WIP NFS RDMA support

Could probably lead to PGs hung in peering states on OSD restart in EC pools,
fixable by primary OSD restart

.gitea/workflows/buildenv.Dockerfile

@@ -22,7 +22,7 @@ RUN apt-get update
 RUN apt-get -y install etcd qemu-system-x86 qemu-block-extra qemu-utils fio libasan5 \
     liburing1 liburing-dev libgoogle-perftools-dev devscripts libjerasure-dev cmake libibverbs-dev libisal-dev
 RUN apt-get -y build-dep fio qemu=`dpkg -s qemu-system-x86|grep ^Version:|awk '{print $2}'`
-RUN apt-get -y install jq lp-solve sudo nfs-common
+RUN apt-get update && apt-get -y install jq lp-solve sudo nfs-common fdisk parted
 RUN apt-get --download-only source fio qemu=`dpkg -s qemu-system-x86|grep ^Version:|awk '{print $2}'`
 
 RUN set -ex; \

.gitea/workflows/test.yml

@@ -828,6 +828,42 @@ jobs:
           echo ""
         done
 
+  test_resize:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_resize.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
+  test_resize_auto:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_resize_auto.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_snapshot_pool2:
     runs-on: ubuntu-latest
     needs: build

CMakeLists.txt

@@ -2,6 +2,6 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VITASTOR_VERSION "1.9.2")
+set(VITASTOR_VERSION "1.9.3")
 
 add_subdirectory(src)

csi/Makefile

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v1.9.2
+VITASTOR_VERSION ?= v1.9.3
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v1.9.2
+          image: vitalif/vitastor-csi:v1.9.3
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/007-csi-provisioner.yaml

@@ -121,7 +121,7 @@ spec:
             privileged: true
             capabilities:
               add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v1.9.2
+          image: vitalif/vitastor-csi:v1.9.3
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "1.9.2"
+    vitastorCSIDriverVersion = "1.9.3"
 )
 
 // Config struct fills the parameters of request or user input

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (1.9.2-1) unstable; urgency=medium
+vitastor (1.9.3-1) unstable; urgency=medium
 
   * Bugfixes
 

docs/intro/architecture.en.md

@@ -6,19 +6,150 @@
 
 # Architecture
 
+- [Server-side components](#server-side-components)
 - [Basic concepts](#basic-concepts)
+- [Client-side components](#client-side-components)
+- [Additional utilities](#additional-utilities)
+- [Overall read/write process](#overall-read-write-process)
+  - [Nuances of request handling](#nuances-of-request-handling)
 - [Similarities to Ceph](#similarities-to-ceph)
 - [Differences from Ceph](#differences-from-ceph)
 - [Implementation Principles](#implementation-principles)
 
+## Server-side components
+
+- **OSD** (Object Storage Daemon) is a process that directly works with the disk, stores data
+  and serves read/write requests. One OSD serves one disk (or one partition). OSDs talk to etcd
+  and to each other — they receive cluster state from etcd, and send read/write requests for
+  secondary copies of data to other OSDs.
+- **etcd** — clustered key/value database, used as a reliable storage for configuration
+  and high-level cluster state. Etcd is the component that prevents splitbrain in the cluster.
+  Data blocks are not stored in etcd, etcd doesn't participate in data write or read path.
+- **Монитор** — a separate node.js based daemon which monitors the cluster, calculates
+  required configuration changes and saves them to etcd, thus commanding OSDs to apply these
+  changes. Monitor also aggregates cluster statistics. OSD don't talk to monitor, monitor
+  only sends and receives data from etcd.
+
 ## Basic concepts
 
-- OSD (Object Storage Daemon) is a process that stores data and serves read/write requests.
+- **Pool** is a container for data that has equal redundancy scheme and disk placement rules.
-- PG (Placement Group) is a "shard" of the cluster, group of data stored on one set of replicas.
+- **PG (Placement Group)** is a "shard" of the cluster, subdivision unit that has its own
-- Pool is a container for data that has equal redundancy scheme and placement rules.
+  set of OSDs for data storage.
-- Monitor is a separate daemon that watches cluster state and handles failures.
+- **Failure Domain** is a group of OSDs, from the simultaneous failure of which you are
-- Failure Domain is a group of OSDs that you allow to fail. It's "host" by default.
+  protected by Vitastor. Default failure domain is "host" (server), but you choose a
-- Placement Tree groups OSDs in a hierarchy to later split them into Failure Domains.
+  larger (for example, a rack of servers) or smaller (a single drive) failure domain
+  for every pool.
+- **Placement Tree** (similar to Ceph CRUSH Tree) groups OSDs in a hierarchy to later
+  split them into Failure Domains.
+
+## Client-side components
+
+- **Client library** encapsulates client I/O logic. Client library connects to etcd and to all OSDs,
+  receives cluster state from etcd, sends read and write requests directly to all OSDs. Due
+  to the symmetric distributed architecture, all data blocks (each 128 KB by default) are placed
+  to different OSDs, but clients always know where each data block is stored and connect directly
+  to the right OSD.
+
+All other client-side components are based on the client library:
+
+- **[vitastor-cli](../usage/cli.en.md)** — command-line utility for cluster management.
+  Allows to view cluster state, manage pools and images, i.e. create, modify and remove
+  virtual disks, their snapshots and clones.
+- **[QEMU driver](../usage/qemu.en.md)** — pluggable QEMU module allowing QEMU/KVM virtual
+  machines work with virtual Vitastor disks directly from userspace through the client library,
+  without the need to attach disks as kernel block devices. However, if you want to attach
+  disks, you can also do that with the same driver and [VDUSE](../usage/qemu.en.md#vduse).
+- **[vitastor-nbd](../usage/nbd.en.md)** — utility that allows to attach Vitastor disks as
+  kernel block devices using NBD (Network Block Device), which works more like "BUSE"
+  (Block Device In Userspace). Vitastor doesn't have Linux kernel modules for the same task
+  (at least by now). NBD is an older, non-recommended way to attach disks — you should use
+  VDUSE whenever you can.
+- **[CSI driver](../installation/kubernetes.en.md)** — driver for attaching Vitastor images
+  as Kubernetes persistent volumes. Works through VDUSE (when available) or NBD — images are
+  attached as kernel block devices and mounted into containers.
+- **Drivers for Proxmox, OpenStack and so on** — pluggable modules for corresponding systems,
+  allowing to use Vitastor as storage in them.
+- **[vitastor-nfs](../usage/nfs.en.md)** — NFS 3.0 server allowing export of two file system variants:
+  the first is a simplified pseudo-FS for file-based access to Vitastor block images (for non-QEMU
+  hypervisors with NFS support), the second is **VitastorFS**, full-featured clustered POSIX FS.
+  Both variants support parallel access from multiple vitastor-nfs servers. In fact, you are
+  not required to setup separate NFS servers at all and use vitastor-nfs mount command on every
+  client node — it starts the NFS server and mounts the FS locally.
+- **[fio driver](../usage/fio.en.md)** — pluggable module for fio disk benchmarking tool for
+  running performance tests on your Vitastor cluster.
+- **vitastor-kv** — client for a key-value DB working over shared block volumes (usual
+  vitastor images). VitastorFS metadata is stored in vitastor-kv.
+
+## Additional utilities
+
+- **vitastor-disk** — a Vitastor OSD disk management tool. You can create, remove,
+  resize and move OSD partitions with it.
+
+## Overall read/write process
+
+- Vitastor stores virtual disks, also named "images" or "inodes".
+- Each image is stored in some pool. Pool specifies storage parameters such as redundancy
+  scheme (replication or EC — erasure codes, i.e. error correction codes), failure domain
+  and restrictions on OSD selection for image data placement. See [Pool configuration](../config/pool.en.md) for details.
+- Each image is split into objects/blocks of fixed size, equal to [block_size](../config/layout-cluster.en.md#block_size)
+  (128 KB by default), multiplied by data part count for EC or 1 for replicas. That is,
+  if a pool uses EC 4+2 coding scheme (4 data parts + 2 parity parts), then, with the
+  default block_size, images are split into 512 KB objects.
+- Client read/write requests are split into parts at object boundaries.
+- Each object is mapped to a PG number it belongs to, by simply taking a remainder of
+  division of its offset by PG count of the image's pool.
+- Client reads primary OSD for all PGs from etcd. Primary OSD for each PG is assigned
+  by the monitor during cluster operation, along with the full PG OSD set.
+- If not already connected, client connects to primary OSDs of all PGs involved in a
+  read/write request and sends parts of the request to them.
+- If a primary OSD is unavailable, client retries connection attempts indefinitely
+  either until it becomes available or until the monitor assigns another OSD as primary
+  for that PG.
+- Client also retries requests if the primary OSD replies with error code EPIPE, meaning
+  that the PG is inactive at this OSD at the moment - for example, when the primary OSD
+  is switched, or if the primary OSD itself loses connection to replicas during request
+  handling.
+- Primary OSD determines where the parts of the object are stored. By default, all objects
+  are assumed to be stored at the target OSD set of a PG, but some of them may be present
+  at a different OSD set if they are degraded or moved, or if the data rebalancing process
+  is active. OSDs doesn't do any network requests, if calculates locations of all objects
+  during PG activation and stores it in memory.
+- Primary OSD handles the request locally when it can - for example, when it's a read
+  from a replicated pool or when it's a read from a EC pool involving only one data part
+  stored on the OSD's local disk.
+- When a request requires reads or writes to additional OSDs, primary OSD uses already
+  established connections to secondary OSDs of the PG to execute these requests. This happens
+  in parallel to local disk operations. All such connections are guaranteed to be already
+  established when the PG is active, and if any of them is dropped, PG is restarted and
+  all current read/write operations to it fail with EPIPE error and are retried by clients.
+- After completing all secondary read/write requests, primary OSD sends the response to
+  the client.
+
+### Nuances of request handling
+
+- If a pool uses erasure codes and some of the OSDs are unavailable, primary OSDs recover
+  data from the remaining parts during read.
+- Each object has a version number. During write, primary OSD first determines the current
+  version of the object. As primary OSD usually stores the object or its part itself, most
+  of the time version is read from the memory of the OSD itself. However, if primary OSD
+  doesn't contain parts of the object, it requests the version number from a secondary OSD
+  which has that part. Such request still doesn't involve reading from the disk though,
+  because object metadata, including version number, is always stored in OSD memory.
+- If a pool uses erasure codes, partial writes of an object require reading other parts of
+  it from secondary OSDs or from the local disk of the primary OSD itself. This is called
+  "read-modify-write" process.
+- If a pool uses erasure codes, two-phase write process is used to get rid of the Write Hole
+  problem: first a new version of object parts is written to all secondary OSDs without
+  removing the previous version, and then, after receiving successful write confirmations
+  from all OSDs, new version is committed and the old one is allowed to be removed.
+- In a pool doesn't use immediate_commit mode, then write requests sent by clients aren't
+  treated as committed to physical media instantly. Clients have to send separate type of
+  requests (SYNC) to commit changes, and before it isn't sent, new versions of data are
+  allowed to be lost if some OSDs die. Thus, when immediate_commit is disabled, clients
+  store copies of all write requests in memory and repeat them from there when the
+  connection to primary OSD is lost. This in-memory copy is removed after a successful
+  SYNC, and to prevent excessive memory usage, clients also do an automatic SYNC
+  every [client_dirty_limit](../config/network.en.md#client_dirty_limit) written bytes.
 
 ## Similarities to Ceph
 

docs/intro/architecture.ru.md

@@ -11,6 +11,7 @@
 - [Серверные компоненты](#серверные-компоненты)
 - [Базовые понятия](#базовые-понятия)
 - [Клиентские компоненты](#клиентские-компоненты)
+- [Дополнительные утилиты](#дополнительные-утилиты)
 - [Общий процесс записи и чтения](#общий-процесс-записи-и-чтения)
   - [Особенности обработки запросов](#особенности-обработки-запросов)
 - [Схожесть с Ceph](#схожесть-с-ceph)
@@ -34,8 +35,9 @@
 - **Пул (Pool)** — контейнер для данных, имеющих одну и ту же схему избыточности и правила распределения по OSD.
 - **PG (Placement Group)** — "шард", единица деления пулов в кластере, которой назначается свой набор
   OSD для хранения данных (копий или частей объектов).
-- **Домен отказа (Failure Domain)** — группа OSD, одновременное падение которых рассматривается
+- **Домен отказа (Failure Domain)** — группа OSD, от одновременного падения которых должен защищать
-  как вероятное. По умолчанию это "host" (сервер).
+  Vitastor. По умолчанию домен отказа — "host" (сервер), но вы можете установить для пула как больший
+  домен отказа (например, стойку серверов), так и меньший (например, отдельный диск).
 - **Дерево распределения** (Placement Tree, в Ceph CRUSH Tree) — иерархическая группировка OSD
   в узлы, которые далее можно использовать как домены отказа.
 
@@ -49,25 +51,39 @@
 
 На базе клиентской библиотеки реализованы все остальные клиенты:
 
-- **vitastor-cli** — утилита командной строки для управления кластером. В данный момент позволяет
+- **[vitastor-cli](../usage/cli.ru.md)** — утилита командной строки для управления кластером.
-  просматривать общее состояние кластера и управлять образами — т.е. создавать, менять и удалять
+  Позволяет просматривать общее состояние кластера, управлять пулами и образами — то есть
-  виртуальные диски, их снимки и клоны.
+  создавать, менять и удалять виртуальные диски, их снимки и клоны.
-- **Драйвер QEMU** — подключаемый модуль QEMU, позволяющий QEMU/KVM виртуальным машинам работать
+- **[Драйвер QEMU](../usage/qemu.ru.md)** — подключаемый модуль QEMU, позволяющий QEMU/KVM
-  с виртуальными дисками Vitastor напрямую из пространства пользователя с помощью клиентской
+  виртуальным машинам работать с виртуальными дисками Vitastor напрямую из пространства пользователя
-  библиотеки, без необходимости отображения дисков в виде блочных устройств. Тот же драйвер
+  с помощью клиентской библиотеки, без необходимости подключения дисков в виде блочных устройств
-  позволяет подключать диски в систему через [VDUSE](../usage/qemu.ru.md#vduse).
+  Linux. Если, однако, вы хотите подключать диски в виде блочных устройств, то вы тоже можете
-- **vitastor-nbd** — утилита, позволяющая монтировать образы Vitastor в виде блочных устройств
+  сделать это с помощью того же самого драйвера и [VDUSE](../usage/qemu.ru.md#vduse).
-  с помощью NBD (Network Block Device), на самом деле скорее работающего как "BUSE"
+- **[vitastor-nbd](../usage/nbd.ru.md)** — утилита, позволяющая монтировать образы Vitastor
-  (Block Device In Userspace). Модуля ядра Linux для выполнения той же задачи в Vitastor нет
+  в виде блочных устройств с помощью NBD (Network Block Device), на самом деле скорее работающего
-  (по крайней мере, пока).
+  как "BUSE" (Block Device In Userspace). Модуля ядра Linux для выполнения той же задачи в
-- **CSI драйвер** — драйвер для подключения Vitastor-образов в виде персистентных томов (PV) Kubernetes.
+  Vitastor нет (по крайней мере, пока). NBD — более старый и нерекомендуемый способ подключения
-  Работает через vitastor-nbd — образы отражаются в виде блочных устройств и монтируются
+  дисков — вам следует использовать VDUSE всегда, когда это возможно.
-  в контейнеры.
+- **[CSI драйвер](../installation/kubernetes.ru.md)** — драйвер для подключения Vitastor-образов
+  в виде персистентных томов (PV) Kubernetes. Работает через VDUSE (если доступно) или через
+  NBD — образы отражаются в виде блочных устройств и монтируются в контейнеры.
 - **Драйвера Proxmox, OpenStack и т.п.** — подключаемые модули для соответствующих систем,
   позволяющие использовать Vitastor как хранилище в оных.
-- **vitastor-nfs** — утилита, предоставляющая файловый доступ к образам в кластере Vitastor
+- **[vitastor-nfs](../usage/nfs.ru.md)** — NFS 3.0 сервер, предоставляющий два варианта файловой системы:
-  по протоколу NFS 3.0. Предназначена для гипервизоров, не основанных на QEMU и Linux, но при
+  первая — упрощённая для файлового доступа к блочным образам (для не-QEMU гипервизоров, поддерживающих NFS),
-  этом поддерживающих NFS.
+  вторая — VitastorFS, полноценная кластерная POSIX ФС. Оба варианта поддерживают параллельный
+  доступ с нескольких vitastor-nfs серверов. На самом деле можно вообще не выделять
+  отдельные NFS-серверы, а вместо этого использовать команду vitastor-nfs mount, запускающую
+  NFS-сервер прямо на клиентской машине и монтирующую ФС локально.
+- **[Драйвер fio](../usage/fio.ru.md)** — подключаемый модуль для утилиты тестирования
+  производительности дисков fio, позволяющий тестировать Vitastor-кластеры.
+- **vitastor-kv** — клиент для key-value базы данных, работающей поверх разделяемого блочного
+  образа (обычного блочного образа vitastor). Метаданные VitastorFS хранятся именно в vitastor-kv.
+
+## Дополнительные утилиты
+
+- **vitastor-disk** — утилита для разметки дисков под Vitastor OSD. С её помощью можно
+  создавать, удалять, менять размеры или перемещать разделы OSD.
 
 ## Общий процесс записи и чтения
 
@@ -98,16 +114,22 @@
   находиться на других OSD, если эти объекты деградированы или перемещены, или идёт процесс
   ребаланса. Запросы для проверки по сети не отправляются, информация о местоположении всех
   объектов рассчитывается первичным OSD при активации PG и хранится в памяти.
-- Первичный OSD соединяется (если ещё не соединён) с вторичными OSD, на которых располагаются
+- Когда это возможно, первичный OSD обрабатывает запрос локально. Например, так происходит
-  части объекта, и отправляет им запросы чтения/записи, а также читает/пишет из/в своё локальное
+  при чтениях объектов из пулов с репликацией или при чтении из EC пула, затрагивающего
-  хранилище, если сам входит в набор.
+  только часть, хранимую на диске самого первичного OSD.
+- Когда запрос требует записи или чтения с вторичных OSD, первичный OSD использует заранее
+  установленные соединения с ними для выполнения этих запросов. Это происходит параллельно
+  локальным операциям чтения/записи с диска самого OSD. Так как соединения к вторичным OSD PG
+  устанавливаются при её запуске, то они уже гарантированно установлены, когда PG активна,
+  и если любое из этих соединений отключается, PG перезапускается, а все текущие запросы чтения
+  и записи в неё завершаются с ошибкой EPIPE, после чего повторяются клиентами.
 - После завершения всех вторичных операций чтения/записи первичный OSD отправляет ответ клиенту.
 
 ### Особенности обработки запросов
 
 - Если в пуле используются коды коррекции ошибок и при этом часть OSD недоступна, первичный
   OSD при чтении восстанавливает данные из оставшихся частей.
-- Каждый объект имеет номер версии. При записи объекта первичный OSD сначала читает из номер
+- Каждый объект имеет номер версии. При записи объекта первичный OSD сначала получает номер
   версии объекта. Так как первичный OSD обычно сам хранит копию или часть объекта, номер
   версии обычно читается из памяти самого OSD. Однако, если ни одна часть обновляемого объекта
   не находится на первичном OSD, для получения номера версии он обращается к одному из вторичных
@@ -115,20 +137,20 @@
   так как метаданные объектов, включая номер версии, все OSD хранят в памяти.
 - Если в пуле используются коды коррекции ошибок, перед частичной записью объекта для вычисления
   чётности зачастую требуется чтение частей объекта с вторичных OSD или с локального диска
-  самого первичного OSD.
+  самого первичного OSD. Это называется процессом "чтение-модификация-запись" (read-modify-write).
-- Также, если в пуле используются коды коррекции ошибок, для закрытия Write Hole применяется
+- Если в пуле используются коды коррекции ошибок, для закрытия Write Hole применяется
   двухфазный алгоритм записи: сначала на все вторичные OSD записывается новая версия частей
   объекта, но при этом старая версия не удаляется, а потом, после получения подтверждения
   успешной записи от всех вторичных OSD, новая версия фиксируется и разрешается удаление старой.
-- Если в кластере не включён режим immediate_commit, то запросы записи, отправляемые клиентами,
+- Если в пуле не включён режим immediate_commit, то запросы записи, отправляемые клиентами,
   не считаются зафиксированными на физических накопителях сразу. Для фиксации данных клиенты
   должны отдельно отправлять запросы SYNC (отдельный от чтения и записи вид запроса),
   а пока такой запрос не отправлен, считается, что записанные данные могут исчезнуть,
   если соответствующий OSD упадёт. Поэтому, когда режим immediate_commit отключён, все
   запросы записи клиенты копируют в памяти и при потере соединения и повторном соединении
-  с OSD повторяют из памяти. Скопированные в память данные удаляются при успешном fsync,
+  с OSD повторяют из памяти. Скопированные в память данные удаляются при успешном SYNC,
   а чтобы хранение этих данных не приводило к чрезмерному потреблению памяти, клиенты
-  автоматически выполняют fsync каждые [client_dirty_limit](../config/network.ru.md#client_dirty_limit)
+  автоматически выполняют SYNC каждые [client_dirty_limit](../config/network.ru.md#client_dirty_limit)
   записанных байт.
 
 ## Схожесть с Ceph

docs/usage/admin.en.md

@@ -171,7 +171,14 @@ to make them use the new version of the client library.
 
 ### 1.7.x to 1.8.0
 
-After upgrading version <= 1.7.x to version >= 1.8.0, BUT <= 1.9.0: restart all clients
+It's recommended to upgrade from version <= 1.7.x to version >= 1.8.0 with full downtime,
+i.e. you should first stop clients and then the cluster (OSDs and monitor), because 1.8.0
+includes a fix for etcd event stream inconsistency which could lead to "incomplete" objects
+appearing in EC pools, and in rare cases, probably, even to data corruption during mass OSD
+restarts. It doesn't mean that you WILL hit this problem if you upgrade without full downtime,
+but it's better to secure yourself against it.
+
+Also, if you upgrade version from <= 1.7.x to version >= 1.8.0, BUT <= 1.9.0: restart all clients
 (VMs and so on), otherwise they will hang when monitor clears old PG configuration key,
 which happens 24 hours after upgrade.
 

docs/usage/admin.ru.md

@@ -168,7 +168,14 @@ done
 
 ### 1.7.x -> 1.8.0
 
-После обновления с версий <= 1.7.x до версий >= 1.8.0, НО <= 1.9.0: перезапустите всех
+Обновляться с версий <= 1.7.x до версий >= 1.8.0 рекомендуется с полной остановкой
+сначала клиентов, а затем кластера, так как в 1.8.0 исправлена проблема (неконсистентность
+потоков событий от etcd), способная приводить к появлению incomplete объектов в EC-пулах
+и, хоть и редко, но даже к повреждению данных при массовых перезапусках OSD. Если вы
+обновляетесь без полной остановки - это не значит, что вы обязательно столкнётесь с этой
+проблемой, но лучше подстраховаться.
+
+Также, если вы обновляетесь с версии <= 1.7.x до версии >= 1.8.0, НО <= 1.9.0: перезапустите всех
 клиентов (процессы виртуальных машин можно перезапустить путём миграции на другой сервер),
 иначе они зависнут, когда монитор удалит старый ключ конфигурации PG, что происходит через
 24 часа после обновления.

docs/usage/disk.en.md

@@ -51,12 +51,16 @@ Options (automatic mode):
 --osd_per_disk <N>
   Create <N> OSDs on each disk (default 1)
 --hybrid
-  Prepare hybrid (HDD+SSD) OSDs using provided devices. SSDs will be used for
+  Prepare hybrid (HDD+SSD, NVMe+SATA or etc) OSDs using provided devices. By default,
-  journals and metadata, HDDs will be used for data. Partitions for journals and
+  any passed SSDs will be used for journals and metadata, HDDs will be used for data,
-  metadata will be created automatically. Whether disks are SSD or HDD is decided
+  but you can override this behaviour with --fast-devices option. Journal and metadata
-  by the `/sys/block/.../queue/rotational` flag. In hybrid mode, default object
+  partitions will be created automatically. In the default mode, SSD and HDD disks
-  size is 1 MB instead of 128 KB, default journal size is 1 GB instead of 32 MB,
+  are distinguished by the `/sys/block/.../queue/rotational` flag. When HDDs are used
-  and throttle_small_writes is enabled by default.
+  for data in hybrid mode, default block_size is 1 MB instead of 128 KB, default journal
+  size is 1 GB instead of 32 MB, and throttle_small_writes is enabled by default.
+--fast-devices /dev/nvmeX,/dev/nvmeY
+  In --hybrid mode, use these devices for journal and metadata instead of auto-detecting
+  and extracting them from the main [devices...] list.
 --disable_data_fsync auto
   Disable data device cache and fsync (1/yes/true = on, default auto)
 --disable_meta_fsync auto

docs/usage/disk.ru.md

@@ -51,12 +51,17 @@ vitastor-disk - инструмент командной строки для уп
 --osd_per_disk <N>
   Создавать по несколько (<N>) OSD на каждом диске (по умолчанию 1)
 --hybrid
-  Инициализировать гибридные (HDD+SSD) OSD на указанных дисках. SSD будут
+  Инициализировать гибридные (HDD+SSD, NVMe+SATA и т.п.) OSD на указанных дисках.
-  использованы для журналов и метаданных, а HDD - для данных. Разделы для журналов
+  По умолчанию, SSD будут использованы для журналов и метаданных, а HDD - для данных,
-  и метаданных будут созданы автоматически. Является ли диск SSD или HDD, определяется
+  но вы можете поменять это поведение опцией --fast-devices. Разделы для журналов
-  по флагу `/sys/block/.../queue/rotational`. В гибридном режиме по умолчанию
+  и метаданных будут созданы автоматически. В режиме по умолчанию SSD и HDD-диски
-  используется размер объекта 1 МБ вместо 128 КБ, размер журнала 1 ГБ вместо 32 МБ
+  различаются по флагу `/sys/block/.../queue/rotational`. Когда в гибридном режиме
-  и включённый throttle_small_writes.
+  для данных используются HDD, по умолчанию размер блока устанавливается 1 МБ вместо
+  128 КБ, размер журнала 1 ГБ вместо 32 МБ, и throttle_small_writes включается по
+  умолчанию.
+--fast-devices /dev/nvmeX,/dev/nvmeY
+  Использовать данные диски для журналов и метаданных в гибридном режиме вместо их
+  автоопределения и извлечения из основного списка [devices...].
 --disable_data_fsync auto
   Отключать кэш и fsync-и для устройств данных. (1/yes/true = да, по умолчанию автоопределение)
 --disable_meta_fsync auto

mon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor-mon",
-  "version": "1.9.2",
+  "version": "1.9.3",
   "description": "Vitastor SDS monitor service",
   "main": "mon-main.js",
   "scripts": {

patches/cinder-vitastor.py

@@ -50,7 +50,7 @@ from cinder.volume import configuration
 from cinder.volume import driver
 from cinder.volume import volume_utils
 
-VITASTOR_VERSION = '1.9.2'
+VITASTOR_VERSION = '1.9.3'
 
 LOG = logging.getLogger(__name__)
 

patches/libvirt-8.0-vitastor.diff

@@ -306,7 +306,7 @@ index e5ff653a60..884ecc79ea 100644
 +        etcd = virBufferContentAndReset(&buf);
 +    }
 +
-+    if (virJSONValueObjectCreate(&ret,
++    if (virJSONValueObjectAdd(&ret,
 +                              "S:etcd-host", etcd,
 +                              "S:etcd-prefix", src->query,
 +                              "S:config-path", src->configFile,

patches/pve-qemu-9.0-vitastor.patch

@@ -0,0 +1,193 @@
+Index: pve-qemu-kvm-9.0.0/block/meson.build
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/block/meson.build
++++ pve-qemu-kvm-9.0.0/block/meson.build
+@@ -126,6 +126,7 @@ foreach m : [
+   [libnfs, 'nfs', files('nfs.c')],
+   [libssh, 'ssh', files('ssh.c')],
+   [rbd, 'rbd', files('rbd.c')],
++  [vitastor, 'vitastor', files('vitastor.c')],
+ ]
+   if m[0].found()
+     module_ss = ss.source_set()
+Index: pve-qemu-kvm-9.0.0/meson.build
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/meson.build
++++ pve-qemu-kvm-9.0.0/meson.build
+@@ -1452,6 +1452,26 @@ if not get_option('rbd').auto() or have_
+   endif
+ endif
+ 
++vitastor = not_found
++if not get_option('vitastor').auto() or have_block
++  libvitastor_client = cc.find_library('vitastor_client', has_headers: ['vitastor_c.h'],
++    required: get_option('vitastor'))
++  if libvitastor_client.found()
++    if cc.links('''
++      #include <vitastor_c.h>
++      int main(void) {
++        vitastor_c_create_qemu(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
++        return 0;
++      }''', dependencies: libvitastor_client)
++      vitastor = declare_dependency(dependencies: libvitastor_client)
++    elif get_option('vitastor').enabled()
++      error('could not link libvitastor_client')
++    else
++      warning('could not link libvitastor_client, disabling')
++    endif
++  endif
++endif
++
+ glusterfs = not_found
+ glusterfs_ftruncate_has_stat = false
+ glusterfs_iocb_has_stat = false
+@@ -2254,6 +2274,7 @@ endif
+ config_host_data.set('CONFIG_OPENGL', opengl.found())
+ config_host_data.set('CONFIG_PLUGIN', get_option('plugins'))
+ config_host_data.set('CONFIG_RBD', rbd.found())
++config_host_data.set('CONFIG_VITASTOR', vitastor.found())
+ config_host_data.set('CONFIG_RDMA', rdma.found())
+ config_host_data.set('CONFIG_RELOCATABLE', get_option('relocatable'))
+ config_host_data.set('CONFIG_SAFESTACK', get_option('safe_stack'))
+@@ -4454,6 +4475,7 @@ summary_info += {'fdt support':       fd
+ summary_info += {'libcap-ng support': libcap_ng}
+ summary_info += {'bpf support':       libbpf}
+ summary_info += {'rbd support':       rbd}
++summary_info += {'vitastor support':  vitastor}
+ summary_info += {'smartcard support': cacard}
+ summary_info += {'U2F support':       u2f}
+ summary_info += {'libusb':            libusb}
+Index: pve-qemu-kvm-9.0.0/meson_options.txt
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/meson_options.txt
++++ pve-qemu-kvm-9.0.0/meson_options.txt
+@@ -194,6 +194,8 @@ option('lzo', type : 'feature', value :
+        description: 'lzo compression support')
+ option('rbd', type : 'feature', value : 'auto',
+        description: 'Ceph block device driver')
++option('vitastor', type : 'feature', value : 'auto',
++       description: 'Vitastor block device driver')
+ option('opengl', type : 'feature', value : 'auto',
+        description: 'OpenGL support')
+ option('rdma', type : 'feature', value : 'auto',
+Index: pve-qemu-kvm-9.0.0/qapi/block-core.json
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/qapi/block-core.json
++++ pve-qemu-kvm-9.0.0/qapi/block-core.json
+@@ -3481,7 +3481,7 @@
+             'raw', 'rbd',
+             { 'name': 'replication', 'if': 'CONFIG_REPLICATION' },
+             'pbs',
+-            'ssh', 'throttle', 'vdi', 'vhdx',
++            'ssh', 'throttle', 'vdi', 'vhdx', 'vitastor',
+             { 'name': 'virtio-blk-vfio-pci', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-user', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-vdpa', 'if': 'CONFIG_BLKIO' },
+@@ -4591,6 +4591,28 @@
+             '*server': ['InetSocketAddressBase'] } }
+ 
+ ##
++# @BlockdevOptionsVitastor:
++#
++# Driver specific block device options for vitastor
++#
++# @image:       Image name
++# @inode:       Inode number
++# @pool:        Pool ID
++# @size:        Desired image size in bytes
++# @config-path: Path to Vitastor configuration
++# @etcd-host:   etcd connection address(es)
++# @etcd-prefix: etcd key/value prefix
++##
++{ 'struct': 'BlockdevOptionsVitastor',
++  'data': { '*inode': 'uint64',
++            '*pool': 'uint64',
++            '*size': 'uint64',
++            '*image': 'str',
++            '*config-path': 'str',
++            '*etcd-host': 'str',
++            '*etcd-prefix': 'str' } }
++
++##
+ # @ReplicationMode:
+ #
+ # An enumeration of replication modes.
+@@ -5053,6 +5075,7 @@
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+       'vhdx':       'BlockdevOptionsGenericFormat',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'virtio-blk-vfio-pci':
+                     { 'type': 'BlockdevOptionsVirtioBlkVfioPci',
+                       'if': 'CONFIG_BLKIO' },
+@@ -5498,6 +5521,20 @@
+             '*encrypt' :        'RbdEncryptionCreateOptions' } }
+ 
+ ##
++# @BlockdevCreateOptionsVitastor:
++#
++# Driver specific image creation options for Vitastor.
++#
++# @location: Where to store the new image file.  This location cannot
++#     point to a snapshot.
++#
++# @size: Size of the virtual disk in bytes
++##
++{ 'struct': 'BlockdevCreateOptionsVitastor',
++  'data': { 'location':         'BlockdevOptionsVitastor',
++            'size':             'size' } }
++
++##
+ # @BlockdevVmdkSubformat:
+ #
+ # Subformat options for VMDK images
+@@ -5719,6 +5753,7 @@
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'vdi':            'BlockdevCreateOptionsVdi',
+       'vhdx':           'BlockdevCreateOptionsVhdx',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'vmdk':           'BlockdevCreateOptionsVmdk',
+       'vpc':            'BlockdevCreateOptionsVpc'
+   } }
+Index: pve-qemu-kvm-9.0.0/scripts/ci/org.centos/stream/8/x86_64/configure
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/scripts/ci/org.centos/stream/8/x86_64/configure
++++ pve-qemu-kvm-9.0.0/scripts/ci/org.centos/stream/8/x86_64/configure
+@@ -30,7 +30,7 @@
+ --with-suffix="qemu-kvm" \
+ --firmwarepath=/usr/share/qemu-firmware \
+ --target-list="x86_64-softmmu" \
+---block-drv-rw-whitelist="qcow2,raw,file,host_device,nbd,iscsi,rbd,blkdebug,luks,null-co,nvme,copy-on-read,throttle,gluster" \
++--block-drv-rw-whitelist="qcow2,raw,file,host_device,nbd,iscsi,rbd,vitastor,blkdebug,luks,null-co,nvme,copy-on-read,throttle,gluster" \
+ --audio-drv-list="" \
+ --block-drv-ro-whitelist="vmdk,vhdx,vpc,https,ssh" \
+ --with-coroutine=ucontext \
+@@ -176,6 +176,7 @@
+ --enable-opengl \
+ --enable-pie \
+ --enable-rbd \
++--enable-vitastor \
+ --enable-rdma \
+ --enable-seccomp \
+ --enable-snappy \
+Index: pve-qemu-kvm-9.0.0/scripts/meson-buildoptions.sh
+===================================================================
+--- pve-qemu-kvm-9.0.0.orig/scripts/meson-buildoptions.sh
++++ pve-qemu-kvm-9.0.0/scripts/meson-buildoptions.sh
+@@ -168,6 +168,7 @@ meson_options_help() {
+   printf "%s\n" '  qed             qed image format support'
+   printf "%s\n" '  qga-vss         build QGA VSS support (broken with MinGW)'
+   printf "%s\n" '  rbd             Ceph block device driver'
++  printf "%s\n" '  vitastor        Vitastor block device driver'
+   printf "%s\n" '  rdma            Enable RDMA-based migration'
+   printf "%s\n" '  replication     replication support'
+   printf "%s\n" '  rutabaga-gfx    rutabaga_gfx support'
+@@ -445,6 +446,8 @@ _meson_option_parse() {
+     --disable-qom-cast-debug) printf "%s" -Dqom_cast_debug=false ;;
+     --enable-rbd) printf "%s" -Drbd=enabled ;;
+     --disable-rbd) printf "%s" -Drbd=disabled ;;
++    --enable-vitastor) printf "%s" -Dvitastor=enabled ;;
++    --disable-vitastor) printf "%s" -Dvitastor=disabled ;;
+     --enable-rdma) printf "%s" -Drdma=enabled ;;
+     --disable-rdma) printf "%s" -Drdma=disabled ;;
+     --enable-relocatable) printf "%s" -Drelocatable=true ;;

patches/qemu-9.1-vitastor.patch

@@ -0,0 +1,172 @@
+diff --git a/block/meson.build b/block/meson.build
+index f1262ec2ba..3cf3e23f16 100644
+--- a/block/meson.build
++++ b/block/meson.build
+@@ -114,6 +114,7 @@ foreach m : [
+   [libnfs, 'nfs', files('nfs.c')],
+   [libssh, 'ssh', files('ssh.c')],
+   [rbd, 'rbd', files('rbd.c')],
++  [vitastor, 'vitastor', files('vitastor.c')],
+ ]
+   if m[0].found()
+     module_ss = ss.source_set()
+diff --git a/meson.build b/meson.build
+index fbda17c987..3edac22aff 100644
+--- a/meson.build
++++ b/meson.build
+@@ -1510,6 +1510,26 @@ if not get_option('rbd').auto() or have_block
+   endif
+ endif
+ 
++vitastor = not_found
++if not get_option('vitastor').auto() or have_block
++  libvitastor_client = cc.find_library('vitastor_client', has_headers: ['vitastor_c.h'],
++    required: get_option('vitastor'))
++  if libvitastor_client.found()
++    if cc.links('''
++      #include <vitastor_c.h>
++      int main(void) {
++        vitastor_c_create_qemu(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
++        return 0;
++      }''', dependencies: libvitastor_client)
++      vitastor = declare_dependency(dependencies: libvitastor_client)
++    elif get_option('vitastor').enabled()
++      error('could not link libvitastor_client')
++    else
++      warning('could not link libvitastor_client, disabling')
++    endif
++  endif
++endif
++
+ glusterfs = not_found
+ glusterfs_ftruncate_has_stat = false
+ glusterfs_iocb_has_stat = false
+@@ -2351,6 +2371,7 @@ endif
+ config_host_data.set('CONFIG_OPENGL', opengl.found())
+ config_host_data.set('CONFIG_PLUGIN', get_option('plugins'))
+ config_host_data.set('CONFIG_RBD', rbd.found())
++config_host_data.set('CONFIG_VITASTOR', vitastor.found())
+ config_host_data.set('CONFIG_RDMA', rdma.found())
+ config_host_data.set('CONFIG_RELOCATABLE', get_option('relocatable'))
+ config_host_data.set('CONFIG_SAFESTACK', get_option('safe_stack'))
+@@ -4510,6 +4531,7 @@ summary_info += {'fdt support':       fdt_opt == 'internal' ? 'internal' : fdt}
+ summary_info += {'libcap-ng support': libcap_ng}
+ summary_info += {'bpf support':       libbpf}
+ summary_info += {'rbd support':       rbd}
++summary_info += {'vitastor support':  vitastor}
+ summary_info += {'smartcard support': cacard}
+ summary_info += {'U2F support':       u2f}
+ summary_info += {'libusb':            libusb}
+diff --git a/meson_options.txt b/meson_options.txt
+index 0269fa0f16..4740ffdc27 100644
+--- a/meson_options.txt
++++ b/meson_options.txt
+@@ -194,6 +194,8 @@ option('lzo', type : 'feature', value : 'auto',
+        description: 'lzo compression support')
+ option('rbd', type : 'feature', value : 'auto',
+        description: 'Ceph block device driver')
++option('vitastor', type : 'feature', value : 'auto',
++       description: 'Vitastor block device driver')
+ option('opengl', type : 'feature', value : 'auto',
+        description: 'OpenGL support')
+ option('rdma', type : 'feature', value : 'auto',
+diff --git a/qapi/block-core.json b/qapi/block-core.json
+index aa40d44f1d..bbee6a0e9c 100644
+--- a/qapi/block-core.json
++++ b/qapi/block-core.json
+@@ -3203,7 +3203,7 @@
+             'parallels', 'preallocate', 'qcow', 'qcow2', 'qed', 'quorum',
+             'raw', 'rbd',
+             { 'name': 'replication', 'if': 'CONFIG_REPLICATION' },
+-            'ssh', 'throttle', 'vdi', 'vhdx',
++            'ssh', 'throttle', 'vdi', 'vhdx', 'vitastor',
+             { 'name': 'virtio-blk-vfio-pci', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-user', 'if': 'CONFIG_BLKIO' },
+             { 'name': 'virtio-blk-vhost-vdpa', 'if': 'CONFIG_BLKIO' },
+@@ -4286,6 +4286,28 @@
+             '*key-secret': 'str',
+             '*server': ['InetSocketAddressBase'] } }
+ 
++##
++# @BlockdevOptionsVitastor:
++#
++# Driver specific block device options for vitastor
++#
++# @image:       Image name
++# @inode:       Inode number
++# @pool:        Pool ID
++# @size:        Desired image size in bytes
++# @config-path: Path to Vitastor configuration
++# @etcd-host:   etcd connection address(es)
++# @etcd-prefix: etcd key/value prefix
++##
++{ 'struct': 'BlockdevOptionsVitastor',
++  'data': { '*inode': 'uint64',
++            '*pool': 'uint64',
++            '*size': 'uint64',
++            '*image': 'str',
++            '*config-path': 'str',
++            '*etcd-host': 'str',
++            '*etcd-prefix': 'str' } }
++
+ ##
+ # @ReplicationMode:
+ #
+@@ -4742,6 +4764,7 @@
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+       'vhdx':       'BlockdevOptionsGenericFormat',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'virtio-blk-vfio-pci':
+                     { 'type': 'BlockdevOptionsVirtioBlkVfioPci',
+                       'if': 'CONFIG_BLKIO' },
+@@ -5183,6 +5206,20 @@
+             '*cluster-size' :   'size',
+             '*encrypt' :        'RbdEncryptionCreateOptions' } }
+ 
++##
++# @BlockdevCreateOptionsVitastor:
++#
++# Driver specific image creation options for Vitastor.
++#
++# @location: Where to store the new image file.  This location cannot
++#     point to a snapshot.
++#
++# @size: Size of the virtual disk in bytes
++##
++{ 'struct': 'BlockdevCreateOptionsVitastor',
++  'data': { 'location':         'BlockdevOptionsVitastor',
++            'size':             'size' } }
++
+ ##
+ # @BlockdevVmdkSubformat:
+ #
+@@ -5405,6 +5442,7 @@
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'vdi':            'BlockdevCreateOptionsVdi',
+       'vhdx':           'BlockdevCreateOptionsVhdx',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'vmdk':           'BlockdevCreateOptionsVmdk',
+       'vpc':            'BlockdevCreateOptionsVpc'
+   } }
+diff --git a/scripts/meson-buildoptions.sh b/scripts/meson-buildoptions.sh
+index c97079a38c..4623f552ec 100644
+--- a/scripts/meson-buildoptions.sh
++++ b/scripts/meson-buildoptions.sh
+@@ -168,6 +168,7 @@ meson_options_help() {
+   printf "%s\n" '  qga-vss         build QGA VSS support (broken with MinGW)'
+   printf "%s\n" '  qpl             Query Processing Library support'
+   printf "%s\n" '  rbd             Ceph block device driver'
++  printf "%s\n" '  vitastor        Vitastor block device driver'
+   printf "%s\n" '  rdma            Enable RDMA-based migration'
+   printf "%s\n" '  replication     replication support'
+   printf "%s\n" '  rutabaga-gfx    rutabaga_gfx support'
+@@ -444,6 +445,8 @@ _meson_option_parse() {
+     --disable-qpl) printf "%s" -Dqpl=disabled ;;
+     --enable-rbd) printf "%s" -Drbd=enabled ;;
+     --disable-rbd) printf "%s" -Drbd=disabled ;;
++    --enable-vitastor) printf "%s" -Dvitastor=enabled ;;
++    --disable-vitastor) printf "%s" -Dvitastor=disabled ;;
+     --enable-rdma) printf "%s" -Drdma=enabled ;;
+     --disable-rdma) printf "%s" -Drdma=disabled ;;
+     --enable-relocatable) printf "%s" -Drelocatable=true ;;

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.9.2
+Version:        1.9.3
 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-1.9.2.el7.tar.gz
+Source0:        vitastor-1.9.3.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.9.2
+Version:        1.9.3
 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-1.9.2.el8.tar.gz
+Source0:        vitastor-1.9.3.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el9.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.9.2
+Version:        1.9.3
 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-1.9.2.el9.tar.gz
+Source0:        vitastor-1.9.3.el9.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -19,7 +19,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVITASTOR_VERSION="1.9.2")
+add_definitions(-DVITASTOR_VERSION="1.9.3")
 add_definitions(-D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64 -Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -fno-omit-frame-pointer -I ${CMAKE_SOURCE_DIR}/src)
 add_link_options(-fno-omit-frame-pointer)
 if (${WITH_ASAN})

src/client/etcd_state_client.cpp

@@ -176,7 +176,7 @@ void etcd_state_client_t::add_etcd_url(std::string addr)
             exit(1);
         }
         if (!local_ips.size())
-            local_ips = getifaddr_list();
+            local_ips = getifaddr_list(std::vector<std::string>(), true);
         std::string check_addr;
         int pos = addr.find('/');
         int pos2 = addr.find(':');

src/client/http_client.cpp

@@ -62,6 +62,7 @@ struct http_co_t
     inline void end() { ended = true; if (!onstack) { delete this; } }
     void run_cb_and_clear();
     void start_connection();
+    void start_ws_connection();
     void close_connection();
     void next_request();
     void handle_events();
@@ -112,7 +113,7 @@ http_co_t* open_websocket(timerfd_manager_t *tfd, const std::string & host, cons
     handler->keepalive = false;
     handler->request = request;
     handler->response_callback = response_callback;
-    handler->start_connection();
+    handler->start_ws_connection();
     return handler;
 }
 
@@ -282,6 +283,27 @@ void http_co_t::close_connection()
     epoll_events = 0;
 }
 
+void http_co_t::start_ws_connection()
+{
+    stackin();
+    start_connection();
+    if (request_timeout > 0)
+    {
+        timeout_id = tfd->set_timer(request_timeout, false, [this](int timer_id)
+        {
+            stackin();
+            if (state != HTTP_CO_WEBSOCKET)
+            {
+                close_connection();
+                parsed = { .error = "Websocket connection timed out" };
+                run_cb_and_clear();
+            }
+            stackout();
+        });
+    }
+    stackout();
+}
+
 void http_co_t::start_connection()
 {
     stackin();

src/client/messenger.cpp

@@ -121,7 +121,7 @@ void osd_messenger_t::init()
     if (use_rdma)
     {
         rdma_context = msgr_rdma_context_t::create(
-            rdma_device != "" ? rdma_device.c_str() : NULL,
+            osd_networks, rdma_device != "" ? rdma_device.c_str() : NULL,
             rdma_port_num, rdma_gid_index, rdma_mtu, rdma_odp, log_level
         );
         if (!rdma_context)
@@ -266,6 +266,7 @@ void osd_messenger_t::parse_config(const json11::Json & config)
     this->rdma_port_num = (uint8_t)config["rdma_port_num"].uint64_value();
     if (!this->rdma_port_num)
         this->rdma_port_num = 1;
+    if (!config["rdma_gid_index"].is_null())
         this->rdma_gid_index = (uint8_t)config["rdma_gid_index"].uint64_value();
     this->rdma_mtu = (uint32_t)config["rdma_mtu"].uint64_value();
     this->rdma_max_sge = config["rdma_max_sge"].uint64_value();
@@ -281,6 +282,15 @@ void osd_messenger_t::parse_config(const json11::Json & config)
     if (!this->rdma_max_msg || this->rdma_max_msg > 128*1024*1024)
         this->rdma_max_msg = 129*1024;
     this->rdma_odp = config["rdma_odp"].bool_value();
+    std::vector<std::string> mask;
+    if (config["bind_address"].is_string())
+        mask.push_back(config["bind_address"].string_value());
+    else if (config["osd_network"].is_string())
+        mask.push_back(config["osd_network"].string_value());
+    else
+        for (auto v: config["osd_network"].array_items())
+            mask.push_back(v.string_value());
+    this->osd_networks = mask;
 #endif
     if (!osd_num)
         this->iothread_count = (uint32_t)config["client_iothread_count"].uint64_value();

src/client/messenger.h

@@ -165,8 +165,9 @@ protected:
 
 #ifdef WITH_RDMA
     bool use_rdma = true;
+    std::vector<std::string> osd_networks;
     std::string rdma_device;
-    uint64_t rdma_port_num = 1, rdma_gid_index = 0, rdma_mtu = 0;
+    uint64_t rdma_port_num = 1, rdma_gid_index = -1, rdma_mtu = 0;
     msgr_rdma_context_t *rdma_context = NULL;
     uint64_t rdma_max_sge = 0, rdma_max_send = 0, rdma_max_recv = 0;
     uint64_t rdma_max_msg = 0;
@@ -177,7 +178,7 @@ protected:
     std::vector<int> read_ready_clients;
     std::vector<int> write_ready_clients;
     // We don't use ringloop->set_immediate here because we may have no ringloop in client :)
-    std::vector<std::function<void()>> set_immediate;
+    std::vector<osd_op_t*> set_immediate_ops;
 
 public:
     timerfd_manager_t *tfd;
@@ -237,6 +238,8 @@ protected:
     void handle_op_hdr(osd_client_t *cl);
     bool handle_reply_hdr(osd_client_t *cl);
     void handle_reply_ready(osd_op_t *op);
+    void handle_immediate_ops();
+    void clear_immediate_ops(int peer_fd);
 
 #ifdef WITH_RDMA
     void try_send_rdma(osd_client_t *cl);

src/client/msgr_rdma.cpp

@@ -3,6 +3,7 @@
 
 #include <stdio.h>
 #include <stdlib.h>
+#include "addr_util.h"
 #include "msgr_rdma.h"
 #include "messenger.h"
 
@@ -69,7 +70,126 @@ msgr_rdma_connection_t::~msgr_rdma_connection_t()
     send_out_size = 0;
 }
 
-msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level)
+static bool is_ipv4_gid(ibv_gid_entry *gidx)
+{
+    return (((uint64_t*)gidx->gid.raw)[0] == 0 &&
+        ((uint32_t*)gidx->gid.raw)[2] == 0xffff0000);
+}
+
+static bool match_gid(ibv_gid_entry *gidx, addr_mask_t *networks, int nnet)
+{
+    if (gidx->gid_type != IBV_GID_TYPE_ROCE_V1 &&
+        gidx->gid_type != IBV_GID_TYPE_ROCE_V2 ||
+        ((uint64_t*)gidx->gid.raw)[0] == 0 &&
+        ((uint64_t*)gidx->gid.raw)[1] == 0)
+    {
+        return false;
+    }
+    if (is_ipv4_gid(gidx))
+    {
+        for (int i = 0; i < nnet; i++)
+        {
+            if (networks[i].family == AF_INET && cidr_match(*(in_addr*)(gidx->gid.raw+12), networks[i].ipv4, networks[i].bits))
+                return true;
+        }
+    }
+    else
+    {
+        for (int i = 0; i < nnet; i++)
+        {
+            if (networks[i].family == AF_INET6 && cidr6_match(*(in6_addr*)gidx->gid.raw, networks[i].ipv6, networks[i].bits))
+                return true;
+        }
+    }
+    return false;
+}
+
+struct matched_dev
+{
+    int dev = -1;
+    int port = -1;
+    int gid = -1;
+    bool rocev2 = false;
+};
+
+static void log_rdma_dev_port_gid(ibv_device *dev, int ib_port, int gid_index, ibv_gid_entry & gidx)
+{
+    bool is4 = ((uint64_t*)gidx.gid.raw)[0] == 0 && ((uint32_t*)gidx.gid.raw)[2] == 0xffff0000;
+    char buf[256];
+    inet_ntop(is4 ? AF_INET : AF_INET6, is4 ? gidx.gid.raw+12 : gidx.gid.raw, buf, sizeof(buf));
+    fprintf(
+        stderr, "Auto-selected RDMA device %s port %d GID %d - ROCEv%d IPv%d %s\n",
+        ibv_get_device_name(dev), ib_port, gid_index,
+        gidx.gid_type == IBV_GID_TYPE_ROCE_V2 ? 2 : 1, is4 ? 4 : 6, buf
+    );
+}
+
+static matched_dev match_device(ibv_device **dev_list, addr_mask_t *networks, int nnet, int log_level)
+{
+    matched_dev best;
+    ibv_device_attr attr;
+    ibv_port_attr portinfo;
+    ibv_gid_entry best_gidx;
+    int res;
+    for (int i = 0; dev_list[i]; ++i)
+    {
+        auto dev = dev_list[i];
+        ibv_context *context = ibv_open_device(dev_list[i]);
+        if ((res = ibv_query_device(context, &attr)) != 0)
+        {
+            fprintf(stderr, "Couldn't query RDMA device %s for its features: %s\n", ibv_get_device_name(dev_list[i]), strerror(res));
+            goto cleanup;
+        }
+        for (int j = 1; j <= attr.phys_port_cnt; j++)
+        {
+            // Try to find a port with matching address
+            if ((res = ibv_query_port(context, j, &portinfo)) != 0)
+            {
+                fprintf(stderr, "Couldn't get RDMA device %s port %d info: %s\n", ibv_get_device_name(dev), j, strerror(res));
+                goto cleanup;
+            }
+            for (int k = 0; k < portinfo.gid_tbl_len; k++)
+            {
+                ibv_gid_entry gidx;
+                if ((res = ibv_query_gid_ex(context, j, k, &gidx, 0)) != 0)
+                {
+                    if (res != ENODATA)
+                    {
+                        fprintf(stderr, "Couldn't read RDMA device %s GID index %d: %s\n", ibv_get_device_name(dev), k, strerror(res));
+                        goto cleanup;
+                    }
+                    else
+                        break;
+                }
+                if (match_gid(&gidx, networks, nnet))
+                {
+                    // Prefer RoCEv2
+                    if (!best.rocev2)
+                    {
+                        best.dev = i;
+                        best.port = j;
+                        best.gid = k;
+                        best.rocev2 = (gidx.gid_type == IBV_GID_TYPE_ROCE_V2);
+                        best_gidx = gidx;
+                    }
+                }
+            }
+        }
+cleanup:
+        ibv_close_device(context);
+        if (best.rocev2)
+        {
+            break;
+        }
+    }
+    if (best.dev >= 0 && log_level > 0)
+    {
+        log_rdma_dev_port_gid(dev_list[best.dev], best.port, best.gid, best_gidx);
+    }
+    return best;
+}
+
+msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_networks, const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level)
 {
     int res;
     ibv_device **dev_list = NULL;
@@ -80,28 +200,23 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t
     clock_gettime(CLOCK_REALTIME, &tv);
     srand48(tv.tv_sec*1000000000 + tv.tv_nsec);
     dev_list = ibv_get_device_list(NULL);
-    if (!dev_list)
+    if (!dev_list || !*dev_list)
     {
         if (errno == -ENOSYS || errno == ENOSYS)
         {
             if (log_level > 0)
                 fprintf(stderr, "No RDMA devices found (RDMA device list returned ENOSYS)\n");
         }
+        else if (!*dev_list)
+        {
+            if (log_level > 0)
+                fprintf(stderr, "No RDMA devices found\n");
+        }
         else
             fprintf(stderr, "Failed to get RDMA device list: %s\n", strerror(errno));
         goto cleanup;
     }
-    if (!ib_devname)
+    if (ib_devname)
-    {
-        ctx->dev = *dev_list;
-        if (!ctx->dev)
-        {
-            if (log_level > 0)
-                fprintf(stderr, "No RDMA devices found\n");
-            goto cleanup;
-        }
-    }
-    else
     {
         int i;
         for (i = 0; dev_list[i]; ++i)
@@ -114,6 +229,31 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t
             goto cleanup;
         }
     }
+    else if (osd_networks.size())
+    {
+        std::vector<addr_mask_t> nets;
+        for (auto & netstr: osd_networks)
+        {
+            nets.push_back(cidr_parse(netstr));
+        }
+        auto best = match_device(dev_list, nets.data(), nets.size(), log_level);
+        if (best.dev < 0)
+        {
+            if (log_level > 0)
+                fprintf(stderr, "RDMA device matching osd_network is not found, using first available device\n");
+            best.dev = 0;
+        }
+        else
+        {
+            ib_port = best.port;
+            gid_index = best.gid;
+        }
+        ctx->dev = dev_list[best.dev];
+    }
+    else
+    {
+        ctx->dev = *dev_list;
+    }
 
     ctx->context = ibv_open_device(ctx->dev);
     if (!ctx->context)
@@ -123,7 +263,6 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t
     }
 
     ctx->ib_port = ib_port;
-    ctx->gid_index = gid_index;
     if ((res = ibv_query_port(ctx->context, ib_port, &ctx->portinfo)) != 0)
     {
         fprintf(stderr, "Couldn't get RDMA device %s port %d info: %s\n", ibv_get_device_name(ctx->dev), ib_port, strerror(res));
@@ -135,11 +274,48 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t
         fprintf(stderr, "RDMA device %s must have local LID because it's not Ethernet, but LID is zero\n", ibv_get_device_name(ctx->dev));
         goto cleanup;
     }
-    if (ibv_query_gid(ctx->context, ib_port, gid_index, &ctx->my_gid))
+
+    if (gid_index != -1)
+    {
+        ctx->gid_index = gid_index;
+        if (ibv_query_gid_ex(ctx->context, ib_port, gid_index, &ctx->my_gid, 0))
         {
             fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), gid_index);
             goto cleanup;
         }
+    }
+    else
+    {
+        // Auto-guess GID
+        for (int k = 0; k < ctx->portinfo.gid_tbl_len; k++)
+        {
+            ibv_gid_entry gidx;
+            if (ibv_query_gid_ex(ctx->context, ib_port, k, &gidx, 0) != 0)
+            {
+                fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), k);
+                goto cleanup;
+            }
+            // Skip empty GID
+            if (((uint64_t*)gidx.gid.raw)[0] == 0 &&
+                ((uint64_t*)gidx.gid.raw)[1] == 0)
+            {
+                continue;
+            }
+            // Prefer IPv4 RoCEv2 GID by default
+            if (gid_index == -1 ||
+                gidx.gid_type == IBV_GID_TYPE_ROCE_V2 &&
+                (ctx->my_gid.gid_type != IBV_GID_TYPE_ROCE_V2 || is_ipv4_gid(&gidx)))
+            {
+                gid_index = k;
+                ctx->my_gid = gidx;
+            }
+        }
+        ctx->gid_index = gid_index = (gid_index == -1 ? 0 : gid_index);
+        if (log_level > 0)
+        {
+            log_rdma_dev_port_gid(ctx->dev, ctx->ib_port, ctx->gid_index, ctx->my_gid);
+        }
+    }
 
     ctx->pd = ibv_alloc_pd(ctx->context);
     if (!ctx->pd)
@@ -255,7 +431,7 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
     }
 
     conn->addr.lid = ctx->my_lid;
-    conn->addr.gid = ctx->my_gid;
+    conn->addr.gid = ctx->my_gid.gid;
     conn->addr.qpn = conn->qp->qp_num;
     conn->addr.psn = lrand48() & 0xffffff;
 
@@ -598,6 +774,7 @@ void osd_messenger_t::handle_rdma_events()
                 }
                 fprintf(stderr, " with status: %s, stopping client\n", ibv_wc_status_str(wc[i].status));
                 stop_client(client_id);
+                clear_immediate_ops(client_id);
                 continue;
             }
             if (!is_send)
@@ -606,6 +783,7 @@ void osd_messenger_t::handle_rdma_events()
                 if (!handle_read_buffer(cl, rc->recv_buffers[rc->next_recv_buf].buf, wc[i].byte_len))
                 {
                     // handle_read_buffer may stop the client
+                    clear_immediate_ops(client_id);
                     continue;
                 }
                 try_recv_rdma_wr(cl, rc->recv_buffers[rc->next_recv_buf]);
@@ -666,9 +844,5 @@ void osd_messenger_t::handle_rdma_events()
             }
         }
     } while (event_count > 0);
-    for (auto cb: set_immediate)
+    handle_immediate_ops();
-    {
-        cb();
-    }
-    set_immediate.clear();
 }

src/client/msgr_rdma.h

@@ -31,12 +31,12 @@ struct msgr_rdma_context_t
     uint8_t ib_port;
     uint8_t gid_index;
     uint16_t my_lid;
-    ibv_gid my_gid;
+    ibv_gid_entry my_gid;
     uint32_t mtu;
     int max_cqe = 0;
     int used_max_cqe = 0;
 
-    static msgr_rdma_context_t *create(const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level);
+    static msgr_rdma_context_t *create(std::vector<std::string> osd_networks, const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level);
     ~msgr_rdma_context_t();
 };
 

src/client/msgr_receive.cpp

@@ -65,6 +65,7 @@ void osd_messenger_t::read_requests()
 bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
 {
     bool ret = false;
+    int peer_fd = cl->peer_fd;
     cl->read_msg.msg_iovlen = 0;
     cl->refs--;
     if (cl->peer_state == PEER_STOPPED)
@@ -101,7 +102,8 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
         {
             if (!handle_read_buffer(cl, cl->in_buf, result))
             {
-                goto fin;
+                clear_immediate_ops(peer_fd);
+                return false;
             }
         }
         else
@@ -113,7 +115,8 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
             {
                 if (!handle_finished_read(cl))
                 {
-                    goto fin;
+                    clear_immediate_ops(peer_fd);
+                    return false;
                 }
             }
         }
@@ -122,15 +125,47 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
             ret = true;
         }
     }
-fin:
+    handle_immediate_ops();
-    for (auto cb: set_immediate)
-    {
-        cb();
-    }
-    set_immediate.clear();
     return ret;
 }
 
+void osd_messenger_t::clear_immediate_ops(int peer_fd)
+{
+    size_t i = 0, j = 0;
+    while (i < set_immediate_ops.size())
+    {
+        if (set_immediate_ops[i]->peer_fd == peer_fd)
+        {
+            delete set_immediate_ops[i];
+        }
+        else
+        {
+            if (i != j)
+                set_immediate_ops[j] = set_immediate_ops[i];
+            j++;
+        }
+        i++;
+    }
+    set_immediate_ops.resize(j);
+}
+
+void osd_messenger_t::handle_immediate_ops()
+{
+    for (auto op: set_immediate_ops)
+    {
+        if (op->op_type == OSD_OP_IN)
+        {
+            exec_op(op);
+        }
+        else
+        {
+            // Copy lambda to be unaffected by `delete op`
+            std::function<void(osd_op_t*)>(op->callback)(op);
+        }
+    }
+    set_immediate_ops.clear();
+}
+
 bool osd_messenger_t::handle_read_buffer(osd_client_t *cl, void *curbuf, int remain)
 {
     // Compose operation(s) from the buffer
@@ -199,7 +234,7 @@ bool osd_messenger_t::handle_finished_read(osd_client_t *cl)
     {
         // Operation is ready
         cl->received_ops.push_back(cl->read_op);
-        set_immediate.push_back([this, op = cl->read_op]() { exec_op(op); });
+        set_immediate_ops.push_back(cl->read_op);
         cl->read_op = NULL;
         cl->read_state = 0;
     }
@@ -295,7 +330,7 @@ void osd_messenger_t::handle_op_hdr(osd_client_t *cl)
     {
         // Operation is ready
         cl->received_ops.push_back(cur_op);
-        set_immediate.push_back([this, cur_op]() { exec_op(cur_op); });
+        set_immediate_ops.push_back(cur_op);
         cl->read_op = NULL;
         cl->read_state = 0;
     }
@@ -416,9 +451,5 @@ void osd_messenger_t::handle_reply_ready(osd_op_t *op)
         (tv_end.tv_sec - op->tv_begin.tv_sec)*1000000 +
         (tv_end.tv_nsec - op->tv_begin.tv_nsec)/1000
     );
-    set_immediate.push_back([op]()
+    set_immediate_ops.push_back(op);
-    {
-        // Copy lambda to be unaffected by `delete op`
-        std::function<void(osd_op_t*)>(op->callback)(op);
-    });
 }

src/client/qemu_driver.c

@@ -16,7 +16,6 @@
 #include "qapi/error.h"
 #include "qapi/qmp/qdict.h"
 #include "qapi/qmp/qerror.h"
-#include "qemu/uri.h"
 #include "qemu/error-report.h"
 #include "qemu/module.h"
 #include "qemu/option.h"
@@ -1021,7 +1020,11 @@ static BlockDriver bdrv_vitastor = {
     // FIXME: Implement it along with per-inode statistics
     //.bdrv_get_allocated_file_size   = vitastor_get_allocated_file_size,
 
+#if QEMU_VERSION_MAJOR > 9 || QEMU_VERSION_MAJOR == 9 && QEMU_VERSION_MINOR > 0
+    .bdrv_open                      = vitastor_file_open,
+#else
     .bdrv_file_open                 = vitastor_file_open,
+#endif
     .bdrv_close                     = vitastor_close,
 
     // Option list for the create operation

src/client/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 1.9.2
+Version: 1.9.3
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

src/cmd/cli_dd.cpp

@@ -431,7 +431,7 @@ struct cli_dd_t
             if (read_op->retval < 0)
             {
                 fprintf(
-                    stderr, "Failed to read bitmap for %lu bytes from image %s at offset %lu: %s (code %d)\n",
+                    stderr, "Failed to read bitmap for %ju bytes from image %s at offset %ju: %s (code %d)\n",
                     read_op->len, iinfo.iimg.c_str(), read_op->offset,
                     strerror(read_op->retval < 0 ? -read_op->retval : EIO), read_op->retval
                 );
@@ -476,7 +476,7 @@ struct cli_dd_t
             if (read_op->retval != read_op->len)
             {
                 fprintf(
-                    stderr, "Failed to read %lu bytes from image %s at offset %lu: %s (code %d)\n",
+                    stderr, "Failed to read %ju bytes from image %s at offset %ju: %s (code %d)\n",
                     read_op->len, iinfo.iimg.c_str(), read_op->offset,
                     strerror(read_op->retval < 0 ? -read_op->retval : EIO), read_op->retval
                 );
@@ -547,7 +547,7 @@ struct cli_dd_t
                 if (data->res < 0)
                 {
                     fprintf(
-                        stderr, "Failed to read %lu bytes from %s at offset %lu: %s (code %d)\n",
+                        stderr, "Failed to read %ju bytes from %s at offset %ju: %s (code %d)\n",
                         data->iov.iov_len, iinfo.ifile == "" ? "stdin" : iinfo.ifile.c_str(), cur_read->offset,
                         strerror(-data->res), data->res
                     );
@@ -644,7 +644,7 @@ struct cli_dd_t
                 if (write_op->retval != write_op->len)
                 {
                     fprintf(
-                        stderr, "Failed to write %lu bytes to image %s at offset %lu: %s (code %d)\n",
+                        stderr, "Failed to write %ju bytes to image %s at offset %ju: %s (code %d)\n",
                         write_op->len, oinfo.oimg.c_str(), write_op->offset,
                         strerror(write_op->retval < 0 ? -write_op->retval : EIO), write_op->retval
                     );
@@ -680,7 +680,7 @@ struct cli_dd_t
                 if (data->res < 0)
                 {
                     fprintf(
-                        stderr, "Failed to write %lu bytes to %s at offset %lu: %s (code %d)\n",
+                        stderr, "Failed to write %ju bytes to %s at offset %ju: %s (code %d)\n",
                         data->iov.iov_len, oinfo.ofile == "" ? "stdout" : oinfo.ofile.c_str(),
                         oinfo.out_seekable ? cur_read->offset+cur_read->len+oseek : 0,
                         strerror(-data->res), data->res
@@ -727,7 +727,7 @@ struct cli_dd_t
         {
             char buf[256];
             snprintf(
-                buf, sizeof(buf), "%lu bytes (%s) copied, %.1f s, %sB/s",
+                buf, sizeof(buf), "%ju bytes (%s) copied, %.1f s, %sB/s",
                 written_size, format_size(written_size).c_str(), sec_total,
                 format_size((uint64_t)(written_size/sec_total), true).c_str()
             );
@@ -749,7 +749,7 @@ struct cli_dd_t
         else
         {
             fprintf(
-                stderr, "\r%lu bytes (%s) copied, %.1f s, %sB/s, avg %sB/s\033[K",
+                stderr, "\r%ju bytes (%s) copied, %.1f s, %sB/s, avg %sB/s\033[K",
                 written_size, format_size(written_size).c_str(), sec_total,
                 format_size((uint64_t)(delta/sec_delta), true).c_str(),
                 format_size((uint64_t)(written_size/sec_total), true).c_str()

src/cmd/cli_osd_tree.cpp

@@ -216,7 +216,7 @@ resume_1:
                 for (uint64_t osd_num: node.child_osds)
                 {
                     auto & osd = placement_tree->osds.at(osd_num);
-                    fmt_items.push_back(json11::Json::object{
+                    auto json_osd = json11::Json::object{
                         { "type", "osd" },
                         { "name", osd.num },
                         { "parent", node.name },
@@ -230,7 +230,16 @@ resume_1:
                         { "bitmap", (uint64_t)osd.bitmap_granularity },
                         { "commit", osd.immediate_commit == IMMEDIATE_NONE ? "none" : (osd.immediate_commit == IMMEDIATE_ALL ? "all" : "small") },
                         { "op_stats", osd_stats[osd_num]["op_stats"] },
-                    });
+                    };
+                    if (osd_stats[osd_num]["slow_ops_primary"].uint64_value() > 0)
+                    {
+                        json_osd["slow_ops_primary"] = osd_stats[osd_num]["slow_ops_primary"];
+                    }
+                    if (osd_stats[osd_num]["slow_ops_secondary"].uint64_value() > 0)
+                    {
+                        json_osd["slow_ops_secondary"] = osd_stats[osd_num]["slow_ops_secondary"];
+                    }
+                    fmt_items.push_back(json_osd);
                 }
             }
             result.data = fmt_items;

src/cmd/cli_status.cpp

@@ -134,6 +134,7 @@ resume_2:
         }
         int osd_count = 0, osd_up = 0;
         uint64_t total_raw = 0, free_raw = 0, free_down_raw = 0, down_raw = 0;
+        std::vector<uint64_t> slow_op_primary_osds, slow_op_secondary_osds;
         parent->iterate_kvs_1(osd_stats, "/osd/stats/", [&](uint64_t stat_osd_num, json11::Json value)
         {
             osd_count++;
@@ -153,6 +154,14 @@ resume_2:
             if (peer_it != parent->cli->st_cli.peer_states.end())
             {
                 osd_up++;
+                if (value["slow_ops_primary"].uint64_value() > 0)
+                {
+                    slow_op_primary_osds.push_back(stat_osd_num);
+                }
+                if (value["slow_ops_secondary"].uint64_value() > 0)
+                {
+                    slow_op_secondary_osds.push_back(stat_osd_num);
+                }
             }
             else
             {
@@ -216,6 +225,10 @@ resume_2:
                 { "mon_master", mon_master },
                 { "osd_up", osd_up },
                 { "osd_count", osd_count },
+                { "osds_full", osds_full },
+                { "osds_nearfull", osds_nearfull },
+                { "osds_primary_slow_ops", slow_op_primary_osds },
+                { "osds_secondary_slow_ops", slow_op_secondary_osds },
                 { "total_raw", total_raw },
                 { "free_raw", free_raw },
                 { "down_raw", down_raw },
@@ -300,6 +313,26 @@ resume_2:
             warning_str += "    "+std::to_string(osds_nearfull)+
                 (osds_nearfull > 1 ? " osds are almost full\n" : " osd is almost full\n");
         }
+        if (slow_op_primary_osds.size() > 0)
+        {
+            warning_str += "    "+std::to_string(slow_op_primary_osds.size());
+            warning_str += (slow_op_primary_osds.size() > 1 ? " osds have" : " osd has");
+            warning_str += " slow client ops: ";
+            for (int i = 0; i < slow_op_primary_osds.size(); i++)
+            {
+                warning_str += (i > 0 ? ", " : "")+std::to_string(slow_op_primary_osds[i])+"\n";
+            }
+        }
+        if (slow_op_secondary_osds.size() > 0)
+        {
+            warning_str += "    "+std::to_string(slow_op_secondary_osds.size());
+            warning_str += (slow_op_secondary_osds.size() > 1 ? " osds have" : " osd has");
+            warning_str += " slow replication ops: ";
+            for (int i = 0; i < slow_op_secondary_osds.size(); i++)
+            {
+                warning_str += (i > 0 ? ", " : "")+std::to_string(slow_op_secondary_osds[i])+"\n";
+            }
+        }
         if (warning_str != "")
         {
             warning_str = "\n  warning:\n"+warning_str;

src/disk_tool/disk_tool.cpp

@@ -27,12 +27,16 @@ static const char *help_text =
     "    --osd_per_disk <N>\n"
     "      Create <N> OSDs on each disk (default 1)\n"
     "    --hybrid\n"
-    "      Prepare hybrid (HDD+SSD) OSDs using provided devices. SSDs will be used for\n"
+    "      Prepare hybrid (HDD+SSD, NVMe+SATA or etc) OSDs using provided devices. By default,\n"
-    "      journals and metadata, HDDs will be used for data. Partitions for journals and\n"
+    "      any passed SSDs will be used for journals and metadata, HDDs will be used for data,\n"
-    "      metadata will be created automatically. Whether disks are SSD or HDD is decided\n"
+    "      but you can override this behaviour with --fast-devices option. Journal and metadata\n"
-    "      by the `/sys/block/.../queue/rotational` flag. In hybrid mode, default object\n"
+    "      partitions will be created automatically. In the default mode, SSD and HDD disks\n"
-    "      size is 1 MB instead of 128 KB, default journal size is 1 GB instead of 32 MB,\n"
+    "      are distinguished by the `/sys/block/.../queue/rotational` flag. When HDDs are used\n"
-    "      and throttle_small_writes is enabled by default.\n"
+    "      for data in hybrid mode, default block_size is 1 MB instead of 128 KB, default journal\n"
+    "      size is 1 GB instead of 32 MB, and throttle_small_writes is enabled by default.\n"
+    "    --fast-devices /dev/nvmeX,/dev/nvmeY\n"
+    "      In --hybrid mode, use these devices for journal and metadata instead of auto-detecting\n"
+    "      and extracting them from the main [devices...] list.\n"
     "    --disable_data_fsync auto\n"
     "      Disable data device cache and fsync (1/yes/true = on, default auto)\n"
     "    --disable_meta_fsync auto\n"
@@ -196,6 +200,7 @@ static const char *help_text =
     "    --device_size 0          Set device size\n"
     "    --format text            Result format: json, options, env, or text\n"
     "\n"
+    "Default I/O mode for commands involving disk I/O is O_DIRECT. If you don't want it, add --io cached.\n"
     "Use vitastor-disk --help <command> for command details or vitastor-disk --help --all for all details.\n"
 ;
 
@@ -220,6 +225,10 @@ int main(int argc, char *argv[])
         cmd.push_back((char*)"dump-journal");
         aliased = true;
     }
+    else if (!strcmp(exe_name, "vitastor-disk-test"))
+    {
+        self.test_mode = true;
+    }
     for (int i = 1; i < argc; i++)
     {
         if (!strcmp(argv[i], "--all"))
@@ -314,6 +323,7 @@ int main(int argc, char *argv[])
             // First argument is an OSD device - take metadata layout parameters from it
             if (self.dump_load_check_superblock(self.new_journal_device))
                 return 1;
+            self.new_journal_device = self.dsk.journal_device;
             self.new_journal_offset = self.dsk.journal_offset;
             self.new_journal_len = self.dsk.journal_len;
         }
@@ -379,6 +389,7 @@ int main(int argc, char *argv[])
             // First argument is an OSD device - take metadata layout parameters from it
             if (self.dump_load_check_superblock(self.new_meta_device))
                 return 1;
+            self.new_meta_device = self.dsk.meta_device;
             self.new_meta_offset = self.dsk.meta_offset;
             self.new_meta_len = self.dsk.meta_len;
         }

src/disk_tool/disk_tool.h

@@ -22,6 +22,7 @@
 #define VITASTOR_DISK_MAX_SB_SIZE 128*1024
 #define VITASTOR_PART_TYPE "e7009fac-a5a1-4d72-af72-53de13059903"
 #define DEFAULT_HYBRID_JOURNAL "1G"
+#define DEFAULT_HYBRID_SSD_JOURNAL "128M"
 
 struct resizer_data_moving_t;
 
@@ -40,6 +41,7 @@ struct disk_tool_t
     /**** Parameters ****/
 
     std::map<std::string, std::string> options;
+    bool test_mode = false;
     bool all, json, now;
     bool dump_with_blocks, dump_with_data;
     blockstore_disk_t dsk;
@@ -126,7 +128,8 @@ struct disk_tool_t
     uint32_t write_osd_superblock(std::string device, json11::Json params);
 
     int prepare_one(std::map<std::string, std::string> options, int is_hdd = -1);
-    int check_existing_partition(const std::string & dev);
+    int check_existing_partition(std::string & dev_by_uuid);
+    int fix_partition_type(std::string & dev_by_uuid);
     int prepare(std::vector<std::string> devices);
     std::vector<vitastor_dev_info_t> collect_devices(const std::vector<std::string> & devices);
     json11::Json add_partitions(vitastor_dev_info_t & devinfo, std::vector<std::string> sizes);
@@ -148,6 +151,6 @@ int write_zero(int fd, uint64_t offset, uint64_t size);
 json11::Json read_parttable(std::string dev);
 uint64_t dev_size_from_parttable(json11::Json pt);
 uint64_t free_from_parttable(json11::Json pt);
-int fix_partition_type(std::string dev_by_uuid);
+int fix_partition_type_uuid(std::string & dev_by_uuid, const std::string & type_uuid);
 std::string csum_type_str(uint32_t data_csum_type);
 uint32_t csum_type_from_str(std::string data_csum_type);

src/disk_tool/disk_tool_journal.cpp

@@ -18,7 +18,7 @@ int disk_tool_t::dump_journal()
         printf("[\n");
     if (all)
     {
-        dsk.journal_fd = open(dsk.journal_device.c_str(), O_DIRECT|O_RDONLY);
+        dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
         if (dsk.journal_fd < 0)
         {
             fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
@@ -121,7 +121,7 @@ int disk_tool_t::dump_journal()
 
 int disk_tool_t::process_journal(std::function<int(void*)> block_fn)
 {
-    dsk.journal_fd = open(dsk.journal_device.c_str(), O_DIRECT|O_RDONLY);
+    dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
     if (dsk.journal_fd < 0)
     {
         fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));

src/disk_tool/disk_tool_meta.cpp

@@ -14,7 +14,7 @@ int disk_tool_t::process_meta(std::function<void(blockstore_meta_header_v2_t *)>
         fprintf(stderr, "Invalid metadata block size: is not a multiple of %d\n", DIRECT_IO_ALIGNMENT);
         return 1;
     }
-    dsk.meta_fd = open(dsk.meta_device.c_str(), O_DIRECT|O_RDONLY);
+    dsk.meta_fd = open(dsk.meta_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
     if (dsk.meta_fd < 0)
     {
         fprintf(stderr, "Failed to open metadata device %s: %s\n", dsk.meta_device.c_str(), strerror(errno));
@@ -159,7 +159,7 @@ int disk_tool_t::dump_load_check_superblock(const std::string & device)
     {
         auto cfg = json_to_string_map(sb["params"].object_items());
         dsk.parse_config(cfg);
-        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "cached";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -315,8 +315,7 @@ int disk_tool_t::write_json_meta(json11::Json meta)
                 fromhexstr(e["data_csum"].string_value(), new_data_csum_size,
                     ((uint8_t*)new_entry) + sizeof(clean_disk_entry) + 2*new_clean_entry_bitmap_size);
             }
-            uint32_t *new_entry_csum = (uint32_t*)(((uint8_t*)new_entry) + sizeof(clean_disk_entry) +
+            uint32_t *new_entry_csum = (uint32_t*)(((uint8_t*)new_entry) + new_clean_entry_size - 4);
-                2*new_clean_entry_bitmap_size + new_data_csum_size);
             *new_entry_csum = crc32c(0, new_entry, new_clean_entry_size - 4);
         }
     }

src/disk_tool/disk_tool_prepare.cpp

@@ -29,18 +29,12 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     };
     if (options.find("force") == options.end())
     {
-        std::vector<std::string> all_devs = { options["data_device"], options["meta_device"], options["journal_device"] };
+        std::string* all_devs[] = { &options["data_device"], &options["meta_device"], &options["journal_device"] };
-        for (int i = 0; i < all_devs.size(); i++)
+        for (int i = 0; i < 3; i++)
         {
-            const auto & dev = all_devs[i];
+            auto & dev = *all_devs[i];
             if (dev == "")
                 continue;
-            if (dev.substr(0, 22) != "/dev/disk/by-partuuid/")
-            {
-                // Partitions should be identified by GPT partition UUID
-                fprintf(stderr, "%s does not start with /dev/disk/by-partuuid/. Partitions should be identified by GPT partition UUIDs\n", dev.c_str());
-                return 1;
-            }
             std::string real_dev = realpath_str(dev, false);
             if (real_dev == "")
                 return 1;
@@ -114,7 +108,11 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     try
     {
         dsk.parse_config(options);
-        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
+        // Set all offsets to 4096 to calculate metadata size with excess
+        dsk.journal_offset = 4096;
+        dsk.meta_offset = 4096;
+        dsk.data_offset = 4096;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = (options["io"] == "cached" ? "cached" : "direct");
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -159,7 +157,11 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
         return 1;
     }
     std::string osd_num_str;
-    if (shell_exec({ "vitastor-cli", "alloc-osd" }, "", &osd_num_str, NULL) != 0)
+    if (test_mode && options.find("osd_num") != options.end())
+    {
+        osd_num_str = options["osd_num"];
+    }
+    else if (shell_exec({ "vitastor-cli", "alloc-osd" }, "", &osd_num_str, NULL) != 0)
     {
         dsk.close_all();
         return 1;
@@ -173,8 +175,8 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     }
     sb["osd_num"] = osd_num;
     // Zero out metadata and journal
-    if (write_zero(dsk.meta_fd, dsk.meta_offset, dsk.meta_len) != 0 ||
+    if (write_zero(dsk.meta_fd, sb["meta_offset"].uint64_value(), dsk.meta_len) != 0 ||
-        write_zero(dsk.journal_fd, dsk.journal_offset, dsk.journal_len) != 0)
+        write_zero(dsk.journal_fd, sb["journal_offset"].uint64_value(), dsk.journal_len) != 0)
     {
         fprintf(stderr, "Failed to zero out metadata or journal: %s\n", strerror(errno));
         dsk.close_all();
@@ -199,15 +201,18 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     if (sep_j)
         desc += (sep_m ? " and journal on " : " with journal on ") + realpath_str(options["journal_device"]);
     fprintf(stderr, "Initialized OSD %ju on %s\n", osd_num, desc.c_str());
+    if (!test_mode || options.find("no_init") == options.end())
+    {
         if (shell_exec({ "systemctl", "enable", "--now", "vitastor-osd@"+std::to_string(osd_num) }, "", NULL, NULL) != 0)
         {
             fprintf(stderr, "Failed to enable systemd unit vitastor-osd@%ju\n", osd_num);
             return 1;
         }
+    }
     return 0;
 }
 
-int disk_tool_t::check_existing_partition(const std::string & dev)
+int disk_tool_t::check_existing_partition(std::string & dev)
 {
     std::string out;
     if (shell_exec({ "wipefs", dev }, "", &out, NULL) != 0 || out != "")
@@ -229,11 +234,27 @@ int disk_tool_t::check_existing_partition(const std::string & dev)
     return 0;
 }
 
+int disk_tool_t::fix_partition_type(std::string & dev)
+{
+    std::string type_uuid = VITASTOR_PART_TYPE;
+    if (test_mode && options.find("part_type_uuid") != options.end())
+    {
+        type_uuid = options["part_type_uuid"];
+    }
+    return fix_partition_type_uuid(dev, type_uuid);
+}
+
 std::vector<vitastor_dev_info_t> disk_tool_t::collect_devices(const std::vector<std::string> & devices)
 {
     std::vector<vitastor_dev_info_t> devinfo;
+    std::set<std::string> seen;
     for (auto & dev: devices)
     {
+        if (seen.find(dev) != seen.end())
+        {
+            fprintf(stderr, "%s is specified multiple times, ignoring\n", dev.c_str());
+            continue;
+        }
         // Check if the device is a whole disk
         if (dev.substr(0, 5) != "/dev/")
         {
@@ -294,10 +315,6 @@ std::vector<vitastor_dev_info_t> disk_tool_t::collect_devices(const std::vector<
             .free = !pt.is_null() ? free_from_parttable(pt) : dev_size,
         });
     }
-    if (!devinfo.size())
-    {
-        fprintf(stderr, "No suitable devices found\n");
-    }
     return devinfo;
 }
 
@@ -348,47 +365,12 @@ json11::Json disk_tool_t::add_partitions(vitastor_dev_info_t & devinfo, std::vec
         fprintf(stderr, "Failed to add %zu partition(s) with sfdisk: new partitions not found in table\n", sizes.size());
         return {};
     }
-    // Check if new nodes exist and run partprobe if not
+    // Check if new devices exist, run partprobe if not, then wait until they appear
     // FIXME: We could use parted instead of sfdisk because partprobe is already a part of parted
-    int iter = 0, r;
-    while (true)
-    {
-        for (const auto & part: new_parts)
-        {
-            std::string link_path = "/dev/disk/by-partuuid/"+strtolower(part["uuid"].string_value());
-            struct stat st;
-            if (lstat(link_path.c_str(), &st) < 0)
-            {
-                if (errno == ENOENT)
-                {
-                    iter++;
-                    // Run partprobe
-                    std::string out;
-                    if (iter > 1 || (r = shell_exec({ "partprobe", devinfo.path }, "", &out, NULL)) != 0)
-                    {
-                        fprintf(
-                            stderr, iter == 1 && r == 255
-                                ? "partprobe utility is required to reread partition table while disk %s is in use\n"
-                                : "partprobe failed to re-read partition table while disk %s is in use\n",
-                            devinfo.path.c_str()
-                        );
-                        return {};
-                    }
-                    break;
-                }
-                else
-                {
-                    fprintf(stderr, "Failed to lstat %s: %s\n", link_path.c_str(), strerror(errno));
-                    return {};
-                }
-            }
-        }
-        break;
-    }
-    // Wait until device symlinks in /dev/disk/by-partuuid/ appear
     bool exists = false;
     const int max_iter = 300; // max 30 sec
-    iter = 0;
+    int iter = 0;
+    int r = 0;
     while (!exists && iter < max_iter)
     {
         exists = true;
@@ -396,28 +378,48 @@ json11::Json disk_tool_t::add_partitions(vitastor_dev_info_t & devinfo, std::vec
         {
             std::string link_path = "/dev/disk/by-partuuid/"+strtolower(part["uuid"].string_value());
             struct stat st;
-            if (lstat(link_path.c_str(), &st) < 0)
+            if (stat(part["node"].string_value().c_str(), &st) < 0 ||
+                lstat(link_path.c_str(), &st) < 0)
             {
                 if (errno == ENOENT)
                 {
                     exists = false;
                     if (iter == 4)
                     {
+                        // Print message after 400ms
                         fprintf(stderr, "Waiting for %s to appear for up to %d sec...\n", link_path.c_str(), max_iter/10);
                     }
                 }
                 else
                 {
-                    fprintf(stderr, "Failed to lstat %s: %s\n", link_path.c_str(), strerror(errno));
+                    fprintf(stderr, "Failed to stat %s or lstat %s: %s\n", part["node"].string_value().c_str(),
+                        link_path.c_str(), strerror(errno));
                     return {};
                 }
             }
         }
-        if (!exists)
+        if (exists)
         {
-            struct timespec ts = { .tv_sec = 0, .tv_nsec = 100000000 }; // 100ms
+            break;
-            iter += (nanosleep(&ts, NULL) == 0);
         }
+        if (!exists && iter == 0)
+        {
+            // Run partprobe
+            std::string out;
+            r = shell_exec({ "partprobe", devinfo.path }, "", &out, NULL);
+            if (r != 0)
+            {
+                fprintf(
+                    stderr, r == 255
+                        ? "partprobe utility is required to reread partition table while disk %s is in use\n"
+                        : "partprobe failed to re-read partition table while disk %s is in use\n",
+                    devinfo.path.c_str()
+                );
+                return {};
+            }
+        }
+        struct timespec ts = { .tv_sec = 0, .tv_nsec = 100000000 }; // 100ms
+        iter += (nanosleep(&ts, NULL) == 0 || !iter);
     }
     devinfo.pt = newpt;
     devinfo.osd_part_count += sizes.size();
@@ -500,7 +502,10 @@ int disk_tool_t::get_meta_partition(std::vector<vitastor_dev_info_t> & ssds, std
     {
         blockstore_disk_t dsk;
         dsk.parse_config(options);
-        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
+        dsk.journal_offset = 4096;
+        dsk.meta_offset = 4096;
+        dsk.data_offset = 4096;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "cached";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -510,6 +515,7 @@ int disk_tool_t::get_meta_partition(std::vector<vitastor_dev_info_t> & ssds, std
     }
     catch (std::exception & e)
     {
+        dsk.close_all();
         fprintf(stderr, "%s\n", e.what());
         return 1;
     }
@@ -564,9 +570,12 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
 {
     if (options.find("data_device") != options.end() && options["data_device"] != "")
     {
-        if (options.find("hybrid") != options.end() || options.find("osd_per_disk") != options.end() || devices.size())
+        if (options.find("hybrid") != options.end() ||
+            options.find("fast_devices") != options.end() ||
+            options.find("osd_per_disk") != options.end() ||
+            devices.size())
         {
-            fprintf(stderr, "Device list (positional arguments) and --hybrid are incompatible with --data_device\n");
+            fprintf(stderr, "Device list (positional arguments), --osd_per_disk, --hybrid and --fast-devices are incompatible with --data_device\n");
             return 1;
         }
         return prepare_one(options, options.find("hdd") != options.end() ? 1 : 0);
@@ -583,8 +592,10 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
     auto devinfo = collect_devices(devices);
     if (!devinfo.size())
     {
+        fprintf(stderr, "No suitable devices found\n");
         return 1;
     }
+    bool explicit_fast = options.find("fast_devices") != options.end();
     uint64_t osd_per_disk = stoull_full(options["osd_per_disk"]);
     if (!osd_per_disk)
         osd_per_disk = 1;
@@ -603,22 +614,56 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
         if (options.find("disable_meta_fsync") == options.end())
             options["disable_meta_fsync"] = "auto";
         options["disable_journal_fsync"] = options["disable_meta_fsync"];
+        if (explicit_fast)
+        {
+            auto fast = explode(",", options["fast_devices"], true);
+            ssds = collect_devices(fast);
+            if (!ssds.size())
+            {
+                fprintf(stderr, "No fast devices found\n");
+                return 1;
+            }
+            if (options["journal_size"] == "")
+            {
+                auto auto_journal_size = DEFAULT_HYBRID_SSD_JOURNAL;
                 for (auto & dev: devinfo)
+                {
+                    if (dev.is_hdd)
+                    {
+                        auto_journal_size = DEFAULT_HYBRID_JOURNAL;
+                        break;
+                    }
+                }
+                options["journal_size"] = auto_journal_size;
+            }
+        }
+        else
+        {
+            std::vector<vitastor_dev_info_t> hdds;
+            for (auto & dev: devinfo)
+            {
                 if (!dev.is_hdd)
                     ssds.push_back(dev);
+                else
+                    hdds.push_back(dev);
+            }
             if (!ssds.size())
             {
                 fprintf(stderr, "No SSDs found\n");
                 return 1;
             }
-        else if (ssds.size() == devinfo.size())
+            if (!hdds.size())
             {
                 fprintf(stderr, "No HDDs found\n");
                 return 1;
             }
+            devinfo = hdds;
             if (options["journal_size"] == "")
+            {
                 options["journal_size"] = DEFAULT_HYBRID_JOURNAL;
             }
+        }
+    }
     else
     {
         options.erase("disable_meta_fsync");
@@ -626,8 +671,6 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
     }
     auto journal_size = options["journal_size"];
     for (auto & dev: devinfo)
-    {
-        if (!hybrid || dev.is_hdd)
     {
         // Select new partitions and create an OSD on each of them
         for (const auto & uuid: get_new_data_parts(dev, osd_per_disk, max_other_percent))
@@ -654,6 +697,5 @@ int disk_tool_t::prepare(std::vector<std::string> devices)
             }
         }
     }
-    }
     return 0;
 }

src/disk_tool/disk_tool_resize.cpp

@@ -91,7 +91,7 @@ int disk_tool_t::resize_parse_params()
     try
     {
         dsk.parse_config(options);
-        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "cached";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -114,7 +114,10 @@ int disk_tool_t::resize_parse_params()
     new_data_offset = options.find("new_data_offset") != options.end()
         ? parse_size(options["new_data_offset"]) : dsk.data_offset;
     new_data_len = options.find("new_data_len") != options.end()
-        ? parse_size(options["new_data_len"]) : dsk.data_len;
+        ? parse_size(options["new_data_len"])
+        : (options.find("new_data_offset") != options.end()
+            ? dsk.data_device_size-new_data_offset
+            : dsk.data_len);
     new_meta_offset = options.find("new_meta_offset") != options.end()
         ? parse_size(options["new_meta_offset"]) : dsk.meta_offset;
     new_meta_len = options.find("new_meta_len") != options.end()
@@ -123,6 +126,14 @@ int disk_tool_t::resize_parse_params()
         ? parse_size(options["new_journal_offset"]) : dsk.journal_offset;
     new_journal_len = options.find("new_journal_len") != options.end()
         ? parse_size(options["new_journal_len"]) : dsk.journal_len;
+    if (new_data_len+new_data_offset > dsk.data_device_size)
+        new_data_len = dsk.data_device_size-new_data_offset;
+    if (new_meta_device == dsk.data_device && new_data_offset < new_meta_offset &&
+        new_data_len+new_data_offset > new_meta_offset)
+        new_data_len = new_meta_offset-new_data_offset;
+    if (new_journal_device == dsk.data_device && new_data_offset < new_journal_offset &&
+        new_data_len+new_data_offset > new_journal_offset)
+        new_data_len = new_journal_offset-new_data_offset;
     if (new_meta_device == dsk.meta_device &&
         new_journal_device == dsk.journal_device &&
         new_data_offset == dsk.data_offset &&
@@ -159,10 +170,10 @@ void disk_tool_t::resize_init(blockstore_meta_header_v2_t *hdr)
         dsk.data_csum_type = hdr->data_csum_type;
         dsk.csum_block_size = hdr->csum_block_size;
     }
-    if (((new_data_len-dsk.data_len) % dsk.data_block_size) ||
+    if (((new_data_offset-dsk.data_offset) % dsk.data_block_size))
-        ((new_data_offset-dsk.data_offset) % dsk.data_block_size))
     {
-        fprintf(stderr, "Data alignment mismatch\n");
+        fprintf(stderr, "Data alignment mismatch: old data offset is 0x%jx, new is 0x%jx, but alignment on %x should be equal\n",
+            dsk.data_offset, new_data_offset, dsk.data_block_size);
         exit(1);
     }
     data_idx_diff = ((int64_t)(dsk.data_offset-new_data_offset)) / dsk.data_block_size;
@@ -220,10 +231,10 @@ int disk_tool_t::resize_remap_blocks()
     }
     for (uint64_t i = 0; i < free_last; i++)
     {
-        if (data_alloc->get(total_blocks-i))
+        if (data_alloc->get(total_blocks-i-1))
-            data_remap[total_blocks-i] = 0;
+            data_remap[total_blocks-i-1] = 0;
         else
-            data_alloc->set(total_blocks-i, true);
+            data_alloc->set(total_blocks-i-1, true);
     }
     for (auto & p: data_remap)
     {
@@ -246,7 +257,7 @@ int disk_tool_t::resize_copy_data()
         iodepth = 32;
     }
     ringloop = new ring_loop_t(iodepth < RINGLOOP_DEFAULT_SIZE ? RINGLOOP_DEFAULT_SIZE : iodepth);
-    dsk.data_fd = open(dsk.data_device.c_str(), O_DIRECT|O_RDWR);
+    dsk.data_fd = open(dsk.data_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (dsk.data_fd < 0)
     {
         fprintf(stderr, "Failed to open data device %s: %s\n", dsk.data_device.c_str(), strerror(errno));
@@ -441,7 +452,7 @@ int disk_tool_t::resize_rewrite_journal()
 
 int disk_tool_t::resize_write_new_journal()
 {
-    new_journal_fd = open(new_journal_device.c_str(), O_DIRECT|O_RDWR);
+    new_journal_fd = open(new_journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (new_journal_fd < 0)
     {
         fprintf(stderr, "Failed to open new journal device %s: %s\n", new_journal_device.c_str(), strerror(errno));
@@ -467,12 +478,13 @@ int disk_tool_t::resize_rewrite_meta()
             blockstore_meta_header_v2_t *new_hdr = (blockstore_meta_header_v2_t *)new_meta_buf;
             new_hdr->zero = 0;
             new_hdr->magic = BLOCKSTORE_META_MAGIC_V1;
-            new_hdr->version = BLOCKSTORE_META_FORMAT_V1;
+            new_hdr->version = BLOCKSTORE_META_FORMAT_V2;
             new_hdr->meta_block_size = dsk.meta_block_size;
             new_hdr->data_block_size = dsk.data_block_size;
             new_hdr->bitmap_granularity = dsk.bitmap_granularity ? dsk.bitmap_granularity : 4096;
             new_hdr->data_csum_type = dsk.data_csum_type;
             new_hdr->csum_block_size = dsk.csum_block_size;
+            new_hdr->header_csum = crc32c(0, new_hdr, sizeof(*new_hdr));
         },
         [this](uint64_t block_num, clean_disk_entry *entry, uint8_t *bitmap)
         {
@@ -481,7 +493,7 @@ int disk_tool_t::resize_rewrite_meta()
                 block_num = remap_it->second;
             if (block_num < free_first || block_num >= total_blocks-free_last)
             {
-                fprintf(stderr, "BUG: remapped block not in range\n");
+                fprintf(stderr, "BUG: remapped block %ju not in range %ju..%ju\n", block_num, free_first, total_blocks-free_last);
                 exit(1);
             }
             block_num += data_idx_diff;
@@ -494,6 +506,8 @@ int disk_tool_t::resize_rewrite_meta()
                 memcpy(new_entry->bitmap, bitmap, 2*new_clean_entry_bitmap_size + new_data_csum_size);
             else
                 memset(new_entry->bitmap, 0xff, 2*new_clean_entry_bitmap_size);
+            uint32_t *new_entry_csum = (uint32_t*)(((uint8_t*)new_entry) + new_clean_entry_size - 4);
+            *new_entry_csum = crc32c(0, new_entry, new_clean_entry_size - 4);
         }
     );
     if (r != 0)
@@ -507,7 +521,7 @@ int disk_tool_t::resize_rewrite_meta()
 
 int disk_tool_t::resize_write_new_meta()
 {
-    new_meta_fd = open(new_meta_device.c_str(), O_DIRECT|O_RDWR);
+    new_meta_fd = open(new_meta_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (new_meta_fd < 0)
     {
         fprintf(stderr, "Failed to open new metadata device %s: %s\n", new_meta_device.c_str(), strerror(errno));

src/disk_tool/disk_tool_resize_auto.cpp

@@ -37,6 +37,8 @@ int disk_tool_t::resize_data(std::string device)
         fprintf(stderr, "%s\n", e.what());
         return 1;
     }
+    // Save FD numbers because calc_lengths() relies on them
+    int old_journal_fd = dsk.journal_fd, old_meta_fd = dsk.meta_fd, old_data_fd = dsk.data_fd;
     dsk.close_all();
     bool dry_run = options.find("dry_run") != options.end();
     auto old_journal_device = dsk.journal_device;
@@ -48,6 +50,22 @@ int disk_tool_t::resize_data(std::string device)
         if (options.find("move_journal") == options.end())
             options["move_journal"] = dsk.journal_device == dsk.data_device ? "" : dsk.journal_device;
     }
+    uint64_t new_data_dev_size = 0;
+    if (options.find("data_size") != options.end())
+    {
+        new_data_dev_size = parse_size(options["data_size"]);
+        new_data_dev_size = options["data_size"] == "max" || new_data_dev_size > dsk.data_device_size
+            ? dsk.data_device_size : new_data_dev_size;
+        dsk.data_device_size = new_data_dev_size;
+        dsk.cfg_data_size = 0;
+        dsk.journal_fd = old_journal_fd;
+        dsk.meta_fd = old_meta_fd;
+        dsk.data_fd = old_data_fd;
+        dsk.calc_lengths(true);
+        dsk.journal_fd = -1;
+        dsk.meta_fd = -1;
+        dsk.data_fd = -1;
+    }
     std::map<std::string, std::string> move_options;
     if (options.find("move_journal") != options.end())
     {
@@ -69,14 +87,8 @@ int disk_tool_t::resize_data(std::string device)
     new_data_offset += ((dsk.data_offset-new_data_offset) % dsk.data_block_size);
     if (new_data_offset != dsk.data_offset)
         move_options["new_data_offset"] = std::to_string(new_data_offset);
-    if (options.find("data_size") != options.end())
+    if (new_data_dev_size != 0)
-    {
-        auto new_data_dev_size = parse_size(options["data_size"]);
-        new_data_dev_size = options["data_size"] == "max" || new_data_dev_size > dsk.data_device_size
-            ? dsk.data_device_size : new_data_dev_size;
-        if (new_data_dev_size-dsk.data_offset != dsk.data_len)
         move_options["new_data_len"] = std::to_string(new_data_dev_size-new_data_offset);
-    }
     new_meta_offset = 4096 + (new_meta_device == new_journal_device ? new_journal_len : 0);
     if (new_meta_offset != dsk.meta_offset)
         move_options["new_meta_offset"] = std::to_string(new_meta_offset);
@@ -188,17 +200,12 @@ int disk_tool_t::resize_parse_move_journal(std::map<std::string, std::string> &
             else
                 options["move_journal"] = "<new journal partition on "+parent_dev+">";
         }
-        else if (options["move_journal"].substr(0, 22) != "/dev/disk/by-partuuid/")
-        {
-            // Partitions should be identified by GPT partition UUID
-            fprintf(stderr, "%s does not start with /dev/disk/by-partuuid/. Partitions should be identified by GPT partition UUIDs\n", options["move_journal"].c_str());
-            return 1;
-        }
         else
         {
             // already a partition - check that it's a GPT partition with correct type
-            if (options.find("force") == options.end() &&
+            if ((options.find("force") == options.end()
-                check_existing_partition(real_dev) != 0)
+                ? check_existing_partition(options["move_journal"])
+                : fix_partition_type(options["move_journal"])) != 0)
             {
                 return 1;
             }
@@ -269,17 +276,12 @@ int disk_tool_t::resize_parse_move_meta(std::map<std::string, std::string> & mov
             else
                 options["move_meta"] = "<new metadata partition on "+parent_dev+">";
         }
-        else if (options["move_meta"].substr(0, 22) != "/dev/disk/by-partuuid/")
-        {
-            // Partitions should be identified by GPT partition UUID
-            fprintf(stderr, "%s does not start with /dev/disk/by-partuuid/. Partitions should be identified by GPT partition UUIDs\n", options["move_meta"].c_str());
-            return 1;
-        }
         else
         {
             // already a partition - check that it's a GPT partition with correct type
-            if (options.find("force") == options.end() &&
+            if ((options.find("force") == options.end()
-                check_existing_partition(real_dev) != 0)
+                ? check_existing_partition(options["move_meta"])
+                : fix_partition_type(options["move_meta"])) != 0)
             {
                 return 1;
             }

src/disk_tool/disk_tool_udev.cpp

@@ -122,7 +122,7 @@ uint32_t disk_tool_t::write_osd_superblock(std::string device, json11::Json para
     sb->size = sb_size;
     memcpy(sb->json_data, json_data.c_str(), json_data.size());
     sb->crc32c = crc32c(0, &sb->size, sb->size - ((uint8_t*)&sb->size - buf));
-    int fd = open(device.c_str(), O_DIRECT|O_RDWR);
+    int fd = open(device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (fd < 0)
     {
         fprintf(stderr, "Failed to open device %s: %s\n", device.c_str(), strerror(errno));
@@ -150,7 +150,7 @@ json11::Json disk_tool_t::read_osd_superblock(std::string device, bool expect_ex
     json11::Json osd_params;
     std::string json_err;
     std::string real_device, device_type, real_data, real_meta, real_journal;
-    int r, fd = open(device.c_str(), O_DIRECT|O_RDWR);
+    int r, fd = open(device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (fd < 0)
     {
         fprintf(stderr, "Failed to open device %s: %s\n", device.c_str(), strerror(errno));
@@ -385,7 +385,7 @@ int disk_tool_t::pre_exec_osd(std::string device)
 int disk_tool_t::clear_osd_superblock(const std::string & dev)
 {
     uint8_t *buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 4096);
-    int fd = -1, r = open(dev.c_str(), O_DIRECT|O_RDWR);
+    int fd = -1, r = open(dev.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDWR);
     if (r >= 0)
     {
         fd = r;

src/disk_tool/disk_tool_utils.cpp

@@ -343,24 +343,43 @@ uint64_t free_from_parttable(json11::Json pt)
     return free;
 }
 
-int fix_partition_type(std::string dev_by_uuid)
+int fix_partition_type_uuid(std::string & dev_by_uuid, const std::string & type_uuid)
 {
-    auto uuid = strtolower(dev_by_uuid.substr(dev_by_uuid.rfind('/')+1));
+    bool is_partuuid = dev_by_uuid.substr(0, 22) == "/dev/disk/by-partuuid/";
-    std::string parent_dev = get_parent_device(realpath_str(dev_by_uuid, false));
+    auto uuid = is_partuuid ? strtolower(dev_by_uuid.substr(22)) : "";
+    auto node = realpath_str(dev_by_uuid, false);
+    std::string parent_dev = get_parent_device(node);
     if (parent_dev == "")
         return 1;
     auto pt = read_parttable(parent_dev);
     if (pt.is_null() || pt.is_bool())
         return 1;
+    bool found = false;
     std::string script = "label: gpt\n\n";
     for (const auto & part: pt["partitions"].array_items())
     {
-        bool this_part = (strtolower(part["uuid"].string_value()) == uuid);
+        bool this_part = (part["node"].string_value() == node) &&
-        if (this_part && strtolower(part["type"].string_value()) == "e7009fac-a5a1-4d72-af72-53de13059903")
+            (!is_partuuid || strtolower(part["uuid"].string_value()) == uuid);
+        if (this_part)
+        {
+            found = true;
+            if (!is_partuuid)
+            {
+                if (part["uuid"] == "")
+                {
+                    fprintf(stderr, "Could not determine partition UUID for %s. Please use GPT partitions\n", dev_by_uuid.c_str());
+                    return 1;
+                }
+                auto new_dev = "/dev/disk/by-partuuid/"+strtolower(part["uuid"].string_value());
+                fprintf(stderr, "Using %s instead of %s\n", new_dev.c_str(), dev_by_uuid.c_str());
+                dev_by_uuid = new_dev;
+            }
+            if (strtolower(part["type"].string_value()) == type_uuid)
             {
                 // Already correct type
                 return 0;
             }
+        }
         script += part["node"].string_value()+": ";
         bool first = true;
         for (const auto & kv: part.object_items())
@@ -369,13 +388,18 @@ int fix_partition_type(std::string dev_by_uuid)
             {
                 script += (first ? "" : ", ")+kv.first+"="+
                     (kv.first == "type" && this_part
-                        ? "e7009fac-a5a1-4d72-af72-53de13059903"
+                        ? type_uuid
                         : (kv.second.is_string() ? kv.second.string_value() : kv.second.dump()));
                 first = false;
             }
         }
         script += "\n";
     }
+    if (!found)
+    {
+        fprintf(stderr, "Could not find partition table entry for %s\n", dev_by_uuid.c_str());
+        return 1;
+    }
     std::string out;
     return shell_exec({ "sfdisk", "--no-reread", "--no-tell-kernel", "--force", parent_dev }, script, &out, NULL);
 }

src/nfs/nfs_proxy_rdma.cpp

@@ -137,7 +137,7 @@ void nfs_proxy_t::handle_rdmacm_events()
         }
         if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
         {
-            rdmacm_accept();
+            rdmacm_accept(ev);
         }
         else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR ||
             ev->event == RDMA_CM_EVENT_REJECTED ||

src/osd/osd.cpp

@@ -535,10 +535,12 @@ void osd_t::print_stats()
 
 void osd_t::print_slow()
 {
-    bool has_slow = false;
+    cur_slow_op_primary = 0;
+    cur_slow_op_secondary = 0;
     char alloc[1024];
     timespec now;
     clock_gettime(CLOCK_REALTIME, &now);
+    // FIXME: Also track slow local blockstore ops and recovery/flush/scrub ops
     for (auto & kv: msgr.clients)
     {
         for (auto op: kv.second->received_ops)
@@ -608,6 +610,7 @@ void osd_t::print_slow()
                     op->req.hdr.opcode == OSD_OP_SEC_STABILIZE || op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK ||
                     op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
                 {
+                    cur_slow_op_secondary++;
                     bufprintf(" state=%d", op->bs_op ? PRIV(op->bs_op)->op_state : -1);
                     int wait_for = op->bs_op ? PRIV(op->bs_op)->wait_for : 0;
                     if (wait_for)
@@ -618,15 +621,19 @@ void osd_t::print_slow()
                 else if (op->req.hdr.opcode == OSD_OP_READ || op->req.hdr.opcode == OSD_OP_WRITE ||
                     op->req.hdr.opcode == OSD_OP_SYNC || op->req.hdr.opcode == OSD_OP_DELETE)
                 {
+                    cur_slow_op_primary++;
                     bufprintf(" state=%d", !op->op_data ? -1 : op->op_data->st);
                 }
+                else
+                {
+                    cur_slow_op_primary++;
+                }
 #undef bufprintf
                 printf("%s\n", alloc);
-                has_slow = true;
             }
         }
     }
-    if (has_slow && bs)
+    if ((cur_slow_op_primary+cur_slow_op_secondary) > 0 && bs)
     {
         bs->dump_diagnostics();
     }

src/osd/osd.h

@@ -150,7 +150,9 @@ class osd_t
     bool pg_config_applied = false;
     bool etcd_reporting_pg_state = false;
     bool etcd_reporting_stats = false;
-    int autosync_timer_id = -1, print_stats_timer_id = -1, slow_log_timer_id = -1;
+    int print_stats_timer_id = -1, slow_log_timer_id = -1;
+    uint64_t cur_slow_op_primary = 0;
+    uint64_t cur_slow_op_secondary = 0;
 
     // peers and PGs
 
@@ -168,6 +170,8 @@ class osd_t
     object_id recovery_last_oid;
     int recovery_pg_done = 0, recovery_done = 0;
     osd_op_t *autosync_op = NULL;
+    int autosync_copies_to_delete = 0;
+    int autosync_timer_id = -1;
 
     // Scrubbing
     uint64_t scrub_nearest_ts = 0;

src/osd/osd_cluster.cpp

@@ -201,6 +201,14 @@ json11::Json osd_t::get_statistics()
     st["immediate_commit"] = immediate_commit == IMMEDIATE_ALL ? "all" : (immediate_commit == IMMEDIATE_SMALL ? "small" : "none");
     st["host"] = self_state["host"];
     st["version"] = VITASTOR_VERSION;
+    if (cur_slow_op_primary > 0)
+    {
+        st["slow_ops_primary"] = cur_slow_op_primary;
+    }
+    if (cur_slow_op_secondary > 0)
+    {
+        st["slow_ops_secondary"] = cur_slow_op_secondary;
+    }
     json11::Json::object op_stats, subop_stats;
     for (int i = OSD_OP_MIN; i <= OSD_OP_MAX; i++)
     {

src/osd/osd_flush.cpp

@@ -13,10 +13,11 @@ void osd_t::submit_pg_flush_ops(pg_t & pg)
     bool first = true;
     while (it != pg.flush_actions.end())
     {
-        if (!first && (it->first.oid.inode != prev_it->first.oid.inode ||
+        if (!first &&
+            (it->first.oid.inode != prev_it->first.oid.inode ||
                 (it->first.oid.stripe & ~STRIPE_MASK) != (prev_it->first.oid.stripe & ~STRIPE_MASK)) &&
-            fb->rollback_lists[it->first.osd_num].size() >= FLUSH_BATCH ||
+            (fb->rollback_lists[it->first.osd_num].size() >= FLUSH_BATCH ||
-            fb->stable_lists[it->first.osd_num].size() >= FLUSH_BATCH)
+                fb->stable_lists[it->first.osd_num].size() >= FLUSH_BATCH))
         {
             // Stop only at the object boundary
             break;
@@ -75,6 +76,7 @@ void osd_t::handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, p
         // Throw the result away
         return;
     }
+    fb->flush_done++;
     if (retval != 0)
     {
         if (peer_osd == this->osd_num)
@@ -92,12 +94,11 @@ void osd_t::handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, p
             auto fd_it = msgr.osd_peer_fds.find(peer_osd);
             if (fd_it != msgr.osd_peer_fds.end())
             {
+                // Will repeer/stop this PG
                 msgr.stop_client(fd_it->second);
             }
-            return;
         }
     }
-    fb->flush_done++;
     if (fb->flush_done == fb->flush_ops)
     {
         // This flush batch is done

src/osd/osd_primary.cpp

@@ -645,6 +645,18 @@ void osd_t::remove_object_from_state(object_id & oid, pg_osd_set_state_t **objec
     {
         throw std::runtime_error("BUG: Invalid object state: "+std::to_string((*object_state)->state));
     }
+    if (changed && immediate_commit != IMMEDIATE_ALL)
+    {
+        // Trigger double automatic sync after changing PG state when we're running with fsyncs.
+        // First autosync commits all written objects and applies copies_to_delete_after_sync;
+        // Second autosync commits all deletions run by the first sync.
+        // Without it, rebalancing in a cluster without load may result in some small amount of
+        // garbage left on "extra" OSDs of the PG, because last deletions are not synced at all.
+        // FIXME: 1000% correct way is to switch PG state only after copies_to_delete_after_sync.
+        // But it's much more complicated.
+        unstable_write_count += autosync_writes;
+        autosync_copies_to_delete = 2;
+    }
     if (changed && report)
     {
         report_pg_state(pg);

src/osd/osd_primary_subops.cpp

@@ -9,6 +9,10 @@ void osd_t::autosync()
 {
     if (immediate_commit != IMMEDIATE_ALL && !autosync_op)
     {
+        if (autosync_copies_to_delete > 0)
+        {
+            autosync_copies_to_delete--;
+        }
         autosync_op = new osd_op_t();
         autosync_op->op_type = OSD_OP_IN;
         autosync_op->peer_fd = SELF_FD;
@@ -29,6 +33,11 @@ void osd_t::autosync()
             }
             delete autosync_op;
             autosync_op = NULL;
+            if (autosync_copies_to_delete > 0)
+            {
+                // Trigger the second "copies_to_delete" autosync
+                autosync();
+            }
         };
         exec_op(autosync_op);
     }

src/osd/osd_primary_sync.cpp

@@ -213,6 +213,15 @@ resume_8:
         {
             goto resume_6;
         }
+        if (immediate_commit == IMMEDIATE_NONE)
+        {
+            // Mark OSDs as dirty because deletions have to be synced too!
+            for (int i = 0; i < op_data->copies_to_delete_count; i++)
+            {
+                auto & chunk = op_data->copies_to_delete[i];
+                this->dirty_osds.insert(chunk.osd_num);
+            }
+        }
     }
     for (int i = 0; i < op_data->dirty_pg_count; i++)
     {
@@ -227,7 +236,7 @@ resume_8:
             start_pg_peering(pg);
         }
     }
-    // FIXME: Free those in the destructor?
+    // FIXME: Free those in the destructor (not here)?
     free(op_data->dirty_pgs);
     op_data->dirty_pgs = NULL;
     op_data->dirty_osds = NULL;

src/osd/osd_primary_write.cpp

@@ -7,6 +7,12 @@
 bool osd_t::check_write_queue(osd_op_t *cur_op, pg_t & pg)
 {
     osd_primary_op_data_t *op_data = cur_op->op_data;
+    // First check if PG is not active anymore
+    if (!(pg.state & PG_ACTIVE))
+    {
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, -EPIPE);
+        return false;
+    }
     // Check if actions are pending for this object
     auto act_it = pg.flush_actions.lower_bound((obj_piece_id_t){
         .oid = op_data->oid,

src/util/addr_util.cpp

@@ -65,7 +65,7 @@ std::string addr_to_string(const sockaddr_storage &addr)
     return std::string(peer_str)+":"+std::to_string(port);
 }
 
-static bool cidr_match(const in_addr &addr, const in_addr &net, uint8_t bits)
+bool cidr_match(const in_addr &addr, const in_addr &net, uint8_t bits)
 {
     if (bits == 0)
     {
@@ -75,7 +75,7 @@ static bool cidr_match(const in_addr &addr, const in_addr &net, uint8_t bits)
     return !((addr.s_addr ^ net.s_addr) & htonl(0xFFFFFFFFu << (32 - bits)));
 }
 
-static bool cidr6_match(const in6_addr &address, const in6_addr &network, uint8_t bits)
+bool cidr6_match(const in6_addr &address, const in6_addr &network, uint8_t bits)
 {
     const uint32_t *a = address.s6_addr32;
     const uint32_t *n = network.s6_addr32;
@@ -93,47 +93,49 @@ static bool cidr6_match(const in6_addr &address, const in6_addr &network, uint8_
     return true;
 }
 
-struct addr_mask_t
+addr_mask_t cidr_parse(std::string mask)
 {
-    sa_family_t family;
+    unsigned bits = 255;
-    in_addr ipv4;
-    in6_addr ipv6;
-    uint8_t bits;
-};
-
-std::vector<std::string> getifaddr_list(std::vector<std::string> mask_cfg, bool include_v6)
-{
-    std::vector<addr_mask_t> masks;
-    for (auto mask: mask_cfg)
-    {
-        unsigned bits = 0;
     int p = mask.find('/');
     if (p != std::string::npos)
     {
         char null_byte = 0;
         if (sscanf(mask.c_str()+p+1, "%u%c", &bits, &null_byte) != 1 || bits > 128)
-            {
+            throw std::runtime_error("Invalid IP address mask: " + mask);
-                throw std::runtime_error((include_v6 ? "Invalid IPv4 address mask: " : "Invalid IP address mask: ") + mask);
-            }
         mask = mask.substr(0, p);
     }
     in_addr ipv4;
     in6_addr ipv6;
     if (inet_pton(AF_INET, mask.c_str(), &ipv4) == 1)
     {
+        if (bits == 255)
+            bits = 32;
         if (bits > 32)
-            {
+            throw std::runtime_error("Invalid IP address mask: " + mask);
-                throw std::runtime_error((include_v6 ? "Invalid IPv4 address mask: " : "Invalid IP address mask: ") + mask);
+        return (addr_mask_t){ .family = AF_INET, .ipv4 = ipv4, .bits = (uint8_t)(bits ? bits : 32) };
     }
-            masks.push_back((addr_mask_t){ .family = AF_INET, .ipv4 = ipv4, .bits = (uint8_t)bits });
+    else if (inet_pton(AF_INET6, mask.c_str(), &ipv6) == 1)
-        }
-        else if (include_v6 && inet_pton(AF_INET6, mask.c_str(), &ipv6) == 1)
     {
-            masks.push_back((addr_mask_t){ .family = AF_INET6, .ipv6 = ipv6, .bits = (uint8_t)bits });
+        if (bits == 255)
+            bits = 128;
+        return (addr_mask_t){ .family = AF_INET6, .ipv6 = ipv6, .bits = (uint8_t)bits };
     }
     else
     {
-            throw std::runtime_error((include_v6 ? "Invalid IPv4 address mask: " : "Invalid IP address mask: ") + mask);
+        throw std::runtime_error("Invalid IP address mask: " + mask);
+    }
+}
+
+std::vector<std::string> getifaddr_list(std::vector<std::string> mask_cfg, bool include_v6)
+{
+    std::vector<addr_mask_t> masks;
+    for (auto mask: mask_cfg)
+    {
+        masks.push_back(cidr_parse(mask));
+        if (masks[masks.size()-1].family == AF_INET6)
+        {
+            // Auto-enable IPv6 addresses
+            include_v6 = true;
         }
     }
     std::set<std::string> addresses;

src/util/addr_util.h

@@ -1,10 +1,22 @@
 #pragma once
 
+#include <netinet/in.h>
 #include <sys/socket.h>
 #include <string>
 #include <vector>
 
+struct addr_mask_t
+{
+    sa_family_t family;
+    in_addr ipv4;
+    in6_addr ipv6;
+    uint8_t bits;
+};
+
 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);
+addr_mask_t cidr_parse(std::string mask);
+bool cidr_match(const in_addr &address, const in_addr &network, uint8_t bits);
+bool cidr6_match(const in6_addr &address, const in6_addr &network, uint8_t bits);
 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/util/timerfd_manager.cpp

@@ -62,7 +62,7 @@ int timerfd_manager_t::set_timer_us(uint64_t micros, bool repeat, std::function<
         .callback = callback,
     });
     inc_timer(timers[timers.size()-1]);
-    set_nearest();
+    set_nearest(false);
     return timer_id;
 }
 
@@ -82,13 +82,13 @@ void timerfd_manager_t::clear_timer(int timer_id)
             {
                 nearest--;
             }
-            set_nearest();
+            set_nearest(false);
             break;
         }
     }
 }
 
-void timerfd_manager_t::set_nearest()
+void timerfd_manager_t::set_nearest(bool trigger_inline)
 {
     if (onstack > 0)
     {
@@ -134,11 +134,14 @@ again:
         }
         if (exp.it_value.tv_sec < 0 || exp.it_value.tv_sec == 0 && exp.it_value.tv_nsec <= 0)
         {
-            // It already happened
+            // It already happened - set minimal timeout
-            // FIXME: Postpone to setImmediate/BH to avoid reenterability problems
+            if (trigger_inline)
+            {
                 trigger_nearest();
                 goto again;
             }
+            exp.it_value = { .tv_sec = 0, .tv_nsec = 1 };
+        }
         if (timerfd_settime(timerfd, 0, &exp, NULL))
         {
             throw std::runtime_error(std::string("timerfd_settime: ") + strerror(errno));
@@ -157,7 +160,7 @@ void timerfd_manager_t::handle_readable()
         trigger_nearest();
     }
     wait_state = 0;
-    set_nearest();
+    set_nearest(true);
 }
 
 void timerfd_manager_t::trigger_nearest()

src/util/timerfd_manager.h

@@ -26,7 +26,7 @@ class timerfd_manager_t
     std::vector<timerfd_timer_t> timers;
 
     void inc_timer(timerfd_timer_t & t);
-    void set_nearest();
+    void set_nearest(bool trigger_inline);
     void trigger_nearest();
     void handle_readable();
 public:

tests/run_tests.sh

@@ -68,6 +68,9 @@ TEST_NAME=csum_4k_dmj  OSD_ARGS="--data_csum_type crc32c --inmemory_metadata fal
 TEST_NAME=csum_4k_dj   OSD_ARGS="--data_csum_type crc32c --inmemory_journal false" OFFSET_ARGS=$OSD_ARGS ./test_heal.sh
 TEST_NAME=csum_4k      OSD_ARGS="--data_csum_type crc32c" OFFSET_ARGS=$OSD_ARGS ./test_heal.sh
 
+./test_resize.sh
+./test_resize_auto.sh
+
 ./test_snapshot_pool2.sh
 
 ./test_osd_tags.sh

tests/test_resize.sh

@@ -3,6 +3,7 @@
 PG_COUNT=${PG_COUNT:-32}
 
 . `dirname $0`/run_3osds.sh
+check_qemu
 
 LD_PRELOAD="build/src/client/libfio_vitastor.so" \
     fio -thread -name=test -ioengine=build/src/client/libfio_vitastor.so -bs=4M -direct=1 -iodepth=4 \
@@ -26,22 +27,22 @@ for i in $(seq 1 $OSD_COUNT); do
     offsets=$(build/src/disk_tool/vitastor-disk simple-offsets --format json ./testdata/bin/test_osd$i.bin)
     meta_offset=$(echo $offsets | jq -r .meta_offset)
     data_offset=$(echo $offsets | jq -r .data_offset)
-    build/src/disk_tool/vitastor-disk dump-journal --json ./testdata/bin/test_osd$i.bin 4096 0 $meta_offset >./testdata/journal_before_resize.json
+    build/src/disk_tool/vitastor-disk dump-journal --io cached --json ./testdata/bin/test_osd$i.bin 4096 0 $meta_offset >./testdata/journal_before_resize.json
-    build/src/disk_tool/vitastor-disk dump-meta ./testdata/bin/test_osd$i.bin 4096 $meta_offset $((data_offset-meta_offset)) >./testdata/meta_before_resize.json
+    build/src/disk_tool/vitastor-disk dump-meta --io cached ./testdata/bin/test_osd$i.bin 4096 $meta_offset $((data_offset-meta_offset)) >./testdata/meta_before_resize.json
-    build/src/disk_tool/vitastor-disk resize \
+    build/src/disk_tool/vitastor-disk raw-resize --io cached \
         $(build/src/disk_tool/vitastor-disk simple-offsets --format options ./testdata/bin/test_osd$i.bin 2>/dev/null) \
         --new_meta_offset 0 \
         --new_meta_len $((1024*1024)) \
         --new_journal_offset $((1024*1024)) \
-        --new_data_offset $((128*1024*1024))
+        --new_data_offset $((128*1024*1024+32768))
-    build/src/disk_tool/vitastor-disk dump-journal --json ./testdata/bin/test_osd$i.bin 4096 $((1024*1024)) $((127*1024*1024)) >./testdata/journal_after_resize.json
+    build/src/disk_tool/vitastor-disk dump-journal --io cached --json ./testdata/bin/test_osd$i.bin 4096 $((1024*1024)) $((127*1024*1024)) >./testdata/journal_after_resize.json
-    build/src/disk_tool/vitastor-disk dump-meta ./testdata/bin/test_osd$i.bin 4096 0 $((1024*1024)) >./testdata/meta_after_resize.json
+    build/src/disk_tool/vitastor-disk dump-meta --io cached ./testdata/bin/test_osd$i.bin 4096 0 $((1024*1024)) >./testdata/meta_after_resize.json
     if ! (cat ./testdata/meta_before_resize.json ./testdata/meta_after_resize.json | \
         jq -e -s 'map([ .entries[] | del(.block) ] | sort_by(.pool, .inode, .stripe)) | .[0] == .[1] and (.[0] | length) > 1000'); then
         format_error "OSD $i metadata corrupted after resizing"
     fi
     if ! (cat ./testdata/journal_before_resize.json ./testdata/journal_after_resize.json | \
-        jq -e -s 'map([ .[].entries[] | del(.crc32, .crc32_prev, .valid, .loc, .start) ]) | .[0] == .[1] and (.[0] | length) > 1'); then
+        jq -e -s 'map([ .[] | del(.crc32, .crc32_prev, .valid, .loc, .start) ]) | .[0] == .[1] and (.[0] | length) > 1'); then
         format_error "OSD $i journal corrupted after resizing"
     fi
 done
@@ -53,7 +54,7 @@ for i in $(seq 1 $OSD_COUNT); do
         --data_device ./testdata/bin/test_osd$i.bin \
         --meta_offset 0 \
         --journal_offset $((1024*1024)) \
-        --data_offset $((128*1024*1024)) >>./testdata/osd$i.log 2>&1 &
+        --data_offset $((128*1024*1024+32768)) >>./testdata/osd$i.log 2>&1 &
     eval OSD${i}_PID=$!
 done
 

tests/test_resize_auto.sh

@@ -0,0 +1,94 @@
+#!/bin/bash -ex
+
+ANTIETCD=1
+. `dirname $0`/common.sh
+
+[[ -e build/src/disk_tool/vitastor-disk-test ]] || ln -s vitastor-disk build/src/disk_tool/vitastor-disk-test
+
+dd if=/dev/zero of=./testdata/bin/test_osd1.bin bs=1 count=1 seek=$((100*1024*1024*1024-1))
+LOOP1=$(sudo losetup --show -f ./testdata/bin/test_osd1.bin)
+trap "kill -9 $(jobs -p) || true; sudo losetup -d $LOOP1"' || true' EXIT
+dd if=/dev/zero of=./testdata/bin/test_meta.bin bs=1 count=1 seek=$((1024*1024*1024-1))
+LOOP2=$(sudo losetup --show -f ./testdata/bin/test_meta.bin)
+trap "kill -9 $(jobs -p) || true; sudo losetup -d $LOOP1 $LOOP2"' || true' EXIT
+
+# also test prepare --hybrid :)
+# non-vitastor random type UUID to prevent udev activation
+mount | grep '/dev type devtmpfs' || sudo mount udev /dev/ -t devtmpfs
+sudo build/src/disk_tool/vitastor-disk-test prepare --no_init 1 --meta_reserve 1x,1M \
+    --block_size 131072 --osd_num 987654 --part_type_uuid 0df42ae0-3695-4395-a957-7d5ff3645c56 \
+    --hybrid --fast-devices $LOOP2 $LOOP1
+
+# write almost empty journal
+node <<EOF > ./testdata/journal.json
+console.log(JSON.stringify([
+    {"type":"start","start":"0x1000"},
+    {"type":"big_write_instant","inode":"0x1000000000001","stripe":"0xc60000","ver":"10","offset":0,"len":131072,"loc":"0x18ffdc0000","bitmap":"ffffffff"}
+]));
+EOF
+sudo build/src/disk_tool/vitastor-disk write-journal ${LOOP1}p1 < ./testdata/journal.json
+sudo build/src/disk_tool/vitastor-disk dump-journal --json --format data ${LOOP1}p1 | jq -S '[ .[] | del(.crc32, .crc32_prev) ]' > ./testdata/j2.json
+jq -S '[ .[] + {"valid":true} ]' < ./testdata/journal.json > ./testdata/j1.json
+diff ./testdata/j1.json ./testdata/j2.json
+
+# write fake metadata items in the end
+DATA_DEV_SIZE=$(sudo blockdev --getsize64 ${LOOP1}p1)
+BLOCK_COUNT=$(((DATA_DEV_SIZE-4096)/128/1024))
+node <<EOF > ./testdata/meta.json
+console.log(JSON.stringify({
+    version: "0.9",
+    meta_block_size: 4096,
+    data_block_size: 131072,
+    bitmap_granularity: 4096,
+    data_csum_type: "none",
+    csum_block_size: 0,
+    entries: [ ...new Array(100).keys() ].map(i => ({
+        block: ($BLOCK_COUNT-100)+i,
+        pool: 1,
+        inode: "0x1",
+        stripe: "0x"+Number(i*0x20000).toString(16),
+        version: 10,
+        bitmap: "ffffffff",
+        ext_bitmap: "ffffffff",
+    })),
+}));
+EOF
+
+# also test write & dump
+sudo build/src/disk_tool/vitastor-disk write-meta ${LOOP1}p1 < ./testdata/meta.json
+sudo build/src/disk_tool/vitastor-disk dump-meta ${LOOP1}p1 > ./testdata/compare.json
+jq -S < ./testdata/meta.json > ./testdata/1.json
+jq -S < ./testdata/compare.json > ./testdata/2.json
+diff ./testdata/1.json ./testdata/2.json
+
+# move journal & meta back, data will become smaller; end indexes should be shifted by -1251
+sudo build/src/disk_tool/vitastor-disk-test resize --move-journal '' --move-meta '' ${LOOP1}p1
+sudo build/src/disk_tool/vitastor-disk dump-meta ${LOOP1}p1 | jq -S > ./testdata/2.json
+jq -S '. + {"entries": [ .entries[] | (. + { "block": (.block-1251) }) ]}' < ./testdata/meta.json > ./testdata/1.json
+diff ./testdata/1.json ./testdata/2.json
+sudo build/src/disk_tool/vitastor-disk dump-journal --json --format data ${LOOP1}p1 | jq -S '[ .[] | del(.crc32, .crc32_prev) ]' > ./testdata/j2.json
+jq -S '[ (.[] + {"valid":true}) | (if .type == "big_write_instant" then . + {"loc":"0x18f6160000"} else . end) ]' < ./testdata/journal.json > ./testdata/j1.json
+diff ./testdata/j1.json ./testdata/j2.json
+
+# move journal & meta out, data will become larger; end indexes should be shifted back by +1251
+sudo build/src/disk_tool/vitastor-disk-test resize --move-journal ${LOOP2}p1 --move-meta ${LOOP2}p2 ${LOOP1}p1
+sudo build/src/disk_tool/vitastor-disk dump-meta ${LOOP1}p1 | jq -S > ./testdata/2.json
+jq -S < ./testdata/meta.json > ./testdata/1.json
+diff ./testdata/1.json ./testdata/2.json
+jq -S '[ .[] + {"valid":true} ]' < ./testdata/journal.json > ./testdata/j1.json
+sudo build/src/disk_tool/vitastor-disk dump-journal --json --format data ${LOOP1}p1 | jq -S '[ .[] | del(.crc32, .crc32_prev) ]' > ./testdata/j2.json
+
+# reduce data device size by exactly 128k * 99 (occupied blocks); exactly 1 should be left in place :)
+sudo build/src/disk_tool/vitastor-disk-test resize --data-size $((DATA_DEV_SIZE-128*1024*99)) ${LOOP1}p1
+sudo build/src/disk_tool/vitastor-disk dump-meta ${LOOP1}p1 | jq -S > ./testdata/2.json
+jq -S '. + {"entries": ([ .entries[] | (. + { "block": (.block | if . > '$BLOCK_COUNT'-100 then .-('$BLOCK_COUNT'-100+1) else '$BLOCK_COUNT'-100 end) }) ] | .[1:] + [ .[0] ])}' < ./testdata/meta.json > ./testdata/1.json
+diff ./testdata/1.json ./testdata/2.json
+jq -S '[ .[] + {"valid":true} ]' < ./testdata/journal.json > ./testdata/j1.json
+sudo build/src/disk_tool/vitastor-disk dump-journal --json --format data ${LOOP1}p1 | jq -S '[ .[] | del(.crc32, .crc32_prev) ]' > ./testdata/j2.json
+
+# extend data device size to maximum
+sudo build/src/disk_tool/vitastor-disk-test resize --data-size max ${LOOP1}p1
+sudo build/src/disk_tool/vitastor-disk dump-meta ${LOOP1}p1 | jq -S > ./testdata/2.json
+diff ./testdata/1.json ./testdata/2.json
+
+format_green OK