Fix missing mutex unlock with zero-copy and iothreads O_o

.gitea/workflows/test.yml

@@ -684,6 +684,24 @@ jobs:
           echo ""
         done
 
+  test_write_iothreads:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: TEST_NAME=iothreads GLOBAL_CONFIG=',"client_iothread_count":4' /root/vitastor/tests/test_write.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_write_no_same:
     runs-on: ubuntu-latest
     needs: build

cpp-btree

@@ -1 +1 @@
-Subproject commit a21350e484cefa5728f23c227323b4b0822e738f
+Subproject commit 8de8b467acbca50cfd8835c20e0e379110f3b32b

docs/intro/author.en.md

@@ -10,8 +10,17 @@ Copyright (c) Vitaliy Filippov (vitalif [at] yourcmc.ru), 2019+
 
 Join Vitastor Telegram Chat: https://t.me/vitastor
 
-All server-side code (OSD, Monitor and so on) is licensed under the terms of
+License: VNPL 1.1 for server-side code and dual VNPL 1.1 + GPL 2.0+ for client tools.
-Vitastor Network Public License 1.1 (VNPL 1.1), a copyleft license based on
+
+Server-side code is licensed only under the terms of VNPL.
+
+Client libraries (cluster_client and so on) are dual-licensed under the same
+VNPL 1.1 and also GNU GPL 2.0 or later to allow for compatibility with GPLed
+software like QEMU and fio.
+
+## VNPL
+
+Vitastor Network Public License 1.1 (VNPL 1.1) is a copyleft license based on
 GNU GPLv3.0 with the additional "Network Interaction" clause which requires
 opensourcing all programs directly or indirectly interacting with Vitastor
 through a computer network and expressly designed to be used in conjunction
@@ -20,18 +29,83 @@ the terms of the same license, but also under the terms of any GPL-Compatible
 Free Software License, as listed by the Free Software Foundation.
 This is a stricter copyleft license than the Affero GPL.
 
-Please note that VNPL doesn't require you to open the code of proprietary
+The idea of VNPL is, in addition to modules linked to Vitastor code in a single
-software running inside a VM if it's not specially designed to be used with
+binary file, to extend copyleft action to micro-service modules only interacting
-Vitastor.
+with it over the network.
 
 Basically, you can't use the software in a proprietary environment to provide
 its functionality to users without opensourcing all intermediary components
 standing between the user and Vitastor or purchasing a commercial license
 from the author 😀.
 
-Client libraries (cluster_client and so on) are dual-licensed under the same
+At the same time, VNPL doesn't impose any restrictions on software *not specially designed*
-VNPL 1.1 and also GNU GPL 2.0 or later to allow for compatibility with GPLed
+to be used with Vitastor, for example, on Windows running inside a VM with a Vitastor disk.
-software like QEMU and fio.
 
-You can find the full text of VNPL-1.1 in the file [VNPL-1.1.txt](../../VNPL-1.1.txt).
+## Explanation
-GPL 2.0 is also included in this repository as [GPL-2.0.txt](../../GPL-2.0.txt).
+
+Network copyleft is governed by the clause **13. Remote Network Interaction** of VNPL.
+
+A program is considered to be a "Proxy Program" if it meets both conditions:
+- It is specially designed to be used with Vitastor. Basically, it means that the program
+  has any functionality specific to Vitastor and thus "knows" that it works with Vitastor,
+  not with something random.
+- It interacts with Vitastor directly or indirectly through any programming interface,
+  including API, CLI, network or any wrapper (also considered a Proxy Program itself).
+
+If, in addition to that:
+- You give any user an apportunity to interact with Vitastor directly or indirectly through
+  any computer interface including the network or any number of wrappers (Proxy Programs).
+
+Then VNPL requires you to publish the code of all above Proxy Programs to all above users
+under the terms of any GPL-compatible license - that is, GPL, LGPL, MIT/BSD or Apache 2,
+because "GPL compatibility" is treated as an ability to legally include licensed code in
+a GPL application.
+
+So, if you have a "Proxy Program", but it's not open to the user who directly or indirectly
+interacts with Vitastor - you are forbidden to use Vitastor under the terms of VNPL and you
+need a commercial license which doesn't contain open-source requirements.
+
+## Examples
+
+- Vitastor Kubernetes CSI driver which creates PersistentVolumes by calling `vitastor-cli create`.
+  - Yes, it interacts with Vitastor through vitastor-cli.
+  - Yes, it is designed specially for use with Vitastor (it has no sense otherwise).
+  - So, CSI driver **definitely IS** a Proxy Program and must be published under the terms of
+    a free software license.
+- Windows, installed in a VM with the system disk on Vitastor storage.
+  - Yes, it interacts with Vitastor indirectly - it reads and writes data through the block
+    device interface, emulated by QEMU.
+  - No, it definitely isn't designed specially for use with Vitastor - Windows was created long
+    ago before Vitastor and doesn't know anything about it.
+  - So, Windows **definitely IS NOT** a Proxy Program and VNPL doesn't require to open it.
+- Cloud control panel which makes requests to Vitastor Kubernetes CSI driver.
+  - Yes, it interacts with Vitastor indirectly through the CSI driver, which is a Proxy Program.
+  - May or may not be designed specially for use with Vitastor. How to determine exactly?
+    Imagine that Vitastor is replaced with any other storage (for example, with a proprietary).
+    Do control panel functions change in any way? If they do (for example, if snapshots stop working),
+    then the panel contains specific functionality and thus is designed specially for use with Vitastor.
+    Otherwise, the panel is universal and isn't designed specially for Vitastor.
+  - So, whether you are required to open-source the panel also **depends** on whether it
+    contains specific functionality or not.
+
+## Why?
+
+Because I believe into the spirit of copyleft (Linux wouldn't become so popular without GPL!)
+and, at the same time, I want to have a way to monetize the product.
+
+Existing licenses including AGPL are useless for it with an SDS - SDS is a very deeply
+internal software which is almost definitely invisible to the user and thus AGPL doesn't
+require anyone to open the code even if they make a proprietary fork.
+
+And, in fact, the current situation in the world where GPL is though to only restrict direct
+linking of programs into a single executable file, isn't much correct. Nowadays, programs
+are more often linked with network API calls, not with /usr/bin/ld, and a software product
+may consist of dozens of microservices interacting with each other over the network.
+
+That's why we need VNPL to keep the license sufficiently copyleft.
+
+## License Texts
+
+- VNPL 1.1 in English: [VNPL-1.1.txt](../../VNPL-1.1.txt)
+- VNPL 1.1 in Russian: [VNPL-1.1-RU.txt](../../VNPL-1.1-RU.txt)
+- GPL 2.0: [GPL-2.0.txt](../../GPL-2.0.txt)

docs/intro/author.ru.md

@@ -12,6 +12,14 @@
 
 Лицензия: VNPL 1.1 на серверный код и двойная VNPL 1.1 + GPL 2.0+ на клиентский.
 
+Серверные компоненты распространяются только на условиях VNPL.
+
+Клиентские библиотеки распространяются на условиях двойной лицензии VNPL 1.0
+и также на условиях GNU GPL 2.0 или более поздней версии. Так сделано в целях
+совместимости с таким ПО, как QEMU и fio.
+
+## VNPL
+
 VNPL - "сетевой копилефт", собственная свободная копилефт-лицензия
 Vitastor Network Public License 1.1, основанная на GNU GPL 3.0 с дополнительным
 условием "Сетевого взаимодействия", требующим распространять все программы,
@@ -29,9 +37,70 @@ Vitastor Network Public License 1.1, основанная на GNU GPL 3.0 с д
 На Windows и любое другое ПО, не разработанное *специально* для использования
 вместе с Vitastor, никакие ограничения не накладываются.
 
-Клиентские библиотеки распространяются на условиях двойной лицензии VNPL 1.0
+## Пояснение
-и также на условиях GNU GPL 2.0 или более поздней версии. Так сделано в целях
-совместимости с таким ПО, как QEMU и fio.
 
-Вы можете найти полный текст VNPL 1.1 на английском языке в файле [VNPL-1.1.txt](../../VNPL-1.1.txt),
+Сетевой копилефт регулируется пунктом лицензии **13. Удалённое сетевое взаимодействие**.
-VNPL 1.1 на русском языке в файле [VNPL-1.1-RU.txt](../../VNPL-1.1-RU.txt), а GPL 2.0 в файле [GPL-2.0.txt](../../GPL-2.0.txt).
+
+Программа считается "прокси-программой", если верны оба условия:
+- Она создана специально для работы вместе с Vitastor. По сути это означает, что программа
+  должна иметь специфичный для Vitastor функционал, то есть, "знать", что она взаимодействует
+  именно с Vitastor.
+- Она прямо или косвенно взаимодействует с Vitastor через абсолютно любой программный
+  интерфейс, включая любые способы вызова: API, CLI, сеть или через какую-то обёртку (в
+  свою очередь тоже являющуюся прокси-программой).
+
+Если в дополнение к этому также:
+- Вы предоставляете любому пользователю возможность взаимодействовать с Vitastor по сети,
+  опять-таки, через любой интерфейс или любую серию "обёрток" (прокси-программ)
+
+То, согласно VNPL, вы должны открыть код "прокси-программ" **таким пользователям** на условиях
+любой GPL-совместимой лицензии - то есть, GPL, LGPL, MIT/BSD или Apache 2 - "совместимость с GPL"
+понимается как возможность включать лицензируемый код в GPL-приложение.
+
+Соответственно, если у вас есть "прокси-программа", но её код не открыт пользователю,
+который прямо или косвенно взаимодействует с Vitastor - вам запрещено использовать Vitastor
+на условиях VNPL и вам нужна коммерческая лицензия, не содержащая требований об открытии кода.
+
+## Примеры
+
+- Kubernetes CSI-драйвер Vitastor, создающий PersistentVolume с помощью вызова `vitastor-cli create`.
+  - Да, взаимодействует с Vitastor через vitastor-cli.
+  - Да, создавался специально для работы с Vitastor (иначе в чём же ещё его смысл).
+  - Значит, CSI-драйвер **точно считается** "прокси-программой" и должен быть открыт под свободной
+    лицензией.
+- Windows, установленный в виртуальную машину на диске Vitastor.
+  - Да, взаимодействует с Vitastor "прямо или косвенно" - пишет и читает данные через интерфейс
+    блочного устройства, эмулируемый QEMU.
+  - Нет, точно не создан *специально для работы с Vitastor* - когда его создавали, никакого
+    Vitastor ещё и в помине не было.
+  - Значит, Windows **точно не считается** "прокси-программой" и на него требования VNPL не распространяются.
+- Панель управления облака, делающая запросы к Kubernetes CSI-драйверу Vitastor.
+  - Да, взаимодействует с Vitastor косвенно через CSI-драйвер, являющийся "прокси-программой".
+  - Сходу не известно, создавалась ли конкретно для работы с Vitastor. Как понять, да или нет?
+    Представьте, что Vitastor заменён на любую другую систему хранения (например, на проприетарную).
+    Работа панели управления изменится? Если да (например, перестанут работать снапшоты) - значит,
+    панель содержит специфичный функционал и "создана специально для работы с Vitastor".
+    Если нет - значит, специфичного функционала панель не содержит и в принципе она универсальна.
+  - Нужно ли открывать панель - **зависит** от того, содержит она специфичный функционал или нет.
+
+## Почему так?
+
+Потому что я одновременно верю в дух копилефт-лицензий (Linux не стал бы так популярен,
+если бы не GPL!) и хочу иметь возможность монетизации продукта.
+
+При этом использовать даже AGPL для программной СХД бессмысленно - это глубоко внутреннее
+ПО, которое пользователь почти наверняка не увидит вообще, поэтому и открывать код никому
+никогда не придётся, даже при создании производного продукта.
+
+Да и в целом сложившаяся в мире ситуация, при которой действие GPL ограничивается только
+прямым связыванием в один исполняемый файл, не очень корректна. В настоящее время программы
+гораздо чаще интегрируют сетевыми вызовами, а не с помощью /usr/bin/ld, и общий программный
+продукт может состоять из нескольких десятков микросервисов, взаимодействующих по сети.
+
+Поэтому для сохранения достаточной "копилефтности" и придумана VNPL.
+
+## Тексты лицензий
+
+- VNPL 1.1 на английском языке: [VNPL-1.1.txt](../../VNPL-1.1.txt)
+- VNPL 1.1 на русском языке: [VNPL-1.1-RU.txt](../../VNPL-1.1-RU.txt)
+- GPL 2.0: [GPL-2.0.txt](../../GPL-2.0.txt)

src/blockstore/blockstore.h

@@ -31,7 +31,6 @@
 #define DEFAULT_DATA_BLOCK_ORDER 17
 #define MIN_DATA_BLOCK_SIZE 4*1024
 #define MAX_DATA_BLOCK_SIZE 128*1024*1024
-#define MAX_META_BLOCK_SIZE 64*1024
 #define DEFAULT_BITMAP_GRANULARITY 4096
 
 #define BS_OP_MIN 1

src/blockstore/blockstore_disk.cpp

@@ -127,9 +127,9 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     {
         throw std::runtime_error("meta_block_size must be a multiple of "+std::to_string(DIRECT_IO_ALIGNMENT));
     }
-    else if (meta_block_size > MAX_META_BLOCK_SIZE)
+    else if (meta_block_size > MAX_DATA_BLOCK_SIZE)
     {
-        throw std::runtime_error("meta_block_size must not exceed "+std::to_string(MAX_META_BLOCK_SIZE));
+        throw std::runtime_error("meta_block_size must not exceed "+std::to_string(MAX_DATA_BLOCK_SIZE));
     }
     if (data_offset % disk_alignment)
     {

src/blockstore/blockstore_flush.cpp

@@ -427,6 +427,13 @@ stop_flusher:
         printf("Flushing %jx:%jx v%ju\n", cur.oid.inode, cur.oid.stripe, cur.version);
 #endif
         flusher->active_flushers++;
+        // Find it in clean_db
+        {
+            auto & clean_db = bs->clean_db_shard(cur.oid);
+            auto clean_it = clean_db.find(cur.oid);
+            old_clean_ver = (clean_it != clean_db.end() ? clean_it->second.version : 0);
+            old_clean_loc = (clean_it != clean_db.end() ? clean_it->second.location : UINT64_MAX);
+        }
         // Scan dirty versions of the object to determine what we need to read
         scan_dirty();
         // Writes and deletes shouldn't happen at the same time
@@ -531,7 +538,7 @@ resume_2:
         {
             // zero out old metadata entry
             {
-                clean_disk_entry *old_entry = (clean_disk_entry*)((uint8_t*)meta_old.buf + meta_old.pos);
+                clean_disk_entry *old_entry = (clean_disk_entry*)((uint8_t*)meta_old.buf + meta_old.pos*bs->dsk.clean_entry_size);
                 if (old_entry->oid.inode != 0 && old_entry->oid != cur.oid)
                 {
                     printf("Fatal error (metadata corruption or bug): tried to wipe metadata entry %ju (%jx:%jx v%ju) as old location of %jx:%jx\n",
@@ -540,7 +547,7 @@ resume_2:
                     exit(1);
                 }
             }
-            memset((uint8_t*)meta_old.buf + meta_old.pos, 0, bs->dsk.clean_entry_size);
+            memset((uint8_t*)meta_old.buf + meta_old.pos*bs->dsk.clean_entry_size, 0, bs->dsk.clean_entry_size);
     resume_20:
             if (meta_old.sector != meta_new.sector && !write_meta_block(meta_old, 20))
                 return false;
@@ -601,7 +608,7 @@ resume_2:
 
 void journal_flusher_co::update_metadata_entry()
 {
-    clean_disk_entry *new_entry = (clean_disk_entry*)((uint8_t*)meta_new.buf + meta_new.pos);
+    clean_disk_entry *new_entry = (clean_disk_entry*)((uint8_t*)meta_new.buf + meta_new.pos*bs->dsk.clean_entry_size);
     if (new_entry->oid.inode != 0 && new_entry->oid != cur.oid)
     {
         printf(
@@ -616,7 +623,7 @@ void journal_flusher_co::update_metadata_entry()
     if (has_delete)
     {
         // Zero out the new metadata entry
-        memset((uint8_t*)meta_new.buf + meta_new.pos, 0, bs->dsk.clean_entry_size);
+        memset((uint8_t*)meta_new.buf + meta_new.pos*bs->dsk.clean_entry_size, 0, bs->dsk.clean_entry_size);
     }
     else
     {
@@ -798,7 +805,7 @@ bool journal_flusher_co::clear_incomplete_csum_block_bits(int wait_base)
             }
         }
         {
-            clean_disk_entry *new_entry = (clean_disk_entry*)((uint8_t*)meta_new.buf + meta_new.pos);
+            clean_disk_entry *new_entry = (clean_disk_entry*)((uint8_t*)meta_new.buf + meta_new.pos*bs->dsk.clean_entry_size);
             if (new_entry->oid != cur.oid)
             {
                 printf(
@@ -905,12 +912,6 @@ void journal_flusher_co::calc_block_checksums(uint32_t *new_data_csums, bool ski
 
 void journal_flusher_co::scan_dirty()
 {
-    // Find it in clean_db
-    auto & clean_db = bs->clean_db_shard(cur.oid);
-    auto clean_it = clean_db.find(cur.oid);
-    old_clean_ver = (clean_it != clean_db.end() ? clean_it->second.version : 0);
-    old_clean_loc = (clean_it != clean_db.end() ? clean_it->second.location : UINT64_MAX);
-    auto old_clean_bitmap = (clean_it != clean_db.end() ? bs->get_clean_entry_bitmap(clean_it, 0) : NULL);
     dirty_it = dirty_start = dirty_end;
     v.clear();
     copy_count = 0;
@@ -1036,12 +1037,13 @@ void journal_flusher_co::scan_dirty()
             read_to_fill_incomplete = 0;
             return;
         }
+        uint8_t *bmp_ptr = bs->get_clean_entry_bitmap(old_clean_loc, 0);
         uint64_t fulfilled = 0;
         int last = v.size()-1;
         while (last >= 0 && (v[last].copy_flags & COPY_BUF_CSUM_FILL))
             last--;
         read_to_fill_incomplete = bs->fill_partial_checksum_blocks(
-            v, fulfilled, old_clean_bitmap, NULL, false, NULL, v[0].offset/bs->dsk.csum_block_size * bs->dsk.csum_block_size,
+            v, fulfilled, bmp_ptr, NULL, false, NULL, v[0].offset/bs->dsk.csum_block_size * bs->dsk.csum_block_size,
             ((v[last].offset+v[last].len-1) / bs->dsk.csum_block_size + 1) * bs->dsk.csum_block_size
         );
     }
@@ -1137,7 +1139,7 @@ bool journal_flusher_co::modify_meta_do_reads(int wait_base)
 resume_0:
     if (!modify_meta_read(clean_loc, meta_new, wait_base+0))
         return false;
-    new_clean_bitmap = (uint8_t*)meta_new.buf + meta_new.pos + sizeof(clean_disk_entry);
+    new_clean_bitmap = (uint8_t*)meta_new.buf + meta_new.pos*bs->dsk.clean_entry_size + sizeof(clean_disk_entry);
     if (old_clean_loc != UINT64_MAX && old_clean_loc != clean_loc)
     {
     resume_1:
@@ -1191,7 +1193,7 @@ bool journal_flusher_co::modify_meta_read(uint64_t meta_loc, flusher_meta_write_
     // so I'll avoid it as long as I can.
     wr.submitted = false;
     wr.sector = ((meta_loc >> bs->dsk.block_order) / (bs->dsk.meta_block_size / bs->dsk.clean_entry_size)) * bs->dsk.meta_block_size;
-    wr.pos = ((meta_loc >> bs->dsk.block_order) % (bs->dsk.meta_block_size / bs->dsk.clean_entry_size)) * bs->dsk.clean_entry_size;
+    wr.pos = ((meta_loc >> bs->dsk.block_order) % (bs->dsk.meta_block_size / bs->dsk.clean_entry_size));
     if (bs->inmemory_meta)
     {
         wr.buf = (uint8_t*)bs->metadata_buffer + wr.sector;

src/blockstore/blockstore_impl.cpp

@@ -42,8 +42,6 @@ blockstore_impl_t::~blockstore_impl_t()
         free(metadata_buffer);
     if (clean_bitmaps)
         free(clean_bitmaps);
-    if (heap_meta.blocks)
-        delete[] heap_meta.blocks;
 }
 
 bool blockstore_impl_t::is_started()
@@ -433,29 +431,13 @@ blockstore_clean_db_t& blockstore_impl_t::clean_db_shard(object_id oid)
 {
     uint64_t pg_num = 0;
     uint64_t pool_id = (oid.inode >> (64-POOL_ID_BITS));
-    auto sett_it = clean_db_settings.find(pool_id);
+    auto sh_it = clean_db_settings.find(pool_id);
-    if (sett_it != clean_db_settings.end())
+    if (sh_it != clean_db_settings.end())
     {
         // like map_to_pg()
-        pg_num = (oid.stripe / sett_it->second.pg_stripe_size) % sett_it->second.pg_count + 1;
+        pg_num = (oid.stripe / sh_it->second.pg_stripe_size) % sh_it->second.pg_count + 1;
     }
-    auto shard_id = (pool_id << (64-POOL_ID_BITS)) | pg_num;
+    return clean_db_shards[(pool_id << (64-POOL_ID_BITS)) | pg_num];
-    if (dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP)
-    {
-        auto sh_it = clean_db_shards.find(shard_id);
-        if (sh_it == clean_db_shards.end())
-        {
-            // clean_db_t stores larger entries with heap_meta, but we disguise it as smaller clean_entry :)
-            // patched cpp-btree with extra_data
-            clean_db_shards[shard_id] = blockstore_clean_db_t(
-                sizeof(clean_entry_heap_t) - sizeof(clean_entry)
-                + (inmemory_meta ? dsk.clean_dyn_size : 2*dsk.clean_entry_bitmap_size)
-            );
-            return clean_db_shards[shard_id];
-        }
-        return sh_it->second;
-    }
-    return clean_db_shards[shard_id];
 }
 
 void blockstore_impl_t::reshard_clean_db(pool_id_t pool, uint32_t pg_count, uint32_t pg_stripe_size)

src/blockstore/blockstore_impl.h

@@ -96,9 +96,6 @@
 #define BLOCKSTORE_META_MAGIC_V1 0x726F747341544956l
 #define BLOCKSTORE_META_FORMAT_V1 1
 #define BLOCKSTORE_META_FORMAT_V2 2
-#define BLOCKSTORE_META_FORMAT_HEAP 3
-#define BLOCKSTORE_META_HEADER_V1_SIZE 36
-#define BLOCKSTORE_META_HEADER_V2_SIZE 48
 
 // metadata header (superblock)
 struct __attribute__((__packed__)) blockstore_meta_header_v1_t
@@ -122,7 +119,6 @@ struct __attribute__((__packed__)) blockstore_meta_header_v2_t
     uint32_t data_csum_type;
     uint32_t csum_block_size;
     uint32_t header_csum;
-    uint32_t block_id_bits; // 32 by default in heap meta
 };
 
 // 32 bytes = 24 bytes + block bitmap (4 bytes by default) + external attributes (also bitmap, 4 bytes by default)
@@ -144,62 +140,6 @@ struct __attribute__((__packed__)) clean_entry
     uint64_t location;
 };
 
-typedef uint32_t heap_block_num_t;
-
-// 50 = 16 (key=object_id) + 26 (value) + 8 (bitmap) + N (checksum) bytes per "clean" entry in memory
-struct __attribute__((__packed__)) clean_entry_heap_t
-{
-    uint64_t version;
-    uint64_t location; // UINT64_MAX = deleted
-    // previous versions invalidated by this version
-    heap_block_num_t prev_versions;
-    // metadata block number
-    heap_block_num_t meta_block;
-    // offset within block
-    uint16_t block_offset;
-    uint8_t bitmap[];
-};
-
-struct __attribute__((__packed__)) heap_meta_block_header_t
-{
-    uint64_t magic;
-    uint64_t seq_num;
-    uint32_t invalidates_blocks;
-};
-
-// 48+checksums = (40+bitmap)+checksums bytes per on-disk "heap" entry
-// for 128 KB block without checksums, it's 48 bytes - 84 entries per 4 kb metadata block
-// for 128 KB block with 4k checksums, it's 176 bytes - 22 entries per 4 kb metadata block
-// for 1 MB block without checksums, it's 80 bytes - 50 entries per 4 kb metadata block
-// for 1 MB block with 4k checksums, it's 1104 bytes O_o - only 3 entries per 4 kb metadata block
-// for 1 MB block with 32k checksums, it's 176 bytes again
-struct __attribute__((__packed__)) heap_meta_entry_t
-{
-    object_id oid;
-    uint64_t version;
-    uint64_t location; // UINT64_MAX = deleted
-    uint64_t reserved;
-    uint8_t bitmap[];
-};
-
-struct heap_meta_block_t
-{
-    heap_block_num_t offset = 0;
-    uint64_t seq_num = 0;
-    uint32_t used_space = 0;
-    std::vector<uint64_t> invalidates_blocks;
-};
-
-struct heap_meta_t
-{
-    heap_block_num_t block_count = 0;
-    heap_meta_block_t *blocks = NULL;
-    // used space => block number
-    std::multimap<uint32_t, heap_block_num_t> used_space_map;
-    heap_block_num_t cur_written_block = 0;
-    uint8_t *written_block_buf = NULL;
-};
-
 // 64 = 24 + 40 bytes per dirty entry in memory (obj_ver_id => dirty_entry). Plus checksums
 struct __attribute__((__packed__)) dirty_entry
 {
@@ -332,8 +272,6 @@ class blockstore_impl_t
 
     struct ring_consumer_t ring_consumer;
 
-    heap_meta_t heap_meta;
-
     std::map<pool_id_t, pool_shard_settings_t> clean_db_settings;
     std::map<pool_pg_id_t, blockstore_clean_db_t> clean_db_shards;
     std::map<uint64_t, int> no_inode_stats;
@@ -379,7 +317,7 @@ class blockstore_impl_t
     void open_data();
     void open_meta();
     void open_journal();
-    uint8_t* get_clean_entry_bitmap(blockstore_clean_db_t::iterator clean_it, int offset);
+    uint8_t* get_clean_entry_bitmap(uint64_t block_loc, int offset);
 
     blockstore_clean_db_t& clean_db_shard(object_id oid);
     void reshard_clean_db(pool_id_t pool_id, uint32_t pg_count, uint32_t pg_stripe_size);
@@ -407,9 +345,9 @@ class blockstore_impl_t
         uint64_t &fulfilled, uint32_t item_start, uint32_t item_end,
         uint32_t item_state, uint64_t item_version, uint64_t item_location,
         uint64_t journal_sector, uint8_t *csum, int *dyn_data);
-    bool fulfill_clean_read_journal(blockstore_op_t *read_op, uint64_t & fulfilled,
+    bool fulfill_clean_read(blockstore_op_t *read_op, uint64_t & fulfilled,
-        uint8_t *clean_entry_bitmap, int *dyn_data, uint32_t item_start, uint32_t item_end, uint64_t clean_loc, uint64_t clean_ver);
+        uint8_t *clean_entry_bitmap, int *dyn_data,
-    bool fulfill_clean_read_meta(blockstore_op_t *read_op, uint64_t & fulfilled, blockstore_clean_db_t::iterator clean_it);
+        uint32_t item_start, uint32_t item_end, uint64_t clean_loc, uint64_t clean_ver);
     int fill_partial_checksum_blocks(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled,
         uint8_t *clean_entry_bitmap, int *dyn_data, bool from_journal, uint8_t *read_buf, uint64_t read_offset, uint64_t read_end);
     int pad_journal_read(std::vector<copy_buffer_t> & rv, copy_buffer_t & cp,
@@ -418,7 +356,7 @@ class blockstore_impl_t
     bool read_range_fulfilled(std::vector<copy_buffer_t> & rv, uint64_t & fulfilled, uint8_t *read_buf,
         uint8_t *clean_entry_bitmap, uint32_t item_start, uint32_t item_end);
     bool read_checksum_block(blockstore_op_t *op, int rv_pos, uint64_t &fulfilled, uint64_t clean_loc);
-    uint8_t* read_clean_meta_block(blockstore_op_t *op, blockstore_clean_db_t::iterator clean_it, int rv_pos);
+    uint8_t* read_clean_meta_block(blockstore_op_t *read_op, uint64_t clean_loc, int rv_pos);
     bool verify_padded_checksums(uint8_t *clean_entry_bitmap, uint8_t *csum_buf, uint32_t offset,
         iovec *iov, int n_iov, std::function<void(uint32_t, uint32_t, uint32_t)> bad_block_cb);
     bool verify_journal_checksums(uint8_t *csums, uint32_t offset,

src/blockstore/blockstore_init.cpp

@@ -54,7 +54,6 @@ int blockstore_init_meta::loop()
     else if (wait_state == 4) goto resume_4;
     else if (wait_state == 5) goto resume_5;
     else if (wait_state == 6) goto resume_6;
-    else if (wait_state == 7) goto resume_7;
     printf("Reading blockstore metadata\n");
     if (bs->inmemory_meta)
         metadata_buffer = bs->metadata_buffer;
@@ -79,7 +78,6 @@ resume_1:
     if (iszero((uint64_t*)metadata_buffer, bs->dsk.meta_block_size / sizeof(uint64_t)))
     {
         {
-            memset(metadata_buffer, 0, bs->dsk.meta_block_size);
             blockstore_meta_header_v2_t *hdr = (blockstore_meta_header_v2_t *)metadata_buffer;
             hdr->zero = 0;
             hdr->magic = BLOCKSTORE_META_MAGIC_V1;
@@ -87,19 +85,12 @@ resume_1:
             hdr->meta_block_size = bs->dsk.meta_block_size;
             hdr->data_block_size = bs->dsk.data_block_size;
             hdr->bitmap_granularity = bs->dsk.bitmap_granularity;
-            if (bs->dsk.meta_format >= BLOCKSTORE_META_FORMAT_HEAP)
-            {
-                hdr->block_id_bits = sizeof(heap_block_num_t);
-            }
             if (bs->dsk.meta_format >= BLOCKSTORE_META_FORMAT_V2)
             {
                 hdr->data_csum_type = bs->dsk.data_csum_type;
                 hdr->csum_block_size = bs->dsk.csum_block_size;
                 hdr->header_csum = 0;
-                hdr->header_csum = crc32c(0, hdr,
+                hdr->header_csum = crc32c(0, hdr, sizeof(*hdr));
-                    bs->dsk.meta_format == BLOCKSTORE_META_FORMAT_V2
-                    ? BLOCKSTORE_META_HEADER_V2_SIZE
-                    : sizeof(*hdr));
             }
         }
         if (bs->readonly)
@@ -137,7 +128,7 @@ resume_1:
             );
             exit(1);
         }
-        if (hdr->version == BLOCKSTORE_META_FORMAT_HEAP)
+        if (hdr->version == BLOCKSTORE_META_FORMAT_V2)
         {
             uint32_t csum = hdr->header_csum;
             hdr->header_csum = 0;
@@ -147,23 +138,6 @@ resume_1:
                 exit(1);
             }
             hdr->header_csum = csum;
-            bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_HEAP;
-            if (hdr->block_id_bits != sizeof(heap_block_num_t))
-            {
-                printf("Heap metadata block ID size (%u) is not supported by this build\n", hdr->block_id_bits);
-                exit(1);
-            }
-        }
-        else if (hdr->version == BLOCKSTORE_META_FORMAT_V2)
-        {
-            uint32_t csum = hdr->header_csum;
-            hdr->header_csum = 0;
-            if (crc32c(0, hdr, BLOCKSTORE_META_HEADER_V2_SIZE) != csum)
-            {
-                printf("Metadata header is corrupt (checksum mismatch).\n");
-                exit(1);
-            }
-            hdr->header_csum = csum;
             if (bs->dsk.meta_format != BLOCKSTORE_META_FORMAT_V2)
             {
                 bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_V2;
@@ -186,11 +160,11 @@ resume_1:
                 printf("Warning: Starting with metadata in the old format without checksums, as stored on disk\n");
             }
         }
-        else
+        else if (hdr->version > BLOCKSTORE_META_FORMAT_V2)
         {
             printf(
                 "Metadata format is too new for me (stored version is %ju, max supported %u).\n",
-                hdr->version, BLOCKSTORE_META_FORMAT_HEAP
+                hdr->version, BLOCKSTORE_META_FORMAT_V2
             );
             exit(1);
         }
@@ -215,12 +189,7 @@ resume_1:
     // Skip superblock
     md_offset = bs->dsk.meta_block_size;
     next_offset = md_offset;
-    entries_per_block = bs->dsk.meta_block_size / bs->dsk.clean_entry_size; // FIXME only array
+    entries_per_block = bs->dsk.meta_block_size / bs->dsk.clean_entry_size;
-    if (bs->dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP)
-    {
-        bs->heap_meta.blocks = new heap_meta_block_t[bs->dsk.meta_len / bs->dsk.meta_block_size];
-        bs->heap_meta.block_count = bs->dsk.meta_len / bs->dsk.meta_block_size;
-    }
     // Read the rest of the metadata
 resume_2:
     if (next_offset < bs->dsk.meta_len && submitted == 0)
@@ -264,10 +233,9 @@ resume_2:
             bool changed = false;
             for (uint64_t sector = 0; sector < bufs[i].size; sector += bs->dsk.meta_block_size)
             {
-                auto this_changed = bs->dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP
+                // handle <count> entries
-                    ? handle_heap_meta_block(bufs[i].buf + sector, bufs[i].offset + sector - md_offset)
+                if (handle_meta_block(bufs[i].buf + sector, entries_per_block,
-                    : handle_array_meta_block(bufs[i].buf + sector, bufs[i].offset + sector - md_offset);
+                    ((bufs[i].offset + sector - md_offset) / bs->dsk.meta_block_size) * entries_per_block))
-                if (this_changed)
                     changed = true;
             }
             if (changed && !bs->inmemory_meta && !bs->readonly)
@@ -294,41 +262,6 @@ resume_2:
         wait_state = 2;
         return 1;
     }
-    if (bs->dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP)
-    {
-        // build used_space index
-        for (heap_block_num_t i = 0; i < bs->heap_meta.block_count; i++)
-        {
-            bs->heap_meta.used_space_map.emplace(std::pair<uint32_t, heap_block_num_t>(bs->heap_meta.blocks[i].used_space, i));
-        }
-    }
-    if (heap_invalidated_block_seq.size() && !bs->readonly)
-    {
-        // zero out invalidated blocks not zeroed during the previous OSD execution
-        for (auto inv_seq: heap_invalidated_block_seq)
-        {
-            auto num_it = heap_block_by_seq.find(inv_seq);
-            if (num_it != heap_block_by_seq.end())
-                heap_invalidated_block_nums.push_back(num_it->second);
-        }
-        memset(metadata_buffer, 0, bs->dsk.meta_block_size);
-        for (i = 0; i < heap_invalidated_block_nums.size(); i++)
-        {
-            GET_SQE();
-            last_read_offset = heap_invalidated_block_nums[i]*bs->dsk.meta_block_size;
-            data->iov = { metadata_buffer, (size_t)bs->dsk.meta_block_size };
-            data->callback = [this](ring_data_t *data) { handle_event(data, -1); };
-            my_uring_prep_writev(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + last_read_offset);
-            bs->ringloop->submit();
-            submitted++;
-resume_7:
-            if (submitted > 0)
-            {
-                wait_state = 7;
-                return 1;
-            }
-        }
-    }
     if (entries_to_zero.size() && !bs->inmemory_meta && !bs->readonly)
     {
         std::sort(entries_to_zero.begin(), entries_to_zero.end());
@@ -396,9 +329,8 @@ resume_6:
     return 0;
 }
 
-bool blockstore_init_meta::handle_array_meta_block(uint8_t *buf, uint64_t block_offset)
+bool blockstore_init_meta::handle_meta_block(uint8_t *buf, uint64_t entries_per_block, uint64_t done_cnt)
 {
-    uint64_t done_cnt = (block_offset / bs->dsk.meta_block_size) * entries_per_block;
     bool updated = false;
     uint64_t max_i = entries_per_block;
     if (max_i > bs->dsk.block_count-done_cnt)
@@ -497,132 +429,6 @@ bool blockstore_init_meta::handle_array_meta_block(uint8_t *buf, uint64_t block_
     return updated;
 }
 
-static int bitmap_count_ones(uint8_t *bitmap, int size)
-{
-    int n = 0, i = 0;
-    for (; i <= size-sizeof(unsigned); i += sizeof(unsigned))
-    {
-        n += __builtin_popcount(*(unsigned*)(bitmap+i));
-    }
-    for (; i < size; i++)
-    {
-        n += __builtin_popcount(*(unsigned char*)(bitmap+i));
-    }
-    return n;
-}
-
-// v3 / heap / "cow" metadata block
-bool blockstore_init_meta::handle_heap_meta_block(uint8_t *buf, uint64_t block_offset)
-{
-    if ((block_offset / bs->dsk.meta_block_size) > (heap_block_num_t)-1)
-    {
-        fprintf(stderr, "Metadata area too large\n");
-        exit(1);
-    }
-    // Validate block CRC
-    uint32_t block_crc = *(uint32_t*)(buf + bs->dsk.meta_block_size - 4);
-    if (crc32c(0, buf, bs->dsk.meta_block_size-4) != block_crc)
-    {
-        return false;
-    }
-    // Validate header
-    heap_meta_block_header_t *hdr = (heap_meta_block_header_t*)buf;
-    if (hdr->magic != BLOCKSTORE_META_MAGIC_V1)
-    {
-        return false;
-    }
-    if (hdr->invalidates_blocks > (bs->dsk.meta_block_size-4-sizeof(heap_meta_block_header_t))/sizeof(uint64_t))
-    {
-        fprintf(stderr, "Metadata block at %jx contains too large invalidates_blocks count: %x\n", block_offset, hdr->invalidates_blocks);
-        exit(1);
-    }
-    if (heap_invalidated_block_seq.find(hdr->seq_num) != heap_invalidated_block_seq.end())
-    {
-        // Check if the block is invalidated and handled after the block that invalidates it
-        return false;
-    }
-    uint64_t hdr_size = sizeof(heap_meta_block_header_t) + hdr->invalidates_blocks*8;
-    heap_meta_block_t & blk = bs->heap_meta.blocks[block_offset/bs->dsk.meta_block_size];
-    blk.offset = block_offset;
-    blk.seq_num = hdr->seq_num;
-    blk.used_space = hdr_size + 4;
-    uint64_t *hdr_inv = (uint64_t*)(hdr + 1);
-    for (int i = 0; i < hdr->invalidates_blocks; i++)
-    {
-        blk.invalidates_blocks.push_back(hdr_inv[i]);
-        heap_invalidated_block_seq.insert(hdr_inv[i]);
-    }
-    heap_block_by_seq[hdr->seq_num] = block_offset;
-    // Process sub-blocks
-    uint64_t heap_entry_size = sizeof(heap_meta_entry_t) + bs->dsk.clean_dyn_size;
-    for (uint64_t pos = sizeof(heap_meta_block_header_t); pos < bs->dsk.meta_block_size-4; pos += heap_entry_size)
-    {
-        heap_meta_entry_t *diskentry = (heap_meta_entry_t*)(buf + pos);
-        if (!diskentry->oid.inode || !diskentry->version)
-        {
-            continue;
-        }
-        auto & clean_db = bs->clean_db_shard(diskentry->oid);
-        auto mementry = (clean_entry_heap_t*)(&clean_db[diskentry->oid]);
-        bool exists = mementry->version != 0;
-        if (exists && mementry->version >= diskentry->version)
-        {
-            if (mementry->version == diskentry->version)
-            {
-                // Voluntarily allow duplicates of in-memory entries with different
-                // bitmaps to support checksum updates with hole-punching
-                int old_count = bitmap_count_ones(mementry->bitmap, bs->dsk.clean_entry_bitmap_size);
-                int new_count = bitmap_count_ones(diskentry->bitmap, bs->dsk.clean_entry_bitmap_size);
-                if (old_count < new_count)
-                {
-                    continue;
-                }
-            }
-            else
-            {
-                continue;
-            }
-        }
-        blk.used_space += heap_entry_size;
-        if (exists && mementry->location != UINT64_MAX)
-        {
-            // free the previous block
-            uint64_t old_clean_loc = mementry->location >> bs->dsk.block_order;
-#ifdef BLOCKSTORE_DEBUG
-            printf("Free block %ju from %jx:%jx v%ju\n", 1+old_clean_loc,
-                diskentry->oid.inode, diskentry->oid.stripe, mementry->version);
-#endif
-            bs->data_alloc->set(old_clean_loc, false);
-            bs->inode_space_stats[diskentry->oid.inode] -= bs->dsk.data_block_size;
-            bs->used_blocks--;
-            bs->heap_meta.blocks[mementry->meta_block].used_space -= heap_entry_size;
-        }
-        if (diskentry->location != UINT64_MAX)
-        {
-            bs->data_alloc->set(diskentry->location >> bs->dsk.block_order, true);
-            bs->inode_space_stats[diskentry->oid.inode] += bs->dsk.data_block_size;
-            bs->used_blocks++;
-#ifdef BLOCKSTORE_DEBUG
-            printf("Allocate block (heap entry) %ju: %jx:%jx v%ju\n", 1 + (diskentry->location >> bs->dsk.block_order),
-                diskentry->oid.inode, diskentry->oid.stripe, diskentry->version);
-#endif
-        }
-        mementry->version = diskentry->version;
-        mementry->location = diskentry->location;
-        mementry->meta_block = block_offset / bs->dsk.meta_block_size;
-        mementry->block_offset = block_offset % bs->dsk.meta_block_size;
-        if (exists)
-        {
-            mementry->prev_versions++;
-        }
-        // Extra data: 2 bitmaps + checksums or just 2 bitmaps if inmemory_meta is disabled
-        memcpy(&mementry->bitmap, &diskentry->bitmap, bs->inmemory_meta ? bs->dsk.clean_dyn_size : 2*bs->dsk.clean_entry_bitmap_size);
-        entries_loaded++;
-    }
-    // We have to zero out headers of invalidated blocks, but we'll do it later
-    return false;
-}
-
 blockstore_init_journal::blockstore_init_journal(blockstore_impl_t *bs)
 {
     this->bs = bs;

src/blockstore/blockstore_init.h

@@ -28,13 +28,7 @@ class blockstore_init_meta
     unsigned entries_per_block = 0;
     int i = 0, j = 0;
     std::vector<uint64_t> entries_to_zero;
-
+    bool handle_meta_block(uint8_t *buf, uint64_t count, uint64_t done_cnt);
-    std::map<uint64_t, heap_block_num_t> heap_block_by_seq;
-    std::set<uint64_t> heap_invalidated_block_seq;
-    std::vector<heap_block_num_t> heap_invalidated_block_nums;
-
-    bool handle_array_meta_block(uint8_t *buf, uint64_t block_offset);
-    bool handle_heap_meta_block(uint8_t *buf, uint64_t block_offset);
     void handle_event(ring_data_t *data, int buf_num);
 public:
     blockstore_init_meta(blockstore_impl_t *bs);

src/blockstore/blockstore_open.cpp

@@ -111,10 +111,6 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     {
         metadata_buf_size = 4*1024*1024;
     }
-    if (metadata_buf_size % dsk.meta_block_size)
-    {
-        metadata_buf_size = ((metadata_buf_size+dsk.meta_block_size-1) / dsk.meta_block_size) * dsk.meta_block_size;
-    }
     if (dsk.meta_device == dsk.data_device)
     {
         disable_meta_fsync = disable_data_fsync;

src/blockstore/blockstore_read.cpp

@@ -148,14 +148,10 @@ int blockstore_impl_t::fulfill_read(blockstore_op_t *read_op,
     return r;
 }
 
-uint8_t* blockstore_impl_t::get_clean_entry_bitmap(blockstore_clean_db_t::iterator clean_it, int offset)
+uint8_t* blockstore_impl_t::get_clean_entry_bitmap(uint64_t block_loc, int offset)
 {
-    if (dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP)
-    {
-        return ((uint8_t*)&clean_it->second) + sizeof(clean_entry_heap_t) + offset;
-    }
     uint8_t *clean_entry_bitmap;
-    uint64_t meta_loc = clean_it->second.location >> dsk.block_order;
+    uint64_t meta_loc = block_loc >> dsk.block_order;
     if (inmemory_meta)
     {
         uint64_t sector = (meta_loc / (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.meta_block_size;
@@ -163,9 +159,7 @@ uint8_t* blockstore_impl_t::get_clean_entry_bitmap(blockstore_clean_db_t::iterat
         clean_entry_bitmap = ((uint8_t*)metadata_buffer + sector + pos*dsk.clean_entry_size + sizeof(clean_disk_entry) + offset);
     }
     else
-    {
         clean_entry_bitmap = (uint8_t*)(clean_bitmaps + meta_loc*2*dsk.clean_entry_bitmap_size + offset);
-    }
     return clean_entry_bitmap;
 }
 
@@ -439,7 +433,7 @@ int blockstore_impl_t::dequeue_read(blockstore_op_t *read_op)
                 if (!IS_JOURNAL(dirty.state))
                 {
                     // Read from data disk, possibly checking checksums
-                    if (!fulfill_clean_read_journal(read_op, fulfilled, bmp_ptr, dyn_data,
+                    if (!fulfill_clean_read(read_op, fulfilled, bmp_ptr, dyn_data,
                         dirty.offset, dirty.offset+dirty.len, dirty.location, dirty_it->first.version))
                     {
                         goto undo_read;
@@ -470,13 +464,14 @@ int blockstore_impl_t::dequeue_read(blockstore_op_t *read_op)
             result_version = clean_it->second.version;
             if (read_op->bitmap)
             {
-                void *bmp_ptr = get_clean_entry_bitmap(clean_it, dsk.clean_entry_bitmap_size);
+                void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, dsk.clean_entry_bitmap_size);
                 memcpy(read_op->bitmap, bmp_ptr, dsk.clean_entry_bitmap_size);
             }
         }
         if (fulfilled < read_op->len)
         {
-            if (!fulfill_clean_read_meta(read_op, fulfilled, clean_it))
+            if (!fulfill_clean_read(read_op, fulfilled, NULL, NULL, 0, dsk.data_block_size,
+                clean_it->second.location, clean_it->second.version))
             {
                 goto undo_read;
             }
@@ -586,22 +581,40 @@ int blockstore_impl_t::pad_journal_read(std::vector<copy_buffer_t> & rv, copy_bu
     return 0;
 }
 
-bool blockstore_impl_t::fulfill_clean_read_journal(blockstore_op_t *read_op, uint64_t & fulfilled,
+bool blockstore_impl_t::fulfill_clean_read(blockstore_op_t *read_op, uint64_t & fulfilled,
     uint8_t *clean_entry_bitmap, int *dyn_data, uint32_t item_start, uint32_t item_end, uint64_t clean_loc, uint64_t clean_ver)
 {
+    bool from_journal = clean_entry_bitmap != NULL;
+    if (!clean_entry_bitmap)
+    {
+        // NULL clean_entry_bitmap means we're reading from data, not from the journal,
+        // and the bitmap location is obvious
+        clean_entry_bitmap = get_clean_entry_bitmap(clean_loc, 0);
+    }
     if (dsk.csum_block_size > dsk.bitmap_granularity)
     {
         auto & rv = PRIV(read_op)->read_vec;
-        int req = fill_partial_checksum_blocks(rv, fulfilled, clean_entry_bitmap, dyn_data, true,
+        int req = fill_partial_checksum_blocks(rv, fulfilled, clean_entry_bitmap, dyn_data, from_journal,
             (uint8_t*)read_op->buf, read_op->offset, read_op->offset+read_op->len);
+        if (!inmemory_meta && !from_journal && req > 0)
+        {
+            // Read checksums from disk
+            uint8_t *csum_buf = read_clean_meta_block(read_op, clean_loc, rv.size()-req);
+            for (int i = req; i > 0; i--)
+            {
+                rv[rv.size()-i].csum_buf = csum_buf;
+            }
+        }
         for (int i = req; i > 0; i--)
         {
             if (!read_checksum_block(read_op, i, fulfilled, clean_loc))
+            {
                 return false;
+            }
         }
         PRIV(read_op)->clean_block_used = req > 0;
     }
-    else
+    else if (from_journal)
     {
         // Don't scan bitmap - journal writes don't have holes (internal bitmap)!
         uint8_t *csum = !dsk.csum_block_size ? 0 : (clean_entry_bitmap + dsk.clean_entry_bitmap_size +
@@ -622,43 +635,6 @@ bool blockstore_impl_t::fulfill_clean_read_journal(blockstore_op_t *read_op, uin
             assert(fulfill_read(read_op, fulfilled, item_end, dsk.data_block_size, (BS_ST_DELETE | BS_ST_STABLE), 0, 0, 0, NULL, NULL));
         }
     }
-    // Increment reference counter if clean data is being read from the disk
-    if (PRIV(read_op)->clean_block_used)
-    {
-        auto & uo = used_clean_objects[clean_loc];
-        uo.refs++;
-        if (dsk.csum_block_size && flusher->is_mutated(clean_loc))
-            uo.was_changed = true;
-        PRIV(read_op)->clean_block_used = clean_loc;
-    }
-    return true;
-}
-
-bool blockstore_impl_t::fulfill_clean_read_meta(blockstore_op_t *read_op, uint64_t & fulfilled, blockstore_clean_db_t::iterator clean_it)
-{
-    uint8_t *clean_entry_bitmap = get_clean_entry_bitmap(clean_it, 0);
-    uint64_t clean_loc = clean_it->second.location;
-    if (dsk.csum_block_size > dsk.bitmap_granularity)
-    {
-        auto & rv = PRIV(read_op)->read_vec;
-        int req = fill_partial_checksum_blocks(rv, fulfilled, clean_entry_bitmap, NULL, false,
-            (uint8_t*)read_op->buf, read_op->offset, read_op->offset+read_op->len);
-        if (!inmemory_meta && req > 0)
-        {
-            // Read checksums from disk
-            uint8_t *csum_buf = read_clean_meta_block(read_op, clean_it, rv.size()-req);
-            for (int i = req; i > 0; i--)
-            {
-                rv[rv.size()-i].csum_buf = csum_buf;
-            }
-        }
-        for (int i = req; i > 0; i--)
-        {
-            if (!read_checksum_block(read_op, i, fulfilled, clean_loc))
-                return false;
-        }
-        PRIV(read_op)->clean_block_used = req > 0;
-    }
     else
     {
         bool csum_done = !dsk.csum_block_size || inmemory_meta;
@@ -686,13 +662,13 @@ bool blockstore_impl_t::fulfill_clean_read_meta(blockstore_op_t *read_op, uint64
                 if (!csum_done)
                 {
                     // Read checksums from disk
-                    csum_buf = read_clean_meta_block(read_op, clean_it, PRIV(read_op)->read_vec.size());
+                    csum_buf = read_clean_meta_block(read_op, clean_loc, PRIV(read_op)->read_vec.size());
                     csum_done = true;
                 }
                 uint8_t *csum = !dsk.csum_block_size ? 0 : (csum_buf + 2*dsk.clean_entry_bitmap_size + bmp_start*(dsk.data_csum_type & 0xFF));
                 if (!fulfill_read(read_op, fulfilled, bmp_start * dsk.bitmap_granularity,
                     bmp_end * dsk.bitmap_granularity, (BS_ST_BIG_WRITE | BS_ST_STABLE), 0,
-                    clean_loc + bmp_start * dsk.bitmap_granularity, 0, csum, NULL))
+                    clean_loc + bmp_start * dsk.bitmap_granularity, 0, csum, dyn_data))
                 {
                     return false;
                 }
@@ -712,22 +688,11 @@ bool blockstore_impl_t::fulfill_clean_read_meta(blockstore_op_t *read_op, uint64
     return true;
 }
 
-uint8_t* blockstore_impl_t::read_clean_meta_block(blockstore_op_t *op, blockstore_clean_db_t::iterator clean_it, int rv_pos)
+uint8_t* blockstore_impl_t::read_clean_meta_block(blockstore_op_t *op, uint64_t clean_loc, int rv_pos)
 {
-    uint64_t sector, pos;
     auto & rv = PRIV(op)->read_vec;
-    if (dsk.meta_format == BLOCKSTORE_META_FORMAT_HEAP)
+    auto sector = ((clean_loc >> dsk.block_order) / (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.meta_block_size;
-    {
+    auto pos = ((clean_loc >> dsk.block_order) % (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.clean_entry_size;
-        auto clean_heap_entry = (clean_entry_heap_t*)(&clean_it->second);
-        sector = clean_heap_entry->meta_block * dsk.meta_block_size;
-        pos = clean_heap_entry->block_offset;
-    }
-    else
-    {
-        auto clean_loc = clean_it->second.location;
-        sector = ((clean_loc >> dsk.block_order) / (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.meta_block_size;
-        pos = ((clean_loc >> dsk.block_order) % (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.clean_entry_size;
-    }
     uint8_t *buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.meta_block_size);
     rv.insert(rv.begin()+rv_pos, (copy_buffer_t){
         .copy_flags = COPY_BUF_META_BLOCK|COPY_BUF_CSUM_FILL,
@@ -842,6 +807,11 @@ bool blockstore_impl_t::verify_clean_padded_checksums(blockstore_op_t *op, uint6
     if (from_journal)
         return verify_padded_checksums(dyn_data, dyn_data + dsk.clean_entry_bitmap_size, offset, iov, n_iov, bad_block_cb);
     clean_loc = (clean_loc >> dsk.block_order) << dsk.block_order;
+    if (!dyn_data)
+    {
+        assert(inmemory_meta);
+        dyn_data = get_clean_entry_bitmap(clean_loc, 0);
+    }
     return verify_padded_checksums(dyn_data, dyn_data + 2*dsk.clean_entry_bitmap_size, offset, iov, n_iov, bad_block_cb);
 }
 
@@ -899,18 +869,8 @@ void blockstore_impl_t::handle_read_event(ring_data_t *data, blockstore_op_t *op
                         auto & uo = used_clean_objects.at((rv[i].disk_offset >> dsk.block_order) << dsk.block_order);
                         if (!uo.was_changed)
                         {
-                            bool from_journal = (rv[i].copy_flags & COPY_BUF_JOURNALED_BIG);
-                            auto csum_buf = rv[i].csum_buf;
-                            if (!from_journal && !csum_buf)
-                            {
-                                assert(inmemory_meta);
-                                auto & clean_db = clean_db_shard(op->oid);
-                                auto clean_it = clean_db.find(op->oid);
-                                assert(clean_it != clean_db.end());
-                                csum_buf = get_clean_entry_bitmap(clean_it, 0);
-                            }
                             verify_clean_padded_checksums(
-                                op, rv[i].disk_offset, csum_buf, from_journal, iov, n_iov,
+                                op, rv[i].disk_offset, rv[i].csum_buf, (rv[i].copy_flags & COPY_BUF_JOURNALED_BIG), iov, n_iov,
                                 [&](uint32_t bad_block, uint32_t calc_csum, uint32_t stored_csum)
                                 {
                                     ok = false;
@@ -1059,7 +1019,7 @@ int blockstore_impl_t::read_bitmap(object_id oid, uint64_t target_version, void
             *result_version = clean_it->second.version;
         if (bitmap)
         {
-            void *bmp_ptr = get_clean_entry_bitmap(clean_it, dsk.clean_entry_bitmap_size);
+            void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, dsk.clean_entry_bitmap_size);
             memcpy(bitmap, bmp_ptr, dsk.clean_entry_bitmap_size);
         }
         return 0;

src/blockstore/blockstore_write.cpp

@@ -57,7 +57,7 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
             version = clean_it->second.version + 1;
             if (!is_del)
             {
-                void *bmp_ptr = get_clean_entry_bitmap(clean_it, dsk.clean_entry_bitmap_size);
+                void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, dsk.clean_entry_bitmap_size);
                 memcpy(dyn_ptr, bmp_ptr, dsk.clean_entry_bitmap_size);
             }
         }
@@ -341,7 +341,7 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             cancel_all_writes(op, dirty_it, -ENOSPC);
             return 2;
         }
-        if (inmemory_meta && dsk.meta_format != BLOCKSTORE_META_FORMAT_HEAP)
+        if (inmemory_meta)
         {
             // Check once more that metadata entry is zeroed (the reverse means a bug or corruption)
             uint64_t sector = (loc / (dsk.meta_block_size / dsk.clean_entry_size)) * dsk.meta_block_size;

src/client/msgr_receive.cpp

@@ -33,8 +33,12 @@ void osd_messenger_t::read_requests()
             auto iothread = iothreads.size() ? iothreads[peer_fd % iothreads.size()] : NULL;
             io_uring_sqe sqe_local;
             ring_data_t data_local;
-            sqe_local.user_data = (uint64_t)&data_local;
             io_uring_sqe* sqe = (iothread ? &sqe_local : ringloop->get_sqe());
+            if (iothread)
+            {
+                sqe_local = { .user_data = (uint64_t)&data_local };
+                data_local = {};
+            }
             if (!sqe)
             {
                 cl->read_msg.msg_iovlen = 0;

src/client/msgr_send.cpp

@@ -194,12 +194,14 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
         auto iothread = iothreads.size() ? iothreads[peer_fd % iothreads.size()] : NULL;
         io_uring_sqe sqe_local;
         ring_data_t data_local;
-        sqe_local.user_data = (uint64_t)&data_local;
         io_uring_sqe* sqe = (iothread ? &sqe_local : ringloop->get_sqe());
-        if (!sqe)
+        if (iothread)
         {
-            return false;
+            sqe_local = { .user_data = (uint64_t)&data_local };
+            data_local = {};
         }
+        if (!sqe)
+            return false;
         cl->write_msg.msg_iov = cl->send_list.data();
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
         cl->refs++;

src/cmd/cli_rm_osd.cpp

@@ -70,7 +70,7 @@ struct rm_osd_t
         {
             if (parent->cli->st_cli.peer_states.find(osd_id) != parent->cli->st_cli.peer_states.end())
             {
-                is_warning = true;
+                is_warning = !allow_up;
                 still_up.push_back(osd_id);
             }
         }

src/osd/osd.h

@@ -278,6 +278,7 @@ class osd_t
     void handle_peers();
     bool check_peer_config(osd_client_t *cl, json11::Json conf);
     void repeer_pgs(osd_num_t osd_num);
+    void repeer_pg(pg_t & pg);
     void start_pg_peering(pg_t & pg);
     void drop_dirty_pg_connections(pool_pg_num_t pg);
     void record_pg_lock(pg_t & pg, osd_num_t peer_osd, uint64_t pg_state);

src/osd/osd_cluster.cpp

@@ -432,9 +432,16 @@ void osd_t::apply_pg_locks_localize_only()
         }
         auto & pool_cfg = pool_it->second;
         auto & pg = pp.second;
+        auto old_disable_pg_locks = pg.disable_pg_locks;
         pg.disable_pg_locks = pg_locks_localize_only &&
             pool_cfg.scheme == POOL_SCHEME_REPLICATED &&
             pool_cfg.local_reads == POOL_LOCAL_READ_PRIMARY;
+        if (!pg.disable_pg_locks && old_disable_pg_locks)
+        {
+            // Relock PG
+            printf("[PG %u/%u] Repeer to enable PG locks\n", pg.pool_id, pg.pg_num);
+            repeer_pg(pg);
+        }
     }
 }
 

src/osd/osd_peering.cpp

@@ -104,21 +104,26 @@ void osd_t::repeer_pgs(osd_num_t peer_osd)
             {
                 // Repeer this pg
                 printf("[PG %u/%u] Repeer because of OSD %ju\n", pg.pool_id, pg.pg_num, peer_osd);
-                if (!(pg.state & (PG_ACTIVE | PG_REPEERING)) || pg.can_repeer())
+                repeer_pg(pg);
-                {
-                    start_pg_peering(pg);
-                }
-                else
-                {
-                    // Stop accepting new operations, wait for current ones to finish or fail
-                    pg.state = pg.state & ~PG_ACTIVE | PG_REPEERING;
-                    report_pg_state(pg);
-                }
             }
         }
     }
 }
 
+void osd_t::repeer_pg(pg_t & pg)
+{
+    if (!(pg.state & (PG_ACTIVE | PG_REPEERING)) || pg.can_repeer())
+    {
+        start_pg_peering(pg);
+    }
+    else
+    {
+        // Stop accepting new operations, wait for current ones to finish or fail
+        pg.state = pg.state & ~PG_ACTIVE | PG_REPEERING;
+        report_pg_state(pg);
+    }
+}
+
 // Reset PG state (when peering or stopping)
 void osd_t::reset_pg(pg_t & pg)
 {

src/util/ringloop.cpp

@@ -125,6 +125,8 @@ void ring_loop_t::loop()
         if (cqe->flags & IORING_CQE_F_MORE)
         {
             // There will be a second notification
+            if (mt)
+                mu.unlock();
             d->res = cqe->res;
             d->more = true;
             if (d->callback)

tests/run_tests.sh

@@ -59,6 +59,7 @@ SCHEME=ec IMMEDIATE_COMMIT=1 ./test_rebalance_verify.sh
 
 ./test_write.sh
 SCHEME=xor ./test_write.sh
+TEST_NAME=iothreads GLOBAL_CONFIG=',"client_iothread_count":4' ./test_write.sh
 
 ./test_write_no_same.sh