Implement simple "flow control" for RDMA (to check hypothesis about slowdowns)

Important fixes:

- Fix possibly incorrect EC parity chunk updates with EC n+k, k > 1 and when
  the first parity chunk is missing

Minor fixes and improvements:

- Fix incorrect EC free space statistics in vitastor-cli df output
- Speedup vitastor-cli startup in clusters with RDMA
- Remove unused PG "peered" state (previously used to update PG epoch)
- Use sfdisk with just --json in vitastor-disk (--dump --json isn't needed)
- Allow trailing comma in sfdisk output (fixes sfdisk 2.36 compatibility)
- Slightly improve RDMA send/receive code
- Reduce RDMA memory consumption by default (rdma_max_recv/send = 16/8)
- Use vitastor-cli instead of direct etcd interaction in the CSI driver

CMakeLists.txt

@@ -1,7 +1,7 @@
-cmake_minimum_required(VERSION 2.8)
+cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VERSION "0.8.3")
+set(VERSION "0.8.6")
 
 add_subdirectory(src)

VNPL-1.1-RU.txt

@@ -48,9 +48,9 @@ Vitastor, составлены для того, чтобы убедиться,
 интерфейс (прокси), опять же, без открытия в свободный публичный доступ как
 самой программы, так и прокси.
 
-  Сетевая Публичная Лицензия Vitastor разработана специально чтобы
+  Сетевая Публичная Лицензия Vitastor разработана специально, чтобы
 гарантировать, что в таких случаях и модифицированная версия программы, и
-прокси оставались доступными сообществу. Для этого лицензия требует от
+прокси останутся доступными сообществу. Для этого лицензия требует от
 операторов сетевых серверов предоставлять исходный код оригинальной программы,
 а также всех других программ, взаимодействующих с ней на их серверах,
 пользователям этих серверов, на условиях свободных лицензий. Таким образом,

csi/Makefile

@@ -1,4 +1,4 @@
-VERSION ?= v0.8.3
+VERSION ?= v0.8.6
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v0.8.3
+          image: vitalif/vitastor-csi:v0.8.6
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/007-csi-provisioner.yaml

@@ -116,7 +116,7 @@ spec:
             privileged: true
             capabilities:
               add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v0.8.3
+          image: vitalif/vitastor-csi:v0.8.6
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "0.8.3"
+    vitastorCSIDriverVersion = "0.8.6"
 )
 
 // Config struct fills the parameters of request or user input

csi/src/controllerserver.go

@@ -6,11 +6,11 @@ package vitastor
 import (
     "context"
     "encoding/json"
+    "fmt"
     "strings"
     "bytes"
     "strconv"
     "time"
-    "fmt"
     "os"
     "os/exec"
     "io/ioutil"
@@ -21,8 +21,6 @@ import (
     "google.golang.org/grpc/codes"
     "google.golang.org/grpc/status"
 
-    "go.etcd.io/etcd/clientv3"
-
     "github.com/container-storage-interface/spec/lib/go/csi"
 )
 
@@ -114,6 +112,34 @@ func GetConnectionParams(params map[string]string) (map[string]string, []string,
     return ctxVars, etcdUrl, etcdPrefix
 }
 
+func invokeCLI(ctxVars map[string]string, args []string) ([]byte, error)
+{
+    if (ctxVars["etcdUrl"] != "")
+    {
+        args = append(args, "--etcd_address", ctxVars["etcdUrl"])
+    }
+    if (ctxVars["etcdPrefix"] != "")
+    {
+        args = append(args, "--etcd_prefix", ctxVars["etcdPrefix"])
+    }
+    if (ctxVars["configPath"] != "")
+    {
+        args = append(args, "--config_path", ctxVars["configPath"])
+    }
+    c := exec.Command("/usr/bin/vitastor-cli", args...)
+    var stdout, stderr bytes.Buffer
+    c.Stdout = &stdout
+    c.Stderr = &stderr
+    err := c.Run()
+    stderrStr := string(stderr.Bytes())
+    if (err != nil)
+    {
+        klog.Errorf("vitastor-cli %s failed: %s, status %s\n", strings.Join(args, " "), stderrStr, err)
+        return nil, status.Error(codes.Internal, stderrStr+" (status "+err.Error()+")")
+    }
+    return stdout.Bytes(), nil
+}
+
 // Create the volume
 func (cs *ControllerServer) CreateVolume(ctx context.Context, req *csi.CreateVolumeRequest) (*csi.CreateVolumeResponse, error)
 {
@@ -146,128 +172,41 @@ func (cs *ControllerServer) CreateVolume(ctx context.Context, req *csi.CreateVol
         volSize = ((capRange.GetRequiredBytes() + MB - 1) / MB) * MB
     }
 
-    // FIXME: The following should PROBABLY be implemented externally in a management tool
+    ctxVars, etcdUrl, _ := GetConnectionParams(req.Parameters)
-
-    ctxVars, etcdUrl, etcdPrefix := GetConnectionParams(req.Parameters)
     if (len(etcdUrl) == 0)
     {
         return nil, status.Error(codes.InvalidArgument, "no etcdUrl in storage class configuration and no etcd_address in vitastor.conf")
     }
 
-    // Connect to etcd
+    // Create image using vitastor-cli
-    cli, err := clientv3.New(clientv3.Config{
+    _, err := invokeCLI(ctxVars, []string{ "create", volName, "-s", fmt.Sprintf("%v", volSize), "--pool", fmt.Sprintf("%v", poolId) })
-        DialTimeout: ETCD_TIMEOUT,
-        Endpoints: etcdUrl,
-    })
     if (err != nil)
     {
-        return nil, status.Error(codes.Internal, "failed to connect to etcd at "+strings.Join(etcdUrl, ",")+": "+err.Error())
+        if (strings.Index(err.Error(), "already exists") > 0)
-    }
-    defer cli.Close()
-
-    var imageId uint64 = 0
-    for
-    {
-        // Check if the image exists
-        ctx, cancel := context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-        resp, err := cli.Get(ctx, etcdPrefix+"/index/image/"+volName)
-        cancel()
-        if (err != nil)
         {
-            return nil, status.Error(codes.Internal, "failed to read key from etcd: "+err.Error())
+            stat, err := invokeCLI(ctxVars, []string{ "ls", "--json", volName })
-        }
-        if (len(resp.Kvs) > 0)
-        {
-            kv := resp.Kvs[0]
-            var v InodeIndex
-            err := json.Unmarshal(kv.Value, &v)
             if (err != nil)
             {
-                return nil, status.Error(codes.Internal, "invalid /index/image/"+volName+" key in etcd: "+err.Error())
+                return nil, err
             }
-            poolId = v.PoolId
+            var inodeCfg []InodeConfig
-            imageId = v.Id
+            err = json.Unmarshal(stat, &inodeCfg)
-            inodeCfgKey := fmt.Sprintf("/config/inode/%d/%d", poolId, imageId)
-            ctx, cancel := context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-            resp, err := cli.Get(ctx, etcdPrefix+inodeCfgKey)
-            cancel()
             if (err != nil)
             {
-                return nil, status.Error(codes.Internal, "failed to read key from etcd: "+err.Error())
+                return nil, status.Error(codes.Internal, "Invalid JSON in vitastor-cli ls: "+err.Error())
             }
-            if (len(resp.Kvs) == 0)
+            if (len(inodeCfg) == 0)
             {
-                return nil, status.Error(codes.Internal, "missing "+inodeCfgKey+" key in etcd")
+                return nil, status.Error(codes.Internal, "vitastor-cli create said that image already exists, but ls can't find it")
             }
-            var inodeCfg InodeConfig
+            if (inodeCfg[0].Size < uint64(volSize))
-            err = json.Unmarshal(resp.Kvs[0].Value, &inodeCfg)
-            if (err != nil)
-            {
-                return nil, status.Error(codes.Internal, "invalid "+inodeCfgKey+" key in etcd: "+err.Error())
-            }
-            if (inodeCfg.Size < uint64(volSize))
             {
                 return nil, status.Error(codes.Internal, "image "+volName+" is already created, but size is less than expected")
             }
         }
         else
         {
-            // Find a free ID
+            return nil, err
-            // Create image metadata in a transaction verifying that the image doesn't exist yet AND ID is still free
-            maxIdKey := fmt.Sprintf("%s/index/maxid/%d", etcdPrefix, poolId)
-            ctx, cancel := context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-            resp, err := cli.Get(ctx, maxIdKey)
-            cancel()
-            if (err != nil)
-            {
-                return nil, status.Error(codes.Internal, "failed to read key from etcd: "+err.Error())
-            }
-            var modRev int64
-            var nextId uint64
-            if (len(resp.Kvs) > 0)
-            {
-                var err error
-                nextId, err = strconv.ParseUint(string(resp.Kvs[0].Value), 10, 64)
-                if (err != nil)
-                {
-                    return nil, status.Error(codes.Internal, maxIdKey+" contains invalid ID")
-                }
-                modRev = resp.Kvs[0].ModRevision
-                nextId++
-            }
-            else
-            {
-                nextId = 1
-            }
-            inodeIdxJson, _ := json.Marshal(InodeIndex{
-                Id: nextId,
-                PoolId: poolId,
-            })
-            inodeCfgJson, _ := json.Marshal(InodeConfig{
-                Name: volName,
-                Size: uint64(volSize),
-            })
-            ctx, cancel = context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-            txnResp, err := cli.Txn(ctx).If(
-                clientv3.Compare(clientv3.ModRevision(fmt.Sprintf("%s/index/maxid/%d", etcdPrefix, poolId)), "=", modRev),
-                clientv3.Compare(clientv3.CreateRevision(fmt.Sprintf("%s/index/image/%s", etcdPrefix, volName)), "=", 0),
-                clientv3.Compare(clientv3.CreateRevision(fmt.Sprintf("%s/config/inode/%d/%d", etcdPrefix, poolId, nextId)), "=", 0),
-            ).Then(
-                clientv3.OpPut(fmt.Sprintf("%s/index/maxid/%d", etcdPrefix, poolId), fmt.Sprintf("%d", nextId)),
-                clientv3.OpPut(fmt.Sprintf("%s/index/image/%s", etcdPrefix, volName), string(inodeIdxJson)),
-                clientv3.OpPut(fmt.Sprintf("%s/config/inode/%d/%d", etcdPrefix, poolId, nextId), string(inodeCfgJson)),
-            ).Commit()
-            cancel()
-            if (err != nil)
-            {
-                return nil, status.Error(codes.Internal, "failed to commit transaction in etcd: "+err.Error())
-            }
-            if (txnResp.Succeeded)
-            {
-                imageId = nextId
-                break
-            }
-            // Start over if the transaction fails
         }
     }
 
@@ -299,97 +238,12 @@ func (cs *ControllerServer) DeleteVolume(ctx context.Context, req *csi.DeleteVol
     }
     volName := ctxVars["name"]
 
-    _, etcdUrl, etcdPrefix := GetConnectionParams(ctxVars)
+    ctxVars, _, _ = GetConnectionParams(ctxVars)
-    if (len(etcdUrl) == 0)
-    {
-        return nil, status.Error(codes.InvalidArgument, "no etcdUrl in storage class configuration and no etcd_address in vitastor.conf")
-    }
 
-    cli, err := clientv3.New(clientv3.Config{
+    _, err = invokeCLI(ctxVars, []string{ "rm", volName })
-        DialTimeout: ETCD_TIMEOUT,
-        Endpoints: etcdUrl,
-    })
     if (err != nil)
     {
-        return nil, status.Error(codes.Internal, "failed to connect to etcd at "+strings.Join(etcdUrl, ",")+": "+err.Error())
+        return nil, err
-    }
-    defer cli.Close()
-
-    // Find inode by name
-    ctx, cancel := context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-    resp, err := cli.Get(ctx, etcdPrefix+"/index/image/"+volName)
-    cancel()
-    if (err != nil)
-    {
-        return nil, status.Error(codes.Internal, "failed to read key from etcd: "+err.Error())
-    }
-    if (len(resp.Kvs) == 0)
-    {
-        return nil, status.Error(codes.NotFound, "volume "+volName+" does not exist")
-    }
-    var idx InodeIndex
-    err = json.Unmarshal(resp.Kvs[0].Value, &idx)
-    if (err != nil)
-    {
-        return nil, status.Error(codes.Internal, "invalid /index/image/"+volName+" key in etcd: "+err.Error())
-    }
-
-    // Get inode config
-    inodeCfgKey := fmt.Sprintf("%s/config/inode/%d/%d", etcdPrefix, idx.PoolId, idx.Id)
-    ctx, cancel = context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-    resp, err = cli.Get(ctx, inodeCfgKey)
-    cancel()
-    if (err != nil)
-    {
-        return nil, status.Error(codes.Internal, "failed to read key from etcd: "+err.Error())
-    }
-    if (len(resp.Kvs) == 0)
-    {
-        return nil, status.Error(codes.NotFound, "volume "+volName+" does not exist")
-    }
-    var inodeCfg InodeConfig
-    err = json.Unmarshal(resp.Kvs[0].Value, &inodeCfg)
-    if (err != nil)
-    {
-        return nil, status.Error(codes.Internal, "invalid "+inodeCfgKey+" key in etcd: "+err.Error())
-    }
-
-    // Delete inode data by invoking vitastor-cli
-    args := []string{
-        "rm-data", "--etcd_address", strings.Join(etcdUrl, ","),
-        "--pool", fmt.Sprintf("%d", idx.PoolId),
-        "--inode", fmt.Sprintf("%d", idx.Id),
-    }
-    if (ctxVars["configPath"] != "")
-    {
-        args = append(args, "--config_path", ctxVars["configPath"])
-    }
-    c := exec.Command("/usr/bin/vitastor-cli", args...)
-    var stderr bytes.Buffer
-    c.Stdout = nil
-    c.Stderr = &stderr
-    err = c.Run()
-    stderrStr := string(stderr.Bytes())
-    if (err != nil)
-    {
-        klog.Errorf("vitastor-cli rm-data failed: %s, status %s\n", stderrStr, err)
-        return nil, status.Error(codes.Internal, stderrStr+" (status "+err.Error()+")")
-    }
-
-    // Delete inode config in etcd
-    ctx, cancel = context.WithTimeout(context.Background(), ETCD_TIMEOUT)
-    txnResp, err := cli.Txn(ctx).Then(
-        clientv3.OpDelete(fmt.Sprintf("%s/index/image/%s", etcdPrefix, volName)),
-        clientv3.OpDelete(fmt.Sprintf("%s/config/inode/%d/%d", etcdPrefix, idx.PoolId, idx.Id)),
-    ).Commit()
-    cancel()
-    if (err != nil)
-    {
-        return nil, status.Error(codes.Internal, "failed to delete keys in etcd: "+err.Error())
-    }
-    if (!txnResp.Succeeded)
-    {
-        return nil, status.Error(codes.Internal, "failed to delete keys in etcd: transaction failed")
     }
 
     return &csi.DeleteVolumeResponse{}, nil

debian/changelog

@@ -1,10 +1,10 @@
-vitastor (0.8.3-1) unstable; urgency=medium
+vitastor (0.8.6-1) unstable; urgency=medium
 
   * Bugfixes
 
  -- Vitaliy Filippov <vitalif@yourcmc.ru>  Fri, 03 Jun 2022 02:09:44 +0300
 
-vitastor (0.8.3-1) unstable; urgency=medium
+vitastor (0.8.6-1) unstable; urgency=medium
 
   * Implement NFS proxy
   * Add documentation

debian/vitastor.Dockerfile

@@ -34,8 +34,8 @@ RUN set -e -x; \
     mkdir -p /root/packages/vitastor-$REL; \
     rm -rf /root/packages/vitastor-$REL/*; \
     cd /root/packages/vitastor-$REL; \
-    cp -r /root/vitastor vitastor-0.8.3; \
+    cp -r /root/vitastor vitastor-0.8.6; \
-    cd vitastor-0.8.3; \
+    cd vitastor-0.8.6; \
     ln -s /root/fio-build/fio-*/ ./fio; \
     FIO=$(head -n1 fio/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     ls /usr/include/linux/raw.h || cp ./debian/raw.h /usr/include/linux/raw.h; \
@@ -48,8 +48,8 @@ RUN set -e -x; \
     rm -rf a b; \
     echo "dep:fio=$FIO" > debian/fio_version; \
     cd /root/packages/vitastor-$REL; \
-    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.8.3.orig.tar.xz vitastor-0.8.3; \
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.8.6.orig.tar.xz vitastor-0.8.6; \
-    cd vitastor-0.8.3; \
+    cd vitastor-0.8.6; \
     V=$(head -n1 debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     DEBFULLNAME="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v "$V""$REL" "Rebuild for $REL"; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \

docs/config/network.en.md

@@ -19,6 +19,7 @@ between clients, OSDs and etcd.
 - [rdma_max_sge](#rdma_max_sge)
 - [rdma_max_msg](#rdma_max_msg)
 - [rdma_max_recv](#rdma_max_recv)
+- [rdma_max_send](#rdma_max_send)
 - [peer_connect_interval](#peer_connect_interval)
 - [peer_connect_timeout](#peer_connect_timeout)
 - [osd_idle_timeout](#osd_idle_timeout)
@@ -74,6 +75,12 @@ to work. For example, Mellanox ConnectX-3 and older adapters don't have
 Implicit ODP, so they're unsupported by Vitastor. Run `ibv_devinfo -v` as
 root to list available RDMA devices and their features.
 
+Remember that you also have to configure your network switches if you use
+RoCE/RoCEv2, otherwise you may experience unstable performance. Refer to
+the manual of your network vendor for details about setting up the switch
+for RoCEv2 correctly. Usually it means setting up Lossless Ethernet with
+PFC (Priority Flow Control) and ECN (Explicit Congestion Notification).
+
 ## rdma_port_num
 
 - Type: integer
@@ -116,20 +123,30 @@ required to change this parameter.
 ## rdma_max_msg
 
 - Type: integer
-- Default: 1048576
+- Default: 132096
 
 Maximum size of a single RDMA send or receive operation in bytes.
 
 ## rdma_max_recv
 
+- Type: integer
+- Default: 16
+
+Maximum number of RDMA receive buffers per connection (RDMA requires
+preallocated buffers to receive data). Each buffer is `rdma_max_msg` bytes
+in size. So this setting directly affects memory usage: a single Vitastor
+RDMA client uses `rdma_max_recv * rdma_max_msg * OSD_COUNT` bytes of memory.
+Default is roughly 2 MB * number of OSDs.
+
+## rdma_max_send
+
 - Type: integer
 - Default: 8
 
-Maximum number of parallel RDMA receive operations. Note that this number
+Maximum number of outstanding RDMA send operations per connection. Should be
-of receive buffers `rdma_max_msg` in size are allocated for each client,
+less than `rdma_max_recv` so the receiving side doesn't run out of buffers.
-so this setting actually affects memory usage. This is because RDMA receive
+Doesn't affect memory usage - additional memory isn't allocated for send
-operations are (sadly) still not zero-copy in Vitastor. It may be fixed in
+operations.
-later versions.
 
 ## peer_connect_interval
 

docs/config/network.ru.md

@@ -19,6 +19,7 @@
 - [rdma_max_sge](#rdma_max_sge)
 - [rdma_max_msg](#rdma_max_msg)
 - [rdma_max_recv](#rdma_max_recv)
+- [rdma_max_send](#rdma_max_send)
 - [peer_connect_interval](#peer_connect_interval)
 - [peer_connect_timeout](#peer_connect_timeout)
 - [osd_idle_timeout](#osd_idle_timeout)
@@ -78,6 +79,13 @@ Implicit On-Demand Paging (Implicit ODP) и Scatter/Gather (SG). Наприме
 суперпользователя, чтобы посмотреть список доступных RDMA-устройств, их
 параметры и возможности.
 
+Обратите внимание, что если вы используете RoCE/RoCEv2, вам также необходимо
+правильно настроить для него коммутаторы, иначе вы можете столкнуться с
+нестабильной производительностью. Подробную информацию о настройке
+коммутатора для RoCEv2 ищите в документации производителя. Обычно это
+подразумевает настройку сети без потерь на основе PFC (Priority Flow
+Control) и ECN (Explicit Congestion Notification).
+
 ## rdma_port_num
 
 - Тип: целое число
@@ -121,22 +129,32 @@ OSD в любом случае согласовывают реальное зн
 ## rdma_max_msg
 
 - Тип: целое число
-- Значение по умолчанию: 1048576
+- Значение по умолчанию: 132096
 
 Максимальный размер одной RDMA-операции отправки или приёма.
 
 ## rdma_max_recv
 
+- Тип: целое число
+- Значение по умолчанию: 16
+
+Максимальное число буферов для RDMA-приёма данных на одно соединение
+(RDMA требует заранее выделенных буферов для приёма данных). Каждый буфер
+имеет размер `rdma_max_msg` байт. Таким образом, настройка прямо влияет на
+потребление памяти - один Vitastor-клиент с RDMA использует
+`rdma_max_recv * rdma_max_msg * ЧИСЛО_OSD` байт памяти, по умолчанию -
+примерно 2 МБ * число OSD.
+
+## rdma_max_send
+
 - Тип: целое число
 - Значение по умолчанию: 8
 
-Максимальное число параллельных RDMA-операций получения данных. Следует
+Максимальное число RDMA-операций отправки, отправляемых в очередь одного
-иметь в виду, что данное число буферов размером `rdma_max_msg` выделяется
+соединения. Желательно, чтобы оно было меньше `rdma_max_recv`, чтобы
-для каждого подключённого клиентского соединения, так что данная настройка
+у принимающей стороны в процессе работы не заканчивались буферы на приём.
-влияет на потребление памяти. Это так потому, что RDMA-приём данных в
+Не влияет на потребление памяти - дополнительная память на операции отправки
-Vitastor, увы, всё равно не является zero-copy, т.е. всё равно 1 раз
+не выделяется.
-копирует данные в памяти. Данная особенность, возможно, будет исправлена в
-более новых версиях Vitastor.
 
 ## peer_connect_interval
 

docs/config/osd.en.md

@@ -17,6 +17,7 @@ initialization and can be changed with an OSD restart.
 - [autosync_interval](#autosync_interval)
 - [autosync_writes](#autosync_writes)
 - [recovery_queue_depth](#recovery_queue_depth)
+- [recovery_pg_switch](#recovery_pg_switch)
 - [recovery_sync_batch](#recovery_sync_batch)
 - [readonly](#readonly)
 - [no_recovery](#no_recovery)
@@ -115,6 +116,16 @@ Maximum recovery operations per one primary OSD at any given moment of time.
 Currently it's the only parameter available to tune the speed or recovery
 and rebalancing, but it's planned to implement more.
 
+## recovery_pg_switch
+
+- Type: integer
+- Default: 128
+
+Number of recovery operations before switching to recovery of the next PG.
+The idea is to mix all PGs during recovery for more even space and load
+distribution but still benefit from recovery queue depth greater than 1.
+Degraded PGs are anyway scanned first.
+
 ## recovery_sync_batch
 
 - Type: integer

docs/config/osd.ru.md

@@ -18,6 +18,7 @@
 - [autosync_interval](#autosync_interval)
 - [autosync_writes](#autosync_writes)
 - [recovery_queue_depth](#recovery_queue_depth)
+- [recovery_pg_switch](#recovery_pg_switch)
 - [recovery_sync_batch](#recovery_sync_batch)
 - [readonly](#readonly)
 - [no_recovery](#no_recovery)
@@ -119,6 +120,17 @@ OSD, чтобы успевать очищать журнал - без них OSD
 для ускорения или замедления восстановления и перебалансировки данных, но
 в планах реализация других параметров.
 
+## recovery_pg_switch
+
+- Тип: целое число
+- Значение по умолчанию: 128
+
+Число операций восстановления перед переключением на восстановление другой PG.
+Идея заключается в том, чтобы восстанавливать все PG одновременно для более
+равномерного распределения места и нагрузки, но при этом всё равно выигрывать
+от глубины очереди восстановления, большей, чем 1. Деградированные PG в любом
+случае сканируются первыми.
+
 ## recovery_sync_batch
 
 - Тип: целое число

docs/config/src/network.yml

@@ -53,6 +53,12 @@
     to work. For example, Mellanox ConnectX-3 and older adapters don't have
     Implicit ODP, so they're unsupported by Vitastor. Run `ibv_devinfo -v` as
     root to list available RDMA devices and their features.
+
+    Remember that you also have to configure your network switches if you use
+    RoCE/RoCEv2, otherwise you may experience unstable performance. Refer to
+    the manual of your network vendor for details about setting up the switch
+    for RoCEv2 correctly. Usually it means setting up Lossless Ethernet with
+    PFC (Priority Flow Control) and ECN (Explicit Congestion Notification).
   info_ru: |
     Название RDMA-устройства для связи с Vitastor OSD (например, "rocep5s0f0").
     Имейте в виду, что поддержка RDMA в Vitastor требует функций устройства
@@ -61,6 +67,13 @@
     потому не поддерживаются в Vitastor. Запустите `ibv_devinfo -v` от имени
     суперпользователя, чтобы посмотреть список доступных RDMA-устройств, их
     параметры и возможности.
+
+    Обратите внимание, что если вы используете RoCE/RoCEv2, вам также необходимо
+    правильно настроить для него коммутаторы, иначе вы можете столкнуться с
+    нестабильной производительностью. Подробную информацию о настройке
+    коммутатора для RoCEv2 ищите в документации производителя. Обычно это
+    подразумевает настройку сети без потерь на основе PFC (Priority Flow
+    Control) и ECN (Explicit Congestion Notification).
 - name: rdma_port_num
   type: int
   default: 1
@@ -114,26 +127,39 @@
     так что менять этот параметр обычно не нужно.
 - name: rdma_max_msg
   type: int
-  default: 1048576
+  default: 132096
   info: Maximum size of a single RDMA send or receive operation in bytes.
   info_ru: Максимальный размер одной RDMA-операции отправки или приёма.
 - name: rdma_max_recv
+  type: int
+  default: 16
+  info: |
+    Maximum number of RDMA receive buffers per connection (RDMA requires
+    preallocated buffers to receive data). Each buffer is `rdma_max_msg` bytes
+    in size. So this setting directly affects memory usage: a single Vitastor
+    RDMA client uses `rdma_max_recv * rdma_max_msg * OSD_COUNT` bytes of memory.
+    Default is roughly 2 MB * number of OSDs.
+  info_ru: |
+    Максимальное число буферов для RDMA-приёма данных на одно соединение
+    (RDMA требует заранее выделенных буферов для приёма данных). Каждый буфер
+    имеет размер `rdma_max_msg` байт. Таким образом, настройка прямо влияет на
+    потребление памяти - один Vitastor-клиент с RDMA использует
+    `rdma_max_recv * rdma_max_msg * ЧИСЛО_OSD` байт памяти, по умолчанию -
+    примерно 2 МБ * число OSD.
+- name: rdma_max_send
   type: int
   default: 8
   info: |
-    Maximum number of parallel RDMA receive operations. Note that this number
+    Maximum number of outstanding RDMA send operations per connection. Should be
-    of receive buffers `rdma_max_msg` in size are allocated for each client,
+    less than `rdma_max_recv` so the receiving side doesn't run out of buffers.
-    so this setting actually affects memory usage. This is because RDMA receive
+    Doesn't affect memory usage - additional memory isn't allocated for send
-    operations are (sadly) still not zero-copy in Vitastor. It may be fixed in
+    operations.
-    later versions.
   info_ru: |
-    Максимальное число параллельных RDMA-операций получения данных. Следует
+    Максимальное число RDMA-операций отправки, отправляемых в очередь одного
-    иметь в виду, что данное число буферов размером `rdma_max_msg` выделяется
+    соединения. Желательно, чтобы оно было меньше `rdma_max_recv`, чтобы
-    для каждого подключённого клиентского соединения, так что данная настройка
+    у принимающей стороны в процессе работы не заканчивались буферы на приём.
-    влияет на потребление памяти. Это так потому, что RDMA-приём данных в
+    Не влияет на потребление памяти - дополнительная память на операции отправки
-    Vitastor, увы, всё равно не является zero-copy, т.е. всё равно 1 раз
+    не выделяется.
-    копирует данные в памяти. Данная особенность, возможно, будет исправлена в
-    более новых версиях Vitastor.
 - name: peer_connect_interval
   type: sec
   min: 1

docs/config/src/osd.yml

@@ -102,6 +102,20 @@
     момент времени. На данный момент единственный параметр, который можно менять
     для ускорения или замедления восстановления и перебалансировки данных, но
     в планах реализация других параметров.
+- name: recovery_pg_switch
+  type: int
+  default: 128
+  info: |
+    Number of recovery operations before switching to recovery of the next PG.
+    The idea is to mix all PGs during recovery for more even space and load
+    distribution but still benefit from recovery queue depth greater than 1.
+    Degraded PGs are anyway scanned first.
+  info_ru: |
+    Число операций восстановления перед переключением на восстановление другой PG.
+    Идея заключается в том, чтобы восстанавливать все PG одновременно для более
+    равномерного распределения места и нагрузки, но при этом всё равно выигрывать
+    от глубины очереди восстановления, большей, чем 1. Деградированные PG в любом
+    случае сканируются первыми.
 - name: recovery_sync_batch
   type: int
   default: 16

docs/installation/packages.en.md

@@ -9,7 +9,7 @@
 ## Debian
 
 - Trust Vitastor package signing key:
-  `wget -q -O - https://vitastor.io/debian/pubkey | sudo apt-key add -`
+  `wget https://vitastor.io/debian/pubkey.gpg -O /etc/apt/trusted.gpg.d/vitastor.gpg`
 - Add Vitastor package repository to your /etc/apt/sources.list:
   - Debian 11 (Bullseye/Sid): `deb https://vitastor.io/debian bullseye main`
   - Debian 10 (Buster): `deb https://vitastor.io/debian buster main`
@@ -20,8 +20,8 @@
 ## CentOS
 
 - Add Vitastor package repository:
-  - CentOS 7: `yum install https://vitastor.io/rpms/centos/7/vitastor-release-1.0-1.el7.noarch.rpm`
+  - CentOS 7: `yum install https://vitastor.io/rpms/centos/7/vitastor-release.rpm`
-  - CentOS 8: `dnf install https://vitastor.io/rpms/centos/8/vitastor-release-1.0-1.el8.noarch.rpm`
+  - CentOS 8: `dnf install https://vitastor.io/rpms/centos/8/vitastor-release.rpm`
 - Enable EPEL: `yum/dnf install epel-release`
 - Enable additional CentOS repositories:
   - CentOS 7: `yum install centos-release-scl`

docs/installation/packages.ru.md

@@ -9,7 +9,7 @@
 ## Debian
 
 - Добавьте ключ репозитория Vitastor:
-  `wget -q -O - https://vitastor.io/debian/pubkey | sudo apt-key add -`
+  `wget https://vitastor.io/debian/pubkey.gpg -O /etc/apt/trusted.gpg.d/vitastor.gpg`
 - Добавьте репозиторий Vitastor в /etc/apt/sources.list:
   - Debian 11 (Bullseye/Sid): `deb https://vitastor.io/debian bullseye main`
   - Debian 10 (Buster): `deb https://vitastor.io/debian buster main`
@@ -20,8 +20,8 @@
 ## CentOS
 
 - Добавьте в систему репозиторий Vitastor:
-  - CentOS 7: `yum install https://vitastor.io/rpms/centos/7/vitastor-release-1.0-1.el7.noarch.rpm`
+  - CentOS 7: `yum install https://vitastor.io/rpms/centos/7/vitastor-release.rpm`
-  - CentOS 8: `dnf install https://vitastor.io/rpms/centos/8/vitastor-release-1.0-1.el8.noarch.rpm`
+  - CentOS 8: `dnf install https://vitastor.io/rpms/centos/8/vitastor-release.rpm`
 - Включите EPEL: `yum/dnf install epel-release`
 - Включите дополнительные репозитории CentOS:
   - CentOS 7: `yum install centos-release-scl`

docs/intro/quickstart.en.md

@@ -70,7 +70,7 @@ For EC pools the configuration should look like the following:
 
 ```
 etcdctl --endpoints=... put /vitastor/config/pools '{"2":{"name":"ecpool",
-  "scheme":"ec","pg_size":4,"parity_chunks":2,"pg_minsize":2,"pg_count":256,"failure_domain":"host"}`
+  "scheme":"ec","pg_size":4,"parity_chunks":2,"pg_minsize":2,"pg_count":256,"failure_domain":"host"}'
 ```
 
 After you do this, one of the monitors will configure PGs and OSDs will start them.

docs/intro/quickstart.ru.md

@@ -71,7 +71,7 @@ etcdctl --endpoints=... put /vitastor/config/pools '{"1":{"name":"testpool",
 
 ```
 etcdctl --endpoints=... put /vitastor/config/pools '{"2":{"name":"ecpool",
-  "scheme":"ec","pg_size":4,"parity_chunks":2,"pg_minsize":2,"pg_count":256,"failure_domain":"host"}`
+  "scheme":"ec","pg_size":4,"parity_chunks":2,"pg_minsize":2,"pg_count":256,"failure_domain":"host"}'
 ```
 
 После этого один из мониторов должен сконфигурировать PG, а OSD должны запустить их.

docs/performance/theoretical.en.md

@@ -35,15 +35,24 @@ Write amplification for 4 KB blocks is usually 3-5 in Vitastor:
 If you manage to get an SSD which handles 512 byte blocks well (Optane?) you may
 lower 1, 3 and 4 to 512 bytes (1/8 of data size) and get WA as low as 2.375.
 
+Implemented NVDIMM support can basically eliminate WA at all - all extra writes will
+go to DRAM memory. But this requires a test cluster with NVDIMM - please contact me
+if you want to provide me with such cluster for tests.
+
 Lazy fsync also reduces WA for parallel workloads because journal blocks are only
 written when they fill up or fsync is requested.
 
 ## In Practice
 
-In practice, using tests from [Understanding Performance](understanding.en.md)
+In practice, using tests from [Understanding Performance](understanding.en.md), decent TCP network,
-and good server-grade SSD/NVMe drives, you should head for:
+good server-grade SSD/NVMe drives and disabled CPU power saving, you should head for:
 - At least 5000 T1Q1 replicated read and write iops (maximum 0.2ms latency)
+- At least 5000 T1Q1 EC read IOPS and at least 2200 EC write IOPS (maximum 0.45ms latency)
 - At least ~80k parallel read iops or ~30k write iops per 1 core (1 OSD)
 - Disk-speed or wire-speed linear reads and writes, whichever is the bottleneck in your case
 
 Lower results may mean that you have bad drives, bad network or some kind of misconfiguration.
+
+Current latency records:
+- 9668 T1Q1 replicated write iops (0.103 ms latency) with TCP and NVMe
+- 9143 T1Q1 replicated read iops (0.109 ms latency) with TCP and NVMe

docs/performance/theoretical.ru.md

@@ -36,6 +36,25 @@ WA (мультипликатор записи) для 4 КБ блоков в Vit
 Если вы найдёте SSD, хорошо работающий с 512-байтными блоками данных (Optane?),
 то 1, 3 и 4 можно снизить до 512 байт (1/8 от размера данных) и получить WA всего 2.375.
 
+Если реализовать поддержку NVDIMM, то WA можно, условно говоря, ликвидировать вообще - все
+дополнительные операции записи смогут обслуживаться DRAM памятью. Но для этого необходим
+тестовый кластер с NVDIMM - пишите, если готовы предоставить такой для тестов.
+
 Кроме того, WA снижается при использовании отложенного/ленивого сброса при параллельной
 нагрузке, т.к. блоки журнала записываются на диск только когда они заполняются или явным
 образом запрашивается fsync.
+
+## На практике
+
+На практике, используя тесты fio со страницы [Понимание сути производительности систем хранения](understanding.ru.md),
+нормальную TCP-сеть, хорошие серверные SSD/NVMe, при отключённом энергосбережении процессоров вы можете рассчитывать на:
+- От 5000 IOPS в 1 поток (T1Q1) и на чтение, и на запись при использовании репликации (задержка до 0.2мс)
+- От 5000 IOPS в 1 поток (T1Q1) на чтение и 2200 IOPS в 1 поток на запись при использовании EC (задержка до 0.45мс)
+- От 80000 IOPS на чтение в параллельном режиме на 1 ядро, от 30000 IOPS на запись на 1 ядро (на 1 OSD)
+- Скорость параллельного линейного чтения и записи, равная меньшему значению из скорости дисков или сети
+
+Худшие результаты означают, что у вас либо медленные диски, либо медленная сеть, либо что-то неправильно настроено.
+
+Зафиксированный на данный момент рекорд задержки:
+- 9668 IOPS (0.103 мс задержка) в 1 поток (T1Q1) на запись с TCP и NVMe при использовании репликации
+- 9143 IOPS (0.109 мс задержка) в 1 поток (T1Q1) на чтение с TCP и NVMe при использовании репликации

docs/usage/cli.en.md

@@ -14,6 +14,7 @@ It supports the following commands:
 - [df](#df)
 - [ls](#ls)
 - [create](#create)
+- [snap-create](#create)
 - [modify](#modify)
 - [rm](#rm)
 - [flatten](#flatten)
@@ -123,6 +124,8 @@ vitastor-cli snap-create [-p|--pool <id|name>] <image>@<snapshot>
 
 Create a snapshot of image `<name>` (either form can be used). May be used live if only a single writer is active.
 
+See also about [how to export snapshots](qemu.en.md#exporting-snapshots).
+
 ## modify
 
 `vitastor-cli modify <name> [--rename <new-name>] [--resize <size>] [--readonly | --readwrite] [-f|--force]`

docs/usage/cli.ru.md

@@ -15,6 +15,7 @@ vitastor-cli - интерфейс командной строки для адм
 - [df](#df)
 - [ls](#ls)
 - [create](#create)
+- [snap-create](#create)
 - [modify](#modify)
 - [rm](#rm)
 - [flatten](#flatten)
@@ -126,6 +127,8 @@ vitastor-cli snap-create [-p|--pool <id|name>] <image>@<snapshot>
 Создать снимок образа `<name>` (можно использовать любую форму команды). Снимок можно создавать без остановки
 клиентов, если пишущий клиент максимум 1.
 
+Смотрите также информацию о том, [как экспортировать снимки](qemu.ru.md#экспорт-снимков).
+
 ## modify
 
 `vitastor-cli modify <name> [--rename <new-name>] [--resize <size>] [--readonly | --readwrite] [-f|--force]`

docs/usage/disk.en.md

@@ -1,4 +1,4 @@
-[Documentation](../../README.md#documentation) → Usage → Disk Tool
+[Documentation](../../README.md#documentation) → Usage → Disk management tool
 
 -----
 

docs/usage/disk.ru.md

@@ -1,4 +1,4 @@
-[Документация](../../README-ru.md#документация) → Использование → Управление дисками
+[Документация](../../README-ru.md#документация) → Использование → Инструмент управления дисками
 
 -----
 

docs/usage/qemu.en.md

@@ -46,3 +46,40 @@ qemu-img convert -f qcow2 debian10.qcow2 -p -O raw 'vitastor:etcd_host=192.168.7
 
 You can also specify `:pool=<POOL>:inode=<INODE>:size=<SIZE>` instead of `:image=<IMAGE>`
 if you don't want to use inode metadata.
+
+### Exporting snapshots
+
+Starting with 0.8.4, you can also export individual layers (snapshot diffs) using `qemu-img`.
+
+Suppose you have an image `testimg` and a snapshot `testimg@0` created with `vitastor-cli snap-create testimg@0`.
+
+Then you can export the `testimg@0` snapshot and the data written to `testimg` after creating
+the snapshot separately using the following commands (key points are using `skip-parents=1` and
+`-B backing_file` option):
+
+```
+qemu-img convert -f raw 'vitastor:etcd_host=192.168.7.2\:2379/v3:image=testimg@0' \
+    -O qcow2 testimg_0.qcow2
+
+qemu-img convert -f raw 'vitastor:etcd_host=192.168.7.2\:2379/v3:image=testimg:skip-parents=1' \
+    -O qcow2 -o 'cluster_size=4k' -B testimg_0.qcow2 testimg.qcow2
+```
+
+In fact, with `cluster_size=4k` any QCOW2 file can be used instead `-B testimg_0.qcow2`, even an empty one.
+
+QCOW2 `cluster_size=4k` option is required if you want `testimg.qcow2` to contain only the data
+overwritten  **exactly** in the child layer. With the default 64 KB QCOW2 cluster size you'll
+get a bit of extra data from parent layers, i.e. a 4 KB overwrite will result in `testimg.qcow2`
+containing 64 KB of data. And this extra data will be taken by `qemu-img` from the file passed
+in `-B` option, so you really need 4 KB cluster if you use an empty image in `-B`.
+
+After this procedure you'll get two chained QCOW2 images. To detach `testimg.qcow2` from
+its parent, run:
+
+```
+qemu-img rebase -u -b '' testimg.qcow2
+```
+
+This can be used for backups. Just note that exporting an image that is currently being written to
+is of course unsafe and doesn't produce a consistent result, so only export snapshots if you do this
+on a live VM.

docs/usage/qemu.ru.md

@@ -50,3 +50,40 @@ qemu-img convert -f qcow2 debian10.qcow2 -p -O raw 'vitastor:etcd_host=10.115.0.
 
 Если вы не хотите обращаться к образу по имени, вместо `:image=<IMAGE>` можно указать номер пула, номер инода и размер:
 `:pool=<POOL>:inode=<INODE>:size=<SIZE>`.
+
+### Экспорт снимков
+
+Начиная с 0.8.4 вы можете экспортировать отдельные слои (изменения в снимках) с помощью `qemu-img`.
+
+Допустим, что у вас есть образ `testimg` и его снимок `testimg@0`, созданный с помощью `vitastor-cli snap-create testimg@0`.
+
+Тогда вы можете выгрузить снимок `testimg@0` и данные, изменённые в `testimg` после создания снимка, отдельно,
+с помощью следующих команд (ключевые моменты - использование `skip-parents=1` и опции `-B backing_file.qcow2`):
+
+```
+qemu-img convert -f raw 'vitastor:etcd_host=192.168.7.2\:2379/v3:image=testimg@0' \
+    -O qcow2 testimg_0.qcow2
+
+qemu-img convert -f raw 'vitastor:etcd_host=192.168.7.2\:2379/v3:image=testimg:skip-parents=1' \
+    -O qcow2 -o 'cluster_size=4k' -B testimg_0.qcow2 testimg.qcow2
+```
+
+На самом деле, с `cluster_size=4k` вместо `-B testimg_0.qcow2` можно использовать любой qcow2-файл,
+даже пустой.
+
+Опция QCOW2 `cluster_size=4k` нужна, если вы хотите, чтобы `testimg.qcow2` содержал **в точности**
+данные, перезаписанные в дочернем слое. С размером кластера QCOW2 по умолчанию, составляющим 64 КБ,
+вы получите немного "лишних" данных из родительских слоёв - перезапись 4 КБ будет приводить к тому,
+что в `testimg.qcow2` будет появляться 64 КБ данных. Причём "лишние" данные qemu-img будет брать
+как раз из файла, указанного в опции `-B`, так что если там указан пустой образ, кластер обязан быть 4 КБ.
+
+После данной процедуры вы получите два QCOW2-образа, связанных в цепочку. Чтобы "отцепить" образ
+`testimg.qcow2` от базового, выполните:
+
+```
+qemu-img rebase -u -b '' testimg.qcow2
+```
+
+Это можно использовать для резервного копирования. Только помните, что экспортировать образ, в который
+в то же время идёт запись, небезопасно - результат чтения не будет целостным. Так что если вы работаете
+с активными виртуальными машинами, экспортируйте только их снимки, но не сам образ.

mon/PGUtil.js

@@ -21,7 +21,7 @@ function add_pg_history(new_pg_history, new_pg, prev_pgs, prev_pg_history, old_p
     {
         for (const pg of oh.osd_sets)
         {
-            nh.osd_sets[pg.join(' ')] = pg;
+            nh.osd_sets[pg.join(' ')] = pg.map(osd_num => Number(osd_num));
         }
     }
     if (oh && oh.all_peers && oh.all_peers.length)

mon/lp-optimizer.js

@@ -550,8 +550,8 @@ function random_combinations(osd_tree, pg_size, count, ordered)
         seed ^= seed << 5;
         return seed + 2147483648;
     };
-    const hosts = Object.keys(osd_tree).sort();
     const osds = Object.keys(osd_tree).reduce((a, c) => { a[c] = Object.keys(osd_tree[c]).sort(); return a; }, {});
+    const hosts = Object.keys(osd_tree).sort().filter(h => osds[h].length > 0);
     const r = {};
     // Generate random combinations including each OSD at least once
     for (let h = 0; h < hosts.length; h++)

mon/make-etcd

@@ -79,7 +79,7 @@ StartLimitInterval=0
 RestartSec=10
 
 [Install]
-WantedBy=local.target
+WantedBy=multi-user.target
 `);
     await system(`useradd etcd`);
     await system(`systemctl daemon-reload`);

mon/mon.js

@@ -70,9 +70,9 @@ const etcd_tree = {
             rdma_gid_index: 0,
             rdma_mtu: 4096,
             rdma_max_sge: 128,
-            rdma_max_send: 32,
+            rdma_max_send: 8,
-            rdma_max_recv: 8,
+            rdma_max_recv: 16,
-            rdma_max_msg: 1048576,
+            rdma_max_msg: 132096,
             log_level: 0,
             block_size: 131072,
             disk_alignment: 4096,
@@ -261,7 +261,7 @@ const etcd_tree = {
             /* <pool_id>: {
                 <pg_id>: {
                     primary: osd_num_t,
-                    state: ("starting"|"peering"|"peered"|"incomplete"|"active"|"repeering"|"stopping"|"offline"|
+                    state: ("starting"|"peering"|"incomplete"|"active"|"repeering"|"stopping"|"offline"|
                         "degraded"|"has_incomplete"|"has_degraded"|"has_misplaced"|"has_unclean"|
                         "has_invalid"|"left_on_dead")[],
                 }
@@ -663,12 +663,15 @@ class Mon
     async save_last_clean()
     {
         // last_clean_pgs is used to avoid extra data move when observing a series of changes in the cluster
+        const new_clean_pgs = { items: {} };
+    next_pool:
         for (const pool_id in this.state.config.pools)
         {
+            new_clean_pgs.items[pool_id] = (this.state.history.last_clean_pgs.items||{})[pool_id];
             const pool_cfg = this.state.config.pools[pool_id];
             if (!this.validate_pool_cfg(pool_id, pool_cfg, false))
             {
-                continue;
+                continue next_pool;
             }
             for (let pg_num = 1; pg_num <= pool_cfg.pg_count; pg_num++)
             {
@@ -677,17 +680,18 @@ class Mon
                     !(this.state.pg.state[pool_id][pg_num].state instanceof Array))
                 {
                     // Unclean
-                    return;
+                    continue next_pool;
                 }
                 let st = this.state.pg.state[pool_id][pg_num].state.join(',');
                 if (st != 'active' && st != 'active,left_on_dead' && st != 'left_on_dead,active')
                 {
                     // Unclean
-                    return;
+                    continue next_pool;
                 }
             }
+            new_clean_pgs.items[pool_id] = this.state.config.pgs.items[pool_id];
         }
-        this.state.history.last_clean_pgs = JSON.parse(JSON.stringify(this.state.config.pgs));
+        this.state.history.last_clean_pgs = new_clean_pgs;
         await this.etcd_call('/kv/txn', {
             success: [ { requestPut: {
                 key: b64(this.etcd_prefix+'/history/last_clean_pgs'),
@@ -1374,16 +1378,14 @@ class Mon
     // This is required for multiple change events to trigger at most 1 recheck in 1s
     schedule_recheck()
     {
-        if (this.recheck_timer)
+        if (!this.recheck_timer)
         {
-            clearTimeout(this.recheck_timer);
+            this.recheck_timer = setTimeout(() =>
-            this.recheck_timer = null;
+            {
+                this.recheck_timer = null;
+                this.recheck_pgs().catch(this.die);
+            }, this.config.mon_change_timeout || 1000);
         }
-        this.recheck_timer = setTimeout(() =>
-        {
-            this.recheck_timer = null;
-            this.recheck_pgs().catch(this.die);
-        }, this.config.mon_change_timeout || 1000);
     }
 
     sum_op_stats(timestamp, prev_stats)
@@ -1719,11 +1721,11 @@ class Mon
         else if (key_parts[0] === 'osd' && key_parts[1] === 'stats')
         {
             // Recheck OSD tree on OSD addition/deletion
-            if ((!old) != (!kv.value) || old && kv.value && (old.size != kv.value.size || old.time != kv.value.time))
+            if ((!old) != (!kv.value) || old && kv.value && old.size != kv.value.size)
             {
                 this.schedule_recheck();
             }
-            // Recheck PGs <osd_out_time> later
+            // Recheck PGs <osd_out_time> after last OSD statistics report
             this.schedule_next_recheck_at(
                 !this.state.osd.stats[key[2]] ? 0 : this.state.osd.stats[key[2]].time+this.config.osd_out_time
             );

patches/VitastorPlugin.pm

@@ -16,6 +16,11 @@ use PVE::Tools qw(run_command);
 
 use base qw(PVE::Storage::Plugin);
 
+if (@PVE::Storage::Plugin::SHARED_STORAGE)
+{
+    push @PVE::Storage::Plugin::SHARED_STORAGE, 'vitastor';
+}
+
 sub api
 {
     # Trick it :)
@@ -133,9 +138,11 @@ sub properties
 sub options
 {
     return {
+        shared => { optional => 1 },
+        content => { optional => 1 },
         nodes => { optional => 1 },
         disable => { optional => 1 },
-        vitastor_etcd_address => { optional => 1},
+        vitastor_etcd_address => { optional => 1 },
         vitastor_etcd_prefix => { optional => 1 },
         vitastor_config_path => { optional => 1 },
         vitastor_prefix => { optional => 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
 
-VERSION = '0.8.3'
+VERSION = '0.8.6'
 
 LOG = logging.getLogger(__name__)
 

rpm/build-tarball.sh

@@ -25,4 +25,4 @@ rm fio
 mv fio-copy fio
 FIO=`rpm -qi fio | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Version[\s:]+(\S+)/ && print $1; /^Release[\s:]+(\S+)/ && print "-$1"; }'`
 perl -i -pe 's/(Requires:\s*fio)([^\n]+)?/$1 = '$FIO'/' $VITASTOR/rpm/vitastor-el$EL.spec
-tar --transform 's#^#vitastor-0.8.3/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.8.3$(rpm --eval '%dist').tar.gz *
+tar --transform 's#^#vitastor-0.8.6/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.8.6$(rpm --eval '%dist').tar.gz *

rpm/vitastor-el7.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.8.3.el7.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.8.6.el7.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el7.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.8.3
+Version:        0.8.6
 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-0.8.3.el7.tar.gz
+Source0:        vitastor-0.8.6.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel
@@ -35,6 +35,7 @@ Summary:        Vitastor - OSD
 Requires:       libJerasure2
 Requires:       libisa-l
 Requires:       liburing >= 0.6
+Requires:       liburing < 2
 Requires:       vitastor-client = %{version}-%{release}
 Requires:       util-linux
 Requires:       parted
@@ -59,6 +60,7 @@ scheduling cluster-level operations.
 %package -n vitastor-client
 Summary:        Vitastor - client
 Requires:       liburing >= 0.6
+Requires:       liburing < 2
 
 
 %description -n vitastor-client

rpm/vitastor-el8.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.8.3.el8.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.8.6.el8.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el8.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.8.3
+Version:        0.8.6
 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-0.8.3.el8.tar.gz
+Source0:        vitastor-0.8.6.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel
@@ -34,6 +34,7 @@ Summary:        Vitastor - OSD
 Requires:       libJerasure2
 Requires:       libisa-l
 Requires:       liburing >= 0.6
+Requires:       liburing < 2
 Requires:       vitastor-client = %{version}-%{release}
 Requires:       util-linux
 Requires:       parted
@@ -57,6 +58,7 @@ scheduling cluster-level operations.
 %package -n vitastor-client
 Summary:        Vitastor - client
 Requires:       liburing >= 0.6
+Requires:       liburing < 2
 
 
 %description -n vitastor-client

src/CMakeLists.txt

@@ -1,8 +1,9 @@
-cmake_minimum_required(VERSION 2.8)
+cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
 include(GNUInstallDirs)
+include(CTest)
 
 set(WITH_QEMU false CACHE BOOL "Build QEMU driver inside Vitastor source tree")
 set(WITH_FIO true CACHE BOOL "Build FIO driver")
@@ -15,7 +16,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVERSION="0.8.3")
+add_definitions(-DVERSION="0.8.6")
 add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -I ${CMAKE_SOURCE_DIR}/src)
 if (${WITH_ASAN})
 	add_definitions(-fsanitize=address -fno-omit-frame-pointer)
@@ -55,6 +56,14 @@ if (ISAL_LIBRARIES)
 	add_definitions(-DWITH_ISAL)
 endif (ISAL_LIBRARIES)
 
+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(
 	../
 	/usr/include/jerasure
@@ -145,7 +154,6 @@ add_library(vitastor_client SHARED
 set_target_properties(vitastor_client PROPERTIES PUBLIC_HEADER "vitastor_c.h")
 target_link_libraries(vitastor_client
 	vitastor_common
-	tcmalloc_minimal
 	${LIBURING_LIBRARIES}
 	${IBVERBS_LIBRARIES}
 )
@@ -235,8 +243,18 @@ add_executable(osd_test osd_test.cpp rw_blocking.cpp addr_util.cpp)
 target_link_libraries(osd_test tcmalloc_minimal)
 
 # osd_rmw_test
-add_executable(osd_rmw_test osd_rmw_test.cpp allocator.cpp)
+add_executable(osd_rmw_test EXCLUDE_FROM_ALL osd_rmw_test.cpp allocator.cpp)
 target_link_libraries(osd_rmw_test Jerasure ${ISAL_LIBRARIES} tcmalloc_minimal)
+add_dependencies(build_tests osd_rmw_test)
+add_test(NAME osd_rmw_test COMMAND osd_rmw_test)
+
+if (ISAL_LIBRARIES)
+	add_executable(osd_rmw_test_je EXCLUDE_FROM_ALL osd_rmw_test.cpp allocator.cpp)
+	target_compile_definitions(osd_rmw_test_je PUBLIC -DNO_ISAL)
+	target_link_libraries(osd_rmw_test_je Jerasure tcmalloc_minimal)
+	add_dependencies(build_tests osd_rmw_test_je)
+	add_test(NAME osd_rmw_test_jerasure COMMAND osd_rmw_test_je)
+endif (ISAL_LIBRARIES)
 
 # stub_uring_osd
 add_executable(stub_uring_osd
@@ -250,11 +268,15 @@ target_link_libraries(stub_uring_osd
 )
 
 # osd_peering_pg_test
-add_executable(osd_peering_pg_test osd_peering_pg_test.cpp osd_peering_pg.cpp)
+add_executable(osd_peering_pg_test EXCLUDE_FROM_ALL osd_peering_pg_test.cpp osd_peering_pg.cpp)
 target_link_libraries(osd_peering_pg_test tcmalloc_minimal)
+add_dependencies(build_tests osd_peering_pg_test)
+add_test(NAME osd_peering_pg_test COMMAND osd_peering_pg_test)
 
 # test_allocator
-add_executable(test_allocator test_allocator.cpp allocator.cpp)
+add_executable(test_allocator EXCLUDE_FROM_ALL test_allocator.cpp allocator.cpp)
+add_dependencies(build_tests test_allocator)
+add_test(NAME test_allocator COMMAND test_allocator)
 
 # test_cas
 add_executable(test_cas
@@ -274,12 +296,15 @@ target_link_libraries(test_crc32
 
 # test_cluster_client
 add_executable(test_cluster_client
+	EXCLUDE_FROM_ALL
 	test_cluster_client.cpp
 	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
 	etcd_state_client.cpp timerfd_manager.cpp ../json11/json11.cpp
 )
 target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)
 target_include_directories(test_cluster_client PUBLIC ${CMAKE_SOURCE_DIR}/src/mock)
+add_dependencies(build_tests test_cluster_client)
+add_test(NAME test_cluster_client COMMAND test_cluster_client)
 
 ## test_blockstore, test_shit
 #add_executable(test_blockstore test_blockstore.cpp)

src/blockstore_flush.cpp

@@ -162,7 +162,8 @@ void journal_flusher_t::mark_trim_possible()
     if (trim_wanted > 0)
     {
         dequeuing = true;
-        journal_trim_counter++;
+        if (!journal_trim_counter)
+            journal_trim_counter = journal_trim_interval;
         bs->ringloop->wakeup();
     }
 }

src/blockstore_impl.cpp

@@ -193,6 +193,7 @@ void blockstore_impl_t::loop()
             }
             if (wr_st == 2)
             {
+                submit_queue[op_idx] = NULL;
                 new_idx--;
             }
             if (wr_st == 0)
@@ -324,7 +325,7 @@ void blockstore_impl_t::enqueue_op(blockstore_op_t *op)
     {
         // Basic verification not passed
         op->retval = -EINVAL;
-        std::function<void (blockstore_op_t*)>(op->callback)(op);
+        ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
         return;
     }
     if (op->opcode == BS_OP_SYNC_STAB_ALL)
@@ -367,7 +368,7 @@ void blockstore_impl_t::enqueue_op(blockstore_op_t *op)
     }
     if ((op->opcode == BS_OP_WRITE || op->opcode == BS_OP_WRITE_STABLE || op->opcode == BS_OP_DELETE) && !enqueue_write(op))
     {
-        std::function<void (blockstore_op_t*)>(op->callback)(op);
+        ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
         return;
     }
     // Call constructor without allocating memory. We'll call destructor before returning op back
@@ -582,7 +583,7 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
                         replace_stable(dirty_it->first.oid, 0, clean_stable_count, stable_count, stable);
                     }
                 }
-                else if (IS_STABLE(dirty_it->second.state))
+                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))

src/blockstore_journal.h

@@ -16,6 +16,7 @@
 // FIXME: This value should be dynamic i.e. Blockstore ideally shouldn't allow
 // writing more than can be stabilized afterwards
 #define JOURNAL_STABILIZE_RESERVATION 65536
+#define JOURNAL_INSTANT_RESERVATION 131072
 
 // Journal entries
 // Journal entries are linked to each other by their crc32 value

src/blockstore_read.cpp

@@ -286,7 +286,10 @@ void blockstore_impl_t::handle_read_event(ring_data_t *data, blockstore_op_t *op
                 {
                     auto used = --journal.used_sectors[rv.journal_sector-1];
                     if (used == 0)
+                    {
                         journal.used_sectors.erase(rv.journal_sector-1);
+                        flusher->mark_trim_possible();
+                    }
                 }
             }
         }

src/blockstore_rollback.cpp

@@ -127,7 +127,6 @@ resume_4:
     {
         mark_rolled_back(*v);
     }
-    flusher->mark_trim_possible();
     // Acknowledge op
     op->retval = 0;
     FINISH_OP(op);
@@ -232,6 +231,7 @@ void blockstore_impl_t::erase_dirty(blockstore_dirty_db_t::iterator dirty_start,
         if (used == 0)
         {
             journal.used_sectors.erase(dirty_it->second.journal_sector);
+            flusher->mark_trim_possible();
         }
         if (dsk.clean_entry_bitmap_size > sizeof(void*))
         {

src/blockstore_write.cpp

@@ -89,6 +89,9 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
         else
         {
             // Invalid version requested
+#ifdef BLOCKSTORE_DEBUG
+            printf("Write %lx:%lx v%lu requested, but we already have v%lu\n", op->oid.inode, op->oid.stripe, op->version, version);
+#endif
             op->retval = -EEXIST;
             if (!is_del && dsk.clean_entry_bitmap_size > sizeof(void*))
             {
@@ -115,8 +118,8 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
     else if (!wait_del)
         printf("Write %lx:%lx v%lu offset=%u len=%u\n", op->oid.inode, op->oid.stripe, op->version, op->offset, op->len);
 #endif
-    // FIXME No strict need to add it into dirty_db here, it's just left
+    // No strict need to add it into dirty_db here except maybe for listings to return
-    // from the previous implementation where reads waited for writes
+    // correct data when there are inflight operations in the queue
     uint32_t state;
     if (is_del)
         state = BS_ST_DELETE | BS_ST_IN_FLIGHT;
@@ -182,9 +185,15 @@ void blockstore_impl_t::cancel_all_writes(blockstore_op_t *op, blockstore_dirty_
     bool found = false;
     for (auto other_op: submit_queue)
     {
-        // <op> may be present in queue multiple times due to moving operations in submit_queue
+        if (!other_op)
-        if (other_op == 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))
         {
@@ -252,7 +261,8 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
     {
         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_entry_bitmap_size, JOURNAL_STABILIZE_RESERVATION))
+            sizeof(journal_entry_big_write) + dsk.clean_entry_bitmap_size,
+            (dirty_it->second.state & BS_ST_INSTANT) ? JOURNAL_INSTANT_RESERVATION : JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
         }
@@ -332,7 +342,8 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             !space_check.check_available(op, unsynced_big_write_count,
                 sizeof(journal_entry_big_write) + dsk.clean_entry_bitmap_size, 0)
             || !space_check.check_available(op, 1,
-                sizeof(journal_entry_small_write) + dsk.clean_entry_bitmap_size, op->len + JOURNAL_STABILIZE_RESERVATION))
+                sizeof(journal_entry_small_write) + dsk.clean_entry_bitmap_size,
+                op->len + ((dirty_it->second.state & BS_ST_INSTANT) ? JOURNAL_INSTANT_RESERVATION : JOURNAL_STABILIZE_RESERVATION)))
         {
             return 0;
         }
@@ -443,18 +454,19 @@ int blockstore_impl_t::continue_write(blockstore_op_t *op)
 resume_2:
     // Only for the immediate_commit mode: prepare and submit big_write journal entry
     {
-        blockstore_journal_check_t space_check(this);
-        if (!space_check.check_available(op, 1,
-            sizeof(journal_entry_big_write) + dsk.clean_entry_bitmap_size, JOURNAL_STABILIZE_RESERVATION))
-        {
-            return 0;
-        }
-        BS_SUBMIT_CHECK_SQES(1);
         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_big_write) + dsk.clean_entry_bitmap_size,
+            ((dirty_it->second.state & BS_ST_INSTANT) ? JOURNAL_INSTANT_RESERVATION : JOURNAL_STABILIZE_RESERVATION)))
+        {
+            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) + dsk.clean_entry_bitmap_size
@@ -641,7 +653,7 @@ int blockstore_impl_t::dequeue_del(blockstore_op_t *op)
     });
     assert(dirty_it != dirty_db.end());
     blockstore_journal_check_t space_check(this);
-    if (!space_check.check_available(op, 1, sizeof(journal_entry_del), JOURNAL_STABILIZE_RESERVATION))
+    if (!space_check.check_available(op, 1, sizeof(journal_entry_del), JOURNAL_INSTANT_RESERVATION))
     {
         return 0;
     }

src/cli_df.cpp

@@ -121,8 +121,7 @@ resume_1:
             }
             if (pool_cfg.scheme != POOL_SCHEME_REPLICATED)
             {
-                uint64_t pg_real_size = pool_stats[pool_cfg.id]["pg_real_size"].uint64_value();
+                pool_avail *= (pool_cfg.pg_size - pool_cfg.parity_chunks);
-                pool_avail = pg_real_size > 0 ? pool_avail * (pool_cfg.pg_size - pool_cfg.parity_chunks) / pg_real_size : 0;
             }
             pool_stats[pool_cfg.id] = json11::Json::object {
                 { "name", pool_cfg.name },

src/cli_merge.cpp

@@ -403,7 +403,7 @@ struct snap_merger_t
         op->opcode = OSD_OP_READ_BITMAP;
         op->inode = target;
         op->offset = offset;
-        op->len = 0;
+        op->len = target_block_size;
         op->callback = [this](cluster_op_t *op)
         {
             if (op->retval < 0)

src/cli_rm_data.cpp

@@ -92,6 +92,7 @@ struct rm_inode_t
 
     void send_ops(rm_pg_t *cur_list)
     {
+        parent->cli->init_msgr();
         if (parent->cli->msgr.osd_peer_fds.find(cur_list->rm_osd_num) ==
             parent->cli->msgr.osd_peer_fds.end())
         {

src/cli_rm_osd.cpp

@@ -5,6 +5,7 @@
 #include "cli.h"
 #include "cluster_client.h"
 #include "str_util.h"
+#include "epoll_manager.h"
 
 #include <algorithm>
 
@@ -14,13 +15,21 @@ struct rm_osd_t
     cli_tool_t *parent;
 
     bool dry_run, force_warning, force_dataloss;
+    uint64_t etcd_tx_retry_ms = 500;
+    uint64_t etcd_tx_retries = 10000;
     std::vector<uint64_t> osd_ids;
 
     int state = 0;
     cli_result_t result;
 
     std::set<uint64_t> to_remove;
+    std::set<uint64_t> to_restart;
     json11::Json::array pool_effects;
+    json11::Json::array history_updates, history_checks;
+    json11::Json new_pgs, new_clean_pgs;
+    uint64_t new_pgs_mod_rev, new_clean_pgs_mod_rev;
+    uint64_t cur_retry = 0;
+    uint64_t retry_wait = 0;
     bool is_warning, is_dataloss;
 
     bool is_done()
@@ -32,6 +41,12 @@ struct rm_osd_t
     {
         if (state == 1)
             goto resume_1;
+        else if (state == 2)
+            goto resume_2;
+        else if (state == 3)
+            goto resume_3;
+        else if (state == 4)
+            goto resume_4;
         if (!osd_ids.size())
         {
             result = (cli_result_t){ .err = EINVAL, .text = "OSD numbers are not specified" };
@@ -152,14 +167,48 @@ struct rm_osd_t
             result.text = error;
             if (dry_run || is_dataloss && !force_dataloss || is_warning && !force_warning)
             {
-                result.err = is_dataloss || is_warning ? EBUSY : 0;
+                result.err = is_dataloss && !force_dataloss || is_warning && !force_warning ? EBUSY : 0;
                 state = 100;
                 return;
             }
         }
+        parent->etcd_txn(json11::Json::object { { "success", json11::Json::array {
+            json11::Json::object {
+                { "request_range", json11::Json::object {
+                    { "key", base64_encode(
+                        parent->cli->st_cli.etcd_prefix+"/config/pgs"
+                    ) },
+                } },
+            },
+            json11::Json::object {
+                { "request_range", json11::Json::object {
+                    { "key", base64_encode(
+                        parent->cli->st_cli.etcd_prefix+"/history/last_clean_pgs"
+                    ) },
+                } },
+            },
+        } } });
+    resume_4:
+        state = 4;
+        if (parent->waiting > 0)
+            return;
+        if (parent->etcd_err.err)
+        {
+            result = parent->etcd_err;
+            state = 100;
+            return;
+        }
+        {
+            auto kv = parent->cli->st_cli.parse_etcd_kv(parent->etcd_result["responses"][0]["response_range"]["kvs"][0]);
+            new_pgs = remove_osds_from_pgs(kv);
+            new_pgs_mod_rev = kv.mod_revision;
+            kv = parent->cli->st_cli.parse_etcd_kv(parent->etcd_result["responses"][1]["response_range"]["kvs"][0]);
+            new_clean_pgs = remove_osds_from_pgs(kv);
+            new_clean_pgs_mod_rev = kv.mod_revision;
+        }
         // Remove keys from etcd
         {
-            json11::Json::array rm_items;
+            json11::Json::array rm_items, rm_checks;
             for (auto osd_id: osd_ids)
             {
                 rm_items.push_back("/config/osd/"+std::to_string(osd_id));
@@ -178,7 +227,39 @@ struct rm_osd_t
                     } },
                 };
             }
-            parent->etcd_txn(json11::Json::object { { "success", rm_items } });
+            if (!new_pgs.is_null())
+            {
+                auto pgs_key = base64_encode(parent->cli->st_cli.etcd_prefix+"/config/pgs");
+                rm_items.push_back(json11::Json::object {
+                    { "request_put", json11::Json::object {
+                        { "key", pgs_key },
+                        { "value", base64_encode(new_pgs.dump()) },
+                    } },
+                });
+                rm_checks.push_back(json11::Json::object {
+                    { "target", "MOD" },
+                    { "key", pgs_key },
+                    { "result", "LESS" },
+                    { "mod_revision", new_pgs_mod_rev+1 },
+                });
+            }
+            if (!new_clean_pgs.is_null())
+            {
+                auto pgs_key = base64_encode(parent->cli->st_cli.etcd_prefix+"/history/last_clean_pgs");
+                rm_items.push_back(json11::Json::object {
+                    { "request_put", json11::Json::object {
+                        { "key", pgs_key },
+                        { "value", base64_encode(new_clean_pgs.dump()) },
+                    } },
+                });
+                rm_checks.push_back(json11::Json::object {
+                    { "target", "MOD" },
+                    { "key", pgs_key },
+                    { "result", "LESS" },
+                    { "mod_revision", new_clean_pgs_mod_rev+1 },
+                });
+            }
+            parent->etcd_txn(json11::Json::object { { "success", rm_items }, { "checks", rm_checks } });
         }
     resume_1:
         state = 1;
@@ -190,6 +271,46 @@ struct rm_osd_t
             state = 100;
             return;
         }
+        // Remove old OSD from PG all_peers to prevent left_on_dead and from
+        // target_history to prevent INCOMPLETE if --allow-data-loss is specified
+        for (auto & rsp: parent->etcd_result["responses"].array_items())
+        {
+            if (rsp["response_delete_range"]["deleted"].uint64_value() > 0)
+            {
+                // Wait for mon_change_timeout before updating PG history, or the monitor's change will likely interfere with ours
+                retry_wait = parent->cli->merged_config["mon_change_timeout"].uint64_value();
+                if (!retry_wait)
+                    retry_wait = 1000;
+                retry_wait += etcd_tx_retry_ms;
+            }
+        }
+        while (1)
+        {
+    resume_2:
+            if (!remove_osds_from_history(2))
+                return;
+    resume_3:
+            state = 3;
+            if (parent->waiting > 0)
+                return;
+            if (parent->etcd_err.err)
+            {
+                result = parent->etcd_err;
+                state = 100;
+                return;
+            }
+            if (parent->etcd_result["succeeded"].bool_value())
+                break;
+            if ((++cur_retry) >= etcd_tx_retries)
+            {
+                result.err = EAGAIN;
+                result.text += "Failed to remove OSDs from PG history due to update conflicts."
+                    " Some PGs may remain left_on_dead or incomplete. Please retry later\n";
+                state = 100;
+                return;
+            }
+            retry_wait = etcd_tx_retry_ms;
+        }
         std::string ids = "";
         for (auto osd_id: osd_ids)
         {
@@ -200,6 +321,138 @@ struct rm_osd_t
         result.text = (result.text != "" ? ids+"\n"+result.text : ids);
         result.err = 0;
     }
+
+    json11::Json remove_osds_from_pgs(const etcd_kv_t & kv)
+    {
+        if (kv.value.is_null())
+        {
+            return kv.value;
+        }
+        json11::Json::object new_pgs;
+        for (auto & pp: kv.value["items"].object_items())
+        {
+            if (pp.second.is_object())
+            {
+                json11::Json::object new_pool;
+                for (auto & pgp: pp.second.object_items())
+                {
+                    json11::Json::array osd_set;
+                    for (auto & osd_json: pgp.second["osd_set"].array_items())
+                    {
+                        uint64_t osd_num = osd_json.uint64_value();
+                        osd_set.push_back(osd_num == 0 || to_remove.find(osd_num) != to_remove.end() ? 0 : osd_num);
+                    }
+                    json11::Json::object new_pg = pgp.second.object_items();
+                    new_pg["osd_set"] = osd_set;
+                    new_pool[pgp.first] = new_pg;
+                }
+                new_pgs[pp.first] = new_pool;
+            }
+            else
+                new_pgs[pp.first] = pp.second;
+        }
+        auto res = kv.value.object_items();
+        res["items"] = new_pgs;
+        return res;
+    }
+
+    bool remove_osds_from_history(int base_state)
+    {
+        if (state == base_state+0)
+            goto resume_0;
+        history_updates.clear();
+        history_checks.clear();
+        for (auto & pp: parent->cli->st_cli.pool_config)
+        {
+            bool update_pg_history = false;
+            auto & pool_cfg = pp.second;
+            for (auto & pgp: pool_cfg.pg_config)
+            {
+                auto pg_num = pgp.first;
+                auto & pg_cfg = pgp.second;
+                for (int i = 0; i < pg_cfg.all_peers.size(); i++)
+                {
+                    if (to_remove.find(pg_cfg.all_peers[i]) != to_remove.end())
+                    {
+                        update_pg_history = true;
+                        pg_cfg.all_peers.erase(pg_cfg.all_peers.begin()+i, pg_cfg.all_peers.begin()+i+1);
+                        i--;
+                    }
+                }
+                for (int i = 0; i < pg_cfg.target_history.size(); i++)
+                {
+                    int hist_size = 0, hist_rm = 0;
+                    for (auto & old_osd: pg_cfg.target_history[i])
+                    {
+                        if (old_osd != 0)
+                        {
+                            hist_size++;
+                            if (to_remove.find(old_osd) != to_remove.end())
+                            {
+                                hist_rm++;
+                                old_osd = 0;
+                            }
+                        }
+                    }
+                    if (hist_rm > 0)
+                    {
+                        if (hist_size-hist_rm == 0)
+                        {
+                            pg_cfg.target_history.erase(pg_cfg.target_history.begin()+i, pg_cfg.target_history.begin()+i+1);
+                            i--;
+                        }
+                        update_pg_history = true;
+                    }
+                }
+                if (update_pg_history)
+                {
+                    std::string history_key = base64_encode(
+                        parent->cli->st_cli.etcd_prefix+"/pg/history/"+
+                        std::to_string(pool_cfg.id)+"/"+std::to_string(pg_num)
+                    );
+                    history_updates.push_back(json11::Json::object {
+                        { "request_put", json11::Json::object {
+                            { "key", history_key },
+                            { "value", base64_encode(json11::Json(json11::Json::object {
+                                { "epoch", pg_cfg.epoch },
+                                { "all_peers", pg_cfg.all_peers },
+                                { "osd_sets", pg_cfg.target_history },
+                            }).dump()) },
+                        } },
+                    });
+                    history_checks.push_back(json11::Json::object {
+                        { "target", "MOD" },
+                        { "key", history_key },
+                        { "result", "LESS" },
+                        { "mod_revision", parent->cli->st_cli.etcd_watch_revision+1 },
+                    });
+                }
+            }
+        }
+        if (history_updates.size())
+        {
+            if (retry_wait)
+            {
+                parent->waiting++;
+                parent->epmgr->tfd->set_timer(retry_wait, false, [this](int timer_id)
+                {
+                    parent->waiting--;
+                    parent->ringloop->wakeup();
+                });
+    resume_0:
+                state = base_state+0;
+                if (parent->waiting > 0)
+                    return false;
+            }
+            parent->etcd_txn(json11::Json::object {
+                { "success", history_updates },
+                { "compare", history_checks },
+            });
+        }
+        else
+            parent->etcd_result = json11::Json::object{ { "succeeded", true } };
+        return true;
+    }
 };
 
 std::function<bool(cli_result_t &)> cli_tool_t::start_rm_osd(json11::Json cfg)
@@ -209,6 +462,14 @@ std::function<bool(cli_result_t &)> cli_tool_t::start_rm_osd(json11::Json cfg)
     rm_osd->dry_run = cfg["dry_run"].bool_value();
     rm_osd->force_dataloss = cfg["allow_data_loss"].bool_value();
     rm_osd->force_warning = rm_osd->force_dataloss || cfg["force"].bool_value();
+    if (!cfg["etcd_tx_retries"].is_null())
+        rm_osd->etcd_tx_retries = cfg["etcd_tx_retries"].uint64_value();
+    if (!cfg["etcd_tx_retry_ms"].is_null())
+    {
+        rm_osd->etcd_tx_retry_ms = cfg["etcd_tx_retry_ms"].uint64_value();
+        if (rm_osd->etcd_tx_retry_ms < 100)
+            rm_osd->etcd_tx_retry_ms = 100;
+    }
     if (cfg["osd_id"].is_number() || cfg["osd_id"].is_string())
         rm_osd->osd_ids.push_back(cfg["osd_id"].uint64_value());
     else

src/cluster_client.cpp

@@ -59,7 +59,6 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
         delete op;
     };
     msgr.parse_config(this->config);
-    msgr.init();
 
     st_cli.tfd = tfd;
     st_cli.on_load_config_hook = [this](json11::Json::object & cfg) { on_load_config_hook(cfg); };
@@ -73,17 +72,6 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
 
     scrap_buffer_size = SCRAP_BUFFER_SIZE;
     scrap_buffer = malloc_or_die(scrap_buffer_size);
-
-    if (ringloop)
-    {
-        consumer.loop = [this]()
-        {
-            msgr.read_requests();
-            msgr.send_replies();
-            this->ringloop->submit();
-        };
-        ringloop->register_consumer(&consumer);
-    }
 }
 
 cluster_client_t::~cluster_client_t()
@@ -115,6 +103,24 @@ cluster_op_t::~cluster_op_t()
     }
 }
 
+void cluster_client_t::init_msgr()
+{
+    if (msgr_initialized)
+        return;
+    msgr.init();
+    msgr_initialized = true;
+    if (ringloop)
+    {
+        consumer.loop = [this]()
+        {
+            msgr.read_requests();
+            msgr.send_replies();
+            this->ringloop->submit();
+        };
+        ringloop->register_consumer(&consumer);
+    }
+}
+
 void cluster_client_t::calc_wait(cluster_op_t *op)
 {
     op->prev_wait = 0;
@@ -143,7 +149,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
         if (!op->prev_wait)
             continue_sync(op);
     }
-    else /* if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP) */
+    else /* if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP) */
     {
         for (auto prev = op_queue_head; prev && prev != op; prev = prev->next)
         {
@@ -151,7 +157,8 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
             {
                 op->prev_wait++;
             }
-            else if (prev->opcode == OSD_OP_WRITE || prev->opcode == OSD_OP_READ || prev->opcode == OSD_OP_READ_BITMAP)
+            else if (prev->opcode == OSD_OP_WRITE || prev->opcode == OSD_OP_READ ||
+                prev->opcode == OSD_OP_READ_BITMAP || prev->opcode == OSD_OP_READ_CHAIN_BITMAP)
             {
                 // Flushes are always in the beginning (we're scanning from the beginning of the queue)
                 break;
@@ -171,7 +178,8 @@ void cluster_client_t::inc_wait(uint64_t opcode, uint64_t flags, cluster_op_t *n
             auto n2 = next->next;
             if (next->opcode == OSD_OP_SYNC && !(flags & OP_IMMEDIATE_COMMIT) ||
                 next->opcode == OSD_OP_WRITE && (flags & OP_FLUSH_BUFFER) && !(next->flags & OP_FLUSH_BUFFER) ||
-                (next->opcode == OSD_OP_READ || next->opcode == OSD_OP_READ_BITMAP) && (flags & OP_FLUSH_BUFFER))
+                (next->opcode == OSD_OP_READ || next->opcode == OSD_OP_READ_BITMAP ||
+                    next->opcode == OSD_OP_READ_CHAIN_BITMAP) && (flags & OP_FLUSH_BUFFER))
             {
                 next->prev_wait += inc;
                 assert(next->prev_wait >= 0);
@@ -221,11 +229,14 @@ void cluster_client_t::erase_op(cluster_op_t *op)
     if (op_queue_tail == op)
         op_queue_tail = op->prev;
     op->next = op->prev = NULL;
+    if (flags & OP_FLUSH_BUFFER)
+        std::function<void(cluster_op_t*)>(op->callback)(op);
     if (!(flags & OP_IMMEDIATE_COMMIT))
         inc_wait(opcode, flags, next, -1);
     // Call callback at the end to avoid inconsistencies in prev_wait
     // if the callback adds more operations itself
-    std::function<void(cluster_op_t*)>(op->callback)(op);
+    if (!(flags & OP_FLUSH_BUFFER))
+        std::function<void(cluster_op_t*)>(op->callback)(op);
 }
 
 void cluster_client_t::continue_ops(bool up_retry)
@@ -337,7 +348,8 @@ void cluster_client_t::on_change_hook(std::map<std::string, etcd_kv_t> & changes
             // And now they have to be resliced!
             for (auto op = op_queue_head; op; op = op->next)
             {
-                if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP) &&
+                if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_READ ||
+                    op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP) &&
                     INODE_POOL(op->cur_inode) == pool_item.first)
                 {
                     op->needs_reslice = true;
@@ -409,7 +421,7 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
 void cluster_client_t::execute(cluster_op_t *op)
 {
     if (op->opcode != OSD_OP_SYNC && op->opcode != OSD_OP_READ &&
-        op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_WRITE)
+        op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_READ_CHAIN_BITMAP && op->opcode != OSD_OP_WRITE)
     {
         op->retval = -EINVAL;
         std::function<void(cluster_op_t*)>(op->callback)(op);
@@ -441,7 +453,7 @@ void cluster_client_t::execute(cluster_op_t *op)
             return;
         }
         // Check alignment
-        if ((op->opcode == OSD_OP_READ || op->opcode == OSD_OP_WRITE) && !op->len ||
+        if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
             op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
         {
             op->retval = -EINVAL;
@@ -702,8 +714,7 @@ resume_3:
         // Finished successfully
         // Even if the PG count has changed in meanwhile we treat it as success
         // because if some operations were invalid for the new PG count we'd get errors
-        bool is_read = op->opcode == OSD_OP_READ;
+        if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
-        if (is_read)
         {
             // Check parent inode
             auto ino_it = st_cli.inode_config.find(op->cur_inode);
@@ -727,6 +738,11 @@ resume_3:
             }
         }
         op->retval = op->len;
+        if (op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
+        {
+            auto & pool_cfg = st_cli.pool_config.at(INODE_POOL(op->inode));
+            op->retval = op->len / pool_cfg.bitmap_granularity;
+        }
         erase_op(op);
         return 1;
     }
@@ -750,7 +766,10 @@ resume_3:
         {
             for (int i = 0; i < op->parts.size(); i++)
             {
-                op->parts[i].flags = PART_RETRY;
+                if (!(op->parts[i].flags & PART_DONE))
+                {
+                    op->parts[i].flags = PART_RETRY;
+                }
             }
             goto resume_2;
         }
@@ -809,23 +828,19 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
     uint64_t last_stripe = op->len > 0 ? ((op->offset + op->len - 1) / pg_block_size) * pg_block_size : first_stripe;
     op->retval = 0;
     op->parts.resize((last_stripe - first_stripe) / pg_block_size + 1);
-    if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP)
+    if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
     {
         // Allocate memory for the bitmap
-        unsigned object_bitmap_size = (((op->opcode == OSD_OP_READ_BITMAP ? pg_block_size : op->len) / pool_cfg.bitmap_granularity + 7) / 8);
+        unsigned object_bitmap_size = ((op->len / pool_cfg.bitmap_granularity + 7) / 8);
         object_bitmap_size = (object_bitmap_size < 8 ? 8 : object_bitmap_size);
         unsigned bitmap_mem = object_bitmap_size + (pool_cfg.data_block_size / pool_cfg.bitmap_granularity / 8 * pg_data_size) * op->parts.size();
-        if (op->bitmap_buf_size < bitmap_mem)
+        if (!op->bitmap_buf || op->bitmap_buf_size < bitmap_mem)
         {
             op->bitmap_buf = realloc_or_die(op->bitmap_buf, bitmap_mem);
-            if (!op->bitmap_buf_size)
-            {
-                // First allocation
-                memset(op->bitmap_buf, 0, object_bitmap_size);
-            }
             op->part_bitmaps = (uint8_t*)op->bitmap_buf + object_bitmap_size;
             op->bitmap_buf_size = bitmap_mem;
         }
+        memset(op->bitmap_buf, 0, bitmap_mem);
     }
     int iov_idx = 0;
     size_t iov_pos = 0;
@@ -876,13 +891,14 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             if (end == begin)
                 op->done_count++;
         }
-        else if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_DELETE)
+        else if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_READ_CHAIN_BITMAP && op->opcode != OSD_OP_DELETE)
         {
             add_iov(end-begin, false, op, iov_idx, iov_pos, op->parts[i].iov, NULL, 0);
         }
         op->parts[i].parent = op;
         op->parts[i].offset = begin;
-        op->parts[i].len = op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_DELETE ? 0 : (uint32_t)(end - begin);
+        op->parts[i].len = op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP ||
+            op->opcode == OSD_OP_DELETE ? 0 : (uint32_t)(end - begin);
         op->parts[i].pg_num = pg_num;
         op->parts[i].osd_num = 0;
         op->parts[i].flags = 0;
@@ -911,6 +927,10 @@ bool cluster_client_t::affects_osd(uint64_t inode, uint64_t offset, uint64_t len
 
 bool cluster_client_t::try_send(cluster_op_t *op, int i)
 {
+    if (!msgr_initialized)
+    {
+        init_msgr();
+    }
     auto part = &op->parts[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);
@@ -929,7 +949,7 @@ bool cluster_client_t::try_send(cluster_op_t *op, int i)
                 pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pool_cfg.pg_size-pool_cfg.parity_chunks
             );
             uint64_t meta_rev = 0;
-            if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_DELETE)
+            if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_READ_CHAIN_BITMAP && op->opcode != OSD_OP_DELETE)
             {
                 auto ino_it = st_cli.inode_config.find(op->inode);
                 if (ino_it != st_cli.inode_config.end())
@@ -942,7 +962,7 @@ bool cluster_client_t::try_send(cluster_op_t *op, int i)
                     .header = {
                         .magic = SECONDARY_OSD_OP_MAGIC,
                         .id = next_op_id(),
-                        .opcode = op->opcode == OSD_OP_READ_BITMAP ? OSD_OP_READ : op->opcode,
+                        .opcode = op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP ? OSD_OP_READ : op->opcode,
                     },
                     .inode = op->cur_inode,
                     .offset = part->offset,
@@ -950,8 +970,10 @@ bool cluster_client_t::try_send(cluster_op_t *op, int i)
                     .meta_revision = meta_rev,
                     .version = op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE ? op->version : 0,
                 } },
-                .bitmap = (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP ? (uint8_t*)op->part_bitmaps + pg_bitmap_size*i : NULL),
+                .bitmap = (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP
-                .bitmap_len = (unsigned)(op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP ? pg_bitmap_size : 0),
+                    ? (uint8_t*)op->part_bitmaps + pg_bitmap_size*i : NULL),
+                .bitmap_len = (unsigned)(op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP
+                    ? pg_bitmap_size : 0),
                 .callback = [this, part](osd_op_t *op_part)
                 {
                     handle_op_part(part);
@@ -1130,11 +1152,11 @@ void cluster_client_t::handle_op_part(cluster_op_part_t *part)
     else
     {
         // OK
-        if (!(op->flags & OP_IMMEDIATE_COMMIT))
+        if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE) && !(op->flags & OP_IMMEDIATE_COMMIT))
             dirty_osds.insert(part->osd_num);
         part->flags |= PART_DONE;
         op->done_count++;
-        if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP)
+        if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
         {
             copy_part_bitmap(op, part);
             op->version = op->parts.size() == 1 ? part->op.reply.rw.version : 0;
@@ -1158,7 +1180,12 @@ void cluster_client_t::copy_part_bitmap(cluster_op_t *op, cluster_op_part_t *par
     );
     uint32_t object_offset = (part->op.req.rw.offset - op->offset) / pool_cfg.bitmap_granularity;
     uint32_t part_offset = (part->op.req.rw.offset % pg_block_size) / pool_cfg.bitmap_granularity;
-    uint32_t part_len = (op->opcode == OSD_OP_READ_BITMAP ? pg_block_size : part->op.req.rw.len) / pool_cfg.bitmap_granularity;
+    uint32_t op_len = op->len / pool_cfg.bitmap_granularity;
+    uint32_t part_len = pg_block_size/pool_cfg.bitmap_granularity - part_offset;
+    if (part_len > op_len-object_offset)
+    {
+        part_len = op_len-object_offset;
+    }
     if (!(object_offset & 0x7) && !(part_offset & 0x7) && (part_len >= 8))
     {
         // Copy bytes

src/cluster_client.h

@@ -11,6 +11,7 @@
 #define INODE_LIST_DONE 1
 #define INODE_LIST_HAS_UNSTABLE 2
 #define OSD_OP_READ_BITMAP OSD_OP_SEC_READ_BMP
+#define OSD_OP_READ_CHAIN_BITMAP 0x102
 
 #define OSD_OP_IGNORE_READONLY 0x08
 
@@ -30,7 +31,7 @@ struct cluster_op_part_t
 
 struct cluster_op_t
 {
-    uint64_t opcode; // OSD_OP_READ, OSD_OP_WRITE, OSD_OP_SYNC, OSD_OP_DELETE, OSD_OP_READ_BITMAP
+    uint64_t opcode; // OSD_OP_READ, OSD_OP_WRITE, OSD_OP_SYNC, OSD_OP_DELETE, OSD_OP_READ_BITMAP, OSD_OP_READ_CHAIN_BITMAP
     uint64_t inode;
     uint64_t offset;
     uint64_t len;
@@ -39,9 +40,13 @@ struct cluster_op_t
     uint64_t version = 0;
     // now only OSD_OP_IGNORE_READONLY is supported
     uint64_t flags = 0;
+    // negative retval is an error number
+    // write and read return len on success
+    // sync and delete return 0 on success
+    // read_bitmap and read_chain_bitmap return the length of bitmap in bits(!)
     int retval;
     osd_op_buf_list_t iov;
-    // READ and READ_BITMAP return the bitmap here
+    // READ, READ_BITMAP, READ_CHAIN_BITMAP return the bitmap here
     void *bitmap_buf = NULL;
     std::function<void(cluster_op_t*)> callback;
     ~cluster_op_t();
@@ -99,10 +104,14 @@ class cluster_client_t
     std::vector<std::function<void(void)>> on_ready_hooks;
     std::vector<inode_list_t*> lists;
     int continuing_ops = 0;
+    bool msgr_initialized = false;
 
 public:
     etcd_state_client_t st_cli;
+
     osd_messenger_t msgr;
+    void init_msgr();
+
     json11::Json config;
     json11::Json::object merged_config;
 

src/disk_tool_journal.cpp

@@ -281,7 +281,7 @@ void disk_tool_t::dump_journal_entry(int num, journal_entry *je, bool json)
         if (je->big_write.size > sizeof(journal_entry_big_write))
         {
             printf(json ? ",\"bitmap\":\"" : " (bitmap: ");
-            for (int i = sizeof(journal_entry_big_write); i < je->small_write.size; i++)
+            for (int i = sizeof(journal_entry_big_write); i < je->big_write.size; i++)
             {
                 printf("%02x", ((uint8_t*)je)[i]);
             }

src/disk_tool_meta.cpp

@@ -26,7 +26,7 @@ int disk_tool_t::process_meta(std::function<void(blockstore_meta_header_v1_t *)>
         buf_size = dsk.meta_len;
     void *data = memalign_or_die(MEM_ALIGNMENT, buf_size);
     lseek64(dsk.meta_fd, dsk.meta_offset, 0);
-    read_blocking(dsk.meta_fd, data, buf_size);
+    read_blocking(dsk.meta_fd, data, dsk.meta_block_size);
     // Check superblock
     blockstore_meta_header_v1_t *hdr = (blockstore_meta_header_v1_t *)data;
     if (hdr->zero == 0 &&
@@ -41,8 +41,11 @@ int disk_tool_t::process_meta(std::function<void(blockstore_meta_header_v1_t *)>
             if (buf_size % dsk.meta_block_size)
             {
                 buf_size = 8*dsk.meta_block_size;
+                void *new_data = memalign_or_die(MEM_ALIGNMENT, buf_size);
+                memcpy(new_data, data, dsk.meta_block_size);
                 free(data);
-                data = memalign_or_die(MEM_ALIGNMENT, buf_size);
+                data = new_data;
+                hdr = (blockstore_meta_header_v1_t *)data;
             }
         }
         dsk.bitmap_granularity = hdr->bitmap_granularity;

src/disk_tool_udev.cpp

@@ -387,6 +387,14 @@ int disk_tool_t::purge_devices(const std::vector<std::string> & devices)
         rm_osd_cli.push_back(std::to_string(osd_num));
     }
     // Check for data loss
+    if (options["force"] != "")
+    {
+        rm_osd_cli.push_back("--force");
+    }
+    else if (options["allow_data_loss"] != "")
+    {
+        rm_osd_cli.push_back("--allow-data-loss");
+    }
     rm_osd_cli.push_back("--dry-run");
     std::string dry_run_ignore_stdout;
     if (shell_exec(rm_osd_cli, "", &dry_run_ignore_stdout, NULL) != 0)
@@ -405,14 +413,6 @@ int disk_tool_t::purge_devices(const std::vector<std::string> & devices)
     }
     // Remove OSD metadata
     rm_osd_cli.pop_back();
-    if (options["force"] != "")
-    {
-        rm_osd_cli.push_back("--force");
-    }
-    else if (options["allow_data_loss"] != "")
-    {
-        rm_osd_cli.push_back("--allow-data-loss");
-    }
     if (shell_exec(rm_osd_cli, "", NULL, NULL) != 0)
     {
         return 1;

src/disk_tool_utils.cpp

@@ -305,10 +305,10 @@ int write_zero(int fd, uint64_t offset, uint64_t size)
 json11::Json read_parttable(std::string dev)
 {
     std::string part_dump;
-    int r = shell_exec({ "sfdisk", "--dump", dev, "--json" }, "", &part_dump, NULL);
+    int r = shell_exec({ "sfdisk", "--json", dev }, "", &part_dump, NULL);
     if (r == 255)
     {
-        fprintf(stderr, "Error running sfdisk --dump %s --json\n", dev.c_str());
+        fprintf(stderr, "Error running sfdisk --json %s\n", dev.c_str());
         return json11::Json(false);
     }
     // Decode partition table
@@ -319,7 +319,7 @@ json11::Json read_parttable(std::string dev)
         pt = json11::Json::parse(part_dump, err);
         if (err != "")
         {
-            fprintf(stderr, "sfdisk --dump %s --json returned bad JSON: %s\n", dev.c_str(), part_dump.c_str());
+            fprintf(stderr, "sfdisk --json %s returned bad JSON: %s\n", dev.c_str(), part_dump.c_str());
             return json11::Json(false);
         }
         pt = pt["partitiontable"];

src/etcd_state_client.cpp

@@ -871,19 +871,33 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             pg_cfg.target_history.clear();
             pg_cfg.all_peers.clear();
             // Refuse to start PG if any set of the <osd_sets> has no live OSDs
-            for (auto hist_item: value["osd_sets"].array_items())
+            for (auto & hist_item: value["osd_sets"].array_items())
             {
                 std::vector<osd_num_t> history_set;
-                for (auto pg_osd: hist_item.array_items())
+                for (auto & pg_osd: hist_item.array_items())
                 {
-                    history_set.push_back(pg_osd.uint64_value());
+                    osd_num_t pg_osd_num = pg_osd.uint64_value();
+                    if (pg_osd_num != 0)
+                    {
+                        auto it = std::lower_bound(history_set.begin(), history_set.end(), pg_osd_num);
+                        if (it == history_set.end() || *it != pg_osd_num)
+                            history_set.insert(it, pg_osd_num);
+                    }
                 }
-                pg_cfg.target_history.push_back(history_set);
+                auto it = std::lower_bound(pg_cfg.target_history.begin(), pg_cfg.target_history.end(), history_set);
+                if (it == pg_cfg.target_history.end() || *it != history_set)
+                    pg_cfg.target_history.insert(it, history_set);
             }
             // Include these additional OSDs when peering the PG
             for (auto pg_osd: value["all_peers"].array_items())
             {
-                pg_cfg.all_peers.push_back(pg_osd.uint64_value());
+                osd_num_t pg_osd_num = pg_osd.uint64_value();
+                if (pg_osd_num != 0)
+                {
+                    auto it = std::lower_bound(pg_cfg.all_peers.begin(), pg_cfg.all_peers.end(), pg_osd_num);
+                    if (it == pg_cfg.all_peers.end() || *it != pg_osd_num)
+                        pg_cfg.all_peers.insert(it, pg_osd_num);
+                }
             }
             // Read epoch
             pg_cfg.epoch = value["epoch"].uint64_value();

src/messenger.cpp

@@ -157,10 +157,10 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         this->rdma_max_sge = 128;
     this->rdma_max_send = config["rdma_max_send"].uint64_value();
     if (!this->rdma_max_send)
-        this->rdma_max_send = 1;
+        this->rdma_max_send = 8;
     this->rdma_max_recv = config["rdma_max_recv"].uint64_value();
     if (!this->rdma_max_recv)
-        this->rdma_max_recv = 128;
+        this->rdma_max_recv = 16;
     this->rdma_max_msg = config["rdma_max_msg"].uint64_value();
     if (!this->rdma_max_msg || this->rdma_max_msg > 128*1024*1024)
         this->rdma_max_msg = 129*1024;

src/messenger.h

@@ -138,6 +138,7 @@ protected:
 
     std::vector<int> read_ready_clients;
     std::vector<int> write_ready_clients;
+    // We don't use ringloop->set_immediate here because we may have no ringloop in client :)
     std::vector<std::function<void()>> set_immediate;
 
 public:

src/msgr_rdma.cpp

@@ -353,8 +353,10 @@ static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
         .wr_id = (uint64_t)(cl->peer_fd*2+1),
         .sg_list = sge,
         .num_sge = op_sge,
-        .opcode = IBV_WR_SEND,
+        .opcode = cl->rdma_conn->avail_recv > 0 ? IBV_WR_SEND_WITH_IMM : IBV_WR_SEND,
         .send_flags = IBV_SEND_SIGNALED,
+        // Notify peer about our available incoming buffers
+        .imm_data = (uint32_t)cl->rdma_conn->avail_recv,
     };
     int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
     if (err || bad_wr)
@@ -363,14 +365,25 @@ static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
         exit(1);
     }
     cl->rdma_conn->cur_send++;
+    cl->rdma_conn->avail_send--;
+    cl->rdma_conn->avail_recv = 0;
 }
 
 bool osd_messenger_t::try_send_rdma(osd_client_t *cl)
 {
     auto rc = cl->rdma_conn;
-    if (!cl->send_list.size() || rc->cur_send > 0)
+    if (rc->cur_send >= rc->max_send)
     {
-        // Only send one batch at a time
+        return true;
+    }
+    if (!cl->send_list.size() || rc->use_flow_control && rc->avail_send <= 0)
+    {
+        if (rc->avail_recv)
+        {
+            // Only notify about available buffers so 2 peers don't lock each other
+            rc->send_sizes.push_back(0);
+            try_send_rdma_wr(cl, NULL, 0);
+        }
         return true;
     }
     uint64_t op_size = 0, op_sge = 0;
@@ -380,6 +393,7 @@ bool osd_messenger_t::try_send_rdma(osd_client_t *cl)
         iovec & iov = cl->send_list[rc->send_pos];
         if (op_size >= rc->max_msg || op_sge >= rc->max_sge)
         {
+            rc->send_sizes.push_back(op_size);
             try_send_rdma_wr(cl, sge, op_sge);
             op_sge = 0;
             op_size = 0;
@@ -405,18 +419,24 @@ bool osd_messenger_t::try_send_rdma(osd_client_t *cl)
     }
     if (op_sge > 0)
     {
+        rc->send_sizes.push_back(op_size);
         try_send_rdma_wr(cl, sge, op_sge);
     }
     return true;
 }
 
-static void try_recv_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
+static void try_recv_rdma_wr(osd_client_t *cl, void *buf)
 {
+    ibv_sge sge = {
+        .addr = (uintptr_t)buf,
+        .length = (uint32_t)cl->rdma_conn->max_msg,
+        .lkey = cl->rdma_conn->ctx->mr->lkey,
+    };
     ibv_recv_wr *bad_wr = NULL;
     ibv_recv_wr wr = {
         .wr_id = (uint64_t)(cl->peer_fd*2),
-        .sg_list = sge,
+        .sg_list = &sge,
-        .num_sge = op_sge,
+        .num_sge = 1,
     };
     int err = ibv_post_recv(cl->rdma_conn->qp, &wr, &bad_wr);
     if (err || bad_wr)
@@ -425,6 +445,7 @@ static void try_recv_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
         exit(1);
     }
     cl->rdma_conn->cur_recv++;
+    cl->rdma_conn->avail_recv++;
 }
 
 bool osd_messenger_t::try_recv_rdma(osd_client_t *cl)
@@ -434,12 +455,7 @@ bool osd_messenger_t::try_recv_rdma(osd_client_t *cl)
     {
         void *buf = malloc_or_die(rc->max_msg);
         rc->recv_buffers.push_back(buf);
-        ibv_sge sge = {
+        try_recv_rdma_wr(cl, buf);
-            .addr = (uintptr_t)buf,
-            .length = (uint32_t)rc->max_msg,
-            .lkey = rc->ctx->mr->lkey,
-        };
-        try_recv_rdma_wr(cl, &sge, 1);
     }
     return true;
 }
@@ -476,6 +492,7 @@ void osd_messenger_t::handle_rdma_events()
                 continue;
             }
             osd_client_t *cl = cl_it->second;
+            auto rc = cl->rdma_conn;
             if (wc[i].status != IBV_WC_SUCCESS)
             {
                 fprintf(stderr, "RDMA work request failed for client %d", client_id);
@@ -489,44 +506,65 @@ void osd_messenger_t::handle_rdma_events()
             }
             if (!is_send)
             {
-                cl->rdma_conn->cur_recv--;
+                rc->cur_recv--;
-                if (!handle_read_buffer(cl, cl->rdma_conn->recv_buffers[0], wc[i].byte_len))
+                if ((wc[i].wc_flags & IBV_WC_WITH_IMM) && wc[i].imm_data > 0)
+                {
+                    rc->avail_send += wc[i].imm_data;
+                    rc->use_flow_control = true;
+                }
+                if (!handle_read_buffer(cl, rc->recv_buffers[rc->next_recv_buf], wc[i].byte_len))
                 {
                     // handle_read_buffer may stop the client
                     continue;
                 }
-                free(cl->rdma_conn->recv_buffers[0]);
+                try_recv_rdma_wr(cl, rc->recv_buffers[rc->next_recv_buf]);
-                cl->rdma_conn->recv_buffers.erase(cl->rdma_conn->recv_buffers.begin(), cl->rdma_conn->recv_buffers.begin()+1);
+                rc->next_recv_buf = (rc->next_recv_buf+1) % rc->recv_buffers.size();
-                try_recv_rdma(cl);
+                try_send_rdma(cl);
             }
             else
             {
-                cl->rdma_conn->cur_send--;
+                rc->cur_send--;
-                if (!cl->rdma_conn->cur_send)
+                uint64_t sent_size = rc->send_sizes.at(0);
+                rc->send_sizes.erase(rc->send_sizes.begin(), rc->send_sizes.begin()+1);
+                int send_pos = 0, send_buf_pos = 0;
+                while (sent_size > 0)
                 {
-                    // Wait for the whole batch
+                    if (sent_size >= cl->send_list.at(send_pos).iov_len)
-                    for (int i = 0; i < cl->rdma_conn->send_pos; i++)
                     {
-                        if (cl->outbox[i].flags & MSGR_SENDP_FREE)
+                        sent_size -= cl->send_list[send_pos].iov_len;
-                        {
+                        send_pos++;
-                            // Reply fully sent
-                            delete cl->outbox[i].op;
-                        }
                     }
-                    if (cl->rdma_conn->send_pos > 0)
+                    else
                     {
-                        cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+cl->rdma_conn->send_pos);
+                        send_buf_pos = sent_size;
-                        cl->outbox.erase(cl->outbox.begin(), cl->outbox.begin()+cl->rdma_conn->send_pos);
+                        sent_size = 0;
-                        cl->rdma_conn->send_pos = 0;
                     }
-                    if (cl->rdma_conn->send_buf_pos > 0)
-                    {
-                        cl->send_list[0].iov_base = (uint8_t*)cl->send_list[0].iov_base + cl->rdma_conn->send_buf_pos;
-                        cl->send_list[0].iov_len -= cl->rdma_conn->send_buf_pos;
-                        cl->rdma_conn->send_buf_pos = 0;
-                    }
-                    try_send_rdma(cl);
                 }
+                assert(rc->send_pos >= send_pos);
+                if (rc->send_pos == send_pos)
+                {
+                    rc->send_buf_pos -= send_buf_pos;
+                }
+                rc->send_pos -= send_pos;
+                for (int i = 0; i < send_pos; i++)
+                {
+                    if (cl->outbox[i].flags & MSGR_SENDP_FREE)
+                    {
+                        // Reply fully sent
+                        delete cl->outbox[i].op;
+                    }
+                }
+                if (send_pos > 0)
+                {
+                    cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+send_pos);
+                    cl->outbox.erase(cl->outbox.begin(), cl->outbox.begin()+send_pos);
+                }
+                if (send_buf_pos > 0)
+                {
+                    cl->send_list[0].iov_base = (uint8_t*)cl->send_list[0].iov_base + send_buf_pos;
+                    cl->send_list[0].iov_len -= send_buf_pos;
+                }
+                try_send_rdma(cl);
             }
         }
     } while (event_count > 0);

src/msgr_rdma.h

@@ -45,12 +45,14 @@ struct msgr_rdma_connection_t
     ibv_qp *qp = NULL;
     msgr_rdma_address_t addr;
     int max_send = 0, max_recv = 0, max_sge = 0;
-    int cur_send = 0, cur_recv = 0;
+    int cur_send = 0, cur_recv = 0, avail_recv = 0, avail_send = 0;
+    bool use_flow_control = false;
     uint64_t max_msg = 0;
 
     int send_pos = 0, send_buf_pos = 0;
-    int recv_pos = 0, recv_buf_pos = 0;
+    int next_recv_buf = 0;
     std::vector<void*> recv_buffers;
+    std::vector<uint64_t> send_sizes;
 
     ~msgr_rdma_connection_t();
     static msgr_rdma_connection_t *create(msgr_rdma_context_t *ctx, uint32_t max_send, uint32_t max_recv, uint32_t max_sge, uint32_t max_msg);

src/osd.cpp

@@ -44,9 +44,10 @@ osd_t::osd_t(const json11::Json & config, ring_loop_t *ringloop)
     // FIXME: Use timerfd_interval based directly on io_uring
     this->tfd = epmgr->tfd;
 
-    auto bs_cfg = json_to_bs(this->config);
+    if (!json_is_true(this->config["disable_blockstore"]))
-    this->bs = new blockstore_t(bs_cfg, ringloop, tfd);
     {
+        auto bs_cfg = json_to_bs(this->config);
+        this->bs = new blockstore_t(bs_cfg, ringloop, tfd);
         // Autosync based on the number of unstable writes to prevent stalls due to insufficient journal space
         uint64_t max_autosync = bs->get_journal_size() / bs->get_block_size() / 2;
         if (autosync_writes > max_autosync)
@@ -93,7 +94,8 @@ osd_t::~osd_t()
 {
     ringloop->unregister_consumer(&consumer);
     delete epmgr;
-    delete bs;
+    if (bs)
+        delete bs;
     close(listen_fd);
     free(zero_buffer);
 }
@@ -163,6 +165,9 @@ void osd_t::parse_config(const json11::Json & config, bool allow_disk_params)
     recovery_queue_depth = config["recovery_queue_depth"].uint64_value();
     if (recovery_queue_depth < 1 || recovery_queue_depth > MAX_RECOVERY_QUEUE)
         recovery_queue_depth = DEFAULT_RECOVERY_QUEUE;
+    recovery_pg_switch = config["recovery_pg_switch"].uint64_value();
+    if (recovery_pg_switch < 1)
+        recovery_pg_switch = DEFAULT_RECOVERY_PG_SWITCH;
     recovery_sync_batch = config["recovery_sync_batch"].uint64_value();
     if (recovery_sync_batch < 1 || recovery_sync_batch > MAX_RECOVERY_QUEUE)
         recovery_sync_batch = DEFAULT_RECOVERY_BATCH;
@@ -472,7 +477,7 @@ void osd_t::print_slow()
             }
         }
     }
-    if (has_slow)
+    if (has_slow && bs)
     {
         bs->dump_diagnostics();
     }

src/osd.h

@@ -34,6 +34,7 @@
 #define DEFAULT_AUTOSYNC_WRITES 128
 #define MAX_RECOVERY_QUEUE 2048
 #define DEFAULT_RECOVERY_QUEUE 4
+#define DEFAULT_RECOVERY_PG_SWITCH 128
 #define DEFAULT_RECOVERY_BATCH 16
 
 //#define OSD_STUB
@@ -108,6 +109,7 @@ class osd_t
     int autosync_interval = DEFAULT_AUTOSYNC_INTERVAL; // "emergency" sync every 5 seconds
     int autosync_writes = DEFAULT_AUTOSYNC_WRITES;
     int recovery_queue_depth = DEFAULT_RECOVERY_QUEUE;
+    int recovery_pg_switch = DEFAULT_RECOVERY_PG_SWITCH;
     int recovery_sync_batch = DEFAULT_RECOVERY_BATCH;
     int inode_vanish_time = 60;
     int log_level = 0;
@@ -135,7 +137,10 @@ class osd_t
     uint64_t misplaced_objects = 0, degraded_objects = 0, incomplete_objects = 0;
     int peering_state = 0;
     std::map<object_id, osd_recovery_op_t> recovery_ops;
-    int recovery_done = 0;
+    bool recovery_last_degraded = true;
+    pool_pg_num_t recovery_last_pg;
+    object_id recovery_last_oid;
+    int recovery_pg_done = 0, recovery_done = 0;
     osd_op_t *autosync_op = NULL;
 
     // Unstable writes
@@ -147,7 +152,7 @@ class osd_t
 
     bool stopping = false;
     int inflight_ops = 0;
-    blockstore_t *bs;
+    blockstore_t *bs = NULL;
     void *zero_buffer = NULL;
     uint64_t zero_buffer_size = 0;
     uint32_t bs_block_size, bs_bitmap_granularity, clean_entry_bitmap_size;
@@ -200,7 +205,6 @@ class osd_t
     bool check_peer_config(osd_client_t *cl, json11::Json conf);
     void repeer_pgs(osd_num_t osd_num);
     void start_pg_peering(pg_t & pg);
-    void submit_sync_and_list_subop(osd_num_t role_osd, pg_peering_state_t *ps);
     void submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps);
     void discard_list_subop(osd_op_t *list_op);
     bool stop_pg(pg_t & pg);

src/osd_cluster.cpp

@@ -132,7 +132,7 @@ bool osd_t::check_peer_config(osd_client_t *cl, json11::Json conf)
                 this->osd_num, immediate_commit == IMMEDIATE_ALL ? "all" : "small",
                 cl->osd_num, conf["immediate_commit"].string_value().c_str()
             );
-            return true;
+            return false;
         }
         else if (conf["block_size"].uint64_value() != (uint64_t)this->bs_block_size)
         {
@@ -140,7 +140,7 @@ bool osd_t::check_peer_config(osd_client_t *cl, json11::Json conf)
                 "[OSD %lu] My block_size is %u, but peer OSD %lu has %lu. We can't work together\n",
                 this->osd_num, this->bs_block_size, cl->osd_num, conf["block_size"].uint64_value()
             );
-            return true;
+            return false;
         }
         else if (conf["bitmap_granularity"].uint64_value() != (uint64_t)this->bs_bitmap_granularity)
         {
@@ -148,7 +148,7 @@ bool osd_t::check_peer_config(osd_client_t *cl, json11::Json conf)
                 "[OSD %lu] My bitmap_granularity is %u, but peer OSD %lu has %lu. We can't work together\n",
                 this->osd_num, this->bs_bitmap_granularity, cl->osd_num, conf["bitmap_granularity"].uint64_value()
             );
-            return true;
+            return false;
         }
     }
     return true;
@@ -182,10 +182,10 @@ json11::Json osd_t::get_statistics()
     char time_str[50] = { 0 };
     sprintf(time_str, "%ld.%03ld", ts.tv_sec, ts.tv_nsec/1000000);
     st["time"] = time_str;
-    st["blockstore_ready"] = bs->is_started();
-    st["data_block_size"] = (uint64_t)bs->get_block_size();
     if (bs)
     {
+        st["blockstore_ready"] = bs->is_started();
+        st["data_block_size"] = (uint64_t)bs->get_block_size();
         st["size"] = bs->get_block_count() * bs->get_block_size();
         st["free"] = bs->get_free_block_count() * bs->get_block_size();
     }
@@ -233,7 +233,8 @@ void osd_t::report_statistics()
     json11::Json::object inode_space;
     json11::Json::object last_stat;
     pool_id_t last_pool = 0;
-    auto & bs_inode_space = bs->get_inode_space_stats();
+    std::map<uint64_t, uint64_t> bs_empty_space;
+    auto & bs_inode_space = bs ? bs->get_inode_space_stats() : bs_empty_space;
     for (auto kv: bs_inode_space)
     {
         pool_id_t pool_id = INODE_POOL(kv.first);
@@ -382,30 +383,6 @@ void osd_t::on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes
     }
 }
 
-void osd_t::on_change_pg_history_hook(pool_id_t pool_id, pg_num_t pg_num)
-{
-    auto pg_it = pgs.find({
-        .pool_id = pool_id,
-        .pg_num = pg_num,
-    });
-    if (pg_it != pgs.end() && pg_it->second.epoch > pg_it->second.reported_epoch &&
-        st_cli.pool_config[pool_id].pg_config[pg_num].epoch >= pg_it->second.epoch)
-    {
-        pg_it->second.reported_epoch = st_cli.pool_config[pool_id].pg_config[pg_num].epoch;
-        object_id oid = { 0 };
-        bool first = true;
-        for (auto op: pg_it->second.write_queue)
-        {
-            if (first || oid != op.first)
-            {
-                oid = op.first;
-                first = false;
-                continue_primary_write(op.second);
-            }
-        }
-    }
-}
-
 void osd_t::on_load_config_hook(json11::Json::object & global_config)
 {
     json11::Json::object osd_config = this->config;
@@ -704,13 +681,16 @@ void osd_t::apply_pg_config()
                         }
                     }
                 }
+                auto vec_all_peers = std::vector<osd_num_t>(all_peers.begin(), all_peers.end());
                 if (currently_taken)
                 {
-                    if (pg_it->second.state & (PG_ACTIVE | PG_INCOMPLETE | PG_PEERING | PG_REPEERING | PG_PEERED))
+                    if (pg_it->second.state & (PG_ACTIVE | PG_INCOMPLETE | PG_PEERING | PG_REPEERING))
                     {
-                        if (pg_it->second.target_set == pg_cfg.target_set)
+                        if (pg_it->second.target_set == pg_cfg.target_set &&
+                            pg_it->second.target_history == pg_cfg.target_history &&
+                            pg_it->second.all_peers == vec_all_peers)
                         {
-                            // No change in osd_set; history changes are ignored
+                            // No change in osd_set and history
                             continue;
                         }
                         else
@@ -761,7 +741,7 @@ void osd_t::apply_pg_config()
                     .pg_num = pg_num,
                     .reported_epoch = pg_cfg.epoch,
                     .target_history = pg_cfg.target_history,
-                    .all_peers = std::vector<osd_num_t>(all_peers.begin(), all_peers.end()),
+                    .all_peers = vec_all_peers,
                     .target_set = pg_cfg.target_set,
                 };
                 if (pg.scheme == POOL_SCHEME_EC)
@@ -984,13 +964,6 @@ void osd_t::report_pg_states()
                         }
                         this->pgs.erase(pg_it);
                     }
-                    else if (pg_it->second.state & PG_PEERED)
-                    {
-                        // Activate PG after PG PEERED state is reported along with history
-                        // (if the state wasn't changed again)
-                        pg_it->second.state = pg_it->second.state & ~PG_PEERED | PG_ACTIVE;
-                        report_pg_state(pg_it->second);
-                    }
                 }
             }
             // Push other PG state updates, if any

src/osd_flush.cpp

@@ -226,42 +226,51 @@ bool osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t
 
 bool osd_t::pick_next_recovery(osd_recovery_op_t &op)
 {
-    if (!no_recovery)
+    if (!pgs.size())
     {
-        for (auto pg_it = pgs.begin(); pg_it != pgs.end(); pg_it++)
+        return false;
-        {
-            if ((pg_it->second.state & (PG_ACTIVE | PG_HAS_DEGRADED)) == (PG_ACTIVE | PG_HAS_DEGRADED))
-            {
-                for (auto obj_it = pg_it->second.degraded_objects.begin(); obj_it != pg_it->second.degraded_objects.end(); obj_it++)
-                {
-                    if (recovery_ops.find(obj_it->first) == recovery_ops.end())
-                    {
-                        op.degraded = true;
-                        op.oid = obj_it->first;
-                        return true;
-                    }
-                }
-            }
-        }
     }
-    if (!no_rebalance)
+    // Restart scanning from the same degraded/misplaced status as the last time
+    for (int tried_degraded = 0; tried_degraded < 2; tried_degraded++)
     {
-        for (auto pg_it = pgs.begin(); pg_it != pgs.end(); pg_it++)
+        if (recovery_last_degraded ? !no_recovery : !no_rebalance)
         {
             // Don't try to "recover" misplaced objects if "recovery" would make them degraded
-            if ((pg_it->second.state & (PG_ACTIVE | PG_DEGRADED | PG_HAS_MISPLACED)) == (PG_ACTIVE | PG_HAS_MISPLACED))
+            auto mask = recovery_last_degraded ? (PG_ACTIVE | PG_HAS_DEGRADED) : (PG_ACTIVE | PG_DEGRADED | PG_HAS_MISPLACED);
+            auto check = recovery_last_degraded ? (PG_ACTIVE | PG_HAS_DEGRADED) : (PG_ACTIVE | PG_HAS_MISPLACED);
+            // Restart scanning from the same PG as the last time
+            for (auto pg_it = pgs.lower_bound(recovery_last_pg); pg_it != pgs.end(); pg_it++)
             {
-                for (auto obj_it = pg_it->second.misplaced_objects.begin(); obj_it != pg_it->second.misplaced_objects.end(); obj_it++)
+                if ((pg_it->second.state & mask) == check)
                 {
-                    if (recovery_ops.find(obj_it->first) == recovery_ops.end())
+                    auto & src = recovery_last_degraded ? pg_it->second.degraded_objects : pg_it->second.misplaced_objects;
+                    assert(src.size() > 0);
+                    // Restart scanning from the next object
+                    for (auto obj_it = src.upper_bound(recovery_last_oid); obj_it != src.end(); obj_it++)
                     {
-                        op.degraded = false;
+                        if (recovery_ops.find(obj_it->first) == recovery_ops.end())
-                        op.oid = obj_it->first;
+                        {
-                        return true;
+                            op.degraded = recovery_last_degraded;
+                            recovery_last_oid = op.oid = obj_it->first;
+                            recovery_pg_done++;
+                            // Switch to another PG after recovery_pg_switch operations
+                            // to always mix all PGs during recovery but still benefit
+                            // from recovery queue depth greater than 1
+                            if (recovery_pg_done >= recovery_pg_switch)
+                            {
+                                recovery_pg_done = 0;
+                                recovery_last_pg.pg_num++;
+                                recovery_last_oid = {};
+                            }
+                            return true;
+                        }
                     }
                 }
             }
         }
+        recovery_last_degraded = !recovery_last_degraded;
+        recovery_last_pg = {};
+        recovery_last_oid = {};
     }
     return false;
 }

src/osd_id.h

@@ -28,3 +28,13 @@ inline bool operator < (const pool_pg_num_t & a, const pool_pg_num_t & b)
 {
     return a.pool_id < b.pool_id || a.pool_id == b.pool_id && a.pg_num < b.pg_num;
 }
+
+inline bool operator == (const pool_pg_num_t & a, const pool_pg_num_t & b)
+{
+    return a.pool_id == b.pool_id && a.pg_num == b.pg_num;
+}
+
+inline bool operator != (const pool_pg_num_t & a, const pool_pg_num_t & b)
+{
+    return a.pool_id != b.pool_id || a.pg_num != b.pg_num;
+}

src/osd_peering.cpp

@@ -32,7 +32,16 @@ void osd_t::handle_peers()
                     if (p.second.state & PG_HAS_UNCLEAN)
                         peering_state = peering_state | OSD_FLUSHING_PGS;
                     else if (p.second.state & (PG_HAS_DEGRADED | PG_HAS_MISPLACED))
+                    {
                         peering_state = peering_state | OSD_RECOVERING;
+                        if (p.second.state & PG_HAS_DEGRADED)
+                        {
+                            // Restart recovery from degraded objects
+                            recovery_last_degraded = true;
+                            recovery_last_pg = {};
+                            recovery_last_oid = {};
+                        }
+                    }
                     ringloop->wakeup();
                     return;
                 }
@@ -41,10 +50,6 @@ void osd_t::handle_peers()
                     still = true;
                 }
             }
-            else if (p.second.state & PG_PEERED)
-            {
-                still = true;
-            }
         }
         if (!still)
         {
@@ -65,10 +70,6 @@ void osd_t::handle_peers()
                 }
                 still = true;
             }
-            else if (p.second.state & PG_PEERED)
-            {
-                still = true;
-            }
         }
         if (!still)
         {
@@ -91,7 +92,7 @@ void osd_t::repeer_pgs(osd_num_t peer_osd)
     {
         auto & pg = p.second;
         bool repeer = false;
-        if (pg.state & (PG_PEERING | PG_PEERED | PG_ACTIVE | PG_INCOMPLETE))
+        if (pg.state & (PG_PEERING | PG_ACTIVE | PG_INCOMPLETE))
         {
             for (osd_num_t pg_osd: pg.all_peers)
             {
@@ -302,82 +303,11 @@ void osd_t::start_pg_peering(pg_t & pg)
         {
             continue;
         }
-        submit_sync_and_list_subop(peer_osd, pg.peering_state);
+        submit_list_subop(peer_osd, pg.peering_state);
     }
     ringloop->wakeup();
 }
 
-void osd_t::submit_sync_and_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
-{
-    // Sync before listing, if not readonly
-    if (readonly)
-    {
-        submit_list_subop(role_osd, ps);
-    }
-    else if (role_osd == this->osd_num)
-    {
-        // Self
-        osd_op_t *op = new osd_op_t();
-        op->op_type = 0;
-        op->peer_fd = SELF_FD;
-        clock_gettime(CLOCK_REALTIME, &op->tv_begin);
-        op->bs_op = new blockstore_op_t();
-        op->bs_op->opcode = BS_OP_SYNC;
-        op->bs_op->callback = [this, ps, op, role_osd](blockstore_op_t *bs_op)
-        {
-            if (bs_op->retval < 0)
-            {
-                printf("Local OP_SYNC failed: %d (%s)\n", bs_op->retval, strerror(-bs_op->retval));
-                force_stop(1);
-                return;
-            }
-            add_bs_subop_stats(op);
-            delete op->bs_op;
-            op->bs_op = NULL;
-            delete op;
-            ps->list_ops.erase(role_osd);
-            submit_list_subop(role_osd, ps);
-        };
-        ps->list_ops[role_osd] = op;
-        bs->enqueue_op(op->bs_op);
-    }
-    else
-    {
-        // Peer
-        auto & cl = msgr.clients.at(msgr.osd_peer_fds.at(role_osd));
-        osd_op_t *op = new osd_op_t();
-        op->op_type = OSD_OP_OUT;
-        op->peer_fd = cl->peer_fd;
-        op->req = (osd_any_op_t){
-            .sec_sync = {
-                .header = {
-                    .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
-                    .opcode = OSD_OP_SEC_SYNC,
-                },
-            },
-        };
-        op->callback = [this, ps, role_osd](osd_op_t *op)
-        {
-            if (op->reply.hdr.retval < 0)
-            {
-                // FIXME: Mark peer as failed and don't reconnect immediately after dropping the connection
-                printf("Failed to sync OSD %lu: %ld (%s), disconnecting peer\n", role_osd, op->reply.hdr.retval, strerror(-op->reply.hdr.retval));
-                int fail_fd = op->peer_fd;
-                ps->list_ops.erase(role_osd);
-                delete op;
-                msgr.stop_client(fail_fd);
-                return;
-            }
-            delete op;
-            ps->list_ops.erase(role_osd);
-            submit_list_subop(role_osd, ps);
-        };
-        ps->list_ops[role_osd] = op;
-        msgr.outbox_push(op);
-    }
-}
-
 void osd_t::submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
 {
     if (role_osd == this->osd_num)
@@ -551,13 +481,17 @@ void osd_t::report_pg_state(pg_t & pg)
         pg.history_changed = true;
         pg.target_history.clear();
         pg.all_peers = pg.target_set;
+        std::sort(pg.all_peers.begin(), pg.all_peers.end());
         pg.cur_peers = pg.target_set;
     }
     else if (pg.state == (PG_ACTIVE|PG_LEFT_ON_DEAD))
     {
         // Clear history of active+left_on_dead PGs, but leave dead OSDs in all_peers
-        pg.history_changed = true;
+        if (pg.target_history.size())
-        pg.target_history.clear();
+        {
+            pg.history_changed = true;
+            pg.target_history.clear();
+        }
         std::set<osd_num_t> dead_peers;
         for (auto pg_osd: pg.all_peers)
         {
@@ -574,8 +508,12 @@ void osd_t::report_pg_state(pg_t & pg)
                 dead_peers.insert(pg_osd);
             }
         }
-        pg.all_peers.clear();
+        auto new_all_peers = std::vector<osd_num_t>(dead_peers.begin(), dead_peers.end());
-        pg.all_peers.insert(pg.all_peers.begin(), dead_peers.begin(), dead_peers.end());
+        if (pg.all_peers != new_all_peers)
+        {
+            pg.history_changed = true;
+            pg.all_peers = new_all_peers;
+        }
         pg.cur_peers.clear();
         for (auto pg_osd: pg.target_set)
         {

src/osd_peering_pg.cpp

@@ -86,24 +86,11 @@ void pg_obj_state_check_t::walk()
     }
     if (pg->pg_cursize < pg->pg_size)
     {
-        // Report PG history and activate
+        // Activate as degraded
-        pg->state |= PG_DEGRADED | PG_PEERED;
+        // Current OSD set will be added into target_history on first write
-        std::vector<osd_num_t> history_set;
+        pg->state |= PG_DEGRADED;
-        for (auto peer_osd: pg->cur_set)
-        {
-            if (peer_osd != 0)
-            {
-                history_set.push_back(peer_osd);
-            }
-        }
-        pg->target_history.push_back(history_set);
-        pg->history_changed = true;
-    }
-    else
-    {
-        // Just activate
-        pg->state |= PG_ACTIVE;
     }
+    pg->state |= PG_ACTIVE;
     if (pg->state == PG_ACTIVE && pg->cur_peers.size() < pg->all_peers.size())
     {
         pg->state |= PG_LEFT_ON_DEAD;
@@ -435,21 +422,44 @@ void pg_t::calc_object_states(int log_level)
     std::sort(st.list.begin(), st.list.end());
     // Walk over it and check object states
     st.walk();
-    if (this->state & (PG_DEGRADED|PG_LEFT_ON_DEAD))
+    if (this->state != PG_ACTIVE)
     {
         assert(epoch != (((uint64_t)1 << PG_EPOCH_BITS)-1));
         epoch++;
     }
+    if (log_level > 0)
+    {
+        std::string osd_set_desc;
+        for (auto & osd_num: target_set)
+        {
+            osd_set_desc += (osd_set_desc == "" ? "" : ", ")+std::to_string(osd_num);
+        }
+        printf(
+            "[PG %u/%u] %lu clean objects on target OSD set %s\n",
+            pool_id, pg_num, clean_count, osd_set_desc.c_str()
+        );
+        for (auto & stp: state_dict)
+        {
+            osd_set_desc = "";
+            for (auto & loc: stp.first)
+            {
+                osd_set_desc += (osd_set_desc == "" ? "" : ", ")+
+                    std::to_string(loc.osd_num)+
+                    (st.replicated ? "" : "("+std::to_string(loc.role)+")")+
+                    (loc.outdated ? "(old)" : "");
+            }
+            printf("[PG %u/%u] %lu objects on OSD set %s\n", pool_id, pg_num, stp.second.object_count, osd_set_desc.c_str());
+        }
+    }
 }
 
 void pg_t::print_state()
 {
     printf(
-        "[PG %u/%u] is %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s (%lu objects)\n", pool_id, pg_num,
+        "[PG %u/%u] is %s%s%s%s%s%s%s%s%s%s%s%s%s%s (%lu objects)\n", pool_id, pg_num,
         (state & PG_STARTING) ? "starting" : "",
         (state & PG_OFFLINE) ? "offline" : "",
         (state & PG_PEERING) ? "peering" : "",
-        (state & PG_PEERED) ? "peered" : "",
         (state & PG_INCOMPLETE) ? "incomplete" : "",
         (state & PG_ACTIVE) ? "active" : "",
         (state & PG_REPEERING) ? "repeering" : "",

src/osd_peering_pg_test.cpp

@@ -54,5 +54,6 @@ int main(int argc, char *argv[])
     {
         printf("dev: state=%lx\n", it.second.state);
     }
+    delete pg.peering_state;
     return 0;
 }

src/osd_primary.cpp

@@ -228,7 +228,7 @@ resume_1:
 resume_2:
     if (op_data->errors > 0)
     {
-        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
+        finish_op(cur_op, op_data->errcode);
         return;
     }
     cur_op->reply.rw.version = op_data->fact_ver;
@@ -350,7 +350,7 @@ resume_2:
 resume_3:
     if (op_data->errors > 0)
     {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
         return;
     }
     // Check CAS version
@@ -371,7 +371,7 @@ resume_4:
 resume_5:
     if (op_data->errors > 0)
     {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
         return;
     }
     // Remove version override

src/osd_primary.h

@@ -24,7 +24,7 @@ struct osd_primary_op_data_t
     uint64_t target_ver;
     uint64_t orig_ver = 0, fact_ver = 0;
     uint64_t scheme = 0;
-    int n_subops = 0, done = 0, errors = 0, epipe = 0;
+    int n_subops = 0, done = 0, errors = 0, errcode = 0;
     int degraded = 0, pg_size, pg_data_size;
     osd_rmw_stripe_t *stripes;
     osd_op_t *subops = NULL;

src/osd_primary_chain.cpp

@@ -42,7 +42,7 @@ resume_4:
     {
         free(op_data->chain_reads);
         op_data->chain_reads = NULL;
-        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
+        finish_op(cur_op, op_data->errcode);
         return;
     }
     send_chained_read_results(pg, cur_op);
@@ -297,7 +297,7 @@ int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
                     // Fail it immediately
                     subop->peer_fd = -1;
                     subop->reply.hdr.retval = -EPIPE;
-                    subop->callback(subop);
+                    ringloop->set_immediate([subop]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
                 }
                 subop_idx++;
             }

src/osd_primary_subops.cpp

@@ -53,7 +53,10 @@ void osd_t::finish_op(osd_op_t *cur_op, int retval)
         inode_stats[cur_op->req.rw.inode].op_count[inode_st_op]++;
         inode_stats[cur_op->req.rw.inode].op_sum[inode_st_op] += usec;
         if (cur_op->req.hdr.opcode == OSD_OP_DELETE)
-            inode_stats[cur_op->req.rw.inode].op_bytes[inode_st_op] += cur_op->op_data->pg_data_size * bs_block_size;
+        {
+            if (cur_op->op_data)
+                inode_stats[cur_op->req.rw.inode].op_bytes[inode_st_op] += cur_op->op_data->pg_data_size * bs_block_size;
+        }
         else
             inode_stats[cur_op->req.rw.inode].op_bytes[inode_st_op] += cur_op->req.rw.len;
     }
@@ -122,7 +125,7 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, const ui
         zero_read = -1;
     osd_op_t *subops = new osd_op_t[n_subops];
     op_data->fact_ver = 0;
-    op_data->done = op_data->errors = 0;
+    op_data->done = op_data->errors = op_data->errcode = 0;
     op_data->n_subops = n_subops;
     op_data->subops = subops;
     int sent = submit_primary_subop_batch(submit_type, op_data->oid.inode, op_version, op_data->stripes, osd_set, cur_op, 0, zero_read);
@@ -235,7 +238,7 @@ int osd_t::submit_primary_subop_batch(int submit_type, inode_t inode, uint64_t o
                     // Fail it immediately
                     subop->peer_fd = -1;
                     subop->reply.hdr.retval = -EPIPE;
-                    subop->callback(subop);
+                    ringloop->set_immediate([subop]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
                 }
             }
             i++;
@@ -263,9 +266,11 @@ void osd_t::handle_primary_bs_subop(osd_op_t *subop)
     blockstore_op_t *bs_op = subop->bs_op;
     int expected = bs_op->opcode == BS_OP_READ || bs_op->opcode == BS_OP_WRITE
         || bs_op->opcode == BS_OP_WRITE_STABLE ? bs_op->len : 0;
-    if (bs_op->retval != expected && bs_op->opcode != BS_OP_READ)
+    if (bs_op->retval != expected && bs_op->opcode != BS_OP_READ &&
+        (bs_op->opcode != BS_OP_WRITE && bs_op->opcode != BS_OP_WRITE_STABLE ||
+        bs_op->retval != -ENOSPC))
     {
-        // die
+        // die on any error except ENOSPC
         throw std::runtime_error(
             "local blockstore modification failed (opcode = "+std::to_string(bs_op->opcode)+
             " retval = "+std::to_string(bs_op->retval)+")"
@@ -276,6 +281,8 @@ void osd_t::handle_primary_bs_subop(osd_op_t *subop)
     subop->reply.hdr.retval = bs_op->retval;
     if (bs_op->opcode == BS_OP_READ || bs_op->opcode == BS_OP_WRITE || bs_op->opcode == BS_OP_WRITE_STABLE)
     {
+        subop->req.sec_rw.oid = bs_op->oid;
+        subop->req.sec_rw.version = bs_op->version;
         subop->req.sec_rw.len = bs_op->len;
         subop->reply.sec_rw.version = bs_op->version;
     }
@@ -337,14 +344,17 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
                 osd_op_names[opcode], subop->peer_fd, retval, expected
             );
         }
-        if (retval == -EPIPE)
+        // Error priority: EIO > ENOSPC > EPIPE
+        if (op_data->errcode == 0 || retval == -EIO ||
+            retval == -ENOSPC && op_data->errcode == -EPIPE)
         {
-            op_data->epipe++;
+            op_data->errcode = retval;
         }
         op_data->errors++;
-        if (subop->peer_fd >= 0)
+        if (subop->peer_fd >= 0 && (opcode != OSD_OP_SEC_WRITE && opcode != OSD_OP_SEC_WRITE_STABLE ||
+            retval != -ENOSPC))
         {
-            // Drop connection on any error
+            // Drop connection on any error expect ENOSPC
             msgr.stop_client(subop->peer_fd);
         }
     }
@@ -408,7 +418,8 @@ void osd_t::cancel_primary_write(osd_op_t *cur_op)
         // are sent to peer OSDs, so we can't just throw them away.
         // Mark them with an extra EPIPE.
         cur_op->op_data->errors++;
-        cur_op->op_data->epipe++;
+        if (cur_op->op_data->errcode == 0)
+            cur_op->op_data->errcode = -EPIPE;
         cur_op->op_data->done--; // Caution: `done` must be signed because may become -1 here
     }
     else
@@ -460,7 +471,7 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
 {
     osd_primary_op_data_t *op_data = cur_op->op_data;
     op_data->n_subops = chunks_to_delete_count;
-    op_data->done = op_data->errors = 0;
+    op_data->done = op_data->errors = op_data->errcode = 0;
     if (!op_data->n_subops)
     {
         return;
@@ -512,7 +523,7 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
                 // Fail it immediately
                 subops[i].peer_fd = -1;
                 subops[i].reply.hdr.retval = -EPIPE;
-                subops[i].callback(&subops[i]);
+                ringloop->set_immediate([subop = &subops[i]]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
             }
         }
     }
@@ -523,7 +534,7 @@ int osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
     osd_primary_op_data_t *op_data = cur_op->op_data;
     int n_osds = op_data->dirty_osd_count;
     osd_op_t *subops = new osd_op_t[n_osds];
-    op_data->done = op_data->errors = 0;
+    op_data->done = op_data->errors = op_data->errcode = 0;
     op_data->n_subops = n_osds;
     op_data->subops = subops;
     std::map<uint64_t, int>::iterator peer_it;
@@ -579,7 +590,7 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
     osd_primary_op_data_t *op_data = cur_op->op_data;
     int n_osds = op_data->unstable_write_osds->size();
     osd_op_t *subops = new osd_op_t[n_osds];
-    op_data->done = op_data->errors = 0;
+    op_data->done = op_data->errors = op_data->errcode = 0;
     op_data->n_subops = n_osds;
     op_data->subops = subops;
     for (int i = 0; i < n_osds; i++)
@@ -627,7 +638,7 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
                 // Fail it immediately
                 subops[i].peer_fd = -1;
                 subops[i].reply.hdr.retval = -EPIPE;
-                subops[i].callback(&subops[i]);
+                ringloop->set_immediate([subop = &subops[i]]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
             }
         }
     }

src/osd_primary_sync.cpp

@@ -240,7 +240,7 @@ resume_8:
     }
     if (op_data->errors > 0)
     {
-        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
+        finish_op(cur_op, op_data->errcode);
     }
     else
     {

src/osd_primary_write.cpp

@@ -93,7 +93,7 @@ resume_2:
 resume_3:
     if (op_data->errors > 0)
     {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
         return;
     }
     // Check CAS version
@@ -155,17 +155,36 @@ resume_3:
     if (pg.epoch > pg.reported_epoch)
     {
         // Report newer epoch before writing
-        // FIXME: We may report only one PG state here...
+        // FIXME: We don't have to report all changed PG states here
         this->pg_state_dirty.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
+        if (pg.state != PG_ACTIVE)
+        {
+            // Check that current OSD set is in history and/or add it there
+            std::vector<osd_num_t> history_set;
+            for (auto peer_osd: pg.cur_set)
+                if (peer_osd != 0)
+                    history_set.push_back(peer_osd);
+            std::sort(history_set.begin(), history_set.end());
+            auto it = std::lower_bound(pg.target_history.begin(), pg.target_history.end(), history_set);
+            if (it == pg.target_history.end() || *it != history_set)
+                pg.target_history.insert(it, history_set);
+        }
         pg.history_changed = true;
         report_pg_states();
 resume_10:
         if (pg.epoch > pg.reported_epoch)
         {
-            op_data->st = 10;
+#define PG_EPOCH_WAIT_STATE 10
+            op_data->st = PG_EPOCH_WAIT_STATE;
             return;
         }
     }
+    // Recheck PG state after reporting history - maybe it's already stopping/restarting
+    if (pg.state & (PG_STOPPING|PG_REPEERING))
+    {
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, -EPIPE);
+        return;
+    }
     submit_primary_subops(SUBMIT_WRITE, op_data->target_ver, pg.cur_set.data(), cur_op);
 resume_4:
     op_data->st = 4;
@@ -178,7 +197,7 @@ resume_5:
     }
     if (op_data->errors > 0)
     {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
         return;
     }
     if (op_data->object_state)
@@ -255,7 +274,7 @@ resume_8:
 resume_9:
                 if (op_data->errors > 0)
                 {
-                    pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+                    pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
                     return;
                 }
             }
@@ -287,6 +306,50 @@ continue_others:
     }
 }
 
+void osd_t::on_change_pg_history_hook(pool_id_t pool_id, pg_num_t pg_num)
+{
+    auto pg_it = pgs.find({
+        .pool_id = pool_id,
+        .pg_num = pg_num,
+    });
+    if (pg_it == pgs.end())
+    {
+        return;
+    }
+    auto & pg = pg_it->second;
+    if (pg.epoch > pg.reported_epoch &&
+        st_cli.pool_config[pool_id].pg_config[pg_num].epoch >= pg.epoch)
+    {
+        pg.reported_epoch = st_cli.pool_config[pool_id].pg_config[pg_num].epoch;
+        std::vector<object_id> resume_oids;
+        for (auto & op: pg.write_queue)
+        {
+            if (op.second->op_data->st == PG_EPOCH_WAIT_STATE)
+            {
+                // Run separately to prevent side effects
+                resume_oids.push_back(op.first);
+            }
+        }
+        for (auto & oid: resume_oids)
+        {
+            auto pg_it = pgs.find({
+                .pool_id = pool_id,
+                .pg_num = pg_num,
+            });
+            if (pg_it != pgs.end())
+            {
+                auto & pg = pg_it->second;
+                auto op_it = pg.write_queue.find(oid);
+                if (op_it != pg.write_queue.end() &&
+                    op_it->second->op_data->st == PG_EPOCH_WAIT_STATE)
+                {
+                    continue_primary_write(op_it->second);
+                }
+            }
+        }
+    }
+}
+
 bool osd_t::remember_unstable_write(osd_op_t *cur_op, pg_t & pg, pg_osd_set_t & loc_set, int base_state)
 {
     osd_primary_op_data_t *op_data = cur_op->op_data;
@@ -337,7 +400,7 @@ resume_7:
             op_data->unstable_write_osds = NULL;
             if (op_data->errors > 0)
             {
-                pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+                pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
                 return false;
             }
         }

src/osd_rmw.cpp

@@ -154,6 +154,8 @@ struct reed_sol_matrix_t
     int refs = 0;
     int *je_data;
     uint8_t *isal_data;
+    // 32 bytes = 256/8 = max pg_size/8
+    std::map<std::array<uint8_t, 32>, void*> subdata;
     std::map<reed_sol_erased_t, void*> decodings;
 };
 
@@ -194,6 +196,12 @@ void use_ec(int pg_size, int pg_minsize, bool use)
         free(rs_it->second.je_data);
         if (rs_it->second.isal_data)
             free(rs_it->second.isal_data);
+        for (auto sub_it = rs_it->second.subdata.begin(); sub_it != rs_it->second.subdata.end();)
+        {
+            void *data = sub_it->second;
+            rs_it->second.subdata.erase(sub_it++);
+            free(data);
+        }
         for (auto dec_it = rs_it->second.decodings.begin(); dec_it != rs_it->second.decodings.end();)
         {
             void *data = dec_it->second;
@@ -294,6 +302,47 @@ static void* get_jerasure_decoding_matrix(osd_rmw_stripe_t *stripes, int pg_size
     return dec_it->second;
 }
 
+#ifndef WITH_ISAL
+#define JERASURE_ALIGNMENT 16
+
+// jerasure requires 16-byte alignment for SSE...
+// FIXME: jerasure/gf-complete should probably be patched to automatically choose non-sse version for unaligned buffers
+static void jerasure_matrix_encode_unaligned(int k, int m, int w, int *matrix, char **data_ptrs, char **coding_ptrs, int size)
+{
+    bool unaligned = false;
+    for (int i = 0; i < k; i++)
+        if (((unsigned long)data_ptrs[i]) % JERASURE_ALIGNMENT)
+            unaligned = true;
+    for (int i = 0; i < m; i++)
+        if (((unsigned long)coding_ptrs[i]) % JERASURE_ALIGNMENT)
+            unaligned = true;
+    if (!unaligned)
+    {
+        jerasure_matrix_encode(k, m, w, matrix, data_ptrs, coding_ptrs, size);
+        return;
+    }
+    int aligned_size = ((size+JERASURE_ALIGNMENT-1)/JERASURE_ALIGNMENT)*JERASURE_ALIGNMENT;
+    int copy_size = aligned_size*(k+m);
+    char local_data[copy_size > 4096 ? 0 : copy_size];
+    char *data_copy = copy_size > 4096 || (unsigned long)local_data % JERASURE_ALIGNMENT
+        ? (char*)memalign_or_die(JERASURE_ALIGNMENT, aligned_size*(k+m))
+        : local_data;
+    char *aligned_ptrs[k+m];
+    for (int i = 0; i < k; i++)
+    {
+        memcpy(data_copy + i*aligned_size, data_ptrs[i], size);
+        aligned_ptrs[i] = data_copy + i*aligned_size;
+    }
+    for (int i = 0; i < m; i++)
+        aligned_ptrs[k+i] = data_copy + (k+i)*aligned_size;
+    jerasure_matrix_encode(k, m, w, matrix, aligned_ptrs, aligned_ptrs+k, size);
+    for (int i = 0; i < m; i++)
+        memcpy(coding_ptrs[i], aligned_ptrs[k+i], size);
+    if (copy_size > 4096 || (unsigned long)local_data % JERASURE_ALIGNMENT)
+        free(data_copy);
+}
+#endif
+
 #ifdef WITH_ISAL
 void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize, uint32_t bitmap_size)
 {
@@ -357,10 +406,12 @@ void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsi
     {
         data_ptrs[role] = NULL;
     }
+    bool recovered = false;
     for (int role = 0; role < pg_minsize; role++)
     {
         if (stripes[role].read_end != 0 && stripes[role].missing)
         {
+            recovered = true;
             if (stripes[role].read_end > stripes[role].read_start)
             {
                 for (int other = 0; other < pg_size; other++)
@@ -378,18 +429,64 @@ void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsi
                     data_ptrs, data_ptrs+pg_minsize, stripes[role].read_end - stripes[role].read_start
                 );
             }
-            for (int other = 0; other < pg_size; other++)
+        }
+    }
+    if (recovered && bitmap_size > 0)
+    {
+        bool unaligned = false;
+        for (int role = 0; role < pg_size; role++)
+        {
+            if (stripes[role].read_end != 0)
             {
-                if (stripes[other].read_end != 0 && !stripes[other].missing)
+                data_ptrs[role] = (char*)stripes[role].bmp_buf;
+                if (((unsigned long)stripes[role].bmp_buf) % JERASURE_ALIGNMENT)
+                    unaligned = true;
+            }
+        }
+        if (!unaligned)
+        {
+            for (int role = 0; role < pg_minsize; role++)
+            {
+                if (stripes[role].read_end != 0 && stripes[role].missing)
                 {
-                    data_ptrs[other] = (char*)(stripes[other].bmp_buf);
+                    jerasure_matrix_dotprod(
+                        pg_minsize, OSD_JERASURE_W, decoding_matrix+(role*pg_minsize), dm_ids, role,
+                        data_ptrs, data_ptrs+pg_minsize, bitmap_size
+                    );
                 }
             }
-            data_ptrs[role] = (char*)stripes[role].bmp_buf;
+        }
-            jerasure_matrix_dotprod(
+        else
-                pg_minsize, OSD_JERASURE_W, decoding_matrix+(role*pg_minsize), dm_ids, role,
+        {
-                data_ptrs, data_ptrs+pg_minsize, bitmap_size
+            // jerasure_matrix_dotprod requires 16-byte alignment for SSE...
-            );
+            int aligned_size = ((bitmap_size+JERASURE_ALIGNMENT-1)/JERASURE_ALIGNMENT)*JERASURE_ALIGNMENT;
+            int copy_size = aligned_size*pg_size;
+            char local_data[copy_size > 4096 ? 0 : copy_size];
+            bool alloc_copy = copy_size > 4096 || (unsigned long)local_data % JERASURE_ALIGNMENT;
+            char *data_copy = alloc_copy
+                ? (char*)memalign_or_die(JERASURE_ALIGNMENT, copy_size)
+                : local_data;
+            for (int role = 0; role < pg_size; role++)
+            {
+                if (stripes[role].read_end != 0)
+                {
+                    data_ptrs[role] = data_copy + role*aligned_size;
+                    memcpy(data_ptrs[role], stripes[role].bmp_buf, bitmap_size);
+                }
+            }
+            for (int role = 0; role < pg_size; role++)
+            {
+                if (stripes[role].read_end != 0 && stripes[role].missing)
+                {
+                    jerasure_matrix_dotprod(
+                        pg_minsize, OSD_JERASURE_W, decoding_matrix+(role*pg_minsize), dm_ids, role,
+                        data_ptrs, data_ptrs+pg_minsize, bitmap_size
+                    );
+                    memcpy(stripes[role].bmp_buf, data_ptrs[role], bitmap_size);
+                }
+            }
+            if (alloc_copy)
+                free(data_copy);
         }
     }
 }
@@ -662,7 +759,18 @@ static void calc_rmw_parity_copy_mod(osd_rmw_stripe_t *stripes, int pg_size, int
     uint64_t *read_osd_set, uint64_t *write_osd_set, uint32_t chunk_size, uint32_t bitmap_granularity,
     uint32_t &start, uint32_t &end)
 {
-    if (write_osd_set[pg_minsize] != 0 || write_osd_set != read_osd_set)
+    bool required = false;
+    for (int role = pg_minsize; role < pg_size; role++)
+    {
+        if (write_osd_set[role] != 0)
+        {
+            // Whole parity chunk is needed when we move the object
+            if (write_osd_set[role] != read_osd_set[role])
+                end = chunk_size;
+            required = true;
+        }
+    }
+    if (required && end != chunk_size)
     {
         // start & end are required for calc_rmw_parity
         for (int role = 0; role < pg_minsize; role++)
@@ -673,14 +781,6 @@ static void calc_rmw_parity_copy_mod(osd_rmw_stripe_t *stripes, int pg_size, int
                 end = std::max(stripes[role].req_end, end);
             }
         }
-        for (int role = pg_minsize; role < pg_size; role++)
-        {
-            if (write_osd_set[role] != 0 && write_osd_set[role] != read_osd_set[role])
-            {
-                start = 0;
-                end = chunk_size;
-            }
-        }
     }
     // Set bitmap bits accordingly
     if (bitmap_granularity > 0)
@@ -808,11 +908,56 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
     if (end != 0)
     {
         int write_parity = 0;
-        for (int i = pg_minsize; i < pg_size; i++)
+        bool is_seq = true;
+        for (int i = pg_size-1; i >= pg_minsize; i--)
+        {
             if (write_osd_set[i] != 0)
                 write_parity++;
+            else if (write_parity != 0)
+                is_seq = false;
+        }
         if (write_parity > 0)
         {
+            // First get the coding matrix or sub-matrix
+            void *matrix_data =
+#ifdef WITH_ISAL
+                matrix->isal_data;
+#else
+                matrix->je_data;
+#endif
+            if (!is_seq)
+            {
+                // We need a coding sub-matrix
+                std::array<uint8_t, 32> missing_parity = {};
+                for (int i = pg_minsize; i < pg_size; i++)
+                {
+                    if (!write_osd_set[i])
+                        missing_parity[(i-pg_minsize) >> 3] |= (1 << ((i-pg_minsize) & 0x7));
+                }
+                auto sub_it = matrix->subdata.find(missing_parity);
+                if (sub_it == matrix->subdata.end())
+                {
+                    int item_size =
+#ifdef WITH_ISAL
+                        32;
+#else
+                        sizeof(int);
+#endif
+                    void *subm = malloc_or_die(item_size * write_parity * pg_minsize);
+                    for (int i = pg_minsize, j = 0; i < pg_size; i++)
+                    {
+                        if (write_osd_set[i])
+                        {
+                            memcpy((uint8_t*)subm + item_size*pg_minsize*j, (uint8_t*)matrix_data + item_size*pg_minsize*(i-pg_minsize), item_size*pg_minsize);
+                            j++;
+                        }
+                    }
+                    matrix->subdata[missing_parity] = subm;
+                    matrix_data = subm;
+                }
+                else
+                    matrix_data = sub_it->second;
+            }
             // Calculate new coding chunks
             buf_len_t bufs[pg_size][3];
             int nbuf[pg_size], curbuf[pg_size];
@@ -841,13 +986,13 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
             while (pos < end)
             {
                 uint32_t next_end = end;
-                for (int i = 0; i < pg_size; i++)
+                for (int i = 0, j = 0; i < pg_size; i++)
                 {
                     if (i < pg_minsize || write_osd_set[i] != 0)
                     {
                         assert(curbuf[i] < nbuf[i]);
                         assert(bufs[i][curbuf[i]].buf);
-                        data_ptrs[i] = (uint8_t*)bufs[i][curbuf[i]].buf + pos-positions[i];
+                        data_ptrs[j++] = (uint8_t*)bufs[i][curbuf[i]].buf + pos-positions[i];
                         uint32_t this_end = bufs[i][curbuf[i]].len + positions[i];
                         if (next_end > this_end)
                             next_end = this_end;
@@ -868,32 +1013,30 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
                 }
 #ifdef WITH_ISAL
                 ec_encode_data(
-                    next_end-pos, pg_minsize, write_parity, matrix->isal_data,
+                    next_end-pos, pg_minsize, write_parity, (uint8_t*)matrix_data,
                     (uint8_t**)data_ptrs, (uint8_t**)data_ptrs+pg_minsize
                 );
 #else
                 jerasure_matrix_encode(
-                    pg_minsize, write_parity, OSD_JERASURE_W, matrix->je_data,
+                    pg_minsize, write_parity, OSD_JERASURE_W, (int*)matrix_data,
                     (char**)data_ptrs, (char**)data_ptrs+pg_minsize, next_end-pos
                 );
 #endif
                 pos = next_end;
             }
-            for (int i = 0; i < pg_size; i++)
+            for (int i = 0, j = 0; i < pg_size; i++)
             {
                 if (i < pg_minsize || write_osd_set[i] != 0)
-                {
+                    data_ptrs[j++] = stripes[i].bmp_buf;
-                    data_ptrs[i] = stripes[i].bmp_buf;
-                }
             }
 #ifdef WITH_ISAL
             ec_encode_data(
-                bitmap_size, pg_minsize, write_parity, matrix->isal_data,
+                bitmap_size, pg_minsize, write_parity, (uint8_t*)matrix_data,
                 (uint8_t**)data_ptrs, (uint8_t**)data_ptrs+pg_minsize
             );
 #else
-            jerasure_matrix_encode(
+            jerasure_matrix_encode_unaligned(
-                pg_minsize, write_parity, OSD_JERASURE_W, matrix->je_data,
+                pg_minsize, write_parity, OSD_JERASURE_W, (int*)matrix_data,
                 (char**)data_ptrs, (char**)data_ptrs+pg_minsize, bitmap_size
             );
 #endif

src/osd_rmw_test.cpp

@@ -3,6 +3,10 @@
 
 #define RMW_DEBUG
 
+#ifdef NO_ISAL
+#undef WITH_ISAL
+#endif
+
 #include <string.h>
 #include "osd_rmw.cpp"
 #include "test_pattern.h"
@@ -20,7 +24,8 @@ void test11();
 void test12();
 void test13();
 void test14();
-void test15();
+void test15(bool second);
+void test16();
 
 int main(int narg, char *args[])
 {
@@ -49,7 +54,10 @@ int main(int narg, char *args[])
     // Test 14
     test14();
     // Test 15
-    test15();
+    test15(false);
+    test15(true);
+    // Test 16
+    test16();
     // End
     printf("all ok\n");
     return 0;
@@ -819,12 +827,11 @@ void test14()
 
 ***/
 
-void test15()
+void test15(bool second)
 {
     const int bmp = 64*1024 / 4096 / 8;
     use_ec(4, 2, true);
-    osd_num_t osd_set[4] = { 1, 2, 3, 0 };
+    osd_num_t osd_set[4] = { 1, 2, (osd_num_t)(second ? 0 : 3), (osd_num_t)(second ? 4 : 0) };
-    osd_num_t write_osd_set[4] = { 1, 2, 3, 0 };
     osd_rmw_stripe_t stripes[4] = {};
     unsigned bitmaps[4] = { 0 };
     // Test 15.0
@@ -835,7 +842,7 @@ void test15()
     assert(stripes[2].req_start == 0 && stripes[2].req_end == 0);
     assert(stripes[3].req_start == 0 && stripes[3].req_end == 0);
     // Test 15.1
-    void *rmw_buf = calc_rmw(write_buf, stripes, osd_set, 4, 2, 3, write_osd_set, 64*1024, bmp);
+    void *rmw_buf = calc_rmw(write_buf, stripes, osd_set, 4, 2, 3, osd_set, 64*1024, bmp);
     for (int i = 0; i < 4; i++)
         stripes[i].bmp_buf = bitmaps+i;
     assert(rmw_buf);
@@ -845,34 +852,139 @@ void test15()
     assert(stripes[3].read_start == 0 && stripes[3].read_end == 0);
     assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
     assert(stripes[1].write_start == 28*1024 && stripes[1].write_end == 32*1024);
-    assert(stripes[2].write_start == 28*1024 && stripes[2].write_end == 32*1024);
+    assert(stripes[2+second].write_start == 28*1024 && stripes[2+second].write_end == 32*1024);
-    assert(stripes[3].write_start == 0 && stripes[3].write_end == 0);
+    assert(stripes[3-second].write_start == 0 && stripes[3-second].write_end == 0);
     assert(stripes[0].read_buf == (uint8_t*)rmw_buf+4*1024);
     assert(stripes[1].read_buf == NULL);
     assert(stripes[2].read_buf == NULL);
     assert(stripes[3].read_buf == NULL);
     assert(stripes[0].write_buf == NULL);
     assert(stripes[1].write_buf == (uint8_t*)write_buf);
-    assert(stripes[2].write_buf == rmw_buf);
+    assert(stripes[2+second].write_buf == rmw_buf);
-    assert(stripes[3].write_buf == NULL);
+    assert(stripes[3-second].write_buf == NULL);
     // Test 15.2 - encode
     set_pattern(write_buf, 4*1024, PATTERN1);
     set_pattern(stripes[0].read_buf, 4*1024, PATTERN2);
     memset(stripes[0].bmp_buf, 0, bmp);
     memset(stripes[1].bmp_buf, 0, bmp);
-    calc_rmw_parity_ec(stripes, 4, 2, osd_set, write_osd_set, 64*1024, bmp);
+    memset(stripes[2+second].write_buf, 0, 4096);
-    assert(*(uint32_t*)stripes[2].bmp_buf == 0x80);
+    calc_rmw_parity_ec(stripes, 4, 2, osd_set, osd_set, 64*1024, bmp);
+    assert(second || *(uint32_t*)stripes[2].bmp_buf == 0x80);
     assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
     assert(stripes[1].write_start == 28*1024 && stripes[1].write_end == 32*1024);
-    assert(stripes[2].write_start == 28*1024 && stripes[2].write_end == 32*1024);
+    assert(stripes[2+second].write_start == 28*1024 && stripes[2+second].write_end == 32*1024);
-    assert(stripes[3].write_start == 0 && stripes[3].write_end == 0);
+    assert(stripes[3-second].write_start == 0 && stripes[3-second].write_end == 0);
     assert(stripes[0].write_buf == NULL);
     assert(stripes[1].write_buf == (uint8_t*)write_buf);
-    assert(stripes[2].write_buf == rmw_buf);
+    assert(stripes[2+second].write_buf == rmw_buf);
-    assert(stripes[3].write_buf == NULL);
+    assert(stripes[3-second].write_buf == NULL);
-    check_pattern(stripes[2].write_buf, 4*1024, PATTERN1^PATTERN2); // first parity is always xor :)
+    // first parity is always xor :), second isn't...
+    check_pattern(stripes[2+second].write_buf, 4*1024, second ? 0xb79a59a0ce8b9b81 : PATTERN1^PATTERN2);
     // Done
     free(rmw_buf);
     free(write_buf);
-    use_ec(3, 2, false);
+    use_ec(4, 2, false);
+}
+
+/***
+
+16. EC 2+2 write one parity block with another missing
+   calc_rmw(offset=0, len=0, osd_set=[1,2,0,0], write_set=[1,2,0,3])
+   = {
+     read: [ [ 0, 128K ], [ 0, 128K ], [ 0, 0 ], [ 0, 0 ] ],
+     write: [ [ 0, 0 ], [ 0, 0 ], [ 0, 0 ], [ 0, 128K ] ],
+     input buffer: [],
+     rmw buffer: [ write3, read0, read1 ],
+   }
+
+***/
+
+void test16()
+{
+    const int bmp = 128*1024 / 4096 / 8;
+    use_ec(4, 2, true);
+    osd_num_t osd_set[4] = { 1, 2, 0, 0 };
+    osd_num_t write_osd_set[4] = { 1, 2, 0, 3 };
+    osd_rmw_stripe_t stripes[4] = {};
+    unsigned bitmaps[4] = { 0 };
+    // Test 16.0
+    void *write_buf = NULL;
+    split_stripes(2, 128*1024, 0, 0, stripes);
+    assert(stripes[0].req_start == 0 && stripes[0].req_end == 0);
+    assert(stripes[1].req_start == 0 && stripes[1].req_end == 0);
+    assert(stripes[2].req_start == 0 && stripes[2].req_end == 0);
+    assert(stripes[3].req_start == 0 && stripes[3].req_end == 0);
+    // Test 16.1
+    void *rmw_buf = calc_rmw(write_buf, stripes, osd_set, 4, 2, 3, write_osd_set, 128*1024, bmp);
+    for (int i = 0; i < 4; i++)
+        stripes[i].bmp_buf = bitmaps+i;
+    assert(rmw_buf);
+    assert(stripes[0].read_start == 0 && stripes[0].read_end == 128*1024);
+    assert(stripes[1].read_start == 0 && stripes[1].read_end == 128*1024);
+    assert(stripes[2].read_start == 0 && stripes[2].read_end == 0);
+    assert(stripes[3].read_start == 0 && stripes[3].read_end == 0);
+    assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
+    assert(stripes[1].write_start == 0 && stripes[1].write_end == 0);
+    assert(stripes[2].write_start == 0 && stripes[2].write_end == 0);
+    assert(stripes[3].write_start == 0 && stripes[3].write_end == 128*1024);
+    assert(stripes[0].read_buf == (uint8_t*)rmw_buf+128*1024);
+    assert(stripes[1].read_buf == (uint8_t*)rmw_buf+256*1024);
+    assert(stripes[2].read_buf == NULL);
+    assert(stripes[3].read_buf == NULL);
+    assert(stripes[0].write_buf == NULL);
+    assert(stripes[1].write_buf == NULL);
+    assert(stripes[2].write_buf == NULL);
+    assert(stripes[3].write_buf == rmw_buf);
+    // Test 16.2 - encode
+    set_pattern(stripes[0].read_buf, 128*1024, PATTERN1);
+    set_pattern(stripes[1].read_buf, 128*1024, PATTERN2);
+    memset(stripes[0].bmp_buf, 0xff, bmp);
+    memset(stripes[1].bmp_buf, 0xff, bmp);
+    calc_rmw_parity_ec(stripes, 4, 2, osd_set, write_osd_set, 128*1024, bmp);
+    assert(*(uint32_t*)stripes[2].bmp_buf == 0);
+    assert(*(uint32_t*)stripes[3].bmp_buf == 0xF1F1F1F1);
+    assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
+    assert(stripes[1].write_start == 0 && stripes[1].write_end == 0);
+    assert(stripes[2].write_start == 0 && stripes[2].write_end == 0);
+    assert(stripes[3].write_start == 0 && stripes[3].write_end == 128*1024);
+    assert(stripes[0].write_buf == NULL);
+    assert(stripes[1].write_buf == NULL);
+    assert(stripes[2].write_buf == NULL);
+    assert(stripes[3].write_buf == rmw_buf);
+    check_pattern(stripes[3].write_buf, 128*1024, 0x7eb9ae9cd8e652c3); // 2nd EC chunk
+    // Test 16.3 - decode and verify
+    osd_num_t read_osd_set[4] = { 0, 2, 0, 3 };
+    memset(stripes, 0, sizeof(stripes));
+    split_stripes(2, 128*1024, 0, 256*1024, stripes);
+    assert(stripes[0].req_start == 0 && stripes[0].req_end == 128*1024);
+    assert(stripes[1].req_start == 0 && stripes[1].req_end == 128*1024);
+    assert(stripes[2].req_start == 0 && stripes[2].req_end == 0);
+    assert(stripes[3].req_start == 0 && stripes[3].req_end == 0);
+    for (int role = 0; role < 4; role++)
+    {
+        stripes[role].read_start = stripes[role].req_start;
+        stripes[role].read_end = stripes[role].req_end;
+    }
+    assert(extend_missing_stripes(stripes, read_osd_set, 2, 4) == 0);
+    assert(stripes[0].read_start == 0 && stripes[0].read_end == 128*1024);
+    assert(stripes[1].read_start == 0 && stripes[1].read_end == 128*1024);
+    assert(stripes[2].read_start == 0 && stripes[2].read_end == 0);
+    assert(stripes[3].read_start == 0 && stripes[3].read_end == 128*1024);
+    void *read_buf = alloc_read_buffer(stripes, 4, 0);
+    for (int i = 0; i < 4; i++)
+        stripes[i].bmp_buf = bitmaps+i;
+    assert(read_buf);
+    assert(stripes[0].read_buf == read_buf);
+    assert(stripes[1].read_buf == (uint8_t*)read_buf+128*1024);
+    assert(stripes[3].read_buf == (uint8_t*)read_buf+2*128*1024);
+    set_pattern(stripes[1].read_buf, 128*1024, PATTERN2);
+    memcpy(stripes[3].read_buf, rmw_buf, 128*1024);
+    reconstruct_stripes_ec(stripes, 4, 2, bmp);
+    assert(bitmaps[0] == 0xFFFFFFFF);
+    check_pattern(stripes[0].read_buf, 128*1024, PATTERN1);
+    free(read_buf);
+    // Done
+    free(rmw_buf);
+    free(write_buf);
+    use_ec(4, 2, false);
 }

src/osd_test.cpp

@@ -150,6 +150,7 @@ int connect_osd(const char *osd_address, int osd_port)
     if (connect(connect_fd, (sockaddr*)&addr, sizeof(addr)) < 0)
     {
         perror("connect");
+        close(connect_fd);
         return -1;
     }
     int one = 1;

src/pg_states.cpp

@@ -3,12 +3,11 @@
 
 #include "pg_states.h"
 
-const int pg_state_bit_count = 16;
+const int pg_state_bit_count = 14;
 
-const int pg_state_bits[16] = {
+const int pg_state_bits[14] = {
     PG_STARTING,
     PG_PEERING,
-    PG_PEERED,
     PG_INCOMPLETE,
     PG_ACTIVE,
     PG_REPEERING,
@@ -23,10 +22,9 @@ const int pg_state_bits[16] = {
     PG_LEFT_ON_DEAD,
 };
 
-const char *pg_state_names[16] = {
+const char *pg_state_names[14] = {
     "starting",
     "peering",
-    "peered",
     "incomplete",
     "active",
     "repeering",

src/pg_states.h

@@ -4,27 +4,25 @@
 #pragma once
 
 // Placement group states
-// STARTING -> [acquire lock] -> PEERING -> PEERED
+// STARTING -> [acquire lock] -> PEERING -> INCOMPLETE|ACTIVE
-// PEERED -> [report history if required!] -> INCOMPLETE|ACTIVE
 // ACTIVE -> REPEERING -> PEERING
 // ACTIVE -> STOPPING -> OFFLINE -> [release lock]
 // Exactly one of these:
 #define PG_STARTING (1<<0)
 #define PG_PEERING (1<<1)
-#define PG_PEERED (1<<2)
+#define PG_INCOMPLETE (1<<2)
-#define PG_INCOMPLETE (1<<3)
+#define PG_ACTIVE (1<<3)
-#define PG_ACTIVE (1<<4)
+#define PG_REPEERING (1<<4)
-#define PG_REPEERING (1<<5)
+#define PG_STOPPING (1<<5)
-#define PG_STOPPING (1<<6)
+#define PG_OFFLINE (1<<6)
-#define PG_OFFLINE (1<<7)
 // Plus any of these:
-#define PG_DEGRADED (1<<8)
+#define PG_DEGRADED (1<<7)
-#define PG_HAS_INCOMPLETE (1<<9)
+#define PG_HAS_INCOMPLETE (1<<8)
-#define PG_HAS_DEGRADED (1<<10)
+#define PG_HAS_DEGRADED (1<<9)
-#define PG_HAS_MISPLACED (1<<11)
+#define PG_HAS_MISPLACED (1<<10)
-#define PG_HAS_UNCLEAN (1<<12)
+#define PG_HAS_UNCLEAN (1<<11)
-#define PG_HAS_INVALID (1<<13)
+#define PG_HAS_INVALID (1<<12)
-#define PG_LEFT_ON_DEAD (1<<14)
+#define PG_LEFT_ON_DEAD (1<<13)
 
 // Lower bits that represent object role (EC 0/1/2... or always 0 with replication)
 // 12 bits is a safe default that doesn't depend on pg_stripe_size or pg_block_size

src/qemu_driver.c

@@ -53,6 +53,7 @@ typedef struct VitastorClient
     char *etcd_host;
     char *etcd_prefix;
     char *image;
+    int skip_parents;
     uint64_t inode;
     uint64_t pool;
     uint64_t size;
@@ -63,6 +64,10 @@ typedef struct VitastorClient
     int rdma_gid_index;
     int rdma_mtu;
     QemuMutex mutex;
+
+    uint64_t last_bitmap_inode, last_bitmap_offset, last_bitmap_len;
+    uint32_t last_bitmap_granularity;
+    uint8_t *last_bitmap;
 } VitastorClient;
 
 typedef struct VitastorRPC
@@ -72,6 +77,9 @@ typedef struct VitastorRPC
     QEMUIOVector *iov;
     long ret;
     int complete;
+    uint64_t inode, offset, len;
+    uint32_t bitmap_granularity;
+    uint8_t *bitmap;
 } VitastorRPC;
 
 static void vitastor_co_init_task(BlockDriverState *bs, VitastorRPC *task);
@@ -147,6 +155,7 @@ static void vitastor_parse_filename(const char *filename, QDict *options, Error
         if (!strcmp(name, "inode") ||
             !strcmp(name, "pool") ||
             !strcmp(name, "size") ||
+            !strcmp(name, "skip-parents") ||
             !strcmp(name, "use-rdma") ||
             !strcmp(name, "rdma-port_num") ||
             !strcmp(name, "rdma-gid-index") ||
@@ -227,13 +236,16 @@ static void vitastor_aio_set_fd_handler(void *ctx, int fd, int unused1, IOHandle
 
 static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, Error **errp)
 {
+    VitastorRPC task;
     VitastorClient *client = bs->opaque;
+    void *image = NULL;
     int64_t ret = 0;
     qemu_mutex_init(&client->mutex);
     client->config_path = g_strdup(qdict_get_try_str(options, "config-path"));
     // FIXME: Rename to etcd_address
     client->etcd_host = g_strdup(qdict_get_try_str(options, "etcd-host"));
     client->etcd_prefix = g_strdup(qdict_get_try_str(options, "etcd-prefix"));
+    client->skip_parents = qdict_get_try_int(options, "skip-parents", 0);
     client->use_rdma = qdict_get_try_int(options, "use-rdma", -1);
     client->rdma_device = g_strdup(qdict_get_try_str(options, "rdma-device"));
     client->rdma_port_num = qdict_get_try_int(options, "rdma-port-num", 0);
@@ -243,23 +255,25 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
         vitastor_aio_set_fd_handler, bdrv_get_aio_context(bs), client->config_path, client->etcd_host, client->etcd_prefix,
         client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
     );
-    client->image = g_strdup(qdict_get_try_str(options, "image"));
+    image = client->image = g_strdup(qdict_get_try_str(options, "image"));
     client->readonly = (flags & BDRV_O_RDWR) ? 1 : 0;
+    // Get image metadata (size and readonly flag) or just wait until the client is ready
+    if (!image)
+        client->image = (char*)"x";
+    task.complete = 0;
+    task.bs = bs;
+    if (qemu_in_coroutine())
+    {
+        vitastor_co_get_metadata(&task);
+    }
+    else
+    {
+        bdrv_coroutine_enter(bs, qemu_coroutine_create((void(*)(void*))vitastor_co_get_metadata, &task));
+        BDRV_POLL_WHILE(bs, !task.complete);
+    }
+    client->image = image;
     if (client->image)
     {
-        // Get image metadata (size and readonly flag)
-        VitastorRPC task;
-        task.complete = 0;
-        task.bs = bs;
-        if (qemu_in_coroutine())
-        {
-            vitastor_co_get_metadata(&task);
-        }
-        else
-        {
-            bdrv_coroutine_enter(bs, qemu_coroutine_create((void(*)(void*))vitastor_co_get_metadata, &task));
-            BDRV_POLL_WHILE(bs, !task.complete);
-        }
         client->watch = (void*)task.ret;
         client->readonly = client->readonly || vitastor_c_inode_get_readonly(client->watch);
         client->size = vitastor_c_inode_get_size(client->watch);
@@ -284,6 +298,7 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
             client->inode = (client->inode & (((uint64_t)1 << (64-POOL_ID_BITS)) - 1)) | (client->pool << (64-POOL_ID_BITS));
         }
         client->size = qdict_get_try_int(options, "size", 0);
+        vitastor_c_close_watch(client->proxy, (void*)task.ret);
     }
     if (!client->size)
     {
@@ -305,6 +320,7 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
     qdict_del(options, "inode");
     qdict_del(options, "pool");
     qdict_del(options, "size");
+    qdict_del(options, "skip-parents");
     return ret;
 }
 
@@ -321,6 +337,8 @@ static void vitastor_close(BlockDriverState *bs)
         g_free(client->etcd_prefix);
     if (client->image)
         g_free(client->image);
+    free(client->last_bitmap);
+    client->last_bitmap = NULL;
 }
 
 #if QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR > 2
@@ -486,6 +504,13 @@ static int coroutine_fn vitastor_co_pwritev(BlockDriverState *bs,
     vitastor_co_init_task(bs, &task);
     task.iov = iov;
 
+    if (client->last_bitmap)
+    {
+        // Invalidate last bitmap on write
+        free(client->last_bitmap);
+        client->last_bitmap = NULL;
+    }
+
     uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;
     qemu_mutex_lock(&client->mutex);
     vitastor_c_write(client->proxy, inode, offset, bytes, 0, iov->iov, iov->niov, vitastor_co_generic_bh_cb, &task);
@@ -499,6 +524,140 @@ static int coroutine_fn vitastor_co_pwritev(BlockDriverState *bs,
     return task.ret;
 }
 
+#if defined VITASTOR_C_API_VERSION && VITASTOR_C_API_VERSION >= 1
+#if QEMU_VERSION_MAJOR >= 2 || QEMU_VERSION_MAJOR == 1 && QEMU_VERSION_MINOR >= 7
+static void vitastor_co_read_bitmap_cb(void *opaque, long retval, uint8_t *bitmap)
+{
+    VitastorRPC *task = opaque;
+    VitastorClient *client = task->bs->opaque;
+    task->ret = retval;
+    task->complete = 1;
+    if (retval >= 0)
+    {
+        task->bitmap = bitmap;
+        if (client->last_bitmap_inode == task->inode &&
+            client->last_bitmap_offset == task->offset &&
+            client->last_bitmap_len == task->len)
+        {
+            free(client->last_bitmap);
+            client->last_bitmap = bitmap;
+        }
+    }
+    if (qemu_coroutine_self() != task->co)
+    {
+#if QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR > 8
+        aio_co_wake(task->co);
+#else
+        qemu_coroutine_enter(task->co, NULL);
+        qemu_aio_release(task);
+#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)
+{
+    // 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`
+    VitastorRPC task;
+    VitastorClient *client = bs->opaque;
+    uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;
+    uint8_t bit = 0;
+    if (client->last_bitmap && client->last_bitmap_inode == inode &&
+        client->last_bitmap_offset <= offset &&
+        client->last_bitmap_offset+client->last_bitmap_len >= (want_zero ? offset+1 : offset+bytes))
+    {
+        // Use the previously read bitmap
+        task.bitmap_granularity = client->last_bitmap_granularity;
+        task.offset = client->last_bitmap_offset;
+        task.len = client->last_bitmap_len;
+        task.bitmap = client->last_bitmap;
+    }
+    else
+    {
+        // Read bitmap from this position, rounding to full inode PG blocks
+        uint32_t block_size = vitastor_c_inode_get_block_size(client->proxy, inode);
+        if (!block_size)
+            return -EAGAIN;
+        // Init coroutine
+        vitastor_co_init_task(bs, &task);
+        free(client->last_bitmap);
+        task.inode = client->last_bitmap_inode = inode;
+        task.bitmap_granularity = client->last_bitmap_granularity = vitastor_c_inode_get_bitmap_granularity(client->proxy, inode);
+        task.offset = client->last_bitmap_offset = offset / block_size * block_size;
+        task.len = client->last_bitmap_len = (offset+bytes+block_size-1) / block_size * block_size - task.offset;
+        task.bitmap = client->last_bitmap = NULL;
+        qemu_mutex_lock(&client->mutex);
+        vitastor_c_read_bitmap(client->proxy, task.inode, task.offset, task.len, !client->skip_parents, vitastor_co_read_bitmap_cb, &task);
+        qemu_mutex_unlock(&client->mutex);
+        while (!task.complete)
+        {
+            qemu_coroutine_yield();
+        }
+        if (task.ret < 0)
+        {
+            // Error
+            return task.ret;
+        }
+    }
+    if (want_zero)
+    {
+        // Get precise mapping with all holes
+        uint64_t bmp_pos = (offset-task.offset) / task.bitmap_granularity;
+        uint64_t bmp_len = task.len / task.bitmap_granularity;
+        uint64_t bmp_end = bmp_pos+1;
+        bit = (task.bitmap[bmp_pos >> 3] >> (bmp_pos & 0x7)) & 1;
+        while (bmp_end < bmp_len && ((task.bitmap[bmp_end >> 3] >> (bmp_end & 0x7)) & 1) == bit)
+        {
+            bmp_end++;
+        }
+        *pnum = (bmp_end-bmp_pos) * task.bitmap_granularity;
+    }
+    else
+    {
+        // Get larger allocated extents, possibly with false positives
+        uint64_t bmp_pos = (offset-task.offset) / task.bitmap_granularity;
+        uint64_t bmp_end = (offset+bytes-task.offset) / task.bitmap_granularity - bmp_pos;
+        while (bmp_pos < bmp_end)
+        {
+            if (!(bmp_pos & 7) && bmp_end >= bmp_pos+8)
+            {
+                bit = bit || task.bitmap[bmp_pos >> 3];
+                bmp_pos += 8;
+            }
+            else
+            {
+                bit = bit || ((task.bitmap[bmp_pos >> 3] >> (bmp_pos & 0x7)) & 1);
+                bmp_pos++;
+            }
+        }
+        *pnum = bytes;
+    }
+    if (bit)
+    {
+        *map = offset;
+        *file = bs;
+    }
+    return (bit ? (BDRV_BLOCK_DATA|BDRV_BLOCK_OFFSET_VALID) : 0);
+}
+#endif
+#if QEMU_VERSION_MAJOR == 1 && QEMU_VERSION_MINOR >= 7 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR < 12
+// QEMU 1.7-2.11
+static int64_t coroutine_fn vitastor_co_get_block_status(BlockDriverState *bs,
+    int64_t sector_num, int nb_sectors, int *pnum, BlockDriverState **file)
+{
+    int64_t map = 0;
+    int64_t pnumbytes = 0;
+    int r = vitastor_co_block_status(bs, 1, sector_num*BDRV_SECTOR_SIZE, nb_sectors*BDRV_SECTOR_SIZE, &pnumbytes, &map, &file);
+    *pnum = pnumbytes/BDRV_SECTOR_SIZE;
+    return r;
+}
+#endif
+#endif
+
 #if !( QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 7 )
 static int coroutine_fn vitastor_co_readv(BlockDriverState *bs, int64_t sector_num, int nb_sectors, QEMUIOVector *iov)
 {
@@ -606,6 +765,15 @@ static BlockDriver bdrv_vitastor = {
     .bdrv_co_truncate               = vitastor_co_truncate,
 #endif
 
+#if defined VITASTOR_C_API_VERSION && VITASTOR_C_API_VERSION >= 1
+#if QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 12
+    // For snapshot export
+    .bdrv_co_block_status           = vitastor_co_block_status,
+#elif QEMU_VERSION_MAJOR == 1 && QEMU_VERSION_MINOR >= 7 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR < 12
+    .bdrv_co_get_block_status       = vitastor_co_get_block_status,
+#endif
+#endif
+
 #if QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 7
     .bdrv_co_preadv                 = vitastor_co_preadv,
     .bdrv_co_pwritev                = vitastor_co_pwritev,

src/ringloop.cpp

@@ -25,7 +25,6 @@ ring_loop_t::ring_loop_t(int qd)
     {
         free_ring_data[i] = i;
     }
-    wait_sqe_id = 1;
 }
 
 ring_loop_t::~ring_loop_t()
@@ -83,17 +82,19 @@ void ring_loop_t::loop()
         }
         io_uring_cqe_seen(&ring, cqe);
     }
-    while (get_sqe_queue.size() > 0)
-    {
-        (get_sqe_queue[0].second)();
-        get_sqe_queue.erase(get_sqe_queue.begin());
-    }
     do
     {
         loop_again = false;
         for (int i = 0; i < consumers.size(); i++)
         {
             consumers[i]->loop();
+            if (immediate_queue.size())
+            {
+                immediate_queue2.swap(immediate_queue);
+                for (auto & cb: immediate_queue2)
+                    cb();
+                immediate_queue2.clear();
+            }
         }
     } while (loop_again);
 }

src/ringloop.h

@@ -119,11 +119,10 @@ struct ring_consumer_t
 
 class ring_loop_t
 {
-    std::vector<std::pair<int,std::function<void()>>> get_sqe_queue;
+    std::vector<std::function<void()>> immediate_queue, immediate_queue2;
     std::vector<ring_consumer_t*> consumers;
     struct ring_data_t *ring_datas;
     int *free_ring_data;
-    int wait_sqe_id;
     unsigned free_ring_data_ptr;
     bool loop_again;
     struct io_uring ring;
@@ -145,20 +144,9 @@ public:
         }
         return sqe;
     }
-    inline int wait_sqe(std::function<void()> cb)
+    inline void set_immediate(const std::function<void()> cb)
     {
-        get_sqe_queue.push_back({ wait_sqe_id, cb });
+        immediate_queue.push_back(cb);
-        return wait_sqe_id++;
-    }
-    inline void cancel_wait_sqe(int wait_id)
-    {
-        for (int i = 0; i < get_sqe_queue.size(); i++)
-        {
-            if (get_sqe_queue[i].first == wait_id)
-            {
-                get_sqe_queue.erase(get_sqe_queue.begin()+i, get_sqe_queue.begin()+i+1);
-            }
-        }
     }
     inline int submit()
     {

src/rw_blocking.cpp

@@ -15,7 +15,7 @@ int read_blocking(int fd, void *read_buf, size_t remaining)
     size_t done = 0;
     while (done < remaining)
     {
-        size_t r = read(fd, read_buf, remaining-done);
+        ssize_t r = read(fd, read_buf, remaining-done);
         if (r <= 0)
         {
             if (!errno)
@@ -41,7 +41,7 @@ int write_blocking(int fd, void *write_buf, size_t remaining)
     size_t done = 0;
     while (done < remaining)
     {
-        size_t r = write(fd, write_buf, remaining-done);
+        ssize_t r = write(fd, write_buf, remaining-done);
         if (r < 0)
         {
             if (errno != EINTR && errno != EAGAIN && errno != EPIPE)

src/stub_bench.cpp

@@ -83,6 +83,7 @@ int connect_stub(const char *server_address, int server_port)
     if (connect(connect_fd, (sockaddr*)&addr, sizeof(addr)) < 0)
     {
         perror("connect");
+        close(connect_fd);
         return -1;
     }
     int one = 1;

src/test_cluster_client.cpp

@@ -8,7 +8,6 @@
 
 void configure_single_pg_pool(cluster_client_t *cli)
 {
-    cli->st_cli.on_load_pgs_hook(true);
     cli->st_cli.parse_state((etcd_kv_t){
         .key = "/config/pools",
         .value = json11::Json::object {
@@ -43,6 +42,7 @@ void configure_single_pg_pool(cluster_client_t *cli)
             { "state", json11::Json::array { "active" } },
         },
     });
+    cli->st_cli.on_load_pgs_hook(true);
     std::map<std::string, etcd_kv_t> changes;
     cli->st_cli.on_change_hook(changes);
 }
@@ -188,7 +188,6 @@ void test1()
     int *r1 = test_write(cli, 0, 4096, 0x55);
     configure_single_pg_pool(cli);
     pretend_connected(cli, 1);
-    cli->continue_ops(true);
     can_complete(r1);
     check_op_count(cli, 1, 1);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
@@ -196,8 +195,6 @@ void test1()
     pretend_disconnected(cli, 1);
     int *r2 = test_sync(cli);
     pretend_connected(cli, 1);
-    check_op_count(cli, 1, 0);
-    cli->continue_ops(true);
     check_op_count(cli, 1, 1);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
     check_op_count(cli, 1, 1);
@@ -303,8 +300,6 @@ void test1()
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), -EPIPE);
     check_disconnected(cli, 1);
     pretend_connected(cli, 1);
-    check_op_count(cli, 1, 0);
-    cli->continue_ops(true);
     check_op_count(cli, 1, 1);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
     check_op_count(cli, 1, 1);

src/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 0.8.3
+Version: 0.8.6
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

src/vitastor_c.cpp

@@ -207,6 +207,28 @@ void vitastor_c_write(vitastor_c *client, uint64_t inode, uint64_t offset, uint6
     client->cli->execute(op);
 }
 
+void vitastor_c_read_bitmap(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len,
+    int with_parents, VitastorReadBitmapHandler cb, void *opaque)
+{
+    cluster_op_t *op = new cluster_op_t;
+    op->opcode = with_parents ? OSD_OP_READ_CHAIN_BITMAP : OSD_OP_READ_BITMAP;
+    op->inode = inode;
+    op->offset = offset;
+    op->len = len;
+    op->callback = [cb, opaque](cluster_op_t *op)
+    {
+        uint8_t *bitmap = NULL;
+        if (op->retval >= 0)
+        {
+            bitmap = (uint8_t*)op->bitmap_buf;
+            op->bitmap_buf = NULL;
+        }
+        cb(opaque, op->retval, bitmap);
+        delete op;
+    };
+    client->cli->execute(op);
+}
+
 void vitastor_c_sync(vitastor_c *client, VitastorIOHandler cb, void *opaque)
 {
     cluster_op_t *op = new cluster_op_t;
@@ -245,6 +267,25 @@ uint64_t vitastor_c_inode_get_num(void *handle)
     return watch->cfg.num;
 }
 
+uint32_t vitastor_c_inode_get_block_size(vitastor_c *client, uint64_t inode_num)
+{
+    auto pool_it = client->cli->st_cli.pool_config.find(INODE_POOL(inode_num));
+    if (pool_it == client->cli->st_cli.pool_config.end())
+        return 0;
+    auto & pool_cfg = pool_it->second;
+    uint32_t pg_data_size = (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pool_cfg.pg_size-pool_cfg.parity_chunks);
+    return pool_cfg.data_block_size * pg_data_size;
+}
+
+uint32_t vitastor_c_inode_get_bitmap_granularity(vitastor_c *client, uint64_t inode_num)
+{
+    auto pool_it = client->cli->st_cli.pool_config.find(INODE_POOL(inode_num));
+    if (pool_it == client->cli->st_cli.pool_config.end())
+        return 0;
+    // FIXME: READ_BITMAP may fails if parent bitmap granularity differs from inode bitmap granularity
+    return pool_it->second.bitmap_granularity;
+}
+
 int vitastor_c_inode_get_readonly(void *handle)
 {
     inode_watch_t *watch = (inode_watch_t*)handle;

src/vitastor_c.h

@@ -6,6 +6,9 @@
 #ifndef VITASTOR_QEMU_PROXY_H
 #define VITASTOR_QEMU_PROXY_H
 
+// C API wrapper version
+#define VITASTOR_C_API_VERSION 1
+
 #ifndef POOL_ID_BITS
 #define POOL_ID_BITS 16
 #endif
@@ -21,6 +24,7 @@ typedef struct vitastor_c vitastor_c;
 
 typedef void VitastorReadHandler(void *opaque, long retval, uint64_t version);
 typedef void VitastorIOHandler(void *opaque, long retval);
+typedef void VitastorReadBitmapHandler(void *opaque, long retval, uint8_t *bitmap);
 
 // QEMU
 typedef void IOHandler(void *opaque);
@@ -42,11 +46,15 @@ void vitastor_c_read(vitastor_c *client, uint64_t inode, uint64_t offset, uint64
     struct iovec *iov, int iovcnt, VitastorReadHandler cb, void *opaque);
 void vitastor_c_write(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len, uint64_t check_version,
     struct iovec *iov, int iovcnt, VitastorIOHandler cb, void *opaque);
+void vitastor_c_read_bitmap(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len,
+    int with_parents, VitastorReadBitmapHandler cb, void *opaque);
 void vitastor_c_sync(vitastor_c *client, VitastorIOHandler cb, void *opaque);
 void vitastor_c_watch_inode(vitastor_c *client, char *image, VitastorIOHandler cb, void *opaque);
 void vitastor_c_close_watch(vitastor_c *client, void *handle);
 uint64_t vitastor_c_inode_get_size(void *handle);
 uint64_t vitastor_c_inode_get_num(void *handle);
+uint32_t vitastor_c_inode_get_block_size(vitastor_c *client, uint64_t inode_num);
+uint32_t vitastor_c_inode_get_bitmap_granularity(vitastor_c *client, uint64_t inode_num);
 int vitastor_c_inode_get_readonly(void *handle);
 
 #ifdef __cplusplus

tests/test_snapshot.sh

@@ -22,6 +22,16 @@ LD_PRELOAD="build/src/libfio_vitastor.so" \
 LD_PRELOAD="build/src/libfio_vitastor.so" \
     fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -rw=read -etcd=$ETCD_URL -pool=1 -inode=3 -size=32M
 
+qemu-img convert -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=2:size=$((32*1024*1024)):skip-parents=1" \
+    -O qcow2 ./testdata/layer0.qcow2
+
+qemu-img create -f qcow2 ./testdata/empty.qcow2 32M
+
+qemu-img convert -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=3:size=$((32*1024*1024)):skip-parents=1" \
+    -O qcow2 -o 'cluster_size=4k' -B empty.qcow2 ./testdata/layer1.qcow2
+
 qemu-img convert -S 4096 -p \
     -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=3:size=$((32*1024*1024))" \
     -O raw ./testdata/merged.bin
@@ -52,4 +62,18 @@ qemu-img convert -S 4096 -p \
 
 cmp ./testdata/merged.bin ./testdata/merged-by-tool.bin
 
+# Test merge by qemu-img
+
+qemu-img rebase -u -b layer0.qcow2 ./testdata/layer1.qcow2
+
+qemu-img convert -S 4096 -f qcow2 ./testdata/layer1.qcow2 -O raw ./testdata/rebased.bin
+
+cmp ./testdata/merged.bin ./testdata/rebased.bin
+
+qemu-img rebase -u -b '' ./testdata/layer1.qcow2
+
+qemu-img convert -S 4096 -f qcow2 ./testdata/layer1.qcow2 -O raw ./testdata/rebased.bin
+
+cmp ./testdata/layer1.bin ./testdata/rebased.bin
+
 format_green OK