Remove n_subops=0 (TODO also needs OSD unit tests)

Followup to mark_partial_write fix - also free subops in other branch to prevent assert on ENOSPC
Fix header checksum after new store write-meta
2026-05-21 13:31:32 +03:00 · 2026-05-20 22:04:28 +03:00 · 2026-05-19 12:30:21 +03:00 · 2026-05-19 02:27:03 +03:00 · 2026-05-18 02:24:22 +03:00 · 2026-05-18 01:17:56 +03:00
106 changed files with 2277 additions and 1337 deletions
@@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 2.8...3.30)

 project(vitastor)

-set(VITASTOR_VERSION "3.0.8")
+set(VITASTOR_VERSION "3.0.12")

 include(CTest)

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v3.0.8
+VITASTOR_VERSION ?= v3.0.12

 all: build push

@@ -49,7 +49,7 @@ spec:
            capabilities:
              add: ["SYS_ADMIN"]
            allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v3.0.8
+          image: vitalif/vitastor-csi:v3.0.12
          args:
            - "--node=$(NODE_ID)"
            - "--endpoint=$(CSI_ENDPOINT)"
@@ -121,7 +121,7 @@ spec:
            privileged: true
            capabilities:
              add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v3.0.8
+          image: vitalif/vitastor-csi:v3.0.12
          args:
            - "--node=$(NODE_ID)"
            - "--endpoint=$(CSI_ENDPOINT)"
@@ -5,7 +5,7 @@ package vitastor

 const (
    vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "3.0.8"
+    vitastorCSIDriverVersion = "3.0.12"
 )

 // Config struct fills the parameters of request or user input
@@ -1,4 +1,4 @@
-vitastor (3.0.8-1) unstable; urgency=medium
+vitastor (3.0.12-1) unstable; urgency=medium

  * Bugfixes

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v3.0.8
+VITASTOR_VERSION ?= v3.0.12

 all: build push

@@ -4,7 +4,7 @@
 #

 # Desired Vitastor version
-VITASTOR_VERSION=v3.0.8
+VITASTOR_VERSION=v3.0.12

 # Additional arguments for all containers
 # For example, you may want to specify a custom logging driver here
@@ -70,6 +70,7 @@ with an OSD restart or, for some of them, even without restarting by updating co
 - [use_atomic_flag](#use_atomic_flag)
 - [pg_reshard_chunk_size](#pg_reshard_chunk_size)
 - [pg_reshard_chunk_pause_ms](#pg_reshard_chunk_pause_ms)
+- [gc_on_start](#gc_on_start)

 ## bind_address

@@ -753,3 +754,9 @@ This option sets the maximum number of object is a chunk. Moving 100k objects us
 - Default: 100

 This option sets the interval between handling two PG count change chunks.
+
+## gc_on_start
+
+- Type: boolean
+
+Forcibly clean all garbage entries in the new store on every OSD restart.
@@ -71,6 +71,7 @@
 - [use_atomic_flag](#use_atomic_flag)
 - [pg_reshard_chunk_size](#pg_reshard_chunk_size)
 - [pg_reshard_chunk_pause_ms](#pg_reshard_chunk_pause_ms)
+- [gc_on_start](#gc_on_start)

 ## bind_address

@@ -793,3 +794,9 @@ pg_minsize OSD во время переключений, что может по
 - Значение по умолчанию: 100

 Данная опция задаёт интервал между обработкой двух порций изменения числа PG пулов.
+
+## gc_on_start
+
+- Тип: булево (да/нет)
+
+Принудительно очищать все мусорные записи в новом хранилище при каждом запуске OSD.
@@ -938,3 +938,7 @@
    This option sets the interval between handling two PG count change chunks.
  info_ru: |
    Данная опция задаёт интервал между обработкой двух порций изменения числа PG пулов.
+- name: gc_on_start
+  type: bool
+  info: Forcibly clean all garbage entries in the new store on every OSD restart.
+  info_ru: Принудительно очищать все мусорные записи в новом хранилище при каждом запуске OSD.
@@ -26,9 +26,9 @@ at Vitastor Kubernetes operator: https://github.com/Antilles7227/vitastor-operat
 The instruction is very simple.

 1. Download a Docker image of the desired version: \
-   `docker pull vitalif/vitastor:v3.0.8`
+   `docker pull vitalif/vitastor:v3.0.12`
 2. Install scripts to the host system: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v3.0.8 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v3.0.12 install.sh`
 3. Reload udev rules: \
   `udevadm control --reload-rules`
 4. Enable the vitastor-host service: \
@@ -25,9 +25,9 @@ Vitastor можно установить в Docker/Podman. При этом etcd,
 Инструкция по установке максимально простая.

 1. Скачайте Docker-образ желаемой версии: \
-   `docker pull vitalif/vitastor:v3.0.8`
+   `docker pull vitalif/vitastor:v3.0.12`
 2. Установите скрипты в хост-систему командой: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v3.0.8 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitalif/vitastor:v3.0.12 install.sh`
 3. Перезагрузите правила udev: \
   `udevadm control --reload-rules`
 4. Включите сервис vitastor-host: \
@@ -16,8 +16,7 @@
  designated initializers support from C++20
 - CMake
 - jerasure headers and libraries
- ISA-L, libibverbs and librdmacm headers and libraries (optional)
- tcmalloc (google-perftools-dev)
+- ISA-L, libibverbs, librdmacm, libnl3 headers and libraries (optional)

 ## Basic instructions

@@ -16,8 +16,7 @@
  назначенных инициализаторов (designated initializers) из C++20
 - CMake
 - Заголовки и библиотеки jerasure
- Опционально - заголовки и библиотеки ISA-L, libibverbs, librdmacm
- tcmalloc (google-perftools-dev)
+- Опционально - заголовки и библиотеки ISA-L, libibverbs, librdmacm, libnl3

 ## Базовая инструкция

@@ -262,3 +262,4 @@ Options:
 | `--logfile <FILE>` | log to the specified file                                |
 | `--enforce 1`      | enforce permissions at the server side (no by default)   |
 | `--foreground 1`   | stay in foreground, do not daemonize                     |
+| `--trace`          | trace all NFS requests                                   |
@@ -274,3 +274,4 @@ VitastorFS из GPUDirect.
 | `--logfile <FILE>` | записывать логи в заданный файл                         |
 | `--enforce 1`      | проверять права доступа на стороне сервера (по умолчанию нет) |
 | `--foreground 1`   | не уходить в фон после запуска                          |
+| `--trace`          | логгировать все запросы NFS                             |
@@ -627,7 +627,7 @@ class Mon
                if (this.state.pg.history[pool_id] &&
                    this.state.pg.history[pool_id][pg])
                {
-                    pg_history[pg-1] = this.state.pg.history[pool_id][pg];
+                    pg_history[pg-1] = JSON.parse(JSON.stringify(this.state.pg.history[pool_id][pg]));
                }
            }
            const real_prev_pgs = [];
@@ -719,7 +719,7 @@ class Mon
                this.next_recheck_timer = null;
                this.next_recheck_at = 0;
                this.schedule_recheck();
-            }, now-this.next_recheck_at);
+            }, (this.next_recheck_at-now)*1000);
        }
    }

@@ -1,6 +1,6 @@
 {
  "name": "vitastor-mon",
-  "version": "3.0.8",
+  "version": "3.0.12",
  "description": "Vitastor SDS monitor service",
  "main": "mon-main.js",
  "scripts": {
@@ -84,7 +84,7 @@ function scale_pg_history(prev_pg_history, prev_pgs, new_pgs)
        finish_pg_history(merged_history[1]);
        for (let i = 0; i < new_pg_count; i++)
        {
-            new_pg_history[i] = { ...merged_history[1] };
+            new_pg_history[i] = JSON.parse(JSON.stringify(merged_history[1]));
        }
    }
    // Mark history keys for removed PGs as removed
@@ -102,7 +102,7 @@ function scale_pg_count(prev_pgs, new_pg_count)
    {
        for (let i = prev_pgs.length; i < new_pg_count; i++)
        {
-            prev_pgs[i] = prev_pgs[i % prev_pgs.length];
+            prev_pgs[i] = [ ...prev_pgs[i % prev_pgs.length] ];
        }
    }
    else if (prev_pgs.length > new_pg_count)
@@ -37,6 +37,7 @@ function derive_osd_stats(st, prev, prev_diff)
        const n = c.count - BigInt(pr && pr.count||0);
        diff.recovery_stats[op] = { ...c, bps: n > 0 ? b*1000n/timediff : 0n, iops: n > 0 ? n*1000n/timediff : 0n };
    }
+    diff.inode_stats = {};
    for (const pool_id in st.inode_stats||{})
    {
        diff.inode_stats[pool_id] = {};
@@ -1,6 +1,6 @@
 {
  "name": "vitastor",
-  "version": "3.0.8",
+  "version": "3.0.12",
  "description": "Low-level native bindings to Vitastor client library",
  "main": "index.js",
  "keywords": [
@@ -50,7 +50,7 @@ from cinder.volume import configuration
 from cinder.volume import driver
 from cinder.volume import volume_utils

-VITASTOR_VERSION = '3.0.8'
+VITASTOR_VERSION = '3.0.12'

 LOG = logging.getLogger(__name__)

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        3.0.8
+Version:        3.0.12
 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-3.0.8.el10.tar.gz
+Source0:        vitastor-3.0.12.el10.tar.gz

 BuildRequires:  gperftools-devel
 BuildRequires:  gcc-c++
@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        3.0.8
+Version:        3.0.12
 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-3.0.8.el7.tar.gz
+Source0:        vitastor-3.0.12.el7.tar.gz

 BuildRequires:  gperftools-devel
 BuildRequires:  devtoolset-9-gcc-c++
@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        3.0.8
+Version:        3.0.12
 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-3.0.8.el8.tar.gz
+Source0:        vitastor-3.0.12.el8.tar.gz

 BuildRequires:  gperftools-devel
 BuildRequires:  gcc-toolset-9-gcc-c++
@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        3.0.8
+Version:        3.0.12
 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-3.0.8.el9.tar.gz
+Source0:        vitastor-3.0.12.el9.tar.gz

 BuildRequires:  gperftools-devel
 BuildRequires:  gcc-c++
@@ -20,7 +20,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 endif()
 set(ENABLE_COVERAGE false CACHE BOOL "Enable code coverage")

-add_definitions(-DVITASTOR_VERSION="3.0.8")
+add_definitions(-DVITASTOR_VERSION="3.0.12")
 add_definitions(-D_GNU_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64 -Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -fno-omit-frame-pointer -fvisibility=hidden -I ${CMAKE_SOURCE_DIR}/src)
 add_link_options(-fno-omit-frame-pointer)
 if (${WITH_ASAN})
@@ -228,4 +228,9 @@ public:
    virtual uint64_t get_journal_size() = 0;

    virtual uint32_t get_bitmap_granularity() = 0;
+
+    virtual uint64_t get_live_entries() = 0;
+    virtual uint64_t get_live_memory() = 0;
+    virtual uint64_t get_garbage_entries() = 0;
+    virtual uint64_t get_garbage_memory() = 0;
 };
@@ -94,6 +94,9 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
    csum_block_size = parse_size(config["csum_block_size"]);
    discard_on_start = config.find("discard_on_start") != config.end() &&
        (config["discard_on_start"] == "true" || config["discard_on_start"] == "1" || config["discard_on_start"] == "yes");
+    gc_on_start = config.find("gc_on_start") == config.end() ||
+        (config["gc_on_start"] == "true" || config["gc_on_start"] == "1" || config["gc_on_start"] == "yes");
+    skip_double_claim = (config["skip_double_claim"] == "true" || config["skip_double_claim"] == "1" || config["skip_double_claim"] == "yes");
    min_discard_size = parse_size(config["min_discard_size"]);
    if (!min_discard_size)
        min_discard_size = 1024*1024;
@@ -57,6 +57,10 @@ struct blockstore_disk_t
    bool inmemory_journal = true;
    // Data discard granularity and minimum size (for the sake of performance)
    bool discard_on_start = false;
+    // GC on start (new store)
+    bool gc_on_start = true;
+    // Skip double claim conflicts on start (new store, temporary until the bug is found)
+    bool skip_double_claim = false;
    uint64_t min_discard_size = 1024*1024;
    uint64_t discard_granularity = 0;

@@ -174,14 +174,18 @@ bool journal_flusher_co::loop()
    else if (wait_state == 19) goto resume_19;
    else if (wait_state == 20) goto resume_20;
    else if (wait_state == 21) goto resume_21;
+    else if (wait_state == 22) goto resume_22;
+    else if (wait_state == 23) goto resume_23;
+    else if (wait_state == 24) goto resume_24;
+    else if (wait_state == 25) goto resume_25;
 resume_0:
    wait_state = 0;
    wait_count = 0;
    cur_oid = {};
    res = bs->heap->get_next_compact(cur_oid);
-    // Advance fsynced_lsn every <journal_trim_interval> intent writes
    if ((bs->intent_write_counter >= bs->journal_trim_interval) && co_id == 0)
    {
+        // Advance fsynced_lsn every <journal_trim_interval> intent writes
        bs->intent_write_counter = 0;
 resume_17:
 resume_18:
@@ -196,6 +200,7 @@ resume_21:
    if (res == ENOENT && flusher->force_start > 0 && co_id == 0 &&
        (!bs->dsk.disable_journal_fsync || !bs->dsk.disable_meta_fsync || !bs->dsk.disable_data_fsync))
    {
+        // When under pressure, do an additional fsync to force entries to be marked compactable
        flusher->active_flushers++;
 resume_14:
 resume_15:
@@ -259,11 +264,9 @@ resume_1:
        if (wr->type() == BS_HEAP_SMALL_WRITE ||
            wr->type() == BS_HEAP_INTENT_WRITE && bs->dsk.csum_block_size > bs->dsk.bitmap_granularity)
        {
-            auto res = bs->prepare_read(read_vec, cur_obj, wr, 0, bs->dsk.data_block_size,
+            bs->prepare_read(read_vec, cur_obj, wr, 0, bs->dsk.data_block_size,
                wr->type() == BS_HEAP_INTENT_WRITE && bs->dsk.csum_block_size > bs->dsk.bitmap_granularity && !bs->perfect_csum_update
                ? COPY_BUF_SKIP_CSUM : 0);
-            if (res > 0)
-                copy_count++;
        }
    });
    if (!compact_info.compact_lsn)
@@ -273,6 +276,25 @@ resume_1:
        bs->heap->unlock_entry(cur_oid);
        goto resume_0;
    }
+    flusher->active_flushers++;
+    for (i = 0; i < read_vec.size(); i++)
+    {
+        if ((read_vec[i].copy_flags & COPY_BUF_JOURNAL) &&
+            !(read_vec[i].copy_flags & COPY_BUF_COALESCED))
+        {
+            copy_count++;
+        }
+    }
+    if (copy_count > 0 && !bs->dsk.disable_data_fsync)
+    {
+        init_fsync_data();
+    }
+    if (bs->log_level > 10)
+    {
+        printf("Compacting %jx:%jx v%ju..v%ju / l%ju..l%ju (%d writes)\n", cur_oid.inode, cur_oid.stripe,
+            compact_info.clean_wr->version, compact_info.compact_version,
+            compact_info.clean_wr->lsn, compact_info.compact_lsn, copy_count);
+    }
    mem_or(new_bmp, compact_info.clean_wr->get_int_bitmap(bs->heap), bs->dsk.clean_entry_bitmap_size);
    if (!bitmap_copied)
    {
@@ -291,13 +313,6 @@ resume_1:
        csum_copy.clear();
    }
    clean_loc = compact_info.clean_wr->big_location(bs->heap);
-    flusher->active_flushers++;
-    if (bs->log_level > 10)
-    {
-        printf("Compacting %jx:%jx v%ju..v%ju / l%ju..l%ju (%d writes)\n", cur_oid.inode, cur_oid.stripe,
-            compact_info.clean_wr->version, compact_info.compact_version,
-            compact_info.clean_wr->lsn, compact_info.compact_lsn, copy_count);
-    }
    overwrite_start = overwrite_end = 0;
    if (read_vec.size() > 0)
    {
@@ -336,6 +351,13 @@ resume_3:
    if (res == ENOENT || res == EDOM)
    {
        // Abort compaction
+abort_compact:
+        if (copy_count > 0 && !bs->dsk.disable_data_fsync)
+        {
+            cur_sync->member_count--;
+            if (cur_sync->member_count > 0)
+                bs->ringloop->wakeup();
+        }
        flusher->flushing.erase(cur_oid);
        bs->heap->unlock_entry(cur_oid);
        flusher->active_flushers--;
@@ -349,10 +371,7 @@ resume_4:
        if (res == ENOENT)
        {
            // Abort compaction
-            flusher->flushing.erase(cur_oid);
-            bs->heap->unlock_entry(cur_oid);
-            flusher->active_flushers--;
-            goto resume_0;
+            goto abort_compact;
        }
        if (res == EAGAIN)
        {
@@ -381,14 +400,14 @@ resume_9:
    for (i = 0; i < read_vec.size(); i++)
    {
        if ((read_vec[i].copy_flags & COPY_BUF_JOURNAL) &&
-            !(read_vec[i].copy_flags & COPY_BUF_COALESCED) ||
-            (read_vec[i].copy_flags & COPY_BUF_PADDED)) // FIXME Shit, simplify these flags
+            !(read_vec[i].copy_flags & COPY_BUF_COALESCED))
        {
            assert(read_vec[i].buf);
            await_sqe(10);
            data->iov = (struct iovec){ read_vec[i].buf + (read_vec[i].copy_flags & COPY_BUF_PADDED
                ? read_vec[i].offset - read_vec[i].disk_offset : 0), (size_t)read_vec[i].len };
            data->callback = simple_callback_w;
+            assert(clean_loc + read_vec[i].offset + data->iov.iov_len <= bs->dsk.block_count*bs->dsk.data_block_size);
            io_uring_prep_writev(sqe, bs->dsk.data_fd, &data->iov, 1, bs->dsk.data_offset + clean_loc + read_vec[i].offset);
            wait_count++;
        }
@@ -399,6 +418,17 @@ resume_11:
        wait_state = 11;
        return false;
    }
+    if (copy_count > 0 && !bs->dsk.disable_data_fsync)
+    {
+resume_22:
+resume_23:
+resume_24:
+resume_25:
+        if (!fsync_data(22))
+        {
+            return false;
+        }
+    }
    // Lock is only needed to prevent freeing the big_write because we overwrite it...
    bs->heap->unlock_entry(cur_oid);
    // Mark the object compacted, but don't free and remove small_writes
@@ -408,12 +438,14 @@ resume_11:
    if (!cur_obj)
    {
        // Abort compaction
+        flusher->active_flushers--;
        flusher->flushing.erase(cur_oid);
        goto resume_0;
    }
    if (!calc_block_checksums())
    {
        // Abort compaction
+        flusher->active_flushers--;
        flusher->flushing.erase(cur_oid);
        goto resume_0;
    }
@@ -422,6 +454,7 @@ resume_11:
    if (res == EBUSY)
    {
        // Abort compaction, object is already overwritten by something else
+        flusher->active_flushers--;
        flusher->flushing.erase(cur_oid);
        goto resume_0;
    }
@@ -586,7 +619,7 @@ int journal_flusher_co::check_and_punch_checksums()
            bs->heap->calc_block_checksums((uint32_t*)(new_csums+csum_off), vec.buf, punch_bmp, vec.offset, vec.offset+vec.len, true, NULL);
        }
    }
-    // Modified, we should add_punch_holes and then write the block to disk
+    // Modified, we should punch_holes and then write the block to disk
    return EBUSY;
 }

@@ -699,6 +732,67 @@ resume_1:
    return true;
 }

+void journal_flusher_co::init_fsync_data()
+{
+    cur_sync = flusher->data_syncs.begin();
+    if (cur_sync == flusher->data_syncs.end() || cur_sync->ready_count > 0)
+    {
+        cur_sync = flusher->data_syncs.emplace(cur_sync);
+    }
+    cur_sync->member_count++;
+}
+
+bool journal_flusher_co::fsync_data(int wait_base)
+{
+    if (wait_state == wait_base)
+        goto resume_0;
+    else if (wait_state == wait_base+1)
+        goto resume_1;
+    else if (wait_state == wait_base+2)
+        goto resume_2;
+    else if (wait_state == wait_base+3)
+        goto resume_3;
+    cur_sync->ready_count++;
+resume_0:
+    if (cur_sync->ready_count < cur_sync->member_count)
+    {
+        wait_state = wait_base;
+        return false;
+    }
+    if (!cur_sync->sent)
+    {
+        // Sync batch is ready. Do it.
+        await_sqe(1);
+        data->iov = { 0 };
+        data->callback = simple_callback_w;
+        io_uring_prep_fsync(sqe, bs->dsk.data_fd, IORING_FSYNC_DATASYNC);
+        cur_sync->sent = true;
+        wait_count++;
+resume_2:
+        if (wait_count > 0)
+        {
+            wait_state = wait_base+2;
+            return false;
+        }
+        cur_sync->done = true;
+        // Wake up other flushers
+        bs->ringloop->wakeup();
+    }
+resume_3:
+    if (!cur_sync->done)
+    {
+        wait_state = wait_base+3;
+        return false;
+    }
+    cur_sync->done_count++;
+    if (cur_sync->done_count >= cur_sync->member_count)
+    {
+        flusher->data_syncs.erase(cur_sync);
+        cur_sync = flusher->data_syncs.end();
+    }
+    return true;
+}
+
 bool journal_flusher_co::fsync_meta(int wait_base)
 {
    if (wait_state == wait_base)        goto resume_0;
@@ -25,6 +25,15 @@ struct flusher_meta_write_t
    std::map<uint64_t, meta_sector_t>::iterator it;
 };

+struct flusher_data_sync_t
+{
+    int member_count = 0;
+    int ready_count = 0;
+    int done_count = 0;
+    bool sent = false;
+    bool done = false;
+};
+
 class journal_flusher_t;

 // Journal flusher coroutine
@@ -58,6 +67,7 @@ class journal_flusher_co
    int i, res;
    bool read_to_fill_incomplete;
    int copy_count;
+    std::list<flusher_data_sync_t>::iterator cur_sync;

    friend class journal_flusher_t;

@@ -68,6 +78,8 @@ class journal_flusher_co
    bool calc_block_checksums();
    bool write_meta_block(int wait_base);
    bool read_buffered(int wait_base);
+    void init_fsync_data();
+    bool fsync_data(int wait_base);
    bool fsync_meta(int wait_base);
    bool fsync_buffer(int wait_base);
    bool trim_lsn(int wait_base);
@@ -88,6 +100,7 @@ class journal_flusher_t

    robin_hood::unordered_flat_set<object_id> flushing;
    int active_flushers = 0;
+    std::list<flusher_data_sync_t> data_syncs;
    int wanting_meta_fsync = 0;
    bool fsyncing_meta = false;
    int syncing_buffer = 0;
@@ -57,11 +57,11 @@ struct __attribute__((__packed__)) heap_entry_t
    inline heap_small_write_t& small() { return *(heap_small_write_t*)this; }
    inline heap_big_write_t& big() { return *(heap_big_write_t*)this; }
    inline heap_big_intent_t& big_intent() { return *(heap_big_intent_t*)this; }
-    bool is_garbage();
+    bool is_garbage() const;
    void set_garbage();
-    bool is_overwrite();
-    bool is_compactable();
-    bool is_before(heap_entry_t *other);
+    bool is_overwrite() const;
+    bool is_compactable() const;
+    bool is_before(const heap_entry_t *other) const;
    uint32_t get_size(blockstore_heap_t *heap);
    uint8_t *get_ext_bitmap(blockstore_heap_t *heap);
    uint8_t *get_int_bitmap(blockstore_heap_t *heap);
@@ -117,10 +117,13 @@ struct heap_object_mvcc_t

 struct heap_block_info_t
 {
-    uint32_t used_space = 0;
+    struct __attribute__((__packed__))
+    {
+        uint32_t used_space = 0;
+        uint32_t garbage_space = 0;
+    };
    uint64_t mod_lsn = 0, mod_lsn_to = 0; // only 1 block write of LSN sequence is allowed at a moment
-    bool is_writing: 1;
-    bool has_garbage: 1;
+    bool is_writing = false;
    std::vector<heap_list_item_t*> entries;
 };

@@ -155,6 +158,16 @@ struct heap_li_equal
    }
 };

+struct heap_recheck_state_t
+{
+    heap_entry_t *obj = NULL;
+    heap_entry_t *next_wr = NULL;
+    size_t total_reads = 0;
+    size_t sent_reads = 0;
+    size_t checked_reads = 0;
+    heap_entry_t *bad_wr = NULL;
+};
+
 using i64hash_t = robin_hood::hash<uint64_t>;
 using heap_inode_map_t = robin_hood::unordered_flat_set<heap_list_item_t*, heap_li_hash, heap_li_equal, 88>;
 using heap_block_index_t = robin_hood::unordered_flat_map<uint64_t,
@@ -184,6 +197,11 @@ class blockstore_heap_t
    uint64_t buffer_area_used_space = 0;
    uint64_t data_used_space = 0;

+    uint64_t live_entries = 0;
+    uint64_t live_memory = 0;
+    uint64_t garbage_entries = 0;
+    uint64_t garbage_memory = 0;
+
    uint64_t next_lsn = 0;
    uint32_t last_allocated_block = UINT32_MAX;
    heap_mvcc_map_t object_mvcc;
@@ -200,9 +218,11 @@ class blockstore_heap_t

    bool marked_used_blocks = false;
    bool recheck_queue_filled = false;
-    std::vector<heap_list_item_t*> loaded_list_items;
+    std::vector<heap_list_item_t*> postponed_items;
    std::set<uint32_t> recheck_modified_blocks;
    std::deque<heap_entry_t*> recheck_queue;
+    std::map<heap_entry_t*, heap_recheck_state_t> recheck_states;
+    size_t recheck_pending_reads = 0;
    int recheck_in_progress = 0;
    bool in_recheck = false;
    std::function<void(bool is_data, uint64_t offset, uint64_t len, uint8_t* buf, std::function<void()>)> recheck_cb;
@@ -211,14 +231,22 @@ class blockstore_heap_t
    uint64_t get_pg_id(inode_t inode, uint64_t stripe);
    bool validate_object(heap_entry_t *obj);
    void fill_recheck_queue();
+    void recheck_drop_entries(heap_entry_t *obj, heap_entry_t *bad_wr);
+    void recheck_start_reads(heap_recheck_state_t *st);
    int mark_used_blocks();
+    void init_free_bad_entry(heap_entry_t *wr);
+    void init_erase_bad_entry(heap_list_item_t *li);
+    bool init_erase_double_claim(heap_list_item_t *prev_li, heap_list_item_t *cur_li);
+    void recheck_full_gc();
    void recheck_buffer(heap_entry_t *cwr, uint8_t *buf);
    void defragment_block(uint32_t block_num);
    void reshard_add(heap_reshard_state_t *st, heap_list_item_t *li);

    void gc_block(heap_block_info_t & inf);
    int allocate_entry(uint32_t entry_size, uint32_t *block_num, bool allow_last_free);
-    void insert_list_item(heap_list_item_t *li);
+    void insert_list_items(heap_list_item_t** v, size_t count, bool postpone);
+    void remove_list_item(heap_list_item_t *li);
+    void unlink_list_item(heap_list_item_t *li);
    int add_entry(uint32_t wr_size, uint32_t *modified_block, bool allow_last_free,
        bool explicit_complete, std::function<void(heap_entry_t *wr)> fill_entry);
    int add_simple(heap_entry_t *obj, uint64_t version, uint32_t *modified_block, uint32_t entry_type);
@@ -345,6 +373,10 @@ public:
    uint32_t get_compact_queue_size();
    uint32_t get_to_compact_count();
    uint64_t get_compacted_count();
+    uint64_t get_live_entries();
+    uint64_t get_live_memory();
+    uint64_t get_garbage_entries();
+    uint64_t get_garbage_memory();

    uint64_t entry_pos(uint32_t block_num, uint32_t offset);
    heap_entry_t *entry_from_pos(uint64_t entry_pos, bool allow_unallocated = false);
@@ -101,6 +101,7 @@ void blockstore_impl_t::loop()
        unsigned initial_ring_space = ringloop->space_left();
        int op_idx = 0, new_idx = 0;
        bool has_unfinished_writes = false;
+        bool has_unfinished_sync = false;
        for (; op_idx < submit_queue.size(); op_idx++, new_idx++)
        {
            auto op = submit_queue[op_idx];
@@ -138,7 +139,13 @@ void blockstore_impl_t::loop()
            else if (op->opcode == BS_OP_SYNC)
            {
                // syncs only completed writes, so doesn't have to be blocked by anything
-                wr_st = continue_sync(op);
+                if (!has_unfinished_sync)
+                {
+                    wr_st = continue_sync(op);
+                    has_unfinished_sync = (wr_st != 2);
+                }
+                else
+                    wr_st = 0;
            }
            else if (op->opcode == BS_OP_STABLE || op->opcode == BS_OP_ROLLBACK)
            {
@@ -154,9 +161,7 @@ void blockstore_impl_t::loop()
                    wr_st = 2;
                }
                else
-                {
                    wr_st = 0;
-                }
            }
            if (wr_st == 2)
            {
@@ -229,4 +229,9 @@ public:
    uint64_t get_free_block_count();
    inline uint32_t get_bitmap_granularity() { return dsk.bitmap_granularity; }
    inline uint64_t get_journal_size() { return dsk.journal_len; }
+
+    inline uint64_t get_live_entries() { return heap->get_live_entries(); }
+    inline uint64_t get_live_memory() { return heap->get_live_memory(); }
+    inline uint64_t get_garbage_entries() { return heap->get_garbage_entries(); }
+    inline uint64_t get_garbage_memory() { return heap->get_garbage_memory(); }
 };
@@ -153,6 +153,14 @@ resume_1:
            );
            exit(1);
        }
+        uint32_t csum = hdr->header_csum;
+        hdr->header_csum = 0;
+        if (crc32c(0, hdr, sizeof(*hdr)) != csum)
+        {
+            printf("Metadata header is corrupt (checksum mismatch).\n");
+            exit(1);
+        }
+        hdr->header_csum = csum;
    }
    bs->heap->start_load(((blockstore_meta_header_v3_t *)bs->meta_superblock)->completed_lsn);
    if (bs->dsk.inmemory_journal)
@@ -240,23 +248,7 @@ resume_4:
    }
    // metadata read finished
    bs->heap->finish_load();
-    printf("Metadata entries loaded: %ju, used blocks: %ju / %ju\n", entries_loaded, bs->heap->get_data_used_space() / bs->dsk.data_block_size, bs->dsk.block_count);
-    if (zero_on_init && !bs->dsk.disable_meta_fsync)
-    {
-        GET_SQE();
-        io_uring_prep_fsync(sqe, bs->dsk.meta_fd, IORING_FSYNC_DATASYNC);
-        last_read_offset = 0;
-        data->iov = { 0 };
-        data->callback = [this](ring_data_t *data) { handle_event(data, -1); };
-        submitted++;
-        bs->ringloop->submit();
-    resume_5:
-        if (submitted > 0)
-        {
-            wait_state = 5;
-            return 1;
-        }
-    }
+    printf("Metadata entries loaded: %ju, rechecking unfinished writes and garbage entries\n", entries_loaded);
    // asynchronous recheck
 resume_6:
    wait_state = 6;
@@ -293,6 +285,11 @@ resume_7:
    if (bs->readonly)
    {
        recheck_mod.clear();
+        printf("Actual metadata entries: %ju\n", bs->heap->get_live_entries());
+    }
+    else
+    {
+        printf("Actual metadata entries: %ju, clearing garbage in %zu metadata blocks\n", bs->heap->get_live_entries(), recheck_mod.size());
    }
    for (i = 0; i < recheck_mod.size(); i++)
    {
@@ -306,7 +303,7 @@ resume_8:
        uint32_t block_num = recheck_mod[i];
        uint64_t block_offset = bs->dsk.meta_offset + (uint64_t)(block_num+1) * bs->dsk.meta_block_size;
        data = ((ring_data_t*)sqe->user_data);
-        uint8_t *buf = (uint8_t*)malloc_or_die(bs->dsk.meta_block_size);
+        uint8_t *buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, bs->dsk.meta_block_size);
        bs->heap->get_meta_block(block_num, buf);
        data->iov = { buf, bs->dsk.meta_block_size };
        data->callback = [this, buf, block_offset](ring_data_t *data)
@@ -332,5 +329,25 @@ resume_9:
    }
    free(metadata_buffer);
    metadata_buffer = NULL;
+    if (!bs->dsk.disable_meta_fsync && !bs->readonly)
+    {
+        GET_SQE();
+        io_uring_prep_fsync(sqe, bs->dsk.meta_fd, IORING_FSYNC_DATASYNC);
+        last_read_offset = 0;
+        data->iov = { 0 };
+        data->callback = [this](ring_data_t *data) { handle_event(data, -1); };
+        submitted++;
+        bs->ringloop->submit();
+    resume_5:
+        if (submitted > 0)
+        {
+            wait_state = 5;
+            return 1;
+        }
+    }
+    printf("Loading finished. Data used: %ju / %ju bytes (%s / %s)\n",
+        bs->heap->get_data_used_space(), bs->dsk.block_count * bs->dsk.data_block_size,
+        format_size(bs->heap->get_data_used_space()).c_str(),
+        format_size(bs->dsk.block_count * bs->dsk.data_block_size).c_str());
    return 0;
 }
@@ -462,6 +462,10 @@ int blockstore_impl_t::read_bitmap(object_id oid, uint64_t target_version, void
        {
            if (target_version >= wr->version)
            {
+                if (wr->type() == BS_HEAP_DELETE)
+                {
+                    return false;
+                }
                found = true;
                if (result_version)
                {
@@ -16,6 +16,7 @@ int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
    else if (priv->op_state == 5) goto resume_5;
    assert(!priv->op_state);
    op->retval = 0;
+    PRIV(op)->lsn = 0;
    priv->modified_block = priv->modified_block2 = UINT32_MAX;
    for (priv->stab_pos = 0; priv->stab_pos < op->len; priv->stab_pos++)
    {
@@ -36,6 +37,12 @@ int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
                FINISH_OP(op);
                return 2;
            }
+            if (res == ENOENT)
+            {
+                op->retval = -ENOENT;
+                FINISH_OP(op);
+                return 2;
+            }
            if (res == ENOSPC)
            {
                if (!heap->get_to_compact_count())
@@ -9,6 +9,7 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
    if (!PRIV(op)->op_state)
    {
        op->retval = 0;
+        PRIV(op)->lsn = 0;
    }
    int res = do_sync(op, 0);
    if (res == 2)
@@ -104,7 +105,8 @@ int blockstore_impl_t::do_sync(blockstore_op_t *op, int base_state)
        unsynced_data_write_count = unsynced_buffer_write_count = unsynced_meta_write_count = 0;
        return 2;
    }
-    PRIV(op)->modified_block = heap->get_completed_lsn();
+    assert(!PRIV(op)->lsn);
+    PRIV(op)->lsn = heap->get_completed_lsn();
    if (!submit_fsyncs(PRIV(op)->pending_ops))
    {
        PRIV(op)->wait_detail = 1;
@@ -118,6 +120,6 @@ resume_1:
        return 1;
    }
 resume_2:
-    heap->mark_lsn_fsynced(PRIV(op)->modified_block);
+    heap->mark_lsn_fsynced(PRIV(op)->lsn);
    return 2;
 }
@@ -37,6 +37,7 @@ void blockstore_impl_t::prepare_meta_block_write(uint32_t modified_block)
        heap->complete_block_write(modified_block);
        ringloop->wakeup();
    };
+    assert(((uint64_t)modified_block+2)*dsk.meta_block_size <= dsk.meta_area_size);
    io_uring_prep_writev(
        sqe, dsk.meta_fd, &data->iov, 1, dsk.meta_offset + ((uint64_t)modified_block+1)*dsk.meta_block_size
    );
@@ -177,6 +178,7 @@ enospc:
        ring_data_t *data = ((ring_data_t*)sqe->user_data);
        data->iov = (struct iovec){ op->buf, op->len };
        data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+        assert(loc+op->offset+op->len <= dsk.block_count*dsk.data_block_size);
        io_uring_prep_writev(sqe, dsk.data_fd, &data->iov, 1, dsk.data_offset + loc + op->offset);
        PRIV(op)->pending_ops++;
        write_iodepth++;
@@ -264,6 +266,7 @@ enospc:
            BS_SUBMIT_GET_SQE(sqe2, data2);
            data2->iov = (struct iovec){ op->buf, op->len };
            data2->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+            assert(loc+op->len <= dsk.journal_len);
            io_uring_prep_writev(sqe2, dsk.journal_fd, &data2->iov, 1, dsk.journal_offset + loc);
            PRIV(op)->pending_ops++;
        }
@@ -453,6 +456,7 @@ resume_10:
    BS_SUBMIT_GET_SQE(sqe, data);
    data->iov = (struct iovec){ op->buf, op->len };
    data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
+    assert(PRIV(op)->location + op->offset <= dsk.block_count*dsk.data_block_size);
    io_uring_prep_writev(sqe, dsk.data_fd, &data->iov, 1, dsk.data_offset + PRIV(op)->location + op->offset);
    if (dsk.use_atomic_flag)
        sqe->rw_flags = RWF_ATOMIC;
@@ -141,7 +141,7 @@ struct __attribute__((__packed__)) journal_entry
 inline uint32_t je_crc32(journal_entry *je)
 {
    // 0x48674bc7 = crc32(4 zero bytes)
-    return crc32c(0x48674bc7, ((uint8_t*)je)+4, je->size-4);
+    return je->size < 4 ? 0 : crc32c(0x48674bc7, ((uint8_t*)je)+4, je->size-4);
 }

 // "VITAstor"
@@ -520,6 +520,7 @@ resume_2:
                await_sqe(15);
                data->iov = (struct iovec){ it->buf, (size_t)it->len };
                data->callback = simple_callback_w;
+                assert(clean_loc+it->offset+it->len <= bs->dsk.block_count*bs->dsk.data_block_size);
                io_uring_prep_writev(
                    sqe, bs->dsk.data_fd, &data->iov, 1, bs->dsk.data_offset + clean_loc + it->offset
                );
@@ -749,6 +750,7 @@ bool journal_flusher_co::write_meta_block(flusher_meta_write_t & meta_block, int
    await_sqe(0);
    data->iov = (struct iovec){ meta_block.buf, (size_t)bs->dsk.meta_block_size };
    data->callback = simple_callback_w;
+    assert(bs->dsk.meta_block_size + meta_block.sector + bs->dsk.meta_block_size <= bs->dsk.meta_area_size);
    io_uring_prep_writev(
        sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + bs->dsk.meta_block_size + meta_block.sector
    );
@@ -855,4 +855,29 @@ std::string blockstore_impl_t::get_op_diag(blockstore_op_t *op)
    return std::string(buf);
 }

+uint64_t blockstore_impl_t::get_live_entries()
+{
+    return used_blocks;
+}
+
+uint64_t blockstore_impl_t::get_live_memory()
+{
+    uint64_t used = 0;
+    for (auto & kv: clean_db_shards)
+    {
+        used += kv.second.size() * sizeof(blockstore_clean_db_t::value_type);
+    }
+    return used;
+}
+
+uint64_t blockstore_impl_t::get_garbage_entries()
+{
+    return dirty_db.size();
+}
+
+uint64_t blockstore_impl_t::get_garbage_memory()
+{
+    return (sizeof(obj_ver_id) + sizeof(dirty_entry) + 32) * dirty_db.size();
+}
+
 } // namespace v1
@@ -332,6 +332,10 @@ public:
    inline uint64_t get_free_block_count() { return dsk.block_count - used_blocks; }
    inline uint32_t get_bitmap_granularity() { return dsk.disk_alignment; }
    inline uint64_t get_journal_size() { return dsk.journal_len; }
+    uint64_t get_live_entries();
+    uint64_t get_live_memory();
+    uint64_t get_garbage_entries();
+    uint64_t get_garbage_memory();
 };

 } // namespace v1
@@ -193,6 +193,7 @@ void blockstore_impl_t::prepare_journal_sector_write(int cur_sector, blockstore_
            (size_t)journal.block_size
        };
        data->callback = [this, flush_id = journal.submit_id](ring_data_t *data) { handle_journal_write(data, flush_id); };
+        assert(journal.sector_info[cur_sector].offset+journal.block_size <= dsk.journal_len);
        io_uring_prep_writev(
            sqe, dsk.journal_fd, &data->iov, 1, journal.offset + journal.sector_info[cur_sector].offset
        );
@@ -368,9 +368,9 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
        }
        data->iov.iov_len = op->len + stripe_offset + stripe_end; // to check it in the callback
        data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-        io_uring_prep_writev(
-            sqe, dsk.data_fd, PRIV(op)->iov_zerofill, vcnt, dsk.data_offset + (loc * dsk.data_block_size) + op->offset - stripe_offset
-        );
+        const uint64_t write_offset = (loc * dsk.data_block_size) + op->offset - stripe_offset;
+        assert(write_offset+op->len+stripe_offset+stripe_end <= dsk.block_count*dsk.data_block_size);
+        io_uring_prep_writev(sqe, dsk.data_fd, PRIV(op)->iov_zerofill, vcnt, dsk.data_offset + write_offset);
        PRIV(op)->pending_ops = 1;
        if (!(dirty_it->second.state & BS_ST_INSTANT))
        {
@@ -495,9 +495,8 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
                .op = op,
            });
            data2->callback = [this, flush_id = journal.submit_id](ring_data_t *data) { handle_journal_write(data, flush_id); };
-            io_uring_prep_writev(
-                sqe2, dsk.journal_fd, &data2->iov, 1, journal.offset + journal.next_free
-            );
+            assert(journal.next_free+op->len <= dsk.journal_len);
+            io_uring_prep_writev(sqe2, dsk.journal_fd, &data2->iov, 1, journal.offset + journal.next_free);
            PRIV(op)->pending_ops++;
        }
        else
@@ -12,7 +12,7 @@ if (RDMACM_LIBRARIES)
 	set(MSGR_RDMACM "msgr_rdmacm.cpp")
 endif (RDMACM_LIBRARIES)
 add_library(vitastor_common STATIC
-	../util/epoll_manager.cpp etcd_state_client.cpp messenger.cpp ../util/addr_util.cpp
+	../util/epoll_manager.cpp etcd_state_client.cpp messenger.cpp msgr_iothread.cpp ../util/addr_util.cpp
 	msgr_stop.cpp msgr_op.cpp msgr_send.cpp msgr_receive.cpp ../util/ringloop.cpp ../../json11/json11.cpp
 	http_client.cpp osd_ops.cpp pg_states.cpp ../util/timerfd_manager.cpp ../util/str_util.cpp ../util/json_util.cpp ${MSGR_RDMA} ${MSGR_RDMACM}
 )
@@ -52,9 +52,6 @@ if (${WITH_FIO})
 		../util/rw_blocking.cpp
 		../util/addr_util.cpp
 	)
-	target_link_libraries(fio_vitastor_sec
-		tcmalloc_minimal
-	)
 endif (${WITH_FIO})

 # vitastor-nbd
@@ -101,6 +98,7 @@ add_executable(test_cluster_client
 	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp cluster_client_wb.cpp msgr_op.cpp ../test/mock/messenger.cpp msgr_stop.cpp
 	etcd_state_client.cpp ../util/timerfd_manager.cpp ../util/addr_util.cpp ../util/str_util.cpp ../util/json_util.cpp ../../json11/json11.cpp
 )
+target_link_libraries(test_cluster_client ${LIBURING_LIBRARIES})
 target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)
 target_include_directories(test_cluster_client BEFORE PUBLIC ${CMAKE_SOURCE_DIR}/src/test/mock)
 add_dependencies(build_tests test_cluster_client)
@@ -590,7 +590,7 @@ void cluster_client_t::on_change_pool_config_hook()
        {
            if (log_level > 2 && pg_counts[pool_item.first])
            {
-                printf("Pool %u (%s) PG count changed from %lu to %lu\n", pool_item.first, pool_item.second.name.c_str(),
+                fprintf(stderr, "Pool %u (%s) PG count changed from %lu to %lu\n", pool_item.first, pool_item.second.name.c_str(),
                    pg_counts[pool_item.first], pool_item.second.real_pg_count);
            }
            // At this point, all pool operations should have been suspended
@@ -1119,6 +1119,13 @@ resume_2:
        // 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
+        if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
+        {
+            // Copy part bitmaps only after finishing all part reads
+            for (auto & part: op->parts)
+                if ((part.flags & (PART_SENT|PART_DONE|PART_VALID)) == (PART_SENT|PART_DONE|PART_VALID))
+                    copy_part_bitmap(op, &part);
+        }
        if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
        {
            // Check parent inode
@@ -1164,7 +1171,7 @@ resume_2:
        erase_op(op);
        return 1;
    }
-    else if (op->retval != 0 && !(op->flags & OP_FLUSH_BUFFER) &&
+    else if (op->retval != 0 && op->opcode != OSD_OP_SYNC && !(op->flags & OP_FLUSH_BUFFER) &&
        op->retval != -EPIPE && (op->retval != -EIO || !client_eio_retry_interval) && (op->retval != -ENOSPC || !client_retry_enospc))
    {
        // Fatal error (neither -EPIPE, -EIO nor -ENOSPC)
@@ -1630,13 +1637,6 @@ void cluster_client_t::handle_op_part(cluster_op_part_t *part)
        }
        if (op->inflight_count == 0 && !op->retry_after)
        {
-            // Copy part bitmaps only after finishing all part reads
-            if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP)
-            {
-                for (auto & part: op->parts)
-                    if (part.flags == (PART_SENT|PART_VALID|PART_DONE))
-                        copy_part_bitmap(op, &part);
-            }
            if (op->opcode == OSD_OP_SYNC)
                continue_sync(op);
            else
@@ -83,9 +83,6 @@ class writeback_cache_t;
 // FIXME: Split into public and private interfaces
 class __attribute__((visibility("default"))) cluster_client_t
 {
-#ifdef __MOCK__
-public:
-#endif
    timerfd_manager_t *tfd = NULL;
    ring_loop_t *ringloop = NULL;

@@ -155,15 +152,9 @@ public:
    void list_inode(inode_t inode, uint64_t min_offset, uint64_t max_offset, int max_parallel_pgs, std::function<void(
        int status, int pgs_left, pg_num_t pg_num, std::set<object_id>&& objects)> pg_callback);

-    //inline uint32_t get_bs_bitmap_granularity() { return st_cli.global_bitmap_granularity; }
-    //inline uint64_t get_bs_block_size() { return st_cli.global_block_size; }
-
-#ifndef __MOCK__
 protected:
-#endif
    void continue_ops(int time_passed = 0);

-protected:
    bool affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd);
    bool affects_pg(uint64_t inode, uint64_t offset, uint64_t len, pool_id_t pool_id, pg_num_t pg_num);

@@ -204,4 +195,5 @@ protected:
    osd_num_t select_nearest_osd(const std::vector<osd_num_t> & osds);

    friend class writeback_cache_t;
+    friend class cluster_client_test_t;
 };
@@ -88,6 +88,11 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flu
    // ...or just save it for writeback if write buffering is enabled
    if (op->len == 0)
    {
+        // FIXME: OSD_OP_DELETEs are currently only sent by vitastor-cli rm/rm-data and
+        // actually have len=0, because delete is actually a delete of the full object
+        // containing the requested offset, not a "punch hole" operation. But here, writeback
+        // cache assumes it IS a "punch hole" operation. I should select one of these
+        // approaches and fix everything accordingly when I decide to implement TRIM.
        return;
    }
    auto dirty_it = find_dirty(op->inode, op->offset);
@@ -244,12 +249,13 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flu
            writeback_queue_size--;
        }
    }
-    if (!is_del)
+    if (!is_del && op->len > 0)
    {
        uint64_t pos = 0, len = op->len, iov_idx = 0;
-        while (len > 0 && iov_idx < op->iov.count)
+        while (iov_idx < op->iov.count)
        {
            auto & iov = op->iov.buf[iov_idx];
+            assert(pos + iov.iov_len <= len);
            memcpy(buf + pos, iov.iov_base, iov.iov_len);
            pos += iov.iov_len;
            iov_idx++;
@@ -443,7 +449,7 @@ void writeback_cache_t::start_writebacks(cluster_client_t *cli, int count)
        started++;
        assert(writeback_queue_size > 0);
        writeback_queue_size--;
-        writeback_bytes -= off - from_it->first.stripe;
+        writeback_bytes -= (is_del ? 0 : off - from_it->first.stripe);
        assert(writeback_queue_size > 0 || !writeback_bytes);
        flush_buffers(cli, from_it, to_it);
    }
@@ -1185,7 +1185,6 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
                if (i >= pg_state_bit_count)
                {
                    fprintf(stderr, "Unexpected pool %u PG %u state keyword in etcd: %s\n", pool_id, pg_num, e.dump().c_str());
-                    return;
                }
            }
            if (!cur_primary || !value["state"].is_array() || !state ||
@@ -1194,7 +1193,6 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
                (state & PG_INCOMPLETE) && state != PG_INCOMPLETE && state != (PG_INCOMPLETE|PG_HAS_INVALID))
            {
                fprintf(stderr, "Unexpected pool %u PG %u state in etcd: primary=%ju, state=%s\n", pool_id, pg_num, cur_primary, value["state"].dump().c_str());
-                return;
            }
            pg_cfg.cur_primary = cur_primary;
            pg_cfg.cur_state = state;
@@ -15,106 +15,6 @@
 #include "msgr_rdma.h"
 #endif

-#include <sys/poll.h>
-
-msgr_iothread_t::msgr_iothread_t():
-    ring(RINGLOOP_DEFAULT_SIZE, true),
-    thread(&msgr_iothread_t::run, this)
-{
-    eventfd = ring.register_eventfd();
-    if (eventfd < 0)
-    {
-        throw std::runtime_error(std::string("failed to register eventfd: ") + strerror(-eventfd));
-    }
-}
-
-msgr_iothread_t::~msgr_iothread_t()
-{
-    stop();
-}
-
-void msgr_iothread_t::add_sqe(io_uring_sqe & sqe)
-{
-    mu.lock();
-    queue.push_back((iothread_sqe_t){ .sqe = sqe, .data = std::move(*(ring_data_t*)sqe.user_data) });
-    if (queue.size() == 1)
-    {
-        cond.notify_all();
-    }
-    mu.unlock();
-}
-
-void msgr_iothread_t::stop()
-{
-    mu.lock();
-    if (stopped)
-    {
-        mu.unlock();
-        return;
-    }
-    stopped = true;
-    if (outer_loop_data)
-    {
-        outer_loop_data->callback = [](ring_data_t*){};
-    }
-    cond.notify_all();
-    close(eventfd);
-    mu.unlock();
-    thread.join();
-}
-
-void msgr_iothread_t::add_to_ringloop(ring_loop_t *outer_loop)
-{
-    assert(!this->outer_loop || this->outer_loop == outer_loop);
-    io_uring_sqe *sqe = outer_loop->get_sqe();
-    assert(sqe != NULL);
-    this->outer_loop = outer_loop;
-    this->outer_loop_data = ((ring_data_t*)sqe->user_data);
-    io_uring_prep_poll_add(sqe, eventfd, POLLIN);
-    outer_loop_data->callback = [this](ring_data_t *data)
-    {
-        if (data->res < 0)
-        {
-            throw std::runtime_error(std::string("eventfd poll failed: ") + strerror(-data->res));
-        }
-        outer_loop_data = NULL;
-        if (stopped)
-        {
-            return;
-        }
-        add_to_ringloop(this->outer_loop);
-        ring.loop();
-    };
-}
-
-void msgr_iothread_t::run()
-{
-    while (true)
-    {
-        {
-            std::unique_lock<std::mutex> lk(mu);
-            while (!stopped && !queue.size())
-                cond.wait(lk);
-            if (stopped)
-                return;
-            int i = 0;
-            for (; i < queue.size(); i++)
-            {
-                io_uring_sqe *sqe = ring.get_sqe();
-                if (!sqe)
-                    break;
-                ring_data_t *data = ((ring_data_t*)sqe->user_data);
-                *data = std::move(queue[i].data);
-                *sqe = queue[i].sqe;
-                sqe->user_data = (uint64_t)data;
-            }
-            queue.erase(queue.begin(), queue.begin()+i);
-        }
-        // We only want to offload sendmsg/recvmsg. Callbacks will be called in main thread
-        ring.submit();
-    }
-}
-
 void osd_messenger_t::init()
 {
 #ifdef WITH_RDMACM
@@ -173,12 +73,7 @@ void osd_messenger_t::init()
    }
    if (ringloop && iothread_count > 0)
    {
-        for (int i = 0; i < iothread_count; i++)
-        {
-            auto iot = new msgr_iothread_t();
-            iothreads.push_back(iot);
-            iot->add_to_ringloop(ringloop);
-        }
+        init_iothreads();
    }
    keepalive_timer_id = tfd->set_timer(1000, true, [this](int)
    {
@@ -220,7 +115,7 @@ void osd_messenger_t::init()
                            .opcode = OSD_OP_PING,
                        },
                    };
-                    op->callback = [this, cl](osd_op_t *op)
+                    op->callback = [this](osd_op_t *op)
                    {
                        auto cl_it = clients.find(op->client_id);
                        if (cl_it == clients.end())
@@ -229,6 +124,7 @@ void osd_messenger_t::init()
                            delete op;
                            return;
                        }
+                        auto cl = cl_it->second;
                        uint64_t fail_client_id = (op->reply.hdr.retval != 0 ? op->client_id : 0);
                        auto fail_osd_num = cl->in_osd_num ? cl->in_osd_num : cl->osd_num;
                        cl->ping_time_remaining = 0;
@@ -271,14 +167,7 @@ osd_messenger_t::~osd_messenger_t()
    {
        stop_client(clients.begin()->first, true);
    }
-    if (iothreads.size())
-    {
-        for (auto iot: iothreads)
-        {
-            delete iot;
-        }
-        iothreads.clear();
-    }
+    destroy_iothreads();
 #ifdef WITH_RDMA
    for (auto rdma_context: rdma_contexts)
    {
@@ -129,43 +129,7 @@ struct osd_op_stats_t
    uint64_t subop_stat_count[OSD_OP_MAX+1] = { 0 };
 };

-#include <mutex>
-#include <condition_variable>
-#include <thread>
-
-#ifdef __MOCK__
 class msgr_iothread_t;
-#else
-struct iothread_sqe_t
-{
-    io_uring_sqe sqe;
-    ring_data_t data;
-};
-
-class msgr_iothread_t
-{
-protected:
-    ring_loop_t ring;
-    ring_loop_t *outer_loop = NULL;
-    ring_data_t *outer_loop_data = NULL;
-    int eventfd = -1;
-    bool stopped = false;
-    std::mutex mu;
-    std::condition_variable cond;
-    std::vector<iothread_sqe_t> queue;
-    std::thread thread;
-
-    void run();
-public:
-
-    msgr_iothread_t();
-    ~msgr_iothread_t();
-
-    void add_sqe(io_uring_sqe & sqe);
-    void stop();
-    void add_to_ringloop(ring_loop_t *outer_loop);
-};
-#endif

 #ifdef WITH_RDMA
 struct rdma_event_channel;
@@ -215,7 +179,7 @@ protected:

 public:
    timerfd_manager_t *tfd = NULL;
-    ring_loop_t *ringloop = NULL;
+    ring_loop_i *ringloop = NULL;
    bool has_sendmsg_zc = false;
    // osd_num_t is only for logging and asserts
    uint64_t next_client_id = 1;
@@ -234,6 +198,7 @@ public:
    osd_op_stats_t stats, recovery_stats;

    void init();
+    void init_iothreads();
    void parse_config(const json11::Json & config);
    void connect_peer(uint64_t osd_num, json11::Json peer_state);
    void stop_client(uint64_t client_id, bool force_delete = false);
@@ -246,6 +211,7 @@ public:
    void read_requests();
    void send_replies();
    void accept_connections(int listen_fd);
+    void destroy_iothreads();
    ~osd_messenger_t();

    static json11::Json::object read_config(const json11::Json & config);
@@ -293,6 +259,7 @@ protected:
    bool init_recv_rdma(osd_client_t *cl);
    void handle_rdma_events(msgr_rdma_context_t *rdma_context);
    msgr_rdma_context_t* choose_rdma_context(osd_client_t *cl);
+    void destroy_rdma_conn(msgr_rdma_connection_t *rdma_conn);
 #endif
 #ifdef WITH_RDMACM
    void handle_rdmacm_events();
@@ -0,0 +1,129 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <stdexcept>
+#include <sys/poll.h>
+#include <unistd.h>
+
+#include "messenger.h"
+#include "msgr_iothread.h"
+
+msgr_iothread_t::msgr_iothread_t():
+    ring(RINGLOOP_DEFAULT_SIZE, true),
+    thread(&msgr_iothread_t::run, this)
+{
+    eventfd = ring.register_eventfd();
+    if (eventfd < 0)
+    {
+        throw std::runtime_error(std::string("failed to register eventfd: ") + strerror(-eventfd));
+    }
+}
+
+msgr_iothread_t::~msgr_iothread_t()
+{
+    stop();
+}
+
+void msgr_iothread_t::add_sqe(io_uring_sqe & sqe)
+{
+    mu.lock();
+    queue.push_back((iothread_sqe_t){ .sqe = sqe, .data = std::move(*(ring_data_t*)sqe.user_data) });
+    if (queue.size() == 1)
+    {
+        cond.notify_all();
+    }
+    mu.unlock();
+}
+
+void msgr_iothread_t::stop()
+{
+    mu.lock();
+    if (stopped)
+    {
+        mu.unlock();
+        return;
+    }
+    stopped = true;
+    if (outer_loop_data)
+    {
+        outer_loop_data->callback = [](ring_data_t*){};
+    }
+    cond.notify_all();
+    close(eventfd);
+    mu.unlock();
+    thread.join();
+}
+
+void msgr_iothread_t::add_to_ringloop(ring_loop_i *outer_loop)
+{
+    assert(!this->outer_loop || this->outer_loop == outer_loop);
+    io_uring_sqe *sqe = outer_loop->get_sqe();
+    assert(sqe != NULL);
+    this->outer_loop = outer_loop;
+    this->outer_loop_data = ((ring_data_t*)sqe->user_data);
+    io_uring_prep_poll_add(sqe, eventfd, POLLIN);
+    outer_loop_data->callback = [this](ring_data_t *data)
+    {
+        if (data->res < 0)
+        {
+            throw std::runtime_error(std::string("eventfd poll failed: ") + strerror(-data->res));
+        }
+        outer_loop_data = NULL;
+        if (stopped)
+        {
+            return;
+        }
+        add_to_ringloop(this->outer_loop);
+        ring.loop();
+    };
+}
+
+void msgr_iothread_t::run()
+{
+    while (true)
+    {
+        {
+            std::unique_lock<std::mutex> lk(mu);
+            while (!stopped && !queue.size())
+                cond.wait(lk);
+            if (stopped)
+                return;
+            int i = 0;
+            for (; i < queue.size(); i++)
+            {
+                io_uring_sqe *sqe = ring.get_sqe();
+                if (!sqe)
+                    break;
+                ring_data_t *data = ((ring_data_t*)sqe->user_data);
+                *data = std::move(queue[i].data);
+                *sqe = queue[i].sqe;
+                sqe->user_data = (uint64_t)data;
+            }
+            queue.erase(queue.begin(), queue.begin()+i);
+        }
+        // We only want to offload sendmsg/recvmsg. Callbacks will be called in main thread
+        ring.submit();
+    }
+}
+
+void osd_messenger_t::init_iothreads()
+{
+    for (int i = 0; i < iothread_count; i++)
+    {
+        auto iot = new msgr_iothread_t();
+        iothreads.push_back(iot);
+        iot->add_to_ringloop(ringloop);
+    }
+}
+
+void osd_messenger_t::destroy_iothreads()
+{
+    if (iothreads.size())
+    {
+        for (auto iot: iothreads)
+        {
+            delete iot;
+        }
+        iothreads.clear();
+    }
+}
@@ -0,0 +1,38 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <mutex>
+#include <condition_variable>
+#include <thread>
+
+#include "ringloop.h"
+
+struct iothread_sqe_t
+{
+    io_uring_sqe sqe;
+    ring_data_t data;
+};
+
+class msgr_iothread_t
+{
+protected:
+    ring_loop_t ring;
+    ring_loop_i *outer_loop = NULL;
+    ring_data_t *outer_loop_data = NULL;
+    int eventfd = -1;
+    bool stopped = false;
+    std::mutex mu;
+    std::condition_variable cond;
+    std::vector<iothread_sqe_t> queue;
+    std::thread thread;
+
+    void run();
+public:
+
+    msgr_iothread_t();
+    ~msgr_iothread_t();
+
+    void add_sqe(io_uring_sqe & sqe);
+    void stop();
+    void add_to_ringloop(ring_loop_i *outer_loop);
+};
@@ -157,6 +157,7 @@ struct __attribute__((visibility("default"))) osd_op_t
    timespec tv_begin = { 0 }, tv_end = { 0 };
    uint64_t op_type = OSD_OP_IN;
    uint64_t client_id = 0;
+    osd_num_t osd_num = 0;
    osd_any_op_t req;
    osd_any_reply_t reply;
    blockstore_op_t *bs_op = NULL;
@@ -187,6 +187,8 @@ std::vector<msgr_rdma_context_t*> msgr_rdma_context_t::create_all(const std::vec
    ibv_device **raw_dev_list = NULL;
    ibv_device **dev_list = NULL;
    ibv_device *single_list[2] = {};
+    int up_ports = 0;
+    int single_port_num = 0;

    raw_dev_list = dev_list = ibv_get_device_list(NULL);
    if (!dev_list || !*dev_list)
@@ -221,6 +223,7 @@ std::vector<msgr_rdma_context_t*> msgr_rdma_context_t::create_all(const std::vec
        dev_list = single_list;
    }

+retry:
    for (int i = 0; dev_list[i]; ++i)
    {
        auto dev = dev_list[i];
@@ -258,6 +261,9 @@ std::vector<msgr_rdma_context_t*> msgr_rdma_context_t::create_all(const std::vec
                fprintf(stderr, "RDMA device %s port %d GID %d does not exist\n", ibv_get_device_name(dev), port_num, sel_gid_index);
                continue;
            }
+            up_ports++;
+            single_port_num = port_num;
+            single_list[0] = dev;
            uint32_t port_mtu = sel_mtu ? sel_mtu : ibv_mtu_to_bytes(portinfo.active_mtu);
 #ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
            if (sel_gid_index < 0)
@@ -298,6 +304,14 @@ cleanup_dev:
        ibv_close_device(context);
    }

+    if (!ret.size() && up_ports == 1 && dev_list != single_list)
+    {
+        // Auto-select the only available device/port if there is only one
+        dev_list = single_list;
+        sel_port_num = single_port_num;
+        goto retry;
+    }
+
 cleanup:
    if (raw_dev_list)
        ibv_free_device_list(raw_dev_list);
@@ -598,7 +612,9 @@ void osd_messenger_t::try_send_rdma(osd_client_t *cl)
    while (!rc->send_out_full && copied > 0 && rc->cur_send < rc->max_send)
    {
        dst = (uint8_t*)rc->send_out.buf + rc->send_out_pos;
-        dst_len = (rc->send_out_pos < rc->send_out_size ? rc->send_out_size-rc->send_out_pos : rc->send_done_pos-rc->send_out_pos);
+        dst_len = (rc->send_out_pos >= rc->send_done_pos
+            ? rc->send_out_size-rc->send_out_pos
+            : rc->send_done_pos-rc->send_out_pos);
        if (dst_len > rc->max_msg)
            dst_len = rc->max_msg;
        copied = try_send_rdma_copy(cl, dst, dst_len);
@@ -608,7 +624,7 @@ void osd_messenger_t::try_send_rdma(osd_client_t *cl)
            if (rc->send_out_pos == rc->send_out_size)
                rc->send_out_pos = 0;
            assert(rc->send_out_pos < rc->send_out_size);
-            if (rc->send_out_pos >= rc->send_done_pos)
+            if (rc->send_out_pos == rc->send_done_pos)
                rc->send_out_full = true;
            ibv_sge sge = {
                .addr = (uintptr_t)dst,
@@ -778,3 +794,16 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
    } while (event_count > 0);
    handle_immediate_ops();
 }
+
+void osd_messenger_t::destroy_rdma_conn(msgr_rdma_connection_t *rdma_conn)
+{
+    if (rdma_conn->cmid)
+    {
+        auto rdma_it = rdmacm_connections.find(rdma_conn->cmid);
+        if (rdma_it != rdmacm_connections.end() && rdma_it->second->rdma_conn == rdma_conn)
+        {
+            rdmacm_connections.erase(rdma_it);
+        }
+    }
+    delete rdma_conn;
+}
@@ -2,6 +2,7 @@
 // License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)

 #include "messenger.h"
+#include "msgr_iothread.h"

 void osd_messenger_t::read_requests()
 {
@@ -348,7 +349,7 @@ void osd_messenger_t::handle_op_hdr(osd_client_t *cl)
 bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
 {
    auto req_it = cl->sent_ops.find(cl->read_op->req.hdr.id);
-    if (req_it == cl->sent_ops.end())
+    if (req_it == cl->sent_ops.end() || req_it->second->req.hdr.opcode != cl->read_op->req.hdr.opcode)
    {
        // Command out of sync. Drop connection
        fprintf(stderr, "Client %ju command out of sync: id %ju\n", cl->client_id, cl->read_op->req.hdr.id);
@@ -372,7 +373,7 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
            stop_client(cl->client_id);
            return false;
        }
-        if (bmp_len > 0)
+        if (op->reply.hdr.retval >= 0 && bmp_len > 0)
        {
            assert(op->bitmap);
            cl->recv_list.push_back(op->bitmap, bmp_len);
@@ -6,6 +6,7 @@
 #include <sys/epoll.h>

 #include "messenger.h"
+#include "msgr_iothread.h"

 void osd_messenger_t::outbox_push(osd_op_t *cur_op)
 {
@@ -99,10 +100,15 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
    if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
    {
        if (cur_op->op_type == OSD_OP_IN && cur_op->reply.hdr.retval > 0)
+        {
            to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->reply.hdr.retval });
+            to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        }
        else if (cur_op->op_type == OSD_OP_OUT && cur_op->req.sec_read_bmp.len > 0)
+        {
            to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len });
-        to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+            to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        }
    }
    if (cur_op->op_type == OSD_OP_IN)
    {
@@ -5,9 +5,6 @@
 #include <assert.h>

 #include "messenger.h"
-#ifdef WITH_RDMA
-#include "msgr_rdma.h"
-#endif

 void osd_client_t::cancel_ops()
 {
@@ -92,23 +89,11 @@ void osd_messenger_t::stop_client(uint64_t client_id, bool force_delete)
            osd_peers.erase(osd_it);
        }
    }
-#ifdef WITH_RDMA
-    if (cl->rdma_conn && cl->rdma_conn->cmid)
-    {
-        auto rdma_it = rdmacm_connections.find(cl->rdma_conn->cmid);
-        if (rdma_it != rdmacm_connections.end() && rdma_it->second == cl)
-        {
-            rdmacm_connections.erase(rdma_it);
-        }
-    }
-#endif
-#ifndef __MOCK__
    if (cl->connect_timeout_id >= 0)
    {
        tfd->clear_timer(cl->connect_timeout_id);
        cl->connect_timeout_id = -1;
    }
-#endif
    if (cl->in_osd_num && break_pg_locks)
    {
        // Break PG locks
@@ -143,9 +128,7 @@ void osd_messenger_t::destroy_client(osd_client_t *cl)
    clients.erase(cl->client_id);
    if (cl->peer_fd >= 0)
    {
-#ifndef __MOCK__
        tfd->set_fd_handler(cl->peer_fd, false, NULL);
-#endif
        for (auto rit = read_ready_clients.begin(); rit != read_ready_clients.end(); rit++)
        {
            if (*rit == cl->client_id)
@@ -164,6 +147,13 @@ void osd_messenger_t::destroy_client(osd_client_t *cl)
        }
        clients_by_fd.erase(cl->peer_fd);
    }
+#ifdef WITH_RDMA
+    if (cl->rdma_conn)
+    {
+        destroy_rdma_conn(cl->rdma_conn);
+        cl->rdma_conn = NULL;
+    }
+#endif
    delete cl;
 }

@@ -196,13 +186,4 @@ osd_client_t::~osd_client_t()
            delete op;
        }
    }
-#ifndef __MOCK__
-#ifdef WITH_RDMA
-    if (rdma_conn)
-    {
-        delete rdma_conn;
-        rdma_conn = NULL;
-    }
-#endif
-#endif
 }
@@ -1049,7 +1049,7 @@ static int coroutine_fn vitastor_co_block_status(BlockDriverState *bs,
    {
        // 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;
+        uint64_t bmp_end = (offset+bytes-task.offset) / task.bitmap_granularity;
        while (bmp_pos < bmp_end)
        {
            if (!(bmp_pos & 7) && bmp_end >= bmp_pos+8)
@@ -282,7 +282,7 @@ help:
                exit(1);
            }
        }
-        const bool writeback = cli->get_immediate_commit(inode);
+        const bool writeback = !cli->get_immediate_commit(inode);
        auto pool_it = cli->st_cli.pool_config.find(INODE_POOL(inode ? inode : watch->cfg.num));
        if (pool_it == cli->st_cli.pool_config.end())
        {
@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@

 Name: Vitastor
 Description: Vitastor client library
-Version: 3.0.8
+Version: 3.0.12
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}

@@ -27,6 +27,7 @@ add_library(vitastor_cli STATIC
 	cli_pool_ls.cpp
 	cli_pool_modify.cpp
 	cli_pool_rm.cpp
+	cli_raw_ls.cpp
 )
 target_compile_options(vitastor_cli PUBLIC -fPIC)

@@ -126,6 +126,11 @@ static const char* help_text =
    "  --min-offset, --max-offset\n"
    "      Restrict listing to specific offsets inside inodes.\n"
    "\n"
+    "vitastor-cli raw-ls [OPTIONS]\n"
+    "  Find object(s) in the cluster using raw secondary listing operations. Options:\n"
+    "  [--min_inode NUM] [--max_inode NUM] [--offset NUM] [--pg_num NUM] [--pg_count COUNT]\n"
+    "  [--pg_stripe_size NUM] [--osds 1,2,3,...]\n"
+    "\n"
    "vitastor-cli fix [--objects <objects>] [--bad-osds <osds>] [--part <part>] [--check no]\n"
    "  Fix inconsistent objects in the cluster by deleting some copies.\n"
    "  --objects <objects>\n"
@@ -459,6 +464,11 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
        // Describe unclean objects
        action_cb = p->start_describe(cfg);
    }
+    else if (cmd[0] == "raw-ls")
+    {
+        // Run raw listings
+        action_cb = p->start_raw_ls(cfg);
+    }
    else if (cmd[0] == "fix")
    {
        // Fix inconsistent objects (by deleting some copies)
@@ -62,6 +62,7 @@ public:
    std::function<bool(cli_result_t &)> start_fix(json11::Json);
    std::function<bool(cli_result_t &)> start_flatten(json11::Json);
    std::function<bool(cli_result_t &)> start_ls(json11::Json);
+    std::function<bool(cli_result_t &)> start_raw_ls(json11::Json cfg);
    std::function<bool(cli_result_t &)> start_merge(json11::Json);
    std::function<bool(cli_result_t &)> start_modify(json11::Json);
    std::function<bool(cli_result_t &)> start_modify_osd(json11::Json);
@@ -0,0 +1,212 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "cli_fix.h"
+#include "cluster_client.h"
+#include "pg_states.h"
+#include "str_util.h"
+#include "json_util.h"
+
+std::vector<uint64_t> parse_uint64_list(json11::Json val);
+
+// Find object(s) in the cluster using raw secondary listing operations.
+// Only for troubleshooting.
+struct cli_raw_ls_t
+{
+    json11::Json cfg;
+    pool_id_t pool_id = 0;
+    pg_num_t pg_num = 0;
+    pg_num_t pg_count = 0;
+    uint32_t pg_stripe_size = 0;
+    std::set<uint64_t> osds;
+    std::vector<osd_num_t> osd_list;
+    uint64_t min_inode = 0, max_inode = 0, min_offset = 0, max_offset = 0;
+    bool offset_set = false;
+
+    cli_tool_t *parent = NULL;
+    int state = 0;
+    cli_result_t result;
+    bool first = true;
+    size_t osd_pos = 0;
+
+    bool is_done()
+    {
+        return state == 100;
+    }
+
+    void loop()
+    {
+        if (state == 1)
+            goto resume_1;
+        if (state == 100)
+            return;
+        pool_id = cfg["pool_id"].uint64_value();
+        pg_num = (pg_num_t)cfg["pg_num"].uint64_value();
+        pg_count = (pg_num_t)cfg["pg_count"].uint64_value();
+        pg_stripe_size = cfg["pg_stripe_size"].uint64_value();
+        if (!pool_id)
+        {
+            result = (cli_result_t){ .err = EINVAL, .text = "Pool ID is required" };
+            state = 100;
+            return;
+        }
+        if (cfg["min_inode"].is_null() && cfg["max_inode"].is_null())
+        {
+            min_inode = INODE_WITH_POOL(cfg["pool_id"].uint64_value(), 0);
+            max_inode = INODE_WITH_POOL(cfg["pool_id"].uint64_value(), UINT64_MAX);
+        }
+        else
+        {
+            min_inode = INODE_WITH_POOL(cfg["pool_id"].uint64_value(), stoull_full(cfg["min_inode"].as_string()));
+            max_inode = INODE_WITH_POOL(cfg["pool_id"].uint64_value(), stoull_full(cfg["max_inode"].as_string()));
+        }
+        if (cfg["offset"].is_null())
+        {
+            if (!pg_num)
+            {
+                result = (cli_result_t){ .err = EINVAL, .text = "PG number is required if offset is not specified" };
+                state = 100;
+                return;
+            }
+            min_offset = 0;
+            max_offset = UINT64_MAX;
+        }
+        else
+        {
+            min_offset = max_offset = stoull_full(cfg["offset"].as_string());
+            offset_set = true;
+        }
+        for (auto osd_num: parse_uint64_list(cfg["osds"]))
+        {
+            osds.insert(osd_num);
+        }
+        if (!pg_count || !pg_stripe_size || !osds.size())
+        {
+            auto pool_it = parent->cli->st_cli.pool_config.find(pool_id);
+            if (pool_it == parent->cli->st_cli.pool_config.end())
+            {
+                result = (cli_result_t){ .err = EINVAL, .text = "pg_count, pg_stripe_size and osds are required if the pool does not exist" };
+                state = 100;
+                return;
+            }
+            if (!pg_count)
+            {
+                pg_count = pool_it->second.real_pg_count;
+            }
+            if (!pg_stripe_size)
+            {
+                pg_stripe_size = pool_it->second.pg_stripe_size;
+            }
+            if (!osds.size())
+            {
+                for (auto & pgp: pool_it->second.pg_config)
+                {
+                    for (osd_num_t pg_osd: pgp.second.target_set)
+                        osds.insert(pg_osd);
+                    for (osd_num_t pg_osd: pgp.second.all_peers)
+                        osds.insert(pg_osd);
+                    for (auto & hist_item: pgp.second.target_history)
+                        for (auto pg_osd: hist_item)
+                            osds.insert(pg_osd);
+                }
+                osds.erase(0);
+            }
+        }
+        if (offset_set)
+        {
+            min_offset = (min_offset / pg_stripe_size) * pg_stripe_size;
+            max_offset = ((max_offset / pg_stripe_size) * pg_stripe_size) | STRIPE_MASK;
+            pg_num = (min_offset/pg_stripe_size) % pg_count + 1; // like map_to_pg()
+            fprintf(stderr, "Selected PG %u\n", pg_num);
+        }
+        parent->cli->init_msgr();
+        osd_list = std::vector<osd_num_t>(osds.begin(), osds.end());
+        osd_pos = 0;
+        state = 1;
+    resume_1:
+        for (; osd_pos < osd_list.size() && parent->waiting < parent->parallel_osds; osd_pos++)
+        {
+            uint64_t osd_num = osd_list[osd_pos];
+            if (parent->cli->st_cli.peer_states[osd_num].is_null())
+            {
+                fprintf(stderr, "OSD %ju is unavailable, skipping\n", osd_num);
+                continue;
+            }
+            osd_op_t *op = new osd_op_t;
+            op->req = (osd_any_op_t){
+                .sec_list = {
+                    .header = {
+                        .magic = SECONDARY_OSD_OP_MAGIC,
+                        .opcode = OSD_OP_SEC_LIST,
+                    },
+                    .list_pg = pg_num,
+                    .pg_count = pg_count,
+                    .pg_stripe_size = pg_stripe_size,
+                    .min_inode = min_inode,
+                    .max_inode = max_inode,
+                    .min_stripe = min_offset,
+                    .max_stripe = max_offset,
+                },
+            };
+            op->callback = [this, osd_num](osd_op_t *op)
+            {
+                if (op->reply.hdr.retval < 0)
+                {
+                    fprintf(stderr, "OSD %ju listing failed: retval=%jd\n", osd_num, op->reply.hdr.retval);
+                }
+                else
+                {
+                    for (uint64_t i = 0; i < op->reply.hdr.retval; i++)
+                    {
+                        auto & ov = ((obj_ver_id*)op->buf)[i];
+                        if (parent->json_output)
+                        {
+                            printf("%s{\"osd\":%ju,\"inode\":\"0x%jx\",\"stripe\":\"0x%jx\",\"version\":%ju,\"stable\":%s}",
+                                first ? "" : ",\n", osd_num, ov.oid.inode, ov.oid.stripe, ov.version,
+                                i < op->reply.sec_list.stable_count ? "true" : "false");
+                            first = false;
+                        }
+                        else
+                        {
+                            printf("OSD %ju - %jx:%jx v%ju%s\n",
+                                osd_num, ov.oid.inode, ov.oid.stripe, ov.version,
+                                i < op->reply.sec_list.stable_count ? " stable" : "");
+                        }
+                    }
+                }
+                parent->waiting--;
+                loop();
+                delete op;
+            };
+            parent->waiting++;
+            parent->cli->execute_raw(osd_num, op);
+        }
+        if (parent->waiting > 0)
+        {
+            return;
+        }
+        if (!first)
+        {
+            printf("\n");
+        }
+        state = 100;
+    }
+};
+
+std::function<bool(cli_result_t &)> cli_tool_t::start_raw_ls(json11::Json cfg)
+{
+    auto raw_ls = new cli_raw_ls_t();
+    raw_ls->parent = this;
+    raw_ls->cfg = cfg;
+    return [raw_ls](cli_result_t & result)
+    {
+        raw_ls->loop();
+        if (raw_ls->is_done())
+        {
+            result = raw_ls->result;
+            delete raw_ls;
+            return true;
+        }
+        return false;
+    };
+}
@@ -286,6 +286,8 @@ struct rm_inode_t
                    .data = data,
                };
            }
+            // Wakeup callers (otherwise they wake up only on next loop())
+            parent->ringloop->wakeup();
        }
        in_continue = false;
    }
@@ -11,7 +11,6 @@ add_executable(vitastor-disk
 	../blockstore/blockstore_disk.cpp ../blockstore/blockstore_heap.cpp ../blockstore/multilist.cpp
 )
 target_link_libraries(vitastor-disk
-	tcmalloc_minimal
 	${LIBURING_LIBRARIES}
 	${ISAL_LIBRARIES}
 )
@@ -427,6 +427,9 @@ int main(int argc, char *argv[])
            self.dsk.open_journal();
            self.dsk.calc_lengths();
            self.dsk.close_all();
+            self.new_meta_device = self.dsk.meta_device;
+            self.new_meta_offset = self.dsk.meta_offset;
+            self.new_meta_len = self.dsk.meta_area_size;
        }
        std::string json_err;
        json11::Json meta = json11::Json::parse(read_all_fd(0), json_err);
@@ -52,7 +52,6 @@ struct disk_tool_t
    bool all = false, json = false, now = false;
    bool dump_with_blocks = false, dump_with_data = false;
    bool dump_as_old = false;
-    bool skip_obsolete = false;
    int log_level = 1;
    double meta_reserve_multiple = 2;
    uint64_t meta_reserve_min_size = (uint64_t)1024*1024*1024;
@@ -138,8 +137,6 @@ struct disk_tool_t
    int resize_write_new_journal();
    void remap_big_write(heap_entry_t *wr);
    void remap_small_write(heap_entry_t *wr);
-    void fill_old_clean_entry(blockstore_heap_t *heap, heap_entry_t *big_wr);
-    void fill_old_journal_entry(blockstore_heap_t *heap, heap_entry_t *wr);
    int resize_rebuild_meta();
    int resize_write_new_meta();
    void free_new_meta();
@@ -589,6 +589,7 @@ int disk_tool_t::write_json_meta(json11::Json meta)

 int disk_tool_t::write_json_heap(json11::Json meta, json11::Json journal)
 {
+    assert(new_meta_len >= sizeof(blockstore_meta_header_v3_t));
    new_meta_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_meta_len);
    memset(new_meta_buf, 0, new_meta_len);
    new_meta_hdr = (blockstore_meta_header_v3_t *)new_meta_buf;
@@ -606,9 +607,9 @@ int disk_tool_t::write_json_heap(json11::Json meta, json11::Json journal)
        : (meta["data_csum_type"].string_value() == "crc32c"
            ? BLOCKSTORE_CSUM_CRC32C
            : BLOCKSTORE_CSUM_NONE);
+    new_meta_hdr->meta_area_size = new_meta_len;
    new_meta_hdr->csum_block_size = meta["csum_block_size"].uint64_value();
    new_meta_hdr->header_csum = crc32c(0, new_meta_hdr, sizeof(blockstore_meta_header_v3_t));
-    new_meta_hdr->meta_area_size = new_meta_len;
    new_clean_entry_bitmap_size = (new_meta_hdr->data_block_size / new_meta_hdr->bitmap_granularity + 7) / 8;
    new_clean_entry_size = 0;
    new_entries_per_block = 0;
@@ -111,7 +111,7 @@ int disk_tool_t::raw_resize()
    fprintf(stderr, "Done\n");
 ret:
    free_new_meta();
-    return 0;
+    return r;
 }

 int disk_tool_t::resize_parse_params()
@@ -154,9 +154,6 @@ int disk_tool_t::resize_parse_params()
        ? parse_size(options["new_journal_offset"]) : dsk.journal_offset;
    new_journal_len = options.find("new_journal_len") != options.end()
        ? parse_size(options["new_journal_len"]) : dsk.journal_len;
-    new_meta_format = options.find("new_meta_format") != options.end()
-        ? stoull_full(options["new_meta_format"]) : 0;
-    skip_obsolete = options.find("skip_obsolete") != options.end();
    if (new_data_len+new_data_offset > dsk.data_device_size)
        new_data_len = dsk.data_device_size-new_data_offset;
    if (new_meta_device == dsk.data_device && new_data_offset < new_meta_offset &&
@@ -205,10 +202,7 @@ void disk_tool_t::resize_init(blockstore_meta_header_v3_t *hdr)
    {
        dsk.meta_format = hdr->version;
    }
-    if (new_meta_format == 0)
-    {
-        new_meta_format = hdr && hdr->version == BLOCKSTORE_META_FORMAT_HEAP ? BLOCKSTORE_META_FORMAT_HEAP : BLOCKSTORE_META_FORMAT_V2;
-    }
+    new_meta_format = hdr && hdr->version == BLOCKSTORE_META_FORMAT_HEAP ? BLOCKSTORE_META_FORMAT_HEAP : BLOCKSTORE_META_FORMAT_V2;
    dsk.calc_lengths();
    if (((new_data_offset-dsk.data_offset) % dsk.data_block_size))
    {
@@ -563,66 +557,6 @@ void disk_tool_t::remap_small_write(heap_entry_t *wr)
    }
 }

-void disk_tool_t::fill_old_clean_entry(blockstore_heap_t *heap, heap_entry_t *big_wr)
-{
-    uint64_t block_num = big_wr->big().block_num;
-    clean_disk_entry *new_entry = (clean_disk_entry*)(new_meta_buf + dsk.meta_block_size +
-        dsk.meta_block_size*(block_num / new_entries_per_block) +
-        new_clean_entry_size*(block_num % new_entries_per_block));
-    new_entry->oid = (object_id){ .inode = big_wr->inode, .stripe = big_wr->stripe };
-    new_entry->version = big_wr->version;
-    memcpy(new_entry->bitmap, big_wr->get_ext_bitmap(heap), new_clean_entry_bitmap_size);
-    memcpy(new_entry->bitmap + new_clean_entry_bitmap_size, big_wr->get_int_bitmap(heap), new_clean_entry_bitmap_size);
-    memcpy(new_entry->bitmap + 2*new_clean_entry_bitmap_size, big_wr->get_checksums(heap), new_data_csum_size);
-    uint32_t *new_entry_csum = (uint32_t*)(((uint8_t*)new_entry) + new_clean_entry_size - 4);
-    *new_entry_csum = crc32c(0, new_entry, new_clean_entry_size - 4);
-}
-
-void disk_tool_t::fill_old_journal_entry(blockstore_heap_t *heap, heap_entry_t *wr)
-{
-    assert(wr->type() == BS_HEAP_SMALL_WRITE ||
-        wr->type() == BS_HEAP_BIG_WRITE ||
-        wr->type() == BS_HEAP_BIG_INTENT);
-    uint32_t je_size = ((wr->entry_type & BS_HEAP_TYPE) == BS_HEAP_SMALL_WRITE
-        ? sizeof(journal_entry_small_write) + dsk.dirty_dyn_size(wr->small().offset, wr->small().len)
-        : sizeof(journal_entry_big_write) + dsk.dirty_dyn_size(0, dsk.data_block_size));
-    choose_journal_block(je_size);
-    journal_entry *je = (journal_entry*)(new_journal_ptr + new_journal_in_pos);
-    je->magic = JOURNAL_MAGIC;
-    je->type = (wr->entry_type & BS_HEAP_STABLE) ? JE_SMALL_WRITE_INSTANT : JE_SMALL_WRITE;
-    je->size = je_size;
-    je->crc32_prev = new_crc32_prev;
-    je->small_write.oid = (object_id){ .inode = wr->inode, .stripe = wr->stripe };
-    je->small_write.version = wr->version;
-    if (wr->type() == BS_HEAP_SMALL_WRITE)
-    {
-        je->small_write.offset = wr->small().offset;
-        je->small_write.len = wr->small().len;
-        je->small_write.data_offset = new_journal_data-new_journal_buf;
-        if (je->small_write.data_offset + je->small_write.len > new_journal_len)
-        {
-            fprintf(stderr, "Error: live entries don't fit to the new journal\n");
-            exit(1);
-        }
-        memcpy(new_journal_data, buffer_area+wr->small().location, je->small_write.len);
-        new_journal_data += je->small_write.len;
-        if (dsk.data_csum_type == 0 && wr->get_checksum(heap))
-            je->small_write.crc32_data = *wr->get_checksum(heap);
-    }
-    else
-    {
-        je->big_write.location = wr->big_location(heap);
-    }
-    memcpy((uint8_t*)je + je->size, wr->get_ext_bitmap(heap), new_clean_entry_bitmap_size);
-    if (dsk.data_csum_type != 0 && wr->get_checksums(heap))
-    {
-        memcpy((uint8_t*)je + je->size + new_clean_entry_bitmap_size, wr->get_checksums(heap), heap->get_csum_size(wr));
-    }
-    je->crc32 = je_crc32(je);
-    new_journal_in_pos += je->size;
-    new_crc32_prev = je->crc32;
-}
-
 int disk_tool_t::resize_rebuild_meta()
 {
    new_meta_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_meta_len);
@@ -630,13 +564,12 @@ int disk_tool_t::resize_rebuild_meta()
    new_meta_hdr = (blockstore_meta_header_v3_t *)new_meta_buf;
    uint64_t new_meta_pos = dsk.meta_block_size;
    uint64_t next_lsn = 0;
-    std::vector<heap_entry_t*> writes;
    int r = process_meta(
        [&](blockstore_meta_header_v3_t *hdr)
        {
            new_meta_hdr->zero = 0;
            new_meta_hdr->magic = BLOCKSTORE_META_MAGIC_V1;
-            new_meta_hdr->version = new_meta_format == 0 ? BLOCKSTORE_META_FORMAT_HEAP : new_meta_format;
+            new_meta_hdr->version = new_meta_format;
            new_meta_hdr->meta_block_size = dsk.meta_block_size;
            new_meta_hdr->data_block_size = dsk.data_block_size;
            new_meta_hdr->bitmap_granularity = dsk.bitmap_granularity ? dsk.bitmap_granularity : 4096;
@@ -654,18 +587,16 @@ int disk_tool_t::resize_rebuild_meta()
        },
        [&](blockstore_heap_t *heap, heap_entry_t *obj, uint32_t meta_block_num)
        {
+            assert(new_meta_format == BLOCKSTORE_META_FORMAT_HEAP);
            if (!obj)
            {
                // Finish
-                if (new_meta_format == BLOCKSTORE_META_FORMAT_HEAP)
+                heap->fill_block_empty_space(new_meta_buf, new_meta_pos);
+                new_meta_pos = (new_meta_pos/dsk.meta_block_size + 1) * dsk.meta_block_size;
+                while (new_meta_pos < new_meta_len)
                {
                    heap->fill_block_empty_space(new_meta_buf, new_meta_pos);
-                    new_meta_pos = (new_meta_pos/dsk.meta_block_size + 1) * dsk.meta_block_size;
-                    while (new_meta_pos < new_meta_len)
-                    {
-                        heap->fill_block_empty_space(new_meta_buf, new_meta_pos);
-                        new_meta_pos += dsk.meta_block_size;
-                    }
+                    new_meta_pos += dsk.meta_block_size;
                }
                return;
            }
@@ -679,72 +610,24 @@ int disk_tool_t::resize_rebuild_meta()
                {
                    remap_small_write(wr);
                }
-                else if (wr->type() != BS_HEAP_DELETE && new_meta_format != BLOCKSTORE_META_FORMAT_HEAP)
+                // New -> New
+                if ((new_meta_pos % dsk.meta_block_size) + wr->size > dsk.meta_block_size)
                {
-                    fprintf(stderr, "Object %jx:%jx can't be converted to the old format because it contains an entry of type 0x%x%s\n",
-                        wr->inode, wr->stripe, wr->entry_type,
-                        (wr->type() == BS_HEAP_INTENT_WRITE ? " (intent_write)" : ""));
-                    exit(1);
-                }
-                if (new_meta_format == BLOCKSTORE_META_FORMAT_HEAP)
-                {
-                    // New -> New
-                    if ((new_meta_pos % dsk.meta_block_size) + wr->size > dsk.meta_block_size)
+                    heap->fill_block_empty_space(new_meta_buf, new_meta_pos);
+                    new_meta_pos = (new_meta_pos/dsk.meta_block_size + 1) * dsk.meta_block_size;
+                    if (new_meta_pos >= new_meta_len)
                    {
-                        heap->fill_block_empty_space(new_meta_buf, new_meta_pos);
-                        new_meta_pos = (new_meta_pos/dsk.meta_block_size + 1) * dsk.meta_block_size;
-                        if (new_meta_pos >= new_meta_len)
-                        {
-                            fprintf(stderr, "New metadata doesn't fit into the provided area\n");
-                            exit(1);
-                        }
-                    }
-                    memcpy(new_meta_buf + new_meta_pos, wr, wr->size);
-                    new_meta_pos += wr->size;
-                    if (skip_obsolete && wr->type() == BS_HEAP_BIG_WRITE && stable)
-                    {
-                        // Skip older writes
-                        return false;
-                    }
-                }
-                else
-                {
-                    // New -> Old
-                    if (wr->type() == BS_HEAP_DELETE && stable)
-                    {
-                        // Object is deleted, skip it
-                        return false;
-                    }
-                    if (wr->type() == BS_HEAP_BIG_WRITE && stable)
-                    {
-                        fill_old_clean_entry(heap, wr);
-                        return false;
-                    }
-                    else
-                    {
-                        writes.push_back(wr);
+                        fprintf(stderr, "New metadata doesn't fit into the provided area\n");
+                        exit(1);
                    }
                }
+                memcpy(new_meta_buf + new_meta_pos, wr, wr->size);
+                new_meta_pos += wr->size;
                return true;
            };
-            if (new_meta_format != BLOCKSTORE_META_FORMAT_HEAP || skip_obsolete)
+            for (auto wr = obj; wr; wr = heap->prev(wr))
            {
-                heap->iterate_with_stable(obj, obj->lsn, handle_write);
-            }
-            else
-            {
-                for (auto wr = obj; wr; wr = heap->prev(wr))
-                {
-                    handle_write(wr, false);
-                }
-            }
-            if (writes.size())
-            {
-                for (size_t i = writes.size(); i > 0; i--)
-                {
-                    fill_old_journal_entry(heap, writes[i-1]);
-                }
-                writes.clear();
+                handle_write(wr, false);
            }
        },
        [&](uint64_t block_num, clean_disk_entry *entry, uint8_t *bitmap)
@@ -121,6 +121,7 @@ static const char* help_text =
    "  --logfile <FILE>  log to the specified file\n"
    "  --enforce 1       enforce permissions at the server side (default is disabled)\n"
    "  --foreground 1    stay in foreground, do not daemonize\n"
+    "  --trace           trace all NFS requests\n"
    "\n"
    "NFS proxy is stateless if you use immediate_commit=all in your cluster and if\n"
    "you do not use client_enable_writeback=true, so you can freely use multiple\n"
@@ -158,7 +159,7 @@ json11::Json::object nfs_proxy_t::parse_args(int narg, const char *args[])
        {
            const char *opt = args[i]+2;
            cfg[str_replace(opt, "-", "_")] = !strcmp(opt, "json") || !strcmp(opt, "block") ||
-                !strcmp(opt, "dry-run") || !strcmp(opt, "recalc-stats") ||
+                !strcmp(opt, "dry-run") || !strcmp(opt, "recalc-stats") || !strcmp(opt, "trace") ||
                !strcmp(opt, "include-empty") || !strcmp(opt, "no-rm") || i == narg-1 ? "1" : args[++i];
        }
        else
@@ -586,8 +586,8 @@ send_again:
        }
        // Estimate reply WR count, create WR and SGE arrays
        xdr_write_chunk *reply_chunk = rop->in_rdma_msg.rdma_body.rdma_msg.rdma_reply;
-        int reply_chunk_wr_count = (reply_chunk ? reply_chunk->target.target_len : 0);
-        uint32_t wr_count = 1 + (chunk_iov ? 1 : 0) + (reply_chunk ? reply_chunk_wr_count : 0);
+        uint32_t reply_chunk_wr_count = (reply_chunk ? reply_chunk->target.target_len : 0);
+        uint32_t wr_count = 1 + (chunk_iov ? 1 : 0) + reply_chunk_wr_count;
        if (wr_count > ctx->max_send_wr)
        {
            fprintf(stderr, "Reply fragmentation (%u) exceeds max_send_wr (%u), sending ERR_CHUNK\n", wr_count, ctx->max_send_wr);
@@ -622,7 +622,7 @@ chunk_error:
        {
            size_t reply_chunk_len = 0;
            size_t left = msg_size;
-            for (uint32_t i = 0; i < reply_chunk->target.target_len; i++)
+            for (uint32_t i = 0; i < reply_chunk_wr_count; i++)
            {
                reply_chunk_len += reply_chunk->target.target_val[i].length;
                if (reply_chunk->target.target_val[i].length > left)
@@ -653,7 +653,7 @@ chunk_error:
        }
        ibv_sge sges[wr_count];
        ibv_send_wr wrs[wr_count];
-        int wr_pos = 0;
+        uint32_t wr_pos = 0;
        // Use a buffer from rdma_malloc for the reply
        assert(!rop->buffer);
        rop->buffer = rdma_malloc_alloc(conn_dev->alloc, hdr_size+msg_size);
@@ -685,7 +685,7 @@ chunk_error:
        if (reply_chunk)
        {
            size_t pos = hdr_size;
-            for (uint32_t i = 0; i < reply_chunk->target.target_len && pos < msg_size; i++)
+            for (uint32_t i = 0; i < reply_chunk_wr_count && pos < msg_size; i++)
            {
                uint32_t len = (reply_chunk->target.target_val[i].length < msg_size-pos
                    ? reply_chunk->target.target_val[i].length : msg_size-pos);
@@ -19,20 +19,19 @@ target_link_libraries(vitastor-osd

 # osd_rmw_test
 add_executable(osd_rmw_test EXCLUDE_FROM_ALL osd_rmw_test.cpp ../util/allocator.cpp)
-target_link_libraries(osd_rmw_test Jerasure ${ISAL_LIBRARIES} tcmalloc_minimal)
+target_link_libraries(osd_rmw_test Jerasure ${ISAL_LIBRARIES})
 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 ../util/allocator.cpp)
 	target_compile_definitions(osd_rmw_test_je PUBLIC -DNO_ISAL)
-	target_link_libraries(osd_rmw_test_je Jerasure tcmalloc_minimal)
+	target_link_libraries(osd_rmw_test_je Jerasure)
 	add_dependencies(build_tests osd_rmw_test_je)
 	add_test(NAME osd_rmw_test_jerasure COMMAND osd_rmw_test_je)
 endif (ISAL_LIBRARIES)

 # osd_peering_pg_test
 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)
@@ -15,22 +15,19 @@
 #include "str_util.h"
 #include "json_util.h"

-osd_t::osd_t(const json11::Json & config, ring_loop_t *ringloop)
+osd_t::osd_t(const json11::Json & config, ring_loop_i *ringloop, timerfd_manager_t *tfd)
 {
    zero_buffer_size = 1<<20;
    zero_buffer = malloc_or_die(zero_buffer_size);
    memset(zero_buffer, 0, zero_buffer_size);

    this->ringloop = ringloop;
+    this->tfd = tfd;

    this->cli_config = config.object_items();
    this->file_config = msgr.read_config(this->cli_config);
    parse_config(true);

-    epmgr = new epoll_manager_t(ringloop);
-    // FIXME: Use timerfd_interval based directly on io_uring
-    this->tfd = epmgr->tfd;
-
    if (json_is_true(this->config["osd_memlock"]))
    {
        // Lock all OSD memory if requested
@@ -99,7 +96,6 @@ osd_t::~osd_t()
    }
    ringloop->unregister_consumer(&consumer);
    ringloop->unregister_consumer(&init_consumer);
-    delete epmgr;
    if (bs)
        delete bs;
 #ifdef WITH_RDMACM
@@ -398,7 +394,7 @@ void osd_t::bind_socket()
        {
            int listen_fd = create_and_bind_socket(bind_address, listening_port ? listening_port : bind_port, listen_backlog, &listening_port);
            fcntl(listen_fd, F_SETFL, fcntl(listen_fd, F_GETFL, 0) | O_NONBLOCK);
-            epmgr->set_fd_handler(listen_fd, false, [this](int fd, int events)
+            tfd->set_fd_handler(listen_fd, false, [this](int fd, int events)
            {
                msgr.accept_connections(fd);
            });
@@ -529,6 +525,11 @@ void osd_t::exec_op(osd_op_t *cur_op)

 void osd_t::print_stats()
 {
+    if (bs && log_level > 1)
+    {
+        printf("[OSD %ju] Live entries: %ju (%ju bytes), garbage entries: %ju (%ju bytes)\n", osd_num,
+            bs->get_live_entries(), bs->get_live_memory(), bs->get_garbage_entries(), bs->get_garbage_memory());
+    }
    for (int i = OSD_OP_MIN; i <= OSD_OP_MAX; i++)
    {
        if (msgr.stats.op_stat_count[i] != prev_stats.op_stat_count[i] && i != OSD_OP_PING)
@@ -19,7 +19,6 @@
 #include "blockstore.h"
 #include "ringloop.h"
 #include "timerfd_manager.h"
-#include "epoll_manager.h"
 #include "osd_peering_pg.h"
 #include "messenger.h"
 #include "etcd_state_client.h"
@@ -210,9 +209,8 @@ class osd_t
    void *zero_buffer = NULL;
    uint64_t zero_buffer_size = 0;
    uint32_t bs_block_size, bs_bitmap_granularity, clean_entry_bitmap_size;
-    ring_loop_t *ringloop = NULL;
+    ring_loop_i *ringloop = NULL;
    timerfd_manager_t *tfd = NULL;
-    epoll_manager_t *epmgr = NULL;

    int listening_port = 0;
    std::vector<std::string> bind_addresses;
@@ -352,8 +350,7 @@ class osd_t
        std::function<int(pg_osd_set_t & new_set)> calc_set);
    pg_osd_set_state_t *mark_object_corrupted(pg_t & pg, object_id oid, pg_osd_set_state_t *prev_object_state,
        osd_rmw_stripe_t *stripes, bool ref);
-    pg_osd_set_state_t *mark_partial_write(pg_t & pg, object_id oid, pg_osd_set_state_t *prev_object_state,
-        osd_rmw_stripe_t *stripes, bool ref);
+    pg_osd_set_state_t *mark_partial_write(pg_t & pg, osd_op_t *cur_op);
    void deref_object_state(pg_t & pg, pg_osd_set_state_t **object_state, bool deref);
    bool remember_unstable_write(osd_op_t *cur_op, pg_t & pg, pg_osd_set_t & loc_set, int base_state);
    void handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op);
@@ -366,9 +363,10 @@ class osd_t
        osd_rmw_stripe_t *stripes, const uint64_t* osd_set, osd_op_t *cur_op, int subop_idx, int zero_read);
    void submit_primary_subop(osd_op_t *cur_op, osd_op_t *subop,
        osd_rmw_stripe_t *si, bool wr, inode_t inode, uint64_t op_version);
+    bool submit_to_osd(osd_op_t *subop, osd_num_t osd_num);
    void submit_primary_del_subops(osd_op_t *cur_op, uint64_t *cur_set, uint64_t set_size, pg_osd_set_t & loc_set);
    void submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_delete, int chunks_to_delete_count);
-    int submit_primary_sync_subops(osd_op_t *cur_op);
+    void submit_primary_sync_subops(osd_op_t *cur_op);
    void submit_primary_stab_subops(osd_op_t *cur_op);
    void submit_primary_rollback_subops(osd_op_t *cur_op, const uint64_t* osd_set);

@@ -392,7 +390,7 @@ class osd_t
    }

 public:
-    osd_t(const json11::Json & config, ring_loop_t *ringloop);
+    osd_t(const json11::Json & config, ring_loop_i *ringloop, timerfd_manager_t *tfd);
    ~osd_t();
    void force_stop(int exitcode);
    bool shutdown();
@@ -97,6 +97,7 @@ void osd_t::handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, p
            {
                // Will repeer/stop this PG
                msgr.stop_client(peer_it->second->client_id);
+                return;
            }
        }
    }
@@ -213,17 +214,8 @@ bool osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t
            handle_flush_op(op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK, pool_id, pg_num, fb, peer_osd, op->reply.hdr.retval);
            delete op;
        };
-        auto peer_it = msgr.osd_peers.find(peer_osd);
-        if (peer_it != msgr.osd_peers.end())
+        if (!submit_to_osd(op, peer_osd))
        {
-            op->client_id = peer_it->second->client_id;
-            msgr.outbox_push(op);
-        }
-        else
-        {
-            // Fail it immediately
-            op->reply.hdr.retval = -EPIPE;
-            op->callback(op);
            return false;
        }
    }
@@ -1,6 +1,7 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)

+#include "epoll_manager.h"
 #include "osd.h"

 #include <sys/prctl.h>
@@ -63,13 +64,15 @@ int main(int narg, char *args[])
    signal(SIGINT, handle_sigint);
    signal(SIGTERM, handle_sigint);
    ring_loop_t *ringloop = new ring_loop_t(RINGLOOP_DEFAULT_SIZE);
-    osd = new osd_t(config, ringloop);
+    epoll_manager_t *epmgr = new epoll_manager_t(ringloop);
+    osd = new osd_t(config, ringloop, epmgr->tfd);
    while (1)
    {
        ringloop->loop();
        ringloop->wait();
    }
    delete osd;
+    delete epmgr;
    delete ringloop;
    return 0;
 }
@@ -140,10 +140,12 @@ void osd_t::reset_pg(pg_t & pg)
    copies_to_delete_after_sync_count -= pg.copies_to_delete_after_sync.size();
    pg.copies_to_delete_after_sync.clear();
    corrupted_objects -= pg.corrupted_count;
+    inconsistent_objects -= pg.inconsistent_objects.size();
    incomplete_objects -= pg.incomplete_objects.size();
    misplaced_objects -= pg.misplaced_objects.size();
    degraded_objects -= pg.degraded_objects.size();
    pg.corrupted_count = 0;
+    pg.inconsistent_objects.clear();
    pg.incomplete_objects.clear();
    pg.misplaced_objects.clear();
    pg.degraded_objects.clear();
@@ -351,6 +353,7 @@ bool osd_t::continue_pg_peering(pg_t & pg)
        pg.calc_object_states(log_level);
        report_pg_state(pg);
        schedule_scrub(pg);
+        inconsistent_objects += pg.inconsistent_objects.size();
        incomplete_objects += pg.incomplete_objects.size();
        misplaced_objects += pg.misplaced_objects.size();
        // FIXME: degraded objects may currently include misplaced, too! Report them separately?
@@ -1,7 +1,9 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)

+#ifndef _LARGEFILE64_SOURCE
 #define _LARGEFILE64_SOURCE
+#endif

 #include "malloc_or_die.h"
 #include "osd_peering_pg.h"
@@ -269,6 +269,7 @@ resume_0:
                finish_op(cur_op, -EIO);
                return;
            }
+            assert(!cur_op->buf);
            cur_op->buf = alloc_read_buffer(op_data->stripes, pg ? pg->pg_data_size : 1, 0);
            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, op_data->prev_set, cur_op);
        }
@@ -281,6 +282,7 @@ resume_0:
            }
            // Submit reads
            op_data->degraded = 1;
+            assert(!cur_op->buf);
            cur_op->buf = alloc_read_buffer(op_data->stripes, pg->pg_size, 0);
            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, op_data->prev_set, cur_op);
        }
@@ -300,7 +302,11 @@ resume_2:
            // FIXME: ref = true ideally... because new_state != state is not necessarily true if it's freed and recreated
            auto new_object_state = mark_object_corrupted(*pg, op_data->oid, op_data->object_state, op_data->stripes, false);
            if (new_object_state != op_data->object_state)
+            {
+                free(cur_op->buf);
+                cur_op->buf = NULL;
                goto resume_0;
+            }
        }
        finish_op(cur_op, op_data->errcode);
        return;
@@ -429,24 +435,33 @@ pg_osd_set_state_t *osd_t::mark_object_corrupted(pg_t & pg, object_id oid,
 }

 // Mark the object as partially updated (probably due to a ENOSPC)
-pg_osd_set_state_t *osd_t::mark_partial_write(pg_t & pg, object_id oid, pg_osd_set_state_t *prev_object_state,
-    osd_rmw_stripe_t *stripes, bool ref)
+pg_osd_set_state_t *osd_t::mark_partial_write(pg_t & pg, osd_op_t *cur_op)
 {
-    return mark_object(pg, oid, prev_object_state, ref, [stripes](pg_osd_set_t & new_set)
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    return mark_object(pg, op_data->oid, op_data->object_state, true, [&](pg_osd_set_t & new_set)
    {
        // Mark object chunk(s) as outdated
        int changes = 0;
-        for (auto chunk_it = new_set.begin(); chunk_it != new_set.end(); )
+        for (auto & chunk: new_set)
        {
-            auto & chunk = *chunk_it;
-            if (stripes[chunk.role].osd_num == chunk.osd_num &&
-                stripes[chunk.role].read_error &&
-                chunk.loc_bad != LOC_OUTDATED)
+            if (chunk.loc_bad != LOC_OUTDATED)
            {
-                changes++;
-                chunk.loc_bad = LOC_OUTDATED;
+                bool success = false;
+                for (int i = 0; i < op_data->n_subops; i++)
+                {
+                    if (op_data->subops[i].osd_num == chunk.osd_num &&
+                        op_data->subops[i].reply.hdr.retval == op_data->subops[i].req.sec_rw.len)
+                    {
+                        success = true;
+                        break;
+                    }
+                }
+                if (!success)
+                {
+                    changes++;
+                    chunk.loc_bad = LOC_OUTDATED;
+                }
            }
-            chunk_it++;
        }
        return changes;
    });
@@ -522,19 +537,19 @@ pg_osd_set_state_t* osd_t::add_object_to_set(pg_t & pg, const object_id oid, con
    {
        this->incomplete_objects++;
        obj_state |= OBJ_INCOMPLETE;
-        pg_state_bits = PG_HAS_INCOMPLETE;
+        pg_state_bits |= PG_HAS_INCOMPLETE;
    }
    else if (n_roles < pg.pg_cursize)
    {
        this->degraded_objects++;
        obj_state |= OBJ_DEGRADED;
-        pg_state_bits = PG_HAS_DEGRADED;
+        pg_state_bits |= PG_HAS_DEGRADED;
    }
    else if (n_misplaced > 0 || n_outdated > 0)
    {
        this->misplaced_objects++;
        obj_state |= OBJ_MISPLACED;
-        pg_state_bits = PG_HAS_MISPLACED;
+        pg_state_bits |= PG_HAS_MISPLACED;
    }
    if (this->log_level >= log_at_level)
    {
@@ -614,7 +629,8 @@ bool osd_t::remove_object_from_state(object_id & oid, pg_osd_set_state_t **objec
    get_object_osd_set(pg, oid, &recheck_state);
    if (recheck_state != *object_state)
    {
-        recheck_state->ref_count++;
+        if (recheck_state)
+            recheck_state->ref_count++;
        (*object_state)->ref_count--;
        *object_state = recheck_state;
        return false;
@@ -769,8 +785,11 @@ resume_3:
        op_data->fact_ver++;
        submit_primary_del_subops(cur_op, NULL, 0, op_data->object_state ? op_data->object_state->osd_set : pg.cur_loc_set);
 resume_4:
-        op_data->st = 4;
-        return;
+        if (op_data->n_subops > 0)
+        {
+            op_data->st = 4;
+            return;
+        }
 resume_5:
        if (op_data->errors > 0)
        {
@@ -780,6 +799,11 @@ resume_5:
        }
        // Remove version override
        pg.ver_override.erase(op_data->oid);
+        // Mark PG and OSDs as dirty
+        for (auto & chunk: (op_data->object_state ? op_data->object_state->osd_set : pg.cur_loc_set))
+        {
+            this->dirty_osds.insert(chunk.osd_num);
+        }
        // Adjust PG stats after "instant stabilize", because we need object_state above
        if (!op_data->object_state)
        {
@@ -790,11 +814,6 @@ resume_5:
            remove_object_from_state(op_data->oid, &op_data->object_state, pg);
            deref_object_state(pg, &op_data->object_state, true);
        }
-        // Mark PG and OSDs as dirty
-        for (auto & chunk: (op_data->object_state ? op_data->object_state->osd_set : pg.cur_loc_set))
-        {
-            this->dirty_osds.insert(chunk.osd_num);
-        }
        for (auto cl_it = msgr.clients.find(cur_op->client_id); cl_it != msgr.clients.end(); )
        {
            cl_it->second->dirty_pgs.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
@@ -82,6 +82,7 @@ int osd_t::read_bitmaps(osd_op_t *cur_op, pg_t *pg, int base_state)
        goto resume_0;
    else if (op_data->st == base_state+1)
        goto resume_1;
+    assert(op_data->st < base_state);
    if (!pg || pg->state == PG_ACTIVE && pg->scheme == POOL_SCHEME_REPLICATED)
    {
        // Happy path for clean replicated PGs (all bitmaps are available locally)
@@ -112,6 +113,12 @@ resume_0:
            return 1;
        }
 resume_1:
+        if (op_data->errors > 0)
+        {
+            // Failure
+            finish_op(cur_op, op_data->errcode);
+            return -1;
+        }
        if (pg->scheme != POOL_SCHEME_REPLICATED)
        {
            for (int chain_num = 0; chain_num < op_data->chain_size; chain_num++)
@@ -183,7 +190,6 @@ int osd_t::collect_bitmap_requests(osd_op_t *cur_op, pg_t & pg, std::vector<bitm
            memcpy(local_stripes, op_data->stripes, sizeof(osd_rmw_stripe_t) * pg.pg_size);
            if (extend_missing_stripes(local_stripes, cur_set, pg.pg_data_size, pg.pg_size) < 0)
            {
-                free(op_data->snapshot_bitmaps);
                return -1;
            }
            int need_at_least = 0;
@@ -312,19 +318,7 @@ int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
                    }
                    handle_primary_subop(subop, cur_op);
                };
-                auto peer_it = msgr.osd_peers.find(subop_osd_num);
-                if (peer_it != msgr.osd_peers.end())
-                {
-                    subop->client_id = peer_it->second->client_id;
-                    msgr.outbox_push(subop);
-                }
-                else
-                {
-                    // Fail it immediately
-                    subop->client_id = 0;
-                    subop->reply.hdr.retval = -EPIPE;
-                    ringloop->set_immediate([subop]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
-                }
+                submit_to_osd(subop, subop_osd_num);
                subop_idx++;
            }
            prev = i+1;
@@ -607,7 +601,7 @@ void osd_t::send_chained_read_results(pg_t *pg, osd_op_t *cur_op)
        {
            if (cur > prev)
            {
-                 // Send buffer in parts to avoid copying
+                // Send buffer in parts to avoid copying
                if (!prev_set)
                {
                    while ((cur-prev) > zero_buffer_size/bs_bitmap_granularity)
@@ -270,22 +270,30 @@ void osd_t::submit_primary_subop(osd_op_t *cur_op, osd_op_t *subop,
        {
            handle_primary_subop(subop, cur_op);
        };
-        auto peer_it = msgr.osd_peers.find(si->osd_num);
-        if (peer_it != msgr.osd_peers.end())
-        {
-            subop->client_id = peer_it->second->client_id;
-            msgr.outbox_push(subop);
-        }
-        else
-        {
-            // Fail it immediately
-            subop->client_id = 0;
-            subop->reply.hdr.retval = -EPIPE;
-            ringloop->set_immediate([subop]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
-        }
+        submit_to_osd(subop, si->osd_num);
    }
 }

+bool osd_t::submit_to_osd(osd_op_t *subop, osd_num_t osd_num)
+{
+    subop->osd_num = osd_num;
+    auto peer_it = msgr.osd_peers.find(osd_num);
+    if (peer_it != msgr.osd_peers.end())
+    {
+        subop->client_id = peer_it->second->client_id;
+        msgr.outbox_push(subop);
+    }
+    else
+    {
+        // Fail it immediately
+        subop->client_id = 0;
+        subop->reply.hdr.retval = -EPIPE;
+        ringloop->set_immediate([subop]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
+        return false;
+    }
+    return true;
+}
+
 static uint64_t bs_op_to_osd_op[] = {
    0,
    OSD_OP_SEC_READ,            // BS_OP_READ = 1
@@ -359,7 +367,7 @@ void osd_t::add_bs_subop_stats(osd_op_t *subop, bool recovery_related)
    uint64_t opcode = bs_op_to_osd_op[subop->bs_op->opcode];
    timespec tv_end;
    clock_gettime(CLOCK_REALTIME, &tv_end);
-    uint64_t len = (opcode == OSD_OP_SEC_READ || opcode == OSD_OP_SEC_WRITE)
+    uint64_t len = (opcode == OSD_OP_SEC_READ || opcode == OSD_OP_SEC_WRITE || opcode == OSD_OP_SEC_WRITE_STABLE)
        ? subop->bs_op->len : 0;
    msgr.inc_op_stats(msgr.stats, opcode, subop->tv_begin, tv_end, len);
    if (recovery_related)
@@ -482,8 +490,11 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
    }
    if ((op_data->errors + op_data->done) >= op_data->n_subops)
    {
-        delete[] op_data->subops;
-        op_data->subops = NULL;
+        if (!op_data->errors || !op_data->done || opcode != OSD_OP_SEC_WRITE && opcode != OSD_OP_SEC_WRITE_STABLE)
+        {
+            delete[] op_data->subops;
+            op_data->subops = NULL;
+        }
        op_data->st++;
        if (cur_op->req.hdr.opcode == OSD_OP_READ)
        {
@@ -616,86 +627,82 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
            {
                handle_primary_subop(subop, cur_op);
            };
-            auto peer_it = msgr.osd_peers.find(chunk.osd_num);
-            if (peer_it != msgr.osd_peers.end())
-            {
-                subops[i].client_id = peer_it->second->client_id;
-                msgr.outbox_push(&subops[i]);
-            }
-            else
-            {
-                // Fail it immediately
-                subops[i].client_id = 0;
-                subops[i].reply.hdr.retval = -EPIPE;
-                ringloop->set_immediate([subop = &subops[i]]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
-            }
+            submit_to_osd(&subops[i], chunk.osd_num);
        }
    }
 }

-int osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
+void 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 = op_data->errcode = 0;
    op_data->n_subops = n_osds;
+    if (op_data->n_subops <= 0)
+    {
+        return;
+    }
+    osd_op_t *subops = new osd_op_t[n_osds];
    op_data->subops = subops;
    robin_hood::unordered_flat_map<uint64_t, osd_client_t*>::iterator peer_it;
+    int subop_idx = 0;
    for (int i = 0; i < n_osds; i++)
    {
        osd_num_t sync_osd = op_data->dirty_osds[i];
+        osd_op_t *subop = &subops[subop_idx];
        if (sync_osd == this->osd_num)
        {
-            clock_gettime(CLOCK_REALTIME, &subops[i].tv_begin);
-            subops[i].op_type = (uint64_t)cur_op;
-            subops[i].bs_op = new blockstore_op_t({
+            clock_gettime(CLOCK_REALTIME, &subop->tv_begin);
+            subop->op_type = (uint64_t)cur_op;
+            subop->bs_op = new blockstore_op_t({
                .opcode = BS_OP_SYNC,
-                .callback = [subop = &subops[i], this](blockstore_op_t *bs_subop)
+                .callback = [subop, this](blockstore_op_t *bs_subop)
                {
                    handle_primary_bs_subop(subop);
                },
            });
-            bs->enqueue_op(subops[i].bs_op);
+            bs->enqueue_op(subop->bs_op);
+            subop_idx++;
        }
        else if ((peer_it = msgr.osd_peers.find(sync_osd)) != msgr.osd_peers.end())
        {
-            subops[i].op_type = OSD_OP_OUT;
-            subops[i].client_id = peer_it->second->client_id;
-            subops[i].req = (osd_any_op_t){ .sec_sync = {
+            subop->op_type = OSD_OP_OUT;
+            subop->osd_num = sync_osd;
+            subop->client_id = peer_it->second->client_id;
+            subop->req = (osd_any_op_t){ .sec_sync = {
                .header = {
                    .magic = SECONDARY_OSD_OP_MAGIC,
                    .opcode = OSD_OP_SEC_SYNC,
                },
                .flags = cur_op->client_id == SELF_CLIENT && cur_op->req.hdr.opcode != OSD_OP_SCRUB ? OSD_OP_RECOVERY_RELATED : 0,
            } };
-            subops[i].callback = [cur_op, this](osd_op_t *subop)
+            subop->callback = [cur_op, this](osd_op_t *subop)
            {
                handle_primary_subop(subop, cur_op);
            };
-            msgr.outbox_push(&subops[i]);
-        }
-        else
-        {
-            op_data->done++;
+            msgr.outbox_push(subop);
+            subop_idx++;
        }
    }
-    if (op_data->done >= op_data->n_subops)
+    op_data->n_subops = subop_idx;
+    if (subop_idx <= 0)
    {
        delete[] op_data->subops;
        op_data->subops = NULL;
-        return 0;
    }
-    return 1;
 }

 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 = op_data->errcode = 0;
    op_data->n_subops = n_osds;
+    if (op_data->n_subops <= 0)
+    {
+        return;
+    }
+    osd_op_t *subops = new osd_op_t[n_osds];
    op_data->subops = subops;
    for (int i = 0; i < n_osds; i++)
    {
@@ -733,19 +740,7 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
            {
                handle_primary_subop(subop, cur_op);
            };
-            auto peer_it = msgr.osd_peers.find(stab_osd.osd_num);
-            if (peer_it != msgr.osd_peers.end())
-            {
-                subops[i].client_id = peer_it->second->client_id;
-                msgr.outbox_push(&subops[i]);
-            }
-            else
-            {
-                // Fail it immediately
-                subops[i].client_id = 0;
-                subops[i].reply.hdr.retval = -EPIPE;
-                ringloop->set_immediate([subop = &subops[i]]() { std::function<void(osd_op_t*)>(subop->callback)(subop); });
-            }
+            submit_to_osd(&subops[i], stab_osd.osd_num);
        }
    }
 }
@@ -831,8 +826,7 @@ void osd_t::submit_primary_rollback_subops(osd_op_t *cur_op, const uint64_t* osd
                    op_data->oid.inode, op_data->oid.stripe | role, op_data->target_ver-1
                );
 #endif
-                subop->client_id = msgr.osd_peers.at(osd_set[role])->client_id;
-                msgr.outbox_push(subop);
+                submit_to_osd(subop, osd_set[role]);
            }
            i++;
        }
@@ -127,13 +127,13 @@ resume_2:
    if (immediate_commit != IMMEDIATE_ALL)
    {
        // SYNC
-        if (!submit_primary_sync_subops(cur_op))
-        {
-            goto resume_4;
-        }
+        submit_primary_sync_subops(cur_op);
 resume_3:
-        op_data->st = 3;
-        return;
+        if (op_data->n_subops > 0)
+        {
+            op_data->st = 3;
+            return;
+        }
 resume_4:
        if (op_data->errors > 0)
        {
@@ -145,8 +145,11 @@ resume_4:
        // Stabilize version sets, if any
        submit_primary_stab_subops(cur_op);
 resume_5:
-        op_data->st = 5;
-        return;
+        if (op_data->n_subops > 0)
+        {
+            op_data->st = 5;
+            return;
+        }
    }
 resume_6:
    if (op_data->errors > 0)
@@ -191,8 +194,11 @@ resume_6:
        // Actually delete copies which we wanted to delete
        submit_primary_del_batch(cur_op, op_data->copies_to_delete, op_data->copies_to_delete_count);
 resume_7:
-        op_data->st = 7;
-        return;
+        if (op_data->n_subops > 0)
+        {
+            op_data->st = 7;
+            return;
+        }
 resume_8:
        if (op_data->errors > 0)
        {
@@ -188,7 +188,7 @@ resume_3:
    }
    else
    {
-        if ((op_data->fact_ver & ((uint64_t)1 << (64-PG_EPOCH_BITS) - 1)) == ((uint64_t)1 << (64-PG_EPOCH_BITS) - 1))
+        if ((op_data->fact_ver & (((uint64_t)1 << (64-PG_EPOCH_BITS)) - 1)) == (((uint64_t)1 << (64-PG_EPOCH_BITS)) - 1))
        {
            assert(pg.epoch != (((uint64_t)1 << PG_EPOCH_BITS)-1));
            pg.epoch++;
@@ -231,8 +231,11 @@ resume_10:
    }
    submit_primary_subops(SUBMIT_WRITE, op_data->target_ver, pg.cur_set.data(), cur_op);
 resume_4:
-    op_data->st = 4;
-    return;
+    if (op_data->n_subops > 0)
+    {
+        op_data->st = 4;
+        return;
+    }
 resume_5:
    if (op_data->errors > 0)
    {
@@ -247,8 +250,11 @@ resume_5:
            {
                submit_primary_rollback_subops(cur_op, pg.cur_set.data());
 resume_11:
-                op_data->st = 11;
-                return;
+                if (op_data->n_subops > 0)
+                {
+                    op_data->st = 11;
+                    return;
+                }
 resume_12:
                // Ignore ROLLBACK errors - submit_primary_subops will drop the connection if it fails
                delete[] op_data->unstable_writes;
@@ -257,11 +263,21 @@ resume_12:
            else
            {
                pg.ver_override.erase(op_data->oid);
-                mark_partial_write(pg, op_data->oid, op_data->object_state, op_data->stripes, true);
+                mark_partial_write(pg, cur_op);
+                if (op_data->subops)
+                {
+                    delete[] op_data->subops;
+                    op_data->subops = NULL;
+                }
                pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
                return;
            }
        }
+        if (op_data->subops)
+        {
+            delete[] op_data->subops;
+            op_data->subops = NULL;
+        }
        pg.ver_override.erase(op_data->oid);
        deref_object_state(pg, &op_data->object_state, true);
        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->errcode);
@@ -323,7 +339,6 @@ resume_12:
    }
 resume_6:
 resume_7:
-    op_data->n_subops = 0;
    if (!remember_unstable_write(cur_op, pg, pg.cur_loc_set, 6))
    {
        return;
@@ -359,16 +374,21 @@ resume_7:
            );
            recovery_stat[recovery_type].usec += usec;
        }
-        if (immediate_commit == IMMEDIATE_ALL)
+        if (immediate_commit != IMMEDIATE_ALL)
+        {
+            // Do not delete anything, it will be deleted as part of copies_to_delete_after_sync
+            deref_object_state(pg, &op_data->object_state, true);
+        }
+        else
        {
            submit_primary_del_subops(cur_op, pg.cur_set.data(), pg.pg_size, op_data->object_state->osd_set);
-        }
-        deref_object_state(pg, &op_data->object_state, true);
-        if (op_data->n_subops > 0)
-        {
+            deref_object_state(pg, &op_data->object_state, true);
+            if (op_data->n_subops > 0)
+            {
 resume_8:
-            op_data->st = 8;
-            return;
+                op_data->st = 8;
+                return;
+            }
 resume_9:
            if (op_data->errors > 0)
            {
@@ -487,8 +507,11 @@ immediate:
            }
            submit_primary_stab_subops(cur_op);
 resume_6:
-            op_data->st = 6;
-            return false;
+            if (op_data->n_subops > 0)
+            {
+                op_data->st = 6;
+                return false;
+            }
 resume_7:
            // FIXME: Free those in the destructor?
            delete op_data->unstable_write_osds;
@@ -417,6 +417,7 @@ void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsi
                    data_ptrs[orig++] = (uint8_t*)stripes[other].read_buf + (read_start - stripes[other].read_start);
                }
            }
+            assert(orig == pg_minsize);
            ec_encode_data(
                read_end-read_start, pg_minsize, wanted, dectable + wanted_base*item_size*pg_minsize,
                data_ptrs, data_ptrs + pg_minsize
@@ -1281,7 +1282,7 @@ std::vector<int> ec_find_good(osd_rmw_stripe_t *stripes, int stripe_count, int p
    std::vector<std::vector<int>> live_variants(pg_size);
    int eq_to[stripe_count];
    int live_roles = 0, live_total = 0;
-    for (int i = 0; i < pg_size; i++)
+    for (int i = 0; i < stripe_count; i++)
    {
        eq_to[i] = i;
    }
@@ -1381,7 +1382,7 @@ std::vector<int> ec_find_good(osd_rmw_stripe_t *stripes, int stripe_count, int p
            for (i = 0; i < pg_minsize; i++)
            {
                subvar[i]++;
-                if (subvar[i] < live_variants[combination[i]].size())
+                if (subvar[i] < live_variants[comb_to_subset[combination[i]]].size())
                    break;
                subvar[i] = 0;
            }
@@ -31,6 +31,7 @@ void test_recover_22_d2();
 void test_ec43_error_bruteforce();
 void test_recover_53_d5();
 void test_recover_22();
+void test_ec_find_good_multi_chunks();

 int main(int narg, char *args[])
 {
@@ -69,6 +70,7 @@ int main(int narg, char *args[])
    test_recover_22_d2();
    // Error bruteforce
    test_ec43_error_bruteforce();
+    test_ec_find_good_multi_chunks();
    // Test 19
    test_recover_53_d5();
    // Test 20
@@ -1345,3 +1347,54 @@ void test_recover_22()
    free(write_buf);
    use_ec(4, 2, false);
 }
+
+void test_ec_find_good_multi_chunks()
+{
+    use_ec(7, 4, true);
+    osd_num_t osd_set[7] = { 1, 2, 3, 4, 5, 6, 7 };
+    osd_rmw_stripe_t stripes[15] = {};
+    split_stripes(4, 4096, 0, 4096 * 4, stripes);
+    uint8_t *write_buf = (uint8_t*)malloc_or_die(4096 * 15);
+    set_pattern(write_buf+0*4096, 4096, PATTERN0);
+    set_pattern(write_buf+1*4096, 4096, PATTERN1);
+    set_pattern(write_buf+2*4096, 4096, PATTERN2);
+    set_pattern(write_buf+3*4096, 4096, PATTERN3);
+    uint8_t *rmw_buf = (uint8_t*)calc_rmw(write_buf, stripes, osd_set, 7, 4, 7, osd_set, 4096, 0);
+    calc_rmw_parity_ec(stripes, 7, 4, osd_set, osd_set, 4096, 0);
+    check_pattern(stripes[4].write_buf, 4096, PATTERN0^PATTERN1^PATTERN2^PATTERN3);
+    check_pattern(stripes[5].write_buf, 4096, 0xfcee568ba36371ac); // 2nd EC chunk
+    check_pattern(stripes[6].write_buf, 4096, 0x139274739ae6f387); // 3rd EC chunk
+    memcpy(write_buf+4*4096, stripes[4].write_buf, 4096);
+    memcpy(write_buf+5*4096, stripes[5].write_buf, 4096);
+    memcpy(write_buf+6*4096, stripes[6].write_buf, 4096);
+    // Make 15 chunks
+    for (int i = 0; i < 15; i++)
+    {
+        stripes[i].read_start = 0;
+        stripes[i].read_end = 4096;
+        stripes[i].read_buf = write_buf+i*4096;
+        stripes[i].write_buf = NULL;
+        stripes[i].role = i;
+        stripes[i].osd_num = i+1;
+    }
+    for (int i = 7; i < 14; i++)
+    {
+        stripes[i].role = i % 7;
+        memcpy(write_buf+i*4096, write_buf + (i%7)*4096, 4096);
+    }
+    stripes[14].role = 6;
+    memcpy(write_buf+14*4096, write_buf + 6*4096, 4096);
+    // Corrupt chunks 0-6 and chunk 13, make role 0 absent
+    stripes[0].read_error = true;
+    stripes[7].read_error = true;
+    for (int i = 0; i < 7; i++)
+        memset(write_buf+i*4096, i*10+1, 4096);
+    memset(write_buf+13*4096, 8*10+1, 4096);
+    // Find errors
+    auto res = ec_find_good(stripes, 15, 7, 4, false, 4096, 0, 100, true);
+    assert_eq_vec(res, std::vector<int>({8, 9, 10, 11, 12, 14}));
+    // Done
+    free(rmw_buf);
+    free(write_buf);
+    use_ec(7, 4, false);
+}
@@ -289,6 +289,8 @@ bool osd_t::continue_scrub()
    if (no_scrub)
    {
        // Return false = no more scrub work to do
+        if (scrub_cur_list.buf)
+            free(scrub_cur_list.buf);
        scrub_cur_list = {};
        scrub_last_pg = {};
        scrub_nearest_ts = 0;
@@ -503,7 +505,7 @@ void osd_t::scrub_check_results(osd_op_t *cur_op)
                {
                    unknown = true;
                }
-                if (votes[role] > 0 && votes[role] < votes[best])
+                if (eq_to[role] != -1 && votes[eq_to[role]] > 0 && votes[eq_to[role]] < votes[best])
                {
                    printf(
                        "[PG %u/%u] Object %jx:%jx v%ju copy on OSD %ju doesn't match %d other copies%s\n",
@@ -192,16 +192,13 @@ void osd_t::exec_secondary_real(osd_op_t *cur_op)
 #ifdef OSD_STUB
        cur_op->bs_op->retval = 0;
 #endif
-        if (enable_pg_locks)
+        for (int i = 0; i < cur_op->bs_op->len; i++)
        {
-            for (int i = 0; i < cur_op->bs_op->len; i++)
+            if (!sec_check_pg_lock(cl->in_osd_num, ((obj_ver_id*)cur_op->buf)[i].oid, cur_op->req.sec_stab.flags))
            {
-                if (!sec_check_pg_lock(cl->in_osd_num, ((obj_ver_id*)cur_op->buf)[i].oid, cur_op->req.sec_stab.flags))
-                {
-                    cur_op->bs_op->retval = -EPIPE;
-                    secondary_op_callback(cur_op);
-                    return;
-                }
+                cur_op->bs_op->retval = -EPIPE;
+                secondary_op_callback(cur_op);
+                return;
            }
        }
    }
@@ -215,8 +212,8 @@ void osd_t::exec_secondary_real(osd_op_t *cur_op)
            secondary_op_callback(cur_op);
            return;
        }
-        auto pool_id = INODE_POOL(cur_op->bs_op->min_oid.inode);
-        if (pool_id && !sec_check_pg_lock(0, (object_id){ .inode = cur_op->bs_op->min_oid.inode }, OSD_OP_IGNORE_PG_LOCK))
+        auto pool_id = INODE_POOL(cur_op->req.sec_list.min_inode);
+        if (!pool_id || !sec_check_pg_lock(0, (object_id){ .inode = cur_op->req.sec_list.min_inode }, OSD_OP_IGNORE_PG_LOCK))
        {
            // Check resharding state of the pool
            cur_op->bs_op->retval = -EPIPE;
@@ -229,7 +226,7 @@ void osd_t::exec_secondary_real(osd_op_t *cur_op)
        cur_op->bs_op->min_oid.inode = cur_op->req.sec_list.min_inode;
        cur_op->bs_op->min_oid.stripe = cur_op->req.sec_list.min_stripe;
        cur_op->bs_op->max_oid.inode = cur_op->req.sec_list.max_inode;
-        if (cur_op->req.sec_list.max_inode && cur_op->req.sec_list.max_stripe != UINT64_MAX)
+        if (cur_op->req.sec_list.max_inode)
        {
            cur_op->bs_op->max_oid.stripe = cur_op->req.sec_list.max_stripe
                ? cur_op->req.sec_list.max_stripe : UINT64_MAX;
@@ -4,13 +4,9 @@ project(vitastor)

 ### Test stubs

-# stub_osd, stub_bench, osd_test
+# stub_osd, stub_bench
 add_executable(stub_osd stub_osd.cpp ../util/rw_blocking.cpp ../util/addr_util.cpp)
-target_link_libraries(stub_osd tcmalloc_minimal)
 add_executable(stub_bench stub_bench.cpp ../util/rw_blocking.cpp ../util/addr_util.cpp)
-target_link_libraries(stub_bench tcmalloc_minimal)
-add_executable(osd_test osd_test.cpp ../util/rw_blocking.cpp ../util/addr_util.cpp)
-target_link_libraries(osd_test tcmalloc_minimal)

 # bindiff
 add_executable(bindiff
@@ -19,6 +15,7 @@ add_executable(bindiff

 # stub_uring_osd
 add_executable(stub_uring_osd
+	EXCLUDE_FROM_ALL
 	stub_uring_osd.cpp
 )
 target_link_libraries(stub_uring_osd
@@ -26,7 +23,6 @@ target_link_libraries(stub_uring_osd
 	${LIBURING_LIBRARIES}
 	${IBVERBS_LIBRARIES}
 	${RDMACM_LIBRARIES}
-	tcmalloc_minimal
 )

 # test_allocator
@@ -57,3 +57,7 @@ json11::Json::object osd_messenger_t::merge_configs(const json11::Json::object &
 {
    return cli_config;
 }
+
+void osd_messenger_t::destroy_rdma_conn(msgr_rdma_connection_t *rdma_conn)
+{
+}
@@ -1,31 +0,0 @@
-// Copyright (c) Vitaliy Filippov, 2019+
-// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
-
-#pragma once
-
-#include <functional>
-
-struct ring_consumer_t
-{
-    std::function<void(void)> loop;
-};
-
-class ring_loop_t
-{
-public:
-    void register_consumer(ring_consumer_t *consumer)
-    {
-    }
-    void unregister_consumer(ring_consumer_t *consumer)
-    {
-    }
-    void submit()
-    {
-    }
-    void wait()
-    {
-    }
-    void loop()
-    {
-    }
-};
@@ -1,389 +0,0 @@
-// Copyright (c) Vitaliy Filippov, 2019+
-// License: VNPL-1.1 (see README.md for details)
-
-#include <sys/types.h>
-#include <sys/socket.h>
-#include <netinet/in.h>
-#include <netinet/tcp.h>
-#include <arpa/inet.h>
-#include <string.h>
-#include <stdio.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <errno.h>
-#include <stdlib.h>
-#include <malloc.h>
-
-#include <stdexcept>
-
-#include "malloc_or_die.h"
-#include "addr_util.h"
-#include "osd_ops.h"
-#include "rw_blocking.h"
-#include "test_pattern.h"
-
-int connect_osd(const char *osd_address, int osd_port);
-
-uint64_t test_read(int connect_fd, uint64_t inode, uint64_t stripe, uint64_t version, uint64_t offset, uint64_t len);
-
-uint64_t test_write(int connect_fd, uint64_t inode, uint64_t stripe, uint64_t version, uint64_t pattern);
-
-void* test_primary_read(int connect_fd, uint64_t inode, uint64_t offset, uint64_t len);
-
-void test_primary_write(int connect_fd, uint64_t inode, uint64_t offset, uint64_t len, uint64_t pattern);
-
-void test_primary_sync(int connect_fd);
-
-void test_sync_stab_all(int connect_fd);
-
-void test_list_stab(int connect_fd);
-
-int main0(int narg, char *args[])
-{
-    int connect_fd;
-    // Prepare data for cluster read
-    connect_fd = connect_osd("127.0.0.1", 11203);
-    test_write(connect_fd, 2, 0, 1, PATTERN0);
-    close(connect_fd);
-    connect_fd = connect_osd("127.0.0.1", 11204);
-    test_write(connect_fd, 2, 1, 1, PATTERN1);
-    close(connect_fd);
-    connect_fd = connect_osd("127.0.0.1", 11205);
-    test_write(connect_fd, 2, 2, 1, PATTERN0^PATTERN1);
-    close(connect_fd);
-    return 0;
-}
-
-int main1(int narg, char *args[])
-{
-    int connect_fd;
-    void *data;
-    // Cluster read
-    connect_fd = connect_osd("127.0.0.1", 11203);
-    data = test_primary_read(connect_fd, 2, 0, 128*1024);
-    if (data)
-    {
-        check_pattern(data, 128*1024, PATTERN0);
-        printf("inode=2 0-128K OK\n");
-        free(data);
-    }
-    data = test_primary_read(connect_fd, 2, 0, 256*1024);
-    if (data)
-    {
-        check_pattern(data, 128*1024, PATTERN0);
-        check_pattern(data+128*1024, 128*1024, PATTERN1);
-        printf("inode=2 0-256K OK\n");
-        free(data);
-    }
-    close(connect_fd);
-    return 0;
-}
-
-int main2(int narg, char *args[])
-{
-    int connect_fd;
-    // Cluster write (sync not implemented yet)
-    connect_fd = connect_osd("127.0.0.1", 11203);
-    test_primary_write(connect_fd, 2, 0, 128*1024, PATTERN0);
-    test_primary_write(connect_fd, 2, 128*1024, 128*1024, PATTERN1);
-    test_sync_stab_all(connect_fd);
-    close(connect_fd);
-    connect_fd = connect_osd("127.0.0.1", 11204);
-    if (connect_fd >= 0)
-    {
-        test_sync_stab_all(connect_fd);
-        close(connect_fd);
-    }
-    connect_fd = connect_osd("127.0.0.1", 11205);
-    if (connect_fd >= 0)
-    {
-        test_sync_stab_all(connect_fd);
-        close(connect_fd);
-    }
-    return 0;
-}
-
-int main3(int narg, char *args[])
-{
-    int connect_fd;
-    connect_fd = connect_osd("127.0.0.1", 11203);
-    test_list_stab(connect_fd);
-    close(connect_fd);
-    return 0;
-}
-
-int main4(int narg, char *args[])
-{
-    int connect_fd;
-    // Cluster write (sync not implemented yet)
-    connect_fd = connect_osd("127.0.0.1", 11203);
-    test_primary_write(connect_fd, 2, 0, 128*1024, PATTERN0);
-    test_primary_write(connect_fd, 2, 128*1024, 128*1024, PATTERN1);
-    test_primary_sync(connect_fd);
-    close(connect_fd);
-    return 0;
-}
-
-int main(int narg, char *args[])
-{
-    int connect_fd;
-    connect_fd = connect_osd("192.168.7.2", 43051);
-    test_read(connect_fd, 1, 1039663104, UINT64_MAX, 0, 128*1024);
-    close(connect_fd);
-    return 0;
-}
-
-int connect_osd(const char *osd_address, int osd_port)
-{
-    struct sockaddr_storage addr;
-    if (!string_to_addr(osd_address, 0, osd_port, &addr))
-    {
-        fprintf(stderr, "server address: %s is not valid\n", osd_address);
-        return -1;
-    }
-
-    int connect_fd = socket(addr.ss_family, SOCK_STREAM, 0);
-    if (connect_fd < 0)
-    {
-        perror("socket");
-        return -1;
-    }
-    if (connect(connect_fd, (sockaddr*)&addr, sizeof(addr)) < 0)
-    {
-        perror("connect");
-        close(connect_fd);
-        return -1;
-    }
-    int one = 1;
-    setsockopt(connect_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-
-    return connect_fd;
-}
-
-bool check_reply(int r, osd_any_op_t & op, osd_any_reply_t & reply, int expected)
-{
-    if (r != OSD_PACKET_SIZE)
-    {
-        printf("read failed\n");
-        return false;
-    }
-    if (reply.hdr.magic != SECONDARY_OSD_REPLY_MAGIC ||
-        reply.hdr.id != op.hdr.id || reply.hdr.opcode != op.hdr.opcode)
-    {
-        printf("bad reply: magic, id or opcode does not match request\n");
-        return false;
-    }
-    if (expected >= 0 && reply.hdr.retval != expected)
-    {
-        printf("operation failed, retval=%jd\n", reply.hdr.retval);
-        return false;
-    }
-    return true;
-}
-
-uint64_t test_read(int connect_fd, uint64_t inode, uint64_t stripe, uint64_t version, uint64_t offset, uint64_t len)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_SEC_READ;
-    op.sec_rw.oid = {
-        .inode = inode,
-        .stripe = stripe,
-    };
-    op.sec_rw.version = version;
-    op.sec_rw.offset = offset;
-    op.sec_rw.len = len;
-    void *data = memalign_or_die(MEM_ALIGNMENT, op.sec_rw.len);
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    if (!check_reply(r, op, reply, op.sec_rw.len))
-    {
-        free(data);
-        return 0;
-    }
-    r = read_blocking(connect_fd, data, len);
-    if (r != len)
-    {
-        free(data);
-        perror("read data");
-        return 0;
-    }
-    free(data);
-    printf("Read %jx:%jx v%ju = v%ju\n", inode, stripe, version, reply.sec_rw.version);
-    op.hdr.opcode = OSD_OP_SEC_LIST;
-    op.sec_list.list_pg = 1;
-    op.sec_list.pg_count = 1;
-    op.sec_list.pg_stripe_size = 4*1024*1024;
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    if (reply.hdr.retval < 0 || !check_reply(r, op, reply, reply.hdr.retval))
-    {
-        return 0;
-    }
-    data = memalign_or_die(MEM_ALIGNMENT, sizeof(obj_ver_id)*reply.hdr.retval);
-    r = read_blocking(connect_fd, data, sizeof(obj_ver_id)*reply.hdr.retval);
-    if (r != sizeof(obj_ver_id)*reply.hdr.retval)
-    {
-        free(data);
-        perror("read data");
-        return 0;
-    }
-    obj_ver_id *ov = (obj_ver_id*)data;
-    for (int i = 0; i < reply.hdr.retval; i++)
-    {
-        if (ov[i].oid.inode == inode && (ov[i].oid.stripe & ~(4096-1)) == (stripe & ~(4096-1)))
-        {
-            printf("list: %jx:%jx v%ju stable=%d\n", ov[i].oid.inode, ov[i].oid.stripe, ov[i].version, i < reply.sec_list.stable_count ? 1 : 0);
-        }
-    }
-    return 0;
-}
-
-uint64_t test_write(int connect_fd, uint64_t inode, uint64_t stripe, uint64_t version, uint64_t pattern)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_SEC_WRITE;
-    op.sec_rw.oid = {
-        .inode = inode,
-        .stripe = stripe,
-    };
-    op.sec_rw.version = version;
-    op.sec_rw.offset = 0;
-    op.sec_rw.len = 128*1024;
-    void *data = memalign_or_die(MEM_ALIGNMENT, op.sec_rw.len);
-    for (int i = 0; i < (op.sec_rw.len)/sizeof(uint64_t); i++)
-        ((uint64_t*)data)[i] = pattern;
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    write_blocking(connect_fd, data, op.sec_rw.len);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    if (!check_reply(r, op, reply, op.sec_rw.len))
-    {
-        free(data);
-        return 0;
-    }
-    version = reply.sec_rw.version;
-    op.hdr.opcode = OSD_OP_TEST_SYNC_STAB_ALL;
-    op.hdr.id = 2;
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    if (!check_reply(r, op, reply, 0))
-    {
-        free(data);
-        return 0;
-    }
-    free(data);
-    return version;
-}
-
-void* test_primary_read(int connect_fd, uint64_t inode, uint64_t offset, uint64_t len)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_READ;
-    op.rw.inode = inode;
-    op.rw.offset = offset;
-    op.rw.len = len;
-    void *data = memalign_or_die(MEM_ALIGNMENT, len);
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    if (!check_reply(r, op, reply, len))
-    {
-        free(data);
-        return NULL;
-    }
-    r = read_blocking(connect_fd, data, len);
-    if (r != len)
-    {
-        free(data);
-        perror("read data");
-        return NULL;
-    }
-    return data;
-}
-
-void test_primary_write(int connect_fd, uint64_t inode, uint64_t offset, uint64_t len, uint64_t pattern)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_WRITE;
-    op.rw.inode = inode;
-    op.rw.offset = offset;
-    op.rw.len = len;
-    void *data = memalign_or_die(MEM_ALIGNMENT, len);
-    set_pattern(data, len, pattern);
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    write_blocking(connect_fd, data, len);
-    free(data);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    assert(check_reply(r, op, reply, len));
-}
-
-void test_primary_sync(int connect_fd)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_SYNC;
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    assert(check_reply(r, op, reply, 0));
-}
-
-void test_sync_stab_all(int connect_fd)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_TEST_SYNC_STAB_ALL;
-    write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    assert(check_reply(r, op, reply, 0));
-}
-
-void test_list_stab(int connect_fd)
-{
-    osd_any_op_t op;
-    osd_any_reply_t reply;
-    op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
-    op.hdr.id = 1;
-    op.hdr.opcode = OSD_OP_SEC_LIST;
-    op.sec_list.pg_count = 0;
-    assert(write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE) == OSD_PACKET_SIZE);
-    int r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-    assert(check_reply(r, op, reply, -1));
-    int total_count = reply.hdr.retval;
-    int stable_count = reply.sec_list.stable_count;
-    obj_ver_id *data = (obj_ver_id*)malloc_or_die(total_count * sizeof(obj_ver_id));
-    assert(data);
-    assert(read_blocking(connect_fd, data, total_count * sizeof(obj_ver_id)) == (total_count * sizeof(obj_ver_id)));
-    int last_start = stable_count;
-    for (int i = stable_count; i <= total_count; i++)
-    {
-        // Stabilize in portions of 32 entries
-        if (i - last_start >= 32 || i == total_count)
-        {
-            op.hdr.opcode = OSD_OP_SEC_STABILIZE;
-            op.sec_stab.len = sizeof(obj_ver_id) * (i - last_start);
-            assert(write_blocking(connect_fd, op.buf, OSD_PACKET_SIZE) == OSD_PACKET_SIZE);
-            assert(write_blocking(connect_fd, data + last_start, op.sec_stab.len) == op.sec_stab.len);
-            r = read_blocking(connect_fd, reply.buf, OSD_PACKET_SIZE);
-            assert(check_reply(r, op, reply, 0));
-            last_start = i;
-        }
-    }
-    obj_ver_id *data2 = (obj_ver_id*)malloc_or_die(sizeof(obj_ver_id) * 32);
-    assert(data2);
-    free(data2);
-    free(data);
-}
@@ -166,8 +166,9 @@ void ring_loop_mock_t::mark_completed(ring_data_t *data)
    wakeup();
 }

-disk_mock_t::disk_mock_t(size_t size, bool buffered)
+disk_mock_t::disk_mock_t(const std::string & name, size_t size, bool buffered)
 {
+    this->name = name;
    this->size = size;
    this->data = (uint8_t*)malloc_or_die(size);
    this->buffered = buffered;
@@ -193,6 +194,7 @@ void disk_mock_t::erase_buffers(uint64_t begin, uint64_t end)
        if (bs >= begin && be <= end)
        {
            // Remove the whole buffer
+            free(it->second.iov_base);
            buffers.erase(it++);
        }
        else if (bs < begin && be > end)
@@ -222,6 +224,7 @@ void disk_mock_t::erase_buffers(uint64_t begin, uint64_t end)
            assert(be > end);
            uint8_t *ce = (uint8_t*)malloc_or_die(be-end);
            memcpy(ce, (uint8_t*)it->second.iov_base + (end-bs), be-end);
+            free(it->second.iov_base);
            buffers[be] = (iovec){ .iov_base = ce, .iov_len = be-end };
            buffers.erase(it);
            break;
@@ -242,7 +245,7 @@ void disk_mock_t::discard_buffers(bool all, uint32_t seed)
    if (all)
    {
        if (trace)
-            printf("disk: discard all buffers (%zu)\n", buffers.size());
+            printf("%s: discard all buffers (%zu)\n", name.c_str(), buffers.size());
        for (auto & b: buffers)
            free(b.second.iov_base);
        buffers.clear();
@@ -250,7 +253,7 @@ void disk_mock_t::discard_buffers(bool all, uint32_t seed)
    else
    {
        if (trace)
-            printf("disk: discard random buffers seed=%u\n", seed);
+            printf("%s: discard random buffers seed=%u\n", name.c_str(), seed);
        std::mt19937 rnd(seed);
        for (auto it = buffers.begin(); it != buffers.end(); )
        {
@@ -279,7 +282,7 @@ ssize_t disk_mock_t::copy_from_sqe(io_uring_sqe *sqe, uint8_t *to, uint64_t base
        }
        size_t cur = (off + v[i].iov_len > size ? size-off : v[i].iov_len);
        if (trace)
-            printf("disk: write %zu+%zu from %jx\n", off, cur, (uint64_t)v[i].iov_base);
+            printf("%s: write %zu+%zu from %jx\n", name.c_str(), off, cur, (uint64_t)v[i].iov_base);
        memcpy(to + off - base_offset, v[i].iov_base, cur);
        off += v[i].iov_len;
    }
@@ -332,7 +335,7 @@ bool disk_mock_t::submit(io_uring_sqe *sqe)
            {
                size_t cur = (off + v[i].iov_len > size ? size-off : v[i].iov_len);
                if (trace)
-                    printf("disk: read %zu+%zu to %jx\n", off, cur, (uint64_t)v[i].iov_base);
+                    printf("%s: read %zu+%zu to %jx\n", name.c_str(), off, cur, (uint64_t)v[i].iov_base);
                if (buffers.size())
                    read_item((uint8_t*)v[i].iov_base, off, cur);
                else
@@ -374,7 +377,7 @@ bool disk_mock_t::submit(io_uring_sqe *sqe)
    else if (sqe->opcode == IORING_OP_FSYNC)
    {
        if (trace)
-            printf("disk: fsync\n");
+            printf("%s: fsync\n", name.c_str());
        if (buffers.size())
        {
            for (auto & b: buffers)
@@ -45,6 +45,7 @@ class disk_mock_t
 {
    uint8_t *data = NULL;
    std::map<uint64_t, iovec> buffers;
+    std::string name;
    size_t size = 0;
    bool buffered = false;

@@ -53,7 +54,7 @@ class disk_mock_t
    void read_item(uint8_t *to, uint64_t offset, uint64_t len);
 public:
    bool trace = false;
-    disk_mock_t(size_t size, bool buffered);
+    disk_mock_t(const std::string & name, size_t size, bool buffered);
    ~disk_mock_t();
    void clear(size_t offset, size_t len);
    void discard_buffers(bool all, uint32_t seed);
@@ -106,13 +106,13 @@ struct bs_test_t
        }
        if (!data_disk)
        {
-            data_disk = new disk_mock_t(parse_size(config["data_device_size"]), config["disable_data_fsync"] != "1");
+            data_disk = new disk_mock_t("data disk", parse_size(config["data_device_size"]), config["disable_data_fsync"] != "1");
            data_disk->clear(0, parse_size(config["data_offset"]));
        }
        uint64_t meta_size = parse_size(config["meta_device_size"]);
        if (meta_size && !meta_disk)
        {
-            meta_disk = new disk_mock_t(meta_size, config["disable_meta_fsync"] != "1");
+            meta_disk = new disk_mock_t("meta disk", meta_size, config["disable_meta_fsync"] != "1");
            meta_disk->clear(0, meta_size);
        }
        if (!bs)
@@ -251,9 +251,9 @@ static void test_fsync(bool separate_meta)
        test.meta_disk->trace = 1;

    // Write
-    printf("writing\n");
+    printf("writing 16K+4K v1\n");
    blockstore_op_t op;
-    op.opcode = BS_OP_WRITE;
+    op.opcode = BS_OP_WRITE_STABLE;
    op.oid = { .inode = 1, .stripe = 0 };
    op.version = 1;
    op.offset = 16384;
@@ -313,6 +313,54 @@ static void test_fsync(bool separate_meta)
    assert(memcmp(op2.buf+16*1024, op.buf, 4*1024) == 0);
    assert(is_zero(op2.buf+20*1024, 108*1024));

+    // Check fsync during compaction - do a small write
+    printf("writing 20K+4K v2\n");
+    op.opcode = BS_OP_WRITE_STABLE;
+    op.oid = { .inode = 1, .stripe = 0 };
+    op.version = 2;
+    op.offset = 20*1024;
+    op.len = 4096;
+    memset(op.buf, 0xab, 4096);
+    test.exec_op(&op);
+    assert(op.retval == op.len);
+
+    op.opcode = BS_OP_SYNC;
+    test.exec_op(&op);
+    assert(op.retval == 0);
+
+    // Check it by a read op
+    op2.version = UINT64_MAX;
+    test.exec_op(&op2);
+    assert(op2.retval == op2.len);
+    assert(is_zero(op2.buf, 16*1024));
+    assert(memcheck(op2.buf+16*1024, 0xaa, 4*1024));
+    assert(memcheck(op2.buf+20*1024, 0xab, 4*1024));
+    assert(is_zero(op2.buf+24*1024, 104*1024));
+
+    // Trigger & wait compaction
+    test.bs->flusher->dump_diagnostics();
+    test.bs->flusher->request_trim();
+    while (test.bs->heap->get_compact_queue_size())
+        test.ringloop->loop();
+    while (test.bs->flusher->is_active())
+        test.ringloop->loop();
+    test.bs->flusher->release_trim();
+    // Check that compaction succeeded
+    assert(!test.bs->heap->get_to_compact_count());
+
+    // Restart and check data again
+    test.destroy_bs();
+    test.data_disk->discard_buffers(true, 0);
+    test.init();
+
+    op2.version = UINT64_MAX;
+    test.exec_op(&op2);
+    assert(op2.retval == op2.len);
+    assert(is_zero(op2.buf, 16*1024));
+    assert(memcheck(op2.buf+16*1024, 0xaa, 4*1024));
+    assert(memcheck(op2.buf+20*1024, 0xab, 4*1024)); // <- would be lost without data device fsync
+    assert(is_zero(op2.buf+24*1024, 104*1024));
+
    free(op.buf);
    free(op2.buf);
 }
@@ -434,6 +482,84 @@ static void test_padded_csum_intent(bool perfect)
    free(op2.buf);
 }

+static void test_perfect_csum_interrupted()
+{
+    printf("\n-- test_perfect_csum_interrupted\n");
+
+    bs_test_t test;
+    test.default_cfg();
+    test.config["csum_block_size"] = "16384";
+    test.config["perfect_csum_update"] = "1";
+    test.config["disable_meta_fsync"] = "1";
+    test.config["meta_device"] = "./test_meta.bin";
+    test.config["meta_device_size"] = "33554432";
+    test.config["meta_device_sect"] = "4096";
+    test.config["data_offset"] = "0";
+    test.init();
+
+    // Write
+    printf("writing\n");
+    blockstore_op_t op;
+    op.opcode = BS_OP_WRITE_STABLE;
+    op.oid = { .inode = 1, .stripe = 0 };
+    op.version = 1;
+    op.offset = 16*1024;
+    op.len = 12*1024;
+    op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 12*1024);
+    memset(op.buf, 0xaa, 12*1024);
+    test.exec_op(&op);
+    assert(op.retval == op.len);
+
+    // Write again
+    printf("writing (small)\n");
+    op.version = 2;
+    op.offset = 20*1024;
+    op.len = 4*1024;
+    memset(op.buf, 0xbb, 4096);
+    test.exec_op(&op);
+    assert(op.retval == op.len);
+
+    // Trigger & block compaction after punch_holes
+    bool modified = false;
+    test.sqe_handler = [&](io_uring_sqe *sqe)
+    {
+        if (sqe->fd == MOCK_META_FD && sqe->opcode == IORING_OP_WRITEV)
+        {
+            bool ok = test.meta_disk->submit(sqe);
+            assert(ok);
+            free(((ring_data_t*)sqe->user_data)->iov.iov_base);
+            modified = true;
+            return true;
+        }
+        return false;
+    };
+    test.bs->flusher->request_trim();
+    while (!modified)
+        test.ringloop->loop();
+    test.destroy_bs();
+    test.init();
+
+    // Read and check
+    printf("rechecking reloaded\n");
+    blockstore_op_t op2;
+    op2.opcode = BS_OP_READ;
+    op2.oid = { .inode = 1, .stripe = 0 };
+    op2.version = UINT64_MAX;
+    op2.offset = 0;
+    op2.len = 128*1024;
+    op2.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 128*1024);
+    test.exec_op(&op2);
+    assert(op2.retval == op2.len);
+    assert(memcheck(op2.buf, 0, 16*1024));
+    assert(memcheck(op2.buf+16*1024, 0xaa, 4*1024));
+    assert(memcheck(op2.buf+20*1024, 0xbb, 4*1024));
+    assert(memcheck(op2.buf+24*1024, 0xaa, 4*1024));
+    assert(memcheck(op2.buf+28*1024, 0, 100*1024));
+
+    free(op.buf);
+    free(op2.buf);
+}
+
 static void test_padded_csum_parallel_read(bool perfect, uint32_t offset)
 {
    printf("\n-- test_padded_csum_parallel_read%s offset=%u\n", perfect ? " perfect_csum_update" : "", offset);
@@ -526,6 +652,7 @@ int main(int narg, char *args[])
    test_intent_over_unstable();
    test_padded_csum_intent(false);
    test_padded_csum_intent(true);
+    test_perfect_csum_interrupted();
    test_padded_csum_parallel_read(false, 8192);
    test_padded_csum_parallel_read(true, 8192);
    test_padded_csum_parallel_read(false, 16384);
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Vitaliy Filippov	ee0bae573f	Remove n_subops=0 (TODO also needs OSD unit tests)	2026-05-21 13:31:32 +03:00
Vitaliy Filippov	27be4ee2fa	Followup to mark_partial_write fix - also free subops in other branch to prevent assert on ENOSPC	2026-05-20 22:04:28 +03:00
Vitaliy Filippov	00517e2bac	Fix header checksum after new store write-meta	2026-05-19 12:30:21 +03:00
Vitaliy Filippov	e0a2615cbc	Fix test build	2026-05-19 02:27:03 +03:00
Vitaliy Filippov	63fe3c323a	Release 3.0.12 Important fixes (except the new store): - Fixed a possible use-after-free in the OSD during error handling of initial commit/rollback of objects in EC pools. - Fixed a possible free of an invalid pointer in the OSD during read errors from snapshot/clone chains in EC pools. - Fixed possibly incorrect handling of commit/rollback operations in EC pools during pool PG count changes. - Fixed the inverted fsync enable parameter in the ublk driver (fsync was not enabled on pools without immediate_commit). - Added the raw-ls command for debugging purposes to find object versions in the cluster using listing operations. New store fixes: - Improved startup speed by using LSN-based sorting only for objects with a large number of intermediate versions. - Added skip_double_claim option as a temporary workaround to fix the rare OSD startup error with the "double claimed block" message, observed by several users. This option does not affect data integrity. - Fixed incorrect rechecking of small writes during startup, which in theory could lead to duplicate small write object entries on the OSD. - Fixed fsync operation for disks with a writeback cache (without capacitors): - Fixed incorrect semantics of consecutive fsyncs (next fsync was not blocked by the previous one). - Added fsync when copying small writes from the buffer to the data device (somehow forgotten during initial development). - Added fsync after the initial garbage collection during OSD startup. - Fixed incorrect cast of LSN from uint64 to uint32, breaking fsync when reaching LSN 2^32. - Added missing verification of the metadata header checksum during startup. - Fixed incorrect updating of object checksums in perfect_csum_update=true mode. - Fixed a possible OSD crash with "assertion failed" when processing a malformed EC STABILIZE operation. - Fixed the accounting of active compactor coroutines. - Removed broken and untested new->old store conversion support. Minor issues fixed: - Incorrect accounting of OSD local operation statistics in replicated pools. - Missing non-zero exitcodes on vitastor-disk resize command errors. - Missing reset of the list of inconsistent objects during PG restarts. - Theoretically possible hangs of various OSD operations when working with completely corrupted objects (without a single available copy), and possibly in some other very rare situations. - Incorrect fsyncs when deleting objects from pools without immediate_commit (on disks with a writeback cache), which previously could leave garbage when deleting misplaced objects. - Possible crash/memory corruption of the NFS server during a targeted attack on NFS-RDMA. - Possibly incorrect handling of ENOSPC/EIO write errors in replicated pools, leading to inability to retry the write later. - Possible crash instead of a clean error exit when starting an OSD with the old storage engine on a disk with corrupted journal data. - Shallow copying of PG configuration in the monitor, however, not related to actual bugs. - Incorrect checking of allocated blocks in the QEMU driver in an unused code branch (without the BDRV_WANT_ZERO flag). - Possible memory leak on read errors of corrupted objects. - Possible incorrect PG states when corrupted objects are detected. - Possible failure to mark all "bad" copies of an object during scrubs without checksums and with a large number of replicas (> 4). - Incorrect checksum calculation in the old storage engine when bitmap_granularity < 4096 (a practically unused configuration). - Theoretically possible OSD crash in rare cases during a scrub and simultaneous object recovery. - Theoretically possible OSD crash when handling PING operation errors. - Slightly suboptimal logic for reusing the RDMA send buffer. - Possible memory leak when canceling an already running scrub via no_scrub. - Possible memory corruption when a client (e.g., QEMU code) passes invalid buffers and the writeback cache is enabled. - Potentially incorrect search for corrupted parts of EC objects (inability to find a "good" combination) during a scrub with checksums disabled. - Possible additional memory usage on the OSD side when handling failed reads from snapshots (not a leak however - the memory was freed upon client disconnection). - Potential sudden write slowdown at certain pg epoch values due to incorrect epoch update logic in etcd.	2026-05-18 02:24:22 +03:00
Vitaliy Filippov	a88465df05	Fix build under older gcc	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	ad24be717a	Wrap delete rdma_conn into destroy_rdma_conn	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	648e3b12f0	Replace public/protected ifdef with a friend class	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	a675993c74	Remove unused osd_test.cpp and mock/ringloop.h, move epmgr out of osd_t	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	c9dfd0f67d	Fix possible buffer overflow in NFS-RDMA	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	84919a10a9	Extract msgr_iothread to separate .{cpp,h}	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	51ae4d6e24	Remove extra assert in heap after reworking recheck	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	572b20fedc	Fix mark_partial_write for replicas (TODO: try to write an OSD unit test for it)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	4e2724b28f	Remove duplicate subop submit/fail code	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	768b1675f8	Do not use ISA-L in test_crc32c	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	38fa722725	Remove the rest of tcmalloc from build	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	e56d83fb7f	Remove heap->v1 conversion code (untested, unused and probably buggy)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	ff95a85875	Add bound check to je_crc32	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	98203568a8	Handle ENOENT from add_commit	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	89df98ee08	Deep copy pg_history and prev_pgs before mutation	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	0007a831b6	Add return to handle_flush_op when stopping the PG	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	40517c335f	Fix absolute bmp_end	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	c9f7308b6a	Wrap this submit_ into n_subops > 0 too	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	85c7e3bde0	Fix inverted writeback in ublk	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	4fb55b3535	Rework recheck_small_writes to free multiple bad entries more correctly and not use garbage as "already freed" flag	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	912aca11a3	Improve test_recheck to check all possible combinations	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	7b454bd16c	Do not try to free snapshot_bitmaps	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	a0c8be46a4	Speedup blockstore_heap loading by postponing only objects with a lot of versions	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	53b4329fac	Add missing active_flushers--	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	a7f41c4a12	Add assert to reconstruct_stripes_ec	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	5d78057ac3	Fix cur_op->buf memory leak on EIO/EDOM reads(?)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	8efc5a353f	Fix pg_state_bits overwriting PG_HAS_CORRUPTED	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	603b26b896	Add a FIXME about delete buffering	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	a3b0fe0deb	Fix op stats not tracked for OP_WRITE_STABLE	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	f504e356d5	Fix schedule_next_recheck_at	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	4ed17b7070	Sync metadata after writing out recheck_modified_blocks	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	1fd2819724	Fix marking only first OSD from non-majority group as corrupted during scrub Note: Could be tested with 5 replicas - 3 good and 2 corrupted if we had OSD unit tests	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	dcdabbc1ec	Fix vitastor-disk resize error return codes missed	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	625d5b7b9e	Fix punch_holes not updating entry crc (bug when perfect_csum_update==true)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	9e507fd333	Fix typo in crc32_pad Didn't affect anything though because it was only skipping < 4096 byte remainders	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	c2b5118127	Add data device fsyncs during compaction (forgotten during initial heap store dev)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	4b926e2223	Add a test for missing data device fsyncs	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	a5d9a6996a	Do not try to increment ref_count if recheck_state is null	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	0ee03e7172	Add reply == request opcode check (also just in case)	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	88b7d9afcd	Expect to receive bitmap only for successfull response	2026-05-18 01:17:56 +03:00
Vitaliy Filippov	f271c8450c	Block sync by all previous syncs	2026-05-11 16:33:49 +03:00
Vitaliy Filippov	f78d7d4efc	Do not capture client pointer in ping lambda	2026-05-11 16:30:01 +03:00
Vitaliy Filippov	fdaf7c88ff	Fix RDMA send buffer wraparound - could only lead to slower sending but not to corruption because the bug was canceled by send_out_pos >= send_done_pos check	2026-05-11 16:28:03 +03:00
Vitaliy Filippov	2fb6eb0c30	Free scrub_cur_list.buf when turning on no_scrub (potential memory leak)	2026-05-11 16:15:48 +03:00
Vitaliy Filippov	36d2b56208	Deep copy merged_history, also just in case	2026-05-11 16:14:03 +03:00
Vitaliy Filippov	14b22f2ba9	Retry even fatal errors for OP_SYNC (no real-world issue, just in case)	2026-05-11 16:08:14 +03:00
Vitaliy Filippov	fe8b1fe0cc	Add assert in cluster_client writeback to prevent overflow when incorrect operation data is passed	2026-05-11 16:06:19 +03:00
Vitaliy Filippov	1ec963e468	Fix ec_find_good for misplaced objects with extra corrupted chunks	2026-05-11 15:57:21 +03:00
Vitaliy Filippov	5100f822d8	Wrap all submit_primary_xxx_subops waits with if (n_subops) > 0	2026-05-11 15:17:50 +03:00
Vitaliy Filippov	7432494e88	Remove _LARGEFILE64_SOURCE warnings	2026-05-11 14:59:20 +03:00
Vitaliy Filippov	d0c0f3ea39	Do not ignore cur_primary change on PG state validation errors	2026-05-11 14:59:20 +03:00
Vitaliy Filippov	f61190f31d	Fix send_list/outbox desync on failed SEC_READ_BMP (possible memory leak)	2026-05-11 14:59:11 +03:00
Vitaliy Filippov	3dc0ab5c33	Clear inconsistent_objects on repeer	2026-05-11 14:59:08 +03:00
Vitaliy Filippov	de96efed2f	Fix 2^(64-PG_EPOCH_BITS) - 1 check	2026-05-11 14:59:08 +03:00
Vitaliy Filippov	87a5230798	Add forgotten header_csum check to the new store	2026-05-11 14:59:08 +03:00
Vitaliy Filippov	0c5e6d4346	Do not use modified_block(ui32) to store lsn(ui64)	2026-05-11 14:59:08 +03:00
Vitaliy Filippov	b278087410	Fix assert in blockstore_heap	2026-05-11 13:01:03 +03:00
Vitaliy Filippov	a8e821b13b	Mark OSDs as dirty during deletes more correctly	2026-05-11 12:56:23 +03:00
Vitaliy Filippov	caa70317fa	Add a workaround for "double claim" bug with the new store	2026-05-10 17:00:36 +03:00
Vitaliy Filippov	b8eaaabfe4	Add raw-ls command for troubleshooting	2026-05-10 17:00:36 +03:00
Vitaliy Filippov	e4d80c415e	Always check pool resharding state also for stabilize/rollback	2026-05-10 13:22:02 +03:00
Vitaliy Filippov	553191c3ff	Add asserts to the new store too O:-)	2026-05-10 11:39:54 +03:00
Vitaliy Filippov	ab385252b5	Add offset asserts before each write	2026-05-10 10:56:32 +03:00
Vitaliy Filippov	041185c673	Release 3.0.11 A single hotfix: - Fix missing error handling in EC snapshot/cloned image reads potentially leading to corrupted reads during OSD/PG restarts	2026-05-09 18:49:36 +03:00
Vitaliy Filippov	b03ac80a57	Fix missing error handling in chained read sec_read_bmp handling	2026-05-07 15:01:49 +03:00
Vitaliy Filippov	2ba56074f9	Run copy_part_bitmaps only after all subops succeed (slightly more correct)	2026-05-07 15:01:40 +03:00
Vitaliy Filippov	4acfe149cb	Release 3.0.10 Important bug fixes (new store): - Fix OSDs possibly refusing to start with "write metadata failed at offset xxx: Invalid argument" (fix buffer alignment during initial garbage collection) - Rollback change from 3.0.4 - on-disk garbage entries are not skipped on start again. This change doesn't have any impact normally, but OSDs originally running 3.0.0-3.0.2 and then upgraded to 3.0.9 may hit a bug where 3.0.9 refuses to start due to entries marked as garbage too early and flushed to disk in 3.0.0-3.0.2. Other changes: - Auto-select the only RDMA device/port if there is only one - Rollback one 3.0.9 change - there was no actual use-after-free :) (the problem was only relevant to an unstable development version) - Fix `vitastor-nfs --trace` option - Fix an unintended 1 second sleep in vitastor-cli rm-data - Fix inode statistics not being cleared for a deleted pool - Fix print to stdout in client	2026-04-27 13:53:48 +03:00
Vitaliy Filippov	008ed5b269	Rollback `25ecca7625` - there was no actual use-after-free :)	2026-04-22 01:35:38 +03:00
Vitaliy Filippov	4fffe0f032	Fix vitastor-nfs trace option	2026-04-20 21:51:47 +03:00
Vitaliy Filippov	a76d5ccc0d	Wakeup callers in rm-data	2026-04-17 13:53:32 +03:00
Vitaliy Filippov	8ed1e180e0	Clear inode_stats in mon	2026-04-14 02:44:46 +03:00
Vitaliy Filippov	8832fc3b14	Fix print to stdout in client	2026-04-14 02:43:50 +03:00
Vitaliy Filippov	0134934c99	Do not skip garbage entries on start (rollback change from 3.0.4)	2026-04-11 12:05:02 +03:00
Vitaliy Filippov	2e36f292bd	Fix buffer alignment during metadata clearing on init	2026-04-10 21:56:12 +03:00
Vitaliy Filippov	bcc6419760	Auto-select the only RDMA device/port if there is only one	2026-04-08 15:46:10 +03:00
Vitaliy Filippov	dd5941b9a4	Release 3.0.9 - Fix broken garbage collection in the new store leading to very high memory usage - Make the new store do full garbage collection on OSD startup by default (can be turned off with a new gc_on_start parameter) - Make startup logging with the new store more informative, allow to log amount of garbage with log_level > 1 - Fix clients reading from snapshots incorrectly on retries (for example when OSDs were restarted during read) - Fix a possible "send-after-free" in OSD which could probably also affect snapshot reads - Fix OSD crashing when handling a read_bitmap operation from a deleted object with the new store	2026-04-04 18:38:40 +03:00
Vitaliy Filippov	4005b88865	Do not use node_type	2026-04-04 17:37:47 +03:00
Vitaliy Filippov	280b5cd675	Add gc_on_start docs	2026-04-04 17:19:32 +03:00
Vitaliy Filippov	e5c505eaf4	Write 1 entry to blocks to collect garbage	2026-04-03 18:08:12 +00:00
Vitaliy Filippov	c1d244d4f0	Fix init metadata entry loading logging	2026-04-03 18:08:12 +00:00
Vitaliy Filippov	9b264a212f	Do full metadata GC on start (new store)	2026-04-03 18:08:12 +00:00
Vitaliy Filippov	ff7f5cb4f4	Report live & garbage entry count / memory with log_level > 1	2026-04-02 23:45:13 +00:00
Vitaliy Filippov	25ecca7625	Fix use-after-free in OSD ops (ooops)	2026-04-02 13:48:17 +00:00
Vitaliy Filippov	99c4244004	Try to write into blocks with more garbage first	2026-04-02 13:48:17 +00:00
Vitaliy Filippov	9949b9fb4e	Fix incorrect reads from snapshots on retries	2026-03-31 02:06:53 +03:00
Vitaliy Filippov	e6881ad1d5	Fix read_bitmap from deleted objects with the new store	2026-03-31 01:59:53 +03:00