WIP Use the disk interface wrapper only in the new blockstore

The original compaction condition only postponed compaction when not under
pressure, and only for BIG_WRITE and DELETE types:
  if (!under_pressure && (BIG_WRITE || DELETE))
This had two bugs:
1. BIG_INTENT overwrites were not included, causing add_compact() to be
   called on BIG_INTENT entries. add_compact() unconditionally rejects
   any stable overwrite with EBUSY, leading to unnecessary retry loops.
2. When under pressure, stable overwrites (BIG_WRITE, BIG_INTENT, DELETE
   with BS_HEAP_STABLE set) were not skipped. add_compact() rejects stable
   overwrites with EBUSY regardless of pressure, so attempting to compact
   them under pressure also always fails.
Fix: change the condition to:
  if ((!under_pressure || stable) && (BIG_WRITE || BIG_INTENT || DELETE))
When under pressure with an unstable overwrite, we now fall through to
`continue`, skipping it and scanning older entries (SMALL_WRITE /
INTENT_WRITE) that can actually be compacted to reclaim space.

docs/usage/cli.en.md

@@ -196,6 +196,7 @@ Options can be specified in classic dd style (`key=value`) or like usual (`--key
 | `iimg=<image>`  | Copy from Vitastor image `<image>`                                      |
 | `if=<file>`     | Copy from file `<file>`                                                 |
 | `oimg=<image>`  | Copy to Vitastor image `<image>`                                        |
+| `opool=<pool>`  | Pool to create output image if it doesn't exist                         |
 | `of=<file>`     | Copy to file `<file>`                                                   |
 | `bs=1M`         | Set copy block size                                                     |
 | `count=N`       | Copy only N input blocks. If N ends in B it counts bytes, not blocks    |

docs/usage/cli.ru.md

@@ -203,6 +203,7 @@ vitastor-cli dd [iimg=<image> | if=<file>] [oimg=<image> | of=<file>] [bs=1M] \
 | `iimg=<image>`  | Копировать из образа Vitastor `<image>`                                 |
 | `if=<file>`     | Копировать из файла `<file>`                                            |
 | `oimg=<image>`  | Копировать в образ Vitastor `<image>`                                   |
+| `opool=<pool>`  | Пул для создания выходного образа, если он не существует                |
 | `of=<file>`     | Копировать в файл `<file>`                                              |
 | `bs=1M`         | Задать размер блока копирования                                         |
 | `count=N`       | Копировать не более N блоков. Если N заканчивается на B - то N байт.    |

src/CMakeLists.txt

@@ -86,6 +86,11 @@ pkg_check_modules(RDMACM librdmacm)
 if (RDMACM_LIBRARIES)
 	add_definitions(-DWITH_RDMACM)
 endif (RDMACM_LIBRARIES)
+pkg_check_modules(VFN libvfn)
+pkg_check_modules(NVME libnvme)
+if (VFN_LIBRARIES AND NVME_LIBRARIES)
+	add_definitions(-DWITH_VFN)
+endif (VFN_LIBRARIES AND NVME_LIBRARIES)
 
 find_package(OpenSSL 3.0 REQUIRED)
 add_definitions(-DWITH_OPENSSL)

src/blockstore/blockstore_disk.cpp

@@ -1,10 +1,6 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#include <sys/file.h>
-#include <sys/ioctl.h>
-#include <unistd.h>
-
 #include <stdexcept>
 
 #include "blockstore.h"
@@ -355,31 +351,22 @@ void blockstore_disk_t::check_lengths()
     }
 }
 
-static int bs_openmode(const std::string & mode)
-{
-    if (mode == "directsync")
-        return O_DIRECT|O_SYNC;
-    else if (mode == "cached")
-        return O_SYNC;
-    else
-        return O_DIRECT;
-}
-
 void blockstore_disk_t::open_data()
 {
-    if (data_fd >= 0)
+    if (data_disk)
     {
         throw std::runtime_error("data device is already opened");
     }
-    data_fd = mock_mode ? MOCK_DATA_FD : open(data_device.c_str(), bs_openmode(data_io) | O_RDWR);
-    if (data_fd == -1)
+    data_disk = mock_mode
+        ? disk_i::create_mock(MOCK_DATA_FD, data_device_size, data_device_sect)
+        : disk_i::create_and_open(data_device, data_io, disable_flock);
+    if (!data_disk)
     {
-        throw std::runtime_error("Failed to open data device "+data_device+": "+std::string(strerror(errno)));
-    }
-    if (!mock_mode)
-    {
-        check_size(data_fd, &data_device_size, &data_device_sect, "data device");
+        throw std::runtime_error("Failed to open data device "+data_device);
     }
+    data_off_ring = data_disk->is_off_ring();
+    data_device_size = data_disk->get_size();
+    data_device_sect = data_disk->get_block_size();
     if (disk_alignment % data_device_sect)
     {
         throw std::runtime_error(
@@ -391,41 +378,33 @@ void blockstore_disk_t::open_data()
     {
         throw std::runtime_error("data_offset exceeds device size = "+std::to_string(data_device_size));
     }
-    if (!mock_mode && !disable_flock && flock(data_fd, LOCK_EX|LOCK_NB) != 0)
-    {
-        throw std::runtime_error(std::string("Failed to lock data device: ") + strerror(errno));
-    }
 }
 
 void blockstore_disk_t::open_meta()
 {
-    if (meta_fd >= 0)
+    if (meta_disk)
     {
         throw std::runtime_error("metadata device is already opened");
     }
     if (meta_device != data_device || meta_io != data_io)
     {
-        meta_fd = mock_mode ? MOCK_META_FD : open(meta_device.c_str(), bs_openmode(meta_io) | O_RDWR);
-        if (meta_fd == -1)
+        meta_disk = mock_mode
+            ? disk_i::create_mock(MOCK_META_FD, meta_device_size, meta_device_sect)
+            : disk_i::create_and_open(meta_device, meta_io, meta_device == data_device || disable_flock);
+        if (!meta_disk)
         {
-            throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
-        }
-        if (!mock_mode)
-        {
-            check_size(meta_fd, &meta_device_size, &meta_device_sect, "metadata device");
+            throw std::runtime_error("Failed to open metadata device "+meta_device);
         }
+        meta_device_size = meta_disk->get_size();
+        meta_device_sect = meta_disk->get_block_size();
         if (meta_offset >= meta_device_size)
         {
             throw std::runtime_error("meta_offset exceeds device size = "+std::to_string(meta_device_size));
         }
-        if (!mock_mode && !disable_flock && meta_device != data_device && flock(meta_fd, LOCK_EX|LOCK_NB) != 0)
-        {
-            throw std::runtime_error(std::string("Failed to lock metadata device: ") + strerror(errno));
-        }
     }
     else
     {
-        meta_fd = data_fd;
+        meta_disk = data_disk;
         meta_device_sect = data_device_sect;
         meta_device_size = 0;
         if (meta_offset >= data_device_size)
@@ -433,6 +412,7 @@ void blockstore_disk_t::open_meta()
             throw std::runtime_error("meta_offset exceeds device size = "+std::to_string(data_device_size));
         }
     }
+    meta_off_ring = meta_disk->is_off_ring();
     if (meta_block_size % meta_device_sect)
     {
         throw std::runtime_error(
@@ -444,33 +424,29 @@ void blockstore_disk_t::open_meta()
 
 void blockstore_disk_t::open_journal()
 {
-    if (journal_fd >= 0)
+    if (journal_disk)
     {
         throw std::runtime_error("journal device is already opened");
     }
     if (journal_device != meta_device || journal_io != meta_io)
     {
-        journal_fd = mock_mode ? MOCK_JOURNAL_FD : open(journal_device.c_str(), bs_openmode(journal_io) | O_RDWR);
-        if (journal_fd == -1)
+        journal_disk = mock_mode
+            ? disk_i::create_mock(MOCK_JOURNAL_FD, journal_device_size, journal_device_sect)
+            : disk_i::create_and_open(journal_device, journal_io, journal_device == meta_device || disable_flock);
+        if (!journal_disk)
         {
-            throw std::runtime_error("Failed to open journal device "+journal_device+": "+std::string(strerror(errno)));
-        }
-        if (!mock_mode)
-        {
-            check_size(journal_fd, &journal_device_size, &journal_device_sect, "journal device");
+            throw std::runtime_error("Failed to open journal device "+journal_device);
         }
+        journal_device_size = journal_disk->get_size();
+        journal_device_sect = journal_disk->get_block_size();
         if (journal_offset >= journal_device_size)
         {
             throw std::runtime_error("journal_offset exceeds device size = "+std::to_string(journal_device_size));
         }
-        if (!mock_mode && !disable_flock && journal_device != meta_device && flock(journal_fd, LOCK_EX|LOCK_NB) != 0)
-        {
-            throw std::runtime_error(std::string("Failed to lock journal device: ") + strerror(errno));
-        }
     }
     else
     {
-        journal_fd = meta_fd;
+        journal_disk = meta_disk;
         journal_device_sect = meta_device_sect;
         journal_device_size = 0;
         if (journal_offset >= (meta_device_size > 0 ? meta_device_size : data_device_size))
@@ -478,6 +454,7 @@ void blockstore_disk_t::open_journal()
             throw std::runtime_error("journal_offset exceeds device size");
         }
     }
+    journal_off_ring = journal_disk->is_off_ring();
     if (journal_block_size % journal_device_sect)
     {
         throw std::runtime_error(
@@ -489,26 +466,19 @@ void blockstore_disk_t::open_journal()
 
 void blockstore_disk_t::close_all()
 {
-    if (!mock_mode)
-    {
-        if (data_fd >= 0)
-            close(data_fd);
-        if (meta_fd >= 0 && meta_fd != data_fd)
-            close(meta_fd);
-        if (journal_fd >= 0 && journal_fd != meta_fd)
-            close(journal_fd);
-    }
-    data_fd = meta_fd = journal_fd = -1;
+    if (journal_disk && journal_disk != meta_disk && journal_disk != data_disk)
+        delete journal_disk;
+    if (meta_disk && meta_disk != data_disk)
+        delete meta_disk;
+    if (data_disk)
+        delete data_disk;
+    data_disk = meta_disk = journal_disk = NULL;
 }
 
 // Sadly DISCARD only works through ioctl(), but it seems to always block the device queue,
 // so it's not a big deal that we can only run it synchronously.
 int blockstore_disk_t::trim_data(std::function<bool(uint64_t)> is_used)
 {
-    if (mock_mode)
-    {
-        return -EINVAL;
-    }
     int r = 0;
     uint64_t j = 0, i = 0;
     uint64_t discarded = 0;
@@ -533,12 +503,12 @@ int blockstore_disk_t::trim_data(std::function<bool(uint64_t)> is_used)
                 }
                 if (range[1] > 0)
                 {
-                    r = ioctl(data_fd, BLKDISCARD, &range);
+                    r = data_disk->discard_sync(range[0], range[1]);
                     if (r != 0)
                     {
-                        fprintf(stderr, "Failed to execute BLKDISCARD %ju+%ju on %s: %s (code %d)\n",
+                        fprintf(stderr, "Failed to discard %ju+%ju bytes on %s: %s (code %d)\n",
                             range[0], range[1], data_device.c_str(), strerror(-r), r);
-                        return -errno;
+                        return r;
                     }
                     discarded += range[1];
                 }

src/blockstore/blockstore_disk.h

@@ -24,6 +24,8 @@
 
 class allocator_t;
 
+class disk_i;
+
 struct blockstore_disk_t
 {
     std::string data_device, meta_device, journal_device;
@@ -66,6 +68,8 @@ struct blockstore_disk_t
     uint64_t discard_granularity = 0;
 
     int meta_fd = -1, data_fd = -1, journal_fd = -1;
+    disk_i *meta_disk = NULL, *data_disk = NULL, *journal_disk = NULL;
+    bool meta_off_ring = false, data_off_ring = false, journal_off_ring = false;
     uint64_t meta_offset = 0, meta_device_sect = 0, meta_device_size = 0, meta_area_size = 0, min_meta_len = 0;
     uint64_t data_offset = 0, data_device_sect = 0, data_device_size = 0, data_len = 0;
     uint64_t journal_offset = 0, journal_device_sect = 0, journal_device_size = 0, journal_len = 0;

src/blockstore/blockstore_flush.cpp

@@ -37,18 +37,18 @@ journal_flusher_t::journal_flusher_t(blockstore_impl_t *bs)
 journal_flusher_co::journal_flusher_co()
 {
     wait_state = 0;
-    simple_callback_r = [this](ring_data_t* data)
+    simple_callback_r = [this](int res)
     {
         bs->live = true;
-        if (data->res != data->iov.iov_len)
-            bs->disk_error_abort("read operation during flush", data->res, data->iov.iov_len);
+        if (res < 0)
+            bs->disk_error_abort("read operation during flush", res);
         wait_count--;
     };
-    simple_callback_w = [this](ring_data_t* data)
+    simple_callback_w = [this](int res)
     {
         bs->live = true;
-        if (data->res != data->iov.iov_len)
-            bs->disk_error_abort("write operation during flush", data->res, data->iov.iov_len);
+        if (res < 0)
+            bs->disk_error_abort("write operation during flush", res);
         wait_count--;
     };
 }
@@ -161,19 +161,8 @@ void journal_flusher_t::loop()
         bs->ringloop->wakeup();
 }
 
-#define await_sqe(label) \
-    resume_##label:\
-        sqe = bs->get_sqe();\
-        if (!sqe)\
-        {\
-            wait_state = wait_base+label;\
-            return false;\
-        }\
-        data = ((ring_data_t*)sqe->user_data);
-
 bool journal_flusher_co::loop()
 {
-    int wait_base = 0;
     // This is much better than implementing the whole function as an FSM
     // Maybe I should consider a coroutine library like https://github.com/hnes/libaco ...
     // Or just C++ coroutines, but they require some wrappers
@@ -213,18 +202,25 @@ resume_0:
     res = bs->heap->get_next_compact(cur_oid);
     if ((bs->intent_write_counter >= bs->journal_trim_interval || flusher->force_fsync > 0) && co_id == 0)
     {
-        // Advance fsynced_lsn every <journal_trim_interval> intent writes
-        bs->intent_write_counter = 0;
 resume_17:
 resume_18:
 resume_19:
         if (!fsync_buffer(17))
+        {
             return false;
+        }
+        if (bs->intent_write_counter >= bs->journal_trim_interval)
+        {
+            // Advance fsynced_lsn every <journal_trim_interval> intent writes
+            bs->intent_write_counter = 0;
 resume_29:
 resume_30:
 resume_31:
-        if (!trim_lsn(29))
-            return false;
+            if (!trim_lsn(29))
+            {
+                return false;
+            }
+        }
     }
     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))
@@ -458,13 +454,22 @@ resume_9:
         if ((read_vec[i].copy_flags & COPY_BUF_JOURNAL) &&
             !(read_vec[i].copy_flags & COPY_BUF_COALESCED))
         {
+resume_10:
             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);
+            assert(clean_loc + read_vec[i].offset + read_vec[i].len <= bs->dsk.block_count*bs->dsk.data_block_size);
+            res = bs->dsk.data_disk->write_async(
+                bs->dsk.data_offset + clean_loc + read_vec[i].offset,
+                (size_t)read_vec[i].len,
+                read_vec[i].buf + (read_vec[i].copy_flags & COPY_BUF_PADDED
+                    ? read_vec[i].offset - read_vec[i].disk_offset : 0),
+                simple_callback_w
+            );
+            if (res == -EAGAIN)
+            {
+                wait_state = 10;
+                return false;
+            }
+            assert(res == 0);
             wait_count++;
         }
     }
@@ -572,7 +577,7 @@ void journal_flusher_co::fill_partial_checksum_blocks()
                 .disk_loc = clean_loc,
                 .disk_offset = blk_begin,
                 .disk_len = blk_end - blk_begin,
-                .buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, blk_end - blk_begin),
+                .buf = (uint8_t*)bs->dsk.data_disk->alloc_iomem(blk_end - blk_begin),
                 .wr = compact_info.clean_wr,
             });
         }
@@ -597,7 +602,8 @@ void journal_flusher_co::free_buffers()
         if (it->buf && !(it->copy_flags & COPY_BUF_COALESCED) &&
             (!bs->dsk.inmemory_journal || it->buf < bs->buffer_area || it->buf >= (uint8_t*)bs->buffer_area + bs->dsk.journal_len))
         {
-            free(it->buf);
+            disk_i *di = ((it->copy_flags & COPY_BUF_JOURNAL) ? bs->dsk.journal_disk : bs->dsk.data_disk);
+            di->free_iomem(it->buf);
         }
     }
     read_vec.clear();
@@ -748,6 +754,8 @@ resume_1:
 
 bool journal_flusher_co::read_buffered(int wait_base)
 {
+    disk_i *di = NULL;
+    copy_buffer_t *vec = NULL;
     if (wait_state == wait_base)
         goto resume_0;
     else if (wait_state == wait_base+1)
@@ -765,19 +773,24 @@ bool journal_flusher_co::read_buffered(int wait_base)
         if (!(read_vec[i].copy_flags & COPY_BUF_COALESCED) &&
             ((read_vec[i].copy_flags & COPY_BUF_JOURNAL) && !bs->dsk.inmemory_journal || (read_vec[i].copy_flags & COPY_BUF_DATA)))
         {
-            await_sqe(0);
-            auto & vec = read_vec[i];
-            if (!vec.buf)
-                vec.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, vec.disk_len);
-            data->iov = (struct iovec){ vec.buf, (size_t)vec.disk_len };
-            wait_count++;
-            io_uring_prep_readv(
-                sqe,
-                (vec.copy_flags & COPY_BUF_JOURNAL) ? bs->dsk.journal_fd : bs->dsk.data_fd,
-                &data->iov, 1,
-                ((vec.copy_flags & COPY_BUF_JOURNAL) ? bs->dsk.journal_offset : bs->dsk.data_offset) + vec.disk_loc + vec.disk_offset
+resume_0:
+            vec = &read_vec[i];
+            di = ((vec->copy_flags & COPY_BUF_JOURNAL) ? bs->dsk.journal_disk : bs->dsk.data_disk);
+            if (!vec->buf)
+                vec->buf = (uint8_t*)di->alloc_iomem(vec->disk_len);
+            res = di->read_async(
+                ((vec->copy_flags & COPY_BUF_JOURNAL) ? bs->dsk.journal_offset : bs->dsk.data_offset) + vec->disk_loc + vec->disk_offset,
+                (size_t)vec->disk_len,
+                vec->buf,
+                simple_callback_r
             );
-            data->callback = simple_callback_r;
+            if (res == -EAGAIN)
+            {
+                wait_state = wait_base+0;
+                return false;
+            }
+            assert(!res);
+            wait_count++;
         }
     }
     // Wait for reads/writes if the journal is not inmemory
@@ -820,18 +833,29 @@ resume_0:
     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)
+resume_1:
+        res = bs->dsk.data_disk->flush_async(simple_callback_w);
+        if (res == -EAGAIN)
         {
-            wait_state = wait_base+2;
+            wait_state = wait_base+1;
             return false;
         }
+        else if (res == -EALREADY)
+        {
+            cur_sync->sent = true;
+        }
+        else
+        {
+            assert(res == 0);
+            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();
@@ -868,17 +892,27 @@ resume_0:
     }
     flusher->fsyncing_meta = true;
     // Sync batch is ready. Do it.
-    await_sqe(1);
-    data->iov = { 0 };
-    data->callback = simple_callback_w;
-    io_uring_prep_fsync(sqe, bs->dsk.meta_fd, IORING_FSYNC_DATASYNC);
-    wait_count++;
-resume_2:
-    if (wait_count > 0)
+resume_1:
+    res = bs->dsk.meta_disk->flush_async(simple_callback_w);
+    if (res == -EAGAIN)
     {
-        wait_state = wait_base+2;
+        wait_state = wait_base+1;
         return false;
     }
+    else if (res == -EALREADY)
+    {
+    }
+    else
+    {
+        assert(res == 0);
+        wait_count++;
+resume_2:
+        if (wait_count > 0)
+        {
+            wait_state = wait_base+2;
+            return false;
+        }
+    }
     // Sync completed. All previous coroutines waiting for it must be resumed
     flusher->fsyncing_meta = false;
     bs->ringloop->wakeup();
@@ -905,7 +939,7 @@ resume_0:
 resume_1:
     assert(!wait_count);
     fsynced_lsn = bs->heap->get_completed_lsn();
-    if (!bs->submit_fsyncs(wait_count))
+    if (bs->submit_fsyncs(wait_count) != 0)
     {
         wait_state = wait_base+1;
         return false;
@@ -944,12 +978,16 @@ resume_0:
     flusher->active_flushers++;
     ((blockstore_meta_header_v3_t*)bs->meta_superblock)->completed_lsn = fsynced_lsn;
     ((blockstore_meta_header_v3_t*)bs->meta_superblock)->set_crc32c();
-    await_sqe(1);
-    data->iov = (struct iovec){ bs->meta_superblock, (size_t)bs->dsk.meta_block_size };
-    data->callback = simple_callback_w;
-    io_uring_prep_writev(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset);
-    // Update superblock with datasync
-    sqe->rw_flags = RWF_DSYNC;
+    // Update superblock with FUA (datasync)
+resume_1:
+    res = bs->dsk.meta_disk->write_async(bs->dsk.meta_offset, bs->dsk.meta_block_size,
+        bs->meta_superblock, simple_callback_w, true);
+    if (res == -EAGAIN)
+    {
+        wait_state = wait_base+1;
+        return false;
+    }
+    assert(res == 0);
     wait_count++;
 resume_2:
     if (wait_count > 0)

src/blockstore/blockstore_flush.h

@@ -43,14 +43,12 @@ class journal_flusher_co
     journal_flusher_t *flusher;
     int co_id;
     int wait_state, wait_count;
-    struct io_uring_sqe *sqe;
-    struct ring_data_t *data;
     uint8_t *new_csums = NULL;
     uint8_t *new_bmp = NULL;
     uint8_t *punch_bmp = NULL;
     uint8_t *new_ext_bmp = NULL;
 
-    std::function<void(ring_data_t*)> simple_callback_r, simple_callback_w;
+    std::function<void(int)> simple_callback_r, simple_callback_w;
 
     object_id cur_oid;
     heap_entry_t *cur_obj;

src/blockstore/blockstore_heap.cpp

@@ -422,7 +422,7 @@ corrupted_object:
             if ((wr->type() == BS_HEAP_BIG_INTENT || wr->type() == BS_HEAP_BIG_WRITE) &&
                 wr->big().block_num >= dsk->block_count)
             {
-                fprintf(stderr, "Error: big_write or big_intent entry %jx:%jx v%ju block_num is too large: %u > %lu. Metadata is incompatible with current parameters. ",
+                fprintf(stderr, "Error: big_write or big_intent entry %jx:%jx v%ju block_num is too large: %u > %ju. Metadata is incompatible with current parameters. ",
                     wr->inode, wr->stripe, wr->version, wr->big_intent().block_num, dsk->block_count);
                 goto corrupted_object;
             }
@@ -2436,9 +2436,11 @@ heap_compact_t blockstore_heap_t::iterate_compaction(heap_entry_t *obj, uint64_t
             res.do_delete = false;
             res.compact_lsn = 0;
             res.compact_version = 0;
-            if (!under_pressure && (wr->type() == BS_HEAP_BIG_WRITE || wr->type() == BS_HEAP_DELETE))
+            if ((!under_pressure || stable) && (wr->type() == BS_HEAP_BIG_WRITE || wr->type() == BS_HEAP_BIG_INTENT || wr->type() == BS_HEAP_DELETE))
             {
                 // We may postpone compaction if we have an unstable overwrite when not under pressure
+                // We MUST postpone if the overwrite is STABLE (even if under pressure), 
+                // because add_compact will reject STABLE overwrites with EBUSY.
                 return res;
             }
             continue;

src/blockstore/blockstore_impl.cpp

@@ -30,12 +30,15 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_i *
         dsk.close_all();
         throw;
     }
-    meta_superblock = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.meta_block_size);
+    dsk.data_disk->set_ringloop(ringloop);
+    dsk.meta_disk->set_ringloop(ringloop);
+    dsk.journal_disk->set_ringloop(ringloop);
+    meta_superblock = (uint8_t*)dsk.meta_disk->alloc_iomem(dsk.meta_block_size);
     memset(meta_superblock, 0, dsk.meta_block_size);
     flusher = new journal_flusher_t(this);
     if (dsk.inmemory_journal)
     {
-        buffer_area = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.journal_len);
+        buffer_area = (uint8_t*)dsk.journal_disk->alloc_iomem(dsk.journal_len);
     }
     heap = new blockstore_heap_t(&dsk, buffer_area, log_level);
     ringloop->wakeup();
@@ -66,9 +69,9 @@ blockstore_impl_t::~blockstore_impl_t()
     if (heap)
         delete heap;
     if (buffer_area)
-        free(buffer_area);
+        dsk.journal_disk->free_iomem(buffer_area);
     if (meta_superblock)
-        free(meta_superblock);
+        dsk.meta_disk->free_iomem(meta_superblock);
     ringloop->unregister_consumer(&ring_consumer);
     dsk.close_all();
 }
@@ -127,9 +130,9 @@ void blockstore_impl_t::loop()
             if (PRIV(op)->wait_for)
             {
                 check_wait(op);
-                if (PRIV(op)->wait_for == WAIT_SQE)
+                if (PRIV(op)->wait_for != WAIT_COMPACTION)
                 {
-                    // ring is full, stop submission
+                    // disk queue(s) are full, stop submission
                     break;
                 }
                 else if (PRIV(op)->wait_for)
@@ -140,7 +143,6 @@ void blockstore_impl_t::loop()
                     continue;
                 }
             }
-            unsigned prev_sqe_pos = ringloop->save();
             // 0 = can't submit
             // 1 = in progress
             // 2 = can be removed from queue
@@ -189,8 +191,7 @@ void blockstore_impl_t::loop()
             if (wr_st == 0)
             {
                 PRIV(op)->pending_ops = 0;
-                ringloop->restore(prev_sqe_pos);
-                if (PRIV(op)->wait_for == WAIT_SQE)
+                if (PRIV(op)->wait_for != WAIT_COMPACTION)
                 {
                     // ring is full, stop submission
                     break;
@@ -261,13 +262,34 @@ bool blockstore_impl_t::is_safe_to_stop()
 
 void blockstore_impl_t::check_wait(blockstore_op_t *op)
 {
-    if (PRIV(op)->wait_for == WAIT_SQE)
+    if (PRIV(op)->wait_for == WAIT_DATA_QUEUE)
     {
-        if (ringloop->space_left() < PRIV(op)->wait_detail)
+        if (dsk.data_disk->get_queue_space() < PRIV(op)->wait_detail)
         {
-            // stop submission if there's still no free space
 #ifdef BLOCKSTORE_DEBUG
-            printf("Still waiting for %ju SQE(s)\n", PRIV(op)->wait_detail);
+            printf("Still waiting for %ju free data disk queue slots\n", PRIV(op)->wait_detail);
+#endif
+            return;
+        }
+        PRIV(op)->wait_for = 0;
+    }
+    else if (PRIV(op)->wait_for == WAIT_META_QUEUE)
+    {
+        if (dsk.meta_disk->get_queue_space() < PRIV(op)->wait_detail)
+        {
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting for %ju free meta disk queue slots\n", PRIV(op)->wait_detail);
+#endif
+            return;
+        }
+        PRIV(op)->wait_for = 0;
+    }
+    else if (PRIV(op)->wait_for == WAIT_JOURNAL_QUEUE)
+    {
+        if (dsk.journal_disk->get_queue_space() < PRIV(op)->wait_detail)
+        {
+#ifdef BLOCKSTORE_DEBUG
+            printf("Still waiting for %ju free journal disk queue slots\n", PRIV(op)->wait_detail);
 #endif
             return;
         }
@@ -404,7 +426,7 @@ void blockstore_meta_header_v3_t::set_crc32c()
     header_csum = calc;
 }
 
-void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expected)
+void blockstore_impl_t::disk_error_abort(const char *op, int retval)
 {
     if (retval == -EAGAIN)
     {
@@ -413,7 +435,7 @@ void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expecte
             " Please upgrade your kernel. Aborting.\n", op);
         exit(1);
     }
-    fprintf(stderr, "Disk %s failed: result is %d, expected %d. Can't continue, sorry :-(\n", op, retval, expected);
+    fprintf(stderr, "Disk %s failed: %s (code %d). Can't continue, sorry :-(\n", op, strerror(-retval), retval);
     exit(1);
 }
 

src/blockstore/blockstore_impl.h

@@ -20,6 +20,7 @@
 #include <unordered_map>
 #include <unordered_set>
 
+#include "disk_util.h"
 #include "malloc_or_die.h"
 
 class blockstore_impl_t;
@@ -132,21 +133,16 @@ public:
 
     bool stop_sync_submitted = false;
 
-    inline struct io_uring_sqe* get_sqe()
-    {
-        return ringloop->get_sqe();
-    }
-
     void open_data();
     void open_meta();
     void open_journal();
 
-    void disk_error_abort(const char *op, int retval, int expected);
+    void disk_error_abort(const char *op, int retval);
 
     // Asynchronous init
-    int initialized;
-    int metadata_buf_size;
-    blockstore_init_meta* metadata_init_reader;
+    int initialized = 0;
+    uint64_t metadata_buf_size = 0;
+    blockstore_init_meta* metadata_init_reader = NULL;
 
     void check_wait(blockstore_op_t *op);
     void init_op(blockstore_op_t *op);
@@ -163,7 +159,7 @@ public:
     void find_holes(std::vector<copy_buffer_t> & read_vec, uint32_t item_start, uint32_t item_end,
         std::function<void(int&, uint32_t, uint32_t)> callback);
     void free_read_buffers(std::vector<copy_buffer_t> & rv);
-    void handle_read_event(ring_data_t *data, blockstore_op_t *op);
+    void handle_read_event(int res, blockstore_op_t *op);
     bool verify_read_checksums(blockstore_op_t *op);
 
     // Write
@@ -173,11 +169,11 @@ public:
     bool intent_write_allowed(blockstore_op_t *op, heap_entry_t *obj);
     int dequeue_write(blockstore_op_t *op);
     int continue_write(blockstore_op_t *op);
-    void handle_write_event(ring_data_t *data, blockstore_op_t *op);
+    void handle_write_event(int res, blockstore_op_t *op);
 
     // Sync
     int continue_sync(blockstore_op_t *op);
-    bool submit_fsyncs(int & wait_count);
+    int submit_fsyncs(int & wait_count);
     int do_sync(blockstore_op_t *op, int base_state);
     bool has_unsynced();
 

src/blockstore/blockstore_init.cpp

@@ -11,12 +11,6 @@
 #define INIT_META_READING 1
 #define INIT_META_READ_DONE 2
 
-#define GET_SQE() \
-    sqe = bs->get_sqe();\
-    if (!sqe)\
-        throw std::runtime_error("io_uring is full during initialization");\
-    data = ((ring_data_t*)sqe->user_data)
-
 struct blockstore_recheck_io_t
 {
     bool is_data;
@@ -36,14 +30,13 @@ blockstore_init_meta::blockstore_init_meta(blockstore_impl_t *bs)
     this->bs = bs;
 }
 
-void blockstore_init_meta::handle_event(ring_data_t *data, int buf_num, const char *op)
+void blockstore_init_meta::handle_event(int res, int buf_num, const char *op)
 {
-    if (data->res != data->iov.iov_len)
+    if (res < 0)
     {
         throw std::runtime_error(strprintf(
-            "%s failed at offset %ju: got %s (code %d), but expected %zu",
-            op, (buf_num >= 0 ? bufs[buf_num].offset : last_read_offset), strerror(-data->res),
-            data->res, data->iov.iov_len
+            "%s failed at offset %ju: %s (code %d)",
+            op, (buf_num >= 0 ? bufs[buf_num].offset : last_read_offset), strerror(-res), res
         ));
     }
     if (buf_num >= 0)
@@ -67,16 +60,16 @@ int blockstore_init_meta::loop()
     else if (wait_state == 7) goto resume_7;
     else if (wait_state == 8) goto resume_8;
     else if (wait_state == 9) goto resume_9;
-    metadata_buffer = memalign(MEM_ALIGNMENT, 2*bs->metadata_buf_size);
+    if (bs->metadata_buf_size > bs->dsk.meta_disk->get_max_io_size())
+        bs->metadata_buf_size = bs->dsk.meta_disk->get_max_io_size();
+    metadata_buffer = bs->dsk.meta_disk->alloc_iomem(2*bs->metadata_buf_size);
     if (!metadata_buffer)
         throw std::runtime_error("Failed to allocate metadata read buffer");
     // Read metadata superblock
-    GET_SQE();
     last_read_offset = 0;
-    data->iov = { bs->meta_superblock, (size_t)bs->dsk.meta_block_size };
-    data->callback = [this](ring_data_t *data) { handle_event(data, -1, "read metadata header"); };
-    io_uring_prep_readv(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset);
-    bs->ringloop->submit();
+    res = bs->dsk.meta_disk->read_async(bs->dsk.meta_offset, bs->dsk.meta_block_size, bs->meta_superblock,
+        [this](int res) { handle_event(res, -1, "read metadata header"); });
+    assert(res == 0);
     submitted++;
 resume_1:
     if (submitted > 0)
@@ -165,16 +158,16 @@ resume_1:
         md_offset = 0;
         while (md_offset < bs->dsk.journal_len)
         {
-            GET_SQE();
-            data->iov = (iovec){
+            submit_len = (bs->dsk.journal_len - md_offset < bs->metadata_buf_size ? bs->dsk.journal_len - md_offset : bs->metadata_buf_size);
+            last_read_offset = md_offset;
+            res = bs->dsk.journal_disk->read_async(
+                bs->dsk.journal_offset + md_offset, (size_t)submit_len,
                 bs->buffer_area + md_offset,
-                (size_t)(bs->dsk.journal_len - md_offset < bs->metadata_buf_size ? bs->dsk.journal_len - md_offset : bs->metadata_buf_size),
-            };
-            data->callback = [this](ring_data_t *data) { handle_event(data, -1, "read buffer area"); };
-            io_uring_prep_readv(sqe, bs->dsk.journal_fd, &data->iov, 1, bs->dsk.journal_offset + md_offset);
-            md_offset += data->iov.iov_len;
+                [this](int res) { handle_event(res, -1, "read buffer area"); }
+            );
+            assert(res == 0);
+            md_offset += submit_len;
             submitted++;
-            bs->ringloop->submit();
 resume_3:
             if (submitted > 0)
             {
@@ -201,26 +194,31 @@ resume_4:
                 bufs[i].size = bs->dsk.meta_area_size-next_offset > bs->metadata_buf_size
                     ? bs->metadata_buf_size : bs->dsk.meta_area_size-next_offset;
                 bufs[i].state = INIT_META_READING;
-                submitted++;
-                next_offset += bufs[i].size;
-                GET_SQE();
                 assert(bufs[i].size <= 0x7fffffff);
-                data->iov = { bufs[i].buf, (size_t)bufs[i].size };
                 if (!zero_on_init)
                 {
-                    data->callback = [this, i](ring_data_t *data) { handle_event(data, i, "read metadata"); };
-                    io_uring_prep_readv(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + bufs[i].offset);
+                    res = bs->dsk.meta_disk->read_async(
+                        bs->dsk.meta_offset + bufs[i].offset,
+                        (size_t)bufs[i].size, bufs[i].buf,
+                        [this, i](int res) { handle_event(res, i, "read metadata"); }
+                    );
+                    assert(res == 0);
                 }
                 else
                 {
                     // Fill metadata with empty block pattern
-                    data->callback = [this, i](ring_data_t *data) { handle_event(data, i, "clear metadata"); };
                     memset(bufs[i].buf, 0, bufs[i].size);
                     for (uint64_t o = 0; o < bufs[i].size; o += bs->dsk.meta_block_size)
                         bs->heap->fill_block_empty_space(bufs[i].buf + o, 0);
-                    io_uring_prep_writev(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + bufs[i].offset);
+                    res = bs->dsk.meta_disk->write_async(
+                        bs->dsk.meta_offset + bufs[i].offset,
+                        (size_t)bufs[i].size, bufs[i].buf,
+                        [this, i](int res) { handle_event(res, i, "clear metadata"); }
+                    );
+                    assert(res == 0);
                 }
-                bs->ringloop->submit();
+                submitted++;
+                next_offset += bufs[i].size;
                 break;
             }
         }
@@ -266,26 +264,27 @@ resume_6:
     while (recheck_ios.size())
     {
         blockstore_recheck_io_t* rc = recheck_ios.back();
-        sqe = bs->get_sqe();
-        if (!sqe)
-            break;
+        disk_i *di = (rc->is_data ? bs->dsk.data_disk : bs->dsk.journal_disk);
         if (!rc->buf)
-            rc->buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, rc->len);
-        data = ((ring_data_t*)sqe->user_data);
-        data->iov = (iovec){ rc->buf, rc->len };
-        data->callback = [this, rc](ring_data_t *data)
-        {
-            if (data->res < 0)
+            rc->buf = (uint8_t*)di->alloc_iomem(rc->len);
+        int res = di->read_async(
+            (rc->is_data ? bs->dsk.data_offset : bs->dsk.journal_offset) + rc->offset,
+            rc->len, rc->buf, [this, rc](int res)
             {
-                fprintf(stderr, "Buffer area read failed at offset %ju: %d\n", rc->offset, data->res);
-                exit(1);
+                if (res < 0)
+                {
+                    fprintf(stderr, "Buffer area read failed at offset %ju: %d\n", rc->offset, res);
+                    exit(1);
+                }
+                rc->cb(rc->buf);
+                disk_i *di = (rc->is_data ? bs->dsk.data_disk : bs->dsk.journal_disk);
+                di->free_iomem(rc->buf);
+                delete rc;
             }
-            rc->cb(rc->buf);
-            free(rc->buf);
-            delete rc;
-        };
-        io_uring_prep_readv(sqe, (rc->is_data ? bs->dsk.data_fd : bs->dsk.journal_fd), &data->iov, 1,
-            (rc->is_data ? bs->dsk.data_offset : bs->dsk.journal_offset) + rc->offset);
+        );
+        if (res == -EAGAIN)
+            break;
+        assert(res == 0);
         recheck_ios.pop_back();
     }
     return 1;
@@ -307,30 +306,35 @@ resume_7:
     for (i = 0; i < recheck_mod.size(); i++)
     {
 resume_8:
-        if (wait_count >= bs->meta_write_recheck_parallelism || !(sqe = bs->get_sqe()))
+        if (wait_count >= bs->meta_write_recheck_parallelism)
         {
-            bs->ringloop->submit();
             wait_state = 8;
             return 1;
         }
         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*)memalign_or_die(MEM_ALIGNMENT, bs->dsk.meta_block_size);
+        uint8_t *buf = (uint8_t*)bs->dsk.meta_disk->alloc_iomem(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)
-        {
-            wait_count--;
-            free(buf);
-            if (data->res != bs->dsk.meta_block_size)
+        res = bs->dsk.meta_disk->write_async(block_offset, bs->dsk.meta_block_size, buf,
+            [this, buf, block_offset](int res)
             {
-                throw std::runtime_error(
-                    "write metadata failed at offset " + std::to_string(block_offset) + ": " + strerror(-data->res)
-                );
+                wait_count--;
+                bs->dsk.meta_disk->free_iomem(buf);
+                if (res < 0)
+                {
+                    throw std::runtime_error(
+                        "write metadata failed at offset " + std::to_string(block_offset) + ": " + strerror(-res)
+                    );
+                }
             }
-        };
-        io_uring_prep_writev(sqe, bs->dsk.meta_fd, &data->iov, 1, block_offset);
+        );
+        if (res == -EAGAIN)
+        {
+            bs->dsk.meta_disk->free_iomem(buf);
+            wait_state = 8;
+            return 1;
+        }
+        assert(!res);
         wait_count++;
     }
 resume_9:
@@ -340,18 +344,18 @@ resume_9:
         wait_state = 9;
         return 1;
     }
-    free(metadata_buffer);
+    bs->dsk.meta_disk->free_iomem(metadata_buffer);
     metadata_buffer = NULL;
 do_fsync:
     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, "fsync metadata"); };
-        submitted++;
-        bs->ringloop->submit();
+        res = bs->dsk.meta_disk->flush_async([this](int res) { handle_event(res, -1, "fsync metadata"); });
+        if (res != -EALREADY)
+        {
+            assert(res == 0);
+            submitted++;
+        }
     resume_5:
         if (submitted > 0)
         {
@@ -361,13 +365,13 @@ do_fsync:
     }
     if (zero_on_init && !header_written && !bs->readonly)
     {
-        GET_SQE();
-        header_written = true;
         last_read_offset = 0;
-        data->iov = (struct iovec){ bs->meta_superblock, (size_t)bs->dsk.meta_block_size };
-        data->callback = [this](ring_data_t *data) { handle_event(data, -1, "write metadata header"); };
-        io_uring_prep_writev(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset);
-        bs->ringloop->submit();
+        res = bs->dsk.meta_disk->write_async(
+            bs->dsk.meta_offset, (size_t)bs->dsk.meta_block_size, bs->meta_superblock,
+            [this](int res) { handle_event(res, -1, "write metadata header"); }
+        );
+        assert(res == 0);
+        header_written = true;
         submitted++;
     resume_2:
         if (submitted > 0)

src/blockstore/blockstore_init.h

@@ -21,19 +21,19 @@ class blockstore_init_meta
     bool zero_on_init = false;
     bool header_written = false;
     void *metadata_buffer = NULL;
+    int res = 0;
     blockstore_init_meta_buf bufs[2] = {};
     int submitted = 0;
-    struct io_uring_sqe *sqe;
-    struct ring_data_t *data;
     uint64_t md_offset = 0;
     uint64_t next_offset = 0;
     uint64_t last_read_offset = 0;
     uint64_t entries_loaded = 0;
+    uint64_t submit_len = 0;
     std::vector<uint32_t> recheck_mod;
     std::vector<blockstore_recheck_io_t*> recheck_ios;
     int i = 0, j = 0;
     bool handle_meta_block(uint8_t *buf, uint64_t count, uint64_t done_cnt);
-    void handle_event(ring_data_t *data, int buf_num, const char *op);
+    void handle_event(int res, int buf_num, const char *op);
 public:
     blockstore_init_meta(blockstore_impl_t *bs);
     int loop();

src/blockstore/blockstore_internal.h

@@ -3,44 +3,40 @@
 
 #pragma once
 
-#define BS_SUBMIT_CHECK_SQES(n) \
-    if (ringloop->space_left() < (n))\
+#define BS_SUBMIT_CHECK_JOURNAL_QUEUE(n) {\
+    if (dsk.journal_disk->get_queue_space() < (n))\
     {\
         /* Pause until there are more requests available */\
         PRIV(op)->wait_detail = (n);\
-        PRIV(op)->wait_for = WAIT_SQE;\
+        PRIV(op)->wait_for = WAIT_JOURNAL_QUEUE;\
         return 0;\
-    }
+    } }
 
-#define BS_SUBMIT_GET_SQE(sqe, data) \
-    BS_SUBMIT_GET_ONLY_SQE(sqe); \
-    struct ring_data_t *data = ((ring_data_t*)sqe->user_data)
-
-#define BS_SUBMIT_GET_ONLY_SQE(sqe) \
-    struct io_uring_sqe *sqe = get_sqe();\
-    if (!sqe)\
+#define BS_SUBMIT_CHECK_DATA_QUEUE(n) {\
+    if (dsk.data_disk->get_queue_space() < (n))\
     {\
         /* Pause until there are more requests available */\
-        PRIV(op)->wait_detail = 1;\
-        PRIV(op)->wait_for = WAIT_SQE;\
+        PRIV(op)->wait_detail = (n);\
+        PRIV(op)->wait_for = WAIT_DATA_QUEUE;\
         return 0;\
-    }
+    } }
 
-#define BS_SUBMIT_GET_SQE_DECL(sqe) \
-    sqe = get_sqe();\
-    if (!sqe)\
+#define BS_SUBMIT_CHECK_META_QUEUE(n) {\
+    if (dsk.meta_disk->get_queue_space() < (n))\
     {\
         /* Pause until there are more requests available */\
-        PRIV(op)->wait_detail = 1;\
-        PRIV(op)->wait_for = WAIT_SQE;\
+        PRIV(op)->wait_detail = (n);\
+        PRIV(op)->wait_for = WAIT_META_QUEUE;\
         return 0;\
-    }
+    } }
 
 #define PRIV(op) ((blockstore_op_private_t*)(op)->private_data)
 #define FINISH_OP(op) PRIV(op)->~blockstore_op_private_t(); std::function<void (blockstore_op_t*)>(op->callback)(op)
 
 // Suspend operation until there are more free SQEs
-#define WAIT_SQE 1
+#define WAIT_DATA_QUEUE 1
+#define WAIT_META_QUEUE 2
+#define WAIT_JOURNAL_QUEUE 3
 // Suspend operation until there are <wait_detail> bytes of free space in the journal on disk
 #define WAIT_COMPACTION 2
 // Suspend operation until LSN <wait_detail> becomes durable
@@ -65,11 +61,3 @@
 #define COPY_BUF_COALESCED  0x10
 #define COPY_BUF_PADDED     0x20
 #define COPY_BUF_SKIP_CSUM  0x40
-
-#ifndef RWF_ATOMIC
-#define RWF_ATOMIC 0x40
-#endif
-
-#ifndef RWF_DSYNC
-#define RWF_DSYNC 0x02
-#endif

src/blockstore/blockstore_read.cpp

@@ -88,7 +88,7 @@ int blockstore_impl_t::dequeue_read(blockstore_op_t *op)
     if (need_wait)
     {
 undo_wait:
-        // Need to wait. undo added requests, unlock lsn
+        // Unlock LSN and wait
         heap->unlock_entry(op->oid);
         free_read_buffers(rv);
         rv.clear();
@@ -106,12 +106,13 @@ undo_wait:
     assert(fulfilled == op->len);
     if (!fulfill_read(op))
     {
+        // Not enough SQEs. Unlock lsn and wait
         goto undo_wait;
     }
     op->version = result_version;
     if (!PRIV(op)->pending_ops)
     {
-        // everything is fulfilled from memory
+        // Everything is fulfilled from memory
         heap->unlock_entry(op->oid);
         op->retval = op->len;
         free_read_buffers(rv);
@@ -125,6 +126,24 @@ undo_wait:
 
 int blockstore_impl_t::fulfill_read(blockstore_op_t *op)
 {
+    int ndata = 0, njournal = 0;
+    uint32_t nodsk = (COPY_BUF_COALESCED|COPY_BUF_CSUM_FILL|COPY_BUF_ZERO) |
+        (dsk.inmemory_journal ? COPY_BUF_JOURNAL : 0);
+    for (auto & vec: PRIV(op)->read_vec)
+    {
+        if (!(vec.copy_flags & nodsk))
+        {
+            auto & n = (vec.copy_flags & COPY_BUF_JOURNAL) ? njournal : ndata;
+            n++;
+        }
+    }
+    if (!dsk.data_off_ring && !dsk.journal_off_ring || dsk.data_disk == dsk.journal_disk)
+    {
+        ndata += njournal;
+        njournal = 0;
+    }
+    BS_SUBMIT_CHECK_DATA_QUEUE(ndata);
+    BS_SUBMIT_CHECK_JOURNAL_QUEUE(njournal);
     for (auto & vec: PRIV(op)->read_vec)
     {
         if (vec.copy_flags & (COPY_BUF_COALESCED|COPY_BUF_CSUM_FILL))
@@ -142,16 +161,14 @@ int blockstore_impl_t::fulfill_read(blockstore_op_t *op)
         }
         else
         {
-            BS_SUBMIT_GET_SQE(sqe, data);
-            data->iov = (struct iovec){ vec.buf ? vec.buf : (op->buf + vec.offset - op->offset), (size_t)vec.disk_len };
-            PRIV(op)->pending_ops++;
-            io_uring_prep_readv(
-                sqe,
-                (vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_fd : dsk.data_fd,
-                &data->iov, 1,
-                ((vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_offset : dsk.data_offset) + vec.disk_loc + vec.disk_offset
+            disk_i *di = ((vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_disk : dsk.data_disk);
+            int res = di->read_async(
+                ((vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_offset : dsk.data_offset) + vec.disk_loc + vec.disk_offset,
+                vec.disk_len, vec.buf ? vec.buf : (op->buf + vec.offset - op->offset),
+                [this, op](int res) { handle_read_event(res, op); }
             );
-            data->callback = [this, op](ring_data_t *data) { handle_read_event(data, op); };
+            assert(res == 0);
+            PRIV(op)->pending_ops++;
         }
     }
     return 1;
@@ -332,7 +349,7 @@ void blockstore_impl_t::prepare_disk_read(std::vector<copy_buffer_t> & read_vec,
         }
         else
         {
-            vec.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, vec.disk_len);
+            vec.buf = (uint8_t*)((vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_disk : dsk.data_disk)->alloc_iomem(vec.disk_len);
         }
     }
     read_vec.insert(read_vec.begin() + pos, vec);
@@ -383,21 +400,22 @@ void blockstore_impl_t::free_read_buffers(std::vector<copy_buffer_t> & rv)
             if (!(vec.copy_flags & COPY_BUF_COALESCED) && vec.buf &&
                 (!buffer_area || vec.buf < buffer_area || vec.buf >= (uint8_t*)buffer_area + dsk.journal_len))
             {
-                free(vec.buf);
+                disk_i *di = (vec.copy_flags & COPY_BUF_JOURNAL) ? dsk.journal_disk : dsk.data_disk;
+                di->free_iomem(vec.buf);
                 vec.buf = NULL;
             }
         }
     }
 }
 
-void blockstore_impl_t::handle_read_event(ring_data_t *data, blockstore_op_t *op)
+void blockstore_impl_t::handle_read_event(int res, blockstore_op_t *op)
 {
     live = true;
     PRIV(op)->pending_ops--;
-    if (data->res != data->iov.iov_len)
+    if (res < 0)
     {
         // read error
-        op->retval = data->res;
+        op->retval = res;
     }
     if (PRIV(op)->pending_ops == 0)
     {

src/blockstore/blockstore_stable.cpp

@@ -86,7 +86,7 @@ resume_1:
         if (priv->modified_block != UINT32_MAX && priv->modified_block2 != priv->modified_block)
         {
             priv->op_state = 1;
-            BS_SUBMIT_CHECK_SQES(1);
+            BS_SUBMIT_CHECK_META_QUEUE(1);
             prepare_meta_block_write(priv->modified_block);
 resume_2:
             if (meta_block_is_pending(priv->modified_block))

src/blockstore/blockstore_sync.cpp

@@ -27,71 +27,72 @@ bool blockstore_impl_t::has_unsynced()
     return data || buffer || meta;
 }
 
-bool blockstore_impl_t::submit_fsyncs(int & wait_count)
+int blockstore_impl_t::submit_fsyncs(int & wait_count)
 {
-    int n = (unsynced_meta_write_count > 0 && !dsk.disable_meta_fsync ? 1 : 0) +
-        (unsynced_buffer_write_count > 0 && !dsk.disable_journal_fsync &&
-            (!unsynced_meta_write_count || dsk.journal_fd != dsk.meta_fd) ? 1 : 0) +
-        (unsynced_data_write_count > 0 && !dsk.disable_data_fsync &&
-            (!unsynced_meta_write_count || dsk.data_fd != dsk.meta_fd) &&
-            (!unsynced_buffer_write_count || dsk.data_fd != dsk.journal_fd) ? 1 : 0);
-    if (ringloop->space_left() < n)
+    int nm = (unsynced_meta_write_count > 0 && !dsk.disable_meta_fsync ? 1 : 0);
+    int nj = (unsynced_buffer_write_count > 0 && !dsk.disable_journal_fsync &&
+        (!unsynced_meta_write_count || dsk.journal_disk != dsk.meta_disk) ? 1 : 0);
+    int nd = (unsynced_data_write_count > 0 && !dsk.disable_data_fsync &&
+        (!unsynced_meta_write_count || dsk.data_disk != dsk.meta_disk) &&
+        (!unsynced_buffer_write_count || dsk.data_disk != dsk.journal_disk) ? 1 : 0);
+    if (!nm && !nj && !nd)
+        return 0;
+    int rnm = nm, rnj = nj, rnd = nd;
+    if (!dsk.meta_off_ring && !dsk.journal_off_ring)
+        rnm += rnj, rnj = 0;
+    if (!dsk.meta_off_ring && !dsk.data_off_ring)
+        rnm += rnd, rnd = 0;
+    if (!dsk.journal_off_ring && !dsk.data_off_ring)
+        rnj += rnd, rnd = 0;
+    if (rnm > 0 && dsk.meta_disk->get_queue_space() < rnm)
+        return WAIT_META_QUEUE;
+    if (rnj > 0 && dsk.journal_disk->get_queue_space() < rnj)
+        return WAIT_JOURNAL_QUEUE;
+    if (rnd > 0 && dsk.data_disk->get_queue_space() < rnd)
+        return WAIT_DATA_QUEUE;
+    auto cb = [this, & wait_count](int res)
     {
-        return false;
-    }
-    if (!n)
-    {
-        return true;
-    }
-    auto cb = [this, & wait_count](ring_data_t *data)
-    {
-        if (data->res != 0)
-            disk_error_abort("sync meta", data->res, 0);
+        if (res != 0)
+            disk_error_abort("sync meta", res);
         wait_count--;
         assert(wait_count >= 0);
         if (!wait_count)
             ringloop->wakeup();
     };
-    if (unsynced_meta_write_count > 0 && !dsk.disable_meta_fsync)
+    if (nm)
     {
         // fsync meta
-        io_uring_sqe *sqe = get_sqe();
-        assert(sqe);
-        ring_data_t *data = ((ring_data_t*)sqe->user_data);
-        io_uring_prep_fsync(sqe, dsk.meta_fd, IORING_FSYNC_DATASYNC);
-        data->iov = { 0 };
-        data->callback = cb;
-        wait_count++;
+        int res = dsk.meta_disk->flush_async(cb);
+        if (res != -EALREADY)
+        {
+            assert(res == 0);
+            wait_count++;
+        }
     }
-    if (unsynced_buffer_write_count > 0 && !dsk.disable_journal_fsync &&
-        (!unsynced_meta_write_count || dsk.journal_fd != dsk.meta_fd))
+    if (nj)
     {
         // fsync buffer
-        io_uring_sqe *sqe = get_sqe();
-        assert(sqe);
-        ring_data_t *data = ((ring_data_t*)sqe->user_data);
-        io_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
-        data->iov = { 0 };
-        data->callback = cb;
-        wait_count++;
+        int res = dsk.journal_disk->flush_async(cb);
+        if (res != -EALREADY)
+        {
+            assert(res == 0);
+            wait_count++;
+        }
     }
-    if (unsynced_data_write_count > 0 && !dsk.disable_data_fsync &&
-        (!unsynced_meta_write_count || dsk.data_fd != dsk.meta_fd) &&
-        (!unsynced_buffer_write_count || dsk.data_fd != dsk.journal_fd))
+    if (nd)
     {
         // fsync data
-        io_uring_sqe *sqe = get_sqe();
-        assert(sqe);
-        ring_data_t *data = ((ring_data_t*)sqe->user_data);
-        io_uring_prep_fsync(sqe, dsk.data_fd, IORING_FSYNC_DATASYNC);
-        data->iov = { 0 };
-        data->callback = cb;
-        wait_count++;
+        int res = dsk.data_disk->flush_async(cb);
+        if (res != -EALREADY)
+        {
+            assert(res == 0);
+            wait_count++;
+        }
     }
     unsynced_data_write_count = 0;
     unsynced_buffer_write_count = 0;
     unsynced_meta_write_count = 0;
-    return true;
+    return 0;
 }
 
 int blockstore_impl_t::do_sync(blockstore_op_t *op, int base_state)
@@ -113,11 +114,10 @@ int blockstore_impl_t::do_sync(blockstore_op_t *op, int base_state)
     }
     assert(!PRIV(op)->lsn);
     PRIV(op)->lsn = heap->get_completed_lsn();
-    if (!submit_fsyncs(PRIV(op)->pending_ops))
+    PRIV(op)->wait_for = submit_fsyncs(PRIV(op)->pending_ops);
+    if (PRIV(op)->wait_for)
     {
         PRIV(op)->lsn = 0;
-        PRIV(op)->wait_detail = 1;
-        PRIV(op)->wait_for = WAIT_SQE;
         return 0;
     }
 resume_1:

src/blockstore/blockstore_write.cpp

@@ -16,32 +16,31 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
 void blockstore_impl_t::prepare_meta_block_write(uint32_t modified_block)
 {
     if (modified_blocks.find(modified_block) != modified_blocks.end())
-        return;
-    io_uring_sqe *sqe = get_sqe();
-    assert(sqe != NULL);
-    ring_data_t *data = ((ring_data_t*)sqe->user_data);
-    uint8_t *buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.meta_block_size);
-    data->iov = (struct iovec){ buf, (size_t)dsk.meta_block_size };
-    data->callback = [this, modified_block](ring_data_t *data)
     {
-        live = true;
-        if (data->res != data->iov.iov_len)
-        {
-            // FIXME: our state becomes corrupted after a write error. maybe do something better than just die
-            disk_error_abort("data write", data->res, data->iov.iov_len);
-        }
-        auto it = modified_blocks.find(modified_block);
-        assert(it != modified_blocks.end());
-        free(it->second.buf);
-        modified_blocks.erase(it);
-        unsynced_meta_write_count++;
-        heap->complete_block_write(modified_block);
-        ringloop->wakeup();
-    };
+        return;
+    }
+    uint8_t *buf = (uint8_t*)dsk.meta_disk->alloc_iomem(dsk.meta_block_size);
     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
+    int res = dsk.meta_disk->write_async(
+        dsk.meta_offset + ((uint64_t)modified_block+1)*dsk.meta_block_size, dsk.meta_block_size, buf,
+        [this, modified_block](int res)
+        {
+            live = true;
+            if (res < 0)
+            {
+                // FIXME: our state becomes corrupted after a write error. maybe do something better than just die
+                disk_error_abort("data write", res);
+            }
+            auto it = modified_blocks.find(modified_block);
+            assert(it != modified_blocks.end());
+            dsk.meta_disk->free_iomem(it->second.buf);
+            modified_blocks.erase(it);
+            unsynced_meta_write_count++;
+            heap->complete_block_write(modified_block);
+            ringloop->wakeup();
+        }
     );
+    assert(res == 0);
     pending_modified_blocks.push_back(modified_block);
     modified_blocks[modified_block] = { .sent = false, .buf = buf };
 }
@@ -135,7 +134,7 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             return 2;
         }
         PRIV(op)->write_type = BS_HEAP_DELETE;
-        BS_SUBMIT_CHECK_SQES(1);
+        BS_SUBMIT_CHECK_META_QUEUE(1);
         int res = heap->add_delete(obj, &PRIV(op)->modified_block);
         assert(res != EAGAIN);
         if (res == ENOSPC)
@@ -152,7 +151,7 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
     {
         // Big (redirect) write
         PRIV(op)->write_type = dsk.disable_data_fsync || op->opcode != BS_OP_WRITE_STABLE ? BS_HEAP_BIG_WRITE : _REDIRECT_INTENT;
-        BS_SUBMIT_CHECK_SQES(1);
+        BS_SUBMIT_CHECK_DATA_QUEUE(1);
         PRIV(op)->location = heap->find_free_data();
         if (PRIV(op)->location == UINT64_MAX)
         {
@@ -177,12 +176,12 @@ enospc:
         );
 #endif
         heap->use_data(op->oid.inode, PRIV(op)->location);
-        io_uring_sqe *sqe = get_sqe();
-        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);
+        int res = dsk.data_disk->write_async(
+            dsk.data_offset + loc + op->offset, op->len, op->buf,
+            [this, op](int res) { handle_write_event(res, op); }
+        );
+        assert(res == 0);
         if (!dsk.disable_data_fsync)
         {
             // use PRIV->lsn for fsync_data_id
@@ -199,7 +198,7 @@ enospc:
     else if (intent_write_allowed(op, obj))
     {
         // Direct intent-write
-        BS_SUBMIT_CHECK_SQES(1);
+        BS_SUBMIT_CHECK_META_QUEUE(1);
         int res = 0;
         if (dsk.csum_block_size <= dsk.bitmap_granularity &&
             (obj->entry_type == (BS_HEAP_BIG_INTENT|BS_HEAP_STABLE) ||
@@ -260,7 +259,15 @@ enospc:
             return 0;
         }
         // There is sufficient space. Check SQE(s)
-        BS_SUBMIT_CHECK_SQES(1 + (op->len > 0 ? 1 : 0));
+        if (!op->len || dsk.journal_disk == dsk.meta_disk || !dsk.journal_off_ring && !dsk.meta_off_ring)
+        {
+            BS_SUBMIT_CHECK_META_QUEUE(1 + (op->len > 0 ? 1 : 0));
+        }
+        else
+        {
+            BS_SUBMIT_CHECK_META_QUEUE(1);
+            BS_SUBMIT_CHECK_JOURNAL_QUEUE(1);
+        }
         int res = heap->add_small_write(op->oid, &obj, (BS_HEAP_SMALL_WRITE | (op->opcode == BS_OP_WRITE_STABLE ? BS_HEAP_STABLE : 0)),
             op->version, op->offset, op->len, loc, op->bitmap, (uint8_t*)op->buf, &PRIV(op)->modified_block);
         BS_SUBMIT_CHECK_PLACEMENT(res, obj->lsn);
@@ -278,11 +285,12 @@ enospc:
             {
                 memcpy((uint8_t*)buffer_area + loc, op->buf, op->len);
             }
-            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);
+            res = dsk.journal_disk->write_async(
+                dsk.journal_offset + loc, op->len, op->buf,
+                [this, op](int res) { handle_write_event(res, op); }
+            );
+            assert(res == 0);
             PRIV(op)->pending_ops++;
         }
         else
@@ -298,6 +306,7 @@ enospc:
 
 int blockstore_impl_t::continue_write(blockstore_op_t *op)
 {
+    int res = 0;
     int op_state = PRIV(op)->op_state;
 again:
     if (op_state == 2)
@@ -350,21 +359,29 @@ resume_11:
         }
         if (PRIV(op)->lsn > data_fsync_sent)
         {
-            BS_SUBMIT_GET_SQE(sqe, data);
-            io_uring_prep_fsync(sqe, dsk.data_fd, IORING_FSYNC_DATASYNC);
-            data->iov = { 0 };
-            data->callback = [this, op, fs = data_fsync_cur](ring_data_t *data)
+            res = dsk.data_disk->flush_async([this, op, fs = data_fsync_cur](int res)
             {
                 live = true;
-                if (data->res != 0)
-                    disk_error_abort("sync data", data->res, 0);
+                if (res != 0)
+                    disk_error_abort("sync data", res);
                 if (fs > data_fsync_done)
                 {
                     data_fsync_done = fs;
                     ringloop->wakeup();
                 }
-            };
-            data_fsync_sent = data_fsync_cur;
+            });
+            if (res == -EAGAIN)
+            {
+                return 1;
+            }
+            else
+            {
+                data_fsync_sent = data_fsync_cur;
+                if (res == -EALREADY)
+                    data_fsync_done = data_fsync_cur;
+                else
+                    assert(res == 0);
+            }
         }
         if (PRIV(op)->lsn > data_fsync_done)
         {
@@ -374,9 +391,8 @@ resume_11:
     }
 resume_4:
     {
-        BS_SUBMIT_CHECK_SQES(1);
+        BS_SUBMIT_CHECK_META_QUEUE(1);
         auto obj = heap->read_entry(op->oid);
-        int res = 0;
         if (PRIV(op)->write_type == _REDIRECT_INTENT)
         {
             res = heap->add_redirect_intent(op->oid, &obj, op->version, op->offset, op->len,
@@ -476,25 +492,28 @@ resume_8:
 resume_10:
     // Direct intent-write
     // LSN is not marked as completed so big_write won't be freed
-    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); };
+    BS_SUBMIT_CHECK_DATA_QUEUE(1);
     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;
+    // FIXME: We tried to use RWF_ATOMIC here, but it's useless because Linux
+    // incorrectly enforces 2^N atomic write alignment which is a blocker for us.
+    // So now it's removed.
+    res = dsk.data_disk->write_async(
+        dsk.data_offset + PRIV(op)->location + op->offset, op->len, op->buf,
+        [this, op](int res) { handle_write_event(res, op); }
+    );
+    assert(res == 0);
     PRIV(op)->pending_ops++;
     PRIV(op)->op_state = 7;
     return 1;
 }
 
-void blockstore_impl_t::handle_write_event(ring_data_t *data, blockstore_op_t *op)
+void blockstore_impl_t::handle_write_event(int res, blockstore_op_t *op)
 {
     live = true;
-    if (data->res != data->iov.iov_len)
+    if (res < 0)
     {
         // FIXME: our state becomes corrupted after a write error. maybe do something better than just die
-        disk_error_abort("data write", data->res, data->iov.iov_len);
+        disk_error_abort("data write", res);
     }
     PRIV(op)->pending_ops--;
     assert(PRIV(op)->pending_ops >= 0);

src/blockstore/v1/impl.cpp

@@ -1,6 +1,7 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
+#include "disk_util.h"
 #include "impl.h"
 #include "internal.h"
 
@@ -22,6 +23,9 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_i *
         dsk.open_meta();
         dsk.open_journal();
         calc_lengths();
+        dsk.data_fd = dsk.data_disk->get_fd();
+        dsk.meta_fd = dsk.meta_disk->get_fd();
+        dsk.journal_fd = dsk.journal_disk->get_fd();
         alloc_dyn_data = dsk.clean_dyn_size > sizeof(void*) || dsk.csum_block_size > 0;
         zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.data_block_size);
         memset(zero_object, 0, dsk.data_block_size);

src/client/cluster_client.cpp

@@ -48,7 +48,7 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
     msgr.exec_op = [this](osd_op_t *op)
     {
         // Garbage in
-        fprintf(stderr, "Can't handle incoming operation from client %lu\n", op->client_id);
+        fprintf(stderr, "Can't handle incoming operation from client %ju\n", op->client_id);
         msgr.stop_client(op->client_id);
         delete op;
     };
@@ -591,7 +591,7 @@ void cluster_client_t::on_change_pool_config_hook()
         {
             if (log_level > 2 && pg_counts[pool_item.first])
             {
-                fprintf(stderr, "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 %ju to %ju\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

src/client/msgr_receive.cpp

@@ -700,15 +700,13 @@ bool osd_messenger_t::handle_hdr(osd_client_t *cl)
     }
     else if (cl->read_op->req.hdr.magic == SECONDARY_OSD_OP_MAGIC)
     {
-        if (cl->check_sequencing)
+        if (cl->check_sequencing && cl->read_op->req.hdr.id != cl->read_op_id)
         {
-            if (cl->read_op->req.hdr.id != cl->read_op_id)
-            {
-                fprintf(stderr, "Warning: operation sequencing is broken on client %d: expected num %ju, got %ju, stopping client\n", cl->peer_fd, cl->read_op_id, cl->read_op->req.hdr.id);
-                return false;
-            }
-            cl->read_op_id++;
+            fprintf(stderr, "Warning: operation sequencing is broken on client %ju: expected num %ju, got %ju, stopping client\n", cl->client_id, cl->read_op_id, cl->read_op->req.hdr.id);
+            return false;
         }
+        // Track expected op_ids before check_sequencing to not break if a PING sneaks before SHOW_CONFIG reply
+        cl->read_op_id = cl->read_op->req.hdr.id + 1;
         if (!allocate_op_buffers(cl))
         {
             return false;
@@ -775,7 +773,7 @@ bool osd_messenger_t::allocate_op_buffers(osd_client_t *cl)
             {
                 if (log_level > 1)
                 {
-                    fprintf(stderr, "Error: peer %ju stabilize request size too large (%lu > %u bytes), stopping\n", cl->client_id,
+                    fprintf(stderr, "Error: peer %ju stabilize request size too large (%ju > %u bytes), stopping\n", cl->client_id,
                         cur_op->req.sec_stab.len, MAX_SIMPLE_PAYLOAD_SIZE);
                 }
                 return false;
@@ -792,7 +790,7 @@ bool osd_messenger_t::allocate_op_buffers(osd_client_t *cl)
             {
                 if (log_level > 1)
                 {
-                    fprintf(stderr, "Error: peer %ju sec_read_bmp request size too large (%lu > %u bytes), stopping\n", cl->client_id,
+                    fprintf(stderr, "Error: peer %ju sec_read_bmp request size too large (%ju > %u bytes), stopping\n", cl->client_id,
                         cur_op->req.sec_read_bmp.len, MAX_SIMPLE_PAYLOAD_SIZE);
                 }
                 return false;
@@ -826,7 +824,7 @@ bool osd_messenger_t::allocate_op_buffers(osd_client_t *cl)
             {
                 if (log_level > 1)
                 {
-                    fprintf(stderr, "Error: peer %ju show_config request length too large (%lu > %u bytes), stopping\n", cl->client_id,
+                    fprintf(stderr, "Error: peer %ju show_config request length too large (%ju > %u bytes), stopping\n", cl->client_id,
                         cur_op->req.show_conf.json_len, MAX_SIMPLE_PAYLOAD_SIZE);
                 }
                 return false;

src/client/qemu_driver.c

@@ -329,7 +329,9 @@ static void coroutine_fn vitastor_co_resize(void *opaque)
     }
 #else
     BdrvChild *c;
+#if QEMU_VERSION_MAJOR >= 8
     WITH_GRAPH_RDLOCK_GUARD() {
+#endif
 #if QEMU_VERSION_MAJOR >= 9
         assert_bdrv_graph_readable();
 #endif
@@ -344,7 +346,9 @@ static void coroutine_fn vitastor_co_resize(void *opaque)
             }
 #endif
         }
+#if QEMU_VERSION_MAJOR >= 8
     }
+#endif
 #endif
     bdrv_dec_in_flight(bs);
 }

src/client/ublk_server.cpp

@@ -706,7 +706,7 @@ protected:
         {
             if (data->res != 0)
             {
-                fprintf(stderr, "Failed to update size of /dev/ublkb%d to %lu bytes: %s (code %d)\n", ublk_dev.dev_id, size, strerror(-data->res), data->res);
+                fprintf(stderr, "Failed to update size of /dev/ublkb%d to %ju bytes: %s (code %d)\n", ublk_dev.dev_id, size, strerror(-data->res), data->res);
             }
         };
     }

src/cmd/cli_merge.cpp

@@ -583,7 +583,7 @@ struct snap_merger_t
                 if (use_cas && subop->retval == -EINTR)
                 {
                     // CAS failure - reread and repeat optimistically
-                    assert(rwo->todo == 1); // initial refcount from read_and_write
+                    assert(!rwo->todo);
                     rwo->error_code = -EINTR;
                     rwo->start = rwo->end = 0;
                     rwo->op.version = 0;

src/disk_tool/CMakeLists.txt

@@ -10,6 +10,7 @@ add_executable(vitastor-disk
 	disk_tool.h
 	disk_tool_journal.cpp
 	disk_tool_meta.cpp
+	disk_tool_nvme.cpp
 	disk_tool_prepare.cpp
 	disk_tool_resize_auto.cpp
 	disk_tool_resize.cpp

src/disk_tool/disk_simple_offsets.cpp

@@ -1,16 +1,11 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
-#include <fcntl.h>
-#include <sys/ioctl.h>
-#include <ctype.h>
-#include <unistd.h>
-#include <sys/stat.h>
-
 #include "json11/json11.hpp"
 #include "str_util.h"
 #include "blockstore.h"
 #include "blockstore_disk.h"
+#include "disk_util.h"
 #include "blockstore_heap.h"
 #include "ondisk_formats.h"
 
@@ -62,39 +57,23 @@ void disk_tool_simple_offsets(json11::Json cfg, bool json_output)
             fprintf(stderr, "Device path is missing\n");
             exit(1);
         }
-        struct stat st;
-        if (stat(device.c_str(), &st) < 0)
+        disk_i *disk = disk_i::create_and_open(device, "cached,ro", true);
+        if (!disk)
         {
-            fprintf(stderr, "Can't stat %s: %s\n", device.c_str(), strerror(errno));
+            fprintf(stderr, "Failed to open %s\n", device.c_str());
             exit(1);
         }
-        if (S_ISBLK(st.st_mode))
+        device_size = disk->get_size();
+        uint64_t blksize = disk->get_block_size();
+        if (blksize < device_block_size)
         {
-            int fd = open(device.c_str(), O_DIRECT|O_RDONLY);
-            if (fd < 0 || ioctl(fd, BLKGETSIZE64, &device_size) < 0)
-            {
-                fprintf(stderr, "Failed to get device size for %s: %s\n", device.c_str(), strerror(errno));
-                exit(1);
-            }
-            close(fd);
-            if (st.st_blksize < device_block_size)
-            {
-                fprintf(
-                    stderr, "Warning: %s reports %ju byte blocks, but we use %ju."
-                    " Set --device_block_size=%ju if you're sure it works well with %ju byte blocks.\n",
-                    device.c_str(), (uint64_t)st.st_blksize, device_block_size, (uint64_t)st.st_blksize, (uint64_t)st.st_blksize
-                );
-            }
-        }
-        else if (S_ISREG(st.st_mode))
-        {
-            device_size = st.st_size;
-        }
-        else
-        {
-            fprintf(stderr, "%s is neither a block device nor a regular file\n", device.c_str());
-            exit(1);
+            fprintf(
+                stderr, "Warning: %s reports %ju byte blocks, but we use %ju."
+                " Set --device_block_size=%ju if you're sure it works well with %ju byte blocks.\n",
+                device.c_str(), blksize, device_block_size, blksize, blksize
+            );
         }
+        delete disk;
     }
     if (!device_size)
     {

src/disk_tool/disk_tool.cpp

@@ -207,6 +207,9 @@ static const char *help_text =
     "  Write metadata from JSON taken from standard input in the same format as produced by `dump-meta`.\n"
     "  You can specify any OSD device (data, metadata or journal), or the layout manually.\n"
     "\n"
+    "vitastor-disk find-pci <serial> [nsid]\n"
+    "  Find PCI ID and attributes of an NVMe device with a specified serial number.\n"
+    "\n"
     "vitastor-disk simple-offsets <device>\n"
     "  Calculate offsets for old simple&stupid (no superblock) OSD deployment. Options:\n"
     "    --object_size 128k       Set blockstore block size\n"
@@ -473,6 +476,16 @@ int main(int argc, char *argv[])
     {
         return self.raw_resize();
     }
+    else if (!strcmp(cmd[0], "find-pci"))
+    {
+        if (cmd.size() != 2 && cmd.size() != 3)
+        {
+            fprintf(stderr, "S/N and optionally NS number is required\n");
+            return 1;
+        }
+        self.find_pci_id(cmd[1], cmd.size() == 3 ? stoull_full(cmd[2]) : 0);
+        return 0;
+    }
     else if (!strcmp(cmd[0], "simple-offsets"))
     {
         // Calculate offsets for simple & stupid OSD deployment without superblock

src/disk_tool/disk_tool.h

@@ -17,6 +17,7 @@
 #include "blockstore_disk.h"
 #include "blockstore.h"
 #include "blockstore_heap.h"
+#include "disk_util.h"
 #include "ondisk_formats.h"
 #include "crc32c.h"
 #include "allocator.h"
@@ -85,7 +86,7 @@ struct disk_tool_t
     uint64_t total_blocks = 0, free_first = 0, free_last = 0;
     uint64_t new_clean_entry_bitmap_size = 0, new_data_csum_size = 0, new_clean_entry_size = 0, new_entries_per_block = 0;
     uint32_t new_meta_format = 0;
-    int new_journal_fd = -1, new_meta_fd = -1;
+    disk_i *new_journal_disk = NULL, *new_meta_disk = NULL;
     resizer_data_moving_t *moving_blocks = NULL;
     bool atomic_warned = false;
 
@@ -152,6 +153,8 @@ struct disk_tool_t
     int clear_osd_superblock(const std::string & dev);
     int trim_data(std::string device);
 
+    std::string find_pci_id(const std::string & nvme_serial, uint32_t nvme_nsid);
+
     json11::Json read_osd_superblock(std::string device, bool expect_exist = true, bool ignore_nonref = false);
     uint32_t write_osd_superblock(std::string device, json11::Json params);
 
@@ -177,7 +180,6 @@ uint64_t get_atomic_write_size(const std::string & dev);
 uint64_t get_device_size(const std::string & dev, bool should_exist = false);
 std::string get_parent_device(std::string dev);
 int shell_exec(const std::vector<std::string> & cmd, const std::string & in, std::string *out, std::string *err);
-int write_zero(int fd, uint64_t offset, uint64_t size);
 json11::Json read_parttable(std::string dev);
 uint64_t dev_size_from_parttable(json11::Json pt);
 uint64_t free_from_parttable(json11::Json pt);

src/disk_tool/disk_tool_journal.cpp

@@ -22,18 +22,22 @@ int disk_tool_t::dump_journal()
         printf("[\n");
     if (all)
     {
-        dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
-        if (dsk.journal_fd < 0)
+        dsk.journal_disk = disk_i::create_and_open(dsk.journal_device, (options["io"] == "cached" ? "cached,ro" : "ro"), true);
+        if (!dsk.journal_disk)
         {
-            fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
+            fprintf(stderr, "Failed to open journal device %s\n", dsk.journal_device.c_str());
             return 1;
         }
-        void *journal_buf = memalign_or_die(MEM_ALIGNMENT, dsk.journal_block_size);
+        void *journal_buf = dsk.journal_disk->alloc_iomem(dsk.journal_block_size);
         journal_pos = 0;
         while (journal_pos < dsk.journal_len)
         {
-            int r = pread(dsk.journal_fd, journal_buf, dsk.journal_block_size, dsk.journal_offset+journal_pos);
-            assert(r == dsk.journal_block_size);
+            int r = dsk.journal_disk->read_sync(dsk.journal_offset+journal_pos, dsk.journal_block_size, journal_buf);
+            if (r != 0)
+            {
+                fprintf(stderr, "Failed to read journal block at %ju: %s (code %u)\n", dsk.journal_offset+journal_pos, strerror(-r), -r);
+                break;
+            }
             uint64_t s;
             for (s = 0; s < dsk.journal_block_size; s += 8)
             {
@@ -90,9 +94,9 @@ int disk_tool_t::dump_journal()
                 journal_pos += dsk.journal_block_size;
             }
         }
-        free(journal_buf);
-        close(dsk.journal_fd);
-        dsk.journal_fd = -1;
+        dsk.journal_disk->free_iomem(journal_buf);
+        delete dsk.journal_disk;
+        dsk.journal_disk = NULL;
     }
     else
     {
@@ -127,23 +131,28 @@ int disk_tool_t::process_journal(std::function<int(void*)> block_fn, bool do_ope
 {
     if (do_open)
     {
-        if (dsk.journal_fd >= 0)
+        if (dsk.journal_disk)
         {
             fprintf(stderr, "Bug: journal device is already opened\n");
             return 1;
         }
-        dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
-        if (dsk.journal_fd < 0)
+        dsk.journal_disk = disk_i::create_and_open(dsk.journal_device, (options["io"] == "cached" ? "cached,ro" : "ro"), true);
+        if (!dsk.journal_disk)
         {
-            fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
+            fprintf(stderr, "Failed to open journal device %s\n", dsk.journal_device.c_str());
             return 1;
         }
     }
-    void *data = memalign_or_die(MEM_ALIGNMENT, dsk.journal_block_size);
+    void *data = dsk.journal_disk->alloc_iomem(dsk.journal_block_size);
     journal_pos = 0;
-    int r = pread(dsk.journal_fd, data, dsk.journal_block_size, dsk.journal_offset+journal_pos);
-    assert(r == dsk.journal_block_size);
-    journal_entry *je = (journal_entry*)(data);
+    journal_entry *je = NULL;
+    int r = dsk.journal_disk->read_sync(dsk.journal_offset+journal_pos, dsk.journal_block_size, data);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to read journal header at %ju: %s (code %u)\n", dsk.journal_offset+journal_pos, strerror(-r), -r);
+        goto ret;
+    }
+    je = (journal_entry*)(data);
     if (je->magic != JOURNAL_MAGIC || je->type != JE_START || je_crc32(je) != je->crc32)
     {
         fprintf(stderr, "offset %08jx: journal superblock is invalid\n", journal_pos);
@@ -175,19 +184,28 @@ int disk_tool_t::process_journal(std::function<int(void*)> block_fn, bool do_ope
         {
             if (journal_pos >= dsk.journal_len)
                 journal_pos = dsk.journal_block_size;
-            r = pread(dsk.journal_fd, data, dsk.journal_block_size, dsk.journal_offset+journal_pos);
-            assert(r == dsk.journal_block_size);
+            r = dsk.journal_disk->read_sync(dsk.journal_offset+journal_pos, dsk.journal_block_size, data);
+            if (r != 0)
+            {
+                fprintf(stderr, "Failed to read journal block at %ju: %s (code %u)\n", dsk.journal_offset+journal_pos, strerror(-r), -r);
+                break;
+            }
             r = block_fn(data);
             if (r <= 0)
                 break;
         }
     }
+ret:
+    if (data)
+    {
+        dsk.journal_disk->free_iomem(data);
+        data = NULL;
+    }
     if (do_open)
     {
-        close(dsk.journal_fd);
-        dsk.journal_fd = -1;
+        delete dsk.journal_disk;
+        dsk.journal_disk = NULL;
     }
-    free(data);
     return r;
 }
 
@@ -227,8 +245,14 @@ int disk_tool_t::process_journal_block(void *buf, std::function<void(int, journa
                 journal_pos = dsk.journal_block_size;
                 wrapped = true;
             }
-            small_write_data = memalign_or_die(MEM_ALIGNMENT, je->small_write.len);
-            assert(pread(dsk.journal_fd, small_write_data, je->small_write.len, dsk.journal_offset+je->small_write.data_offset) == je->small_write.len);
+            small_write_data = dsk.journal_disk->alloc_iomem(je->small_write.len);
+            int r = dsk.journal_disk->read_sync(dsk.journal_offset+je->small_write.data_offset, je->small_write.len, small_write_data);
+            if (r != 0)
+            {
+                fprintf(stderr, "Failed to read %u journal bytes at %ju: %s (code %u)\n", je->small_write.len,
+                    dsk.journal_offset+je->small_write.data_offset, strerror(-r), -r);
+                break;
+            }
             data_crc32 = je_start.csum_block_size ? 0 : crc32c(0, small_write_data, je->small_write.len);
             data_csum_valid = (data_crc32 == je->small_write.crc32_data);
             if (je_start.csum_block_size && je->small_write.len > 0)
@@ -276,7 +300,7 @@ int disk_tool_t::process_journal_block(void *buf, std::function<void(int, journa
         iter_fn(entry, je);
         if (je->type == JE_SMALL_WRITE || je->type == JE_SMALL_WRITE_INSTANT)
         {
-            free(small_write_data);
+            dsk.journal_disk->free_iomem(small_write_data);
             small_write_data = NULL;
         }
         pos += je->size;
@@ -438,7 +462,7 @@ void disk_tool_t::dump_journal_entry(int num, journal_entry *je, bool json)
 
 int disk_tool_t::write_json_journal(json11::Json entries)
 {
-    new_journal_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_journal_len);
+    new_journal_buf = (uint8_t*)dsk.journal_disk->alloc_iomem(new_journal_len);
     new_journal_ptr = new_journal_buf;
     new_journal_data = new_journal_ptr + dsk.journal_block_size;
     new_journal_in_pos = 0;
@@ -497,7 +521,7 @@ int disk_tool_t::write_json_journal(json11::Json entries)
             if (new_journal_ptr-new_journal_buf >= new_journal_len)
             {
                 fprintf(stderr, "Error: entries don't fit to the new journal\n");
-                free(new_journal_buf);
+                dsk.journal_disk->free_iomem(new_journal_buf);
                 return 1;
             }
             new_journal_data = new_journal_ptr+dsk.journal_block_size;
@@ -505,7 +529,7 @@ int disk_tool_t::write_json_journal(json11::Json entries)
             if (dsk.journal_block_size < entry_size)
             {
                 fprintf(stderr, "Error: journal entry too large (%u bytes)\n", entry_size);
-                free(new_journal_buf);
+                dsk.journal_disk->free_iomem(new_journal_buf);
                 return 1;
             }
         }
@@ -515,7 +539,7 @@ int disk_tool_t::write_json_journal(json11::Json entries)
             if (new_journal_data - new_journal_buf + len > new_journal_len)
             {
                 fprintf(stderr, "Error: entries don't fit to the new journal\n");
-                free(new_journal_buf);
+                dsk.journal_disk->free_iomem(new_journal_buf);
                 return 1;
             }
             *((journal_entry_small_write*)ne) = (journal_entry_small_write){
@@ -538,7 +562,7 @@ int disk_tool_t::write_json_journal(json11::Json entries)
             if (ne->small_write.len > 0 && !rec["data"].is_string())
             {
                 fprintf(stderr, "Error: entry data is missing, please generate the dump with --json --format data\n");
-                free(new_journal_buf);
+                dsk.journal_disk->free_iomem(new_journal_buf);
                 return 1;
             }
             if (ne->small_write.len > 0)
@@ -599,6 +623,6 @@ int disk_tool_t::write_json_journal(json11::Json entries)
         new_journal_in_pos += ne->size;
     }
     int r = resize_write_new_journal();
-    free(new_journal_buf);
+    dsk.journal_disk->free_iomem(new_journal_buf);
     return r;
 }

src/disk_tool/disk_tool_meta.cpp

@@ -22,31 +22,36 @@ int disk_tool_t::process_meta(std::function<void(blockstore_meta_header_v3_t *)>
     if (buf_size % dsk.meta_block_size)
         buf_size = 8*dsk.meta_block_size;
     uint8_t *data = NULL;
-    data = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, buf_size);
-    blockstore_meta_header_v3_t *hdr = (blockstore_meta_header_v3_t *)memalign_or_die(MEM_ALIGNMENT, dsk.meta_block_size);
+    blockstore_meta_header_v3_t *hdr = NULL;
     if (do_open)
     {
-        if (dsk.meta_fd >= 0)
+        if (dsk.meta_disk)
         {
             fprintf(stderr, "Bug: Metadata device is already opened\n");
 close_error:
             r = 1;
             goto close_free;
         }
-        dsk.meta_fd = open(dsk.meta_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
-        if (dsk.meta_fd < 0)
+        dsk.meta_disk = disk_i::create_and_open(dsk.meta_device, (options["io"] == "cached" ? "cached,ro" : "ro"), true);
+        if (!dsk.meta_disk)
         {
-            fprintf(stderr, "Failed to open metadata device %s: %s\n", dsk.meta_device.c_str(), strerror(errno));
+            fprintf(stderr, "Failed to open metadata device %s\n", dsk.meta_device.c_str());
             goto close_error;
         }
     }
-    else if (dsk.meta_fd < 0)
+    else if (!dsk.meta_disk)
     {
         fprintf(stderr, "Bug: Metadata device is not opened\n");
         goto close_error;
     }
+    data = (uint8_t*)dsk.meta_disk->alloc_iomem(buf_size);
+    hdr = (blockstore_meta_header_v3_t *)dsk.meta_disk->alloc_iomem(dsk.meta_block_size);
     // Check superblock
-    pread_blocking(dsk.meta_fd, hdr, dsk.meta_block_size, dsk.meta_offset);
+    if ((r = dsk.meta_disk->read_sync(dsk.meta_offset, dsk.meta_block_size, hdr)) != 0)
+    {
+        fprintf(stderr, "Failed to read metadata header: %s (code %d)\n", strerror(-r), -r);
+        goto close_error;
+    }
     if (hdr->zero == 0 && hdr->magic == BLOCKSTORE_META_MAGIC_V1 && hdr->version == BLOCKSTORE_META_FORMAT_HEAP)
     {
         if (hdr->data_csum_type != 0 &&
@@ -62,35 +67,40 @@ close_error:
         // Load buffer_area
         if (with_data)
         {
-            buffer_area = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.journal_len);
             if (dsk.journal_device == dsk.meta_device || dsk.journal_device == "")
             {
-                dsk.journal_fd = dsk.meta_fd;
+                dsk.journal_disk = dsk.meta_disk;
             }
             else if (do_open)
             {
-                if (dsk.journal_fd >= 0)
+                if (dsk.journal_disk)
                 {
                     fprintf(stderr, "Bug: Metadata device is already opened\n");
                     goto close_error;
                 }
-                dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
-                if (dsk.journal_fd < 0)
+                dsk.journal_disk = disk_i::create_and_open(dsk.journal_device, (options["io"] == "cached" ? "cached,ro" : "ro"), true);
+                if (!dsk.journal_disk)
                 {
-                    fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
+                    fprintf(stderr, "Failed to open journal device %s\n", dsk.journal_device.c_str());
                     goto close_error;
                 }
             }
-            else if (dsk.journal_fd < 0)
+            else if (!dsk.journal_disk)
             {
                 fprintf(stderr, "Bug: journal device is not opened\n");
                 goto close_error;
             }
+            buffer_area = (uint8_t*)dsk.journal_disk->alloc_iomem(dsk.journal_len);
             uint64_t journal_pos = 0;
             while (journal_pos < dsk.journal_len)
             {
                 uint64_t read_len = buf_size < dsk.journal_len-journal_pos ? buf_size : dsk.journal_len-journal_pos;
-                pread_blocking(dsk.journal_fd, buffer_area+journal_pos, read_len, dsk.journal_offset+journal_pos);
+                if ((r = dsk.journal_disk->read_sync(dsk.journal_offset+journal_pos, read_len, buffer_area+journal_pos)) != 0)
+                {
+                    fprintf(stderr, "Failed to read %ju bytes at %ju from journal device %s: %s (code %d)\n",
+                        read_len, dsk.journal_offset+journal_pos, dsk.journal_device.c_str(), strerror(-r), -r);
+                    goto close_error;
+                }
                 journal_pos += read_len;
             }
         }
@@ -108,7 +118,12 @@ close_error:
         while (meta_pos < hdr->meta_area_size)
         {
             uint64_t read_len = buf_size < hdr->meta_area_size-meta_pos ? buf_size : hdr->meta_area_size-meta_pos;
-            pread_blocking(dsk.meta_fd, data, read_len, dsk.meta_offset+meta_pos);
+            if ((r = dsk.meta_disk->read_sync(dsk.meta_offset+meta_pos, read_len, data)) != 0)
+            {
+                fprintf(stderr, "Failed to read %ju bytes at %ju from metadata device %s: %s (code %d)\n",
+                    read_len, dsk.meta_offset+meta_pos, dsk.meta_device.c_str(), strerror(-r), -r);
+                goto close_error;
+            }
             r = heap->load_blocks(meta_pos-dsk.meta_block_size, read_len, data, true, entries_loaded);
             meta_pos += read_len;
         }
@@ -170,7 +185,12 @@ csum_unknown:
         while (meta_pos < dsk.min_meta_len)
         {
             uint64_t read_len = buf_size < dsk.min_meta_len-meta_pos ? buf_size : dsk.min_meta_len-meta_pos;
-            pread_blocking(dsk.meta_fd, data, read_len, dsk.meta_offset+meta_pos);
+            if ((r = dsk.meta_disk->read_sync(dsk.meta_offset+meta_pos, read_len, data)) != 0)
+            {
+                fprintf(stderr, "Failed to read %ju bytes at %ju from metadata device %s: %s (code %d)\n",
+                    read_len, dsk.meta_offset+meta_pos, dsk.meta_device.c_str(), strerror(-r), -r);
+                goto close_error;
+            }
             meta_pos += read_len;
             for (uint64_t blk = 0; blk < read_len; blk += dsk.meta_block_size)
             {
@@ -184,7 +204,7 @@ csum_unknown:
                             uint32_t *entry_csum = (uint32_t*)((uint8_t*)entry + dsk.clean_entry_size - 4);
                             if (*entry_csum != crc32c(0, entry, dsk.clean_entry_size - 4))
                             {
-                                fprintf(stderr, "Metadata entry %lu is corrupt (checksum mismatch), skipping\n", block_num);
+                                fprintf(stderr, "Metadata entry %ju is corrupt (checksum mismatch), skipping\n", block_num);
                                 continue;
                             }
                         }
@@ -205,7 +225,12 @@ csum_unknown:
         while (meta_pos < dsk.min_meta_len)
         {
             uint64_t read_len = buf_size < dsk.min_meta_len-meta_pos ? buf_size : dsk.min_meta_len-meta_pos;
-            pread_blocking(dsk.meta_fd, data, read_len, dsk.meta_offset+meta_pos);
+            if ((r = dsk.meta_disk->read_sync(dsk.meta_offset+meta_pos, read_len, data)) != 0)
+            {
+                fprintf(stderr, "Failed to read %ju bytes at %ju from metadata device %s: %s (code %d)\n",
+                    read_len, dsk.meta_offset+meta_pos, dsk.meta_device.c_str(), strerror(-r), -r);
+                goto close_error;
+            }
             meta_pos += read_len;
             for (uint64_t blk = 0; blk < read_len; blk += dsk.meta_block_size)
             {
@@ -221,23 +246,29 @@ csum_unknown:
         }
     }
 close_free:
-    free(data);
-    free(hdr);
+    if (data)
+    {
+        dsk.meta_disk->free_iomem(data);
+        data = NULL;
+    }
+    if (hdr)
+    {
+        dsk.meta_disk->free_iomem(hdr);
+        hdr = NULL;
+    }
     if (buffer_area)
     {
-        free(buffer_area);
+        dsk.journal_disk->free_iomem(buffer_area);
         buffer_area = NULL;
     }
     if (do_open)
     {
-        close(dsk.meta_fd);
-        dsk.meta_fd = -1;
-        if (dsk.journal_fd >= 0)
-        {
-            if (dsk.journal_fd != dsk.meta_fd)
-                close(dsk.journal_fd);
-            dsk.journal_fd = -1;
-        }
+        if (dsk.journal_disk && dsk.journal_disk != dsk.meta_disk)
+            delete dsk.journal_disk;
+        if (dsk.meta_disk)
+            delete dsk.meta_disk;
+        dsk.meta_disk = NULL;
+        dsk.journal_disk = NULL;
     }
     return r;
 }
@@ -520,7 +551,7 @@ int disk_tool_t::write_json_meta(json11::Json meta)
     {
         return write_json_heap(meta, meta["journal"]);
     }
-    new_meta_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_meta_len);
+    new_meta_buf = (uint8_t*)dsk.meta_disk->alloc_iomem(new_meta_len);
     memset(new_meta_buf, 0, new_meta_len);
     blockstore_meta_header_v2_t *new_hdr = (blockstore_meta_header_v2_t *)new_meta_buf;
     new_hdr->zero = 0;
@@ -558,7 +589,7 @@ int disk_tool_t::write_json_meta(json11::Json meta)
         uint64_t mb = 1 + data_block/new_entries_per_block;
         if (mb >= new_meta_len/new_hdr->meta_block_size)
         {
-            free(new_meta_buf);
+            dsk.meta_disk->free_iomem(new_meta_buf);
             new_meta_buf = NULL;
             fprintf(stderr, "Metadata (data block %ju) doesn't fit into the new area\n", data_block);
             return 1;
@@ -592,7 +623,7 @@ 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);
+    new_meta_buf = (uint8_t*)dsk.meta_disk->alloc_iomem(new_meta_len);
     memset(new_meta_buf, 0, new_meta_len);
     new_meta_hdr = (blockstore_meta_header_v3_t *)new_meta_buf;
     new_meta_hdr->zero = 0;
@@ -627,7 +658,7 @@ int disk_tool_t::write_json_heap(json11::Json meta, json11::Json journal)
     new_journal_buf = NULL;
     if (new_journal_len)
     {
-        new_journal_buf = (uint8_t*)memalign(MEM_ALIGNMENT, new_journal_len);
+        new_journal_buf = (uint8_t*)dsk.journal_disk->alloc_iomem(new_journal_len);
         memset(new_journal_buf, 0, new_journal_len);
     }
     uint64_t total_used_space = 0;
@@ -684,8 +715,7 @@ int disk_tool_t::write_json_heap(json11::Json meta, json11::Json journal)
                     fprintf(stderr, "Write entry %s has invalid type: %s, aborting\n",
                         write_entry.dump().c_str(), write_entry["type"].dump().c_str());
 close_err0:
-                    free(new_meta_buf);
-                    new_meta_buf = NULL;
+                    free_new_meta();
                     return 1;
                 }
                 uint64_t wr_offset = write_entry["offset"].uint64_value();
@@ -772,8 +802,7 @@ close_err0:
         {
             fprintf(stderr, "Metadata should include journal in you want to convert it to the \"heap\" format\n");
 close_err:
-            free(new_meta_buf);
-            new_meta_buf = NULL;
+            free_new_meta();
             return 1;
         }
         std::map<object_id, std::vector<json11::Json::object>> journal_by_object;

src/disk_tool/disk_tool_nvme.cpp

@@ -0,0 +1,158 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#ifdef WITH_VFN
+
+#include "disk_tool.h"
+#include "str_util.h"
+#include "json_util.h"
+#include "disk_util.h"
+#include "malloc_or_die.h"
+
+#define NVME_CLASS 0x010800
+#define NVME_CLASS_MASK 0xffff00
+
+const char *oncs_bit_names[] = {
+    "compare",
+    "write uncorrectable",
+    "dsm (deallocate)",
+    "write zeroes",
+    "save features",
+    "reservations",
+    "timestamp",
+    "verify",
+    "copy",
+    "copies are atomic",
+    "copies are fast",
+    "write zeroes with deallocate",
+    "namespaces zeroes"
+};
+
+bool bind_to_vfio(const std::string & dev)
+{
+    int r;
+    if (access("/sys/bus/pci/drivers/vfio-pci", F_OK) != 0 && errno == ENOENT)
+    {
+        r = system("modprobe vfio-pci");
+        if (r != 0)
+        {
+            fprintf(stderr, "Failed to load vfio-pci kernel module: %s (code %u)\n", strerror(errno), errno);
+            return false;
+        }
+    }
+    r = write_file("/sys/bus/pci/drivers/vfio-pci/bind", dev);
+    if (r == -ENODEV)
+    {
+        auto vid = str_replace(trim(read_file("/sys/bus/pci/devices/"+dev+"/vendor")), "0x", "");
+        auto pid = str_replace(trim(read_file("/sys/bus/pci/devices/"+dev+"/device")), "0x", "");
+        r = write_file("/sys/bus/pci/drivers/vfio-pci/new_id", vid+" "+pid);
+        if (r != 0)
+        {
+            fprintf(stderr, "Failed to register VID:PID %s:%s in vfio-pci: %s (code %u)\n",
+                vid.c_str(), pid.c_str(), strerror(-r), -r);
+            return false;
+        }
+        r = write_file("/sys/bus/pci/drivers/vfio-pci/bind", dev);
+    }
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to bind %s to vfio-pci driver: %s (code %u)\n",
+            dev.c_str(), strerror(-r), -r);
+        return false;
+    }
+    return true;
+}
+
+std::string disk_tool_t::find_pci_id(const std::string & nvme_serial, uint32_t nvme_nsid)
+{
+    auto dev = vfn_disk_i::find_pci_id(nvme_serial, nvme_nsid);
+    if (dev == "")
+    {
+        return dev;
+    }
+    std::unique_ptr<vfn_disk_i> dsk(vfn_disk_i::create());
+    if (!dsk->open_dev(dev))
+    {
+        continue;
+    }
+    const nvme_id_ctrl* id_ctrl = dsk->get_id_ctrl();
+    std::string serial = trim(std::string(id_ctrl->sn, 20));
+    fprintf(stderr, "NVMe device %s found - PCI ID %s\n", serial.c_str(), dev.c_str());
+    // Interesting fields: mdts, tnvmcap, nn, vwc, awupf, sgls, msdbd, cap.cmbs, hmpre, hmmin, oncs
+    fprintf(stderr, "Maximum transfer size: %ju bytes\n", dsk->get_mdts());
+    fprintf(stderr, "Total capacity: %ju bytes\n", *(uint64_t*)id_ctrl->tnvmcap);
+    fprintf(stderr, "Max number of namespaces: %u\n", id_ctrl->nn);
+    fprintf(stderr, "Volatile cache: %s\n", (id_ctrl->vwc & NVME_CTRL_VWC_PRESENT) ? "yes" : "no");
+    fprintf(stderr, "Atomic write unit power fail: %u blocks (probably 4k)\n", id_ctrl->awupf+1);
+    fprintf(stderr, "SGL support: %s",
+        (id_ctrl->sgls & NVME_CTRL_SGLS_SUPPORTED) == 1 ? "unaligned" :
+        ((id_ctrl->sgls & NVME_CTRL_SGLS_SUPPORTED) == 2 ? "8 byte aligned" : "no support"));
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_KEYED)
+        fprintf(stderr, ", keyed (RDMA)");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_BIT_BUCKET)
+        fprintf(stderr, ", bit bucket");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_MPTR_BYTE_ALIGNED)
+        fprintf(stderr, ", byte-aligned mptr");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_OVERSIZE)
+        fprintf(stderr, ", oversized");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_MPTR_SGL)
+        fprintf(stderr, ", SGL mptr");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_OFFSET)
+        fprintf(stderr, ", offset");
+    if (id_ctrl->sgls & NVME_CTRL_SGLS_TPORT)
+        fprintf(stderr, ", transport data block");
+    fprintf(stderr, "\nMax SGL data block descriptors: %u\n", id_ctrl->msdbd);
+    fprintf(stderr, "Optional command support:");
+    int oncs_n = 0;
+    for (int i = 0; i < sizeof(oncs_bit_names)/sizeof(oncs_bit_names[0]); i++)
+    {
+        if ((id_ctrl->oncs >> i) & 1)
+            fprintf(stderr, "%s%s", oncs_n++ > 0 ? ", " : " ", oncs_bit_names[i]);
+    }
+    fprintf(stderr, "%s\n", oncs_n > 0 ? "" : "none");
+    fprintf(stderr, "CMB: %s\n", NVME_CAP_CMBS(dsk->get_cap()) ? "supported" : "not supported");
+    fprintf(stderr, "HMB: min %u bytes, preferred %u bytes\n", id_ctrl->hmmin*4096, id_ctrl->hmpre*4096);
+    if (nvme_nsid > 0)
+    {
+        if (!dsk->open_ns(nvme_nsid))
+        {
+            return "";
+        }
+        // Interesting fields: nsze, nsfeat (bit 3 - nguid, bit 1 - nawupf), nlbaf,
+        // flbas, mc (bit 1 - OK, bit 0 - not OK), nawupf, nabspf, nabo, nguid, lbaf0-lbaf15
+        const nvme_id_ns *id_ns = dsk->get_id_ns();
+        uint8_t fmt = NVME_FLBAS_LOWER(id_ns->flbas);
+        uint64_t blksize = 1 << id_ns->lbaf[fmt].ds;
+        uint64_t ms = id_ns->lbaf[fmt].ms;
+        fprintf(stderr, "Namespace size: %ju bytes\n", (uint64_t)id_ns->nsze * blksize);
+        fprintf(stderr, "Format %u/%u: %ju byte sectors + %ju byte metadata\n", id_ns->flbas, id_ns->nlbaf+1, blksize, ms);
+        if (id_ns->nsfeat & NVME_NS_FEAT_ID_REUSE)
+        {
+            fprintf(stderr, "Namespace GUID: ");
+            for (int i = 0; i < sizeof(id_ns->nguid); i++)
+                fprintf(stderr, "%02x", id_ns->nguid[i]);
+            fprintf(stderr, "\n");
+        }
+        if (id_ns->nsfeat & NVME_NS_FEAT_NATOMIC)
+        {
+            fprintf(stderr, "Namespace atomic write unit power fail: %u blocks\n", id_ns->nawupf+1);
+            fprintf(stderr, "Namespace atomic boundary power fail: %u blocks\n", id_ns->nabspf);
+            fprintf(stderr, "Namespace atomic boundary offset: %u blocks\n", id_ns->nabo);
+        }
+    }
+    return dev;
+}
+
+#else
+
+#include "disk_tool.h"
+#include <string>
+
+std::string disk_tool_t::find_pci_id(const std::string & nvme_serial, uint32_t nvme_nsid)
+{
+    fprintf(stderr, "Vitastor is built without userspace NVMe (libvfn) support\n");
+    exit(1);
+    return "";
+}
+
+#endif

src/disk_tool/disk_tool_prepare.cpp

@@ -4,6 +4,7 @@
 #include "disk_tool.h"
 #include "str_util.h"
 #include "json_util.h"
+#include "malloc_or_die.h"
 
 void disk_tool_t::parse_meta_reserve()
 {
@@ -272,10 +273,24 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
         }
     }
     // Zero out the first block of metadata and journal - OSD will zero the rest on the first run
-    if (write_zero(dsk.meta_fd, sb["meta_offset"].uint64_value(), 4096) != 0 ||
-        write_zero(dsk.journal_fd, sb["journal_offset"].uint64_value(), 4096) != 0)
+    int rz = 0;
+    if (dsk.meta_disk)
     {
-        fprintf(stderr, "Failed to zero out metadata or journal: %s\n", strerror(errno));
+        void *zero_buf = dsk.meta_disk->alloc_iomem(4096);
+        memset(zero_buf, 0, 4096);
+        rz = dsk.meta_disk->write_sync(sb["meta_offset"].uint64_value(), 4096, zero_buf);
+        dsk.meta_disk->free_iomem(zero_buf);
+    }
+    if (!rz && dsk.journal_disk)
+    {
+        void *zero_buf = dsk.journal_disk->alloc_iomem(4096);
+        memset(zero_buf, 0, 4096);
+        rz = dsk.journal_disk->write_sync(sb["journal_offset"].uint64_value(), 4096, zero_buf);
+        dsk.journal_disk->free_iomem(zero_buf);
+    }
+    if (rz != 0)
+    {
+        fprintf(stderr, "Failed to zero out metadata or journal header: %s (code %d)\n", strerror(-rz), -rz);
         dsk.close_all();
         return 1;
     }

src/disk_tool/disk_tool_resize.cpp

@@ -303,16 +303,17 @@ int disk_tool_t::resize_copy_data()
         iodepth = 32;
     }
     ringloop = new ring_loop_t(iodepth < RINGLOOP_DEFAULT_SIZE ? RINGLOOP_DEFAULT_SIZE : iodepth);
-    dsk.data_fd = open(dsk.data_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
-    if (dsk.data_fd < 0)
+    dsk.data_disk = disk_i::create_and_open(dsk.data_device, (options["io"] == "cached" ? "cached" : ""), true);
+    if (!dsk.data_disk)
     {
-        fprintf(stderr, "Failed to open data device %s: %s\n", dsk.data_device.c_str(), strerror(errno));
+        fprintf(stderr, "Failed to open data device %s\n", dsk.data_device.c_str());
         delete ringloop;
         ringloop = NULL;
         return 1;
     }
+    dsk.data_disk->set_ringloop(ringloop);
     moving_blocks = new resizer_data_moving_t[iodepth];
-    moving_blocks[0].buf = memalign_or_die(MEM_ALIGNMENT, iodepth*dsk.data_block_size);
+    moving_blocks[0].buf = dsk.data_disk->alloc_iomem(iodepth*dsk.data_block_size);
     for (int i = 1; i < iodepth; i++)
     {
         moving_blocks[i].buf = (uint8_t*)moving_blocks[0].buf + i*dsk.data_block_size;
@@ -334,52 +335,76 @@ int disk_tool_t::resize_copy_data()
             }
             if (moving_blocks[i].state == DM_ST_TO_READ)
             {
-                struct io_uring_sqe *sqe = ringloop->get_sqe();
-                if (sqe)
-                {
-                    moving_blocks[i].state = DM_ST_READING;
-                    struct ring_data_t *data = ((ring_data_t*)sqe->user_data);
-                    data->iov = (struct iovec){ moving_blocks[i].buf, dsk.data_block_size };
-                    io_uring_prep_readv(sqe, dsk.data_fd, &data->iov, 1, dsk.data_offset + moving_blocks[i].old_loc*dsk.data_block_size);
-                    data->callback = [this, i](ring_data_t *data)
+                int res = dsk.data_disk->read_async(
+                    dsk.data_offset + moving_blocks[i].old_loc*dsk.data_block_size,
+                    dsk.data_block_size, moving_blocks[i].buf, [this, i](int res)
                     {
-                        if (data->res != dsk.data_block_size)
+                        if (res != 0)
                         {
                             fprintf(
                                 stderr, "Failed to read %ju bytes at %ju from %s: %s\n", dsk.data_block_size,
                                 dsk.data_offset + moving_blocks[i].old_loc*dsk.data_block_size, dsk.data_device.c_str(),
-                                data->res < 0 ? strerror(-data->res) : "short read"
+                                strerror(-res)
                             );
                             exit(1);
                         }
                         moving_blocks[i].state = DM_ST_TO_WRITE;
                         ringloop->wakeup();
-                    };
-                 }
+                    }
+                );
+                if (res == 0)
+                {
+                    moving_blocks[i].state = DM_ST_READING;
+                }
+                else if (res == -EAGAIN)
+                {
+                    break;
+                }
+                else
+                {
+                    fprintf(
+                        stderr, "Failed to submit read of %ju bytes at %ju from %s: %s\n", dsk.data_block_size,
+                        dsk.data_offset + moving_blocks[i].old_loc*dsk.data_block_size, dsk.data_device.c_str(),
+                        strerror(-res)
+                    );
+                    exit(1);
+                }
             }
             if (moving_blocks[i].state == DM_ST_TO_WRITE)
             {
-                struct io_uring_sqe *sqe = ringloop->get_sqe();
-                if (sqe)
-                {
-                    moving_blocks[i].state = DM_ST_WRITING;
-                    struct ring_data_t *data = ((ring_data_t*)sqe->user_data);
-                    data->iov = (struct iovec){ moving_blocks[i].buf, dsk.data_block_size };
-                    io_uring_prep_writev(sqe, dsk.data_fd, &data->iov, 1, dsk.data_offset + moving_blocks[i].new_loc*dsk.data_block_size);
-                    data->callback = [this, i](ring_data_t *data)
+                int res = dsk.data_disk->write_async(
+                    dsk.data_offset + moving_blocks[i].new_loc*dsk.data_block_size,
+                    dsk.data_block_size, moving_blocks[i].buf, [this, i](int res)
                     {
-                        if (data->res != dsk.data_block_size)
+                        if (res != 0)
                         {
                             fprintf(
                                 stderr, "Failed to write %ju bytes at %ju to %s: %s\n", dsk.data_block_size,
                                 dsk.data_offset + moving_blocks[i].new_loc*dsk.data_block_size, dsk.data_device.c_str(),
-                                data->res < 0 ? strerror(-data->res) : "short write"
+                                strerror(-res)
                             );
                             exit(1);
                         }
                         moving_blocks[i].state = DM_ST_EMPTY;
                         ringloop->wakeup();
-                    };
+                    }
+                );
+                if (res == 0)
+                {
+                    moving_blocks[i].state = DM_ST_WRITING;
+                }
+                else if (res == -EAGAIN)
+                {
+                    break;
+                }
+                else
+                {
+                    fprintf(
+                        stderr, "Failed to submit write of %ju bytes at %ju to %s: %s\n", dsk.data_block_size,
+                        dsk.data_offset + moving_blocks[i].new_loc*dsk.data_block_size, dsk.data_device.c_str(),
+                        strerror(-res)
+                    );
+                    exit(1);
                 }
             }
             remap_active += moving_blocks[i].state != DM_ST_EMPTY ? 1 : 0;
@@ -395,11 +420,11 @@ int disk_tool_t::resize_copy_data()
         ringloop->wait();
     }
     ringloop->unregister_consumer(&ring_consumer);
-    free(moving_blocks[0].buf);
+    dsk.data_disk->free_iomem(moving_blocks[0].buf);
     delete[] moving_blocks;
     moving_blocks = NULL;
-    close(dsk.data_fd);
-    dsk.data_fd = -1;
+    delete dsk.data_disk;
+    dsk.data_disk = NULL;
     delete ringloop;
     ringloop = NULL;
     return 0;
@@ -407,7 +432,7 @@ int disk_tool_t::resize_copy_data()
 
 void disk_tool_t::resize_alloc_journal()
 {
-    new_journal_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_journal_len);
+    new_journal_buf = (uint8_t*)dsk.journal_disk->alloc_iomem(new_journal_len);
     memset(new_journal_buf, 0, new_journal_len);
     new_journal_ptr = new_journal_buf;
     new_journal_data = new_journal_ptr + dsk.journal_block_size;
@@ -512,17 +537,24 @@ int disk_tool_t::resize_rebuild_journal()
 
 int disk_tool_t::resize_write_new_journal()
 {
-    new_journal_fd = open(new_journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
-    if (new_journal_fd < 0)
+    new_journal_disk = disk_i::create_and_open(new_journal_device, (options["io"] == "cached" ? "cached" : ""), true);
+    if (!new_journal_disk)
     {
-        fprintf(stderr, "Failed to open new journal device %s: %s\n", new_journal_device.c_str(), strerror(errno));
+        fprintf(stderr, "Failed to open new journal device %s\n", new_journal_device.c_str());
         return 1;
     }
-    pwrite_blocking(new_journal_fd, new_journal_buf, new_journal_len, new_journal_offset);
-    fsync(new_journal_fd);
-    close(new_journal_fd);
-    new_journal_fd = -1;
-    return 0;
+    int r = new_journal_disk->write_sync(new_journal_offset, new_journal_len, new_journal_buf);
+    if (r == 0)
+    {
+        r = new_journal_disk->flush_sync();
+    }
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to write new journal to %s: %s (code %d)\n", new_journal_device.c_str(), strerror(-r), -r);
+    }
+    delete new_journal_disk;
+    new_journal_disk = NULL;
+    return r;
 }
 
 void disk_tool_t::remap_big_write(blockstore_heap_t *heap, heap_entry_t *wr)
@@ -559,7 +591,7 @@ void disk_tool_t::remap_small_write(blockstore_heap_t *heap, heap_entry_t *wr)
 
 int disk_tool_t::resize_rebuild_meta()
 {
-    new_meta_buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, new_meta_len);
+    new_meta_buf = (uint8_t*)dsk.meta_disk->alloc_iomem(new_meta_len);
     memset(new_meta_buf, 0, new_meta_len);
     new_meta_hdr = (blockstore_meta_header_v3_t *)new_meta_buf;
     uint64_t new_meta_pos = dsk.meta_block_size;
@@ -691,34 +723,41 @@ int disk_tool_t::resize_rebuild_meta()
 
 int disk_tool_t::resize_write_new_meta()
 {
-    new_meta_fd = open(new_meta_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
-    if (new_meta_fd < 0)
+    new_meta_disk = disk_i::create_and_open(new_meta_device, (options["io"] == "cached" ? "cached" : ""), true);
+    if (!new_meta_disk)
     {
-        fprintf(stderr, "Failed to open new metadata device %s: %s\n", new_meta_device.c_str(), strerror(errno));
+        fprintf(stderr, "Failed to open new metadata device %s\n", new_meta_device.c_str());
         return 1;
     }
-    pwrite_blocking(new_meta_fd, new_meta_buf, new_meta_len, new_meta_offset);
-    fsync(new_meta_fd);
-    close(new_meta_fd);
-    new_meta_fd = -1;
-    return 0;
+    int r = new_meta_disk->write_sync(new_meta_offset, new_meta_len, new_meta_buf);
+    if (r == 0)
+    {
+        r = new_meta_disk->flush_sync();
+    }
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to write new metadata to %s: %s (code %d)\n", new_meta_device.c_str(), strerror(-r), -r);
+    }
+    delete new_meta_disk;
+    new_meta_disk = NULL;
+    return r;
 }
 
 void disk_tool_t::free_new_meta()
 {
     if ((uint8_t*)new_meta_hdr != new_meta_buf)
     {
-        free(new_meta_hdr);
+        dsk.meta_disk->free_iomem(new_meta_hdr);
         new_meta_hdr = NULL;
     }
     if (new_meta_buf)
     {
-        free(new_meta_buf);
+        dsk.meta_disk->free_iomem(new_meta_buf);
         new_meta_buf = NULL;
     }
     if (new_journal_buf)
     {
-        free(new_journal_buf);
+        dsk.journal_disk->free_iomem(new_journal_buf);
         new_journal_buf = NULL;
     }
 }

src/disk_tool/disk_tool_resize_auto.cpp

@@ -37,8 +37,6 @@ int disk_tool_t::resize_data(std::string device)
         fprintf(stderr, "%s\n", e.what());
         return 1;
     }
-    // Save FD numbers because calc_lengths() relies on them
-    int old_journal_fd = dsk.journal_fd, old_meta_fd = dsk.meta_fd, old_data_fd = dsk.data_fd;
     dsk.close_all();
     bool dry_run = options.find("dry_run") != options.end();
     auto old_journal_device = dsk.journal_device;
@@ -58,13 +56,7 @@ int disk_tool_t::resize_data(std::string device)
             ? dsk.data_device_size : new_data_dev_size;
         dsk.data_device_size = new_data_dev_size;
         dsk.cfg_data_size = 0;
-        dsk.journal_fd = old_journal_fd;
-        dsk.meta_fd = old_meta_fd;
-        dsk.data_fd = old_data_fd;
         dsk.calc_lengths(true);
-        dsk.journal_fd = -1;
-        dsk.meta_fd = -1;
-        dsk.data_fd = -1;
     }
     std::map<std::string, std::string> move_options;
     if (options.find("move_journal") != options.end())

src/disk_tool/disk_tool_udev.cpp

@@ -155,15 +155,16 @@ json11::Json disk_tool_t::read_osd_superblock(std::string device, bool expect_ex
     json11::Json osd_params;
     std::string json_err;
     std::string real_device, device_type, real_data, real_meta, real_journal;
-    int r, fd = open(device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
-    if (fd < 0)
+    int r = 0;
+    auto di = disk_i::create_and_open(device, (options["io"] == "cached" ? "cached,ro" : "ro"), true);
+    if (!di)
     {
-        fprintf(stderr, "Failed to open device %s: %s\n", device.c_str(), strerror(errno));
+        fprintf(stderr, "Failed to open device %s\n", device.c_str());
         return osd_params;
     }
-    buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 4096);
-    r = pread_blocking(fd, buf, 4096, 0);
-    if (r != 4096)
+    buf = (uint8_t*)di->alloc_iomem(4096);
+    r = di->read_sync(0, 4096, buf);
+    if (r != 0)
     {
         fprintf(stderr, "Failed to read OSD superblock from %s: %s\n", device.c_str(), strerror(errno));
         goto ex;
@@ -186,10 +187,10 @@ json11::Json disk_tool_t::read_osd_superblock(std::string device, bool expect_ex
     if (sb->size > 4096)
     {
         uint64_t sb_size = ((sb->size+4095)/4096)*4096;
-        free(buf);
-        buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, sb_size);
-        r = pread_blocking(fd, buf, sb_size, 0);
-        if (r != sb_size)
+        di->free_iomem(buf);
+        buf = (uint8_t*)di->alloc_iomem(sb_size);
+        r = di->read_sync(0, sb_size, buf);
+        if (r != 0)
         {
             fprintf(stderr, "Failed to read OSD superblock from %s: %s\n", device.c_str(), strerror(errno));
             goto ex;
@@ -224,12 +225,12 @@ json11::Json disk_tool_t::read_osd_superblock(std::string device, bool expect_ex
         osd_params = json11::Json();
         goto ex;
     }
-    real_device = realpath_str(device);
-    real_data = realpath_str(osd_params["data_device"].string_value());
+    real_device = disk_i::realpath(device);
+    real_data = disk_i::realpath(osd_params["data_device"].string_value());
     real_meta = osd_params["meta_device"].string_value() != "" && osd_params["meta_device"] != osd_params["data_device"]
-        ? realpath_str(osd_params["meta_device"].string_value()) : "";
+        ? disk_i::realpath(osd_params["meta_device"].string_value()) : "";
     real_journal = osd_params["journal_device"].string_value() != "" && osd_params["journal_device"] != osd_params["meta_device"]
-        ? realpath_str(osd_params["journal_device"].string_value()) : "";
+        ? disk_i::realpath(osd_params["journal_device"].string_value()) : "";
     if (real_journal == real_meta)
     {
         real_journal = "";
@@ -265,8 +266,12 @@ json11::Json disk_tool_t::read_osd_superblock(std::string device, bool expect_ex
         { "real_journal_device", real_journal },
     };
 ex:
-    free(buf);
-    close(fd);
+    if (buf)
+    {
+        di->free_iomem(buf);
+        buf = NULL;
+    }
+    delete di;
     return osd_params;
 }
 

src/disk_tool/disk_tool_utils.cpp

@@ -3,7 +3,6 @@
 
 #include <sys/wait.h>
 #include <sys/ioctl.h>
-#include <dirent.h>
 
 #include "disk_tool.h"
 #include "rw_blocking.h"
@@ -234,30 +233,6 @@ err_pipe1:
     return 255;
 }
 
-int write_zero(int fd, uint64_t offset, uint64_t size)
-{
-    uint64_t buf_len = 1024*1024;
-    void *zero_buf = memalign_or_die(MEM_ALIGNMENT, buf_len);
-    memset(zero_buf, 0, buf_len);
-    ssize_t r;
-    while (size > 0)
-    {
-        r = pwrite(fd, zero_buf, size > buf_len ? buf_len : size, offset);
-        if (r > 0)
-        {
-            size -= r;
-            offset += r;
-        }
-        else if (errno != EAGAIN && errno != EINTR)
-        {
-            free(zero_buf);
-            return -1;
-        }
-    }
-    free(zero_buf);
-    return 0;
-}
-
 // Returns false in case of an error
 // Returns null if there is no partition table
 json11::Json read_parttable(std::string dev)

src/osd/osd_primary.cpp

@@ -485,24 +485,26 @@ pg_osd_set_state_t *osd_t::mark_partial_write(pg_t & pg, osd_op_t *cur_op)
         int changes = 0;
         for (auto & chunk: new_set)
         {
-            if (chunk.loc_bad != LOC_OUTDATED)
+            bool success = false;
+            for (int i = 0; i < op_data->n_subops; i++)
             {
-                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)
                 {
-                    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;
+                    success = true;
+                    break;
                 }
             }
+            if (!success && !(chunk.loc_bad & LOC_OUTDATED))
+            {
+                changes++;
+                chunk.loc_bad = LOC_OUTDATED;
+            }
+            else if (success && (chunk.loc_bad & LOC_OUTDATED))
+            {
+                changes++;
+                chunk.loc_bad &= ~LOC_OUTDATED;
+            }
         }
         return changes;
     });

src/osd/osd_secondary.cpp

@@ -370,7 +370,6 @@ void osd_t::exec_show_config(osd_op_t *cur_op)
     if (req_json["features"]["check_sequencing"].bool_value())
     {
         cl->check_sequencing = true;
-        cl->read_op_id = cur_op->req.hdr.id + 1;
     }
     auto features = json11::Json::object{ { "pg_locks", true } };
     if (msgr.use_proto_checksums)

src/test/osd_test.cpp

@@ -144,6 +144,20 @@ void test_replicated_write()
     printf("test_replicated_write passed\n");
 }
 
+// Find the object's current pg_osd_set_state_t in whatever "not clean" index
+// the PG keeps it in. NULL means the object is considered clean.
+static pg_osd_set_state_t *find_object_state(pg_t & pg, object_id oid)
+{
+    for (auto *idx: { &pg.inconsistent_objects, &pg.incomplete_objects,
+        &pg.degraded_objects, &pg.misplaced_objects })
+    {
+        auto it = idx->find(oid);
+        if (it != idx->end())
+            return it->second;
+    }
+    return NULL;
+}
+
 // Regression test for a bug where a second scrub on a still-failing replica
 // (already marked LOC_CORRUPTED) would clear the corruption flag.
 void test_scrub_corruption_persists()
@@ -2035,6 +2049,113 @@ void test_ec42_inconsistent_read()
     printf("test_ec42_inconsistent_read passed\n");
 }
 
+// Regression test for a bug where a repeated partially failed write to the same
+// object didn't clear LOC_OUTDATED from the replicas that succeeded this time —
+// so a copy stayed marked outdated even though it had just received the newest
+// version, and every failing write only ever added outdated copies.
+//
+// Layout: pool 1, replicated x3, primary = OSD 1 (us), secondaries = OSD 2 and 3.
+//
+// Write 1 fails with ENOSPC on OSD 3     -> OSD 3 is marked outdated.
+// Write 2 fails with ENOSPC on OSD 1 (us) but succeeds on OSD 2 and OSD 3
+//   -> OSD 1 must become outdated AND OSD 3 must stop being outdated.
+void test_repeated_partial_write_unmarks_outdated()
+{
+    printf("test_repeated_partial_write_unmarks_outdated\n");
+
+    osd_test_fixture_t f;
+    f.configure_replicated_pool(/*pool_id*/ 1, /*pg_size*/ 3, /*pg_minsize*/ 1, /*pg_count*/ 1,
+        { { 1, 2, 3 } });
+    f.start(json11::Json::object {
+        { "osd_num", 1 },
+        { "etcd_address", "127.0.0.1:2379" },
+        { "immediate_commit", "all" },
+        { "block_size", 131072 },
+        { "bitmap_granularity", 4096 },
+    });
+    f.connect_peer(2);
+    f.connect_peer(3);
+    f.complete_peering_empty();
+    assert(f.pg(1, 1).state & PG_ACTIVE);
+
+    inode_t inode = INODE_WITH_POOL(1, 1);
+    object_id oid = { .inode = inode, .stripe = 0 };
+
+    // Is the copy on <osd_num> marked outdated? A copy that vanished from the
+    // set entirely counts as outdated, too.
+    auto outdated_on = [&](osd_num_t osd_num)
+    {
+        auto *st = find_object_state(f.pg(1, 1), oid);
+        assert(st);
+        for (auto & chunk: st->osd_set)
+            if (chunk.osd_num == osd_num)
+                return (chunk.loc_bad & LOC_OUTDATED) != 0;
+        return true;
+    };
+
+    // ============ WRITE 1 — ENOSPC on OSD 3 ============
+
+    // Full-block writes on purpose: a sub-block write to a degraded object
+    // would turn into a read-modify-write and only obscure what's tested here.
+    auto *write1 = make_write_op(inode, 0, 131072, 0xab);
+    int retval1 = -1;
+    write1->callback = [&retval1](osd_op_t *op) { retval1 = op->reply.hdr.retval; };
+    f.exec(write1);
+
+    // Zero-length read resolving the (not yet existing) object version
+    f.bs_zero_read_ok(0);
+
+    // One local write + one write per peer
+    assert(f.bs->queued.size() == 1);
+    f.bs_write_ok(BS_OP_WRITE_STABLE, 1);
+    f.peer_write_ok(2, OSD_OP_SEC_WRITE_STABLE, 1);
+    auto *peer3_w1 = f.peer_take(3, OSD_OP_SEC_WRITE_STABLE);
+    assert(retval1 == -1); // still waiting for OSD 3
+    f.peer_complete(peer3_w1, -ENOSPC);
+
+    assert(retval1 == -ENOSPC);
+    assert(f.pg(1, 1).state & PG_HAS_DEGRADED);
+    assert(!outdated_on(1));
+    assert(!outdated_on(2));
+    assert(outdated_on(3));
+    printf("test_repeated_partial_write_unmarks_outdated: write 1 -> OSD 3 outdated\n");
+    delete write1;
+
+    // ============ WRITE 2 — ENOSPC locally, OK on both peers ============
+
+    auto *write2 = make_write_op(inode, 0, 131072, 0xcd);
+    int retval2 = -1;
+    write2->callback = [&retval2](osd_op_t *op) { retval2 = op->reply.hdr.retval; };
+    f.exec(write2);
+
+    // The object exists at version 1 now
+    f.bs_zero_read_ok(1);
+
+    assert(f.bs->queued.size() == 1);
+    auto *local_w2 = f.bs->take(BS_OP_WRITE_STABLE);
+    assert(local_w2->version == 2);
+    local_w2->retval = -ENOSPC;
+    local_w2->callback(local_w2);
+    f.peer_write_ok(2, OSD_OP_SEC_WRITE_STABLE, 2);
+    f.peer_write_ok(3, OSD_OP_SEC_WRITE_STABLE, 2);
+
+    assert(retval2 == -ENOSPC);
+
+    // THE KEY ASSERTION — OSD 3 took the new version, it isn't outdated anymore
+    if (outdated_on(3))
+        printf("BUG: OSD 3 is still marked outdated after successfully taking the new version\n");
+    assert(!outdated_on(3));
+    // ... and the OSD that missed this write is
+    assert(outdated_on(1));
+    assert(!outdated_on(2));
+
+    assert(f.pg(1, 1).inflight == 0);
+    assert(f.pg(1, 1).write_queue.empty());
+
+    delete write2;
+    printf("test_repeated_partial_write_unmarks_outdated passed\n");
+}
+
 int main(int narg, char *args[])
 {
     test_load_global_config();
@@ -2058,5 +2179,6 @@ int main(int narg, char *args[])
     test_no_new_write_on_non_active_pg();
     test_pg_epoch_bump_blocks_write_until_reported();
     test_ec42_inconsistent_read();
+    test_repeated_partial_write_unmarks_outdated();
     return 0;
 }

src/test/test_heap.cpp

@@ -1010,6 +1010,211 @@ void test_iterate_compaction()
         assert(small_writes == 1);
     }
 
+    {
+        blockstore_heap_t heap(&dsk, buffer_area.data());
+        heap.finish_recheck();
+
+        // Case: BIG_STABLE(v1 l1) SMALL_STABLE(v2 l2) BIG_STABLE(v3 l3) un-fsynced
+        // -> skip compaction of l2 into l1 EVEN WHEN under pressure,
+        // because add_compact rejects stable overwrites with EBUSY
+        _test_big_write(heap, dsk, 1, 0, 1, 0, true, 0, 4096, buffer_area.data());
+        _test_small_write(heap, dsk, 1, 0, 2, 0, 4096, 0, true, buffer_area.data(), false);
+
+        heap.use_data(INODE_WITH_POOL(1, 1), 128*1024);
+        uint32_t mblock = 999999;
+        res = _test_do_big_write(heap, dsk, 1, 0, 3, 128*1024, true, 0, 0, buffer_area.data(), &mblock);
+        assert(res == 0);
+        heap.start_block_write(mblock);
+
+        object_id oid = { .inode = INODE_WITH_POOL(1, 1), .stripe = 0 };
+        auto obj = heap.read_entry(oid);
+        assert(obj->lsn == 3);
+        assert(obj->is_overwrite());
+        assert(heap.get_fsynced_lsn() == 2);
+
+        // When not under pressure: postpone compaction
+        int small_writes = 0;
+        auto compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // When under pressure: MUST ALSO postpone compaction because obj is STABLE!
+        // Before the fix, this would incorrectly compact l2 (compact_lsn == 2)
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), true, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // Verify that add_compact() rejects stable overwrites with EBUSY
+        uint32_t compact_mblock = 999999;
+        uint8_t ref_int_bitmap[dsk.clean_entry_bitmap_size];
+        memset(ref_int_bitmap, 0, dsk.clean_entry_bitmap_size);
+        res = heap.add_compact(obj, 2, 2, 0, false, &compact_mblock, ref_int_bitmap, ref_int_bitmap, NULL);
+        assert(res == EBUSY);
+
+        heap.complete_block_write(mblock);
+        assert(heap.get_fsynced_lsn() == 3);
+
+        // Once fsynced, l3 is durable and becomes the clean_wr base
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.clean_wr && compact_info.clean_wr->lsn == 3);
+        assert(small_writes == 0);
+    }
+
+    {
+        blockstore_heap_t heap(&dsk, buffer_area.data());
+        heap.finish_recheck();
+
+        // Case: BIG_STABLE(v1 l1) SMALL_STABLE(v2 l2) BIG_INTENT(v3 l3) un-fsynced
+        // -> skip compaction of l2 into l1 both when not under pressure and under pressure
+        // Before the fix, BIG_INTENT was missing from the check and was never skipped.
+        _test_big_write(heap, dsk, 1, 0, 1, 0, true, 0, 4096, buffer_area.data());
+        _test_small_write(heap, dsk, 1, 0, 2, 0, 4096, 0, true, buffer_area.data(), false);
+
+        object_id oid = { .inode = INODE_WITH_POOL(1, 1), .stripe = 0 };
+        auto obj = heap.read_entry(oid);
+        uint8_t ext_bitmap[dsk.clean_entry_bitmap_size];
+        memset(ext_bitmap, 0x8e, dsk.clean_entry_bitmap_size);
+        heap.use_data(INODE_WITH_POOL(1, 1), 128*1024);
+        uint32_t mblock = 999999;
+        res = heap.add_redirect_intent(oid, &obj, 3, 0, dsk.data_block_size, 128*1024, ext_bitmap, buffer_area.data(), &mblock);
+        assert(res == 0);
+        heap.start_block_write(mblock);
+
+        obj = heap.read_entry(oid);
+        assert(obj->lsn == 3);
+        assert(obj->type() == BS_HEAP_BIG_INTENT);
+        assert(obj->is_overwrite());
+        assert(heap.get_fsynced_lsn() == 2);
+
+        // When not under pressure: MUST postpone compaction!
+        // Before the fix, BIG_INTENT was not checked, so it incorrectly compacted l2 (compact_lsn == 2)
+        int small_writes = 0;
+        auto compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // When under pressure: MUST ALSO postpone compaction!
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), true, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // Verify add_compact() rejects stable overwrite with EBUSY
+        uint32_t compact_mblock = 999999;
+        uint8_t ref_int_bitmap[dsk.clean_entry_bitmap_size];
+        memset(ref_int_bitmap, 0, dsk.clean_entry_bitmap_size);
+        res = heap.add_compact(obj, 2, 2, 0, false, &compact_mblock, ref_int_bitmap, ref_int_bitmap, NULL);
+        assert(res == EBUSY);
+
+        // Complete block write
+        heap.complete_block_write(mblock);
+        assert(heap.get_fsynced_lsn() == 3);
+
+        // Once fsynced, BIG_INTENT is clean_wr base
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.clean_wr && compact_info.clean_wr->lsn == 3);
+        assert(small_writes == 0);
+    }
+
+    {
+        blockstore_heap_t heap(&dsk, buffer_area.data());
+        heap.finish_recheck();
+
+        // Case: BIG_STABLE(v1 l1) SMALL_STABLE(v2 l2) DELETE_STABLE(l3) un-fsynced
+        // -> skip compaction of l2 into l1 EVEN WHEN under pressure,
+        // because add_compact rejects stable overwrites with EBUSY
+        _test_big_write(heap, dsk, 1, 0, 1, 0, true, 0, 4096, buffer_area.data());
+        _test_small_write(heap, dsk, 1, 0, 2, 0, 4096, 0, true, buffer_area.data(), false);
+
+        object_id oid = { .inode = INODE_WITH_POOL(1, 1), .stripe = 0 };
+        auto obj = heap.read_entry(oid);
+        uint32_t mblock = 999999;
+        res = heap.add_delete(obj, &mblock);
+        assert(res == 0);
+        heap.start_block_write(mblock);
+
+        obj = heap.read_entry(oid);
+        assert(obj->lsn == 3);
+        assert(obj->type() == BS_HEAP_DELETE);
+        assert(obj->is_overwrite());
+        assert(heap.get_fsynced_lsn() == 2);
+
+        // When not under pressure: postpone compaction
+        int small_writes = 0;
+        auto compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // When under pressure: MUST ALSO postpone compaction because DELETE is STABLE!
+        // Before the fix, this would incorrectly compact l2 (compact_lsn == 2)
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), true, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.compact_lsn == 0);
+        assert(compact_info.compact_version == 0);
+        assert(!compact_info.clean_wr);
+        assert(!compact_info.do_delete);
+        assert(small_writes == 0);
+
+        // Verify add_compact() rejects stable overwrite with EBUSY
+        uint32_t compact_mblock = 999999;
+        uint8_t ref_int_bitmap[dsk.clean_entry_bitmap_size];
+        memset(ref_int_bitmap, 0, dsk.clean_entry_bitmap_size);
+        res = heap.add_compact(obj, 2, 2, 0, false, &compact_mblock, ref_int_bitmap, ref_int_bitmap, NULL);
+        assert(res == EBUSY);
+
+        // Complete block write
+        heap.complete_block_write(mblock);
+        assert(heap.get_fsynced_lsn() == 3);
+
+        // Once fsynced, DELETE is processed
+        compact_info = heap.iterate_compaction(obj, heap.get_fsynced_lsn(), false, [&](heap_entry_t *wr)
+        {
+            small_writes++;
+        });
+        assert(compact_info.do_delete);
+        assert(!compact_info.clean_wr);
+        assert(small_writes == 0);
+    }
+
     {
         blockstore_heap_t heap(&dsk, buffer_area.data());
         heap.finish_recheck();

src/util/CMakeLists.txt

@@ -19,20 +19,27 @@ add_library(util STATIC
 	../../json11/json11.cpp
 	xxh_x86dispatch.c
 )
-if (RDMACM_LIBRARIES)
+if (RDMACM_LIBRARIES OR (VFN_LIBRARIES AND NVME_LIBRARIES))
 	target_sources(util PRIVATE
 		dma_alloc_impl.cpp
 	)
-endif (RDMACM_LIBRARIES)
+endif (RDMACM_LIBRARIES OR (VFN_LIBRARIES AND NVME_LIBRARIES))
 if (RDMACM_LIBRARIES)
 	target_sources(util PRIVATE
 		rdma_alloc.cpp
 	)
 endif (RDMACM_LIBRARIES)
+if (VFN_LIBRARIES AND NVME_LIBRARIES)
+	target_sources(util PRIVATE
+		vfn_alloc.cpp
+	)
+endif (VFN_LIBRARIES AND NVME_LIBRARIES)
 target_compile_options(util PUBLIC -fPIC)
 target_link_libraries(util
 	${LIBURING_LIBRARIES}
 	${OPENSSL_LIBRARIES}
+	${VFN_LIBRARIES}
+	${NVME_LIBRARIES}
 )
 
 # libcrc32c.a to isolate the isa-l dependency

src/util/disk_util.h

@@ -4,6 +4,76 @@
 #pragma once
 
 #include <string>
+#include <map>
+#include <functional>
 #include <stdint.h>
+#include <time.h>
+
+#ifdef WITH_VFN
+#include <nvme/types.h>
+#endif
+
+class ring_loop_i;
+
+class disk_i
+{
+public:
+    static disk_i *create_and_open(const std::string & path, const std::string & mode, bool disable_flock = false);
+    static disk_i *create_mock(int mock_fd, uint64_t mock_size, uint64_t mock_sectsize);
+    static std::string realpath(const std::string & path);
+    virtual ~disk_i() = default;
+    virtual void set_ringloop(ring_loop_i *ringloop) = 0;
+    virtual bool is_off_ring() const = 0;
+    virtual int get_fd() const = 0;
+    virtual void close() = 0;
+    virtual uint64_t get_size() const = 0;
+    virtual uint64_t get_block_size() const = 0;
+    virtual uint64_t get_max_io_size() const = 0;
+    virtual void *alloc_iomem(size_t size) = 0;
+    virtual void free_iomem(void *buf) = 0;
+    virtual uint32_t get_queue_space() const = 0;
+    // all errors are passed as negative errno's
+    virtual int rw_sync(bool is_wr, uint64_t offset, size_t len, void *buf) = 0;
+    virtual int flush_sync() = 0;
+    virtual int discard_sync(uint64_t offset, uint64_t len) = 0;
+    virtual int rw_async(bool is_wr, uint64_t offset, size_t len, void *buf, bool fua, std::function<void(int res)> cb) = 0;
+    virtual int flush_async(std::function<void(int res)> cb) = 0;
+    virtual int discard_async(uint64_t offset, uint64_t len, std::function<void(int res)> cb) = 0;
+    inline int read_sync(uint64_t offset, size_t len, void *buf)
+    {
+        return rw_sync(false, offset, len, buf);
+    }
+    inline int write_sync(uint64_t offset, size_t len, void *buf)
+    {
+        return rw_sync(true, offset, len, buf);
+    }
+    inline int read_async(uint64_t offset, size_t len, void *buf, std::function<void(int res)> cb)
+    {
+        return rw_async(false, offset, len, buf, false, cb);
+    }
+    inline int write_async(uint64_t offset, size_t len, void *buf, std::function<void(int res)> cb, bool fua = false)
+    {
+        return rw_async(true, offset, len, buf, fua, cb);
+    }
+};
 
 void check_size(int fd, uint64_t *size, uint64_t *sectsize, const std::string & name);
+
+#ifdef WITH_VFN
+class vfn_disk_i: public disk_i
+{
+public:
+    static vfn_disk_i *create();
+    virtual ~vfn_disk_i() = default;
+    virtual bool open_dev(const std::string & pci_id) = 0;
+    virtual bool open_ns(uint32_t ns_num) = 0;
+    virtual const nvme_id_ctrl* get_id_ctrl() const = 0;
+    virtual const nvme_id_ns* get_id_ns() const = 0;
+    virtual uint64_t get_mdts() const = 0;
+    virtual uint64_t get_cap() const = 0;
+
+    static bool is_vfn_path(const std::string & path);
+    static bool bind_to_vfio(const std::string & dev);
+    static std::string find_pci_id(const std::string & nvme_serial, uint32_t nvme_nsid);
+};
+#endif

src/util/ringloop.cpp

@@ -52,6 +52,8 @@ void ring_loop_t::register_consumer(ring_consumer_t *consumer)
 {
     unregister_consumer(consumer);
     consumers.push_back(consumer);
+    if (consumer->is_idle)
+        idle_consumers++;
 }
 
 void ring_loop_t::wakeup()
@@ -65,6 +67,8 @@ void ring_loop_t::unregister_consumer(ring_consumer_t *consumer)
     {
         if (consumers[i] == consumer)
         {
+            if (consumer->is_idle)
+                idle_consumers--;
             consumers.erase(consumers.begin()+i, consumers.begin()+i+1);
             break;
         }
@@ -154,10 +158,30 @@ void ring_loop_t::loop()
     }
     do
     {
-        loop_again = false;
-        for (int i = 0; i < consumers.size(); i++)
+        do
         {
-            consumers[i]->loop();
+            loop_again = false;
+            for (int i = 0; i < consumers.size(); i++)
+            {
+                if (!consumers[i]->is_idle)
+                    consumers[i]->loop();
+            }
+            if (immediate_queue.size())
+            {
+                immediate_queue2.swap(immediate_queue);
+                for (auto & cb: immediate_queue2)
+                    cb();
+                immediate_queue2.clear();
+            }
+        } while (loop_again);
+        if (idle_consumers > 0)
+        {
+            // idle consumers are those who poll devices when everything is idle
+            for (int i = 0; i < consumers.size(); i++)
+            {
+                if (consumers[i]->is_idle)
+                    consumers[i]->loop();
+            }
             if (immediate_queue.size())
             {
                 immediate_queue2.swap(immediate_queue);

src/util/ringloop.h

@@ -32,6 +32,7 @@ struct ring_data_t
 struct ring_consumer_t
 {
     std::function<void(void)> loop;
+    bool is_idle = false;
 };
 
 class __attribute__((visibility("default"))) ring_loop_i
@@ -60,6 +61,7 @@ class __attribute__((visibility("default"))) ring_loop_t: public ring_loop_i
     std::vector<ring_consumer_t*> consumers;
     std::vector<ring_data_t> ring_datas;
     std::vector<int> free_ring_data;
+    int idle_consumers = 0;
     std::mutex mu;
     bool mt;
     bool in_loop;

src/util/vfn_alloc.cpp

@@ -0,0 +1,56 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// libvfn (IOMMU) variant of our simple allocator
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <string.h>
+#include <vfn/iommu.h>
+
+#include "dma_alloc_impl.h"
+
+static_assert(sizeof(iova_t) == sizeof(void*));
+
+class vfn_allocator_t: public dma_allocator_t
+{
+protected:
+    iommu_ctx *ctx = NULL;
+
+    void* reg_mr(void *buf, size_t len) override
+    {
+        iova_t iova = 0;
+        if (iommu_map_vaddr(ctx, buf, len, &iova, 0))
+        {
+            fprintf(stderr, "iommu_map_vaddr %zu bytes failed: %s\n", len, strerror(errno));
+            exit(1);
+        }
+        return (void*)iova;
+    }
+
+    void dereg_mr(void *buf, size_t len, void *handle) override
+    {
+        if (iommu_unmap_vaddr(ctx, buf, &len))
+        {
+            fprintf(stderr, "iommu_unmap_vaddr failed: %s\n", strerror(errno));
+            exit(1);
+        }
+    }
+public:
+    vfn_allocator_t(iommu_ctx *ctx, size_t region_size, size_t max_unused)
+        : dma_allocator_t(region_size, max_unused, 1), ctx(ctx)
+    {
+        alignment = page_size;
+    }
+
+    ~vfn_allocator_t()
+    {
+        free_unused_buffers(0, true);
+    }
+};
+
+dma_allocator_i *vfn_malloc_create(iommu_ctx *ctx, size_t region_size, size_t max_unused)
+{
+    return new vfn_allocator_t(ctx, region_size, max_unused);
+}

src/util/vfn_alloc.h

@@ -0,0 +1,11 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// libvfn (IOMMU) variant of our simple allocator
+
+#pragma once
+
+#include "rdma_alloc.h"
+
+struct iommu_ctx;
+dma_allocator_i *vfn_malloc_create(iommu_ctx *ctx, size_t rdma_alloc_size, size_t rdma_max_unused);

tests/test_snapshot.sh

@@ -72,4 +72,19 @@ qemu-img convert -S 4096 -f qcow2 ./testdata/layer1.qcow2 -O raw ./testdata/bin/
 
 cmp ./testdata/bin/layer1.bin ./testdata/bin/rebased.bin
 
+# Test CAS merge
+
+$VITASTOR_CLI create -s 1M microimg
+
+$VITASTOR_FIO -bs=1M -direct=1 -iodepth=1 -rw=write -image=microimg
+
+$VITASTOR_CLI snap-create microimg@0
+
+$VITASTOR_FIO -bs=16k -direct=1 -iodepth=1 -rw=randwrite -image=microimg -size=16k -loops=1000000 -runtime=10 &
+FIO_PID=$!
+
+sleep 5
+$VITASTOR_CLI rm microimg@0
+wait $FIO_PID
+
 format_green OK