Do not allow reweight > 1

New features:

- Support separate OSD cluster network - [osd_cluster_network](https://vitastor.io/docs/config/network.html#osd_cluster_network)
  and, in general, multiple OSD networks, including RDMA
- Add an alternative RDMA implementation via RDMA-CM - [use_rdmacm](https://vitastor.io/docs/config/network.html#use_rdmacm),
  required for iWARP and, maybe, for some IB setups (but not for RoCE)
- Change default PG behaviour to wait for all "up" OSDs to be connected before starting it.
  The old behaviour may be returned by enabling a new [allow_net_split](https://vitastor.io/docs/config/osd.html#allow_net_split)
  option.
- Add a patch for QEMU 9.2

Bug fixes:

- Fix incorrect "has_xxx" PG state names in ls-pgs
- Fix possible QEMU crashes after detaching of Vitastor disks (and update all QEMU builds in Vitastor repos)
- Fix clients sometimes spamming OSDs with infinite reconnections when some PGs are offline
- Fall back to TCP on RDMA connection failures
- Add missing logging of RDMA ibv_modify_qp() errors
- Add a minimum interval for etcd_state_client to reload state

CMakeLists.txt

@@ -2,6 +2,6 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VITASTOR_VERSION "1.11.0")
+set(VITASTOR_VERSION "2.1.0")
 
 add_subdirectory(src)

README-ru.md

@@ -6,7 +6,7 @@
 
 Вернём былую скорость кластерному блочному хранилищу!
 
-Vitastor - распределённая блочная и файловая SDS (программная СХД), прямой аналог Ceph RBD и CephFS,
+Vitastor - распределённая блочная, файловая и объектная SDS (программная СХД), прямой аналог Ceph RBD, CephFS и RGW,
 а также внутренних СХД популярных облачных провайдеров. Однако, в отличие от них, Vitastor
 быстрый и при этом простой. Только пока маленький :-).
 
@@ -46,6 +46,7 @@ Vitastor поддерживает QEMU-драйвер, протоколы NBD и
   - [OpenNebula](docs/installation/opennebula.ru.md)
   - [OpenStack](docs/installation/openstack.ru.md)
   - [Kubernetes CSI](docs/installation/kubernetes.ru.md)
+  - [S3](docs/installation/s3.ru.md)
   - [Сборка из исходных кодов](docs/installation/source.ru.md)
 - Конфигурация
   - [Обзор](docs/config.ru.md)

README.md

@@ -6,7 +6,7 @@
 
 Make Clustered Block Storage Fast Again.
 
-Vitastor is a distributed block and file SDS, direct replacement of Ceph RBD and CephFS,
+Vitastor is a distributed block, file and object SDS, direct replacement of Ceph RBD, CephFS and RGW,
 and also internal SDS's of public clouds. However, in contrast to them, Vitastor is fast
 and simple at the same time. The only thing is it's slightly young :-).
 
@@ -46,6 +46,7 @@ Read more details in the documentation. You can start from here: [Quick Start](d
   - [OpenNebula](docs/installation/opennebula.en.md)
   - [OpenStack](docs/installation/openstack.en.md)
   - [Kubernetes CSI](docs/installation/kubernetes.en.md)
+  - [S3](docs/installation/s3.en.md)
   - [Building from Source](docs/installation/source.en.md)
 - Configuration
   - [Overview](docs/config.en.md)

csi/Dockerfile

@@ -37,8 +37,8 @@ RUN (echo deb http://vitastor.io/debian bookworm main > /etc/apt/sources.list.d/
     wget -q -O /etc/apt/trusted.gpg.d/vitastor.gpg https://vitastor.io/debian/pubkey.gpg && \
     apt-get update && \
     apt-get install -y vitastor-client && \
-    wget https://vitastor.io/archive/qemu/qemu-bookworm-8.1.2%2Bds-1%2Bvitastor1/qemu-utils_8.1.2%2Bds-1%2Bvitastor1_amd64.deb && \
-    wget https://vitastor.io/archive/qemu/qemu-bookworm-8.1.2%2Bds-1%2Bvitastor1/qemu-block-extra_8.1.2%2Bds-1%2Bvitastor1_amd64.deb && \
+    wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-utils_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
+    wget https://vitastor.io/archive/qemu/qemu-bookworm-9.2.2%2Bds-1%2Bvitastor4/qemu-block-extra_9.2.2%2Bds-1%2Bvitastor4_amd64.deb && \
     dpkg -x qemu-utils*.deb tmp1 && \
     dpkg -x qemu-block-extra*.deb tmp1 && \
     cp -a tmp1/usr/bin/qemu-storage-daemon /usr/bin/ && \

csi/Makefile

@@ -1,4 +1,4 @@
-VITASTOR_VERSION ?= v1.11.0
+VITASTOR_VERSION ?= v2.1.0
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v1.11.0
+          image: vitalif/vitastor-csi:v2.1.0
           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:v1.11.0
+          image: vitalif/vitastor-csi:v2.1.0
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "1.11.0"
+    vitastorCSIDriverVersion = "2.1.0"
 )
 
 // Config struct fills the parameters of request or user input

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (1.11.0-1) unstable; urgency=medium
+vitastor (2.1.0-1) unstable; urgency=medium
 
   * Bugfixes
 

debian/patched-qemu.Dockerfile

@@ -10,10 +10,14 @@ ARG REL=
 WORKDIR /root
 
 RUN if [ "$REL" = "buster" -o "$REL" = "bullseye" -o "$REL" = "bookworm" ]; then \
-        echo "deb http://deb.debian.org/debian $REL-backports main" >> /etc/apt/sources.list; \
+        if [ "$REL" = "buster" ]; then \
+            echo "deb http://archive.debian.org/debian $REL-backports main" >> /etc/apt/sources.list; \
+        else \
+            echo "deb http://deb.debian.org/debian $REL-backports main" >> /etc/apt/sources.list; \
+        fi; \
         echo >> /etc/apt/preferences; \
         echo 'Package: *' >> /etc/apt/preferences; \
-        echo "Pin: release a=$REL-backports" >> /etc/apt/preferences; \
+        echo "Pin: release n=$REL-backports" >> /etc/apt/preferences; \
         echo 'Pin-Priority: 500' >> /etc/apt/preferences; \
     fi; \
     grep '^deb ' /etc/apt/sources.list | perl -pe 's/^deb/deb-src/' >> /etc/apt/sources.list; \
@@ -56,7 +60,7 @@ RUN set -e; \
     quilt add block/vitastor.c; \
     cp /root/qemu_driver.c block/vitastor.c; \
     quilt refresh; \
-    V=$(head -n1 debian/changelog | perl -pe 's/5\.2\+dfsg-9/5.2+dfsg-11/; s/^.*\((.*?)(~bpo[\d\+]*)?\).*$/$1/')+vitastor4; \
+    V=$(head -n1 debian/changelog | perl -pe 's/5\.2\+dfsg-9/5.2+dfsg-11/; s/^.*\((.*?)(\+deb\d+u\d+)?(~bpo[\d\+]*)?\).*$/$1/')+vitastor5; \
     if [ "$REL" = bullseye ]; then V=${V}bullseye; fi; \
     DEBEMAIL="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v $V 'Plug Vitastor block driver'; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \

docker/Dockerfile

@@ -3,7 +3,7 @@
 FROM debian:bookworm
 
 ADD etc/apt /etc/apt/
-RUN apt-get update && apt-get -y install vitastor qemu-system-x86 qemu-system-common qemu-block-extra qemu-utils jq nfs-common && apt-get clean
+RUN apt-get update && apt-get -y install vitastor udev systemd qemu-system-x86 qemu-system-common qemu-block-extra qemu-utils jq nfs-common && apt-get clean
 ADD sleep.sh /usr/bin/
 ADD install.sh /usr/bin/
 ADD scripts /opt/scripts/

docker/Makefile

@@ -1,9 +1,9 @@
-VITASTOR_VERSION ?= v1.11.0
+VITASTOR_VERSION ?= v2.1.0
 
 all: build push
 
 build:
-	@docker build --rm -t vitalif/vitastor:$(VITASTOR_VERSION) .
+	@docker build --no-cache --rm -t vitalif/vitastor:$(VITASTOR_VERSION) .
 
 push:
 	@docker push vitalif/vitastor:$(VITASTOR_VERSION)

docker/etc/apt/sources.list.d/vitastor.list

@@ -1 +1,2 @@
 deb http://vitastor.io/debian bookworm main
+deb http://http.debian.net/debian/ bookworm-backports main

docker/etc/systemd/system/vitastor-host.service

@@ -7,8 +7,8 @@ PartOf=vitastor.target
 [Service]
 Restart=always
 EnvironmentFile=/etc/vitastor/docker.conf
-ExecStart=bash -c 'docker run --rm -i -v /etc/vitastor:/etc/vitastor -v /dev:/dev \
-    --privileged --log-driver none --network host --name vitastor vitastor:$VITASTOR_VERSION \
+ExecStart=bash -c 'docker run --rm -i -v /etc/vitastor:/etc/vitastor -v /dev:/dev -v /run:/run \
+    --security-opt seccomp=unconfined --privileged --pid=host --log-driver none --network host --name vitastor vitastor:$VITASTOR_VERSION \
     sleep.sh'
 ExecStartPost=udevadm trigger
 ExecStop=docker stop vitastor

docker/etc/systemd/system/vitastor-osd@.service

@@ -12,7 +12,8 @@ EnvironmentFile=/etc/vitastor/docker.conf
 SyslogIdentifier=vitastor-osd%i
 ExecStart=bash -c 'docker run --rm -i -v /etc/vitastor:/etc/vitastor -v /dev:/dev \
     $(for i in $(ls /dev/vitastor/osd%i-*); do echo --device $i:$i; done) \
-    --log-driver none --network host --ulimit nofile=1048576 --ulimit memlock=-1 $CONTAINER_OPTIONS --name vitastor-osd%i \
+    --log-driver none --network host --ulimit nofile=1048576 --ulimit memlock=-1 \
+    --security-opt seccomp=unconfined $CONTAINER_OPTIONS --name vitastor-osd%i \
     vitastor:$VITASTOR_VERSION vitastor-disk exec-osd /dev/vitastor/osd%i-data'
 ExecStartPre=+docker exec vitastor vitastor-disk pre-exec /dev/vitastor/osd%i-data
 ExecStop=docker stop vitastor-etcd%i

docker/etc/vitastor/docker.conf

@@ -4,7 +4,7 @@
 #
 
 # Desired Vitastor version
-VITASTOR_VERSION=1.11.0
+VITASTOR_VERSION=v2.1.0
 
 # Additional arguments for all containers
 # For example, you may want to specify a custom logging driver here

docs/config/client.en.md

@@ -13,6 +13,7 @@ affect their interaction with the cluster.
 - [client_retry_interval](#client_retry_interval)
 - [client_eio_retry_interval](#client_eio_retry_interval)
 - [client_retry_enospc](#client_retry_enospc)
+- [client_wait_up_timeout](#client_wait_up_timeout)
 - [client_max_dirty_bytes](#client_max_dirty_bytes)
 - [client_max_dirty_ops](#client_max_dirty_ops)
 - [client_enable_writeback](#client_enable_writeback)
@@ -70,6 +71,19 @@ and clients are not blocked and just get EIO error code instead.
 Retry writes on out of space errors to wait until some space is freed on
 OSDs.
 
+## client_wait_up_timeout
+
+- Type: seconds
+- Default: 16
+- Can be changed online: yes
+
+Wait for this number of seconds until PGs are up when doing operations
+which require all PGs to be up. Currently only used by object listings
+in delete and merge-based commands ([vitastor-cli rm](../usage/cli.en.md#rm), merge and so on).
+
+The default value is calculated as `1 + OSD lease timeout`, which is
+`1 + etcd_report_interval + max_etcd_attempts*2*etcd_quick_timeout`.
+
 ## client_max_dirty_bytes
 
 - Type: integer

docs/config/client.ru.md

@@ -13,6 +13,7 @@
 - [client_retry_interval](#client_retry_interval)
 - [client_eio_retry_interval](#client_eio_retry_interval)
 - [client_retry_enospc](#client_retry_enospc)
+- [client_wait_up_timeout](#client_wait_up_timeout)
 - [client_max_dirty_bytes](#client_max_dirty_bytes)
 - [client_max_dirty_ops](#client_max_dirty_ops)
 - [client_enable_writeback](#client_enable_writeback)
@@ -72,6 +73,19 @@ RDMA и хотите повысить пиковую производитель
 Повторять запросы записи, завершившиеся с ошибками нехватки места, т.е.
 ожидать, пока на OSD не освободится место.
 
+## client_wait_up_timeout
+
+- Тип: секунды
+- Значение по умолчанию: 16
+- Можно менять на лету: да
+
+Время ожидания поднятия PG при операциях, требующих активности всех PG.
+В данный момент используется листингами объектов в командах, использующих
+удаление и слияние ([vitastor-cli rm](../usage/cli.ru.md#rm), merge и подобные).
+
+Значение по умолчанию вычисляется как `1 + время lease OSD`, равное
+`1 + etcd_report_interval + max_etcd_attempts*2*etcd_quick_timeout`.
+
 ## client_max_dirty_bytes
 
 - Тип: целое число

docs/config/monitor.en.md

@@ -74,13 +74,13 @@ Grafana dashboard suitable for this exporter is here: [Vitastor-Grafana-6+.json]
 - Type: integer
 - Default: 8060
 
-HTTP port for monitors to listen on (including metrics exporter)
+HTTP port for monitors to listen to (including metrics exporter)
 
 ## mon_http_ip
 
 - Type: string
 
-IP address for monitors to listen on (all addresses by default)
+IP address for monitors to listen to (all addresses by default)
 
 ## mon_https_cert
 

docs/config/network.en.md

@@ -9,9 +9,11 @@
 These parameters apply to clients and OSDs and affect network connection logic
 between clients, OSDs and etcd.
 
-- [tcp_header_buffer_size](#tcp_header_buffer_size)
-- [use_sync_send_recv](#use_sync_send_recv)
+- [osd_network](#osd_network)
+- [osd_cluster_network](#osd_cluster_network)
 - [use_rdma](#use_rdma)
+- [use_rdmacm](#use_rdmacm)
+- [disable_tcp](#disable_tcp)
 - [rdma_device](#rdma_device)
 - [rdma_port_num](#rdma_port_num)
 - [rdma_gid_index](#rdma_gid_index)
@@ -30,38 +32,62 @@ between clients, OSDs and etcd.
 - [etcd_slow_timeout](#etcd_slow_timeout)
 - [etcd_keepalive_timeout](#etcd_keepalive_timeout)
 - [etcd_ws_keepalive_interval](#etcd_ws_keepalive_interval)
+- [etcd_min_reload_interval](#etcd_min_reload_interval)
+- [tcp_header_buffer_size](#tcp_header_buffer_size)
+- [use_sync_send_recv](#use_sync_send_recv)
 
-## tcp_header_buffer_size
+## osd_network
 
-- Type: integer
-- Default: 65536
+- Type: string or array of strings
 
-Size of the buffer used to read data using an additional copy. Vitastor
-packet headers are 128 bytes, payload is always at least 4 KB, so it is
-usually beneficial to try to read multiple packets at once even though
-it requires to copy the data an additional time. The rest of each packet
-is received without an additional copy. You can try to play with this
-parameter and see how it affects random iops and linear bandwidth if you
-want.
+Network mask of public OSD network(s) (IPv4 or IPv6). Each OSD listens to all
+addresses of UP + RUNNING interfaces matching one of these networks, on the
+same port. Port is auto-selected except if [bind_port](osd.en.md#bind_port) is
+explicitly specified. Bind address(es) may also be overridden manually by
+specifying [bind_address](osd.en.md#bind_address). If OSD networks are not specified
+at all, OSD just listens to a wildcard address (0.0.0.0).
 
-## use_sync_send_recv
+## osd_cluster_network
 
-- Type: boolean
-- Default: false
+- Type: string or array of strings
 
-If true, synchronous send/recv syscalls are used instead of io_uring for
-socket communication. Useless for OSDs because they require io_uring anyway,
-but may be required for clients with old kernel versions.
+Network mask of separate network(s) (IPv4 or IPv6) to use for OSD
+cluster connections. I.e. OSDs will always attempt to use these networks
+to connect to other OSDs, while clients will attempt to use networks from
+[osd_network](#osd_network).
 
 ## use_rdma
 
 - Type: boolean
 - Default: true
 
-Try to use RDMA for communication if it's available. Disable if you don't
-want Vitastor to use RDMA. TCP-only clients can also talk to an RDMA-enabled
-cluster, so disabling RDMA may be needed if clients have RDMA devices,
-but they are not connected to the cluster.
+Try to use RDMA through libibverbs for communication if it's available.
+Disable if you don't want Vitastor to use RDMA. TCP-only clients can also
+talk to an RDMA-enabled cluster, so disabling RDMA may be needed if clients
+have RDMA devices, but they are not connected to the cluster.
+
+`use_rdma` works with RoCEv1/RoCEv2 networks, but not with iWARP and,
+maybe, with some Infiniband configurations which require RDMA-CM.
+Consider `use_rdmacm` for such networks.
+
+## use_rdmacm
+
+- Type: boolean
+- Default: true
+
+Use an alternative implementation of RDMA through RDMA-CM (Connection
+Manager). Works with all RDMA networks: Infiniband, iWARP and
+RoCEv1/RoCEv2, and even allows to disable TCP and run only with RDMA.
+OSDs always use random port numbers for RDMA-CM listeners, different
+from their TCP ports. `use_rdma` is automatically disabled when
+`use_rdmacm` is enabled.
+
+## disable_tcp
+
+- Type: boolean
+- Default: true
+
+Fully disable TCP and only use RDMA-CM for OSD communication.
 
 ## rdma_device
 
@@ -92,12 +118,13 @@ PFC (Priority Flow Control) and ECN (Explicit Congestion Notification).
 ## rdma_port_num
 
 - Type: integer
-- Default: 1
 
 RDMA device port number to use. Only for devices that have more than 1 port.
 See `phys_port_cnt` in `ibv_devinfo -v` output to determine how many ports
 your device has.
 
+Not relevant for RDMA-CM (use_rdmacm).
+
 ## rdma_gid_index
 
 - Type: integer
@@ -113,13 +140,14 @@ GID auto-selection is unsupported with libibverbs < v32.
 
 A correct rdma_gid_index for RoCEv2 is usually 1 (IPv6) or 3 (IPv4).
 
+Not relevant for RDMA-CM (use_rdmacm).
+
 ## rdma_mtu
 
 - Type: integer
-- Default: 4096
 
-RDMA Path MTU to use. Must be 1024, 2048 or 4096. There is usually no
-sense to change it from the default 4096.
+RDMA Path MTU to use. Must be 1024, 2048 or 4096. Default is to use the
+RDMA device's MTU.
 
 ## rdma_max_sge
 
@@ -261,3 +289,35 @@ etcd_report_interval to guarantee that keepalive actually works.
 
 etcd websocket ping interval required to keep the connection alive and
 detect disconnections quickly.
+
+## etcd_min_reload_interval
+
+- Type: milliseconds
+- Default: 1000
+- Can be changed online: yes
+
+Minimum interval for full etcd state reload. Introduced to prevent
+excessive load on etcd during outages when etcd can't keep up with event
+streams and cancels them.
+
+## tcp_header_buffer_size
+
+- Type: integer
+- Default: 65536
+
+Size of the buffer used to read data using an additional copy. Vitastor
+packet headers are 128 bytes, payload is always at least 4 KB, so it is
+usually beneficial to try to read multiple packets at once even though
+it requires to copy the data an additional time. The rest of each packet
+is received without an additional copy. You can try to play with this
+parameter and see how it affects random iops and linear bandwidth if you
+want.
+
+## use_sync_send_recv
+
+- Type: boolean
+- Default: false
+
+If true, synchronous send/recv syscalls are used instead of io_uring for
+socket communication. Useless for OSDs because they require io_uring anyway,
+but may be required for clients with old kernel versions.

docs/config/network.ru.md

@@ -9,9 +9,11 @@
 Данные параметры используются клиентами и OSD и влияют на логику сетевого
 взаимодействия между клиентами, OSD, а также etcd.
 
-- [tcp_header_buffer_size](#tcp_header_buffer_size)
-- [use_sync_send_recv](#use_sync_send_recv)
+- [osd_network](#osd_network)
+- [osd_cluster_network](#osd_cluster_network)
 - [use_rdma](#use_rdma)
+- [use_rdmacm](#use_rdmacm)
+- [disable_tcp](#disable_tcp)
 - [rdma_device](#rdma_device)
 - [rdma_port_num](#rdma_port_num)
 - [rdma_gid_index](#rdma_gid_index)
@@ -30,41 +32,62 @@
 - [etcd_slow_timeout](#etcd_slow_timeout)
 - [etcd_keepalive_timeout](#etcd_keepalive_timeout)
 - [etcd_ws_keepalive_interval](#etcd_ws_keepalive_interval)
+- [etcd_min_reload_interval](#etcd_min_reload_interval)
+- [tcp_header_buffer_size](#tcp_header_buffer_size)
+- [use_sync_send_recv](#use_sync_send_recv)
 
-## tcp_header_buffer_size
+## osd_network
 
-- Тип: целое число
-- Значение по умолчанию: 65536
+- Тип: строка или массив строк
 
-Размер буфера для чтения данных с дополнительным копированием. Пакеты
-Vitastor содержат 128-байтные заголовки, за которыми следуют данные размером
-от 4 КБ и для мелких операций ввода-вывода обычно выгодно за 1 вызов читать
-сразу несколько пакетов, даже не смотря на то, что это требует лишний раз
-скопировать данные. Часть каждого пакета за пределами значения данного
-параметра читается без дополнительного копирования. Вы можете попробовать
-поменять этот параметр и посмотреть, как он влияет на производительность
-случайного и линейного доступа.
+Маски подсетей (IPv4 или IPv6) публичной сети или сетей OSD. Каждый OSD слушает
+один и тот же порт на всех адресах поднятых (UP + RUNNING) сетевых интерфейсов,
+соответствующих одной из указанных сетей. Порт выбирается автоматически, если
+только [bind_port](osd.ru.md#bind_port) не задан явно. Адреса для подключений можно
+также переопределить явно, задав [bind_address](osd.ru.md#bind_address). Если сети OSD
+не заданы вообще, OSD слушает все адреса (0.0.0.0).
 
-## use_sync_send_recv
+## osd_cluster_network
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+- Тип: строка или массив строк
 
-Если установлено в истину, то вместо io_uring для передачи данных по сети
-будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
-это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
-принципе, это может применяться для клиентов со старыми версиями ядра.
+Маски подсетей (IPv4 или IPv6) отдельной кластерной сети или сетей OSD.
+То есть, OSD будут всегда стараться использовать эти сети для соединений
+с другими OSD, а клиенты будут стараться использовать сети из [osd_network](#osd_network).
 
 ## use_rdma
 
 - Тип: булево (да/нет)
 - Значение по умолчанию: true
 
-Пытаться использовать RDMA для связи при наличии доступных устройств.
-Отключите, если вы не хотите, чтобы Vitastor использовал RDMA.
-TCP-клиенты также могут работать с RDMA-кластером, так что отключать
-RDMA может быть нужно только если у клиентов есть RDMA-устройства,
-но они не имеют соединения с кластером Vitastor.
+Попробовать использовать RDMA через libibverbs для связи при наличии
+доступных устройств. Отключите, если вы не хотите, чтобы Vitastor
+использовал RDMA. TCP-клиенты также могут работать с RDMA-кластером,
+так что отключать RDMA может быть нужно, только если у клиентов есть
+RDMA-устройства, но они не имеют соединения с кластером Vitastor.
+
+`use_rdma` работает с RoCEv1/RoCEv2 сетями, но не работает с iWARP и
+может не работать с частью конфигураций Infiniband, требующих RDMA-CM.
+Рассмотрите включение `use_rdmacm` для таких сетей.
+
+## use_rdmacm
+
+- Тип: булево (да/нет)
+- Значение по умолчанию: true
+
+Использовать альтернативную реализацию RDMA на основе RDMA-CM (Connection
+Manager). Работает со всеми типами RDMA-сетей: Infiniband, iWARP и
+RoCEv1/RoCEv2, и даже позволяет полностью отключить TCP и работать
+только на RDMA. OSD используют случайные номера портов для ожидания
+соединений через RDMA-CM, отличающиеся от их TCP-портов. Также при
+включении `use_rdmacm` автоматически отключается опция `use_rdma`.
+
+## disable_tcp
+
+- Тип: булево (да/нет)
+- Значение по умолчанию: true
+
+Полностью отключить TCP и использовать только RDMA-CM для соединений с OSD.
 
 ## rdma_device
 
@@ -96,13 +119,14 @@ Control) и ECN (Explicit Congestion Notification).
 ## rdma_port_num
 
 - Тип: целое число
-- Значение по умолчанию: 1
 
 Номер порта RDMA-устройства, который следует использовать. Имеет смысл
 только для устройств, у которых более 1 порта. Чтобы узнать, сколько портов
 у вашего адаптера, посмотрите `phys_port_cnt` в выводе команды
 `ibv_devinfo -v`.
 
+Опция неприменима к RDMA-CM (use_rdmacm).
+
 ## rdma_gid_index
 
 - Тип: целое число
@@ -119,13 +143,14 @@ libibverbs < v32.
 
 Правильный rdma_gid_index для RoCEv2, как правило, 1 (IPv6) или 3 (IPv4).
 
+Опция неприменима к RDMA-CM (use_rdmacm).
+
 ## rdma_mtu
 
 - Тип: целое число
-- Значение по умолчанию: 4096
 
 Максимальная единица передачи (Path MTU) для RDMA. Должно быть равно 1024,
-2048 или 4096. Обычно нет смысла менять значение по умолчанию, равное 4096.
+2048 или 4096. По умолчанию используется значение MTU RDMA-устройства.
 
 ## rdma_max_sge
 
@@ -271,3 +296,37 @@ etcd_report_interval, чтобы keepalive гарантированно рабо
 - Можно менять на лету: да
 
 Интервал проверки живости вебсокет-подключений к etcd.
+
+## etcd_min_reload_interval
+
+- Тип: миллисекунды
+- Значение по умолчанию: 1000
+- Можно менять на лету: да
+
+Минимальный интервал полной перезагрузки состояния из etcd. Добавлено для
+предотвращения избыточной нагрузки на etcd во время отказов, когда etcd не
+успевает рассылать потоки событий и отменяет их.
+
+## tcp_header_buffer_size
+
+- Тип: целое число
+- Значение по умолчанию: 65536
+
+Размер буфера для чтения данных с дополнительным копированием. Пакеты
+Vitastor содержат 128-байтные заголовки, за которыми следуют данные размером
+от 4 КБ и для мелких операций ввода-вывода обычно выгодно за 1 вызов читать
+сразу несколько пакетов, даже не смотря на то, что это требует лишний раз
+скопировать данные. Часть каждого пакета за пределами значения данного
+параметра читается без дополнительного копирования. Вы можете попробовать
+поменять этот параметр и посмотреть, как он влияет на производительность
+случайного и линейного доступа.
+
+## use_sync_send_recv
+
+- Тип: булево (да/нет)
+- Значение по умолчанию: false
+
+Если установлено в истину, то вместо io_uring для передачи данных по сети
+будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
+это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
+принципе, это может применяться для клиентов со старыми версиями ядра.

docs/config/osd.en.md

@@ -7,16 +7,15 @@
 # Runtime OSD Parameters
 
 These parameters only apply to OSDs, are not fixed at the moment of OSD drive
-initialization and can be changed - either with an OSD restart or, for some of
-them, even without restarting by updating configuration in etcd.
+initialization and can be changed - in /etc/vitastor/vitastor.conf or [vitastor-disk update-sb](../usage/disk.en.md#update-sb)
+with an OSD restart or, for some of them, even without restarting by updating configuration in etcd.
 
+- [bind_address](#bind_address)
+- [bind_port](#bind_port)
 - [osd_iothread_count](#osd_iothread_count)
 - [etcd_report_interval](#etcd_report_interval)
 - [etcd_stats_interval](#etcd_stats_interval)
 - [run_primary](#run_primary)
-- [osd_network](#osd_network)
-- [bind_address](#bind_address)
-- [bind_port](#bind_port)
 - [autosync_interval](#autosync_interval)
 - [autosync_writes](#autosync_writes)
 - [recovery_queue_depth](#recovery_queue_depth)
@@ -61,6 +60,26 @@ them, even without restarting by updating configuration in etcd.
 - [recovery_tune_agg_interval](#recovery_tune_agg_interval)
 - [recovery_tune_sleep_min_us](#recovery_tune_sleep_min_us)
 - [recovery_tune_sleep_cutoff_us](#recovery_tune_sleep_cutoff_us)
+- [discard_on_start](#discard_on_start)
+- [min_discard_size](#min_discard_size)
+- [allow_net_split](#allow_net_split)
+
+## bind_address
+
+- Type: string or array of strings
+
+Instead of the network masks ([osd_network](network.en.md#osd_network) and
+[osd_cluster_network](network.en.md#osd_cluster_network)), you can also set
+OSD listen addresses explicitly using this parameter. May be useful if you
+want to start OSDs on interfaces that are not UP + RUNNING.
+
+## bind_port
+
+- Type: integer
+
+By default, OSDs pick random ports to use for incoming connections
+automatically. With this option you can set a specific port for a specific
+OSD by hand.
 
 ## osd_iothread_count
 
@@ -104,34 +123,6 @@ debugging purposes. It's possible to implement additional feature for the
 monitor which may allow to separate primary and secondary OSDs, but it's
 unclear why anyone could need it, so it's not implemented.
 
-## osd_network
-
-- Type: string or array of strings
-
-Network mask of the network (IPv4 or IPv6) to use for OSDs. Note that
-although it's possible to specify multiple networks here, this does not
-mean that OSDs will create multiple listening sockets - they'll only
-pick the first matching address of an UP + RUNNING interface. Separate
-networks for cluster and client connections are also not implemented, but
-they are mostly useless anyway, so it's not a big deal.
-
-## bind_address
-
-- Type: string
-- Default: 0.0.0.0
-
-Instead of the network mask, you can also set OSD listen address explicitly
-using this parameter. May be useful if you want to start OSDs on interfaces
-that are not UP + RUNNING.
-
-## bind_port
-
-- Type: integer
-
-By default, OSDs pick random ports to use for incoming connections
-automatically. With this option you can set a specific port for a specific
-OSD by hand.
-
 ## autosync_interval
 
 - Type: seconds
@@ -316,7 +307,7 @@ for hot data and slower disks - HDDs and maybe SATA SSDs - but will slightly
 decrease write performance for fast disks because page cache is an overhead
 itself.
 
-Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+Choose "directsync" to use [immediate_commit](layout-cluster.en.md#immediate_commit)
 (which requires disable_data_fsync) with drives having write-back cache
 which can't be turned off, for example, Intel Optane. Also note that *some*
 desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
@@ -629,3 +620,30 @@ are changed to 0.
 
 Maximum possible value for auto-tuned recovery_sleep_us. Higher values
 are treated as outliers and ignored in aggregation.
+
+## discard_on_start
+
+- Type: boolean
+
+Discard (SSD TRIM) unused data device blocks on every OSD startup.
+
+## min_discard_size
+
+- Type: integer
+- Default: 1048576
+
+Minimum consecutive block size to TRIM it.
+
+## allow_net_split
+
+- Type: boolean
+- Default: false
+
+Allow "safe" cases of network splits/partitions - allow to start PGs without
+connections to some OSDs currently registered as alive in etcd, if the number
+of actually connected PG OSDs is at least pg_minsize. That is, allow some OSDs to lose
+connectivity with some other OSDs as long as it doesn't break pg_minsize guarantees.
+The downside is that it increases the probability of writing data into just pg_minsize
+OSDs during failover which can lead to PGs becoming incomplete after additional outages.
+
+The old behaviour in versions up to 2.0.0 was equal to enabled allow_net_split.

docs/config/osd.ru.md

@@ -8,16 +8,15 @@
 
 Данные параметры используются только OSD, но, в отличие от дисковых параметров,
 не фиксируются в момент инициализации дисков OSD и могут быть изменены в любой
-момент с помощью перезапуска OSD, а некоторые и без перезапуска, с помощью
-изменения конфигурации в etcd.
+момент с перезапуском OSD в /etc/vitastor/vitastor.conf или [vitastor-disk update-sb](../usage/disk.ru.md#update-sb),
+а некоторые и без перезапуска, с помощью изменения конфигурации в etcd.
 
+- [bind_address](#bind_address)
+- [bind_port](#bind_port)
 - [osd_iothread_count](#osd_iothread_count)
 - [etcd_report_interval](#etcd_report_interval)
 - [etcd_stats_interval](#etcd_stats_interval)
 - [run_primary](#run_primary)
-- [osd_network](#osd_network)
-- [bind_address](#bind_address)
-- [bind_port](#bind_port)
 - [autosync_interval](#autosync_interval)
 - [autosync_writes](#autosync_writes)
 - [recovery_queue_depth](#recovery_queue_depth)
@@ -62,6 +61,26 @@
 - [recovery_tune_agg_interval](#recovery_tune_agg_interval)
 - [recovery_tune_sleep_min_us](#recovery_tune_sleep_min_us)
 - [recovery_tune_sleep_cutoff_us](#recovery_tune_sleep_cutoff_us)
+- [discard_on_start](#discard_on_start)
+- [min_discard_size](#min_discard_size)
+- [allow_net_split](#allow_net_split)
+
+## bind_address
+
+- Тип: строка или массив строк
+
+Вместо использования масок подсети ([osd_network](network.ru.md#osd_network) и
+[osd_cluster_network](network.ru.md#osd_cluster_network)), вы также можете явно
+задать адрес(а), на которых будут ожидать соединений OSD, с помощью данного
+параметра. Это может быть полезно, например, чтобы запускать OSD на неподнятых
+интерфейсах (не UP + RUNNING).
+
+## bind_port
+
+- Тип: целое число
+
+По умолчанию OSD сами выбирают случайные порты для входящих подключений.
+С помощью данной опции вы можете задать порт для отдельного OSD вручную.
 
 ## osd_iothread_count
 
@@ -107,34 +126,6 @@ max_etcd_attempts * etcd_quick_timeout.
 первичные OSD от вторичных, но пока не понятно, зачем это может кому-то
 понадобиться, поэтому это не реализовано.
 
-## osd_network
-
-- Тип: строка или массив строк
-
-Маска подсети (IPv4 или IPv6) для использования для соединений с OSD.
-Имейте в виду, что хотя сейчас и можно передать в этот параметр несколько
-подсетей, это не означает, что OSD будут создавать несколько слушающих
-сокетов - они лишь будут выбирать адрес первого поднятого (состояние UP +
-RUNNING), подходящий под заданную маску. Также не реализовано разделение
-кластерной и публичной сетей OSD. Правда, от него обычно всё равно довольно
-мало толку, так что особенной проблемы в этом нет.
-
-## bind_address
-
-- Тип: строка
-- Значение по умолчанию: 0.0.0.0
-
-Этим параметром можно явным образом задать адрес, на котором будет ожидать
-соединений OSD (вместо использования маски подсети). Может быть полезно,
-например, чтобы запускать OSD на неподнятых интерфейсах (не UP + RUNNING).
-
-## bind_port
-
-- Тип: целое число
-
-По умолчанию OSD сами выбирают случайные порты для входящих подключений.
-С помощью данной опции вы можете задать порт для отдельного OSD вручную.
-
 ## autosync_interval
 
 - Тип: секунды
@@ -660,3 +651,31 @@ EC (кодов коррекции ошибок) с более, чем 1 диск
 Максимальное возможное значение авто-подстроенного recovery_sleep_us.
 Большие значения считаются случайными выбросами и игнорируются в
 усреднении.
+
+## discard_on_start
+
+- Тип: булево (да/нет)
+
+Освобождать (SSD TRIM) неиспользуемые блоки диска данных при каждом запуске OSD.
+
+## min_discard_size
+
+- Тип: целое число
+- Значение по умолчанию: 1048576
+
+Минимальный размер последовательного блока данных, чтобы освобождать его через TRIM.
+
+## allow_net_split
+
+- Тип: булево (да/нет)
+- Значение по умолчанию: false
+
+Разрешить "безопасные" случаи разделений сети - разрешить активировать PG без
+соединений к некоторым OSD, помеченным активными в etcd, если общее число активных
+OSD в PG составляет как минимум pg_minsize. То есть, разрешать некоторым OSD терять
+соединения с некоторыми другими OSD, если это не нарушает гарантий pg_minsize.
+Минус такого разрешения в том, что оно повышает вероятность записи данных ровно в
+pg_minsize OSD во время переключений, что может потом привести к тому, что PG станут
+неполными (incomplete), если упадут ещё какие-то OSD.
+
+Старое поведение в версиях до 2.0.0 было идентично включённому allow_net_split.

docs/config/pool.en.md

@@ -43,7 +43,7 @@ Parameters:
 - [osd_tags](#osd_tags)
 - [primary_affinity_tags](#primary_affinity_tags)
 - [scrub_interval](#scrub_interval)
-- [used_for_fs](#used_for_fs)
+- [used_for_app](#used_for_app)
 
 Examples:
 
@@ -189,6 +189,9 @@ So, pg_minsize regulates the number of failures that a pool can tolerate
 without temporary downtime for [osd_out_time](monitor.en.md#osd_out_time),
 but at a cost of slightly reduced storage reliability.
 
+See also [allow_net_split](osd.en.md#allow_net_split) and
+[PG state descriptions](../usage/admin.en.md#pg-states).
+
 FIXME: pg_minsize behaviour may be changed in the future to only make PGs
 read-only instead of deactivating them.
 
@@ -377,24 +380,37 @@ of the OSDs containing a data chunk for a PG.
 Automatic scrubbing interval for this pool. Overrides
 [global scrub_interval setting](osd.en.md#scrub_interval).
 
-## used_for_fs
+## used_for_app
 
 - Type: string
 
-If non-empty, the pool is marked as used for VitastorFS with metadata stored
-in block image (regular Vitastor volume) named as the value of this pool parameter.
+If non-empty, the pool is marked as used for a separate application, for example,
+VitastorFS or S3, which allocates Vitastor volume IDs by itself and does not use
+image/inode metadata in etcd.
 
-When a pool is marked as used for VitastorFS, regular block volume creation in it
+When a pool is marked as used for such app, regular block volume creation in it
 is disabled (vitastor-cli refuses to create images without --force) to protect
-the user from block volume and FS file ID collisions and data loss.
+the user from block volume and FS/S3 volume ID collisions and data loss.
 
-[vitastor-nfs](../usage/nfs.ru.md), in its turn, refuses to use pools not marked
+Also such pools do not calculate per-inode space usage statistics in etcd because
+using it for an external application implies that it may contain a very large
+number of volumes and their statistics may take too much space in etcd.
+
+Setting used_for_app to `fs:<name>` tells Vitastor that the pool is used for VitastorFS
+with VitastorKV metadata base stored in a block image (regular Vitastor volume) named
+`<name>`.
+
+[vitastor-nfs](../usage/nfs.en.md), in its turn, refuses to use pools not marked
 for the corresponding FS when starting. This also implies that you can use one
 pool only for one VitastorFS.
 
-The second thing that is disabled for VitastorFS pools is reporting per-inode space
-usage statistics in etcd because a FS pool may store a very large number of files
-and statistics for them all would take a lot of space in etcd.
+If you plan to use the pool for S3, set its used_for_app to `s3:<name>`. `<name>` may
+be basically anything you want (for example, `s3:standard`) - it's not validated
+by Vitastor S3 components in any way.
+
+All other values except prefixed with `fs:` or `s3:` may be used freely and don't
+mean anything special for Vitastor core components. For now, you can use them as
+you wish.
 
 # Examples
 

docs/config/pool.ru.md

@@ -42,7 +42,7 @@
 - [osd_tags](#osd_tags)
 - [primary_affinity_tags](#primary_affinity_tags)
 - [scrub_interval](#scrub_interval)
-- [used_for_fs](#used_for_fs)
+- [used_for_app](#used_for_app)
 
 Примеры:
 
@@ -256,7 +256,7 @@ PG в Vitastor эферемерны, то есть вы можете менят
 
 ## raw_placement
 
-- Type: string
+- Тип: строка
 
 Низкоуровневые правила генерации PG в форме DSL (доменно-специфичного языка).
 Используйте, только если действительно знаете, зачем вам это надо :)
@@ -383,26 +383,42 @@ OSD с "all".
 Интервал скраба, то есть, автоматической фоновой проверки данных для данного пула.
 Переопределяет [глобальную настройку scrub_interval](osd.ru.md#scrub_interval).
 
-## used_for_fs
+## used_for_app
 
-- Type: string
+- Тип: строка
 
-Если непусто, пул помечается как используемый для файловой системы VitastorFS с
-метаданными, хранимыми в блочном образе Vitastor с именем, равным значению
-этого параметра.
+Если непусто, пул помечается как используемый для отдельного приложения, например,
+для VitastorFS или S3, которое распределяет ID образов в пуле само и не использует
+метаданные образов/инодов в etcd.
 
-Когда пул помечается как используемый для VitastorFS, создание обычных блочных
-образов в нём отключается (vitastor-cli отказывается создавать образы без --force),
-чтобы защитить пользователя от коллизий ID файлов и блочных образов и, таким
-образом, от потери данных.
+Когда пул помечается используемым для такого приложения, создание обычных блочных
+образов в нём запрещается (vitastor-cli отказывается создавать образы без --force),
+чтобы защитить пользователя от коллизий ID блочных образов и томов ФС/S3, и,
+таким образом, от потери данных.
+
+Также для таких пулов отключается передача статистики в etcd по отдельным инодам,
+так как использование для внешнего приложения подразумевает, что пул может содержать
+очень много томов и их статистика может занять слишком много места в etcd.
+
+Установка used_for_app в значение `fs:<name>` сообщает о том, что пул используется
+для VitastorFS с базой метаданных VitastorKV, хранимой в блочном образе с именем
+`<name>`.
 
 [vitastor-nfs](../usage/nfs.ru.md), в свою очередь, при запуске отказывается
-использовать для ФС пулы, не выделенные для неё. Это также означает, что один
-пул может использоваться только для одной VitastorFS.
+использовать для ФС пулы, не помеченные, как используемые для неё. Это также
+означает, что один пул может использоваться только для одной VitastorFS.
 
-Также для ФС-пулов отключается передача статистики в etcd по отдельным инодам,
-так как ФС-пул может содержать очень много файлов и статистика по ним всем
-заняла бы очень много места в etcd.
+Если же вы планируете использовать пул для данных S3, установите его used_for_app
+в значение `s3:<name>`, где `<name>` - любое название по вашему усмотрению
+(например, `s3:standard`) - конкретное содержимое `<name>` пока никак не проверяется
+компонентами Vitastor S3.
+
+Смотрите также [allow_net_split](osd.ru.md#allow_net_split) и
+[документацию по состояниям PG](../usage/admin.ru.md#состояния-pg).
+
+Все остальные значения used_for_app, кроме начинающихся на `fs:` или `s3:`, не
+означают ничего особенного для основных компонентов Vitastor. Поэтому сейчас вы
+можете использовать их свободно любым желаемым способом.
 
 # Примеры
 

docs/config/src/monitor.yml

@@ -75,11 +75,11 @@
 - name: mon_http_port
   type: int
   default: 8060
-  info: HTTP port for monitors to listen on (including metrics exporter)
+  info: HTTP port for monitors to listen to (including metrics exporter)
   info_ru: Порт, на котором мониторы принимают HTTP-соединения (в том числе для отдачи метрик)
 - name: mon_http_ip
   type: string
-  info: IP address for monitors to listen on (all addresses by default)
+  info: IP address for monitors to listen to (all addresses by default)
   info_ru: IP-адрес, на котором мониторы принимают HTTP-соединения (по умолчанию все адреса)
 - name: mon_https_cert
   type: string

docs/config/src/network.yml

@@ -1,49 +1,78 @@
-- name: tcp_header_buffer_size
-  type: int
-  default: 65536
+- name: osd_network
+  type: string or array of strings
+  type_ru: строка или массив строк
   info: |
-    Size of the buffer used to read data using an additional copy. Vitastor
-    packet headers are 128 bytes, payload is always at least 4 KB, so it is
-    usually beneficial to try to read multiple packets at once even though
-    it requires to copy the data an additional time. The rest of each packet
-    is received without an additional copy. You can try to play with this
-    parameter and see how it affects random iops and linear bandwidth if you
-    want.
+    Network mask of public OSD network(s) (IPv4 or IPv6). Each OSD listens to all
+    addresses of UP + RUNNING interfaces matching one of these networks, on the
+    same port. Port is auto-selected except if [bind_port](osd.en.md#bind_port) is
+    explicitly specified. Bind address(es) may also be overridden manually by
+    specifying [bind_address](osd.en.md#bind_address). If OSD networks are not specified
+    at all, OSD just listens to a wildcard address (0.0.0.0).
   info_ru: |
-    Размер буфера для чтения данных с дополнительным копированием. Пакеты
-    Vitastor содержат 128-байтные заголовки, за которыми следуют данные размером
-    от 4 КБ и для мелких операций ввода-вывода обычно выгодно за 1 вызов читать
-    сразу несколько пакетов, даже не смотря на то, что это требует лишний раз
-    скопировать данные. Часть каждого пакета за пределами значения данного
-    параметра читается без дополнительного копирования. Вы можете попробовать
-    поменять этот параметр и посмотреть, как он влияет на производительность
-    случайного и линейного доступа.
-- name: use_sync_send_recv
-  type: bool
-  default: false
+    Маски подсетей (IPv4 или IPv6) публичной сети или сетей OSD. Каждый OSD слушает
+    один и тот же порт на всех адресах поднятых (UP + RUNNING) сетевых интерфейсов,
+    соответствующих одной из указанных сетей. Порт выбирается автоматически, если
+    только [bind_port](osd.ru.md#bind_port) не задан явно. Адреса для подключений можно
+    также переопределить явно, задав [bind_address](osd.ru.md#bind_address). Если сети OSD
+    не заданы вообще, OSD слушает все адреса (0.0.0.0).
+- name: osd_cluster_network
+  type: string or array of strings
+  type_ru: строка или массив строк
   info: |
-    If true, synchronous send/recv syscalls are used instead of io_uring for
-    socket communication. Useless for OSDs because they require io_uring anyway,
-    but may be required for clients with old kernel versions.
+    Network mask of separate network(s) (IPv4 or IPv6) to use for OSD
+    cluster connections. I.e. OSDs will always attempt to use these networks
+    to connect to other OSDs, while clients will attempt to use networks from
+    [osd_network](#osd_network).
   info_ru: |
-    Если установлено в истину, то вместо io_uring для передачи данных по сети
-    будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
-    это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
-    принципе, это может применяться для клиентов со старыми версиями ядра.
+    Маски подсетей (IPv4 или IPv6) отдельной кластерной сети или сетей OSD.
+    То есть, OSD будут всегда стараться использовать эти сети для соединений
+    с другими OSD, а клиенты будут стараться использовать сети из [osd_network](#osd_network).
 - name: use_rdma
   type: bool
   default: true
   info: |
-    Try to use RDMA for communication if it's available. Disable if you don't
-    want Vitastor to use RDMA. TCP-only clients can also talk to an RDMA-enabled
-    cluster, so disabling RDMA may be needed if clients have RDMA devices,
-    but they are not connected to the cluster.
+    Try to use RDMA through libibverbs for communication if it's available.
+    Disable if you don't want Vitastor to use RDMA. TCP-only clients can also
+    talk to an RDMA-enabled cluster, so disabling RDMA may be needed if clients
+    have RDMA devices, but they are not connected to the cluster.
+
+    `use_rdma` works with RoCEv1/RoCEv2 networks, but not with iWARP and,
+    maybe, with some Infiniband configurations which require RDMA-CM.
+    Consider `use_rdmacm` for such networks.
   info_ru: |
-    Пытаться использовать RDMA для связи при наличии доступных устройств.
-    Отключите, если вы не хотите, чтобы Vitastor использовал RDMA.
-    TCP-клиенты также могут работать с RDMA-кластером, так что отключать
-    RDMA может быть нужно только если у клиентов есть RDMA-устройства,
-    но они не имеют соединения с кластером Vitastor.
+    Попробовать использовать RDMA через libibverbs для связи при наличии
+    доступных устройств. Отключите, если вы не хотите, чтобы Vitastor
+    использовал RDMA. TCP-клиенты также могут работать с RDMA-кластером,
+    так что отключать RDMA может быть нужно, только если у клиентов есть
+    RDMA-устройства, но они не имеют соединения с кластером Vitastor.
+
+    `use_rdma` работает с RoCEv1/RoCEv2 сетями, но не работает с iWARP и
+    может не работать с частью конфигураций Infiniband, требующих RDMA-CM.
+    Рассмотрите включение `use_rdmacm` для таких сетей.
+- name: use_rdmacm
+  type: bool
+  default: true
+  info: |
+    Use an alternative implementation of RDMA through RDMA-CM (Connection
+    Manager). Works with all RDMA networks: Infiniband, iWARP and
+    RoCEv1/RoCEv2, and even allows to disable TCP and run only with RDMA.
+    OSDs always use random port numbers for RDMA-CM listeners, different
+    from their TCP ports. `use_rdma` is automatically disabled when
+    `use_rdmacm` is enabled.
+  info_ru: |
+    Использовать альтернативную реализацию RDMA на основе RDMA-CM (Connection
+    Manager). Работает со всеми типами RDMA-сетей: Infiniband, iWARP и
+    RoCEv1/RoCEv2, и даже позволяет полностью отключить TCP и работать
+    только на RDMA. OSD используют случайные номера портов для ожидания
+    соединений через RDMA-CM, отличающиеся от их TCP-портов. Также при
+    включении `use_rdmacm` автоматически отключается опция `use_rdma`.
+- name: disable_tcp
+  type: bool
+  default: true
+  info: |
+    Fully disable TCP and only use RDMA-CM for OSD communication.
+  info_ru: |
+    Полностью отключить TCP и использовать только RDMA-CM для соединений с OSD.
 - name: rdma_device
   type: string
   info: |
@@ -93,16 +122,19 @@
     Control) и ECN (Explicit Congestion Notification).
 - name: rdma_port_num
   type: int
-  default: 1
   info: |
     RDMA device port number to use. Only for devices that have more than 1 port.
     See `phys_port_cnt` in `ibv_devinfo -v` output to determine how many ports
     your device has.
+
+    Not relevant for RDMA-CM (use_rdmacm).
   info_ru: |
     Номер порта RDMA-устройства, который следует использовать. Имеет смысл
     только для устройств, у которых более 1 порта. Чтобы узнать, сколько портов
     у вашего адаптера, посмотрите `phys_port_cnt` в выводе команды
     `ibv_devinfo -v`.
+
+    Опция неприменима к RDMA-CM (use_rdmacm).
 - name: rdma_gid_index
   type: int
   info: |
@@ -116,6 +148,8 @@
     GID auto-selection is unsupported with libibverbs < v32.
 
     A correct rdma_gid_index for RoCEv2 is usually 1 (IPv6) or 3 (IPv4).
+
+    Not relevant for RDMA-CM (use_rdmacm).
   info_ru: |
     Номер глобального идентификатора адреса RDMA-устройства, который следует
     использовать. Разным gid_index могут соответствовать разные протоколы связи:
@@ -128,15 +162,16 @@
     libibverbs < v32.
 
     Правильный rdma_gid_index для RoCEv2, как правило, 1 (IPv6) или 3 (IPv4).
+
+    Опция неприменима к RDMA-CM (use_rdmacm).
 - name: rdma_mtu
   type: int
-  default: 4096
   info: |
-    RDMA Path MTU to use. Must be 1024, 2048 or 4096. There is usually no
-    sense to change it from the default 4096.
+    RDMA Path MTU to use. Must be 1024, 2048 or 4096. Default is to use the
+    RDMA device's MTU.
   info_ru: |
     Максимальная единица передачи (Path MTU) для RDMA. Должно быть равно 1024,
-    2048 или 4096. Обычно нет смысла менять значение по умолчанию, равное 4096.
+    2048 или 4096. По умолчанию используется значение MTU RDMA-устройства.
 - name: rdma_max_sge
   type: int
   default: 128
@@ -306,3 +341,47 @@
     detect disconnections quickly.
   info_ru: |
     Интервал проверки живости вебсокет-подключений к etcd.
+- name: etcd_min_reload_interval
+  type: ms
+  default: 1000
+  online: true
+  info: |
+    Minimum interval for full etcd state reload. Introduced to prevent
+    excessive load on etcd during outages when etcd can't keep up with event
+    streams and cancels them.
+  info_ru: |
+    Минимальный интервал полной перезагрузки состояния из etcd. Добавлено для
+    предотвращения избыточной нагрузки на etcd во время отказов, когда etcd не
+    успевает рассылать потоки событий и отменяет их.
+- name: tcp_header_buffer_size
+  type: int
+  default: 65536
+  info: |
+    Size of the buffer used to read data using an additional copy. Vitastor
+    packet headers are 128 bytes, payload is always at least 4 KB, so it is
+    usually beneficial to try to read multiple packets at once even though
+    it requires to copy the data an additional time. The rest of each packet
+    is received without an additional copy. You can try to play with this
+    parameter and see how it affects random iops and linear bandwidth if you
+    want.
+  info_ru: |
+    Размер буфера для чтения данных с дополнительным копированием. Пакеты
+    Vitastor содержат 128-байтные заголовки, за которыми следуют данные размером
+    от 4 КБ и для мелких операций ввода-вывода обычно выгодно за 1 вызов читать
+    сразу несколько пакетов, даже не смотря на то, что это требует лишний раз
+    скопировать данные. Часть каждого пакета за пределами значения данного
+    параметра читается без дополнительного копирования. Вы можете попробовать
+    поменять этот параметр и посмотреть, как он влияет на производительность
+    случайного и линейного доступа.
+- name: use_sync_send_recv
+  type: bool
+  default: false
+  info: |
+    If true, synchronous send/recv syscalls are used instead of io_uring for
+    socket communication. Useless for OSDs because they require io_uring anyway,
+    but may be required for clients with old kernel versions.
+  info_ru: |
+    Если установлено в истину, то вместо io_uring для передачи данных по сети
+    будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
+    это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
+    принципе, это может применяться для клиентов со старыми версиями ядра.

docs/config/src/osd.en.md

@@ -1,5 +1,5 @@
 # Runtime OSD Parameters
 
 These parameters only apply to OSDs, are not fixed at the moment of OSD drive
-initialization and can be changed - either with an OSD restart or, for some of
-them, even without restarting by updating configuration in etcd.
+initialization and can be changed - in /etc/vitastor/vitastor.conf or [vitastor-disk update-sb](../usage/disk.en.md#update-sb)
+with an OSD restart or, for some of them, even without restarting by updating configuration in etcd.

docs/config/src/osd.ru.md

@@ -2,5 +2,5 @@
 
 Данные параметры используются только OSD, но, в отличие от дисковых параметров,
 не фиксируются в момент инициализации дисков OSD и могут быть изменены в любой
-момент с помощью перезапуска OSD, а некоторые и без перезапуска, с помощью
-изменения конфигурации в etcd.
+момент с перезапуском OSD в /etc/vitastor/vitastor.conf или [vitastor-disk update-sb](../usage/disk.ru.md#update-sb),
+а некоторые и без перезапуска, с помощью изменения конфигурации в etcd.

docs/config/src/osd.yml

@@ -1,3 +1,26 @@
+- name: bind_address
+  type: string or array of strings
+  type_ru: строка или массив строк
+  info: |
+    Instead of the network masks ([osd_network](network.en.md#osd_network) and
+    [osd_cluster_network](network.en.md#osd_cluster_network)), you can also set
+    OSD listen addresses explicitly using this parameter. May be useful if you
+    want to start OSDs on interfaces that are not UP + RUNNING.
+  info_ru: |
+    Вместо использования масок подсети ([osd_network](network.ru.md#osd_network) и
+    [osd_cluster_network](network.ru.md#osd_cluster_network)), вы также можете явно
+    задать адрес(а), на которых будут ожидать соединений OSD, с помощью данного
+    параметра. Это может быть полезно, например, чтобы запускать OSD на неподнятых
+    интерфейсах (не UP + RUNNING).
+- name: bind_port
+  type: int
+  info: |
+    By default, OSDs pick random ports to use for incoming connections
+    automatically. With this option you can set a specific port for a specific
+    OSD by hand.
+  info_ru: |
+    По умолчанию OSD сами выбирают случайные порты для входящих подключений.
+    С помощью данной опции вы можете задать порт для отдельного OSD вручную.
 - name: osd_iothread_count
   type: int
   default: 0
@@ -56,44 +79,6 @@
     реализовать дополнительный режим для монитора, который позволит отделять
     первичные OSD от вторичных, но пока не понятно, зачем это может кому-то
     понадобиться, поэтому это не реализовано.
-- name: osd_network
-  type: string or array of strings
-  type_ru: строка или массив строк
-  info: |
-    Network mask of the network (IPv4 or IPv6) to use for OSDs. Note that
-    although it's possible to specify multiple networks here, this does not
-    mean that OSDs will create multiple listening sockets - they'll only
-    pick the first matching address of an UP + RUNNING interface. Separate
-    networks for cluster and client connections are also not implemented, but
-    they are mostly useless anyway, so it's not a big deal.
-  info_ru: |
-    Маска подсети (IPv4 или IPv6) для использования для соединений с OSD.
-    Имейте в виду, что хотя сейчас и можно передать в этот параметр несколько
-    подсетей, это не означает, что OSD будут создавать несколько слушающих
-    сокетов - они лишь будут выбирать адрес первого поднятого (состояние UP +
-    RUNNING), подходящий под заданную маску. Также не реализовано разделение
-    кластерной и публичной сетей OSD. Правда, от него обычно всё равно довольно
-    мало толку, так что особенной проблемы в этом нет.
-- name: bind_address
-  type: string
-  default: "0.0.0.0"
-  info: |
-    Instead of the network mask, you can also set OSD listen address explicitly
-    using this parameter. May be useful if you want to start OSDs on interfaces
-    that are not UP + RUNNING.
-  info_ru: |
-    Этим параметром можно явным образом задать адрес, на котором будет ожидать
-    соединений OSD (вместо использования маски подсети). Может быть полезно,
-    например, чтобы запускать OSD на неподнятых интерфейсах (не UP + RUNNING).
-- name: bind_port
-  type: int
-  info: |
-    By default, OSDs pick random ports to use for incoming connections
-    automatically. With this option you can set a specific port for a specific
-    OSD by hand.
-  info_ru: |
-    По умолчанию OSD сами выбирают случайные порты для входящих подключений.
-    С помощью данной опции вы можете задать порт для отдельного OSD вручную.
 - name: autosync_interval
   type: sec
   default: 5
@@ -315,7 +300,7 @@
     decrease write performance for fast disks because page cache is an overhead
     itself.
 
-    Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+    Choose "directsync" to use [immediate_commit](layout-cluster.en.md#immediate_commit)
     (which requires disable_data_fsync) with drives having write-back cache
     which can't be turned off, for example, Intel Optane. Also note that *some*
     desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
@@ -765,3 +750,34 @@
     Максимальное возможное значение авто-подстроенного recovery_sleep_us.
     Большие значения считаются случайными выбросами и игнорируются в
     усреднении.
+- name: discard_on_start
+  type: bool
+  info: Discard (SSD TRIM) unused data device blocks on every OSD startup.
+  info_ru: Освобождать (SSD TRIM) неиспользуемые блоки диска данных при каждом запуске OSD.
+- name: min_discard_size
+  type: int
+  default: 1048576
+  info: Minimum consecutive block size to TRIM it.
+  info_ru: Минимальный размер последовательного блока данных, чтобы освобождать его через TRIM.
+- name: allow_net_split
+  type: bool
+  default: false
+  info: |
+    Allow "safe" cases of network splits/partitions - allow to start PGs without
+    connections to some OSDs currently registered as alive in etcd, if the number
+    of actually connected PG OSDs is at least pg_minsize. That is, allow some OSDs to lose
+    connectivity with some other OSDs as long as it doesn't break pg_minsize guarantees.
+    The downside is that it increases the probability of writing data into just pg_minsize
+    OSDs during failover which can lead to PGs becoming incomplete after additional outages.
+
+    The old behaviour in versions up to 2.0.0 was equal to enabled allow_net_split.
+  info_ru: |
+    Разрешить "безопасные" случаи разделений сети - разрешить активировать PG без
+    соединений к некоторым OSD, помеченным активными в etcd, если общее число активных
+    OSD в PG составляет как минимум pg_minsize. То есть, разрешать некоторым OSD терять
+    соединения с некоторыми другими OSD, если это не нарушает гарантий pg_minsize.
+    Минус такого разрешения в том, что оно повышает вероятность записи данных ровно в
+    pg_minsize OSD во время переключений, что может потом привести к тому, что PG станут
+    неполными (incomplete), если упадут ещё какие-то OSD.
+
+    Старое поведение в версиях до 2.0.0 было идентично включённому allow_net_split.

docs/installation/docker.en.md

@@ -26,9 +26,9 @@ at Vitastor Kubernetes operator: https://github.com/Antilles7227/vitastor-operat
 The instruction is very simple.
 
 1. Download a Docker image of the desired version: \
-   `docker pull vitastor:1.10.2`
+   `docker pull vitastor:2.1.0`
 2. Install scripts to the host system: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:1.10.2 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:2.1.0 install.sh`
 3. Reload udev rules: \
    `udevadm control --reload-rules`
 

docs/installation/docker.ru.md

@@ -25,9 +25,9 @@ Vitastor можно установить в Docker/Podman. При этом etcd,
 Инструкция по установке максимально простая.
 
 1. Скачайте Docker-образ желаемой версии: \
-   `docker pull vitastor:1.10.2`
+   `docker pull vitastor:2.1.0`
 2. Установите скрипты в хост-систему командой: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:1.10.2 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:2.1.0 install.sh`
 3. Перезагрузите правила udev: \
    `udevadm control --reload-rules`
 

docs/installation/s3.en.md

@@ -0,0 +1,191 @@
+[Documentation](../../README.md#documentation) → Installation → S3 for Vitastor
+
+-----
+
+[Читать на русском](s3.ru.md)
+
+# S3 for Vitastor
+
+The moment has come - Vitastor S3 implementation based on Zenko CloudServer is released.
+
+## Highlights
+
+- Zenko CloudServer is implemented in node.js.
+- Object metadata is stored in MongoDB.
+- Modified Zenko CloudServer version is used for Vitastor. It is slightly different from
+  the original, has an optimised build and unneeded dependencies are stripped off.
+- Object data is stored in Vitastor block volumes, but the volume metadata is stored in
+  the same MongoDB, not in Vitastor etcd.
+- Objects are written to volumes sequentially one after another. The space is allocated
+  with rounding to the sector size (4 KB), so each object takes at least 4 KB.
+- An important property of such storage scheme is that small objects aren't chunked into
+  parts in Vitastor EC N+K pools and thus don't require reads from all N disks when
+  downloading.
+- Deleted objects are marked as deleted, but the space is only actually freed during
+  asynchronously executed "defragmentation" process. Defragmentation runs automatically
+  in the background when a volume reaches configured amount of "garbage" (20% by default).
+  Defragmentation copies actual objects to new volume(s) and then removes the old volume.
+  Defragmentation can be configured in locationConfig.json.
+
+## Plans for future development
+
+- User account storage in the DB instead of a static file. Original Zenko uses
+  a separate closed-source "Scality Vault" service for it, that's why we use
+  a static file for now.
+- More detailed documentation.
+- Support for other (and faster) key-value DBMS for object metadata storage.
+- Other performance optimisations, for example, related to the used hash function -
+  MD5 used for Amazon compatibility purposes is relatively slow.
+- Object Lifecycle support. There is a Lifecycle implementation for Zenko called
+  [Backbeat](https://github.com/scality/backbeat) but it's not adapted for Vitastor yet.
+- Quota support. Original Zenko uses a separate "SCUBA" service for quotas, but
+  it's also proprietary and not available publicly.
+
+## Installation
+
+In a few words:
+
+- Install MongoDB, create a user for S3 metadata DB.
+- Create a Vitastor pool for S3 data.
+- Download and setup the Docker container `vitalif/vitastor-zenko`.
+
+### Setup MongoDB
+
+You can setup MongoDB yourself, following the [MongoDB manual](https://www.mongodb.com/docs/manual/installation/).
+
+Or you can follow the instructions below - it describes a simple example of MongoDB setup
+in Docker (through docker-compose) with 3 replicas.
+
+1. On each host, create a file `docker-compose.yml` with the content listed below.
+   Replace `<YOUR_PASSWORD>` with your future mongodb administrator password, and optionally
+   replace `0.0.0.0` with `localhost,<server_IP>`. It's recommended to either use a private IP
+   or [setup TLS](https://www.mongodb.com/docs/manual/tutorial/configure-ssl/) afterwards.
+
+```
+version: '3.1'
+
+services:
+
+  mongo:
+    container_name: mongo
+    image: mongo:7-jammy
+    restart: always
+    environment:
+      MONGO_INITDB_ROOT_USERNAME: root
+      MONGO_INITDB_ROOT_PASSWORD: <YOUR_PASSWORD>
+    network_mode: host
+    volumes:
+      - ./keyfile:/opt/keyfile
+      - ./mongo-data/db:/data/db
+      - ./mongo-data/configdb:/data/configdb
+    entrypoint: /bin/bash -c
+    command: [ "chown mongodb /opt/keyfile && chmod 600 /opt/keyfile && . /usr/local/bin/docker-entrypoint.sh mongod --replSet rs0 --keyFile /opt/keyfile --bind_ip 0.0.0.0" ]
+```
+
+2. Generate a shared cluster key using `openssl rand -base64 756 > ./keyfile` and copy
+   that `keyfile` to all hosts.
+
+3. Start MongoDB on all hosts with `docker compose up -d mongo`.
+
+4. Enter Mongo Shell with `docker exec -it mongo mongosh -u root -p <YOUR_PASSWORD> localhost/admin`
+   and execute the following command (replace IP addresses `10.10.10.{1,2,3}` with your host IPs):
+
+`rs.initiate({ _id: 'rs0', members: [
+  { _id: 1, host: '10.10.10.1:27017' },
+  { _id: 2, host: '10.10.10.2:27017' },
+  { _id: 3, host: '10.10.10.3:27017' }
+] })`
+
+5. Stay in Mongo Shell and create a user for the future S3 database:
+
+`db.createUser({ user: 's3', pwd: '<YOUR_S3_PASSWORD>', roles: [
+  { role: 'readWrite', db: 's3' },
+  { role: 'dbAdmin', db: 's3' },
+  { role: 'readWrite', db: 'vitastor' },
+  { role: 'dbAdmin', db: 'vitastor' }
+] })`
+
+### Setup Vitastor
+
+Create a pool in Vitastor for S3 object data, for example:
+
+`vitastor-cli create-pool --ec 2+1 -n 512 s3-data --used_for_app s3:standard`
+
+The `--used_for_app` options works as fool-proofing and prevents you from
+accidentally creating a regular block volume in the S3 pool and overwriting some S3 data.
+Also it hides inode space statistics from Vitastor etcd.
+
+Retrieve the ID of your pool with `vitastor-cli ls-pools s3-data --detail`.
+
+### Setup Vitastor S3
+
+1. Add the following lines to `docker-compose.yml` (instead of `network_mode: host`,
+   you can use `ports: [ "8000:8000", "8002:8002" ]`):
+
+```
+  zenko:
+    container_name: zenko
+    image: vitalif/vitastor-zenko
+    restart: always
+    security_opt:
+      - seccomp:unconfined
+    ulimits:
+      memlock: -1
+    network_mode: host
+    volumes:
+      - /etc/vitastor:/etc/vitastor
+      - /etc/vitastor/s3:/conf
+```
+
+2. Download Docker image: `docker pull vitalif/vitastor-zenko`
+
+3. Extract configuration file examples from the Docker image:
+   ```
+   docker run --rm -it -v /etc/vitastor:/etc/vitastor -v /etc/vitastor/s3:/conf vitalif/vitastor-zenko configure.sh
+   ```
+
+4. Edit configuration files in `/etc/vitastor/s3/`:
+   - `config.json` - common settings.
+   - `authdata.json` - user accounts and access keys.
+   - `locationConfig.json` - S3 storage class list with placement settings.
+     Note: it actually contains storage classes (like STANDARD, COLD, etc)
+     instead of "locations" (zones like us-east-1) as in the original Zenko CloudServer.
+   - Put your MongoDB connection data into `config.json` and `locationConfig.json`.
+   - Put your Vitastor pool ID into `locationConfig.json`.
+   - For now, the complete list of Vitastor backend settings is only available [in the code](https://git.yourcmc.ru/vitalif/zenko-arsenal/src/branch/master/lib/storage/data/vitastor/VitastorBackend.ts#L94).
+
+### Start Zenko
+
+Start the S3 server with:
+
+```
+docker run --restart always --security-opt seccomp:unconfined --ulimit memlock=-1 --network=host \
+    -v /etc/vitastor:/etc/vitastor -v /etc/vitastor/s3:/conf --name zenko vitalif/vitastor-zenko
+```
+
+If you use default settings, Zenko CloudServer starts on port 8000.
+The default access key is `accessKey1` with a secret key of `verySecretKey1`.
+
+Now you can access your S3 with, for example, [s3cmd](https://s3tools.org/s3cmd):
+
+```
+s3cmd --access_key=accessKey1 --secret_key=verySecretKey1 --host=http://localhost:8000 mb s3://testbucket
+```
+
+Or even mount it with [GeeseFS](https://github.com/yandex-cloud/geesefs):
+
+```
+AWS_ACCESS_KEY_ID=accessKey1 \
+    AWS_SECRET_ACCESS_KEY=verySecretKey1 \
+    geesefs --endpoint http://localhost:8000 testbucket mountdir
+```
+
+## Author & License
+
+- [Zenko CloudServer](https://s3-server.readthedocs.io/en/latest/) author is Scality,
+  licensed under [Apache License, version 2.0](https://www.apache.org/licenses/LICENSE-2.0)
+- [Vitastor](https://git.yourcmc.ru/vitalif/vitastor/) and Zenko Vitastor backend author is
+  Vitaliy Filippov, licensed under [VNPL-1.1](https://git.yourcmc.ru/vitalif/vitastor/src/branch/master/VNPL-1.1.txt)
+  (a "network copyleft" license based on AGPL/SSPL, but worded in a better way)
+- Vitastor S3 repository: https://git.yourcmc.ru/vitalif/zenko-cloudserver-vitastor
+- Vitastor S3 backend code: https://git.yourcmc.ru/vitalif/zenko-arsenal/src/branch/master/lib/storage/data/vitastor/VitastorBackend.ts

docs/installation/s3.ru.md

@@ -0,0 +1,171 @@
+[Документация](../../README-ru.md#документация) → Установка → S3 на базе Vitastor
+
+-----
+
+[Read in English](s3.en.md)
+
+# S3 на базе Vitastor
+
+Итак, свершилось - реализация Vitastor S3 на базе Zenko CloudServer достигла
+состояния готовности к публикации и использованию.
+
+## Ключевые особенности
+
+- Zenko CloudServer реализован на node.js.
+- Метаданные объектов хранятся в MongoDB.
+- Поставляется модифицированная версия Zenko CloudServer, отвязанная от лишних зависимостей,
+  с оптимизированной сборкой и немного отличающаяся от оригинала.
+- Данные объектов хранятся в блочных томах Vitastor, однако информация о самих томах
+  сохраняется не в etcd Vitastor, а тоже в БД на основе MongoDB.
+- Объекты записываются в тома последовательно друг за другом. Место выделяется с округлением
+  до размера сектора (до 4 килобайт), поэтому каждый объект занимает как минимум 4 КБ.
+- Благодаря такой схеме записи объектов мелкие объекты не нарезаются на части и поэтому не
+  требуют чтения с N дисков данных в EC N+K пулах Vitastor.
+- При удалении объекты помечаются удалёнными, но место освобождается не сразу, а при
+  запускаемой асинхронно "дефрагментации". Дефрагментация запускается автоматически в фоне
+  при достижении заданного объёма "мусора" в томе (по умолчанию 20%), копирует актуальные
+  объекты в новые тома, после чего очищает старый том полностью. Дефрагментацию можно
+  настраивать в locationConfig.json.
+
+## Планы развития
+
+- Хранение учётных записей в БД, а не в статическом файле (в оригинальном Zenko для
+  этого используется отдельный закрытый сервис "Scality Vault").
+- Более подробная документация.
+- Поддержка других (и более производительных) key-value СУБД для хранения метаданных.
+- Другие оптимизации производительности, например, в области используемой хеш-функции
+  (хеш MD5, используемый в целях совместимости, относительно медленный).
+- Поддержка Object Lifecycle. Реализация Lifecycle для Zenko существует и называется
+  [Backbeat](https://github.com/scality/backbeat), но она ещё не адаптирована для Vitastor.
+- Квоты. В оригинальном Zenko для этого используется отдельный сервис "SCUBA", однако
+  он тоже является закрытым и недоступен для публичного использования.
+
+## Установка
+
+Кратко:
+
+- Установите MongoDB, создайте пользователя для БД метаданных S3.
+- Создайте в Vitastor пул для хранения данных объектов.
+- Скачайте и настройте Docker-контейнер `vitalif/vitastor-zenko`.
+
+### Установка MongoDB
+
+Вы можете установить MongoDB сами, следуя [официальному руководству MongoDB](https://www.mongodb.com/docs/manual/installation/).
+
+Либо вы можете последовать инструкции, приведённой ниже - здесь описан простейший пример
+установки MongoDB в Docker (docker-compose) в конфигурации с 3 репликами.
+
+1. На всех 3 серверах создайте файл `docker-compose.yml`, заменив `<ВАШ_ПАРОЛЬ>`
+   на собственный будущий пароль администратора mongodb, а `0.0.0.0` по желанию
+   заменив на на `localhost,<IP_сервера>` - желательно либо использовать публично не доступный IP,
+   либо потом [настроить TLS](https://www.mongodb.com/docs/manual/tutorial/configure-ssl/).
+
+```
+version: '3.1'
+
+services:
+
+  mongo:
+    container_name: mongo
+    image: mongo:7-jammy
+    restart: always
+    environment:
+      MONGO_INITDB_ROOT_USERNAME: root
+      MONGO_INITDB_ROOT_PASSWORD: <ВАШ_ПАРОЛЬ>
+    network_mode: host
+    volumes:
+      - ./keyfile:/opt/keyfile
+      - ./mongo-data/db:/data/db
+      - ./mongo-data/configdb:/data/configdb
+    entrypoint: /bin/bash -c
+    command: [ "chown mongodb /opt/keyfile && chmod 600 /opt/keyfile && . /usr/local/bin/docker-entrypoint.sh mongod --replSet rs0 --keyFile /opt/keyfile --bind_ip 0.0.0.0" ]
+```
+
+2. В той же директории сгенерируйте общий ключ кластера командой `openssl rand -base64 756 > ./keyfile`
+   и скопируйте этот файл на все 3 сервера.
+
+3. На всех 3 серверах запустите MongoDB командой `docker compose up -d mongo`.
+
+4. Зайдите в Mongo Shell с помощью команды `docker exec -it mongo mongosh -u root -p <ВАШ_ПАРОЛЬ> localhost/admin`
+   и там выполните команду (заменив IP-адреса `10.10.10.{1,2,3}` на адреса своих серверов):
+
+`rs.initiate({ _id: 'rs0', members: [
+  { _id: 1, host: '10.10.10.1:27017' },
+  { _id: 2, host: '10.10.10.2:27017' },
+  { _id: 3, host: '10.10.10.3:27017' }
+] })`
+
+5. Находясь там же, в Mongo Shell, создайте пользователя с доступом к будущей базе данных S3:
+
+`db.createUser({ user: 's3', pwd: '<ВАШ_ПАРОЛЬ_S3>', roles: [
+  { role: 'readWrite', db: 's3' },
+  { role: 'dbAdmin', db: 's3' },
+  { role: 'readWrite', db: 'vitastor' },
+  { role: 'dbAdmin', db: 'vitastor' }
+] })`
+
+### Настройка Vitastor
+
+Создайте в Vitastor отдельный пул для данных объектов S3, например:
+
+`vitastor-cli create-pool --ec 2+1 -n 512 s3-data --used_for_app s3:standard`
+
+Опция `--used_for_app` работает как "защита от дурака" и не даёт вам случайно создать
+в этом пуле обычный блочный том и перезаписать им какие-то данные S3, а также скрывает
+статистику занятого места по томам S3 из etcd.
+
+Получите ID своего пула с помощью команды `vitastor-cli ls-pools --detail`.
+
+### Установка Vitastor S3
+
+1. Добавьте в `docker-compose.yml` строки (альтернативно вместо `network_mode: host`
+   можно использовать `ports: [ "8000:8000", "8002:8002" ]`):
+
+```
+  zenko:
+    container_name: zenko
+    image: vitalif/vitastor-zenko
+    restart: always
+    security_opt:
+      - seccomp:unconfined
+    ulimits:
+      memlock: -1
+    network_mode: host
+    volumes:
+      - /etc/vitastor:/etc/vitastor
+      - /etc/vitastor/s3:/conf
+```
+
+2. Извлеките из Docker-образа Vitastor примеры файлов конфигурации:
+   `docker run --rm -it -v /etc/vitastor:/etc/vitastor -v /etc/vitastor/s3:/conf vitalif/vitastor-zenko configure.sh`
+
+3. Отредактируйте файлы конфигурации в `/etc/vitastor/s3/`:
+   - `config.json` - общие настройки.
+   - `authdata.json` - учётные записи и ключи доступа.
+   - `locationConfig.json` - список классов хранения S3 с настройками расположения.
+     Внимание: в данной версии это именно список S3 storage class-ов (STANDARD, COLD и т.п.),
+     а не зон (подобных us-east-1), как в оригинальном Zenko CloudServer.
+   - В `config.json` и в `locationConfig.json` пропишите свои данные подключения к MongoDB.
+   - В `locationConfig.json` укажите ID пула Vitastor для хранения данных.
+   - Полный перечень настроек Vitastor-бэкенда пока можно посмотреть [в коде](https://git.yourcmc.ru/vitalif/zenko-arsenal/src/branch/master/lib/storage/data/vitastor/VitastorBackend.ts#L94).
+
+### Запуск
+
+Запустите S3-сервер: `docker-compose up -d zenko`
+
+Готово! Вы получили S3-сервер, работающий на порту 8000.
+
+Можете попробовать обратиться к нему с помощью, например, [s3cmd](https://s3tools.org/s3cmd):
+
+`s3cmd --host-bucket= --no-ssl --access_key=accessKey1 --secret_key=verySecretKey1 --host=http://localhost:8000 mb s3://testbucket`
+
+Или смонтировать его с помощью [GeeseFS](https://github.com/yandex-cloud/geesefs):
+
+`AWS_ACCESS_KEY_ID=accessKey1 AWS_SECRET_ACCESS_KEY=verySecretKey1 geesefs --endpoint http://localhost:8000 testbucket /mnt/geesefs`
+
+## Лицензия
+
+- Автор [Zenko CloudServer](https://s3-server.readthedocs.io/en/latest/) - Scality, лицензия [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
+- Vitastor-бэкенд для S3, как и сам Vitastor, лицензируется на условиях [VNPL 1.1](https://git.yourcmc.ru/vitalif/vitastor/src/branch/master/VNPL-1.1.txt)
+- Репозиторий сборки: https://git.yourcmc.ru/vitalif/zenko-cloudserver-vitastor
+- Бэкенд хранения данных: https://git.yourcmc.ru/vitalif/zenko-arsenal/src/branch/master/lib/storage/data/vitastor/VitastorBackend.ts

docs/installation/source.en.md

@@ -16,7 +16,7 @@
   designated initializers support from C++20
 - CMake
 - liburing, jerasure headers and libraries
-- ISA-L, libibverbs headers and libraries (optional)
+- ISA-L, libibverbs and librdmacm headers and libraries (optional)
 - tcmalloc (google-perftools-dev)
 
 ## Basic instructions

docs/installation/source.ru.md

@@ -16,7 +16,7 @@
   назначенных инициализаторов (designated initializers) из C++20
 - CMake
 - Заголовки и библиотеки liburing, jerasure
-- Опционально - заголовки и библиотеки ISA-L, libibverbs
+- Опционально - заголовки и библиотеки ISA-L, libibverbs, librdmacm
 - tcmalloc (google-perftools-dev)
 
 ## Базовая инструкция

docs/intro/features.en.md

@@ -28,7 +28,7 @@
 - Per-OSD and per-image I/O and space usage statistics in etcd
 - Snapshots and copy-on-write image clones
 - [Write throttling to smooth random write workloads in SSD+HDD configurations](../config/osd.en.md#throttle_small_writes)
-- [RDMA/RoCEv2 support via libibverbs](../config/network.en.md#rdma_device)
+- RDMA/RoCEv2 support [via libibverbs](../config/network.en.md#use_rdma) or [RDMA-CM](../config/network.en.md#use_rdmacm)
 - [Scrubbing](../config/osd.en.md#auto_scrub) (verification of copies)
 - [Checksums](../config/layout-osd.en.md#data_csum_type)
 - [Client write-back cache](../config/client.en.md#client_enable_writeback)
@@ -37,6 +37,7 @@
 - [Experimental internal etcd replacement - antietcd](../config/monitor.en.md#use_antietcd)
 - [Built-in Prometheus metric exporter](../config/monitor.en.md#enable_prometheus)
 - [NFS RDMA support](../usage/nfs.en.md#rdma) (probably also usable for GPUDirect)
+- [S3](../installation/s3.en.md)
 
 ## Plugins and tools
 
@@ -63,7 +64,6 @@ The following features are planned for the future:
 - iSCSI and NVMeoF gateways
 - Multi-threaded client
 - Faster failover
-- S3
 - Tiered storage (SSD caching)
 - NVDIMM support
 - Compression (possibly)

docs/intro/features.ru.md

@@ -30,7 +30,7 @@
 - Именование инодов через хранение их метаданных в etcd
 - Снапшоты и copy-on-write клоны
 - [Сглаживание производительности случайной записи в SSD+HDD конфигурациях](../config/osd.ru.md#throttle_small_writes)
-- [Поддержка RDMA/RoCEv2 через libibverbs](../config/network.ru.md#rdma_device)
+- Поддержка RDMA/RoCEv2 [через libibverbs](../config/network.ru.md#use_rdma) или [RDMA-CM](../config/network.ru.md#use_rdmacm)
 - [Фоновая проверка целостности](../config/osd.ru.md#auto_scrub) (сверка копий)
 - [Контрольные суммы](../config/layout-osd.ru.md#data_csum_type)
 - [Буферизация записи на стороне клиента](../config/client.ru.md#client_enable_writeback)
@@ -39,6 +39,7 @@
 - [Экспериментальная встроенная замена etcd - antietcd](../config/monitor.ru.md#use_antietcd)
 - [Встроенный Prometheus-экспортер метрик](../config/monitor.ru.md#enable_prometheus)
 - [Поддержка NFS RDMA](../usage/nfs.ru.md#rdma) (вероятно, также подходящая для GPUDirect)
+- [S3](../installation/s3.ru.md)
 
 ## Драйверы и инструменты
 
@@ -63,7 +64,6 @@
 - iSCSI и NVMeoF прокси
 - Многопоточный клиент
 - Более быстрое переключение при отказах
-- S3
 - Поддержка SSD-кэширования (tiered storage)
 - Поддержка NVDIMM
 - Возможно, сжатие

docs/intro/quickstart.en.md

@@ -50,7 +50,7 @@ On the monitor hosts:
 
 ## Configure OSDs
 
-- Put etcd_address and osd_network into `/etc/vitastor/vitastor.conf`. Example:
+- Put etcd_address and [osd_network](../config/network.en.md#osd_network) into `/etc/vitastor/vitastor.conf`. Example:
   ```
   {
     "etcd_address": ["10.200.1.10:2379","10.200.1.11:2379","10.200.1.12:2379"],

docs/intro/quickstart.ru.md

@@ -50,7 +50,7 @@
 
 ## Настройте OSD
 
-- Пропишите etcd_address и osd_network в `/etc/vitastor/vitastor.conf`. Например:
+- Пропишите etcd_address и [osd_network](../config/network.ru.md#osd_network) в `/etc/vitastor/vitastor.conf`. Например:
   ```
   {
     "etcd_address": ["10.200.1.10:2379","10.200.1.11:2379","10.200.1.12:2379"],

docs/usage/admin.en.md

@@ -35,10 +35,19 @@ PG state consists of exactly 1 base state and an arbitrary number of additional
 
 PG state always includes exactly 1 of the following base states:
 - **active** — PG is active and handles user I/O.
-- **incomplete** — Not enough OSDs are available to activate this PG. That is, more disks
-  are lost than it's allowed by the pool's redundancy scheme. For example, if the pool has
-  pg_size=3 and pg_minsize=1, part of the data may be written only to 1 OSD. If that exact
-  OSD is lost, PG will become **incomplete**.
+- **incomplete** — Not enough OSDs are available to activate this PG. More exactly, that
+  means one of the following:
+  - Less than pg_minsize current target OSDs are available for the PG. I.e. more disks
+    are lost than allowed by the pool's redundancy scheme.
+  - All OSDs of some of PG's history records are unavailable, or, for EC pools, less
+    than (pg_size-parity_chunks) OSDs are available in one of the history records.
+    In other words it means that some data in this PG was written to an OSD set such that
+    it's currently impossible to read it back because these OSDs are down. For example,
+    if the pool has pg_size=3 and pg_minsize=1, part of the data may be written only to
+    1 OSD. If that exact OSD is lost, PG becomes **incomplete**.
+  - [allow_net_split](../config/osd.en.md#allow_net_split) is disabled (default) and
+    primary OSD of the PG can't connect to some secondary OSDs marked as alive in etcd.
+    I.e. a network partition happened: OSDs can talk to etcd, but not to some other OSDs.
 - **offline** — PG isn't activated by any OSD at all. Either primary OSD isn't set for
   this PG at all (if the pool is just created), or an unavailable OSD is set as primary,
   or the primary OSD refuses to start this PG (for example, because of wrong block_size),

docs/usage/admin.ru.md

@@ -35,10 +35,20 @@
 
 Состояние PG включает в себя ровно 1 флаг из следующих:
 - **active** — PG активна и обрабатывает запросы ввода-вывода от пользователей.
-- **incomplete** — Недостаточно живых OSD, чтобы включить эту PG.
-  То есть, дисков потеряно больше, чем разрешено схемой отказоустойчивости пула и pg_minsize.
-  Например, если у пула pg_size=3 и pg_minsize=1, то часть данных может записаться всего на 1 OSD.
-  Если потом конкретно этот OSD упадёт, PG окажется **incomplete**.
+- **incomplete** — Недостаточно живых OSD, чтобы включить эту PG. Если точнее, то это
+  означает один из следующих вариантов:
+  - Доступно менее, чем pg_minsize текущих целевых OSD данной PG. Иными словами, потеряно
+    больше дисков, чем это разрешает схема отказоустойчивости пула.
+  - Все OSD одной из исторических записей PG недоступны, или, для EC-пулов, в одной
+    из исторических записей PG доступно менее, чем (pg_size-parity_chunks) OSD. Другими
+    словами это означает, что часть данных этой PG была записана в такой набор OSD, из
+    которого их сейчас невозможно прочитать обратно, так как OSD не включены. Например,
+    если у пула pg_size=3 и pg_minsize=1, то часть данных может записаться всего на 1 OSD.
+    Если потом конкретно этот OSD упадёт, PG окажется **incomplete**.
+  - [allow_net_split](../config/osd.ru.md#allow_net_split) отключено (по умолчанию) и
+    первичный OSD данной PG не может соединиться с частью вторичных OSD этой PG, помеченных
+    как живых в etcd. Это означает, что произошло разделение сети: OSD могут общаться с etcd,
+    но не могут общаться с частью других OSD.
 - **offline** — PG вообще не активирована ни одним OSD. Либо первичный OSD не назначен вообще
   (если пул только создан), либо в качестве первичного назначен недоступный OSD, либо
   назначенный OSD отказывается запускать эту PG (например, из-за несовпадения block_size),

docs/usage/cli.en.md

@@ -355,7 +355,7 @@ Set OSD reweight, tags or noout flag. See detail description in [OSD config docu
 
 ## pg-list
 
-`vitastor-cli pg-list|pg-ls|list-pg|ls-pg|ls-pgs [OPTIONS] [state1+state2] [^state3] [...]`
+`vitastor-cli pg-list|pg-ls|list-pg|ls-pg|ls-pgs|pgs [OPTIONS] [state1+state2] [^state3] [...]`
 
 List PGs with any of listed state filters (^ or ! in the beginning is negation). Options:
 
@@ -363,6 +363,7 @@ List PGs with any of listed state filters (^ or ! in the beginning is negation).
 --pool <pool name or number>  Only list PGs of the given pool.
 --min <min pg number>         Only list PGs with number >= min.
 --max <max pg number>         Only list PGs with number <= max.
+--osd 1,2,...                 Only list PGs with some data on specified OSD(s).
 ```
 
 Examples:
@@ -377,11 +378,11 @@ Examples:
 
 Create a pool. Required parameters:
 
-| <!-- -->                 | <!-- -->                                                                              |
-|--------------------------|---------------------------------------------------------------------------------------|
-| `-s R` or `--pg_size R`  | Number of replicas for replicated pools                                               |
-| `--ec N+K`               | Number of data (N) and parity (K) chunks for erasure-coded pools                      |
-| `-n N` or `--pg_count N` | PG count for the new pool (start with 10*<OSD count>/pg_size rounded to a power of 2) |
+| <!-- -->                 | <!-- -->                                                                                |
+|--------------------------|-----------------------------------------------------------------------------------------|
+| `-s R` or `--pg_size R`  | Number of replicas for replicated pools                                                 |
+| `--ec N+K`               | Number of data (N) and parity (K) chunks for erasure-coded pools                        |
+| `-n N` or `--pg_count N` | PG count for the new pool (start with 10*\<OSD count\>/pg_size rounded to a power of 2) |
 
 Optional parameters:
 
@@ -398,7 +399,8 @@ Optional parameters:
 | `--raw_placement <rules>`      | Specify raw PG generation rules ([details](../config/pool.en.md#raw_placement)) |
 | `--primary_affinity_tags tags` | Prefer to put primary copies on OSDs with all specified tags               |
 | `--scrub_interval <time>`      | Enable regular scrubbing for this pool. Format: number + unit s/m/h/d/M/y  |
-| `--used_for_fs <name>`         | Mark pool as used for VitastorFS with metadata in image <name>             |
+| `--used_for_app fs:<name>`     | Mark pool as used for VitastorFS with metadata in image `<name>`           |
+| `--used_for_app s3:<name>`     | Mark pool as used for S3 location with name `<name>`                       |
 | `--pg_stripe_size <number>`    | Increase object grouping stripe                                            |
 | `--max_osd_combinations 10000` | Maximum number of random combinations for LP solver input                  |
 | `--wait`                       | Wait for the new pool to come online                                       |

docs/usage/cli.ru.md

@@ -22,6 +22,8 @@ vitastor-cli - интерфейс командной строки для адм
 - [flatten](#flatten)
 - [rm-data](#rm-data)
 - [merge-data](#merge-data)
+- [describe](#describe)
+- [fix](#fix)
 - [alloc-osd](#alloc-osd)
 - [rm-osd](#rm-osd)
 - [osd-tree](#osd-tree)
@@ -375,9 +377,10 @@ OSD  PARENT            UP    SIZE  USED%    TAGS          WEIGHT  BLOCK  BITMAP
 в начале фильтра означает отрицание). Опции:
 
 ```
---pool <pool name or number>  Only list PGs of the given pool.
---min <min pg number>         Only list PGs with number >= min.
---max <max pg number>         Only list PGs with number <= max.
+--pool <pool name or number>  Вывести только PG в заданном пуле.
+--min <min pg number>         Вывести только PG с номерами >= min.
+--max <max pg number>         Вывести только PG с номерами <= max.
+--osd 1,2,...                 Вывести только PG с данными на заданных OSD.
 ```
 
 Примеры:
@@ -392,11 +395,11 @@ OSD  PARENT            UP    SIZE  USED%    TAGS          WEIGHT  BLOCK  BITMAP
 
 Создать пул. Обязательные параметры:
 
-| <!-- -->                  | <!-- -->                                                                                    |
-|---------------------------|---------------------------------------------------------------------------------------------|
-| `-s R` или `--pg_size R`  | Число копий данных для реплицированных пулов                                                |
-| `--ec N+K`                | Число частей данных (N) и чётности (K) для пулов с кодами коррекции ошибок                  |
-| `-n N` или `--pg_count N` | Число PG для нового пула (начните с 10*<число OSD>/pg_size, округлённого до степени двойки) |
+| <!-- -->                  | <!-- -->                                                                                      |
+|---------------------------|-----------------------------------------------------------------------------------------------|
+| `-s R` или `--pg_size R`  | Число копий данных для реплицированных пулов                                                  |
+| `--ec N+K`                | Число частей данных (N) и чётности (K) для пулов с кодами коррекции ошибок                    |
+| `-n N` или `--pg_count N` | Число PG для нового пула (начните с 10*\<число OSD\>/pg_size, округлённого до степени двойки) |
 
 Необязательные параметры:
 

docs/usage/disk.en.md

@@ -14,6 +14,7 @@ It supports the following commands:
 - [upgrade-simple](#upgrade-simple)
 - [resize](#resize)
 - [raw-resize](#raw-resize)
+- [trim](#trim)
 - [start/stop/restart/enable/disable](#start/stop/restart/enable/disable)
 - [purge](#purge)
 - [read-sb](#read-sb)
@@ -97,6 +98,9 @@ Options (both modes):
 --data_device_block 4k     Override data device block size
 --meta_device_block 4k     Override metadata device block size
 --journal_device_block 4k  Override journal device block size
+--discard_on_start 0       TRIM unused data device blocks every OSD start (default off)
+--min_discard_size 1M      Minimum TRIM block size
+--json                     Enable JSON output
 ```
 
 [immediate_commit](../config/layout-cluster.en.md#immediate_commit) setting is
@@ -179,6 +183,19 @@ parameters from OSD command line (i.e. from systemd unit or superblock).
 SIZE may include k/m/g/t suffixes. If any of the new layout parameter
 options are not specified, old values will be used.
 
+## trim
+
+`vitastor-disk trim <osd_num>|<osd_device> [<osd_num>|<osd_device>...]`
+
+Try to discard unused blocks (SSD TRIM) on the data device of each of the OSD(s).
+
+May only be used on stopped OSDs. Options:
+
+```
+--min_discard_size 1M      Minimum TRIM block size
+--discard_granularity 0    Override device's discard granularity
+```
+
 ## start/stop/restart/enable/disable
 
 `vitastor-disk start|stop|restart|enable|disable [--now] <device> [device2 device3 ...]`

docs/usage/disk.ru.md

@@ -99,6 +99,9 @@ vitastor-disk - инструмент командной строки для уп
 --data_device_block 4k     Задать размер блока устройства данных
 --meta_device_block 4k     Задать размер блока метаданных
 --journal_device_block 4k  Задать размер блока журнала
+--discard_on_start 0       Выполнять TRIM пустых блоков данных при запуске OSD (по умолчанию нет)
+--min_discard_size 1M      Минимальный размер блока для TRIM
+--json                     Включить JSON-вывод
 ```
 
 Настройка [immediate_commit](../config/layout-cluster.ru.md#immediate_commit)
@@ -182,6 +185,20 @@ throttle_target_mbs, throttle_target_parallelism, throttle_threshold_us.
 `РАЗМЕР` может быть указан с суффиксами k/m/g/t. Если любой из новых параметров
 расположения не указан, он принимается равным старому значению.
 
+## trim
+
+`vitastor-disk trim <osd_num>|<osd_device> [<osd_num>|<osd_device>...]`
+
+Попробовать пометить пустые блоки дисков данных всех указанных OSD неиспользуемыми
+(выполнить команду SSD TRIM).
+
+Можно использовать только с остановленными OSD. Опции:
+
+```
+--min_discard_size 1M      Минимальный размер блока для TRIM
+--discard_granularity 0    Кратность размера блока для TRIM
+```
+
 ## start/stop/restart/enable/disable
 
 `vitastor-disk start|stop|restart|enable|disable [--now] <device> [device2 device3 ...]`

docs/usage/nfs.en.md

@@ -58,7 +58,7 @@ To use VitastorFS:
 2. Create an image for FS metadata, preferably in a faster (SSD or replica-HDD) pool,
    but you can create it in the data pool too if you want (image size doesn't matter):
    `vitastor-cli create -s 10G -p fastpool testfs`
-3. Mark data pool as an FS pool: `vitastor-cli modify-pool --used-for-fs testfs data-pool`
+3. Mark data pool as an FS pool: `vitastor-cli modify-pool --used-for-app fs:testfs data-pool`
 4. Either mount the FS: `vitastor-nfs mount --fs testfs --pool data-pool /mnt/vita`
 5. Or start the NFS server: `vitastor-nfs start --fs testfs --pool data-pool`
 

docs/usage/nfs.ru.md

@@ -60,7 +60,7 @@ JSON-формате :-). Для инспекции содержимого БД
    или по крайней мере на HDD, но без EC), но можно и в том же пуле, что данные
    (размер образа значения не имеет):
    `vitastor-cli create -s 10G -p fastpool testfs`
-3. Пометьте пул данных как ФС-пул: `vitastor-cli modify-pool --used-for-fs testfs data-pool`
+3. Пометьте пул данных как ФС-пул: `vitastor-cli modify-pool --used-for-app fs:testfs data-pool`
 4. Либо примонтируйте ФС: `vitastor-nfs mount --fs testfs --pool data-pool /mnt/vita`
 5. Либо запустите сетевой NFS-сервер: `vitastor-nfs start --fs testfs --pool data-pool`
 

mon/osd_tree.js

@@ -15,7 +15,7 @@ function get_osd_tree(global_config, state)
         const stat = state.osd.stats[osd_num];
         const osd_cfg = state.config.osd[osd_num];
         let reweight = osd_cfg == null ? 1 : Number(osd_cfg.reweight);
-        if (reweight < 0 || isNaN(reweight))
+        if (isNaN(reweight) || reweight < 0 || reweight > 0)
             reweight = 1;
         if (stat && stat.size && reweight && (state.osd.state[osd_num] || Number(stat.time) >= down_time ||
             osd_cfg && osd_cfg.noout))

mon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor-mon",
-  "version": "1.11.0",
+  "version": "2.1.0",
   "description": "Vitastor SDS monitor service",
   "main": "mon-main.js",
   "scripts": {
@@ -19,6 +19,6 @@
     "eslint-plugin-node": "^11.1.0"
   },
   "engines": {
-    "node": ">=12.0.0"
+    "node": ">=12.1.0"
   }
 }

node-binding/client.cc

@@ -19,11 +19,16 @@
 class NodeVitastorRequest: public Nan::AsyncResource
 {
 public:
-    NodeVitastorRequest(NodeVitastor *cli, v8::Local<v8::Function> cb): Nan::AsyncResource("NodeVitastorRequest")
+    NodeVitastorRequest(NodeVitastor *cli, const v8::Local<v8::Function> & cb): Nan::AsyncResource("NodeVitastorRequest")
     {
         this->cli = cli;
         callback.Reset(cb);
     }
+    ~NodeVitastorRequest()
+    {
+        callback.Reset();
+        buffer_ref.Reset();
+    }
 
     iovec iov;
     std::vector<iovec> iov_list;
@@ -33,6 +38,7 @@ public:
     uint64_t offset = 0, len = 0, version = 0;
     bool with_parents = false;
     Nan::Persistent<v8::Function> callback;
+    Nan::Persistent<v8::Value> buffer_ref;
 };
 
 static uint64_t get_ui64(const v8::Local<v8::Value> & val)
@@ -129,8 +135,8 @@ NodeVitastorRequest* NodeVitastor::get_read_request(const Nan::FunctionCallbackI
         Nan::ThrowError("failed to allocate memory");
         return NULL;
     }
-    v8::Local<v8::Function> callback = info[argpos+2].As<v8::Function>();
-    auto req = new NodeVitastorRequest(this, callback);
+
+    auto req = new NodeVitastorRequest(this, info[argpos+2].As<v8::Function>());
 
     req->offset = offset;
     req->len = len;
@@ -155,7 +161,9 @@ NAN_METHOD(NodeVitastor::Read)
 
     self->Ref();
     vitastor_c_read(self->c, ((pool << (64-POOL_ID_BITS)) | inode), req->offset, req->len, &req->iov, 1, on_read_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 NodeVitastorRequest* NodeVitastor::get_write_request(const Nan::FunctionCallbackInfo<v8::Value> & info, int argpos)
@@ -175,11 +183,11 @@ NodeVitastorRequest* NodeVitastor::get_write_request(const Nan::FunctionCallback
         argpos++;
     }
 
-    v8::Local<v8::Function> callback = info[argpos+2].As<v8::Function>();
-    auto req = new NodeVitastorRequest(this, callback);
+    auto req = new NodeVitastorRequest(this, info[argpos+2].As<v8::Function>());
 
     req->offset = offset;
     req->version = version;
+    req->buffer_ref.Reset(bufarg);
 
     if (bufarg->IsArray())
     {
@@ -224,7 +232,9 @@ NAN_METHOD(NodeVitastor::Write)
         req->iov_list.size() ? req->iov_list.data() : &req->iov,
         req->iov_list.size() ? req->iov_list.size() : 1,
         on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 NodeVitastorRequest* NodeVitastor::get_delete_request(const Nan::FunctionCallbackInfo<v8::Value> & info, int argpos)
@@ -244,8 +254,7 @@ NodeVitastorRequest* NodeVitastor::get_delete_request(const Nan::FunctionCallbac
         argpos++;
     }
 
-    v8::Local<v8::Function> callback = info[argpos+2].As<v8::Function>();
-    auto req = new NodeVitastorRequest(this, callback);
+    auto req = new NodeVitastorRequest(this, info[argpos+2].As<v8::Function>());
 
     req->offset = offset;
     req->len = len;
@@ -271,7 +280,9 @@ NAN_METHOD(NodeVitastor::Delete)
     self->Ref();
     vitastor_c_delete(self->c, ((pool << (64-POOL_ID_BITS)) | inode), req->offset, req->len, req->version,
         on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 // sync(callback(err))
@@ -283,12 +294,13 @@ NAN_METHOD(NodeVitastor::Sync)
 
     NodeVitastor* self = Nan::ObjectWrap::Unwrap<NodeVitastor>(info.This());
 
-    v8::Local<v8::Function> callback = info[0].As<v8::Function>();
-    auto req = new NodeVitastorRequest(self, callback);
+    auto req = new NodeVitastorRequest(self, info[0].As<v8::Function>());
 
     self->Ref();
     vitastor_c_sync(self->c, on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 // read_bitmap(pool, inode, offset, length, with_parents, callback(err, bitmap_buffer))
@@ -305,12 +317,13 @@ NAN_METHOD(NodeVitastor::ReadBitmap)
     uint64_t offset = get_ui64(info[2]);
     uint64_t len = get_ui64(info[3]);
     bool with_parents = Nan::To<bool>(info[4]).FromJust();
-    v8::Local<v8::Function> callback = info[5].As<v8::Function>();
 
-    auto req = new NodeVitastorRequest(self, callback);
+    auto req = new NodeVitastorRequest(self, info[5].As<v8::Function>());
     self->Ref();
     vitastor_c_read_bitmap(self->c, ((pool << (64-POOL_ID_BITS)) | inode), offset, len, with_parents, on_read_bitmap_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 static void on_error(NodeVitastorRequest *req, Nan::Callback & nanCallback, long retval)
@@ -331,11 +344,12 @@ NAN_METHOD(NodeVitastor::OnReady)
     if (info.Length() < 1)
         Nan::ThrowError("Not enough arguments to on_ready(callback(err))");
     NodeVitastor* self = Nan::ObjectWrap::Unwrap<NodeVitastor>(info.This());
-    v8::Local<v8::Function> callback = info[0].As<v8::Function>();
-    auto req = new NodeVitastorRequest(self, callback);
+    auto req = new NodeVitastorRequest(self, info[0].As<v8::Function>());
     self->Ref();
     vitastor_c_on_ready(self->c, on_ready_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(self->c);
+#endif
 }
 
 void NodeVitastor::on_ready_finish(void *opaque, long retval)
@@ -480,7 +494,9 @@ NAN_METHOD(NodeVitastorImage::Create)
     img->Ref();
     cli->Ref();
     vitastor_c_watch_inode(cli->c, (char*)img->name.c_str(), on_watch_start, img);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(cli->c);
+#endif
 
     info.GetReturnValue().Set(info.This());
 }
@@ -552,8 +568,7 @@ NAN_METHOD(NodeVitastorImage::Sync)
 
     NodeVitastorImage* img = Nan::ObjectWrap::Unwrap<NodeVitastorImage>(info.This());
 
-    v8::Local<v8::Function> callback = info[0].As<v8::Function>();
-    auto req = new NodeVitastorRequest(img->cli, callback);
+    auto req = new NodeVitastorRequest(img->cli, info[0].As<v8::Function>());
     req->img = img;
     req->op = NODE_VITASTOR_SYNC;
 
@@ -572,9 +587,8 @@ NAN_METHOD(NodeVitastorImage::ReadBitmap)
     uint64_t offset = get_ui64(info[0]);
     uint64_t len = get_ui64(info[1]);
     bool with_parents = Nan::To<bool>(info[2]).FromJust();
-    v8::Local<v8::Function> callback = info[3].As<v8::Function>();
 
-    auto req = new NodeVitastorRequest(img->cli, callback);
+    auto req = new NodeVitastorRequest(img->cli, info[3].As<v8::Function>());
     req->img = img;
     req->op = NODE_VITASTOR_READ_BITMAP;
     req->offset = offset;
@@ -593,8 +607,7 @@ NAN_METHOD(NodeVitastorImage::GetInfo)
 
     NodeVitastorImage* img = Nan::ObjectWrap::Unwrap<NodeVitastorImage>(info.This());
 
-    v8::Local<v8::Function> callback = info[0].As<v8::Function>();
-    auto req = new NodeVitastorRequest(img->cli, callback);
+    auto req = new NodeVitastorRequest(img->cli, info[0].As<v8::Function>());
     req->img = img;
     req->op = NODE_VITASTOR_GET_INFO;
 
@@ -621,7 +634,9 @@ void NodeVitastorImage::exec_request(NodeVitastorRequest *req)
         uint64_t ino = vitastor_c_inode_get_num(watch);
         cli->Ref();
         vitastor_c_read(cli->c, ino, req->offset, req->len, &req->iov, 1, NodeVitastor::on_read_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
         vitastor_c_uring_handle_events(cli->c);
+#endif
     }
     else if (req->op == NODE_VITASTOR_WRITE)
     {
@@ -631,7 +646,9 @@ void NodeVitastorImage::exec_request(NodeVitastorRequest *req)
             req->iov_list.size() ? req->iov_list.data() : &req->iov,
             req->iov_list.size() ? req->iov_list.size() : 1,
             NodeVitastor::on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
         vitastor_c_uring_handle_events(cli->c);
+#endif
     }
     else if (req->op == NODE_VITASTOR_DELETE)
     {
@@ -639,7 +656,9 @@ void NodeVitastorImage::exec_request(NodeVitastorRequest *req)
         cli->Ref();
         vitastor_c_delete(cli->c, ino, req->offset, req->len, req->version,
             NodeVitastor::on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
         vitastor_c_uring_handle_events(cli->c);
+#endif
     }
     else if (req->op == NODE_VITASTOR_SYNC)
     {
@@ -649,7 +668,9 @@ void NodeVitastorImage::exec_request(NodeVitastorRequest *req)
         if (imm != IMMEDIATE_ALL)
         {
             vitastor_c_sync(cli->c, NodeVitastor::on_write_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
             vitastor_c_uring_handle_events(cli->c);
+#endif
         }
         else
         {
@@ -661,7 +682,9 @@ void NodeVitastorImage::exec_request(NodeVitastorRequest *req)
         uint64_t ino = vitastor_c_inode_get_num(watch);
         cli->Ref();
         vitastor_c_read_bitmap(cli->c, ino, req->offset, req->len, req->with_parents, NodeVitastor::on_read_bitmap_finish, req);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
         vitastor_c_uring_handle_events(cli->c);
+#endif
     }
     else if (req->op == NODE_VITASTOR_GET_INFO)
     {
@@ -779,8 +802,7 @@ NAN_METHOD(NodeVitastorKV::Open)
         cfg[std::string(*Nan::Utf8String(key))] = std::string(*Nan::Utf8String(Nan::Get(jsParams, key).ToLocalChecked()));
     }
 
-    v8::Local<v8::Function> callback = info[3].As<v8::Function>();
-    auto req = new NodeVitastorRequest(kv->cli, callback);
+    auto req = new NodeVitastorRequest(kv->cli, info[3].As<v8::Function>());
 
     kv->Ref();
     kv->dbw->open(inode_id, cfg, [kv, req](int res)
@@ -793,7 +815,9 @@ NAN_METHOD(NodeVitastorKV::Open)
         delete req;
         kv->Unref();
     });
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(kv->cli->c);
+#endif
 }
 
 // close(callback(err))
@@ -805,8 +829,7 @@ NAN_METHOD(NodeVitastorKV::Close)
 
     NodeVitastorKV* kv = Nan::ObjectWrap::Unwrap<NodeVitastorKV>(info.This());
 
-    v8::Local<v8::Function> callback = info[0].As<v8::Function>();
-    auto req = new NodeVitastorRequest(kv->cli, callback);
+    auto req = new NodeVitastorRequest(kv->cli, info[0].As<v8::Function>());
 
     kv->Ref();
     kv->dbw->close([kv, req]()
@@ -817,7 +840,9 @@ NAN_METHOD(NodeVitastorKV::Close)
         delete req;
         kv->Unref();
     });
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(kv->cli->c);
+#endif
 }
 
 // set_config({ ...config })
@@ -861,8 +886,7 @@ void NodeVitastorKV::get_impl(const Nan::FunctionCallbackInfo<v8::Value> & info,
     // FIXME: Handle Buffer too
     std::string key(*Nan::Utf8String(info[0].As<v8::String>()));
 
-    v8::Local<v8::Function> callback = info[1].As<v8::Function>();
-    auto req = new NodeVitastorRequest(kv->cli, callback);
+    auto req = new NodeVitastorRequest(kv->cli, info[1].As<v8::Function>());
 
     kv->Ref();
     kv->dbw->get(key, [kv, req](int res, const std::string & value)
@@ -876,7 +900,9 @@ void NodeVitastorKV::get_impl(const Nan::FunctionCallbackInfo<v8::Value> & info,
         delete req;
         kv->Unref();
     }, allow_cache);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(kv->cli->c);
+#endif
 }
 
 // get(key, callback(err, value))
@@ -925,14 +951,12 @@ NAN_METHOD(NodeVitastorKV::Set)
     std::string key(*Nan::Utf8String(info[0].As<v8::String>()));
     std::string value(*Nan::Utf8String(info[1].As<v8::String>()));
 
-    v8::Local<v8::Function> callback = info[2].As<v8::Function>();
-    NodeVitastorRequest *req = new NodeVitastorRequest(kv->cli, callback), *cas_req = NULL;
+    NodeVitastorRequest *req = new NodeVitastorRequest(kv->cli, info[2].As<v8::Function>()), *cas_req = NULL;
 
     std::function<bool(int, const std::string &)> cas_cb;
     if (info.Length() > 3 && info[3]->IsObject())
     {
-        v8::Local<v8::Function> cas_callback = info[3].As<v8::Function>();
-        cas_req = new NodeVitastorRequest(kv->cli, cas_callback);
+        cas_req = new NodeVitastorRequest(kv->cli, info[3].As<v8::Function>());
         cas_cb = make_cas_callback(cas_req);
     }
 
@@ -949,7 +973,9 @@ NAN_METHOD(NodeVitastorKV::Set)
             delete cas_req;
         kv->Unref();
     }, cas_cb);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(kv->cli->c);
+#endif
 }
 
 // del(key, callback(err), cas_compare(old_value)?)
@@ -964,14 +990,12 @@ NAN_METHOD(NodeVitastorKV::Del)
     // FIXME: Handle Buffer too
     std::string key(*Nan::Utf8String(info[0].As<v8::String>()));
 
-    v8::Local<v8::Function> callback = info[1].As<v8::Function>();
-    NodeVitastorRequest *req = new NodeVitastorRequest(kv->cli, callback), *cas_req = NULL;
+    NodeVitastorRequest *req = new NodeVitastorRequest(kv->cli, info[1].As<v8::Function>()), *cas_req = NULL;
 
     std::function<bool(int, const std::string &)> cas_cb;
     if (info.Length() > 2 && info[2]->IsObject())
     {
-        v8::Local<v8::Function> cas_callback = info[2].As<v8::Function>();
-        cas_req = new NodeVitastorRequest(kv->cli, cas_callback);
+        cas_req = new NodeVitastorRequest(kv->cli, info[2].As<v8::Function>());
         cas_cb = make_cas_callback(cas_req);
     }
 
@@ -988,7 +1012,9 @@ NAN_METHOD(NodeVitastorKV::Del)
             delete cas_req;
         kv->Unref();
     }, cas_cb);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(kv->cli->c);
+#endif
 }
 
 // list(start_key?)
@@ -1068,12 +1094,11 @@ NAN_METHOD(NodeVitastorKVListing::Next)
 
     if (info.Length() > 0)
     {
-        v8::Local<v8::Function> callback = info[0].As<v8::Function>();
         if (list->iter)
         {
             delete list->iter;
         }
-        list->iter = new NodeVitastorRequest(list->kv->cli, callback);
+        list->iter = new NodeVitastorRequest(list->kv->cli, info[0].As<v8::Function>());
     }
     if (!list->handle)
     {
@@ -1109,7 +1134,9 @@ NAN_METHOD(NodeVitastorKVListing::Next)
             list->iter = req;
         list->kv->Unref();
     });
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_c_uring_handle_events(list->kv->cli->c);
+#endif
 }
 
 // close()

node-binding/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor",
-  "version": "1.11.0",
+  "version": "2.1.0",
   "description": "Low-level native bindings to Vitastor client library",
   "main": "index.js",
   "keywords": [
@@ -16,7 +16,7 @@
     "build": "node-gyp rebuild"
   },
   "author": "Vitaliy Filippov",
-  "license": "VNPL-2.0",
+  "license": "VNPL-1.1",
   "dependencies": {
     "bindings": "1.5.0",
     "nan": "^2.19.0"

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 = '1.11.0'
+VITASTOR_VERSION = '2.1.0'
 
 LOG = logging.getLogger(__name__)
 

patches/pve-qemu-9.2-vitastor.patch

@@ -0,0 +1,172 @@
+Index: pve-qemu-kvm-9.2.0/block/meson.build
+===================================================================
+--- pve-qemu-kvm-9.2.0.orig/block/meson.build
++++ pve-qemu-kvm-9.2.0/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-9.2.0/meson.build
+===================================================================
+--- pve-qemu-kvm-9.2.0.orig/meson.build
++++ pve-qemu-kvm-9.2.0/meson.build
+@@ -1590,6 +1590,26 @@ 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
+@@ -2478,6 +2498,7 @@ 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'))
+@@ -4789,6 +4810,7 @@ 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-9.2.0/meson_options.txt
+===================================================================
+--- pve-qemu-kvm-9.2.0.orig/meson_options.txt
++++ pve-qemu-kvm-9.2.0/meson_options.txt
+@@ -200,6 +200,8 @@ 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-9.2.0/qapi/block-core.json
+===================================================================
+--- pve-qemu-kvm-9.2.0.orig/qapi/block-core.json
++++ pve-qemu-kvm-9.2.0/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' },
+@@ -4592,6 +4592,28 @@
+             '*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.
+@@ -5054,6 +5076,7 @@
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+       'vhdx':       'BlockdevOptionsGenericFormat',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'virtio-blk-vfio-pci':
+                     { 'type': 'BlockdevOptionsVirtioBlkVfioPci',
+                       'if': 'CONFIG_BLKIO' },
+@@ -5501,6 +5524,20 @@
+             '*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
+@@ -5722,6 +5759,7 @@
+       'ssh':            'BlockdevCreateOptionsSsh',
+       'vdi':            'BlockdevCreateOptionsVdi',
+       'vhdx':           'BlockdevCreateOptionsVhdx',
++      'vitastor':       'BlockdevCreateOptionsVitastor',
+       'vmdk':           'BlockdevCreateOptionsVmdk',
+       'vpc':            'BlockdevCreateOptionsVpc'
+   } }
+Index: pve-qemu-kvm-9.2.0/scripts/meson-buildoptions.sh
+===================================================================
+--- pve-qemu-kvm-9.2.0.orig/scripts/meson-buildoptions.sh
++++ pve-qemu-kvm-9.2.0/scripts/meson-buildoptions.sh
+@@ -174,6 +174,7 @@ 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" '  rust            Rust support'
+@@ -455,6 +456,8 @@ _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 ;;

patches/qemu-9.2-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 147097c652..2486b3aeb5 100644
+--- a/meson.build
++++ b/meson.build
+@@ -1590,6 +1590,26 @@ 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
+@@ -2474,6 +2494,7 @@ 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'))
+@@ -4778,6 +4799,7 @@ 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 5eeaf3eee5..b04eda29f9 100644
+--- a/meson_options.txt
++++ b/meson_options.txt
+@@ -200,6 +200,8 @@ 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 fd3bcc1c17..41571ac3f9 100644
+--- a/qapi/block-core.json
++++ b/qapi/block-core.json
+@@ -3212,7 +3212,7 @@
+             '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' },
+@@ -4295,6 +4295,28 @@
+             '*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:
+ #
+@@ -4757,6 +4779,7 @@
+       'throttle':   'BlockdevOptionsThrottle',
+       'vdi':        'BlockdevOptionsGenericFormat',
+       'vhdx':       'BlockdevOptionsGenericFormat',
++      'vitastor':   'BlockdevOptionsVitastor',
+       'virtio-blk-vfio-pci':
+                     { 'type': 'BlockdevOptionsVirtioBlkVfioPci',
+                       'if': 'CONFIG_BLKIO' },
+@@ -5198,6 +5221,20 @@
+             '*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:
+ #
+@@ -5420,6 +5457,7 @@
+       '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 a8066aab03..12e650e7d4 100644
+--- a/scripts/meson-buildoptions.sh
++++ b/scripts/meson-buildoptions.sh
+@@ -174,6 +174,7 @@ 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" '  rust            Rust support'
+@@ -455,6 +456,8 @@ _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:        1.11.0
+Version:        2.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.11.0.el7.tar.gz
+Source0:        vitastor-2.1.0.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.11.0
+Version:        2.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.11.0.el8.tar.gz
+Source0:        vitastor-2.1.0.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el9.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.11.0
+Version:        2.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.11.0.el9.tar.gz
+Source0:        vitastor-2.1.0.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="1.11.0")
+add_definitions(-DVITASTOR_VERSION="2.1.0")
 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/blockstore/blockstore_disk.cpp

@@ -8,6 +8,7 @@
 #include "blockstore_impl.h"
 #include "blockstore_disk.h"
 #include "str_util.h"
+#include "allocator.h"
 
 static uint32_t is_power_of_two(uint64_t value)
 {
@@ -83,6 +84,12 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
         throw std::runtime_error("data_csum_type="+config["data_csum_type"]+" is unsupported, only \"crc32c\" and \"none\" are supported");
     }
     csum_block_size = parse_size(config["csum_block_size"]);
+    discard_on_start = config.find("discard_on_start") != config.end() &&
+        (config["discard_on_start"] == "true" || config["discard_on_start"] == "1" || config["discard_on_start"] == "yes");
+    min_discard_size = parse_size(config["min_discard_size"]);
+    if (!min_discard_size)
+        min_discard_size = 1024*1024;
+    discard_granularity = parse_size(config["discard_granularity"]);
     // Validate
     if (!data_block_size)
     {
@@ -172,10 +179,6 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     {
         throw std::runtime_error("journal_offset must be a multiple of journal_block_size = "+std::to_string(journal_block_size));
     }
-    clean_entry_bitmap_size = data_block_size / bitmap_granularity / 8;
-    clean_dyn_size = clean_entry_bitmap_size*2 + (csum_block_size
-        ? data_block_size/csum_block_size*(data_csum_type & 0xFF) : 0);
-    clean_entry_size = sizeof(clean_disk_entry) + clean_dyn_size + 4 /*entry_csum*/;
 }
 
 void blockstore_disk_t::calc_lengths(bool skip_meta_check)
@@ -224,9 +227,13 @@ void blockstore_disk_t::calc_lengths(bool skip_meta_check)
     }
     // required metadata size
     block_count = data_len / data_block_size;
+    clean_entry_bitmap_size = data_block_size / bitmap_granularity / 8;
+    clean_dyn_size = clean_entry_bitmap_size*2 + (csum_block_size
+        ? data_block_size/csum_block_size*(data_csum_type & 0xFF) : 0);
+    clean_entry_size = sizeof(clean_disk_entry) + clean_dyn_size + 4 /*entry_csum*/;
     meta_len = (1 + (block_count - 1 + meta_block_size / clean_entry_size) / (meta_block_size / clean_entry_size)) * meta_block_size;
-    if (meta_format == BLOCKSTORE_META_FORMAT_V1 ||
-        !meta_format && !skip_meta_check && meta_area_size < meta_len && !data_csum_type)
+    bool new_doesnt_fit = (!meta_format && !skip_meta_check && meta_area_size < meta_len && !data_csum_type);
+    if (meta_format == BLOCKSTORE_META_FORMAT_V1 || new_doesnt_fit)
     {
         uint64_t clean_entry_v0_size = sizeof(clean_disk_entry) + 2*clean_entry_bitmap_size;
         uint64_t meta_v0_len = (1 + (block_count - 1 + meta_block_size / clean_entry_v0_size)
@@ -234,7 +241,11 @@ void blockstore_disk_t::calc_lengths(bool skip_meta_check)
         if (meta_format == BLOCKSTORE_META_FORMAT_V1 || meta_area_size >= meta_v0_len)
         {
             // Old metadata fits.
-            printf("Warning: Using old metadata format without checksums because the new format doesn't fit into provided area\n");
+            if (new_doesnt_fit)
+            {
+                printf("Warning: Using old metadata format without checksums because the new format"
+                    " doesn't fit into provided area (%ju bytes required, %ju bytes available)\n", meta_len, meta_area_size);
+            }
             clean_entry_size = clean_entry_v0_size;
             meta_len = meta_v0_len;
             meta_format = BLOCKSTORE_META_FORMAT_V1;
@@ -246,7 +257,7 @@ void blockstore_disk_t::calc_lengths(bool skip_meta_check)
         meta_format = BLOCKSTORE_META_FORMAT_V2;
     if (!skip_meta_check && meta_area_size < meta_len)
     {
-        throw std::runtime_error("Metadata area is too small, need at least "+std::to_string(meta_len)+" bytes");
+        throw std::runtime_error("Metadata area is too small, need at least "+std::to_string(meta_len)+" bytes, have only "+std::to_string(meta_area_size)+" bytes");
     }
     // requested journal size
     if (!skip_meta_check && cfg_journal_size > journal_len)
@@ -415,3 +426,44 @@ void blockstore_disk_t::close_all()
         close(journal_fd);
     data_fd = meta_fd = journal_fd = -1;
 }
+
+// Sadly DISCARD only works through ioctl(), but it seems to always block the device queue,
+// so it's not a big deal that we can only run it synchronously.
+int blockstore_disk_t::trim_data(allocator_t *alloc)
+{
+    int r = 0;
+    uint64_t j = 0, i = 0;
+    uint64_t discarded = 0;
+    for (; i <= block_count; i++)
+    {
+        if (i >= block_count || alloc->get(i))
+        {
+            if (i > j && (i-j)*data_block_size >= min_discard_size)
+            {
+                uint64_t range[2] = { data_offset + j*data_block_size, (i-j)*data_block_size };
+                if (discard_granularity)
+                {
+                    range[1] += range[0];
+                    if (range[1] % discard_granularity)
+                        range[1] = range[1] - (range[1] % discard_granularity);
+                    if (range[0] % discard_granularity)
+                        range[0] = range[0] + discard_granularity - (range[0] % discard_granularity);
+                    if (range[0] >= range[1])
+                        continue;
+                    range[1] -= range[0];
+                }
+                r = ioctl(data_fd, BLKDISCARD, &range);
+                if (r != 0)
+                {
+                    fprintf(stderr, "Failed to execute BLKDISCARD %ju+%ju on %s: %s (code %d)\n",
+                        range[0], range[1], data_device.c_str(), strerror(-r), r);
+                    return -errno;
+                }
+                discarded += range[1];
+            }
+            j = i+1;
+        }
+    }
+    fprintf(stderr, "%s (%ju bytes) of unused data discarded on %s\n", format_size(discarded).c_str(), discarded, data_device.c_str());
+    return 0;
+}

src/blockstore/blockstore_disk.h

@@ -12,6 +12,8 @@
 // Lower byte of checksum type is its length
 #define BLOCKSTORE_CSUM_CRC32C 0x104
 
+class allocator_t;
+
 struct blockstore_disk_t
 {
     std::string data_device, meta_device, journal_device;
@@ -34,14 +36,18 @@ struct blockstore_disk_t
     // I/O modes for data, metadata and journal: direct or "" = O_DIRECT, cached = O_SYNC, directsync = O_DIRECT|O_SYNC
     // O_SYNC without O_DIRECT = use Linux page cache for reads and writes
     std::string data_io, meta_io, journal_io;
+    // Data discard granularity and minimum size (for the sake of performance)
+    bool discard_on_start = false;
+    uint64_t min_discard_size = 1024*1024;
+    uint64_t discard_granularity = 0;
 
     int meta_fd = -1, data_fd = -1, journal_fd = -1;
     uint64_t meta_offset, meta_device_sect, meta_device_size, meta_len, meta_format = 0;
     uint64_t data_offset, data_device_sect, data_device_size, data_len;
     uint64_t journal_offset, journal_device_sect, journal_device_size, journal_len;
 
-    uint32_t block_order;
-    uint64_t block_count;
+    uint32_t block_order = 0;
+    uint64_t block_count = 0;
     uint32_t clean_entry_bitmap_size = 0, clean_entry_size = 0, clean_dyn_size = 0;
 
     void parse_config(std::map<std::string, std::string> & config);
@@ -50,6 +56,7 @@ struct blockstore_disk_t
     void open_journal();
     void calc_lengths(bool skip_meta_check = false);
     void close_all();
+    int trim_data(allocator_t *alloc);
 
     inline uint64_t dirty_dyn_size(uint64_t offset, uint64_t len)
     {

src/blockstore/blockstore_impl.cpp

@@ -12,15 +12,15 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *
     ringloop->register_consumer(&ring_consumer);
     initialized = 0;
     parse_config(config, true);
-    zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.data_block_size);
-    alloc_dyn_data = dsk.clean_dyn_size > sizeof(void*) || dsk.csum_block_size > 0;
     try
     {
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
         calc_lengths();
-        data_alloc = new allocator(dsk.block_count);
+        alloc_dyn_data = dsk.clean_dyn_size > sizeof(void*) || dsk.csum_block_size > 0;
+        zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, dsk.data_block_size);
+        data_alloc = new allocator_t(dsk.block_count);
     }
     catch (std::exception & e)
     {
@@ -34,7 +34,8 @@ blockstore_impl_t::~blockstore_impl_t()
 {
     delete data_alloc;
     delete flusher;
-    free(zero_object);
+    if (zero_object)
+        free(zero_object);
     ringloop->unregister_consumer(&ring_consumer);
     dsk.close_all();
     if (metadata_buffer)
@@ -83,14 +84,20 @@ void blockstore_impl_t::loop()
             {
                 delete journal_init_reader;
                 journal_init_reader = NULL;
-                if (journal.flush_journal)
-                    initialized = 3;
-                else
-                    initialized = 10;
+                initialized = 3;
                 ringloop->wakeup();
             }
         }
         if (initialized == 3)
+        {
+            if (!readonly && dsk.discard_on_start)
+                dsk.trim_data(data_alloc);
+            if (journal.flush_journal)
+                initialized = 4;
+            else
+                initialized = 10;
+        }
+        if (initialized == 4)
         {
             if (readonly)
             {

src/blockstore/blockstore_impl.h

@@ -279,9 +279,9 @@ class blockstore_impl_t
     std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
     int unsynced_big_write_count = 0, unstable_unsynced = 0;
     int unsynced_queued_ops = 0;
-    allocator *data_alloc = NULL;
+    allocator_t *data_alloc = NULL;
     uint64_t used_blocks = 0;
-    uint8_t *zero_object;
+    uint8_t *zero_object = NULL;
 
     void *metadata_buffer = NULL;
 

src/blockstore/blockstore_init.cpp

@@ -138,7 +138,11 @@ resume_1:
                 exit(1);
             }
             hdr->header_csum = csum;
-            bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_V2;
+            if (bs->dsk.meta_format != BLOCKSTORE_META_FORMAT_V2)
+            {
+                bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_V2;
+                bs->dsk.calc_lengths();
+            }
         }
         else if (hdr->version == BLOCKSTORE_META_FORMAT_V1)
         {
@@ -146,11 +150,15 @@ resume_1:
             hdr->csum_block_size = 0;
             hdr->header_csum = 0;
             // Enable compatibility mode - entries without checksums
-            bs->dsk.clean_entry_size = sizeof(clean_disk_entry) + bs->dsk.clean_entry_bitmap_size*2;
-            bs->dsk.meta_len = (1 + (bs->dsk.block_count - 1 + bs->dsk.meta_block_size / bs->dsk.clean_entry_size)
-                / (bs->dsk.meta_block_size / bs->dsk.clean_entry_size)) * bs->dsk.meta_block_size;
-            bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_V1;
-            printf("Warning: Starting with metadata in the old format without checksums, as stored on disk\n");
+            if (bs->dsk.meta_format != BLOCKSTORE_META_FORMAT_V1 ||
+                bs->dsk.data_csum_type != 0 || bs->dsk.csum_block_size != 0)
+            {
+                bs->dsk.data_csum_type = 0;
+                bs->dsk.csum_block_size = 0;
+                bs->dsk.meta_format = BLOCKSTORE_META_FORMAT_V1;
+                bs->dsk.calc_lengths();
+                printf("Warning: Starting with metadata in the old format without checksums, as stored on disk\n");
+            }
         }
         else if (hdr->version > BLOCKSTORE_META_FORMAT_V2)
         {
@@ -338,7 +346,7 @@ bool blockstore_init_meta::handle_meta_block(uint8_t *buf, uint64_t entries_per_
                 uint32_t *entry_csum = (uint32_t*)((uint8_t*)entry + bs->dsk.clean_entry_size - 4);
                 if (*entry_csum != crc32c(0, entry, bs->dsk.clean_entry_size - 4))
                 {
-                    printf("Metadata entry %ju is corrupt (checksum mismatch), skipping\n", done_cnt+i);
+                    printf("Metadata entry %ju is corrupt (checksum mismatch: %08x vs %08x), skipping\n", done_cnt+i, *entry_csum, crc32c(0, entry, bs->dsk.clean_entry_size - 4));
                     // zero out the invalid entry, otherwise we'll hit "tried to overwrite non-zero metadata entry" later
                     if (bs->inmemory_meta)
                     {

src/client/CMakeLists.txt

@@ -7,10 +7,14 @@ set(MSGR_RDMA "")
 if (IBVERBS_LIBRARIES)
 	set(MSGR_RDMA "msgr_rdma.cpp")
 endif (IBVERBS_LIBRARIES)
+set(MSGR_RDMACM "")
+if (RDMACM_LIBRARIES)
+	set(MSGR_RDMACM "msgr_rdmacm.cpp")
+endif (RDMACM_LIBRARIES)
 add_library(vitastor_common STATIC
 	../util/epoll_manager.cpp etcd_state_client.cpp messenger.cpp ../util/addr_util.cpp
 	msgr_stop.cpp msgr_op.cpp msgr_send.cpp msgr_receive.cpp ../util/ringloop.cpp ../../json11/json11.cpp
-	http_client.cpp osd_ops.cpp pg_states.cpp ../util/timerfd_manager.cpp ../util/str_util.cpp ../util/json_util.cpp ${MSGR_RDMA}
+	http_client.cpp osd_ops.cpp pg_states.cpp ../util/timerfd_manager.cpp ../util/str_util.cpp ../util/json_util.cpp ${MSGR_RDMA} ${MSGR_RDMACM}
 )
 target_link_libraries(vitastor_common pthread)
 target_compile_options(vitastor_common PUBLIC -fPIC)
@@ -28,6 +32,7 @@ target_link_libraries(vitastor_client
 	vitastor_cli
 	${LIBURING_LIBRARIES}
 	${IBVERBS_LIBRARIES}
+	${RDMACM_LIBRARIES}
 )
 set_target_properties(vitastor_client PROPERTIES VERSION ${VITASTOR_VERSION} SOVERSION 0)
 configure_file(vitastor.pc.in vitastor.pc @ONLY)

src/client/cluster_client.cpp

@@ -29,8 +29,7 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
         if (msgr.osd_peer_fds.find(peer_osd) != msgr.osd_peer_fds.end())
         {
             // peer_osd just connected
-            // retry operations waiting for connection immediately
-            continue_ops(client_retry_interval);
+            continue_ops();
             continue_lists();
             continue_raw_ops(peer_osd);
         }

src/client/cluster_client.h

@@ -83,6 +83,9 @@ class writeback_cache_t;
 // FIXME: Split into public and private interfaces
 class cluster_client_t
 {
+#ifdef __MOCK__
+public:
+#endif
     timerfd_manager_t *tfd;
     ring_loop_t *ringloop;
 
@@ -144,8 +147,6 @@ public:
 
     bool get_immediate_commit(uint64_t inode);
 
-    void continue_ops(int time_passed = 0);
-
     void list_inode(inode_t inode, uint64_t min_offset, uint64_t max_offset, int max_parallel_pgs, std::function<void(
         int status, int pgs_left, pg_num_t pg_num, std::set<object_id>&& objects)> pg_callback);
 
@@ -153,6 +154,11 @@ public:
     //inline uint64_t get_bs_block_size() { return st_cli.global_block_size; }
     uint64_t next_op_id();
 
+#ifndef __MOCK__
+protected:
+#endif
+    void continue_ops(int time_passed = 0);
+
 protected:
     bool affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd);
     bool affects_pg(uint64_t inode, uint64_t offset, uint64_t len, pool_id_t pool_id, pg_num_t pg_num);

src/client/cluster_client_impl.h

@@ -52,6 +52,7 @@ public:
     bool read_from_cache(cluster_op_t *op, uint32_t bitmap_granularity);
     void flush_buffers(cluster_client_t *cli, dirty_buf_it_t from_it, dirty_buf_it_t to_it);
     void mark_flush_written(uint64_t inode, uint64_t offset, uint64_t len, uint64_t flush_id);
+    void delete_flush(uint64_t inode, uint64_t offset, uint64_t len, uint64_t flush_id);
     void fsync_start();
     void fsync_error();
     void fsync_ok();

src/client/cluster_client_wb.cpp

@@ -9,7 +9,7 @@ writeback_cache_t::~writeback_cache_t()
 {
     for (auto & bp: dirty_buffers)
     {
-        if (!--(*bp.second.refcnt))
+        if (bp.second.buf && !--(*bp.second.refcnt))
         {
             free(bp.second.refcnt); // refcnt is allocated with the buffer
         }
@@ -115,7 +115,10 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flu
                     .flush_id = dirty_it->second.flush_id,
                     .refcnt = dirty_it->second.refcnt,
                 });
-                (*dirty_it->second.refcnt)++;
+                if (dirty_it->second.buf)
+                {
+                    (*dirty_it->second.refcnt)++;
+                }
                 if (dirty_it->second.state == CACHE_DIRTY)
                 {
                     if (dirty_it->second.buf)
@@ -193,7 +196,7 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flu
                     writeback_queue_size++;
                 }
             }
-            if (!--(*dirty_it->second.refcnt))
+            if (dirty_it->second.buf && !--(*dirty_it->second.refcnt))
             {
                 free(dirty_it->second.refcnt);
             }
@@ -204,7 +207,10 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flu
     bool is_del = op->opcode == OSD_OP_DELETE;
     uint64_t *refcnt = is_del ? NULL : (uint64_t*)malloc_or_die(sizeof(uint64_t) + op->len);
     uint8_t *buf = is_del ? NULL : ((uint8_t*)refcnt + sizeof(uint64_t));
-    *refcnt = 1;
+    if (!is_del)
+    {
+        *refcnt = 1;
+    }
     dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
         .inode = op->inode,
         .stripe = op->offset,
@@ -326,7 +332,14 @@ void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from
             }
             flushed_buffers.erase(fl_it++);
         }
-        mark_flush_written(op->inode, op->offset, op->len, flush_id);
+        if (op->flags & OP_IMMEDIATE_COMMIT)
+        {
+            delete_flush(op->inode, op->offset, op->len, flush_id);
+        }
+        else
+        {
+            mark_flush_written(op->inode, op->offset, op->len, flush_id);
+        }
         delete op;
         writebacks_active--;
         // We can't call execute_internal because it affects an invalid copy of the list here
@@ -344,6 +357,25 @@ void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from
     }
 }
 
+void writeback_cache_t::delete_flush(uint64_t inode, uint64_t offset, uint64_t len, uint64_t flush_id)
+{
+    for (auto dirty_it = find_dirty(inode, offset);
+        dirty_it != dirty_buffers.end() && dirty_it->first.inode == inode &&
+        dirty_it->first.stripe < offset+len; )
+    {
+        if (dirty_it->second.flush_id == flush_id && dirty_it->second.state == CACHE_REPEATING)
+        {
+            if (dirty_it->second.buf && !--(*dirty_it->second.refcnt))
+            {
+                free(dirty_it->second.refcnt);
+            }
+            dirty_buffers.erase(dirty_it++);
+        }
+        else
+            dirty_it++;
+    }
+}
+
 void writeback_cache_t::mark_flush_written(uint64_t inode, uint64_t offset, uint64_t len, uint64_t flush_id)
 {
     for (auto dirty_it = find_dirty(inode, offset);
@@ -560,8 +592,10 @@ void writeback_cache_t::fsync_ok()
     {
         if (uw_it->second.state == CACHE_FLUSHING)
         {
-            if (!--(*uw_it->second.refcnt))
+            if (uw_it->second.buf && !--(*uw_it->second.refcnt))
+            {
                 free(uw_it->second.refcnt);
+            }
             dirty_buffers.erase(uw_it++);
         }
         else

src/client/etcd_state_client.cpp

@@ -31,6 +31,11 @@ etcd_state_client_t::~etcd_state_client_t()
         keepalive_client = NULL;
     }
 #endif
+    if (load_pgs_timer_id >= 0)
+    {
+        tfd->clear_timer(load_pgs_timer_id);
+        load_pgs_timer_id = -1;
+    }
 }
 
 #ifndef __MOCK__
@@ -143,6 +148,7 @@ void etcd_state_client_t::etcd_call(std::string api, json11::Json payload, int t
                 }
                 if (interval > 0)
                 {
+                    // FIXME: Prevent destruction of etcd_state_client if timers or requests are active
                     tfd->set_timer(interval, false, [this, api, payload, timeout, retries, interval, callback](int)
                     {
                         etcd_call(api, payload, timeout, retries, interval, callback);
@@ -176,7 +182,7 @@ void etcd_state_client_t::add_etcd_url(std::string addr)
             exit(1);
         }
         if (!local_ips.size())
-            local_ips = getifaddr_list(std::vector<std::string>(), true);
+            local_ips = getifaddr_list(std::vector<addr_mask_t>(), true);
         std::string check_addr;
         int pos = addr.find('/');
         int pos2 = addr.find(':');
@@ -271,6 +277,11 @@ void etcd_state_client_t::parse_config(const json11::Json & config)
     {
         this->etcd_quick_timeout = 1000;
     }
+    this->etcd_min_reload_interval = config["etcd_min_reload_interval"].uint64_value();
+    if (this->etcd_min_reload_interval <= 0)
+    {
+        this->etcd_min_reload_interval = 50;
+    }
     if (this->etcd_ws_keepalive_interval != old_etcd_ws_keepalive_interval && ws_keepalive_timer >= 0)
     {
 #ifndef __MOCK__
@@ -603,6 +614,23 @@ void etcd_state_client_t::load_global_config()
 
 void etcd_state_client_t::load_pgs()
 {
+    timespec tv;
+    clock_gettime(CLOCK_REALTIME, &tv);
+    uint64_t ms_passed = (tv.tv_sec-etcd_last_reload.tv_sec)*1000 + (tv.tv_nsec-etcd_last_reload.tv_nsec)/1000000;
+    if (ms_passed < etcd_min_reload_interval)
+    {
+        if (load_pgs_timer_id < 0)
+        {
+            load_pgs_timer_id = tfd->set_timer(etcd_min_reload_interval+50-ms_passed, false, [this](int) { load_pgs(); });
+        }
+        return;
+    }
+    etcd_last_reload = tv;
+    if (load_pgs_timer_id >= 0)
+    {
+        tfd->clear_timer(load_pgs_timer_id);
+        load_pgs_timer_id = -1;
+    }
     json11::Json::array txn = {
         json11::Json::object {
             { "request_range", json11::Json::object {
@@ -889,7 +917,11 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             if (!pc.scrub_interval)
                 pc.scrub_interval = 0;
             // Mark pool as VitastorFS pool (disable per-inode stats and block volume creation)
-            pc.used_for_fs = pool_item.second["used_for_fs"].as_string();
+            pc.used_for_app = pool_item.second["used_for_fs"].as_string();
+            if (pc.used_for_app != "")
+                pc.used_for_app = "fs:"+pc.used_for_app;
+            else
+                pc.used_for_app = pool_item.second["used_for_app"].as_string();
             // Immediate Commit Mode
             pc.immediate_commit = pool_item.second["immediate_commit"].is_string()
                 ? parse_immediate_commit(pool_item.second["immediate_commit"].string_value(), IMMEDIATE_ALL)

src/client/etcd_state_client.h

@@ -61,7 +61,7 @@ struct pool_config_t
     uint64_t pg_stripe_size;
     std::map<pg_num_t, pg_config_t> pg_config;
     uint64_t scrub_interval;
-    std::string used_for_fs;
+    std::string used_for_app;
     int backfillfull;
 };
 
@@ -108,6 +108,7 @@ public:
     int max_etcd_attempts = 5;
     int etcd_quick_timeout = 1000;
     int etcd_slow_timeout = 5000;
+    int etcd_min_reload_interval = 1000;
     bool infinite_start = true;
     uint64_t global_block_size = DEFAULT_BLOCK_SIZE;
     uint32_t global_bitmap_granularity = DEFAULT_BITMAP_GRANULARITY;
@@ -122,6 +123,8 @@ public:
     uint64_t etcd_watch_revision_config = 0;
     uint64_t etcd_watch_revision_osd = 0;
     uint64_t etcd_watch_revision_pg = 0;
+    timespec etcd_last_reload = {};
+    int load_pgs_timer_id = -1;
     std::map<pool_id_t, pool_config_t> pool_config;
     std::map<osd_num_t, json11::Json> peer_states;
     std::set<osd_num_t> seen_peers;

src/client/fio_cluster.cpp

@@ -377,7 +377,7 @@ static void io_callback(void *opaque, long retval)
     bsd->completed.push_back(io);
     if (bsd->trace)
     {
-        printf("--- %s 0x%jx retval=%ld\n", io->ddir == DDIR_READ ? "READ" :
+        printf("--- %s 0x%jx retval=%jd\n", io->ddir == DDIR_READ ? "READ" :
             (io->ddir == DDIR_WRITE ? "WRITE" : "SYNC"), (uint64_t)io, retval);
     }
 }

src/client/messenger.cpp

@@ -117,29 +117,53 @@ void msgr_iothread_t::run()
 
 void osd_messenger_t::init()
 {
+#ifdef WITH_RDMACM
+    if (use_rdmacm)
+    {
+        // RDMA-CM only requires the event channel. All the remaining work is done separately
+        rdmacm_evch = rdma_create_event_channel();
+        if (!rdmacm_evch)
+        {
+            // ENODEV means that the client doesn't have RDMA devices available
+            if (errno != ENODEV || log_level > 0)
+                fprintf(stderr, "Failed to initialize RDMA-CM event channel: %s (code %d)\n", strerror(errno), errno);
+        }
+        else
+        {
+            fcntl(rdmacm_evch->fd, F_SETFL, fcntl(rdmacm_evch->fd, F_GETFL, 0) | O_NONBLOCK);
+            tfd->set_fd_handler(rdmacm_evch->fd, false, [this](int rdmacm_eventfd, int epoll_events)
+            {
+                handle_rdmacm_events();
+            });
+        }
+    }
+    else
+#endif
 #ifdef WITH_RDMA
     if (use_rdma)
     {
-        rdma_context = msgr_rdma_context_t::create(
-            osd_networks, rdma_device != "" ? rdma_device.c_str() : NULL,
+        rdma_contexts = msgr_rdma_context_t::create_all(
+            osd_num && osd_cluster_network_masks.size() ? osd_cluster_network_masks : osd_network_masks,
+            rdma_device != "" ? rdma_device.c_str() : NULL,
             rdma_port_num, rdma_gid_index, rdma_mtu, rdma_odp, log_level
         );
-        if (!rdma_context)
+        if (!rdma_contexts.size())
         {
             if (log_level > 0)
                 fprintf(stderr, "[OSD %ju] Couldn't initialize RDMA, proceeding with TCP only\n", osd_num);
         }
         else
         {
-            rdma_max_sge = rdma_max_sge < rdma_context->attrx.orig_attr.max_sge
-                ? rdma_max_sge : rdma_context->attrx.orig_attr.max_sge;
             fprintf(stderr, "[OSD %ju] RDMA initialized successfully\n", osd_num);
-            fcntl(rdma_context->channel->fd, F_SETFL, fcntl(rdma_context->channel->fd, F_GETFL, 0) | O_NONBLOCK);
-            tfd->set_fd_handler(rdma_context->channel->fd, false, [this](int notify_fd, int epoll_events)
+            for (msgr_rdma_context_t* rdma_context: rdma_contexts)
             {
-                handle_rdma_events();
-            });
-            handle_rdma_events();
+                fcntl(rdma_context->channel->fd, F_SETFL, fcntl(rdma_context->channel->fd, F_GETFL, 0) | O_NONBLOCK);
+                tfd->set_fd_handler(rdma_context->channel->fd, false, [this, rdma_context](int notify_fd, int epoll_events)
+                {
+                    handle_rdma_events(rdma_context);
+                });
+                handle_rdma_events(rdma_context);
+            }
         }
     }
 #endif
@@ -247,10 +271,19 @@ osd_messenger_t::~osd_messenger_t()
         iothreads.clear();
     }
 #ifdef WITH_RDMA
-    if (rdma_context)
+    for (auto rdma_context: rdma_contexts)
     {
         delete rdma_context;
     }
+    rdma_contexts.clear();
+#endif
+#ifdef WITH_RDMACM
+    if (rdmacm_evch)
+    {
+        tfd->set_fd_handler(rdmacm_evch->fd, false, NULL);
+        rdma_destroy_event_channel(rdmacm_evch);
+        rdmacm_evch = NULL;
+    }
 #endif
 }
 
@@ -262,10 +295,14 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         // RDMA is on by default in RDMA-enabled builds
         this->use_rdma = config["use_rdma"].bool_value() || config["use_rdma"].uint64_value() != 0;
     }
+#ifdef WITH_RDMACM
+    // Use RDMA CM? (required for iWARP and may be useful for IB)
+    // FIXME: Only parse during start
+    this->use_rdmacm = config["use_rdmacm"].bool_value() || config["use_rdmacm"].uint64_value() != 0;
+    this->disable_tcp = this->use_rdmacm && (config["disable_tcp"].bool_value() || config["disable_tcp"].uint64_value() != 0);
+#endif
     this->rdma_device = config["rdma_device"].string_value();
     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();
@@ -282,15 +319,6 @@ 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();
@@ -314,23 +342,87 @@ void osd_messenger_t::parse_config(const json11::Json & config)
     if (!this->osd_ping_timeout)
         this->osd_ping_timeout = 5;
     this->log_level = config["log_level"].uint64_value();
+    // OSD public & cluster networks
+    this->osd_networks.clear();
+    if (config["osd_network"].is_string())
+        this->osd_networks.push_back(config["osd_network"].string_value());
+    else
+        for (auto v: config["osd_network"].array_items())
+            this->osd_networks.push_back(v.string_value());
+    this->osd_cluster_networks.clear();
+    if (config["osd_cluster_network"].is_string())
+        this->osd_cluster_networks.push_back(config["osd_cluster_network"].string_value());
+    else
+        for (auto v: config["osd_cluster_network"].array_items())
+            this->osd_cluster_networks.push_back(v.string_value());
+    if (this->osd_cluster_networks.size())
+        for (auto & net: this->osd_cluster_networks)
+            for (int i = this->osd_networks.size()-1; i >= 0; i--)
+                if (this->osd_networks[i] == net)
+                    this->osd_networks.erase(this->osd_networks.begin()+i, this->osd_networks.begin()+i+1);
+    this->osd_network_masks.clear();
+    for (auto & netstr: this->osd_networks)
+        this->osd_network_masks.push_back(cidr_parse(netstr));
+    this->osd_cluster_network_masks.clear();
+    for (auto & netstr: this->osd_cluster_networks)
+        this->osd_cluster_network_masks.push_back(cidr_parse(netstr));
+    this->all_osd_networks.clear();
+    this->all_osd_networks.insert(this->all_osd_networks.end(), this->osd_networks.begin(), this->osd_networks.end());
+    this->all_osd_networks.insert(this->all_osd_networks.end(), this->osd_cluster_networks.begin(), this->osd_cluster_networks.end());
+    this->all_osd_network_masks.clear();
+    this->all_osd_network_masks.insert(this->all_osd_network_masks.end(), this->osd_network_masks.begin(), this->osd_network_masks.end());
+    this->all_osd_network_masks.insert(this->all_osd_network_masks.end(), this->osd_cluster_network_masks.begin(), this->osd_cluster_network_masks.end());
+    if (!this->osd_networks.size())
+    {
+        this->osd_networks = this->osd_cluster_networks;
+        this->osd_network_masks = this->osd_cluster_network_masks;
+    }
 }
 
 void osd_messenger_t::connect_peer(uint64_t peer_osd, json11::Json peer_state)
 {
-    if (wanted_peers.find(peer_osd) == wanted_peers.end())
+    if (wanted_peers[peer_osd].raw_address_list != peer_state["addresses"])
     {
-        wanted_peers[peer_osd] = (osd_wanted_peer_t){
-            .address_list = peer_state["addresses"],
-            .port = (int)peer_state["port"].int64_value(),
-        };
+        wanted_peers[peer_osd].raw_address_list = peer_state["addresses"];
+        // We are an OSD -> try to select a cluster address
+        // We are a client -> try to select a public address
+        // OSD only has 1 address -> don't try anything, it's pointless
+        // FIXME: Maybe support optional fallback from cluster to public network?
+        auto & match_masks = (this->osd_num ? osd_cluster_network_masks : osd_network_masks);
+        if (peer_state["addresses"].array_items().size() > 1 && match_masks.size())
+        {
+            json11::Json::array address_list;
+            for (auto json_addr: peer_state["addresses"].array_items())
+            {
+                struct sockaddr_storage addr;
+                auto ok = string_to_addr(json_addr.string_value(), false, 0, &addr);
+                if (ok)
+                {
+                    bool matches = false;
+                    for (auto & mask: match_masks)
+                    {
+                        if (cidr_sockaddr_match(addr, mask))
+                        {
+                            matches = true;
+                            break;
+                        }
+                    }
+                    if (matches)
+                        address_list.push_back(json_addr);
+                }
+            }
+            if (!address_list.size())
+                address_list = peer_state["addresses"].array_items();
+            wanted_peers[peer_osd].address_list = address_list;
+        }
+        else
+            wanted_peers[peer_osd].address_list = peer_state["addresses"];
+        wanted_peers[peer_osd].address_changed = true;
     }
-    else
-    {
-        wanted_peers[peer_osd].address_list = peer_state["addresses"];
-        wanted_peers[peer_osd].port = (int)peer_state["port"].int64_value();
-    }
-    wanted_peers[peer_osd].address_changed = true;
+#ifdef WITH_RDMACM
+    wanted_peers[peer_osd].rdmacm_port = (int)peer_state["rdmacm_port"].int64_value();
+#endif
+    wanted_peers[peer_osd].port = (int)peer_state["port"].int64_value();
     try_connect_peer(peer_osd);
 }
 
@@ -355,12 +447,24 @@ void osd_messenger_t::try_connect_peer(uint64_t peer_osd)
     wp.cur_addr = wp.address_list[wp.address_index].string_value();
     wp.cur_port = wp.port;
     wp.connecting = true;
-    try_connect_peer_addr(peer_osd, wp.cur_addr.c_str(), wp.cur_port);
+#ifdef WITH_RDMACM
+    if (use_rdmacm && wp.rdmacm_port)
+        rdmacm_try_connect_peer(peer_osd, wp.cur_addr.c_str(), wp.rdmacm_port, wp.cur_port);
+    else
+#endif
+        try_connect_peer_tcp(peer_osd, wp.cur_addr.c_str(), wp.cur_port);
 }
 
-void osd_messenger_t::try_connect_peer_addr(osd_num_t peer_osd, const char *peer_host, int peer_port)
+void osd_messenger_t::try_connect_peer_tcp(osd_num_t peer_osd, const char *peer_host, int peer_port)
 {
     assert(peer_osd != this->osd_num);
+#ifdef WITH_RDMACM
+    if (disable_tcp)
+    {
+        on_connect_peer(peer_osd, -EINVAL);
+        return;
+    }
+#endif
     struct sockaddr_storage addr;
     if (!string_to_addr(peer_host, 0, peer_port, &addr))
     {
@@ -524,20 +628,30 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
         },
     };
 #ifdef WITH_RDMA
-    if (rdma_context)
+    if (!use_rdmacm && rdma_contexts.size())
     {
-        cl->rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_max_msg);
-        if (cl->rdma_conn)
+        // Choose the right context for the selected network
+        msgr_rdma_context_t *selected_ctx = choose_rdma_context(cl);
+        if (!selected_ctx)
         {
-            json11::Json payload = json11::Json::object {
-                { "connect_rdma", cl->rdma_conn->addr.to_string() },
-                { "rdma_max_msg", cl->rdma_conn->max_msg },
-            };
-            std::string payload_str = payload.dump();
-            op->req.show_conf.json_len = payload_str.size();
-            op->buf = malloc_or_die(payload_str.size());
-            op->iov.push_back(op->buf, payload_str.size());
-            memcpy(op->buf, payload_str.c_str(), payload_str.size());
+            if (log_level > 0)
+                fprintf(stderr, "No RDMA context for OSD %ju connection (peer %d), using only TCP\n", cl->osd_num, cl->peer_fd);
+        }
+        else
+        {
+            cl->rdma_conn = msgr_rdma_connection_t::create(selected_ctx, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_max_msg);
+            if (cl->rdma_conn)
+            {
+                json11::Json payload = json11::Json::object {
+                    { "connect_rdma", cl->rdma_conn->addr.to_string() },
+                    { "rdma_max_msg", cl->rdma_conn->max_msg },
+                };
+                std::string payload_str = payload.dump();
+                op->req.show_conf.json_len = payload_str.size();
+                op->buf = malloc_or_die(payload_str.size());
+                op->iov.push_back(op->buf, payload_str.size());
+                memcpy(op->buf, payload_str.c_str(), payload_str.size());
+            }
         }
     }
 #endif
@@ -582,29 +696,23 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
         {
             osd_num_t peer_osd = cl->osd_num;
             stop_client(op->peer_fd);
-            on_connect_peer(peer_osd, -1);
+            on_connect_peer(peer_osd, -EINVAL);
             delete op;
             return;
         }
 #ifdef WITH_RDMA
-        if (config["rdma_address"].is_string())
+        if (!use_rdmacm && cl->rdma_conn && config["rdma_address"].is_string())
         {
             msgr_rdma_address_t addr;
             if (!msgr_rdma_address_t::from_string(config["rdma_address"].string_value().c_str(), &addr) ||
                 cl->rdma_conn->connect(&addr) != 0)
             {
                 fprintf(
-                    stderr, "Failed to connect to OSD %ju (address %s) using RDMA\n",
+                    stderr, "Failed to connect to OSD %ju (address %s) using RDMA, switching back to TCP\n",
                     cl->osd_num, config["rdma_address"].string_value().c_str()
                 );
                 delete cl->rdma_conn;
                 cl->rdma_conn = NULL;
-                // FIXME: Keep TCP connection in this case
-                osd_num_t peer_osd = cl->osd_num;
-                stop_client(cl->peer_fd);
-                on_connect_peer(peer_osd, -1);
-                delete op;
-                return;
             }
             else
             {
@@ -673,9 +781,29 @@ void osd_messenger_t::accept_connections(int listen_fd)
 }
 
 #ifdef WITH_RDMA
+msgr_rdma_context_t* osd_messenger_t::choose_rdma_context(osd_client_t *cl)
+{
+    // Choose the right context for the selected network
+    msgr_rdma_context_t *selected_ctx = NULL;
+    for (auto rdma_ctx: rdma_contexts)
+    {
+        if (!rdma_ctx->net_mask.family && !selected_ctx ||
+            rdma_ctx->net_mask.family && cidr_sockaddr_match(cl->peer_addr, rdma_ctx->net_mask))
+        {
+            selected_ctx = rdma_ctx;
+        }
+    }
+    return selected_ctx;
+}
+
 bool osd_messenger_t::is_rdma_enabled()
 {
-    return rdma_context != NULL;
+    return rdma_contexts.size() > 0;
+}
+
+bool osd_messenger_t::is_use_rdmacm()
+{
+    return use_rdmacm;
 }
 #endif
 

src/client/messenger.h

@@ -16,6 +16,7 @@
 #include "json11/json11.hpp"
 #include "msgr_op.h"
 #include "timerfd_manager.h"
+#include "addr_util.h"
 #include <ringloop.h>
 
 #define CL_READ_HDR 1
@@ -49,10 +50,10 @@ struct osd_client_t
 {
     int refs = 0;
 
-    sockaddr_storage peer_addr;
-    int peer_port;
+    sockaddr_storage peer_addr = {};
+    int peer_port = 0;
     int peer_fd = -1;
-    int peer_state;
+    int peer_state = 0;
     int connect_timeout_id = -1;
     int ping_time_remaining = 0;
     int idle_time_remaining = 0;
@@ -93,13 +94,17 @@ struct osd_client_t
 
 struct osd_wanted_peer_t
 {
+    json11::Json raw_address_list;
     json11::Json address_list;
-    int port;
-    time_t last_connect_attempt;
-    bool connecting, address_changed;
-    int address_index;
+    int port = 0;
+#ifdef WITH_RDMACM
+    int rdmacm_port = 0;
+#endif
+    time_t last_connect_attempt = 0;
+    bool connecting = false, address_changed = false;
+    int address_index = 0;
     std::string cur_addr;
-    int cur_port;
+    int cur_port = 0;
 };
 
 struct osd_op_stats_t
@@ -149,6 +154,15 @@ public:
 };
 #endif
 
+#ifdef WITH_RDMA
+struct rdma_event_channel;
+struct rdma_cm_id;
+struct rdma_cm_event;
+struct ibv_context;
+struct osd_messenger_t;
+struct rdmacm_connecting_t;
+#endif
+
 struct osd_messenger_t
 {
 protected:
@@ -165,14 +179,19 @@ protected:
 
 #ifdef WITH_RDMA
     bool use_rdma = true;
-    std::vector<std::string> osd_networks;
+    bool use_rdmacm = false;
+    bool disable_tcp = false;
     std::string rdma_device;
-    uint64_t rdma_port_num = 1, rdma_mtu = 0;
+    uint64_t rdma_port_num = 1;
+    int rdma_mtu = 0;
     int rdma_gid_index = -1;
-    msgr_rdma_context_t *rdma_context = NULL;
+    std::vector<msgr_rdma_context_t *> rdma_contexts;
     uint64_t rdma_max_sge = 0, rdma_max_send = 0, rdma_max_recv = 0;
     uint64_t rdma_max_msg = 0;
     bool rdma_odp = false;
+    rdma_event_channel *rdmacm_evch = NULL;
+    std::map<rdma_cm_id*, osd_client_t*> rdmacm_connections;
+    std::map<rdma_cm_id*, rdmacm_connecting_t*> rdmacm_connecting;
 #endif
 
     std::vector<msgr_iothread_t*> iothreads;
@@ -190,6 +209,12 @@ public:
     std::map<int, osd_client_t*> clients;
     std::map<osd_num_t, osd_wanted_peer_t> wanted_peers;
     std::map<uint64_t, int> osd_peer_fds;
+    std::vector<std::string> osd_networks;
+    std::vector<addr_mask_t> osd_network_masks;
+    std::vector<std::string> osd_cluster_networks;
+    std::vector<addr_mask_t> osd_cluster_network_masks;
+    std::vector<std::string> all_osd_networks;
+    std::vector<addr_mask_t> all_osd_network_masks;
     // op statistics
     osd_op_stats_t stats, recovery_stats;
 
@@ -216,13 +241,18 @@ public:
     bool is_rdma_enabled();
     bool connect_rdma(int peer_fd, std::string rdma_address, uint64_t client_max_msg);
 #endif
+#ifdef WITH_RDMACM
+    bool is_use_rdmacm();
+    rdma_cm_id *rdmacm_listen(const std::string & bind_address, int rdmacm_port, int *bound_port, int log_level);
+    void rdmacm_destroy_listener(rdma_cm_id *listener);
+#endif
 
     void inc_op_stats(osd_op_stats_t & stats, uint64_t opcode, timespec & tv_begin, timespec & tv_end, uint64_t len);
     void measure_exec(osd_op_t *cur_op);
 
 protected:
     void try_connect_peer(uint64_t osd_num);
-    void try_connect_peer_addr(osd_num_t peer_osd, const char *peer_host, int peer_port);
+    void try_connect_peer_tcp(osd_num_t peer_osd, const char *peer_host, int peer_port);
     void handle_peer_epoll(int peer_fd, int epoll_events);
     void handle_connect_epoll(int peer_fd);
     void on_connect_peer(osd_num_t peer_osd, int peer_fd);
@@ -247,6 +277,18 @@ protected:
     void try_send_rdma_odp(osd_client_t *cl);
     void try_send_rdma_nodp(osd_client_t *cl);
     bool try_recv_rdma(osd_client_t *cl);
-    void handle_rdma_events();
+    void handle_rdma_events(msgr_rdma_context_t *rdma_context);
+    msgr_rdma_context_t* choose_rdma_context(osd_client_t *cl);
+#endif
+#ifdef WITH_RDMACM
+    void handle_rdmacm_events();
+    msgr_rdma_context_t* rdmacm_get_context(ibv_context *verbs);
+    msgr_rdma_context_t* rdmacm_create_qp(rdma_cm_id *cmid);
+    void rdmacm_accept(rdma_cm_event *ev);
+    void rdmacm_try_connect_peer(uint64_t peer_osd, const std::string & addr, int rdmacm_port, int fallback_tcp_port);
+    void rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res);
+    void rdmacm_address_resolved(rdma_cm_event *ev);
+    void rdmacm_route_resolved(rdma_cm_event *ev);
+    void rdmacm_established(rdma_cm_event *ev);
 #endif
 };

src/client/msgr_rdma.cpp

@@ -3,10 +3,35 @@
 
 #include <stdio.h>
 #include <stdlib.h>
-#include "addr_util.h"
 #include "msgr_rdma.h"
 #include "messenger.h"
 
+static uint32_t ibv_mtu_to_bytes(ibv_mtu mtu)
+{
+    switch (mtu)
+    {
+        case IBV_MTU_256:  return 256;
+        case IBV_MTU_512:  return 512;
+        case IBV_MTU_1024: return 1024;
+        case IBV_MTU_2048: return 2048;
+        case IBV_MTU_4096: return 4096;
+    }
+    return 4096;
+}
+
+static ibv_mtu bytes_to_ibv_mtu(uint32_t mtu)
+{
+    switch (mtu)
+    {
+        case 256:  return IBV_MTU_256;
+        case 512:  return IBV_MTU_512;
+        case 1024: return IBV_MTU_1024;
+        case 2048: return IBV_MTU_2048;
+        case 4096: return IBV_MTU_4096;
+    }
+    return IBV_MTU_4096;
+}
+
 std::string msgr_rdma_address_t::to_string()
 {
     char msg[sizeof "0000:00000000:00000000:00000000000000000000000000000000"];
@@ -38,15 +63,22 @@ msgr_rdma_context_t::~msgr_rdma_context_t()
         ibv_dereg_mr(mr);
     if (pd)
         ibv_dealloc_pd(pd);
-    if (context)
+    if (context && !is_cm)
         ibv_close_device(context);
 }
 
 msgr_rdma_connection_t::~msgr_rdma_connection_t()
 {
-    ctx->used_max_cqe -= max_send+max_recv;
-    if (qp)
+    ctx->reserve_cqe(-max_send-max_recv);
+    if (qp && !cmid)
         ibv_destroy_qp(qp);
+    if (cmid)
+    {
+        ctx->cm_refs--;
+        if (cmid->qp)
+            rdma_destroy_qp(cmid);
+        rdma_destroy_id(cmid);
+    }
     if (recv_buffers.size())
     {
         for (auto b: recv_buffers)
@@ -77,21 +109,21 @@ static bool is_ipv4_gid(ibv_gid_entry *gidx)
         ((uint32_t*)gidx->gid.raw)[2] == 0xffff0000);
 }
 
-static bool match_gid(ibv_gid_entry *gidx, addr_mask_t *networks, int nnet)
+static int match_gid(ibv_gid_entry *gidx, const 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;
+        return -1;
     }
     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;
+                return i;
         }
     }
     else
@@ -99,119 +131,67 @@ static bool match_gid(ibv_gid_entry *gidx, addr_mask_t *networks, int nnet)
         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 i;
         }
     }
-    return false;
+    return -1;
 }
 
-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)
+static void log_rdma_dev_port_gid(ibv_device *dev, int ib_port, int gid_index, int mtu, 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",
+        stderr, "Selected RDMA device %s port %d GID %d - ROCEv%d IPv%d %s, MTU %d\n",
         ibv_get_device_name(dev), ib_port, gid_index,
-        gidx.gid_type == IBV_GID_TYPE_ROCE_V2 ? 2 : 1, is4 ? 4 : 6, buf
+        gidx.gid_type == IBV_GID_TYPE_ROCE_V2 ? 2 : 1, is4 ? 4 : 6, buf, mtu
     );
 }
 
-static matched_dev match_device(ibv_device **dev_list, addr_mask_t *networks, int nnet, int log_level)
+static int match_port_gid(const std::vector<addr_mask_t> & osd_network_masks, ibv_context *context,
+    int port_num, int gid_count, int log_level, ibv_gid_entry *best_gidx, int *net_num)
 {
-    matched_dev best;
-    ibv_device_attr attr;
-    ibv_port_attr portinfo;
-    ibv_gid_entry best_gidx;
-    int res;
-    bool have_non_roce = false, have_roce = false;
-    for (int i = 0; dev_list[i]; ++i)
+    // Try to find a port with matching address
+    int best_gid_idx = -1, res = 0;
+    for (int k = 0; k < gid_count; k++)
     {
-        auto dev = dev_list[i];
-        ibv_context *context = ibv_open_device(dev_list[i]);
-        if ((res = ibv_query_device(context, &attr)) != 0)
+        ibv_gid_entry gidx;
+        if ((res = ibv_query_gid_ex(context, port_num, k, &gidx, 0)) != 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)
+            if (res != ENODATA)
             {
-                fprintf(stderr, "Couldn't get RDMA device %s port %d info: %s\n", ibv_get_device_name(dev), j, strerror(res));
-                goto cleanup;
+                fprintf(stderr, "Couldn't read RDMA device %s GID index %d: %s\n", ibv_get_device_name(context->device), k, strerror(res));
+                continue;
             }
-            for (int k = 0; k < portinfo.gid_tbl_len; k++)
+            else
+                break;
+        }
+        if ((res = match_gid(&gidx, osd_network_masks.data(), osd_network_masks.size())) >= 0)
+        {
+            // Prefer RoCEv2
+            if (best_gid_idx < 0 || best_gidx->gid_type != IBV_GID_TYPE_ROCE_V2 && gidx.gid_type == IBV_GID_TYPE_ROCE_V2)
             {
-                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 (gidx.gid_type != IBV_GID_TYPE_ROCE_V1 &&
-                    gidx.gid_type != IBV_GID_TYPE_ROCE_V2)
-                    have_non_roce = true;
-                else
-                    have_roce = true;
-                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;
-                    }
-                }
+                best_gid_idx = k;
+                *best_gidx = gidx;
+                *net_num = res;
             }
         }
-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);
-    }
-    if (best.dev < 0 && have_non_roce && !have_roce)
-    {
-        best.dev = -2;
-    }
-    return best;
+    return best_gid_idx;
 }
 #endif
 
-msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_networks, const char *ib_devname, uint8_t ib_port, int gid_index, uint32_t mtu, bool odp, int log_level)
+std::vector<msgr_rdma_context_t*> msgr_rdma_context_t::create_all(const std::vector<addr_mask_t> & osd_network_masks,
+    const char *sel_dev_name, int sel_port_num, int sel_gid_index, uint32_t sel_mtu, bool odp, int log_level)
 {
     int res;
+    std::vector<msgr_rdma_context_t*> ret;
+    ibv_device **raw_dev_list = NULL;
     ibv_device **dev_list = NULL;
-    msgr_rdma_context_t *ctx = new msgr_rdma_context_t();
-    ctx->mtu = mtu;
+    ibv_device *single_list[2] = {};
 
-    timespec tv;
-    clock_gettime(CLOCK_REALTIME, &tv);
-    srand48(tv.tv_sec*1000000000 + tv.tv_nsec);
-    dev_list = ibv_get_device_list(NULL);
+    raw_dev_list = dev_list = ibv_get_device_list(NULL);
     if (!dev_list || !*dev_list)
     {
         if (errno == -ENOSYS || errno == ENOSYS)
@@ -228,121 +208,131 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
             fprintf(stderr, "Failed to get RDMA device list: %s\n", strerror(errno));
         goto cleanup;
     }
-    if (ib_devname)
+
+    if (sel_dev_name)
     {
         int i;
         for (i = 0; dev_list[i]; ++i)
-            if (!strcmp(ibv_get_device_name(dev_list[i]), ib_devname))
+            if (!strcmp(ibv_get_device_name(dev_list[i]), sel_dev_name))
                 break;
-        ctx->dev = dev_list[i];
-        if (!ctx->dev)
+        if (!dev_list[i])
         {
-            fprintf(stderr, "RDMA device %s not found\n", ib_devname);
+            fprintf(stderr, "RDMA device %s not found\n", sel_dev_name);
             goto cleanup;
         }
-    }
-#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
-    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 == -2)
-        {
-            best.dev = 0;
-            if (log_level > 0)
-                fprintf(stderr, "No RoCE devices found, using first available RDMA device %s\n", ibv_get_device_name(*dev_list));
-        }
-        else if (best.dev < 0)
-        {
-            if (log_level > 0)
-                fprintf(stderr, "RDMA device matching osd_network is not found, disabling RDMA\n");
-            goto cleanup;
-        }
-        else
-        {
-            ib_port = best.port;
-            gid_index = best.gid;
-        }
-        ctx->dev = dev_list[best.dev];
-    }
-#endif
-    else
-    {
-        ctx->dev = *dev_list;
+        single_list[0] = dev_list[i];
+        dev_list = single_list;
     }
 
-    ctx->context = ibv_open_device(ctx->dev);
-    if (!ctx->context)
+    for (int i = 0; dev_list[i]; ++i)
     {
-        fprintf(stderr, "Couldn't get RDMA context for %s\n", ibv_get_device_name(ctx->dev));
-        goto cleanup;
-    }
-
-    ctx->ib_port = ib_port;
-    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));
-        goto cleanup;
-    }
-    ctx->my_lid = ctx->portinfo.lid;
-    if (ctx->portinfo.link_layer != IBV_LINK_LAYER_ETHERNET && !ctx->my_lid)
-    {
-        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;
-    }
-
-#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
-    if (gid_index != -1)
-#endif
-    {
-        ctx->gid_index = gid_index < 0 ? 0 : gid_index;
-        if (ibv_query_gid(ctx->context, ib_port, gid_index, &ctx->my_gid))
+        auto dev = dev_list[i];
+        ibv_context *context = ibv_open_device(dev);
+        if (!context)
         {
-            fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), gid_index);
-            goto cleanup;
+            fprintf(stderr, "Couldn't get RDMA context for %s\n", ibv_get_device_name(dev));
+            continue;
         }
-    }
-#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
-    else
-    {
-        // Auto-guess GID
-        ibv_gid_entry best_gidx;
-        for (int k = 0; k < ctx->portinfo.gid_tbl_len; k++)
+        ibv_device_attr attr;
+        if ((res = ibv_query_device(context, &attr)) != 0)
         {
-            ibv_gid_entry gidx;
-            if (ibv_query_gid_ex(ctx->context, ib_port, k, &gidx, 0) != 0)
+            fprintf(stderr, "Couldn't query RDMA device %s for its features: %s\n", ibv_get_device_name(dev), strerror(res));
+            goto cleanup_dev;
+        }
+        if (sel_port_num && sel_port_num > attr.phys_port_cnt)
+        {
+            fprintf(stderr, "RDMA device %s port %d does not exist\n", ibv_get_device_name(dev), sel_port_num);
+            goto cleanup_dev;
+        }
+        for (int port_num = (sel_port_num ? sel_port_num : 1); port_num <= (sel_port_num ? sel_port_num : attr.phys_port_cnt); port_num++)
+        {
+            ibv_port_attr portinfo;
+            if ((res = ibv_query_port(context, port_num, &portinfo)) != 0)
             {
-                fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), k);
-                goto cleanup;
+                fprintf(stderr, "Couldn't get RDMA device %s port %d info: %s\n", ibv_get_device_name(dev), port_num, strerror(res));
+                continue;
             }
-            // Skip empty GID
-            if (((uint64_t*)gidx.gid.raw)[0] == 0 &&
-                ((uint64_t*)gidx.gid.raw)[1] == 0)
+            if (portinfo.state != IBV_PORT_ACTIVE)
             {
                 continue;
             }
-            // Prefer IPv4 RoCEv2 -> IPv6 RoCEv2 -> IPv4 RoCEv1 -> IPv6 RoCEv1 -> IB
-            if (gid_index == -1 ||
-                gidx.gid_type == IBV_GID_TYPE_ROCE_V2 && best_gidx.gid_type != IBV_GID_TYPE_ROCE_V2 ||
-                gidx.gid_type == IBV_GID_TYPE_ROCE_V1 && best_gidx.gid_type == IBV_GID_TYPE_IB ||
-                gidx.gid_type == best_gidx.gid_type && is_ipv4_gid(&gidx))
+            if (sel_gid_index >= (int)portinfo.gid_tbl_len)
             {
-                gid_index = k;
-                best_gidx = gidx;
+                fprintf(stderr, "RDMA device %s port %d GID %d does not exist\n", ibv_get_device_name(dev), port_num, sel_gid_index);
+                continue;
+            }
+            uint32_t port_mtu = sel_mtu ? sel_mtu : ibv_mtu_to_bytes(portinfo.active_mtu);
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
+            if (sel_gid_index < 0)
+            {
+                ibv_gid_entry best_gidx;
+                int net_num = 0;
+                int best_gid_idx = match_port_gid(osd_network_masks, context, port_num, portinfo.gid_tbl_len, log_level, &best_gidx, &net_num);
+                if (best_gid_idx >= 0)
+                {
+                    if (log_level > 0)
+                        log_rdma_dev_port_gid(dev, port_num, best_gid_idx, port_mtu, best_gidx);
+                    auto ctx = msgr_rdma_context_t::create(dev, portinfo, port_num, best_gid_idx, port_mtu, odp, log_level);
+                    if (ctx)
+                    {
+                        ctx->net_mask = osd_network_masks[net_num];
+                        ret.push_back(ctx);
+                    }
+                }
+            }
+            else
+#endif
+            {
+                int best_gid_idx = sel_gid_index >= 0 ? sel_gid_index : 0;
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
+                if (log_level > 0)
+                {
+                    ibv_gid_entry gidx;
+                    ibv_query_gid_ex(context, port_num, best_gid_idx, &gidx, 0);
+                    log_rdma_dev_port_gid(dev, port_num, best_gid_idx, port_mtu, gidx);
+                }
+#endif
+                auto ctx = msgr_rdma_context_t::create(dev, portinfo, port_num, best_gid_idx, port_mtu, odp, log_level);
+                if (ctx)
+                    ret.push_back(ctx);
             }
         }
-        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, best_gidx);
-        }
-        ctx->my_gid = best_gidx.gid;
+cleanup_dev:
+        ibv_close_device(context);
+    }
+
+cleanup:
+    if (raw_dev_list)
+        ibv_free_device_list(raw_dev_list);
+    return ret;
+}
+
+msgr_rdma_context_t *msgr_rdma_context_t::create(ibv_device *dev, ibv_port_attr & portinfo, int ib_port, int gid_index, uint32_t mtu, bool odp, int log_level)
+{
+    msgr_rdma_context_t *ctx = new msgr_rdma_context_t();
+    ibv_context *context = ibv_open_device(dev);
+    if (!context)
+    {
+        fprintf(stderr, "Couldn't get RDMA context for %s\n", ibv_get_device_name(dev));
+        goto cleanup;
+    }
+
+    ctx->mtu = mtu;
+    ctx->context = context;
+    ctx->ib_port = ib_port;
+    ctx->portinfo = portinfo;
+    ctx->my_lid = ctx->portinfo.lid;
+    if (ctx->portinfo.link_layer != IBV_LINK_LAYER_ETHERNET && !ctx->my_lid)
+    {
+        fprintf(stderr, "RDMA device %s must have local LID because it's not Ethernet, but LID is zero\n", ibv_get_device_name(dev));
+        goto cleanup;
+    }
+    ctx->gid_index = gid_index;
+    if (ibv_query_gid(ctx->context, ib_port, gid_index, &ctx->my_gid))
+    {
+        fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(dev), gid_index);
+        goto cleanup;
     }
-#endif
 
     ctx->pd = ibv_alloc_pd(ctx->context);
     if (!ctx->pd)
@@ -351,18 +341,19 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
         goto cleanup;
     }
 
+    if (ibv_query_device_ex(ctx->context, NULL, &ctx->attrx))
     {
-        if (ibv_query_device_ex(ctx->context, NULL, &ctx->attrx))
-        {
-            fprintf(stderr, "Couldn't query RDMA device for its features\n");
-            goto cleanup;
-        }
-        ctx->odp = odp;
-        if (ctx->odp &&
-            (!(ctx->attrx.odp_caps.general_caps & IBV_ODP_SUPPORT) ||
+        fprintf(stderr, "Couldn't query RDMA device for its features\n");
+        goto cleanup;
+    }
+
+    ctx->odp = odp;
+    if (ctx->odp)
+    {
+        if (!(ctx->attrx.odp_caps.general_caps & IBV_ODP_SUPPORT) ||
             !(ctx->attrx.odp_caps.general_caps & IBV_ODP_SUPPORT_IMPLICIT) ||
             !(ctx->attrx.odp_caps.per_transport_caps.rc_odp_caps & IBV_ODP_SUPPORT_SEND) ||
-            !(ctx->attrx.odp_caps.per_transport_caps.rc_odp_caps & IBV_ODP_SUPPORT_RECV)))
+            !(ctx->attrx.odp_caps.per_transport_caps.rc_odp_caps & IBV_ODP_SUPPORT_RECV))
         {
             ctx->odp = false;
             if (log_level > 0)
@@ -395,20 +386,43 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
         goto cleanup;
     }
 
-    if (dev_list)
-        ibv_free_device_list(dev_list);
     return ctx;
 
 cleanup:
+    if (context)
+        ibv_close_device(context);
     delete ctx;
-    if (dev_list)
-        ibv_free_device_list(dev_list);
     return NULL;
 }
 
+bool msgr_rdma_context_t::reserve_cqe(int n)
+{
+    this->used_max_cqe += n;
+    if (this->used_max_cqe > this->max_cqe)
+    {
+        // Resize CQ
+        // Mellanox ConnectX-4 supports up to 4194303 CQEs, so it's fine to put everything into a single CQ
+        int new_max_cqe = this->max_cqe;
+        while (this->used_max_cqe > new_max_cqe)
+        {
+            new_max_cqe *= 2;
+        }
+        if (ibv_resize_cq(this->cq, new_max_cqe) != 0)
+        {
+            fprintf(stderr, "Couldn't resize RDMA completion queue to %d entries\n", new_max_cqe);
+            return false;
+        }
+        this->max_cqe = new_max_cqe;
+    }
+    return true;
+}
+
 msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx, uint32_t max_send,
     uint32_t max_recv, uint32_t max_sge, uint32_t max_msg)
 {
+    if (!ctx->reserve_cqe(max_send+max_recv))
+        return NULL;
+
     msgr_rdma_connection_t *conn = new msgr_rdma_connection_t;
 
     max_sge = max_sge > ctx->attrx.orig_attr.max_sge ? ctx->attrx.orig_attr.max_sge : max_sge;
@@ -419,25 +433,6 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
     conn->max_sge = max_sge;
     conn->max_msg = max_msg;
 
-    ctx->used_max_cqe += max_send+max_recv;
-    if (ctx->used_max_cqe > ctx->max_cqe)
-    {
-        // Resize CQ
-        // Mellanox ConnectX-4 supports up to 4194303 CQEs, so it's fine to put everything into a single CQ
-        int new_max_cqe = ctx->max_cqe;
-        while (ctx->used_max_cqe > new_max_cqe)
-        {
-            new_max_cqe *= 2;
-        }
-        if (ibv_resize_cq(ctx->cq, new_max_cqe) != 0)
-        {
-            fprintf(stderr, "Couldn't resize RDMA completion queue to %d entries\n", new_max_cqe);
-            delete conn;
-            return NULL;
-        }
-        ctx->max_cqe = new_max_cqe;
-    }
-
     ibv_qp_init_attr init_attr = {
         .send_cq = ctx->cq,
         .recv_cq = ctx->cq,
@@ -469,9 +464,10 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
         .port_num        = ctx->ib_port,
     };
 
-    if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT | IBV_QP_ACCESS_FLAGS))
+    int r = 0;
+    if ((r = ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT | IBV_QP_ACCESS_FLAGS)) != 0)
     {
-        fprintf(stderr, "Failed to switch RDMA queue pair to INIT state\n");
+        fprintf(stderr, "Failed to switch RDMA queue pair to INIT state: %s (code %d)\n", strerror(r), r);
         delete conn;
         return NULL;
     }
@@ -479,25 +475,12 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
     return conn;
 }
 
-static ibv_mtu mtu_to_ibv_mtu(uint32_t mtu)
-{
-    switch (mtu)
-    {
-        case 256:  return IBV_MTU_256;
-        case 512:  return IBV_MTU_512;
-        case 1024: return IBV_MTU_1024;
-        case 2048: return IBV_MTU_2048;
-        case 4096: return IBV_MTU_4096;
-    }
-    return IBV_MTU_4096;
-}
-
 int msgr_rdma_connection_t::connect(msgr_rdma_address_t *dest)
 {
     auto conn = this;
     ibv_qp_attr attr = {
         .qp_state       = IBV_QPS_RTR,
-        .path_mtu       = mtu_to_ibv_mtu(conn->ctx->mtu),
+        .path_mtu       = bytes_to_ibv_mtu(conn->ctx->mtu),
         .rq_psn         = dest->psn,
         .sq_psn         = conn->addr.psn,
         .dest_qp_num    = dest->qpn,
@@ -522,18 +505,19 @@ int msgr_rdma_connection_t::connect(msgr_rdma_address_t *dest)
         .rnr_retry      = 7,
     };
     // FIXME No idea if ibv_modify_qp is a blocking operation or not. No idea if it has a timeout and what it is.
-    if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU |
-        IBV_QP_DEST_QPN | IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER))
+    int r;
+    if ((r = ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU |
+        IBV_QP_DEST_QPN | IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER)) != 0)
     {
-        fprintf(stderr, "Failed to switch RDMA queue pair to RTR (ready-to-receive) state\n");
-        return 1;
+        fprintf(stderr, "Failed to switch RDMA queue pair to RTR (ready-to-receive) state: %s (code %d)\n", strerror(r), r);
+        return -r;
     }
     attr.qp_state = IBV_QPS_RTS;
-    if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_TIMEOUT |
-        IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | IBV_QP_SQ_PSN | IBV_QP_MAX_QP_RD_ATOMIC))
+    if ((r = ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_TIMEOUT |
+        IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | IBV_QP_SQ_PSN | IBV_QP_MAX_QP_RD_ATOMIC)) != 0)
     {
-        fprintf(stderr, "Failed to switch RDMA queue pair to RTS (ready-to-send) state\n");
-        return 1;
+        fprintf(stderr, "Failed to switch RDMA queue pair to RTS (ready-to-send) state: %s (code %d)\n", strerror(r), r);
+        return -r;
     }
     return 0;
 }
@@ -548,7 +532,15 @@ bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address, uint64
         {
             client_max_msg = rdma_max_msg;
         }
-        auto rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, client_max_msg);
+        auto cl = clients.at(peer_fd);
+        msgr_rdma_context_t *selected_ctx = choose_rdma_context(cl);
+        if (!selected_ctx)
+        {
+            if (log_level > 0)
+                fprintf(stderr, "No RDMA context for peer %d, using only TCP\n", cl->peer_fd);
+            return false;
+        }
+        msgr_rdma_connection_t *rdma_conn = msgr_rdma_connection_t::create(selected_ctx, rdma_max_send, rdma_max_recv, rdma_max_sge, client_max_msg);
         if (rdma_conn)
         {
             int r = rdma_conn->connect(&addr);
@@ -586,7 +578,7 @@ static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
     int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
     if (err || bad_wr)
     {
-        fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
+        fprintf(stderr, "RDMA send failed: %s (code %d)\n", strerror(err), err);
         exit(1);
     }
     cl->rdma_conn->cur_send++;
@@ -667,9 +659,9 @@ void osd_messenger_t::try_send_rdma_nodp(osd_client_t *cl)
         // Allocate send ring buffer, if not yet
         rc->send_out_size = rc->max_msg*rdma_max_send;
         rc->send_out.buf = malloc_or_die(rc->send_out_size);
-        if (!rdma_context->odp)
+        if (!rc->ctx->odp)
         {
-            rc->send_out.mr = ibv_reg_mr(rdma_context->pd, rc->send_out.buf, rc->send_out_size, 0);
+            rc->send_out.mr = ibv_reg_mr(rc->ctx->pd, rc->send_out.buf, rc->send_out_size, 0);
             if (!rc->send_out.mr)
             {
                 fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno));
@@ -699,7 +691,7 @@ void osd_messenger_t::try_send_rdma_nodp(osd_client_t *cl)
             ibv_sge sge = {
                 .addr = (uintptr_t)dst,
                 .length = (uint32_t)copied,
-                .lkey = rdma_context->odp ? rdma_context->mr->lkey : rc->send_out.mr->lkey,
+                .lkey = rc->ctx->odp ? rc->ctx->mr->lkey : rc->send_out.mr->lkey,
             };
             try_send_rdma_wr(cl, &sge, 1);
             rc->send_sizes.push_back(copied);
@@ -709,7 +701,7 @@ void osd_messenger_t::try_send_rdma_nodp(osd_client_t *cl)
 
 void osd_messenger_t::try_send_rdma(osd_client_t *cl)
 {
-    if (rdma_context->odp)
+    if (cl->rdma_conn->ctx->odp)
         try_send_rdma_odp(cl);
     else
         try_send_rdma_nodp(cl);
@@ -744,9 +736,9 @@ bool osd_messenger_t::try_recv_rdma(osd_client_t *cl)
     {
         msgr_rdma_buf_t b;
         b.buf = malloc_or_die(rc->max_msg);
-        if (!rdma_context->odp)
+        if (!rc->ctx->odp)
         {
-            b.mr = ibv_reg_mr(rdma_context->pd, b.buf, rc->max_msg, IBV_ACCESS_LOCAL_WRITE);
+            b.mr = ibv_reg_mr(rc->ctx->pd, b.buf, rc->max_msg, IBV_ACCESS_LOCAL_WRITE);
             if (!b.mr)
             {
                 fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno));
@@ -761,7 +753,7 @@ bool osd_messenger_t::try_recv_rdma(osd_client_t *cl)
 
 #define RDMA_EVENTS_AT_ONCE 32
 
-void osd_messenger_t::handle_rdma_events()
+void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
 {
     // Request next notification
     ibv_cq *ev_cq;

src/client/msgr_rdma.h

@@ -2,9 +2,13 @@
 // License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
 
 #pragma once
+#ifdef WITH_RDMACM
+#include <rdma/rdma_cma.h>
+#endif
 #include <infiniband/verbs.h>
 #include <string>
 #include <vector>
+#include "addr_util.h"
 
 struct msgr_rdma_address_t
 {
@@ -20,7 +24,6 @@ struct msgr_rdma_address_t
 struct msgr_rdma_context_t
 {
     ibv_context *context = NULL;
-    ibv_device *dev = NULL;
     ibv_device_attr_ex attrx;
     ibv_pd *pd = NULL;
     bool odp = false;
@@ -35,8 +38,17 @@ struct msgr_rdma_context_t
     uint32_t mtu;
     int max_cqe = 0;
     int used_max_cqe = 0;
+    addr_mask_t net_mask = {};
+    bool is_cm = false;
+    int cm_refs = 0;
+
+    static std::vector<msgr_rdma_context_t*> create_all(const std::vector<addr_mask_t> & osd_network_masks,
+        const char *sel_dev_name, int sel_port_num, int sel_gid_index, uint32_t sel_mtu, bool odp, int log_level);
+    static msgr_rdma_context_t *create(ibv_device *dev, ibv_port_attr & portinfo,
+        int ib_port, int gid_index, uint32_t mtu, bool odp, int log_level);
+    static msgr_rdma_context_t* create_cm(ibv_context *ctx);
+    bool reserve_cqe(int n);
 
-    static msgr_rdma_context_t *create(std::vector<std::string> osd_networks, const char *ib_devname, uint8_t ib_port, int gid_index, uint32_t mtu, bool odp, int log_level);
     ~msgr_rdma_context_t();
 };
 
@@ -50,11 +62,14 @@ struct msgr_rdma_connection_t
 {
     msgr_rdma_context_t *ctx = NULL;
     ibv_qp *qp = NULL;
+#ifdef WITH_RDMACM
+    rdma_cm_id *cmid = NULL;
+#endif
     msgr_rdma_address_t addr;
     int max_send = 0, max_recv = 0, max_sge = 0;
-    int cur_send = 0, cur_recv = 0;
     uint64_t max_msg = 0;
 
+    int cur_send = 0, cur_recv = 0;
     int send_pos = 0, send_buf_pos = 0;
     int next_recv_buf = 0;
     std::vector<msgr_rdma_buf_t> recv_buffers;

src/client/msgr_rdmacm.cpp

@@ -0,0 +1,526 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include "msgr_rdma.h"
+#include "messenger.h"
+
+struct rdmacm_connecting_t
+{
+    rdma_cm_id *cmid = NULL;
+    int peer_fd = -1;
+    osd_num_t peer_osd = 0;
+    std::string addr;
+    sockaddr_storage parsed_addr = {};
+    int rdmacm_port = 0;
+    int tcp_port = 0;
+    int timeout_ms = 0;
+    int timeout_id = -1;
+    msgr_rdma_context_t *rdma_context = NULL;
+};
+
+rdma_cm_id *osd_messenger_t::rdmacm_listen(const std::string & bind_address, int rdmacm_port, int *bound_port, int log_level)
+{
+    sockaddr_storage addr = {};
+    rdma_cm_id *listener = NULL;
+    int r = rdma_create_id(rdmacm_evch, &listener, NULL, RDMA_PS_TCP);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to create RDMA-CM ID: %s (code %d)\n", strerror(errno), errno);
+        goto fail;
+    }
+    if (!string_to_addr(bind_address, 0, rdmacm_port, &addr))
+    {
+        fprintf(stderr, "Server address: %s is not valid\n", bind_address.c_str());
+        goto fail;
+    }
+    r = rdma_bind_addr(listener, (sockaddr*)&addr);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to bind RDMA-CM to %s:%d: %s (code %d)\n", bind_address.c_str(), rdmacm_port, strerror(errno), errno);
+        goto fail;
+    }
+    r = rdma_listen(listener, 128);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to listen to RDMA-CM address %s:%d: %s (code %d)\n", bind_address.c_str(), rdmacm_port, strerror(errno), errno);
+        goto fail;
+    }
+    if (bound_port)
+    {
+        *bound_port = ntohs(rdma_get_src_port(listener));
+    }
+    if (log_level > 0)
+    {
+        fprintf(stderr, "Listening to RDMA-CM address %s port %d\n", bind_address.c_str(), *bound_port);
+    }
+    return listener;
+fail:
+    rdma_destroy_id(listener);
+    return NULL;
+}
+
+void osd_messenger_t::rdmacm_destroy_listener(rdma_cm_id *listener)
+{
+    rdma_destroy_id(listener);
+}
+
+void osd_messenger_t::handle_rdmacm_events()
+{
+    // rdma_destroy_id infinitely waits for pthread_cond if called before all events are acked :-(
+    std::vector<rdma_cm_event> events_copy;
+    while (1)
+    {
+        rdma_cm_event *ev = NULL;
+        int r = rdma_get_cm_event(rdmacm_evch, &ev);
+        if (r != 0)
+        {
+            if (errno == EAGAIN || errno == EINTR)
+                break;
+            fprintf(stderr, "Failed to get RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
+            exit(1);
+        }
+        events_copy.push_back(*ev);
+        r = rdma_ack_cm_event(ev);
+        if (r != 0)
+        {
+            fprintf(stderr, "Failed to ack (free) RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
+            exit(1);
+        }
+    }
+    for (auto & evl: events_copy)
+    {
+        auto ev = &evl;
+        if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
+        {
+            rdmacm_accept(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR ||
+            ev->event == RDMA_CM_EVENT_REJECTED ||
+            ev->event == RDMA_CM_EVENT_DISCONNECTED ||
+            ev->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
+        {
+            auto event_type_name = ev->event == RDMA_CM_EVENT_CONNECT_ERROR ? "RDMA_CM_EVENT_CONNECT_ERROR" : (
+                ev->event == RDMA_CM_EVENT_REJECTED ? "RDMA_CM_EVENT_REJECTED" : (
+                ev->event == RDMA_CM_EVENT_DISCONNECTED ? "RDMA_CM_EVENT_DISCONNECTED" : "RDMA_CM_EVENT_DEVICE_REMOVAL"));
+            auto cli_it = rdmacm_connections.find(ev->id);
+            if (cli_it != rdmacm_connections.end())
+            {
+                fprintf(stderr, "Received %s event for peer %d, closing connection\n",
+                    event_type_name, cli_it->second->peer_fd);
+                stop_client(cli_it->second->peer_fd);
+            }
+            else if (rdmacm_connecting.find(ev->id) != rdmacm_connecting.end())
+            {
+                fprintf(stderr, "Received %s event for RDMA-CM OSD %ju connection\n",
+                    event_type_name, rdmacm_connecting[ev->id]->peer_osd);
+                rdmacm_established(ev);
+            }
+            else
+            {
+                fprintf(stderr, "Received %s event for an unknown RDMA-CM connection 0x%jx - ignoring\n",
+                    event_type_name, (uint64_t)ev->id);
+            }
+        }
+        else if (ev->event == RDMA_CM_EVENT_ADDR_RESOLVED || ev->event == RDMA_CM_EVENT_ADDR_ERROR)
+        {
+            rdmacm_address_resolved(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_ROUTE_RESOLVED || ev->event == RDMA_CM_EVENT_ROUTE_ERROR)
+        {
+            rdmacm_route_resolved(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_CONNECT_RESPONSE)
+        {
+            // Just OK
+        }
+        else if (ev->event == RDMA_CM_EVENT_UNREACHABLE || ev->event == RDMA_CM_EVENT_REJECTED)
+        {
+            // Handle error
+            rdmacm_established(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_ESTABLISHED)
+        {
+            rdmacm_established(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_ADDR_CHANGE || ev->event == RDMA_CM_EVENT_TIMEWAIT_EXIT)
+        {
+            // Do nothing
+        }
+        else
+        {
+            // Other events are unexpected
+            fprintf(stderr, "Unexpected RDMA-CM event type: %d\n", ev->event);
+        }
+    }
+}
+
+msgr_rdma_context_t* msgr_rdma_context_t::create_cm(ibv_context *ctx)
+{
+    auto rdma_context = new msgr_rdma_context_t;
+    rdma_context->is_cm = true;
+    rdma_context->context = ctx;
+    rdma_context->pd = ibv_alloc_pd(ctx);
+    if (!rdma_context->pd)
+    {
+        fprintf(stderr, "Couldn't allocate RDMA protection domain\n");
+        delete rdma_context;
+        return NULL;
+    }
+    rdma_context->odp = false;
+    rdma_context->channel = ibv_create_comp_channel(rdma_context->context);
+    if (!rdma_context->channel)
+    {
+        fprintf(stderr, "Couldn't create RDMA completion channel\n");
+        delete rdma_context;
+        return NULL;
+    }
+    rdma_context->max_cqe = 4096;
+    rdma_context->cq = ibv_create_cq(rdma_context->context, rdma_context->max_cqe, NULL, rdma_context->channel, 0);
+    if (!rdma_context->cq)
+    {
+        fprintf(stderr, "Couldn't create RDMA completion queue\n");
+        delete rdma_context;
+        return NULL;
+    }
+    if (ibv_query_device_ex(rdma_context->context, NULL, &rdma_context->attrx))
+    {
+        fprintf(stderr, "Couldn't query RDMA device for its features\n");
+        delete rdma_context;
+        return NULL;
+    }
+    return rdma_context;
+}
+
+msgr_rdma_context_t* osd_messenger_t::rdmacm_get_context(ibv_context *verbs)
+{
+    // Find the context by device
+    // We assume that RDMA_CM ev->id->verbs is always the same for the same device (but PD for example isn't)
+    msgr_rdma_context_t *rdma_context = NULL;
+    for (auto ctx: rdma_contexts)
+    {
+        if (ctx->context == verbs)
+        {
+            rdma_context = ctx;
+            break;
+        }
+    }
+    if (!rdma_context)
+    {
+        // Wrap into a new msgr_rdma_context_t
+        rdma_context = msgr_rdma_context_t::create_cm(verbs);
+        if (!rdma_context)
+            return NULL;
+        fcntl(rdma_context->channel->fd, F_SETFL, fcntl(rdma_context->channel->fd, F_GETFL, 0) | O_NONBLOCK);
+        tfd->set_fd_handler(rdma_context->channel->fd, false, [this, rdma_context](int notify_fd, int epoll_events)
+        {
+            handle_rdma_events(rdma_context);
+        });
+        handle_rdma_events(rdma_context);
+        rdma_contexts.push_back(rdma_context);
+    }
+    return rdma_context;
+}
+
+msgr_rdma_context_t* osd_messenger_t::rdmacm_create_qp(rdma_cm_id *cmid)
+{
+    auto rdma_context = rdmacm_get_context(cmid->verbs);
+    if (!rdma_context)
+    {
+        return NULL;
+    }
+    rdma_context->reserve_cqe(rdma_max_send+rdma_max_recv);
+    auto max_sge = rdma_max_sge > rdma_context->attrx.orig_attr.max_sge
+        ? rdma_context->attrx.orig_attr.max_sge : rdma_max_sge;
+    ibv_qp_init_attr init_attr = {
+        .send_cq = rdma_context->cq,
+        .recv_cq = rdma_context->cq,
+        .cap     = {
+            .max_send_wr  = (uint32_t)rdma_max_send,
+            .max_recv_wr  = (uint32_t)rdma_max_recv,
+            .max_send_sge = (uint32_t)max_sge,
+            .max_recv_sge = (uint32_t)max_sge,
+        },
+        .qp_type = IBV_QPT_RC,
+    };
+    int r = rdma_create_qp(cmid, rdma_context->pd, &init_attr);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to create a queue pair via RDMA-CM: %s (code %d)\n", strerror(errno), errno);
+        rdma_context->reserve_cqe(-rdma_max_send-rdma_max_recv);
+        return NULL;
+    }
+    return rdma_context;
+}
+
+void osd_messenger_t::rdmacm_accept(rdma_cm_event *ev)
+{
+    // Make a fake FD (FIXME: do not use FDs for identifying clients!)
+    int fake_fd = socket(AF_INET, SOCK_STREAM, 0);
+    if (fake_fd < 0)
+    {
+        fprintf(stderr, "Failed to allocate a fake socket for RDMA-CM client: %s (code %d)\n", strerror(errno), errno);
+        rdma_destroy_id(ev->id);
+        return;
+    }
+    auto rdma_context = rdmacm_create_qp(ev->id);
+    if (!rdma_context)
+    {
+        rdma_destroy_id(ev->id);
+        return;
+    }
+    // We don't need private_data, RDMA_READ or ATOMIC so use default 1
+    rdma_conn_param conn_params = {
+        .responder_resources = 1,
+        .initiator_depth = 1,
+        .retry_count = 7,
+        .rnr_retry_count = 7,
+    };
+    if (rdma_accept(ev->id, &conn_params) != 0)
+    {
+        fprintf(stderr, "Failed to accept RDMA-CM connection: %s (code %d)\n", strerror(errno), errno);
+        rdma_context->reserve_cqe(-rdma_max_send-rdma_max_recv);
+        rdma_destroy_qp(ev->id);
+        rdma_destroy_id(ev->id);
+        return;
+    }
+    rdma_context->cm_refs++;
+    // Wrap into a new msgr_rdma_connection_t
+    msgr_rdma_connection_t *conn = new msgr_rdma_connection_t;
+    conn->ctx = rdma_context;
+    conn->max_send = rdma_max_send;
+    conn->max_recv = rdma_max_recv;
+    conn->max_sge = rdma_max_sge > rdma_context->attrx.orig_attr.max_sge
+        ? rdma_context->attrx.orig_attr.max_sge : rdma_max_sge;
+    conn->max_msg = rdma_max_msg;
+    conn->cmid = ev->id;
+    conn->qp = ev->id->qp;
+    auto cl = new osd_client_t();
+    cl->peer_fd = fake_fd;
+    cl->peer_state = PEER_RDMA;
+    cl->peer_addr = *(sockaddr_storage*)rdma_get_peer_addr(ev->id);
+    cl->in_buf = malloc_or_die(receive_buffer_size);
+    cl->rdma_conn = conn;
+    clients[fake_fd] = cl;
+    rdmacm_connections[ev->id] = cl;
+    // Add initial receive request(s)
+    try_recv_rdma(cl);
+    fprintf(stderr, "[OSD %ju] new client %d: connection from %s via RDMA-CM\n", this->osd_num, fake_fd,
+        addr_to_string(cl->peer_addr).c_str());
+}
+
+void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res)
+{
+    auto conn = rdmacm_connecting.at(cmid);
+    auto addr = conn->addr;
+    auto tcp_port = conn->tcp_port;
+    auto peer_osd = conn->peer_osd;
+    if (conn->timeout_id >= 0)
+        tfd->clear_timer(conn->timeout_id);
+    if (conn->peer_fd >= 0)
+        close(conn->peer_fd);
+    if (conn->rdma_context)
+        conn->rdma_context->reserve_cqe(-rdma_max_send-rdma_max_recv);
+    if (conn->cmid)
+    {
+        if (conn->cmid->qp)
+            rdma_destroy_qp(conn->cmid);
+        rdma_destroy_id(conn->cmid);
+    }
+    rdmacm_connecting.erase(cmid);
+    delete conn;
+    if (!disable_tcp)
+    {
+        // Fall back to TCP instead of just reporting the error to on_connect_peer()
+        try_connect_peer_tcp(peer_osd, addr.c_str(), tcp_port);
+    }
+    else
+    {
+        // TCP is disabled
+        on_connect_peer(peer_osd, res == 0 ? -EINVAL : (res > 0 ? -res : res));
+    }
+}
+
+void osd_messenger_t::rdmacm_try_connect_peer(uint64_t peer_osd, const std::string & addr, int rdmacm_port, int fallback_tcp_port)
+{
+    struct sockaddr_storage sa = {};
+    if (!string_to_addr(addr, false, rdmacm_port, &sa))
+    {
+        fprintf(stderr, "Address %s is invalid\n", addr.c_str());
+        on_connect_peer(peer_osd, -EINVAL);
+        return;
+    }
+    rdma_cm_id *cmid = NULL;
+    if (rdma_create_id(rdmacm_evch, &cmid, NULL, RDMA_PS_TCP) != 0)
+    {
+        int res = -errno;
+        fprintf(stderr, "Failed to create RDMA-CM ID: %s (code %d), using TCP\n", strerror(errno), errno);
+        if (!disable_tcp)
+            try_connect_peer_tcp(peer_osd, addr.c_str(), fallback_tcp_port);
+        else
+            on_connect_peer(peer_osd, res);
+        return;
+    }
+    // Make a fake FD (FIXME: do not use FDs for identifying clients!)
+    int fake_fd = socket(AF_INET, SOCK_STREAM, 0);
+    if (fake_fd < 0)
+    {
+        int res = -errno;
+        rdma_destroy_id(cmid);
+        // Can't create socket, pointless to try TCP
+        on_connect_peer(peer_osd, res);
+        return;
+    }
+    auto conn = new rdmacm_connecting_t;
+    rdmacm_connecting[cmid] = conn;
+    conn->cmid = cmid;
+    conn->peer_fd = fake_fd;
+    conn->peer_osd = peer_osd;
+    conn->addr = addr;
+    conn->parsed_addr = sa;
+    conn->rdmacm_port = rdmacm_port;
+    conn->tcp_port = fallback_tcp_port;
+    conn->timeout_ms = peer_connect_timeout*1000;
+    conn->timeout_id = -1;
+    if (peer_connect_timeout > 0)
+    {
+        conn->timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, cmid](int timer_id)
+        {
+            auto conn = rdmacm_connecting.at(cmid);
+            conn->timeout_id = -1;
+            fprintf(stderr, "RDMA-CM connection to %s timed out\n", conn->addr.c_str());
+            rdmacm_on_connect_peer_error(cmid, -EPIPE);
+            return;
+        });
+    }
+    if (rdma_resolve_addr(cmid, NULL, (sockaddr*)&conn->parsed_addr, conn->timeout_ms) != 0)
+    {
+        auto res = -errno;
+        // ENODEV means that the client doesn't have an RDMA device for this address
+        if (res != -ENODEV || log_level > 0)
+            fprintf(stderr, "Failed to resolve address %s via RDMA-CM: %s (code %d)\n", addr.c_str(), strerror(errno), errno);
+        rdmacm_on_connect_peer_error(cmid, res);
+        return;
+    }
+}
+
+void osd_messenger_t::rdmacm_address_resolved(rdma_cm_event *ev)
+{
+    auto cmid = ev->id;
+    auto conn_it = rdmacm_connecting.find(cmid);
+    if (conn_it == rdmacm_connecting.end())
+    {
+        // Silently ignore unknown IDs
+        return;
+    }
+    auto conn = conn_it->second;
+    if (ev->event != RDMA_CM_EVENT_ADDR_RESOLVED || ev->status != 0)
+    {
+        fprintf(stderr, "Failed to resolve address %s via RDMA-CM: %s (code %d)\n", conn->addr.c_str(),
+            ev->status > 0 ? "unknown error" : strerror(-ev->status), ev->status);
+        rdmacm_on_connect_peer_error(cmid, ev->status);
+        return;
+    }
+    auto rdma_context = rdmacm_create_qp(cmid);
+    if (!rdma_context)
+    {
+        rdmacm_on_connect_peer_error(cmid, -EIO);
+        return;
+    }
+    conn->rdma_context = rdma_context;
+    if (rdma_resolve_route(cmid, conn->timeout_ms) != 0)
+    {
+        int res = -errno;
+        fprintf(stderr, "Failed to resolve route to %s via RDMA-CM: %s (code %d)\n", conn->addr.c_str(), strerror(errno), errno);
+        rdmacm_on_connect_peer_error(cmid, res);
+        return;
+    }
+}
+
+void osd_messenger_t::rdmacm_route_resolved(rdma_cm_event *ev)
+{
+    auto cmid = ev->id;
+    auto conn_it = rdmacm_connecting.find(cmid);
+    if (conn_it == rdmacm_connecting.end())
+    {
+        // Silently ignore unknown IDs
+        return;
+    }
+    auto conn = conn_it->second;
+    if (ev->event != RDMA_CM_EVENT_ROUTE_RESOLVED || ev->status != 0)
+    {
+        fprintf(stderr, "Failed to resolve route to %s via RDMA-CM: %s (code %d)\n", conn->addr.c_str(),
+            ev->status > 0 ? "unknown error" : strerror(-ev->status), ev->status);
+        rdmacm_on_connect_peer_error(cmid, ev->status);
+        return;
+    }
+    // We don't need private_data, RDMA_READ or ATOMIC so use default 1
+    rdma_conn_param conn_params = {
+        .responder_resources = 1,
+        .initiator_depth = 1,
+        .retry_count = 7,
+        .rnr_retry_count = 7,
+    };
+    if (rdma_connect(cmid, &conn_params) != 0)
+    {
+        int res = -errno;
+        fprintf(stderr, "Failed to connect to %s:%d via RDMA-CM: %s (code %d)\n", conn->addr.c_str(), conn->rdmacm_port, strerror(errno), errno);
+        rdmacm_on_connect_peer_error(cmid, res);
+        return;
+    }
+}
+
+void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
+{
+    auto cmid = ev->id;
+    auto conn_it = rdmacm_connecting.find(cmid);
+    if (conn_it == rdmacm_connecting.end())
+    {
+        // Silently ignore unknown IDs
+        return;
+    }
+    auto conn = conn_it->second;
+    auto peer_osd = conn->peer_osd;
+    if (ev->event != RDMA_CM_EVENT_ESTABLISHED || ev->status != 0)
+    {
+        fprintf(stderr, "Failed to connect to %s:%d via RDMA-CM: %s (code %d)\n", conn->addr.c_str(), conn->rdmacm_port,
+            ev->status > 0 ? "unknown error" : strerror(-ev->status), ev->status);
+        rdmacm_on_connect_peer_error(cmid, ev->status);
+        return;
+    }
+    // Wrap into a new msgr_rdma_connection_t
+    msgr_rdma_connection_t *rc = new msgr_rdma_connection_t;
+    rc->ctx = conn->rdma_context;
+    rc->ctx->cm_refs++;
+    rc->max_send = rdma_max_send;
+    rc->max_recv = rdma_max_recv;
+    rc->max_sge = rdma_max_sge > rc->ctx->attrx.orig_attr.max_sge
+        ? rc->ctx->attrx.orig_attr.max_sge : rdma_max_sge;
+    rc->max_msg = rdma_max_msg;
+    rc->cmid = conn->cmid;
+    rc->qp = conn->cmid->qp;
+    // And an osd_client_t
+    auto cl = new osd_client_t();
+    cl->peer_addr = conn->parsed_addr;
+    cl->peer_port = conn->rdmacm_port;
+    cl->peer_fd = conn->peer_fd;
+    cl->peer_state = PEER_RDMA;
+    cl->connect_timeout_id = -1;
+    cl->osd_num = peer_osd;
+    cl->in_buf = malloc_or_die(receive_buffer_size);
+    cl->rdma_conn = rc;
+    clients[conn->peer_fd] = cl;
+    if (conn->timeout_id >= 0)
+        tfd->clear_timer(conn->timeout_id);
+    delete conn;
+    rdmacm_connecting.erase(cmid);
+    rdmacm_connections[cmid] = cl;
+    if (log_level > 0)
+        fprintf(stderr, "Successfully connected with OSD %ju using RDMA-CM\n", peer_osd);
+    // Add initial receive request(s)
+    try_recv_rdma(cl);
+    check_peer_config(cl);
+}

src/client/msgr_send.cpp

@@ -187,6 +187,7 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
     {
         return true;
     }
+    assert(cl->peer_state != PEER_RDMA);
     if (ringloop && !use_sync_send_recv)
     {
         auto iothread = iothreads.size() ? iothreads[peer_fd % iothreads.size()] : NULL;

src/client/nbd_proxy.cpp

@@ -1194,7 +1194,7 @@ protected:
         }
         else
         {
-            if (cur_op->opcode == OSD_OP_WRITE && watch->cfg.readonly)
+            if (cur_op->opcode == OSD_OP_WRITE && !inode && watch->cfg.readonly)
             {
                 cur_op->retval = -EROFS;
                 std::function<void(cluster_op_t*)>(cur_op->callback)(cur_op);

src/client/qemu_driver.c

@@ -294,7 +294,9 @@ static void coroutine_fn vitastor_co_get_metadata(VitastorRPC *task)
 
     qemu_mutex_lock(&client->mutex);
     vitastor_c_watch_inode(client->proxy, client->image, vitastor_co_generic_cb, task);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_schedule_uring_handler(client);
+#endif
     qemu_mutex_unlock(&client->mutex);
 
     while (!task->complete)
@@ -566,6 +568,22 @@ static int vitastor_file_open(BlockDriverState *bs, QDict *options, int flags, E
 static void vitastor_close(BlockDriverState *bs)
 {
     VitastorClient *client = bs->opaque;
+    if (client->uring_eventfd >= 0)
+    {
+        // clear the eventfd handler
+        universal_aio_set_fd_handler(client->ctx, client->uring_eventfd, NULL, NULL, NULL);
+        int wait_bh = 0;
+        qemu_mutex_lock(&client->mutex);
+        // clear uring_eventfd itself to prevent future scheduling of new B/H
+        client->uring_eventfd = -1;
+        wait_bh = client->bh_uring_scheduled;
+        qemu_mutex_unlock(&client->mutex);
+        if (wait_bh)
+        {
+            // wait until existing scheduled B/H is ran
+            BDRV_POLL_WHILE(bs, client->bh_uring_scheduled);
+        }
+    }
     vitastor_c_destroy(client->proxy);
     if (client->fds)
     {
@@ -749,7 +767,9 @@ static int coroutine_fn vitastor_co_preadv(BlockDriverState *bs,
     uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;
     qemu_mutex_lock(&client->mutex);
     vitastor_c_read(client->proxy, inode, offset, bytes, iov->iov, iov->niov, vitastor_co_read_cb, &task);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_schedule_uring_handler(client);
+#endif
     qemu_mutex_unlock(&client->mutex);
 
     while (!task.complete)
@@ -783,7 +803,9 @@ static int coroutine_fn vitastor_co_pwritev(BlockDriverState *bs,
     uint64_t inode = client->watch ? vitastor_c_inode_get_num(client->watch) : client->inode;
     qemu_mutex_lock(&client->mutex);
     vitastor_c_write(client->proxy, inode, offset, bytes, 0, iov->iov, iov->niov, vitastor_co_generic_cb, &task);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_schedule_uring_handler(client);
+#endif
     qemu_mutex_unlock(&client->mutex);
 
     while (!task.complete)
@@ -863,7 +885,9 @@ static int coroutine_fn vitastor_co_block_status(
         task.bitmap = client->last_bitmap = NULL;
         qemu_mutex_lock(&client->mutex);
         vitastor_c_read_bitmap(client->proxy, task.inode, task.offset, task.len, !client->skip_parents, vitastor_co_read_bitmap_cb, &task);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
         vitastor_schedule_uring_handler(client);
+#endif
         qemu_mutex_unlock(&client->mutex);
         while (!task.complete)
         {
@@ -950,7 +974,9 @@ static int coroutine_fn vitastor_co_flush(BlockDriverState *bs)
 
     qemu_mutex_lock(&client->mutex);
     vitastor_c_sync(client->proxy, vitastor_co_generic_cb, &task);
+#if !defined VITASTOR_C_API_VERSION || VITASTOR_C_API_VERSION < 5
     vitastor_schedule_uring_handler(client);
+#endif
     qemu_mutex_unlock(&client->mutex);
 
     while (!task.complete)

src/client/vitastor.pc.in

@@ -6,7 +6,7 @@ includedir=${prefix}/@CMAKE_INSTALL_INCLUDEDIR@
 
 Name: Vitastor
 Description: Vitastor client library
-Version: 1.11.0
+Version: 2.1.0
 Libs: -L${libdir} -lvitastor_client
 Cflags: -I${includedir}
 

src/client/vitastor_c.cpp

@@ -127,6 +127,7 @@ vitastor_c *vitastor_c_create_qemu_uring(QEMUSetFDHandler *aio_set_fd_handler, v
     auto self = vitastor_c_create_qemu_common(aio_set_fd_handler, aio_context);
     self->ringloop = ringloop;
     self->cli = new cluster_client_t(self->ringloop, self->tfd, cfg_json);
+    ringloop->loop();
     return self;
 }
 
@@ -150,6 +151,7 @@ vitastor_c *vitastor_c_create_uring(const char *config_path, const char *etcd_ho
     self->ringloop = ringloop;
     self->epmgr = new epoll_manager_t(self->ringloop);
     self->cli = new cluster_client_t(self->ringloop, self->epmgr->tfd, cfg_json);
+    ringloop->loop();
     return self;
 }
 
@@ -183,6 +185,7 @@ vitastor_c *vitastor_c_create_uring_json(const char **options, int options_len)
     self->ringloop = ringloop;
     self->epmgr = new epoll_manager_t(self->ringloop);
     self->cli = new cluster_client_t(self->ringloop, self->epmgr->tfd, cfg_json);
+    ringloop->loop();
     return self;
 }
 
@@ -228,6 +231,10 @@ void vitastor_c_on_ready(vitastor_c *client, VitastorIOHandler cb, void *opaque)
     {
         cb(opaque, 0);
     });
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_uring_wait_ready(vitastor_c *client)
@@ -284,6 +291,10 @@ void vitastor_c_read(vitastor_c *client, uint64_t inode, uint64_t offset, uint64
         delete op;
     };
     client->cli->execute(op);
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_write(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len, uint64_t check_version,
@@ -305,6 +316,10 @@ void vitastor_c_write(vitastor_c *client, uint64_t inode, uint64_t offset, uint6
         delete op;
     };
     client->cli->execute(op);
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_delete(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len, uint64_t check_version,
@@ -322,6 +337,10 @@ void vitastor_c_delete(vitastor_c *client, uint64_t inode, uint64_t offset, uint
         delete op;
     };
     client->cli->execute(op);
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_read_bitmap(vitastor_c *client, uint64_t inode, uint64_t offset, uint64_t len,
@@ -344,6 +363,10 @@ void vitastor_c_read_bitmap(vitastor_c *client, uint64_t inode, uint64_t offset,
         delete op;
     };
     client->cli->execute(op);
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_sync(vitastor_c *client, VitastorIOHandler cb, void *opaque)
@@ -356,6 +379,10 @@ void vitastor_c_sync(vitastor_c *client, VitastorIOHandler cb, void *opaque)
         delete op;
     };
     client->cli->execute(op);
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_watch_inode(vitastor_c *client, char *image, VitastorIOHandler cb, void *opaque)
@@ -365,6 +392,10 @@ void vitastor_c_watch_inode(vitastor_c *client, char *image, VitastorIOHandler c
         auto watch = client->cli->st_cli.watch_inode(std::string(image));
         cb(opaque, (long)watch);
     });
+    if (client->ringloop)
+    {
+        client->ringloop->loop();
+    }
 }
 
 void vitastor_c_close_watch(vitastor_c *client, void *handle)

src/client/vitastor_c.h

@@ -7,7 +7,7 @@
 #define VITASTOR_QEMU_PROXY_H
 
 // C API wrapper version
-#define VITASTOR_C_API_VERSION 4
+#define VITASTOR_C_API_VERSION 5
 
 #ifndef POOL_ID_BITS
 #define POOL_ID_BITS 16

src/cmd/cli.cpp

@@ -160,11 +160,12 @@ static const char* help_text =
     "vitastor-cli modify-osd [--tags tag1,tag2,...] [--reweight <number>] [--noout true/false] <osd_number>\n"
     "  Set OSD reweight, tags or noout flag.\n"
     "\n"
-    "vitastor-cli pg-list|pg-ls|list-pg|ls-pg|ls-pgs [OPTIONS] [state1+state2] [^state3] [...]\n"
+    "vitastor-cli pg-list|pg-ls|list-pg|ls-pg|ls-pgs|pgs [OPTIONS] [state1+state2] [^state3] [...]\n"
     "  List PGs with any of listed state filters (^ or ! in the beginning is negation). Options:\n"
     "    --pool <pool name or number>  Only list PGs of the given pool.\n"
     "    --min <min pg number>         Only list PGs with number >= min.\n"
     "    --max <max pg number>         Only list PGs with number <= max.\n"
+    "    --osd 1,2,...                 Only list PGs with some data on specified OSD(s).\n"
     "  Examples:\n"
     "    vitastor-cli pg-list active+degraded\n"
     "    vitastor-cli pg-list ^active\n"
@@ -186,7 +187,8 @@ static const char* help_text =
     "    --raw_placement <rules>       Specify raw PG generation rules (see documentation for details)\n"
     "    --primary_affinity_tags tags  Prefer to put primary copies on OSDs with all specified tags\n"
     "    --scrub_interval <time>       Enable regular scrubbing for this pool. Format: number + unit s/m/h/d/M/y\n"
-    "    --used_for_fs <name>          Mark pool as used for VitastorFS with metadata in image <name>\n"
+    "    --used_for_app fs:<name>      Mark pool as used for VitastorFS with metadata in image <name>\n"
+    "    --used_for_app s3:<name>      Mark pool as used for S3 location with name <name>\n"
     "    --pg_stripe_size <number>     Increase object grouping stripe\n"
     "    --max_osd_combinations 10000  Maximum number of random combinations for LP solver input\n"
     "    --wait                        Wait for the new pool to come online\n"
@@ -198,7 +200,7 @@ static const char* help_text =
     "vitastor-cli modify-pool|pool-modify <id|name> [--name <new_name>] [PARAMETERS...]\n"
     "  Modify an existing pool. Modifiable parameters:\n"
     "    [-s|--pg_size <number>] [--pg_minsize <number>] [-n|--pg_count <count>]\n"
-    "    [--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>] [--used_for_fs <name>]\n"
+    "    [--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>] [--used_for_app <type>:<name>]\n"
     "    [--max_osd_combinations <number>] [--primary_affinity_tags <tags>] [--scrub_interval <time>]\n"
     "    [--level_placement <rules>] [--raw_placement <rules>]\n"
     "  Non-modifiable parameters (changing them WILL lead to data loss):\n"
@@ -482,7 +484,7 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
             cfg["osd_num"] = cmd[1];
         action_cb = p->start_modify_osd(cfg);
     }
-    else if (cmd[0] == "pg-list" || cmd[0] == "pg-ls" || cmd[0] == "list-pg" || cmd[0] == "ls-pg" || cmd[0] == "ls-pgs")
+    else if (cmd[0] == "pg-list" || cmd[0] == "pg-ls" || cmd[0] == "list-pg" || cmd[0] == "ls-pg" || cmd[0] == "ls-pgs" || cmd[0] == "pgs")
     {
         // Modify OSD configuration
         if (cmd.size() > 1)

src/cmd/cli.h

@@ -98,5 +98,3 @@ std::string format_lat(uint64_t lat);
 std::string format_q(double depth);
 
 bool stupid_glob(const std::string str, const std::string glob);
-
-std::string implode(const std::string & sep, json11::Json array);

src/cmd/cli_create.cpp

@@ -92,12 +92,12 @@ struct image_creator_t
         {
             new_pool_id = pools.begin()->first;
         }
-        if (new_pool_id && !pools.at(new_pool_id).used_for_fs.empty() && !force)
+        if (new_pool_id && !pools.at(new_pool_id).used_for_app.empty() && !force)
         {
             result = (cli_result_t){
                 .err = EINVAL,
                 .text = "Pool "+pools.at(new_pool_id).name+
-                    " is used for VitastorFS "+pools.at(new_pool_id).used_for_fs+
+                    " is used for application "+pools.at(new_pool_id).used_for_app+
                     ". Use --force if you really know what you are doing",
             };
             state = 100;

src/cmd/cli_modify_osd.cpp

@@ -58,6 +58,12 @@ struct osd_changer_t
             state = 100;
             return;
         }
+        if (set_reweight && new_reweight > 1)
+        {
+            result = (cli_result_t){ .err = EINVAL, .text = "Reweight can't be larger than 1" };
+            state = 100;
+            return;
+        }
         parent->etcd_txn(json11::Json::object {
             { "success", json11::Json::array {
                 json11::Json::object {

src/cmd/cli_osd_tree.cpp

@@ -7,6 +7,7 @@
 #include "epoll_manager.h"
 #include "pg_states.h"
 #include "str_util.h"
+#include "json_util.h"
 
 struct placement_osd_t
 {

src/cmd/cli_pg_ls.cpp

@@ -5,12 +5,14 @@
 #include "cluster_client.h"
 #include "pg_states.h"
 #include "str_util.h"
+#include "json_util.h"
 
 struct pg_lister_t
 {
     cli_tool_t *parent;
 
     uint64_t pool_id = 0;
+    std::set<osd_num_t> osd_nums;
     std::string pool_name;
     std::vector<std::string> pg_state;
     uint64_t min_pg_num = 0;
@@ -103,7 +105,7 @@ resume_1:
                                 {
                                     if (pg_state_names[i] == bit)
                                     {
-                                        mask |= (uint64_t)1 << i;
+                                        mask |= pg_state_bits[i];
                                         found = true;
                                         break;
                                     }
@@ -137,6 +139,22 @@ resume_1:
                     {
                         continue;
                     }
+                    if (osd_nums.size())
+                    {
+                        bool found = false;
+                        for (int i = 0; !found && i < pgp.second.target_set.size(); i++)
+                            if (osd_nums.find(pgp.second.target_set[i]) != osd_nums.end())
+                                found = true;
+                        for (int i = 0; !found && i < pgp.second.target_history.size(); i++)
+                            for (int j = 0; !found && j < pgp.second.target_history[i].size(); j++)
+                                if (osd_nums.find(pgp.second.target_history[i][j]) != osd_nums.end())
+                                    found = true;
+                        for (int i = 0; !found && i < pgp.second.all_peers.size(); i++)
+                            if (osd_nums.find(pgp.second.all_peers[i]) != osd_nums.end())
+                                found = true;
+                        if (!found)
+                            continue;
+                    }
                     if (masks.size())
                     {
                         bool found = false;
@@ -156,7 +174,7 @@ resume_1:
                     json11::Json::array state_names;
                     for (int i = 0; i < pg_state_bit_count; i++)
                     {
-                        if (pgp.second.cur_state & (1 << i))
+                        if (pgp.second.cur_state & pg_state_bits[i])
                         {
                             state_names.push_back(std::string(pg_state_names[i]));
                         }
@@ -274,6 +292,14 @@ std::function<bool(cli_result_t &)> cli_tool_t::start_pg_list(json11::Json cfg)
         pg_lister->pg_state.push_back(cfg["pg_state"].string_value());
     pg_lister->min_pg_num = cfg["min"].uint64_value();
     pg_lister->max_pg_num = cfg["max"].uint64_value();
+    if (cfg["osd"].is_array())
+        for (auto & osd_num_json: cfg["osd"].array_items())
+            pg_lister->osd_nums.insert(osd_num_json.uint64_value());
+    else if (cfg["osd"].is_string())
+        for (auto & osd_num_str: explode(",", cfg["osd"].string_value(), true))
+            pg_lister->osd_nums.insert(stoull_full(osd_num_str));
+    else if (cfg["osd"].uint64_value())
+        pg_lister->osd_nums.insert(cfg["osd"].uint64_value());
     return [pg_lister](cli_result_t & result)
     {
         pg_lister->loop();

src/cmd/cli_pool_cfg.cpp

@@ -90,8 +90,8 @@ std::string validate_pool_config(json11::Json::object & new_cfg, json11::Json ol
             value = sz;
         }
         else if (key == "name" || key == "scheme" || key == "immediate_commit" ||
-            key == "failure_domain" || key == "root_node" || key == "scrub_interval" || key == "used_for_fs" ||
-            key == "raw_placement")
+            key == "failure_domain" || key == "root_node" || key == "scrub_interval" || key == "used_for_app" ||
+            key == "used_for_fs" || key == "raw_placement")
         {
             if (!value.is_string())
             {
@@ -156,8 +156,13 @@ std::string validate_pool_config(json11::Json::object & new_cfg, json11::Json ol
     {
         new_cfg.erase("parity_chunks");
     }
-    if (new_cfg.find("used_for_fs") != new_cfg.end() && new_cfg["used_for_fs"].string_value() == "")
+    if (new_cfg.find("used_for_app") != new_cfg.end() && new_cfg["used_for_app"].string_value() == "")
     {
+        new_cfg.erase("used_for_app");
+    }
+    if (new_cfg.find("used_for_app") == new_cfg.end() && new_cfg.find("used_for_fs") != new_cfg.end())
+    {
+        new_cfg["used_for_app"] = "fs:"+new_cfg["used_for_fs"].string_value();
         new_cfg.erase("used_for_fs");
     }
 

src/cmd/cli_pool_create.cpp

@@ -10,6 +10,7 @@
 #include "epoll_manager.h"
 #include "pg_states.h"
 #include "str_util.h"
+#include "json_util.h"
 
 struct pool_creator_t
 {

src/cmd/cli_pool_ls.cpp

@@ -5,6 +5,7 @@
 #include "cli.h"
 #include "cluster_client.h"
 #include "str_util.h"
+#include "json_util.h"
 #include "pg_states.h"
 
 // List pools with space statistics
@@ -199,7 +200,9 @@ resume_1:
             auto & st = pool_stats[pool_id];
             for (auto & kv: pp.second.object_items())
             {
-                if (st.find(kv.first) == st.end())
+                if (kv.first == "used_for_fs" && st.find("used_for_app") == st.end())
+                    st["used_for_app"] = "fs:"+kv.second.string_value();
+                else if (st.find(kv.first) == st.end())
                     st[kv.first] = kv.second;
             }
         }
@@ -493,7 +496,7 @@ resume_3:
                 { "name", "Name" },
                 { "id", "ID" },
                 { "scheme_name", "Scheme" },
-                { "used_for_fs", "Used for VitastorFS" },
+                { "used_for_app", "Used for app" },
                 { "status", "Status" },
                 { "backfillfull_fmt", "Backfillfull" },
                 { "pg_count_fmt", "PGs" },
@@ -663,19 +666,3 @@ std::function<bool(cli_result_t &)> cli_tool_t::start_pool_ls(json11::Json cfg)
         return false;
     };
 }
-
-std::string implode(const std::string & sep, json11::Json array)
-{
-    if (array.is_number() || array.is_bool() || array.is_string())
-    {
-        return array.as_string();
-    }
-    std::string res;
-    bool first = true;
-    for (auto & item: array.array_items())
-    {
-        res += (first ? item.as_string() : sep+item.as_string());
-        first = false;
-    }
-    return res;
-}

src/cmd/cli_pool_modify.cpp

@@ -112,19 +112,21 @@ resume_1:
                 return;
             }
 
-            if (new_cfg.find("used_for_fs") != new_cfg.end() && !force)
+            if (new_cfg.find("used_for_app") != new_cfg.end() && !force)
             {
                 // Check that pool doesn't have images
                 auto img_it = parent->cli->st_cli.inode_config.lower_bound(INODE_WITH_POOL(pool_id, 0));
-                if (img_it != parent->cli->st_cli.inode_config.end() && INODE_POOL(img_it->first) == pool_id &&
-                    img_it->second.name == new_cfg["used_for_fs"].string_value())
+                if (img_it != parent->cli->st_cli.inode_config.end() &&
+                    INODE_POOL(img_it->first) == pool_id &&
+                    new_cfg["used_for_app"].string_value().substr(0, 3) == "fs:" &&
+                    img_it->second.name == new_cfg["used_for_app"].string_value().substr(3))
                 {
                     // Only allow metadata image to exist in the FS pool
                     img_it++;
                 }
                 if (img_it != parent->cli->st_cli.inode_config.end() && INODE_POOL(img_it->first) == pool_id)
                 {
-                    result = (cli_result_t){ .err = ENOENT, .text = "Pool "+pool_name+" has block images, delete them before using it for VitastorFS" };
+                    result = (cli_result_t){ .err = ENOENT, .text = "Pool "+pool_name+" has block images, delete them before using it for VitastorFS, S3 or another app" };
                     state = 100;
                     return;
                 }

src/cmd/cli_rm_data.cpp

@@ -69,11 +69,11 @@ struct rm_inode_t
                 });
                 if (min_offset == 0 && max_offset == 0)
                 {
-                    total_count += objects.size();
+                    total_count += rm->objects.size();
                 }
                 else
                 {
-                    for (object_id oid: objects)
+                    for (object_id oid: rm->objects)
                     {
                         if (oid.stripe >= min_offset && (!max_offset || oid.stripe < max_offset))
                         {

src/cmd/cli_rm_wildcard.cpp

@@ -138,7 +138,7 @@ struct wildcard_remover_t
                     if (inode_it != parent->cli->st_cli.inode_config.end())
                         fprintf(stderr, "Warning: image %s modified by someone else during deletion, restarting wildcard deletion\n", inode_it->second.name.c_str());
                     else
-                        fprintf(stderr, "Warning: inode %lx modified by someone else during deletion, retrying wildcard deletion\n", irev.inode_num);
+                        fprintf(stderr, "Warning: inode %jx modified by someone else during deletion, retrying wildcard deletion\n", irev.inode_num);
                     goto resume_0;
                 }
             }

src/disk_tool/CMakeLists.txt

@@ -5,7 +5,7 @@ project(vitastor)
 # vitastor-disk
 add_executable(vitastor-disk
 	disk_tool.cpp disk_simple_offsets.cpp
-	disk_tool_journal.cpp disk_tool_meta.cpp disk_tool_prepare.cpp disk_tool_resize.cpp
+	disk_tool_discard.cpp disk_tool_journal.cpp disk_tool_meta.cpp disk_tool_prepare.cpp disk_tool_resize.cpp
 	disk_tool_resize_auto.cpp disk_tool_udev.cpp disk_tool_utils.cpp disk_tool_upgrade.cpp
 	../util/crc32c.c ../util/str_util.cpp ../util/json_util.cpp ../../json11/json11.cpp ../util/rw_blocking.cpp ../util/allocator.cpp ../util/ringloop.cpp ../blockstore/blockstore_disk.cpp
 )

src/disk_tool/disk_tool.cpp

@@ -68,6 +68,8 @@ static const char *help_text =
     "    --data_device_block 4k     Override data device block size\n"
     "    --meta_device_block 4k     Override metadata device block size\n"
     "    --journal_device_block 4k  Override journal device block size\n"
+    "    --discard_on_start 0       TRIM unused data device blocks every OSD start (default off)\n"
+    "    --min_discard_size 1M      Minimum TRIM block size\n"
     "    --json                     Enable JSON output\n"
     "  \n"
     "  immediate_commit setting is automatically derived from \"disable fsync\" options.\n"
@@ -128,6 +130,12 @@ static const char *help_text =
     "  SIZE may include k/m/g/t suffixes. If any of the new layout parameter\n"
     "  options are not specified, old values will be used.\n"
     "\n"
+    "vitastor-disk trim <osd_num>|<osd_device> [<osd_num>|<osd_device>...]\n"
+    "  Try to discard unused blocks (SSD TRIM) on the data device of each of the OSD(s).\n"
+    "  May only be used on stopped OSDs. Options:\n"
+    "    --min_discard_size 1M      Minimum TRIM block size\n"
+    "    --discard_granularity 0    Override device's discard granularity\n"
+    "\n"
     "vitastor-disk start|stop|restart|enable|disable [--now] <device> [device2 device3 ...]\n"
     "  Manipulate Vitastor OSDs using systemd by their device paths.\n"
     "  Commands are passed to systemctl with vitastor-osd@<num> units as arguments.\n"
@@ -428,6 +436,19 @@ int main(int argc, char *argv[])
         disk_tool_simple_offsets(self.options, self.json);
         return 0;
     }
+    else if (!strcmp(cmd[0], "trim"))
+    {
+        if (cmd.size() < 2)
+        {
+            fprintf(stderr, "OSD number(s) or device path(s) are required\n");
+            return 1;
+        }
+        for (int i = 1; i < cmd.size(); i++)
+        {
+            self.trim_data(cmd[i]);
+        }
+        return 0;
+    }
     else if (!strcmp(cmd[0], "udev"))
     {
         if (cmd.size() != 2)

src/disk_tool/disk_tool.h

@@ -60,7 +60,7 @@ struct disk_tool_t
 
     bool first_block, first_entry;
 
-    allocator *data_alloc;
+    allocator_t *data_alloc;
     std::map<uint64_t, uint64_t> data_remap;
     std::map<uint64_t, uint64_t>::iterator remap_it;
     ring_loop_t *ringloop;
@@ -86,10 +86,10 @@ struct disk_tool_t
 
     int dump_journal();
     void dump_journal_entry(int num, journal_entry *je, bool json);
-    int process_journal(std::function<int(void*)> block_fn);
+    int process_journal(std::function<int(void*)> block_fn, bool do_open = true);
     int process_journal_block(void *buf, std::function<void(int, journal_entry*)> iter_fn);
     int process_meta(std::function<void(blockstore_meta_header_v2_t *)> hdr_fn,
-        std::function<void(uint64_t, clean_disk_entry*, uint8_t*)> record_fn);
+        std::function<void(uint64_t, clean_disk_entry*, uint8_t*)> record_fn, bool do_open = true);
 
     int dump_meta();
     void dump_meta_header(blockstore_meta_header_v2_t *hdr);
@@ -123,6 +123,7 @@ struct disk_tool_t
     int pre_exec_osd(std::string device);
     int purge_devices(const std::vector<std::string> & devices);
     int clear_osd_superblock(const std::string & dev);
+    int trim_data(std::string device);
 
     json11::Json read_osd_superblock(std::string device, bool expect_exist = true, bool ignore_nonref = false);
     uint32_t write_osd_superblock(std::string device, json11::Json params);

src/disk_tool/disk_tool_discard.cpp

@@ -0,0 +1,89 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "disk_tool.h"
+#include "str_util.h"
+#include "json_util.h"
+
+int disk_tool_t::trim_data(std::string device)
+{
+    int r;
+    // Parse parameters
+    if (stoull_full(device))
+        device = "/dev/vitastor/osd"+device+"-data";
+    json11::Json sb = read_osd_superblock(device, true, false);
+    if (sb.is_null())
+        return 1;
+    auto sb_params = json_to_string_map(sb["params"].object_items());
+    if (options["discard_granularity"] != "")
+        sb_params["discard_granularity"] = options["discard_granularity"];
+    if (options["min_discard_size"] != "")
+        sb_params["min_discard_size"] = options["min_discard_size"];
+    try
+    {
+        dsk.parse_config(sb_params);
+    }
+    catch (std::exception & e)
+    {
+        fprintf(stderr, "Error: %s\n", e.what());
+        return 1;
+    }
+    if (!dsk.discard_granularity && sb["real_data_device"].string_value().substr(0, 5) == "/dev/")
+    {
+        auto dg = read_file("/sys/block/"+sb["real_data_device"].string_value().substr(5)+"/queue/discard_granularity", true);
+        if (dg != "")
+            dsk.discard_granularity = parse_size(trim(dg));
+    }
+    // Open devices
+    try
+    {
+        if (options["io"] != "")
+            dsk.data_io = dsk.meta_io = dsk.journal_io = options["io"];
+        dsk.open_data();
+        dsk.open_meta();
+        dsk.open_journal();
+        dsk.calc_lengths();
+    }
+    catch (std::exception & e)
+    {
+        dsk.close_all();
+        fprintf(stderr, "Error: %s\n", e.what());
+        return 1;
+    }
+    // Fill allocator
+    fprintf(stderr, "Reading metadata\n");
+    data_alloc = new allocator_t(dsk.block_count);
+    r = process_meta(
+        [this](blockstore_meta_header_v2_t *hdr) {},
+        [this](uint64_t block_num, clean_disk_entry *entry, uint8_t *bitmap)
+        {
+            data_alloc->set(block_num, true);
+        },
+        false
+    );
+    if (r != 0)
+    {
+        dsk.close_all();
+        return r;
+    }
+    fprintf(stderr, "Reading journal\n");
+    r = process_journal([this](void *buf)
+    {
+        return process_journal_block(buf, [this](int num, journal_entry *je)
+        {
+            if (je->type == JE_BIG_WRITE || je->type == JE_BIG_WRITE_INSTANT)
+            {
+                data_alloc->set(je->big_write.location / dsk.data_block_size, true);
+            }
+        });
+    }, false);
+    if (r != 0)
+    {
+        dsk.close_all();
+        return r;
+    }
+    // Trim
+    r = dsk.trim_data(data_alloc);
+    dsk.close_all();
+    return r == 0;
+}

src/disk_tool/disk_tool_journal.cpp

@@ -119,13 +119,21 @@ int disk_tool_t::dump_journal()
     return 0;
 }
 
-int disk_tool_t::process_journal(std::function<int(void*)> block_fn)
+int disk_tool_t::process_journal(std::function<int(void*)> block_fn, bool do_open)
 {
-    dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
-    if (dsk.journal_fd < 0)
+    if (do_open)
     {
-        fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
-        return 1;
+        if (dsk.journal_fd >= 0)
+        {
+            fprintf(stderr, "Bug: journal device is already opened\n");
+            return 1;
+        }
+        dsk.journal_fd = open(dsk.journal_device.c_str(), (options["io"] == "cached" ? 0 : O_DIRECT) | O_RDONLY);
+        if (dsk.journal_fd < 0)
+        {
+            fprintf(stderr, "Failed to open journal device %s: %s\n", dsk.journal_device.c_str(), strerror(errno));
+            return 1;
+        }
     }
     void *data = memalign_or_die(MEM_ALIGNMENT, dsk.journal_block_size);
     journal_pos = 0;
@@ -170,8 +178,11 @@ int disk_tool_t::process_journal(std::function<int(void*)> block_fn)
                 break;
         }
     }
-    close(dsk.journal_fd);
-    dsk.journal_fd = -1;
+    if (do_open)
+    {
+        close(dsk.journal_fd);
+        dsk.journal_fd = -1;
+    }
     free(data);
     return r;
 }