(blockstore_heap_t *heap, uint64_t location);
+};
+
+struct __attribute__((__packed__)) heap_object_t
+{
+    // size should have top bit cleared
+    uint16_t size = 0;
+    // linked list of write entries...
+    // newest entries are stored first to simplify scanning
+    int16_t write_pos = 0;
+    uint8_t entry_type = 0; // BS_HEAP_*
+    uint32_t crc32c = 0;
+    uint64_t inode = 0;
+    uint64_t stripe = 0;
+
+    heap_write_t *get_writes();
+    uint32_t calc_crc32c();
+};
+
+struct heap_object_lsn_t
+{
+    object_id oid;
+    uint64_t lsn;
+};
+
+inline bool operator < (const heap_object_lsn_t & a, const heap_object_lsn_t & b)
+{
+    return a.oid < b.oid || a.oid == b.oid && a.lsn < b.lsn;
+}
+
+struct tmp_compact_item_t
+{
+    object_id oid;
+    uint64_t lsn;
+    bool compact;
+};
+
+struct heap_mvcc_copy_id_t
+{
+    object_id oid;
+    uint64_t copy_id;
+};
+
+inline bool operator == (const heap_mvcc_copy_id_t & a, const heap_mvcc_copy_id_t & b)
+{
+    return a.oid.inode == b.oid.inode && a.oid.stripe == b.oid.stripe && a.copy_id == b.copy_id;
+}
+
+namespace std
+{
+    template<> struct hash<heap_mvcc_copy_id_t>
+    {
+        inline size_t operator()(const heap_mvcc_copy_id_t &s) const
+        {
+            size_t seed = std::hash<object_id>()(s.oid);
+            // Copy-pasted from spp::hash_combine()
+            seed ^= (s.copy_id + 0xc6a4a7935bd1e995 + (seed << 6) + (seed >> 2));
+            return seed;
+        }
+    };
+};
+
+struct heap_object_mvcc_t
+{
+    uint32_t readers = 0;
+    heap_object_t *entry_copy = NULL;
+};
+
+struct __attribute__((__packed__)) heap_block_info_t
+{
+    uint32_t used_space = 0;
+    uint32_t free_pos = 0;
+    uint8_t *data = NULL;
+};
+
+struct heap_inflight_lsn_t
+{
+    object_id oid;
+    uint64_t flags;
+};
+
+struct heap_refqi_t
+{
+    uint64_t lsn;
+    uint64_t inode;
+    uint64_t location;
+    uint32_t len;
+    bool is_data;
+};
+
+using i64hash_t = robin_hood::hash<uint64_t>;
+using heap_block_index_t = robin_hood::unordered_flat_map<uint64_t,
+    robin_hood::unordered_flat_map<inode_t, robin_hood::unordered_flat_map<uint64_t, uint64_t, i64hash_t, std::equal_to<uint64_t>, 88>, i64hash_t>, i64hash_t>;
+using heap_mvcc_map_t = robin_hood::unordered_flat_map<heap_mvcc_copy_id_t, heap_object_mvcc_t>;
+
+class blockstore_heap_t
+{
+    friend class heap_write_t;
+    friend class heap_object_t;
+
+    blockstore_disk_t *dsk = NULL;
+    uint8_t* buffer_area = NULL;
+    bool abort_on_corruption = false;
+    bool abort_on_overlap = true;
+    int log_level = 0;
+
+    const uint32_t meta_block_count = 0;
+    uint32_t target_block_free_space = 800;
+
+    uint64_t next_lsn = 0;
+    robin_hood::unordered_flat_map<pool_id_t, pool_shard_settings_t> pool_shard_settings;
+    // PG => inode => stripe => block number
+    heap_block_index_t block_index;
+    std::vector<heap_block_info_t> block_info;
+    allocator_t *data_alloc = NULL;
+    multilist_index_t *meta_alloc = NULL;
+    uint32_t meta_alloc_count = 0;
+    uint64_t meta_used_space = 0;
+    multilist_alloc_t *buffer_alloc = NULL;
+    heap_mvcc_map_t object_mvcc;
+    std::unordered_map<uint64_t, uint32_t> mvcc_data_refs;
+    std::unordered_map<uint64_t, uint32_t> mvcc_buffer_refs;
+    std::map<uint64_t, uint64_t> inode_space_stats;
+    uint64_t buffer_area_used_space = 0;
+    uint64_t data_used_space = 0;
+
+    // LSN queue: inflight (writing) -> completed [-> fsynced] -> compactable -> compacted [-> fsynced] -> trimmed and removed
+    std::deque<heap_inflight_lsn_t> inflight_lsn;
+    uint32_t to_compact_count = 0;
+    uint64_t first_inflight_lsn = 0;
+    uint64_t completed_lsn = 0;
+    uint64_t fsynced_lsn = 0;
+    uint64_t compacted_lsn = 0;
+    uint64_t next_compact_lsn = 0;
+    std::deque<heap_refqi_t> overwrite_ref_queue;
+
+    std::vector<tmp_compact_item_t> tmp_compact_queue;
+    std::deque<object_id> recheck_queue;
+    int recheck_in_progress = 0;
+    bool in_recheck = false;
+    std::function<void(bool is_data, uint64_t offset, uint64_t len, uint8_t* buf, std::function<void()>)> recheck_cb;
+    int recheck_queue_depth = 0;
+
+    const uint32_t max_write_entry_size;
+
+    uint64_t get_pg_id(inode_t inode, uint64_t stripe);
+    void defragment_block(uint32_t block_num);
+    uint32_t find_block_run(heap_block_info_t & block, uint32_t space);
+    uint32_t find_block_space(uint32_t block_num, uint32_t space);
+    uint32_t block_has_compactable(uint8_t *data);
+    uint32_t compact_object_to(heap_object_t *obj, uint64_t lsn, uint8_t *new_csums, bool do_free);
+    void copy_full_object(uint8_t *dst, heap_object_t *obj);
+    bool mvcc_save_copy(heap_object_t *obj);
+    bool mvcc_check_tracking(object_id oid);
+    void free_mvcc(heap_mvcc_map_t::iterator mvcc_it);
+    void allocate_block(heap_block_info_t & inf);
+    int allocate_new_object(object_id oid, uint32_t full_object_size, uint32_t *modified_block, heap_object_t **new_obj);
+    int add_object(object_id oid, heap_write_t *wr, uint32_t *modified_block);
+    void mark_overwritten(uint64_t over_lsn, uint64_t inode, heap_write_t *wr, heap_write_t *end_wr, bool tracking_active);
+    int update_object(uint32_t block_num, heap_object_t *obj, heap_write_t *wr, uint32_t *modified_block, uint32_t *moved_from_block);
+    void init_erase(uint32_t block_num, heap_object_t *obj);
+    void erase_object(uint32_t block_num, heap_object_t *obj, uint64_t lsn, bool tracking_active);
+    void reindex_block(uint32_t block_num, heap_object_t *from_obj);
+    void erase_block_index(inode_t inode, uint64_t stripe);
+    void deref_data(uint64_t inode, uint64_t location, bool free_at_0);
+    void deref_buffer(uint64_t inode, uint64_t location, uint32_t len, bool free_at_0);
+    void deref_overwrites(uint64_t lsn);
+    void free_object_space(inode_t inode, heap_write_t *from, heap_write_t *to, int mode = 0);
+    void add_used_space(uint32_t block_num, int32_t used_delta);
+    void push_inflight_lsn(object_id oid, uint64_t lsn, uint64_t flags);
+
+public:
+    blockstore_heap_t(blockstore_disk_t *dsk, uint8_t *buffer_area, int log_level = 0);
+    ~blockstore_heap_t();
+    // set initially compacted lsn - should be done before loading
+    void set_compacted_lsn(uint64_t compacted_lsn);
+    uint64_t get_compacted_lsn();
+    // load data from the disk, returns count of loaded write entries
+    void read_blocks(uint64_t disk_offset, uint64_t size, uint8_t *buf,
+        std::function<void(heap_object_t*)> handle_object, std::function<void(uint32_t, uint32_t, uint8_t*)> handle_block);
+    uint64_t load_blocks(uint64_t disk_offset, uint64_t size, uint8_t *buf);
+    // finish loading
+    void finish_load();
+    // recheck small write data after reading the database from disk
+    bool recheck_small_writes(std::function<void(bool is_data, uint64_t offset, uint64_t len, uint8_t* buf, std::function<void()>)> read_buffer, int queue_depth);
+    // initialize metadata area (fill it with empty data)
+    // returns 0 when done, EAGAIN when the caller has to wait more
+    int initialize();
+    // read from the metadata area
+    // returns 0 when done, EAGAIN when the caller has to wait more
+    int read();
+    // reshard database according to the pool's PG count
+    void reshard(pool_id_t pool, uint32_t pg_count, uint32_t pg_stripe_size);
+    // read an object entry and lock it against removal
+    // in the future, may become asynchronous
+    heap_object_t *lock_and_read_entry(object_id oid, uint64_t & copy_id);
+    // re-read a locked object entry with the given lsn (pointer may be invalidated)
+    heap_object_t *read_locked_entry(object_id oid, uint64_t copy_id);
+    // read an object entry without locking it
+    heap_object_t *read_entry(object_id oid, uint32_t *block_num_ptr, bool for_update = false);
+    // unlock an entry
+    bool unlock_entry(object_id oid, uint64_t copy_id);
+    // set or verify checksums in a write request
+    bool calc_checksums(heap_write_t *wr, uint8_t *data, bool set, uint32_t offset = 0, uint32_t len = 0);
+    // set or verify raw block checksums
+    bool calc_block_checksums(uint32_t *block_csums, uint8_t *data, uint8_t *bitmap, uint32_t start, uint32_t end,
+        bool set, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
+    bool calc_block_checksums(uint32_t *block_csums, uint8_t *bitmap,
+        uint32_t start, uint32_t end, std::function<uint8_t*(uint32_t start, uint32_t & len)> next,
+        bool set, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
+    // copy an object as is
+    int copy_object(heap_object_t *obj, uint32_t *modified_block);
+    // auto-compacts the object, then adds a write entry to it and to the compaction queue
+    // return 0 if OK, or maybe ENOSPC
+    int post_write(object_id oid, heap_write_t *wr, uint32_t *modified_block, uint32_t *moved_from_block);
+    int post_write(uint32_t & block_num, object_id oid, heap_object_t *obj, heap_write_t *wr, uint32_t *moved_from_block);
+    // stabilize an unstable object version
+    // return 0 if OK, ENOENT if not exists
+    int post_stabilize(object_id oid, uint64_t version, uint32_t *modified_block, uint64_t *new_lsn, uint64_t *new_to_lsn);
+    // rollback an unstable object version
+    // return 0 if OK, ENOENT if not exists, EBUSY if already stable
+    int post_rollback(object_id oid, uint64_t version, uint64_t *new_lsn, uint32_t *modified_block);
+    // forget an object
+    // return error code
+    int post_delete(object_id oid, uint64_t *new_lsn, uint32_t *modified_block);
+    int post_delete(uint32_t block_num, heap_object_t *obj, uint64_t *new_lsn);
+    // get the next object to compact
+    // guaranteed to return objects in min lsn order
+    // returns 0 if OK, ENOENT if nothing to compact
+    int get_next_compact(object_id & oid);
+    // get the range of an object eligible for compaction
+    void get_compact_range(heap_object_t *obj, uint64_t max_lsn, heap_write_t **begin_wr, heap_write_t **end_wr);
+    // mark an object as compacted up to the given lsn
+    int compact_object(object_id oid, uint64_t lsn, uint8_t *new_csums);
+    // retrieve object listing from a PG
+    int list_objects(uint32_t pg_num, object_id min_oid, object_id max_oid,
+        obj_ver_id **result_list, size_t *stable_count, size_t *unstable_count);
+    // set a block number for a new object and returns error status: 0, EAGAIN or ENOSPC
+    int get_block_for_new_object(uint32_t & out_block_num, uint32_t size = 0);
+
+    // inflight write tracking
+    void mark_lsn_completed(uint64_t lsn);
+    void mark_lsn_fsynced(uint64_t lsn);
+    void mark_lsn_compacted(uint64_t lsn, bool allow_undone = false);
+    void mark_object_compacted(heap_object_t *obj, uint64_t max_lsn);
+    void mark_lsn_trimmed(uint64_t lsn);
+    uint64_t get_completed_lsn();
+    uint64_t get_fsynced_lsn();
+
+    // data device block allocator functions
+    uint64_t find_free_data();
+    bool is_data_used(uint64_t location);
+    void use_data(inode_t inode, uint64_t location);
+    void free_data(inode_t inode, uint64_t location);
+
+    // buffer device allocator functions
+    uint64_t find_free_buffer_area(uint64_t size);
+    bool is_buffer_area_free(uint64_t location, uint64_t size);
+    void use_buffer_area(inode_t inode, uint64_t location, uint64_t size);
+    void free_buffer_area(inode_t inode, uint64_t location, uint64_t size);
+    uint64_t get_buffer_area_used_space();
+
+    // get metadata block data buffer and used space
+    uint8_t *get_meta_block(uint32_t block_num);
+    uint32_t get_meta_block_used_space(uint32_t block_num);
+
+    // get space usage statistics
+    uint64_t get_data_used_space();
+    const std::map<uint64_t, uint64_t> & get_inode_space_stats();
+    uint64_t get_meta_total_space();
+    uint64_t get_meta_used_space();
+    uint32_t get_meta_nearfull_blocks();
+    uint32_t get_inflight_queue_size();
+    uint32_t get_compact_queue_size();
+    uint32_t get_to_compact_count();
+
+    // get maximum size for a temporary heap_write_t buffer
+    uint32_t get_max_write_entry_size();
+
+    // only for tests
+    void set_abort_on_corruption(bool fail);
+    void set_abort_on_overlap(bool fail);
+};

src/blockstore/blockstore_impl.cpp

@@ -1,13 +1,18 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#include "blockstore_impl.h"
+#include <stdexcept>
 
-blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd)
+#include "blockstore_impl.h"
+#include "blockstore_internal.h"
+#include "crc32c.h"
+
+blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd, bool mock_mode)
 {
     assert(sizeof(blockstore_op_private_t) <= BS_OP_PRIVATE_DATA_SIZE);
     this->tfd = tfd;
     this->ringloop = ringloop;
+    dsk.mock_mode = mock_mode;
     ring_consumer.loop = [this]() { loop(); };
     ringloop->register_consumer(&ring_consumer);
     initialized = 0;
@@ -17,31 +22,43 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
-        calc_lengths();
-        alloc_dyn_data = dsk.clean_dyn_size > sizeof(void*) || dsk.csum_block_size > 0;
+        dsk.calc_lengths();
         zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.data_block_size);
-        data_alloc = new allocator_t(dsk.block_count);
     }
     catch (std::exception & e)
     {
         dsk.close_all();
         throw;
     }
+    memset(zero_object, 0, dsk.data_block_size);
+    meta_superblock = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.meta_block_size);
+    memset(meta_superblock, 0, dsk.meta_block_size);
+}
+
+void blockstore_impl_t::init()
+{
     flusher = new journal_flusher_t(this);
+    if (dsk.inmemory_journal)
+    {
+        buffer_area = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.journal_len);
+    }
+    heap = new blockstore_heap_t(&dsk, buffer_area, log_level);
 }
 
 blockstore_impl_t::~blockstore_impl_t()
 {
-    delete data_alloc;
-    delete flusher;
+    if (flusher)
+        delete flusher;
+    if (heap)
+        delete heap;
+    if (buffer_area)
+        free(buffer_area);
+    if (meta_superblock)
+        free(meta_superblock);
     if (zero_object)
         free(zero_object);
     ringloop->unregister_consumer(&ring_consumer);
     dsk.close_all();
-    if (metadata_buffer)
-        free(metadata_buffer);
-    if (clean_bitmaps)
-        free(clean_bitmaps);
 }
 
 bool blockstore_impl_t::is_started()
@@ -57,10 +74,9 @@ bool blockstore_impl_t::is_stalled()
 // main event loop - produce requests
 void blockstore_impl_t::loop()
 {
-    // FIXME: initialized == 10 is ugly
     if (initialized != 10)
     {
-        // read metadata, then journal
+        // read metadata
         if (initialized == 0)
         {
             metadata_init_reader = new blockstore_init_meta(this);
@@ -73,69 +89,41 @@ void blockstore_impl_t::loop()
             {
                 delete metadata_init_reader;
                 metadata_init_reader = NULL;
-                journal_init_reader = new blockstore_init_journal(this);
-                initialized = 2;
-            }
-        }
-        if (initialized == 2)
-        {
-            int res = journal_init_reader->loop();
-            if (!res)
-            {
-                delete journal_init_reader;
-                journal_init_reader = NULL;
                 initialized = 3;
-                ringloop->wakeup();
             }
         }
         if (initialized == 3)
         {
             if (!readonly && dsk.discard_on_start)
-                dsk.trim_data(data_alloc);
-            if (journal.flush_journal)
-                initialized = 4;
-            else
-                initialized = 10;
-        }
-        if (initialized == 4)
-        {
-            if (readonly)
             {
-                printf("Can't flush the journal in readonly mode\n");
-                exit(1);
+                dsk.trim_data([this](uint64_t block_num){ return heap->is_data_used(block_num * dsk.data_block_size); });
             }
-            flusher->loop();
-            ringloop->submit();
+            initialized = 10;
         }
     }
     else
     {
         // try to submit ops
         unsigned initial_ring_space = ringloop->space_left();
-        // has_writes == 0 - no writes before the current queue item
-        // has_writes == 1 - some writes in progress
-        // has_writes == 2 - tried to submit some writes, but failed
-        int has_writes = 0, op_idx = 0, new_idx = 0;
+        int op_idx = 0, new_idx = 0;
+        bool has_unfinished_writes = false;
         for (; op_idx < submit_queue.size(); op_idx++, new_idx++)
         {
             auto op = submit_queue[op_idx];
             submit_queue[new_idx] = op;
-            // FIXME: This needs some simplification
-            // Writes should not block reads if the ring is not full and reads don't depend on them
-            // In all other cases we should stop submission
             if (PRIV(op)->wait_for)
             {
                 check_wait(op);
                 if (PRIV(op)->wait_for == WAIT_SQE)
                 {
+                    // ring is full, stop submission
                     break;
                 }
                 else if (PRIV(op)->wait_for)
                 {
-                    if (op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE)
-                    {
-                        has_writes = 2;
-                    }
+                    has_unfinished_writes = has_unfinished_writes || op->opcode == BS_OP_WRITE ||
+                        op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE ||
+                        op->opcode == BS_OP_STABLE || op->opcode == BS_OP_ROLLBACK;
                     continue;
                 }
             }
@@ -148,46 +136,33 @@ void blockstore_impl_t::loop()
             {
                 wr_st = dequeue_read(op);
             }
-            else if (op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE)
+            else if (op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE)
             {
-                if (has_writes == 2)
-                {
-                    // Some writes already could not be submitted
-                    continue;
-                }
                 wr_st = dequeue_write(op);
-                has_writes = wr_st > 0 ? 1 : 2;
-            }
-            else if (op->opcode == BS_OP_DELETE)
-            {
-                if (has_writes == 2)
-                {
-                    // Some writes already could not be submitted
-                    continue;
-                }
-                wr_st = dequeue_del(op);
-                has_writes = wr_st > 0 ? 1 : 2;
+                has_unfinished_writes = has_unfinished_writes || (wr_st != 2);
             }
             else if (op->opcode == BS_OP_SYNC)
             {
-                // sync only completed writes?
-                // wait for the data device fsync to complete, then submit journal writes for big writes
-                // then submit an fsync operation
+                // syncs only completed writes, so doesn't have to be blocked by anything
                 wr_st = continue_sync(op);
             }
-            else if (op->opcode == BS_OP_STABLE)
+            else if (op->opcode == BS_OP_STABLE || op->opcode == BS_OP_ROLLBACK)
             {
                 wr_st = dequeue_stable(op);
-            }
-            else if (op->opcode == BS_OP_ROLLBACK)
-            {
-                wr_st = dequeue_rollback(op);
+                has_unfinished_writes = has_unfinished_writes || (wr_st != 2);
             }
             else if (op->opcode == BS_OP_LIST)
             {
-                // LIST doesn't have to be blocked by previous modifications
-                process_list(op);
-                wr_st = 2;
+                // LIST has to be blocked by previous writes and commits/rollbacks
+                if (!has_unfinished_writes)
+                {
+                    process_list(op);
+                    wr_st = 2;
+                }
+                else
+                {
+                    wr_st = 0;
+                }
             }
             if (wr_st == 2)
             {
@@ -196,16 +171,13 @@ void blockstore_impl_t::loop()
             }
             if (wr_st == 0)
             {
+                PRIV(op)->pending_ops = 0;
                 ringloop->restore(prev_sqe_pos);
                 if (PRIV(op)->wait_for == WAIT_SQE)
                 {
                     // ring is full, stop submission
                     break;
                 }
-                else if (PRIV(op)->wait_for == WAIT_JOURNAL)
-                {
-                    PRIV(op)->wait_detail2 = (unstable_writes.size()+unstable_unsynced);
-                }
             }
         }
         if (op_idx != new_idx)
@@ -225,12 +197,6 @@ void blockstore_impl_t::loop()
         {
             throw std::runtime_error(std::string("io_uring_submit: ") + strerror(-ret));
         }
-        for (auto s: journal.submitting_sectors)
-        {
-            // Mark journal sector writes as submitted
-            journal.sector_info[s].submit_id = 0;
-        }
-        journal.submitting_sectors.clear();
         if ((initial_ring_space - ringloop->space_left()) > 0)
         {
             live = true;
@@ -248,7 +214,7 @@ bool blockstore_impl_t::is_safe_to_stop()
     {
         return false;
     }
-    if (unsynced_big_writes.size() > 0 || unsynced_small_writes.size() > 0)
+    if (unsynced_big_write_count > 0 || unsynced_small_write_count > 0)
     {
         if (!readonly && !stop_sync_submitted)
         {
@@ -272,7 +238,7 @@ void blockstore_impl_t::check_wait(blockstore_op_t *op)
 {
     if (PRIV(op)->wait_for == WAIT_SQE)
     {
-        if (ringloop->sqes_left() < PRIV(op)->wait_detail)
+        if (ringloop->space_left() < PRIV(op)->wait_detail)
         {
             // stop submission if there's still no free space
 #ifdef BLOCKSTORE_DEBUG
@@ -282,40 +248,13 @@ void blockstore_impl_t::check_wait(blockstore_op_t *op)
         }
         PRIV(op)->wait_for = 0;
     }
-    else if (PRIV(op)->wait_for == WAIT_JOURNAL)
+    else if (PRIV(op)->wait_for == WAIT_COMPACTION)
     {
-        if (journal.used_start == PRIV(op)->wait_detail &&
-            (unstable_writes.size()+unstable_unsynced) == PRIV(op)->wait_detail2)
+        if (flusher->get_compact_counter() <= PRIV(op)->wait_detail)
         {
             // do not submit
 #ifdef BLOCKSTORE_DEBUG
-            printf("Still waiting to flush journal offset %08jx\n", PRIV(op)->wait_detail);
-#endif
-            return;
-        }
-        flusher->release_trim();
-        PRIV(op)->wait_for = 0;
-    }
-    else if (PRIV(op)->wait_for == WAIT_JOURNAL_BUFFER)
-    {
-        int next = ((journal.cur_sector + 1) % journal.sector_count);
-        if (journal.sector_info[next].flush_count > 0 ||
-            journal.sector_info[next].dirty)
-        {
-            // do not submit
-#ifdef BLOCKSTORE_DEBUG
-            printf("Still waiting for a journal buffer\n");
-#endif
-            return;
-        }
-        PRIV(op)->wait_for = 0;
-    }
-    else if (PRIV(op)->wait_for == WAIT_FREE)
-    {
-        if (!data_alloc->get_free_count() && big_to_flush > 0)
-        {
-#ifdef BLOCKSTORE_DEBUG
-            printf("Still waiting for free space on the data device\n");
+            printf("Still waiting for more flushes\n");
 #endif
             return;
         }
@@ -361,75 +300,11 @@ void blockstore_impl_t::init_op(blockstore_op_t *op)
 {
     // Call constructor without allocating memory. We'll call destructor before returning op back
     new ((void*)op->private_data) blockstore_op_private_t;
-    PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
     PRIV(op)->wait_for = 0;
     PRIV(op)->op_state = 0;
     PRIV(op)->pending_ops = 0;
 }
 
-static bool replace_stable(object_id oid, uint64_t version, int search_start, int search_end, obj_ver_id* list)
-{
-    while (search_start < search_end)
-    {
-        int pos = search_start+(search_end-search_start)/2;
-        if (oid < list[pos].oid)
-        {
-            search_end = pos;
-        }
-        else if (list[pos].oid < oid)
-        {
-            search_start = pos+1;
-        }
-        else
-        {
-            list[pos].version = version;
-            return true;
-        }
-    }
-    return false;
-}
-
-blockstore_clean_db_t& blockstore_impl_t::clean_db_shard(object_id oid)
-{
-    uint64_t pg_num = 0;
-    uint64_t pool_id = (oid.inode >> (64-POOL_ID_BITS));
-    auto sh_it = clean_db_settings.find(pool_id);
-    if (sh_it != clean_db_settings.end())
-    {
-        // like map_to_pg()
-        pg_num = (oid.stripe / sh_it->second.pg_stripe_size) % sh_it->second.pg_count + 1;
-    }
-    return clean_db_shards[(pool_id << (64-POOL_ID_BITS)) | pg_num];
-}
-
-void blockstore_impl_t::reshard_clean_db(pool_id_t pool, uint32_t pg_count, uint32_t pg_stripe_size)
-{
-    uint64_t pool_id = (uint64_t)pool;
-    std::map<pool_pg_id_t, blockstore_clean_db_t> new_shards;
-    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
-    while (sh_it != clean_db_shards.end() &&
-        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
-    {
-        for (auto & pair: sh_it->second)
-        {
-            // like map_to_pg()
-            uint64_t pg_num = (pair.first.stripe / pg_stripe_size) % pg_count + 1;
-            uint64_t shard_id = (pool_id << (64-POOL_ID_BITS)) | pg_num;
-            new_shards[shard_id][pair.first] = pair.second;
-        }
-        clean_db_shards.erase(sh_it++);
-    }
-    for (sh_it = new_shards.begin(); sh_it != new_shards.end(); sh_it++)
-    {
-        auto & to = clean_db_shards[sh_it->first];
-        to.swap(sh_it->second);
-    }
-    clean_db_settings[pool_id] = (pool_shard_settings_t){
-        .pg_count = pg_count,
-        .pg_stripe_size = pg_stripe_size,
-    };
-}
-
 void blockstore_impl_t::process_list(blockstore_op_t *op)
 {
     uint32_t list_pg = op->pg_number+1;
@@ -438,7 +313,8 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
     uint64_t min_inode = op->min_oid.inode;
     uint64_t max_inode = op->max_oid.inode;
     // Check PG
-    if (pg_count != 0 && (pg_stripe_size < MIN_DATA_BLOCK_SIZE || list_pg > pg_count))
+    if (!pg_count || (pg_stripe_size < MIN_DATA_BLOCK_SIZE || list_pg > pg_count) ||
+        !INODE_POOL(min_inode) || INODE_POOL(min_inode) != INODE_POOL(max_inode))
     {
         op->retval = -EINVAL;
         FINISH_OP(op);
@@ -446,250 +322,44 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
     }
     // Check if the DB needs resharding
     // (we don't know about PGs from the beginning, we only create "shards" here)
-    uint64_t first_shard = 0, last_shard = UINT64_MAX;
-    if (min_inode != 0 &&
-        // Check if min_inode == max_inode == pool_id<<N, i.e. this is a pool listing
-        (min_inode >> (64-POOL_ID_BITS)) == (max_inode >> (64-POOL_ID_BITS)))
+    heap->reshard(INODE_POOL(min_inode), pg_count, pg_stripe_size);
+    obj_ver_id *result = NULL;
+    size_t stable_count = 0, unstable_count = 0;
+    int res = heap->list_objects(list_pg, op->min_oid, op->max_oid, &result, &stable_count, &unstable_count);
+    if (op->list_stable_limit)
     {
-        pool_id_t pool_id = (min_inode >> (64-POOL_ID_BITS));
-        if (pg_count > 1)
+        // Ordered result is expected - used by scrub
+        // We use an unordered map
+        std::sort(result, result + stable_count);
+        if (stable_count > op->list_stable_limit)
         {
-            // Per-pg listing
-            auto sh_it = clean_db_settings.find(pool_id);
-            if (sh_it == clean_db_settings.end() ||
-                sh_it->second.pg_count != pg_count ||
-                sh_it->second.pg_stripe_size != pg_stripe_size)
-            {
-                reshard_clean_db(pool_id, pg_count, pg_stripe_size);
-            }
-            first_shard = last_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS)) | list_pg;
-        }
-        else
-        {
-            // Per-pool listing
-            first_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS));
-            last_shard = ((uint64_t)(pool_id+1) << (64-POOL_ID_BITS)) - 1;
+            memmove(result + op->list_stable_limit, result + stable_count, unstable_count);
+            stable_count = op->list_stable_limit;
         }
     }
-    // Copy clean_db entries
-    int stable_count = 0, stable_alloc = 0;
-    if (min_inode != max_inode)
-    {
-        for (auto shard_it = clean_db_shards.lower_bound(first_shard);
-            shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
-            shard_it++)
-        {
-            auto & clean_db = shard_it->second;
-            stable_alloc += clean_db.size();
-        }
-    }
-    if (op->list_stable_limit > 0)
-    {
-        stable_alloc = op->list_stable_limit;
-        if (stable_alloc > 1024*1024)
-            stable_alloc = 1024*1024;
-    }
-    if (stable_alloc < 32768)
-    {
-        stable_alloc = 32768;
-    }
-    obj_ver_id *stable = (obj_ver_id*)malloc(sizeof(obj_ver_id) * stable_alloc);
-    if (!stable)
-    {
-        op->retval = -ENOMEM;
-        FINISH_OP(op);
-        return;
-    }
-    auto max_oid = op->max_oid;
-    bool limited = false;
-    pool_pg_id_t last_shard_id = 0;
-    for (auto shard_it = clean_db_shards.lower_bound(first_shard);
-        shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
-        shard_it++)
-    {
-        auto & clean_db = shard_it->second;
-        auto clean_it = clean_db.begin(), clean_end = clean_db.end();
-        if (op->min_oid.inode != 0 || op->min_oid.stripe != 0)
-        {
-            clean_it = clean_db.lower_bound(op->min_oid);
-        }
-        if ((max_oid.inode != 0 || max_oid.stripe != 0) && !(max_oid < op->min_oid))
-        {
-            clean_end = clean_db.upper_bound(max_oid);
-        }
-        for (; clean_it != clean_end; clean_it++)
-        {
-            if (stable_count >= stable_alloc)
-            {
-                stable_alloc *= 2;
-                obj_ver_id* nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
-                if (!nst)
-                {
-                    op->retval = -ENOMEM;
-                    FINISH_OP(op);
-                    return;
-                }
-                stable = nst;
-            }
-            stable[stable_count++] = {
-                .oid = clean_it->first,
-                .version = clean_it->second.version,
-            };
-            if (op->list_stable_limit > 0 && stable_count >= op->list_stable_limit)
-            {
-                if (!limited)
-                {
-                    limited = true;
-                    max_oid = stable[stable_count-1].oid;
-                }
-                break;
-            }
-        }
-        if (op->list_stable_limit > 0)
-        {
-            // To maintain the order, we have to include objects in the same range from other shards
-            if (last_shard_id != 0 && last_shard_id != shard_it->first)
-                std::sort(stable, stable+stable_count);
-            if (stable_count > op->list_stable_limit)
-                stable_count = op->list_stable_limit;
-        }
-        last_shard_id = shard_it->first;
-    }
-    if (op->list_stable_limit == 0 && first_shard != last_shard)
-    {
-        // If that's not a per-PG listing, sort clean entries (already sorted if list_stable_limit != 0)
-        std::sort(stable, stable+stable_count);
-    }
-    int clean_stable_count = stable_count;
-    // Copy dirty_db entries (sorted, too)
-    int unstable_count = 0, unstable_alloc = 0;
-    obj_ver_id *unstable = NULL;
-    {
-        auto dirty_it = dirty_db.begin(), dirty_end = dirty_db.end();
-        if (op->min_oid.inode != 0 || op->min_oid.stripe != 0)
-        {
-            dirty_it = dirty_db.lower_bound({
-                .oid = op->min_oid,
-                .version = 0,
-            });
-        }
-        if ((max_oid.inode != 0 || max_oid.stripe != 0) && !(max_oid < op->min_oid))
-        {
-            dirty_end = dirty_db.upper_bound({
-                .oid = max_oid,
-                .version = UINT64_MAX,
-            });
-        }
-        for (; dirty_it != dirty_end; dirty_it++)
-        {
-            if (!pg_count || ((dirty_it->first.oid.stripe / pg_stripe_size) % pg_count + 1) == list_pg) // like map_to_pg()
-            {
-                if (IS_DELETE(dirty_it->second.state))
-                {
-                    // Deletions are always stable, so try to zero out two possible entries
-                    if (!replace_stable(dirty_it->first.oid, 0, 0, clean_stable_count, stable))
-                    {
-                        replace_stable(dirty_it->first.oid, 0, clean_stable_count, stable_count, stable);
-                    }
-                }
-                else if (IS_STABLE(dirty_it->second.state) || (dirty_it->second.state & BS_ST_INSTANT))
-                {
-                    // First try to replace a clean stable version in the first part of the list
-                    if (!replace_stable(dirty_it->first.oid, dirty_it->first.version, 0, clean_stable_count, stable))
-                    {
-                        // Then try to replace the last dirty stable version in the second part of the list
-                        if (stable_count > 0 && stable[stable_count-1].oid == dirty_it->first.oid)
-                        {
-                            stable[stable_count-1].version = dirty_it->first.version;
-                        }
-                        else
-                        {
-                            if (stable_count >= stable_alloc)
-                            {
-                                stable_alloc += 32768;
-                                obj_ver_id *nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
-                                if (!nst)
-                                {
-                                    if (unstable)
-                                        free(unstable);
-                                    op->retval = -ENOMEM;
-                                    FINISH_OP(op);
-                                    return;
-                                }
-                                stable = nst;
-                            }
-                            stable[stable_count++] = dirty_it->first;
-                        }
-                    }
-                    if (op->list_stable_limit > 0 && stable_count >= op->list_stable_limit)
-                    {
-                        // Stop here
-                        break;
-                    }
-                }
-                else
-                {
-                    if (unstable_count >= unstable_alloc)
-                    {
-                        unstable_alloc += 32768;
-                        obj_ver_id *nst = (obj_ver_id*)realloc(unstable, sizeof(obj_ver_id) * unstable_alloc);
-                        if (!nst)
-                        {
-                            if (stable)
-                                free(stable);
-                            op->retval = -ENOMEM;
-                            FINISH_OP(op);
-                            return;
-                        }
-                        unstable = nst;
-                    }
-                    unstable[unstable_count++] = dirty_it->first;
-                }
-            }
-        }
-    }
-    // Remove zeroed out stable entries
-    int j = 0;
-    for (int i = 0; i < stable_count; i++)
-    {
-        if (stable[i].version != 0)
-        {
-            stable[j++] = stable[i];
-        }
-    }
-    stable_count = j;
-    if (stable_count+unstable_count > stable_alloc)
-    {
-        stable_alloc = stable_count+unstable_count;
-        obj_ver_id *nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
-        if (!nst)
-        {
-            if (unstable)
-                free(unstable);
-            op->retval = -ENOMEM;
-            FINISH_OP(op);
-            return;
-        }
-        stable = nst;
-    }
-    // Copy unstable entries
-    for (int i = 0; i < unstable_count; i++)
-    {
-        stable[j++] = unstable[i];
-    }
-    free(unstable);
     op->version = stable_count;
-    op->retval = stable_count+unstable_count;
-    op->buf = stable;
+    op->retval = res == 0 ? stable_count+unstable_count : -res;
+    op->buf = (uint8_t*)result;
     FINISH_OP(op);
 }
 
+void blockstore_impl_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
+{
+}
+
 void blockstore_impl_t::dump_diagnostics()
 {
-    journal.dump_diagnostics();
     flusher->dump_diagnostics();
 }
 
+void blockstore_meta_header_v3_t::set_crc32c()
+{
+    header_csum = 0;
+    uint32_t calc = crc32c(0, this, version == BLOCKSTORE_META_FORMAT_HEAP
+        ? sizeof(blockstore_meta_header_v3_t) : sizeof(blockstore_meta_header_v2_t));
+    header_csum = calc;
+}
+
 void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expected)
 {
     if (retval == -EAGAIN)
@@ -703,85 +373,18 @@ void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expecte
     exit(1);
 }
 
-void blockstore_impl_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
+uint64_t blockstore_impl_t::get_free_block_count()
 {
-    for (auto & np: no_inode_stats)
-    {
-        np.second = 2;
-    }
-    for (auto pool_id: pool_ids)
-    {
-        if (!no_inode_stats[pool_id])
-            recalc_inode_space_stats(pool_id, false);
-        no_inode_stats[pool_id] = 1;
-    }
-    for (auto np_it = no_inode_stats.begin(); np_it != no_inode_stats.end(); )
-    {
-        if (np_it->second == 2)
-        {
-            recalc_inode_space_stats(np_it->first, true);
-            no_inode_stats.erase(np_it++);
-        }
-        else
-            np_it++;
-    }
+    return dsk.block_count - heap->get_data_used_space()/dsk.data_block_size;
 }
 
-void blockstore_impl_t::recalc_inode_space_stats(uint64_t pool_id, bool per_inode)
+std::string blockstore_impl_t::get_op_diag(blockstore_op_t *op)
 {
-    auto sp_begin = inode_space_stats.lower_bound((pool_id << (64-POOL_ID_BITS)));
-    auto sp_end = inode_space_stats.lower_bound(((pool_id+1) << (64-POOL_ID_BITS)));
-    inode_space_stats.erase(sp_begin, sp_end);
-    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
-    while (sh_it != clean_db_shards.end() &&
-        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
-    {
-        for (auto & pair: sh_it->second)
-        {
-            uint64_t space_id = per_inode ? pair.first.inode : (pool_id << (64-POOL_ID_BITS));
-            inode_space_stats[space_id] += dsk.data_block_size;
-        }
-        sh_it++;
-    }
-    object_id last_oid = {};
-    bool last_exists = false;
-    auto dirty_it = dirty_db.lower_bound((obj_ver_id){ .oid = { .inode = (pool_id << (64-POOL_ID_BITS)) } });
-    while (dirty_it != dirty_db.end() && (dirty_it->first.oid.inode >> (64-POOL_ID_BITS)) == pool_id)
-    {
-        if (IS_STABLE(dirty_it->second.state) && (IS_BIG_WRITE(dirty_it->second.state) || IS_DELETE(dirty_it->second.state)))
-        {
-            bool exists = false;
-            if (last_oid == dirty_it->first.oid)
-            {
-                exists = last_exists;
-            }
-            else
-            {
-                auto & clean_db = clean_db_shard(dirty_it->first.oid);
-                auto clean_it = clean_db.find(dirty_it->first.oid);
-                exists = clean_it != clean_db.end();
-            }
-            uint64_t space_id = per_inode ? dirty_it->first.oid.inode : (pool_id << (64-POOL_ID_BITS));
-            if (IS_BIG_WRITE(dirty_it->second.state))
-            {
-                if (!exists)
-                    inode_space_stats[space_id] += dsk.data_block_size;
-                last_exists = true;
-            }
-            else
-            {
-                if (exists)
-                {
-                    auto & sp = inode_space_stats[space_id];
-                    if (sp > dsk.data_block_size)
-                        sp -= dsk.data_block_size;
-                    else
-                        inode_space_stats.erase(space_id);
-                }
-                last_exists = false;
-            }
-            last_oid = dirty_it->first.oid;
-        }
-        dirty_it++;
-    }
+    char buf[256];
+    auto priv = PRIV(op);
+    if (priv->wait_for)
+        snprintf(buf, sizeof(buf), "state=%d wait=%d (detail=%ju)", priv->op_state, priv->wait_for, priv->wait_detail);
+    else
+        snprintf(buf, sizeof(buf), "state=%d", priv->op_state);
+    return std::string(buf);
 }

src/blockstore/blockstore_impl.h

@@ -5,6 +5,8 @@
 
 #include "blockstore.h"
 #include "blockstore_disk.h"
+#include "blockstore_heap.h"
+#include "ondisk_formats.h"
 
 #include <sys/types.h>
 #include <sys/ioctl.h>
@@ -21,240 +23,73 @@
 #include <unordered_map>
 #include <unordered_set>
 
-#include "cpp-btree/btree_map.h"
-
 #include "malloc_or_die.h"
-#include "allocator.h"
+
+class blockstore_impl_t;
 
 //#define BLOCKSTORE_DEBUG
 
-// States are not stored on disk. Instead, they're deduced from the journal
-
-#define BS_ST_SMALL_WRITE 0x01
-#define BS_ST_BIG_WRITE 0x02
-#define BS_ST_DELETE 0x03
-
-#define BS_ST_WAIT_DEL 0x10
-#define BS_ST_WAIT_BIG 0x20
-#define BS_ST_IN_FLIGHT 0x30
-#define BS_ST_SUBMITTED 0x40
-#define BS_ST_WRITTEN 0x50
-#define BS_ST_SYNCED 0x60
-#define BS_ST_STABLE 0x70
-
-#define BS_ST_INSTANT 0x100
-
-#define IMMEDIATE_NONE 0
-#define IMMEDIATE_SMALL 1
-#define IMMEDIATE_ALL 2
-
-#define BS_ST_TYPE_MASK 0x0F
-#define BS_ST_WORKFLOW_MASK 0xF0
-#define IS_IN_FLIGHT(st) (((st) & 0xF0) <= BS_ST_SUBMITTED)
-#define IS_STABLE(st) (((st) & 0xF0) == BS_ST_STABLE)
-#define IS_SYNCED(st) (((st) & 0xF0) >= BS_ST_SYNCED)
-#define IS_JOURNAL(st) (((st) & 0x0F) == BS_ST_SMALL_WRITE)
-#define IS_BIG_WRITE(st) (((st) & 0x0F) == BS_ST_BIG_WRITE)
-#define IS_DELETE(st) (((st) & 0x0F) == BS_ST_DELETE)
-#define IS_INSTANT(st) (((st) & BS_ST_TYPE_MASK) == BS_ST_DELETE || ((st) & BS_ST_INSTANT))
-
-#define BS_SUBMIT_CHECK_SQES(n) \
-    if (ringloop->sqes_left() < (n))\
-    {\
-        /* Pause until there are more requests available */\
-        PRIV(op)->wait_detail = (n);\
-        PRIV(op)->wait_for = WAIT_SQE;\
-        return 0;\
-    }
-
-#define BS_SUBMIT_GET_SQE(sqe, data) \
-    BS_SUBMIT_GET_ONLY_SQE(sqe); \
-    struct ring_data_t *data = ((ring_data_t*)sqe->user_data)
-
-#define BS_SUBMIT_GET_ONLY_SQE(sqe) \
-    struct io_uring_sqe *sqe = get_sqe();\
-    if (!sqe)\
-    {\
-        /* Pause until there are more requests available */\
-        PRIV(op)->wait_detail = 1;\
-        PRIV(op)->wait_for = WAIT_SQE;\
-        return 0;\
-    }
-
-#define BS_SUBMIT_GET_SQE_DECL(sqe) \
-    sqe = get_sqe();\
-    if (!sqe)\
-    {\
-        /* Pause until there are more requests available */\
-        PRIV(op)->wait_detail = 1;\
-        PRIV(op)->wait_for = WAIT_SQE;\
-        return 0;\
-    }
-
-#include "blockstore_journal.h"
-
-// "VITAstor"
-#define BLOCKSTORE_META_MAGIC_V1 0x726F747341544956l
-#define BLOCKSTORE_META_FORMAT_V1 1
-#define BLOCKSTORE_META_FORMAT_V2 2
-
-// metadata header (superblock)
-struct __attribute__((__packed__)) blockstore_meta_header_v1_t
-{
-    uint64_t zero;
-    uint64_t magic;
-    uint64_t version;
-    uint32_t meta_block_size;
-    uint32_t data_block_size;
-    uint32_t bitmap_granularity;
-};
-
-struct __attribute__((__packed__)) blockstore_meta_header_v2_t
-{
-    uint64_t zero;
-    uint64_t magic;
-    uint64_t version;
-    uint32_t meta_block_size;
-    uint32_t data_block_size;
-    uint32_t bitmap_granularity;
-    uint32_t data_csum_type;
-    uint32_t csum_block_size;
-    uint32_t header_csum;
-};
-
-// 32 bytes = 24 bytes + block bitmap (4 bytes by default) + external attributes (also bitmap, 4 bytes by default)
-// per "clean" entry on disk with fixed metadata tables
-struct __attribute__((__packed__)) clean_disk_entry
-{
-    object_id oid;
-    uint64_t version;
-    uint8_t bitmap[];
-    // Two more fields come after bitmap in metadata version 2:
-    // uint32_t data_csum[];
-    // uint32_t entry_csum;
-};
-
-// 32 = 16 + 16 bytes per "clean" entry in memory (object_id => clean_entry)
-struct __attribute__((__packed__)) clean_entry
-{
-    uint64_t version;
-    uint64_t location;
-};
-
-// 64 = 24 + 40 bytes per dirty entry in memory (obj_ver_id => dirty_entry). Plus checksums
-struct __attribute__((__packed__)) dirty_entry
-{
-    uint32_t state;
-    uint32_t flags;    // unneeded, but present for alignment
-    uint64_t location; // location in either journal or data -> in BYTES
-    uint32_t offset;   // data offset within object (stripe)
-    uint32_t len;      // data length
-    uint64_t journal_sector; // journal sector used for this entry
-    void* dyn_data;    // dynamic data: external bitmap and data block checksums. may be a pointer to the in-memory journal
-};
-
 // - Sync must be submitted after previous writes/deletes (not before!)
-// - Reads to the same object must be submitted after previous writes/deletes
-//   are written (not necessarily synced) in their location. This is because we
-//   rely on read-modify-write for erasure coding and we must return new data
-//   to calculate parity for subsequent writes
+// - Reads may be submitted in parallel with writes/deletes because we use MVCC
 // - Writes may be submitted in any order, because they don't overlap. Each write
 //   goes into a new location - either on the journal device or on the data device
 // - Stable (stabilize) must be submitted after sync of that object is completed
 //   It's even OK to return an error to the caller if that object is not synced yet
-// - Journal trim may be processed only after all versions are moved to
-//   the main storage AND after all read operations for older versions complete
+// - compacted_lsn should be moved forward only after all versions are moved to the main storage
 // - If an operation can not be submitted because the ring is full
 //   we should stop submission of other operations. Otherwise some "scatter" reads
 //   may end up blocked for a long time.
-// Otherwise, the submit order is free, that is all operations may be submitted immediately
-// In fact, adding a write operation must immediately result in dirty_db being populated
-
-// Suspend operation until there are more free SQEs
-#define WAIT_SQE 1
-// Suspend operation until there are <wait_detail> bytes of free space in the journal on disk
-#define WAIT_JOURNAL 3
-// Suspend operation until the next journal sector buffer is free
-#define WAIT_JOURNAL_BUFFER 4
-// Suspend operation until there is some free space on the data device
-#define WAIT_FREE 5
-
-struct used_clean_obj_t
-{
-    int refs;
-    bool was_freed; // was freed by a parallel flush?
-    bool was_changed; // was changed by a parallel flush?
-};
-
-// https://github.com/algorithm-ninja/cpp-btree
-// https://github.com/greg7mdp/sparsepp/ was used previously, but it was TERRIBLY slow after resizing
-// with sparsepp, random reads dropped to ~700 iops very fast with just as much as ~32k objects in the DB
-typedef btree::btree_map<object_id, clean_entry> blockstore_clean_db_t;
-typedef std::map<obj_ver_id, dirty_entry> blockstore_dirty_db_t;
+// Otherwise, the submission order is free.
 
 #include "blockstore_init.h"
 
 #include "blockstore_flush.h"
 
-#define PRIV(op) ((blockstore_op_private_t*)(op)->private_data)
-#define FINISH_OP(op) PRIV(op)->~blockstore_op_private_t(); std::function<void (blockstore_op_t*)>(op->callback)(op)
-
 struct blockstore_op_private_t
 {
     // Wait status
     int wait_for;
-    uint64_t wait_detail, wait_detail2;
+    uint64_t wait_detail;
     int pending_ops;
     int op_state;
 
+    // Read, write, sync, stabilize
+    uint64_t lsn;
+
     // Read
-    uint64_t clean_block_used;
     std::vector<copy_buffer_t> read_vec;
 
-    // Sync, write
-    uint64_t min_flushed_journal_sector, max_flushed_journal_sector;
+    // Write
+    uint64_t location;
+    bool is_big;
+
+    // Stabilize, rollback
+    int stab_pos;
+
+    // Stabilize
+    uint64_t to_lsn;
 
     // Write
     struct iovec iov_zerofill[3];
-    // Warning: must not have a default value here because it's written to before calling constructor in blockstore_write.cpp O_o
-    uint64_t real_version;
     timespec tv_begin;
-
-    // Sync
-    std::vector<obj_ver_id> sync_big_writes, sync_small_writes;
 };
 
-typedef uint32_t pool_id_t;
-typedef uint64_t pool_pg_id_t;
-
-#define POOL_ID_BITS 16
-
-struct pool_shard_settings_t
-{
-    uint32_t pg_count;
-    uint32_t pg_stripe_size;
-};
-
-#define STAB_SPLIT_DONE 1
-#define STAB_SPLIT_WAIT 2
-#define STAB_SPLIT_SYNC 3
-#define STAB_SPLIT_TODO 4
-
-class blockstore_impl_t
+class blockstore_impl_t: public blockstore_i
 {
+public:
     blockstore_disk_t dsk;
 
     /******* OPTIONS *******/
     bool readonly = false;
-    // It is safe to disable fsync() if drive write cache is writethrough
-    bool disable_data_fsync = false, disable_meta_fsync = false, disable_journal_fsync = false;
     // Enable if you want every operation to be executed with an "implicit fsync"
     // Suitable only for server SSDs with capacitors, requires disabled data and journal fsyncs
     int immediate_commit = IMMEDIATE_NONE;
     bool inmemory_meta = false;
+    uint32_t meta_write_recheck_parallelism = 0;
     // Maximum and minimum flusher count
-    unsigned max_flusher_count, min_flusher_count;
-    unsigned journal_trim_interval;
+    unsigned max_flusher_count = 0, min_flusher_count = 0;
+    unsigned journal_trim_interval = 0;
+    unsigned flusher_start_threshold = 0;
     // Maximum queue depth
     unsigned max_write_iodepth = 128;
     // Enable small (journaled) write throttling, useful for the SSD+HDD case
@@ -269,141 +104,92 @@ class blockstore_impl_t
     uint64_t autosync_writes = 128;
     // Log level (0-10)
     int log_level = 0;
+    // Enable correct block checksum validation on objects updated with small writes when checksum block
+    // is larger than bitmap_granularity, at the expense of extra metadata fsyncs during compaction
+    bool perfect_csum_update = false;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;
 
-    std::map<pool_id_t, pool_shard_settings_t> clean_db_settings;
-    std::map<pool_pg_id_t, blockstore_clean_db_t> clean_db_shards;
-    std::map<uint64_t, int> no_inode_stats;
-    uint8_t *clean_bitmaps = NULL;
-    blockstore_dirty_db_t dirty_db;
+    blockstore_heap_t *heap = NULL;
+    uint8_t* meta_superblock = NULL;
+    uint8_t *buffer_area = NULL;
     std::vector<blockstore_op_t*> submit_queue;
-    std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
-    int unsynced_big_write_count = 0, unstable_unsynced = 0;
+    int unsynced_big_write_count = 0, unsynced_small_write_count = 0, unsynced_meta_write_count = 0;
     int unsynced_queued_ops = 0;
-    allocator_t *data_alloc = NULL;
-    uint64_t used_blocks = 0;
     uint8_t *zero_object = NULL;
 
-    void *metadata_buffer = NULL;
-
-    struct journal_t journal;
     journal_flusher_t *flusher;
-    int big_to_flush = 0;
     int write_iodepth = 0;
-    bool alloc_dyn_data = false;
-
-    // clean data blocks referenced by read operations
-    std::map<uint64_t, used_clean_obj_t> used_clean_objects;
+    int inflight_big = 0;
+    bool fsyncing_data = false;
 
     bool live = false, queue_stall = false;
-    ring_loop_t *ringloop;
-    timerfd_manager_t *tfd;
+    ring_loop_i *ringloop = NULL;
+    timerfd_manager_t *tfd = NULL;
 
-    bool stop_sync_submitted;
+    bool stop_sync_submitted = false;
 
     inline struct io_uring_sqe* get_sqe()
     {
         return ringloop->get_sqe();
     }
 
-    friend class blockstore_init_meta;
-    friend class blockstore_init_journal;
-    friend struct blockstore_journal_check_t;
-    friend class journal_flusher_t;
-    friend class journal_flusher_co;
-
-    void calc_lengths();
     void open_data();
     void open_meta();
     void open_journal();
-    uint8_t* get_clean_entry_bitmap(uint64_t block_loc, int offset);
 
-    blockstore_clean_db_t& clean_db_shard(object_id oid);
-    void reshard_clean_db(pool_id_t pool_id, uint32_t pg_count, uint32_t pg_stripe_size);
-    void recalc_inode_space_stats(uint64_t pool_id, bool per_inode);
-
-    // Journaling
-    void prepare_journal_sector_write(int sector, blockstore_op_t *op);
-    void handle_journal_write(ring_data_t *data, uint64_t flush_id);
     void disk_error_abort(const char *op, int retval, int expected);
 
     // Asynchronous init
     int initialized;
     int metadata_buf_size;
     blockstore_init_meta* metadata_init_reader;
-    blockstore_init_journal* journal_init_reader;
 
+    void init();
     void check_wait(blockstore_op_t *op);
     void init_op(blockstore_op_t *op);
 
     // Read
-    int dequeue_read(blockstore_op_t *read_op);
+    int dequeue_read(blockstore_op_t *op);
+    int fulfill_read(blockstore_op_t *op);
+    uint32_t prepare_read(std::vector<copy_buffer_t> & read_vec, heap_object_t *obj, heap_write_t *wr, uint32_t start, uint32_t end);
+    uint32_t prepare_read_with_bitmaps(std::vector<copy_buffer_t> & read_vec, heap_object_t *obj, heap_write_t *wr, uint32_t start, uint32_t end);
+    uint32_t prepare_read_zero(std::vector<copy_buffer_t> & read_vec, uint32_t start, uint32_t end);
+    uint32_t prepare_read_simple(std::vector<copy_buffer_t> & read_vec, heap_object_t *obj, heap_write_t *wr, uint32_t start, uint32_t end);
+    void prepare_disk_read(std::vector<copy_buffer_t> & read_vec, int pos, heap_object_t *obj, heap_write_t *wr,
+        uint32_t blk_start, uint32_t blk_end, uint32_t start, uint32_t end, uint32_t copy_flags);
     void find_holes(std::vector<copy_buffer_t> & read_vec, uint32_t item_start, uint32_t item_end,
-        std::function<int(int, bool, uint32_t, uint32_t)> callback);
-    int fulfill_read(blockstore_op_t *read_op,
-        uint64_t &fulfilled, uint32_t item_start, uint32_t item_end,
-        uint32_t item_state, uint64_t item_version, uint64_t item_location,
-        uint64_t journal_sector, uint8_t *csum, int *dyn_data);
-    bool fulfill_clean_read(blockstore_op_t *read_op, uint64_t & fulfilled,
-        uint8_t *clean_entry_bitmap, int *dyn_data,
-        uint32_t item_start, uint32_t item_end, uint64_t clean_loc, uint64_t clean_ver);
-    int fill_partial_checksum_blocks(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled,
-        uint8_t *clean_entry_bitmap, int *dyn_data, bool from_journal, uint8_t *read_buf, uint64_t read_offset, uint64_t read_end);
-    int pad_journal_read(std::vector<copy_buffer_t> & rv, copy_buffer_t & cp,
-        uint64_t dirty_offset, uint64_t dirty_end, uint64_t dirty_loc, uint8_t *csum_ptr, int *dyn_data,
-        uint64_t offset, uint64_t submit_len, uint64_t & blk_begin, uint64_t & blk_end, uint8_t* & blk_buf);
-    bool read_range_fulfilled(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled, uint8_t *read_buf,
-        uint8_t *clean_entry_bitmap, uint32_t item_start, uint32_t item_end);
-    bool read_checksum_block(blockstore_op_t *op, int rv_pos, uint64_t &fulfilled, uint64_t clean_loc);
-    uint8_t* read_clean_meta_block(blockstore_op_t *read_op, uint64_t clean_loc, int rv_pos);
-    bool verify_padded_checksums(uint8_t *clean_entry_bitmap, uint8_t *csum_buf, uint32_t offset,
-        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
-    bool verify_journal_checksums(uint8_t *csums, uint32_t offset,
-        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
-    bool verify_clean_padded_checksums(blockstore_op_t *op, uint64_t clean_loc, uint8_t *dyn_data, bool from_journal,
-        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
-    int fulfill_read_push(blockstore_op_t *op, void *buf, uint64_t offset, uint64_t len,
-        uint32_t item_state, uint64_t item_version);
+        std::function<void(int&, uint32_t, uint32_t)> callback);
+    void free_read_buffers(std::vector<copy_buffer_t> & rv);
     void handle_read_event(ring_data_t *data, blockstore_op_t *op);
+    bool verify_read_checksums(blockstore_op_t *op);
 
     // Write
     bool enqueue_write(blockstore_op_t *op);
-    void cancel_all_writes(blockstore_op_t *op, blockstore_dirty_db_t::iterator dirty_it, int retval);
+    void prepare_meta_block_write(blockstore_op_t *op, uint64_t modified_block, io_uring_sqe *sqe = NULL);
     int dequeue_write(blockstore_op_t *op);
-    int dequeue_del(blockstore_op_t *op);
+    int make_big_write(blockstore_op_t *op, uint32_t offset, uint32_t len, uint32_t *modified_block, uint32_t *moved_from_block);
     int continue_write(blockstore_op_t *op);
-    void release_journal_sectors(blockstore_op_t *op);
     void handle_write_event(ring_data_t *data, blockstore_op_t *op);
 
     // Sync
     int continue_sync(blockstore_op_t *op);
-    void ack_sync(blockstore_op_t *op);
+    bool submit_fsyncs(int & wait_count);
+    int do_sync(blockstore_op_t *op, int base_state);
 
     // Stabilize
     int dequeue_stable(blockstore_op_t *op);
-    int continue_stable(blockstore_op_t *op);
-    void mark_stable(obj_ver_id ov, bool forget_dirty = false);
-    void stabilize_object(object_id oid, uint64_t max_ver);
-    blockstore_op_t* selective_sync(blockstore_op_t *op);
-    int split_stab_op(blockstore_op_t *op, std::function<int(obj_ver_id v)> decider);
-
-    // Rollback
-    int dequeue_rollback(blockstore_op_t *op);
-    int continue_rollback(blockstore_op_t *op);
-    void mark_rolled_back(const obj_ver_id & ov);
-    void erase_dirty(blockstore_dirty_db_t::iterator dirty_start, blockstore_dirty_db_t::iterator dirty_end, uint64_t clean_loc);
-    void free_dirty_dyn_data(dirty_entry & e);
 
     // List
     void process_list(blockstore_op_t *op);
 
-public:
+/*public:*/
 
-    blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd);
+    blockstore_impl_t(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd, bool mock_mode = false);
     ~blockstore_impl_t();
 
+    void parse_config(blockstore_config_t & config);
     void parse_config(blockstore_config_t & config, bool init);
 
     // Event loop
@@ -427,21 +213,19 @@ public:
     // Simplified synchronous operation: get object bitmap & current version
     int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL);
 
-    // Unstable writes are added here (map of object_id -> version)
-    std::unordered_map<object_id, uint64_t> unstable_writes;
-
-    // Space usage statistics
-    std::map<uint64_t, uint64_t> inode_space_stats;
-
     // Set per-pool no_inode_stats
     void set_no_inode_stats(const std::vector<uint64_t> & pool_ids);
 
     // Print diagnostics to stdout
     void dump_diagnostics();
 
+    // Get diagnostic string for an operation
+    std::string get_op_diag(blockstore_op_t *op);
+
+    const std::map<uint64_t, uint64_t> & get_inode_space_stats() { return heap->get_inode_space_stats(); }
     inline uint32_t get_block_size() { return dsk.data_block_size; }
     inline uint64_t get_block_count() { return dsk.block_count; }
-    inline uint64_t get_free_block_count() { return dsk.block_count - used_blocks; }
+    uint64_t get_free_block_count();
     inline uint32_t get_bitmap_granularity() { return dsk.disk_alignment; }
     inline uint64_t get_journal_size() { return dsk.journal_len; }
 };

src/blockstore/blockstore_init.h

@@ -25,47 +25,10 @@ class blockstore_init_meta
     uint64_t next_offset = 0;
     uint64_t last_read_offset = 0;
     uint64_t entries_loaded = 0;
-    unsigned entries_per_block = 0;
     int i = 0, j = 0;
-    std::vector<uint64_t> entries_to_zero;
     bool handle_meta_block(uint8_t *buf, uint64_t count, uint64_t done_cnt);
     void handle_event(ring_data_t *data, int buf_num);
 public:
     blockstore_init_meta(blockstore_impl_t *bs);
     int loop();
 };
-
-struct bs_init_journal_done
-{
-    void *buf;
-    uint64_t pos, len;
-};
-
-class blockstore_init_journal
-{
-    blockstore_impl_t *bs;
-    int wait_state = 0, wait_count = 0, handle_res = 0;
-    uint64_t entries_loaded = 0;
-    uint32_t crc32_last = 0;
-    bool started = false;
-    uint64_t next_free;
-    std::vector<bs_init_journal_done> done;
-    std::vector<obj_ver_id> double_allocs;
-    std::vector<iovec> small_write_data;
-    uint64_t journal_pos = 0;
-    uint64_t continue_pos = 0;
-    void *init_write_buf = NULL;
-    uint64_t init_write_sector = 0;
-    bool wrapped = false;
-    void *submitted_buf;
-    struct io_uring_sqe *sqe;
-    struct ring_data_t *data;
-    journal_entry_start *je_start;
-    std::function<void(ring_data_t*)> simple_callback;
-    int handle_journal_part(void *buf, uint64_t done_pos, uint64_t len);
-    void handle_event(ring_data_t *data);
-    void erase_dirty_object(blockstore_dirty_db_t::iterator dirty_it);
-public:
-    blockstore_init_journal(blockstore_impl_t* bs);
-    int loop();
-};

src/blockstore/blockstore_internal.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#define BS_SUBMIT_CHECK_SQES(n) \
+    if (ringloop->space_left() < (n))\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = (n);\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define BS_SUBMIT_GET_SQE(sqe, data) \
+    BS_SUBMIT_GET_ONLY_SQE(sqe); \
+    struct ring_data_t *data = ((ring_data_t*)sqe->user_data)
+
+#define BS_SUBMIT_GET_ONLY_SQE(sqe) \
+    struct io_uring_sqe *sqe = get_sqe();\
+    if (!sqe)\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define BS_SUBMIT_GET_SQE_DECL(sqe) \
+    sqe = get_sqe();\
+    if (!sqe)\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define PRIV(op) ((blockstore_op_private_t*)(op)->private_data)
+#define FINISH_OP(op) PRIV(op)->~blockstore_op_private_t(); std::function<void (blockstore_op_t*)>(op->callback)(op)
+
+// Suspend operation until there are more free SQEs
+#define WAIT_SQE 1
+// Suspend operation until there are <wait_detail> bytes of free space in the journal on disk
+#define WAIT_COMPACTION 2
+
+#define COPY_BUF_JOURNAL    0x01
+#define COPY_BUF_DATA       0x02
+#define COPY_BUF_ZERO       0x04
+#define COPY_BUF_CSUM_FILL  0x08
+#define COPY_BUF_COALESCED  0x10
+#define COPY_BUF_PADDED     0x20
+#define COPY_BUF_SKIP_CSUM  0x40

src/blockstore/blockstore_open.cpp

@@ -2,8 +2,14 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include <sys/file.h>
+#include <stdexcept>
 #include "blockstore_impl.h"
 
+void blockstore_impl_t::parse_config(blockstore_config_t & config)
+{
+    return parse_config(config, false);
+}
+
 void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
 {
     // Online-configurable options:
@@ -14,12 +20,14 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     }
     min_flusher_count = strtoull(config["min_flusher_count"].c_str(), NULL, 10);
     journal_trim_interval = strtoull(config["journal_trim_interval"].c_str(), NULL, 10);
+    flusher_start_threshold = strtoull(config["flusher_start_threshold"].c_str(), NULL, 10);
     max_write_iodepth = strtoull(config["max_write_iodepth"].c_str(), NULL, 10);
     throttle_small_writes = config["throttle_small_writes"] == "true" || config["throttle_small_writes"] == "1" || config["throttle_small_writes"] == "yes";
     throttle_target_iops = strtoull(config["throttle_target_iops"].c_str(), NULL, 10);
     throttle_target_mbs = strtoull(config["throttle_target_mbs"].c_str(), NULL, 10);
     throttle_target_parallelism = strtoull(config["throttle_target_parallelism"].c_str(), NULL, 10);
     throttle_threshold_us = strtoull(config["throttle_threshold_us"].c_str(), NULL, 10);
+    perfect_csum_update = config["perfect_csum_update"] == "true" || config["perfect_csum_update"] == "1" || config["perfect_csum_update"] == "yes";
     if (config["autosync_writes"] != "")
     {
         autosync_writes = strtoull(config["autosync_writes"].c_str(), NULL, 10);
@@ -28,13 +36,17 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     {
         max_flusher_count = 256;
     }
-    if (!min_flusher_count || journal.flush_journal)
+    if (!min_flusher_count)
     {
         min_flusher_count = 1;
     }
     if (!journal_trim_interval)
     {
-        journal_trim_interval = 512;
+        journal_trim_interval = 1024;
+    }
+    if (!flusher_start_threshold)
+    {
+        flusher_start_threshold = 32;
     }
     if (!max_write_iodepth)
     {
@@ -68,23 +80,6 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     {
         readonly = true;
     }
-    if (config["disable_data_fsync"] == "true" || config["disable_data_fsync"] == "1" || config["disable_data_fsync"] == "yes")
-    {
-        disable_data_fsync = true;
-    }
-    if (config["disable_meta_fsync"] == "true" || config["disable_meta_fsync"] == "1" || config["disable_meta_fsync"] == "yes")
-    {
-        disable_meta_fsync = true;
-    }
-    if (config["disable_journal_fsync"] == "true" || config["disable_journal_fsync"] == "1" || config["disable_journal_fsync"] == "yes")
-    {
-        disable_journal_fsync = true;
-    }
-    if (config["flush_journal"] == "true" || config["flush_journal"] == "1" || config["flush_journal"] == "yes")
-    {
-        // Only flush journal and exit
-        journal.flush_journal = true;
-    }
     if (config["immediate_commit"] == "all")
     {
         immediate_commit = IMMEDIATE_ALL;
@@ -94,85 +89,23 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
         immediate_commit = IMMEDIATE_SMALL;
     }
     metadata_buf_size = strtoull(config["meta_buf_size"].c_str(), NULL, 10);
-    inmemory_meta = config["inmemory_metadata"] != "false" && config["inmemory_metadata"] != "0" &&
-        config["inmemory_metadata"] != "no";
-    journal.sector_count = strtoull(config["journal_sector_buffer_count"].c_str(), NULL, 10);
-    journal.no_same_sector_overwrites = config["journal_no_same_sector_overwrites"] == "true" ||
-        config["journal_no_same_sector_overwrites"] == "1" || config["journal_no_same_sector_overwrites"] == "yes";
-    journal.inmemory = config["inmemory_journal"] != "false" && config["inmemory_journal"] != "0" &&
-        config["inmemory_journal"] != "no";
+    meta_write_recheck_parallelism = strtoull(config["meta_write_recheck_parallelism"].c_str(), NULL, 10);
     log_level = strtoull(config["log_level"].c_str(), NULL, 10);
     // Validate
-    if (journal.sector_count < 2)
-    {
-        journal.sector_count = 32;
-    }
     if (metadata_buf_size < 65536)
     {
         metadata_buf_size = 4*1024*1024;
     }
-    if (dsk.meta_device == dsk.data_device)
+    if (!meta_write_recheck_parallelism)
     {
-        disable_meta_fsync = disable_data_fsync;
+        meta_write_recheck_parallelism = 16;
     }
-    if (dsk.journal_device == dsk.meta_device)
-    {
-        disable_journal_fsync = disable_meta_fsync;
-    }
-    if (immediate_commit != IMMEDIATE_NONE && !disable_journal_fsync)
+    if (immediate_commit != IMMEDIATE_NONE && !dsk.disable_journal_fsync)
     {
         throw std::runtime_error("immediate_commit requires disable_journal_fsync");
     }
-    if (immediate_commit == IMMEDIATE_ALL && !disable_data_fsync)
+    if (immediate_commit == IMMEDIATE_ALL && !dsk.disable_data_fsync)
     {
         throw std::runtime_error("immediate_commit=all requires disable_journal_fsync and disable_data_fsync");
     }
-    // init some fields
-    journal.block_size = dsk.journal_block_size;
-    journal.next_free = dsk.journal_block_size;
-    journal.used_start = dsk.journal_block_size;
-    // no free space because sector is initially unmapped
-    journal.in_sector_pos = dsk.journal_block_size;
-}
-
-void blockstore_impl_t::calc_lengths()
-{
-    dsk.calc_lengths();
-    journal.len = dsk.journal_len;
-    journal.block_size = dsk.journal_block_size;
-    journal.offset = dsk.journal_offset;
-    if (inmemory_meta)
-    {
-        metadata_buffer = memalign(MEM_ALIGNMENT, dsk.meta_len);
-        if (!metadata_buffer)
-            throw std::runtime_error("Failed to allocate memory for the metadata ("+std::to_string(dsk.meta_len/1024/1024)+" MB)");
-    }
-    else if (dsk.clean_entry_bitmap_size || dsk.data_csum_type)
-    {
-        clean_bitmaps = (uint8_t*)malloc(dsk.block_count * 2 * dsk.clean_entry_bitmap_size);
-        if (!clean_bitmaps)
-        {
-            throw std::runtime_error(
-                "Failed to allocate memory for the metadata sparse write bitmap ("+
-                std::to_string(dsk.block_count * 2 * dsk.clean_entry_bitmap_size / 1024 / 1024)+" MB)"
-            );
-        }
-    }
-    if (journal.inmemory)
-    {
-        journal.buffer = memalign(MEM_ALIGNMENT, journal.len);
-        if (!journal.buffer)
-            throw std::runtime_error("Failed to allocate memory for journal ("+std::to_string(journal.len/1024/1024)+" MB)");
-    }
-    else
-    {
-        journal.sector_buf = (uint8_t*)memalign(MEM_ALIGNMENT, journal.sector_count * dsk.journal_block_size);
-        if (!journal.sector_buf)
-            throw std::bad_alloc();
-    }
-    journal.sector_info = (journal_sector_info_t*)calloc(journal.sector_count, sizeof(journal_sector_info_t));
-    if (!journal.sector_info)
-    {
-        throw std::bad_alloc();
-    }
 }

src/blockstore/blockstore_stable.cpp

@@ -2,560 +2,89 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include "blockstore_impl.h"
+#include "blockstore_internal.h"
 
-// Stabilize small write:
-// 1) Copy data from the journal to the data device
-// 2) Increase version on the metadata device and sync it
-// 3) Advance clean_db entry's version, clear previous journal entries
-//
-// This makes 1 4K small write+sync look like:
-// 512b+4K (journal) + sync + 512b (journal) + sync + 4K (data) [+ sync?] + 512b (metadata) + sync.
-// WA = 2.375. It's not the best, SSD FTL-like redirect-write could probably be lower
-// even with defragmentation. But it's fixed and it's still better than in Ceph. :)
-// except for HDD-only clusters, because each write results in 3 seeks.
-
-// Stabilize big write:
-// 1) Copy metadata from the journal to the metadata device
-// 2) Move dirty_db entry to clean_db and clear previous journal entries
-//
-// This makes 1 128K big write+sync look like:
-// 128K (data) + sync + 512b (journal) + sync + 512b (journal) + sync + 512b (metadata) + sync.
-// WA = 1.012. Very good :)
-
-// Stabilize delete:
-// 1) Remove metadata entry and sync it
-// 2) Remove dirty_db entry and clear previous journal entries
-// We have 2 problems here:
-// - In the cluster environment, we must store the "tombstones" of deleted objects until
-//   all replicas (not just quorum) agrees about their deletion. That is, "stabilize" is
-//   not possible for deletes in degraded placement groups
-// - With simple "fixed" metadata tables we can't just clear the metadata entry of the latest
-//   object version. We must clear all previous entries, too.
-// FIXME Fix both problems - probably, by switching from "fixed" metadata tables to "dynamic"
-
-// AND We must do it in batches, for the sake of reduced fsync call count
-// AND We must know what we stabilize. Basic workflow is like:
-// 1) primary OSD receives sync request
-// 2) it submits syncs to blockstore and peers
-// 3) after everyone acks sync it acks sync to the client
-// 4) after a while it takes his synced object list and sends stabilize requests
-//    to peers and to its own blockstore, thus freeing the old version
-
-struct ver_vector_t
-{
-    obj_ver_id *items = NULL;
-    uint64_t alloc = 0, size = 0;
-};
-
-static void init_versions(ver_vector_t & vec, obj_ver_id *start, obj_ver_id *end, uint64_t len)
-{
-    if (!vec.items)
-    {
-        vec.alloc = len;
-        vec.items = (obj_ver_id*)malloc_or_die(sizeof(obj_ver_id) * vec.alloc);
-        for (auto sv = start; sv < end; sv++)
-        {
-            vec.items[vec.size++] = *sv;
-        }
-    }
-}
-
-static void append_version(ver_vector_t & vec, obj_ver_id ov)
-{
-    if (vec.size >= vec.alloc)
-    {
-        vec.alloc = !vec.alloc ? 4 : vec.alloc*2;
-        vec.items = (obj_ver_id*)realloc_or_die(vec.items, sizeof(obj_ver_id) * vec.alloc);
-    }
-    vec.items[vec.size++] = ov;
-}
-
-static bool check_unsynced(std::vector<obj_ver_id> & check, obj_ver_id ov, std::vector<obj_ver_id> & to, int *count)
-{
-    bool found = false;
-    int j = 0, k = 0;
-    while (j < check.size())
-    {
-        if (check[j] == ov)
-            found = true;
-        if (check[j].oid == ov.oid && check[j].version <= ov.version)
-        {
-            to.push_back(check[j++]);
-            if (count)
-                (*count)--;
-        }
-        else
-            check[k++] = check[j++];
-    }
-    check.resize(k);
-    return found;
-}
-
-blockstore_op_t* blockstore_impl_t::selective_sync(blockstore_op_t *op)
-{
-    unsynced_big_write_count -= unsynced_big_writes.size();
-    unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
-    unsynced_big_write_count += unsynced_big_writes.size();
-    unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
-    // Create a sync operation, insert into the end of the queue
-    // And move ourselves into the end too!
-    // Rather hacky but that's what we need...
-    blockstore_op_t *sync_op = new blockstore_op_t;
-    sync_op->opcode = BS_OP_SYNC;
-    sync_op->buf = NULL;
-    sync_op->callback = [](blockstore_op_t *sync_op)
-    {
-        delete sync_op;
-    };
-    init_op(sync_op);
-    int sync_res = continue_sync(sync_op);
-    if (sync_res != 2)
-    {
-        // Put SYNC into the queue if it's not finished yet
-        submit_queue.push_back(sync_op);
-    }
-    // Restore unsynced_writes
-    unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
-    unsynced_big_write_count -= unsynced_big_writes.size();
-    unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
-    unsynced_big_write_count += unsynced_big_writes.size();
-    if (sync_res == 2)
-    {
-        // Sync is immediately completed
-        return NULL;
-    }
-    return sync_op;
-}
-
-// Returns: 2 = stop processing and dequeue, 0 = stop processing and do not dequeue, 1 = proceed with op itself
-int blockstore_impl_t::split_stab_op(blockstore_op_t *op, std::function<int(obj_ver_id v)> decider)
-{
-    bool add_sync = false;
-    ver_vector_t good_vers, bad_vers;
-    obj_ver_id* v;
-    int i, todo = 0;
-    for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
-    {
-        int action = decider(*v);
-        if (action < 0)
-        {
-            // Rollback changes
-            for (auto & ov: PRIV(op)->sync_big_writes)
-            {
-                unsynced_big_writes.push_back(ov);
-                unsynced_big_write_count++;
-            }
-            for (auto & ov: PRIV(op)->sync_small_writes)
-            {
-                unsynced_small_writes.push_back(ov);
-            }
-            free(good_vers.items);
-            good_vers.items = NULL;
-            free(bad_vers.items);
-            bad_vers.items = NULL;
-            // Error
-            op->retval = action;
-            FINISH_OP(op);
-            return 2;
-        }
-        else if (action == STAB_SPLIT_DONE)
-        {
-            // Already done
-            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
-        }
-        else if (action == STAB_SPLIT_WAIT)
-        {
-            // Already in progress, we just have to wait until it finishes
-            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
-            append_version(bad_vers, *v);
-        }
-        else if (action == STAB_SPLIT_SYNC)
-        {
-            // Needs a SYNC, we have to send a SYNC if not already in progress
-            //
-            // If the object is not present in unsynced_(big|small)_writes then
-            // it's currently being synced. If it's present then we can initiate
-            // its sync ourselves.
-            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
-            append_version(bad_vers, *v);
-            if (!add_sync)
-            {
-                PRIV(op)->sync_big_writes.clear();
-                PRIV(op)->sync_small_writes.clear();
-                add_sync = true;
-            }
-            check_unsynced(unsynced_small_writes, *v, PRIV(op)->sync_small_writes, NULL);
-            check_unsynced(unsynced_big_writes, *v, PRIV(op)->sync_big_writes, &unsynced_big_write_count);
-        }
-        else /* if (action == STAB_SPLIT_TODO) */
-        {
-            if (good_vers.items)
-            {
-                // If we're selecting versions then append it
-                // Main idea is that 99% of the time all versions passed to BS_OP_STABLE are synced
-                // And we don't want to select/allocate anything in that optimistic case
-                append_version(good_vers, *v);
-            }
-            todo++;
-        }
-    }
-    // In a pessimistic scenario, an operation may be split into 3:
-    // - Stabilize synced entries
-    // - Sync unsynced entries
-    // - Continue for unsynced entries after sync
-    add_sync = add_sync && (PRIV(op)->sync_big_writes.size() || PRIV(op)->sync_small_writes.size());
-    if (!todo && !bad_vers.size)
-    {
-        // Already stable
-        op->retval = 0;
-        FINISH_OP(op);
-        return 2;
-    }
-    op->retval = 0;
-    if (!todo && !add_sync)
-    {
-        // Only wait for inflight writes or current in-progress syncs
-        return 0;
-    }
-    blockstore_op_t *sync_op = NULL, *split_stab_op = NULL;
-    if (add_sync)
-    {
-        // Initiate a selective sync for PRIV(op)->sync_(big|small)_writes
-        sync_op = selective_sync(op);
-    }
-    if (bad_vers.size)
-    {
-        // Split part of the request into a separate operation
-        split_stab_op = new blockstore_op_t;
-        split_stab_op->opcode = op->opcode;
-        split_stab_op->buf = bad_vers.items;
-        split_stab_op->len = bad_vers.size;
-        init_op(split_stab_op);
-        submit_queue.push_back(split_stab_op);
-    }
-    if (sync_op || split_stab_op || good_vers.items)
-    {
-        void *orig_buf = op->buf;
-        if (good_vers.items)
-        {
-            op->buf = good_vers.items;
-            op->len = good_vers.size;
-        }
-        // Make a wrapped callback
-        int *split_op_counter = (int*)malloc_or_die(sizeof(int));
-        *split_op_counter = (sync_op ? 1 : 0) + (split_stab_op ? 1 : 0) + (todo ? 1 : 0);
-        auto cb = [op, good_items = good_vers.items,
-            bad_items = bad_vers.items, split_op_counter,
-            orig_buf, real_cb = op->callback](blockstore_op_t *split_op)
-        {
-            if (split_op->retval != 0)
-                op->retval = split_op->retval;
-            (*split_op_counter)--;
-            assert((*split_op_counter) >= 0);
-            if (op != split_op)
-                delete split_op;
-            if (!*split_op_counter)
-            {
-                free(good_items);
-                free(bad_items);
-                free(split_op_counter);
-                op->buf = orig_buf;
-                real_cb(op);
-            }
-        };
-        if (sync_op)
-        {
-            sync_op->callback = cb;
-        }
-        if (split_stab_op)
-        {
-            split_stab_op->callback = cb;
-        }
-        op->callback = cb;
-    }
-    if (!todo)
-    {
-        // All work is postponed
-        op->callback = NULL;
-        return 2;
-    }
-    return 1;
-}
-
+// Handles both stabilize (commit) and rollback
 int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
 {
-    if (PRIV(op)->op_state)
-    {
-        return continue_stable(op);
-    }
-    int r = split_stab_op(op, [this](obj_ver_id ov)
-    {
-        auto dirty_it = dirty_db.find(ov);
-        if (dirty_it == dirty_db.end())
-        {
-            auto & clean_db = clean_db_shard(ov.oid);
-            auto clean_it = clean_db.find(ov.oid);
-            if (clean_it == clean_db.end() || clean_it->second.version < ov.version)
-            {
-                // No such object version
-                printf("Error: %jx:%jx v%ju not found while stabilizing\n", ov.oid.inode, ov.oid.stripe, ov.version);
-                return -ENOENT;
-            }
-            else
-            {
-                // Already stable
-                return STAB_SPLIT_DONE;
-            }
-        }
-        else if (IS_STABLE(dirty_it->second.state))
-        {
-            // Already stable
-            return STAB_SPLIT_DONE;
-        }
-        while (true)
-        {
-            if (IS_IN_FLIGHT(dirty_it->second.state))
-            {
-                // Object write is still in progress. Wait until the write request completes
-                return STAB_SPLIT_WAIT;
-            }
-            else if (!IS_SYNCED(dirty_it->second.state))
-            {
-                // Object not synced yet - sync it
-                // In previous versions we returned EBUSY here and required
-                // the caller (OSD) to issue a global sync first. But a global sync
-                // waits for all writes in the queue including inflight writes. And
-                // inflight writes may themselves be blocked by unstable writes being
-                // still present in the journal and not flushed away from it.
-                // So we must sync specific objects here.
-                //
-                // Even more, we have to process "stabilize" request in parts. That is,
-                // we must stabilize all objects which are already synced. Otherwise
-                // they may block objects which are NOT synced yet.
-                return STAB_SPLIT_SYNC;
-            }
-            else if (IS_STABLE(dirty_it->second.state))
-            {
-                break;
-            }
-            // Check previous versions too
-            if (dirty_it == dirty_db.begin())
-            {
-                break;
-            }
-            dirty_it--;
-            if (dirty_it->first.oid != ov.oid)
-            {
-                break;
-            }
-        }
-        return STAB_SPLIT_TODO;
-    });
-    if (r != 1)
-    {
-        return r;
-    }
-    // Check journal space
-    blockstore_journal_check_t space_check(this);
-    if (!space_check.check_available(op, op->len, sizeof(journal_entry_stable), 0))
-    {
-        return 0;
-    }
-    // There is sufficient space. Check SQEs
-    BS_SUBMIT_CHECK_SQES(space_check.sectors_to_write);
-    // Prepare and submit journal entries
-    int s = 0;
-    auto v = (obj_ver_id*)op->buf;
-    for (int i = 0; i < op->len; i++, v++)
-    {
-        if (!journal.entry_fits(sizeof(journal_entry_stable)) &&
-            journal.sector_info[journal.cur_sector].dirty)
-        {
-            prepare_journal_sector_write(journal.cur_sector, op);
-            s++;
-        }
-        journal_entry_stable *je = (journal_entry_stable*)
-            prefill_single_journal_entry(journal, JE_STABLE, sizeof(journal_entry_stable));
-        je->oid = v->oid;
-        je->version = v->version;
-        je->crc32 = je_crc32((journal_entry*)je);
-        journal.crc32_last = je->crc32;
-    }
-    prepare_journal_sector_write(journal.cur_sector, op);
-    s++;
-    assert(s == space_check.sectors_to_write);
-    PRIV(op)->op_state = 1;
-    return 1;
-}
-
-int blockstore_impl_t::continue_stable(blockstore_op_t *op)
-{
-    if (PRIV(op)->op_state == 2)
-        goto resume_2;
-    else if (PRIV(op)->op_state == 4)
-        goto resume_4;
-    else
-        return 1;
-resume_2:
-    if (!disable_journal_fsync)
-    {
-        BS_SUBMIT_GET_SQE(sqe, data);
-        io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
-        data->iov = { 0 };
-        data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-        PRIV(op)->pending_ops = 1;
-        PRIV(op)->op_state = 3;
-        return 1;
-    }
-resume_4:
-    // Mark dirty_db entries as stable, acknowledge op completion
-    obj_ver_id* v;
-    int i;
-    for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
-    {
-        // Mark all dirty_db entries up to op->version as stable
-#ifdef BLOCKSTORE_DEBUG
-        printf("Stabilize %jx:%jx v%ju\n", v->oid.inode, v->oid.stripe, v->version);
-#endif
-        mark_stable(*v);
-    }
-    // Acknowledge op
+    obj_ver_id *v = (obj_ver_id*)op->buf;
+    auto priv = PRIV(op);
+    if (priv->op_state == 1)      goto resume_1;
+    else if (priv->op_state == 2) goto resume_2;
+    else if (priv->op_state == 3) goto resume_3;
+    else if (priv->op_state == 4) goto resume_4;
+    assert(!priv->op_state);
+    // Modify in-memory state and assign contiguous LSNs
+    priv->stab_pos = 0;
+    priv->lsn = priv->to_lsn = 0;
     op->retval = 0;
+    while (priv->stab_pos < op->len)
+    {
+        uint32_t modified_block = 0;
+        uint64_t new_lsn = 0;
+        uint64_t new_to_lsn = 0;
+        int res = op->opcode == BS_OP_STABLE
+            ? heap->post_stabilize(v[priv->stab_pos].oid, v[priv->stab_pos].version, &modified_block, &new_lsn, &new_to_lsn)
+            : heap->post_rollback(v[priv->stab_pos].oid, v[priv->stab_pos].version, &new_lsn, &modified_block);
+        if (res != 0)
+        {
+            assert(res == ENOENT || res == EBUSY);
+            op->retval = -res;
+        }
+        if (new_lsn)
+        {
+            assert(priv->lsn == 0 || priv->to_lsn == new_lsn-1);
+            if (!priv->lsn)
+                priv->lsn = new_lsn;
+            priv->to_lsn = op->opcode == BS_OP_STABLE ? new_to_lsn : new_lsn;
+        }
+        priv->stab_pos++;
+    }
+    // Submit metadata writes
+    priv->stab_pos = 0;
+resume_1:
+    priv->op_state = 1;
+    while (priv->stab_pos < op->len)
+    {
+        uint32_t block_num = 0;
+        heap_object_t *obj = heap->read_entry(v[priv->stab_pos].oid, &block_num);
+        if (obj)
+        {
+            io_uring_sqe *sqe = get_sqe();
+            if (!sqe)
+            {
+                if (priv->pending_ops > 0)
+                    return 1;
+                priv->wait_detail = 1;
+                priv->wait_for = WAIT_SQE;
+                return 0;
+            }
+            prepare_meta_block_write(op, block_num, sqe);
+        }
+        priv->stab_pos++;
+    }
+    if (priv->pending_ops > 0)
+    {
+        priv->op_state = 1;
+        return 1;
+    }
+    // Mark writes as completed to allow compaction
+    for (uint64_t lsn = priv->lsn; lsn <= priv->to_lsn; lsn++)
+    {
+        heap->mark_lsn_completed(lsn);
+    }
+    unsynced_meta_write_count++;
+    // Fsync, just because our semantics imply that commit (stabilize) is immediately fsynced
+    priv->op_state = 2;
+resume_2:
+resume_3:
+resume_4:
+    int res = do_sync(op, 2);
+    if (res != 2)
+    {
+        return res;
+    }
+    // Done. Don't touch op->retval - if anything resulted in ENOENT, return it as is
     FINISH_OP(op);
     return 2;
 }
-
-void blockstore_impl_t::mark_stable(obj_ver_id v, bool forget_dirty)
-{
-    auto dirty_it = dirty_db.find(v);
-    if (dirty_it != dirty_db.end())
-    {
-        if (IS_INSTANT(dirty_it->second.state))
-        {
-            // 'Instant' (non-EC) operations may complete and try to become stable out of order. Prevent it.
-            auto back_it = dirty_it;
-            while (back_it != dirty_db.begin())
-            {
-                back_it--;
-                if (back_it->first.oid != v.oid)
-                {
-                    break;
-                }
-                if (!IS_STABLE(back_it->second.state))
-                {
-                    // There are preceding unstable versions, can't flush <v>
-                    return;
-                }
-            }
-            while (true)
-            {
-                dirty_it++;
-                if (dirty_it == dirty_db.end() || dirty_it->first.oid != v.oid ||
-                    !IS_SYNCED(dirty_it->second.state))
-                {
-                    dirty_it--;
-                    break;
-                }
-                v.version = dirty_it->first.version;
-            }
-        }
-        while (1)
-        {
-            bool was_stable = IS_STABLE(dirty_it->second.state);
-            if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_SYNCED)
-            {
-                dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_STABLE;
-                // Allocations and deletions are counted when they're stabilized
-                if (IS_BIG_WRITE(dirty_it->second.state))
-                {
-                    int exists = -1;
-                    if (dirty_it != dirty_db.begin())
-                    {
-                        auto prev_it = dirty_it;
-                        prev_it--;
-                        if (prev_it->first.oid == v.oid)
-                        {
-                            exists = IS_DELETE(prev_it->second.state) ? 0 : 1;
-                        }
-                    }
-                    if (exists == -1)
-                    {
-                        auto & clean_db = clean_db_shard(v.oid);
-                        auto clean_it = clean_db.find(v.oid);
-                        exists = clean_it != clean_db.end() ? 1 : 0;
-                    }
-                    if (!exists)
-                    {
-                        uint64_t space_id = dirty_it->first.oid.inode;
-                        if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
-                            space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
-                        inode_space_stats[space_id] += dsk.data_block_size;
-                        used_blocks++;
-                    }
-                    big_to_flush++;
-                }
-                else if (IS_DELETE(dirty_it->second.state))
-                {
-                    uint64_t space_id = dirty_it->first.oid.inode;
-                    if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
-                        space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
-                    auto & sp = inode_space_stats[space_id];
-                    if (sp > dsk.data_block_size)
-                        sp -= dsk.data_block_size;
-                    else
-                        inode_space_stats.erase(space_id);
-                    used_blocks--;
-                    big_to_flush++;
-                }
-            }
-            else if (IS_IN_FLIGHT(dirty_it->second.state))
-            {
-                // mark_stable should never be called for in-flight or submitted writes
-                printf(
-                    "BUG: Attempt to mark_stable object %jx:%jx v%ju state of which is %x\n",
-                    dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
-                    dirty_it->second.state
-                );
-                exit(1);
-            }
-            if (forget_dirty && (IS_BIG_WRITE(dirty_it->second.state) ||
-                IS_DELETE(dirty_it->second.state)))
-            {
-                // Big write overrides all previous dirty entries
-                auto erase_end = dirty_it;
-                while (dirty_it != dirty_db.begin())
-                {
-                    dirty_it--;
-                    if (dirty_it->first.oid != v.oid)
-                    {
-                        dirty_it++;
-                        break;
-                    }
-                }
-                auto & clean_db = clean_db_shard(v.oid);
-                auto clean_it = clean_db.find(v.oid);
-                uint64_t clean_loc = clean_it != clean_db.end()
-                    ? clean_it->second.location : UINT64_MAX;
-                erase_dirty(dirty_it, erase_end, clean_loc);
-                break;
-            }
-            if (was_stable || dirty_it == dirty_db.begin())
-            {
-                break;
-            }
-            dirty_it--;
-            if (dirty_it->first.oid != v.oid)
-            {
-                break;
-            }
-        }
-        flusher->enqueue_flush(v);
-    }
-    auto unstab_it = unstable_writes.find(v.oid);
-    if (unstab_it != unstable_writes.end() &&
-        unstab_it->second <= v.version)
-    {
-        unstable_writes.erase(unstab_it);
-    }
-}

src/blockstore/blockstore_sync.cpp

@@ -2,232 +2,114 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include "blockstore_impl.h"
-
-#define SYNC_HAS_SMALL 1
-#define SYNC_HAS_BIG 2
-#define SYNC_DATA_SYNC_SENT 3
-#define SYNC_DATA_SYNC_DONE 4
-#define SYNC_JOURNAL_WRITE_SENT 5
-#define SYNC_JOURNAL_WRITE_DONE 6
-#define SYNC_JOURNAL_SYNC_SENT 7
-#define SYNC_DONE 8
+#include "blockstore_internal.h"
 
 int blockstore_impl_t::continue_sync(blockstore_op_t *op)
 {
-    if (immediate_commit == IMMEDIATE_ALL)
+    if (!PRIV(op)->op_state)
     {
-        // We can return immediately because sync is only dequeued after all previous writes
         op->retval = 0;
+    }
+    int res = do_sync(op, 0);
+    if (res == 2)
+    {
         FINISH_OP(op);
-        return 2;
     }
-    if (PRIV(op)->op_state == 0)
-    {
-        stop_sync_submitted = false;
-        unsynced_big_write_count -= unsynced_big_writes.size();
-        PRIV(op)->sync_big_writes.swap(unsynced_big_writes);
-        PRIV(op)->sync_small_writes.swap(unsynced_small_writes);
-        unsynced_big_writes.clear();
-        unsynced_small_writes.clear();
-        if (PRIV(op)->sync_big_writes.size() > 0)
-            PRIV(op)->op_state = SYNC_HAS_BIG;
-        else if (PRIV(op)->sync_small_writes.size() > 0)
-            PRIV(op)->op_state = SYNC_HAS_SMALL;
-        else
-            PRIV(op)->op_state = SYNC_DONE;
-    }
-    if (PRIV(op)->op_state == SYNC_HAS_SMALL)
-    {
-        // No big writes, just fsync the journal
-        if (journal.sector_info[journal.cur_sector].dirty)
-        {
-            // Write out the last journal sector if it happens to be dirty
-            BS_SUBMIT_CHECK_SQES(1);
-            prepare_journal_sector_write(journal.cur_sector, op);
-            PRIV(op)->op_state = SYNC_JOURNAL_WRITE_SENT;
-            return 1;
-        }
-        else
-        {
-            PRIV(op)->op_state = SYNC_JOURNAL_WRITE_DONE;
-        }
-    }
-    if (PRIV(op)->op_state == SYNC_HAS_BIG)
-    {
-        // 1st step: fsync data
-        if (!disable_data_fsync)
-        {
-            BS_SUBMIT_GET_SQE(sqe, data);
-            io_uring_prep_fsync(sqe, dsk.data_fd, IORING_FSYNC_DATASYNC);
-            data->iov = { 0 };
-            data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-            PRIV(op)->pending_ops = 1;
-            PRIV(op)->op_state = SYNC_DATA_SYNC_SENT;
-            return 1;
-        }
-        else
-        {
-            PRIV(op)->op_state = SYNC_DATA_SYNC_DONE;
-        }
-    }
-    if (PRIV(op)->op_state == SYNC_DATA_SYNC_DONE)
-    {
-        // 2nd step: Data device is synced, prepare & write journal entries
-        // Check space in the journal and journal memory buffers
-        blockstore_journal_check_t space_check(this);
-        if (dsk.csum_block_size)
-        {
-            // More complex check because all journal entries have different lengths
-            int left = PRIV(op)->sync_big_writes.size();
-            for (auto & sbw: PRIV(op)->sync_big_writes)
-            {
-                left--;
-                auto & dirty_entry = dirty_db.at(sbw);
-                uint64_t dyn_size = dsk.dirty_dyn_size(dirty_entry.offset, dirty_entry.len);
-                if (!space_check.check_available(op, 1, sizeof(journal_entry_big_write) + dyn_size, 0))
-                {
-                    return 0;
-                }
-            }
-        }
-        else if (!space_check.check_available(op, PRIV(op)->sync_big_writes.size(),
-            sizeof(journal_entry_big_write) + dsk.clean_entry_bitmap_size, 0))
-        {
-            return 0;
-        }
-        // Check SQEs. Don't bother about merging, submit each journal sector as a separate request
-        BS_SUBMIT_CHECK_SQES(space_check.sectors_to_write);
-        // Prepare and submit journal entries
-        auto it = PRIV(op)->sync_big_writes.begin();
-        int s = 0;
-        while (it != PRIV(op)->sync_big_writes.end())
-        {
-            auto & dirty_entry = dirty_db.at(*it);
-            uint64_t dyn_size = dsk.dirty_dyn_size(dirty_entry.offset, dirty_entry.len);
-            if (!journal.entry_fits(sizeof(journal_entry_big_write) + dyn_size) &&
-                journal.sector_info[journal.cur_sector].dirty)
-            {
-                prepare_journal_sector_write(journal.cur_sector, op);
-                s++;
-            }
-            journal_entry_big_write *je = (journal_entry_big_write*)prefill_single_journal_entry(
-                journal, (dirty_entry.state & BS_ST_INSTANT) ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
-                sizeof(journal_entry_big_write) + dyn_size
-            );
-            auto jsec = dirty_entry.journal_sector = journal.sector_info[journal.cur_sector].offset;
-            assert(journal.next_free >= journal.used_start
-                ? (jsec >= journal.used_start && jsec < journal.next_free)
-                : (jsec >= journal.used_start || jsec < journal.next_free));
-            journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
-#ifdef BLOCKSTORE_DEBUG
-            printf(
-                "journal offset %08jx is used by %jx:%jx v%ju (%ju refs)\n",
-                dirty_entry.journal_sector, it->oid.inode, it->oid.stripe, it->version,
-                journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
-            );
-#endif
-            je->oid = it->oid;
-            je->version = it->version;
-            je->offset = dirty_entry.offset;
-            je->len = dirty_entry.len;
-            je->location = dirty_entry.location;
-            memcpy((void*)(je+1), (alloc_dyn_data
-                ? (uint8_t*)dirty_entry.dyn_data+sizeof(int) : (uint8_t*)&dirty_entry.dyn_data), dyn_size);
-            je->crc32 = je_crc32((journal_entry*)je);
-            journal.crc32_last = je->crc32;
-            it++;
-        }
-        prepare_journal_sector_write(journal.cur_sector, op);
-        s++;
-        assert(s == space_check.sectors_to_write);
-        PRIV(op)->op_state = SYNC_JOURNAL_WRITE_SENT;
-        return 1;
-    }
-    if (PRIV(op)->op_state == SYNC_JOURNAL_WRITE_DONE)
-    {
-        if (!disable_journal_fsync)
-        {
-            BS_SUBMIT_GET_SQE(sqe, data);
-            io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
-            data->iov = { 0 };
-            data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-            PRIV(op)->pending_ops = 1;
-            PRIV(op)->op_state = SYNC_JOURNAL_SYNC_SENT;
-            return 1;
-        }
-        else
-        {
-            PRIV(op)->op_state = SYNC_DONE;
-        }
-    }
-    if (PRIV(op)->op_state == SYNC_DONE)
-    {
-        ack_sync(op);
-        return 2;
-    }
-    return 1;
+    return res;
 }
 
-void blockstore_impl_t::ack_sync(blockstore_op_t *op)
+bool blockstore_impl_t::submit_fsyncs(int & wait_count)
 {
-    // Handle states
-    for (auto it = PRIV(op)->sync_big_writes.begin(); it != PRIV(op)->sync_big_writes.end(); it++)
+    int n = ((unsynced_small_write_count > 0 || unsynced_big_write_count > 0 || unsynced_meta_write_count > 0) && !dsk.disable_meta_fsync) +
+        (unsynced_small_write_count > 0 && !dsk.disable_journal_fsync && dsk.journal_fd != dsk.meta_fd) +
+        (unsynced_big_write_count > 0 && !dsk.disable_data_fsync && dsk.data_fd != dsk.meta_fd && dsk.data_fd != dsk.journal_fd);
+    if (ringloop->space_left() < n)
     {
-#ifdef BLOCKSTORE_DEBUG
-        printf("Ack sync big %jx:%jx v%ju\n", it->oid.inode, it->oid.stripe, it->version);
-#endif
-        auto & unstab = unstable_writes[it->oid];
-        unstab = unstab < it->version ? it->version : unstab;
-        auto dirty_it = dirty_db.find(*it);
-        dirty_it->second.state = ((dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SYNCED);
-        if (dirty_it->second.state & BS_ST_INSTANT)
-        {
-            mark_stable(dirty_it->first);
-        }
-        else
-        {
-            unstable_unsynced--;
-            assert(unstable_unsynced >= 0);
-        }
-        dirty_it++;
-        while (dirty_it != dirty_db.end() && dirty_it->first.oid == it->oid)
-        {
-            if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG)
-            {
-                dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_IN_FLIGHT;
-            }
-            dirty_it++;
-        }
+        return false;
     }
-    for (auto it = PRIV(op)->sync_small_writes.begin(); it != PRIV(op)->sync_small_writes.end(); it++)
+    if (!n)
     {
-#ifdef BLOCKSTORE_DEBUG
-        printf("Ack sync small %jx:%jx v%ju\n", it->oid.inode, it->oid.stripe, it->version);
-#endif
-        auto & unstab = unstable_writes[it->oid];
-        unstab = unstab < it->version ? it->version : unstab;
-        if (dirty_db[*it].state == (BS_ST_DELETE | BS_ST_WRITTEN))
-        {
-            dirty_db[*it].state = (BS_ST_DELETE | BS_ST_SYNCED);
-            // Deletions are treated as immediately stable
-            mark_stable(*it);
-        }
-        else /* (BS_ST_INSTANT?) | BS_ST_SMALL_WRITE | BS_ST_WRITTEN */
-        {
-            dirty_db[*it].state = (dirty_db[*it].state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SYNCED;
-            if (dirty_db[*it].state & BS_ST_INSTANT)
-            {
-                mark_stable(*it);
-            }
-            else
-            {
-                unstable_unsynced--;
-                assert(unstable_unsynced >= 0);
-            }
-        }
+        return true;
     }
-    op->retval = 0;
-    FINISH_OP(op);
+    auto cb = [this, & wait_count](ring_data_t *data)
+    {
+        if (data->res != 0)
+            disk_error_abort("sync meta", data->res, 0);
+        wait_count--;
+        assert(wait_count >= 0);
+        if (!wait_count)
+            ringloop->wakeup();
+    };
+    if ((unsynced_small_write_count > 0 || unsynced_big_write_count > 0 || unsynced_meta_write_count > 0) && !dsk.disable_meta_fsync)
+    {
+        // fsync meta
+        io_uring_sqe *sqe = get_sqe();
+        assert(sqe);
+        ring_data_t *data = ((ring_data_t*)sqe->user_data);
+        io_uring_prep_fsync(sqe, dsk.meta_fd, IORING_FSYNC_DATASYNC);
+        data->iov = { 0 };
+        data->callback = cb;
+        wait_count++;
+    }
+    if (unsynced_small_write_count > 0 && !dsk.disable_journal_fsync && dsk.meta_fd != dsk.journal_fd)
+    {
+        // fsync buffer
+        io_uring_sqe *sqe = get_sqe();
+        assert(sqe);
+        ring_data_t *data = ((ring_data_t*)sqe->user_data);
+        io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
+        data->iov = { 0 };
+        data->callback = cb;
+        wait_count++;
+    }
+    if (unsynced_big_write_count > 0 && !dsk.disable_data_fsync && dsk.data_fd != dsk.meta_fd && dsk.data_fd != dsk.journal_fd)
+    {
+        // fsync data
+        io_uring_sqe *sqe = get_sqe();
+        assert(sqe);
+        ring_data_t *data = ((ring_data_t*)sqe->user_data);
+        io_uring_prep_fsync(sqe, dsk.data_fd, IORING_FSYNC_DATASYNC);
+        data->iov = { 0 };
+        data->callback = cb;
+        wait_count++;
+    }
+    unsynced_big_write_count = 0;
+    unsynced_small_write_count = 0;
+    unsynced_meta_write_count = 0;
+    return true;
+}
+
+int blockstore_impl_t::do_sync(blockstore_op_t *op, int base_state)
+{
+    int op_state = PRIV(op)->op_state - base_state;
+    if (op_state == 1) goto resume_1;
+    if (op_state == 2) goto resume_2;
+    assert(!op_state);
+    if (flusher->get_syncing_buffer())
+    {
+        // Wait for flusher-initiated sync
+        return 0;
+    }
+    if (dsk.disable_journal_fsync && dsk.disable_meta_fsync && dsk.disable_data_fsync || !unsynced_big_write_count && !unsynced_small_write_count)
+    {
+        // We can return immediately because sync only syncs previous writes
+        unsynced_big_write_count = unsynced_small_write_count = unsynced_meta_write_count = 0;
+        return 2;
+    }
+    PRIV(op)->lsn = heap->get_completed_lsn();
+    if (!submit_fsyncs(PRIV(op)->pending_ops))
+    {
+        PRIV(op)->wait_detail = 1;
+        PRIV(op)->wait_for = WAIT_SQE;
+        return 0;
+    }
+resume_1:
+    if (PRIV(op)->pending_ops > 0)
+    {
+        PRIV(op)->op_state = base_state+1;
+        return 1;
+    }
+resume_2:
+    heap->mark_lsn_fsynced(PRIV(op)->lsn);
+    return 2;
 }

src/blockstore/fio_engine.cpp

@@ -12,7 +12,7 @@
 // [LD_PRELOAD=libasan.so.8] \
 // fio -name=test -thread -ioengine=../build/src/blockstore/libfio_vitastor_blk.so \
 //     -bs=4k -direct=1 -rw=randwrite -iodepth=16 -size=900M -loops=10 \
-//     -bs_config='{"data_device":"./test_data.bin","meta_offset":0,"journal_offset":16777216,"data_offset":33554432,"disable_data_fsync":true,"immediate_commit":"all","journal_no_same_sector_overwrites":true}'
+//     -bs_config='{"data_device":"./test_data.bin","meta_offset":0,"journal_offset":16777216,"data_offset":33554432,"disable_data_fsync":true,"meta_format":3,"immediate_commit":"all","log_level":100,"journal_no_same_sector_overwrites":true,"journal_sector_buffer_count":1024}'
 //
 // Linear write:
 //
@@ -26,17 +26,20 @@
 
 #include "blockstore.h"
 #include "epoll_manager.h"
+#include "malloc_or_die.h"
 #include "json11/json11.hpp"
 #include "fio_headers.h"
 
 struct bs_data
 {
-    blockstore_t *bs;
+    blockstore_i *bs;
     epoll_manager_t *epmgr;
     ring_loop_t *ringloop;
     /* The list of completed io_u structs. */
     std::vector<io_u*> completed;
     int op_n = 0, inflight = 0;
+    bool ec = false;
+    bool imm = true;
     bool last_sync = false;
     bool trace = false;
 };
@@ -45,6 +48,7 @@ struct bs_options
 {
     int __pad;
     char *json_config = NULL;
+    int ec = 0;
     int trace = 0;
 };
 
@@ -58,6 +62,16 @@ static struct fio_option options[] = {
         .category = FIO_OPT_C_ENGINE,
         .group  = FIO_OPT_G_FILENAME,
     },
+    {
+        .name   = "ec",
+        .lname  = "Use EC write method",
+        .type   = FIO_OPT_BOOL,
+        .off1   = offsetof(struct bs_options, ec),
+        .help   = "Use EC write method",
+        .def    = "0",
+        .category = FIO_OPT_C_ENGINE,
+        .group  = FIO_OPT_G_FILENAME,
+    },
     {
         .name   = "bs_trace",
         .lname  = "trace",
@@ -88,6 +102,7 @@ static int bs_setup(struct thread_data *td)
         return 1;
     }
     td->io_ops_data = bsd;
+    bsd->ec = o->ec;
 
     if (!td->files_index)
     {
@@ -147,7 +162,9 @@ static int bs_init(struct thread_data *td)
     }
     bsd->ringloop = new ring_loop_t(RINGLOOP_DEFAULT_SIZE);
     bsd->epmgr = new epoll_manager_t(bsd->ringloop);
-    bsd->bs = new blockstore_t(config, bsd->ringloop, bsd->epmgr->tfd);
+    bsd->bs = blockstore_i::create(config, bsd->ringloop, bsd->epmgr->tfd);
+    bsd->imm = config.find("immediate_commit") == config.end() ||
+        config["immediate_commit"] == "all";
     while (1)
     {
         bsd->ringloop->loop();
@@ -183,7 +200,7 @@ static enum fio_q_status bs_queue(struct thread_data *td, struct io_u *io)
     {
     case DDIR_READ:
         op->opcode = BS_OP_READ;
-        op->buf = io->xfer_buf;
+        op->buf = (uint8_t*)io->xfer_buf;
         op->oid = {
             .inode = 1,
             .stripe = io->offset / bsd->bs->get_block_size(),
@@ -203,8 +220,8 @@ static enum fio_q_status bs_queue(struct thread_data *td, struct io_u *io)
         };
         break;
     case DDIR_WRITE:
-        op->opcode = BS_OP_WRITE_STABLE;
-        op->buf = io->xfer_buf;
+        op->opcode = bsd->ec ? BS_OP_WRITE : BS_OP_WRITE_STABLE;
+        op->buf = (uint8_t*)io->xfer_buf;
         op->oid = {
             .inode = 1,
             .stripe = io->offset / bsd->bs->get_block_size(),
@@ -212,16 +229,57 @@ static enum fio_q_status bs_queue(struct thread_data *td, struct io_u *io)
         op->version = 0; // assign automatically
         op->offset = io->offset % bsd->bs->get_block_size();
         op->len = io->xfer_buflen;
-        op->callback = [io, n = bsd->op_n](blockstore_op_t *op)
+        if (bsd->ec)
         {
-            io->error = op->retval < 0 ? -op->retval : 0;
-            bs_data *bsd = (bs_data*)io->engine_data;
-            bsd->inflight--;
-            bsd->completed.push_back(io);
-            if (bsd->trace)
-                printf("--- OP_WRITE %zx n=%d retval=%d\n", (size_t)op, n, op->retval);
-            delete op;
-        };
+            op->callback = [io, n = bsd->op_n](blockstore_op_t *op)
+            {
+                bs_data *bsd = (bs_data*)io->engine_data;
+                if (bsd->trace)
+                    printf("--- OP_WRITE %zx n=%d retval=%d\n", (size_t)op, n, op->retval);
+                if (op->retval < 0)
+                {
+                    io->error = op->retval < 0 ? -op->retval : 0;
+                    bsd->inflight--;
+                    bsd->completed.push_back(io);
+                    delete op;
+                }
+                else
+                {
+                    auto stab_op = new blockstore_op_t;
+                    stab_op->opcode = BS_OP_STABLE;
+                    stab_op->buf = (uint8_t*)malloc_or_die(sizeof(obj_ver_id));
+                    obj_ver_id *ver = (obj_ver_id *)stab_op->buf;
+                    ver[0].oid = op->oid;
+                    ver[0].version = op->version;
+                    stab_op->len = 1;
+                    stab_op->callback = [io, n](blockstore_op_t *op)
+                    {
+                        bs_data *bsd = (bs_data*)io->engine_data;
+                        if (bsd->trace)
+                            printf("--- OP_STABLE %zx n=%d retval=%d\n", (size_t)op, n, op->retval);
+                        io->error = op->retval < 0 ? -op->retval : 0;
+                        bsd->inflight--;
+                        bsd->completed.push_back(io);
+                        delete op;
+                    };
+                    bsd->bs->enqueue_op(stab_op);
+                    delete op;
+                }
+            };
+        }
+        else
+        {
+            op->callback = [io, n = bsd->op_n](blockstore_op_t *op)
+            {
+                bs_data *bsd = (bs_data*)io->engine_data;
+                if (bsd->trace)
+                    printf("--- OP_WRITE_STABLE %zx n=%d retval=%d\n", (size_t)op, n, op->retval);
+                io->error = op->retval < 0 ? -op->retval : 0;
+                bsd->inflight--;
+                bsd->completed.push_back(io);
+                delete op;
+            };
+        }
         bsd->last_sync = false;
         break;
     case DDIR_SYNC:

src/blockstore/multilist.cpp

@@ -0,0 +1,338 @@
+// Variable-length O(1) disk space allocator
+// Copyright (c) Vitaliy Filippov, 2025+
+// License: VNPL-1.1 (see README.md for details)
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <set>
+#include "multilist.h"
+
+multilist_alloc_t::multilist_alloc_t(uint32_t count, uint32_t maxn):
+    count(count), maxn(maxn)
+{
+    // not-so-memory-efficient: 16 MB memory per 1 GB buffer space, but buffer spaces are small, so OK
+    assert(count > 1 && count < 0x80000000);
+    sizes.resize(count);
+    nexts.resize(count); // nexts[i] = 0 -> area is used; nexts[i] = 1 -> no next; nexts[i] >= 2 -> next item
+    prevs.resize(count);
+    heads.resize(maxn); // heads[i] = 0 -> empty list; heads[i] >= 1 -> list head
+    sizes[0] = count;
+    sizes[count-1] = -count; // end
+    nexts[0] = 1;
+    heads[maxn-1] = 1;
+#ifdef MULTILIST_TRACE
+    print();
+#endif
+}
+
+bool multilist_alloc_t::is_free(uint32_t pos)
+{
+    assert(pos < count);
+    if (sizes[pos] < 0)
+        pos += sizes[pos]+1;
+    while (pos > 0 && !sizes[pos])
+        pos--;
+    return nexts[pos] > 0;
+}
+
+uint32_t multilist_alloc_t::find(uint32_t size)
+{
+    assert(size > 0);
+    assert(size <= maxn);
+    for (uint32_t i = size-1; i < maxn; i++)
+    {
+        if (heads[i])
+        {
+            return heads[i]-1;
+        }
+    }
+    return UINT32_MAX;
+}
+
+void multilist_alloc_t::verify()
+{
+    std::set<uint32_t> reachable;
+    for (int i = 0; i < maxn; i++)
+    {
+        uint32_t cur = heads[i];
+        while (cur)
+        {
+            if (!nexts[cur-1])
+            {
+                fprintf(stderr, "ERROR: item %d from freelist %d is not free\n", cur-1, i);
+                print();
+                abort();
+            }
+            if (nexts[cur-1] >= count+2)
+            {
+                fprintf(stderr, "ERROR: next out of range at %d: %d\n", cur-1, nexts[cur-1]);
+                print();
+                abort();
+            }
+            if (!(i < maxn-1 ? sizes[cur-1] == i+1 : (sizes[cur-1] >= i+1)))
+            {
+                fprintf(stderr, "ERROR: item %d is in wrong freelist: expected size %d, but actual size is %d\n", cur-1, i+1, sizes[cur-1]);
+                print();
+                abort();
+            }
+            if (reachable.find(cur-1) != reachable.end())
+            {
+                fprintf(stderr, "ERROR: doubly-claimed item %d\n", cur-1);
+                print();
+                abort();
+            }
+            reachable.insert(cur-1);
+            cur = nexts[cur-1]-1;
+        }
+    }
+    for (int i = 0; i < count; )
+    {
+        if (sizes[i])
+        {
+            assert(i+sizes[i] <= count);
+            if (sizes[i] > 1 && sizes[i+sizes[i]-1] != -sizes[i])
+            {
+                fprintf(stderr, "ERROR: start/end mismatch at %d: sizes[%d] should be %d, but is %d\n", i, i+sizes[i]-1, -sizes[i], sizes[i+sizes[i]-1]);
+                print();
+                abort();
+            }
+            for (int j = i+1; j < i+sizes[i]-1; j++)
+            {
+                if (sizes[j])
+                {
+                    fprintf(stderr, "ERROR: internal non-zero at %d: %d\n", j, sizes[j]);
+                    print();
+                    abort();
+                }
+            }
+            if (nexts[i] && reachable.find(i) == reachable.end())
+            {
+                fprintf(stderr, "ERROR: %d is unreachable from heads\n", i);
+                print();
+                abort();
+            }
+            if (nexts[i] >= 2)
+            {
+                if (nexts[i] >= 2+count)
+                {
+                    fprintf(stderr, "ERROR: next out of range at %d: %d\n", i, nexts[i]);
+                    print();
+                    abort();
+                }
+                if (prevs[nexts[i]-2] != i+1)
+                {
+                    fprintf(stderr, "ERROR: prev[next] (%d) != this (%d) at %d", prevs[nexts[i]-2], i+1, i);
+                    print();
+                    abort();
+                }
+            }
+            i += (sizes[i] > 1 ? sizes[i] : 1);
+        }
+        else
+            i++;
+    }
+}
+
+void multilist_alloc_t::print()
+{
+    printf("heads:");
+    for (int i = 0; i < maxn; i++)
+        if (heads[i])
+            printf(" %u=%u", i, heads[i]);
+    printf("\n");
+    printf("sizes:");
+    for (int i = 0; i < count; i++)
+        if (sizes[i])
+            printf(" %d=%d", i, sizes[i]);
+    printf("\n");
+    printf("prevs:");
+    for (int i = 0; i < count; i++)
+        if (prevs[i])
+            printf(" %d=%d", i, prevs[i]);
+    printf("\n");
+    printf("nexts:");
+    for (int i = 0; i < count; i++)
+        if (nexts[i])
+            printf(" %d=%d", i, nexts[i]);
+    printf("\n");
+    printf("items:");
+    for (int i = 0; i < count; )
+    {
+        if (sizes[i])
+        {
+            printf(" %u=(s:%d,n:%u,p:%u)", i, sizes[i], nexts[i], prevs[i]);
+            assert(i+sizes[i] <= count);
+            i += (sizes[i] > 1 ? sizes[i] : 1);
+        }
+        else
+            i++;
+    }
+    printf("\n");
+}
+
+void multilist_alloc_t::use(uint32_t pos, uint32_t size)
+{
+    assert(pos+size <= count && size > 0);
+    if (sizes[pos] <= 0)
+    {
+        uint32_t start = pos;
+        if (sizes[start] < 0)
+            start += sizes[start]+1;
+        else
+            while (start > 0 && !sizes[start])
+                start--;
+        assert(sizes[start] >= size);
+        use_full(start);
+        uint32_t full = sizes[start];
+        sizes[pos-1] = -pos+start;
+        sizes[start] = pos-start;
+        free(start);
+        sizes[pos+size-1] = -size;
+        sizes[pos] = size;
+        if (pos+size < start+full)
+        {
+            sizes[start+full-1] = -(start+full-pos-size);
+            sizes[pos+size] = start+full-pos-size;
+            free(pos+size);
+        }
+    }
+    else
+    {
+        assert(sizes[pos] >= size);
+        use_full(pos);
+        if (sizes[pos] > size)
+        {
+            uint32_t full = sizes[pos];
+            sizes[pos+size-1] = -size;
+            sizes[pos] = size;
+            sizes[pos+full-1] = -full+size;
+            sizes[pos+size] = full-size;
+            free(pos+size);
+        }
+    }
+#ifdef MULTILIST_TRACE
+    print();
+#endif
+}
+
+void multilist_alloc_t::use_full(uint32_t pos)
+{
+    uint32_t prevsize = sizes[pos];
+    assert(prevsize);
+    assert(nexts[pos]);
+    uint32_t pi = (prevsize < maxn ? prevsize : maxn)-1;
+    if (heads[pi] == pos+1)
+        heads[pi] = nexts[pos]-1;
+    if (prevs[pos])
+        nexts[prevs[pos]-1] = nexts[pos];
+    if (nexts[pos] >= 2)
+        prevs[nexts[pos]-2] = prevs[pos];
+    prevs[pos] = 0;
+    nexts[pos] = 0;
+}
+
+void multilist_alloc_t::free(uint32_t pos)
+{
+    do_free(pos);
+#ifdef MULTILIST_TRACE
+    print();
+#endif
+}
+
+void multilist_alloc_t::do_free(uint32_t pos)
+{
+    assert(!nexts[pos]);
+    uint32_t size = sizes[pos];
+    assert(size > 0);
+    // merge with previous?
+    if (pos > 0 && nexts[pos+(sizes[pos-1] == 1 ? -1 : sizes[pos-1])] > 0)
+    {
+        assert(sizes[pos-1] < 0 || sizes[pos-1] == 1);
+        uint32_t prevsize = sizes[pos-1] < 0 ? -sizes[pos-1] : 1;
+        use_full(pos-prevsize);
+        sizes[pos] = 0;
+        sizes[pos-1] = 0;
+        size += prevsize;
+        pos -= prevsize;
+        sizes[pos+size-1] = -size;
+        sizes[pos] = size;
+    }
+    // merge with next?
+    if (pos+size < count && nexts[pos+size] >= 1)
+    {
+        uint32_t nextsize = sizes[pos+size];
+        use_full(pos+size);
+        sizes[pos+size] = 0;
+        sizes[pos+size-1] = 0;
+        size += nextsize;
+        sizes[pos+size-1] = -size;
+        sizes[pos] = size;
+    }
+    uint32_t ni = (size < maxn ? size : maxn)-1;
+    nexts[pos] = heads[ni]+1;
+    prevs[pos] = 0;
+    if (heads[ni])
+        prevs[heads[ni]-1] = pos+1;
+    heads[ni] = pos+1;
+}
+
+multilist_index_t::multilist_index_t(uint32_t count, uint32_t max_used, uint32_t init_used):
+    count(count), max_used(max_used)
+{
+    assert(init_used < max_used);
+    nexts.resize(count, UINT32_MAX);
+    prevs.resize(count, UINT32_MAX);
+    heads.resize(max_used, UINT32_MAX);
+    for (size_t i = 0; i < count-1; i++)
+    {
+        nexts[i] = i+1;
+        prevs[i+1] = i;
+    }
+    prevs[0] = UINT32_MAX;
+    nexts[count-1] = UINT32_MAX;
+    heads[init_used] = 0;
+}
+
+uint32_t multilist_index_t::find(uint32_t wanted_used)
+{
+    assert(wanted_used < max_used);
+    return heads[wanted_used];
+}
+
+void multilist_index_t::change(uint32_t pos, uint32_t old_used, uint32_t new_used)
+{
+    if (new_used == old_used)
+        return;
+    assert(old_used < max_used && new_used < max_used);
+    if (prevs[pos] != UINT32_MAX)
+        nexts[prevs[pos]] = nexts[pos];
+    if (nexts[pos] != UINT32_MAX)
+        prevs[nexts[pos]] = prevs[pos];
+    if (heads[old_used] == pos)
+        heads[old_used] = nexts[pos];
+    prevs[pos] = UINT32_MAX;
+    if (heads[new_used] != UINT32_MAX)
+        prevs[heads[new_used]] = pos;
+    nexts[pos] = heads[new_used];
+    heads[new_used] = pos;
+}
+
+void multilist_index_t::print()
+{
+    printf("heads:");
+    for (int i = 0; i < max_used; i++)
+        if (heads[i] != UINT32_MAX)
+            printf(" %u=%u", i, heads[i]);
+    printf("\n");
+    printf("prevs:");
+    for (int i = 0; i < count; i++)
+        if (prevs[i] != UINT32_MAX)
+            printf(" %d=%d", i, prevs[i]);
+    printf("\n");
+    printf("nexts:");
+    for (int i = 0; i < count; i++)
+        if (nexts[i] != UINT32_MAX)
+            printf(" %d=%d", i, nexts[i]);
+    printf("\n");
+}

src/blockstore/multilist.h

@@ -0,0 +1,37 @@
+// Variable-length O(1) disk space allocator
+// Copyright (c) Vitaliy Filippov, 2025+
+// License: VNPL-1.1 (see README.md for details)
+
+#pragma once
+
+#include <stdint.h>
+#include <vector>
+
+struct multilist_alloc_t
+{
+    const uint32_t count, maxn;
+    std::vector<int32_t> sizes;
+    std::vector<uint32_t> nexts, prevs, heads;
+
+    multilist_alloc_t(uint32_t count, uint32_t maxn);
+    bool is_free(uint32_t pos);
+    uint32_t find(uint32_t size);
+    void use_full(uint32_t pos);
+    void use(uint32_t pos, uint32_t size);
+    void do_free(uint32_t pos);
+    void free(uint32_t pos);
+    void verify();
+    void print();
+};
+
+struct multilist_index_t
+{
+    const uint32_t count, max_used;
+    std::vector<uint32_t> nexts, prevs, heads;
+
+    // used should be always < max_used
+    multilist_index_t(uint32_t count, uint32_t max_used, uint32_t init_used);
+    uint32_t find(uint32_t wanted_used);
+    void change(uint32_t pos, uint32_t old_used, uint32_t new_used);
+    void print();
+};

src/blockstore/ondisk_formats.h

@@ -1,12 +1,11 @@
+// Metadata on-disk structures
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
 #pragma once
 
 #include "crc32c.h"
-#include <set>
 
-#define MIN_JOURNAL_SIZE 4*1024*1024
 #define JOURNAL_MAGIC 0x4A33
 #define JOURNAL_VERSION_V1 1
 #define JOURNAL_VERSION_V2 2
@@ -145,74 +144,60 @@ inline uint32_t je_crc32(journal_entry *je)
     return crc32c(0x48674bc7, ((uint8_t*)je)+4, je->size-4);
 }
 
-struct journal_sector_info_t
+// "VITAstor"
+#define BLOCKSTORE_META_MAGIC_V1 0x726F747341544956l
+#define BLOCKSTORE_META_FORMAT_V1 1
+#define BLOCKSTORE_META_FORMAT_V2 2
+#define BLOCKSTORE_META_FORMAT_HEAP 3
+
+// metadata header (superblock)
+struct __attribute__((__packed__)) blockstore_meta_header_v1_t
 {
-    uint64_t offset;
-    uint64_t flush_count;
-    bool written;
-    bool dirty;
-    uint64_t submit_id;
+    uint64_t zero;
+    uint64_t magic;
+    uint64_t version;
+    uint32_t meta_block_size;
+    uint32_t data_block_size;
+    uint32_t bitmap_granularity;
 };
 
-struct pending_journaling_t
+struct __attribute__((__packed__)) blockstore_meta_header_v2_t
 {
-    int pending;
-    int sector;
-    blockstore_op_t *op;
+    uint64_t zero;
+    uint64_t magic;
+    uint64_t version;
+    uint32_t meta_block_size;
+    uint32_t data_block_size;
+    uint32_t bitmap_granularity;
+    uint32_t data_csum_type;
+    uint32_t csum_block_size;
+    uint32_t header_csum;
 };
 
-struct journal_t
+struct __attribute__((__packed__)) blockstore_meta_header_v3_t
 {
-    int fd;
-    bool inmemory = false;
-    bool flush_journal = false;
-    void *buffer = NULL;
+    uint64_t zero;
+    uint64_t magic;
+    uint64_t version;
+    uint32_t meta_block_size;
+    uint32_t data_block_size;
+    uint32_t bitmap_granularity;
+    uint32_t data_csum_type;
+    uint32_t csum_block_size;
+    uint32_t header_csum;
+    uint64_t compacted_lsn;
 
-    uint64_t block_size;
-    uint64_t offset, len;
-    // Next free block offset
-    uint64_t next_free = 0;
-    // First occupied block offset
-    uint64_t used_start = 0;
-    // End of the last block not used for writing anymore
-    uint64_t dirty_start = 0;
-    uint32_t crc32_last = 0;
-
-    // Current sector(s) used for writing
-    void *sector_buf = NULL;
-    journal_sector_info_t *sector_info = NULL;
-    uint64_t sector_count;
-    bool no_same_sector_overwrites = false;
-    int cur_sector = 0;
-    int in_sector_pos = 0;
-    std::vector<int> submitting_sectors;
-    std::multimap<uint64_t, pending_journaling_t> flushing_ops;
-    uint64_t submit_id = 0;
-
-    // Used sector map
-    // May use ~ 80 MB per 1 GB of used journal space in the worst case
-    std::map<uint64_t, uint64_t> used_sectors;
-
-    ~journal_t();
-    bool trim();
-    uint64_t get_trim_pos();
-    void dump_diagnostics();
-    inline bool entry_fits(int size)
-    {
-        return !(block_size - in_sector_pos < size ||
-            no_same_sector_overwrites && sector_info[cur_sector].written);
-    }
+    void set_crc32c();
 };
 
-struct blockstore_journal_check_t
+// 32 bytes = 24 bytes + block bitmap (4 bytes by default) + external attributes (also bitmap, 4 bytes by default)
+// per "clean" entry on disk with fixed metadata tables
+struct __attribute__((__packed__)) clean_disk_entry
 {
-    blockstore_impl_t *bs;
-    uint64_t next_pos, next_sector, next_in_pos;
-    int sectors_to_write, first_sector;
-    bool right_dir; // writing to the end or the beginning of the ring buffer
-
-    blockstore_journal_check_t(blockstore_impl_t *bs);
-    int check_available(blockstore_op_t *op, int required, int size, int data_after);
+    object_id oid;
+    uint64_t version;
+    uint8_t bitmap[];
+    // Two more fields come after bitmap in metadata version 2:
+    // uint32_t data_csum[];
+    // uint32_t entry_csum;
 };
-
-journal_entry* prefill_single_journal_entry(journal_t & journal, uint16_t type, uint32_t size);

src/blockstore/v1/flush.h

@@ -0,0 +1,134 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+struct copy_buffer_t
+{
+    int copy_flags;
+    uint64_t offset, len, disk_offset;
+    uint64_t journal_sector; // only for reads: sector+1 if used and !journal.inmemory, otherwise 0
+    void *buf;
+    uint8_t *csum_buf;
+    int *dyn_data;
+};
+
+struct meta_sector_t
+{
+    uint64_t offset, len;
+    int state;
+    void *buf;
+    int usage_count;
+};
+
+struct flusher_sync_t
+{
+    bool fsync_meta;
+    int ready_count;
+    int state;
+};
+
+struct flusher_meta_write_t
+{
+    uint64_t sector, pos;
+    bool submitted;
+    void *buf;
+    std::map<uint64_t, meta_sector_t>::iterator it;
+};
+
+class journal_flusher_t;
+
+// Journal flusher coroutine
+class journal_flusher_co
+{
+    blockstore_impl_t *bs;
+    journal_flusher_t *flusher;
+    int wait_state, wait_count, wait_journal_count;
+    struct io_uring_sqe *sqe;
+    struct ring_data_t *data;
+
+    std::list<flusher_sync_t>::iterator cur_sync;
+
+    obj_ver_id cur;
+    std::map<obj_ver_id, dirty_entry>::iterator dirty_it, dirty_start, dirty_end;
+    std::map<object_id, uint64_t>::iterator repeat_it;
+    std::function<void(ring_data_t*)> simple_callback_r, simple_callback_rj, simple_callback_w;
+
+    bool try_trim = false;
+    bool skip_copy, has_delete, has_writes;
+    std::vector<copy_buffer_t> v;
+    std::vector<copy_buffer_t>::iterator it;
+    int i;
+    bool fill_incomplete, cleared_incomplete;
+    int read_to_fill_incomplete;
+    int copy_count;
+    uint64_t clean_loc, clean_ver, old_clean_loc, old_clean_ver;
+    flusher_meta_write_t meta_old, meta_new;
+    bool clean_init_bitmap;
+    uint64_t clean_bitmap_offset, clean_bitmap_len;
+    uint8_t *clean_init_dyn_ptr;
+    uint8_t *new_clean_bitmap;
+
+    uint64_t new_trim_pos;
+
+    friend class journal_flusher_t;
+    void scan_dirty();
+    bool read_dirty(int wait_base);
+    bool modify_meta_do_reads(int wait_base);
+    bool wait_meta_reads(int wait_base);
+    bool modify_meta_read(uint64_t meta_loc, flusher_meta_write_t &wr, int wait_base);
+    bool clear_incomplete_csum_block_bits(int wait_base);
+    void calc_block_checksums(uint32_t *new_data_csums, bool skip_overwrites);
+    void update_metadata_entry();
+    bool write_meta_block(flusher_meta_write_t & meta_block, int wait_base);
+    void update_clean_db();
+    void free_data_blocks();
+    bool fsync_batch(bool fsync_meta, int wait_base);
+    bool trim_journal(int wait_base);
+    void free_buffers();
+public:
+    journal_flusher_co();
+    bool loop();
+};
+
+// Journal flusher itself
+class journal_flusher_t
+{
+    int trim_wanted = 0;
+    bool dequeuing;
+    int min_flusher_count, max_flusher_count, cur_flusher_count, target_flusher_count;
+    int flusher_start_threshold;
+    journal_flusher_co *co;
+    blockstore_impl_t *bs;
+    friend class journal_flusher_co;
+
+    int journal_trim_counter;
+    bool trimming;
+    void* journal_superblock;
+
+    int active_flushers;
+    int syncing_flushers;
+    std::list<flusher_sync_t> syncs;
+    std::map<object_id, uint64_t> sync_to_repeat;
+
+    std::map<uint64_t, meta_sector_t> meta_sectors;
+    std::deque<object_id> flush_queue;
+    std::unordered_map<object_id, uint64_t> flush_versions;
+    std::unordered_set<uint64_t> inflight_meta_sectors;
+
+    bool try_find_older(std::map<obj_ver_id, dirty_entry>::iterator & dirty_end, obj_ver_id & cur);
+    bool try_find_other(std::map<obj_ver_id, dirty_entry>::iterator & dirty_end, obj_ver_id & cur);
+
+public:
+    journal_flusher_t(blockstore_impl_t *bs);
+    ~journal_flusher_t();
+    void loop();
+    bool is_trim_wanted() { return trim_wanted; }
+    bool is_active();
+    void mark_trim_possible();
+    void request_trim();
+    void release_trim();
+    void enqueue_flush(obj_ver_id oid);
+    void unshift_flush(obj_ver_id oid, bool force);
+    void remove_flush(object_id oid);
+    void dump_diagnostics();
+    bool is_mutated(uint64_t clean_loc);
+};

src/blockstore/v1/impl.cpp

@@ -0,0 +1,810 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
+
+blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd)
+{
+    assert(sizeof(blockstore_op_private_t) <= BS_OP_PRIVATE_DATA_SIZE);
+    this->tfd = tfd;
+    this->ringloop = ringloop;
+    ring_consumer.loop = [this]() { loop(); };
+    ringloop->register_consumer(&ring_consumer);
+    initialized = 0;
+    parse_config(config, true);
+    try
+    {
+        dsk.open_data();
+        dsk.open_meta();
+        dsk.open_journal();
+        calc_lengths();
+        alloc_dyn_data = dsk.clean_dyn_size > sizeof(void*) || dsk.csum_block_size > 0;
+        zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.data_block_size);
+        data_alloc = new allocator_t(dsk.block_count);
+    }
+    catch (std::exception & e)
+    {
+        dsk.close_all();
+        throw;
+    }
+    flusher = new journal_flusher_t(this);
+}
+
+blockstore_impl_t::~blockstore_impl_t()
+{
+    delete data_alloc;
+    delete flusher;
+    if (zero_object)
+        free(zero_object);
+    ringloop->unregister_consumer(&ring_consumer);
+    dsk.close_all();
+    if (metadata_buffer)
+        free(metadata_buffer);
+    if (clean_bitmaps)
+        free(clean_bitmaps);
+}
+
+bool blockstore_impl_t::is_started()
+{
+    return initialized == 10;
+}
+
+bool blockstore_impl_t::is_stalled()
+{
+    return queue_stall;
+}
+
+// main event loop - produce requests
+void blockstore_impl_t::loop()
+{
+    // FIXME: initialized == 10 is ugly
+    if (initialized != 10)
+    {
+        // read metadata, then journal
+        if (initialized == 0)
+        {
+            metadata_init_reader = new blockstore_init_meta(this);
+            initialized = 1;
+        }
+        if (initialized == 1)
+        {
+            int res = metadata_init_reader->loop();
+            if (!res)
+            {
+                delete metadata_init_reader;
+                metadata_init_reader = NULL;
+                journal_init_reader = new blockstore_init_journal(this);
+                initialized = 2;
+            }
+        }
+        if (initialized == 2)
+        {
+            int res = journal_init_reader->loop();
+            if (!res)
+            {
+                delete journal_init_reader;
+                journal_init_reader = NULL;
+                initialized = 3;
+                ringloop->wakeup();
+            }
+        }
+        if (initialized == 3)
+        {
+            if (!readonly && dsk.discard_on_start)
+                dsk.trim_data([this](uint64_t block_num){ return data_alloc->get(block_num); });
+            if (journal.flush_journal)
+                initialized = 4;
+            else
+                initialized = 10;
+        }
+        if (initialized == 4)
+        {
+            if (readonly)
+            {
+                printf("Can't flush the journal in readonly mode\n");
+                exit(1);
+            }
+            flusher->loop();
+            ringloop->submit();
+        }
+    }
+    else
+    {
+        // try to submit ops
+        unsigned initial_ring_space = ringloop->space_left();
+        // has_writes == 0 - no writes before the current queue item
+        // has_writes == 1 - some writes in progress
+        // has_writes == 2 - tried to submit some writes, but failed
+        int has_writes = 0, op_idx = 0, new_idx = 0;
+        for (; op_idx < submit_queue.size(); op_idx++, new_idx++)
+        {
+            auto op = submit_queue[op_idx];
+            submit_queue[new_idx] = op;
+            // FIXME: This needs some simplification
+            // Writes should not block reads if the ring is not full and reads don't depend on them
+            // In all other cases we should stop submission
+            if (PRIV(op)->wait_for)
+            {
+                check_wait(op);
+                if (PRIV(op)->wait_for == WAIT_SQE)
+                {
+                    break;
+                }
+                else if (PRIV(op)->wait_for)
+                {
+                    if (op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE)
+                    {
+                        has_writes = 2;
+                    }
+                    continue;
+                }
+            }
+            unsigned prev_sqe_pos = ringloop->save();
+            // 0 = can't submit
+            // 1 = in progress
+            // 2 = can be removed from queue
+            int wr_st = 0;
+            if (op->opcode == BS_OP_READ)
+            {
+                wr_st = dequeue_read(op);
+            }
+            else if (op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE)
+            {
+                if (has_writes == 2)
+                {
+                    // Some writes already could not be submitted
+                    continue;
+                }
+                wr_st = dequeue_write(op);
+                has_writes = wr_st > 0 ? 1 : 2;
+            }
+            else if (op->opcode == BS_OP_DELETE)
+            {
+                if (has_writes == 2)
+                {
+                    // Some writes already could not be submitted
+                    continue;
+                }
+                wr_st = dequeue_del(op);
+                has_writes = wr_st > 0 ? 1 : 2;
+            }
+            else if (op->opcode == BS_OP_SYNC)
+            {
+                // sync only completed writes?
+                // wait for the data device fsync to complete, then submit journal writes for big writes
+                // then submit an fsync operation
+                wr_st = continue_sync(op);
+            }
+            else if (op->opcode == BS_OP_STABLE)
+            {
+                wr_st = dequeue_stable(op);
+            }
+            else if (op->opcode == BS_OP_ROLLBACK)
+            {
+                wr_st = dequeue_rollback(op);
+            }
+            else if (op->opcode == BS_OP_LIST)
+            {
+                // LIST doesn't have to be blocked by previous modifications
+                process_list(op);
+                wr_st = 2;
+            }
+            if (wr_st == 2)
+            {
+                submit_queue[op_idx] = NULL;
+                new_idx--;
+            }
+            if (wr_st == 0)
+            {
+                ringloop->restore(prev_sqe_pos);
+                if (PRIV(op)->wait_for == WAIT_SQE)
+                {
+                    // ring is full, stop submission
+                    break;
+                }
+                else if (PRIV(op)->wait_for == WAIT_JOURNAL)
+                {
+                    PRIV(op)->wait_detail2 = (unstable_writes.size()+unstable_unsynced);
+                }
+            }
+        }
+        if (op_idx != new_idx)
+        {
+            while (op_idx < submit_queue.size())
+            {
+                submit_queue[new_idx++] = submit_queue[op_idx++];
+            }
+            submit_queue.resize(new_idx);
+        }
+        if (!readonly)
+        {
+            flusher->loop();
+        }
+        int ret = ringloop->submit();
+        if (ret < 0)
+        {
+            throw std::runtime_error(std::string("io_uring_submit: ") + strerror(-ret));
+        }
+        for (auto s: journal.submitting_sectors)
+        {
+            // Mark journal sector writes as submitted
+            if (journal.sector_info[s].submit_id)
+                journal.sector_info[s].written = true;
+            journal.sector_info[s].submit_id = 0;
+        }
+        journal.submitting_sectors.clear();
+        if ((initial_ring_space - ringloop->space_left()) > 0)
+        {
+            live = true;
+        }
+        queue_stall = !live && !ringloop->has_work();
+        live = false;
+    }
+}
+
+bool blockstore_impl_t::is_safe_to_stop()
+{
+    // It's safe to stop blockstore when there are no in-flight operations,
+    // no in-progress syncs and flusher isn't doing anything
+    if (submit_queue.size() > 0 || !readonly && flusher->is_active())
+    {
+        return false;
+    }
+    if (unsynced_big_writes.size() > 0 || unsynced_small_writes.size() > 0)
+    {
+        if (!readonly && !stop_sync_submitted)
+        {
+            // We should sync the blockstore before unmounting
+            blockstore_op_t *op = new blockstore_op_t;
+            op->opcode = BS_OP_SYNC;
+            op->buf = NULL;
+            op->callback = [](blockstore_op_t *op)
+            {
+                delete op;
+            };
+            enqueue_op(op);
+            stop_sync_submitted = true;
+        }
+        return false;
+    }
+    return true;
+}
+
+void blockstore_impl_t::check_wait(blockstore_op_t *op)
+{
+    if (PRIV(op)->wait_for == WAIT_SQE)
+    {
+        if (ringloop->space_left() < PRIV(op)->wait_detail)
+        {
+            // stop submission if there's still no free space
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting for %ju SQE(s)\n", PRIV(op)->wait_detail);
+#endif
+            return;
+        }
+        PRIV(op)->wait_for = 0;
+    }
+    else if (PRIV(op)->wait_for == WAIT_JOURNAL)
+    {
+        if (journal.used_start == PRIV(op)->wait_detail &&
+            (unstable_writes.size()+unstable_unsynced) == PRIV(op)->wait_detail2)
+        {
+            // do not submit
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting to flush journal offset %08jx\n", PRIV(op)->wait_detail);
+#endif
+            return;
+        }
+        flusher->release_trim();
+        PRIV(op)->wait_for = 0;
+    }
+    else if (PRIV(op)->wait_for == WAIT_JOURNAL_BUFFER)
+    {
+        int next = ((journal.cur_sector + 1) % journal.sector_count);
+        if (journal.sector_info[next].flush_count > 0 ||
+            journal.sector_info[next].dirty)
+        {
+            // do not submit
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting for a journal buffer\n");
+#endif
+            return;
+        }
+        PRIV(op)->wait_for = 0;
+    }
+    else if (PRIV(op)->wait_for == WAIT_FREE)
+    {
+        if (!data_alloc->get_free_count() && big_to_flush > 0)
+        {
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting for free space on the data device\n");
+#endif
+            return;
+        }
+        flusher->release_trim();
+        PRIV(op)->wait_for = 0;
+    }
+    else
+    {
+        throw std::runtime_error("BUG: op->wait_for value is unexpected");
+    }
+}
+
+void blockstore_impl_t::enqueue_op(blockstore_op_t *op)
+{
+    if (op->opcode < BS_OP_MIN || op->opcode > BS_OP_MAX ||
+        ((op->opcode == BS_OP_READ || op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE) && (
+            op->offset >= dsk.data_block_size ||
+            op->len > dsk.data_block_size-op->offset ||
+            (op->len % dsk.disk_alignment)
+        )) ||
+        readonly && op->opcode != BS_OP_READ && op->opcode != BS_OP_LIST)
+    {
+        // Basic verification not passed
+        op->retval = -EINVAL;
+        ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
+        return;
+    }
+    if ((op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE) && !enqueue_write(op))
+    {
+        ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
+        return;
+    }
+    if (op->opcode == BS_OP_SYNC)
+    {
+        unsynced_queued_ops = 0;
+    }
+    init_op(op);
+    submit_queue.push_back(op);
+    ringloop->wakeup();
+}
+
+void blockstore_impl_t::init_op(blockstore_op_t *op)
+{
+    // Call constructor without allocating memory. We'll call destructor before returning op back
+    new ((void*)op->private_data) blockstore_op_private_t;
+    PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
+    PRIV(op)->wait_for = 0;
+    PRIV(op)->op_state = 0;
+    PRIV(op)->pending_ops = 0;
+}
+
+static bool replace_stable(object_id oid, uint64_t version, int search_start, int search_end, obj_ver_id* list)
+{
+    while (search_start < search_end)
+    {
+        int pos = search_start+(search_end-search_start)/2;
+        if (oid < list[pos].oid)
+        {
+            search_end = pos;
+        }
+        else if (list[pos].oid < oid)
+        {
+            search_start = pos+1;
+        }
+        else
+        {
+            list[pos].version = version;
+            return true;
+        }
+    }
+    return false;
+}
+
+blockstore_clean_db_t& blockstore_impl_t::clean_db_shard(object_id oid)
+{
+    uint64_t pg_num = 0;
+    uint64_t pool_id = (oid.inode >> (64-POOL_ID_BITS));
+    auto sh_it = clean_db_settings.find(pool_id);
+    if (sh_it != clean_db_settings.end())
+    {
+        // like map_to_pg()
+        pg_num = (oid.stripe / sh_it->second.pg_stripe_size) % sh_it->second.pg_count + 1;
+    }
+    return clean_db_shards[(pool_id << (64-POOL_ID_BITS)) | pg_num];
+}
+
+void blockstore_impl_t::reshard_clean_db(pool_id_t pool, uint32_t pg_count, uint32_t pg_stripe_size)
+{
+    uint64_t pool_id = (uint64_t)pool;
+    std::map<pool_pg_id_t, blockstore_clean_db_t> new_shards;
+    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    while (sh_it != clean_db_shards.end() &&
+        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        for (auto & pair: sh_it->second)
+        {
+            // like map_to_pg()
+            uint64_t pg_num = (pair.first.stripe / pg_stripe_size) % pg_count + 1;
+            uint64_t shard_id = (pool_id << (64-POOL_ID_BITS)) | pg_num;
+            new_shards[shard_id][pair.first] = pair.second;
+        }
+        clean_db_shards.erase(sh_it++);
+    }
+    for (sh_it = new_shards.begin(); sh_it != new_shards.end(); sh_it++)
+    {
+        auto & to = clean_db_shards[sh_it->first];
+        to.swap(sh_it->second);
+    }
+    clean_db_settings[pool_id] = (pool_shard_settings_t){
+        .pg_count = pg_count,
+        .pg_stripe_size = pg_stripe_size,
+    };
+}
+
+void blockstore_impl_t::process_list(blockstore_op_t *op)
+{
+    uint32_t list_pg = op->pg_number+1;
+    uint32_t pg_count = op->pg_count;
+    uint64_t pg_stripe_size = op->pg_alignment;
+    uint64_t min_inode = op->min_oid.inode;
+    uint64_t max_inode = op->max_oid.inode;
+    // Check PG
+    if (pg_count != 0 && (pg_stripe_size < MIN_DATA_BLOCK_SIZE || list_pg > pg_count))
+    {
+        op->retval = -EINVAL;
+        FINISH_OP(op);
+        return;
+    }
+    // Check if the DB needs resharding
+    // (we don't know about PGs from the beginning, we only create "shards" here)
+    uint64_t first_shard = 0, last_shard = UINT64_MAX;
+    if (min_inode != 0 &&
+        // Check if min_inode == max_inode == pool_id<<N, i.e. this is a pool listing
+        (min_inode >> (64-POOL_ID_BITS)) == (max_inode >> (64-POOL_ID_BITS)))
+    {
+        pool_id_t pool_id = (min_inode >> (64-POOL_ID_BITS));
+        if (pg_count > 1)
+        {
+            // Per-pg listing
+            auto sh_it = clean_db_settings.find(pool_id);
+            if (sh_it == clean_db_settings.end() ||
+                sh_it->second.pg_count != pg_count ||
+                sh_it->second.pg_stripe_size != pg_stripe_size)
+            {
+                reshard_clean_db(pool_id, pg_count, pg_stripe_size);
+            }
+            first_shard = last_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS)) | list_pg;
+        }
+        else
+        {
+            // Per-pool listing
+            first_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS));
+            last_shard = ((uint64_t)(pool_id+1) << (64-POOL_ID_BITS)) - 1;
+        }
+    }
+    // Copy clean_db entries
+    int stable_count = 0, stable_alloc = 0;
+    if (min_inode != max_inode)
+    {
+        for (auto shard_it = clean_db_shards.lower_bound(first_shard);
+            shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
+            shard_it++)
+        {
+            auto & clean_db = shard_it->second;
+            stable_alloc += clean_db.size();
+        }
+    }
+    if (op->list_stable_limit > 0)
+    {
+        stable_alloc = op->list_stable_limit;
+        if (stable_alloc > 1024*1024)
+            stable_alloc = 1024*1024;
+    }
+    if (stable_alloc < 32768)
+    {
+        stable_alloc = 32768;
+    }
+    obj_ver_id *stable = (obj_ver_id*)malloc(sizeof(obj_ver_id) * stable_alloc);
+    if (!stable)
+    {
+        op->retval = -ENOMEM;
+        FINISH_OP(op);
+        return;
+    }
+    auto max_oid = op->max_oid;
+    bool limited = false;
+    pool_pg_id_t last_shard_id = 0;
+    for (auto shard_it = clean_db_shards.lower_bound(first_shard);
+        shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
+        shard_it++)
+    {
+        auto & clean_db = shard_it->second;
+        auto clean_it = clean_db.begin(), clean_end = clean_db.end();
+        if (op->min_oid.inode != 0 || op->min_oid.stripe != 0)
+        {
+            clean_it = clean_db.lower_bound(op->min_oid);
+        }
+        if ((max_oid.inode != 0 || max_oid.stripe != 0) && !(max_oid < op->min_oid))
+        {
+            clean_end = clean_db.upper_bound(max_oid);
+        }
+        for (; clean_it != clean_end; clean_it++)
+        {
+            if (stable_count >= stable_alloc)
+            {
+                stable_alloc *= 2;
+                obj_ver_id* nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
+                if (!nst)
+                {
+                    op->retval = -ENOMEM;
+                    FINISH_OP(op);
+                    return;
+                }
+                stable = nst;
+            }
+            stable[stable_count++] = {
+                .oid = clean_it->first,
+                .version = clean_it->second.version,
+            };
+            if (op->list_stable_limit > 0 && stable_count >= op->list_stable_limit)
+            {
+                if (!limited)
+                {
+                    limited = true;
+                    max_oid = stable[stable_count-1].oid;
+                }
+                break;
+            }
+        }
+        if (op->list_stable_limit > 0)
+        {
+            // To maintain the order, we have to include objects in the same range from other shards
+            if (last_shard_id != 0 && last_shard_id != shard_it->first)
+                std::sort(stable, stable+stable_count);
+            if (stable_count > op->list_stable_limit)
+                stable_count = op->list_stable_limit;
+        }
+        last_shard_id = shard_it->first;
+    }
+    if (op->list_stable_limit == 0 && first_shard != last_shard)
+    {
+        // If that's not a per-PG listing, sort clean entries (already sorted if list_stable_limit != 0)
+        std::sort(stable, stable+stable_count);
+    }
+    int clean_stable_count = stable_count;
+    // Copy dirty_db entries (sorted, too)
+    int unstable_count = 0, unstable_alloc = 0;
+    obj_ver_id *unstable = NULL;
+    {
+        auto dirty_it = dirty_db.begin(), dirty_end = dirty_db.end();
+        if (op->min_oid.inode != 0 || op->min_oid.stripe != 0)
+        {
+            dirty_it = dirty_db.lower_bound({
+                .oid = op->min_oid,
+                .version = 0,
+            });
+        }
+        if ((max_oid.inode != 0 || max_oid.stripe != 0) && !(max_oid < op->min_oid))
+        {
+            dirty_end = dirty_db.upper_bound({
+                .oid = max_oid,
+                .version = UINT64_MAX,
+            });
+        }
+        for (; dirty_it != dirty_end; dirty_it++)
+        {
+            if (!pg_count || ((dirty_it->first.oid.stripe / pg_stripe_size) % pg_count + 1) == list_pg) // like map_to_pg()
+            {
+                if (IS_DELETE(dirty_it->second.state))
+                {
+                    // Deletions are always stable, so try to zero out two possible entries
+                    if (!replace_stable(dirty_it->first.oid, 0, 0, clean_stable_count, stable))
+                    {
+                        replace_stable(dirty_it->first.oid, 0, clean_stable_count, stable_count, stable);
+                    }
+                }
+                else if (IS_STABLE(dirty_it->second.state) || (dirty_it->second.state & BS_ST_INSTANT))
+                {
+                    // First try to replace a clean stable version in the first part of the list
+                    if (!replace_stable(dirty_it->first.oid, dirty_it->first.version, 0, clean_stable_count, stable))
+                    {
+                        // Then try to replace the last dirty stable version in the second part of the list
+                        if (stable_count > 0 && stable[stable_count-1].oid == dirty_it->first.oid)
+                        {
+                            stable[stable_count-1].version = dirty_it->first.version;
+                        }
+                        else
+                        {
+                            if (stable_count >= stable_alloc)
+                            {
+                                stable_alloc += 32768;
+                                obj_ver_id *nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
+                                if (!nst)
+                                {
+                                    if (unstable)
+                                        free(unstable);
+                                    op->retval = -ENOMEM;
+                                    FINISH_OP(op);
+                                    return;
+                                }
+                                stable = nst;
+                            }
+                            stable[stable_count++] = dirty_it->first;
+                        }
+                    }
+                    if (op->list_stable_limit > 0 && stable_count >= op->list_stable_limit)
+                    {
+                        // Stop here
+                        break;
+                    }
+                }
+                else
+                {
+                    if (unstable_count >= unstable_alloc)
+                    {
+                        unstable_alloc += 32768;
+                        obj_ver_id *nst = (obj_ver_id*)realloc(unstable, sizeof(obj_ver_id) * unstable_alloc);
+                        if (!nst)
+                        {
+                            if (stable)
+                                free(stable);
+                            op->retval = -ENOMEM;
+                            FINISH_OP(op);
+                            return;
+                        }
+                        unstable = nst;
+                    }
+                    unstable[unstable_count++] = dirty_it->first;
+                }
+            }
+        }
+    }
+    // Remove zeroed out stable entries
+    int j = 0;
+    for (int i = 0; i < stable_count; i++)
+    {
+        if (stable[i].version != 0)
+        {
+            stable[j++] = stable[i];
+        }
+    }
+    stable_count = j;
+    if (stable_count+unstable_count > stable_alloc)
+    {
+        stable_alloc = stable_count+unstable_count;
+        obj_ver_id *nst = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
+        if (!nst)
+        {
+            if (unstable)
+                free(unstable);
+            op->retval = -ENOMEM;
+            FINISH_OP(op);
+            return;
+        }
+        stable = nst;
+    }
+    // Copy unstable entries
+    for (int i = 0; i < unstable_count; i++)
+    {
+        stable[j++] = unstable[i];
+    }
+    free(unstable);
+    op->version = stable_count;
+    op->retval = stable_count+unstable_count;
+    op->buf = (uint8_t*)stable;
+    FINISH_OP(op);
+}
+
+void blockstore_impl_t::dump_diagnostics()
+{
+    journal.dump_diagnostics();
+    flusher->dump_diagnostics();
+}
+
+void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expected)
+{
+    if (retval == -EAGAIN)
+    {
+        fprintf(stderr, "EAGAIN error received from a disk %s during flush."
+            " It must never happen with io_uring and indicates a kernel bug."
+            " Please upgrade your kernel. Aborting.\n", op);
+        exit(1);
+    }
+    fprintf(stderr, "Disk %s failed: result is %d, expected %d. Can't continue, sorry :-(\n", op, retval, expected);
+    exit(1);
+}
+
+const std::map<uint64_t, uint64_t> & blockstore_impl_t::get_inode_space_stats()
+{
+    return inode_space_stats;
+}
+
+void blockstore_impl_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
+{
+    for (auto & np: no_inode_stats)
+    {
+        np.second = 2;
+    }
+    for (auto pool_id: pool_ids)
+    {
+        if (!no_inode_stats[pool_id])
+            recalc_inode_space_stats(pool_id, false);
+        no_inode_stats[pool_id] = 1;
+    }
+    for (auto np_it = no_inode_stats.begin(); np_it != no_inode_stats.end(); )
+    {
+        if (np_it->second == 2)
+        {
+            recalc_inode_space_stats(np_it->first, true);
+            no_inode_stats.erase(np_it++);
+        }
+        else
+            np_it++;
+    }
+}
+
+void blockstore_impl_t::recalc_inode_space_stats(uint64_t pool_id, bool per_inode)
+{
+    auto sp_begin = inode_space_stats.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    auto sp_end = inode_space_stats.lower_bound(((pool_id+1) << (64-POOL_ID_BITS)));
+    inode_space_stats.erase(sp_begin, sp_end);
+    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    while (sh_it != clean_db_shards.end() &&
+        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        for (auto & pair: sh_it->second)
+        {
+            uint64_t space_id = per_inode ? pair.first.inode : (pool_id << (64-POOL_ID_BITS));
+            inode_space_stats[space_id] += dsk.data_block_size;
+        }
+        sh_it++;
+    }
+    object_id last_oid = {};
+    bool last_exists = false;
+    auto dirty_it = dirty_db.lower_bound((obj_ver_id){ .oid = { .inode = (pool_id << (64-POOL_ID_BITS)) } });
+    while (dirty_it != dirty_db.end() && (dirty_it->first.oid.inode >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        if (IS_STABLE(dirty_it->second.state) && (IS_BIG_WRITE(dirty_it->second.state) || IS_DELETE(dirty_it->second.state)))
+        {
+            bool exists = false;
+            if (last_oid == dirty_it->first.oid)
+            {
+                exists = last_exists;
+            }
+            else
+            {
+                auto & clean_db = clean_db_shard(dirty_it->first.oid);
+                auto clean_it = clean_db.find(dirty_it->first.oid);
+                exists = clean_it != clean_db.end();
+            }
+            uint64_t space_id = per_inode ? dirty_it->first.oid.inode : (pool_id << (64-POOL_ID_BITS));
+            if (IS_BIG_WRITE(dirty_it->second.state))
+            {
+                if (!exists)
+                    inode_space_stats[space_id] += dsk.data_block_size;
+                last_exists = true;
+            }
+            else
+            {
+                if (exists)
+                {
+                    auto & sp = inode_space_stats[space_id];
+                    if (sp > dsk.data_block_size)
+                        sp -= dsk.data_block_size;
+                    else
+                        inode_space_stats.erase(space_id);
+                }
+                last_exists = false;
+            }
+            last_oid = dirty_it->first.oid;
+        }
+        dirty_it++;
+    }
+}
+
+std::string blockstore_impl_t::get_op_diag(blockstore_op_t *op)
+{
+    char buf[256];
+    auto priv = PRIV(op);
+    if (priv->wait_for)
+        snprintf(buf, sizeof(buf), "state=%d wait=%d (detail=%ju)", priv->op_state, priv->wait_for, priv->wait_detail);
+    else
+        snprintf(buf, sizeof(buf), "state=%d", priv->op_state);
+    return std::string(buf);
+}
+
+} // namespace v1

src/blockstore/v1/impl.h

@@ -0,0 +1,334 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#pragma once
+
+#include "blockstore.h"
+#include "blockstore_disk.h"
+#include "ondisk_formats.h"
+
+#include <sys/types.h>
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <time.h>
+#include <unistd.h>
+#include <linux/fs.h>
+
+#include <vector>
+#include <list>
+#include <deque>
+#include <new>
+#include <unordered_map>
+#include <unordered_set>
+
+#include "cpp-btree/btree_map.h"
+
+#include "malloc_or_die.h"
+#include "allocator.h"
+#include "crc32c.h"
+
+//#define BLOCKSTORE_DEBUG
+
+namespace v1 {
+
+#include "journal.h"
+
+// 32 = 16 + 16 bytes per "clean" entry in memory (object_id => clean_entry)
+struct __attribute__((__packed__)) clean_entry
+{
+    uint64_t version;
+    uint64_t location;
+};
+
+// 64 = 24 + 40 bytes per dirty entry in memory (obj_ver_id => dirty_entry). Plus checksums
+struct __attribute__((__packed__)) dirty_entry
+{
+    uint32_t state;
+    uint32_t flags;    // unneeded, but present for alignment
+    uint64_t location; // location in either journal or data -> in BYTES
+    uint32_t offset;   // data offset within object (stripe)
+    uint32_t len;      // data length
+    uint64_t journal_sector; // journal sector used for this entry
+    void* dyn_data;    // dynamic data: external bitmap and data block checksums. may be a pointer to the in-memory journal
+};
+
+// - Sync must be submitted after previous writes/deletes (not before!)
+// - Reads to the same object must be submitted after previous writes/deletes
+//   are written (not necessarily synced) in their location. This is because we
+//   rely on read-modify-write for erasure coding and we must return new data
+//   to calculate parity for subsequent writes
+// - Writes may be submitted in any order, because they don't overlap. Each write
+//   goes into a new location - either on the journal device or on the data device
+// - Stable (stabilize) must be submitted after sync of that object is completed
+//   It's even OK to return an error to the caller if that object is not synced yet
+// - Journal trim may be processed only after all versions are moved to
+//   the main storage AND after all read operations for older versions complete
+// - If an operation can not be submitted because the ring is full
+//   we should stop submission of other operations. Otherwise some "scatter" reads
+//   may end up blocked for a long time.
+// Otherwise, the submit order is free, that is all operations may be submitted immediately
+// In fact, adding a write operation must immediately result in dirty_db being populated
+
+struct used_clean_obj_t
+{
+    int refs;
+    bool was_freed; // was freed by a parallel flush?
+    bool was_changed; // was changed by a parallel flush?
+};
+
+// https://github.com/algorithm-ninja/cpp-btree
+// https://github.com/greg7mdp/sparsepp/ was used previously, but it was TERRIBLY slow after resizing
+// with sparsepp, random reads dropped to ~700 iops very fast with just as much as ~32k objects in the DB
+typedef btree::btree_map<object_id, clean_entry> blockstore_clean_db_t;
+typedef std::map<obj_ver_id, dirty_entry> blockstore_dirty_db_t;
+
+#include "init.h"
+
+#include "flush.h"
+
+struct blockstore_op_private_t
+{
+    // Wait status
+    int wait_for;
+    uint64_t wait_detail, wait_detail2;
+    int pending_ops;
+    int op_state;
+
+    // Read
+    uint64_t clean_block_used;
+    std::vector<copy_buffer_t> read_vec;
+
+    // Sync, write
+    uint64_t min_flushed_journal_sector, max_flushed_journal_sector;
+
+    // Write
+    struct iovec iov_zerofill[3];
+    // Warning: must not have a default value here because it's written to before calling constructor in blockstore_write.cpp O_o
+    uint64_t real_version;
+    timespec tv_begin;
+
+    // Sync
+    std::vector<obj_ver_id> sync_big_writes, sync_small_writes;
+};
+
+struct pool_shard_settings_t
+{
+    uint32_t pg_count;
+    uint32_t pg_stripe_size;
+};
+
+typedef uint64_t pool_pg_id_t;
+
+class blockstore_impl_t: public blockstore_i
+{
+    blockstore_disk_t dsk;
+
+    /******* OPTIONS *******/
+    bool readonly = false;
+    // It is safe to disable fsync() if drive write cache is writethrough
+    bool disable_data_fsync = false, disable_meta_fsync = false, disable_journal_fsync = false;
+    // Enable if you want every operation to be executed with an "implicit fsync"
+    // Suitable only for server SSDs with capacitors, requires disabled data and journal fsyncs
+    int immediate_commit = IMMEDIATE_NONE;
+    bool inmemory_meta = false;
+    // Maximum and minimum flusher count
+    unsigned max_flusher_count, min_flusher_count;
+    unsigned journal_trim_interval;
+    // Maximum queue depth
+    unsigned max_write_iodepth = 128;
+    // Enable small (journaled) write throttling, useful for the SSD+HDD case
+    bool throttle_small_writes = false;
+    // Target data device iops, bandwidth and parallelism for throttling (100/100/1 is the default for HDD)
+    int throttle_target_iops = 100;
+    int throttle_target_mbs = 100;
+    int throttle_target_parallelism = 1;
+    // Minimum difference in microseconds between target and real execution times to throttle the response
+    int throttle_threshold_us = 50;
+    // Maximum writes between automatically added fsync operations
+    uint64_t autosync_writes = 128;
+    // Log level (0-10)
+    int log_level = 0;
+    /******* END OF OPTIONS *******/
+
+    struct ring_consumer_t ring_consumer;
+
+    std::map<pool_id_t, pool_shard_settings_t> clean_db_settings;
+    std::map<pool_pg_id_t, blockstore_clean_db_t> clean_db_shards;
+    std::map<uint64_t, int> no_inode_stats;
+    std::map<uint64_t, uint64_t> inode_space_stats;
+    uint8_t *clean_bitmaps = NULL;
+    blockstore_dirty_db_t dirty_db;
+    std::vector<blockstore_op_t*> submit_queue;
+    std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
+    int unsynced_big_write_count = 0, unstable_unsynced = 0;
+    int unsynced_queued_ops = 0;
+    allocator_t *data_alloc = NULL;
+    uint64_t used_blocks = 0;
+    uint8_t *zero_object = NULL;
+
+    void *metadata_buffer = NULL;
+
+    struct journal_t journal;
+    journal_flusher_t *flusher;
+    int big_to_flush = 0;
+    int write_iodepth = 0;
+    bool alloc_dyn_data = false;
+
+    // clean data blocks referenced by read operations
+    std::map<uint64_t, used_clean_obj_t> used_clean_objects;
+
+    bool live = false, queue_stall = false;
+    ring_loop_i *ringloop;
+    timerfd_manager_t *tfd;
+
+    bool stop_sync_submitted;
+
+    inline struct io_uring_sqe* get_sqe()
+    {
+        return ringloop->get_sqe();
+    }
+
+    friend class blockstore_init_meta;
+    friend class blockstore_init_journal;
+    friend struct blockstore_journal_check_t;
+    friend class journal_flusher_t;
+    friend class journal_flusher_co;
+
+    void calc_lengths();
+    void open_data();
+    void open_meta();
+    void open_journal();
+    uint8_t* get_clean_entry_bitmap(uint64_t block_loc, int offset);
+
+    blockstore_clean_db_t& clean_db_shard(object_id oid);
+    void reshard_clean_db(pool_id_t pool_id, uint32_t pg_count, uint32_t pg_stripe_size);
+    void recalc_inode_space_stats(uint64_t pool_id, bool per_inode);
+
+    // Journaling
+    void prepare_journal_sector_write(int sector, blockstore_op_t *op);
+    void handle_journal_write(ring_data_t *data, uint64_t flush_id);
+    void disk_error_abort(const char *op, int retval, int expected);
+
+    // Asynchronous init
+    int initialized;
+    int metadata_buf_size;
+    blockstore_init_meta* metadata_init_reader;
+    blockstore_init_journal* journal_init_reader;
+
+    void check_wait(blockstore_op_t *op);
+    void init_op(blockstore_op_t *op);
+
+    // Read
+    int dequeue_read(blockstore_op_t *read_op);
+    void find_holes(std::vector<copy_buffer_t> & read_vec, uint32_t item_start, uint32_t item_end,
+        std::function<int(int, bool, uint32_t, uint32_t)> callback);
+    int fulfill_read(blockstore_op_t *read_op,
+        uint64_t &fulfilled, uint32_t item_start, uint32_t item_end,
+        uint32_t item_state, uint64_t item_version, uint64_t item_location,
+        uint64_t journal_sector, uint8_t *csum, int *dyn_data);
+    bool fulfill_clean_read(blockstore_op_t *read_op, uint64_t & fulfilled,
+        uint8_t *clean_entry_bitmap, int *dyn_data,
+        uint32_t item_start, uint32_t item_end, uint64_t clean_loc, uint64_t clean_ver);
+    int fill_partial_checksum_blocks(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled,
+        uint8_t *clean_entry_bitmap, int *dyn_data, bool from_journal, uint8_t *read_buf, uint64_t read_offset, uint64_t read_end);
+    int pad_journal_read(std::vector<copy_buffer_t> & rv, copy_buffer_t & cp,
+        uint64_t dirty_offset, uint64_t dirty_end, uint64_t dirty_loc, uint8_t *csum_ptr, int *dyn_data,
+        uint64_t offset, uint64_t submit_len, uint64_t & blk_begin, uint64_t & blk_end, uint8_t* & blk_buf);
+    bool read_range_fulfilled(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled, uint8_t *read_buf,
+        uint8_t *clean_entry_bitmap, uint32_t item_start, uint32_t item_end);
+    bool read_checksum_block(blockstore_op_t *op, int rv_pos, uint64_t &fulfilled, uint64_t clean_loc);
+    uint8_t* read_clean_meta_block(blockstore_op_t *read_op, uint64_t clean_loc, int rv_pos);
+    bool verify_padded_checksums(uint8_t *clean_entry_bitmap, uint8_t *csum_buf, uint32_t offset,
+        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
+    bool verify_journal_checksums(uint8_t *csums, uint32_t offset,
+        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
+    bool verify_clean_padded_checksums(blockstore_op_t *op, uint64_t clean_loc, uint8_t *dyn_data, bool from_journal,
+        iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
+    int fulfill_read_push(blockstore_op_t *op, void *buf, uint64_t offset, uint64_t len,
+        uint32_t item_state, uint64_t item_version);
+    void handle_read_event(ring_data_t *data, blockstore_op_t *op);
+
+    // Write
+    bool enqueue_write(blockstore_op_t *op);
+    void cancel_all_writes(blockstore_op_t *op, blockstore_dirty_db_t::iterator dirty_it, int retval);
+    int dequeue_write(blockstore_op_t *op);
+    int dequeue_del(blockstore_op_t *op);
+    int continue_write(blockstore_op_t *op);
+    void release_journal_sectors(blockstore_op_t *op);
+    void handle_write_event(ring_data_t *data, blockstore_op_t *op);
+
+    // Sync
+    int continue_sync(blockstore_op_t *op);
+    void ack_sync(blockstore_op_t *op);
+
+    // Stabilize
+    int dequeue_stable(blockstore_op_t *op);
+    int continue_stable(blockstore_op_t *op);
+    void mark_stable(obj_ver_id ov, bool forget_dirty = false);
+    void stabilize_object(object_id oid, uint64_t max_ver);
+    blockstore_op_t* selective_sync(blockstore_op_t *op);
+    int split_stab_op(blockstore_op_t *op, std::function<int(obj_ver_id v)> decider);
+
+    // Rollback
+    int dequeue_rollback(blockstore_op_t *op);
+    int continue_rollback(blockstore_op_t *op);
+    void mark_rolled_back(const obj_ver_id & ov);
+    void erase_dirty(blockstore_dirty_db_t::iterator dirty_start, blockstore_dirty_db_t::iterator dirty_end, uint64_t clean_loc);
+    void free_dirty_dyn_data(dirty_entry & e);
+
+    // List
+    void process_list(blockstore_op_t *op);
+
+public:
+
+    blockstore_impl_t(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd);
+    ~blockstore_impl_t();
+
+    void parse_config(blockstore_config_t & config);
+    void parse_config(blockstore_config_t & config, bool init);
+
+    // Event loop
+    void loop();
+
+    // Returns true when blockstore is ready to process operations
+    // (Although you're free to enqueue them before that)
+    bool is_started();
+
+    // Returns true when it's safe to destroy the instance. If destroying the instance
+    // requires to purge some queues, starts that process. Should be called in the event
+    // loop until it returns true.
+    bool is_safe_to_stop();
+
+    // Returns true if stalled
+    bool is_stalled();
+
+    // Submission
+    void enqueue_op(blockstore_op_t *op);
+
+    // Simplified synchronous operation: get object bitmap & current version
+    int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL);
+
+    // Unstable writes are added here (map of object_id -> version)
+    std::unordered_map<object_id, uint64_t> unstable_writes;
+
+    // Get space usage statistics
+    const std::map<uint64_t, uint64_t> & get_inode_space_stats();
+
+    // Set per-pool no_inode_stats
+    void set_no_inode_stats(const std::vector<uint64_t> & pool_ids);
+
+    // Print diagnostics to stdout
+    void dump_diagnostics();
+
+    // Get diagnostic string for an operation
+    std::string get_op_diag(blockstore_op_t *op);
+
+    inline uint32_t get_block_size() { return dsk.data_block_size; }
+    inline uint64_t get_block_count() { return dsk.block_count; }
+    inline uint64_t get_free_block_count() { return dsk.block_count - used_blocks; }
+    inline uint32_t get_bitmap_granularity() { return dsk.disk_alignment; }
+    inline uint64_t get_journal_size() { return dsk.journal_len; }
+};
+
+} // namespace v1

src/blockstore/v1/init.h

@@ -0,0 +1,71 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#pragma once
+
+struct blockstore_init_meta_buf
+{
+    uint8_t *buf = NULL;
+    uint64_t size = 0;
+    uint64_t offset = 0;
+    int state = 0;
+};
+
+class blockstore_init_meta
+{
+    blockstore_impl_t *bs;
+    int wait_state = 0;
+    bool zero_on_init = false;
+    void *metadata_buffer = NULL;
+    blockstore_init_meta_buf bufs[2] = {};
+    int submitted = 0;
+    struct io_uring_sqe *sqe;
+    struct ring_data_t *data;
+    uint64_t md_offset = 0;
+    uint64_t next_offset = 0;
+    uint64_t last_read_offset = 0;
+    uint64_t entries_loaded = 0;
+    unsigned entries_per_block = 0;
+    int i = 0, j = 0;
+    std::vector<uint64_t> entries_to_zero;
+    bool handle_meta_block(uint8_t *buf, uint64_t count, uint64_t done_cnt);
+    void handle_event(ring_data_t *data, int buf_num);
+public:
+    blockstore_init_meta(blockstore_impl_t *bs);
+    int loop();
+};
+
+struct bs_init_journal_done
+{
+    void *buf;
+    uint64_t pos, len;
+};
+
+class blockstore_init_journal
+{
+    blockstore_impl_t *bs;
+    int wait_state = 0, wait_count = 0, handle_res = 0;
+    uint64_t entries_loaded = 0;
+    uint32_t crc32_last = 0;
+    bool started = false;
+    uint64_t next_free;
+    std::vector<bs_init_journal_done> done;
+    std::vector<obj_ver_id> double_allocs;
+    std::vector<iovec> small_write_data;
+    uint64_t journal_pos = 0;
+    uint64_t continue_pos = 0;
+    void *init_write_buf = NULL;
+    uint64_t init_write_sector = 0;
+    bool wrapped = false;
+    void *submitted_buf;
+    struct io_uring_sqe *sqe;
+    struct ring_data_t *data;
+    journal_entry_start *je_start;
+    std::function<void(ring_data_t*)> simple_callback;
+    int handle_journal_part(void *buf, uint64_t done_pos, uint64_t len);
+    void handle_event(ring_data_t *data);
+    void erase_dirty_object(blockstore_dirty_db_t::iterator dirty_it);
+public:
+    blockstore_init_journal(blockstore_impl_t* bs);
+    int loop();
+};

src/blockstore/v1/internal.h

@@ -0,0 +1,85 @@
+#pragma once
+
+// States are not stored on disk. Instead, they're deduced from the journal
+
+#define BS_ST_SMALL_WRITE 0x01
+#define BS_ST_BIG_WRITE 0x02
+#define BS_ST_DELETE 0x03
+
+#define BS_ST_WAIT_DEL 0x10
+#define BS_ST_WAIT_BIG 0x20
+#define BS_ST_IN_FLIGHT 0x30
+#define BS_ST_SUBMITTED 0x40
+#define BS_ST_WRITTEN 0x50
+#define BS_ST_SYNCED 0x60
+#define BS_ST_STABLE 0x70
+
+#define BS_ST_INSTANT 0x100
+
+#define BS_ST_TYPE_MASK 0x0F
+#define BS_ST_WORKFLOW_MASK 0xF0
+#define IS_IN_FLIGHT(st) (((st) & 0xF0) <= BS_ST_SUBMITTED)
+#define IS_STABLE(st) (((st) & 0xF0) == BS_ST_STABLE)
+#define IS_SYNCED(st) (((st) & 0xF0) >= BS_ST_SYNCED)
+#define IS_JOURNAL(st) (((st) & 0x0F) == BS_ST_SMALL_WRITE)
+#define IS_BIG_WRITE(st) (((st) & 0x0F) == BS_ST_BIG_WRITE)
+#define IS_DELETE(st) (((st) & 0x0F) == BS_ST_DELETE)
+#define IS_INSTANT(st) (((st) & BS_ST_TYPE_MASK) == BS_ST_DELETE || ((st) & BS_ST_INSTANT))
+
+#define BS_SUBMIT_CHECK_SQES(n) \
+    if (ringloop->space_left() < (n))\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = (n);\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define BS_SUBMIT_GET_SQE(sqe, data) \
+    BS_SUBMIT_GET_ONLY_SQE(sqe); \
+    struct ring_data_t *data = ((ring_data_t*)sqe->user_data)
+
+#define BS_SUBMIT_GET_ONLY_SQE(sqe) \
+    struct io_uring_sqe *sqe = get_sqe();\
+    if (!sqe)\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define BS_SUBMIT_GET_SQE_DECL(sqe) \
+    sqe = get_sqe();\
+    if (!sqe)\
+    {\
+        /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
+        PRIV(op)->wait_for = WAIT_SQE;\
+        return 0;\
+    }
+
+#define PRIV(op) ((blockstore_op_private_t*)(op)->private_data)
+#define FINISH_OP(op) PRIV(op)->~blockstore_op_private_t(); std::function<void (blockstore_op_t*)>(op->callback)(op)
+
+// Suspend operation until there are more free SQEs
+#define WAIT_SQE 1
+// Suspend operation until there are <wait_detail> bytes of free space in the journal on disk
+#define WAIT_JOURNAL 3
+// Suspend operation until the next journal sector buffer is free
+#define WAIT_JOURNAL_BUFFER 4
+// Suspend operation until there is some free space on the data device
+#define WAIT_FREE 5
+
+#define COPY_BUF_JOURNAL 1
+#define COPY_BUF_DATA 2
+#define COPY_BUF_ZERO 4
+#define COPY_BUF_CSUM_FILL 8
+#define COPY_BUF_COALESCED 16
+#define COPY_BUF_META_BLOCK 32
+#define COPY_BUF_JOURNALED_BIG 64
+
+#define STAB_SPLIT_DONE 1
+#define STAB_SPLIT_WAIT 2
+#define STAB_SPLIT_SYNC 3
+#define STAB_SPLIT_TODO 4

src/blockstore/v1/journal.cpp

@@ -1,7 +1,10 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#include "blockstore_impl.h"
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
 
 blockstore_journal_check_t::blockstore_journal_check_t(blockstore_impl_t *bs)
 {
@@ -178,7 +181,7 @@ void blockstore_impl_t::prepare_journal_sector_write(int cur_sector, blockstore_
         // Caller must ensure availability of an SQE
         assert(sqe != NULL);
         ring_data_t *data = ((ring_data_t*)sqe->user_data);
-        journal.sector_info[cur_sector].written = true;
+        // <written> flag will be set at the moment of actual submission
         journal.sector_info[cur_sector].submit_id = ++journal.submit_id;
         assert(journal.submit_id != 0); // check overflow
         journal.submitting_sectors.push_back(cur_sector);
@@ -326,3 +329,5 @@ void journal_t::dump_diagnostics()
         journal_used_it == used_sectors.end() ? 0 : journal_used_it->second
     );
 }
+
+} // namespace v1

src/blockstore/v1/journal.h

@@ -0,0 +1,78 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#pragma once
+
+class blockstore_impl_t;
+
+struct journal_sector_info_t
+{
+    uint64_t offset;
+    uint64_t flush_count;
+    bool written;
+    bool dirty;
+    uint64_t submit_id;
+};
+
+struct pending_journaling_t
+{
+    int pending;
+    int sector;
+    blockstore_op_t *op;
+};
+
+struct journal_t
+{
+    int fd;
+    bool inmemory = false;
+    bool flush_journal = false;
+    void *buffer = NULL;
+
+    uint64_t block_size;
+    uint64_t offset, len;
+    // Next free block offset
+    uint64_t next_free = 0;
+    // First occupied block offset
+    uint64_t used_start = 0;
+    // End of the last block not used for writing anymore
+    uint64_t dirty_start = 0;
+    uint32_t crc32_last = 0;
+
+    // Current sector(s) used for writing
+    void *sector_buf = NULL;
+    journal_sector_info_t *sector_info = NULL;
+    uint64_t sector_count;
+    bool no_same_sector_overwrites = false;
+    int cur_sector = 0;
+    int in_sector_pos = 0;
+    std::vector<int> submitting_sectors;
+    std::multimap<uint64_t, pending_journaling_t> flushing_ops;
+    uint64_t submit_id = 0;
+
+    // Used sector map
+    // May use ~ 80 MB per 1 GB of used journal space in the worst case
+    std::map<uint64_t, uint64_t> used_sectors;
+
+    ~journal_t();
+    bool trim();
+    uint64_t get_trim_pos();
+    void dump_diagnostics();
+    inline bool entry_fits(int size)
+    {
+        return !(block_size - in_sector_pos < size ||
+            no_same_sector_overwrites && sector_info[cur_sector].written);
+    }
+};
+
+struct blockstore_journal_check_t
+{
+    blockstore_impl_t *bs;
+    uint64_t next_pos, next_sector, next_in_pos;
+    int sectors_to_write, first_sector;
+    bool right_dir; // writing to the end or the beginning of the ring buffer
+
+    blockstore_journal_check_t(blockstore_impl_t *bs);
+    int check_available(blockstore_op_t *op, int required, int size, int data_after);
+};
+
+journal_entry* prefill_single_journal_entry(journal_t & journal, uint16_t type, uint32_t size);

src/blockstore/v1/open.cpp

@@ -0,0 +1,187 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include <sys/file.h>
+#include "impl.h"
+
+namespace v1 {
+
+void blockstore_impl_t::parse_config(blockstore_config_t & config)
+{
+    return parse_config(config, false);
+}
+
+void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
+{
+    // Online-configurable options:
+    max_flusher_count = strtoull(config["max_flusher_count"].c_str(), NULL, 10);
+    if (!max_flusher_count)
+    {
+        max_flusher_count = strtoull(config["flusher_count"].c_str(), NULL, 10);
+    }
+    min_flusher_count = strtoull(config["min_flusher_count"].c_str(), NULL, 10);
+    journal_trim_interval = strtoull(config["journal_trim_interval"].c_str(), NULL, 10);
+    max_write_iodepth = strtoull(config["max_write_iodepth"].c_str(), NULL, 10);
+    throttle_small_writes = config["throttle_small_writes"] == "true" || config["throttle_small_writes"] == "1" || config["throttle_small_writes"] == "yes";
+    throttle_target_iops = strtoull(config["throttle_target_iops"].c_str(), NULL, 10);
+    throttle_target_mbs = strtoull(config["throttle_target_mbs"].c_str(), NULL, 10);
+    throttle_target_parallelism = strtoull(config["throttle_target_parallelism"].c_str(), NULL, 10);
+    throttle_threshold_us = strtoull(config["throttle_threshold_us"].c_str(), NULL, 10);
+    if (config["autosync_writes"] != "")
+    {
+        autosync_writes = strtoull(config["autosync_writes"].c_str(), NULL, 10);
+    }
+    if (!max_flusher_count)
+    {
+        max_flusher_count = 256;
+    }
+    if (!min_flusher_count || journal.flush_journal)
+    {
+        min_flusher_count = 1;
+    }
+    if (!journal_trim_interval)
+    {
+        journal_trim_interval = 512;
+    }
+    if (!max_write_iodepth)
+    {
+        max_write_iodepth = 128;
+    }
+    if (!throttle_target_iops)
+    {
+        throttle_target_iops = 100;
+    }
+    if (!throttle_target_mbs)
+    {
+        throttle_target_mbs = 100;
+    }
+    if (!throttle_target_parallelism)
+    {
+        throttle_target_parallelism = 1;
+    }
+    if (!throttle_threshold_us)
+    {
+        throttle_threshold_us = 50;
+    }
+    if (!init)
+    {
+        return;
+    }
+    // Offline-configurable options:
+    // Common disk options
+    dsk.parse_config(config);
+    // Parse
+    if (config["readonly"] == "true" || config["readonly"] == "1" || config["readonly"] == "yes")
+    {
+        readonly = true;
+    }
+    if (config["disable_data_fsync"] == "true" || config["disable_data_fsync"] == "1" || config["disable_data_fsync"] == "yes")
+    {
+        disable_data_fsync = true;
+    }
+    if (config["disable_meta_fsync"] == "true" || config["disable_meta_fsync"] == "1" || config["disable_meta_fsync"] == "yes")
+    {
+        disable_meta_fsync = true;
+    }
+    if (config["disable_journal_fsync"] == "true" || config["disable_journal_fsync"] == "1" || config["disable_journal_fsync"] == "yes")
+    {
+        disable_journal_fsync = true;
+    }
+    if (config["flush_journal"] == "true" || config["flush_journal"] == "1" || config["flush_journal"] == "yes")
+    {
+        // Only flush journal and exit
+        journal.flush_journal = true;
+    }
+    if (config["immediate_commit"] == "all")
+    {
+        immediate_commit = IMMEDIATE_ALL;
+    }
+    else if (config["immediate_commit"] == "small")
+    {
+        immediate_commit = IMMEDIATE_SMALL;
+    }
+    metadata_buf_size = strtoull(config["meta_buf_size"].c_str(), NULL, 10);
+    inmemory_meta = config["inmemory_metadata"] != "false" && config["inmemory_metadata"] != "0" &&
+        config["inmemory_metadata"] != "no";
+    journal.sector_count = strtoull(config["journal_sector_buffer_count"].c_str(), NULL, 10);
+    journal.no_same_sector_overwrites = config["journal_no_same_sector_overwrites"] == "true" ||
+        config["journal_no_same_sector_overwrites"] == "1" || config["journal_no_same_sector_overwrites"] == "yes";
+    journal.inmemory = config["inmemory_journal"] != "false" && config["inmemory_journal"] != "0" &&
+        config["inmemory_journal"] != "no";
+    log_level = strtoull(config["log_level"].c_str(), NULL, 10);
+    // Validate
+    if (journal.sector_count < 2)
+    {
+        journal.sector_count = 32;
+    }
+    if (metadata_buf_size < 65536)
+    {
+        metadata_buf_size = 4*1024*1024;
+    }
+    if (dsk.meta_device == dsk.data_device)
+    {
+        disable_meta_fsync = disable_data_fsync;
+    }
+    if (dsk.journal_device == dsk.meta_device)
+    {
+        disable_journal_fsync = disable_meta_fsync;
+    }
+    if (immediate_commit != IMMEDIATE_NONE && !disable_journal_fsync)
+    {
+        throw std::runtime_error("immediate_commit requires disable_journal_fsync");
+    }
+    if (immediate_commit == IMMEDIATE_ALL && !disable_data_fsync)
+    {
+        throw std::runtime_error("immediate_commit=all requires disable_journal_fsync and disable_data_fsync");
+    }
+    // init some fields
+    journal.block_size = dsk.journal_block_size;
+    journal.next_free = dsk.journal_block_size;
+    journal.used_start = dsk.journal_block_size;
+    // no free space because sector is initially unmapped
+    journal.in_sector_pos = dsk.journal_block_size;
+}
+
+void blockstore_impl_t::calc_lengths()
+{
+    dsk.calc_lengths();
+    journal.len = dsk.journal_len;
+    journal.block_size = dsk.journal_block_size;
+    journal.offset = dsk.journal_offset;
+    if (inmemory_meta)
+    {
+        metadata_buffer = memalign(MEM_ALIGNMENT, dsk.meta_area_size);
+        if (!metadata_buffer)
+            throw std::runtime_error("Failed to allocate memory for the metadata ("+std::to_string(dsk.meta_area_size/1024/1024)+" MB)");
+    }
+    else if (dsk.clean_entry_bitmap_size || dsk.data_csum_type)
+    {
+        clean_bitmaps = (uint8_t*)malloc(dsk.block_count * 2 * dsk.clean_entry_bitmap_size);
+        if (!clean_bitmaps)
+        {
+            throw std::runtime_error(
+                "Failed to allocate memory for the metadata sparse write bitmap ("+
+                std::to_string(dsk.block_count * 2 * dsk.clean_entry_bitmap_size / 1024 / 1024)+" MB)"
+            );
+        }
+    }
+    if (journal.inmemory)
+    {
+        journal.buffer = memalign(MEM_ALIGNMENT, journal.len);
+        if (!journal.buffer)
+            throw std::runtime_error("Failed to allocate memory for journal ("+std::to_string(journal.len/1024/1024)+" MB)");
+    }
+    else
+    {
+        journal.sector_buf = (uint8_t*)memalign(MEM_ALIGNMENT, journal.sector_count * dsk.journal_block_size);
+        if (!journal.sector_buf)
+            throw std::bad_alloc();
+    }
+    journal.sector_info = (journal_sector_info_t*)calloc(journal.sector_count, sizeof(journal_sector_info_t));
+    if (!journal.sector_info)
+    {
+        throw std::bad_alloc();
+    }
+}
+
+}

src/blockstore/v1/rollback.cpp

@@ -1,7 +1,10 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#include "blockstore_impl.h"
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
 
 int blockstore_impl_t::dequeue_rollback(blockstore_op_t *op)
 {
@@ -210,10 +213,10 @@ void blockstore_impl_t::erase_dirty(blockstore_dirty_db_t::iterator dirty_start,
             dirty_it->second.location != UINT64_MAX)
         {
 #ifdef BLOCKSTORE_DEBUG
-            printf("Free block %ju from %jx:%jx v%ju\n", dirty_it->second.location >> dsk.block_order,
+            printf("Free block %ju from %jx:%jx v%ju\n", dirty_it->second.location / dsk.data_block_size,
                 dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version);
 #endif
-            data_alloc->set(dirty_it->second.location >> dsk.block_order, false);
+            data_alloc->set(dirty_it->second.location / dsk.data_block_size, false);
         }
         auto used = --journal.used_sectors.at(dirty_it->second.journal_sector);
 #ifdef BLOCKSTORE_DEBUG
@@ -256,3 +259,5 @@ void blockstore_impl_t::free_dirty_dyn_data(dirty_entry & e)
         e.dyn_data = NULL;
     }
 }
+
+} // namespace v1

src/blockstore/v1/stable.cpp

@@ -0,0 +1,566 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
+
+// Stabilize small write:
+// 1) Copy data from the journal to the data device
+// 2) Increase version on the metadata device and sync it
+// 3) Advance clean_db entry's version, clear previous journal entries
+//
+// This makes 1 4K small write+sync look like:
+// 512b+4K (journal) + sync + 512b (journal) + sync + 4K (data) [+ sync?] + 512b (metadata) + sync.
+// WA = 2.375. It's not the best, SSD FTL-like redirect-write could probably be lower
+// even with defragmentation. But it's fixed and it's still better than in Ceph. :)
+// except for HDD-only clusters, because each write results in 3 seeks.
+
+// Stabilize big write:
+// 1) Copy metadata from the journal to the metadata device
+// 2) Move dirty_db entry to clean_db and clear previous journal entries
+//
+// This makes 1 128K big write+sync look like:
+// 128K (data) + sync + 512b (journal) + sync + 512b (journal) + sync + 512b (metadata) + sync.
+// WA = 1.012. Very good :)
+
+// Stabilize delete:
+// 1) Remove metadata entry and sync it
+// 2) Remove dirty_db entry and clear previous journal entries
+// We have 2 problems here:
+// - In the cluster environment, we must store the "tombstones" of deleted objects until
+//   all replicas (not just quorum) agrees about their deletion. That is, "stabilize" is
+//   not possible for deletes in degraded placement groups
+// - With simple "fixed" metadata tables we can't just clear the metadata entry of the latest
+//   object version. We must clear all previous entries, too.
+// FIXME Fix both problems - probably, by switching from "fixed" metadata tables to "dynamic"
+
+// AND We must do it in batches, for the sake of reduced fsync call count
+// AND We must know what we stabilize. Basic workflow is like:
+// 1) primary OSD receives sync request
+// 2) it submits syncs to blockstore and peers
+// 3) after everyone acks sync it acks sync to the client
+// 4) after a while it takes his synced object list and sends stabilize requests
+//    to peers and to its own blockstore, thus freeing the old version
+
+struct ver_vector_t
+{
+    obj_ver_id *items = NULL;
+    uint64_t alloc = 0, size = 0;
+};
+
+static void init_versions(ver_vector_t & vec, obj_ver_id *start, obj_ver_id *end, uint64_t len)
+{
+    if (!vec.items)
+    {
+        vec.alloc = len;
+        vec.items = (obj_ver_id*)malloc_or_die(sizeof(obj_ver_id) * vec.alloc);
+        for (auto sv = start; sv < end; sv++)
+        {
+            vec.items[vec.size++] = *sv;
+        }
+    }
+}
+
+static void append_version(ver_vector_t & vec, obj_ver_id ov)
+{
+    if (vec.size >= vec.alloc)
+    {
+        vec.alloc = !vec.alloc ? 4 : vec.alloc*2;
+        vec.items = (obj_ver_id*)realloc_or_die(vec.items, sizeof(obj_ver_id) * vec.alloc);
+    }
+    vec.items[vec.size++] = ov;
+}
+
+static bool check_unsynced(std::vector<obj_ver_id> & check, obj_ver_id ov, std::vector<obj_ver_id> & to, int *count)
+{
+    bool found = false;
+    int j = 0, k = 0;
+    while (j < check.size())
+    {
+        if (check[j] == ov)
+            found = true;
+        if (check[j].oid == ov.oid && check[j].version <= ov.version)
+        {
+            to.push_back(check[j++]);
+            if (count)
+                (*count)--;
+        }
+        else
+            check[k++] = check[j++];
+    }
+    check.resize(k);
+    return found;
+}
+
+blockstore_op_t* blockstore_impl_t::selective_sync(blockstore_op_t *op)
+{
+    unsynced_big_write_count -= unsynced_big_writes.size();
+    unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
+    unsynced_big_write_count += unsynced_big_writes.size();
+    unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
+    // Create a sync operation, insert into the end of the queue
+    // And move ourselves into the end too!
+    // Rather hacky but that's what we need...
+    blockstore_op_t *sync_op = new blockstore_op_t;
+    sync_op->opcode = BS_OP_SYNC;
+    sync_op->buf = NULL;
+    sync_op->callback = [](blockstore_op_t *sync_op)
+    {
+        delete sync_op;
+    };
+    init_op(sync_op);
+    int sync_res = continue_sync(sync_op);
+    if (sync_res != 2)
+    {
+        // Put SYNC into the queue if it's not finished yet
+        submit_queue.push_back(sync_op);
+    }
+    // Restore unsynced_writes
+    unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
+    unsynced_big_write_count -= unsynced_big_writes.size();
+    unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
+    unsynced_big_write_count += unsynced_big_writes.size();
+    if (sync_res == 2)
+    {
+        // Sync is immediately completed
+        return NULL;
+    }
+    return sync_op;
+}
+
+// Returns: 2 = stop processing and dequeue, 0 = stop processing and do not dequeue, 1 = proceed with op itself
+int blockstore_impl_t::split_stab_op(blockstore_op_t *op, std::function<int(obj_ver_id v)> decider)
+{
+    bool add_sync = false;
+    ver_vector_t good_vers, bad_vers;
+    obj_ver_id* v;
+    int i, todo = 0;
+    for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
+    {
+        int action = decider(*v);
+        if (action < 0)
+        {
+            // Rollback changes
+            for (auto & ov: PRIV(op)->sync_big_writes)
+            {
+                unsynced_big_writes.push_back(ov);
+                unsynced_big_write_count++;
+            }
+            for (auto & ov: PRIV(op)->sync_small_writes)
+            {
+                unsynced_small_writes.push_back(ov);
+            }
+            free(good_vers.items);
+            good_vers.items = NULL;
+            free(bad_vers.items);
+            bad_vers.items = NULL;
+            // Error
+            op->retval = action;
+            FINISH_OP(op);
+            return 2;
+        }
+        else if (action == STAB_SPLIT_DONE)
+        {
+            // Already done
+            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
+        }
+        else if (action == STAB_SPLIT_WAIT)
+        {
+            // Already in progress, we just have to wait until it finishes
+            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
+            append_version(bad_vers, *v);
+        }
+        else if (action == STAB_SPLIT_SYNC)
+        {
+            // Needs a SYNC, we have to send a SYNC if not already in progress
+            //
+            // If the object is not present in unsynced_(big|small)_writes then
+            // it's currently being synced. If it's present then we can initiate
+            // its sync ourselves.
+            init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
+            append_version(bad_vers, *v);
+            if (!add_sync)
+            {
+                PRIV(op)->sync_big_writes.clear();
+                PRIV(op)->sync_small_writes.clear();
+                add_sync = true;
+            }
+            check_unsynced(unsynced_small_writes, *v, PRIV(op)->sync_small_writes, NULL);
+            check_unsynced(unsynced_big_writes, *v, PRIV(op)->sync_big_writes, &unsynced_big_write_count);
+        }
+        else /* if (action == STAB_SPLIT_TODO) */
+        {
+            if (good_vers.items)
+            {
+                // If we're selecting versions then append it
+                // Main idea is that 99% of the time all versions passed to BS_OP_STABLE are synced
+                // And we don't want to select/allocate anything in that optimistic case
+                append_version(good_vers, *v);
+            }
+            todo++;
+        }
+    }
+    // In a pessimistic scenario, an operation may be split into 3:
+    // - Stabilize synced entries
+    // - Sync unsynced entries
+    // - Continue for unsynced entries after sync
+    add_sync = add_sync && (PRIV(op)->sync_big_writes.size() || PRIV(op)->sync_small_writes.size());
+    if (!todo && !bad_vers.size)
+    {
+        // Already stable
+        op->retval = 0;
+        FINISH_OP(op);
+        return 2;
+    }
+    op->retval = 0;
+    if (!todo && !add_sync)
+    {
+        // Only wait for inflight writes or current in-progress syncs
+        return 0;
+    }
+    blockstore_op_t *sync_op = NULL, *split_stab_op = NULL;
+    if (add_sync)
+    {
+        // Initiate a selective sync for PRIV(op)->sync_(big|small)_writes
+        sync_op = selective_sync(op);
+    }
+    if (bad_vers.size)
+    {
+        // Split part of the request into a separate operation
+        split_stab_op = new blockstore_op_t;
+        split_stab_op->opcode = op->opcode;
+        split_stab_op->buf = (uint8_t*)bad_vers.items;
+        split_stab_op->len = bad_vers.size;
+        init_op(split_stab_op);
+        submit_queue.push_back(split_stab_op);
+    }
+    if (sync_op || split_stab_op || good_vers.items)
+    {
+        uint8_t *orig_buf = op->buf;
+        if (good_vers.items)
+        {
+            op->buf = (uint8_t*)good_vers.items;
+            op->len = good_vers.size;
+        }
+        // Make a wrapped callback
+        int *split_op_counter = (int*)malloc_or_die(sizeof(int));
+        *split_op_counter = (sync_op ? 1 : 0) + (split_stab_op ? 1 : 0) + (todo ? 1 : 0);
+        auto cb = [op, good_items = good_vers.items,
+            bad_items = bad_vers.items, split_op_counter,
+            orig_buf, real_cb = op->callback](blockstore_op_t *split_op)
+        {
+            if (split_op->retval != 0)
+                op->retval = split_op->retval;
+            (*split_op_counter)--;
+            assert((*split_op_counter) >= 0);
+            if (op != split_op)
+                delete split_op;
+            if (!*split_op_counter)
+            {
+                free(good_items);
+                free(bad_items);
+                free(split_op_counter);
+                op->buf = orig_buf;
+                real_cb(op);
+            }
+        };
+        if (sync_op)
+        {
+            sync_op->callback = cb;
+        }
+        if (split_stab_op)
+        {
+            split_stab_op->callback = cb;
+        }
+        op->callback = cb;
+    }
+    if (!todo)
+    {
+        // All work is postponed
+        op->callback = NULL;
+        return 2;
+    }
+    return 1;
+}
+
+int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
+{
+    if (PRIV(op)->op_state)
+    {
+        return continue_stable(op);
+    }
+    int r = split_stab_op(op, [this](obj_ver_id ov)
+    {
+        auto dirty_it = dirty_db.find(ov);
+        if (dirty_it == dirty_db.end())
+        {
+            auto & clean_db = clean_db_shard(ov.oid);
+            auto clean_it = clean_db.find(ov.oid);
+            if (clean_it == clean_db.end() || clean_it->second.version < ov.version)
+            {
+                // No such object version
+                printf("Error: %jx:%jx v%ju not found while stabilizing\n", ov.oid.inode, ov.oid.stripe, ov.version);
+                return -ENOENT;
+            }
+            else
+            {
+                // Already stable
+                return STAB_SPLIT_DONE;
+            }
+        }
+        else if (IS_STABLE(dirty_it->second.state))
+        {
+            // Already stable
+            return STAB_SPLIT_DONE;
+        }
+        while (true)
+        {
+            if (IS_IN_FLIGHT(dirty_it->second.state))
+            {
+                // Object write is still in progress. Wait until the write request completes
+                return STAB_SPLIT_WAIT;
+            }
+            else if (!IS_SYNCED(dirty_it->second.state))
+            {
+                // Object not synced yet - sync it
+                // In previous versions we returned EBUSY here and required
+                // the caller (OSD) to issue a global sync first. But a global sync
+                // waits for all writes in the queue including inflight writes. And
+                // inflight writes may themselves be blocked by unstable writes being
+                // still present in the journal and not flushed away from it.
+                // So we must sync specific objects here.
+                //
+                // Even more, we have to process "stabilize" request in parts. That is,
+                // we must stabilize all objects which are already synced. Otherwise
+                // they may block objects which are NOT synced yet.
+                return STAB_SPLIT_SYNC;
+            }
+            else if (IS_STABLE(dirty_it->second.state))
+            {
+                break;
+            }
+            // Check previous versions too
+            if (dirty_it == dirty_db.begin())
+            {
+                break;
+            }
+            dirty_it--;
+            if (dirty_it->first.oid != ov.oid)
+            {
+                break;
+            }
+        }
+        return STAB_SPLIT_TODO;
+    });
+    if (r != 1)
+    {
+        return r;
+    }
+    // Check journal space
+    blockstore_journal_check_t space_check(this);
+    if (!space_check.check_available(op, op->len, sizeof(journal_entry_stable), 0))
+    {
+        return 0;
+    }
+    // There is sufficient space. Check SQEs
+    BS_SUBMIT_CHECK_SQES(space_check.sectors_to_write);
+    // Prepare and submit journal entries
+    int s = 0;
+    auto v = (obj_ver_id*)op->buf;
+    for (int i = 0; i < op->len; i++, v++)
+    {
+        if (!journal.entry_fits(sizeof(journal_entry_stable)) &&
+            journal.sector_info[journal.cur_sector].dirty)
+        {
+            prepare_journal_sector_write(journal.cur_sector, op);
+            s++;
+        }
+        journal_entry_stable *je = (journal_entry_stable*)
+            prefill_single_journal_entry(journal, JE_STABLE, sizeof(journal_entry_stable));
+        je->oid = v->oid;
+        je->version = v->version;
+        je->crc32 = je_crc32((journal_entry*)je);
+        journal.crc32_last = je->crc32;
+    }
+    prepare_journal_sector_write(journal.cur_sector, op);
+    s++;
+    assert(s == space_check.sectors_to_write);
+    PRIV(op)->op_state = 1;
+    return 1;
+}
+
+int blockstore_impl_t::continue_stable(blockstore_op_t *op)
+{
+    if (PRIV(op)->op_state == 2)
+        goto resume_2;
+    else if (PRIV(op)->op_state == 4)
+        goto resume_4;
+    else
+        return 1;
+resume_2:
+    if (!disable_journal_fsync)
+    {
+        BS_SUBMIT_GET_SQE(sqe, data);
+        io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
+        data->iov = { 0 };
+        data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
+        PRIV(op)->pending_ops = 1;
+        PRIV(op)->op_state = 3;
+        return 1;
+    }
+resume_4:
+    // Mark dirty_db entries as stable, acknowledge op completion
+    obj_ver_id* v;
+    int i;
+    for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
+    {
+        // Mark all dirty_db entries up to op->version as stable
+#ifdef BLOCKSTORE_DEBUG
+        printf("Stabilize %jx:%jx v%ju\n", v->oid.inode, v->oid.stripe, v->version);
+#endif
+        mark_stable(*v);
+    }
+    // Acknowledge op
+    op->retval = 0;
+    FINISH_OP(op);
+    return 2;
+}
+
+void blockstore_impl_t::mark_stable(obj_ver_id v, bool forget_dirty)
+{
+    auto dirty_it = dirty_db.find(v);
+    if (dirty_it != dirty_db.end())
+    {
+        if (IS_INSTANT(dirty_it->second.state))
+        {
+            // 'Instant' (non-EC) operations may complete and try to become stable out of order. Prevent it.
+            auto back_it = dirty_it;
+            while (back_it != dirty_db.begin())
+            {
+                back_it--;
+                if (back_it->first.oid != v.oid)
+                {
+                    break;
+                }
+                if (!IS_STABLE(back_it->second.state))
+                {
+                    // There are preceding unstable versions, can't flush <v>
+                    return;
+                }
+            }
+            while (true)
+            {
+                dirty_it++;
+                if (dirty_it == dirty_db.end() || dirty_it->first.oid != v.oid ||
+                    !IS_SYNCED(dirty_it->second.state))
+                {
+                    dirty_it--;
+                    break;
+                }
+                v.version = dirty_it->first.version;
+            }
+        }
+        while (1)
+        {
+            bool was_stable = IS_STABLE(dirty_it->second.state);
+            if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_SYNCED)
+            {
+                dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_STABLE;
+                // Allocations and deletions are counted when they're stabilized
+                if (IS_BIG_WRITE(dirty_it->second.state))
+                {
+                    int exists = -1;
+                    if (dirty_it != dirty_db.begin())
+                    {
+                        auto prev_it = dirty_it;
+                        prev_it--;
+                        if (prev_it->first.oid == v.oid)
+                        {
+                            exists = IS_DELETE(prev_it->second.state) ? 0 : 1;
+                        }
+                    }
+                    if (exists == -1)
+                    {
+                        auto & clean_db = clean_db_shard(v.oid);
+                        auto clean_it = clean_db.find(v.oid);
+                        exists = clean_it != clean_db.end() ? 1 : 0;
+                    }
+                    if (!exists)
+                    {
+                        uint64_t space_id = dirty_it->first.oid.inode;
+                        if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
+                            space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
+                        inode_space_stats[space_id] += dsk.data_block_size;
+                        used_blocks++;
+                    }
+                    big_to_flush++;
+                }
+                else if (IS_DELETE(dirty_it->second.state))
+                {
+                    uint64_t space_id = dirty_it->first.oid.inode;
+                    if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
+                        space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
+                    auto & sp = inode_space_stats[space_id];
+                    if (sp > dsk.data_block_size)
+                        sp -= dsk.data_block_size;
+                    else
+                        inode_space_stats.erase(space_id);
+                    used_blocks--;
+                    big_to_flush++;
+                }
+            }
+            else if (IS_IN_FLIGHT(dirty_it->second.state))
+            {
+                // mark_stable should never be called for in-flight or submitted writes
+                printf(
+                    "BUG: Attempt to mark_stable object %jx:%jx v%ju state of which is %x\n",
+                    dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
+                    dirty_it->second.state
+                );
+                exit(1);
+            }
+            if (forget_dirty && (IS_BIG_WRITE(dirty_it->second.state) ||
+                IS_DELETE(dirty_it->second.state)))
+            {
+                // Big write overrides all previous dirty entries
+                auto erase_end = dirty_it;
+                while (dirty_it != dirty_db.begin())
+                {
+                    dirty_it--;
+                    if (dirty_it->first.oid != v.oid)
+                    {
+                        dirty_it++;
+                        break;
+                    }
+                }
+                auto & clean_db = clean_db_shard(v.oid);
+                auto clean_it = clean_db.find(v.oid);
+                uint64_t clean_loc = clean_it != clean_db.end()
+                    ? clean_it->second.location : UINT64_MAX;
+                erase_dirty(dirty_it, erase_end, clean_loc);
+                break;
+            }
+            if (was_stable || dirty_it == dirty_db.begin())
+            {
+                break;
+            }
+            dirty_it--;
+            if (dirty_it->first.oid != v.oid)
+            {
+                break;
+            }
+        }
+        flusher->enqueue_flush(v);
+    }
+    auto unstab_it = unstable_writes.find(v.oid);
+    if (unstab_it != unstable_writes.end() &&
+        unstab_it->second <= v.version)
+    {
+        unstable_writes.erase(unstab_it);
+    }
+}
+
+} // namespace v1

src/blockstore/v1/sync.cpp

@@ -0,0 +1,238 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
+
+#define SYNC_HAS_SMALL 1
+#define SYNC_HAS_BIG 2
+#define SYNC_DATA_SYNC_SENT 3
+#define SYNC_DATA_SYNC_DONE 4
+#define SYNC_JOURNAL_WRITE_SENT 5
+#define SYNC_JOURNAL_WRITE_DONE 6
+#define SYNC_JOURNAL_SYNC_SENT 7
+#define SYNC_DONE 8
+
+int blockstore_impl_t::continue_sync(blockstore_op_t *op)
+{
+    if (immediate_commit == IMMEDIATE_ALL)
+    {
+        // We can return immediately because sync is only dequeued after all previous writes
+        op->retval = 0;
+        FINISH_OP(op);
+        return 2;
+    }
+    if (PRIV(op)->op_state == 0)
+    {
+        stop_sync_submitted = false;
+        unsynced_big_write_count -= unsynced_big_writes.size();
+        PRIV(op)->sync_big_writes.swap(unsynced_big_writes);
+        PRIV(op)->sync_small_writes.swap(unsynced_small_writes);
+        unsynced_big_writes.clear();
+        unsynced_small_writes.clear();
+        if (PRIV(op)->sync_big_writes.size() > 0)
+            PRIV(op)->op_state = SYNC_HAS_BIG;
+        else if (PRIV(op)->sync_small_writes.size() > 0)
+            PRIV(op)->op_state = SYNC_HAS_SMALL;
+        else
+            PRIV(op)->op_state = SYNC_DONE;
+    }
+    if (PRIV(op)->op_state == SYNC_HAS_SMALL)
+    {
+        // No big writes, just fsync the journal
+        if (journal.sector_info[journal.cur_sector].dirty)
+        {
+            // Write out the last journal sector if it happens to be dirty
+            BS_SUBMIT_CHECK_SQES(1);
+            prepare_journal_sector_write(journal.cur_sector, op);
+            PRIV(op)->op_state = SYNC_JOURNAL_WRITE_SENT;
+            return 1;
+        }
+        else
+        {
+            PRIV(op)->op_state = SYNC_JOURNAL_WRITE_DONE;
+        }
+    }
+    if (PRIV(op)->op_state == SYNC_HAS_BIG)
+    {
+        // 1st step: fsync data
+        if (!disable_data_fsync)
+        {
+            BS_SUBMIT_GET_SQE(sqe, data);
+            io_uring_prep_fsync(sqe, dsk.data_fd, IORING_FSYNC_DATASYNC);
+            data->iov = { 0 };
+            data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
+            PRIV(op)->pending_ops = 1;
+            PRIV(op)->op_state = SYNC_DATA_SYNC_SENT;
+            return 1;
+        }
+        else
+        {
+            PRIV(op)->op_state = SYNC_DATA_SYNC_DONE;
+        }
+    }
+    if (PRIV(op)->op_state == SYNC_DATA_SYNC_DONE)
+    {
+        // 2nd step: Data device is synced, prepare & write journal entries
+        // Check space in the journal and journal memory buffers
+        blockstore_journal_check_t space_check(this);
+        if (dsk.csum_block_size)
+        {
+            // More complex check because all journal entries have different lengths
+            int left = PRIV(op)->sync_big_writes.size();
+            for (auto & sbw: PRIV(op)->sync_big_writes)
+            {
+                left--;
+                auto & dirty_entry = dirty_db.at(sbw);
+                uint64_t dyn_size = dsk.dirty_dyn_size(dirty_entry.offset, dirty_entry.len);
+                if (!space_check.check_available(op, 1, sizeof(journal_entry_big_write) + dyn_size, 0))
+                {
+                    return 0;
+                }
+            }
+        }
+        else if (!space_check.check_available(op, PRIV(op)->sync_big_writes.size(),
+            sizeof(journal_entry_big_write) + dsk.clean_entry_bitmap_size, 0))
+        {
+            return 0;
+        }
+        // Check SQEs. Don't bother about merging, submit each journal sector as a separate request
+        BS_SUBMIT_CHECK_SQES(space_check.sectors_to_write);
+        // Prepare and submit journal entries
+        auto it = PRIV(op)->sync_big_writes.begin();
+        int s = 0;
+        while (it != PRIV(op)->sync_big_writes.end())
+        {
+            auto & dirty_entry = dirty_db.at(*it);
+            uint64_t dyn_size = dsk.dirty_dyn_size(dirty_entry.offset, dirty_entry.len);
+            if (!journal.entry_fits(sizeof(journal_entry_big_write) + dyn_size) &&
+                journal.sector_info[journal.cur_sector].dirty)
+            {
+                prepare_journal_sector_write(journal.cur_sector, op);
+                s++;
+            }
+            journal_entry_big_write *je = (journal_entry_big_write*)prefill_single_journal_entry(
+                journal, (dirty_entry.state & BS_ST_INSTANT) ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
+                sizeof(journal_entry_big_write) + dyn_size
+            );
+            auto jsec = dirty_entry.journal_sector = journal.sector_info[journal.cur_sector].offset;
+            assert(journal.next_free >= journal.used_start
+                ? (jsec >= journal.used_start && jsec < journal.next_free)
+                : (jsec >= journal.used_start || jsec < journal.next_free));
+            journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
+#ifdef BLOCKSTORE_DEBUG
+            printf(
+                "journal offset %08jx is used by %jx:%jx v%ju (%ju refs)\n",
+                dirty_entry.journal_sector, it->oid.inode, it->oid.stripe, it->version,
+                journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
+            );
+#endif
+            je->oid = it->oid;
+            je->version = it->version;
+            je->offset = dirty_entry.offset;
+            je->len = dirty_entry.len;
+            je->location = dirty_entry.location;
+            memcpy((void*)(je+1), (alloc_dyn_data
+                ? (uint8_t*)dirty_entry.dyn_data+sizeof(int) : (uint8_t*)&dirty_entry.dyn_data), dyn_size);
+            je->crc32 = je_crc32((journal_entry*)je);
+            journal.crc32_last = je->crc32;
+            it++;
+        }
+        prepare_journal_sector_write(journal.cur_sector, op);
+        s++;
+        assert(s == space_check.sectors_to_write);
+        PRIV(op)->op_state = SYNC_JOURNAL_WRITE_SENT;
+        return 1;
+    }
+    if (PRIV(op)->op_state == SYNC_JOURNAL_WRITE_DONE)
+    {
+        if (!disable_journal_fsync)
+        {
+            BS_SUBMIT_GET_SQE(sqe, data);
+            io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
+            data->iov = { 0 };
+            data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
+            PRIV(op)->pending_ops = 1;
+            PRIV(op)->op_state = SYNC_JOURNAL_SYNC_SENT;
+            return 1;
+        }
+        else
+        {
+            PRIV(op)->op_state = SYNC_DONE;
+        }
+    }
+    if (PRIV(op)->op_state == SYNC_DONE)
+    {
+        ack_sync(op);
+        return 2;
+    }
+    return 1;
+}
+
+void blockstore_impl_t::ack_sync(blockstore_op_t *op)
+{
+    // Handle states
+    for (auto it = PRIV(op)->sync_big_writes.begin(); it != PRIV(op)->sync_big_writes.end(); it++)
+    {
+#ifdef BLOCKSTORE_DEBUG
+        printf("Ack sync big %jx:%jx v%ju\n", it->oid.inode, it->oid.stripe, it->version);
+#endif
+        auto & unstab = unstable_writes[it->oid];
+        unstab = unstab < it->version ? it->version : unstab;
+        auto dirty_it = dirty_db.find(*it);
+        dirty_it->second.state = ((dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SYNCED);
+        if (dirty_it->second.state & BS_ST_INSTANT)
+        {
+            mark_stable(dirty_it->first);
+        }
+        else
+        {
+            unstable_unsynced--;
+            assert(unstable_unsynced >= 0);
+        }
+        dirty_it++;
+        while (dirty_it != dirty_db.end() && dirty_it->first.oid == it->oid)
+        {
+            if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG)
+            {
+                dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_IN_FLIGHT;
+            }
+            dirty_it++;
+        }
+    }
+    for (auto it = PRIV(op)->sync_small_writes.begin(); it != PRIV(op)->sync_small_writes.end(); it++)
+    {
+#ifdef BLOCKSTORE_DEBUG
+        printf("Ack sync small %jx:%jx v%ju\n", it->oid.inode, it->oid.stripe, it->version);
+#endif
+        auto & unstab = unstable_writes[it->oid];
+        unstab = unstab < it->version ? it->version : unstab;
+        if (dirty_db[*it].state == (BS_ST_DELETE | BS_ST_WRITTEN))
+        {
+            dirty_db[*it].state = (BS_ST_DELETE | BS_ST_SYNCED);
+            // Deletions are treated as immediately stable
+            mark_stable(*it);
+        }
+        else /* (BS_ST_INSTANT?) | BS_ST_SMALL_WRITE | BS_ST_WRITTEN */
+        {
+            dirty_db[*it].state = (dirty_db[*it].state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SYNCED;
+            if (dirty_db[*it].state & BS_ST_INSTANT)
+            {
+                mark_stable(*it);
+            }
+            else
+            {
+                unstable_unsynced--;
+                assert(unstable_unsynced >= 0);
+            }
+        }
+    }
+    op->retval = 0;
+    FINISH_OP(op);
+}
+
+} // namespace v1

src/blockstore/v1/write.cpp

@@ -0,0 +1,828 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "impl.h"
+#include "internal.h"
+
+namespace v1 {
+
+bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
+{
+    // Check or assign version number
+    bool found = false, deleted = false, unsynced = false, is_del = (op->opcode == BS_OP_DELETE);
+    bool wait_big = false, wait_del = false;
+    void *dyn = NULL;
+    if (is_del)
+    {
+        op->len = 0;
+    }
+    size_t dyn_size = dsk.dirty_dyn_size(op->offset, op->len);
+    if (!is_del && alloc_dyn_data)
+    {
+        // FIXME: Working with `dyn_data` has to be refactored somehow but I first have to decide how :)
+        // +sizeof(int) = refcount
+        dyn = calloc_or_die(1, dyn_size+sizeof(int));
+        *((int*)dyn) = 1;
+    }
+    uint8_t *dyn_ptr = (alloc_dyn_data ? (uint8_t*)dyn+sizeof(int) : (uint8_t*)&dyn);
+    uint64_t version = 1;
+    if (dirty_db.size() > 0)
+    {
+        auto dirty_it = dirty_db.upper_bound((obj_ver_id){
+            .oid = op->oid,
+            .version = UINT64_MAX,
+        });
+        dirty_it--; // segfaults when dirty_db is empty
+        if (dirty_it != dirty_db.end() && dirty_it->first.oid == op->oid)
+        {
+            found = true;
+            version = dirty_it->first.version + 1;
+            deleted = IS_DELETE(dirty_it->second.state);
+            unsynced = !IS_SYNCED(dirty_it->second.state);
+            wait_del = ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_DEL);
+            wait_big = (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE
+                ? !IS_SYNCED(dirty_it->second.state)
+                : ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG);
+            if (!is_del && !deleted)
+            {
+                void *dyn_from = alloc_dyn_data
+                    ? (uint8_t*)dirty_it->second.dyn_data + sizeof(int) : (uint8_t*)&dirty_it->second.dyn_data;
+                memcpy(dyn_ptr, dyn_from, dsk.clean_entry_bitmap_size);
+            }
+        }
+    }
+    if (!found)
+    {
+        auto & clean_db = clean_db_shard(op->oid);
+        auto clean_it = clean_db.find(op->oid);
+        if (clean_it != clean_db.end())
+        {
+            version = clean_it->second.version + 1;
+            if (!is_del)
+            {
+                void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, dsk.clean_entry_bitmap_size);
+                memcpy(dyn_ptr, bmp_ptr, dsk.clean_entry_bitmap_size);
+            }
+        }
+        else
+        {
+            deleted = true;
+        }
+    }
+    if (deleted && is_del)
+    {
+        // Already deleted
+        op->retval = 0;
+        return false;
+    }
+    PRIV(op)->real_version = 0;
+    if (op->version == 0)
+    {
+        op->version = version;
+    }
+    else if (op->version < version)
+    {
+        // Implicit operations must be added like that: DEL [FLUSH] BIG [SYNC] SMALL SMALL
+        if (deleted || wait_del)
+        {
+            // It's allowed to write versions with low numbers over deletes
+            // However, we have to flush those deletes first as we use version number for ordering
+#ifdef BLOCKSTORE_DEBUG
+            printf("Write %jx:%jx v%ju over delete (real v%ju) offset=%u len=%u\n", op->oid.inode, op->oid.stripe, version, op->version, op->offset, op->len);
+#endif
+            wait_del = true;
+            PRIV(op)->real_version = op->version;
+            op->version = version;
+            if (unsynced)
+            {
+                // Issue an additional sync so the delete reaches the journal
+                blockstore_op_t *sync_op = new blockstore_op_t;
+                sync_op->opcode = BS_OP_SYNC;
+                sync_op->oid = op->oid;
+                sync_op->version = op->version;
+                sync_op->callback = [this](blockstore_op_t *sync_op)
+                {
+                    flusher->unshift_flush((obj_ver_id){
+                        .oid = sync_op->oid,
+                        .version = sync_op->version-1,
+                    }, true);
+                    delete sync_op;
+                };
+                enqueue_op(sync_op);
+            }
+            else
+            {
+                flusher->unshift_flush((obj_ver_id){
+                    .oid = op->oid,
+                    .version = version-1,
+                }, true);
+            }
+        }
+        else
+        {
+            // Invalid version requested
+#ifdef BLOCKSTORE_DEBUG
+            printf("Write %jx:%jx v%ju requested, but we already have v%ju\n", op->oid.inode, op->oid.stripe, op->version, version);
+#endif
+            op->retval = -EEXIST;
+            if (!is_del && alloc_dyn_data)
+            {
+                free(dyn);
+            }
+            return false;
+        }
+    }
+    bool imm = (op->len < dsk.data_block_size ? (immediate_commit != IMMEDIATE_NONE) : (immediate_commit == IMMEDIATE_ALL));
+    if (wait_big && !is_del && !deleted && op->len < dsk.data_block_size && !imm ||
+        !imm && autosync_writes && unsynced_queued_ops >= autosync_writes)
+    {
+        // Issue an additional sync so that the previous big write can reach the journal
+        blockstore_op_t *sync_op = new blockstore_op_t;
+        sync_op->opcode = BS_OP_SYNC;
+        sync_op->callback = [](blockstore_op_t *sync_op)
+        {
+            delete sync_op;
+        };
+        enqueue_op(sync_op);
+    }
+    else if (!imm)
+        unsynced_queued_ops++;
+#ifdef BLOCKSTORE_DEBUG
+    if (is_del)
+        printf("Delete %jx:%jx v%ju\n", op->oid.inode, op->oid.stripe, op->version);
+    else if (!wait_del)
+        printf("Write %jx:%jx v%ju offset=%u len=%u\n", op->oid.inode, op->oid.stripe, op->version, op->offset, op->len);
+#endif
+    // No strict need to add it into dirty_db here except maybe for listings to return
+    // correct data when there are inflight operations in the queue
+    uint32_t state;
+    if (is_del)
+        state = BS_ST_DELETE | BS_ST_IN_FLIGHT;
+    else
+    {
+        state = (op->len == dsk.data_block_size || deleted ? BS_ST_BIG_WRITE : BS_ST_SMALL_WRITE);
+        if (state == BS_ST_SMALL_WRITE && throttle_small_writes)
+            clock_gettime(CLOCK_REALTIME, &PRIV(op)->tv_begin);
+        if (wait_del)
+            state |= BS_ST_WAIT_DEL;
+        else if (state == BS_ST_SMALL_WRITE && wait_big)
+            state |= BS_ST_WAIT_BIG;
+        else
+            state |= BS_ST_IN_FLIGHT;
+        if (op->opcode == BS_OP_WRITE_STABLE)
+            state |= BS_ST_INSTANT;
+        if (op->bitmap)
+            memcpy(dyn_ptr, op->bitmap, dsk.clean_entry_bitmap_size);
+    }
+    // Calculate checksums
+    // FIXME: Allow to receive checksums from outside?
+    if (!is_del && dsk.data_csum_type && op->len > 0)
+    {
+        uint32_t *data_csums = (uint32_t*)(dyn_ptr + dsk.clean_entry_bitmap_size);
+        uint32_t start = op->offset / dsk.csum_block_size;
+        uint32_t end = (op->offset+op->len-1) / dsk.csum_block_size;
+        auto fn = state & BS_ST_BIG_WRITE ? crc32c_pad : crc32c_nopad;
+        if (start == end)
+            data_csums[0] = fn(0, op->buf, op->len, op->offset - start*dsk.csum_block_size, end*dsk.csum_block_size - (op->offset+op->len));
+        else
+        {
+            // First block
+            data_csums[0] = fn(0, op->buf, dsk.csum_block_size*(start+1)-op->offset, op->offset - start*dsk.csum_block_size, 0);
+            // Intermediate blocks
+            for (uint32_t i = start+1; i < end; i++)
+                data_csums[i-start] = crc32c(0, (uint8_t*)op->buf + dsk.csum_block_size*i-op->offset, dsk.csum_block_size);
+            // Last block
+            data_csums[end-start] = fn(
+                0, (uint8_t*)op->buf + end*dsk.csum_block_size - op->offset,
+                op->offset+op->len - end*dsk.csum_block_size,
+                0, (end+1)*dsk.csum_block_size - (op->offset+op->len)
+            );
+        }
+    }
+    dirty_db.emplace((obj_ver_id){
+        .oid = op->oid,
+        .version = op->version,
+    }, (dirty_entry){
+        .state = state,
+        .flags = 0,
+        .location = 0,
+        .offset = is_del ? 0 : op->offset,
+        .len = is_del ? 0 : op->len,
+        .journal_sector = 0,
+        .dyn_data = dyn,
+    });
+    return true;
+}
+
+void blockstore_impl_t::cancel_all_writes(blockstore_op_t *op, blockstore_dirty_db_t::iterator dirty_it, int retval)
+{
+    while (dirty_it != dirty_db.end() && dirty_it->first.oid == op->oid)
+    {
+        free_dirty_dyn_data(dirty_it->second);
+        dirty_db.erase(dirty_it++);
+    }
+    bool found = false;
+    for (auto other_op: submit_queue)
+    {
+        if (!other_op)
+        {
+            // freed operations during submitting are zeroed
+        }
+        else if (other_op == op)
+        {
+            // <op> may be present in queue multiple times due to moving operations in submit_queue
+            found = true;
+        }
+        else if (found && other_op->oid == op->oid &&
+            (other_op->opcode == BS_OP_WRITE || other_op->opcode == BS_OP_WRITE_STABLE))
+        {
+            // Mark operations to cancel them
+            PRIV(other_op)->real_version = UINT64_MAX;
+            other_op->retval = retval;
+        }
+    }
+    op->retval = retval;
+    FINISH_OP(op);
+}
+
+// First step of the write algorithm: dequeue operation and submit initial write(s)
+int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
+{
+    if (PRIV(op)->op_state)
+    {
+        return continue_write(op);
+    }
+    auto dirty_it = dirty_db.find((obj_ver_id){
+        .oid = op->oid,
+        .version = op->version,
+    });
+    assert(dirty_it != dirty_db.end());
+    if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) < BS_ST_IN_FLIGHT)
+    {
+        // Don't dequeue
+        return 0;
+    }
+    if (PRIV(op)->real_version != 0)
+    {
+        if (PRIV(op)->real_version == UINT64_MAX)
+        {
+            // This is the flag value used to cancel operations
+            FINISH_OP(op);
+            return 2;
+        }
+        // Restore original low version number for unblocked operations
+#ifdef BLOCKSTORE_DEBUG
+        printf("Restoring %jx:%jx version: v%ju -> v%ju\n", op->oid.inode, op->oid.stripe, op->version, PRIV(op)->real_version);
+#endif
+        auto prev_it = dirty_it;
+        if (prev_it != dirty_db.begin())
+        {
+            prev_it--;
+            if (prev_it->first.oid == op->oid && prev_it->first.version >= PRIV(op)->real_version)
+            {
+                // Original version is still invalid
+                // All subsequent writes to the same object must be canceled too
+                printf("Tried to write %jx:%jx v%ju after delete (old version v%ju), but already have v%ju\n",
+                    op->oid.inode, op->oid.stripe, PRIV(op)->real_version, op->version, prev_it->first.version);
+                cancel_all_writes(op, dirty_it, -EEXIST);
+                return 2;
+            }
+        }
+        op->version = PRIV(op)->real_version;
+        PRIV(op)->real_version = 0;
+        dirty_entry e = dirty_it->second;
+        dirty_db.erase(dirty_it);
+        dirty_it = dirty_db.emplace((obj_ver_id){
+            .oid = op->oid,
+            .version = op->version,
+        }, e).first;
+    }
+    if (write_iodepth >= max_write_iodepth)
+    {
+        return 0;
+    }
+    if ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE)
+    {
+        blockstore_journal_check_t space_check(this);
+        if (!space_check.check_available(op, unsynced_big_write_count + 1,
+            sizeof(journal_entry_big_write) + dsk.clean_dyn_size,
+            (unstable_writes.size()+unstable_unsynced+((dirty_it->second.state & BS_ST_INSTANT) ? 0 : 1))*journal.block_size))
+        {
+            return 0;
+        }
+        // Big (redirect) write
+        uint64_t loc = data_alloc->find_free();
+        if (loc == UINT64_MAX)
+        {
+            // no space
+            if (big_to_flush > 0)
+            {
+                // hope that some space will be available after flush
+                flusher->request_trim();
+                PRIV(op)->wait_for = WAIT_FREE;
+                return 0;
+            }
+            cancel_all_writes(op, dirty_it, -ENOSPC);
+            return 2;
+        }
+        if (inmemory_meta)
+        {
+            // Check once more that metadata entry is zeroed (the reverse means a bug or corruption)
+            uint64_t sector = (loc / (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.meta_block_size;
+            uint64_t pos = (loc % (dsk.meta_block_size / dsk.clean_entry_size));
+            clean_disk_entry *entry = (clean_disk_entry*)((uint8_t*)metadata_buffer + sector + pos*dsk.clean_entry_size);
+            if (entry->oid.inode || entry->oid.stripe || entry->version)
+            {
+                printf(
+                    "Fatal error (metadata corruption or bug): tried to write object %jx:%jx v%ju"
+                    " over a non-zero metadata entry %ju with %jx:%jx v%ju\n", op->oid.inode,
+                    op->oid.stripe, op->version, loc, entry->oid.inode, entry->oid.stripe, entry->version
+                );
+                exit(1);
+            }
+        }
+        BS_SUBMIT_GET_SQE(sqe, data);
+        write_iodepth++;
+        dirty_it->second.location = loc * dsk.data_block_size;
+        dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SUBMITTED;
+#ifdef BLOCKSTORE_DEBUG
+        printf(
+            "Allocate block %ju for %jx:%jx v%ju\n",
+            loc, op->oid.inode, op->oid.stripe, op->version
+        );
+#endif
+        data_alloc->set(loc, true);
+        uint64_t stripe_offset = (op->offset % dsk.bitmap_granularity);
+        uint64_t stripe_end = (op->offset + op->len) % dsk.bitmap_granularity;
+        // Zero fill up to dsk.bitmap_granularity
+        int vcnt = 0;
+        if (stripe_offset)
+        {
+            PRIV(op)->iov_zerofill[vcnt++] = (struct iovec){ zero_object, (size_t)stripe_offset };
+        }
+        PRIV(op)->iov_zerofill[vcnt++] = (struct iovec){ op->buf, op->len };
+        if (stripe_end)
+        {
+            stripe_end = dsk.bitmap_granularity - stripe_end;
+            PRIV(op)->iov_zerofill[vcnt++] = (struct iovec){ zero_object, (size_t)stripe_end };
+        }
+        data->iov.iov_len = op->len + stripe_offset + stripe_end; // to check it in the callback
+        data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+        io_uring_prep_writev(
+            sqe, dsk.data_fd, PRIV(op)->iov_zerofill, vcnt, dsk.data_offset + (loc * dsk.data_block_size) + op->offset - stripe_offset
+        );
+        PRIV(op)->pending_ops = 1;
+        if (!(dirty_it->second.state & BS_ST_INSTANT))
+        {
+            unstable_unsynced++;
+        }
+        if (immediate_commit != IMMEDIATE_ALL)
+        {
+            // Increase the counter, but don't save into unsynced_writes yet (can't sync until the write is finished)
+            unsynced_big_write_count++;
+            PRIV(op)->op_state = 3;
+        }
+        else
+        {
+            PRIV(op)->op_state = 1;
+        }
+    }
+    else /* if ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_SMALL_WRITE) */
+    {
+        // Small (journaled) write
+        // First check if the journal has sufficient space
+        uint64_t dyn_size = dsk.dirty_dyn_size(op->offset, op->len);
+        blockstore_journal_check_t space_check(this);
+        if (unsynced_big_write_count &&
+            !space_check.check_available(op, unsynced_big_write_count,
+                sizeof(journal_entry_big_write) + dsk.clean_dyn_size, 0)
+            || !space_check.check_available(op, 1,
+                sizeof(journal_entry_small_write) + dyn_size,
+                op->len + (unstable_writes.size()+unstable_unsynced+((dirty_it->second.state & BS_ST_INSTANT) ? 0 : 1))*journal.block_size))
+        {
+            return 0;
+        }
+        // There is sufficient space. Check SQE(s)
+        BS_SUBMIT_CHECK_SQES(
+            // Write current journal sector only if it's dirty and full, or in the immediate_commit mode
+            (immediate_commit != IMMEDIATE_NONE ||
+                !journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size) ? 1 : 0) +
+            (op->len > 0 ? 1 : 0)
+        );
+        write_iodepth++;
+        // Got SQEs. Prepare previous journal sector write if required
+        if (immediate_commit == IMMEDIATE_NONE &&
+            !journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size))
+        {
+            prepare_journal_sector_write(journal.cur_sector, op);
+        }
+        // Then pre-fill journal entry
+        journal_entry_small_write *je = (journal_entry_small_write*)prefill_single_journal_entry(
+            journal, op->opcode == BS_OP_WRITE_STABLE ? JE_SMALL_WRITE_INSTANT : JE_SMALL_WRITE,
+            sizeof(journal_entry_small_write) + dyn_size
+        );
+        auto jsec = dirty_it->second.journal_sector = journal.sector_info[journal.cur_sector].offset;
+        if (!(journal.next_free >= journal.used_start
+            ? (jsec >= journal.used_start && jsec < journal.next_free)
+            : (jsec >= journal.used_start || jsec < journal.next_free)))
+        {
+            printf(
+                "BUG: journal offset %08jx is used by %jx:%jx v%ju (%ju refs) BUT used_start=%jx next_free=%jx\n",
+                dirty_it->second.journal_sector, dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
+                journal.used_sectors[journal.sector_info[journal.cur_sector].offset],
+                journal.used_start, journal.next_free
+            );
+            abort();
+        }
+        journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
+#ifdef BLOCKSTORE_DEBUG
+        printf(
+            "journal offset %08jx is used by %jx:%jx v%ju (%ju refs)\n",
+            dirty_it->second.journal_sector, dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
+            journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
+        );
+#endif
+        // Figure out where data will be
+        auto next_next_free = (journal.next_free + op->len) <= journal.len ? journal.next_free : dsk.journal_block_size;
+        if (op->len > 0)
+        {
+            auto journal_used_it = journal.used_sectors.lower_bound(next_next_free);
+            if (journal_used_it != journal.used_sectors.end() &&
+                journal_used_it->first < next_next_free + op->len)
+            {
+                printf(
+                    "BUG: Attempt to overwrite used offset (%jx, %ju refs) of the journal with the object %jx:%jx v%ju: data at %jx, len %x!"
+                    " Journal used_start=%08jx (%ju refs), next_free=%08jx, dirty_start=%08jx\n",
+                    journal_used_it->first, journal_used_it->second, op->oid.inode, op->oid.stripe, op->version, next_next_free, op->len,
+                    journal.used_start, journal.used_sectors[journal.used_start], journal.next_free, journal.dirty_start
+                );
+                exit(1);
+            }
+        }
+        // double check that next_free doesn't cross used_start from the left
+        assert(journal.next_free >= journal.used_start && next_next_free >= journal.next_free || next_next_free < journal.used_start);
+        journal.next_free = next_next_free;
+        je->oid = op->oid;
+        je->version = op->version;
+        je->offset = op->offset;
+        je->len = op->len;
+        je->data_offset = journal.next_free;
+        je->crc32_data = dsk.csum_block_size ? 0 : crc32c(0, op->buf, op->len);
+        memcpy((void*)(je+1), (alloc_dyn_data
+            ? (uint8_t*)dirty_it->second.dyn_data+sizeof(int) : (uint8_t*)&dirty_it->second.dyn_data), dyn_size);
+        je->crc32 = je_crc32((journal_entry*)je);
+        journal.crc32_last = je->crc32;
+        if (immediate_commit != IMMEDIATE_NONE)
+        {
+            prepare_journal_sector_write(journal.cur_sector, op);
+        }
+        if (op->len > 0)
+        {
+            // Prepare journal data write
+            if (journal.inmemory)
+            {
+                // Copy data
+                memcpy((uint8_t*)journal.buffer + journal.next_free, op->buf, op->len);
+            }
+            BS_SUBMIT_GET_SQE(sqe2, data2);
+            data2->iov = (struct iovec){ op->buf, op->len };
+            ++journal.submit_id;
+            assert(journal.submit_id != 0); // check overflow
+            // Make subsequent journal writes wait for our data write
+            journal.flushing_ops.emplace(journal.submit_id, (pending_journaling_t){
+                .pending = 1,
+                .sector = -1,
+                .op = op,
+            });
+            data2->callback = [this, flush_id = journal.submit_id](ring_data_t *data) { handle_journal_write(data, flush_id); };
+            io_uring_prep_writev(
+                sqe2, dsk.journal_fd, &data2->iov, 1, journal.offset + journal.next_free
+            );
+            PRIV(op)->pending_ops++;
+        }
+        else
+        {
+            // Zero-length overwrite. Allowed to bump object version in EC placement groups without actually writing data
+        }
+        dirty_it->second.location = journal.next_free;
+        dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SUBMITTED;
+        next_next_free = journal.next_free + op->len;
+        if (next_next_free >= journal.len)
+            next_next_free = dsk.journal_block_size;
+        // double check that next_free doesn't cross used_start from the left
+        assert(journal.next_free >= journal.used_start && next_next_free >= journal.next_free || next_next_free < journal.used_start);
+        journal.next_free = next_next_free;
+        if (!(dirty_it->second.state & BS_ST_INSTANT))
+        {
+            unstable_unsynced++;
+        }
+        if (!PRIV(op)->pending_ops)
+        {
+            PRIV(op)->op_state = 4;
+            return continue_write(op);
+        }
+        else
+        {
+            PRIV(op)->op_state = 3;
+        }
+    }
+    return 1;
+}
+
+int blockstore_impl_t::continue_write(blockstore_op_t *op)
+{
+    int op_state = PRIV(op)->op_state;
+    if (op_state == 2)
+        goto resume_2;
+    else if (op_state == 4)
+        goto resume_4;
+    else if (op_state == 6)
+        goto resume_6;
+    else
+    {
+        // In progress
+        return 1;
+    }
+resume_2:
+    // Only for the immediate_commit mode: prepare and submit big_write journal entry
+    {
+        auto dirty_it = dirty_db.find((obj_ver_id){
+            .oid = op->oid,
+            .version = op->version,
+        });
+        assert(dirty_it != dirty_db.end());
+        uint64_t dyn_size = dsk.dirty_dyn_size(op->offset, op->len);
+        blockstore_journal_check_t space_check(this);
+        if (!space_check.check_available(op, 1, sizeof(journal_entry_big_write) + dyn_size,
+            (unstable_writes.size()+unstable_unsynced+((dirty_it->second.state & BS_ST_INSTANT) ? 0 : 1))*journal.block_size))
+        {
+            return 0;
+        }
+        BS_SUBMIT_CHECK_SQES(1);
+        journal_entry_big_write *je = (journal_entry_big_write*)prefill_single_journal_entry(
+            journal, op->opcode == BS_OP_WRITE_STABLE ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
+            sizeof(journal_entry_big_write) + dyn_size
+        );
+        auto jsec = dirty_it->second.journal_sector = journal.sector_info[journal.cur_sector].offset;
+        if (!(journal.next_free >= journal.used_start
+            ? (jsec >= journal.used_start && jsec < journal.next_free)
+            : (jsec >= journal.used_start || jsec < journal.next_free)))
+        {
+            printf(
+                "BUG: journal offset %08jx is used by %jx:%jx v%ju (%ju refs) BUT used_start=%jx next_free=%jx\n",
+                dirty_it->second.journal_sector, dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
+                journal.used_sectors[journal.sector_info[journal.cur_sector].offset],
+                journal.used_start, journal.next_free
+            );
+            abort();
+        }
+        journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
+#ifdef BLOCKSTORE_DEBUG
+        printf(
+            "journal offset %08jx is used by %jx:%jx v%ju (%ju refs)\n",
+            journal.sector_info[journal.cur_sector].offset, op->oid.inode, op->oid.stripe, op->version,
+            journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
+        );
+#endif
+        je->oid = op->oid;
+        je->version = op->version;
+        je->offset = op->offset;
+        je->len = op->len;
+        je->location = dirty_it->second.location;
+        memcpy((void*)(je+1), (alloc_dyn_data
+            ? (uint8_t*)dirty_it->second.dyn_data+sizeof(int) : (uint8_t*)&dirty_it->second.dyn_data), dyn_size);
+        je->crc32 = je_crc32((journal_entry*)je);
+        journal.crc32_last = je->crc32;
+        prepare_journal_sector_write(journal.cur_sector, op);
+        PRIV(op)->op_state = 3;
+        return 1;
+    }
+resume_4:
+    // Switch object state
+    {
+        auto dirty_it = dirty_db.find((obj_ver_id){
+            .oid = op->oid,
+            .version = op->version,
+        });
+        assert(dirty_it != dirty_db.end());
+#ifdef BLOCKSTORE_DEBUG
+        printf("Ack write %jx:%jx v%ju = state 0x%x\n", op->oid.inode, op->oid.stripe, op->version, dirty_it->second.state);
+#endif
+        bool is_big = (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE;
+        bool imm = is_big ? (immediate_commit == IMMEDIATE_ALL) : (immediate_commit != IMMEDIATE_NONE);
+        bool is_instant = IS_INSTANT(dirty_it->second.state);
+        if (imm)
+        {
+            auto & unstab = unstable_writes[op->oid];
+            unstab = unstab < op->version ? op->version : unstab;
+            if (!is_instant)
+            {
+                unstable_unsynced--;
+                assert(unstable_unsynced >= 0);
+            }
+        }
+        dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK)
+            | (imm ? BS_ST_SYNCED : BS_ST_WRITTEN);
+        if (imm && is_instant)
+        {
+            // Deletions and 'instant' operations are treated as immediately stable
+            mark_stable(dirty_it->first);
+        }
+        if (!imm)
+        {
+            if (is_big)
+            {
+                // Remember big write as unsynced
+                unsynced_big_writes.push_back((obj_ver_id){
+                    .oid = op->oid,
+                    .version = op->version,
+                });
+            }
+            else
+            {
+                // Remember small write as unsynced
+                unsynced_small_writes.push_back((obj_ver_id){
+                    .oid = op->oid,
+                    .version = op->version,
+                });
+            }
+        }
+        if (imm && (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE)
+        {
+            // Unblock small writes
+            dirty_it++;
+            while (dirty_it != dirty_db.end() && dirty_it->first.oid == op->oid)
+            {
+                if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG)
+                {
+                    dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_IN_FLIGHT;
+                }
+                dirty_it++;
+            }
+        }
+        // Apply throttling to not fill the journal too fast for the SSD+HDD case
+        if (!is_big && throttle_small_writes)
+        {
+            // Apply throttling
+            timespec tv_end;
+            clock_gettime(CLOCK_REALTIME, &tv_end);
+            uint64_t exec_us =
+                (tv_end.tv_sec - PRIV(op)->tv_begin.tv_sec)*1000000 +
+                (tv_end.tv_nsec - PRIV(op)->tv_begin.tv_nsec)/1000;
+            // Compare with target execution time
+            // 100% free -> target time = 0
+            // 0% free -> target time = iodepth/parallelism * (iops + size/bw) / write per second
+            uint64_t used_start = journal.get_trim_pos();
+            uint64_t journal_free_space = journal.next_free < used_start
+                ? (used_start - journal.next_free)
+                : (journal.len - journal.next_free + used_start - journal.block_size);
+            uint64_t ref_us =
+                (write_iodepth <= throttle_target_parallelism ? 100 : 100*write_iodepth/throttle_target_parallelism)
+                * (1000000/throttle_target_iops + op->len*1000000/throttle_target_mbs/1024/1024)
+                / 100;
+            ref_us -= ref_us * journal_free_space / journal.len;
+            if (ref_us > exec_us + throttle_threshold_us)
+            {
+                // Pause reply
+                PRIV(op)->op_state = 5;
+                // Remember that the timer can in theory be called right here
+                tfd->set_timer_us(ref_us-exec_us, false, [this, op](int timer_id)
+                {
+                    PRIV(op)->op_state++;
+                    ringloop->wakeup();
+                });
+                return 1;
+            }
+        }
+    }
+resume_6:
+    // Acknowledge write
+    op->retval = op->len;
+    write_iodepth--;
+    FINISH_OP(op);
+    return 2;
+}
+
+void blockstore_impl_t::handle_write_event(ring_data_t *data, blockstore_op_t *op)
+{
+    live = true;
+    if (data->res != data->iov.iov_len)
+    {
+        // FIXME: our state becomes corrupted after a write error. maybe do something better than just die
+        disk_error_abort("data write", data->res, data->iov.iov_len);
+    }
+    PRIV(op)->pending_ops--;
+    assert(PRIV(op)->pending_ops >= 0);
+    if (PRIV(op)->pending_ops == 0)
+    {
+        release_journal_sectors(op);
+        PRIV(op)->op_state++;
+        ringloop->wakeup();
+    }
+}
+
+void blockstore_impl_t::release_journal_sectors(blockstore_op_t *op)
+{
+    // Release flushed journal sectors
+    if (PRIV(op)->min_flushed_journal_sector > 0 &&
+        PRIV(op)->max_flushed_journal_sector > 0)
+    {
+        uint64_t s = PRIV(op)->min_flushed_journal_sector;
+        while (1)
+        {
+            if (!journal.sector_info[s-1].dirty && journal.sector_info[s-1].flush_count == 0)
+            {
+                if (s == (1+journal.cur_sector))
+                {
+                    // Forcibly move to the next sector and move dirty position
+                    journal.in_sector_pos = journal.block_size;
+                }
+                // We know for sure that we won't write into this sector anymore
+                uint64_t new_ds = journal.sector_info[s-1].offset + journal.block_size;
+                if (new_ds >= journal.len)
+                {
+                    new_ds = journal.block_size;
+                }
+                if ((journal.dirty_start + (journal.dirty_start >= journal.used_start ? 0 : journal.len)) <
+                    (new_ds + (new_ds >= journal.used_start ? 0 : journal.len)))
+                {
+                    journal.dirty_start = new_ds;
+                }
+            }
+            if (s == PRIV(op)->max_flushed_journal_sector)
+                break;
+            s = 1 + s % journal.sector_count;
+        }
+        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
+    }
+}
+
+int blockstore_impl_t::dequeue_del(blockstore_op_t *op)
+{
+    if (PRIV(op)->op_state)
+    {
+        return continue_write(op);
+    }
+    auto dirty_it = dirty_db.find((obj_ver_id){
+        .oid = op->oid,
+        .version = op->version,
+    });
+    assert(dirty_it != dirty_db.end());
+    blockstore_journal_check_t space_check(this);
+    if (!space_check.check_available(op, 1, sizeof(journal_entry_del), (unstable_writes.size()+unstable_unsynced)*journal.block_size))
+    {
+        return 0;
+    }
+    // Write current journal sector only if it's dirty and full, or in the immediate_commit mode
+    BS_SUBMIT_CHECK_SQES(
+        (immediate_commit != IMMEDIATE_NONE ||
+            (dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
+            journal.sector_info[journal.cur_sector].dirty) ? 1 : 0
+    );
+    if (write_iodepth >= max_write_iodepth)
+    {
+        return 0;
+    }
+    write_iodepth++;
+    // Prepare journal sector write
+    if (immediate_commit == IMMEDIATE_NONE &&
+        (dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
+        journal.sector_info[journal.cur_sector].dirty)
+    {
+        prepare_journal_sector_write(journal.cur_sector, op);
+    }
+    // Pre-fill journal entry
+    journal_entry_del *je = (journal_entry_del*)prefill_single_journal_entry(
+        journal, JE_DELETE, sizeof(struct journal_entry_del)
+    );
+    dirty_it->second.journal_sector = journal.sector_info[journal.cur_sector].offset;
+    journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
+#ifdef BLOCKSTORE_DEBUG
+    printf(
+        "journal offset %08jx is used by %jx:%jx v%ju (%ju refs)\n",
+        dirty_it->second.journal_sector, dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
+        journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
+    );
+#endif
+    je->oid = op->oid;
+    je->version = op->version;
+    je->crc32 = je_crc32((journal_entry*)je);
+    journal.crc32_last = je->crc32;
+    dirty_it->second.state = BS_ST_DELETE | BS_ST_SUBMITTED;
+    if (immediate_commit != IMMEDIATE_NONE)
+    {
+        prepare_journal_sector_write(journal.cur_sector, op);
+    }
+    if (!PRIV(op)->pending_ops)
+    {
+        PRIV(op)->op_state = 4;
+        return continue_write(op);
+    }
+    else
+    {
+        PRIV(op)->op_state = 3;
+    }
+    return 1;
+}
+
+} // namespace v1

src/client/cluster_client.cpp

@@ -1317,7 +1317,8 @@ int cluster_client_t::try_send(cluster_op_t *op, int i)
     auto & pool_cfg = st_cli.pool_config.at(INODE_POOL(op->cur_inode));
     auto pg_it = pool_cfg.pg_config.find(part->pg_num);
     if (pg_it != pool_cfg.pg_config.end() &&
-        !pg_it->second.pause && pg_it->second.cur_primary)
+        !pg_it->second.pause && pg_it->second.cur_primary &&
+        (pg_it->second.cur_state & PG_ACTIVE))
     {
         osd_num_t primary_osd = pg_it->second.cur_primary;
         if (pool_cfg.local_reads != POOL_LOCAL_READ_PRIMARY &&

src/client/cluster_client.h

@@ -127,7 +127,7 @@ public:
     ring_consumer_t consumer;
     std::vector<std::function<void(void)>> on_ready_hooks;
     int list_retry_timeout_id = -1;
-    timespec list_retry_time;
+    timespec list_retry_time = {};
     std::vector<inode_list_t*> lists;
     std::multimap<osd_num_t, osd_op_t*> raw_ops;
     int continuing_ops = 0;

src/client/cluster_client_list.cpp

@@ -31,7 +31,7 @@ struct inode_list_pg_t
     osd_num_t cur_primary = 0;
     int state = 0;
     int inflight_ops = 0;
-    timespec wait_until;
+    timespec wait_until = {};
     std::vector<inode_list_osd_t> list_osds;
 
     bool has_unstable = false;
@@ -177,6 +177,9 @@ void cluster_client_t::retry_start_pg_listing(inode_list_pg_t *pg)
         if (tv.tv_sec < pg->wait_until.tv_sec ||
             tv.tv_sec == pg->wait_until.tv_sec && tv.tv_nsec < pg->wait_until.tv_nsec)
         {
+            // Ensure that the retry timer is set
+            set_list_retry_timeout(pg->wait_until.tv_sec*1000 - tv.tv_sec*1000 +
+                (pg->wait_until.tv_nsec - tv.tv_nsec)/1000000, pg->wait_until);
             return;
         }
     }

src/client/etcd_state_client.cpp

@@ -1167,7 +1167,7 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             if (!cur_primary || !value["state"].is_array() || !state ||
                 (state & PG_OFFLINE) && state != PG_OFFLINE ||
                 (state & PG_PEERING) && state != PG_PEERING ||
-                (state & PG_INCOMPLETE) && state != PG_INCOMPLETE)
+                (state & PG_INCOMPLETE) && state != PG_INCOMPLETE && state != (PG_INCOMPLETE|PG_HAS_INVALID))
             {
                 fprintf(stderr, "Unexpected pool %u PG %u state in etcd: primary=%ju, state=%s\n", pool_id, pg_num, cur_primary, value["state"].dump().c_str());
                 return;

src/client/messenger.cpp

@@ -276,6 +276,7 @@ osd_messenger_t::~osd_messenger_t()
 #ifdef WITH_RDMA
     for (auto rdma_context: rdma_contexts)
     {
+        tfd->set_fd_handler(rdma_context->channel->fd, false, NULL);
         delete rdma_context;
     }
     rdma_contexts.clear();

src/client/msgr_receive.cpp

@@ -109,6 +109,7 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
             if (!handle_read_buffer(cl, cl->in_buf, result))
             {
                 clear_immediate_ops(peer_fd);
+                handle_immediate_ops();
                 return false;
             }
         }
@@ -122,6 +123,7 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
                 if (!handle_finished_read(cl))
                 {
                     clear_immediate_ops(peer_fd);
+                    handle_immediate_ops();
                     return false;
                 }
             }
@@ -140,7 +142,7 @@ void osd_messenger_t::clear_immediate_ops(int peer_fd)
     size_t i = 0, j = 0;
     while (i < set_immediate_ops.size())
     {
-        if (set_immediate_ops[i]->peer_fd == peer_fd)
+        if (set_immediate_ops[i]->peer_fd == peer_fd && set_immediate_ops[i]->op_type == OSD_OP_IN)
         {
             delete set_immediate_ops[i];
         }

src/client/msgr_send.cpp

@@ -323,8 +323,12 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
         }
         if (more)
         {
-            auto expected = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
-            assert(done == expected);
+            int expected = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+            if (done != expected)
+            {
+                fprintf(stderr, "BUG (maybe kernel): Expected to send %d iovecs with MSG_WAITALL but sent %d\n", expected, done);
+                exit(1);
+            }
             cl->zc_free_list.push_back(NULL); // end marker
         }
         if (done > 0)

src/client/msgr_stop.cpp

@@ -52,6 +52,7 @@ void osd_messenger_t::stop_client(int peer_fd, bool force, bool force_delete)
         return;
     }
     osd_client_t *cl = it->second;
+    // FIXME: This 'force' flag is probably an ugly reenterability hack - check its logic and maybe remove it
     if (cl->peer_state == PEER_CONNECTING && !force || cl->peer_state == PEER_STOPPED)
     {
         return;
@@ -73,6 +74,7 @@ void osd_messenger_t::stop_client(int peer_fd, bool force, bool force_delete)
     }
     // First set state to STOPPED so another stop_client() call doesn't try to free it again
     cl->refs++;
+    int prev_state = cl->peer_state;
     cl->peer_state = PEER_STOPPED;
     if (cl->osd_num)
     {
@@ -113,10 +115,12 @@ void osd_messenger_t::stop_client(int peer_fd, bool force, bool force_delete)
         // Break PG locks
         break_pg_locks(cl->in_osd_num);
     }
-    if (cl->osd_num)
+    if (cl->osd_num && prev_state != PEER_CONNECTING)
     {
         // Then repeer PGs because cancel_op() callbacks can try to perform
         // some actions and we need correct PG states to not do something silly
+        // PEER_CONNECTING has neither 'just dropped the connection' nor 'just connected'
+        // so do not repeer on it.
         repeer_pgs(cl->osd_num);
     }
     // Find the item again because it can be invalidated at this point

src/client/nbd_proxy.cpp

@@ -272,6 +272,8 @@ const char *help_text =
     "  --dev_num N\n"
     "    Use the specified device /dev/nbdN instead of automatic selection (alternative syntax\n"
     "    to /dev/nbdN positional parameter).\n"
+    "  --readonly\n"
+    "    Make the device read-only.\n"
     "  --foreground 1\n"
     "    Stay in foreground, do not daemonize.\n"
     "\n"
@@ -372,7 +374,7 @@ public:
             else if (args[i][0] == '-' && args[i][1] == '-')
             {
                 const char *opt = args[i]+2;
-                cfg[opt] = !strcmp(opt, "json") || !strcmp(opt, "all") ||
+                cfg[opt] = !strcmp(opt, "json") || !strcmp(opt, "all") || !strcmp(opt, "readonly") ||
                     !strcmp(opt, "force") || i == narg-1 ? "1" : args[++i];
             }
             else if (pos == 0)
@@ -695,20 +697,6 @@ help:
             ringloop->loop();
             ringloop->wait();
         }
-        stop = false;
-        cluster_op_t *close_sync = new cluster_op_t;
-        close_sync->opcode = OSD_OP_SYNC;
-        close_sync->callback = [&stop](cluster_op_t *op)
-        {
-            stop = true;
-            delete op;
-        };
-        cli->execute(close_sync);
-        while (!stop)
-        {
-            ringloop->loop();
-            ringloop->wait();
-        }
         cli->flush();
         delete cli;
         delete epmgr;
@@ -912,7 +900,7 @@ protected:
             int r, nbd = open(path, O_RDWR), qd_fd;
             if (nbd < 0)
             {
-                write(notifyfd[1], &errno, sizeof(errno));
+                (void)write(notifyfd[1], &errno, sizeof(errno));
                 exit(1);
             }
             r = ioctl(nbd, NBD_SET_SOCK, sockfd[1]);
@@ -954,7 +942,7 @@ protected:
             close(qd_fd);
             // Notify parent
             errno = 0;
-            write(notifyfd[1], &errno, sizeof(errno));
+            (void)write(notifyfd[1], &errno, sizeof(errno));
             close(notifyfd[1]);
             close(sockfd[0]);
             if (bg)
@@ -971,11 +959,11 @@ protected:
             ioctl(nbd, NBD_CLEAR_SOCK);
             exit(0);
     end_close:
-            write(notifyfd[1], &errno, sizeof(errno));
+            (void)write(notifyfd[1], &errno, sizeof(errno));
             close(nbd);
             exit(2);
     end_unmap:
-            write(notifyfd[1], &errno, sizeof(errno));
+            (void)write(notifyfd[1], &errno, sizeof(errno));
             ioctl(nbd, NBD_CLEAR_SOCK);
             close(nbd);
             exit(3);

src/client/object_id.h

@@ -6,8 +6,20 @@
 #include <stdint.h>
 #include <functional>
 
+#define POOL_SCHEME_REPLICATED 1
+#define POOL_SCHEME_XOR 2
+#define POOL_SCHEME_EC 3
+#define POOL_ID_MAX 0x10000
+#define POOL_ID_BITS 16
+#define INODE_POOL(inode) (pool_id_t)((inode) >> (64 - POOL_ID_BITS))
+#define INODE_NO_POOL(inode) (inode_t)((inode) & (((uint64_t)1 << (64-POOL_ID_BITS)) - 1))
+#define INODE_WITH_POOL(pool_id, inode) (((inode_t)(pool_id) << (64-POOL_ID_BITS)) | INODE_NO_POOL(inode))
+
 typedef uint64_t inode_t;
 
+// Pool ID is 16 bits long
+typedef uint32_t pool_id_t;
+
 // 16 bytes per object/stripe id
 // stripe = (start of the parity stripe + peer role)
 // i.e. for example (256KB + one of 0,1,2)

src/client/osd_ops.h

@@ -44,11 +44,6 @@
 #define DIRECT_IO_ALIGNMENT 512
 #endif
 
-// Memory allocation alignment (page size is usually optimal)
-#ifndef MEM_ALIGNMENT
-#define MEM_ALIGNMENT 4096
-#endif
-
 // Constants for osd_reply_describe_item_t.loc_bad
 #define LOC_OUTDATED 1
 #define LOC_CORRUPTED 2

src/client/qemu_driver.c

@@ -974,14 +974,21 @@ static void vitastor_co_read_bitmap_cb(void *opaque, long retval, uint8_t *bitma
 #endif
 }
 
-static int coroutine_fn vitastor_co_block_status(
-    BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
-    int64_t *pnum, int64_t *map, BlockDriverState **file)
+static int coroutine_fn vitastor_co_block_status(BlockDriverState *bs,
+#if QEMU_VERSION_MAJOR > 10 || QEMU_VERSION_MAJOR == 10 && QEMU_VERSION_MINOR >= 1
+    unsigned int mode,
+#else
+    bool want_zero,
+#endif
+    int64_t offset, int64_t bytes, int64_t *pnum, int64_t *map, BlockDriverState **file)
 {
     // Allocated => return BDRV_BLOCK_DATA|BDRV_BLOCK_OFFSET_VALID
     // Not allocated => return 0
     // Error => return -errno
     // Set pnum to length of the extent, `*map` = `offset`, `*file` = `bs`
+#if QEMU_VERSION_MAJOR > 10 || QEMU_VERSION_MAJOR == 10 && QEMU_VERSION_MINOR >= 1
+    int want_zero = (mode == BDRV_WANT_PRECISE);
+#endif
     VitastorRPC task;
     VitastorClient *client = bs->opaque;
     uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;

src/client/ublk_server.cpp

@@ -40,12 +40,14 @@ const char *help_text =
     "    Make the device read-only.\n"
     "  --hdd\n"
     "    Mark the device as rotational.\n"
-    "  --logfile /path/to/log/file.txt\n"
-    "    Write log messages to the specified file instead of dropping them (in background mode)\n"
-    "    or printing them to the standard output (in foreground mode).\n"
     "  --dev_num N\n"
     "    Use the specified device /dev/ublkbN instead of automatic selection (alternative syntax\n"
     "    to /dev/ublkbN positional parameter).\n"
+    "  --pidfile /run/ublk_pid_file.pid\n"
+    "    Write process ID to the specified file.\n"
+    "  --logfile /path/to/log/file.txt\n"
+    "    Write log messages to the specified file instead of dropping them (in background mode)\n"
+    "    or printing them to the standard output (in foreground mode).\n"
     "  --foreground 1\n"
     "    Stay in foreground, do not daemonize.\n"
     "\n"
@@ -79,6 +81,7 @@ protected:
     inode_watch_t *watch = NULL;
 
     std::string logfile = "/dev/null";
+    std::string pidfile;
 
 public:
     ublk_server()
@@ -300,10 +303,8 @@ help:
         load_module();
 
         bool bg = cfg["foreground"].is_null();
-        if (cfg["logfile"].string_value() != "")
-        {
-            logfile = cfg["logfile"].string_value();
-        }
+        logfile = cfg["logfile"].string_value();
+        pidfile = cfg["pidfile"].string_value();
 
         open_control();
         if (recover)
@@ -331,11 +332,13 @@ help:
             close(notifyfd[0]);
         }
         start_device(recover);
+        if (pidfile != "")
+            write_pid();
         if (bg)
         {
             daemonize_reopen_stdio();
             int ok = 0;
-            write(notifyfd[1], &ok, sizeof(ok));
+            (void)write(notifyfd[1], &ok, sizeof(ok));
             close(notifyfd[1]);
         }
         else
@@ -346,19 +349,6 @@ help:
             ringloop->loop();
             ringloop->wait();
         }
-        cluster_op_t *close_sync = new cluster_op_t;
-        close_sync->opcode = OSD_OP_SYNC;
-        close_sync->callback = [this](cluster_op_t *op)
-        {
-            stop = true;
-            delete op;
-        };
-        cli->execute(close_sync);
-        while (!stop)
-        {
-            ringloop->loop();
-            ringloop->wait();
-        }
         cli->flush();
         delete cli;
         delete epmgr;
@@ -390,7 +380,7 @@ help:
             // Parent - check status
             close(notifyfd[1]);
             int child_errno = 1;
-            read(notifyfd[0], &child_errno, sizeof(child_errno));
+            (void)read(notifyfd[0], &child_errno, sizeof(child_errno));
             if (!child_errno)
                 printf("/dev/ublkb%d\n", ublk_dev.dev_id);
             exit(child_errno);
@@ -412,6 +402,22 @@ help:
             fprintf(stderr, "Warning: Failed to chdir into /\n");
     }
 
+    void write_pid()
+    {
+        int fd = open(pidfile.c_str(), O_WRONLY|O_CREAT|O_TRUNC, 0666);
+        if (fd < 0)
+        {
+            fprintf(stderr, "Failed to create pid file %s: %s (code %d)\n", pidfile.c_str(), strerror(errno), errno);
+            return;
+        }
+        auto pid = std::to_string(getpid());
+        if (write(fd, pid.c_str(), pid.size()) < 0)
+        {
+            fprintf(stderr, "Failed to write pid to %s: %s (code %d)\n", pidfile.c_str(), strerror(errno), errno);
+        }
+        close(fd);
+    }
+
     json11::Json::object list_mapped()
     {
         int n_in_dev = 0;
@@ -618,13 +624,13 @@ protected:
             .types = UBLK_PARAM_TYPE_BASIC,
             .basic = {
                 .attrs = attrs, // UBLK_ATTR_READ_ONLY | UBLK_ATTR_ROTATIONAL | UBLK_ATTR_VOLATILE_CACHE | UBLK_ATTR_FUA
-                .logical_bs_shift = 9,
+                .logical_bs_shift = phys_shift,
                 .physical_bs_shift = phys_shift,
                 .io_opt_shift = io_opt_shift,
                 .io_min_shift = phys_shift,
-                .max_sectors = max_io_buf_bytes / phys_block_size,
+                .max_sectors = max_io_buf_bytes / 512,
                 .chunk_sectors = 0,
-                .dev_sectors = device_size / phys_block_size,
+                .dev_sectors = device_size / 512,
                 .virt_boundary_mask = 0,
             },
             .discard = {

src/client/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 2.3.0
+Version: 2.4.3
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

src/cmd/cli.cpp

@@ -37,6 +37,7 @@ static const char* help_text =
     "  --sort FIELD    Sort by specified field (name, size, used_size, <read|write|delete>_<iops|bps|lat|queue>)\n"
     "  -r|--reverse    Sort in descending order\n"
     "  -n|--count N    Only list first N items\n"
+    "  --tree          Show image snapshot/clone tree\n"
     "\n"
     "vitastor-cli create -s|--size <size> [-p|--pool <id|name>] [--parent <parent_name>[@<snapshot>]] <name>\n"
     "  Create an image. You may use K/M/G/T suffixes for <size>. If --parent is specified,\n"
@@ -287,7 +288,7 @@ static json11::Json::object parse_args(int narg, const char *args[])
                 !strcmp(opt, "down-ok") || !strcmp(opt, "down_ok") ||
                 !strcmp(opt, "dry-run") || !strcmp(opt, "dry_run") ||
                 !strcmp(opt, "help") || !strcmp(opt, "all") ||
-                !strcmp(opt, "exact") || !strcmp(opt, "matching") ||
+                !strcmp(opt, "exact") || !strcmp(opt, "matching") || !strcmp(opt, "tree") ||
                 !strcmp(opt, "writers-stopped") || !strcmp(opt, "writers_stopped"))
             {
                 cfg[opt] = "1";