WIP Send additional fsyncs from secondary OSDs to primaries to prevent lockups

- Disabled by default, enable with client_enable_writeback=true
- Even then only enabled in FIO when -direct is disabled and in QEMU when
  block device cache is enabled in settings
- Can also be enabled in other clients like vitastor-cli using parameter
  client_writeback_allowed=true, but not recommended

CMakeLists.txt

@@ -2,6 +2,6 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VERSION "0.9.3")
+set(VERSION "1.0.0")
 
 add_subdirectory(src)

csi/Makefile

@@ -1,4 +1,4 @@
-VERSION ?= v0.9.3
+VERSION ?= v1.0.0
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v0.9.3
+          image: vitalif/vitastor-csi:v1.0.0
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/007-csi-provisioner.yaml

@@ -116,7 +116,7 @@ spec:
             privileged: true
             capabilities:
               add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v0.9.3
+          image: vitalif/vitastor-csi:v1.0.0
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "0.9.3"
+    vitastorCSIDriverVersion = "1.0.0"
 )
 
 // Config struct fills the parameters of request or user input

debian/changelog

@@ -1,10 +1,10 @@
-vitastor (0.9.3-1) unstable; urgency=medium
+vitastor (1.0.0-1) unstable; urgency=medium
 
   * Bugfixes
 
  -- Vitaliy Filippov <vitalif@yourcmc.ru>  Fri, 03 Jun 2022 02:09:44 +0300
 
-vitastor (0.9.3-1) unstable; urgency=medium
+vitastor (1.0.0-1) unstable; urgency=medium
 
   * Implement NFS proxy
   * Add documentation

debian/patched-qemu.Dockerfile

@@ -28,13 +28,19 @@ RUN apt-get --download-only source qemu
 
 ADD patches /root/vitastor/patches
 ADD src/qemu_driver.c /root/vitastor/src/qemu_driver.c
+
+#RUN set -e; \
+#    apt-get install -y wget; \
+#    wget -q -O /etc/apt/trusted.gpg.d/vitastor.gpg https://vitastor.io/debian/pubkey.gpg; \
+#    (echo deb http://vitastor.io/debian $REL main > /etc/apt/sources.list.d/vitastor.list); \
+#    (echo "APT::Install-Recommends false;" > /etc/apt/apt.conf) && \
+#    apt-get update; \
+#    apt-get install -y vitastor-client vitastor-client-dev quilt
+
 RUN set -e; \
-    apt-get install -y wget; \
-    wget -q -O /etc/apt/trusted.gpg.d/vitastor.gpg https://vitastor.io/debian/pubkey.gpg; \
-    (echo deb http://vitastor.io/debian $REL main > /etc/apt/sources.list.d/vitastor.list); \
-    (echo "APT::Install-Recommends false;" > /etc/apt/apt.conf) && \
+    dpkg -i /root/packages/vitastor-$REL/vitastor-client_*.deb /root/packages/vitastor-$REL/vitastor-client-dev_*.deb; \
     apt-get update; \
-    apt-get install -y vitastor-client vitastor-client-dev quilt; \
+    apt-get install -y quilt; \
     mkdir -p /root/packages/qemu-$REL; \
     rm -rf /root/packages/qemu-$REL/*; \
     cd /root/packages/qemu-$REL; \

debian/vitastor.Dockerfile

@@ -35,8 +35,8 @@ RUN set -e -x; \
     mkdir -p /root/packages/vitastor-$REL; \
     rm -rf /root/packages/vitastor-$REL/*; \
     cd /root/packages/vitastor-$REL; \
-    cp -r /root/vitastor vitastor-0.9.3; \
-    cd vitastor-0.9.3; \
+    cp -r /root/vitastor vitastor-1.0.0; \
+    cd vitastor-1.0.0; \
     ln -s /root/fio-build/fio-*/ ./fio; \
     FIO=$(head -n1 fio/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     ls /usr/include/linux/raw.h || cp ./debian/raw.h /usr/include/linux/raw.h; \
@@ -49,8 +49,8 @@ RUN set -e -x; \
     rm -rf a b; \
     echo "dep:fio=$FIO" > debian/fio_version; \
     cd /root/packages/vitastor-$REL; \
-    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.9.3.orig.tar.xz vitastor-0.9.3; \
-    cd vitastor-0.9.3; \
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_1.0.0.orig.tar.xz vitastor-1.0.0; \
+    cd vitastor-1.0.0; \
     V=$(head -n1 debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     DEBFULLNAME="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v "$V""$REL" "Rebuild for $REL"; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \

docs/config/layout-cluster.en.md

@@ -96,8 +96,9 @@ SSD cache or "media-cache" - for example, a lot of Seagate EXOS drives have
 it (they have internal SSD cache even though it's not stated in datasheets).
 
 Setting this parameter to "all" or "small" in OSD parameters requires enabling
-disable_journal_fsync and disable_meta_fsync, setting it to "all" also requires
-enabling disable_data_fsync.
+[disable_journal_fsync](layout-osd.en.yml#disable_journal_fsync) and
+[disable_meta_fsync](layout-osd.en.yml#disable_meta_fsync), setting it to
+"all" also requires enabling [disable_data_fsync](layout-osd.en.yml#disable_data_fsync).
 
 TLDR: For optimal performance, set immediate_commit to "all" if you only use
 SSDs with supercapacitor-based power loss protection (nonvolatile

docs/config/layout-cluster.ru.md

@@ -103,8 +103,9 @@ HDD-дисках с внутренним SSD или "медиа" кэшем - н
 указано в спецификациях).
 
 Указание "all" или "small" в настройках / командной строке OSD требует
-включения disable_journal_fsync и disable_meta_fsync, значение "all" также
-требует включения disable_data_fsync.
+включения [disable_journal_fsync](layout-osd.ru.yml#disable_journal_fsync) и
+[disable_meta_fsync](layout-osd.ru.yml#disable_meta_fsync), значение "all"
+также требует включения [disable_data_fsync](layout-osd.ru.yml#disable_data_fsync).
 
 Итого, вкратце: для оптимальной производительности установите
 immediate_commit в значение "all", если вы используете в кластере только SSD

docs/config/layout-osd.en.md

@@ -197,21 +197,22 @@ Must be equal or a multiple of [bitmap_granularity](layout-cluster.en.md#bitmap_
 
 Checksums increase metadata size by 4 bytes per each csum_block_size of data.
 
-Checksums are always a compromise:
+Checksums are always a tradeoff:
 1. You either sacrifice +1 GB RAM per 1 TB of data
 2. Or you raise csum_block_size, for example, to 32k and sacrifice
    50% random write iops due to checksum read-modify-write
 3. Or you turn off [inmemory_metadata](osd.en.md#inmemory_metadata) and
    sacrifice 50% random read iops due to checksum reads
 
-Option 1 (default) is recommended for all-flash setups because these usually
-have enough RAM.
+All-flash clusters usually have enough RAM to use default csum_block_size,
+which uses 1 GB RAM per 1 TB of data. HDD clusters usually don't.
 
-Option 2 is recommended for HDD-only setups. HDD-only setups usually do NOT
-have enough RAM for the default 4 KB csum_block_size.
+Thus, recommended setups are:
+1. All-flash, 1 GB RAM per 1 TB data: default (csum_block_size=4k)
+2. All-flash, less RAM: csum_block_size=4k + inmemory_metadata=false
+3. Hybrid HDD+SSD: csum_block_size=4k + inmemory_metadata=false
+4. HDD-only, faster random read: csum_block_size=32k
+5. HDD-only, faster random write: csum_block_size=4k +
+   inmemory_metadata=false + meta_io=cached
 
-Option 3 is recommended for SSD+HDD setups (because metadata SSDs will handle
-extra reads without any performance drop) and also *maybe* for NVMe all-flash
-setups when you don't have enough RAM (because NVMe drives have plenty
-of read iops to spare). You may also consider enabling
-[cached_read_meta](osd.en.md#cached_read_meta) in this case.
+See also [meta_io](osd.en.md#meta_io).

docs/config/layout-osd.ru.md

@@ -220,17 +220,12 @@ csum_block_size данных.
    жертвуете 50% скорости случайного чтения из-за чтения контрольных сумм
    с диска
 
-Вариант 1 (при настройках по умолчанию) рекомендуется для SSD (All-Flash)
-кластеров, потому что памяти в них обычно хватает.
+Таким образом, рекомендуются следующие варианты настроек:
+1. All-flash, 1 ГБ памяти на 1 ТБ данных: по умолчанию (csum_block_size=4k)
+2. All-flash, меньше памяти: csum_block_size=4k + inmemory_metadata=false
+3. Гибридные HDD+SSD: csum_block_size=4k + inmemory_metadata=false
+4. Только HDD, быстрее случайное чтение: csum_block_size=32k
+5. Только HDD, быстрее случайная запись: csum_block_size=4k +
+   inmemory_metadata=false + meta_io=cached
 
-Вариант 2 рекомендуется для кластеров на одних жёстких дисках (без SSD
-под метаданные). На 4 кб блок контрольной суммы памяти в таких кластерах
-обычно НЕ хватает.
-
-Вариант 3 рекомендуется для гибридных кластеров (SSD+HDD), потому что
-скорости SSD под метаданными хватит, чтобы обработать дополнительные чтения
-без снижения производительности. Также вариант 3 *может* рекомендоваться
-для All-Flash кластеров на основе NVMe-дисков, когда памяти НЕ достаточно,
-потому что NVMe-диски имеют огромный запас производительности по чтению.
-В таких случаях, возможно, также имеет смысл включать параметр
-[cached_read_meta](osd.ru.md#cached_read_meta).
+Смотрите также [meta_io](osd.ru.md#meta_io).

docs/config/osd.en.md

@@ -31,9 +31,9 @@ them, even without restarting by updating configuration in etcd.
 - [max_flusher_count](#max_flusher_count)
 - [inmemory_metadata](#inmemory_metadata)
 - [inmemory_journal](#inmemory_journal)
-- [cached_read_data](#cached_read_data)
-- [cached_read_meta](#cached_read_meta)
-- [cached_read_journal](#cached_read_journal)
+- [data_io](#data_io)
+- [meta_io](#meta_io)
+- [journal_io](#journal_io)
 - [journal_sector_buffer_count](#journal_sector_buffer_count)
 - [journal_no_same_sector_overwrites](#journal_no_same_sector_overwrites)
 - [throttle_small_writes](#throttle_small_writes)
@@ -258,45 +258,59 @@ is typically very small because it's sufficient to have 16-32 MB journal
 for SSD OSDs. However, in theory it's possible that you'll want to turn it
 off for hybrid (HDD+SSD) OSDs with large journals on quick devices.
 
-## cached_read_data
+## data_io
 
-- Type: boolean
-- Default: false
+- Type: string
+- Default: direct
 
-Read data through Linux page cache, i.e. use a file descriptor opened without
-O_DIRECT for data reads. May improve read performance for frequently accessed
-data if it fits in RAM. Memory in page cache is shared by all processes and
-not accounted in OSD memory consumption.
+I/O mode for *data*. One of "direct", "cached" or "directsync". Corresponds
+to O_DIRECT, O_SYNC and O_DIRECT|O_SYNC, respectively.
 
-## cached_read_meta
+Choose "cached" to use Linux page cache. This may improve read performance
+for hot data and slower disks - HDDs and maybe SATA SSDs - but will slightly
+decrease write performance for fast disks because page cache is an overhead
+itself.
 
-- Type: boolean
-- Default: false
+Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+(which requires disable_data_fsync) with drives having write-back cache
+which can't be turned off, for example, Intel Optane. Also note that *some*
+desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
+disable_data_fsync unsafe even with "directsync".
 
-Read metadata through Linux page cache. May be beneficial when checksums
-are enabled and [inmemory_metadata](#inmemory_metadata) is disabled, because
-in this case metadata blocks are read from disk to verify checksums on every
-read request and caching them may reduce this extra read load.
+## meta_io
 
-Absolutely pointless to enable with enabled inmemory_metadata because all
-metadata is kept in memory anyway, and likely pointless without checksums,
-because in that case, metadata blocks are read from disk only during journal
+- Type: string
+- Default: direct
+
+I/O mode for *metadata*. One of "direct", "cached" or "directsync".
+
+"cached" may improve read performance, but only under the following conditions:
+1. your drives are relatively slow (HDD, SATA SSD), and
+2. checksums are enabled, and
+3. [inmemory_metadata](#inmemory_metadata) is disabled.
+Under all these conditions, metadata blocks are read from disk on every
+read request to verify checksums and caching them may reduce this extra
+read load. Without (3) metadata is never read from the disk after starting,
+and without (2) metadata blocks are read from disk only during journal
 flushing.
 
-If the same device is used for data and metadata, enabling [cached_read_data](#cached_read_data)
-also enables this parameter, given that it isn't turned off explicitly.
+"directsync" is the same as above.
 
-## cached_read_journal
+If the same device is used for data and metadata, meta_io by default is set
+to the same value as [data_io](#data_io).
 
-- Type: boolean
-- Default: false
+## journal_io
 
-Read buffered data from journal through Linux page cache. Does not have sense
-without disabling [inmemory_journal](#inmemory_journal), which, again, is
-enabled by default.
+- Type: string
+- Default: direct
 
-If the same device is used for metadata and journal, enabling [cached_read_meta](#cached_read_meta)
-also enables this parameter, given that it isn't turned off explicitly.
+I/O mode for *journal*. One of "direct", "cached" or "directsync".
+
+Here, "cached" may only improve read performance for recent writes and
+only if [inmemory_journal](#inmemory_journal) is turned off.
+
+If the same device is used for metadata and journal, journal_io by default
+is set to the same value as [meta_io](#meta_io).
 
 ## journal_sector_buffer_count
 

docs/config/osd.ru.md

@@ -32,9 +32,9 @@
 - [max_flusher_count](#max_flusher_count)
 - [inmemory_metadata](#inmemory_metadata)
 - [inmemory_journal](#inmemory_journal)
-- [cached_read_data](#cached_read_data)
-- [cached_read_meta](#cached_read_meta)
-- [cached_read_journal](#cached_read_journal)
+- [data_io](#data_io)
+- [meta_io](#meta_io)
+- [journal_io](#journal_io)
 - [journal_sector_buffer_count](#journal_sector_buffer_count)
 - [journal_no_same_sector_overwrites](#journal_no_same_sector_overwrites)
 - [throttle_small_writes](#throttle_small_writes)
@@ -266,50 +266,62 @@ Flusher - это микро-поток (корутина), которая коп
 параметра может оказаться полезным для гибридных OSD (HDD+SSD) с большими
 журналами, расположенными на быстром по сравнению с HDD устройстве.
 
-## cached_read_data
+## data_io
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+- Тип: строка
+- Значение по умолчанию: direct
 
-Читать данные через системный кэш Linux (page cache), то есть, использовать
-для чтения данных файловый дескриптор, открытый без флага O_DIRECT. Может
-улучшить производительность чтения для часто используемых данных, если они
-помещаются в память. Память кэша разделяется между всеми процессами в
-системе и не учитывается в потреблении памяти процессом OSD.
+Режим ввода-вывода для *данных*. Одно из значений "direct", "cached" или
+"directsync", означающих O_DIRECT, O_SYNC и O_DIRECT|O_SYNC, соответственно.
 
-## cached_read_meta
+Выберите "cached", чтобы использовать системный кэш Linux (page cache) при
+чтении и записи. Это может улучшить скорость чтения горячих данных с
+относительно медленных дисков - HDD и, возможно, SATA SSD - но немного
+снижает производительность записи для быстрых дисков, так как кэш сам по
+себе тоже добавляет накладные расходы.
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+Выберите "directsync", если хотите задействовать
+[immediate_commit](layout-cluster.ru.md#immediate_commit) (требующий
+включенияd disable_data_fsync) на дисках с неотключаемым кэшем. Пример таких
+дисков - Intel Optane. При этом также стоит иметь в виду, что *некоторые*
+настольные SSD (например, HP EX950) игнорируют флаг O_SYNC, делая отключение
+fsync небезопасным даже с режимом "directsync".
 
-Читать метаданные через системный кэш Linux. Может быть полезно, когда
-включены контрольные суммы, а параметр [inmemory_metadata](#inmemory_metadata)
-отключён, так как в этом случае блоки метаданных читаются с диска при каждом
-запросе чтения для проверки контрольных сумм и их кэширование может снизить
-дополнительную нагрузку на диск.
+## meta_io
 
-Абсолютно бессмысленно включать данный параметр, если параметр
-inmemory_metadata включён (по умолчанию это так), и также вероятно
-бессмысленно включать его, если не включены контрольные суммы, так как в
-этом случае блоки метаданных читаются с диска только во время сброса
-журнала.
+- Тип: строка
+- Значение по умолчанию: direct
 
-Если одно и то же устройство используется для данных и метаданных, включение
-[cached_read_data](#cached_read_data) также включает данный параметр, при
-условии, что он не отключён явным образом.
+Режим ввода-вывода для *метаданных*. Одно из значений "direct", "cached" или
+"directsync".
 
-## cached_read_journal
+"cached" может улучшить скорость чтения, если:
+1. у вас медленные диски (HDD, SATA SSD)
+2. контрольные суммы включены
+3. параметр [inmemory_metadata](#inmemory_metadata) отключён.
+При этих условиях блоки метаданных читаются с диска при каждом запросе чтения
+для проверки контрольных сумм и их кэширование может снизить дополнительную
+нагрузку на диск. Без (3) метаданные никогда не читаются с диска после
+запуска OSD, а без (2) блоки метаданных читаются только при сбросе журнала.
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+Если одно и то же устройство используется для данных и метаданных, режим
+ввода-вывода метаданных по умолчанию устанавливается равным [data_io](#data_io).
 
-Читать буферизованные в журнале данные через системный кэш Linux. Не имеет
-смысла без отключения параметра [inmemory_journal](#inmemory_journal),
-который, опять же, по умолчанию включён.
+## journal_io
+
+- Тип: строка
+- Значение по умолчанию: direct
+
+Режим ввода-вывода для *журнала*. Одно из значений "direct", "cached" или
+"directsync".
+
+Здесь "cached" может улучшить скорость чтения только недавно записанных
+данных и только если параметр [inmemory_journal](#inmemory_journal)
+отключён.
 
 Если одно и то же устройство используется для метаданных и журнала,
-включение [cached_read_meta](#cached_read_meta) также включает данный
-параметр, при условии, что он не отключён явным образом.
+режим ввода-вывода журнала по умолчанию устанавливается равным
+[meta_io](#meta_io).
 
 ## journal_sector_buffer_count
 

docs/config/pool.en.md

@@ -205,9 +205,8 @@ This parameter usually doesn't require to be changed.
 - Default: 131072
 
 Block size for this pool. The value from /vitastor/config/global is used when
-unspecified. If your cluster has OSDs with different block sizes then pool must
-be restricted by [osd_tags](#osd_tags) to only include OSDs with matching block
-size.
+unspecified. Only OSDs with matching block_size are used for each pool. If you
+want to further restrict OSDs for the pool, use [osd_tags](#osd_tags).
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#block_size).
 
@@ -216,10 +215,9 @@ Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-c
 - Type: integer
 - Default: 4096
 
-"Sector" size of virtual disks in this pool. The value from
-/vitastor/config/global is used when unspecified. Similar to block_size, the
-pool must be restricted by [osd_tags](#osd_tags) to only include OSDs with
-matching bitmap_granularity.
+"Sector" size of virtual disks in this pool. The value from /vitastor/config/global
+is used when unspecified. Similarly to block_size, only OSDs with matching
+bitmap_granularity are used for each pool.
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#bitmap_granularity).
 
@@ -229,10 +227,11 @@ Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-c
 - Default: none
 
 Immediate commit setting for this pool. The value from /vitastor/config/global
-is used when unspecified. Similar to block_size, the pool must be restricted by
-[osd_tags](#osd_tags) to only include OSDs with compatible immediate_commit.
-Compatible means that a pool with non-immediate commit will work with OSDs with
-immediate commit enabled, but not vice versa.
+is used when unspecified. Similarly to block_size, only OSDs with compatible
+bitmap_granularity are used for each pool. "Compatible" means that a pool with
+non-immediate commit will use OSDs with immediate commit enabled, but not vice
+versa. I.e., pools with "none" use all OSDs, pools with "small" only use OSDs
+with "all" or "small", and pools with "all" only use OSDs with "all".
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#immediate_commit).
 

docs/config/pool.ru.md

@@ -208,8 +208,9 @@ PG в Vitastor эферемерны, то есть вы можете менят
 
 Размер блока для данного пула. Если не задан, используется значение из
 /vitastor/config/global. Если в вашем кластере есть OSD с разными размерами
-блока, пул должен быть ограничен только OSD, блок которых равен блоку пула,
-с помощью [osd_tags](#osd_tags).
+блока, пул будет использовать только OSD с размером блока, равным размеру блока
+пула. Если вы хотите сильнее ограничить набор используемых для пула OSD -
+используйте [osd_tags](#osd_tags).
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#block_size).
 
@@ -219,9 +220,8 @@ PG в Vitastor эферемерны, то есть вы можете менят
 - По умолчанию: 4096
 
 Размер "сектора" виртуальных дисков в данном пуле. Если не задан, используется
-значение из /vitastor/config/global. Аналогично block_size, пул должен быть
-ограничен OSD со значением bitmap_granularity, равным значению пула, с помощью
-[osd_tags](#osd_tags).
+значение из /vitastor/config/global. Аналогично block_size, каждый пул будет
+использовать только OSD с совпадающей с пулом настройкой bitmap_granularity.
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#bitmap_granularity).
 
@@ -231,11 +231,13 @@ PG в Vitastor эферемерны, то есть вы можете менят
 - По умолчанию: none
 
 Настройка мгновенного коммита для данного пула. Если не задана, используется
-значение из /vitastor/config/global. Аналогично block_size, пул должен быть
-ограничен OSD со значением bitmap_granularity, совместимым со значением пула, с
-помощью [osd_tags](#osd_tags). Совместимость означает, что пул с отключенным
-мгновенным коммитом может работать на OSD с включённым мгновенным коммитом, но
-не наоборот.
+значение из /vitastor/config/global. Аналогично block_size, каждый пул будет
+использовать только OSD с *совместимыми* настройками immediate_commit.
+"Совместимыми" означает, что пул с отключенным мгновенным коммитом будет
+использовать OSD с включённым мгновенным коммитом, но не наоборот. То есть,
+пул со значением "none" будет использовать все OSD, пул со "small" будет
+использовать OSD с "all" или "small", а пул с "all" будет использовать только
+OSD с "all".
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#immediate_commit).
 

docs/config/src/layout-cluster.yml

@@ -87,8 +87,9 @@
     it (they have internal SSD cache even though it's not stated in datasheets).
 
     Setting this parameter to "all" or "small" in OSD parameters requires enabling
-    disable_journal_fsync and disable_meta_fsync, setting it to "all" also requires
-    enabling disable_data_fsync.
+    [disable_journal_fsync](layout-osd.en.yml#disable_journal_fsync) and
+    [disable_meta_fsync](layout-osd.en.yml#disable_meta_fsync), setting it to
+    "all" also requires enabling [disable_data_fsync](layout-osd.en.yml#disable_data_fsync).
 
     TLDR: For optimal performance, set immediate_commit to "all" if you only use
     SSDs with supercapacitor-based power loss protection (nonvolatile
@@ -140,8 +141,9 @@
     указано в спецификациях).
 
     Указание "all" или "small" в настройках / командной строке OSD требует
-    включения disable_journal_fsync и disable_meta_fsync, значение "all" также
-    требует включения disable_data_fsync.
+    включения [disable_journal_fsync](layout-osd.ru.yml#disable_journal_fsync) и
+    [disable_meta_fsync](layout-osd.ru.yml#disable_meta_fsync), значение "all"
+    также требует включения [disable_data_fsync](layout-osd.ru.yml#disable_data_fsync).
 
     Итого, вкратце: для оптимальной производительности установите
     immediate_commit в значение "all", если вы используете в кластере только SSD

docs/config/src/layout-osd.yml

@@ -228,24 +228,25 @@
 
     Checksums increase metadata size by 4 bytes per each csum_block_size of data.
 
-    Checksums are always a compromise:
+    Checksums are always a tradeoff:
     1. You either sacrifice +1 GB RAM per 1 TB of data
     2. Or you raise csum_block_size, for example, to 32k and sacrifice
        50% random write iops due to checksum read-modify-write
     3. Or you turn off [inmemory_metadata](osd.en.md#inmemory_metadata) and
        sacrifice 50% random read iops due to checksum reads
 
-    Option 1 (default) is recommended for all-flash setups because these usually
-    have enough RAM.
+    All-flash clusters usually have enough RAM to use default csum_block_size,
+    which uses 1 GB RAM per 1 TB of data. HDD clusters usually don't.
 
-    Option 2 is recommended for HDD-only setups. HDD-only setups usually do NOT
-    have enough RAM for the default 4 KB csum_block_size.
+    Thus, recommended setups are:
+    1. All-flash, 1 GB RAM per 1 TB data: default (csum_block_size=4k)
+    2. All-flash, less RAM: csum_block_size=4k + inmemory_metadata=false
+    3. Hybrid HDD+SSD: csum_block_size=4k + inmemory_metadata=false
+    4. HDD-only, faster random read: csum_block_size=32k
+    5. HDD-only, faster random write: csum_block_size=4k +
+       inmemory_metadata=false + meta_io=cached
 
-    Option 3 is recommended for SSD+HDD setups (because metadata SSDs will handle
-    extra reads without any performance drop) and also *maybe* for NVMe all-flash
-    setups when you don't have enough RAM (because NVMe drives have plenty
-    of read iops to spare). You may also consider enabling
-    [cached_read_meta](osd.en.md#cached_read_meta) in this case.
+    See also [meta_io](osd.en.md#meta_io).
   info_ru: |
     Размер блока расчёта контрольных сумм.
 
@@ -264,17 +265,12 @@
        жертвуете 50% скорости случайного чтения из-за чтения контрольных сумм
        с диска
 
-    Вариант 1 (при настройках по умолчанию) рекомендуется для SSD (All-Flash)
-    кластеров, потому что памяти в них обычно хватает.
+    Таким образом, рекомендуются следующие варианты настроек:
+    1. All-flash, 1 ГБ памяти на 1 ТБ данных: по умолчанию (csum_block_size=4k)
+    2. All-flash, меньше памяти: csum_block_size=4k + inmemory_metadata=false
+    3. Гибридные HDD+SSD: csum_block_size=4k + inmemory_metadata=false
+    4. Только HDD, быстрее случайное чтение: csum_block_size=32k
+    5. Только HDD, быстрее случайная запись: csum_block_size=4k +
+       inmemory_metadata=false + meta_io=cached
 
-    Вариант 2 рекомендуется для кластеров на одних жёстких дисках (без SSD
-    под метаданные). На 4 кб блок контрольной суммы памяти в таких кластерах
-    обычно НЕ хватает.
-
-    Вариант 3 рекомендуется для гибридных кластеров (SSD+HDD), потому что
-    скорости SSD под метаданными хватит, чтобы обработать дополнительные чтения
-    без снижения производительности. Также вариант 3 *может* рекомендоваться
-    для All-Flash кластеров на основе NVMe-дисков, когда памяти НЕ достаточно,
-    потому что NVMe-диски имеют огромный запас производительности по чтению.
-    В таких случаях, возможно, также имеет смысл включать параметр
-    [cached_read_meta](osd.ru.md#cached_read_meta).
+    Смотрите также [meta_io](osd.ru.md#meta_io).

docs/config/src/osd.yml

@@ -260,70 +260,96 @@
     достаточно 16- или 32-мегабайтного журнала. Однако в теории отключение
     параметра может оказаться полезным для гибридных OSD (HDD+SSD) с большими
     журналами, расположенными на быстром по сравнению с HDD устройстве.
-- name: cached_read_data
-  type: bool
-  default: false
+- name: data_io
+  type: string
+  default: direct
   info: |
-    Read data through Linux page cache, i.e. use a file descriptor opened without
-    O_DIRECT for data reads. May improve read performance for frequently accessed
-    data if it fits in RAM. Memory in page cache is shared by all processes and
-    not accounted in OSD memory consumption.
-  info_ru: |
-    Читать данные через системный кэш Linux (page cache), то есть, использовать
-    для чтения данных файловый дескриптор, открытый без флага O_DIRECT. Может
-    улучшить производительность чтения для часто используемых данных, если они
-    помещаются в память. Память кэша разделяется между всеми процессами в
-    системе и не учитывается в потреблении памяти процессом OSD.
-- name: cached_read_meta
-  type: bool
-  default: false
-  info: |
-    Read metadata through Linux page cache. May be beneficial when checksums
-    are enabled and [inmemory_metadata](#inmemory_metadata) is disabled, because
-    in this case metadata blocks are read from disk to verify checksums on every
-    read request and caching them may reduce this extra read load.
+    I/O mode for *data*. One of "direct", "cached" or "directsync". Corresponds
+    to O_DIRECT, O_SYNC and O_DIRECT|O_SYNC, respectively.
 
-    Absolutely pointless to enable with enabled inmemory_metadata because all
-    metadata is kept in memory anyway, and likely pointless without checksums,
-    because in that case, metadata blocks are read from disk only during journal
+    Choose "cached" to use Linux page cache. This may improve read performance
+    for hot data and slower disks - HDDs and maybe SATA SSDs - but will slightly
+    decrease write performance for fast disks because page cache is an overhead
+    itself.
+
+    Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+    (which requires disable_data_fsync) with drives having write-back cache
+    which can't be turned off, for example, Intel Optane. Also note that *some*
+    desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
+    disable_data_fsync unsafe even with "directsync".
+  info_ru: |
+    Режим ввода-вывода для *данных*. Одно из значений "direct", "cached" или
+    "directsync", означающих O_DIRECT, O_SYNC и O_DIRECT|O_SYNC, соответственно.
+
+    Выберите "cached", чтобы использовать системный кэш Linux (page cache) при
+    чтении и записи. Это может улучшить скорость чтения горячих данных с
+    относительно медленных дисков - HDD и, возможно, SATA SSD - но немного
+    снижает производительность записи для быстрых дисков, так как кэш сам по
+    себе тоже добавляет накладные расходы.
+
+    Выберите "directsync", если хотите задействовать
+    [immediate_commit](layout-cluster.ru.md#immediate_commit) (требующий
+    включенияd disable_data_fsync) на дисках с неотключаемым кэшем. Пример таких
+    дисков - Intel Optane. При этом также стоит иметь в виду, что *некоторые*
+    настольные SSD (например, HP EX950) игнорируют флаг O_SYNC, делая отключение
+    fsync небезопасным даже с режимом "directsync".
+- name: meta_io
+  type: string
+  default: direct
+  info: |
+    I/O mode for *metadata*. One of "direct", "cached" or "directsync".
+
+    "cached" may improve read performance, but only under the following conditions:
+    1. your drives are relatively slow (HDD, SATA SSD), and
+    2. checksums are enabled, and
+    3. [inmemory_metadata](#inmemory_metadata) is disabled.
+    Under all these conditions, metadata blocks are read from disk on every
+    read request to verify checksums and caching them may reduce this extra
+    read load. Without (3) metadata is never read from the disk after starting,
+    and without (2) metadata blocks are read from disk only during journal
     flushing.
 
-    If the same device is used for data and metadata, enabling [cached_read_data](#cached_read_data)
-    also enables this parameter, given that it isn't turned off explicitly.
+    "directsync" is the same as above.
+
+    If the same device is used for data and metadata, meta_io by default is set
+    to the same value as [data_io](#data_io).
   info_ru: |
-    Читать метаданные через системный кэш Linux. Может быть полезно, когда
-    включены контрольные суммы, а параметр [inmemory_metadata](#inmemory_metadata)
-    отключён, так как в этом случае блоки метаданных читаются с диска при каждом
-    запросе чтения для проверки контрольных сумм и их кэширование может снизить
-    дополнительную нагрузку на диск.
+    Режим ввода-вывода для *метаданных*. Одно из значений "direct", "cached" или
+    "directsync".
 
-    Абсолютно бессмысленно включать данный параметр, если параметр
-    inmemory_metadata включён (по умолчанию это так), и также вероятно
-    бессмысленно включать его, если не включены контрольные суммы, так как в
-    этом случае блоки метаданных читаются с диска только во время сброса
-    журнала.
+    "cached" может улучшить скорость чтения, если:
+    1. у вас медленные диски (HDD, SATA SSD)
+    2. контрольные суммы включены
+    3. параметр [inmemory_metadata](#inmemory_metadata) отключён.
+    При этих условиях блоки метаданных читаются с диска при каждом запросе чтения
+    для проверки контрольных сумм и их кэширование может снизить дополнительную
+    нагрузку на диск. Без (3) метаданные никогда не читаются с диска после
+    запуска OSD, а без (2) блоки метаданных читаются только при сбросе журнала.
 
-    Если одно и то же устройство используется для данных и метаданных, включение
-    [cached_read_data](#cached_read_data) также включает данный параметр, при
-    условии, что он не отключён явным образом.
-- name: cached_read_journal
-  type: bool
-  default: false
+    Если одно и то же устройство используется для данных и метаданных, режим
+    ввода-вывода метаданных по умолчанию устанавливается равным [data_io](#data_io).
+- name: journal_io
+  type: string
+  default: direct
   info: |
-    Read buffered data from journal through Linux page cache. Does not have sense
-    without disabling [inmemory_journal](#inmemory_journal), which, again, is
-    enabled by default.
+    I/O mode for *journal*. One of "direct", "cached" or "directsync".
 
-    If the same device is used for metadata and journal, enabling [cached_read_meta](#cached_read_meta)
-    also enables this parameter, given that it isn't turned off explicitly.
+    Here, "cached" may only improve read performance for recent writes and
+    only if [inmemory_journal](#inmemory_journal) is turned off.
+
+    If the same device is used for metadata and journal, journal_io by default
+    is set to the same value as [meta_io](#meta_io).
   info_ru: |
-    Читать буферизованные в журнале данные через системный кэш Linux. Не имеет
-    смысла без отключения параметра [inmemory_journal](#inmemory_journal),
-    который, опять же, по умолчанию включён.
+    Режим ввода-вывода для *журнала*. Одно из значений "direct", "cached" или
+    "directsync".
+
+    Здесь "cached" может улучшить скорость чтения только недавно записанных
+    данных и только если параметр [inmemory_journal](#inmemory_journal)
+    отключён.
 
     Если одно и то же устройство используется для метаданных и журнала,
-    включение [cached_read_meta](#cached_read_meta) также включает данный
-    параметр, при условии, что он не отключён явным образом.
+    режим ввода-вывода журнала по умолчанию устанавливается равным
+    [meta_io](#meta_io).
 - name: journal_sector_buffer_count
   type: int
   default: 32

docs/installation/packages.en.md

@@ -14,6 +14,8 @@
   - Debian 12 (Bookworm/Sid): `deb https://vitastor.io/debian bookworm main`
   - Debian 11 (Bullseye): `deb https://vitastor.io/debian bullseye main`
   - Debian 10 (Buster): `deb https://vitastor.io/debian buster main`
+  - Add `-oldstable` to bookworm/bullseye/buster in this line to install the last
+    stable version from 0.9.x branch instead of 1.x
 - For Debian 10 (Buster) also enable backports repository:
   `deb http://deb.debian.org/debian buster-backports main`
 - Install packages: `apt update; apt install vitastor lp-solve etcd linux-image-amd64 qemu`

docs/installation/packages.ru.md

@@ -14,6 +14,8 @@
   - Debian 12 (Bookworm/Sid): `deb https://vitastor.io/debian bookworm main`
   - Debian 11 (Bullseye): `deb https://vitastor.io/debian bullseye main`
   - Debian 10 (Buster): `deb https://vitastor.io/debian buster main`
+  - Добавьте `-oldstable` к слову bookworm/bullseye/buster в этой строке, чтобы
+    установить последнюю стабильную версию из ветки 0.9.x вместо 1.x
 - Для Debian 10 (Buster) также включите репозиторий backports:
   `deb http://deb.debian.org/debian buster-backports main`
 - Установите пакеты: `apt update; apt install vitastor lp-solve etcd linux-image-amd64 qemu`

docs/installation/source.en.md

@@ -21,7 +21,7 @@
 
 ## Basic instructions
 
-Download source, for example using git: `git clone --recurse-submodules https://yourcmc.ru/git/vitalif/vitastor/`
+Download source, for example using git: `git clone --recurse-submodules https://git.yourcmc.ru/vitalif/vitastor/`
 
 Get `fio` source and symlink it into `<vitastor>/fio`. If you don't want to build fio engine,
 you can disable it by passing `-DWITH_FIO=no` to cmake.
@@ -41,7 +41,7 @@ It's recommended to build the QEMU driver (qemu_driver.c) in-tree, as a part of
 QEMU build process. To do that:
 - Install vitastor client library headers (from source or from vitastor-client-dev package)
 - Take a corresponding patch from `patches/qemu-*-vitastor.patch` and apply it to QEMU source
-- Copy `src/qemu_driver.c` to QEMU source directory as `block/block-vitastor.c`
+- Copy `src/qemu_driver.c` to QEMU source directory as `block/vitastor.c`
 - Build QEMU as usual
 
 But it is also possible to build it out-of-tree. To do that:

docs/installation/source.ru.md

@@ -21,7 +21,7 @@
 
 ## Базовая инструкция
 
-Скачайте исходные коды, например, из git: `git clone --recurse-submodules https://yourcmc.ru/git/vitalif/vitastor/`
+Скачайте исходные коды, например, из git: `git clone --recurse-submodules https://git.yourcmc.ru/vitalif/vitastor/`
 
 Скачайте исходные коды пакета `fio`, распакуйте их и создайте символическую ссылку на них
 в директории исходников Vitastor: `<vitastor>/fio`. Либо, если вы не хотите собирать плагин fio,
@@ -41,7 +41,7 @@ cmake .. && make -j8 install
 Драйвер QEMU (qemu_driver.c) рекомендуется собирать вместе с самим QEMU. Для этого:
 - Установите заголовки клиентской библиотеки Vitastor (из исходников или из пакета vitastor-client-dev)
 - Возьмите соответствующий патч из `patches/qemu-*-vitastor.patch` и примените его к исходникам QEMU
-- Скопируйте [src/qemu_driver.c](../../src/qemu_driver.c) в директорию исходников QEMU как `block/block-vitastor.c`
+- Скопируйте [src/qemu_driver.c](../../src/qemu_driver.c) в директорию исходников QEMU как `block/vitastor.c`
 - Соберите QEMU как обычно
 
 Однако в целях отладки драйвер также можно собирать отдельно от QEMU. Для этого:
@@ -60,7 +60,7 @@ cmake .. && make -j8 install
       * Для QEMU 2.0+: `<qemu>/qapi-types.h` &rarr; `<vitastor>/qemu/b/qemu/qapi-types.h`
    - `config-host.h` и `qapi` нужны, т.к. в них содержатся автогенерируемые заголовки
 - Сконфигурируйте cmake Vitastor с `WITH_QEMU=yes` (`cmake .. -DWITH_QEMU=yes`) и, если вы
-  используете RHEL-подобый дистрибутив, также с `QEMU_PLUGINDIR=qemu-kvm`.
+  используете RHEL-подобный дистрибутив, также с `QEMU_PLUGINDIR=qemu-kvm`.
 - После этого в процессе сборки Vitastor также будет собираться подходящий для вашей
   версии QEMU `block-vitastor.so`.
 - Таким образом можно использовать драйвер даже с немодифицированным QEMU, но в этом случае

docs/intro/features.en.md

@@ -29,7 +29,7 @@
 - Snapshots and copy-on-write image clones
 - [Write throttling to smooth random write workloads in SSD+HDD configurations](../config/osd.en.md#throttle_small_writes)
 - [RDMA/RoCEv2 support via libibverbs](../config/network.en.md#rdma_device)
-- [Scrubbing without checksums](../config/osd.en.md#auto_scrub) (verification of copies)
+- [Scrubbing](../config/osd.en.md#auto_scrub) (verification of copies)
 - [Checksums](../config/layout-osd.en.md#data_csum_type)
 
 ## Plugins and tools

docs/intro/features.ru.md

@@ -31,7 +31,7 @@
 - Снапшоты и copy-on-write клоны
 - [Сглаживание производительности случайной записи в SSD+HDD конфигурациях](../config/osd.ru.md#throttle_small_writes)
 - [Поддержка RDMA/RoCEv2 через libibverbs](../config/network.ru.md#rdma_device)
-- [Фоновая проверка целостности без контрольных сумм](../config/osd.ru.md#auto_scrub) (сверка копий)
+- [Фоновая проверка целостности](../config/osd.ru.md#auto_scrub) (сверка копий)
 - [Контрольные суммы](../config/layout-osd.ru.md#data_csum_type)
 
 ## Драйверы и инструменты

docs/usage/disk.en.md

@@ -102,7 +102,7 @@ checks the device cache status on start and tries to disable cache for SATA/SAS
 If it doesn't succeed it issues a warning in the system log.
 
 You can also pass other OSD options here as arguments and they'll be persisted
-in the superblock: cached_read_data, cached_read_meta, cached_read_journal,
+in the superblock: cached_io_data, cached_io_meta, cached_io_journal,
 inmemory_metadata, inmemory_journal, max_write_iodepth,
 min_flusher_count, max_flusher_count, journal_sector_buffer_count,
 journal_no_same_sector_overwrites, throttle_small_writes, throttle_target_iops,

docs/usage/disk.ru.md

@@ -103,8 +103,8 @@ vitastor-disk - инструмент командной строки для уп
 это не удаётся, в системный журнал выводится предупреждение.
 
 Вы можете передать данной команде и некоторые другие опции OSD в качестве аргументов
-и они тоже будут сохранены в суперблок: cached_read_data, cached_read_meta,
-cached_read_journal, inmemory_metadata, inmemory_journal, max_write_iodepth,
+и они тоже будут сохранены в суперблок: cached_io_data, cached_io_meta,
+cached_io_journal, inmemory_metadata, inmemory_journal, max_write_iodepth,
 min_flusher_count, max_flusher_count, journal_sector_buffer_count,
 journal_no_same_sector_overwrites, throttle_small_writes, throttle_target_iops,
 throttle_target_mbs, throttle_target_parallelism, throttle_threshold_us.

docs/usage/qemu.en.md

@@ -107,7 +107,8 @@ disabled by now, so if you want to try it on Debian, use a kernel from Ubuntu
 Commands to attach Vitastor image as a VDUSE device:
 
 ```
-modprobe vduse virtio-vdpa
+modprobe vduse
+modprobe virtio-vdpa
 qemu-storage-daemon --daemonize --blockdev '{"node-name":"test1","driver":"vitastor",\
   "etcd-host":"192.168.7.2:2379/v3","image":"testosd1","cache":{"direct":true,"no-flush":false},"discard":"unmap"}' \
   --export vduse-blk,id=test1,node-name=test1,name=test1,num-queues=16,queue-size=128,writable=true

docs/usage/qemu.ru.md

@@ -111,7 +111,8 @@ VDUSE (CONFIG_VIRTIO_VDPA=m и CONFIG_VDPA_USER=m). В ядрах в Debian Linu
 Команды для подключения виртуального диска через VDUSE:
 
 ```
-modprobe vduse virtio-vdpa
+modprobe vduse
+modprobe virtio-vdpa
 qemu-storage-daemon --daemonize --blockdev '{"node-name":"test1","driver":"vitastor",\
   "etcd-host":"192.168.7.2:2379/v3","image":"testosd1","cache":{"direct":true,"no-flush":false},"discard":"unmap"}' \
   --export vduse-blk,id=test1,node-name=test1,name=test1,num-queues=16,queue-size=128,writable=true

mon/mon.js

@@ -78,9 +78,15 @@ const etcd_tree = {
             disk_alignment: 4096,
             bitmap_granularity: 4096,
             immediate_commit: false, // 'all' or 'small'
+            // client - configurable online
+            client_max_dirty_bytes: 33554432,
+            client_max_dirty_ops: 1024,
+            client_enable_writeback: false,
+            client_max_buffered_bytes: 33554432,
+            client_max_buffered_ops: 1024,
+            client_max_writeback_iodepth: 256,
             // client and osd - configurable online
             log_level: 0,
-            client_dirty_limit: 33554432,
             peer_connect_interval: 5, // seconds. min: 1
             peer_connect_timeout: 5, // seconds. min: 1
             osd_idle_timeout: 5, // seconds. min: 1
@@ -539,10 +545,18 @@ class Mon
         {
             retries = 1;
         }
+        const tried = {};
         while (retries < 0 || retry < retries)
         {
             const cur_addr = this.pick_next_etcd();
             const base = 'ws'+cur_addr.substr(4);
+            let now = Date.now();
+            if (tried[base] && now-tried[base] < timeout)
+            {
+                await new Promise(ok => setTimeout(ok, timeout-(now-tried[base])));
+                now = Date.now();
+            }
+            tried[base] = now;
             const ok = await new Promise((ok, no) =>
             {
                 const timer_id = setTimeout(() =>
@@ -1148,6 +1162,33 @@ class Mon
         }
     }
 
+    filter_osds_by_block_layout(flat_tree, block_size, bitmap_granularity, immediate_commit)
+    {
+        for (const host in flat_tree)
+        {
+            let found = 0;
+            for (const osd in flat_tree[host])
+            {
+                const osd_stat = this.state.osd.stats[osd];
+                if (osd_stat && (osd_stat.bs_block_size && osd_stat.bs_block_size != block_size ||
+                    osd_stat.bitmap_granularity && osd_stat.bitmap_granularity != bitmap_granularity ||
+                    osd_stat.immediate_commit == 'small' && immediate_commit == 'all' ||
+                    osd_stat.immediate_commit == 'none' && immediate_commit != 'none'))
+                {
+                    delete flat_tree[host][osd];
+                }
+                else
+                {
+                    found++;
+                }
+            }
+            if (!found)
+            {
+                delete flat_tree[host];
+            }
+        }
+    }
+
     get_affinity_osds(pool_cfg, up_osds, osd_tree)
     {
         let aff_osds = up_osds;
@@ -1208,6 +1249,12 @@ class Mon
                 pool_tree = pool_tree ? pool_tree.children : [];
                 pool_tree = LPOptimizer.flatten_tree(pool_tree, levels, pool_cfg.failure_domain, 'osd');
                 this.filter_osds_by_tags(osd_tree, pool_tree, pool_cfg.osd_tags);
+                this.filter_osds_by_block_layout(
+                    pool_tree,
+                    pool_cfg.block_size || this.config.block_size || 131072,
+                    pool_cfg.bitmap_granularity || this.config.bitmap_granularity || 4096,
+                    pool_cfg.immediate_commit || this.config.immediate_commit || 'none'
+                );
                 // These are for the purpose of building history.osd_sets
                 const real_prev_pgs = [];
                 let pg_history = [];
@@ -1497,10 +1544,14 @@ class Mon
                     break;
                 }
             }
+            const pool_cfg = (this.state.config.pools[pool_id]||{});
             if (!object_size)
             {
-                object_size = (this.state.config.pools[pool_id]||{}).block_size ||
-                    this.config.block_size || 131072;
+                object_size = pool_cfg.block_size || this.config.block_size || 131072;
+            }
+            if (pool_cfg.scheme !== 'replicated')
+            {
+                object_size *= ((pool_cfg.pg_size||0) - (pool_cfg.parity_chunks||0));
             }
             object_size = BigInt(object_size);
             for (const pg_num in this.state.pg.stats[pool_id])
@@ -1784,10 +1835,18 @@ class Mon
         {
             retries = 1;
         }
+        const tried = {};
         while (retries < 0 || retry < retries)
         {
             retry++;
             const base = this.pick_next_etcd();
+            let now = Date.now();
+            if (tried[base] && now-tried[base] < timeout)
+            {
+                await new Promise(ok => setTimeout(ok, timeout-(now-tried[base])));
+                now = Date.now();
+            }
+            tried[base] = now;
             const res = await POST(base+path, body, timeout);
             if (res.error)
             {

patches/cinder-vitastor.py

@@ -50,7 +50,7 @@ from cinder.volume import configuration
 from cinder.volume import driver
 from cinder.volume import volume_utils
 
-VERSION = '0.9.3'
+VERSION = '1.0.0'
 
 LOG = logging.getLogger(__name__)
 

patches/qemu-2.12-vitastor.patch

@@ -0,0 +1,176 @@
+diff --git a/block/Makefile.objs b/block/Makefile.objs
+index d644bac60a..e404236291 100644
+--- a/block/Makefile.objs
++++ b/block/Makefile.objs
+@@ -19,6 +19,7 @@ block-obj-$(if $(CONFIG_LIBISCSI),y,n) += iscsi-opts.o
+ block-obj-$(CONFIG_LIBNFS) += nfs.o
+ block-obj-$(CONFIG_CURL) += curl.o
+ block-obj-$(CONFIG_RBD) += rbd.o
++block-obj-$(CONFIG_VITASTOR) += vitastor.o
+ block-obj-$(CONFIG_GLUSTERFS) += gluster.o
+ block-obj-$(CONFIG_VXHS) += vxhs.o
+ block-obj-$(CONFIG_LIBSSH2) += ssh.o
+@@ -39,6 +40,8 @@ curl.o-cflags      := $(CURL_CFLAGS)
+ curl.o-libs        := $(CURL_LIBS)
+ rbd.o-cflags       := $(RBD_CFLAGS)
+ rbd.o-libs         := $(RBD_LIBS)
++vitastor.o-cflags  := $(VITASTOR_CFLAGS)
++vitastor.o-libs    := $(VITASTOR_LIBS)
+ gluster.o-cflags   := $(GLUSTERFS_CFLAGS)
+ gluster.o-libs     := $(GLUSTERFS_LIBS)
+ vxhs.o-libs        := $(VXHS_LIBS)
+diff --git a/configure b/configure
+index 0a19b033bc..58b7fbf24c 100755
+--- a/configure
++++ b/configure
+@@ -398,6 +398,7 @@ trace_backends="log"
+ trace_file="trace"
+ spice=""
+ rbd=""
++vitastor=""
+ smartcard=""
+ libusb=""
+ usb_redir=""
+@@ -1213,6 +1214,10 @@ for opt do
+   ;;
+   --enable-rbd) rbd="yes"
+   ;;
++  --disable-vitastor) vitastor="no"
++  ;;
++  --enable-vitastor) vitastor="yes"
++  ;;
+   --disable-xfsctl) xfs="no"
+   ;;
+   --enable-xfsctl) xfs="yes"
+@@ -1601,6 +1606,7 @@ disabled with --disable-FEATURE, default is enabled if available:
+   vhost-crypto    vhost-crypto acceleration support
+   spice           spice
+   rbd             rados block device (rbd)
++  vitastor        vitastor block device
+   libiscsi        iscsi support
+   libnfs          nfs support
+   smartcard       smartcard support (libcacard)
+@@ -3594,6 +3600,27 @@ EOF
+   fi
+ fi
+ 
++##########################################
++# vitastor probe
++if test "$vitastor" != "no" ; then
++  cat > $TMPC <<EOF
++#include <vitastor_c.h>
++int main(void) {
++  vitastor_c_create_qemu(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
++  return 0;
++}
++EOF
++  vitastor_libs="-lvitastor_client"
++  if compile_prog "" "$vitastor_libs" ; then
++    vitastor=yes
++  else
++    if test "$vitastor" = "yes" ; then
++      feature_not_found "vitastor block device" "Install vitastor-client-dev"
++    fi
++    vitastor=no
++  fi
++fi
++
+ ##########################################
+ # libssh2 probe
+ min_libssh2_version=1.2.8
+@@ -5837,6 +5864,7 @@ echo "Trace output file $trace_file-<pid>"
+ fi
+ echo "spice support     $spice $(echo_version $spice $spice_protocol_version/$spice_server_version)"
+ echo "rbd support       $rbd"
++echo "vitastor support  $vitastor"
+ echo "xfsctl support    $xfs"
+ echo "smartcard support $smartcard"
+ echo "libusb            $libusb"
+@@ -6416,6 +6444,11 @@ if test "$rbd" = "yes" ; then
+   echo "RBD_CFLAGS=$rbd_cflags" >> $config_host_mak
+   echo "RBD_LIBS=$rbd_libs" >> $config_host_mak
+ fi
++if test "$vitastor" = "yes" ; then
++  echo "CONFIG_VITASTOR=m" >> $config_host_mak
++  echo "VITASTOR_CFLAGS=$vitastor_cflags" >> $config_host_mak
++  echo "VITASTOR_LIBS=$vitastor_libs" >> $config_host_mak
++fi
+ 
+ echo "CONFIG_COROUTINE_BACKEND=$coroutine" >> $config_host_mak
+ if test "$coroutine_pool" = "yes" ; then
+diff --git a/qapi/block-core.json b/qapi/block-core.json
+index c50517bff3..c780bb2c1c 100644
+--- a/qapi/block-core.json
++++ b/qapi/block-core.json
+@@ -2514,7 +2514,7 @@
+             'dmg', 'file', 'ftp', 'ftps', 'gluster', 'host_cdrom',
+             'host_device', 'http', 'https', 'iscsi', 'luks', 'nbd', 'nfs',
+             'null-aio', 'null-co', 'nvme', 'parallels', 'qcow', 'qcow2', 'qed',
+-            'quorum', 'raw', 'rbd', 'replication', 'sheepdog', 'ssh',
++            'quorum', 'raw', 'rbd', 'vitastor', 'replication', 'sheepdog', 'ssh',
+             'throttle', 'vdi', 'vhdx', 'vmdk', 'vpc', 'vvfat', 'vxhs' ] }
+ 
+ ##
+@@ -3217,6 +3217,28 @@
+             '*snap-id': 'uint32',
+             '*tag': 'str' } }
+ 
++##
++# @BlockdevOptionsVitastor:
++#
++# Driver specific block device options for vitastor
++#
++# @image:       Image name
++# @inode:       Inode number
++# @pool:        Pool ID
++# @size:        Desired image size in bytes
++# @config-path: Path to Vitastor configuration
++# @etcd-host:   etcd connection address(es)
++# @etcd-prefix: etcd key/value prefix
++##
++{ 'struct': 'BlockdevOptionsVitastor',
++  'data': { '*inode': 'uint64',
++            '*pool': 'uint64',
++            '*size': 'uint64',
++            '*image': 'str',
++            '*config-path': 'str',
++            '*etcd-host': 'str',
++            '*etcd-prefix': 'str' } }
++
+ ##
+ # @ReplicationMode:
+ #
+@@ -3547,6 +3569,7 @@
+       'rbd':        'BlockdevOptionsRbd',
+       'replication':'BlockdevOptionsReplication',
+       'sheepdog':   'BlockdevOptionsSheepdog',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'ssh':        'BlockdevOptionsSsh',
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+@@ -3991,6 +4014,17 @@
+             '*subformat':           'BlockdevVhdxSubformat',
+             '*block-state-zero':    'bool' } }
+ 
++##
++# @BlockdevCreateOptionsVitastor:
++#
++# Driver specific image creation options for Vitastor.
++#
++# @size: Size of the virtual disk in bytes
++##
++{ 'struct': 'BlockdevCreateOptionsVitastor',
++  'data': { 'location':         'BlockdevOptionsVitastor',
++            'size':             'size' } }
++
+ ##
+ # @BlockdevVpcSubformat:
+ #
+@@ -4074,6 +4108,7 @@
+       'rbd':            'BlockdevCreateOptionsRbd',
+       'replication':    'BlockdevCreateNotSupported',
+       'sheepdog':       'BlockdevCreateOptionsSheepdog',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'throttle':       'BlockdevCreateNotSupported',
+       'vdi':            'BlockdevCreateOptionsVdi',

patches/qemu-5.2-vitastor.patch

@@ -0,0 +1,181 @@
+Index: qemu-5.2+dfsg/qapi/block-core.json
+===================================================================
+--- qemu-5.2+dfsg.orig/qapi/block-core.json
++++ qemu-5.2+dfsg/qapi/block-core.json
+@@ -2831,7 +2831,7 @@
+             'luks', 'nbd', 'nfs', 'null-aio', 'null-co', 'nvme', 'parallels',
+             'qcow', 'qcow2', 'qed', 'quorum', 'raw', 'rbd',
+             { 'name': 'replication', 'if': 'defined(CONFIG_REPLICATION)' },
+-            'sheepdog',
++            'sheepdog', 'vitastor',
+             'ssh', 'throttle', 'vdi', 'vhdx', 'vmdk', 'vpc', 'vvfat' ] }
+ 
+ ##
+@@ -3668,6 +3668,28 @@
+             '*tag': 'str' } }
+ 
+ ##
++# @BlockdevOptionsVitastor:
++#
++# Driver specific block device options for vitastor
++#
++# @image:       Image name
++# @inode:       Inode number
++# @pool:        Pool ID
++# @size:        Desired image size in bytes
++# @config-path: Path to Vitastor configuration
++# @etcd-host:   etcd connection address(es)
++# @etcd-prefix: etcd key/value prefix
++##
++{ 'struct': 'BlockdevOptionsVitastor',
++  'data': { '*inode': 'uint64',
++            '*pool': 'uint64',
++            '*size': 'uint64',
++            '*image': 'str',
++            '*config-path': 'str',
++            '*etcd-host': 'str',
++            '*etcd-prefix': 'str' } }
++
++##
+ # @ReplicationMode:
+ #
+ # An enumeration of replication modes.
+@@ -4015,6 +4037,7 @@
+       'replication': { 'type': 'BlockdevOptionsReplication',
+                        'if': 'defined(CONFIG_REPLICATION)' },
+       'sheepdog':   'BlockdevOptionsSheepdog',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'ssh':        'BlockdevOptionsSsh',
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+@@ -4404,6 +4427,17 @@
+             '*cluster-size' :   'size' } }
+ 
+ ##
++# @BlockdevCreateOptionsVitastor:
++#
++# Driver specific image creation options for Vitastor.
++#
++# @size: Size of the virtual disk in bytes
++##
++{ 'struct': 'BlockdevCreateOptionsVitastor',
++  'data': { 'location':         'BlockdevOptionsVitastor',
++            'size':             'size' } }
++
++##
+ # @BlockdevVmdkSubformat:
+ #
+ # Subformat options for VMDK images
+@@ -4665,6 +4699,7 @@
+       'qed':            'BlockdevCreateOptionsQed',
+       'rbd':            'BlockdevCreateOptionsRbd',
+       'sheepdog':       'BlockdevCreateOptionsSheepdog',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'vdi':            'BlockdevCreateOptionsVdi',
+       'vhdx':           'BlockdevCreateOptionsVhdx',
+Index: qemu-5.2+dfsg/block/meson.build
+===================================================================
+--- qemu-5.2+dfsg.orig/block/meson.build
++++ qemu-5.2+dfsg/block/meson.build
+@@ -76,6 +76,7 @@ foreach m : [
+   ['CONFIG_LIBNFS', 'nfs', libnfs, 'nfs.c'],
+   ['CONFIG_LIBSSH', 'ssh', libssh, 'ssh.c'],
+   ['CONFIG_RBD', 'rbd', rbd, 'rbd.c'],
++  ['CONFIG_VITASTOR', 'vitastor', vitastor, 'vitastor.c'],
+ ]
+   if config_host.has_key(m[0])
+     if enable_modules
+Index: qemu-5.2+dfsg/configure
+===================================================================
+--- qemu-5.2+dfsg.orig/configure
++++ qemu-5.2+dfsg/configure
+@@ -372,6 +372,7 @@ trace_backends="log"
+ trace_file="trace"
+ spice=""
+ rbd=""
++vitastor=""
+ smartcard=""
+ u2f="auto"
+ libusb=""
+@@ -1263,6 +1264,10 @@ for opt do
+   ;;
+   --enable-rbd) rbd="yes"
+   ;;
++  --disable-vitastor) vitastor="no"
++  ;;
++  --enable-vitastor) vitastor="yes"
++  ;;
+   --disable-xfsctl) xfs="no"
+   ;;
+   --enable-xfsctl) xfs="yes"
+@@ -1827,6 +1832,7 @@ disabled with --disable-FEATURE, default
+   vhost-vdpa      vhost-vdpa kernel backend support
+   spice           spice
+   rbd             rados block device (rbd)
++  vitastor        vitastor block device
+   libiscsi        iscsi support
+   libnfs          nfs support
+   smartcard       smartcard support (libcacard)
+@@ -3719,6 +3725,27 @@ EOF
+ fi
+ 
+ ##########################################
++# vitastor probe
++if test "$vitastor" != "no" ; then
++  cat > $TMPC <<EOF
++#include <vitastor_c.h>
++int main(void) {
++  vitastor_c_create_qemu(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
++  return 0;
++}
++EOF
++  vitastor_libs="-lvitastor_client"
++  if compile_prog "" "$vitastor_libs" ; then
++    vitastor=yes
++  else
++    if test "$vitastor" = "yes" ; then
++      feature_not_found "vitastor block device" "Install vitastor-client-dev"
++    fi
++    vitastor=no
++  fi
++fi
++
++##########################################
+ # libssh probe
+ if test "$libssh" != "no" ; then
+   if $pkg_config --exists libssh; then
+@@ -6456,6 +6483,10 @@ if test "$rbd" = "yes" ; then
+   echo "CONFIG_RBD=y" >> $config_host_mak
+   echo "RBD_LIBS=$rbd_libs" >> $config_host_mak
+ fi
++if test "$vitastor" = "yes" ; then
++  echo "CONFIG_VITASTOR=y" >> $config_host_mak
++  echo "VITASTOR_LIBS=$vitastor_libs" >> $config_host_mak
++fi
+ 
+ echo "CONFIG_COROUTINE_BACKEND=$coroutine" >> $config_host_mak
+ if test "$coroutine_pool" = "yes" ; then
+Index: qemu-5.2+dfsg/meson.build
+===================================================================
+--- qemu-5.2+dfsg.orig/meson.build
++++ qemu-5.2+dfsg/meson.build
+@@ -596,6 +596,10 @@ rbd = not_found
+ if 'CONFIG_RBD' in config_host
+   rbd = declare_dependency(link_args: config_host['RBD_LIBS'].split())
+ endif
++vitastor = not_found
++if 'CONFIG_VITASTOR' in config_host
++  vitastor = declare_dependency(link_args: config_host['VITASTOR_LIBS'].split())
++endif
+ glusterfs = not_found
+ if 'CONFIG_GLUSTERFS' in config_host
+   glusterfs = declare_dependency(compile_args: config_host['GLUSTERFS_CFLAGS'].split(),
+@@ -2145,6 +2149,7 @@ endif
+ # TODO: add back protocol and server version
+ summary_info += {'spice support':     config_host.has_key('CONFIG_SPICE')}
+ summary_info += {'rbd support':       config_host.has_key('CONFIG_RBD')}
++summary_info += {'vitastor support':  config_host.has_key('CONFIG_VITASTOR')}
+ summary_info += {'xfsctl support':    config_host.has_key('CONFIG_XFS')}
+ summary_info += {'smartcard support': config_host.has_key('CONFIG_SMARTCARD')}
+ summary_info += {'U2F support':       u2f.found()}

rpm/build-tarball.sh

@@ -24,4 +24,4 @@ rm fio
 mv fio-copy fio
 FIO=`rpm -qi fio | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Version[\s:]+(\S+)/ && print $1; /^Release[\s:]+(\S+)/ && print "-$1"; }'`
 perl -i -pe 's/(Requires:\s*fio)([^\n]+)?/$1 = '$FIO'/' $VITASTOR/rpm/vitastor-el$EL.spec
-tar --transform 's#^#vitastor-0.9.3/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.9.3$(rpm --eval '%dist').tar.gz *
+tar --transform 's#^#vitastor-1.0.0/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-1.0.0$(rpm --eval '%dist').tar.gz *

rpm/qemu-kvm-4.2-el7.spec.patch

@@ -22,7 +22,7 @@
  Name: qemu-kvm
  Version: 4.2.0
 -Release: 29.vitastor%{?dist}.6
-+Release: 32.vitastor%{?dist}.6
++Release: 34.vitastor%{?dist}.6
  # Epoch because we pushed a qemu-1.0 package. AIUI this can't ever be dropped
  Epoch: 15
  License: GPLv2 and GPLv2+ and CC-BY

rpm/qemu-kvm-4.2-el8.spec.patch

@@ -13,7 +13,7 @@
  Name: qemu-kvm
  Version: 4.2.0
 -Release: 29%{?dist}.6
-+Release: 32.vitastor%{?dist}.6
++Release: 33.vitastor%{?dist}.6
  # Epoch because we pushed a qemu-1.0 package. AIUI this can't ever be dropped
  Epoch: 15
  License: GPLv2 and GPLv2+ and CC-BY

rpm/qemu-kvm-6.2-el8.spec.patch

@@ -0,0 +1,103 @@
+--- qemu-kvm-6.2.spec.orig	2023-07-18 13:52:57.636625440 +0000
++++ qemu-kvm-6.2.spec	2023-07-18 13:52:19.011683886 +0000
+@@ -73,6 +73,7 @@ Requires: %{name}-hw-usbredir = %{epoch}
+ %endif                                                           \
+ Requires: %{name}-block-iscsi = %{epoch}:%{version}-%{release}   \
+ Requires: %{name}-block-rbd = %{epoch}:%{version}-%{release}     \
++Requires: %{name}-block-vitastor = %{epoch}:%{version}-%{release}\
+ Requires: %{name}-block-ssh = %{epoch}:%{version}-%{release}
+ 
+ # Macro to properly setup RHEL/RHEV conflict handling
+@@ -83,7 +84,7 @@ Obsoletes: %1-rhev <= %{epoch}:%{version
+ Summary: QEMU is a machine emulator and virtualizer
+ Name: qemu-kvm
+ Version: 6.2.0
+-Release: 32%{?rcrel}%{?dist}
++Release: 32.vitastor%{?rcrel}%{?dist}
+ # Epoch because we pushed a qemu-1.0 package. AIUI this can't ever be dropped
+ Epoch: 15
+ License: GPLv2 and GPLv2+ and CC-BY
+@@ -122,6 +123,7 @@ Source37: tests_data_acpi_pc_SSDT.dimmpx
+ Source38: tests_data_acpi_q35_FACP.slic
+ Source39: tests_data_acpi_q35_SSDT.dimmpxm
+ Source40: tests_data_acpi_virt_SSDT.memhp
++Source41: qemu-vitastor.c
+ 
+ Patch0001: 0001-redhat-Adding-slirp-to-the-exploded-tree.patch
+ Patch0005: 0005-Initial-redhat-build.patch
+@@ -652,6 +654,7 @@ Patch255: kvm-scsi-protect-req-aiocb-wit
+ Patch256: kvm-dma-helpers-prevent-dma_blk_cb-vs-dma_aio_cancel-rac.patch
+ # For bz#2090990 - qemu crash with error scsi_req_unref(SCSIRequest *): Assertion `req->refcount > 0' failed or scsi_dma_complete(void *, int): Assertion `r->req.aiocb != NULL' failed [8.7.0]
+ Patch257: kvm-virtio-scsi-reset-SCSI-devices-from-main-loop-thread.patch
++Patch258: qemu-6.2-vitastor.patch
+ 
+ BuildRequires: wget
+ BuildRequires: rpm-build
+@@ -689,6 +692,7 @@ BuildRequires: libcurl-devel
+ BuildRequires: libssh-devel
+ BuildRequires: librados-devel
+ BuildRequires: librbd-devel
++BuildRequires: vitastor-client-devel
+ %if %{have_gluster}
+ # For gluster block driver
+ BuildRequires: glusterfs-api-devel
+@@ -926,6 +930,14 @@ Install this package if you want to acce
+ using the rbd protocol.
+ 
+ 
++%package  block-vitastor
++Summary: QEMU Vitastor block driver
++Requires: %{name}-common%{?_isa} = %{epoch}:%{version}-%{release}
++
++%description block-vitastor
++This package provides the additional Vitastor block driver for QEMU.
++
++
+ %package  block-ssh
+ Summary: QEMU SSH block driver
+ Requires: %{name}-common%{?_isa} = %{epoch}:%{version}-%{release}
+@@ -979,6 +991,7 @@ This package provides usbredir support.
+ rm -fr slirp
+ mkdir slirp
+ %autopatch -p1
++cp %{SOURCE41} ./block/vitastor.c
+ 
+ %global qemu_kvm_build qemu_kvm_build
+ mkdir -p %{qemu_kvm_build}
+@@ -994,7 +1007,7 @@ cp -f %{SOURCE40} tests/data/acpi/virt/S
+ # --build-id option is used for giving info to the debug packages.
+ buildldflags="VL_LDFLAGS=-Wl,--build-id"
+ 
+-%global block_drivers_list qcow2,raw,file,host_device,nbd,iscsi,rbd,blkdebug,luks,null-co,nvme,copy-on-read,throttle
++%global block_drivers_list qcow2,raw,file,host_device,nbd,iscsi,rbd,vitastor,blkdebug,luks,null-co,nvme,copy-on-read,throttle
+ 
+ %if 0%{have_gluster}
+     %global block_drivers_list %{block_drivers_list},gluster
+@@ -1149,9 +1162,7 @@ pushd %{qemu_kvm_build}
+   --firmwarepath=%{_prefix}/share/qemu-firmware \
+   --meson="git" \
+   --target-list="%{buildarch}" \
+-  --block-drv-rw-whitelist=%{block_drivers_list} \
+   --audio-drv-list= \
+-  --block-drv-ro-whitelist=vmdk,vhdx,vpc,https,ssh \
+   --with-coroutine=ucontext \
+   --with-git=git \
+   --tls-priority=@QEMU,SYSTEM \
+@@ -1197,6 +1208,7 @@ pushd %{qemu_kvm_build}
+ %endif
+   --enable-pie \
+   --enable-rbd \
++  --enable-vitastor \
+ %if 0%{have_librdma}
+   --enable-rdma \
+ %endif
+@@ -1794,6 +1806,9 @@ sh %{_sysconfdir}/sysconfig/modules/kvm.
+ %files block-rbd
+ %{_libdir}/qemu-kvm/block-rbd.so
+ 
++%files block-vitastor
++%{_libdir}/qemu-kvm/block-vitastor.so
++
+ %files block-ssh
+ %{_libdir}/qemu-kvm/block-ssh.so
+ 

rpm/qemu-kvm-7.2-el9.spec.patch

@@ -0,0 +1,93 @@
+--- qemu-kvm-7.2.spec.orig	2023-06-22 13:56:19.000000000 +0000
++++ qemu-kvm-7.2.spec	2023-07-18 07:55:22.347090196 +0000
+@@ -100,8 +100,6 @@
+ %endif
+ 
+ %global target_list %{kvm_target}-softmmu
+-%global block_drivers_rw_list qcow2,raw,file,host_device,nbd,iscsi,rbd,blkdebug,luks,null-co,nvme,copy-on-read,throttle,compress
+-%global block_drivers_ro_list vdi,vmdk,vhdx,vpc,https
+ %define qemudocdir %{_docdir}/%{name}
+ %global firmwaredirs "%{_datadir}/qemu-firmware:%{_datadir}/ipxe/qemu:%{_datadir}/seavgabios:%{_datadir}/seabios"
+ 
+@@ -126,6 +124,7 @@ Requires: %{name}-device-usb-host = %{ep
+ Requires: %{name}-device-usb-redirect = %{epoch}:%{version}-%{release}   \
+ %endif                                                           \
+ Requires: %{name}-block-rbd = %{epoch}:%{version}-%{release}     \
++Requires: %{name}-block-vitastor = %{epoch}:%{version}-%{release}\
+ Requires: %{name}-audio-pa = %{epoch}:%{version}-%{release}
+ 
+ # Since SPICE is removed from RHEL-9, the following Obsoletes:
+@@ -148,7 +147,7 @@ Obsoletes: %{name}-block-ssh <= %{epoch}
+ Summary: QEMU is a machine emulator and virtualizer
+ Name: qemu-kvm
+ Version: 7.2.0
+-Release: 14%{?rcrel}%{?dist}%{?cc_suffix}.1
++Release: 14.vitastor%{?rcrel}%{?dist}%{?cc_suffix}.1
+ # Epoch because we pushed a qemu-1.0 package. AIUI this can't ever be dropped
+ # Epoch 15 used for RHEL 8
+ # Epoch 17 used for RHEL 9 (due to release versioning offset in RHEL 8.5)
+@@ -171,6 +170,7 @@ Source28: 95-kvm-memlock.conf
+ Source30: kvm-s390x.conf
+ Source31: kvm-x86.conf
+ Source36: README.tests
++Source37: qemu-vitastor.c
+ 
+ 
+ Patch0004: 0004-Initial-redhat-build.patch
+@@ -418,6 +418,7 @@ Patch134: kvm-target-i386-Fix-BZHI-instr
+ Patch135: kvm-intel-iommu-fail-DEVIOTLB_UNMAP-without-dt-mode.patch
+ # For bz#2203745 - Disk detach is unsuccessful while the guest is still booting [rhel-9.2.0.z]
+ Patch136: kvm-acpi-pcihp-allow-repeating-hot-unplug-requests.patch
++Patch137: qemu-7.2-vitastor.patch
+ 
+ %if %{have_clang}
+ BuildRequires: clang
+@@ -449,6 +450,7 @@ BuildRequires: libcurl-devel
+ %if %{have_block_rbd}
+ BuildRequires: librbd-devel
+ %endif
++BuildRequires: vitastor-client-devel
+ # We need both because the 'stap' binary is probed for by configure
+ BuildRequires: systemtap
+ BuildRequires: systemtap-sdt-devel
+@@ -642,6 +644,14 @@ using the rbd protocol.
+ %endif
+ 
+ 
++%package  block-vitastor
++Summary: QEMU Vitastor block driver
++Requires: %{name}-common%{?_isa} = %{epoch}:%{version}-%{release}
++
++%description block-vitastor
++This package provides the additional Vitastor block driver for QEMU.
++
++
+ %package  audio-pa
+ Summary: QEMU PulseAudio audio driver
+ Requires: %{name}-common%{?_isa} = %{epoch}:%{version}-%{release}
+@@ -719,6 +729,7 @@ This package provides usbredir support.
+ %prep
+ %setup -q -n qemu-%{version}%{?rcstr}
+ %autopatch -p1
++cp %{SOURCE37} ./block/vitastor.c
+ 
+ %global qemu_kvm_build qemu_kvm_build
+ mkdir -p %{qemu_kvm_build}
+@@ -946,6 +957,7 @@ run_configure \
+ %if %{have_block_rbd}
+   --enable-rbd \
+ %endif
++  --enable-vitastor \
+ %if %{have_librdma}
+   --enable-rdma \
+ %endif
+@@ -1426,6 +1438,9 @@ useradd -r -u 107 -g qemu -G kvm -d / -s
+ %files block-rbd
+ %{_libdir}/%{name}/block-rbd.so
+ %endif
++%files block-vitastor
++%{_libdir}/%{name}/block-vitastor.so
++
+ %files audio-pa
+ %{_libdir}/%{name}/audio-pa.so
+ 

rpm/vitastor-el7.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.9.3.el7.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.0.0.el7.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el7.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.9.3
+Version:        1.0.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-0.9.3.el7.tar.gz
+Source0:        vitastor-1.0.0.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.9.3.el8.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.0.0.el8.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el8.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.9.3
+Version:        1.0.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-0.9.3.el8.tar.gz
+Source0:        vitastor-1.0.0.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el9.Dockerfile

@@ -18,7 +18,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.9.3.el9.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.0.0.el9.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el9.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el9.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.9.3
+Version:        1.0.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-0.9.3.el9.tar.gz
+Source0:        vitastor-1.0.0.el9.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -16,7 +16,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVERSION="0.9.3")
+add_definitions(-DVERSION="1.0.0")
 add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -I ${CMAKE_SOURCE_DIR}/src)
 if (${WITH_ASAN})
 	add_definitions(-fsanitize=address -fno-omit-frame-pointer)
@@ -137,6 +137,7 @@ endif (${WITH_FIO})
 add_library(vitastor_client SHARED
 	cluster_client.cpp
 	cluster_client_list.cpp
+	cluster_client_wb.cpp
 	vitastor_c.cpp
 	cli_common.cpp
 	cli_alloc_osd.cpp
@@ -300,7 +301,7 @@ target_link_libraries(test_crc32
 add_executable(test_cluster_client
 	EXCLUDE_FROM_ALL
 	test_cluster_client.cpp
-	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
+	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp cluster_client_wb.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
 	etcd_state_client.cpp timerfd_manager.cpp str_util.cpp ../json11/json11.cpp
 )
 target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)

src/addr_util.cpp

@@ -19,8 +19,8 @@ bool string_to_addr(std::string str, bool parse_port, int default_port, struct s
         if (p != std::string::npos && !(str.length() > 0 && str[p-1] == ']')) // "[ipv6]" which contains ':'
         {
             char null_byte = 0;
-            int n = sscanf(str.c_str()+p+1, "%d%c", &default_port, &null_byte);
-            if (n != 1 || default_port >= 0x10000)
+            int scanned = sscanf(str.c_str()+p+1, "%d%c", &default_port, &null_byte);
+            if (scanned != 1 || default_port >= 0x10000)
                 return false;
             str = str.substr(0, p);
         }

src/blockstore.cpp

@@ -82,3 +82,8 @@ uint32_t blockstore_t::get_bitmap_granularity()
 {
     return impl->get_bitmap_granularity();
 }
+
+bool blockstore_t::wants_fsync()
+{
+    return impl->wants_fsync();
+}

src/blockstore.h

@@ -226,4 +226,7 @@ public:
     uint64_t get_journal_size();
 
     uint32_t get_bitmap_granularity();
+
+    // Returns true if writing can stall due to a lack of fsync
+    bool wants_fsync();
 };

src/blockstore_disk.cpp

@@ -45,13 +45,31 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     meta_block_size = parse_size(config["meta_block_size"]);
     bitmap_granularity = parse_size(config["bitmap_granularity"]);
     meta_format = stoull_full(config["meta_format"]);
-    cached_read_data = config["cached_read_data"] == "true" || config["cached_read_data"] == "yes" || config["cached_read_data"] == "1";
-    cached_read_meta = cached_read_data && (meta_device == data_device || meta_device == "") &&
-        config.find("cached_read_meta") == config.end() ||
-        config["cached_read_meta"] == "true" || config["cached_read_meta"] == "yes" || config["cached_read_meta"] == "1";
-    cached_read_journal = cached_read_meta && (journal_device == meta_device || journal_device == "") &&
-        config.find("cached_read_journal") == config.end() ||
-        config["cached_read_journal"] == "true" || config["cached_read_journal"] == "yes" || config["cached_read_journal"] == "1";
+    if (config.find("data_io") == config.end() &&
+        config.find("meta_io") == config.end() &&
+        config.find("journal_io") == config.end())
+    {
+        bool cached_io_data = config["cached_io_data"] == "true" || config["cached_io_data"] == "yes" || config["cached_io_data"] == "1";
+        bool cached_io_meta = cached_io_data && (meta_device == data_device || meta_device == "") &&
+            config.find("cached_io_meta") == config.end() ||
+            config["cached_io_meta"] == "true" || config["cached_io_meta"] == "yes" || config["cached_io_meta"] == "1";
+        bool cached_io_journal = cached_io_meta && (journal_device == meta_device || journal_device == "") &&
+            config.find("cached_io_journal") == config.end() ||
+            config["cached_io_journal"] == "true" || config["cached_io_journal"] == "yes" || config["cached_io_journal"] == "1";
+        data_io = cached_io_data ? "cached" : "direct";
+        meta_io = cached_io_meta ? "cached" : "direct";
+        journal_io = cached_io_journal ? "cached" : "direct";
+    }
+    else
+    {
+        data_io = config.find("data_io") != config.end() ? config["data_io"] : "direct";
+        meta_io = config.find("meta_io") != config.end()
+            ? config["meta_io"]
+            : (meta_device == data_device || meta_device == "" ? data_io : "direct");
+        journal_io = config.find("journal_io") != config.end()
+            ? config["journal_io"]
+            : (journal_device == meta_device || journal_device == "" ? meta_io : "direct");
+    }
     if (config["data_csum_type"] == "crc32c")
     {
         data_csum_type = BLOCKSTORE_CSUM_CRC32C;
@@ -272,9 +290,19 @@ static void check_size(int fd, uint64_t *size, uint64_t *sectsize, std::string n
     }
 }
 
+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()
 {
-    data_fd = open(data_device.c_str(), O_DIRECT|O_RDWR);
+    data_fd = open(data_device.c_str(), bs_openmode(data_io) | O_RDWR);
     if (data_fd == -1)
     {
         throw std::runtime_error("Failed to open data device "+data_device+": "+std::string(strerror(errno)));
@@ -295,25 +323,13 @@ void blockstore_disk_t::open_data()
     {
         throw std::runtime_error(std::string("Failed to lock data device: ") + strerror(errno));
     }
-    if (cached_read_data)
-    {
-        read_data_fd = open(data_device.c_str(), O_RDWR);
-        if (read_data_fd == -1)
-        {
-            throw std::runtime_error("Failed to open data device "+data_device+": "+std::string(strerror(errno)));
-        }
-    }
-    else
-    {
-        read_data_fd = data_fd;
-    }
 }
 
 void blockstore_disk_t::open_meta()
 {
-    if (meta_device != data_device)
+    if (meta_device != data_device || meta_io != data_io)
     {
-        meta_fd = open(meta_device.c_str(), O_DIRECT|O_RDWR);
+        meta_fd = open(meta_device.c_str(), bs_openmode(meta_io) | O_RDWR);
         if (meta_fd == -1)
         {
             throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
@@ -323,22 +339,10 @@ void blockstore_disk_t::open_meta()
         {
             throw std::runtime_error("meta_offset exceeds device size = "+std::to_string(meta_device_size));
         }
-        if (!disable_flock && flock(meta_fd, LOCK_EX|LOCK_NB) != 0)
+        if (!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));
         }
-        if (cached_read_meta)
-        {
-            read_meta_fd = open(meta_device.c_str(), O_RDWR);
-            if (read_meta_fd == -1)
-            {
-                throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
-            }
-        }
-        else
-        {
-            read_meta_fd = meta_fd;
-        }
     }
     else
     {
@@ -357,35 +361,19 @@ void blockstore_disk_t::open_meta()
             ") is not a multiple of data device sector size ("+std::to_string(meta_device_sect)+")"
         );
     }
-    if (!cached_read_meta)
-    {
-        read_meta_fd = meta_fd;
-    }
-    else if (meta_device == data_device && cached_read_data)
-    {
-        read_meta_fd = read_data_fd;
-    }
-    else
-    {
-        read_meta_fd = open(meta_device.c_str(), O_RDWR);
-        if (read_meta_fd == -1)
-        {
-            throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
-        }
-    }
 }
 
 void blockstore_disk_t::open_journal()
 {
-    if (journal_device != meta_device)
+    if (journal_device != meta_device || journal_io != meta_io)
     {
-        journal_fd = open(journal_device.c_str(), O_DIRECT|O_RDWR);
+        journal_fd = open(journal_device.c_str(), bs_openmode(journal_io) | O_RDWR);
         if (journal_fd == -1)
         {
             throw std::runtime_error("Failed to open journal device "+journal_device+": "+std::string(strerror(errno)));
         }
         check_size(journal_fd, &journal_device_size, &journal_device_sect, "journal device");
-        if (!disable_flock && flock(journal_fd, LOCK_EX|LOCK_NB) != 0)
+        if (!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));
         }
@@ -407,26 +395,6 @@ void blockstore_disk_t::open_journal()
             ") is not a multiple of journal device sector size ("+std::to_string(journal_device_sect)+")"
         );
     }
-    if (!cached_read_journal)
-    {
-        read_journal_fd = journal_fd;
-    }
-    else if (journal_device == meta_device && cached_read_meta)
-    {
-        read_journal_fd = read_meta_fd;
-    }
-    else if (journal_device == data_device && cached_read_data)
-    {
-        read_journal_fd = read_data_fd;
-    }
-    else
-    {
-        read_journal_fd = open(journal_device.c_str(), O_RDWR);
-        if (read_journal_fd == -1)
-        {
-            throw std::runtime_error("Failed to open journal device "+journal_device+": "+std::string(strerror(errno)));
-        }
-    }
 }
 
 void blockstore_disk_t::close_all()
@@ -437,12 +405,5 @@ void blockstore_disk_t::close_all()
         close(meta_fd);
     if (journal_fd >= 0 && journal_fd != meta_fd)
         close(journal_fd);
-    if (read_data_fd >= 0 && read_data_fd != data_fd)
-        close(read_data_fd);
-    if (read_meta_fd >= 0 && read_meta_fd != meta_fd)
-        close(read_meta_fd);
-    if (read_journal_fd >= 0 && read_journal_fd != journal_fd)
-        close(read_journal_fd);
     data_fd = meta_fd = journal_fd = -1;
-    read_data_fd = read_meta_fd = read_journal_fd = -1;
 }

src/blockstore_disk.h

@@ -31,11 +31,11 @@ struct blockstore_disk_t
     uint32_t csum_block_size = 4096;
     // By default, Blockstore locks all opened devices exclusively. This option can be used to disable locking
     bool disable_flock = false;
-    // Use linux page cache for reads. If enabled, separate buffered FDs will be opened for reading
-    bool cached_read_data = false, cached_read_meta = false, cached_read_journal = false;
+    // I/O modes for data, metadata and journal: direct or "" = O_DIRECT, cached = O_SYNC, directsync = O_DIRECT|O_SYNC
+    // O_SYNC without O_DIRECT = use Linux page cache for reads and writes
+    std::string data_io, meta_io, journal_io;
 
     int meta_fd = -1, data_fd = -1, journal_fd = -1;
-    int read_meta_fd = -1, read_data_fd = -1, read_journal_fd = -1;
     uint64_t meta_offset, meta_device_sect, meta_device_size, meta_len, meta_format = 0;
     uint64_t data_offset, data_device_sect, data_device_size, data_len;
     uint64_t journal_offset, journal_device_sect, journal_device_size, journal_len;

src/blockstore_flush.cpp

@@ -1087,7 +1087,7 @@ bool journal_flusher_co::read_dirty(int wait_base)
         data->iov = (struct iovec){ vi.buf, vi.len };
         data->callback = simple_callback_r;
         my_uring_prep_readv(
-            sqe, bs->dsk.read_data_fd, &data->iov, 1, bs->dsk.data_offset + old_clean_loc + vi.offset
+            sqe, bs->dsk.data_fd, &data->iov, 1, bs->dsk.data_offset + old_clean_loc + vi.offset
         );
         wait_count++;
         bs->find_holes(v, vi.offset, vi.offset+vi.len, [this, buf = (uint8_t*)vi.buf-vi.offset](int pos, bool alloc, uint32_t cur_start, uint32_t cur_end)
@@ -1119,7 +1119,7 @@ bool journal_flusher_co::read_dirty(int wait_base)
                 data->iov = (struct iovec){ v[i].buf, (size_t)v[i].len };
                 data->callback = simple_callback_rj;
                 my_uring_prep_readv(
-                    sqe, bs->dsk.read_journal_fd, &data->iov, 1, bs->journal.offset + v[i].disk_offset
+                    sqe, bs->dsk.journal_fd, &data->iov, 1, bs->journal.offset + v[i].disk_offset
                 );
                 wait_journal_count++;
             }
@@ -1212,7 +1212,7 @@ bool journal_flusher_co::modify_meta_read(uint64_t meta_loc, flusher_meta_write_
         data->callback = simple_callback_r;
         wr.submitted = true;
         my_uring_prep_readv(
-            sqe, bs->dsk.read_meta_fd, &data->iov, 1, bs->dsk.meta_offset + bs->dsk.meta_block_size + wr.sector
+            sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + bs->dsk.meta_block_size + wr.sector
         );
         wait_count++;
     }

src/blockstore_impl.cpp

@@ -167,7 +167,7 @@ void blockstore_impl_t::loop()
                 // wait for all big writes to complete, submit data device fsync
                 // wait for the data device fsync to complete, then submit journal writes for big writes
                 // then submit an fsync operation
-                if (has_writes)
+                if (0 && has_writes)
                 {
                     // Can't submit SYNC before previous writes
                     continue;
@@ -384,6 +384,10 @@ void blockstore_impl_t::enqueue_op(blockstore_op_t *op)
         ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
         return;
     }
+    if (op->opcode == BS_OP_SYNC)
+    {
+        unsynced_queued_ops = 0;
+    }
     init_op(op);
     submit_queue.push_back(op);
     ringloop->wakeup();
@@ -393,6 +397,7 @@ void blockstore_impl_t::init_op(blockstore_op_t *op)
 {
     // Call constructor without allocating memory. We'll call destructor before returning op back
     new ((void*)op->private_data) blockstore_op_private_t;
+    PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
     PRIV(op)->wait_for = 0;
     PRIV(op)->op_state = 0;
     PRIV(op)->pending_ops = 0;
@@ -729,3 +734,15 @@ void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expecte
     fprintf(stderr, "Disk %s failed: result is %d, expected %d. Can't continue, sorry :-(\n", op, retval, expected);
     exit(1);
 }
+
+bool blockstore_impl_t::wants_fsync()
+{
+    if (!unstable_writes.size())
+    {
+        return false;
+    }
+    uint64_t journal_free_space = journal.next_free < journal.used_start
+        ? (journal.used_start - journal.next_free)
+        : (journal.len - journal.next_free + journal.used_start - journal.block_size);
+    return journal_fsync_feedback_limit > 0 && journal.len-journal_free_space >= journal_fsync_feedback_limit;
+}

src/blockstore_impl.h

@@ -210,7 +210,7 @@ struct blockstore_op_private_t
     std::vector<copy_buffer_t> read_vec;
 
     // Sync, write
-    int min_flushed_journal_sector, max_flushed_journal_sector;
+    uint64_t min_flushed_journal_sector, max_flushed_journal_sector;
 
     // Write
     struct iovec iov_zerofill[3];
@@ -220,7 +220,6 @@ struct blockstore_op_private_t
 
     // Sync
     std::vector<obj_ver_id> sync_big_writes, sync_small_writes;
-    int sync_small_checked, sync_big_checked;
 };
 
 typedef uint32_t pool_id_t;
@@ -263,6 +262,10 @@ class blockstore_impl_t
     int throttle_target_parallelism = 1;
     // Minimum difference in microseconds between target and real execution times to throttle the response
     int throttle_threshold_us = 50;
+    // Maximum writes between automatically added fsync operations
+    uint64_t autosync_writes = 128;
+    // Maximum free space in the journal in bytes to start sending fsync feedback to primary OSDs
+    uint64_t journal_fsync_feedback_limit = 0;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;
@@ -274,6 +277,7 @@ class blockstore_impl_t
     std::vector<blockstore_op_t*> submit_queue;
     std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
     int unsynced_big_write_count = 0;
+    int unsynced_queued_ops = 0;
     allocator *data_alloc = NULL;
     uint8_t *zero_object;
 
@@ -431,4 +435,6 @@ public:
     inline uint64_t get_free_block_count() { return data_alloc->get_free_count(); }
     inline uint32_t get_bitmap_granularity() { return dsk.disk_alignment; }
     inline uint64_t get_journal_size() { return dsk.journal_len; }
+
+    bool wants_fsync();
 };

src/blockstore_init.cpp

@@ -65,7 +65,7 @@ int blockstore_init_meta::loop()
     GET_SQE();
     data->iov = { metadata_buffer, bs->dsk.meta_block_size };
     data->callback = [this](ring_data_t *data) { handle_event(data, -1); };
-    my_uring_prep_readv(sqe, bs->dsk.read_meta_fd, &data->iov, 1, bs->dsk.meta_offset);
+    my_uring_prep_readv(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset);
     bs->ringloop->submit();
     submitted++;
 resume_1:
@@ -202,7 +202,7 @@ resume_2:
                 data->iov = { bufs[i].buf, bufs[i].size };
                 data->callback = [this, i](ring_data_t *data) { handle_event(data, i); };
                 if (!zero_on_init)
-                    my_uring_prep_readv(sqe, bs->dsk.read_meta_fd, &data->iov, 1, bs->dsk.meta_offset + bufs[i].offset);
+                    my_uring_prep_readv(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + bufs[i].offset);
                 else
                 {
                     // Fill metadata with zeroes
@@ -259,7 +259,7 @@ resume_2:
             GET_SQE();
             data->iov = { metadata_buffer, bs->dsk.meta_block_size };
             data->callback = [this](ring_data_t *data) { handle_event(data, -1); };
-            my_uring_prep_readv(sqe, bs->dsk.read_meta_fd, &data->iov, 1, bs->dsk.meta_offset + (1+next_offset)*bs->dsk.meta_block_size);
+            my_uring_prep_readv(sqe, bs->dsk.meta_fd, &data->iov, 1, bs->dsk.meta_offset + (1+next_offset)*bs->dsk.meta_block_size);
             submitted++;
 resume_5:
             if (submitted > 0)
@@ -467,7 +467,7 @@ int blockstore_init_journal::loop()
     data = ((ring_data_t*)sqe->user_data);
     data->iov = { submitted_buf, bs->journal.block_size };
     data->callback = simple_callback;
-    my_uring_prep_readv(sqe, bs->dsk.read_journal_fd, &data->iov, 1, bs->journal.offset);
+    my_uring_prep_readv(sqe, bs->dsk.journal_fd, &data->iov, 1, bs->journal.offset);
     bs->ringloop->submit();
     wait_count = 1;
 resume_1:
@@ -607,7 +607,7 @@ resume_1:
                     end - journal_pos < JOURNAL_BUFFER_SIZE ? end - journal_pos : JOURNAL_BUFFER_SIZE,
                 };
                 data->callback = [this](ring_data_t *data1) { handle_event(data1); };
-                my_uring_prep_readv(sqe, bs->dsk.read_journal_fd, &data->iov, 1, bs->journal.offset + journal_pos);
+                my_uring_prep_readv(sqe, bs->dsk.journal_fd, &data->iov, 1, bs->journal.offset + journal_pos);
                 bs->ringloop->submit();
             }
             while (done.size() > 0)

src/blockstore_journal.cpp

@@ -198,6 +198,7 @@ void blockstore_impl_t::prepare_journal_sector_write(int cur_sector, blockstore_
     priv->pending_ops++;
     if (!priv->min_flushed_journal_sector)
         priv->min_flushed_journal_sector = 1+cur_sector;
+    assert(priv->min_flushed_journal_sector <= journal.sector_count);
     priv->max_flushed_journal_sector = 1+cur_sector;
 }
 

src/blockstore_open.cpp

@@ -4,6 +4,25 @@
 #include <sys/file.h>
 #include "blockstore_impl.h"
 
+static uint64_t parse_fsync_feedback(blockstore_config_t & config, uint64_t journal_len)
+{
+    uint64_t journal_fsync_feedback_limit = 0;
+    if (config.find("journal_min_free_bytes") == config.end() &&
+        config.find("journal_min_free_percent") == config.end())
+    {
+        journal_fsync_feedback_limit = 90 * journal_len / 100;
+    }
+    else
+    {
+        journal_fsync_feedback_limit = strtoull(config["journal_min_free_bytes"].c_str(), NULL, 10);
+        if (!journal_fsync_feedback_limit)
+        {
+            journal_fsync_feedback_limit = strtoull(config["journal_min_free_percent"].c_str(), NULL, 10) * journal_len / 100;
+        }
+    }
+    return journal_fsync_feedback_limit;
+}
+
 void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
 {
     // Online-configurable options:
@@ -19,6 +38,10 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     throttle_target_mbs = strtoull(config["throttle_target_mbs"].c_str(), NULL, 10);
     throttle_target_parallelism = strtoull(config["throttle_target_parallelism"].c_str(), NULL, 10);
     throttle_threshold_us = strtoull(config["throttle_threshold_us"].c_str(), NULL, 10);
+    if (config.find("autosync_writes") != config.end())
+    {
+        autosync_writes = strtoull(config["autosync_writes"].c_str(), NULL, 10);
+    }
     if (!max_flusher_count)
     {
         max_flusher_count = 256;
@@ -49,6 +72,8 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     }
     if (!init)
     {
+        // has to be parsed after dsk.parse_config(), thus repeated here for online update
+        journal_fsync_feedback_limit = parse_fsync_feedback(config, journal.len);
         return;
     }
     // Offline-configurable options:
@@ -85,11 +110,14 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
         immediate_commit = IMMEDIATE_SMALL;
     }
     metadata_buf_size = strtoull(config["meta_buf_size"].c_str(), NULL, 10);
-    inmemory_meta = config["inmemory_metadata"] != "false";
+    inmemory_meta = config["inmemory_metadata"] != "false" && config["inmemory_metadata"] != "0" &&
+        config["inmemory_metadata"] != "no";
     journal.sector_count = strtoull(config["journal_sector_buffer_count"].c_str(), NULL, 10);
     journal.no_same_sector_overwrites = config["journal_no_same_sector_overwrites"] == "true" ||
         config["journal_no_same_sector_overwrites"] == "1" || config["journal_no_same_sector_overwrites"] == "yes";
-    journal.inmemory = config["inmemory_journal"] != "false";
+    journal.inmemory = config["inmemory_journal"] != "false" && config["inmemory_journal"] != "0" &&
+        config["inmemory_journal"] != "no";
+    journal_fsync_feedback_limit = parse_fsync_feedback(config, journal.len);
     // Validate
     if (journal.sector_count < 2)
     {

src/blockstore_read.cpp

@@ -29,7 +29,7 @@ int blockstore_impl_t::fulfill_read_push(blockstore_op_t *op, void *buf, uint64_
     PRIV(op)->pending_ops++;
     my_uring_prep_readv(
         sqe,
-        IS_JOURNAL(item_state) ? dsk.read_journal_fd : dsk.read_data_fd,
+        IS_JOURNAL(item_state) ? dsk.journal_fd : dsk.data_fd,
         &data->iov, 1,
         (IS_JOURNAL(item_state) ? dsk.journal_offset : dsk.data_offset) + offset
     );
@@ -348,7 +348,7 @@ bool blockstore_impl_t::read_checksum_block(blockstore_op_t *op, int rv_pos, uin
         .csum_buf = vi->csum_buf,
         .dyn_data = vi->dyn_data,
     };
-    int submit_fd = (vi->copy_flags & COPY_BUF_JOURNAL ? dsk.read_journal_fd : dsk.read_data_fd);
+    int submit_fd = (vi->copy_flags & COPY_BUF_JOURNAL ? dsk.journal_fd : dsk.data_fd);
     uint64_t submit_offset = (vi->copy_flags & COPY_BUF_JOURNAL ? journal.offset : dsk.data_offset);
     uint32_t d_pos = 0;
     for (int n_pos = 0; n_pos < n_iov; n_pos += IOV_MAX)
@@ -702,7 +702,7 @@ uint8_t* blockstore_impl_t::read_clean_meta_block(blockstore_op_t *op, uint64_t
     BS_SUBMIT_GET_SQE(sqe, data);
     data->iov = (struct iovec){ buf, dsk.meta_block_size };
     PRIV(op)->pending_ops++;
-    my_uring_prep_readv(sqe, dsk.read_meta_fd, &data->iov, 1, dsk.meta_offset + dsk.meta_block_size + sector);
+    my_uring_prep_readv(sqe, dsk.meta_fd, &data->iov, 1, dsk.meta_offset + dsk.meta_block_size + sector);
     data->callback = [this, op](ring_data_t *data) { handle_read_event(data, op); };
     // return pointer to checksums + bitmap
     return buf + pos + sizeof(clean_disk_entry);

src/blockstore_sync.cpp

@@ -16,7 +16,6 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
 {
     if (immediate_commit == IMMEDIATE_ALL)
     {
-        // We can return immediately because sync is only dequeued after all previous writes
         op->retval = 0;
         FINISH_OP(op);
         return 2;
@@ -27,8 +26,6 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
         unsynced_big_write_count -= unsynced_big_writes.size();
         PRIV(op)->sync_big_writes.swap(unsynced_big_writes);
         PRIV(op)->sync_small_writes.swap(unsynced_small_writes);
-        PRIV(op)->sync_small_checked = 0;
-        PRIV(op)->sync_big_checked = 0;
         unsynced_big_writes.clear();
         unsynced_small_writes.clear();
         if (PRIV(op)->sync_big_writes.size() > 0)

src/blockstore_write.cpp

@@ -127,8 +127,9 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
             return false;
         }
     }
-    if (wait_big && !is_del && !deleted && op->len < dsk.data_block_size &&
-        immediate_commit != IMMEDIATE_ALL)
+    bool imm = (op->len < dsk.data_block_size ? (immediate_commit != IMMEDIATE_NONE) : (immediate_commit == IMMEDIATE_ALL));
+    if (wait_big && !is_del && !deleted && op->len < dsk.data_block_size && !imm ||
+        !imm && unsynced_queued_ops >= autosync_writes)
     {
         // Issue an additional sync so that the previous big write can reach the journal
         blockstore_op_t *sync_op = new blockstore_op_t;
@@ -139,6 +140,8 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
         };
         enqueue_op(sync_op);
     }
+    else if (!imm)
+        unsynced_queued_ops++;
 #ifdef BLOCKSTORE_DEBUG
     if (is_del)
         printf("Delete %lx:%lx v%lu\n", op->oid.inode, op->oid.stripe, op->version);
@@ -378,7 +381,6 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             sqe, dsk.data_fd, PRIV(op)->iov_zerofill, vcnt, dsk.data_offset + (loc << dsk.block_order) + op->offset - stripe_offset
         );
         PRIV(op)->pending_ops = 1;
-        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
         if (immediate_commit != IMMEDIATE_ALL)
         {
             // Increase the counter, but don't save into unsynced_writes yet (can't sync until the write is finished)
@@ -415,16 +417,10 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         write_iodepth++;
         // Got SQEs. Prepare previous journal sector write if required
         auto cb = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-        if (immediate_commit == IMMEDIATE_NONE)
+        if (immediate_commit == IMMEDIATE_NONE &&
+            !journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size))
         {
-            if (!journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size))
-            {
-                prepare_journal_sector_write(journal.cur_sector, op);
-            }
-            else
-            {
-                PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-            }
+            prepare_journal_sector_write(journal.cur_sector, op);
         }
         // Then pre-fill journal entry
         journal_entry_small_write *je = (journal_entry_small_write*)prefill_single_journal_entry(
@@ -750,17 +746,11 @@ int blockstore_impl_t::dequeue_del(blockstore_op_t *op)
     }
     write_iodepth++;
     // Prepare journal sector write
-    if (immediate_commit == IMMEDIATE_NONE)
+    if (immediate_commit == IMMEDIATE_NONE &&
+        (dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
+        journal.sector_info[journal.cur_sector].dirty)
     {
-        if ((dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
-            journal.sector_info[journal.cur_sector].dirty)
-        {
-            prepare_journal_sector_write(journal.cur_sector, op);
-        }
-        else
-        {
-            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-        }
+        prepare_journal_sector_write(journal.cur_sector, op);
     }
     // Pre-fill journal entry
     journal_entry_del *je = (journal_entry_del*)prefill_single_journal_entry(

src/cli.cpp

@@ -349,6 +349,7 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
                 p->ringloop->wait();
         }
         // Destroy the client
+        p->cli->flush();
         delete p->cli;
         delete p->epmgr;
         delete p->ringloop;
@@ -357,6 +358,8 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
         p->ringloop = NULL;
     }
     // Print result
+    fflush(stderr);
+    fflush(stdout);
     if (p->json_output && !result.data.is_null())
     {
         printf("%s\n", result.data.dump().c_str());

src/cli_alloc_osd.cpp

@@ -77,8 +77,8 @@ struct alloc_osd_t
                     std::string key = base64_decode(kv["key"].string_value());
                     osd_num_t cur_osd;
                     char null_byte = 0;
-                    sscanf(key.c_str() + parent->cli->st_cli.etcd_prefix.length(), "/osd/stats/%lu%c", &cur_osd, &null_byte);
-                    if (!cur_osd || null_byte != 0)
+                    int scanned = sscanf(key.c_str() + parent->cli->st_cli.etcd_prefix.length(), "/osd/stats/%lu%c", &cur_osd, &null_byte);
+                    if (scanned != 1 || !cur_osd)
                     {
                         fprintf(stderr, "Invalid key in etcd: %s\n", key.c_str());
                         continue;

src/cli_df.cpp

@@ -67,8 +67,8 @@ resume_1:
             // pool ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -82,8 +82,8 @@ resume_1:
             // osd ID
             osd_num_t osd_num;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/osd/stats/%lu%c", &osd_num, &null_byte);
-            if (!osd_num || osd_num >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/osd/stats/%lu%c", &osd_num, &null_byte);
+            if (scanned != 1 || !osd_num || osd_num >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;

src/cli_ls.cpp

@@ -56,14 +56,15 @@ struct image_lister_t
             {
                 continue;
             }
-            auto & pool_cfg = parent->cli->st_cli.pool_config.at(INODE_POOL(ic.second.num));
+            auto pool_it = parent->cli->st_cli.pool_config.find(INODE_POOL(ic.second.num));
+            bool good_pool = pool_it != parent->cli->st_cli.pool_config.end();
             auto item = json11::Json::object {
                 { "name", ic.second.name },
                 { "size", ic.second.size },
                 { "used_size", 0 },
                 { "readonly", ic.second.readonly },
                 { "pool_id", (uint64_t)INODE_POOL(ic.second.num) },
-                { "pool_name", pool_cfg.name },
+                { "pool_name", good_pool ? pool_it->second.name : "? (ID:"+std::to_string(INODE_POOL(ic.second.num))+")" },
                 { "inode_num", INODE_NO_POOL(ic.second.num) },
                 { "inode_id", ic.second.num },
             };
@@ -132,8 +133,8 @@ resume_1:
             // pool ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -148,9 +149,9 @@ resume_1:
             pool_id_t pool_id;
             inode_t only_inode_num;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(),
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(),
                 "/inode/stats/%u/%lu%c", &pool_id, &only_inode_num, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || INODE_POOL(only_inode_num) != 0 || null_byte != 0)
+            if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || INODE_POOL(only_inode_num) != 0)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -247,6 +248,8 @@ resume_1:
         if (state == 1)
             goto resume_1;
         get_list();
+        if (state == 100)
+            return;
         if (show_stats)
         {
 resume_1:
@@ -269,7 +272,7 @@ resume_1:
             { "key", "name" },
             { "title", "NAME" },
         });
-        if (!list_pool_id)
+        if (list_pool_name == "")
         {
             cols.push_back(json11::Json::object{
                 { "key", "pool_name" },
@@ -376,16 +379,18 @@ resume_1:
 
 std::string print_table(json11::Json items, json11::Json header, bool use_esc)
 {
+    int header_sizes[header.array_items().size()];
     std::vector<int> sizes;
     for (int i = 0; i < header.array_items().size(); i++)
     {
-        sizes.push_back(header[i]["title"].string_value().length());
+        header_sizes[i] = utf8_length(header[i]["title"].string_value());
+        sizes.push_back(header_sizes[i]);
     }
     for (auto & item: items.array_items())
     {
         for (int i = 0; i < header.array_items().size(); i++)
         {
-            int l = item[header[i]["key"].string_value()].as_string().length();
+            int l = utf8_length(item[header[i]["key"].string_value()].as_string());
             sizes[i] = sizes[i] < l ? l : sizes[i];
         }
     }
@@ -397,7 +402,7 @@ std::string print_table(json11::Json items, json11::Json header, bool use_esc)
             // Separator
             str += "  ";
         }
-        int pad = sizes[i]-header[i]["title"].string_value().length();
+        int pad = sizes[i]-header_sizes[i];
         if (header[i]["right"].bool_value())
         {
             // Align right
@@ -425,7 +430,7 @@ std::string print_table(json11::Json items, json11::Json header, bool use_esc)
                 // Separator
                 str += "  ";
             }
-            int pad = sizes[i] - item[header[i]["key"].string_value()].as_string().length();
+            int pad = sizes[i] - utf8_length(item[header[i]["key"].string_value()].as_string());
             if (header[i]["right"].bool_value())
             {
                 // Align right

src/cli_merge.cpp

@@ -53,6 +53,7 @@ struct snap_merger_t
     std::map<inode_t, std::vector<uint64_t>> layer_lists;
     std::map<inode_t, uint64_t> layer_block_size;
     std::map<inode_t, uint64_t> layer_list_pos;
+    std::vector<snap_rw_op_t*> continue_rwo, continue_rwo2;
     int in_flight = 0;
     uint64_t last_fsync_offset = 0;
     uint64_t last_written_offset = 0;
@@ -304,6 +305,12 @@ struct snap_merger_t
         oit = merge_offsets.begin();
     resume_5:
         // Now read, overwrite and optionally delete offsets one by one
+        continue_rwo2.swap(continue_rwo);
+        for (auto rwo: continue_rwo2)
+        {
+            next_write(rwo);
+        }
+        continue_rwo2.clear();
         while (in_flight < parent->iodepth*parent->parallel_osds &&
             oit != merge_offsets.end() && !rwo_error.size())
         {
@@ -464,7 +471,8 @@ struct snap_merger_t
                 rwo->error_offset = op->offset;
                 rwo->error_read = true;
             }
-            next_write(rwo);
+            continue_rwo.push_back(rwo);
+            parent->ringloop->wakeup();
         };
         parent->cli->execute(op);
     }
@@ -544,11 +552,9 @@ struct snap_merger_t
             }
             // Increment CAS version
             rwo->op.version = subop->version;
-            if (use_cas)
-                next_write(rwo);
-            else
-                autofree_op(rwo);
             delete subop;
+            continue_rwo.push_back(rwo);
+            parent->ringloop->wakeup();
         };
         parent->cli->execute(subop);
     }

src/cli_modify.cpp

@@ -13,7 +13,7 @@ struct image_changer_t
     std::string image_name;
     std::string new_name;
     uint64_t new_size = 0;
-    bool force_size = false;
+    bool force_size = false, inc_size = false;
     bool set_readonly = false, set_readwrite = false, force = false;
     // interval between fsyncs
     int fsync_interval = 128;
@@ -81,14 +81,14 @@ struct image_changer_t
         }
         if ((!set_readwrite || !cfg.readonly) &&
             (!set_readonly || cfg.readonly) &&
-            (!new_size && !force_size || cfg.size == new_size) &&
+            (!new_size && !force_size || cfg.size == new_size || cfg.size >= new_size && inc_size) &&
             (new_name == "" || new_name == image_name))
         {
             result = (cli_result_t){ .text = "No change" };
             state = 100;
             return;
         }
-        if (new_size != 0 || force_size)
+        if ((new_size != 0 || force_size) && (cfg.size < new_size || !inc_size))
         {
             if (cfg.size >= new_size)
             {
@@ -233,6 +233,7 @@ std::function<bool(cli_result_t &)> cli_tool_t::start_modify(json11::Json cfg)
     changer->new_name = cfg["rename"].string_value();
     changer->new_size = parse_size(cfg["resize"].as_string());
     changer->force_size = cfg["force_size"].bool_value();
+    changer->inc_size = cfg["inc_size"].bool_value();
     changer->force = cfg["force"].bool_value();
     changer->set_readonly = cfg["readonly"].bool_value();
     changer->set_readwrite = cfg["readwrite"].bool_value();

src/cli_rm.cpp

@@ -384,8 +384,8 @@ resume_100:
                 pool_id_t pool_id = 0;
                 inode_t inode = 0;
                 char null_byte = 0;
-                sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode, &null_byte);
-                if (!inode || null_byte != 0)
+                int scanned = sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode, &null_byte);
+                if (scanned != 2 || !inode)
                 {
                     result = (cli_result_t){ .err = EIO, .text = "Bad key returned from etcd: "+kv.key };
                     state = 100;

src/cli_status.cpp

@@ -132,8 +132,8 @@ resume_2:
             auto kv = parent->cli->st_cli.parse_etcd_kv(osd_stats[i]);
             osd_num_t stat_osd_num = 0;
             char null_byte = 0;
-            sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.size(), "/osd/stats/%lu%c", &stat_osd_num, &null_byte);
-            if (!stat_osd_num || null_byte != 0)
+            int scanned = sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.size(), "/osd/stats/%lu%c", &stat_osd_num, &null_byte);
+            if (scanned != 1 || !stat_osd_num)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;

src/cluster_client.cpp

@@ -3,21 +3,13 @@
 
 #include <stdexcept>
 #include <assert.h>
-#include "cluster_client.h"
-
-#define SCRAP_BUFFER_SIZE 4*1024*1024
-#define PART_SENT 1
-#define PART_DONE 2
-#define PART_ERROR 4
-#define PART_RETRY 8
-#define CACHE_DIRTY 1
-#define CACHE_FLUSHING 2
-#define CACHE_REPEATING 3
-#define OP_FLUSH_BUFFER 0x02
-#define OP_IMMEDIATE_COMMIT 0x04
+#include "cluster_client_impl.h"
+#include "http_client.h" // json_is_true
 
 cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config)
 {
+    wb = new writeback_cache_t();
+
     cli_config = config.object_items();
     file_config = osd_messenger_t::read_config(config);
     config = osd_messenger_t::merge_configs(cli_config, file_config, etcd_global_config, {});
@@ -37,20 +29,14 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
             continue_lists();
             continue_raw_ops(peer_osd);
         }
-        else if (dirty_buffers.size())
+        else
         {
             // peer_osd just dropped connection
             // determine WHICH dirty_buffers are now obsolete and repeat them
-            for (auto & wr: dirty_buffers)
+            if (wb->repeat_ops_for(this, peer_osd) > 0)
             {
-                if (affects_osd(wr.first.inode, wr.first.stripe, wr.second.len, peer_osd) &&
-                    wr.second.state != CACHE_REPEATING)
-                {
-                    // FIXME: Flush in larger parts
-                    flush_buffer(wr.first, &wr.second);
-                }
+                continue_ops();
             }
-            continue_ops();
         }
     };
     msgr.exec_op = [this](osd_op_t *op)
@@ -78,16 +64,14 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
 
 cluster_client_t::~cluster_client_t()
 {
-    for (auto bp: dirty_buffers)
-    {
-        free(bp.second.buf);
-    }
-    dirty_buffers.clear();
+    msgr.repeer_pgs = [this](osd_num_t){};
     if (ringloop)
     {
         ringloop->unregister_consumer(&consumer);
     }
     free(scrap_buffer);
+    delete wb;
+    wb = NULL;
 }
 
 cluster_op_t::~cluster_op_t()
@@ -136,6 +120,19 @@ void cluster_client_t::init_msgr()
     }
 }
 
+void cluster_client_t::unshift_op(cluster_op_t *op)
+{
+    op->next = op_queue_head;
+    if (op_queue_head)
+    {
+        op_queue_head->prev = op;
+        op_queue_head = op;
+    }
+    else
+        op_queue_tail = op_queue_head = op;
+    inc_wait(op->opcode, op->flags, op->next, 1);
+}
+
 void cluster_client_t::calc_wait(cluster_op_t *op)
 {
     op->prev_wait = 0;
@@ -156,7 +153,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
     {
         for (auto prev = op->prev; prev; prev = prev->prev)
         {
-            if (prev->opcode == OSD_OP_SYNC || prev->opcode == OSD_OP_WRITE && !(prev->flags & OP_IMMEDIATE_COMMIT))
+            if (prev->opcode == OSD_OP_SYNC || prev->opcode == OSD_OP_WRITE && (!(prev->flags & OP_IMMEDIATE_COMMIT) || enable_writeback))
             {
                 op->prev_wait++;
             }
@@ -166,21 +163,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
     }
     else /* if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP) */
     {
-        for (auto prev = op_queue_head; prev && prev != op; prev = prev->next)
-        {
-            if (prev->opcode == OSD_OP_WRITE && (prev->flags & OP_FLUSH_BUFFER))
-            {
-                op->prev_wait++;
-            }
-            else if (prev->opcode == OSD_OP_WRITE || prev->opcode == OSD_OP_READ ||
-                prev->opcode == OSD_OP_READ_BITMAP || prev->opcode == OSD_OP_READ_CHAIN_BITMAP)
-            {
-                // Flushes are always in the beginning (we're scanning from the beginning of the queue)
-                break;
-            }
-        }
-        if (!op->prev_wait)
-            continue_rw(op);
+        continue_rw(op);
     }
 }
 
@@ -191,10 +174,8 @@ void cluster_client_t::inc_wait(uint64_t opcode, uint64_t flags, cluster_op_t *n
         while (next)
         {
             auto n2 = next->next;
-            if (next->opcode == OSD_OP_SYNC && !(flags & OP_IMMEDIATE_COMMIT) ||
-                next->opcode == OSD_OP_WRITE && (flags & OP_FLUSH_BUFFER) && !(next->flags & OP_FLUSH_BUFFER) ||
-                (next->opcode == OSD_OP_READ || next->opcode == OSD_OP_READ_BITMAP ||
-                    next->opcode == OSD_OP_READ_CHAIN_BITMAP) && (flags & OP_FLUSH_BUFFER))
+            if (next->opcode == OSD_OP_SYNC && (!(flags & OP_IMMEDIATE_COMMIT) || enable_writeback) ||
+                next->opcode == OSD_OP_WRITE && (flags & OP_FLUSH_BUFFER) && !(next->flags & OP_FLUSH_BUFFER))
             {
                 next->prev_wait += inc;
                 assert(next->prev_wait >= 0);
@@ -245,13 +226,37 @@ void cluster_client_t::erase_op(cluster_op_t *op)
         op_queue_tail = op->prev;
     op->next = op->prev = NULL;
     if (flags & OP_FLUSH_BUFFER)
+    {
+        // Completed flushes change writeback buffer states,
+        // so the callback should be run before inc_wait()
+        // which may continue following SYNCs, but these SYNCs
+        // should know about the changed buffer state
+        // This is ugly but this is the way we do it
         std::function<void(cluster_op_t*)>(op->callback)(op);
-    if (!(flags & OP_IMMEDIATE_COMMIT))
+    }
+    if (!(flags & OP_IMMEDIATE_COMMIT) || enable_writeback)
+    {
         inc_wait(opcode, flags, next, -1);
-    // Call callback at the end to avoid inconsistencies in prev_wait
-    // if the callback adds more operations itself
+    }
     if (!(flags & OP_FLUSH_BUFFER))
+    {
+        // Call callback at the end to avoid inconsistencies in prev_wait
+        // if the callback adds more operations itself
         std::function<void(cluster_op_t*)>(op->callback)(op);
+    }
+    if (flags & OP_FLUSH_BUFFER)
+    {
+        int i = 0;
+        while (i < wb->writeback_overflow.size() && wb->writebacks_active < client_max_writeback_iodepth)
+        {
+            execute_internal(wb->writeback_overflow[i]);
+            i++;
+        }
+        if (i > 0)
+        {
+            wb->writeback_overflow.erase(wb->writeback_overflow.begin(), wb->writeback_overflow.begin()+i);
+        }
+    }
 }
 
 void cluster_client_t::continue_ops(bool up_retry)
@@ -295,6 +300,7 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & etcd_global_co
 {
     this->etcd_global_config = etcd_global_config;
     config = osd_messenger_t::merge_configs(cli_config, file_config, etcd_global_config, {});
+    // client_max_dirty_bytes/client_dirty_limit
     if (config.find("client_max_dirty_bytes") != config.end())
     {
         client_max_dirty_bytes = config["client_max_dirty_bytes"].uint64_value();
@@ -310,11 +316,34 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & etcd_global_co
     {
         client_max_dirty_bytes = DEFAULT_CLIENT_MAX_DIRTY_BYTES;
     }
+    // client_max_dirty_ops
     client_max_dirty_ops = config["client_max_dirty_ops"].uint64_value();
     if (!client_max_dirty_ops)
     {
         client_max_dirty_ops = DEFAULT_CLIENT_MAX_DIRTY_OPS;
     }
+    // client_enable_writeback
+    enable_writeback = json_is_true(config["client_enable_writeback"]) &&
+        json_is_true(config["client_writeback_allowed"]);
+    // client_max_buffered_bytes
+    client_max_buffered_bytes = config["client_max_buffered_bytes"].uint64_value();
+    if (!client_max_buffered_bytes)
+    {
+        client_max_buffered_bytes = DEFAULT_CLIENT_MAX_BUFFERED_BYTES;
+    }
+    // client_max_buffered_ops
+    client_max_buffered_ops = config["client_max_buffered_ops"].uint64_value();
+    if (!client_max_buffered_ops)
+    {
+        client_max_buffered_ops = DEFAULT_CLIENT_MAX_BUFFERED_OPS;
+    }
+    // client_max_writeback_iodepth
+    client_max_writeback_iodepth = config["client_max_writeback_iodepth"].uint64_value();
+    if (!client_max_writeback_iodepth)
+    {
+        client_max_writeback_iodepth = DEFAULT_CLIENT_MAX_WRITEBACK_IODEPTH;
+    }
+    // up_wait_retry_interval
     up_wait_retry_interval = config["up_wait_retry_interval"].uint64_value();
     if (!up_wait_retry_interval)
     {
@@ -374,6 +403,8 @@ void cluster_client_t::on_change_hook(std::map<std::string, etcd_kv_t> & changes
 
 bool cluster_client_t::get_immediate_commit(uint64_t inode)
 {
+    if (enable_writeback)
+        return false;
     pool_id_t pool_id = INODE_POOL(inode);
     if (!pool_id)
         return true;
@@ -408,6 +439,41 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
     }
 }
 
+bool cluster_client_t::flush()
+{
+    if (!ringloop)
+    {
+        if (wb->writeback_queue.size())
+        {
+            wb->start_writebacks(this, 0);
+            cluster_op_t *sync = new cluster_op_t;
+            sync->opcode = OSD_OP_SYNC;
+            sync->callback = [this](cluster_op_t *sync)
+            {
+                delete sync;
+            };
+            execute(sync);
+        }
+        return op_queue_head == NULL;
+    }
+    bool sync_done = false;
+    cluster_op_t *sync = new cluster_op_t;
+    sync->opcode = OSD_OP_SYNC;
+    sync->callback = [this, &sync_done](cluster_op_t *sync)
+    {
+        delete sync;
+        sync_done = true;
+    };
+    execute(sync);
+    while (!sync_done)
+    {
+        ringloop->loop();
+        if (!sync_done)
+            ringloop->wait();
+    }
+    return true;
+}
+
 /**
  * How writes are synced when immediate_commit is false
  *
@@ -428,6 +494,9 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
  * 3) if yes, send all SYNCs. otherwise, leave current SYNC as is.
  * 4) if any of them fail due to disconnected peers, repeat SYNC after repeating all writes
  * 5) if any of them fail due to other errors, fail the SYNC operation
+ *
+ * If writeback caching is turned on and writeback limit is not exhausted:
+ * data is just copied and the write is confirmed to the client.
  */
 void cluster_client_t::execute(cluster_op_t *op)
 {
@@ -443,67 +512,73 @@ void cluster_client_t::execute(cluster_op_t *op)
         offline_ops.push_back(op);
         return;
     }
+    op->flags = op->flags & OSD_OP_IGNORE_READONLY; // the only allowed flag
+    execute_internal(op);
+}
+
+void cluster_client_t::execute_internal(cluster_op_t *op)
+{
     op->cur_inode = op->inode;
     op->retval = 0;
-    op->flags = op->flags & OSD_OP_IGNORE_READONLY; // single allowed flag
-    if (op->opcode != OSD_OP_SYNC)
+    // check alignment, readonly flag and so on
+    if (!check_rw(op))
     {
-        pool_id_t pool_id = INODE_POOL(op->cur_inode);
-        if (!pool_id)
+        return;
+    }
+    if (op->opcode == OSD_OP_WRITE && enable_writeback && !(op->flags & OP_FLUSH_BUFFER) &&
+        !op->version /* FIXME no CAS writeback */)
+    {
+        if (wb->writebacks_active >= client_max_writeback_iodepth)
         {
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
+            // Writeback queue is full, postpone the operation
+            wb->writeback_overflow.push_back(op);
             return;
         }
-        auto pool_it = st_cli.pool_config.find(pool_id);
-        if (pool_it == st_cli.pool_config.end() || pool_it->second.real_pg_count == 0)
+        // Just copy and acknowledge the operation
+        wb->copy_write(op, CACHE_DIRTY);
+        while (wb->writeback_bytes + op->len > client_max_buffered_bytes || wb->writeback_queue_size > client_max_buffered_ops)
         {
-            // Pools are loaded, but this one is unknown
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
-            return;
-        }
-        // Check alignment
-        if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
-            op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
-        {
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
-            return;
-        }
-        if (pool_it->second.immediate_commit == IMMEDIATE_ALL)
-        {
-            op->flags |= OP_IMMEDIATE_COMMIT;
+            // Initiate some writeback (asynchronously)
+            wb->start_writebacks(this, 1);
         }
+        op->retval = op->len;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return;
     }
     if (op->opcode == OSD_OP_WRITE && !(op->flags & OP_IMMEDIATE_COMMIT))
     {
+        if (!(op->flags & OP_FLUSH_BUFFER))
+        {
+            wb->copy_write(op, CACHE_WRITTEN);
+        }
         if (dirty_bytes >= client_max_dirty_bytes || dirty_ops >= client_max_dirty_ops)
         {
             // Push an extra SYNC operation to flush previous writes
             cluster_op_t *sync_op = new cluster_op_t;
             sync_op->opcode = OSD_OP_SYNC;
+            sync_op->flags = OP_FLUSH_BUFFER;
             sync_op->callback = [](cluster_op_t* sync_op)
             {
                 delete sync_op;
             };
-            sync_op->prev = op_queue_tail;
-            if (op_queue_tail)
-            {
-                op_queue_tail->next = sync_op;
-                op_queue_tail = sync_op;
-            }
-            else
-                op_queue_tail = op_queue_head = sync_op;
-            dirty_bytes = 0;
-            dirty_ops = 0;
-            calc_wait(sync_op);
+            execute_internal(sync_op);
         }
         dirty_bytes += op->len;
         dirty_ops++;
     }
     else if (op->opcode == OSD_OP_SYNC)
     {
+        // Flush the whole write-back queue first
+        if (!(op->flags & OP_FLUSH_BUFFER) && wb->writeback_overflow.size() > 0)
+        {
+            // Writeback queue is full, postpone the operation
+            wb->writeback_overflow.push_back(op);
+            return;
+        }
+        if (wb->writeback_queue.size())
+        {
+            wb->start_writebacks(this, 0);
+        }
         dirty_bytes = 0;
         dirty_ops = 0;
     }
@@ -515,7 +590,7 @@ void cluster_client_t::execute(cluster_op_t *op)
     }
     else
         op_queue_tail = op_queue_head = op;
-    if (!(op->flags & OP_IMMEDIATE_COMMIT))
+    if (!(op->flags & OP_IMMEDIATE_COMMIT) || enable_writeback)
         calc_wait(op);
     else
     {
@@ -526,6 +601,52 @@ void cluster_client_t::execute(cluster_op_t *op)
     }
 }
 
+bool cluster_client_t::check_rw(cluster_op_t *op)
+{
+    if (op->opcode == OSD_OP_SYNC)
+    {
+        return true;
+    }
+    pool_id_t pool_id = INODE_POOL(op->cur_inode);
+    if (!pool_id)
+    {
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    auto pool_it = st_cli.pool_config.find(pool_id);
+    if (pool_it == st_cli.pool_config.end() || pool_it->second.real_pg_count == 0)
+    {
+        // Pools are loaded, but this one is unknown
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    // Check alignment
+    if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
+        op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
+    {
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    if (pool_it->second.immediate_commit == IMMEDIATE_ALL)
+    {
+        op->flags |= OP_IMMEDIATE_COMMIT;
+    }
+    if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE) && !(op->flags & OSD_OP_IGNORE_READONLY))
+    {
+        auto ino_it = st_cli.inode_config.find(op->inode);
+        if (ino_it != st_cli.inode_config.end() && ino_it->second.readonly)
+        {
+            op->retval = -EROFS;
+            std::function<void(cluster_op_t*)>(op->callback)(op);
+            return false;
+        }
+    }
+    return true;
+}
+
 void cluster_client_t::execute_raw(osd_num_t osd_num, osd_op_t *op)
 {
     auto fd_it = msgr.osd_peer_fds.find(osd_num);
@@ -543,114 +664,6 @@ void cluster_client_t::execute_raw(osd_num_t osd_num, osd_op_t *op)
     }
 }
 
-void cluster_client_t::copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers)
-{
-    // Save operation for replay when one of PGs goes out of sync
-    // (primary OSD drops our connection in this case)
-    auto dirty_it = dirty_buffers.lower_bound((object_id){
-        .inode = op->inode,
-        .stripe = op->offset,
-    });
-    while (dirty_it != dirty_buffers.begin())
-    {
-        dirty_it--;
-        if (dirty_it->first.inode != op->inode ||
-            (dirty_it->first.stripe + dirty_it->second.len) <= op->offset)
-        {
-            dirty_it++;
-            break;
-        }
-    }
-    uint64_t pos = op->offset, len = op->len, iov_idx = 0, iov_pos = 0;
-    while (len > 0)
-    {
-        uint64_t new_len = 0;
-        if (dirty_it == dirty_buffers.end())
-        {
-            new_len = len;
-        }
-        else if (dirty_it->first.inode != op->inode || dirty_it->first.stripe > pos)
-        {
-            new_len = dirty_it->first.stripe - pos;
-            if (new_len > len)
-            {
-                new_len = len;
-            }
-        }
-        if (new_len > 0)
-        {
-            dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
-                .inode = op->inode,
-                .stripe = pos,
-            }, (cluster_buffer_t){
-                .buf = malloc_or_die(new_len),
-                .len = new_len,
-            });
-        }
-        // FIXME: Split big buffers into smaller ones on overwrites. But this will require refcounting
-        dirty_it->second.state = CACHE_DIRTY;
-        uint64_t cur_len = (dirty_it->first.stripe + dirty_it->second.len - pos);
-        if (cur_len > len)
-        {
-            cur_len = len;
-        }
-        while (cur_len > 0 && iov_idx < op->iov.count)
-        {
-            unsigned iov_len = (op->iov.buf[iov_idx].iov_len - iov_pos);
-            if (iov_len <= cur_len)
-            {
-                memcpy((uint8_t*)dirty_it->second.buf + pos - dirty_it->first.stripe,
-                    (uint8_t*)op->iov.buf[iov_idx].iov_base + iov_pos, iov_len);
-                pos += iov_len;
-                len -= iov_len;
-                cur_len -= iov_len;
-                iov_pos = 0;
-                iov_idx++;
-            }
-            else
-            {
-                memcpy((uint8_t*)dirty_it->second.buf + pos - dirty_it->first.stripe,
-                    (uint8_t*)op->iov.buf[iov_idx].iov_base + iov_pos, cur_len);
-                pos += cur_len;
-                len -= cur_len;
-                iov_pos += cur_len;
-                cur_len = 0;
-            }
-        }
-        dirty_it++;
-    }
-}
-
-void cluster_client_t::flush_buffer(const object_id & oid, cluster_buffer_t *wr)
-{
-    wr->state = CACHE_REPEATING;
-    cluster_op_t *op = new cluster_op_t;
-    op->flags = OSD_OP_IGNORE_READONLY|OP_FLUSH_BUFFER;
-    op->opcode = OSD_OP_WRITE;
-    op->cur_inode = op->inode = oid.inode;
-    op->offset = oid.stripe;
-    op->len = wr->len;
-    op->iov.push_back(wr->buf, wr->len);
-    op->callback = [wr](cluster_op_t* op)
-    {
-        if (wr->state == CACHE_REPEATING)
-        {
-            wr->state = CACHE_DIRTY;
-        }
-        delete op;
-    };
-    op->next = op_queue_head;
-    if (op_queue_head)
-    {
-        op_queue_head->prev = op;
-        op_queue_head = op;
-    }
-    else
-        op_queue_tail = op_queue_head = op;
-    inc_wait(op->opcode, op->flags, op->next, 1);
-    continue_rw(op);
-}
-
 int cluster_client_t::continue_rw(cluster_op_t *op)
 {
     if (op->state == 0)
@@ -659,27 +672,7 @@ int cluster_client_t::continue_rw(cluster_op_t *op)
         goto resume_1;
     else if (op->state == 2)
         goto resume_2;
-    else if (op->state == 3)
-        goto resume_3;
 resume_0:
-    if (op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE)
-    {
-        if (!(op->flags & OSD_OP_IGNORE_READONLY))
-        {
-            auto ino_it = st_cli.inode_config.find(op->inode);
-            if (ino_it != st_cli.inode_config.end() && ino_it->second.readonly)
-            {
-                op->retval = -EINVAL;
-                erase_op(op);
-                return 1;
-            }
-        }
-        if (op->opcode == OSD_OP_WRITE && !(op->flags & OP_IMMEDIATE_COMMIT) && !(op->flags & OP_FLUSH_BUFFER))
-        {
-            copy_write(op, dirty_buffers);
-        }
-    }
-resume_1:
     // Slice the operation into parts
     slice_rw(op);
     op->needs_reslice = false;
@@ -690,9 +683,9 @@ resume_1:
         erase_op(op);
         return 1;
     }
-resume_2:
+resume_1:
     // Send unsent parts, if they're not subject to change
-    op->state = 3;
+    op->state = 2;
     if (op->needs_reslice)
     {
         for (int i = 0; i < op->parts.size(); i++)
@@ -702,7 +695,7 @@ resume_2:
                 op->retval = -EPIPE;
             }
         }
-        goto resume_3;
+        goto resume_2;
     }
     for (int i = 0; i < op->parts.size(); i++)
     {
@@ -723,18 +716,18 @@ resume_2:
                         });
                     }
                 }
-                op->state = 2;
+                op->state = 1;
             }
         }
     }
-    if (op->state == 2)
+    if (op->state == 1)
     {
         return 0;
     }
-resume_3:
+resume_2:
     if (op->inflight_count > 0)
     {
-        op->state = 3;
+        op->state = 2;
         return 0;
     }
     if (op->done_count >= op->parts.size())
@@ -762,7 +755,7 @@ resume_3:
                 op->cur_inode = ino_it->second.parent_id;
                 op->parts.clear();
                 op->done_count = 0;
-                goto resume_1;
+                goto resume_0;
             }
         }
         op->retval = op->len;
@@ -774,7 +767,8 @@ resume_3:
         erase_op(op);
         return 1;
     }
-    else if (op->retval != 0 && op->retval != -EPIPE && op->retval != -EIO && op->retval != -ENOSPC)
+    else if (op->retval != 0 && !(op->flags & OP_FLUSH_BUFFER) &&
+        op->retval != -EPIPE && op->retval != -EIO && op->retval != -ENOSPC)
     {
         // Fatal error (neither -EPIPE, -EIO nor -ENOSPC)
         // FIXME: Add a parameter to allow to not wait for EIOs (incomplete or corrupted objects) to heal
@@ -789,7 +783,7 @@ resume_3:
         {
             op->parts.clear();
             op->done_count = 0;
-            goto resume_1;
+            goto resume_0;
         }
         else
         {
@@ -800,7 +794,7 @@ resume_3:
                     op->parts[i].flags = PART_RETRY;
                 }
             }
-            goto resume_2;
+            goto resume_1;
         }
     }
     return 0;
@@ -874,6 +868,11 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
     int iov_idx = 0;
     size_t iov_pos = 0;
     int i = 0;
+    // We also have to return reads from CACHE_REPEATING buffers - they are not
+    // guaranteed to be present on target OSDs at the moment of repeating
+    // And we're also free to return data from other cached buffers just
+    // because it's faster
+    bool dirty_copied = wb->read_from_cache(op, pool_cfg.bitmap_granularity);
     for (uint64_t stripe = first_stripe; stripe <= last_stripe; stripe += pg_block_size)
     {
         pg_num_t pg_num = (stripe/pool_cfg.pg_stripe_size) % pool_cfg.real_pg_count + 1; // like map_to_pg()
@@ -881,7 +880,8 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
         uint64_t end = (op->offset + op->len) > (stripe + pg_block_size)
             ? (stripe + pg_block_size) : (op->offset + op->len);
         op->parts[i].iov.reset();
-        if (op->cur_inode != op->inode)
+        op->parts[i].flags = 0;
+        if (op->cur_inode != op->inode || op->opcode == OSD_OP_READ && dirty_copied)
         {
             // Read remaining parts from upper layers
             uint64_t prev = begin, cur = begin;
@@ -918,7 +918,10 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             else
                 add_iov(cur-prev, skip_prev, op, iov_idx, iov_pos, op->parts[i].iov, scrap_buffer, scrap_buffer_size);
             if (end == begin)
+            {
                 op->done_count++;
+                op->parts[i].flags = PART_DONE;
+            }
         }
         else if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_READ_CHAIN_BITMAP && op->opcode != OSD_OP_DELETE)
         {
@@ -930,7 +933,6 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             op->opcode == OSD_OP_DELETE ? 0 : (uint32_t)(end - begin);
         op->parts[i].pg_num = pg_num;
         op->parts[i].osd_num = 0;
-        op->parts[i].flags = 0;
         i++;
     }
 }
@@ -1042,13 +1044,7 @@ int cluster_client_t::continue_sync(cluster_op_t *op)
             do_it++;
     }
     // Post sync to affected OSDs
-    for (auto & prev_op: dirty_buffers)
-    {
-        if (prev_op.second.state == CACHE_DIRTY)
-        {
-            prev_op.second.state = CACHE_FLUSHING;
-        }
-    }
+    wb->fsync_start();
     op->parts.resize(dirty_osds.size());
     op->retval = 0;
     {
@@ -1073,13 +1069,7 @@ resume_1:
     }
     if (op->retval != 0)
     {
-        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); uw_it++)
-        {
-            if (uw_it->second.state == CACHE_FLUSHING)
-            {
-                uw_it->second.state = CACHE_DIRTY;
-            }
-        }
+        wb->fsync_error();
         if (op->retval == -EPIPE || op->retval == -EIO || op->retval == -ENOSPC)
         {
             // Retry later
@@ -1093,16 +1083,7 @@ resume_1:
     }
     else
     {
-        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); )
-        {
-            if (uw_it->second.state == CACHE_FLUSHING)
-            {
-                free(uw_it->second.buf);
-                dirty_buffers.erase(uw_it++);
-            }
-            else
-                uw_it++;
-        }
+        wb->fsync_ok();
     }
     erase_op(op);
     return 1;

src/cluster_client.h

@@ -8,6 +8,9 @@
 
 #define DEFAULT_CLIENT_MAX_DIRTY_BYTES 32*1024*1024
 #define DEFAULT_CLIENT_MAX_DIRTY_OPS 1024
+#define DEFAULT_CLIENT_MAX_BUFFERED_BYTES 32*1024*1024
+#define DEFAULT_CLIENT_MAX_BUFFERED_OPS 1024
+#define DEFAULT_CLIENT_MAX_WRITEBACK_IODEPTH 256
 #define INODE_LIST_DONE 1
 #define INODE_LIST_HAS_UNSTABLE 2
 #define OSD_OP_READ_BITMAP OSD_OP_SEC_READ_BMP
@@ -64,17 +67,12 @@ protected:
     cluster_op_t *prev = NULL, *next = NULL;
     int prev_wait = 0;
     friend class cluster_client_t;
-};
-
-struct cluster_buffer_t
-{
-    void *buf;
-    uint64_t len;
-    int state;
+    friend class writeback_cache_t;
 };
 
 struct inode_list_t;
 struct inode_list_osd_t;
+class writeback_cache_t;
 
 // FIXME: Split into public and private interfaces
 class cluster_client_t
@@ -83,16 +81,23 @@ class cluster_client_t
     ring_loop_t *ringloop;
 
     std::map<pool_id_t, uint64_t> pg_counts;
-    // FIXME: Implement inmemory_commit mode. Note that it requires to return overlapping reads from memory.
+    // client_max_dirty_* is actually "max unsynced", for the case when immediate_commit is off
     uint64_t client_max_dirty_bytes = 0;
     uint64_t client_max_dirty_ops = 0;
+    // writeback improves (1) small consecutive writes and (2) Q1 writes without fsync
+    bool enable_writeback = false;
+    // client_max_buffered_* is the real "dirty limit" - maximum amount of writes buffered in memory
+    uint64_t client_max_buffered_bytes = 0;
+    uint64_t client_max_buffered_ops = 0;
+    uint64_t client_max_writeback_iodepth = 0;
+
     int log_level;
     int up_wait_retry_interval = 500; // ms
 
     int retry_timeout_id = 0;
     std::vector<cluster_op_t*> offline_ops;
     cluster_op_t *op_queue_head = NULL, *op_queue_tail = NULL;
-    std::map<object_id, cluster_buffer_t> dirty_buffers;
+    writeback_cache_t *wb = NULL;
     std::set<osd_num_t> dirty_osds;
     uint64_t dirty_bytes = 0, dirty_ops = 0;
 
@@ -122,10 +127,10 @@ public:
     void execute_raw(osd_num_t osd_num, osd_op_t *op);
     bool is_ready();
     void on_ready(std::function<void(void)> fn);
+    bool flush();
 
     bool get_immediate_commit(uint64_t inode);
 
-    static void copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers);
     void continue_ops(bool up_retry = false);
     inode_list_t *list_inode_start(inode_t inode,
         std::function<void(inode_list_t* lst, std::set<object_id>&& objects, pg_num_t pg_num, osd_num_t primary_osd, int status)> callback);
@@ -138,12 +143,14 @@ public:
 
 protected:
     bool affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd);
-    void flush_buffer(const object_id & oid, cluster_buffer_t *wr);
     void on_load_config_hook(json11::Json::object & config);
     void on_load_pgs_hook(bool success);
     void on_change_hook(std::map<std::string, etcd_kv_t> & changes);
     void on_change_osd_state_hook(uint64_t peer_osd);
+    void execute_internal(cluster_op_t *op);
+    void unshift_op(cluster_op_t *op);
     int continue_rw(cluster_op_t *op);
+    bool check_rw(cluster_op_t *op);
     void slice_rw(cluster_op_t *op);
     bool try_send(cluster_op_t *op, int i);
     int continue_sync(cluster_op_t *op);
@@ -157,4 +164,6 @@ protected:
     void continue_listing(inode_list_t *lst);
     void send_list(inode_list_osd_t *cur_list);
     void continue_raw_ops(osd_num_t peer_osd);
+
+    friend class writeback_cache_t;
 };

src/cluster_client_impl.h

@@ -0,0 +1,57 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#pragma once
+
+#include "cluster_client.h"
+
+#define SCRAP_BUFFER_SIZE 4*1024*1024
+#define PART_SENT 1
+#define PART_DONE 2
+#define PART_ERROR 4
+#define PART_RETRY 8
+#define CACHE_DIRTY 1
+#define CACHE_WRITTEN 2
+#define CACHE_FLUSHING 3
+#define CACHE_REPEATING 4
+#define OP_FLUSH_BUFFER 0x02
+#define OP_IMMEDIATE_COMMIT 0x04
+
+struct cluster_buffer_t
+{
+    uint8_t *buf;
+    uint64_t len;
+    int state;
+    uint64_t flush_id;
+    uint64_t *refcnt;
+};
+
+typedef std::map<object_id, cluster_buffer_t>::iterator dirty_buf_it_t;
+
+class writeback_cache_t
+{
+public:
+    uint64_t writeback_bytes = 0;
+    int writeback_queue_size = 0;
+    int writebacks_active = 0;
+    uint64_t last_flush_id = 0;
+
+    std::map<object_id, cluster_buffer_t> dirty_buffers;
+    std::vector<cluster_op_t*> writeback_overflow;
+    std::vector<object_id> writeback_queue;
+    std::multimap<uint64_t, uint64_t*> flushed_buffers; // flush_id => refcnt
+
+    ~writeback_cache_t();
+    dirty_buf_it_t find_dirty(uint64_t inode, uint64_t offset);
+    bool is_left_merged(dirty_buf_it_t dirty_it);
+    bool is_right_merged(dirty_buf_it_t dirty_it);
+    bool is_merged(const dirty_buf_it_t & dirty_it);
+    void copy_write(cluster_op_t *op, int state);
+    int repeat_ops_for(cluster_client_t *cli, osd_num_t peer_osd);
+    void start_writebacks(cluster_client_t *cli, int count);
+    bool read_from_cache(cluster_op_t *op, uint32_t bitmap_granularity);
+    void flush_buffers(cluster_client_t *cli, dirty_buf_it_t from_it, dirty_buf_it_t to_it);
+    void fsync_start();
+    void fsync_error();
+    void fsync_ok();
+};

src/cluster_client_wb.cpp

@@ -0,0 +1,498 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <cassert>
+
+#include "cluster_client_impl.h"
+
+writeback_cache_t::~writeback_cache_t()
+{
+    for (auto & bp: dirty_buffers)
+    {
+        if (!--(*bp.second.refcnt))
+        {
+            free(bp.second.refcnt); // refcnt is allocated with the buffer
+        }
+    }
+    dirty_buffers.clear();
+}
+
+dirty_buf_it_t writeback_cache_t::find_dirty(uint64_t inode, uint64_t offset)
+{
+    auto dirty_it = dirty_buffers.lower_bound((object_id){
+        .inode = inode,
+        .stripe = offset,
+    });
+    while (dirty_it != dirty_buffers.begin())
+    {
+        dirty_it--;
+        if (dirty_it->first.inode != inode ||
+            (dirty_it->first.stripe + dirty_it->second.len) <= offset)
+        {
+            dirty_it++;
+            break;
+        }
+    }
+    return dirty_it;
+}
+
+bool writeback_cache_t::is_left_merged(dirty_buf_it_t dirty_it)
+{
+    if (dirty_it != dirty_buffers.begin())
+    {
+        auto prev_it = dirty_it;
+        prev_it--;
+        if (prev_it->first.inode == dirty_it->first.inode &&
+            prev_it->first.stripe+prev_it->second.len == dirty_it->first.stripe &&
+            prev_it->second.state == CACHE_DIRTY)
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+bool writeback_cache_t::is_right_merged(dirty_buf_it_t dirty_it)
+{
+    auto next_it = dirty_it;
+    next_it++;
+    if (next_it != dirty_buffers.end() &&
+        next_it->first.inode == dirty_it->first.inode &&
+        next_it->first.stripe == dirty_it->first.stripe+dirty_it->second.len &&
+        next_it->second.state == CACHE_DIRTY)
+    {
+        return true;
+    }
+    return false;
+}
+
+bool writeback_cache_t::is_merged(const dirty_buf_it_t & dirty_it)
+{
+    return is_left_merged(dirty_it) || is_right_merged(dirty_it);
+}
+
+void writeback_cache_t::copy_write(cluster_op_t *op, int state)
+{
+    // Save operation for replay when one of PGs goes out of sync
+    // (primary OSD drops our connection in this case)
+    // ...or just save it for writeback if write buffering is enabled
+    if (op->len == 0)
+    {
+        return;
+    }
+    auto dirty_it = find_dirty(op->inode, op->offset);
+    auto new_end = op->offset + op->len;
+    while (dirty_it != dirty_buffers.end() &&
+        dirty_it->first.inode == op->inode &&
+        dirty_it->first.stripe < op->offset+op->len)
+    {
+        assert(dirty_it->first.stripe + dirty_it->second.len > op->offset);
+        // Remove overlapping part(s) of buffers
+        auto old_end = dirty_it->first.stripe + dirty_it->second.len;
+        if (dirty_it->first.stripe < op->offset)
+        {
+            if (old_end > new_end)
+            {
+                // Split into end and start
+                dirty_it->second.len = op->offset - dirty_it->first.stripe;
+                dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+                    .inode = op->inode,
+                    .stripe = new_end,
+                }, (cluster_buffer_t){
+                    .buf = dirty_it->second.buf + new_end - dirty_it->first.stripe,
+                    .len = old_end - new_end,
+                    .state = dirty_it->second.state,
+                    .flush_id = dirty_it->second.flush_id,
+                    .refcnt = dirty_it->second.refcnt,
+                });
+                (*dirty_it->second.refcnt)++;
+                if (dirty_it->second.state == CACHE_DIRTY)
+                {
+                    writeback_bytes -= op->len;
+                    writeback_queue_size++;
+                }
+                break;
+            }
+            else
+            {
+                // Only leave the beginning
+                if (dirty_it->second.state == CACHE_DIRTY)
+                {
+                    writeback_bytes -= old_end - op->offset;
+                    if (is_left_merged(dirty_it) && !is_right_merged(dirty_it))
+                    {
+                        writeback_queue_size++;
+                    }
+                }
+                dirty_it->second.len = op->offset - dirty_it->first.stripe;
+                dirty_it++;
+            }
+        }
+        else if (old_end > new_end)
+        {
+            // Only leave the end
+            if (dirty_it->second.state == CACHE_DIRTY)
+            {
+                writeback_bytes -= new_end - dirty_it->first.stripe;
+                if (!is_left_merged(dirty_it) && is_right_merged(dirty_it))
+                {
+                    writeback_queue_size++;
+                }
+            }
+            auto new_dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+                .inode = op->inode,
+                .stripe = new_end,
+            }, (cluster_buffer_t){
+                .buf = dirty_it->second.buf + new_end - dirty_it->first.stripe,
+                .len = old_end - new_end,
+                .state = dirty_it->second.state,
+                .flush_id = dirty_it->second.flush_id,
+                .refcnt = dirty_it->second.refcnt,
+            });
+            dirty_buffers.erase(dirty_it);
+            dirty_it = new_dirty_it;
+            break;
+        }
+        else
+        {
+            // Remove the whole buffer
+            if (dirty_it->second.state == CACHE_DIRTY && !is_merged(dirty_it))
+            {
+                writeback_bytes -= dirty_it->second.len;
+                assert(writeback_queue_size > 0);
+                writeback_queue_size--;
+            }
+            if (!--(*dirty_it->second.refcnt))
+            {
+                free(dirty_it->second.refcnt);
+            }
+            dirty_buffers.erase(dirty_it++);
+        }
+    }
+    // Overlapping buffers are removed, just insert the new one
+    uint64_t *refcnt = (uint64_t*)malloc_or_die(sizeof(uint64_t) + op->len);
+    uint8_t *buf = (uint8_t*)refcnt + sizeof(uint64_t);
+    *refcnt = 1;
+    dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+        .inode = op->inode,
+        .stripe = op->offset,
+    }, (cluster_buffer_t){
+        .buf = buf,
+        .len = op->len,
+        .state = state,
+        .refcnt = refcnt,
+    });
+    if (state == CACHE_DIRTY)
+    {
+        writeback_bytes += op->len;
+        // Track consecutive write-back operations
+        if (!is_merged(dirty_it))
+        {
+            // <writeback_queue> is OK to contain more than actual number of consecutive
+            // requests as long as it doesn't miss anything. But <writeback_queue_size>
+            // is always calculated correctly.
+            writeback_queue_size++;
+            writeback_queue.push_back((object_id){
+                .inode = op->inode,
+                .stripe = op->offset,
+            });
+        }
+    }
+    uint64_t pos = 0, len = op->len, iov_idx = 0;
+    while (len > 0 && iov_idx < op->iov.count)
+    {
+        auto & iov = op->iov.buf[iov_idx];
+        memcpy(buf + pos, iov.iov_base, iov.iov_len);
+        pos += iov.iov_len;
+        iov_idx++;
+    }
+}
+
+int writeback_cache_t::repeat_ops_for(cluster_client_t *cli, osd_num_t peer_osd)
+{
+    int repeated = 0;
+    if (dirty_buffers.size())
+    {
+        // peer_osd just dropped connection
+        // determine WHICH dirty_buffers are now obsolete and repeat them
+        for (auto wr_it = dirty_buffers.begin(), flush_it = wr_it, last_it = wr_it; ; )
+        {
+            bool end = wr_it == dirty_buffers.end();
+            bool flush_this = !end && wr_it->second.state != CACHE_REPEATING &&
+                cli->affects_osd(wr_it->first.inode, wr_it->first.stripe, wr_it->second.len, peer_osd);
+            if (flush_it != wr_it && (end || !flush_this ||
+                wr_it->first.inode != flush_it->first.inode ||
+                wr_it->first.stripe != last_it->first.stripe+last_it->second.len))
+            {
+                repeated++;
+                flush_buffers(cli, flush_it, wr_it);
+                flush_it = wr_it;
+            }
+            if (end)
+                break;
+            last_it = wr_it;
+            wr_it++;
+            if (!flush_this)
+                flush_it = wr_it;
+        }
+    }
+    return repeated;
+}
+
+void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from_it, dirty_buf_it_t to_it)
+{
+    auto prev_it = to_it;
+    prev_it--;
+    bool is_writeback = from_it->second.state == CACHE_DIRTY;
+    cluster_op_t *op = new cluster_op_t;
+    op->flags = OSD_OP_IGNORE_READONLY|OP_FLUSH_BUFFER;
+    op->opcode = OSD_OP_WRITE;
+    op->cur_inode = op->inode = from_it->first.inode;
+    op->offset = from_it->first.stripe;
+    op->len = prev_it->first.stripe + prev_it->second.len - from_it->first.stripe;
+    uint32_t calc_len = 0;
+    uint64_t flush_id = ++last_flush_id;
+    for (auto it = from_it; it != to_it; it++)
+    {
+        it->second.state = CACHE_REPEATING;
+        it->second.flush_id = flush_id;
+        (*it->second.refcnt)++;
+        flushed_buffers.emplace(flush_id, it->second.refcnt);
+        op->iov.push_back(it->second.buf, it->second.len);
+        calc_len += it->second.len;
+    }
+    assert(calc_len == op->len);
+    writebacks_active++;
+    op->callback = [this, cli, flush_id](cluster_op_t* op)
+    {
+        // Buffer flushes should be always retried, regardless of the error,
+        // so they should never result in an error here
+        assert(op->retval == op->len);
+        for (auto fl_it = flushed_buffers.find(flush_id);
+            fl_it != flushed_buffers.end() && fl_it->first == flush_id; )
+        {
+            if (!--(*fl_it->second)) // refcnt
+            {
+                free(fl_it->second);
+            }
+            flushed_buffers.erase(fl_it++);
+        }
+        for (auto dirty_it = find_dirty(op->inode, op->offset);
+            dirty_it != dirty_buffers.end() && dirty_it->first.inode == op->inode &&
+            dirty_it->first.stripe < op->offset+op->len; dirty_it++)
+        {
+            if (dirty_it->second.flush_id == flush_id && dirty_it->second.state == CACHE_REPEATING)
+            {
+                dirty_it->second.flush_id = 0;
+                dirty_it->second.state = CACHE_WRITTEN;
+            }
+        }
+        delete op;
+        writebacks_active--;
+        // We can't call execute_internal because it affects an invalid copy of the list here
+        // (erase_op remembers `next` after writeback callback)
+    };
+    if (is_writeback)
+    {
+        cli->execute_internal(op);
+    }
+    else
+    {
+        // Insert repeated flushes into the beginning
+        cli->unshift_op(op);
+        cli->continue_rw(op);
+    }
+}
+
+void writeback_cache_t::start_writebacks(cluster_client_t *cli, int count)
+{
+    if (!writeback_queue.size())
+    {
+        return;
+    }
+    std::vector<object_id> queue_copy;
+    queue_copy.swap(writeback_queue);
+    int started = 0, i = 0;
+    for (i = 0; i < queue_copy.size() && (!count || started < count); i++)
+    {
+        object_id & req = queue_copy[i];
+        auto dirty_it = find_dirty(req.inode, req.stripe);
+        if (dirty_it == dirty_buffers.end() ||
+            dirty_it->first.inode != req.inode ||
+            dirty_it->second.state != CACHE_DIRTY)
+        {
+            continue;
+        }
+        auto from_it = dirty_it;
+        uint64_t off = dirty_it->first.stripe;
+        while (from_it != dirty_buffers.begin())
+        {
+            from_it--;
+            if (from_it->second.state != CACHE_DIRTY ||
+                from_it->first.inode != req.inode ||
+                from_it->first.stripe+from_it->second.len != off)
+            {
+                from_it++;
+                break;
+            }
+            off = from_it->first.stripe;
+        }
+        off = dirty_it->first.stripe + dirty_it->second.len;
+        auto to_it = dirty_it;
+        to_it++;
+        while (to_it != dirty_buffers.end())
+        {
+            if (to_it->second.state != CACHE_DIRTY ||
+                to_it->first.inode != req.inode ||
+                to_it->first.stripe != off)
+            {
+                break;
+            }
+            off = to_it->first.stripe + to_it->second.len;
+            to_it++;
+        }
+        started++;
+        assert(writeback_queue_size > 0);
+        writeback_queue_size--;
+        writeback_bytes -= off - from_it->first.stripe;
+        flush_buffers(cli, from_it, to_it);
+    }
+    queue_copy.erase(queue_copy.begin(), queue_copy.begin()+i);
+    if (writeback_queue.size())
+    {
+        queue_copy.insert(queue_copy.end(), writeback_queue.begin(), writeback_queue.end());
+    }
+    queue_copy.swap(writeback_queue);
+}
+
+static void copy_to_op(cluster_op_t *op, uint64_t offset, uint8_t *buf, uint64_t len, uint32_t bitmap_granularity)
+{
+    if (op->opcode == OSD_OP_READ)
+    {
+        // Not OSD_OP_READ_BITMAP or OSD_OP_READ_CHAIN_BITMAP
+        int iov_idx = 0;
+        uint64_t cur_offset = op->offset;
+        while (iov_idx < op->iov.count && cur_offset+op->iov.buf[iov_idx].iov_len <= offset)
+        {
+            cur_offset += op->iov.buf[iov_idx].iov_len;
+            iov_idx++;
+        }
+        while (iov_idx < op->iov.count && cur_offset < offset+len)
+        {
+            auto & v = op->iov.buf[iov_idx];
+            auto begin = (cur_offset < offset ? offset : cur_offset);
+            auto end = (cur_offset+v.iov_len > offset+len ? offset+len : cur_offset+v.iov_len);
+            memcpy(
+                v.iov_base + begin - cur_offset,
+                buf + (cur_offset <= offset ? 0 : cur_offset-offset),
+                end - begin
+            );
+            cur_offset += v.iov_len;
+            iov_idx++;
+        }
+    }
+    // Set bitmap bits
+    int start_bit = (offset-op->offset)/bitmap_granularity;
+    int end_bit = (offset-op->offset+len)/bitmap_granularity;
+    for (int bit = start_bit; bit < end_bit;)
+    {
+        if (!(bit%8) && bit <= end_bit-8)
+        {
+            ((uint8_t*)op->bitmap_buf)[bit/8] = 0xFF;
+            bit += 8;
+        }
+        else
+        {
+            ((uint8_t*)op->bitmap_buf)[bit/8] |= (1 << (bit%8));
+            bit++;
+        }
+    }
+}
+
+bool writeback_cache_t::read_from_cache(cluster_op_t *op, uint32_t bitmap_granularity)
+{
+    bool dirty_copied = false;
+    if (dirty_buffers.size() && (op->opcode == OSD_OP_READ ||
+        op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP))
+    {
+        // We also have to return reads from CACHE_REPEATING buffers - they are not
+        // guaranteed to be present on target OSDs at the moment of repeating
+        // And we're also free to return data from other cached buffers just
+        // because it's faster
+        auto dirty_it = find_dirty(op->cur_inode, op->offset);
+        while (dirty_it != dirty_buffers.end() && dirty_it->first.inode == op->cur_inode &&
+            dirty_it->first.stripe < op->offset+op->len)
+        {
+            uint64_t begin = dirty_it->first.stripe, end = dirty_it->first.stripe + dirty_it->second.len;
+            if (begin < op->offset)
+                begin = op->offset;
+            if (end > op->offset+op->len)
+                end = op->offset+op->len;
+            bool skip_prev = true;
+            uint64_t cur = begin, prev = begin;
+            while (cur < end)
+            {
+                unsigned bmp_loc = (cur - op->offset)/bitmap_granularity;
+                bool skip = (((*((uint8_t*)op->bitmap_buf + bmp_loc/8)) >> (bmp_loc%8)) & 0x1);
+                if (skip_prev != skip)
+                {
+                    if (cur > prev && !skip)
+                    {
+                        // Copy data
+                        dirty_copied = true;
+                        copy_to_op(op, prev, dirty_it->second.buf + prev - dirty_it->first.stripe, cur-prev, bitmap_granularity);
+                    }
+                    skip_prev = skip;
+                    prev = cur;
+                }
+                cur += bitmap_granularity;
+            }
+            assert(cur > prev);
+            if (!skip_prev)
+            {
+                // Copy data
+                dirty_copied = true;
+                copy_to_op(op, prev, dirty_it->second.buf + prev - dirty_it->first.stripe, cur-prev, bitmap_granularity);
+            }
+            dirty_it++;
+        }
+    }
+    return dirty_copied;
+}
+
+void writeback_cache_t::fsync_start()
+{
+    for (auto & prev_op: dirty_buffers)
+    {
+        if (prev_op.second.state == CACHE_WRITTEN)
+        {
+            prev_op.second.state = CACHE_FLUSHING;
+        }
+    }
+}
+
+void writeback_cache_t::fsync_error()
+{
+    for (auto & prev_op: dirty_buffers)
+    {
+        if (prev_op.second.state == CACHE_FLUSHING)
+        {
+            prev_op.second.state = CACHE_WRITTEN;
+        }
+    }
+}
+
+void writeback_cache_t::fsync_ok()
+{
+    for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); )
+    {
+        if (uw_it->second.state == CACHE_FLUSHING)
+        {
+            if (!--(*uw_it->second.refcnt))
+                free(uw_it->second.refcnt);
+            dirty_buffers.erase(uw_it++);
+        }
+        else
+            uw_it++;
+    }
+}

src/disk_tool.cpp

@@ -74,7 +74,7 @@ static const char *help_text =
     "  If it doesn't succeed it issues a warning in the system log.\n"
     "  \n"
     "  You can also pass other OSD options here as arguments and they'll be persisted\n"
-    "  in the superblock: cached_read_data, cached_read_meta, cached_read_journal,\n"
+    "  in the superblock: data_io, meta_io, journal_io,\n"
     "  inmemory_metadata, inmemory_journal, max_write_iodepth,\n"
     "  min_flusher_count, max_flusher_count, journal_sector_buffer_count,\n"
     "  journal_no_same_sector_overwrites, throttle_small_writes, throttle_target_iops,\n"

src/disk_tool_prepare.cpp

@@ -8,9 +8,9 @@
 int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_hdd)
 {
     static const char *allow_additional_params[] = {
-        "cached_read_data",
-        "cached_read_meta",
-        "cached_read_journal",
+        "data_io",
+        "meta_io",
+        "journal_io",
         "max_write_iodepth",
         "max_write_iodepth",
         "min_flusher_count",
@@ -119,7 +119,7 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     try
     {
         dsk.parse_config(options);
-        dsk.cached_read_data = dsk.cached_read_meta = dsk.cached_read_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -151,7 +151,7 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
         for (int i = 0; i < sizeof(allow_additional_params)/sizeof(allow_additional_params[0]); i++)
         {
             auto it = options.find(allow_additional_params[i]);
-            if (it != options.end())
+            if (it != options.end() && it->second != "")
             {
                 sb[it->first] = it->second;
             }
@@ -483,7 +483,7 @@ int disk_tool_t::get_meta_partition(std::vector<vitastor_dev_info_t> & ssds, std
     {
         blockstore_disk_t dsk;
         dsk.parse_config(options);
-        dsk.cached_read_data = dsk.cached_read_meta = dsk.cached_read_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -626,7 +626,7 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
                     }
                 }
                 // Treat all disks as SSDs if not in the hybrid mode
-                prepare_one(options, hybrid && dev.is_hdd ? 1 : 0);
+                prepare_one(options, dev.is_hdd ? 1 : 0);
                 if (hybrid)
                 {
                     options.erase("journal_device");

src/disk_tool_resize.cpp

@@ -91,7 +91,7 @@ int disk_tool_t::resize_parse_params()
     try
     {
         dsk.parse_config(options);
-        dsk.cached_read_data = dsk.cached_read_meta = dsk.cached_read_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();

src/disk_tool_utils.cpp

@@ -264,6 +264,7 @@ 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)
     {

src/etcd_state_client.cpp

@@ -684,8 +684,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             // ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Pool ID %s is invalid (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
@@ -829,8 +829,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         {
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Pool ID %s is invalid in PG configuration (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
@@ -838,8 +838,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             for (auto & pg_item: pool_item.second.object_items())
             {
                 pg_num_t pg_num = 0;
-                sscanf(pg_item.first.c_str(), "%u%c", &pg_num, &null_byte);
-                if (!pg_num || null_byte != 0)
+                int scanned = sscanf(pg_item.first.c_str(), "%u%c", &pg_num, &null_byte);
+                if (scanned != 1 || !pg_num)
                 {
                     fprintf(stderr, "Bad key in pool %u PG configuration: %s (must be a number), skipped\n", pool_id, pg_item.first.c_str());
                     continue;
@@ -889,8 +889,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         pg_num_t pg_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+12, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !pg_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+12, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !pg_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -944,8 +944,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         pg_num_t pg_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !pg_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !pg_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -1015,8 +1015,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         uint64_t pool_id = 0;
         uint64_t inode_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+14, "%lu/%lu%c", &pool_id, &inode_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !inode_num || (inode_num >> (64-POOL_ID_BITS)) || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+14, "%lu/%lu%c", &pool_id, &inode_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !inode_num || (inode_num >> (64-POOL_ID_BITS)))
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }

src/fio_cluster.cpp

@@ -24,6 +24,7 @@
 #include <netinet/tcp.h>
 
 #include <vector>
+#include <string>
 
 #include "vitastor_c.h"
 #include "fio_headers.h"
@@ -203,6 +204,15 @@ static void watch_callback(void *opaque, long watch)
     bsd->watch = (void*)watch;
 }
 
+static void opt_push(std::vector<char *> & options, const char *opt, const char *value)
+{
+    if (value)
+    {
+        options.push_back(strdup(opt));
+        options.push_back(strdup(value));
+    }
+}
+
 static int sec_setup(struct thread_data *td)
 {
     sec_options *o = (sec_options*)td->eo;
@@ -254,8 +264,27 @@ static int sec_setup(struct thread_data *td)
     {
         o->inode = 0;
     }
-    bsd->cli = vitastor_c_create_uring(o->config_path, o->etcd_host, o->etcd_prefix,
-        o->use_rdma, o->rdma_device, o->rdma_port_num, o->rdma_gid_index, o->rdma_mtu, o->cluster_log);
+    std::vector<char *> options;
+    opt_push(options, "config_path", o->config_path);
+    opt_push(options, "etcd_address", o->etcd_host);
+    opt_push(options, "etcd_prefix", o->etcd_prefix);
+    if (o->use_rdma != -1)
+        opt_push(options, "use_rdma", std::to_string(o->use_rdma).c_str());
+    opt_push(options, "rdma_device", o->rdma_device);
+    if (o->rdma_port_num)
+        opt_push(options, "rdma_port_num", std::to_string(o->rdma_port_num).c_str());
+    if (o->rdma_gid_index)
+        opt_push(options, "rdma_gid_index", std::to_string(o->rdma_gid_index).c_str());
+    if (o->rdma_mtu)
+        opt_push(options, "rdma_mtu", std::to_string(o->rdma_mtu).c_str());
+    if (o->cluster_log)
+        opt_push(options, "log_level", std::to_string(o->cluster_log).c_str());
+    // allow writeback caching if -direct is not set
+    opt_push(options, "client_writeback_allowed", td->o.odirect ? "0" : "1");
+    bsd->cli = vitastor_c_create_uring_json((const char**)options.data(), options.size());
+    for (auto opt: options)
+        free(opt);
+    options.clear();
     if (o->image)
     {
         bsd->watch = NULL;

src/fio_sec_osd.cpp

@@ -242,6 +242,7 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
             op.sec_rw.version = UINT64_MAX; // last unstable
             op.sec_rw.offset = io->offset % bsd->block_size;
             op.sec_rw.len = io->xfer_buflen;
+            op.sec_rw.attr_len = 0;
         }
         else
         {
@@ -263,6 +264,7 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
             op.sec_rw.version = 0; // assign automatically
             op.sec_rw.offset = io->offset % bsd->block_size;
             op.sec_rw.len = io->xfer_buflen;
+            op.sec_rw.attr_len = 0;
         }
         else
         {

src/messenger.cpp

@@ -11,6 +11,9 @@
 
 #include "addr_util.h"
 #include "messenger.h"
+#ifdef WITH_RDMA
+#include "msgr_rdma.h"
+#endif
 
 void osd_messenger_t::init()
 {
@@ -392,24 +395,27 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
             },
         },
     };
+    json11::Json::object payload;
+    if (this->osd_num)
+    {
+        payload["osd_num"] = this->osd_num;
+    }
 #ifdef WITH_RDMA
     if (rdma_context)
     {
         cl->rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_max_msg);
         if (cl->rdma_conn)
         {
-            json11::Json payload = json11::Json::object {
-                { "connect_rdma", cl->rdma_conn->addr.to_string() },
-                { "rdma_max_msg", cl->rdma_conn->max_msg },
-            };
-            std::string payload_str = payload.dump();
-            op->req.show_conf.json_len = payload_str.size();
-            op->buf = malloc_or_die(payload_str.size());
-            op->iov.push_back(op->buf, payload_str.size());
-            memcpy(op->buf, payload_str.c_str(), payload_str.size());
+            payload["connect_rdma"] = cl->rdma_conn->addr.to_string();
+            payload["rdma_max_msg"] = cl->rdma_conn->max_msg;
         }
     }
 #endif
+    std::string payload_str = json11::Json(payload).dump();
+    op->req.show_conf.json_len = payload_str.size();
+    op->buf = malloc_or_die(payload_str.size());
+    op->iov.push_back(op->buf, payload_str.size());
+    memcpy(op->buf, payload_str.c_str(), payload_str.size());
     op->callback = [this, cl](osd_op_t *op)
     {
         std::string json_err;

src/messenger.h

@@ -18,10 +18,6 @@
 #include "timerfd_manager.h"
 #include <ringloop.h>
 
-#ifdef WITH_RDMA
-#include "msgr_rdma.h"
-#endif
-
 #define CL_READ_HDR 1
 #define CL_READ_DATA 2
 #define CL_READ_REPLY_DATA 3
@@ -44,6 +40,11 @@ struct msgr_sendp_t
     int flags;
 };
 
+#ifdef WITH_RDMA
+struct msgr_rdma_connection_t;
+struct msgr_rdma_context_t;
+#endif
+
 struct osd_client_t
 {
     int refs = 0;

src/mock/messenger.cpp

@@ -55,3 +55,10 @@ json11::Json::object osd_messenger_t::merge_configs(const json11::Json::object &
 {
     return cli_config;
 }
+
+bool json_is_true(const json11::Json & val)
+{
+    if (val.is_string())
+        return val == "true" || val == "yes" || val == "1";
+    return val.bool_value();
+}

src/mock/ringloop.h

@@ -22,4 +22,10 @@ public:
     void submit()
     {
     }
+    void wait()
+    {
+    }
+    void loop()
+    {
+    }
 };

src/msgr_rdma.cpp

@@ -19,12 +19,12 @@ std::string msgr_rdma_address_t::to_string()
 bool msgr_rdma_address_t::from_string(const char *str, msgr_rdma_address_t *dest)
 {
     uint64_t* gid = (uint64_t*)&dest->gid;
-    int n = sscanf(
+    int scanned = sscanf(
         str, "%hx:%x:%x:%16lx%16lx", &dest->lid, &dest->qpn, &dest->psn, gid, gid+1
     );
     gid[0] = be64toh(gid[0]);
     gid[1] = be64toh(gid[1]);
-    return n == 5;
+    return scanned == 5;
 }
 
 msgr_rdma_context_t::~msgr_rdma_context_t()

src/msgr_stop.cpp

@@ -5,6 +5,9 @@
 #include <assert.h>
 
 #include "messenger.h"
+#ifdef WITH_RDMA
+#include "msgr_rdma.h"
+#endif
 
 void osd_messenger_t::cancel_osd_ops(osd_client_t *cl)
 {

src/nbd_proxy.cpp

@@ -341,6 +341,7 @@ public:
             ringloop->loop();
             ringloop->wait();
         }
+        cli->flush();
         delete cli;
         delete epmgr;
         delete ringloop;

src/nfs_conn.cpp

@@ -190,7 +190,15 @@ static int nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
     {
         if (handle == "roothandle" || self->parent->dir_by_hash.find(handle) != self->parent->dir_by_hash.end())
         {
-            *reply = (SETATTR3res){ .status = NFS3ERR_ISDIR };
+            if (args->new_attributes.size.set_it)
+            {
+                *reply = (SETATTR3res){ .status = NFS3ERR_ISDIR };
+            }
+            else
+            {
+                // Silently ignore mode, uid, gid, atime, mtime changes
+                *reply = (SETATTR3res){ .status = NFS3_OK };
+            }
         }
         else
         {
@@ -358,7 +366,6 @@ static int nfs3_read_proc(void *opaque, rpc_op_t *rop)
 }
 
 static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t new_size, uint64_t offset, uint64_t count, void *buf);
-static void nfs_do_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t offset, uint64_t count, void *buf);
 
 static int nfs3_write_proc(void *opaque, rpc_op_t *rop)
 {
@@ -392,7 +399,6 @@ static int nfs3_write_proc(void *opaque, rpc_op_t *rop)
         .resok = (WRITE3resok){
             //.file_wcc = ...,
             .count = (unsigned)count,
-            .committed = args->stable,
         },
     };
     if ((args->offset % alignment) != 0 || (count % alignment) != 0)
@@ -436,42 +442,101 @@ static int nfs3_write_proc(void *opaque, rpc_op_t *rop)
     return 1;
 }
 
-static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t new_size, uint64_t offset, uint64_t count, void *buf)
+static void complete_extend_write(nfs_client_t *self, rpc_op_t *rop, inode_t inode, int res)
 {
-    // Check if we have to resize the inode before writing
+    WRITE3args *args = (WRITE3args*)rop->request;
+    WRITE3res *reply = (WRITE3res*)rop->reply;
+    if (res < 0)
+    {
+        *reply = (WRITE3res){ .status = vitastor_nfs_map_err(res) };
+        rpc_queue_reply(rop);
+        return;
+    }
+    bool imm = self->parent->cli->get_immediate_commit(inode);
+    reply->resok.committed = args->stable != UNSTABLE || imm ? FILE_SYNC : UNSTABLE;
+    *(uint64_t*)reply->resok.verf = self->parent->server_id;
+    if (args->stable != UNSTABLE && !imm)
+    {
+        // Client requested a stable write. Add an fsync
+        auto op = new cluster_op_t;
+        op->opcode = OSD_OP_SYNC;
+        op->callback = [rop](cluster_op_t *op)
+        {
+            if (op->retval != 0)
+            {
+                WRITE3res *reply = (WRITE3res*)rop->reply;
+                *reply = (WRITE3res){ .status = vitastor_nfs_map_err(-op->retval) };
+            }
+            delete op;
+            rpc_queue_reply(rop);
+        };
+        self->parent->cli->execute(op);
+    }
+    else
+    {
+        rpc_queue_reply(rop);
+    }
+}
+
+static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size, int err)
+{
+    auto ext_it = self->extend_writes.lower_bound((extend_size_t){ .inode = inode, .new_size = 0 });
+    while (ext_it != self->extend_writes.end() &&
+        ext_it->first.inode == inode &&
+        ext_it->first.new_size <= new_size)
+    {
+        ext_it->second.resize_res = err;
+        if (ext_it->second.write_res <= 0)
+        {
+            complete_extend_write(self, ext_it->second.rop, inode, ext_it->second.write_res < 0
+                ? ext_it->second.write_res : ext_it->second.resize_res);
+            self->extend_writes.erase(ext_it++);
+        }
+        else
+            ext_it++;
+    }
+}
+
+static void extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size)
+{
+    // Send an extend request
+    auto & ext = self->extends[inode];
+    ext.cur_extend = new_size;
     auto inode_it = self->parent->cli->st_cli.inode_config.find(inode);
     if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
         inode_it->second.size < new_size)
     {
         self->parent->cmd->loop_and_wait(self->parent->cmd->start_modify(json11::Json::object {
+            // FIXME: Resizing by ID is probably more correct
             { "image", inode_it->second.name },
             { "resize", new_size },
+            { "inc_size", true },
             { "force_size", true },
         }), [=](const cli_result_t & r)
         {
+            auto & ext = self->extends[inode];
             if (r.err)
             {
-                if (r.err == EAGAIN)
-                {
-                    // Multiple concurrent resize requests received, try to repeat
-                    nfs_resize_write(self, rop, inode, new_size, offset, count, buf);
-                    return;
-                }
-                WRITE3res *reply = (WRITE3res*)rop->reply;
-                *reply = (WRITE3res){ .status = vitastor_nfs_map_err(r.err) };
-                rpc_queue_reply(rop);
+                fprintf(stderr, "Error extending inode %lu to %lu bytes: %s\n", inode, new_size, r.text.c_str());
+            }
+            if (r.err == EAGAIN || ext.next_extend > ext.cur_extend)
+            {
+                // Multiple concurrent resize requests received, try to repeat
+                extend_inode(self, inode, ext.next_extend > ext.cur_extend ? ext.next_extend : ext.cur_extend);
                 return;
             }
-            nfs_do_write(self, rop, inode, offset, count, buf);
+            ext.cur_extend = ext.next_extend = 0;
+            complete_extend_inode(self, inode, new_size, r.err);
         });
     }
     else
     {
-        nfs_do_write(self, rop, inode, offset, count, buf);
+        complete_extend_inode(self, inode, new_size, 0);
     }
 }
 
-static void nfs_do_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t offset, uint64_t count, void *buf)
+static void nfs_do_write(nfs_client_t *self, std::multimap<extend_size_t, extend_write_t>::iterator ewr_it,
+    rpc_op_t *rop, uint64_t inode, uint64_t offset, uint64_t count, void *buf)
 {
     cluster_op_t *op = new cluster_op_t;
     op->opcode = OSD_OP_WRITE;
@@ -479,48 +544,61 @@ static void nfs_do_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint
     op->offset = offset;
     op->len = count;
     op->iov.push_back(buf, count);
-    op->callback = [self, rop](cluster_op_t *op)
+    op->callback = [self, ewr_it, rop](cluster_op_t *op)
     {
-        uint64_t inode = op->inode;
-        WRITE3args *args = (WRITE3args*)rop->request;
-        WRITE3res *reply = (WRITE3res*)rop->reply;
-        if (op->retval != op->len)
+        auto inode = op->inode;
+        int write_res = op->retval < 0 ? op->retval : (op->retval != op->len ? -ERANGE : 0);
+        if (ewr_it == self->extend_writes.end())
         {
-            *reply = (WRITE3res){ .status = vitastor_nfs_map_err(-op->retval) };
-            delete op;
-            rpc_queue_reply(rop);
+            complete_extend_write(self, rop, inode, write_res);
         }
         else
         {
-            *(uint64_t*)reply->resok.verf = self->parent->server_id;
-            delete op;
-            if (args->stable != UNSTABLE &&
-                !self->parent->cli->get_immediate_commit(inode))
+            ewr_it->second.write_res = write_res;
+            if (ewr_it->second.resize_res <= 0)
             {
-                // Client requested a stable write. Add an fsync
-                op = new cluster_op_t;
-                op->opcode = OSD_OP_SYNC;
-                op->callback = [rop](cluster_op_t *op)
-                {
-                    if (op->retval != 0)
-                    {
-                        WRITE3res *reply = (WRITE3res*)rop->reply;
-                        *reply = (WRITE3res){ .status = vitastor_nfs_map_err(-op->retval) };
-                    }
-                    delete op;
-                    rpc_queue_reply(rop);
-                };
-                self->parent->cli->execute(op);
-            }
-            else
-            {
-                rpc_queue_reply(rop);
+                complete_extend_write(self, rop, inode, write_res < 0 ? write_res : ewr_it->second.resize_res);
+                self->extend_writes.erase(ewr_it);
             }
         }
     };
     self->parent->cli->execute(op);
 }
 
+static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t new_size, uint64_t offset, uint64_t count, void *buf)
+{
+    // Check if we have to resize the inode during write
+    auto inode_it = self->parent->cli->st_cli.inode_config.find(inode);
+    if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
+        inode_it->second.size < new_size)
+    {
+        auto ewr_it = self->extend_writes.emplace((extend_size_t){
+            .inode = inode,
+            .new_size = new_size,
+        }, (extend_write_t){
+            .rop = rop,
+            .resize_res = 1,
+            .write_res = 1,
+        });
+        auto & ext = self->extends[inode];
+        if (ext.cur_extend > 0)
+        {
+            // Already resizing, just wait
+            if (ext.next_extend < new_size)
+                ext.next_extend = new_size;
+        }
+        else
+        {
+            extend_inode(self, inode, new_size);
+        }
+        nfs_do_write(self, ewr_it, rop, inode, offset, count, buf);
+    }
+    else
+    {
+        nfs_do_write(self, self->extend_writes.end(), rop, inode, offset, count, buf);
+    }
+}
+
 static int nfs3_create_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
@@ -881,6 +959,27 @@ static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
+static void fill_dir_entry(nfs_client_t *self, rpc_op_t *rop,
+    std::map<std::string, nfs_dir_t>::iterator dir_id_it, struct entryplus3 *entry, bool is_plus)
+{
+    if (dir_id_it == self->parent->dir_info.end())
+    {
+        return;
+    }
+    entry->fileid = dir_id_it->second.id;
+    if (is_plus)
+    {
+        entry->name_attributes = (post_op_attr){
+            .attributes_follow = 1,
+            .attributes = get_dir_attributes(self, dir_id_it->first),
+        };
+        entry->name_handle = (post_op_fh3){
+            .handle_follows = 1,
+            .handle = xdr_copy_string(rop->xdrs, "S"+base64_encode(sha256(dir_id_it->first))),
+        };
+    }
+}
+
 static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
@@ -958,17 +1057,17 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
             continue;
         std::string subname = dir_id_it->first.substr(prefix.size());
         // for directories, fileid changes when the user restarts proxy
-        entries[subname].fileid = dir_id_it->second.id;
-        if (is_plus)
+        fill_dir_entry(self, rop, dir_id_it, &entries[subname], is_plus);
+    }
+    // Add . and ..
+    {
+        auto dir_id_it = self->parent->dir_info.find(dir);
+        fill_dir_entry(self, rop, dir_id_it, &entries["."], is_plus);
+        auto sl = dir.rfind("/");
+        if (sl != std::string::npos)
         {
-            entries[subname].name_attributes = (post_op_attr){
-                .attributes_follow = 1,
-                .attributes = get_dir_attributes(self, dir_id_it->first),
-            };
-            entries[subname].name_handle = (post_op_fh3){
-                .handle_follows = 1,
-                .handle = xdr_copy_string(rop->xdrs, "S"+base64_encode(sha256(dir_id_it->first))),
-            };
+            auto dir_id_it = self->parent->dir_info.find(dir.substr(0, sl));
+            fill_dir_entry(self, rop, dir_id_it, &entries[".."], is_plus);
         }
     }
     // Offset results by the continuation cookie (equal to index in the listing)
@@ -1193,10 +1292,11 @@ static int nfs3_commit_proc(void *opaque, rpc_op_t *rop)
     cluster_op_t *op = new cluster_op_t;
     // fsync. we don't know how to fsync a single inode, so just fsync everything
     op->opcode = OSD_OP_SYNC;
-    op->callback = [rop](cluster_op_t *op)
+    op->callback = [self, rop](cluster_op_t *op)
     {
         COMMIT3res *reply = (COMMIT3res*)rop->reply;
         *reply = (COMMIT3res){ .status = vitastor_nfs_map_err(op->retval) };
+        *(uint64_t*)reply->resok.verf = self->parent->server_id;
         rpc_queue_reply(rop);
     };
     self->parent->cli->execute(op);

src/nfs_proxy.cpp

@@ -34,7 +34,10 @@ nfs_proxy_t::~nfs_proxy_t()
     if (cmd)
         delete cmd;
     if (cli)
+    {
+        cli->flush();
         delete cli;
+    }
     if (epmgr)
         delete epmgr;
     if (ringloop)
@@ -261,16 +264,8 @@ void nfs_proxy_t::run(json11::Json cfg)
         ringloop->loop();
         ringloop->wait();
     }
-    /*// Sync at the end
-    cluster_op_t *close_sync = new cluster_op_t;
-    close_sync->opcode = OSD_OP_SYNC;
-    close_sync->callback = [&stop](cluster_op_t *op)
-    {
-        stop = true;
-        delete op;
-    };
-    cli->execute(close_sync);*/
     // Destroy the client
+    cli->flush();
     delete cli;
     delete epmgr;
     delete ringloop;
@@ -346,8 +341,8 @@ void nfs_proxy_t::parse_stats(etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         inode_t inode_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !inode_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !inode_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -360,8 +355,8 @@ void nfs_proxy_t::parse_stats(etcd_kv_t & kv)
     {
         pool_id_t pool_id = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+12, "%u%c", &pool_id, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX)
+        int scanned = sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+12, "%u%c", &pool_id, &null_byte);
+        if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }

src/nfs_proxy.h

@@ -86,6 +86,28 @@ struct rpc_free_buffer_t
     unsigned size;
 };
 
+struct extend_size_t
+{
+    inode_t inode;
+    uint64_t new_size;
+};
+
+inline bool operator < (const extend_size_t &a, const extend_size_t &b)
+{
+    return a.inode < b.inode || a.inode == b.inode && a.new_size < b.new_size;
+}
+
+struct extend_write_t
+{
+    rpc_op_t *rop;
+    int resize_res, write_res; // 1 = started, 0 = completed OK, -errno = completed with error
+};
+
+struct extend_inode_t
+{
+    uint64_t cur_extend = 0, next_extend = 0;
+};
+
 class nfs_client_t
 {
 public:
@@ -100,6 +122,8 @@ public:
     rpc_cur_buffer_t cur_buffer = { 0 };
     std::map<uint8_t*, rpc_used_buffer_t> used_buffers;
     std::vector<rpc_free_buffer_t> free_buffers;
+    std::map<inode_t, extend_inode_t> extends;
+    std::multimap<extend_size_t, extend_write_t> extend_writes;
 
     iovec read_iov;
     msghdr read_msg = { 0 };

src/osd.cpp

@@ -184,6 +184,14 @@ void osd_t::parse_config(bool init)
         // Allow to set it to 0
         autosync_writes = config["autosync_writes"].uint64_value();
     }
+    if (!config["fsync_feedback_repeat_interval"].is_null())
+    {
+        fsync_feedback_repeat_interval = config["fsync_feedback_repeat_interval"].uint64_value();
+    }
+    if (!fsync_feedback_repeat_interval)
+    {
+        fsync_feedback_repeat_interval = 500; // ms
+    }
     if (!config["client_queue_depth"].is_null())
     {
         client_queue_depth = config["client_queue_depth"].uint64_value();

src/osd.h

@@ -122,6 +122,7 @@ class osd_t
     uint32_t scrub_list_limit = 1000;
     bool scrub_find_best = true;
     uint64_t scrub_ec_max_bruteforce = 100;
+    uint64_t fsync_feedback_repeat_interval = 500;
 
     // cluster state
 
@@ -166,6 +167,8 @@ class osd_t
     uint64_t unstable_write_count = 0;
     std::map<osd_object_id_t, uint64_t> unstable_writes;
     std::deque<osd_op_t*> syncs_in_progress;
+    std::map<int, timespec> unstable_write_osds;
+    int fsync_feedback_timer_id = -1;
 
     // client & peer I/O
 
@@ -257,6 +260,7 @@ class osd_t
     void exec_show_config(osd_op_t *cur_op);
     void exec_secondary(osd_op_t *cur_op);
     void secondary_op_callback(osd_op_t *cur_op);
+    void fsync_feedback();
 
     // primary ops
     void autosync();

src/osd_cluster.cpp

@@ -186,10 +186,12 @@ json11::Json osd_t::get_statistics()
     if (bs)
     {
         st["blockstore_ready"] = bs->is_started();
-        st["data_block_size"] = (uint64_t)bs->get_block_size();
         st["size"] = bs->get_block_count() * bs->get_block_size();
         st["free"] = bs->get_free_block_count() * bs->get_block_size();
     }
+    st["data_block_size"] = (uint64_t)bs_block_size;
+    st["bitmap_granularity"] = (uint64_t)bs_bitmap_granularity;
+    st["immediate_commit"] = immediate_commit == IMMEDIATE_ALL ? "all" : (immediate_commit == IMMEDIATE_SMALL ? "small" : "none");
     st["host"] = self_state["host"];
     json11::Json::object op_stats, subop_stats;
     for (int i = OSD_OP_MIN; i <= OSD_OP_MAX; i++)
@@ -649,7 +651,7 @@ void osd_t::apply_pg_config()
             auto pg_it = this->pgs.find({ .pool_id = pool_id, .pg_num = pg_num });
             bool currently_taken = pg_it != this->pgs.end() && pg_it->second.state != PG_OFFLINE;
             // Check pool block size and bitmap granularity
-            if (this->bs_block_size != pool_item.second.data_block_size ||
+            if (take && this->bs_block_size != pool_item.second.data_block_size ||
                 this->bs_bitmap_granularity != pool_item.second.bitmap_granularity)
             {
                 if (!warned_block_size)
@@ -967,8 +969,8 @@ void osd_t::report_pg_states()
                         pool_id_t pool_id = 0;
                         pg_num_t pg_num = 0;
                         char null_byte = 0;
-                        sscanf(kv.key.c_str() + st_cli.etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-                        if (null_byte == 0)
+                        int scanned = sscanf(kv.key.c_str() + st_cli.etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+                        if (scanned == 2)
                         {
                             auto pg_it = pgs.find({ .pool_id = pool_id, .pg_num = pg_num });
                             if (pg_it != pgs.end() && pg_it->second.state != PG_OFFLINE && pg_it->second.state != PG_STARTING &&

src/osd_secondary.cpp

@@ -2,6 +2,9 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include "osd.h"
+#ifdef WITH_RDMA
+#include "msgr_rdma.h"
+#endif
 
 #include "json11/json11.hpp"
 
@@ -26,6 +29,23 @@ void osd_t::secondary_op_callback(osd_op_t *op)
         if (op->bs_op->retval > 0)
             op->iov.push_back(op->buf, op->bs_op->retval);
     }
+    else if (op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
+        op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE)
+    {
+#ifndef OSD_STUB
+        fsync_feedback();
+#endif
+        if (op->req.hdr.opcode == OSD_OP_SEC_WRITE)
+        {
+            auto & u = unstable_write_osds[op->peer_fd];
+            u = u;
+        }
+    }
+    else if (op->req.hdr.opcode == OSD_OP_SEC_SYNC)
+    {
+        // FIXME It would be more correct to track STABILIZE ops, not just reset on SYNC
+        unstable_write_osds.erase(op->peer_fd);
+    }
     else if (op->req.hdr.opcode == OSD_OP_SEC_LIST)
     {
         // allocated by blockstore
@@ -42,6 +62,71 @@ void osd_t::secondary_op_callback(osd_op_t *op)
     finish_op(op, retval);
 }
 
+void osd_t::fsync_feedback()
+{
+    if (!unstable_write_osds.size() || !bs->wants_fsync())
+    {
+        return;
+    }
+    bool postpone = false;
+    // Broadcast fsync feedback
+    timespec now;
+    clock_gettime(CLOCK_REALTIME, &now);
+    for (auto up_it = unstable_write_osds.begin(); up_it != unstable_write_osds.end(); )
+    {
+        auto & peer_fd = up_it->first;
+        auto & last_feedback = up_it->second;
+        if (msgr.clients.find(peer_fd) == msgr.clients.end() ||
+            !msgr.clients.at(peer_fd)->osd_num)
+        {
+            unstable_write_osds.erase(up_it++);
+            continue;
+        }
+        auto diff = (now.tv_sec-last_feedback.tv_sec)*1000 + (now.tv_nsec-last_feedback.tv_nsec)/1000000;
+        if (diff > fsync_feedback_repeat_interval)
+        {
+            last_feedback = now;
+            // Request fsync from the primary OSD
+            // Note: primary OSD should NOT divide syncs by clients or this logic will break
+            osd_op_t *fb_op = new osd_op_t();
+            fb_op->op_type = OSD_OP_OUT;
+            fb_op->req = (osd_any_op_t){
+                .sync = {
+                    .header = {
+                        .magic = SECONDARY_OSD_OP_MAGIC,
+                        .id = msgr.next_subop_id++,
+                        .opcode = OSD_OP_SYNC,
+                    },
+                },
+            };
+            fb_op->callback = [this](osd_op_t *op)
+            {
+                delete op;
+            };
+            fb_op->peer_fd = peer_fd;
+            msgr.outbox_push(fb_op);
+        }
+        else
+        {
+            postpone = true;
+        }
+        up_it++;
+    }
+    if (fsync_feedback_timer_id >= 0)
+    {
+        tfd->clear_timer(fsync_feedback_timer_id);
+        fsync_feedback_timer_id = -1;
+    }
+    if (postpone)
+    {
+        fsync_feedback_timer_id = tfd->set_timer(fsync_feedback_repeat_interval, false, [this](int timer_id)
+        {
+            fsync_feedback_timer_id = -1;
+            fsync_feedback();
+        });
+    }
+}
+
 void osd_t::exec_secondary(osd_op_t *cur_op)
 {
     if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
@@ -155,6 +240,7 @@ void osd_t::exec_show_config(osd_op_t *cur_op)
     json11::Json req_json = cur_op->req.show_conf.json_len > 0
         ? json11::Json::parse(std::string((char *)cur_op->buf), json_err)
         : json11::Json();
+    msgr.clients.at(cur_op->peer_fd)->osd_num = req_json["osd_num"].uint64_value();
     // Expose sensitive configuration values so peers can check them
     json11::Json::object wire_config = json11::Json::object {
         { "osd_num", osd_num },

src/qemu_driver.c

@@ -197,7 +197,11 @@ static void vitastor_parse_filename(const char *filename, QDict *options, Error
             !strcmp(name, "rdma-mtu"))
         {
             unsigned long long num_val;
+#if QEMU_VERSION_MAJOR < 8 || QEMU_VERSION_MAJOR == 8 && QEMU_VERSION_MINOR < 1
             if (parse_uint_full(value, &num_val, 0))
+#else
+            if (parse_uint_full(value, 0, &num_val))
+#endif
             {
                 error_setg(errp, "Illegal %s: %s", name, value);
                 goto out;
@@ -234,15 +238,13 @@ out:
     return;
 }
 
+#if defined VITASTOR_C_API_VERSION && VITASTOR_C_API_VERSION >= 2
 static void vitastor_uring_handler(void *opaque)
 {
     VitastorClient *client = (VitastorClient*)opaque;
     qemu_mutex_lock(&client->mutex);
     client->bh_uring_scheduled = 0;
-    do
-    {
-        vitastor_c_uring_handle_events(client->proxy);
-    } while (vitastor_c_uring_has_work(client->proxy));
+    vitastor_c_uring_handle_events(client->proxy);
     qemu_mutex_unlock(&client->mutex);
 }
 
@@ -266,20 +268,23 @@ static void vitastor_schedule_uring_handler(VitastorClient *client)
         replay_bh_schedule_oneshot_event(client->ctx, vitastor_uring_handler, opaque);
 #elif QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 8
         aio_bh_schedule_oneshot(client->ctx, vitastor_uring_handler, opaque);
-#elif QEMU_VERSION_MAJOR >= 2
+#else
         VitastorBH *vbh = (VitastorBH*)malloc(sizeof(VitastorBH));
         vbh->cli = client;
+#if QEMU_VERSION_MAJOR >= 2
         vbh->bh = aio_bh_new(bdrv_get_aio_context(task->bs), vitastor_bh_uring_handler, vbh);
-        qemu_bh_schedule(vbh->bh);
 #else
-        client->bh_uring_scheduled = 0;
-        do
-        {
-            vitastor_c_uring_handle_events(client->proxy);
-        } while (vitastor_c_uring_has_work(client->proxy));
+        vbh->bh = qemu_bh_new(vitastor_bh_uring_handler, vbh);
+#endif
+        qemu_bh_schedule(vbh->bh);
 #endif
     }
 }
+#else
+static void vitastor_schedule_uring_handler(VitastorClient *client)
+{
+}
+#endif
 
 static void coroutine_fn vitastor_co_get_metadata(VitastorRPC *task)
 {
@@ -319,7 +324,7 @@ static void vitastor_aio_fd_write(void *fddv)
 static void universal_aio_set_fd_handler(AioContext *ctx, int fd, IOHandler *fd_read, IOHandler *fd_write, void *opaque)
 {
     aio_set_fd_handler(ctx, fd,
-#if QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 5 || QEMU_VERSION_MAJOR >= 3
+#if QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 5 || QEMU_VERSION_MAJOR >= 3 && (QEMU_VERSION_MAJOR < 8 || QEMU_VERSION_MAJOR == 8 && QEMU_VERSION_MINOR < 1)
         0 /*is_external*/,
 #endif
         fd_read,
@@ -383,6 +388,43 @@ static void vitastor_aio_set_fd_handler(void *vcli, int fd, int unused1, IOHandl
     );
 }
 
+typedef struct str_array
+{
+    const char **items;
+    int len, alloc;
+} str_array;
+
+static void strarray_push(str_array *a, const char *str)
+{
+    if (a->len >= a->alloc)
+    {
+        a->alloc = !a->alloc ? 4 : 2*a->alloc;
+        a->items = (const char**)realloc(a->items, a->alloc*sizeof(char*));
+        if (!a->items)
+        {
+            fprintf(stderr, "bad alloc\n");
+            abort();
+        }
+    }
+    a->items[a->len++] = str;
+}
+
+static void strarray_push_kv(str_array *a, const char *key, const char *value)
+{
+    if (key && value)
+    {
+        strarray_push(a, key);
+        strarray_push(a, value);
+    }
+}
+
+static void strarray_free(str_array *a)
+{
+    free(a->items);
+    a->items = NULL;
+    a->len = a->alloc = 0;
+}
+
 static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, Error **errp)
 {
     VitastorRPC task;
@@ -401,27 +443,19 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
     client->rdma_gid_index = qdict_get_try_int(options, "rdma-gid-index", 0);
     client->rdma_mtu = qdict_get_try_int(options, "rdma-mtu", 0);
     client->ctx = bdrv_get_aio_context(bs);
-#if defined VITASTOR_C_API_VERSION && VITASTOR_C_API_VERSION >= 2
-    client->proxy = vitastor_c_create_qemu_uring(
-        vitastor_aio_set_fd_handler, client, client->config_path, client->etcd_host, client->etcd_prefix,
-        client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
-    );
-#else
-    client->proxy = vitastor_c_create_uring(
-        client->config_path, client->etcd_host, client->etcd_prefix,
-        client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
-    );
-#endif
-    if (!client->proxy)
-    {
-        fprintf(stderr, "vitastor: failed to create io_uring: %s - I/O will be slower\n", strerror(errno));
-        client->uring_eventfd = -1;
-        client->proxy = vitastor_c_create_qemu(
-            vitastor_aio_set_fd_handler, client, client->config_path, client->etcd_host, client->etcd_prefix,
-            client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
-        );
-    }
-    else
+    str_array opt = {};
+    strarray_push_kv(&opt, "config_path", qdict_get_try_str(options, "config-path"));
+    strarray_push_kv(&opt, "etcd_address", qdict_get_try_str(options, "etcd-host"));
+    strarray_push_kv(&opt, "etcd_prefix", qdict_get_try_str(options, "etcd-prefix"));
+    strarray_push_kv(&opt, "use_rdma", qdict_get_try_str(options, "use-rdma"));
+    strarray_push_kv(&opt, "rdma_device", qdict_get_try_str(options, "rdma-device"));
+    strarray_push_kv(&opt, "rdma_port_num", qdict_get_try_str(options, "rdma-port-num"));
+    strarray_push_kv(&opt, "rdma_gid_index", qdict_get_try_str(options, "rdma-gid-index"));
+    strarray_push_kv(&opt, "rdma_mtu", qdict_get_try_str(options, "rdma-mtu"));
+    strarray_push_kv(&opt, "client_writeback_allowed", (flags & BDRV_O_NOCACHE) ? "0" : "1");
+    client->proxy = vitastor_c_create_uring_json(opt.items, opt.len);
+    strarray_free(&opt);
+    if (client->proxy)
     {
         client->uring_eventfd = vitastor_c_uring_register_eventfd(client->proxy);
         if (client->uring_eventfd < 0)
@@ -433,6 +467,45 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
         }
         universal_aio_set_fd_handler(client->ctx, client->uring_eventfd, vitastor_uring_handler, NULL, client);
     }
+    else
+    {
+        // Writeback cache is unusable without io_uring because the client can't correctly flush on exit
+        fprintf(stderr, "vitastor: failed to create io_uring: %s - I/O will be slower%s\n",
+            strerror(errno), (flags & BDRV_O_NOCACHE ? "" : " and writeback cache will be disabled"));
+        client->uring_eventfd = -1;
+#if defined VITASTOR_C_API_VERSION && VITASTOR_C_API_VERSION >= 2
+        client->proxy = vitastor_c_create_qemu_uring(
+            vitastor_aio_set_fd_handler, client, client->config_path, client->etcd_host, client->etcd_prefix,
+            client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
+        );
+        if (!client->proxy)
+        {
+            fprintf(stderr, "vitastor: failed to create io_uring: %s - I/O will be slower\n", strerror(errno));
+            client->uring_eventfd = -1;
+            client->proxy = vitastor_c_create_qemu(
+                vitastor_aio_set_fd_handler, client, client->config_path, client->etcd_host, client->etcd_prefix,
+                client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
+            );
+        }
+        else
+        {
+            client->uring_eventfd = vitastor_c_uring_register_eventfd(client->proxy);
+            if (client->uring_eventfd < 0)
+            {
+                fprintf(stderr, "vitastor: failed to create io_uring eventfd: %s\n", strerror(errno));
+                error_setg(errp, "failed to create io_uring eventfd");
+                vitastor_close(bs);
+                return -1;
+            }
+            universal_aio_set_fd_handler(client->ctx, client->uring_eventfd, vitastor_uring_handler, NULL, client);
+        }
+#else
+        client->proxy = vitastor_c_create_qemu(
+            vitastor_aio_set_fd_handler, client, client->config_path, client->etcd_host, client->etcd_prefix,
+            client->use_rdma, client->rdma_device, client->rdma_port_num, client->rdma_gid_index, client->rdma_mtu, 0
+        );
+#endif
+    }
     image = client->image = g_strdup(qdict_get_try_str(options, "image"));
     client->readonly = (flags & BDRV_O_RDWR) ? 1 : 0;
     // Get image metadata (size and readonly flag) or just wait until the client is ready
@@ -491,6 +564,10 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
         return -1;
     }
     bs->total_sectors = client->size / BDRV_SECTOR_SIZE;
+#if QEMU_VERSION_MAJOR > 5 || QEMU_VERSION_MAJOR == 5 && QEMU_VERSION_MINOR >= 1
+    /* When extending regular files, we get zeros from the OS */
+    bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
+#endif
     //client->aio_context = bdrv_get_aio_context(bs);
     qdict_del(options, "use-rdma");
     qdict_del(options, "rdma-mtu");
@@ -587,7 +664,11 @@ static int coroutine_fn vitastor_co_truncate(BlockDriverState *bs, int64_t offse
     }
 
     // TODO: Resize inode to <offset> bytes
-    client->size = offset / BDRV_SECTOR_SIZE;
+#if QEMU_VERSION_MAJOR >= 4
+    client->size = exact || client->size < offset ? offset : client->size;
+#else
+    client->size = offset;
+#endif
 
     return 0;
 }
@@ -664,7 +745,8 @@ static void vitastor_co_generic_cb(void *opaque, long retval)
     task->bh = aio_bh_new(bdrv_get_aio_context(task->bs), vitastor_co_generic_bh_cb, opaque);
     qemu_bh_schedule(task->bh);
 #else
-    vitastor_co_generic_bh_cb(opaque);
+    task->bh = qemu_bh_new(vitastor_co_generic_bh_cb, opaque);
+    qemu_bh_schedule(task->bh);
 #endif
 }
 
@@ -741,7 +823,6 @@ static void vitastor_co_read_bitmap_cb(void *opaque, long retval, uint8_t *bitma
     VitastorRPC *task = opaque;
     VitastorClient *client = task->bs->opaque;
     task->ret = retval;
-    task->complete = 1;
     if (retval >= 0)
     {
         task->bitmap = bitmap;
@@ -753,15 +834,17 @@ static void vitastor_co_read_bitmap_cb(void *opaque, long retval, uint8_t *bitma
             client->last_bitmap = bitmap;
         }
     }
-    if (qemu_coroutine_self() != task->co)
-    {
-#if QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR > 8
-        aio_co_wake(task->co);
+#if QEMU_VERSION_MAJOR > 4 || QEMU_VERSION_MAJOR == 4 && QEMU_VERSION_MINOR >= 2
+    replay_bh_schedule_oneshot_event(bdrv_get_aio_context(task->bs), vitastor_co_generic_bh_cb, opaque);
+#elif QEMU_VERSION_MAJOR >= 3 || QEMU_VERSION_MAJOR == 2 && QEMU_VERSION_MINOR >= 8
+    aio_bh_schedule_oneshot(bdrv_get_aio_context(task->bs), vitastor_co_generic_bh_cb, opaque);
+#elif QEMU_VERSION_MAJOR >= 2
+    task->bh = aio_bh_new(bdrv_get_aio_context(task->bs), vitastor_co_generic_bh_cb, opaque);
+    qemu_bh_schedule(task->bh);
 #else
-        qemu_coroutine_enter(task->co, NULL);
-        qemu_aio_release(task);
+    task->bh = qemu_bh_new(vitastor_co_generic_bh_cb, opaque);
+    qemu_bh_schedule(task->bh);
 #endif
-    }
 }
 
 static int coroutine_fn vitastor_co_block_status(

src/ringloop.cpp

@@ -66,6 +66,16 @@ void ring_loop_t::unregister_consumer(ring_consumer_t *consumer)
 
 void ring_loop_t::loop()
 {
+    if (ring_eventfd >= 0)
+    {
+        // Reset eventfd counter
+        uint64_t ctr = 0;
+        int r = read(ring_eventfd, &ctr, 8);
+        if (r < 0 && errno != EAGAIN && errno != EINTR)
+        {
+            fprintf(stderr, "Error resetting eventfd: %s\n", strerror(errno));
+        }
+    }
     struct io_uring_cqe *cqe;
     while (!io_uring_peek_cqe(&ring, &cqe))
     {
@@ -84,7 +94,7 @@ void ring_loop_t::loop()
         }
         else
         {
-            printf("Warning: empty callback in SQE\n");
+            fprintf(stderr, "Warning: empty callback in SQE\n");
             free_ring_data[free_ring_data_ptr++] = d - ring_datas;
         }
         io_uring_cqe_seen(&ring, cqe);

src/str_util.cpp

@@ -308,3 +308,19 @@ std::string str_repeat(const std::string & str, int times)
         r += str;
     return r;
 }
+
+size_t utf8_length(const std::string & s)
+{
+    size_t len = 0;
+    for (size_t i = 0; i < s.size(); i++)
+        len += (s[i] & 0xC0) != 0x80;
+    return len;
+}
+
+size_t utf8_length(const char *s)
+{
+    size_t len = 0;
+    for (; *s; s++)
+        len += (*s & 0xC0) != 0x80;
+    return len;
+}

src/str_util.h

@@ -18,3 +18,5 @@ void print_help(const char *help_text, std::string exe_name, std::string cmd, bo
 uint64_t parse_time(std::string time_str, bool *ok = NULL);
 std::string read_all_fd(int fd);
 std::string str_repeat(const std::string & str, int times);
+size_t utf8_length(const std::string & s);
+size_t utf8_length(const char *s);

src/test_cluster_client.cpp

@@ -4,7 +4,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <assert.h>
-#include "cluster_client.h"
+#include "cluster_client_impl.h"
 
 void configure_single_pg_pool(cluster_client_t *cli)
 {
@@ -47,11 +47,11 @@ void configure_single_pg_pool(cluster_client_t *cli)
     cli->st_cli.on_change_hook(changes);
 }
 
-int *test_write(cluster_client_t *cli, uint64_t offset, uint64_t len, uint8_t c, std::function<void()> cb = NULL)
+int *test_write(cluster_client_t *cli, uint64_t offset, uint64_t len, uint8_t c, std::function<void()> cb = NULL, bool instant = false)
 {
     printf("Post write %lx+%lx\n", offset, len);
     int *r = new int;
-    *r = -1;
+    *r = instant ? -2 : -1;
     cluster_op_t *op = new cluster_op_t();
     op->opcode = OSD_OP_WRITE;
     op->inode = 0x1000000000001;
@@ -72,6 +72,13 @@ int *test_write(cluster_client_t *cli, uint64_t offset, uint64_t len, uint8_t c,
             cb();
     };
     cli->execute(op);
+    if (instant)
+    {
+        long res = *r;
+        assert(*r >= 0);
+        delete r;
+        return (int*)res;
+    }
     return r;
 }
 
@@ -160,6 +167,13 @@ osd_op_t *find_op(cluster_client_t *cli, osd_num_t osd_num, uint64_t opcode, uin
         }
         op_it++;
     }
+    op_it = cli->msgr.clients[peer_fd]->sent_ops.begin();
+    while (op_it != cli->msgr.clients[peer_fd]->sent_ops.end())
+    {
+        printf("Found opcode %lu offset %lx size %x\n", op_it->second->req.hdr.opcode, op_it->second->req.rw.offset, op_it->second->req.rw.len);
+        op_it++;
+    }
+    printf("Not found opcode %lu offset %lx size %lx\n", opcode, offset, len);
     return NULL;
 }
 
@@ -192,11 +206,16 @@ void test1()
     check_op_count(cli, 1, 1);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
     check_completed(r1);
+    r1 = test_write(cli, 4096, 4096, 0x56);
+    can_complete(r1);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 4096, 4096), 0);
+    check_completed(r1);
     pretend_disconnected(cli, 1);
     int *r2 = test_sync(cli);
     pretend_connected(cli, 1);
     check_op_count(cli, 1, 1);
-    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 8192), 0);
     check_op_count(cli, 1, 1);
     can_complete(r2);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
@@ -321,9 +340,8 @@ void test1()
     check_disconnected(cli, 1);
     pretend_connected(cli, 1);
     cli->continue_ops(true);
-    check_op_count(cli, 1, 2);
-    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
-    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0x1000, 0x1000), 0);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x2000), 0);
     check_op_count(cli, 1, 1);
     can_complete(r2);
     pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0x1000, 0x1000), 0);
@@ -337,7 +355,7 @@ void test1()
 
 void test2()
 {
-    std::map<object_id, cluster_buffer_t> unsynced_writes;
+    writeback_cache_t *wb = new writeback_cache_t();
     cluster_op_t *op = new cluster_op_t();
     op->opcode = OSD_OP_WRITE;
     op->inode = 1;
@@ -346,19 +364,19 @@ void test2()
     op->iov.push_back(malloc_or_die(4096*1024), 4096);
     // 0-4k = 0x55
     memset(op->iov.buf[0].iov_base, 0x55, op->iov.buf[0].iov_len);
-    cluster_client_t::copy_write(op, unsynced_writes);
+    wb->copy_write(op, CACHE_WRITTEN);
     // 8k-12k = 0x66
     op->offset = 8192;
     memset(op->iov.buf[0].iov_base, 0x66, op->iov.buf[0].iov_len);
-    cluster_client_t::copy_write(op, unsynced_writes);
+    wb->copy_write(op, CACHE_WRITTEN);
     // 4k-1M+4k = 0x77
     op->len = op->iov.buf[0].iov_len = 1048576;
     op->offset = 4096;
     memset(op->iov.buf[0].iov_base, 0x77, op->iov.buf[0].iov_len);
-    cluster_client_t::copy_write(op, unsynced_writes);
+    wb->copy_write(op, CACHE_WRITTEN);
     // check it
-    assert(unsynced_writes.size() == 4);
-    auto uit = unsynced_writes.begin();
+    assert(wb->dirty_buffers.size() == 2);
+    auto uit = wb->dirty_buffers.begin();
     int i;
     assert(uit->first.inode == 1);
     assert(uit->first.stripe == 0);
@@ -368,35 +386,106 @@ void test2()
     uit++;
     assert(uit->first.inode == 1);
     assert(uit->first.stripe == 4096);
-    assert(uit->second.len == 4096);
-    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
-    assert(i == uit->second.len);
-    uit++;
-    assert(uit->first.inode == 1);
-    assert(uit->first.stripe == 8192);
-    assert(uit->second.len == 4096);
-    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
-    assert(i == uit->second.len);
-    uit++;
-    assert(uit->first.inode == 1);
-    assert(uit->first.stripe == 12*1024);
-    assert(uit->second.len == 1016*1024);
+    assert(uit->second.len == 1048576);
     for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
     assert(i == uit->second.len);
     uit++;
     // free memory
     free(op->iov.buf[0].iov_base);
     delete op;
-    for (auto p: unsynced_writes)
-    {
-        free(p.second.buf);
-    }
+    delete wb;
     printf("[ok] copy_write test\n");
 }
 
+void test_writeback()
+{
+    json11::Json config = json11::Json::object {
+        { "client_enable_writeback", true },
+        { "client_writeback_allowed", true },
+        { "client_max_buffered_bytes", 1024*1024 },
+        { "client_max_buffered_ops", 2 },
+        { "client_max_writeback_iodepth", 2 },
+        { "client_max_dirty_bytes", 1024*1024 },
+        { "client_max_dirty_ops", 2 },
+    };
+    timerfd_manager_t *tfd = new timerfd_manager_t([](int fd, bool wr, std::function<void(int, int)> callback){});
+    cluster_client_t *cli = new cluster_client_t(NULL, tfd, config);
+
+    configure_single_pg_pool(cli);
+    pretend_connected(cli, 1);
+
+    // Check that 3 consecutive writes are merged by writeback
+    assert((long)test_write(cli, 0, 4096, 0x55, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    assert((long)test_write(cli, 4096, 4096, 0x55, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    assert((long)test_write(cli, 8192, 4096, 0x55, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+
+    assert((long)test_write(cli, 1024*1024, 4096, 0x66, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+
+    // 3rd and 4th writes should trigger 1 writeback each
+    assert((long)test_write(cli, 2*1024*1024, 4096, 0x66, NULL, true) == 1);
+    check_op_count(cli, 1, 1);
+    assert((long)test_write(cli, 3*1024*1024, 4096, 0x66, NULL, true) == 1);
+    check_op_count(cli, 1, 2);
+
+    // 5th write should be postponed until at least 1 writeback is completed
+    int *r1 = test_write(cli, 4*1024*1024, 4096, 0x67, NULL);
+    check_op_count(cli, 1, 2);
+    can_complete(r1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 3*4096), 0);
+    check_completed(r1);
+    // autosync because max_dirty_ops=2, flush waits for sync
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 1024*1024, 4096), 0);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 2*1024*1024, 4096), 0);
+    check_op_count(cli, 1, 0);
+
+    int *r2 = test_sync(cli);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 3*1024*1024, 4096), 0);
+    check_op_count(cli, 1, 1);
+    // autosync because max_dirty_ops=2, flush waits for sync
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 4*1024*1024, 4096), 0);
+    check_op_count(cli, 1, 1);
+    can_complete(r2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_completed(r2);
+
+    // Check cutting of the beginning and end
+    assert((long)test_write(cli, 0, 32768, 0x55, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    assert((long)test_write(cli, 32768, 32768, 0x56, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    assert((long)test_write(cli, 16384, 32768, 0x57, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    assert((long)test_write(cli, 16384+4096, 32768-4096, 0x58, NULL, true) == 1);
+    check_op_count(cli, 1, 0);
+    r2 = test_sync(cli);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 65536), 0);
+    check_op_count(cli, 1, 1);
+    can_complete(r2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_completed(r2);
+
+    // Free client
+    delete cli;
+    delete tfd;
+    printf("[ok] writeback test\n");
+}
+
 int main(int narg, char *args[])
 {
     test1();
     test2();
+    test_writeback();
     return 0;
 }

src/timerfd_manager.cpp

@@ -129,6 +129,7 @@ again:
         if (exp.it_value.tv_sec < 0 || exp.it_value.tv_sec == 0 && exp.it_value.tv_nsec <= 0)
         {
             // It already happened
+            // FIXME: Postpone to setImmediate/BH to avoid reenterability problems
             trigger_nearest();
             goto again;
         }

src/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 0.9.3
+Version: 1.0.0
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

src/vitastor_c.cpp

@@ -215,7 +215,7 @@ void vitastor_c_uring_wait_events(vitastor_c *client)
     client->ringloop->wait();
 }
 
-bool vitastor_c_uring_has_work(vitastor_c *client)
+int vitastor_c_uring_has_work(vitastor_c *client)
 {
     return client->ringloop->has_work();
 }