Release 0.6.16

- Implement `vitastor-cli status` (print cluster status) command
- Add a new `make-osd-hybrid.js` script to quickly prepare a lot of hybrid (HDD+SSD) OSDs
- Implement snapshot deletion for Cinder driver (only works in a healthy cluster)
- Fix a huge :) bug causing reads to return all zeroes during rebalance. Add a test to prevent it in the future
- Disconnect NBD proxy correctly without leaving a zombie [vitastor-nbd] process in D state
- Fix a rare write hang appearing with small write throttling enabled

CMakeLists.txt

@@ -2,6 +2,6 @@ cmake_minimum_required(VERSION 2.8)
 
 project(vitastor)
 
-set(VERSION "0.6.12")
+set(VERSION "0.6.16")
 
 add_subdirectory(src)

README-ru.md

@@ -55,11 +55,11 @@ Vitastor на данный момент находится в статусе п
 
 ## Планы развития
 
-- Поддержка удаления снапшотов (слияния слоёв)
 - Более корректные скрипты разметки дисков и автоматического запуска OSD
 - Другие инструменты администрирования
 - Плагины для OpenNebula и других облачных систем
 - iSCSI-прокси
+- Упрощённый NFS прокси
 - Более быстрое переключение при отказах
 - Фоновая проверка целостности без контрольных сумм (сверка реплик)
 - Контрольные суммы
@@ -414,7 +414,14 @@ Vitastor с однопоточной NBD прокси на том же стен
     "osd_network": "10.200.1.0/24"
   }
   ```
-- Создайте юниты systemd для OSD: `/usr/lib/vitastor/make-osd.sh /dev/disk/by-partuuid/XXX [/dev/disk/by-partuuid/YYY ...]`
+- Инициализуйте OSD:
+  - SSD: `/usr/lib/vitastor/make-osd.sh /dev/disk/by-partuuid/XXX [/dev/disk/by-partuuid/YYY ...]`
+  - Гибридные, HDD+SSD: `/usr/lib/vitastor/mon/make-osd-hybrid.js /dev/sda /dev/sdb ...` - передайте
+    все ваши SSD и HDD скрипту в командной строке подряд, скрипт автоматически выделит разделы под
+    журналы на SSD и данные на HDD. Скрипт пропускает HDD, на которых уже есть разделы
+    или вообще какие-то данные, поэтому если диски непустые, сначала очистите их с помощью
+    `wipefs -a`. SSD с таблицей разделов не пропускаются, но так как скрипт создаёт новые разделы
+    для журналов, на SSD должно быть доступно свободное нераспределённое место.
 - Вы можете менять параметры OSD в юнитах systemd или в `vitastor.conf`. Смысл некоторых параметров:
   - `disable_data_fsync 1` - отключает fsync, используется с SSD с конденсаторами.
   - `immediate_commit all` - используется с SSD с конденсаторами.

README.md

@@ -49,11 +49,11 @@ breaking changes in the future. However, the following is implemented:
 
 ## Roadmap
 
-- Snapshot deletion (layer merge) support
 - Better OSD creation and auto-start tools
 - Other administrative tools
-- Plugins for OpenNebula, Proxmox and other cloud systems
+- Plugins for OpenNebula and other cloud systems
 - iSCSI proxy
+- Simplified NFS proxy
 - Faster failover
 - Scrubbing without checksums (verification of replicas)
 - Checksums
@@ -367,7 +367,13 @@ and calculate disk offsets almost by hand. This will be fixed in near future.
     "osd_network": "10.200.1.0/24"
   }
   ```
-- Create systemd units for your OSDs: `/usr/lib/vitastor/mon/make-osd.sh /dev/disk/by-partuuid/XXX [/dev/disk/by-partuuid/YYY ...]`
+- Initialize OSDs:
+  - Simplest, SSD-only: `/usr/lib/vitastor/mon/make-osd.sh /dev/disk/by-partuuid/XXX [/dev/disk/by-partuuid/YYY ...]`
+  - Hybrid, HDD+SSD: `/usr/lib/vitastor/mon/make-osd-hybrid.js /dev/sda /dev/sdb ...` - pass all your
+    devices (HDD and SSD) to this script - it will partition disks and initialize journals on its own.
+    This script skips HDDs which are already partitioned so if you want to use non-empty disks for
+    Vitastor you should first wipe them with `wipefs -a`. SSDs with GPT partition table are not skipped,
+    but some free unpartitioned space must be available because the script creates new partitions for journals.
 - You can change OSD configuration in units or in `vitastor.conf`. Notable configuration variables:
   - `disable_data_fsync 1` - only safe with server-grade drives with capacitors.
   - `immediate_commit all` - use this if all your drives are server-grade.

csi/Makefile

@@ -1,4 +1,4 @@
-VERSION ?= v0.6.12
+VERSION ?= v0.6.16
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v0.6.12
+          image: vitalif/vitastor-csi:v0.6.16
           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:v0.6.12
+          image: vitalif/vitastor-csi:v0.6.16
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "0.6.12"
+    vitastorCSIDriverVersion = "0.6.16"
 )
 
 // Config struct fills the parameters of request or user input

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (0.6.12-1) unstable; urgency=medium
+vitastor (0.6.16-1) unstable; urgency=medium
 
   * RDMA support
   * Bugfixes

debian/vitastor.Dockerfile

@@ -33,8 +33,8 @@ RUN set -e -x; \
     mkdir -p /root/packages/vitastor-$REL; \
     rm -rf /root/packages/vitastor-$REL/*; \
     cd /root/packages/vitastor-$REL; \
-    cp -r /root/vitastor vitastor-0.6.12; \
-    cd vitastor-0.6.12; \
+    cp -r /root/vitastor vitastor-0.6.16; \
+    cd vitastor-0.6.16; \
     ln -s /root/fio-build/fio-*/ ./fio; \
     FIO=$(head -n1 fio/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     ls /usr/include/linux/raw.h || cp ./debian/raw.h /usr/include/linux/raw.h; \
@@ -47,8 +47,8 @@ RUN set -e -x; \
     rm -rf a b; \
     echo "dep:fio=$FIO" > debian/fio_version; \
     cd /root/packages/vitastor-$REL; \
-    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.6.12.orig.tar.xz vitastor-0.6.12; \
-    cd vitastor-0.6.12; \
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.6.16.orig.tar.xz vitastor-0.6.16; \
+    cd vitastor-0.6.16; \
     V=$(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/params/common.yml

@@ -0,0 +1,35 @@
+- name: config_path
+  type: string
+  default: "/etc/vitastor/vitastor.conf"
+  info: |
+    Path to the JSON configuration file. Configuration file is optional,
+    a non-existing configuration file does not prevent Vitastor from
+    running if required parameters are specified.
+  info_ru: |
+    Путь к файлу конфигурации в формате JSON. Файл конфигурации необязателен,
+    без него Vitastor тоже будет работать, если переданы необходимые параметры.
+- name: etcd_address
+  type: string or array of strings
+  type_ru: строка или массив строк
+  info: |
+    etcd connection endpoint(s). Multiple endpoints may be delimited by "," or
+    specified in a JSON array `["10.0.115.10:2379/v3","10.0.115.11:2379/v3"]`.
+    Note that https is not supported for etcd connections yet.
+  info_ru: |
+    Адрес(а) подключения к etcd. Несколько адресов могут разделяться запятой
+    или указываться в виде JSON-массива `["10.0.115.10:2379/v3","10.0.115.11:2379/v3"]`.
+- name: etcd_prefix
+  type: string
+  default: "/vitastor"
+  info: |
+    Prefix for all keys in etcd used by Vitastor. You can change prefix and, for
+    example, use a single etcd cluster for multiple Vitastor clusters.
+  info_ru: |
+    Префикс для ключей etcd, которые использует Vitastor. Вы можете задать другой
+    префикс, например, чтобы запустить несколько кластеров Vitastor с одним
+    кластером etcd.
+- name: log_level
+  type: int
+  default: 0
+  info: Log level. Raise if you want more verbose output.
+  info_ru: Уровень логгирования. Повысьте, если хотите более подробный вывод.

docs/params/layout-cluster.yml

@@ -0,0 +1,200 @@
+- name: block_size
+  type: int
+  default: 131072
+  info: |
+    Size of objects (data blocks) into which all physical and virtual drives are
+    subdivided in Vitastor. One of current main settings in Vitastor, affects
+    memory usage, write amplification and I/O load distribution effectiveness.
+
+    Recommended default block size is 128 KB for SSD and 4 MB for HDD. In fact,
+    it's possible to use 4 MB for SSD too - it will lower memory usage, but
+    may increase average WA and reduce linear performance.
+
+    OSDs with different block sizes (for example, SSD and SSD+HDD OSDs) can
+    currently coexist in one etcd instance only within separate Vitastor
+    clusters with different etcd_prefix'es.
+
+    Also block size can't be changed after OSD initialization without losing
+    data.
+
+    You must always specify block_size in etcd in /vitastor/config/global if
+    you change it so all clients can know about it.
+
+    OSD memory usage is roughly (SIZE / BLOCK * 68 bytes) which is roughly
+    544 MB per 1 TB of used disk space with the default 128 KB block size.
+  info_ru: |
+    Размер объектов (блоков данных), на которые делятся физические и виртуальные
+    диски в Vitastor. Одна из ключевых на данный момент настроек, влияет на
+    потребление памяти, объём избыточной записи (write amplification) и
+    эффективность распределения нагрузки по OSD.
+
+    Рекомендуемые по умолчанию размеры блока - 128 килобайт для SSD и 4
+    мегабайта для HDD. В принципе, для SSD можно тоже использовать 4 мегабайта,
+    это понизит использование памяти, но ухудшит распределение нагрузки и в
+    среднем увеличит WA.
+
+    OSD с разными размерами блока (например, SSD и SSD+HDD OSD) на данный
+    момент могут сосуществовать в рамках одного etcd только в виде двух независимых
+    кластеров Vitastor с разными etcd_prefix.
+
+    Также размер блока нельзя менять после инициализации OSD без потери данных.
+
+    Если вы меняете размер блока, обязательно прописывайте его в etcd в
+    /vitastor/config/global, дабы все клиенты его знали.
+
+    Потребление памяти OSD составляет примерно (РАЗМЕР / БЛОК * 68 байт),
+    т.е. примерно 544 МБ памяти на 1 ТБ занятого места на диске при
+    стандартном 128 КБ блоке.
+- name: bitmap_granularity
+  type: int
+  default: 4096
+  info: |
+    Required virtual disk write alignment ("sector size"). Must be a multiple
+    of disk_alignment. It's called bitmap granularity because Vitastor tracks
+    an allocation bitmap for each object containing 2 bits per each
+    (bitmap_granularity) bytes.
+
+    This parameter can't be changed after OSD initialization without losing
+    data. Also it's fixed for the whole Vitastor cluster i.e. two different
+    values can't be used in a single Vitastor cluster.
+
+    Clients MUST be aware of this parameter value, so put it into etcd key
+    /vitastor/config/global if you change it for any reason.
+  info_ru: |
+    Требуемое выравнивание записи на виртуальные диски (размер их "сектора").
+    Должен быть кратен disk_alignment. Называется гранулярностью битовой карты
+    потому, что Vitastor хранит битовую карту для каждого объекта, содержащую
+    по 2 бита на каждые (bitmap_granularity) байт.
+
+    Данный параметр нельзя менять после инициализации OSD без потери данных.
+    Также он фиксирован для всего кластера Vitastor, т.е. разные значения
+    не могут сосуществовать в одном кластере.
+
+    Клиенты ДОЛЖНЫ знать правильное значение этого параметра, так что если вы
+    его меняете, обязательно прописывайте изменённое значение в etcd в ключ
+    /vitastor/config/global.
+- name: immediate_commit
+  type: string
+  default: false
+  info: |
+    Another parameter which is really important for performance.
+
+    Desktop SSDs are very fast (100000+ iops) for simple random writes
+    without cache flush. However, they are really slow (only around 1000 iops)
+    if you try to fsync() each write, that is, when you want to guarantee that
+    each change gets immediately persisted to the physical media.
+
+    Server-grade SSDs with "Advanced/Enhanced Power Loss Protection" or with
+    "Supercapacitor-based Power Loss Protection", on the other hand, are equally
+    fast with and without fsync because their cache is protected from sudden
+    power loss by a built-in supercapacitor-based "UPS".
+
+    Some software-defined storage systems always fsync each write and thus are
+    really slow when used with desktop SSDs. Vitastor, however, can also
+    efficiently utilize desktop SSDs by postponing fsync until the client calls
+    it explicitly.
+
+    This is what this parameter regulates. When it's set to "all" the whole
+    Vitastor cluster commits each change to disks immediately and clients just
+    ignore fsyncs because they know for sure that they're unneeded. This reduces
+    the amount of network roundtrips performed by clients and improves
+    performance. So it's always better to use server grade SSDs with
+    supercapacitors even with Vitastor, especially given that they cost only
+    a bit more than desktop models.
+
+    There is also a common SATA SSD (and HDD too!) firmware bug (or feature)
+    that makes server SSDs which have supercapacitors slow with fsync. To check
+    if your SSDs are affected, compare benchmark results from `fio -name=test
+    -ioengine=libaio -direct=1 -bs=4k -rw=randwrite -iodepth=1` with and without
+    `-fsync=1`. Results should be the same. If fsync=1 result is worse you can
+    try to work around this bug by "disabling" drive write-back cache by running
+    `hdparm -W 0 /dev/sdXX` or `echo write through > /sys/block/sdXX/device/scsi_disk/*/cache_type`
+    (IMPORTANT: don't mistake it with `/sys/block/sdXX/queue/write_cache` - it's
+    unsafe to change by hand). The same may apply to newer HDDs with internal
+    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).
+
+    This parameter must be set both in etcd in /vitastor/config/global and in
+    OSD command line or configuration. Setting it to "all" or "small" requires
+    enabling disable_journal_fsync and disable_meta_fsync, setting it to "all"
+    also requires enabling disable_data_fsync.
+
+    TLDR: For optimal performance, set immediate_commit to "all" if you only use
+    SSDs with supercapacitor-based power loss protection (nonvolatile
+    write-through cache) for both data and journals in the whole Vitastor
+    cluster. Set it to "small" if you only use such SSDs for journals. Leave
+    empty if your drives have write-back cache.
+  info_ru: |
+    Ещё один важный для производительности параметр.
+
+    Модели SSD для настольных компьютеров очень быстрые (100000+ операций в
+    секунду) при простой случайной записи без сбросов кэша. Однако они очень
+    медленные (всего порядка 1000 iops), если вы пытаетесь сбрасывать кэш после
+    каждой записи, то есть, если вы пытаетесь гарантировать, что каждое
+    изменение физически записывается в энергонезависимую память.
+
+    С другой стороны, серверные SSD с конденсаторами - функцией, называемой
+    "Advanced/Enhanced Power Loss Protection" или просто "Supercapacitor-based
+    Power Loss Protection" - одинаково быстрые и со сбросом кэша, и без
+    него, потому что их кэш защищён от потери питания встроенным "источником
+    бесперебойного питания" на основе суперконденсаторов и на самом деле они
+    его никогда не сбрасывают.
+
+    Некоторые программные СХД всегда сбрасывают кэши дисков при каждой записи
+    и поэтому работают очень медленно с настольными SSD. Vitastor, однако, может
+    откладывать fsync до явного его вызова со стороны клиента и таким образом
+    эффективно утилизировать настольные SSD.
+
+    Данный параметр влияет как раз на это. Когда он установлен в значение "all",
+    весь кластер Vitastor мгновенно фиксирует каждое изменение на физические
+    носители и клиенты могут просто игнорировать запросы fsync, т.к. они точно
+    знают, что fsync-и не нужны. Это уменьшает число необходимых обращений к OSD
+    по сети и улучшает производительность. Поэтому даже с Vitastor лучше всегда
+    использовать только серверные модели SSD с суперконденсаторами, особенно
+    учитывая то, что стоят они ненамного дороже настольных.
+
+    Также в прошивках SATA SSD (и даже HDD!) очень часто встречается либо баг,
+    либо просто особенность логики, из-за которой серверные SSD, имеющие
+    конденсаторы и защиту от потери питания, всё равно медленно работают с
+    fsync. Чтобы понять, подвержены ли этой проблеме ваши SSD, сравните
+    результаты тестов `fio -name=test -ioengine=libaio -direct=1 -bs=4k
+    -rw=randwrite -iodepth=1` без и с опцией `-fsync=1`. Результаты должны
+    быть одинаковые. Если результат с `fsync=1` хуже, вы можете попробовать
+    обойти проблему, "отключив" кэш записи диска командой `hdparm -W 0 /dev/sdXX`
+    либо `echo write through > /sys/block/sdXX/device/scsi_disk/*/cache_type`
+    (ВАЖНО: не перепутайте с `/sys/block/sdXX/queue/write_cache` - этот параметр
+    менять руками небезопасно). Такая же проблема может встречаться и в новых
+    HDD-дисках с внутренним SSD или "медиа" кэшем - например, она встречается во
+    многих дисках Seagate EXOS (у них есть внутренний SSD-кэш, хотя это и не
+    указано в спецификациях).
+
+    Данный параметр нужно указывать и в etcd в /vitastor/config/global, и в
+    командной строке или конфигурации OSD. Значения "all" и "small" требуют
+    включения disable_journal_fsync и disable_meta_fsync, значение "all" также
+    требует включения disable_data_fsync.
+
+    Итого, вкратце: для оптимальной производительности установите
+    immediate_commit в значение "all", если вы используете в кластере только SSD
+    с суперконденсаторами и для данных, и для журналов. Если вы используете
+    такие SSD для всех журналов, но не для данных - можете установить параметр
+    в "small". Если и какие-то из дисков журналов имеют волатильный кэш записи -
+    оставьте параметр пустым.
+- name: client_dirty_limit
+  type: int
+  default: 33554432
+  info: |
+    Without immediate_commit=all this parameter sets the limit of "dirty"
+    (not committed by fsync) data allowed by the client before forcing an
+    additional fsync and committing the data. Also note that the client always
+    holds a copy of uncommitted data in memory so this setting also affects
+    RAM usage of clients.
+
+    This parameter doesn't affect OSDs themselves.
+  info_ru: |
+    При работе без immediate_commit=all - это лимит объёма "грязных" (не
+    зафиксированных fsync-ом) данных, при достижении которого клиент будет
+    принудительно вызывать fsync и фиксировать данные. Также стоит иметь в виду,
+    что в этом случае до момента fsync клиент хранит копию незафиксированных
+    данных в памяти, то есть, настройка влияет на потребление памяти клиентами.
+
+    Параметр не влияет на сами OSD.

docs/params/layout-osd.yml

@@ -0,0 +1,205 @@
+- name: data_device
+  type: string
+  info: |
+    Path to the block device to use for data. It's highly recommendded to use
+    stable paths for all device names: `/dev/disk/by-partuuid/xxx...` instead
+    of just `/dev/sda` or `/dev/nvme0n1` to not mess up after server restart.
+    Files can also be used instead of block devices, but this is implemented
+    only for testing purposes and not for production.
+  info_ru: |
+    Путь к диску (блочному устройству) для хранения данных. Крайне рекомендуется
+    использовать стабильные пути: `/dev/disk/by-partuuid/xxx...` вместо простых
+    `/dev/sda` или `/dev/nvme0n1`, чтобы пути не могли спутаться после
+    перезагрузки сервера. Также вместо блочных устройств можно указывать файлы,
+    но это реализовано только для тестирования, а не для боевой среды.
+- name: meta_device
+  type: string
+  info: |
+    Path to the block device to use for the metadata. Metadata must be on a fast
+    SSD or performance will suffer. If this option is skipped, `data_device` is
+    used for the metadata.
+  info_ru: |
+    Путь к диску метаданных. Метаданные должны располагаться на быстром
+    SSD-диске, иначе производительность пострадает. Если эта опция не указана,
+    для метаданных используется `data_device`.
+- name: journal_device
+  type: string
+  info: |
+    Path to the block device to use for the journal. Journal must be on a fast
+    SSD or performance will suffer. If this option is skipped, `meta_device` is
+    used for the journal, and if it's also empty, journal is put on
+    `data_device`. It's almost always fine to put metadata and journal on the
+    same device, in this case you only need to set `meta_device`.
+  info_ru: |
+    Путь к диску журнала. Журнал должен располагаться на быстром SSD-диске,
+    иначе производительность пострадает. Если эта опция не указана,
+    для журнала используется `meta_device`, если же пуста и она, журнал
+    располагается на `data_device`. Нормально располагать журнал и метаданные
+    на одном устройстве, в этом случае достаточно указать только `meta_device`.
+- name: journal_offset
+  type: int
+  default: 0
+  info: Offset on the device in bytes where the journal is stored.
+  info_ru: Смещение на устройстве в байтах, по которому располагается журнал.
+- name: journal_size
+  type: int
+  info: |
+    Journal size in bytes. Doesn't have to be large, 16-32 MB is usually fine.
+    By default, the whole journal device will be used for the journal. You must
+    set it to some value manually (or use make-osd.sh) if you colocate the
+    journal with data or metadata.
+  info_ru: |
+    Размер журнала в байтах. Большим быть не обязан, 16-32 МБ обычно достаточно.
+    По умолчанию для журнала используется всё устройство журнала. Если же вы
+    размещаете журнал на устройстве данных или метаданных, то вы должны
+    установить эту опцию в какое-то значение сами (или использовать скрипт
+    make-osd.sh).
+- name: meta_offset
+  type: int
+  default: 0
+  info: |
+    Offset on the device in bytes where the metadata area is stored.
+    Again, set it to something if you colocate metadata with journal or data.
+  info_ru: |
+    Смещение на устройстве в байтах, по которому располагаются метаданные.
+    Эту опцию нужно задать, если метаданные у вас хранятся на том же
+    устройстве, что данные или журнал.
+- name: data_offset
+  type: int
+  default: 0
+  info: |
+    Offset on the device in bytes where the data area is stored.
+    Again, set it to something if you colocate data with journal or metadata.
+  info_ru: |
+    Смещение на устройстве в байтах, по которому располагаются данные.
+    Эту опцию нужно задать, если данные у вас хранятся на том же
+    устройстве, что метаданные или журнал.
+- name: data_size
+  type: int
+  info: |
+    Data area size in bytes. By default, the whole data device up to the end
+    will be used for the data area, but you can restrict it if you want to use
+    a smaller part. Note that there is no option to set metadata area size -
+    it's derived from the data area size.
+  info_ru: |
+    Размер области данных в байтах. По умолчанию под данные будет использована
+    вся доступная область устройства данных до конца устройства, но вы можете
+    использовать эту опцию, чтобы ограничить её меньшим размером. Заметьте, что
+    опции размера области метаданных нет - она вычисляется из размера области
+    данных автоматически.
+- name: meta_block_size
+  type: int
+  default: 4096
+  info: |
+    Physical block size of the metadata device. 4096 for most current
+    HDDs and SSDs.
+  info_ru: |
+    Размер физического блока устройства метаданных. 4096 для большинства
+    современных SSD и HDD.
+- name: journal_block_size
+  type: int
+  default: 4096
+  info: |
+    Physical block size of the journal device. Must be a multiple of
+    `disk_alignment`. 4096 for most current HDDs and SSDs.
+  info_ru: |
+    Размер физического блока устройства журнала. Должен быть кратен
+    `disk_alignment`. 4096 для большинства современных SSD и HDD.
+- name: disable_data_fsync
+  type: bool
+  default: false
+  info: |
+    Do not issue fsyncs to the data device, i.e. do not flush its cache.
+    Safe ONLY if your data device has write-through cache. If you disable
+    the cache yourself using `hdparm` or `scsi_disk/cache_type` then make sure
+    that the cache disable command is run every time before starting Vitastor
+    OSD, for example, in the systemd unit. See also `immediate_commit` option
+    for the instructions to disable cache and how to benefit from it.
+  info_ru: |
+    Не отправлять fsync-и устройству данных, т.е. не сбрасывать его кэш.
+    Безопасно, ТОЛЬКО если ваше устройство данных имеет кэш со сквозной
+    записью (write-through). Если вы отключаете кэш через `hdparm` или
+    `scsi_disk/cache_type`, то удостоверьтесь, что команда отключения кэша
+    выполняется перед каждым запуском Vitastor OSD, например, в systemd unit-е.
+    Смотрите также опцию `immediate_commit` для инструкций по отключению кэша
+    и о том, как из этого извлечь выгоду.
+- name: disable_meta_fsync
+  type: bool
+  default: false
+  info: |
+    Same as disable_data_fsync, but for the metadata device. If the metadata
+    device is not set or if the data device is used for the metadata the option
+    is ignored and disable_data_fsync value is used instead of it.
+  info_ru: |
+    То же, что disable_data_fsync, но для устройства метаданных. Если устройство
+    метаданных не задано или если оно равно устройству данных, значение опции
+    игнорируется и вместо него используется значение опции disable_data_fsync.
+- name: disable_journal_fsync
+  type: bool
+  default: false
+  info: |
+    Same as disable_data_fsync, but for the journal device. If the journal
+    device is not set or if the metadata device is used for the journal the
+    option is ignored and disable_meta_fsync value is used instead of it. If
+    the same device is used for data, metadata and journal the option is also
+    ignored and disable_data_fsync value is used instead of it.
+  info_ru: |
+    То же, что disable_data_fsync, но для устройства журнала. Если устройство
+    журнала не задано или если оно равно устройству метаданных, значение опции
+    игнорируется и вместо него используется значение опции disable_meta_fsync.
+    Если одно и то же устройство используется и под данные, и под журнал, и под
+    метаданные - значение опции также игнорируется и вместо него используется
+    значение опции disable_data_fsync.
+- name: disable_device_lock
+  type: bool
+  default: false
+  info: |
+    Do not lock data, metadata and journal block devices exclusively with
+    flock(). Though it's not recommended, but you can use it you want to run
+    multiple OSD with a single device and different offsets, without using
+    partitions.
+  info_ru: |
+    Не блокировать устройства данных, метаданных и журнала от открытия их
+    другими OSD с помощью flock(). Так делать не рекомендуется, но теоретически
+    вы можете это использовать, чтобы запускать несколько OSD на одном
+    устройстве с разными смещениями и без использования разделов.
+- name: disk_alignment
+  type: int
+  default: 4096
+  info: |
+    Required physical disk write alignment. Most current SSD and HDD drives
+    use 4 KB physical sectors even if they report 512 byte logical sector
+    size, so 4 KB is a good default setting.
+
+    Note, however, that physical sector size also affects WA, because with block
+    devices it's impossible to write anything smaller than a block. So, when
+    Vitastor has to write a single metadata entry that's only about 32 bytes in
+    size, it actually has to write the whole 4 KB sector.
+
+    Because of this it can actually be beneficial to use SSDs which work well
+    with 512 byte sectors and use 512 byte disk_alignment, journal_block_size
+    and meta_block_size. But the only SSD that may fit into this category is
+    Intel Optane (probably, not tested yet).
+
+    Clients don't need to be aware of disk_alignment, so it's not required to
+    put a modified value into etcd key /vitastor/config/global.
+  info_ru: |
+    Требуемое выравнивание записи на физические диски. Почти все современные
+    SSD и HDD диски используют 4 КБ физические секторы, даже если показывают
+    логический размер сектора 512 байт, поэтому 4 КБ - хорошее значение по
+    умолчанию.
+
+    Однако стоит понимать, что физический размер сектора тоже влияет на
+    избыточную запись (WA), потому что ничего меньше блока (сектора) на блочное
+    устройство записать невозможно. Таким образом, когда Vitastor-у нужно
+    записать на диск всего лишь одну 32-байтную запись метаданных, фактически
+    приходится перезаписывать 4 КБ сектор целиком.
+
+    Поэтому, на самом деле, может быть выгодно найти SSD, хорошо работающие с
+    меньшими, 512-байтными, блоками и использовать 512-байтные disk_alignment,
+    journal_block_size и meta_block_size. Однако единственные SSD, которые
+    теоретически могут попасть в эту категорию - это Intel Optane (но и это
+    пока не проверялось автором).
+
+    Клиентам не обязательно знать про disk_alignment, так что помещать значение
+    этого параметра в etcd в /vitastor/config/global не нужно.

docs/params/monitor.yml

@@ -0,0 +1,65 @@
+- name: etcd_mon_ttl
+  type: sec
+  min: 10
+  default: 30
+  info: Monitor etcd lease refresh interval in seconds
+  info_ru: Интервал обновления etcd резервации (lease) монитором
+- name: etcd_mon_timeout
+  type: ms
+  default: 1000
+  info: etcd request timeout used by monitor
+  info_ru: Таймаут выполнения запросов к etcd от монитора
+- name: etcd_mon_retries
+  type: int
+  default: 5
+  info: Maximum number of attempts for one monitor etcd request
+  info_ru: Максимальное число попыток выполнения запросов к etcd монитором
+- name: mon_change_timeout
+  type: ms
+  min: 100
+  default: 1000
+  info: Optimistic retry interval for monitor etcd modification requests
+  info_ru: Время повтора при коллизиях при запросах модификации в etcd, производимых монитором
+- name: mon_stats_timeout
+  type: ms
+  min: 100
+  default: 1000
+  info: |
+    Interval for monitor to wait before updating aggregated statistics in
+    etcd after receiving OSD statistics updates
+  info_ru: |
+    Интервал, который монитор ожидает при изменении статистики по отдельным
+    OSD перед обновлением агрегированной статистики в etcd
+- name: osd_out_time
+  type: sec
+  default: 600
+  info: |
+    Time after which a failed OSD is removed from the data distribution.
+    I.e. time which the monitor waits before attempting to restore data
+    redundancy using other OSDs.
+  info_ru: |
+    Время, через которое отключенный OSD исключается из распределения данных.
+    То есть, время, которое монитор ожидает перед попыткой переместить данные
+    на другие OSD и таким образом восстановить избыточность хранения.
+- name: placement_levels
+  type: json
+  default: '`{"host":100,"osd":101}`'
+  info: |
+    Levels for the placement tree. You can define arbitrary tree levels by
+    defining them in this parameter. The configuration parameter value should
+    contain a JSON object with level names as keys and integer priorities as
+    values.  Smaller priority means higher level in tree. For example,
+    "datacenter" should have smaller priority than "osd". "host" and "osd"
+    levels are always predefined and can't be removed. If one of them is not
+    present in the configuration, then it is defined with the default priority
+    (100 for "host", 101 for "osd").
+  info_ru: |
+    Определения уровней для дерева размещения OSD. Вы можете определять
+    произвольные уровни, помещая их в данный параметр конфигурации. Значение
+    параметра должно содержать JSON-объект, ключи которого будут являться
+    названиями уровней, а значения - целочисленными приоритетами. Меньшие
+    приоритеты соответствуют верхним уровням дерева. Например, уровень
+    "датацентр" должен иметь меньший приоритет, чем "OSD". Уровни с названиями
+    "host" и "osd" являются предопределёнными и не могут быть удалены. Если
+    один из них отсутствует в конфигурации, он доопределяется с приоритетом по
+    умолчанию (100 для уровня "host", 101 для "osd").

docs/params/network.yml

@@ -0,0 +1,225 @@
+- name: tcp_header_buffer_size
+  type: int
+  default: 65536
+  info: |
+    Size of the buffer used to read data using an additional copy. Vitastor
+    packet headers are 128 bytes, payload is always at least 4 KB, so it is
+    usually beneficial to try to read multiple packets at once even though
+    it requires to copy the data an additional time. The rest of each packet
+    is received without an additional copy. You can try to play with this
+    parameter and see how it affects random iops and linear bandwidth if you
+    want.
+  info_ru: |
+    Размер буфера для чтения данных с дополнительным копированием. Пакеты
+    Vitastor содержат 128-байтные заголовки, за которыми следуют данные размером
+    от 4 КБ и для мелких операций ввода-вывода обычно выгодно за 1 вызов читать
+    сразу несколько пакетов, даже не смотря на то, что это требует лишний раз
+    скопировать данные. Часть каждого пакета за пределами значения данного
+    параметра читается без дополнительного копирования. Вы можете попробовать
+    поменять этот параметр и посмотреть, как он влияет на производительность
+    случайного и линейного доступа.
+- name: use_sync_send_recv
+  type: bool
+  default: false
+  info: |
+    If true, synchronous send/recv syscalls are used instead of io_uring for
+    socket communication. Useless for OSDs because they require io_uring anyway,
+    but may be required for clients with old kernel versions.
+  info_ru: |
+    Если установлено в истину, то вместо io_uring для передачи данных по сети
+    будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
+    это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
+    принципе, это может применяться для клиентов со старыми версиями ядра.
+- name: use_rdma
+  type: bool
+  default: true
+  info: |
+    Try to use RDMA for communication if it's available. Disable if you don't
+    want Vitastor to use RDMA. RDMA increases the performance, but TCP-only
+    clients can still talk to an RDMA-enabled cluster, so you don't need to
+    make sure that all clients support RDMA when enabling it.
+  info_ru: |
+    Пытаться использовать RDMA для связи при наличии доступных устройств.
+    Отключите, если вы не хотите, чтобы Vitastor использовал RDMA.
+    RDMA улучшает производительность, но 
+    Клиенты и клиентов and TCP-only clients in the cluster at the
+    same time - TCP-only clients are still able to use an RDMA-enabled cluster.
+- name: rdma_device
+  type: string
+  info: |
+    RDMA device name to use for Vitastor OSD communications (for example,
+    "rocep5s0f0"). Please note that Vitastor RDMA requires Implicit On-Demand
+    Paging (Implicit ODP) and Scatter/Gather (SG) support from the RDMA device
+    to work. For example, Mellanox ConnectX-3 and older adapters don't have
+    Implicit ODP, so they're unsupported by Vitastor. Run `ibv_devinfo -v` as
+    root to list available RDMA devices and their features.
+  info_ru: |
+    Название RDMA-устройства для связи с Vitastor OSD (например, "rocep5s0f0").
+    Имейте в виду, что поддержка RDMA в Vitastor требует функций устройства
+    Implicit On-Demand Paging (Implicit ODP) и Scatter/Gather (SG). Например,
+    адаптеры Mellanox ConnectX-3 и более старые не поддерживают Implicit ODP и
+    потому не поддерживаются в Vitastor. Запустите `ibv_devinfo -v` от имени
+    суперпользователя, чтобы посмотреть список доступных RDMA-устройств, их
+    параметры и возможности.
+- name: rdma_port_num
+  type: int
+  default: 1
+  info: |
+    RDMA device port number to use. Only for devices that have more than 1 port.
+    See `phys_port_cnt` in `ibv_devinfo -v` output to determine how many ports
+    your device has.
+  info_ru: |
+    Номер порта RDMA-устройства, который следует использовать. Имеет смысл
+    только для устройств, у которых более 1 порта. Чтобы узнать, сколько портов
+    у вашего адаптера, посмотрите `phys_port_cnt` в выводе команды
+    `ibv_devinfo -v`.
+- name: rdma_gid_index
+  type: int
+  default: 0
+  info: |
+    Global address identifier index of the RDMA device to use. Different GID
+    indexes may correspond to different protocols like RoCEv1, RoCEv2 and iWARP.
+    Search for "GID" in `ibv_devinfo -v` output to determine which GID index
+    you need.
+
+    **IMPORTANT:** If you want to use RoCEv2 (as recommended) then the correct
+    rdma_gid_index is usually 1 (IPv6) or 3 (IPv4).
+  info_ru: |
+    Номер глобального идентификатора адреса RDMA-устройства, который следует
+    использовать. Разным gid_index могут соответствовать разные протоколы связи:
+    RoCEv1, RoCEv2, iWARP. Чтобы понять, какой нужен вам - смотрите строчки со
+    словом "GID" в выводе команды `ibv_devinfo -v`.
+
+    **ВАЖНО:** Если вы хотите использовать RoCEv2 (как мы и рекомендуем), то
+    правильный rdma_gid_index, как правило, 1 (IPv6) или 3 (IPv4).
+- name: rdma_mtu
+  type: int
+  default: 4096
+  info: |
+    RDMA Path MTU to use. Must be 1024, 2048 or 4096. There is usually no
+    sense to change it from the default 4096.
+  info_ru: |
+    Максимальная единица передачи (Path MTU) для RDMA. Должно быть равно 1024,
+    2048 или 4096. Обычно нет смысла менять значение по умолчанию, равное 4096.
+- name: rdma_max_sge
+  type: int
+  default: 128
+  info: |
+    Maximum number of scatter/gather entries to use for RDMA. OSDs negotiate
+    the actual value when establishing connection anyway, so it's usually not
+    required to change this parameter.
+  info_ru: |
+    Максимальное число записей разделения/сборки (scatter/gather) для RDMA.
+    OSD в любом случае согласовывают реальное значение при установке соединения,
+    так что менять этот параметр обычно не нужно.
+- name: rdma_max_msg
+  type: int
+  default: 1048576
+  info: Maximum size of a single RDMA send or receive operation in bytes.
+  info_ru: Максимальный размер одной RDMA-операции отправки или приёма.
+- name: rdma_max_recv
+  type: int
+  default: 8
+  info: |
+    Maximum number of parallel RDMA receive operations. Note that this number
+    of receive buffers `rdma_max_msg` in size are allocated for each client,
+    so this setting actually affects memory usage. This is because RDMA receive
+    operations are (sadly) still not zero-copy in Vitastor. It may be fixed in
+    later versions.
+  info_ru: |
+    Максимальное число параллельных RDMA-операций получения данных. Следует
+    иметь в виду, что данное число буферов размером `rdma_max_msg` выделяется
+    для каждого подключённого клиентского соединения, так что данная настройка
+    влияет на потребление памяти. Это так потому, что RDMA-приём данных в
+    Vitastor, увы, всё равно не является zero-copy, т.е. всё равно 1 раз
+    копирует данные в памяти. Данная особенность, возможно, будет исправлена в
+    более новых версиях Vitastor.
+- name: peer_connect_interval
+  type: sec
+  min: 1
+  default: 5
+  info: Interval before attempting to reconnect to an unavailable OSD.
+  info_ru: Время ожидания перед повторной попыткой соединиться с недоступным OSD.
+- name: peer_connect_timeout
+  type: sec
+  min: 1
+  default: 5
+  info: Timeout for OSD connection attempts.
+  info_ru: Максимальное время ожидания попытки соединения с OSD.
+- name: osd_idle_timeout
+  type: sec
+  min: 1
+  default: 5
+  info: |
+    OSD connection inactivity time after which clients and other OSDs send
+    keepalive requests to check state of the connection.
+  info_ru: |
+    Время неактивности соединения с OSD, после которого клиенты или другие OSD
+    посылают запрос проверки состояния соединения.
+- name: osd_ping_timeout
+  type: sec
+  min: 1
+  default: 5
+  info: |
+    Maximum time to wait for OSD keepalive responses. If an OSD doesn't respond
+    within this time, the connection to it is dropped and a reconnection attempt
+    is scheduled.
+  info_ru: |
+    Максимальное время ожидания ответа на запрос проверки состояния соединения.
+    Если OSD не отвечает за это время, соединение отключается и производится
+    повторная попытка соединения.
+- name: up_wait_retry_interval
+  type: ms
+  min: 50
+  default: 500
+  info: |
+    OSDs respond to clients with a special error code when they receive I/O
+    requests for a PG that's not synchronized and started. This parameter sets
+    the time for the clients to wait before re-attempting such I/O requests.
+  info_ru: |
+    Когда OSD получают от клиентов запросы ввода-вывода, относящиеся к не
+    поднятым на данный момент на них PG, либо к PG в процессе синхронизации,
+    они отвечают клиентам специальным кодом ошибки, означающим, что клиент
+    должен некоторое время подождать перед повторением запроса. Именно это время
+    ожидания задаёт данный параметр.
+- name: max_etcd_attempts
+  type: int
+  default: 5
+  info: |
+    Maximum number of attempts for etcd requests which can't be retried
+    indefinitely.
+  info_ru: |
+    Максимальное число попыток выполнения запросов к etcd для тех запросов,
+    которые нельзя повторять бесконечно.
+- name: etcd_quick_timeout
+  type: ms
+  default: 1000
+  info: |
+    Timeout for etcd requests which should complete quickly, like lease refresh.
+  info_ru: |
+    Максимальное время выполнения запросов к etcd, которые должны завершаться
+    быстро, таких, как обновление резервации (lease).
+- name: etcd_slow_timeout
+  type: ms
+  default: 5000
+  info: Timeout for etcd requests which are allowed to wait for some time.
+  info_ru: |
+    Максимальное время выполнения запросов к etcd, для которых не обязательно
+    гарантировать быстрое выполнение.
+- name: etcd_keepalive_timeout
+  type: sec
+  default: max(30, etcd_report_interval*2)
+  info: |
+    Timeout for etcd connection HTTP Keep-Alive. Should be higher than
+    etcd_report_interval to guarantee that keepalive actually works.
+  info_ru: |
+    Таймаут для HTTP Keep-Alive в соединениях к etcd. Должен быть больше, чем
+    etcd_report_interval, чтобы keepalive гарантированно работал.
+- name: etcd_ws_keepalive_timeout
+  type: sec
+  default: 30
+  info: |
+    etcd websocket ping interval required to keep the connection alive and
+    detect disconnections quickly.
+  info_ru: |
+    Интервал проверки живости вебсокет-подключений к etcd.

docs/params/osd.yml

@@ -0,0 +1,341 @@
+- name: etcd_report_interval
+  type: sec
+  default: 5
+  info: |
+    Interval at which OSDs report their state to etcd. Affects OSD lease time
+    and thus the failover speed. Lease time is equal to this parameter value
+    plus max_etcd_attempts * etcd_quick_timeout because it should be guaranteed
+    that every OSD always refreshes its lease in time.
+  info_ru: |
+    Интервал, с которым OSD обновляет своё состояние в etcd. Значение параметра
+    влияет на время резервации (lease) OSD и поэтому на скорость переключения
+    при падении OSD. Время lease равняется значению этого параметра плюс
+    max_etcd_attempts * etcd_quick_timeout.
+- name: run_primary
+  type: bool
+  default: true
+  info: |
+    Start primary OSD logic on this OSD. As of now, can be turned off only for
+    debugging purposes. It's possible to implement additional feature for the
+    monitor which may allow to separate primary and secondary OSDs, but it's
+    unclear why anyone could need it, so it's not implemented.
+  info_ru: |
+    Запускать логику первичного OSD на данном OSD. На данный момент отключать
+    эту опцию может иметь смысл только в целях отладки. В теории, можно
+    реализовать дополнительный режим для монитора, который позволит отделять
+    первичные OSD от вторичных, но пока не понятно, зачем это может кому-то
+    понадобиться, поэтому это не реализовано.
+- name: osd_network
+  type: string or array of strings
+  type_ru: строка или массив строк
+  info: |
+    Network mask of the network (IPv4 or IPv6) to use for OSDs. Note that
+    although it's possible to specify multiple networks here, this does not
+    mean that OSDs will create multiple listening sockets - they'll only
+    pick the first matching address of an UP + RUNNING interface. Separate
+    networks for cluster and client connections are also not implemented, but
+    they are mostly useless anyway, so it's not a big deal.
+  info_ru: |
+    Маска подсети (IPv4 или IPv6) для использования для соединений с OSD.
+    Имейте в виду, что хотя сейчас и можно передать в этот параметр несколько
+    подсетей, это не означает, что OSD будут создавать несколько слушающих
+    сокетов - они лишь будут выбирать адрес первого поднятого (состояние UP +
+    RUNNING), подходящий под заданную маску. Также не реализовано разделение
+    кластерной и публичной сетей OSD. Правда, от него обычно всё равно довольно
+    мало толку, так что особенной проблемы в этом нет.
+- name: bind_address
+  type: string
+  default: "0.0.0.0"
+  info: |
+    Instead of the network mask, you can also set OSD listen address explicitly
+    using this parameter. May be useful if you want to start OSDs on interfaces
+    that are not UP + RUNNING.
+  info_ru: |
+    Этим параметром можно явным образом задать адрес, на котором будет ожидать
+    соединений OSD (вместо использования маски подсети). Может быть полезно,
+    например, чтобы запускать OSD на неподнятых интерфейсах (не UP + RUNNING).
+- name: bind_port
+  type: int
+  info: |
+    By default, OSDs pick random ports to use for incoming connections
+    automatically. With this option you can set a specific port for a specific
+    OSD by hand.
+  info_ru: |
+    По умолчанию OSD сами выбирают случайные порты для входящих подключений.
+    С помощью данной опции вы можете задать порт для отдельного OSD вручную.
+- name: autosync_interval
+  type: sec
+  default: 5
+  info: |
+    Time interval at which automatic fsyncs/flushes are issued by each OSD when
+    the immediate_commit mode if disabled. fsyncs are required because without
+    them OSDs quickly fill their journals, become unable to clear them and
+    stall. Also this option limits the amount of recent uncommitted changes
+    which OSDs may lose in case of a power outage in case when clients don't
+    issue fsyncs at all.
+  info_ru: |
+    Временной интервал отправки автоматических fsync-ов (операций очистки кэша)
+    каждым OSD для случая, когда режим immediate_commit отключён. fsync-и нужны
+    OSD, чтобы успевать очищать журнал - без них OSD быстро заполняют журналы и
+    перестают обрабатывать операции записи. Также эта опция ограничивает объём
+    недавних незафиксированных изменений, которые OSD могут терять при
+    отключении питания, если клиенты вообще не отправляют fsync.
+- name: autosync_writes
+  type: int
+  default: 128
+  info: |
+    Same as autosync_interval, but sets the maximum number of uncommitted write
+    operations before issuing an fsync operation internally.
+  info_ru: |
+    Аналогично autosync_interval, но задаёт не временной интервал, а
+    максимальное количество незафиксированных операций записи перед
+    принудительной отправкой fsync-а.
+- name: recovery_queue_depth
+  type: int
+  default: 4
+  info: |
+    Maximum recovery operations per one primary OSD at any given moment of time.
+    Currently it's the only parameter available to tune the speed or recovery
+    and rebalancing, but it's planned to implement more.
+  info_ru: |
+    Максимальное число операций восстановления на одном первичном OSD в любой
+    момент времени. На данный момент единственный параметр, который можно менять
+    для ускорения или замедления восстановления и перебалансировки данных, но
+    в планах реализация других параметров.
+- name: recovery_sync_batch
+  type: int
+  default: 16
+  info: Maximum number of recovery operations before issuing an additional fsync.
+  info_ru: Максимальное число операций восстановления перед дополнительным fsync.
+- name: readonly
+  type: bool
+  default: false
+  info: |
+    Read-only mode. If this is enabled, an OSD will never issue any writes to
+    the underlying device. This may be useful for recovery purposes.
+  info_ru: |
+    Режим "только чтение". Если включить этот режим, OSD не будет писать ничего
+    на диск. Может быть полезно в целях восстановления.
+- name: no_recovery
+  type: bool
+  default: false
+  info: |
+    Disable automatic background recovery of objects. Note that it doesn't
+    affect implicit recovery of objects happening during writes - a write is
+    always made to a full set of at least pg_minsize OSDs.
+  info_ru: |
+    Отключить автоматическое фоновое восстановление объектов. Обратите внимание,
+    что эта опция не отключает восстановление объектов, происходящее при
+    записи - запись всегда производится в полный набор из как минимум pg_minsize
+    OSD.
+- name: no_rebalance
+  type: bool
+  default: false
+  info: |
+    Disable background movement of data between different OSDs. Disabling it
+    means that PGs in the `has_misplaced` state will be left in it indefinitely.
+  info_ru: |
+    Отключить фоновое перемещение объектов между разными OSD. Отключение
+    означает, что PG, находящиеся в состоянии `has_misplaced`, будут оставлены
+    в нём на неопределённый срок.
+- name: print_stats_interval
+  type: sec
+  default: 3
+  info: |
+    Time interval at which OSDs print simple human-readable operation
+    statistics on stdout.
+  info_ru: |
+    Временной интервал, с которым OSD печатают простую человекочитаемую
+    статистику выполнения операций в стандартный вывод.
+- name: slow_log_interval
+  type: sec
+  default: 10
+  info: |
+    Time interval at which OSDs dump slow or stuck operations on stdout, if
+    they're any. Also it's the time after which an operation is considered
+    "slow".
+  info_ru: |
+    Временной интервал, с которым OSD выводят в стандартный вывод список
+    медленных или зависших операций, если таковые имеются. Также время, при
+    превышении которого операция считается "медленной".
+- name: max_write_iodepth
+  type: int
+  default: 128
+  info: |
+    Parallel client write operation limit per one OSD. Operations that exceed
+    this limit are pushed to a temporary queue instead of being executed
+    immediately.
+  info_ru: |
+    Максимальное число одновременных клиентских операций записи на один OSD.
+    Операции, превышающие этот лимит, не исполняются сразу, а сохраняются во
+    временной очереди.
+- name: min_flusher_count
+  type: int
+  default: 1
+  info: |
+    Flusher is a micro-thread that moves data from the journal to the data
+    area of the device. Their number is auto-tuned between minimum and maximum.
+    Minimum number is set by this parameter.
+  info_ru: |
+    Flusher - это микро-поток (корутина), которая копирует данные из журнала в
+    основную область устройства данных. Их число настраивается динамически между
+    минимальным и максимальным значением. Этот параметр задаёт минимальное число.
+- name: max_flusher_count
+  type: int
+  default: 256
+  info: |
+    Maximum number of journal flushers (see above min_flusher_count).
+  info_ru: |
+    Максимальное число микро-потоков очистки журнала (см. выше min_flusher_count).
+- name: inmemory_metadata
+  type: bool
+  default: true
+  info: |
+    This parameter makes Vitastor always keep metadata area of the block device
+    in memory. It's required for good performance because it allows to avoid
+    additional read-modify-write cycles during metadata modifications. Metadata
+    area size is currently roughly 224 MB per 1 TB of data. You can turn it off
+    to reduce memory usage by this value, but it will hurt performance. This
+    restriction is likely to be removed in the future along with the upgrade
+    of the metadata storage scheme.
+  info_ru: |
+    Данный параметр заставляет Vitastor всегда держать область метаданных диска
+    в памяти. Это нужно, чтобы избегать дополнительных операций чтения с диска
+    при записи. Размер области метаданных на данный момент составляет примерно
+    224 МБ на 1 ТБ данных. При включении потребление памяти снизится примерно
+    на эту величину, но при этом также снизится и производительность. В будущем,
+    после обновления схемы хранения метаданных, это ограничение, скорее всего,
+    будет ликвидировано.
+- name: inmemory_journal
+  type: bool
+  default: true
+  info: |
+    This parameter make Vitastor always keep journal area of the block
+    device in memory. Turning it off will, again, reduce memory usage, but
+    hurt performance because flusher coroutines will have to read data from
+    the disk back before copying it into the main area. The memory usage benefit
+    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.
+  info_ru: |
+    Данный параметр заставляет Vitastor всегда держать в памяти журналы OSD.
+    Отключение параметра, опять же, снижает потребление памяти, но ухудшает
+    производительность, так как для копирования данных из журнала в основную
+    область устройства OSD будут вынуждены читать их обратно с диска. Выигрыш
+    по памяти при этом обычно крайне низкий, так как для SSD OSD обычно
+    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: journal_sector_buffer_count
+  type: int
+  default: 32
+  info: |
+    Maximum number of buffers that can be used for writing journal metadata
+    blocks. The only situation when you should increase it to a larger value
+    is when you enable journal_no_same_sector_overwrites. In this case set
+    it to, for example, 1024.
+  info_ru: |
+    Максимальное число буферов, разрешённых для использования под записываемые
+    в журнал блоки метаданных. Единственная ситуация, в которой этот параметр
+    нужно менять - это если вы включаете journal_no_same_sector_overwrites. В
+    этом случае установите данный параметр, например, в 1024.
+- name: journal_no_same_sector_overwrites
+  type: bool
+  default: false
+  info: |
+    Enable this option for SSDs like Intel D3-S4510 and D3-S4610 which REALLY
+    don't like when a program overwrites the same sector multiple times in a
+    row and slow down significantly (from 25000+ iops to ~3000 iops). When
+    this option is set, Vitastor will always move to the next sector of the
+    journal after writing it instead of possibly overwriting it the second time.
+  info_ru: |
+    Включайте данную опцию для SSD вроде Intel D3-S4510 и D3-S4610, которые
+    ОЧЕНЬ не любят, когда ПО перезаписывает один и тот же сектор несколько раз
+    подряд. Такие SSD при многократной перезаписи одного и того же сектора
+    сильно замедляются - условно, с 25000 и более iops до 3000 iops. Когда
+    данная опция установлена, Vitastor всегда переходит к следующему сектору
+    журнала после записи вместо потенциально повторной перезаписи того же
+    самого сектора.
+- name: throttle_small_writes
+  type: bool
+  default: false
+  info: |
+    Enable soft throttling of small journaled writes. Useful for hybrid OSDs
+    with fast journal/metadata devices and slow data devices. The idea is that
+    small writes complete very quickly because they're first written to the
+    journal device, but moving them to the main device is slow. So if an OSD
+    allows clients to issue a lot of small writes it will perform very good
+    for several seconds and then the journal will fill up and the performance
+    will drop to almost zero. Throttling is meant to prevent this problem by
+    artifically slowing quick writes down based on the amount of free space in
+    the journal. When throttling is used, the performance of small writes will
+    decrease smoothly instead of abrupt drop at the moment when the journal
+    fills up.
+  info_ru: |
+    Разрешить мягкое ограничение скорости журналируемой записи. Полезно для
+    гибридных OSD с быстрыми устройствами метаданных и медленными устройствами
+    данных. Идея заключается в том, что мелкие записи в этой ситуации могут
+    завершаться очень быстро, так как они изначально записываются на быстрое
+    журнальное устройство (SSD). Но перемещать их потом на основное медленное
+    устройство долго. Поэтому если OSD быстро примет от клиентов очень много
+    мелких операций записи, он быстро заполнит свой журнал, после чего
+    производительность записи резко упадёт практически до нуля. Ограничение
+    скорости записи призвано решить эту проблему с помощью искусственного
+    замедления операций записи на основании объёма свободного места в журнале.
+    Когда эта опция включена, производительность мелких операций записи будет
+    снижаться плавно, а не резко в момент окончательного заполнения журнала.
+- name: throttle_target_iops
+  type: int
+  default: 100
+  info: |
+    Target maximum number of throttled operations per second under the condition
+    of full journal. Set it to approximate random write iops of your data devices
+    (HDDs).
+  info_ru: |
+    Расчётное максимальное число ограничиваемых операций в секунду при условии
+    отсутствия свободного места в журнале. Устанавливайте приблизительно равным
+    максимальной производительности случайной записи ваших устройств данных
+    (HDD) в операциях в секунду.
+- name: throttle_target_mbs
+  type: int
+  default: 100
+  info: |
+    Target maximum bandwidth in MB/s of throttled operations per second under
+    the condition of full journal. Set it to approximate linear write
+    performance of your data devices (HDDs).
+  info_ru: |
+    Расчётный максимальный размер в МБ/с ограничиваемых операций в секунду при
+    условии отсутствия свободного места в журнале. Устанавливайте приблизительно
+    равным максимальной производительности линейной записи ваших устройств
+    данных (HDD).
+- name: throttle_target_parallelism
+  type: int
+  default: 1
+  info: |
+    Target maximum parallelism of throttled operations under the condition of
+    full journal. Set it to approximate internal parallelism of your data
+    devices (1 for HDDs, 4-8 for SSDs).
+  info_ru: |
+    Расчётный максимальный параллелизм ограничиваемых операций в секунду при
+    условии отсутствия свободного места в журнале. Устанавливайте приблизительно
+    равным внутреннему параллелизму ваших устройств данных (1 для HDD, 4-8
+    для SSD).
+- name: throttle_threshold_us
+  type: us
+  default: 50
+  info: |
+    Minimal computed delay to be applied to throttled operations. Usually
+    doesn't need to be changed.
+  info_ru: |
+    Минимальная применимая к ограничиваемым операциям задержка. Обычно не
+    требует изменений.
+- name: osd_memlock
+  type: bool
+  default: false
+  info: >
+    Lock all OSD memory to prevent it from being unloaded into swap with
+    mlockall(). Requires sufficient ulimit -l (max locked memory).
+  info_ru: >
+    Блокировать всю память OSD с помощью mlockall, чтобы запретить её выгрузку
+    в пространство подкачки. Требует достаточного значения ulimit -l (лимита
+    заблокированной памяти).

mon/make-osd-hybrid.js

@@ -0,0 +1,414 @@
+#!/usr/bin/nodejs
+// systemd unit generator for hybrid (HDD+SSD) vitastor OSDs
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1
+
+// USAGE: nodejs make-osd-hybrid.js [--disable_ssd_cache 0] [--disable_hdd_cache 0] /dev/sda /dev/sdb /dev/sdc /dev/sdd ...
+// I.e. - just pass all HDDs and SSDs mixed, the script will decide where
+// to put journals on its own
+
+const fs = require('fs');
+const fsp = fs.promises;
+const child_process = require('child_process');
+
+const options = {
+    debug: 1,
+    journal_size: 1024*1024*1024,
+    min_meta_size: 1024*1024*1024,
+    object_size: 1024*1024,
+    bitmap_granularity: 4096,
+    device_block_size: 4096,
+    disable_ssd_cache: 1,
+    disable_hdd_cache: 1,
+};
+
+run().catch(console.fatal);
+
+async function run()
+{
+    const device_list = parse_options();
+    await system_or_die("mkdir -p /var/log/vitastor; chown vitastor /var/log/vitastor");
+    // Collect devices
+    const all_devices = await collect_devices(device_list);
+    const ssds = all_devices.filter(d => d.ssd);
+    const hdds = all_devices.filter(d => !d.ssd);
+    // Collect existing OSD units
+    const osd_units = await collect_osd_units();
+    // Count assigned HDD journals and unallocated space for each SSD
+    await check_journal_count(ssds, osd_units);
+    // Create new OSDs
+    await create_new_hybrid_osds(hdds, ssds, osd_units);
+    process.exit(0);
+}
+
+function parse_options()
+{
+    const devices = [];
+    const opt = {};
+    for (let i = 2; i < process.argv.length; i++)
+    {
+        const arg = process.argv[i];
+        if (arg == '--help' || arg == '-h')
+        {
+            opt.help = true;
+            break;
+        }
+        else if (arg.substr(0, 2) == '--')
+            opt[arg.substr(2)] = process.argv[++i];
+        else
+            devices.push(arg);
+    }
+    if (opt.help || !devices.length)
+    {
+        console.log(
+            'Prepare hybrid (HDD+SSD) Vitastor OSDs\n'+
+            '(c) Vitaliy Filippov, 2019+, license: VNPL-1.1\n\n'+
+            'USAGE: nodejs make-osd-hybrid.js [OPTIONS] /dev/sda /dev/sdb /dev/sdc ...\n'+
+            'Just pass all your SSDs and HDDs in any order, the script will distribute OSDs for you.\n\n'+
+            'OPTIONS (with defaults):\n'+
+            Object.keys(options).map(k => `  --${k} ${options[k]}`).join('\n')
+        );
+        process.exit(0);
+    }
+    for (const k in opt)
+        options[k] = opt[k];
+    return devices;
+}
+
+// Collect devices
+async function collect_devices(devices_to_check)
+{
+    const devices = [];
+    for (const dev of devices_to_check)
+    {
+        if (dev.substr(0, 5) != '/dev/')
+        {
+            console.log(`${dev} does not start with /dev/, skipping`);
+            continue;
+        }
+        if (!await file_exists('/sys/block/'+dev.substr(5)))
+        {
+            console.log(`${dev} is a partition, skipping`);
+            continue;
+        }
+        // Check if the device is an SSD
+        const rot = '/sys/block/'+dev.substr(5)+'/queue/rotational';
+        if (!await file_exists(rot))
+        {
+            console.log(`${dev} does not have ${rot} to check whether it's an SSD, skipping`);
+            continue;
+        }
+        const ssd = !parseInt(await fsp.readFile(rot, { encoding: 'utf-8' }));
+        // Check if the device has partition table
+        let [ has_partition_table, parts ] = await system(`sfdisk --dump ${dev} --json`);
+        if (has_partition_table != 0)
+        {
+            // Check if the device has any data
+            const [ has_data, out ] = await system(`blkid ${dev}`);
+            if (has_data == 0)
+            {
+                console.log(`${dev} contains data, skipping:\n  ${out.trim().replace(/\n/g, '\n  ')}`);
+                continue;
+            }
+        }
+        parts = parts ? JSON.parse(parts).partitiontable : null;
+        if (parts && parts.label != 'gpt')
+        {
+            console.log(`${dev} contains "${parts.label}" partition table, only GPT is supported, skipping`);
+            continue;
+        }
+        devices.push({
+            path: dev,
+            ssd,
+            parts,
+        });
+    }
+    return devices;
+}
+
+// Collect existing OSD units
+async function collect_osd_units()
+{
+    const units = [];
+    for (const unit of (await system("ls /etc/systemd/system/vitastor-osd*.service"))[1].trim().split('\n'))
+    {
+        if (!unit)
+        {
+            continue;
+        }
+        let cmd = /^ExecStart\s*=\s*(([^\n]*\\\n)*[^\n]*)/.exec(await fsp.readFile(unit, { encoding: 'utf-8' }));
+        if (!cmd)
+        {
+            console.log('ExecStart= not found in '+unit+', skipping')
+            continue;
+        }
+        let kv = {}, key;
+        cmd = cmd[1].replace(/^bash\s+-c\s+'/, '')
+            .replace(/>>\s*\S+2>\s*&1\s*'$/, '')
+            .replace(/\s*\\\n\s*/g, ' ')
+            .replace(/([^\s']+)|'([^']+)'/g, (m, m1, m2) =>
+            {
+                m1 = m1||m2;
+                if (key == null)
+                {
+                    if (m1.substr(0, 2) != '--')
+                    {
+                        console.log('Strange command line in '+unit+', stopping');
+                        process.exit(1);
+                    }
+                    key = m1.substr(2);
+                }
+                else
+                {
+                    kv[key] = m1;
+                    key = null;
+                }
+            });
+        units.push(kv);
+    }
+    return units;
+}
+
+// Count assigned HDD journals and unallocated space for each SSD
+async function check_journal_count(ssds, osd_units)
+{
+    const units_by_journal = osd_units.reduce((a, c) =>
+    {
+        if (c.journal_device)
+            a[c.journal_device] = c;
+        return a;
+    }, {});
+    for (const dev of ssds)
+    {
+        dev.journals = 0;
+        if (dev.parts)
+        {
+            for (const part of dev.parts.partitions)
+            {
+                if (part.uuid && units_by_journal['/dev/disk/by-partuuid/'+part.uuid.toLowerCase()])
+                {
+                    dev.journals++;
+                }
+            }
+            dev.free = free_from_parttable(dev.parts);
+        }
+        else
+        {
+            dev.free = parseInt(await system_or_die("blockdev --getsize64 "+dev.path));
+        }
+    }
+}
+
+async function create_new_hybrid_osds(hdds, ssds, osd_units)
+{
+    const units_by_disk = osd_units.reduce((a, c) => { a[c.data_device] = c; return a; }, {});
+    for (const dev of hdds)
+    {
+        if (!dev.parts)
+        {
+            // HDD is not partitioned yet, create a single partition
+            // + is the "default value" for sfdisk
+            await system_or_die('sfdisk '+dev.path, 'label: gpt\n\n+ +\n');
+            dev.parts = JSON.parse(await system_or_die('sfdisk --dump '+dev.path+' --json')).partitiontable;
+        }
+        if (dev.parts.partitions.length != 1)
+        {
+            console.log(dev.path+' has more than 1 partition, skipping');
+        }
+        else if ((dev.parts.partitions[0].start + dev.parts.partitions[0].size) != (1 + dev.parts.lastlba))
+        {
+            console.log(dev.path+'1 is not a whole-disk partition, skipping');
+        }
+        else if (!dev.parts.partitions[0].uuid)
+        {
+            console.log(dev.parts.partitions[0].node+' does not have UUID. Please repartition '+dev.path+' with GPT');
+        }
+        else if (!units_by_disk['/dev/disk/by-partuuid/'+dev.parts.partitions[0].uuid.toLowerCase()])
+        {
+            await create_hybrid_osd(dev, ssds);
+        }
+    }
+}
+
+async function create_hybrid_osd(dev, ssds)
+{
+    // Create a new OSD
+    // Calculate metadata size
+    const data_device = '/dev/disk/by-partuuid/'+dev.parts.partitions[0].uuid.toLowerCase();
+    const data_size = dev.parts.partitions[0].size * dev.parts.sectorsize;
+    const meta_entry_size = 24 + 2*options.object_size/options.bitmap_granularity/8;
+    const entries_per_block = Math.floor(options.device_block_size / meta_entry_size);
+    const object_count = Math.floor(data_size / options.object_size);
+    let meta_size = Math.ceil(1 + object_count / entries_per_block) * options.device_block_size;
+    // Leave some extra space for future metadata formats and round metadata area size to multiples of 1 MB
+    meta_size = 2*meta_size;
+    meta_size = Math.ceil(meta_size/1024/1024) * 1024*1024;
+    if (meta_size < options.min_meta_size)
+        meta_size = options.min_meta_size;
+    let journal_size = Math.ceil(options.journal_size/1024/1024) * 1024*1024;
+    // Pick an SSD for journal, balancing the number of journals across SSDs
+    let selected_ssd;
+    for (const ssd of ssds)
+        if (ssd.free >= (meta_size+journal_size) && (!selected_ssd || selected_ssd.journals > ssd.journals))
+            selected_ssd = ssd;
+    if (!selected_ssd)
+    {
+        console.error('Could not find free space for SSD journal and metadata for '+dev.path);
+        process.exit(1);
+    }
+    // Allocate an OSD number
+    const osd_num = (await system_or_die("vitastor-cli alloc-osd")).trim();
+    if (!osd_num)
+    {
+        console.error('Failed to run vitastor-cli alloc-osd');
+        process.exit(1);
+    }
+    console.log('Creating OSD '+osd_num+' on '+dev.path+' (HDD) with journal and metadata on '+selected_ssd.path+' (SSD)');
+    // Add two partitions: journal and metadata
+    const new_parts = await add_partitions(selected_ssd, [ journal_size, meta_size ]);
+    selected_ssd.journals++;
+    const journal_device = '/dev/disk/by-partuuid/'+new_parts[0].uuid.toLowerCase();
+    const meta_device = '/dev/disk/by-partuuid/'+new_parts[1].uuid.toLowerCase();
+    // Wait until the device symlinks appear
+    while (!await file_exists(journal_device))
+    {
+        await new Promise(ok => setTimeout(ok, 100));
+    }
+    while (!await file_exists(meta_device))
+    {
+        await new Promise(ok => setTimeout(ok, 100));
+    }
+    // Zero out metadata and journal
+    await system_or_die("dd if=/dev/zero of="+journal_device+" bs=1M count="+(journal_size/1024/1024)+" oflag=direct");
+    await system_or_die("dd if=/dev/zero of="+meta_device+" bs=1M count="+(meta_size/1024/1024)+" oflag=direct");
+    // Create unit file for the OSD
+    const has_scsi_cache_type = options.disable_ssd_cache &&
+        (await system("ls /sys/block/"+selected_ssd.path.substr(5)+"/device/scsi_disk/*/cache_type"))[0] == 0;
+    const write_through = options.disable_ssd_cache && (
+        has_scsi_cache_type || selected_ssd.path.substr(5, 4) == 'nvme'
+        && (await system_or_die("/sys/block/"+selected_ssd.path.substr(5)+"/queue/write_cache")).trim() == "write through");
+    await fsp.writeFile('/etc/systemd/system/vitastor-osd'+osd_num+'.service',
+`[Unit]
+Description=Vitastor object storage daemon osd.${osd_num}
+After=network-online.target local-fs.target time-sync.target
+Wants=network-online.target local-fs.target time-sync.target
+PartOf=vitastor.target
+
+[Service]
+LimitNOFILE=1048576
+LimitNPROC=1048576
+LimitMEMLOCK=infinity
+ExecStart=bash -c '/usr/bin/vitastor-osd \\
+    --osd_num ${osd_num} ${write_through
+        ? "--disable_meta_fsync 1 --disable_journal_fsync 1 --immediate_commit "+(options.disable_hdd_cache ? "all" : "small")
+        : ""} \\
+    --throttle_small_writes 1 \\
+    --disk_alignment ${options.device_block_size} \\
+    --journal_block_size ${options.device_block_size} \\
+    --meta_block_size ${options.device_block_size} \\
+    --journal_no_same_sector_overwrites true \\
+    --journal_sector_buffer_count 1024 \\
+    --block_size ${options.object_size} \\
+    --data_device ${data_device} \\
+    --journal_device ${journal_device} \\
+    --meta_device ${meta_device} >>/var/log/vitastor/osd${osd_num}.log 2>&1'
+WorkingDirectory=/
+ExecStartPre=+chown vitastor:vitastor ${data_device}
+ExecStartPre=+chown vitastor:vitastor ${journal_device}
+ExecStartPre=+chown vitastor:vitastor ${meta_device}${
+    has_scsi_cache_type
+    ? "\nExecStartPre=+bash -c 'D=$$$(readlink "+journal_device+"); echo write through > $$$(dirname /sys/block/*/$$\${D##*/})/device/scsi_disk/*/cache_type'"
+    : ""}${
+    options.disable_hdd_cache
+    ? "\nExecStartPre=+bash -c 'D=$$$(readlink "+data_device+"); echo write through > $$$(dirname /sys/block/*/$$\${D##*/})/device/scsi_disk/*/cache_type'"
+    : ""}
+User=vitastor
+PrivateTmp=false
+TasksMax=infinity
+Restart=always
+StartLimitInterval=0
+RestartSec=10
+
+[Install]
+WantedBy=vitastor.target
+`);
+    await system_or_die("systemctl enable vitastor-osd"+osd_num);
+}
+
+async function add_partitions(dev, sizes)
+{
+    let script = 'label: gpt\n\n';
+    if (dev.parts)
+    {
+        // Old partitions
+        for (const part of dev.parts.partitions)
+        {
+            script += part.node+': '+Object.keys(part).map(k => k == 'node' ? '' : k+'='+part[k]).filter(k => k).join(', ')+'\n';
+        }
+    }
+    // New partitions
+    for (const size of sizes)
+    {
+        script += '+ '+Math.ceil(size/1024)+'KiB\n';
+    }
+    await system_or_die('sfdisk '+dev.path, script);
+    // Get new partition table and find the new partition
+    const newpt = JSON.parse(await system_or_die('sfdisk --dump '+dev.path+' --json')).partitiontable;
+    const old_nodes = dev.parts ? dev.parts.partitions.reduce((a, c) => { a[c.uuid] = true; return a; }, {}) : {};
+    const new_nodes = newpt.partitions.filter(part => !old_nodes[part.uuid]);
+    if (new_nodes.length != sizes.length)
+    {
+        console.error('Failed to partition '+dev.path+': new partitions not found in table');
+        process.exit(1);
+    }
+    dev.parts = newpt;
+    dev.free = free_from_parttable(newpt);
+    return new_nodes;
+}
+
+function free_from_parttable(pt)
+{
+    let free = pt.lastlba + 1 - pt.firstlba;
+    for (const part of pt.partitions)
+    {
+        free -= part.size;
+    }
+    free *= pt.sectorsize;
+    return free;
+}
+
+async function system_or_die(cmd, input = '')
+{
+    let [ exitcode, stdout, stderr ] = await system(cmd, input);
+    if (exitcode != 0)
+    {
+        console.error(cmd+' failed: '+stderr);
+        process.exit(1);
+    }
+    return stdout;
+}
+
+async function system(cmd, input = '')
+{
+    if (options.debug)
+    {
+        process.stderr.write('+ '+cmd+(input ? " <<EOF\n"+input.replace(/\s*$/, '\n')+"EOF" : '')+'\n');
+    }
+    const cp = child_process.spawn(cmd, { shell: true });
+    let stdout = '', stderr = '', finish_cb;
+    cp.stdout.on('data', buf => stdout += buf.toString());
+    cp.stderr.on('data', buf => stderr += buf.toString());
+    cp.on('exit', () => finish_cb && finish_cb());
+    cp.stdin.write(input);
+    cp.stdin.end();
+    if (cp.exitCode == null)
+    {
+        await new Promise(ok => finish_cb = ok);
+    }
+    return [ cp.exitCode, stdout, stderr ];
+}
+
+async function file_exists(filename)
+{
+    return new Promise((ok, no) => fs.access(filename, fs.constants.R_OK, err => ok(!err)));
+}

mon/make-osd.sh

@@ -25,6 +25,10 @@ OPT=$(vitastor-cli simple-offsets --format options $DEV | tr '\n' ' ')
 META=$(vitastor-cli simple-offsets --format json $DEV | jq .data_offset)
 dd if=/dev/zero of=$DEV bs=1048576 count=$(((META+1048575)/1048576)) oflag=direct
 
+mkdir -p /var/log/vitastor
+id vitastor &>/dev/null || useradd vitastor
+chown vitastor /var/log/vitastor
+
 cat >/etc/systemd/system/vitastor-osd$OSD_NUM.service <<EOF
 [Unit]
 Description=Vitastor object storage daemon osd.$OSD_NUM
@@ -36,14 +40,14 @@ PartOf=vitastor.target
 LimitNOFILE=1048576
 LimitNPROC=1048576
 LimitMEMLOCK=infinity
-ExecStart=/usr/bin/vitastor-osd \\
+ExecStart=bash -c '/usr/bin/vitastor-osd \\
     --osd_num $OSD_NUM \\
     --disable_data_fsync 1 \\
     --immediate_commit all \\
     --disk_alignment 4096 --journal_block_size 4096 --meta_block_size 4096 \\
     --journal_no_same_sector_overwrites true \\
     --journal_sector_buffer_count 1024 \\
-    $OPT
+    $OPT >>/var/log/vitastor/osd$OSD_NUM.log 2>&1'
 WorkingDirectory=/
 ExecStartPre=+chown vitastor:vitastor $DEV
 User=vitastor

mon/mon.js

@@ -31,6 +31,7 @@ const etcd_allow = new RegExp('^'+[
     'osd/inodestats/[1-9]\\d*',
     'osd/space/[1-9]\\d*',
     'mon/master',
+    'mon/member/[a-f0-9]+',
     'pg/state/[1-9]\\d*/[1-9]\\d*',
     'pg/stats/[1-9]\\d*/[1-9]\\d*',
     'pg/history/[1-9]\\d*/[1-9]\\d*',
@@ -83,13 +84,13 @@ const etcd_tree = {
             osd_idle_timeout: 5, // seconds. min: 1
             osd_ping_timeout: 5, // seconds. min: 1
             up_wait_retry_interval: 500, // ms. min: 50
-            // osd
-            etcd_report_interval: 5, // seconds
             max_etcd_attempts: 5,
             etcd_quick_timeout: 1000, // ms
             etcd_slow_timeout: 5000, // ms
-            etcd_keepalive_timeout: 30, // seconds, default is min(30, etcd_report_interval*2)
+            etcd_keepalive_timeout: 30, // seconds, default is max(30, etcd_report_interval*2)
             etcd_ws_keepalive_interval: 30, // seconds
+            // osd
+            etcd_report_interval: 5, // seconds
             run_primary: true,
             osd_network: null, // "192.168.7.0/24" or an array of masks
             bind_address: "0.0.0.0",
@@ -104,6 +105,7 @@ const etcd_tree = {
             no_rebalance: false,
             print_stats_interval: 3,
             slow_log_interval: 10,
+            osd_memlock: false,
             // blockstore - fixed in superblock
             block_size,
             disk_alignment,
@@ -130,6 +132,11 @@ const etcd_tree = {
             inmemory_journal,
             journal_sector_buffer_count,
             journal_no_same_sector_overwrites,
+            throttle_small_writes: false,
+            throttle_target_iops: 100,
+            throttle_target_mbs: 100,
+            throttle_target_parallelism: 1,
+            throttle_threshold_us: 50,
         }, */
         global: {},
         /* node_placement: {
@@ -231,7 +238,10 @@ const etcd_tree = {
     },
     mon: {
         master: {
-            /* ip: [ string ], */
+            /* ip: [ string ], id: uint64_t */
+        },
+        standby: {
+            /* <uint64_t>: { ip: [ string ] }, */
         },
     },
     pg: {
@@ -667,11 +677,25 @@ class Mon
         }, this.etcd_start_timeout, 0);
     }
 
+    get_mon_state()
+    {
+        return { ip: this.local_ips(), hostname: os.hostname() };
+    }
+
     async get_lease()
     {
         const max_ttl = this.config.etcd_mon_ttl + this.config.etcd_mon_timeout/1000*this.config.etcd_mon_retries;
-        const res = await this.etcd_call('/lease/grant', { TTL: max_ttl }, this.config.etcd_mon_timeout, -1);
+        // Get lease
+        let res = await this.etcd_call('/lease/grant', { TTL: max_ttl }, this.config.etcd_mon_timeout, -1);
         this.etcd_lease_id = res.ID;
+        // Register in /mon/member, just for the information
+        const state = this.get_mon_state();
+        res = await this.etcd_call('/kv/put', {
+            key: b64(this.etcd_prefix+'/mon/member/'+this.etcd_lease_id),
+            value: b64(JSON.stringify(state)),
+            lease: ''+this.etcd_lease_id
+        }, this.etcd_start_timeout, 0);
+        // Set refresh timer
         this.lease_timer = setInterval(async () =>
         {
             const res = await this.etcd_call('/lease/keepalive', { ID: this.etcd_lease_id }, this.config.etcd_mon_timeout, this.config.etcd_mon_retries);
@@ -697,7 +721,7 @@ class Mon
 
     async become_master()
     {
-        const state = { ip: this.local_ips() };
+        const state = { ...this.get_mon_state(), id: ''+this.etcd_lease_id };
         while (1)
         {
             const res = await this.etcd_call('/kv/txn', {
@@ -1339,21 +1363,30 @@ class Mon
         const tm = prev_stats ? BigInt(timestamp - prev_stats.timestamp) : 0;
         for (const op in op_stats)
         {
-            op_stats[op].bps = prev_stats ? (op_stats[op].bytes - prev_stats.op_stats[op].bytes) * 1000n / tm : 0;
-            op_stats[op].iops = prev_stats ? (op_stats[op].count - prev_stats.op_stats[op].count) * 1000n / tm : 0;
-            op_stats[op].lat = prev_stats ? (op_stats[op].usec - prev_stats.op_stats[op].usec)
-                / ((op_stats[op].count - prev_stats.op_stats[op].count) || 1n) : 0;
+            if (prev_stats && prev_stats.op_stats && prev_stats.op_stats[op])
+            {
+                op_stats[op].bps = (op_stats[op].bytes - prev_stats.op_stats[op].bytes) * 1000n / tm;
+                op_stats[op].iops = (op_stats[op].count - prev_stats.op_stats[op].count) * 1000n / tm;
+                op_stats[op].lat = (op_stats[op].usec - prev_stats.op_stats[op].usec)
+                    / ((op_stats[op].count - prev_stats.op_stats[op].count) || 1n);
+            }
         }
         for (const op in subop_stats)
         {
-            subop_stats[op].iops = prev_stats ? (subop_stats[op].count - prev_stats.subop_stats[op].count) * 1000n / tm : 0;
-            subop_stats[op].lat = prev_stats ? (subop_stats[op].usec - prev_stats.subop_stats[op].usec)
-                / ((subop_stats[op].count - prev_stats.subop_stats[op].count) || 1n) : 0;
+            if (prev_stats && prev_stats.subop_stats && prev_stats.subop_stats[op])
+            {
+                subop_stats[op].iops = (subop_stats[op].count - prev_stats.subop_stats[op].count) * 1000n / tm;
+                subop_stats[op].lat = (subop_stats[op].usec - prev_stats.subop_stats[op].usec)
+                    / ((subop_stats[op].count - prev_stats.subop_stats[op].count) || 1n);
+            }
         }
         for (const op in recovery_stats)
         {
-            recovery_stats[op].bps = prev_stats ? (recovery_stats[op].bytes - prev_stats.recovery_stats[op].bytes) * 1000n / tm : 0;
-            recovery_stats[op].iops = prev_stats ? (recovery_stats[op].count - prev_stats.recovery_stats[op].count) * 1000n / tm : 0;
+            if (prev_stats && prev_stats.recovery_stats && prev_stats.recovery_stats[op])
+            {
+                recovery_stats[op].bps = (recovery_stats[op].bytes - prev_stats.recovery_stats[op].bytes) * 1000n / tm;
+                recovery_stats[op].iops = (recovery_stats[op].count - prev_stats.recovery_stats[op].count) * 1000n / tm;
+            }
         }
         return { op_stats, subop_stats, recovery_stats };
     }

mon/simple-offsets.js

@@ -49,7 +49,8 @@ async function run()
     }
     options.journal_offset = Math.ceil(options.journal_offset/options.device_block_size)*options.device_block_size;
     const meta_offset = options.journal_offset + Math.ceil(options.journal_size/options.device_block_size)*options.device_block_size;
-    const entries_per_block = Math.floor(options.device_block_size / (24 + 2*options.object_size/options.bitmap_granularity/8));
+    const meta_entry_size = 24 + 2*options.object_size/options.bitmap_granularity/8;
+    const entries_per_block = Math.floor(options.device_block_size / meta_entry_size);
     const object_count = Math.floor((device_size-meta_offset)/options.object_size);
     const meta_size = Math.ceil(1 + object_count / entries_per_block) * options.device_block_size;
     const data_offset = meta_offset + meta_size;

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 = '0.6.12'
+VERSION = '0.6.16'
 
 LOG = logging.getLogger(__name__)
 
@@ -355,7 +355,25 @@ class VitastorDriver(driver.CloneableImageVD,
     def revert_to_snapshot(self, context, volume, snapshot):
         """Revert a volume to a given snapshot."""
 
-        # FIXME Delete the image, then recreate it from the snapshot
+        vol_name = utils.convert_str(snapshot.volume_name)
+        snap_name = utils.convert_str(snapshot.name)
+
+        # Delete the image and recreate it from the snapshot
+        args = [ 'vitastor-cli', 'rm', vol_name, *(self._vitastor_args()) ]
+        try:
+            self._execute(*args)
+        except processutils.ProcessExecutionError as exc:
+            LOG.error("Failed to delete image "+vol_name+": "+exc)
+            raise exception.VolumeBackendAPIException(data = exc.stderr)
+        args = [
+            'vitastor-cli', 'create', '--parent', vol_name+'@'+snap_name,
+            vol_name, *(self._vitastor_args())
+        ]
+        try:
+            self._execute(*args)
+        except processutils.ProcessExecutionError as exc:
+            LOG.error("Failed to recreate image "+vol_name+" from "+vol_name+"@"+snap_name+": "+exc)
+            raise exception.VolumeBackendAPIException(data = exc.stderr)
 
     def delete_snapshot(self, snapshot):
         """Deletes a snapshot."""
@@ -363,24 +381,15 @@ class VitastorDriver(driver.CloneableImageVD,
         vol_name = utils.convert_str(snapshot.volume_name)
         snap_name = utils.convert_str(snapshot.name)
 
-        # Find the snapshot
-        resp = self._etcd_txn({ 'success': [
-            { 'request_range': { 'key': 'index/image/'+vol_name+'@'+snap_name } },
-        ] })
-        if len(resp['responses'][0]['kvs']) == 0:
-            raise exception.SnapshotNotFound(snapshot_id = snap_name)
-        inode_id = int(resp['responses'][0]['kvs'][0]['value']['id'])
-        pool_id = int(resp['responses'][0]['kvs'][0]['value']['pool_id'])
-        parents = {}
-        parents[(pool_id << 48) | (inode_id & 0xffffffffffff)] = True
-
-        # Check if there are child volumes
-        children = self._child_count(parents)
-        if children > 0:
-            raise exception.SnapshotIsBusy(snapshot_name = snap_name)
-
-        # FIXME: We can't delete snapshots because we can't merge layers yet
-        raise exception.VolumeBackendAPIException(data = 'Snapshot delete (layer merge) is not implemented yet')
+        args = [
+            'vitastor-cli', 'rm', vol_name+'@'+snap_name,
+            *(self._vitastor_args())
+        ]
+        try:
+            self._execute(*args)
+        except processutils.ProcessExecutionError as exc:
+            LOG.error("Failed to remove snapshot "+vol_name+'@'+snap_name+": "+exc)
+            raise exception.VolumeBackendAPIException(data = exc.stderr)
 
     def _child_count(self, parents):
         children = 0

rpm/build-tarball.sh

@@ -25,4 +25,4 @@ rm fio
 mv fio-copy fio
 FIO=`rpm -qi fio | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Version[\s:]+(\S+)/ && print $1; /^Release[\s:]+(\S+)/ && print "-$1"; }'`
 perl -i -pe 's/(Requires:\s*fio)([^\n]+)?/$1 = '$FIO'/' $VITASTOR/rpm/vitastor-el$EL.spec
-tar --transform 's#^#vitastor-0.6.12/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.6.12$(rpm --eval '%dist').tar.gz *
+tar --transform 's#^#vitastor-0.6.16/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.6.16$(rpm --eval '%dist').tar.gz *

rpm/vitastor-el7.Dockerfile

@@ -34,7 +34,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.6.12.el7.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.6.16.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:        0.6.12
+Version:        0.6.16
 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-0.6.12.el7.tar.gz
+Source0:        vitastor-0.6.16.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.Dockerfile

@@ -33,7 +33,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.6.12.el8.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.6.16.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:        0.6.12
+Version:        0.6.16
 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-0.6.12.el8.tar.gz
+Source0:        vitastor-0.6.16.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -15,7 +15,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVERSION="0.6.12")
+add_definitions(-DVERSION="0.6.16")
 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)
@@ -154,8 +154,8 @@ target_link_libraries(vitastor-nbd
 
 # vitastor-cli
 add_executable(vitastor-cli
-	cli.cpp cli_alloc_osd.cpp cli_simple_offsets.cpp cli_df.cpp
-	cli_ls.cpp cli_create.cpp cli_modify.cpp cli_flatten.cpp cli_merge.cpp cli_rm.cpp cli_snap_rm.cpp
+	cli.cpp cli_alloc_osd.cpp cli_simple_offsets.cpp cli_status.cpp cli_df.cpp
+	cli_ls.cpp cli_create.cpp cli_modify.cpp cli_flatten.cpp cli_merge.cpp cli_rm_data.cpp cli_rm.cpp
 )
 target_link_libraries(vitastor-cli
 	vitastor_client

src/addr_util.cpp

@@ -1,3 +1,5 @@
+#include <sys/socket.h>
+#include <unistd.h>
 #include <arpa/inet.h>
 #include <net/if.h>
 #include <sys/types.h>
@@ -9,7 +11,7 @@
 
 #include "addr_util.h"
 
-bool string_to_addr(std::string str, bool parse_port, int default_port, struct sockaddr *addr)
+bool string_to_addr(std::string str, bool parse_port, int default_port, struct sockaddr_storage *addr)
 {
     if (parse_port)
     {
@@ -25,7 +27,7 @@ bool string_to_addr(std::string str, bool parse_port, int default_port, struct s
     }
     if (inet_pton(AF_INET, str.c_str(), &((struct sockaddr_in*)addr)->sin_addr) == 1)
     {
-        addr->sa_family = AF_INET;
+        addr->ss_family = AF_INET;
         ((struct sockaddr_in*)addr)->sin_port = htons(default_port);
         return true;
     }
@@ -33,30 +35,30 @@ bool string_to_addr(std::string str, bool parse_port, int default_port, struct s
         str = str.substr(1, str.length()-2);
     if (inet_pton(AF_INET6, str.c_str(), &((struct sockaddr_in6*)addr)->sin6_addr) == 1)
     {
-        addr->sa_family = AF_INET6;
+        addr->ss_family = AF_INET6;
         ((struct sockaddr_in6*)addr)->sin6_port = htons(default_port);
         return true;
     }
     return false;
 }
 
-std::string addr_to_string(const sockaddr &addr)
+std::string addr_to_string(const sockaddr_storage &addr)
 {
     char peer_str[256];
     bool ok = false;
     int port;
-    if (addr.sa_family == AF_INET)
+    if (addr.ss_family == AF_INET)
     {
         ok = !!inet_ntop(AF_INET, &((sockaddr_in*)&addr)->sin_addr, peer_str, 256);
         port = ntohs(((sockaddr_in*)&addr)->sin_port);
     }
-    else if (addr.sa_family == AF_INET6)
+    else if (addr.ss_family == AF_INET6)
     {
         ok = !!inet_ntop(AF_INET6, &((sockaddr_in6*)&addr)->sin6_addr, peer_str, 256);
         port = ntohs(((sockaddr_in6*)&addr)->sin6_port);
     }
     else
-        throw std::runtime_error("Unknown address family "+std::to_string(addr.sa_family));
+        throw std::runtime_error("Unknown address family "+std::to_string(addr.ss_family));
     if (!ok)
         throw std::runtime_error(std::string("inet_ntop: ") + strerror(errno));
     return std::string(peer_str)+":"+std::to_string(port);
@@ -186,3 +188,51 @@ std::vector<std::string> getifaddr_list(std::vector<std::string> mask_cfg, bool
     freeifaddrs(list);
     return addresses;
 }
+
+int create_and_bind_socket(std::string bind_address, int bind_port, int listen_backlog, int *listening_port)
+{
+    sockaddr_storage addr;
+    if (!string_to_addr(bind_address, 0, bind_port, &addr))
+    {
+        throw std::runtime_error("bind address "+bind_address+" is not valid");
+    }
+
+    int listen_fd = socket(addr.ss_family, SOCK_STREAM, 0);
+    if (listen_fd < 0)
+    {
+        throw std::runtime_error(std::string("socket: ") + strerror(errno));
+    }
+    int enable = 1;
+    setsockopt(listen_fd, SOL_SOCKET, SO_REUSEADDR, &enable, sizeof(enable));
+
+    if (bind(listen_fd, (sockaddr*)&addr, sizeof(addr)) < 0)
+    {
+        close(listen_fd);
+        throw std::runtime_error(std::string("bind: ") + strerror(errno));
+    }
+    if (listening_port)
+    {
+        if (bind_port == 0)
+        {
+            socklen_t len = sizeof(addr);
+            if (getsockname(listen_fd, (sockaddr *)&addr, &len) == -1)
+            {
+                close(listen_fd);
+                throw std::runtime_error(std::string("getsockname: ") + strerror(errno));
+            }
+            *listening_port = ntohs(((sockaddr_in*)&addr)->sin_port);
+        }
+        else
+        {
+            *listening_port = bind_port;
+        }
+    }
+
+    if (listen(listen_fd, listen_backlog ? listen_backlog : 128) < 0)
+    {
+        close(listen_fd);
+        throw std::runtime_error(std::string("listen: ") + strerror(errno));
+    }
+
+    return listen_fd;
+}

src/addr_util.h

@@ -4,6 +4,7 @@
 #include <string>
 #include <vector>
 
-bool string_to_addr(std::string str, bool parse_port, int default_port, struct sockaddr *addr);
-std::string addr_to_string(const sockaddr &addr);
+bool string_to_addr(std::string str, bool parse_port, int default_port, struct sockaddr_storage *addr);
+std::string addr_to_string(const sockaddr_storage &addr);
 std::vector<std::string> getifaddr_list(std::vector<std::string> mask_cfg = std::vector<std::string>(), bool include_v6 = false);
+int create_and_bind_socket(std::string bind_address, int bind_port, int listen_backlog, int *listening_port);

src/blockstore.h

@@ -21,7 +21,7 @@
 // Memory alignment for direct I/O (usually 512 bytes)
 // All other alignments must be a multiple of this one
 #ifndef MEM_ALIGNMENT
-#define MEM_ALIGNMENT 512
+#define MEM_ALIGNMENT 4096
 #endif
 
 // Default block size is 128 KB, current allowed range is 4K - 128M

src/blockstore_flush.cpp

@@ -415,8 +415,11 @@ stop_flusher:
         flusher->active_flushers++;
 resume_1:
         // Find it in clean_db
-        clean_it = bs->clean_db.find(cur.oid);
-        old_clean_loc = (clean_it != bs->clean_db.end() ? clean_it->second.location : UINT64_MAX);
+        {
+            auto & clean_db = bs->clean_db_shard(cur.oid);
+            auto clean_it = clean_db.find(cur.oid);
+            old_clean_loc = (clean_it != clean_db.end() ? clean_it->second.location : UINT64_MAX);
+        }
         // Scan dirty versions of the object
         if (!scan_dirty(1))
         {
@@ -870,10 +873,11 @@ void journal_flusher_co::update_clean_db()
 #endif
         bs->data_alloc->set(old_clean_loc >> bs->block_order, false);
     }
+    auto & clean_db = bs->clean_db_shard(cur.oid);
     if (has_delete)
     {
-        auto clean_it = bs->clean_db.find(cur.oid);
-        bs->clean_db.erase(clean_it);
+        auto clean_it = clean_db.find(cur.oid);
+        clean_db.erase(clean_it);
 #ifdef BLOCKSTORE_DEBUG
         printf("Free block %lu from %lx:%lx v%lu (delete)\n",
             clean_loc >> bs->block_order,
@@ -884,7 +888,7 @@ void journal_flusher_co::update_clean_db()
     }
     else
     {
-        bs->clean_db[cur.oid] = {
+        clean_db[cur.oid] = {
             .version = cur.version,
             .location = clean_loc,
         };

src/blockstore_flush.h

@@ -49,7 +49,6 @@ class journal_flusher_co
     std::function<void(ring_data_t*)> simple_callback_r, simple_callback_w;
 
     bool skip_copy, has_delete, has_writes;
-    blockstore_clean_db_t::iterator clean_it;
     std::vector<copy_buffer_t> v;
     std::vector<copy_buffer_t>::iterator it;
     int copy_count;

src/blockstore_impl.cpp

@@ -118,7 +118,7 @@ void blockstore_impl_t::loop()
         // has_writes == 0 - no writes before the current queue item
         // has_writes == 1 - some writes in progress
         // has_writes == 2 - tried to submit some writes, but failed
-        int has_writes = 0, op_idx = 0, new_idx = 0;
+        int has_writes = 0, op_idx = 0, new_idx = 0, done_lists = 0;
         for (; op_idx < submit_queue.size(); op_idx++, new_idx++)
         {
             auto op = submit_queue[op_idx];
@@ -142,7 +142,6 @@ void blockstore_impl_t::loop()
                     continue;
                 }
             }
-            unsigned ring_space = ringloop->space_left();
             unsigned prev_sqe_pos = ringloop->save();
             // 0 = can't submit
             // 1 = in progress
@@ -199,9 +198,14 @@ void blockstore_impl_t::loop()
             }
             else if (op->opcode == BS_OP_LIST)
             {
-                // LIST doesn't need to be blocked by previous modifications
-                process_list(op);
-                wr_st = 2;
+                // LIST doesn't have to be blocked by previous modifications
+                // But don't do a lot of LISTs at once, because they're blocking and potentially slow
+                if (single_tick_list_limit <= 0 || done_lists < single_tick_list_limit)
+                {
+                    process_list(op);
+                    done_lists++;
+                    wr_st = 2;
+                }
             }
             if (wr_st == 2)
             {
@@ -212,7 +216,6 @@ void blockstore_impl_t::loop()
                 ringloop->restore(prev_sqe_pos);
                 if (PRIV(op)->wait_for == WAIT_SQE)
                 {
-                    PRIV(op)->wait_detail = 1 + ring_space;
                     // ring is full, stop submission
                     break;
                 }
@@ -282,7 +285,7 @@ void blockstore_impl_t::check_wait(blockstore_op_t *op)
 {
     if (PRIV(op)->wait_for == WAIT_SQE)
     {
-        if (ringloop->space_left() < PRIV(op)->wait_detail)
+        if (ringloop->sqes_left() < PRIV(op)->wait_detail)
         {
             // stop submission if there's still no free space
 #ifdef BLOCKSTORE_DEBUG
@@ -425,22 +428,104 @@ static bool replace_stable(object_id oid, uint64_t version, int search_start, in
     return false;
 }
 
+blockstore_clean_db_t& blockstore_impl_t::clean_db_shard(object_id oid)
+{
+    uint64_t pg_num = 0;
+    uint64_t pool_id = (oid.inode >> (64-POOL_ID_BITS));
+    auto sh_it = clean_db_settings.find(pool_id);
+    if (sh_it != clean_db_settings.end())
+    {
+        // like map_to_pg()
+        pg_num = (oid.stripe / sh_it->second.pg_stripe_size) % sh_it->second.pg_count + 1;
+    }
+    return clean_db_shards[(pool_id << (64-POOL_ID_BITS)) | pg_num];
+}
+
+void blockstore_impl_t::reshard_clean_db(pool_id_t pool, uint32_t pg_count, uint32_t pg_stripe_size)
+{
+    uint64_t pool_id = (uint64_t)pool;
+    std::map<pool_pg_id_t, blockstore_clean_db_t> new_shards;
+    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    while (sh_it != clean_db_shards.end() &&
+        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        for (auto & pair: sh_it->second)
+        {
+            // like map_to_pg()
+            uint64_t pg_num = (pair.first.stripe / pg_stripe_size) % pg_count + 1;
+            uint64_t shard_id = (pool_id << (64-POOL_ID_BITS)) | pg_num;
+            new_shards[shard_id][pair.first] = pair.second;
+        }
+        clean_db_shards.erase(sh_it++);
+    }
+    for (sh_it = new_shards.begin(); sh_it != new_shards.end(); sh_it++)
+    {
+        auto & to = clean_db_shards[sh_it->first];
+        to.swap(sh_it->second);
+    }
+    clean_db_settings[pool_id] = (pool_shard_settings_t){
+        .pg_count = pg_count,
+        .pg_stripe_size = pg_stripe_size,
+    };
+}
+
 void blockstore_impl_t::process_list(blockstore_op_t *op)
 {
-    uint32_t list_pg = op->offset;
+    uint32_t list_pg = op->offset+1;
     uint32_t pg_count = op->len;
     uint64_t pg_stripe_size = op->oid.stripe;
     uint64_t min_inode = op->oid.inode;
     uint64_t max_inode = op->version;
     // Check PG
-    if (pg_count != 0 && (pg_stripe_size < MIN_BLOCK_SIZE || list_pg >= pg_count))
+    if (pg_count != 0 && (pg_stripe_size < MIN_BLOCK_SIZE || list_pg > pg_count))
     {
         op->retval = -EINVAL;
         FINISH_OP(op);
         return;
     }
-    // Copy clean_db entries (sorted)
-    int stable_count = 0, stable_alloc = clean_db.size() / (pg_count ? pg_count : 1);
+    // Check if the DB needs resharding
+    // (we don't know about PGs from the beginning, we only create "shards" here)
+    uint64_t first_shard = 0, last_shard = UINT64_MAX;
+    if (min_inode != 0 &&
+        // Check if min_inode == max_inode == pool_id<<N, i.e. this is a pool listing
+        (min_inode >> (64-POOL_ID_BITS)) == (max_inode >> (64-POOL_ID_BITS)))
+    {
+        pool_id_t pool_id = (min_inode >> (64-POOL_ID_BITS));
+        if (pg_count > 1)
+        {
+            // Per-pg listing
+            auto sh_it = clean_db_settings.find(pool_id);
+            if (sh_it == clean_db_settings.end() ||
+                sh_it->second.pg_count != pg_count ||
+                sh_it->second.pg_stripe_size != pg_stripe_size)
+            {
+                reshard_clean_db(pool_id, pg_count, pg_stripe_size);
+            }
+            first_shard = last_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS)) | list_pg;
+        }
+        else
+        {
+            // Per-pool listing
+            first_shard = ((uint64_t)pool_id << (64-POOL_ID_BITS));
+            last_shard = ((uint64_t)(pool_id+1) << (64-POOL_ID_BITS)) - 1;
+        }
+    }
+    // Copy clean_db entries
+    int stable_count = 0, stable_alloc = 0;
+    if (min_inode != max_inode)
+    {
+        for (auto shard_it = clean_db_shards.lower_bound(first_shard);
+            shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
+            shard_it++)
+        {
+            auto & clean_db = shard_it->second;
+            stable_alloc += clean_db.size();
+        }
+    }
+    else
+    {
+        stable_alloc = 32768;
+    }
     obj_ver_id *stable = (obj_ver_id*)malloc(sizeof(obj_ver_id) * stable_alloc);
     if (!stable)
     {
@@ -448,7 +533,11 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
         FINISH_OP(op);
         return;
     }
+    for (auto shard_it = clean_db_shards.lower_bound(first_shard);
+        shard_it != clean_db_shards.end() && shard_it->first <= last_shard;
+        shard_it++)
     {
+        auto & clean_db = shard_it->second;
         auto clean_it = clean_db.begin(), clean_end = clean_db.end();
         if ((min_inode != 0 || max_inode != 0) && min_inode <= max_inode)
         {
@@ -463,26 +552,28 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
         }
         for (; clean_it != clean_end; clean_it++)
         {
-            if (!pg_count || ((clean_it->first.stripe / pg_stripe_size) % pg_count) == list_pg) // like map_to_pg()
+            if (stable_count >= stable_alloc)
             {
-                if (stable_count >= stable_alloc)
+                stable_alloc *= 2;
+                stable = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
+                if (!stable)
                 {
-                    stable_alloc += 32768;
-                    stable = (obj_ver_id*)realloc(stable, sizeof(obj_ver_id) * stable_alloc);
-                    if (!stable)
-                    {
-                        op->retval = -ENOMEM;
-                        FINISH_OP(op);
-                        return;
-                    }
+                    op->retval = -ENOMEM;
+                    FINISH_OP(op);
+                    return;
                 }
-                stable[stable_count++] = {
-                    .oid = clean_it->first,
-                    .version = clean_it->second.version,
-                };
             }
+            stable[stable_count++] = {
+                .oid = clean_it->first,
+                .version = clean_it->second.version,
+            };
         }
     }
+    if (first_shard != last_shard)
+    {
+        // If that's not a per-PG listing, sort clean entries
+        std::sort(stable, stable+stable_count);
+    }
     int clean_stable_count = stable_count;
     // Copy dirty_db entries (sorted, too)
     int unstable_count = 0, unstable_alloc = 0;
@@ -508,7 +599,7 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
         }
         for (; dirty_it != dirty_end; dirty_it++)
         {
-            if (!pg_count || ((dirty_it->first.oid.stripe / pg_stripe_size) % pg_count) == list_pg) // like map_to_pg()
+            if (!pg_count || ((dirty_it->first.oid.stripe / pg_stripe_size) % pg_count + 1) == list_pg) // like map_to_pg()
             {
                 if (IS_DELETE(dirty_it->second.state))
                 {

src/blockstore_impl.h

@@ -55,9 +55,10 @@
 #define IS_DELETE(st) (((st) & 0x0F) == BS_ST_DELETE)
 
 #define BS_SUBMIT_CHECK_SQES(n) \
-    if (ringloop->space_left() < (n))\
+    if (ringloop->sqes_left() < (n))\
     {\
         /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = (n);\
         PRIV(op)->wait_for = WAIT_SQE;\
         return 0;\
     }
@@ -71,6 +72,7 @@
     if (!sqe)\
     {\
         /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
         PRIV(op)->wait_for = WAIT_SQE;\
         return 0;\
     }
@@ -80,6 +82,7 @@
     if (!sqe)\
     {\
         /* Pause until there are more requests available */\
+        PRIV(op)->wait_detail = 1;\
         PRIV(op)->wait_for = WAIT_SQE;\
         return 0;\
     }
@@ -201,6 +204,17 @@ typedef std::map<obj_ver_id, dirty_entry> blockstore_dirty_db_t;
 
 #include "blockstore_flush.h"
 
+typedef uint32_t pool_id_t;
+typedef uint64_t pool_pg_id_t;
+
+#define POOL_ID_BITS 16
+
+struct pool_shard_settings_t
+{
+    uint32_t pg_count;
+    uint32_t pg_stripe_size;
+};
+
 class blockstore_impl_t
 {
     /******* OPTIONS *******/
@@ -238,11 +252,14 @@ 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 number of LIST operations to be processed between
+    int single_tick_list_limit = 1;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;
 
-    blockstore_clean_db_t clean_db;
+    std::map<pool_id_t, pool_shard_settings_t> clean_db_settings;
+    std::map<pool_pg_id_t, blockstore_clean_db_t> clean_db_shards;
     uint8_t *clean_bitmap = NULL;
     blockstore_dirty_db_t dirty_db;
     std::vector<blockstore_op_t*> submit_queue;
@@ -291,6 +308,9 @@ class blockstore_impl_t
     void open_journal();
     uint8_t* get_clean_entry_bitmap(uint64_t block_loc, int offset);
 
+    blockstore_clean_db_t& clean_db_shard(object_id oid);
+    void reshard_clean_db(pool_id_t pool_id, uint32_t pg_count, uint32_t pg_stripe_size);
+
     // Journaling
     void prepare_journal_sector_write(int sector, blockstore_op_t *op);
     void handle_journal_write(ring_data_t *data, uint64_t flush_id);

src/blockstore_init.cpp

@@ -222,10 +222,11 @@ void blockstore_init_meta::handle_entries(void* entries, unsigned count, int blo
         }
         if (entry->oid.inode > 0)
         {
-            auto clean_it = bs->clean_db.find(entry->oid);
-            if (clean_it == bs->clean_db.end() || clean_it->second.version < entry->version)
+            auto & clean_db = bs->clean_db_shard(entry->oid);
+            auto clean_it = clean_db.find(entry->oid);
+            if (clean_it == clean_db.end() || clean_it->second.version < entry->version)
             {
-                if (clean_it != bs->clean_db.end())
+                if (clean_it != clean_db.end())
                 {
                     // free the previous block
 #ifdef BLOCKSTORE_DEBUG
@@ -245,7 +246,7 @@ void blockstore_init_meta::handle_entries(void* entries, unsigned count, int blo
                 printf("Allocate block (clean entry) %lu: %lx:%lx v%lu\n", done_cnt+i, entry->oid.inode, entry->oid.stripe, entry->version);
 #endif
                 bs->data_alloc->set(done_cnt+i, true);
-                bs->clean_db[entry->oid] = (struct clean_entry){
+                clean_db[entry->oid] = (struct clean_entry){
                     .version = entry->version,
                     .location = (done_cnt+i) << block_order,
                 };
@@ -656,8 +657,9 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     init_write_sector = proc_pos;
                     return 0;
                 }
-                auto clean_it = bs->clean_db.find(je->small_write.oid);
-                if (clean_it == bs->clean_db.end() ||
+                auto & clean_db = bs->clean_db_shard(je->small_write.oid);
+                auto clean_it = clean_db.find(je->small_write.oid);
+                if (clean_it == clean_db.end() ||
                     clean_it->second.version < je->small_write.version)
                 {
                     obj_ver_id ov = {
@@ -735,8 +737,9 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         erase_dirty_object(dirty_it);
                     }
                 }
-                auto clean_it = bs->clean_db.find(je->big_write.oid);
-                if (clean_it == bs->clean_db.end() ||
+                auto & clean_db = bs->clean_db_shard(je->big_write.oid);
+                auto clean_it = clean_db.find(je->big_write.oid);
+                if (clean_it == clean_db.end() ||
                     clean_it->second.version < je->big_write.version)
                 {
                     // oid, version, block
@@ -841,8 +844,9 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     dirty_it--;
                     dirty_exists = dirty_it->first.oid == je->del.oid;
                 }
-                auto clean_it = bs->clean_db.find(je->del.oid);
-                bool clean_exists = (clean_it != bs->clean_db.end() &&
+                auto & clean_db = bs->clean_db_shard(je->del.oid);
+                auto clean_it = clean_db.find(je->del.oid);
+                bool clean_exists = (clean_it != clean_db.end() &&
                     clean_it->second.version < je->del.version);
                 if (!clean_exists && dirty_exists)
                 {
@@ -901,8 +905,9 @@ void blockstore_init_journal::erase_dirty_object(blockstore_dirty_db_t::iterator
             break;
         }
     }
-    auto clean_it = bs->clean_db.find(oid);
-    uint64_t clean_loc = clean_it != bs->clean_db.end()
+    auto & clean_db = bs->clean_db_shard(oid);
+    auto clean_it = clean_db.find(oid);
+    uint64_t clean_loc = clean_it != clean_db.end()
         ? clean_it->second.location : UINT64_MAX;
     if (exists && clean_loc == UINT64_MAX)
     {

src/blockstore_read.cpp

@@ -111,6 +111,7 @@ uint8_t* blockstore_impl_t::get_clean_entry_bitmap(uint64_t block_loc, int offse
 
 int blockstore_impl_t::dequeue_read(blockstore_op_t *read_op)
 {
+    auto & clean_db = clean_db_shard(read_op->oid);
     auto clean_it = clean_db.find(read_op->oid);
     auto dirty_it = dirty_db.upper_bound((obj_ver_id){
         .oid = read_op->oid,
@@ -297,6 +298,7 @@ int blockstore_impl_t::read_bitmap(object_id oid, uint64_t target_version, void
             dirty_it--;
         }
     }
+    auto & clean_db = clean_db_shard(oid);
     auto clean_it = clean_db.find(oid);
     if (clean_it != clean_db.end())
     {

src/blockstore_stable.cpp

@@ -54,6 +54,7 @@ int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
         auto dirty_it = dirty_db.find(*v);
         if (dirty_it == dirty_db.end())
         {
+            auto & clean_db = clean_db_shard(v->oid);
             auto clean_it = clean_db.find(v->oid);
             if (clean_it == clean_db.end() || clean_it->second.version < v->version)
             {
@@ -188,6 +189,7 @@ void blockstore_impl_t::mark_stable(const obj_ver_id & v, bool forget_dirty)
                     }
                     if (exists == -1)
                     {
+                        auto & clean_db = clean_db_shard(v.oid);
                         auto clean_it = clean_db.find(v.oid);
                         exists = clean_it != clean_db.end() ? 1 : 0;
                     }
@@ -215,6 +217,7 @@ void blockstore_impl_t::mark_stable(const obj_ver_id & v, bool forget_dirty)
                         break;
                     }
                 }
+                auto & clean_db = clean_db_shard(v.oid);
                 auto clean_it = clean_db.find(v.oid);
                 uint64_t clean_loc = clean_it != clean_db.end()
                     ? clean_it->second.location : UINT64_MAX;

src/blockstore_write.cpp

@@ -41,6 +41,7 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
     }
     if (!found)
     {
+        auto & clean_db = clean_db_shard(op->oid);
         auto clean_it = clean_db.find(op->oid);
         if (clean_it != clean_db.end())
         {
@@ -543,12 +544,13 @@ resume_4:
             if (ref_us > exec_us + throttle_threshold_us)
             {
                 // Pause reply
+                PRIV(op)->op_state = 5;
+                // Remember that the timer can in theory be called right here
                 tfd->set_timer_us(ref_us-exec_us, false, [this, op](int timer_id)
                 {
                     PRIV(op)->op_state++;
                     ringloop->wakeup();
                 });
-                PRIV(op)->op_state = 5;
                 return 1;
             }
         }

src/cli.cpp

@@ -86,6 +86,9 @@ void cli_tool_t::help()
         "(c) Vitaliy Filippov, 2019+ (VNPL-1.1)\n"
         "\n"
         "USAGE:\n"
+        "%s status\n"
+        "  Show cluster status\n"
+        "\n"
         "%s df\n"
         "  Show pool space statistics\n"
         "\n"
@@ -155,7 +158,7 @@ void cli_tool_t::help()
         "  --no-color          Disable colored output\n"
         "  --json              JSON output\n"
         ,
-        exe_name, exe_name, exe_name, exe_name, exe_name, exe_name,
+        exe_name, exe_name, exe_name, exe_name, exe_name, exe_name, exe_name,
         exe_name, exe_name, exe_name, exe_name, exe_name, exe_name
     );
     exit(0);
@@ -266,6 +269,11 @@ void cli_tool_t::run(json11::Json cfg)
         fprintf(stderr, "command is missing\n");
         exit(1);
     }
+    else if (cmd[0] == "status")
+    {
+        // Show cluster status
+        action_cb = start_status(cfg);
+    }
     else if (cmd[0] == "df")
     {
         // Show pool space stats
@@ -340,6 +348,8 @@ void cli_tool_t::run(json11::Json cfg)
     ringloop = new ring_loop_t(512);
     epmgr = new epoll_manager_t(ringloop);
     cli = new cluster_client_t(ringloop, epmgr->tfd, cfg);
+    // Smaller timeout by default for more interactiveness
+    cli->st_cli.etcd_slow_timeout = cli->st_cli.etcd_quick_timeout;
     cli->on_ready([this]()
     {
         // Initialize job
@@ -365,6 +375,13 @@ void cli_tool_t::run(json11::Json cfg)
         if (action_cb != NULL)
             ringloop->wait();
     }
+    // Destroy the client
+    delete cli;
+    delete epmgr;
+    delete ringloop;
+    cli = NULL;
+    epmgr = NULL;
+    ringloop = NULL;
 }
 
 int main(int narg, const char *args[])
@@ -374,5 +391,6 @@ int main(int narg, const char *args[])
     exe_name = args[0];
     cli_tool_t *p = new cli_tool_t();
     p->run(cli_tool_t::parse_args(narg, args));
+    delete p;
     return 0;
 }

src/cli.h

@@ -51,6 +51,7 @@ public:
     friend struct snap_flattener_t;
     friend struct snap_remover_t;
 
+    std::function<bool(void)> start_status(json11::Json cfg);
     std::function<bool(void)> start_df(json11::Json);
     std::function<bool(void)> start_ls(json11::Json);
     std::function<bool(void)> start_create(json11::Json);
@@ -69,7 +70,7 @@ uint64_t parse_size(std::string size_str);
 
 std::string print_table(json11::Json items, json11::Json header, bool use_esc);
 
-std::string format_size(uint64_t size);
+std::string format_size(uint64_t size, bool nobytes = false);
 
 std::string format_lat(uint64_t lat);
 

src/cli_df.cpp

@@ -108,9 +108,14 @@ resume_1:
                     pool_avail = pg_free;
                 }
             }
+            if (pool_avail == UINT64_MAX)
+            {
+                pool_avail = 0;
+            }
             if (pool_cfg.scheme != POOL_SCHEME_REPLICATED)
             {
-                pool_avail = pool_avail * (pool_cfg.pg_size - pool_cfg.parity_chunks) / pool_stats[pool_cfg.id]["pg_real_size"].uint64_value();
+                uint64_t pg_real_size = pool_stats[pool_cfg.id]["pg_real_size"].uint64_value();
+                pool_avail = pg_real_size > 0 ? pool_avail * (pool_cfg.pg_size - pool_cfg.parity_chunks) / pg_real_size : 0;
             }
             pool_stats[pool_cfg.id] = json11::Json::object {
                 { "name", pool_cfg.name },
@@ -189,11 +194,16 @@ resume_1:
         json11::Json::array list;
         for (auto & kv: pool_stats)
         {
-            kv.second["total_fmt"] = format_size(kv.second["total_raw"].uint64_value() / kv.second["raw_to_usable"].number_value());
-            kv.second["used_fmt"] = format_size(kv.second["used_raw"].uint64_value() / kv.second["raw_to_usable"].number_value());
+            double raw_to = kv.second["raw_to_usable"].number_value();
+            if (raw_to < 0.000001 && raw_to > -0.000001)
+                raw_to = 1;
+            kv.second["total_fmt"] = format_size(kv.second["total_raw"].uint64_value() / raw_to);
+            kv.second["used_fmt"] = format_size(kv.second["used_raw"].uint64_value() / raw_to);
             kv.second["max_avail_fmt"] = format_size(kv.second["max_available"].uint64_value());
-            kv.second["used_pct"] = format_q(100 - 100*kv.second["max_available"].uint64_value() *
-                kv.second["raw_to_usable"].number_value() / kv.second["total_raw"].uint64_value())+"%";
+            kv.second["used_pct"] = format_q(kv.second["total_raw"].uint64_value()
+                ? (100 - 100*kv.second["max_available"].uint64_value() *
+                    kv.second["raw_to_usable"].number_value() / kv.second["total_raw"].uint64_value())
+                : 100)+"%";
             kv.second["eff_fmt"] = format_q(kv.second["space_efficiency"].number_value()*100)+"%";
         }
         printf("%s", print_table(to_list(), cols, parent->color).c_str());

src/cli_ls.cpp

@@ -154,7 +154,7 @@ resume_1:
             if (pool_it != parent->cli->st_cli.pool_config.end())
             {
                 auto & pool_cfg = pool_it->second;
-                used_size = used_size / pool_pg_real_size[pool_id]
+                used_size = used_size / (pool_pg_real_size[pool_id] ? pool_pg_real_size[pool_id] : 1)
                     * (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pool_cfg.pg_size-pool_cfg.parity_chunks);
             }
             auto stat_it = stats.find(inode_num);
@@ -437,22 +437,25 @@ std::string print_table(json11::Json items, json11::Json header, bool use_esc)
 }
 
 static uint64_t size_thresh[] = { 1024l*1024*1024*1024, 1024l*1024*1024, 1024l*1024, 1024, 0 };
+static uint64_t size_thresh_d[] = { 1000000000000l, 1000000000l, 1000000l, 1000l, 0 };
+static const int size_thresh_n = sizeof(size_thresh)/sizeof(size_thresh[0]);
 static const char *size_unit = "TGMKB";
 
-std::string format_size(uint64_t size)
+std::string format_size(uint64_t size, bool nobytes)
 {
+    uint64_t *thr = nobytes ? size_thresh_d : size_thresh;
     char buf[256];
-    for (int i = 0; i < sizeof(size_thresh)/sizeof(size_thresh[0]); i++)
+    for (int i = 0; i < size_thresh_n; i++)
     {
-        if (size >= size_thresh[i] || i >= sizeof(size_thresh)/sizeof(size_thresh[0])-1)
+        if (size >= thr[i] || i >= size_thresh_n-1)
         {
-            double value = size_thresh[i] ? (double)size/size_thresh[i] : size;
+            double value = thr[i] ? (double)size/thr[i] : size;
             int l = snprintf(buf, sizeof(buf), "%.1f", value);
             assert(l < sizeof(buf)-2);
             if (buf[l-1] == '0')
                 l -= 2;
-            buf[l] = ' ';
-            buf[l+1] = size_unit[i];
+            buf[l] = i == size_thresh_n-1 && nobytes ? 0 : ' ';
+            buf[l+1] = i == size_thresh_n-1 && nobytes ? 0 : size_unit[i];
             buf[l+2] = 0;
             break;
         }

src/cli_rm.cpp

@@ -1,211 +1,566 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 
+#include <fcntl.h>
 #include "cli.h"
 #include "cluster_client.h"
+#include "base64.h"
 
-#define RM_LISTING 1
-#define RM_REMOVING 2
-#define RM_END 3
-
-struct rm_pg_t
+// Remove layer(s): similar to merge, but alters metadata and processes multiple merge targets
+//
+// Exactly one child of the requested layers may be merged using the "inverted" workflow,
+// where we merge it "down" into one of the "to-be-removed" layers and then rename the
+// "to-be-removed" layer to the child. It may be done either if all writers are stopped
+// before trying to delete layers (which is signaled by --writers-stopped) or if that child
+// is a read-only layer (snapshot) itself.
+//
+// This "inverted" workflow trades copying data of one of the deleted layers for copying
+// data of one child of the chain which is also a child of the "traded" layer. So we
+// choose the (parent,child) pair which has the largest difference between "parent" and
+// "child" inode sizes.
+//
+// All other children of the chain are processed by iterating though them, merging removed
+// parents into them and rebasing them to the last layer which isn't a member of the removed
+// chain.
+//
+// Example:
+//
+// <parent> - <from> - <layer 2> - <to> - <child 1>
+//                 \           \       \- <child 2>
+//                  \           \- <child 3>
+//                   \-<child 4>
+//
+// 1) Find optimal pair for the "reverse" scenario
+//    Imagine that it's (<layer 2>, <child 1>) in this example
+// 2) Process all children except <child 1>:
+//    - Merge <from>..<to> to <child 2>
+//    - Set <child 2> parent to <parent>
+//    - Repeat for others
+// 3) Process <child 1>:
+//    - Merge <from>..<child 1> to <layer 2>
+//    - Set <layer 2> parent to <parent>
+//    - Rename <layer 2> to <child 1>
+// 4) Delete other layers of the chain (<from>, <to>)
+struct snap_remover_t
 {
-    pg_num_t pg_num;
-    osd_num_t rm_osd_num;
-    std::set<object_id> objects;
-    std::set<object_id>::iterator obj_pos;
-    uint64_t obj_count = 0, obj_done = 0;
+    cli_tool_t *parent;
+
+    // remove from..to
+    std::string from_name, to_name;
+    // writers are stopped, we can safely change writable layers
+    bool writers_stopped = false;
+    // use CAS writes (0 = never, 1 = auto, 2 = always)
+    int use_cas = 1;
+    // interval between fsyncs
+    int fsync_interval = 128;
+
+    std::map<inode_t,int> sources;
+    std::map<inode_t,uint64_t> inode_used;
+    std::vector<inode_t> merge_children;
+    std::vector<inode_t> chain_list;
+    std::map<inode_t,int> inverse_candidates;
+    inode_t inverse_parent = 0, inverse_child = 0;
+    inode_t new_parent = 0;
     int state = 0;
-    int in_flight = 0;
-};
+    int current_child = 0;
+    std::function<bool(void)> cb;
 
-struct rm_inode_t
-{
-    uint64_t inode = 0;
-    pool_id_t pool_id = 0;
-    uint64_t min_offset = 0;
-
-    cli_tool_t *parent = NULL;
-    inode_list_t *lister = NULL;
-    std::vector<rm_pg_t*> lists;
-    uint64_t total_count = 0, total_done = 0, total_prev_pct = 0;
-    uint64_t pgs_to_list = 0;
-    bool lists_done = false;
-    int state = 0;
-
-    void start_delete()
+    bool is_done()
     {
-        lister = parent->cli->list_inode_start(inode, [this](inode_list_t *lst,
-            std::set<object_id>&& objects, pg_num_t pg_num, osd_num_t primary_osd, int status)
+        return state == 9;
+    }
+
+    void loop()
+    {
+        if (state == 1)
+            goto resume_1;
+        else if (state == 2)
+            goto resume_2;
+        else if (state == 3)
+            goto resume_3;
+        else if (state == 4)
+            goto resume_4;
+        else if (state == 5)
+            goto resume_5;
+        else if (state == 6)
+            goto resume_6;
+        else if (state == 7)
+            goto resume_7;
+        else if (state == 8)
+            goto resume_8;
+        else if (state == 9)
+            goto resume_9;
+        // Get children to merge
+        get_merge_children();
+        // Try to select an inode for the "inverse" optimized scenario
+        // Read statistics from etcd to do it
+        read_stats();
+        state = 1;
+resume_1:
+        if (parent->waiting > 0)
+            return;
+        choose_inverse_candidate();
+        // Merge children one by one, except our "inverse" child
+        for (current_child = 0; current_child < merge_children.size(); current_child++)
         {
-            rm_pg_t *rm = new rm_pg_t((rm_pg_t){
-                .pg_num = pg_num,
-                .rm_osd_num = primary_osd,
-                .objects = objects,
-                .obj_count = objects.size(),
-                .obj_done = 0,
-            });
-            if (min_offset == 0)
+            if (merge_children[current_child] == inverse_child)
+                continue;
+            start_merge_child(merge_children[current_child], merge_children[current_child]);
+resume_2:
+            while (!cb())
             {
-                total_count += objects.size();
+                state = 2;
+                return;
             }
-            else
-            {
-                for (object_id oid: objects)
-                {
-                    if (oid.stripe >= min_offset)
-                    {
-                        total_count++;
-                    }
-                }
-            }
-            rm->obj_pos = rm->objects.begin();
-            lists.push_back(rm);
-            if (parent->list_first)
-            {
-                parent->cli->list_inode_next(lister, 1);
-            }
-            if (status & INODE_LIST_DONE)
-            {
-                lists_done = true;
-            }
-            pgs_to_list--;
-            continue_delete();
-        });
-        if (!lister)
+            cb = NULL;
+            parent->change_parent(merge_children[current_child], new_parent);
+            state = 3;
+resume_3:
+            if (parent->waiting > 0)
+                return;
+        }
+        // Merge our "inverse" child into our "inverse" parent
+        if (inverse_child != 0)
         {
-            fprintf(stderr, "Failed to list inode %lu from pool %u objects\n", INODE_NO_POOL(inode), INODE_POOL(inode));
+            start_merge_child(inverse_child, inverse_parent);
+resume_4:
+            while (!cb())
+            {
+                state = 4;
+                return;
+            }
+            cb = NULL;
+            // Delete "inverse" child data
+            start_delete_source(inverse_child);
+resume_5:
+            while (!cb())
+            {
+                state = 5;
+                return;
+            }
+            cb = NULL;
+            // Delete "inverse" child metadata, rename parent over it,
+            // and also change parent links of the previous "inverse" child
+            rename_inverse_parent();
+            state = 6;
+resume_6:
+            if (parent->waiting > 0)
+                return;
+        }
+        // Delete parents, except the "inverse" one
+        for (current_child = 0; current_child < chain_list.size(); current_child++)
+        {
+            if (chain_list[current_child] == inverse_parent)
+                continue;
+            start_delete_source(chain_list[current_child]);
+resume_7:
+            while (!cb())
+            {
+                state = 7;
+                return;
+            }
+            cb = NULL;
+            delete_inode_config(chain_list[current_child]);
+            state = 8;
+resume_8:
+            if (parent->waiting > 0)
+                return;
+        }
+        state = 9;
+resume_9:
+        // Done
+        return;
+    }
+
+    void get_merge_children()
+    {
+        // Get all children of from..to
+        inode_config_t *from_cfg = parent->get_inode_cfg(from_name);
+        inode_config_t *to_cfg = parent->get_inode_cfg(to_name);
+        // Check that to_cfg is actually a child of from_cfg
+        // FIXME de-copypaste the following piece of code with snap_merger_t
+        inode_config_t *cur = to_cfg;
+        chain_list.push_back(cur->num);
+        while (cur->num != from_cfg->num && cur->parent_id != 0)
+        {
+            auto it = parent->cli->st_cli.inode_config.find(cur->parent_id);
+            if (it == parent->cli->st_cli.inode_config.end())
+            {
+                fprintf(stderr, "Parent inode of layer %s (id %ld) not found\n", cur->name.c_str(), cur->parent_id);
+                exit(1);
+            }
+            cur = &it->second;
+            chain_list.push_back(cur->num);
+        }
+        if (cur->num != from_cfg->num)
+        {
+            fprintf(stderr, "Layer %s is not a child of %s\n", to_name.c_str(), from_name.c_str());
             exit(1);
         }
-        pgs_to_list = parent->cli->list_pg_count(lister);
-        parent->cli->list_inode_next(lister, parent->parallel_osds);
-    }
-
-    void send_ops(rm_pg_t *cur_list)
-    {
-        if (parent->cli->msgr.osd_peer_fds.find(cur_list->rm_osd_num) ==
-            parent->cli->msgr.osd_peer_fds.end())
+        new_parent = from_cfg->parent_id;
+        // Calculate ranks
+        int i = chain_list.size()-1;
+        for (inode_t item: chain_list)
         {
-            // Initiate connection
-            parent->cli->msgr.connect_peer(cur_list->rm_osd_num, parent->cli->st_cli.peer_states[cur_list->rm_osd_num]);
-            return;
+            sources[item] = i--;
         }
-        while (cur_list->in_flight < parent->iodepth && cur_list->obj_pos != cur_list->objects.end())
+        for (auto & ic: parent->cli->st_cli.inode_config)
         {
-            if (cur_list->obj_pos->stripe >= min_offset)
+            if (!ic.second.parent_id)
             {
-                osd_op_t *op = new osd_op_t();
-                op->op_type = OSD_OP_OUT;
-                op->peer_fd = parent->cli->msgr.osd_peer_fds[cur_list->rm_osd_num];
-                op->req = (osd_any_op_t){
-                    .rw = {
-                        .header = {
-                            .magic = SECONDARY_OSD_OP_MAGIC,
-                            .id = parent->cli->next_op_id(),
-                            .opcode = OSD_OP_DELETE,
-                        },
-                        .inode = cur_list->obj_pos->inode,
-                        .offset = cur_list->obj_pos->stripe,
-                        .len = 0,
-                    },
-                };
-                op->callback = [this, cur_list](osd_op_t *op)
-                {
-                    cur_list->in_flight--;
-                    if (op->reply.hdr.retval < 0)
-                    {
-                        fprintf(stderr, "Failed to remove object %lx:%lx from PG %u (OSD %lu) (retval=%ld)\n",
-                            op->req.rw.inode, op->req.rw.offset,
-                            cur_list->pg_num, cur_list->rm_osd_num, op->reply.hdr.retval);
-                    }
-                    delete op;
-                    cur_list->obj_done++;
-                    total_done++;
-                    continue_delete();
-                };
-                cur_list->in_flight++;
-                parent->cli->msgr.outbox_push(op);
+                continue;
             }
-            cur_list->obj_pos++;
-        }
-    }
-
-    void continue_delete()
-    {
-        if (parent->list_first && !lists_done)
-        {
-            return;
-        }
-        for (int i = 0; i < lists.size(); i++)
-        {
-            if (!lists[i]->in_flight && lists[i]->obj_pos == lists[i]->objects.end())
+            auto it = sources.find(ic.second.parent_id);
+            if (it != sources.end() && sources.find(ic.second.num) == sources.end())
             {
-                delete lists[i];
-                lists.erase(lists.begin()+i, lists.begin()+i+1);
-                i--;
-                if (!lists_done)
+                merge_children.push_back(ic.second.num);
+                if (ic.second.readonly || writers_stopped)
                 {
-                    parent->cli->list_inode_next(lister, 1);
+                    inverse_candidates[ic.second.num] = it->second;
                 }
             }
-            else
-            {
-                send_ops(lists[i]);
-            }
-        }
-        if (parent->progress && total_count > 0 && total_done*1000/total_count != total_prev_pct)
-        {
-            printf("\rRemoved %lu/%lu objects, %lu more PGs to list...", total_done, total_count, pgs_to_list);
-            total_prev_pct = total_done*1000/total_count;
-        }
-        if (lists_done && !lists.size())
-        {
-            printf("Done, inode %lu in pool %u data removed\n", INODE_NO_POOL(inode), pool_id);
-            state = 2;
         }
     }
 
-    bool loop()
+    void read_stats()
     {
-        if (state == 0)
+        if (inverse_candidates.size() == 0)
         {
-            start_delete();
-            state = 1;
+            return;
         }
-        else if (state == 1)
+        json11::Json::array reads;
+        for (auto cp: inverse_candidates)
         {
-            continue_delete();
+            inode_t inode = cp.first;
+            reads.push_back(json11::Json::object {
+                { "request_range", json11::Json::object {
+                    { "key", base64_encode(
+                        parent->cli->st_cli.etcd_prefix+
+                        "/inode/stats/"+std::to_string(INODE_POOL(inode))+
+                        "/"+std::to_string(INODE_NO_POOL(inode))
+                    ) },
+                } }
+            });
         }
-        else if (state == 2)
+        for (auto cp: sources)
         {
-            return true;
+            inode_t inode = cp.first;
+            reads.push_back(json11::Json::object {
+                { "request_range", json11::Json::object {
+                    { "key", base64_encode(
+                        parent->cli->st_cli.etcd_prefix+
+                        "/inode/stats/"+std::to_string(INODE_POOL(inode))+
+                        "/"+std::to_string(INODE_NO_POOL(inode))
+                    ) },
+                } }
+            });
         }
-        return false;
+        parent->waiting++;
+        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
+            { "success", reads },
+        }, [this](std::string err, json11::Json data)
+        {
+            parent->waiting--;
+            if (err != "")
+            {
+                fprintf(stderr, "Error reading layer statistics from etcd: %s\n", err.c_str());
+                exit(1);
+            }
+            for (auto inode_result: data["responses"].array_items())
+            {
+                auto kv = parent->cli->st_cli.parse_etcd_kv(inode_result["kvs"][0]);
+                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)
+                {
+                    fprintf(stderr, "Bad key returned from etcd: %s\n", kv.key.c_str());
+                    exit(1);
+                }
+                auto pool_cfg_it = parent->cli->st_cli.pool_config.find(pool_id);
+                if (pool_cfg_it == parent->cli->st_cli.pool_config.end())
+                {
+                    fprintf(stderr, "Pool %u does not exist\n", pool_id);
+                    exit(1);
+                }
+                inode = INODE_WITH_POOL(pool_id, inode);
+                auto & pool_cfg = pool_cfg_it->second;
+                uint64_t used_bytes = kv.value["raw_used"].uint64_value() / pool_cfg.pg_size;
+                if (pool_cfg.scheme != POOL_SCHEME_REPLICATED)
+                {
+                    used_bytes *= (pool_cfg.pg_size - pool_cfg.parity_chunks);
+                }
+                inode_used[inode] = used_bytes;
+            }
+            parent->ringloop->wakeup();
+        });
+    }
+
+    void choose_inverse_candidate()
+    {
+        uint64_t max_diff = 0;
+        for (auto cp: inverse_candidates)
+        {
+            inode_t child = cp.first;
+            uint64_t child_used = inode_used[child];
+            int rank = cp.second;
+            for (int i = chain_list.size()-rank; i < chain_list.size(); i++)
+            {
+                inode_t parent = chain_list[i];
+                uint64_t parent_used = inode_used[parent];
+                if (parent_used > child_used && (!max_diff || max_diff < (parent_used-child_used)))
+                {
+                    max_diff = (parent_used-child_used);
+                    inverse_parent = parent;
+                    inverse_child = child;
+                }
+            }
+        }
+    }
+
+    void rename_inverse_parent()
+    {
+        auto child_it = parent->cli->st_cli.inode_config.find(inverse_child);
+        if (child_it == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode %ld disappeared\n", inverse_child);
+            exit(1);
+        }
+        auto target_it = parent->cli->st_cli.inode_config.find(inverse_parent);
+        if (target_it == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode %ld disappeared\n", inverse_parent);
+            exit(1);
+        }
+        inode_config_t *child_cfg = &child_it->second;
+        inode_config_t *target_cfg = &target_it->second;
+        std::string child_name = child_cfg->name;
+        std::string target_name = target_cfg->name;
+        std::string child_cfg_key = base64_encode(
+            parent->cli->st_cli.etcd_prefix+
+            "/config/inode/"+std::to_string(INODE_POOL(inverse_child))+
+            "/"+std::to_string(INODE_NO_POOL(inverse_child))
+        );
+        std::string target_cfg_key = base64_encode(
+            parent->cli->st_cli.etcd_prefix+
+            "/config/inode/"+std::to_string(INODE_POOL(inverse_parent))+
+            "/"+std::to_string(INODE_NO_POOL(inverse_parent))
+        );
+        // Fill new configuration
+        inode_config_t new_cfg = *child_cfg;
+        new_cfg.num = target_cfg->num;
+        new_cfg.parent_id = new_parent;
+        json11::Json::array cmp = json11::Json::array {
+            json11::Json::object {
+                { "target", "MOD" },
+                { "key", child_cfg_key },
+                { "result", "LESS" },
+                { "mod_revision", child_cfg->mod_revision+1 },
+            },
+            json11::Json::object {
+                { "target", "MOD" },
+                { "key", target_cfg_key },
+                { "result", "LESS" },
+                { "mod_revision", target_cfg->mod_revision+1 },
+            },
+        };
+        json11::Json::array txn = json11::Json::array {
+            json11::Json::object {
+                { "request_delete_range", json11::Json::object {
+                    { "key", child_cfg_key },
+                } },
+            },
+            json11::Json::object {
+                { "request_put", json11::Json::object {
+                    { "key", target_cfg_key },
+                    { "value", base64_encode(json11::Json(parent->cli->st_cli.serialize_inode_cfg(&new_cfg)).dump()) },
+                } },
+            },
+            json11::Json::object {
+                { "request_put", json11::Json::object {
+                    { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/index/image/"+child_cfg->name) },
+                    { "value", base64_encode(json11::Json({
+                        { "id", INODE_NO_POOL(inverse_parent) },
+                        { "pool_id", (uint64_t)INODE_POOL(inverse_parent) },
+                    }).dump()) },
+                } },
+            },
+        };
+        // Reparent children of inverse_child
+        for (auto & cp: parent->cli->st_cli.inode_config)
+        {
+            if (cp.second.parent_id == child_cfg->num)
+            {
+                auto cp_cfg = cp.second;
+                cp_cfg.parent_id = inverse_parent;
+                auto cp_key = base64_encode(
+                    parent->cli->st_cli.etcd_prefix+
+                    "/config/inode/"+std::to_string(INODE_POOL(cp.second.num))+
+                    "/"+std::to_string(INODE_NO_POOL(cp.second.num))
+                );
+                cmp.push_back(json11::Json::object {
+                    { "target", "MOD" },
+                    { "key", cp_key },
+                    { "result", "LESS" },
+                    { "mod_revision", cp.second.mod_revision+1 },
+                });
+                txn.push_back(json11::Json::object {
+                    { "request_put", json11::Json::object {
+                        { "key", cp_key },
+                        { "value", base64_encode(json11::Json(parent->cli->st_cli.serialize_inode_cfg(&cp_cfg)).dump()) },
+                    } },
+                });
+            }
+        }
+        parent->waiting++;
+        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
+            { "compare", cmp },
+            { "success", txn },
+        }, [this, target_name, child_name](std::string err, json11::Json res)
+        {
+            parent->waiting--;
+            if (err != "")
+            {
+                fprintf(stderr, "Error renaming %s to %s: %s\n", target_name.c_str(), child_name.c_str(), err.c_str());
+                exit(1);
+            }
+            if (!res["succeeded"].bool_value())
+            {
+                fprintf(
+                    stderr, "Parent (%s), child (%s), or one of its children"
+                    " configuration was modified during rename\n", target_name.c_str(), child_name.c_str()
+                );
+                exit(1);
+            }
+            printf("Layer %s renamed to %s\n", target_name.c_str(), child_name.c_str());
+            parent->ringloop->wakeup();
+        });
+    }
+
+    void delete_inode_config(inode_t cur)
+    {
+        auto cur_cfg_it = parent->cli->st_cli.inode_config.find(cur);
+        if (cur_cfg_it == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode 0x%lx disappeared\n", cur);
+            exit(1);
+        }
+        inode_config_t *cur_cfg = &cur_cfg_it->second;
+        std::string cur_name = cur_cfg->name;
+        std::string cur_cfg_key = base64_encode(
+            parent->cli->st_cli.etcd_prefix+
+            "/config/inode/"+std::to_string(INODE_POOL(cur))+
+            "/"+std::to_string(INODE_NO_POOL(cur))
+        );
+        parent->waiting++;
+        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
+            { "compare", json11::Json::array {
+                json11::Json::object {
+                    { "target", "MOD" },
+                    { "key", cur_cfg_key },
+                    { "result", "LESS" },
+                    { "mod_revision", cur_cfg->mod_revision+1 },
+                },
+            } },
+            { "success", json11::Json::array {
+                json11::Json::object {
+                    { "request_delete_range", json11::Json::object {
+                        { "key", cur_cfg_key },
+                    } },
+                },
+                json11::Json::object {
+                    { "request_delete_range", json11::Json::object {
+                        { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/index/image/"+cur_name) },
+                    } },
+                },
+            } },
+        }, [this, cur_name](std::string err, json11::Json res)
+        {
+            parent->waiting--;
+            if (err != "")
+            {
+                fprintf(stderr, "Error deleting %s: %s\n", cur_name.c_str(), err.c_str());
+                exit(1);
+            }
+            if (!res["succeeded"].bool_value())
+            {
+                fprintf(stderr, "Layer %s configuration was modified during deletion\n", cur_name.c_str());
+                exit(1);
+            }
+            printf("Layer %s deleted\n", cur_name.c_str());
+            parent->ringloop->wakeup();
+        });
+    }
+
+    void start_merge_child(inode_t child_inode, inode_t target_inode)
+    {
+        auto child_it = parent->cli->st_cli.inode_config.find(child_inode);
+        if (child_it == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode %ld disappeared\n", child_inode);
+            exit(1);
+        }
+        auto target_it = parent->cli->st_cli.inode_config.find(target_inode);
+        if (target_it == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode %ld disappeared\n", target_inode);
+            exit(1);
+        }
+        cb = parent->start_merge(json11::Json::object {
+            { "command", json11::Json::array{ "merge-data", from_name, child_it->second.name } },
+            { "target", target_it->second.name },
+            { "delete-source", false },
+            { "cas", use_cas },
+            { "fsync-interval", fsync_interval },
+        });
+    }
+
+    void start_delete_source(inode_t inode)
+    {
+        auto source = parent->cli->st_cli.inode_config.find(inode);
+        if (source == parent->cli->st_cli.inode_config.end())
+        {
+            fprintf(stderr, "Inode %ld disappeared\n", inode);
+            exit(1);
+        }
+        cb = parent->start_rm(json11::Json::object {
+            { "inode", inode },
+            { "pool", (uint64_t)INODE_POOL(inode) },
+            { "fsync-interval", fsync_interval },
+        });
     }
 };
 
-std::function<bool(void)> cli_tool_t::start_rm(json11::Json cfg)
+std::function<bool(void)> cli_tool_t::start_snap_rm(json11::Json cfg)
 {
-    auto remover = new rm_inode_t();
-    remover->parent = this;
-    remover->inode = cfg["inode"].uint64_value();
-    remover->pool_id = cfg["pool"].uint64_value();
-    if (remover->pool_id)
+    json11::Json::array cmd = cfg["command"].array_items();
+    auto snap_remover = new snap_remover_t();
+    snap_remover->parent = this;
+    snap_remover->from_name = cmd.size() > 1 ? cmd[1].string_value() : "";
+    snap_remover->to_name = cmd.size() > 2 ? cmd[2].string_value() : "";
+    if (snap_remover->from_name == "")
     {
-        remover->inode = (remover->inode & ((1l << (64-POOL_ID_BITS)) - 1)) | (((uint64_t)remover->pool_id) << (64-POOL_ID_BITS));
-    }
-    remover->pool_id = INODE_POOL(remover->inode);
-    if (!remover->pool_id)
-    {
-        fprintf(stderr, "pool is missing\n");
+        fprintf(stderr, "Layer to remove argument is missing\n");
         exit(1);
     }
-    remover->min_offset = cfg["min-offset"].uint64_value();
-    return [remover]()
+    if (snap_remover->to_name == "")
     {
-        if (remover->loop())
+        snap_remover->to_name = snap_remover->from_name;
+    }
+    snap_remover->fsync_interval = cfg["fsync-interval"].uint64_value();
+    if (!snap_remover->fsync_interval)
+        snap_remover->fsync_interval = 128;
+    if (!cfg["cas"].is_null())
+        snap_remover->use_cas = cfg["cas"].uint64_value() ? 2 : 0;
+    if (!cfg["writers_stopped"].is_null())
+        snap_remover->writers_stopped = true;
+    return [snap_remover]()
+    {
+        snap_remover->loop();
+        if (snap_remover->is_done())
         {
-            delete remover;
+            delete snap_remover;
             return true;
         }
         return false;

src/cli_rm_data.cpp

@@ -0,0 +1,214 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "cli.h"
+#include "cluster_client.h"
+
+#define RM_LISTING 1
+#define RM_REMOVING 2
+#define RM_END 3
+
+struct rm_pg_t
+{
+    pg_num_t pg_num;
+    osd_num_t rm_osd_num;
+    std::set<object_id> objects;
+    std::set<object_id>::iterator obj_pos;
+    uint64_t obj_count = 0, obj_done = 0;
+    int state = 0;
+    int in_flight = 0;
+};
+
+struct rm_inode_t
+{
+    uint64_t inode = 0;
+    pool_id_t pool_id = 0;
+    uint64_t min_offset = 0;
+
+    cli_tool_t *parent = NULL;
+    inode_list_t *lister = NULL;
+    std::vector<rm_pg_t*> lists;
+    uint64_t total_count = 0, total_done = 0, total_prev_pct = 0;
+    uint64_t pgs_to_list = 0;
+    bool lists_done = false;
+    int state = 0;
+
+    void start_delete()
+    {
+        lister = parent->cli->list_inode_start(inode, [this](inode_list_t *lst,
+            std::set<object_id>&& objects, pg_num_t pg_num, osd_num_t primary_osd, int status)
+        {
+            rm_pg_t *rm = new rm_pg_t((rm_pg_t){
+                .pg_num = pg_num,
+                .rm_osd_num = primary_osd,
+                .objects = objects,
+                .obj_count = objects.size(),
+                .obj_done = 0,
+            });
+            if (min_offset == 0)
+            {
+                total_count += objects.size();
+            }
+            else
+            {
+                for (object_id oid: objects)
+                {
+                    if (oid.stripe >= min_offset)
+                    {
+                        total_count++;
+                    }
+                }
+            }
+            rm->obj_pos = rm->objects.begin();
+            lists.push_back(rm);
+            if (parent->list_first)
+            {
+                parent->cli->list_inode_next(lister, 1);
+            }
+            if (status & INODE_LIST_DONE)
+            {
+                lists_done = true;
+            }
+            pgs_to_list--;
+            continue_delete();
+        });
+        if (!lister)
+        {
+            fprintf(stderr, "Failed to list inode %lu from pool %u objects\n", INODE_NO_POOL(inode), INODE_POOL(inode));
+            exit(1);
+        }
+        pgs_to_list = parent->cli->list_pg_count(lister);
+        parent->cli->list_inode_next(lister, parent->parallel_osds);
+    }
+
+    void send_ops(rm_pg_t *cur_list)
+    {
+        if (parent->cli->msgr.osd_peer_fds.find(cur_list->rm_osd_num) ==
+            parent->cli->msgr.osd_peer_fds.end())
+        {
+            // Initiate connection
+            parent->cli->msgr.connect_peer(cur_list->rm_osd_num, parent->cli->st_cli.peer_states[cur_list->rm_osd_num]);
+            return;
+        }
+        while (cur_list->in_flight < parent->iodepth && cur_list->obj_pos != cur_list->objects.end())
+        {
+            if (cur_list->obj_pos->stripe >= min_offset)
+            {
+                osd_op_t *op = new osd_op_t();
+                op->op_type = OSD_OP_OUT;
+                // Already checked that it exists above, but anyway
+                op->peer_fd = parent->cli->msgr.osd_peer_fds.at(cur_list->rm_osd_num);
+                op->req = (osd_any_op_t){
+                    .rw = {
+                        .header = {
+                            .magic = SECONDARY_OSD_OP_MAGIC,
+                            .id = parent->cli->next_op_id(),
+                            .opcode = OSD_OP_DELETE,
+                        },
+                        .inode = cur_list->obj_pos->inode,
+                        .offset = cur_list->obj_pos->stripe,
+                        .len = 0,
+                    },
+                };
+                op->callback = [this, cur_list](osd_op_t *op)
+                {
+                    cur_list->in_flight--;
+                    if (op->reply.hdr.retval < 0)
+                    {
+                        fprintf(stderr, "Failed to remove object %lx:%lx from PG %u (OSD %lu) (retval=%ld)\n",
+                            op->req.rw.inode, op->req.rw.offset,
+                            cur_list->pg_num, cur_list->rm_osd_num, op->reply.hdr.retval);
+                    }
+                    delete op;
+                    cur_list->obj_done++;
+                    total_done++;
+                    continue_delete();
+                };
+                cur_list->in_flight++;
+                parent->cli->msgr.outbox_push(op);
+            }
+            cur_list->obj_pos++;
+        }
+    }
+
+    void continue_delete()
+    {
+        if (parent->list_first && !lists_done)
+        {
+            return;
+        }
+        for (int i = 0; i < lists.size(); i++)
+        {
+            if (!lists[i]->in_flight && lists[i]->obj_pos == lists[i]->objects.end())
+            {
+                delete lists[i];
+                lists.erase(lists.begin()+i, lists.begin()+i+1);
+                i--;
+                if (!lists_done)
+                {
+                    parent->cli->list_inode_next(lister, 1);
+                }
+            }
+            else
+            {
+                send_ops(lists[i]);
+            }
+        }
+        if (parent->progress && total_count > 0 && total_done*1000/total_count != total_prev_pct)
+        {
+            printf("\rRemoved %lu/%lu objects, %lu more PGs to list...", total_done, total_count, pgs_to_list);
+            total_prev_pct = total_done*1000/total_count;
+        }
+        if (lists_done && !lists.size())
+        {
+            printf("Done, inode %lu in pool %u data removed\n", INODE_NO_POOL(inode), pool_id);
+            state = 2;
+        }
+    }
+
+    bool loop()
+    {
+        if (state == 0)
+        {
+            start_delete();
+            state = 1;
+        }
+        else if (state == 1)
+        {
+            continue_delete();
+        }
+        else if (state == 2)
+        {
+            return true;
+        }
+        return false;
+    }
+};
+
+std::function<bool(void)> cli_tool_t::start_rm(json11::Json cfg)
+{
+    auto remover = new rm_inode_t();
+    remover->parent = this;
+    remover->inode = cfg["inode"].uint64_value();
+    remover->pool_id = cfg["pool"].uint64_value();
+    if (remover->pool_id)
+    {
+        remover->inode = (remover->inode & ((1l << (64-POOL_ID_BITS)) - 1)) | (((uint64_t)remover->pool_id) << (64-POOL_ID_BITS));
+    }
+    remover->pool_id = INODE_POOL(remover->inode);
+    if (!remover->pool_id)
+    {
+        fprintf(stderr, "pool is missing\n");
+        exit(1);
+    }
+    remover->min_offset = cfg["min-offset"].uint64_value();
+    return [remover]()
+    {
+        if (remover->loop())
+        {
+            delete remover;
+            return true;
+        }
+        return false;
+    };
+}

src/cli_snap_rm.cpp

@@ -1,568 +0,0 @@
-// Copyright (c) Vitaliy Filippov, 2019+
-// License: VNPL-1.1 (see README.md for details)
-
-#include <fcntl.h>
-#include "cli.h"
-#include "cluster_client.h"
-#include "base64.h"
-
-// Remove layer(s): similar to merge, but alters metadata and processes multiple merge targets
-//
-// Exactly one child of the requested layers may be merged using the "inverted" workflow,
-// where we merge it "down" into one of the "to-be-removed" layers and then rename the
-// "to-be-removed" layer to the child. It may be done either if all writers are stopped
-// before trying to delete layers (which is signaled by --writers-stopped) or if that child
-// is a read-only layer (snapshot) itself.
-//
-// This "inverted" workflow trades copying data of one of the deleted layers for copying
-// data of one child of the chain which is also a child of the "traded" layer. So we
-// choose the (parent,child) pair which has the largest difference between "parent" and
-// "child" inode sizes.
-//
-// All other children of the chain are processed by iterating though them, merging removed
-// parents into them and rebasing them to the last layer which isn't a member of the removed
-// chain.
-//
-// Example:
-//
-// <parent> - <from> - <layer 2> - <to> - <child 1>
-//                 \           \       \- <child 2>
-//                  \           \- <child 3>
-//                   \-<child 4>
-//
-// 1) Find optimal pair for the "reverse" scenario
-//    Imagine that it's (<layer 2>, <child 1>) in this example
-// 2) Process all children except <child 1>:
-//    - Merge <from>..<to> to <child 2>
-//    - Set <child 2> parent to <parent>
-//    - Repeat for others
-// 3) Process <child 1>:
-//    - Merge <from>..<child 1> to <layer 2>
-//    - Set <layer 2> parent to <parent>
-//    - Rename <layer 2> to <child 1>
-// 4) Delete other layers of the chain (<from>, <to>)
-struct snap_remover_t
-{
-    cli_tool_t *parent;
-
-    // remove from..to
-    std::string from_name, to_name;
-    // writers are stopped, we can safely change writable layers
-    bool writers_stopped = false;
-    // use CAS writes (0 = never, 1 = auto, 2 = always)
-    int use_cas = 1;
-    // interval between fsyncs
-    int fsync_interval = 128;
-
-    std::map<inode_t,int> sources;
-    std::map<inode_t,uint64_t> inode_used;
-    std::vector<inode_t> merge_children;
-    std::vector<inode_t> chain_list;
-    std::map<inode_t,int> inverse_candidates;
-    inode_t inverse_parent = 0, inverse_child = 0;
-    inode_t new_parent = 0;
-    int state = 0;
-    int current_child = 0;
-    std::function<bool(void)> cb;
-
-    bool is_done()
-    {
-        return state == 9;
-    }
-
-    void loop()
-    {
-        if (state == 1)
-            goto resume_1;
-        else if (state == 2)
-            goto resume_2;
-        else if (state == 3)
-            goto resume_3;
-        else if (state == 4)
-            goto resume_4;
-        else if (state == 5)
-            goto resume_5;
-        else if (state == 6)
-            goto resume_6;
-        else if (state == 7)
-            goto resume_7;
-        else if (state == 8)
-            goto resume_8;
-        else if (state == 9)
-            goto resume_9;
-        // Get children to merge
-        get_merge_children();
-        // Try to select an inode for the "inverse" optimized scenario
-        // Read statistics from etcd to do it
-        read_stats();
-        state = 1;
-resume_1:
-        if (parent->waiting > 0)
-            return;
-        choose_inverse_candidate();
-        // Merge children one by one, except our "inverse" child
-        for (current_child = 0; current_child < merge_children.size(); current_child++)
-        {
-            if (merge_children[current_child] == inverse_child)
-                continue;
-            start_merge_child(merge_children[current_child], merge_children[current_child]);
-resume_2:
-            while (!cb())
-            {
-                state = 2;
-                return;
-            }
-            cb = NULL;
-            parent->change_parent(merge_children[current_child], new_parent);
-            state = 3;
-resume_3:
-            if (parent->waiting > 0)
-                return;
-        }
-        // Merge our "inverse" child into our "inverse" parent
-        if (inverse_child != 0)
-        {
-            start_merge_child(inverse_child, inverse_parent);
-resume_4:
-            while (!cb())
-            {
-                state = 4;
-                return;
-            }
-            cb = NULL;
-            // Delete "inverse" child data
-            start_delete_source(inverse_child);
-resume_5:
-            while (!cb())
-            {
-                state = 5;
-                return;
-            }
-            cb = NULL;
-            // Delete "inverse" child metadata, rename parent over it,
-            // and also change parent links of the previous "inverse" child
-            rename_inverse_parent();
-            state = 6;
-resume_6:
-            if (parent->waiting > 0)
-                return;
-        }
-        // Delete parents, except the "inverse" one
-        for (current_child = 0; current_child < chain_list.size(); current_child++)
-        {
-            if (chain_list[current_child] == inverse_parent)
-                continue;
-            start_delete_source(chain_list[current_child]);
-resume_7:
-            while (!cb())
-            {
-                state = 7;
-                return;
-            }
-            cb = NULL;
-            delete_inode_config(chain_list[current_child]);
-            state = 8;
-resume_8:
-            if (parent->waiting > 0)
-                return;
-        }
-        state = 9;
-resume_9:
-        // Done
-        return;
-    }
-
-    void get_merge_children()
-    {
-        // Get all children of from..to
-        inode_config_t *from_cfg = parent->get_inode_cfg(from_name);
-        inode_config_t *to_cfg = parent->get_inode_cfg(to_name);
-        // Check that to_cfg is actually a child of from_cfg
-        // FIXME de-copypaste the following piece of code with snap_merger_t
-        inode_config_t *cur = to_cfg;
-        chain_list.push_back(cur->num);
-        while (cur->num != from_cfg->num && cur->parent_id != 0)
-        {
-            auto it = parent->cli->st_cli.inode_config.find(cur->parent_id);
-            if (it == parent->cli->st_cli.inode_config.end())
-            {
-                fprintf(stderr, "Parent inode of layer %s (id %ld) not found\n", cur->name.c_str(), cur->parent_id);
-                exit(1);
-            }
-            cur = &it->second;
-            chain_list.push_back(cur->num);
-        }
-        if (cur->num != from_cfg->num)
-        {
-            fprintf(stderr, "Layer %s is not a child of %s\n", to_name.c_str(), from_name.c_str());
-            exit(1);
-        }
-        new_parent = from_cfg->parent_id;
-        // Calculate ranks
-        int i = chain_list.size()-1;
-        for (inode_t item: chain_list)
-        {
-            sources[item] = i--;
-        }
-        for (auto & ic: parent->cli->st_cli.inode_config)
-        {
-            if (!ic.second.parent_id)
-            {
-                continue;
-            }
-            auto it = sources.find(ic.second.parent_id);
-            if (it != sources.end() && sources.find(ic.second.num) == sources.end())
-            {
-                merge_children.push_back(ic.second.num);
-                if (ic.second.readonly || writers_stopped)
-                {
-                    inverse_candidates[ic.second.num] = it->second;
-                }
-            }
-        }
-    }
-
-    void read_stats()
-    {
-        if (inverse_candidates.size() == 0)
-        {
-            return;
-        }
-        json11::Json::array reads;
-        for (auto cp: inverse_candidates)
-        {
-            inode_t inode = cp.first;
-            reads.push_back(json11::Json::object {
-                { "request_range", json11::Json::object {
-                    { "key", base64_encode(
-                        parent->cli->st_cli.etcd_prefix+
-                        "/inode/stats/"+std::to_string(INODE_POOL(inode))+
-                        "/"+std::to_string(INODE_NO_POOL(inode))
-                    ) },
-                } }
-            });
-        }
-        for (auto cp: sources)
-        {
-            inode_t inode = cp.first;
-            reads.push_back(json11::Json::object {
-                { "request_range", json11::Json::object {
-                    { "key", base64_encode(
-                        parent->cli->st_cli.etcd_prefix+
-                        "/inode/stats/"+std::to_string(INODE_POOL(inode))+
-                        "/"+std::to_string(INODE_NO_POOL(inode))
-                    ) },
-                } }
-            });
-        }
-        parent->waiting++;
-        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
-            { "success", reads },
-        }, [this](std::string err, json11::Json data)
-        {
-            parent->waiting--;
-            if (err != "")
-            {
-                fprintf(stderr, "Error reading layer statistics from etcd: %s\n", err.c_str());
-                exit(1);
-            }
-            for (auto inode_result: data["responses"].array_items())
-            {
-                auto kv = parent->cli->st_cli.parse_etcd_kv(inode_result["kvs"][0]);
-                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)
-                {
-                    fprintf(stderr, "Bad key returned from etcd: %s\n", kv.key.c_str());
-                    exit(1);
-                }
-                auto pool_cfg_it = parent->cli->st_cli.pool_config.find(pool_id);
-                if (pool_cfg_it == parent->cli->st_cli.pool_config.end())
-                {
-                    fprintf(stderr, "Pool %u does not exist\n", pool_id);
-                    exit(1);
-                }
-                inode = INODE_WITH_POOL(pool_id, inode);
-                auto & pool_cfg = pool_cfg_it->second;
-                uint64_t used_bytes = kv.value["raw_used"].uint64_value() / pool_cfg.pg_size;
-                if (pool_cfg.scheme != POOL_SCHEME_REPLICATED)
-                {
-                    used_bytes *= (pool_cfg.pg_size - pool_cfg.parity_chunks);
-                }
-                inode_used[inode] = used_bytes;
-            }
-            parent->ringloop->wakeup();
-        });
-    }
-
-    void choose_inverse_candidate()
-    {
-        uint64_t max_diff = 0;
-        for (auto cp: inverse_candidates)
-        {
-            inode_t child = cp.first;
-            uint64_t child_used = inode_used[child];
-            int rank = cp.second;
-            for (int i = chain_list.size()-rank; i < chain_list.size(); i++)
-            {
-                inode_t parent = chain_list[i];
-                uint64_t parent_used = inode_used[parent];
-                if (parent_used > child_used && (!max_diff || max_diff < (parent_used-child_used)))
-                {
-                    max_diff = (parent_used-child_used);
-                    inverse_parent = parent;
-                    inverse_child = child;
-                }
-            }
-        }
-    }
-
-    void rename_inverse_parent()
-    {
-        auto child_it = parent->cli->st_cli.inode_config.find(inverse_child);
-        if (child_it == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode %ld disappeared\n", inverse_child);
-            exit(1);
-        }
-        auto target_it = parent->cli->st_cli.inode_config.find(inverse_parent);
-        if (target_it == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode %ld disappeared\n", inverse_parent);
-            exit(1);
-        }
-        inode_config_t *child_cfg = &child_it->second;
-        inode_config_t *target_cfg = &target_it->second;
-        std::string child_name = child_cfg->name;
-        std::string target_name = target_cfg->name;
-        std::string child_cfg_key = base64_encode(
-            parent->cli->st_cli.etcd_prefix+
-            "/config/inode/"+std::to_string(INODE_POOL(inverse_child))+
-            "/"+std::to_string(INODE_NO_POOL(inverse_child))
-        );
-        std::string target_cfg_key = base64_encode(
-            parent->cli->st_cli.etcd_prefix+
-            "/config/inode/"+std::to_string(INODE_POOL(inverse_parent))+
-            "/"+std::to_string(INODE_NO_POOL(inverse_parent))
-        );
-        // Fill new configuration
-        inode_config_t new_cfg = *child_cfg;
-        new_cfg.num = target_cfg->num;
-        new_cfg.parent_id = new_parent;
-        json11::Json::array cmp = json11::Json::array {
-            json11::Json::object {
-                { "target", "MOD" },
-                { "key", child_cfg_key },
-                { "result", "LESS" },
-                { "mod_revision", child_cfg->mod_revision+1 },
-            },
-            json11::Json::object {
-                { "target", "MOD" },
-                { "key", target_cfg_key },
-                { "result", "LESS" },
-                { "mod_revision", target_cfg->mod_revision+1 },
-            },
-        };
-        json11::Json::array txn = json11::Json::array {
-            json11::Json::object {
-                { "request_delete_range", json11::Json::object {
-                    { "key", child_cfg_key },
-                } },
-            },
-            json11::Json::object {
-                { "request_put", json11::Json::object {
-                    { "key", target_cfg_key },
-                    { "value", base64_encode(json11::Json(parent->cli->st_cli.serialize_inode_cfg(&new_cfg)).dump()) },
-                } },
-            },
-            json11::Json::object {
-                { "request_put", json11::Json::object {
-                    { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/index/image/"+child_cfg->name) },
-                    { "value", base64_encode(json11::Json({
-                        { "id", INODE_NO_POOL(inverse_parent) },
-                        { "pool_id", (uint64_t)INODE_POOL(inverse_parent) },
-                    }).dump()) },
-                } },
-            },
-        };
-        // Reparent children of inverse_child
-        for (auto & cp: parent->cli->st_cli.inode_config)
-        {
-            if (cp.second.parent_id == child_cfg->num)
-            {
-                auto cp_cfg = cp.second;
-                cp_cfg.parent_id = inverse_parent;
-                auto cp_key = base64_encode(
-                    parent->cli->st_cli.etcd_prefix+
-                    "/config/inode/"+std::to_string(INODE_POOL(cp.second.num))+
-                    "/"+std::to_string(INODE_NO_POOL(cp.second.num))
-                );
-                cmp.push_back(json11::Json::object {
-                    { "target", "MOD" },
-                    { "key", cp_key },
-                    { "result", "LESS" },
-                    { "mod_revision", cp.second.mod_revision+1 },
-                });
-                txn.push_back(json11::Json::object {
-                    { "request_put", json11::Json::object {
-                        { "key", cp_key },
-                        { "value", base64_encode(json11::Json(parent->cli->st_cli.serialize_inode_cfg(&cp_cfg)).dump()) },
-                    } },
-                });
-            }
-        }
-        parent->waiting++;
-        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
-            { "compare", cmp },
-            { "success", txn },
-        }, [this, target_name, child_name](std::string err, json11::Json res)
-        {
-            parent->waiting--;
-            if (err != "")
-            {
-                fprintf(stderr, "Error renaming %s to %s: %s\n", target_name.c_str(), child_name.c_str(), err.c_str());
-                exit(1);
-            }
-            if (!res["succeeded"].bool_value())
-            {
-                fprintf(
-                    stderr, "Parent (%s), child (%s), or one of its children"
-                    " configuration was modified during rename\n", target_name.c_str(), child_name.c_str()
-                );
-                exit(1);
-            }
-            printf("Layer %s renamed to %s\n", target_name.c_str(), child_name.c_str());
-            parent->ringloop->wakeup();
-        });
-    }
-
-    void delete_inode_config(inode_t cur)
-    {
-        auto cur_cfg_it = parent->cli->st_cli.inode_config.find(cur);
-        if (cur_cfg_it == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode 0x%lx disappeared\n", cur);
-            exit(1);
-        }
-        inode_config_t *cur_cfg = &cur_cfg_it->second;
-        std::string cur_name = cur_cfg->name;
-        std::string cur_cfg_key = base64_encode(
-            parent->cli->st_cli.etcd_prefix+
-            "/config/inode/"+std::to_string(INODE_POOL(cur))+
-            "/"+std::to_string(INODE_NO_POOL(cur))
-        );
-        parent->waiting++;
-        parent->cli->st_cli.etcd_txn_slow(json11::Json::object {
-            { "compare", json11::Json::array {
-                json11::Json::object {
-                    { "target", "MOD" },
-                    { "key", cur_cfg_key },
-                    { "result", "LESS" },
-                    { "mod_revision", cur_cfg->mod_revision+1 },
-                },
-            } },
-            { "success", json11::Json::array {
-                json11::Json::object {
-                    { "request_delete_range", json11::Json::object {
-                        { "key", cur_cfg_key },
-                    } },
-                },
-                json11::Json::object {
-                    { "request_delete_range", json11::Json::object {
-                        { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/index/image/"+cur_name) },
-                    } },
-                },
-            } },
-        }, [this, cur_name](std::string err, json11::Json res)
-        {
-            parent->waiting--;
-            if (err != "")
-            {
-                fprintf(stderr, "Error deleting %s: %s\n", cur_name.c_str(), err.c_str());
-                exit(1);
-            }
-            if (!res["succeeded"].bool_value())
-            {
-                fprintf(stderr, "Layer %s configuration was modified during deletion\n", cur_name.c_str());
-                exit(1);
-            }
-            printf("Layer %s deleted\n", cur_name.c_str());
-            parent->ringloop->wakeup();
-        });
-    }
-
-    void start_merge_child(inode_t child_inode, inode_t target_inode)
-    {
-        auto child_it = parent->cli->st_cli.inode_config.find(child_inode);
-        if (child_it == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode %ld disappeared\n", child_inode);
-            exit(1);
-        }
-        auto target_it = parent->cli->st_cli.inode_config.find(target_inode);
-        if (target_it == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode %ld disappeared\n", target_inode);
-            exit(1);
-        }
-        cb = parent->start_merge(json11::Json::object {
-            { "command", json11::Json::array{ "merge-data", from_name, child_it->second.name } },
-            { "target", target_it->second.name },
-            { "delete-source", false },
-            { "cas", use_cas },
-            { "fsync-interval", fsync_interval },
-        });
-    }
-
-    void start_delete_source(inode_t inode)
-    {
-        auto source = parent->cli->st_cli.inode_config.find(inode);
-        if (source == parent->cli->st_cli.inode_config.end())
-        {
-            fprintf(stderr, "Inode %ld disappeared\n", inode);
-            exit(1);
-        }
-        cb = parent->start_rm(json11::Json::object {
-            { "inode", inode },
-            { "pool", (uint64_t)INODE_POOL(inode) },
-            { "fsync-interval", fsync_interval },
-        });
-    }
-};
-
-std::function<bool(void)> cli_tool_t::start_snap_rm(json11::Json cfg)
-{
-    json11::Json::array cmd = cfg["command"].array_items();
-    auto snap_remover = new snap_remover_t();
-    snap_remover->parent = this;
-    snap_remover->from_name = cmd.size() > 1 ? cmd[1].string_value() : "";
-    snap_remover->to_name = cmd.size() > 2 ? cmd[2].string_value() : "";
-    if (snap_remover->from_name == "")
-    {
-        fprintf(stderr, "Layer to remove argument is missing\n");
-        exit(1);
-    }
-    if (snap_remover->to_name == "")
-    {
-        snap_remover->to_name = snap_remover->from_name;
-    }
-    snap_remover->fsync_interval = cfg["fsync-interval"].uint64_value();
-    if (!snap_remover->fsync_interval)
-        snap_remover->fsync_interval = 128;
-    if (!cfg["cas"].is_null())
-        snap_remover->use_cas = cfg["cas"].uint64_value() ? 2 : 0;
-    if (!cfg["writers_stopped"].is_null())
-        snap_remover->writers_stopped = true;
-    return [snap_remover]()
-    {
-        snap_remover->loop();
-        if (snap_remover->is_done())
-        {
-            delete snap_remover;
-            return true;
-        }
-        return false;
-    };
-}

src/cli_status.cpp

@@ -0,0 +1,295 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "cli.h"
+#include "cluster_client.h"
+#include "base64.h"
+#include "pg_states.h"
+
+// Print cluster status:
+// etcd, mon, osd states
+// raw/used space, object states, pool states, pg states
+// client io, recovery io, rebalance io
+struct status_printer_t
+{
+    cli_tool_t *parent;
+
+    int state = 0;
+    json11::Json::array mon_members, osd_stats;
+    json11::Json agg_stats;
+    std::map<pool_id_t, json11::Json::object> pool_stats;
+    json11::Json::array etcd_states;
+
+    bool is_done()
+    {
+        return state == 100;
+    }
+
+    void loop()
+    {
+        if (state == 1)
+            goto resume_1;
+        else if (state == 2)
+            goto resume_2;
+        // etcd states
+        {
+            auto addrs = parent->cli->st_cli.get_addresses();
+            etcd_states.resize(addrs.size());
+            for (int i = 0; i < etcd_states.size(); i++)
+            {
+                parent->waiting++;
+                parent->cli->st_cli.etcd_call_oneshot(
+                    addrs[i], "/maintenance/status", json11::Json::object(),
+                    parent->cli->st_cli.etcd_quick_timeout, [this, i](std::string err, json11::Json res)
+                    {
+                        parent->waiting--;
+                        etcd_states[i] = err != "" ? json11::Json::object{ { "error", err } } : res;
+                        parent->ringloop->wakeup();
+                    }
+                );
+            }
+        }
+        state = 1;
+resume_1:
+        if (parent->waiting > 0)
+            return;
+        // Monitors, OSD states
+        parent->etcd_txn(json11::Json::object {
+            { "success", json11::Json::array {
+                json11::Json::object {
+                    { "request_range", json11::Json::object {
+                        { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/mon/") },
+                        { "range_end", base64_encode(parent->cli->st_cli.etcd_prefix+"/mon0") },
+                    } },
+                },
+                json11::Json::object {
+                    { "request_range", json11::Json::object {
+                        { "key", base64_encode(
+                            parent->cli->st_cli.etcd_prefix+"/osd/stats/"
+                        ) },
+                        { "range_end", base64_encode(
+                            parent->cli->st_cli.etcd_prefix+"/osd/stats0"
+                        ) },
+                    } },
+                },
+                json11::Json::object {
+                    { "request_range", json11::Json::object {
+                        { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/stats") },
+                    } },
+                },
+            } },
+        });
+        state = 2;
+resume_2:
+        if (parent->waiting > 0)
+            return;
+        mon_members = parent->etcd_result["responses"][0]["response_range"]["kvs"].array_items();
+        osd_stats = parent->etcd_result["responses"][1]["response_range"]["kvs"].array_items();
+        if (parent->etcd_result["responses"][2]["response_range"]["kvs"].array_items().size() > 0)
+        {
+            agg_stats = parent->cli->st_cli.parse_etcd_kv(parent->etcd_result["responses"][2]["response_range"]["kvs"][0]).value;
+        }
+        int etcd_alive = 0;
+        uint64_t etcd_db_size = 0;
+        std::string etcd_detail;
+        for (int i = 0; i < etcd_states.size(); i++)
+        {
+            if (etcd_states[i]["error"].is_null())
+            {
+                etcd_alive++;
+                etcd_db_size = etcd_states[i]["dbSizeInUse"].uint64_value();
+            }
+        }
+        int mon_count = 0;
+        std::string mon_master;
+        for (int i = 0; i < mon_members.size(); i++)
+        {
+            auto kv = parent->cli->st_cli.parse_etcd_kv(mon_members[i]);
+            kv.key = kv.key.substr(parent->cli->st_cli.etcd_prefix.size());
+            if (kv.key.substr(0, 12) == "/mon/member/")
+                mon_count++;
+            else if (kv.key == "/mon/master")
+            {
+                if (kv.value["hostname"].is_string())
+                    mon_master = kv.value["hostname"].string_value();
+                else
+                    mon_master = kv.value["ip"][0].string_value();
+            }
+        }
+        int osd_count = 0, osd_up = 0;
+        uint64_t total_raw = 0, free_raw = 0, free_down_raw = 0, down_raw = 0;
+        for (int i = 0; i < osd_stats.size(); i++)
+        {
+            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)
+            {
+                fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
+                continue;
+            }
+            osd_count++;
+            total_raw += kv.value["size"].uint64_value();
+            free_raw += kv.value["free"].uint64_value();
+            auto peer_it = parent->cli->st_cli.peer_states.find(stat_osd_num);
+            if (peer_it != parent->cli->st_cli.peer_states.end())
+            {
+                osd_up++;
+            }
+            else
+            {
+                down_raw += kv.value["size"].uint64_value();
+                free_down_raw += kv.value["size"].uint64_value();
+            }
+        }
+        int pool_count = 0, pools_active = 0;
+        std::map<std::string, int> pgs_by_state;
+        std::string pgs_by_state_str;
+        for (auto & pool_pair: parent->cli->st_cli.pool_config)
+        {
+            auto & pool_cfg = pool_pair.second;
+            bool active = true;
+            if (pool_cfg.pg_config.size() != pool_cfg.pg_count)
+            {
+                active = false;
+                pgs_by_state["offline"] += pool_cfg.pg_count-pool_cfg.pg_config.size();
+            }
+            pool_count++;
+            for (auto pg_it = pool_cfg.pg_config.begin(); pg_it != pool_cfg.pg_config.end(); pg_it++)
+            {
+                if (!(pg_it->second.cur_state & PG_ACTIVE))
+                {
+                    active = false;
+                }
+                std::string pg_state_str;
+                for (int i = 0; i < pg_state_bit_count; i++)
+                {
+                    if (pg_it->second.cur_state & pg_state_bits[i])
+                    {
+                        pg_state_str += "+";
+                        pg_state_str += pg_state_names[i];
+                    }
+                }
+                if (pg_state_str.size())
+                    pgs_by_state[pg_state_str.substr(1)]++;
+                else
+                    pgs_by_state["offline"]++;
+            }
+            if (active)
+            {
+                pools_active++;
+            }
+        }
+        for (auto & kv: pgs_by_state)
+        {
+            if (pgs_by_state_str.size())
+            {
+                pgs_by_state_str += "\n           ";
+            }
+            pgs_by_state_str += std::to_string(kv.second)+" "+kv.first;
+        }
+        uint64_t object_size = parent->cli->get_bs_block_size();
+        std::string more_states;
+        uint64_t obj_n;
+        obj_n = agg_stats["object_counts"]["misplaced"].uint64_value();
+        if (obj_n > 0)
+            more_states += ", "+format_size(obj_n*object_size)+" misplaced";
+        obj_n = agg_stats["object_counts"]["degraded"].uint64_value();
+        if (obj_n > 0)
+            more_states += ", "+format_size(obj_n*object_size)+" degraded";
+        obj_n = agg_stats["object_counts"]["incomplete"].uint64_value();
+        if (obj_n > 0)
+            more_states += ", "+format_size(obj_n*object_size)+" incomplete";
+        std::string recovery_io;
+        {
+            uint64_t deg_bps = agg_stats["recovery_stats"]["degraded"]["bps"].uint64_value();
+            uint64_t deg_iops = agg_stats["recovery_stats"]["degraded"]["iops"].uint64_value();
+            uint64_t misp_bps = agg_stats["recovery_stats"]["misplaced"]["bps"].uint64_value();
+            uint64_t misp_iops = agg_stats["recovery_stats"]["misplaced"]["iops"].uint64_value();
+            if (deg_iops > 0 || deg_bps > 0)
+                recovery_io += "    recovery:  "+format_size(deg_bps)+"/s, "+format_size(deg_iops, true)+" op/s\n";
+            if (misp_iops > 0 || misp_bps > 0)
+                recovery_io += "    rebalance: "+format_size(misp_bps)+"/s, "+format_size(misp_iops, true)+" op/s\n";
+        }
+        if (parent->json_output)
+        {
+            // JSON output
+            printf("%s\n", json11::Json(json11::Json::object {
+                { "etcd_alive", etcd_alive },
+                { "etcd_count", etcd_states.size() },
+                { "etcd_db_size", etcd_db_size },
+                { "mon_count", mon_count },
+                { "mon_master", mon_master },
+                { "osd_up", osd_up },
+                { "osd_count", osd_count },
+                { "total_raw", total_raw },
+                { "free_raw", free_raw },
+                { "down_raw", down_raw },
+                { "free_down_raw", free_down_raw },
+                { "clean_data", agg_stats["object_counts"]["clean"].uint64_value() * object_size },
+                { "misplaced_data", agg_stats["object_counts"]["misplaced"].uint64_value() * object_size },
+                { "degraded_data", agg_stats["object_counts"]["degraded"].uint64_value() * object_size },
+                { "incomplete_data", agg_stats["object_counts"]["incomplete"].uint64_value() * object_size },
+                { "pool_count", pool_count },
+                { "active_pool_count", pools_active },
+                { "pg_states", pgs_by_state },
+                { "op_stats", agg_stats["op_stats"] },
+                { "recovery_stats", agg_stats["recovery_stats"] },
+                { "object_counts", agg_stats["object_counts"] },
+            }).dump().c_str());
+            state = 100;
+            return;
+        }
+        printf(
+            "  cluster:\n"
+            "    etcd: %d / %ld up, %s database size\n"
+            "    mon:  %d up%s\n"
+            "    osd:  %d / %d up\n"
+            "  \n"
+            "  data:\n"
+            "    raw:   %s used, %s / %s available%s\n"
+            "    state: %s clean%s\n"
+            "    pools: %d / %d active\n"
+            "    pgs:   %s\n"
+            "  \n"
+            "  io:\n"
+            "    client:%s %s/s rd, %s op/s rd, %s/s wr, %s op/s wr\n"
+            "%s",
+            etcd_alive, etcd_states.size(), format_size(etcd_db_size).c_str(),
+            mon_count, mon_master == "" ? "" : (", master "+mon_master).c_str(),
+            osd_up, osd_count,
+            format_size(total_raw-free_raw).c_str(),
+            format_size(free_raw-free_down_raw).c_str(),
+            format_size(total_raw-down_raw).c_str(),
+            (down_raw > 0 ? (", "+format_size(down_raw)+" down").c_str() : ""),
+            format_size(agg_stats["object_counts"]["clean"].uint64_value() * object_size).c_str(), more_states.c_str(),
+            pools_active, pool_count,
+            pgs_by_state_str.c_str(),
+            recovery_io.size() > 0 ? "   " : "",
+            format_size(agg_stats["op_stats"]["primary_read"]["bps"].uint64_value()).c_str(),
+            format_size(agg_stats["op_stats"]["primary_read"]["iops"].uint64_value(), true).c_str(),
+            format_size(agg_stats["op_stats"]["primary_write"]["bps"].uint64_value()).c_str(),
+            format_size(agg_stats["op_stats"]["primary_write"]["iops"].uint64_value(), true).c_str(),
+            recovery_io.c_str()
+        );
+        state = 100;
+    }
+};
+
+std::function<bool(void)> cli_tool_t::start_status(json11::Json cfg)
+{
+    json11::Json::array cmd = cfg["command"].array_items();
+    auto printer = new status_printer_t();
+    printer->parent = this;
+    return [printer]()
+    {
+        printer->loop();
+        if (printer->is_done())
+        {
+            delete printer;
+            return true;
+        }
+        return false;
+    };
+}

src/cluster_client.cpp

@@ -143,7 +143,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
     }
     else /* if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP) */
     {
-        for (auto prev = op->prev; prev; prev = prev->prev)
+        for (auto prev = op_queue_head; prev && prev != op; prev = prev->next)
         {
             if (prev->opcode == OSD_OP_WRITE && prev->flags & OP_FLUSH_BUFFER)
             {
@@ -151,7 +151,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
             }
             else if (prev->opcode == OSD_OP_WRITE || prev->opcode == OSD_OP_READ || prev->opcode == OSD_OP_READ_BITMAP)
             {
-                // Flushes are always in the beginning
+                // Flushes are always in the beginning (we're scanning from the beginning of the queue)
                 break;
             }
         }
@@ -172,6 +172,7 @@ void cluster_client_t::inc_wait(uint64_t opcode, uint64_t flags, cluster_op_t *n
                 (next->opcode == OSD_OP_READ || next->opcode == OSD_OP_READ_BITMAP) && (flags & OP_FLUSH_BUFFER))
             {
                 next->prev_wait += inc;
+                assert(next->prev_wait >= 0);
                 if (!next->prev_wait)
                 {
                     if (next->opcode == OSD_OP_SYNC)
@@ -191,6 +192,7 @@ void cluster_client_t::inc_wait(uint64_t opcode, uint64_t flags, cluster_op_t *n
             if (next->opcode == OSD_OP_SYNC || next->opcode == OSD_OP_WRITE)
             {
                 next->prev_wait += inc;
+                assert(next->prev_wait >= 0);
                 if (!next->prev_wait)
                 {
                     if (next->opcode == OSD_OP_SYNC)

src/cluster_client_list.cpp

@@ -200,7 +200,8 @@ void cluster_client_t::send_list(inode_list_osd_t *cur_list)
     auto & pool_cfg = st_cli.pool_config[cur_list->pg->lst->pool_id];
     osd_op_t *op = new osd_op_t();
     op->op_type = OSD_OP_OUT;
-    op->peer_fd = msgr.osd_peer_fds[cur_list->osd_num];
+    // Already checked that it exists above, but anyway
+    op->peer_fd = msgr.osd_peer_fds.at(cur_list->osd_num);
     op->req = (osd_any_op_t){
         .sec_list = {
             .header = {

src/epoll_manager.cpp

@@ -13,6 +13,7 @@
 epoll_manager_t::epoll_manager_t(ring_loop_t *ringloop)
 {
     this->ringloop = ringloop;
+    this->pending = false;
 
     epoll_fd = epoll_create(1);
     if (epoll_fd < 0)
@@ -22,11 +23,19 @@ epoll_manager_t::epoll_manager_t(ring_loop_t *ringloop)
 
     tfd = new timerfd_manager_t([this](int fd, bool wr, std::function<void(int, int)> handler) { set_fd_handler(fd, wr, handler); });
 
+    consumer.loop = [this]()
+    {
+        if (pending)
+            handle_epoll_events();
+    };
+    ringloop->register_consumer(&consumer);
+
     handle_epoll_events();
 }
 
 epoll_manager_t::~epoll_manager_t()
 {
+    ringloop->unregister_consumer(&consumer);
     if (tfd)
     {
         delete tfd;
@@ -64,8 +73,13 @@ void epoll_manager_t::handle_epoll_events()
     io_uring_sqe *sqe = ringloop->get_sqe();
     if (!sqe)
     {
-        throw std::runtime_error("can't get SQE, will fall out of sync with EPOLLET");
+        // Don't handle epoll events until we manage to post the next event handler
+        // otherwise we'll fall out of sync with EPOLLET
+        pending = true;
+        ringloop->wakeup();
+        return;
     }
+    pending = false;
     ring_data_t *data = ((ring_data_t*)sqe->user_data);
     my_uring_prep_poll_add(sqe, epoll_fd, POLLIN);
     data->callback = [this](ring_data_t *data)

src/epoll_manager.h

@@ -11,6 +11,8 @@
 class epoll_manager_t
 {
     int epoll_fd;
+    bool pending;
+    ring_consumer_t consumer;
     ring_loop_t *ringloop;
     std::map<int, std::function<void(int, int)>> epoll_handlers;
 public:

src/etcd_state_client.cpp

@@ -64,6 +64,42 @@ void etcd_state_client_t::etcd_txn_slow(json11::Json txn, std::function<void(std
     etcd_call("/kv/txn", txn, etcd_slow_timeout, max_etcd_attempts, 0, callback);
 }
 
+std::vector<std::string> etcd_state_client_t::get_addresses()
+{
+    auto addrs = etcd_local;
+    addrs.insert(addrs.end(), etcd_addresses.begin(), etcd_addresses.end());
+    return addrs;
+}
+
+void etcd_state_client_t::etcd_call_oneshot(std::string etcd_address, std::string api, json11::Json payload,
+    int timeout, std::function<void(std::string, json11::Json)> callback)
+{
+    std::string etcd_api_path;
+    int pos = etcd_address.find('/');
+    if (pos >= 0)
+    {
+        etcd_api_path = etcd_address.substr(pos);
+        etcd_address = etcd_address.substr(0, pos);
+    }
+    std::string req = payload.dump();
+    req = "POST "+etcd_api_path+api+" HTTP/1.1\r\n"
+        "Host: "+etcd_address+"\r\n"
+        "Content-Type: application/json\r\n"
+        "Content-Length: "+std::to_string(req.size())+"\r\n"
+        "Connection: close\r\n"
+        "\r\n"+req;
+    auto http_cli = http_init(tfd);
+    auto cb = [this, http_cli, callback](const http_response_t *response)
+    {
+        std::string err;
+        json11::Json data;
+        response->parse_json_response(err, data);
+        callback(err, data);
+        http_close(http_cli);
+    };
+    http_request(http_cli, etcd_address, req, { .timeout = timeout }, cb);
+}
+
 void etcd_state_client_t::etcd_call(std::string api, json11::Json payload, int timeout,
     int retries, int interval, std::function<void(std::string, json11::Json)> callback)
 {

src/etcd_state_client.h

@@ -112,6 +112,8 @@ public:
 
     json11::Json::object serialize_inode_cfg(inode_config_t *cfg);
     etcd_kv_t parse_etcd_kv(const json11::Json & kv_json);
+    std::vector<std::string> get_addresses();
+    void etcd_call_oneshot(std::string etcd_address, std::string api, json11::Json payload, int timeout, std::function<void(std::string, json11::Json)> callback);
     void etcd_call(std::string api, json11::Json payload, int timeout, int retries, int interval, std::function<void(std::string, json11::Json)> callback);
     void etcd_txn(json11::Json txn, int timeout, int retries, int interval, std::function<void(std::string, json11::Json)> callback);
     void etcd_txn_slow(json11::Json txn, std::function<void(std::string, json11::Json)> callback);

src/fio_cluster.cpp

@@ -351,9 +351,9 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
         }
         else
         {
-            printf("+++ %s 0x%lx 0x%llx+%llx\n",
+            printf("+++ %s 0x%lx 0x%llx+%lx\n",
                 io->ddir == DDIR_READ ? "READ" : "WRITE",
-                (uint64_t)io, io->offset, io->xfer_buflen);
+                (uint64_t)io, io->offset, (uint64_t)io->xfer_buflen);
         }
     }
 

src/fio_headers.h

@@ -1,4 +1,3 @@
-extern "C" {
 // Kill atomics in fio headers
 #define _STDATOMIC_H
 #include "fio/arch/arch.h"
@@ -11,6 +10,7 @@ extern "C" {
 #define CONFIG_HAVE_GETTID
 #define CONFIG_SYNC_FILE_RANGE
 #define CONFIG_PWRITEV2
+extern "C" {
 #include "fio/fio.h"
 #include "fio/optgroup.h"
 }

src/fio_sec_osd.cpp

@@ -170,14 +170,14 @@ static int sec_init(struct thread_data *td)
     bsd->block_order = o->block_order == 0 ? 17 : o->block_order;
     bsd->block_size = 1 << o->block_order;
 
-    sockaddr addr;
+    sockaddr_storage addr;
     if (!string_to_addr(std::string(o->host ? o->host : "127.0.0.1"), false, o->port > 0 ? o->port : 11203, &addr))
     {
         fprintf(stderr, "server address: %s is not valid\n", o->host ? o->host : "127.0.0.1");
         return 1;
     }
 
-    bsd->connect_fd = socket(addr.sa_family, SOCK_STREAM, 0);
+    bsd->connect_fd = socket(addr.ss_family, SOCK_STREAM, 0);
     if (bsd->connect_fd < 0)
     {
         perror("socket");
@@ -192,11 +192,16 @@ static int sec_init(struct thread_data *td)
     setsockopt(bsd->connect_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
     if (o->zerocopy_send)
     {
+#ifndef SO_ZEROCOPY
+        perror("zerocopy send not supported on your system (socket.h misses SO_ZEROCOPY)");
+        return 1;
+#else
         if (setsockopt(bsd->connect_fd, SOL_SOCKET, SO_ZEROCOPY, &one, sizeof(one)) < 0)
         {
             perror("setsockopt zerocopy");
             return 1;
         }
+#endif
     }
 
     // FIXME: read config (block size) from OSD
@@ -306,7 +311,13 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
         iov[iovcnt++] = { .iov_base = io->xfer_buf, .iov_len = io->xfer_buflen };
         wtotal += io->xfer_buflen;
     }
-    if (sendv_blocking(bsd->connect_fd, iov, iovcnt, opt->zerocopy_send ? MSG_ZEROCOPY : 0) != wtotal)
+    if (sendv_blocking(bsd->connect_fd, iov, iovcnt,
+#ifdef SO_ZEROCOPY
+        opt->zerocopy_send ? MSG_ZEROCOPY : 0
+#else
+        0
+#endif
+    ) != wtotal)
     {
         perror("sendmsg");
         exit(1);
@@ -344,7 +355,7 @@ static int sec_getevents(struct thread_data *td, unsigned int min, unsigned int
         {
             if (reply.hdr.retval != io->xfer_buflen)
             {
-                fprintf(stderr, "Short read: retval = %ld instead of %llu\n", reply.hdr.retval, io->xfer_buflen);
+                fprintf(stderr, "Short read: retval = %ld instead of %lu\n", reply.hdr.retval, (uint64_t)io->xfer_buflen);
                 exit(1);
             }
             // Support bitmap
@@ -369,7 +380,7 @@ static int sec_getevents(struct thread_data *td, unsigned int min, unsigned int
         {
             if (reply.hdr.retval != io->xfer_buflen)
             {
-                fprintf(stderr, "Short write: retval = %ld instead of %llu\n", reply.hdr.retval, io->xfer_buflen);
+                fprintf(stderr, "Short write: retval = %ld instead of %lu\n", reply.hdr.retval, (uint64_t)io->xfer_buflen);
                 exit(1);
             }
         }

src/http_client.cpp

@@ -62,9 +62,10 @@ struct http_co_t
     void run_cb_and_clear();
     void start_connection();
     void close_connection();
+    void next_request();
     void handle_events();
     void handle_connect_result();
-    void submit_read();
+    void submit_read(bool check_timeout);
     void submit_send();
     bool handle_read();
     void post_message(int type, const std::string & msg);
@@ -128,6 +129,7 @@ void http_co_t::run_cb_and_clear()
     // Call callback after clearing it because otherwise we may hit reenterability problems
     if (cb != NULL)
         cb(&parsed);
+    next_request();
 }
 
 void http_co_t::send_request(const std::string & host, const std::string & request,
@@ -161,17 +163,6 @@ void http_co_t::send_request(const std::string & host, const std::string & reque
     this->sent = 0;
     this->response_callback = response_callback;
     this->parsed = {};
-    if (request_timeout > 0)
-    {
-        timeout_id = tfd->set_timer(request_timeout, false, [this](int timer_id)
-        {
-            stackin();
-            close_connection();
-            parsed = { .error = "HTTP request timed out" };
-            run_cb_and_clear();
-            stackout();
-        });
-    }
     if (state == HTTP_CO_KEEPALIVE)
     {
         state = HTTP_CO_SENDING_REQUEST;
@@ -181,6 +172,28 @@ void http_co_t::send_request(const std::string & host, const std::string & reque
     {
         start_connection();
     }
+    // Do it _after_ state assignment because set_timer() can actually trigger
+    // other timers and requests (reenterability is our friend)
+    if (request_timeout > 0)
+    {
+        timeout_id = tfd->set_timer(request_timeout, false, [this](int timer_id)
+        {
+            stackin();
+            if (state == HTTP_CO_REQUEST_SENT)
+            {
+                // In case of high CPU load, we may not handle etcd responses in time
+                // For this case, first check the socket and only then terminate request with the timeout
+                submit_read(true);
+            }
+            else
+            {
+                close_connection();
+                parsed = { .error = "HTTP request timed out" };
+                run_cb_and_clear();
+            }
+            stackout();
+        });
+    }
     stackout();
 }
 
@@ -271,17 +284,19 @@ void http_co_t::close_connection()
 void http_co_t::start_connection()
 {
     stackin();
-    struct sockaddr addr;
+    struct sockaddr_storage addr;
     if (!string_to_addr(host.c_str(), 1, 80, &addr))
     {
+        close_connection();
         parsed = { .error = "Invalid address: "+host };
         run_cb_and_clear();
         stackout();
         return;
     }
-    peer_fd = socket(addr.sa_family, SOCK_STREAM, 0);
+    peer_fd = socket(addr.ss_family, SOCK_STREAM, 0);
     if (peer_fd < 0)
     {
+        close_connection();
         parsed = { .error = std::string("socket: ")+strerror(errno) };
         run_cb_and_clear();
         stackout();
@@ -323,10 +338,12 @@ void http_co_t::handle_events()
             epoll_events &= ~EPOLLOUT;
             if (epoll_events & EPOLLIN)
             {
-                submit_read();
+                submit_read(false);
             }
             else if (epoll_events & (EPOLLRDHUP|EPOLLERR))
             {
+                if (state == HTTP_CO_HEADERS_RECEIVED)
+                    std::swap(parsed.body, response);
                 close_connection();
                 run_cb_and_clear();
                 break;
@@ -410,10 +427,11 @@ again:
     stackout();
 }
 
-void http_co_t::submit_read()
+void http_co_t::submit_read(bool check_timeout)
 {
     stackin();
     int res;
+again:
     if (rbuf.size() != READ_BUFFER_SIZE)
     {
         rbuf.resize(READ_BUFFER_SIZE);
@@ -428,12 +446,29 @@ void http_co_t::submit_read()
     }
     if (res == -EAGAIN || res == -EINTR)
     {
-        epoll_events = epoll_events & ~EPOLLIN;
+        if (check_timeout)
+        {
+            if (res == -EINTR)
+                goto again;
+            else
+            {
+                // Timeout happened and there is no data to read
+                close_connection();
+                parsed = { .error = "HTTP request timed out" };
+                run_cb_and_clear();
+            }
+        }
+        else
+        {
+            epoll_events = epoll_events & ~EPOLLIN;
+        }
     }
     else if (res <= 0)
     {
         // < 0 means error, 0 means EOF
         epoll_events = epoll_events & ~EPOLLIN;
+        if (state == HTTP_CO_HEADERS_RECEIVED)
+            std::swap(parsed.body, response);
         close_connection();
         if (res < 0)
             parsed = { .error = std::string("recvmsg: ")+strerror(-res) };
@@ -501,8 +536,11 @@ bool http_co_t::handle_read()
     if (state == HTTP_CO_HEADERS_RECEIVED && target_response_size > 0 && response.size() >= target_response_size)
     {
         std::swap(parsed.body, response);
-        response_callback(&parsed);
-        parsed.eof = true;
+        if (!keepalive)
+            close_connection();
+        else
+            state = HTTP_CO_KEEPALIVE;
+        run_cb_and_clear();
     }
     else if (state == HTTP_CO_CHUNKED && response.size() > 0)
     {
@@ -533,10 +571,14 @@ bool http_co_t::handle_read()
             response_callback(&parsed);
             parsed.body = "";
         }
-        if (parsed.eof && !want_streaming)
+        else if (parsed.eof)
         {
             // Normal response
-            response_callback(&parsed);
+            if (!keepalive)
+                close_connection();
+            else
+                state = HTTP_CO_KEEPALIVE;
+            run_cb_and_clear();
         }
     }
     else if (state == HTTP_CO_WEBSOCKET && response.size() > 0)
@@ -547,29 +589,20 @@ bool http_co_t::handle_read()
             parsed.body = "";
         }
     }
-    if (parsed.eof)
-    {
-        response_callback = NULL;
-        parsed = {};
-        if (!keepalive)
-        {
-            close_connection();
-        }
-        else
-        {
-            state = HTTP_CO_KEEPALIVE;
-            if (keepalive_queue.size() > 0)
-            {
-                auto next = keepalive_queue[0];
-                keepalive_queue.erase(keepalive_queue.begin(), keepalive_queue.begin()+1);
-                next();
-            }
-        }
-    }
     stackout();
     return true;
 }
 
+void http_co_t::next_request()
+{
+    if (keepalive_queue.size() > 0)
+    {
+        auto next = keepalive_queue[0];
+        keepalive_queue.erase(keepalive_queue.begin(), keepalive_queue.begin()+1);
+        next();
+    }
+}
+
 uint64_t stoull_full(const std::string & str, int base)
 {
     if (isspace(str[0]))

src/messenger.cpp

@@ -222,13 +222,13 @@ void osd_messenger_t::try_connect_peer(uint64_t peer_osd)
 void osd_messenger_t::try_connect_peer_addr(osd_num_t peer_osd, const char *peer_host, int peer_port)
 {
     assert(peer_osd != this->osd_num);
-    struct sockaddr addr;
+    struct sockaddr_storage addr;
     if (!string_to_addr(peer_host, 0, peer_port, &addr))
     {
         on_connect_peer(peer_osd, -EINVAL);
         return;
     }
-    int peer_fd = socket(addr.sa_family, SOCK_STREAM, 0);
+    int peer_fd = socket(addr.ss_family, SOCK_STREAM, 0);
     if (peer_fd < 0)
     {
         on_connect_peer(peer_osd, -errno);
@@ -484,10 +484,10 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
 void osd_messenger_t::accept_connections(int listen_fd)
 {
     // Accept new connections
-    sockaddr addr;
+    sockaddr_storage addr;
     socklen_t peer_addr_size = sizeof(addr);
     int peer_fd;
-    while ((peer_fd = accept(listen_fd, &addr, &peer_addr_size)) >= 0)
+    while ((peer_fd = accept(listen_fd, (sockaddr*)&addr, &peer_addr_size)) >= 0)
     {
         assert(peer_fd != 0);
         fprintf(stderr, "[OSD %lu] new client %d: connection from %s\n", this->osd_num, peer_fd,

src/messenger.h

@@ -49,7 +49,7 @@ struct osd_client_t
 {
     int refs = 0;
 
-    sockaddr peer_addr;
+    sockaddr_storage peer_addr;
     int peer_port;
     int peer_fd;
     int peer_state;

src/msgr_stop.cpp

@@ -111,6 +111,10 @@ void osd_messenger_t::stop_client(int peer_fd, bool force, bool force_delete)
         {
             delete cl->read_op;
         }
+        else
+        {
+            cancel_op(cl->read_op);
+        }
         cl->read_op = NULL;
     }
     if (cl->osd_num)

src/nbd_proxy.cpp

@@ -54,7 +54,18 @@ protected:
     msghdr read_msg = { 0 }, send_msg = { 0 };
     iovec read_iov = { 0 };
 
+    std::string logfile = "/dev/null";
+
 public:
+    ~nbd_proxy()
+    {
+        if (recv_buf)
+        {
+            free(recv_buf);
+            recv_buf = NULL;
+        }
+    }
+
     static json11::Json::object parse_args(int narg, const char *args[])
     {
         json11::Json::object cfg;
@@ -269,6 +280,10 @@ public:
                 }
             }
         }
+        if (cfg["logfile"].is_string())
+        {
+            logfile = cfg["logfile"].string_value();
+        }
         if (bg)
         {
             daemonize();
@@ -322,6 +337,9 @@ public:
         delete cli;
         delete epmgr;
         delete ringloop;
+        cli = NULL;
+        epmgr = NULL;
+        ringloop = NULL;
     }
 
     void load_module()
@@ -351,13 +369,14 @@ public:
         setsid();
         if (fork())
             exit(0);
-        chdir("/");
         close(0);
         close(1);
         close(2);
         open("/dev/null", O_RDONLY);
-        open("/dev/null", O_WRONLY);
-        open("/dev/null", O_WRONLY);
+        open(logfile.c_str(), O_WRONLY|O_APPEND|O_CREAT, 0666);
+        open(logfile.c_str(), O_WRONLY|O_APPEND|O_CREAT, 0666);
+        if (chdir("/") != 0)
+            fprintf(stderr, "Warning: Failed to chdir into /\n");
     }
 
     json11::Json::object list_mapped()
@@ -498,7 +517,7 @@ protected:
             goto end_unmap;
         }
         ioctl(nbd, NBD_SET_FLAGS, flags);
-        if (timeout >= 0)
+        if (timeout > 0)
         {
             r = ioctl(nbd, NBD_SET_TIMEOUT, (unsigned long)timeout);
             if (r < 0)
@@ -513,7 +532,11 @@ protected:
         {
             goto end_unmap;
         }
-        write(qd_fd, "32768", 5);
+        r = write(qd_fd, "32768", 5);
+        if (r != 5)
+        {
+            fprintf(stderr, "Warning: Failed to configure max_sectors_kb\n");
+        }
         close(qd_fd);
         if (!fork())
         {
@@ -667,6 +690,7 @@ protected:
             {
                 assert(result <= cur_left);
                 cur_left -= result;
+                cur_buf = (uint8_t*)cur_buf + result;
                 result = 0;
             }
             if (cur_left <= 0)
@@ -681,6 +705,12 @@ protected:
         if (read_state == CL_READ_HDR)
         {
             int req_type = be32toh(cur_req.type);
+            if (be32toh(cur_req.magic) == NBD_REQUEST_MAGIC && req_type == NBD_CMD_DISC)
+            {
+                // Disconnect
+                close(nbd_fd);
+                exit(0);
+            }
             if (be32toh(cur_req.magic) != NBD_REQUEST_MAGIC ||
                 req_type != NBD_CMD_READ && req_type != NBD_CMD_WRITE && req_type != NBD_CMD_FLUSH)
             {

src/osd.cpp

@@ -3,6 +3,7 @@
 
 #include <sys/socket.h>
 #include <sys/poll.h>
+#include <sys/mman.h>
 #include <netinet/in.h>
 #include <netinet/tcp.h>
 #include <arpa/inet.h>
@@ -53,6 +54,20 @@ osd_t::osd_t(const json11::Json & config, ring_loop_t *ringloop)
             autosync_writes = max_autosync;
     }
 
+    if (this->config["osd_memlock"] == "true" || this->config["osd_memlock"] == "1" || this->config["osd_memlock"] == "yes")
+    {
+        // Lock all OSD memory if requested
+        if (mlockall(MCL_CURRENT|MCL_FUTURE
+#ifdef MCL_ONFAULT
+            | MCL_ONFAULT
+#endif
+            ) != 0)
+        {
+            fprintf(stderr, "osd_memlock is set to true, but mlockall() failed: %s\n", strerror(errno));
+            exit(-1);
+        }
+    }
+
     this->tfd->set_timer(print_stats_interval*1000, true, [this](int timer_id)
     {
         print_stats();
@@ -185,46 +200,7 @@ void osd_t::bind_socket()
 
     // FIXME Support multiple listening sockets
 
-    sockaddr addr;
-    if (!string_to_addr(bind_address, 0, bind_port, &addr))
-    {
-        throw std::runtime_error("bind address "+bind_address+" is not valid");
-    }
-
-    listen_fd = socket(addr.sa_family, SOCK_STREAM, 0);
-    if (listen_fd < 0)
-    {
-        throw std::runtime_error(std::string("socket: ") + strerror(errno));
-    }
-    int enable = 1;
-    setsockopt(listen_fd, SOL_SOCKET, SO_REUSEADDR, &enable, sizeof(enable));
-
-    if (bind(listen_fd, &addr, sizeof(addr)) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("bind: ") + strerror(errno));
-    }
-    if (bind_port == 0)
-    {
-        socklen_t len = sizeof(addr);
-        if (getsockname(listen_fd, (sockaddr *)&addr, &len) == -1)
-        {
-            close(listen_fd);
-            throw std::runtime_error(std::string("getsockname: ") + strerror(errno));
-        }
-        listening_port = ntohs(((sockaddr_in*)&addr)->sin_port);
-    }
-    else
-    {
-        listening_port = bind_port;
-    }
-
-    if (listen(listen_fd, listen_backlog) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("listen: ") + strerror(errno));
-    }
-
+    listen_fd = create_and_bind_socket(bind_address, bind_port, listen_backlog, &listening_port);
     fcntl(listen_fd, F_SETFL, fcntl(listen_fd, F_GETFL, 0) | O_NONBLOCK);
 
     epmgr->set_fd_handler(listen_fd, false, [this](int fd, int events)

src/osd.h

@@ -211,7 +211,7 @@ class osd_t
     // flushing, recovery and backfill
     void submit_pg_flush_ops(pg_t & pg);
     void handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t *fb, osd_num_t peer_osd, int retval);
-    void submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t *fb, bool rollback, osd_num_t peer_osd, int count, obj_ver_id *data);
+    bool submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t *fb, bool rollback, osd_num_t peer_osd, int count, obj_ver_id *data);
     bool pick_next_recovery(osd_recovery_op_t &op);
     void submit_recovery_op(osd_recovery_op_t *op);
     bool continue_recovery();

src/osd_cluster.cpp

@@ -457,7 +457,8 @@ void osd_t::renew_lease()
         if (err == "" && data["result"]["TTL"].string_value() == "")
         {
             // Die
-            throw std::runtime_error("etcd lease has expired");
+            fprintf(stderr, "Error refreshing etcd lease\n");
+            force_stop(1);
         }
         if (err != "")
         {
@@ -466,7 +467,8 @@ void osd_t::renew_lease()
             if (etcd_failed_attempts > st_cli.max_etcd_attempts)
             {
                 // Die
-                throw std::runtime_error("Cluster connection failed");
+                fprintf(stderr, "Cluster connection failed\n");
+                force_stop(1);
             }
             // Retry
             tfd->set_timer(st_cli.etcd_quick_timeout, false, [this](int timer_id)

src/osd_flush.cpp

@@ -47,7 +47,8 @@ void osd_t::submit_pg_flush_ops(pg_t & pg)
         if (l.second.size() > 0)
         {
             fb->flush_ops++;
-            submit_flush_op(pg.pool_id, pg.pg_num, fb, true, l.first, l.second.size(), l.second.data());
+            if (!submit_flush_op(pg.pool_id, pg.pg_num, fb, true, l.first, l.second.size(), l.second.data()))
+                return;
         }
     }
     for (auto & l: fb->stable_lists)
@@ -55,7 +56,8 @@ void osd_t::submit_pg_flush_ops(pg_t & pg)
         if (l.second.size() > 0)
         {
             fb->flush_ops++;
-            submit_flush_op(pg.pool_id, pg.pg_num, fb, false, l.first, l.second.size(), l.second.data());
+            if (!submit_flush_op(pg.pool_id, pg.pg_num, fb, false, l.first, l.second.size(), l.second.data()))
+                return;
         }
     }
 }
@@ -160,7 +162,7 @@ void osd_t::handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, p
     }
 }
 
-void osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t *fb, bool rollback, osd_num_t peer_osd, int count, obj_ver_id *data)
+bool osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t *fb, bool rollback, osd_num_t peer_osd, int count, obj_ver_id *data)
 {
     osd_op_t *op = new osd_op_t();
     // Copy buffer so it gets freed along with the operation
@@ -188,10 +190,8 @@ void osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t
     else
     {
         // Peer
-        int peer_fd = msgr.osd_peer_fds[peer_osd];
         op->op_type = OSD_OP_OUT;
         op->iov.push_back(op->buf, count * sizeof(obj_ver_id));
-        op->peer_fd = peer_fd;
         op->req = (osd_any_op_t){
             .sec_stab = {
                 .header = {
@@ -207,8 +207,21 @@ void osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t
             handle_flush_op(op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK, pool_id, pg_num, fb, peer_osd, op->reply.hdr.retval);
             delete op;
         };
-        msgr.outbox_push(op);
+        auto peer_fd_it = msgr.osd_peer_fds.find(peer_osd);
+        if (peer_fd_it != msgr.osd_peer_fds.end())
+        {
+            op->peer_fd = peer_fd_it->second;
+            msgr.outbox_push(op);
+        }
+        else
+        {
+            // Fail it immediately
+            op->reply.hdr.retval = -EPIPE;
+            op->callback(op);
+            return false;
+        }
     }
+    return true;
 }
 
 bool osd_t::pick_next_recovery(osd_recovery_op_t &op)

src/osd_peering.cpp

@@ -29,8 +29,10 @@ void osd_t::handle_peers()
                     degraded_objects += p.second.degraded_objects.size();
                     if (p.second.state & PG_HAS_UNCLEAN)
                         peering_state = peering_state | OSD_FLUSHING_PGS;
-                    else if (p.second.state & PG_HAS_DEGRADED)
+                    else if (p.second.state & (PG_HAS_DEGRADED | PG_HAS_MISPLACED))
                         peering_state = peering_state | OSD_RECOVERING;
+                    ringloop->wakeup();
+                    return;
                 }
                 else
                 {
@@ -340,7 +342,7 @@ void osd_t::submit_sync_and_list_subop(osd_num_t role_osd, pg_peering_state_t *p
     else
     {
         // Peer
-        auto & cl = msgr.clients.at(msgr.osd_peer_fds[role_osd]);
+        auto & cl = msgr.clients.at(msgr.osd_peer_fds.at(role_osd));
         osd_op_t *op = new osd_op_t();
         op->op_type = OSD_OP_OUT;
         op->peer_fd = cl->peer_fd;
@@ -394,7 +396,9 @@ void osd_t::submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
         {
             if (op->bs_op->retval < 0)
             {
-                throw std::runtime_error("local OP_LIST failed");
+                printf("Local OP_LIST failed: retval=%d\n", op->bs_op->retval);
+                force_stop(1);
+                return;
             }
             add_bs_subop_stats(op);
             printf(
@@ -419,7 +423,7 @@ void osd_t::submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
         // Peer
         osd_op_t *op = new osd_op_t();
         op->op_type = OSD_OP_OUT;
-        op->peer_fd = msgr.osd_peer_fds[role_osd];
+        op->peer_fd = msgr.osd_peer_fds.at(role_osd);
         op->req = (osd_any_op_t){
             .sec_list = {
                 .header = {

src/osd_primary.cpp

@@ -194,18 +194,22 @@ void osd_t::continue_primary_read(osd_op_t *cur_op)
         // Determine version
         auto vo_it = pg.ver_override.find(op_data->oid);
         op_data->target_ver = vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX;
+        op_data->prev_set = pg.cur_set.data();
+        if (pg.state != PG_ACTIVE)
+        {
+            // PG may be degraded or have misplaced objects
+            op_data->prev_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
+        }
         if (pg.state == PG_ACTIVE || op_data->scheme == POOL_SCHEME_REPLICATED)
         {
             // Fast happy-path
             cur_op->buf = alloc_read_buffer(op_data->stripes, op_data->pg_data_size, 0);
-            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, pg.cur_set.data(), cur_op);
+            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, op_data->prev_set, cur_op);
             op_data->st = 1;
         }
         else
         {
-            // PG may be degraded or have misplaced objects
-            uint64_t* cur_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
-            if (extend_missing_stripes(op_data->stripes, cur_set, op_data->pg_data_size, pg.pg_size) < 0)
+            if (extend_missing_stripes(op_data->stripes, op_data->prev_set, op_data->pg_data_size, pg.pg_size) < 0)
             {
                 finish_op(cur_op, -EIO);
                 return;
@@ -215,7 +219,7 @@ void osd_t::continue_primary_read(osd_op_t *cur_op)
             op_data->scheme = pg.scheme;
             op_data->degraded = 1;
             cur_op->buf = alloc_read_buffer(op_data->stripes, pg.pg_size, 0);
-            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, cur_set, cur_op);
+            submit_primary_subops(SUBMIT_RMW_READ, op_data->target_ver, op_data->prev_set, cur_op);
             op_data->st = 1;
         }
     }

src/osd_primary_chain.cpp

@@ -246,7 +246,6 @@ int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
                 // Send to a remote OSD
                 osd_op_t *subop = op_data->subops+subop_idx;
                 subop->op_type = OSD_OP_OUT;
-                subop->peer_fd = msgr.osd_peer_fds.at(subop_osd_num);
                 // FIXME: Use the pre-allocated buffer
                 subop->buf = malloc_or_die(sizeof(obj_ver_id)*(i+1-prev));
                 subop->req = (osd_any_op_t){
@@ -287,7 +286,18 @@ int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
                     }
                     handle_primary_subop(subop, cur_op);
                 };
-                msgr.outbox_push(subop);
+                auto peer_fd_it = msgr.osd_peer_fds.find(subop_osd_num);
+                if (peer_fd_it != msgr.osd_peer_fds.end())
+                {
+                    subop->peer_fd = peer_fd_it->second;
+                    msgr.outbox_push(subop);
+                }
+                else
+                {
+                    // Fail it immediately
+                    subop->reply.hdr.retval = -EPIPE;
+                    subop->callback(subop);
+                }
                 subop_idx++;
             }
             prev = i+1;
@@ -390,18 +400,21 @@ int osd_t::submit_chained_read_requests(pg_t & pg, osd_op_t *cur_op)
             stripes[role].read_end = stripes[role].req_end;
         }
         uint64_t *cur_set = pg.cur_set.data();
-        if (pg.state != PG_ACTIVE && op_data->scheme != POOL_SCHEME_REPLICATED)
+        if (pg.state != PG_ACTIVE)
         {
             pg_osd_set_state_t *object_state;
             cur_set = get_object_osd_set(pg, cur_oid, pg.cur_set.data(), &object_state);
-            if (extend_missing_stripes(stripes, cur_set, pg.pg_data_size, pg.pg_size) < 0)
+            if (op_data->scheme != POOL_SCHEME_REPLICATED)
             {
-                free(op_data->chain_reads);
-                op_data->chain_reads = NULL;
-                finish_op(cur_op, -EIO);
-                return -1;
+                if (extend_missing_stripes(stripes, cur_set, pg.pg_data_size, pg.pg_size) < 0)
+                {
+                    free(op_data->chain_reads);
+                    op_data->chain_reads = NULL;
+                    finish_op(cur_op, -EIO);
+                    return -1;
+                }
+                op_data->degraded = 1;
             }
-            op_data->degraded = 1;
         }
         if (op_data->scheme == POOL_SCHEME_REPLICATED)
         {
@@ -455,7 +468,7 @@ int osd_t::submit_chained_read_requests(pg_t & pg, osd_op_t *cur_op)
         auto vo_it = pg.ver_override.find(cur_oid);
         uint64_t target_ver = vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX;
         uint64_t *cur_set = pg.cur_set.data();
-        if (pg.state != PG_ACTIVE && op_data->scheme != POOL_SCHEME_REPLICATED)
+        if (pg.state != PG_ACTIVE)
         {
             pg_osd_set_state_t *object_state;
             cur_set = get_object_osd_set(pg, cur_oid, pg.cur_set.data(), &object_state);

src/osd_primary_subops.cpp

@@ -182,7 +182,6 @@ int osd_t::submit_primary_subop_batch(int submit_type, inode_t inode, uint64_t o
             else
             {
                 subop->op_type = OSD_OP_OUT;
-                subop->peer_fd = msgr.osd_peer_fds.at(role_osd_num);
                 subop->bitmap = stripes[stripe_num].bmp_buf;
                 subop->bitmap_len = clean_entry_bitmap_size;
                 subop->req.sec_rw = {
@@ -225,7 +224,18 @@ int osd_t::submit_primary_subop_batch(int submit_type, inode_t inode, uint64_t o
                 {
                     handle_primary_subop(subop, cur_op);
                 };
-                msgr.outbox_push(subop);
+                auto peer_fd_it = msgr.osd_peer_fds.find(role_osd_num);
+                if (peer_fd_it != msgr.osd_peer_fds.end())
+                {
+                    subop->peer_fd = peer_fd_it->second;
+                    msgr.outbox_push(subop);
+                }
+                else
+                {
+                    // Fail it immediately
+                    subop->reply.hdr.retval = -EPIPE;
+                    subop->callback(subop);
+                }
             }
             i++;
         }
@@ -463,7 +473,6 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
         else
         {
             subops[i].op_type = OSD_OP_OUT;
-            subops[i].peer_fd = msgr.osd_peer_fds.at(chunk.osd_num);
             subops[i].req = (osd_any_op_t){ .sec_del = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
@@ -477,7 +486,18 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
             {
                 handle_primary_subop(subop, cur_op);
             };
-            msgr.outbox_push(&subops[i]);
+            auto peer_fd_it = msgr.osd_peer_fds.find(chunk.osd_num);
+            if (peer_fd_it != msgr.osd_peer_fds.end())
+            {
+                subops[i].peer_fd = peer_fd_it->second;
+                msgr.outbox_push(&subops[i]);
+            }
+            else
+            {
+                // Fail it immediately
+                subops[i].reply.hdr.retval = -EPIPE;
+                subops[i].callback(&subops[i]);
+            }
         }
     }
 }
@@ -567,7 +587,6 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
         else
         {
             subops[i].op_type = OSD_OP_OUT;
-            subops[i].peer_fd = msgr.osd_peer_fds.at(stab_osd.osd_num);
             subops[i].req = (osd_any_op_t){ .sec_stab = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
@@ -581,7 +600,18 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
             {
                 handle_primary_subop(subop, cur_op);
             };
-            msgr.outbox_push(&subops[i]);
+            auto peer_fd_it = msgr.osd_peer_fds.find(stab_osd.osd_num);
+            if (peer_fd_it != msgr.osd_peer_fds.end())
+            {
+                subops[i].peer_fd = peer_fd_it->second;
+                msgr.outbox_push(&subops[i]);
+            }
+            else
+            {
+                // Fail it immediately
+                subops[i].reply.hdr.retval = -EPIPE;
+                subops[i].callback(&subops[i]);
+            }
         }
     }
 }

src/osd_rmw.h

@@ -8,7 +8,7 @@
 #include "osd_id.h"
 
 #ifndef MEM_ALIGNMENT
-#define MEM_ALIGNMENT 512
+#define MEM_ALIGNMENT 4096
 #endif
 
 struct buf_len_t

src/osd_test.cpp

@@ -134,14 +134,14 @@ int main(int narg, char *args[])
 
 int connect_osd(const char *osd_address, int osd_port)
 {
-    struct sockaddr addr;
+    struct sockaddr_storage addr;
     if (!string_to_addr(osd_address, 0, osd_port, &addr))
     {
         fprintf(stderr, "server address: %s is not valid\n", osd_address);
         return -1;
     }
 
-    int connect_fd = socket(addr.sa_family, SOCK_STREAM, 0);
+    int connect_fd = socket(addr.ss_family, SOCK_STREAM, 0);
     if (connect_fd < 0)
     {
         perror("socket");

src/ringloop.cpp

@@ -112,3 +112,17 @@ void ring_loop_t::restore(unsigned sqe_tail)
     }
     ring.sq.sqe_tail = sqe_tail;
 }
+
+int ring_loop_t::sqes_left()
+{
+    struct io_uring_sq *sq = &ring.sq;
+    unsigned int head = io_uring_smp_load_acquire(sq->khead);
+    unsigned int next = sq->sqe_tail + 1;
+    int left = *sq->kring_entries - (next - head);
+    if (left > free_ring_data_ptr)
+    {
+        // return min(sqes left, ring_datas left)
+        return free_ring_data_ptr;
+    }
+    return left;
+}

src/ringloop.h

@@ -17,15 +17,12 @@
 
 static inline void my_uring_prep_rw(int op, struct io_uring_sqe *sqe, int fd, const void *addr, unsigned len, off_t offset)
 {
-    sqe->opcode = op;
-    sqe->flags = 0;
-    sqe->ioprio = 0;
-    sqe->fd = fd;
-    sqe->off = offset;
-    sqe->addr = (unsigned long) addr;
-    sqe->len = len;
-    sqe->rw_flags = 0;
-    sqe->__pad2[0] = sqe->__pad2[1] = sqe->__pad2[2] = 0;
+    // Prepare a read/write operation without clearing user_data
+    // Very recently, 22 Dec 2021, liburing finally got this change too (8ecd3fd959634df81d66af8b3a69c16202a014e8)
+    // But all versions prior to it (sadly) clear user_data
+    __u64 user_data = sqe->user_data;
+    io_uring_prep_rw(op, sqe, fd, addr, len, offset);
+    sqe->user_data = user_data;
 }
 
 static inline void my_uring_prep_readv(struct io_uring_sqe *sqe, int fd, const struct iovec *iovecs, unsigned nr_vecs, off_t offset)
@@ -172,6 +169,7 @@ public:
         struct io_uring_cqe *cqe;
         return io_uring_wait_cqe(&ring, &cqe);
     }
+    int sqes_left();
     inline unsigned space_left()
     {
         return free_ring_data_ptr;

src/stub_bench.cpp

@@ -67,14 +67,14 @@ int main(int narg, char *args[])
 
 int connect_stub(const char *server_address, int server_port)
 {
-    struct sockaddr addr;
+    struct sockaddr_storage addr;
     if (!string_to_addr(server_address, 0, server_port, &addr))
     {
         fprintf(stderr, "server address: %s is not valid\n", server_address);
         return -1;
     }
 
-    int connect_fd = socket(addr.sa_family, SOCK_STREAM, 0);
+    int connect_fd = socket(addr.ss_family, SOCK_STREAM, 0);
     if (connect_fd < 0)
     {
         perror("socket");

src/stub_osd.cpp

@@ -41,21 +41,19 @@
 #include "rw_blocking.h"
 #include "osd_ops.h"
 
-int bind_stub(std::string bind_address, int bind_port);
-
 void run_stub(int peer_fd);
 
 int main(int narg, char *args[])
 {
-    int listen_fd = bind_stub("0.0.0.0", 11203);
+    int listen_fd = create_and_bind_socket("0.0.0.0", 11203, 128, NULL);
     // Accept new connections
-    sockaddr addr;
+    sockaddr_storage addr;
     socklen_t peer_addr_size = sizeof(addr);
     int peer_fd;
     while (1)
     {
         printf("stub_osd: waiting for 1 client\n");
-        peer_fd = accept(listen_fd, &addr, &peer_addr_size);
+        peer_fd = accept(listen_fd, (sockaddr*)&addr, &peer_addr_size);
         if (peer_fd == -1)
         {
             if (errno == EAGAIN)
@@ -76,39 +74,6 @@ int main(int narg, char *args[])
     return 0;
 }
 
-int bind_stub(std::string bind_address, int bind_port)
-{
-    int listen_backlog = 128;
-
-    sockaddr addr;
-    if (!string_to_addr(bind_address, 0, bind_port, &addr))
-    {
-        throw std::runtime_error("bind address "+bind_address+" is not valid");
-    }
-
-    int listen_fd = socket(addr.sa_family, SOCK_STREAM, 0);
-    if (listen_fd < 0)
-    {
-        throw std::runtime_error(std::string("socket: ") + strerror(errno));
-    }
-    int enable = 1;
-    setsockopt(listen_fd, SOL_SOCKET, SO_REUSEADDR, &enable, sizeof(enable));
-
-    if (bind(listen_fd, &addr, sizeof(addr)) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("bind: ") + strerror(errno));
-    }
-
-    if (listen(listen_fd, listen_backlog) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("listen: ") + strerror(errno));
-    }
-
-    return listen_fd;
-}
-
 void run_stub(int peer_fd)
 {
     osd_any_op_t op;

src/stub_uring_osd.cpp

@@ -25,8 +25,6 @@
 #include "epoll_manager.h"
 #include "messenger.h"
 
-int bind_stub(std::string bind_address, int bind_port);
-
 void stub_exec_op(osd_messenger_t *msgr, osd_op_t *op);
 
 int main(int narg, char *args[])
@@ -43,7 +41,8 @@ int main(int narg, char *args[])
     json11::Json config = json11::Json::object { { "log_level", 1 } };
     msgr->parse_config(config);
     // Accept new connections
-    int listen_fd = bind_stub("0.0.0.0", 11203);
+    int listen_fd = create_and_bind_socket("0.0.0.0", 11203, 128, NULL);
+    fcntl(listen_fd, F_SETFL, fcntl(listen_fd, F_GETFL, 0) | O_NONBLOCK);
     epmgr->set_fd_handler(listen_fd, false, [listen_fd, msgr](int fd, int events)
     {
         msgr->accept_connections(listen_fd);
@@ -67,41 +66,6 @@ int main(int narg, char *args[])
     return 0;
 }
 
-int bind_stub(std::string bind_address, int bind_port)
-{
-    int listen_backlog = 128;
-
-    sockaddr addr;
-    if (!string_to_addr(bind_address, 0, bind_port, &addr))
-    {
-        throw std::runtime_error("bind address "+bind_address+" is not valid");
-    }
-
-    int listen_fd = socket(addr.sa_family, SOCK_STREAM, 0);
-    if (listen_fd < 0)
-    {
-        throw std::runtime_error(std::string("socket: ") + strerror(errno));
-    }
-    int enable = 1;
-    setsockopt(listen_fd, SOL_SOCKET, SO_REUSEADDR, &enable, sizeof(enable));
-
-    if (bind(listen_fd, &addr, sizeof(addr)) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("bind: ") + strerror(errno));
-    }
-
-    if (listen(listen_fd, listen_backlog) < 0)
-    {
-        close(listen_fd);
-        throw std::runtime_error(std::string("listen: ") + strerror(errno));
-    }
-
-    fcntl(listen_fd, F_SETFL, fcntl(listen_fd, F_GETFL, 0) | O_NONBLOCK);
-
-    return listen_fd;
-}
-
 void stub_exec_op(osd_messenger_t *msgr, osd_op_t *op)
 {
     op->reply.hdr.magic = SECONDARY_OSD_REPLY_MAGIC;

src/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 0.6.12
+Version: 0.6.16
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

tests/run_7osds.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+
+. `dirname $0`/common.sh
+
+if [ "$IMMEDIATE_COMMIT" != "" ]; then
+    NO_SAME="--journal_no_same_sector_overwrites true --journal_sector_buffer_count 1024 --disable_data_fsync 1 --immediate_commit all --log_level 1"
+    $ETCDCTL put /vitastor/config/global '{"recovery_queue_depth":1,"osd_out_time":5,"immediate_commit":"all"}'
+else
+    NO_SAME="--journal_sector_buffer_count 1024 --log_level 1"
+    $ETCDCTL put /vitastor/config/global '{"recovery_queue_depth":1,"osd_out_time":5}'
+fi
+
+OSD_SIZE=1024
+OSD_COUNT=7
+OSD_ARGS=
+for i in $(seq 1 $OSD_COUNT); do
+    dd if=/dev/zero of=./testdata/test_osd$i.bin bs=1024 count=1 seek=$((OSD_SIZE*1024-1))
+    build/src/vitastor-osd --osd_num $i --bind_address 127.0.0.1 $NO_SAME $OSD_ARGS --etcd_address $ETCD_URL $(build/src/vitastor-cli simple-offsets --format options ./testdata/test_osd$i.bin 2>/dev/null) &>./testdata/osd$i.log &
+    eval OSD${i}_PID=$!
+done
+
+cd mon
+npm install
+cd ..
+node mon/mon-main.js --etcd_url $ETCD_URL --etcd_prefix "/vitastor" --verbose 1 &>./testdata/mon.log &
+MON_PID=$!
+
+if [ "$EC" != "" ]; then
+    POOLCFG='"scheme":"xor","pg_size":3,"pg_minsize":2,"parity_chunks":1'
+    PG_SIZE=3
+else
+    POOLCFG='"scheme":"replicated","pg_size":2,"pg_minsize":2'
+    PG_SIZE=2
+fi
+$ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool",'$POOLCFG',"pg_count":32,"failure_domain":"osd"}}'
+
+sleep 2
+
+if ! ($ETCDCTL get /vitastor/config/pgs --print-value-only | jq -s -e '(.[0].items["1"] | map((.osd_set | select(. > 0)) | length == '$PG_SIZE') | length) == 32'); then
+    format_error "FAILED: 32 PGS NOT CONFIGURED"
+fi
+
+if ! ($ETCDCTL get --prefix /vitastor/pg/state/ --print-value-only | jq -s -e '([ .[] | select(.state == ["active"]) ] | length) == 32'); then
+    format_error "FAILED: 32 PGS NOT UP"
+fi
+
+try_reweight()
+{
+    osd=$1
+    w=$2
+    $ETCDCTL put /vitastor/config/osd/$osd '{"reweight":'$w'}'
+    sleep 3
+}
+
+wait_finish_rebalance()
+{
+    sec=$1
+    i=0
+    while [[ $i -lt $sec ]]; do
+        ($ETCDCTL get --prefix /vitastor/pg/state/ --print-value-only | jq -s -e '([ .[] | select(.state == ["active"]) ] | length) == 32') && \
+            break
+        if [ $i -eq 60 ]; then
+            format_error "Rebalance couldn't finish in $sec seconds"
+        fi
+        sleep 1
+        i=$((i+1))
+    done
+}

tests/run_tests.sh

@@ -0,0 +1,43 @@
+#!/bin/bash -ex
+# Run all possible tests
+
+cd $(dirname $0)
+
+./test_add_osd.sh
+
+./test_cas.sh
+
+./test_change_pg_count.sh
+EC=1 ./test_change_pg_count.sh
+
+./test_change_pg_size.sh
+
+./test_etcd_fail.sh
+
+./test_failure_domain.sh
+
+./test_interrupted_rebalance.sh
+IMMEDIATE_COMMIT=1 ./test_interrupted_rebalance.sh
+EC=1 ./test_interrupted_rebalance.sh
+EC=1 IMMEDIATE_COMMIT=1 ./test_interrupted_rebalance.sh
+
+./test_minsize_1.sh
+
+./test_move_reappear.sh
+
+./test_rebalance_verify.sh
+IMMEDIATE_COMMIT=1 ./test_rebalance_verify.sh
+EC=1 ./test_rebalance_verify.sh
+EC=1 IMMEDIATE_COMMIT=1 ./test_rebalance_verify.sh
+
+./test_rm.sh
+
+./test_snapshot.sh
+SCHEME=replicated ./test_snapshot.sh
+
+./test_splitbrain.sh
+
+./test_write.sh
+SCHEME=replicated ./test_write.sh
+
+./test_write_no_same.sh

tests/test_change_pg_count.sh

@@ -78,7 +78,16 @@ try_change()
     fi
 
     # Check that no objects are lost !
-    nobj=`$ETCDCTL get --prefix '/vitastor/pg/stats' --print-value-only | jq -s '[ .[].object_count ] | reduce .[] as $num (0; .+$num)'`
+    # But note that reporting this information may take up to <etcd_report_interval+1> seconds
+    nobj=0
+    waittime=0
+    while [[ $nobj -ne $NOBJ && $waittime -lt 7 ]]; do
+        nobj=`$ETCDCTL get --prefix '/vitastor/pg/stats' --print-value-only | jq -s '[ .[].object_count ] | reduce .[] as $num (0; .+$num)'`
+        if [[ $nobj -ne $NOBJ ]]; then
+            waittime=$((waittime+1))
+            sleep 1
+        fi
+    done
     if [ "$nobj" -ne $NOBJ ]; then
         format_error "Data lost after changing PG count to $n: $NOBJ objects expected, but got $nobj"
     fi

tests/test_interrupted_rebalance.sh

@@ -1,41 +1,6 @@
 #!/bin/bash -ex
 
-. `dirname $0`/common.sh
-
-if [ "$IMMEDIATE_COMMIT" != "" ]; then
-    NO_SAME="--journal_no_same_sector_overwrites true --journal_sector_buffer_count 1024 --disable_data_fsync 1 --immediate_commit all --log_level 1"
-    $ETCDCTL put /vitastor/config/global '{"recovery_queue_depth":1,"osd_out_time":5,"immediate_commit":"all"}'
-else
-    NO_SAME="--journal_sector_buffer_count 1024 --log_level 1"
-    $ETCDCTL put /vitastor/config/global '{"recovery_queue_depth":1,"osd_out_time":5}'
-fi
-
-OSD_SIZE=1024
-OSD_COUNT=7
-OSD_ARGS=
-for i in $(seq 1 $OSD_COUNT); do
-    dd if=/dev/zero of=./testdata/test_osd$i.bin bs=1024 count=1 seek=$((OSD_SIZE*1024-1))
-    build/src/vitastor-osd --osd_num $i --bind_address 127.0.0.1 $OSD_ARGS --etcd_address $ETCD_URL $(build/src/vitastor-cli simple-offsets --format options ./testdata/test_osd$i.bin 2>/dev/null) &>./testdata/osd$i.log &
-    eval OSD${i}_PID=$!
-done
-
-cd mon
-npm install
-cd ..
-node mon/mon-main.js --etcd_url $ETCD_URL --etcd_prefix "/vitastor" --verbose 1 &>./testdata/mon.log &
-MON_PID=$!
-
-$ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool","scheme":"replicated","pg_size":2,"pg_minsize":1,"pg_count":32,"failure_domain":"osd"}}'
-
-sleep 2
-
-if ! ($ETCDCTL get /vitastor/config/pgs --print-value-only | jq -s -e '(.[0].items["1"] | map((.osd_set | select(. > 0)) | length == 2) | length) == 32'); then
-    format_error "FAILED: 32 PGS NOT CONFIGURED"
-fi
-
-if ! ($ETCDCTL get --prefix /vitastor/pg/state/ --print-value-only | jq -s -e '([ .[] | select(.state == ["active"]) ] | length) == 32'); then
-    format_error "FAILED: 32 PGS NOT UP"
-fi
+. `dirname $0`/run_7osds.sh
 
 IMG_SIZE=960
 
@@ -43,14 +8,6 @@ LD_PRELOAD=libasan.so.5 \
 fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=16 -fsync=16 -rw=write \
     -etcd=$ETCD_URL -pool=1 -inode=2 -size=${IMG_SIZE}M -cluster_log_level=10
 
-try_reweight()
-{
-    osd=$1
-    w=$2
-    $ETCDCTL put /vitastor/config/osd/$osd '{"reweight":'$w'}'
-    sleep 3
-}
-
 try_reweight 1 0
 
 try_reweight 2 0
@@ -72,14 +29,7 @@ try_reweight 4 1
 try_reweight 5 1
 
 # Wait for the rebalance to finish
-for i in {1..60}; do
-    ($ETCDCTL get --prefix /vitastor/pg/state/ --print-value-only | jq -s -e '([ .[] | select(.state == ["active"]) ] | length) == 32') && \
-        break
-    if [ $i -eq 60 ]; then
-        format_error "Rebalance couldn't finish in 60 seconds"
-    fi
-    sleep 1
-done
+wait_finish_rebalance 60
 
 # Check that PGs never had degraded objects !
 if grep has_degraded ./testdata/mon.log; then

tests/test_rebalance_verify.sh

@@ -0,0 +1,57 @@
+#!/bin/bash -ex
+
+. `dirname $0`/run_7osds.sh
+
+IMG_SIZE=256
+
+$ETCDCTL put /vitastor/config/inode/1/1 '{"name":"testimg","size":'$((IMG_SIZE*1024*1024))'}'
+
+NBD_DEV=$(sudo build/src/vitastor-nbd map --etcd_address $ETCD_URL --image testimg --logfile ./testdata/nbd.log &)
+
+trap "sudo build/src/vitastor-nbd unmap $NBD_DEV"'; kill -9 $(jobs -p)' EXIT
+
+sudo chown $(id -u) $NBD_DEV
+
+dd if=/dev/urandom of=./testdata/img1.bin bs=1M count=$IMG_SIZE
+
+dd if=./testdata/img1.bin of=$NBD_DEV bs=1M count=$IMG_SIZE oflag=direct
+
+verify() {
+    echo "Verifying before rebalance"
+    dd if=$NBD_DEV of=./testdata/img2.bin bs=1M count=$IMG_SIZE iflag=direct
+    diff ./testdata/img1.bin ./testdata/img2.bin
+
+    $ETCDCTL put /vitastor/config/osd/1 '{"reweight":'$1'}'
+    $ETCDCTL put /vitastor/config/osd/2 '{"reweight":'$1'}'
+    $ETCDCTL put /vitastor/config/osd/3 '{"reweight":'$1'}'
+    sleep 1
+
+    for i in {1..10000}; do
+        O=$(((RANDOM*RANDOM) % (IMG_SIZE*128)))
+        dd if=$NBD_DEV of=./testdata/img2.bin bs=4k seek=$O skip=$O count=1 iflag=direct conv=notrunc
+    done
+
+    echo "Verifying during rebalance"
+    diff ./testdata/img1.bin ./testdata/img2.bin
+
+    # Wait for the rebalance to finish
+    wait_finish_rebalance 60
+
+    echo "Verifying after rebalance"
+    dd if=$NBD_DEV of=./testdata/img2.bin bs=1M count=$IMG_SIZE iflag=direct
+    diff ./testdata/img1.bin ./testdata/img2.bin
+}
+
+# Verify with regular reads
+
+verify 0
+
+# Same with chained reads
+
+$ETCDCTL put /vitastor/config/inode/1/1 '{"name":"testimg0","size":'$((IMG_SIZE*1024*1024))'}'
+$ETCDCTL put /vitastor/config/inode/1/2 '{"name":"testimg","size":'$((IMG_SIZE*1024*1024))',"parent_id":1}'
+sleep 1
+
+verify 1
+
+format_green OK