Move gethostname_str to utils

This greatly speeds up PG placement and makes it more uniform both because the LP task
becomes simpler and because the distribution of individual OSDs is optimised manually

docker/Makefile

@@ -3,7 +3,7 @@ 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

docs/config/network.en.md

@@ -34,6 +34,7 @@ between clients, OSDs and etcd.
 - [etcd_ws_keepalive_interval](#etcd_ws_keepalive_interval)
 - [etcd_min_reload_interval](#etcd_min_reload_interval)
 - [tcp_header_buffer_size](#tcp_header_buffer_size)
+- [min_zerocopy_send_size](#min_zerocopy_send_size)
 - [use_sync_send_recv](#use_sync_send_recv)
 
 ## osd_network
@@ -313,6 +314,34 @@ 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.
 
+## min_zerocopy_send_size
+
+- Type: integer
+- Default: 32768
+
+OSDs and clients will attempt to use io_uring-based zero-copy TCP send
+for buffers larger than this number of bytes. Zero-copy send with io_uring is
+supported since Linux kernel version 6.1. Support is auto-detected and disabled
+automatically when not available. It can also be disabled explicitly by setting
+this parameter to a negative value.
+
+⚠️ Warning! Zero-copy send performance may vary greatly from CPU to CPU and from
+one kernel version to another. Generally, it tends to only make benefit with larger
+messages. With smaller messages (say, 4 KB), it may actually be slower. 32 KB is
+enough for almost all CPUs, but even smaller values are optimal for some of them.
+For example, 4 KB is OK for EPYC Milan/Genoa and 12 KB is OK for Xeon Ice Lake
+(but verify it yourself please).
+
+Verification instructions:
+1. Add `iommu=pt` into your Linux kernel command line and reboot.
+2. Upgrade your kernel. For example, it's very important to use 6.11+ with recent AMD EPYCs.
+3. Run some tests with the [send-zerocopy liburing example](https://github.com/axboe/liburing/blob/master/examples/send-zerocopy.c)
+   to find the minimal message size for which zero-copy is optimal.
+   Use `./send-zerocopy tcp -4 -R` at the server side and
+   `time ./send-zerocopy tcp -4 -b 0 -s BUFFER_SIZE -D SERVER_IP` at the client side with
+   `-z 0` (no zero-copy) and `-z 1` (zero-copy), and compare MB/s and used CPU time
+   (user+system).
+
 ## use_sync_send_recv
 
 - Type: boolean

docs/config/network.ru.md

@@ -34,6 +34,7 @@
 - [etcd_ws_keepalive_interval](#etcd_ws_keepalive_interval)
 - [etcd_min_reload_interval](#etcd_min_reload_interval)
 - [tcp_header_buffer_size](#tcp_header_buffer_size)
+- [min_zerocopy_send_size](#min_zerocopy_send_size)
 - [use_sync_send_recv](#use_sync_send_recv)
 
 ## osd_network
@@ -321,6 +322,34 @@ Vitastor содержат 128-байтные заголовки, за котор
 поменять этот параметр и посмотреть, как он влияет на производительность
 случайного и линейного доступа.
 
+## min_zerocopy_send_size
+
+- Тип: целое число
+- Значение по умолчанию: 32768
+
+OSD и клиенты будут пробовать использовать TCP-отправку без копирования (zero-copy) на
+основе io_uring для буферов, больших, чем это число байт. Отправка без копирования
+поддерживается в io_uring, начиная с версии ядра Linux 6.1. Наличие поддержки
+проверяется автоматически и zero-copy отключается, когда поддержки нет. Также
+её можно отключить явно, установив данный параметр в отрицательное значение.
+
+⚠️ Внимание! Производительность данной функции может сильно отличаться на разных
+процессорах и на разных версиях ядра Linux. В целом, zero-copy обычно быстрее с
+большими сообщениями, а с мелкими (например, 4 КБ) zero-copy может быть даже
+медленнее. 32 КБ достаточно почти для всех процессоров, но для каких-то можно
+использовать даже меньшие значения. Например, для EPYC Milan/Genoa подходит 4 КБ,
+а для Xeon Ice Lake - 12 КБ (но, пожалуйста, перепроверьте это сами).
+
+Инструкция по проверке:
+1. Добавьте `iommu=pt` в командную строку загрузки вашего ядра Linux и перезагрузитесь.
+2. Обновите ядро. Например, для AMD EPYC очень важно использовать версию 6.11+.
+3. Позапускайте тесты с помощью [send-zerocopy из примеров liburing](https://github.com/axboe/liburing/blob/master/examples/send-zerocopy.c),
+   чтобы найти минимальный размер сообщения, для которого zero-copy отправка оптимальна.
+   Запускайте `./send-zerocopy tcp -4 -R` на стороне сервера и
+   `time ./send-zerocopy tcp -4 -b 0 -s РАЗМЕР_БУФЕРА -D АДРЕС_СЕРВЕРА` на стороне клиента
+   с опцией `-z 0` (обычная отправка) и `-z 1` (отправка без копирования), и сравнивайте
+   скорость в МБ/с и занятое процессорное время (user+system).
+
 ## use_sync_send_recv
 
 - Тип: булево (да/нет)

docs/config/src/network.yml

@@ -373,6 +373,55 @@
     параметра читается без дополнительного копирования. Вы можете попробовать
     поменять этот параметр и посмотреть, как он влияет на производительность
     случайного и линейного доступа.
+- name: min_zerocopy_send_size
+  type: int
+  default: 32768
+  info: |
+    OSDs and clients will attempt to use io_uring-based zero-copy TCP send
+    for buffers larger than this number of bytes. Zero-copy send with io_uring is
+    supported since Linux kernel version 6.1. Support is auto-detected and disabled
+    automatically when not available. It can also be disabled explicitly by setting
+    this parameter to a negative value.
+
+    ⚠️ Warning! Zero-copy send performance may vary greatly from CPU to CPU and from
+    one kernel version to another. Generally, it tends to only make benefit with larger
+    messages. With smaller messages (say, 4 KB), it may actually be slower. 32 KB is
+    enough for almost all CPUs, but even smaller values are optimal for some of them.
+    For example, 4 KB is OK for EPYC Milan/Genoa and 12 KB is OK for Xeon Ice Lake
+    (but verify it yourself please).
+
+    Verification instructions:
+    1. Add `iommu=pt` into your Linux kernel command line and reboot.
+    2. Upgrade your kernel. For example, it's very important to use 6.11+ with recent AMD EPYCs.
+    3. Run some tests with the [send-zerocopy liburing example](https://github.com/axboe/liburing/blob/master/examples/send-zerocopy.c)
+       to find the minimal message size for which zero-copy is optimal.
+       Use `./send-zerocopy tcp -4 -R` at the server side and
+       `time ./send-zerocopy tcp -4 -b 0 -s BUFFER_SIZE -D SERVER_IP` at the client side with
+       `-z 0` (no zero-copy) and `-z 1` (zero-copy), and compare MB/s and used CPU time
+       (user+system).
+  info_ru: |
+    OSD и клиенты будут пробовать использовать TCP-отправку без копирования (zero-copy) на
+    основе io_uring для буферов, больших, чем это число байт. Отправка без копирования
+    поддерживается в io_uring, начиная с версии ядра Linux 6.1. Наличие поддержки
+    проверяется автоматически и zero-copy отключается, когда поддержки нет. Также
+    её можно отключить явно, установив данный параметр в отрицательное значение.
+
+    ⚠️ Внимание! Производительность данной функции может сильно отличаться на разных
+    процессорах и на разных версиях ядра Linux. В целом, zero-copy обычно быстрее с
+    большими сообщениями, а с мелкими (например, 4 КБ) zero-copy может быть даже
+    медленнее. 32 КБ достаточно почти для всех процессоров, но для каких-то можно
+    использовать даже меньшие значения. Например, для EPYC Milan/Genoa подходит 4 КБ,
+    а для Xeon Ice Lake - 12 КБ (но, пожалуйста, перепроверьте это сами).
+
+    Инструкция по проверке:
+    1. Добавьте `iommu=pt` в командную строку загрузки вашего ядра Linux и перезагрузитесь.
+    2. Обновите ядро. Например, для AMD EPYC очень важно использовать версию 6.11+.
+    3. Позапускайте тесты с помощью [send-zerocopy из примеров liburing](https://github.com/axboe/liburing/blob/master/examples/send-zerocopy.c),
+       чтобы найти минимальный размер сообщения, для которого zero-copy отправка оптимальна.
+       Запускайте `./send-zerocopy tcp -4 -R` на стороне сервера и
+       `time ./send-zerocopy tcp -4 -b 0 -s РАЗМЕР_БУФЕРА -D АДРЕС_СЕРВЕРА` на стороне клиента
+       с опцией `-z 0` (обычная отправка) и `-z 1` (отправка без копирования), и сравнивайте
+       скорость в МБ/с и занятое процессорное время (user+system).
 - name: use_sync_send_recv
   type: bool
   default: false

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:2.1.0`
+   `docker pull vitastor:v2.1.0`
 2. Install scripts to the host system: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:2.1.0 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:v2.1.0 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:2.1.0`
+   `docker pull vitastor:v2.1.0`
 2. Установите скрипты в хост-систему командой: \
-   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:2.1.0 install.sh`
+   `docker run --rm -it -v /etc:/host-etc -v /usr/bin:/host-bin vitastor:v2.1.0 install.sh`
 3. Перезагрузите правила udev: \
    `udevadm control --reload-rules`
 

docs/usage/admin.en.md

@@ -14,6 +14,7 @@
 - [Removing a failed disk](#removing-a-failed-disk)
 - [Adding a disk](#adding-a-disk)
 - [Restoring from lost pool configuration](#restoring-from-lost-pool-configuration)
+- [Incompatibility problems](#Incompatibility-problems)
 - [Upgrading Vitastor](#upgrading-vitastor)
 - [OSD memory usage](#osd-memory-usage)
 
@@ -166,6 +167,17 @@ done
 
 After that all PGs should peer and find all previous data.
 
+## Incompatibility problems
+
+### ISA-L 2.31
+
+⚠ It is FORBIDDEN to use Vitastor 2.1.0 and earlier versions with ISA-L 2.31 and newer if
+you use EC N+K pools and K > 1 on a CPU with GF-NI instruction support, because it WILL
+lead to **data loss** during EC recovery.
+
+If you accidentally upgraded ISA-L to 2.31 but didn't upgrade Vitastor and restarted OSDs,
+then stop them as soon as possible and either update Vitastor or roll back ISA-L.
+
 ## Upgrading Vitastor
 
 Every upcoming Vitastor version is usually compatible with previous both forward

docs/usage/admin.ru.md

@@ -14,6 +14,7 @@
 - [Удаление неисправного диска](#удаление-неисправного-диска)
 - [Добавление диска](#добавление-диска)
 - [Восстановление потерянной конфигурации пулов](#восстановление-потерянной-конфигурации-пулов)
+- [Проблемы несовместимости](#проблемы-несовместимости)
 - [Обновление Vitastor](#обновление-vitastor)
 - [Потребление памяти OSD](#потребление-памяти-osd)
 
@@ -163,6 +164,17 @@ done
 
 После этого все PG должны пройти peering и найти все предыдущие данные.
 
+## Проблемы несовместимости
+
+### ISA-L 2.31
+
+⚠ ЗАПРЕЩЕНО использовать Vitastor 2.1.0 и более ранних версий с библиотекой ISA-L версии 2.31
+или более новой, если вы используете EC-пулы N+K и K > 1 на CPU с поддержкой инструкций GF-NI,
+так как это приведёт к **потере данных** при восстановлении из EC.
+
+Если вы случайно обновили ISA-L до 2.31, но не обновили Vitastor, и успели перезапустить OSD,
+то как можно скорее остановите их все и либо обновите Vitastor, либо откатите ISA-L.
+
 ## Обновление Vitastor
 
 Обычно каждая следующая версия Vitastor совместима с предыдущими и "вперёд", и "назад"

docs/usage/nfs.en.md

@@ -14,6 +14,9 @@ Commands:
 - [upgrade](#upgrade)
 - [defrag](#defrag)
 
+⚠️ Important: follow the instructions from [Linux NFS write size](#linux-nfs-write-size)
+for optimal Vitastor NFS performance if you use EC and HDD and mount your NFS from Linux.
+
 ## Pseudo-FS
 
 Simplified pseudo-FS proxy is used for file-based image access emulation. It's not
@@ -100,6 +103,62 @@ Other notable missing features which should be addressed in the future:
   in the DB. The FS is implemented is such way that this garbage doesn't affect its
   function, but having a tool to clean it up still seems a right thing to do.
 
+## Linux NFS write size
+
+Linux NFS client (nfs/nfsv3/nfsv4 kernel modules) has a hard-coded maximum I/O size,
+currently set to 1 MB - see `rsize` and `wsize` in [man 5 nfs](https://linux.die.net/man/5/nfs).
+
+This means that when you write to a file in an FS mounted over NFS, the maximum write
+request size is 1 MB, even in the O_DIRECT mode and even if the original write request
+is larger.
+
+However, for optimal linear write performance in Vitastor EC (erasure-coded) pools,
+the size of write requests should be a multiple of [block_size](../config/layout-cluster.en.md#block_size),
+multiplied by the data chunk count of the pool ([pg_size](../config/pool.en.md#pg_size)-[parity_chunks](../config/pool.en.md#parity_chunks)).
+When write requests are smaller or not a multiple of this number, Vitastor has to first
+read paired data blocks from disks, calculate new parity blocks and only then write them
+back. Obviously this is 2-3 times slower than a simple disk write.
+
+Vitastor HDD setups use 1 MB block_size by default. So, for optimal performance, if
+you use EC 2+1 and HDD, you need your NFS client to send 2 MB write requests, if you
+use EC 4+1 - 4 MB and so on.
+
+But Linux NFS client only writes in 1 MB chunks. 😢
+
+The good news is that you can fix it by rebuilding Linux NFS kernel modules 😉 🤩!
+You need to change NFS_MAX_FILE_IO_SIZE in nfs_xdr.h and then rebuild and reload modules.
+
+The instruction, using Debian as an example (should be ran under root):
+
+```
+# download current Linux kernel headers required to build modules
+apt-get install linux-headers-`uname -r`
+
+# replace NFS_MAX_FILE_IO_SIZE with a desired number (here it's 4194304 - 4 MB)
+sed -i 's/NFS_MAX_FILE_IO_SIZE\s*.*/NFS_MAX_FILE_IO_SIZE\t(4194304U)/' /lib/modules/`uname -r`/source/include/linux/nfs_xdr.h
+
+# download current Linux kernel source
+mkdir linux_src
+cd linux_src
+apt-get source linux-image-`uname -r`-unsigned
+
+# build NFS modules
+cd linux-*/fs/nfs
+make -C /lib/modules/`uname -r`/build M=$PWD -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD modules_install
+
+# move default NFS modules away
+mv /lib/modules/`uname -r`/kernel/fs/nfs ~/nfs_orig_`uname -r`
+depmod -a
+
+# unload old modules and load the new ones
+rmmod nfsv3 nfs
+modprobe nfsv3
+```
+
+After these (not much complicated 🙂) manipulations NFS begins to be mounted
+with new wsize and rsize by default and it fixes Vitastor-NFS linear write performance.
+
 ## Horizontal scaling
 
 Linux NFS 3.0 client doesn't support built-in scaling or failover, i.e. you can't

docs/usage/nfs.ru.md

@@ -14,6 +14,9 @@
 - [upgrade](#upgrade)
 - [defrag](#defrag)
 
+⚠️ Важно: для оптимальной производительности Vitastor NFS в Linux при использовании
+HDD и EC (erasure кодов) выполните инструкции из раздела [Размер записи Linux NFS](#размер-записи-linux-nfs).
+
 ## Псевдо-ФС
 
 Упрощённая реализация псевдо-ФС используется для эмуляции файлового доступа к блочным
@@ -104,6 +107,66 @@ JSON-формате :-). Для инспекции содержимого БД
   записи. ФС устроена так, что на работу они не влияют, но для порядка и их стоит
   уметь подчищать.
 
+## Размер записи Linux NFS
+
+Клиент Linux NFS (модули ядра nfs/nfsv3/nfsv4) имеет фиксированный в коде максимальный
+размер запроса ввода-вывода, равный 1 МБ - см. `rsize` и `wsize` в [man 5 nfs](https://linux.die.net/man/5/nfs).
+
+Это означает, что когда вы записываете в файл в примонтированной по NFS файловой системе,
+максимальный размер запроса записи составляет 1 МБ, даже в режиме O_DIRECT и даже если
+исходный запрос записи был больше.
+
+Однако для оптимальной скорости линейной записи в Vitastor при использовании EC-пулов
+(пулов с кодами коррекции ошибок) запросы записи должны быть по размеру кратны
+[block_size](../config/layout-cluster.ru.md#block_size), умноженному на число частей
+данных пула ([pg_size](../config/pool.ru.md#pg_size)-[parity_chunks](../config/pool.ru.md#parity_chunks)).
+Если запросы записи меньше или не кратны, то Vitastor приходится сначала прочитать
+с дисков старые версии парных блоков данных, рассчитать новые блоки чётности и только
+после этого записать их на диски. Естественно, это в 2-3 раза медленнее простой записи
+на диск.
+
+При этом block_size на жёстких дисках по умолчанию устанавливается равным 1 МБ.
+Таким образом, если вы используете EC 2+1 и HDD, для оптимальной скорости записи вам
+нужно, чтобы NFS-клиент писал по 2 МБ, если EC 4+1 и HDD - то по 4 МБ, и т.п.
+
+А Linux NFS-клиент пишет только по 1 МБ. 😢
+
+Но это можно исправить, пересобрав модули ядра Linux NFS 😉 🤩! Для этого нужно
+поменять значение переменной NFS_MAX_FILE_IO_SIZE в заголовочном файле nfs_xdr.h,
+после чего пересобрать модули NFS.
+
+Инструкция по пересборке на примере Debian (выполнять под root):
+
+```
+# скачиваем заголовки для сборки модулей для текущего ядра Linux
+apt-get install linux-headers-`uname -r`
+
+# заменяем в заголовках NFS_MAX_FILE_IO_SIZE на желаемый (здесь 4194304 - 4 МБ)
+sed -i 's/NFS_MAX_FILE_IO_SIZE\s*.*/NFS_MAX_FILE_IO_SIZE\t(4194304U)/' /lib/modules/`uname -r`/source/include/linux/nfs_xdr.h
+
+# скачиваем исходный код текущего ядра
+mkdir linux_src
+cd linux_src
+apt-get source linux-image-`uname -r`-unsigned
+
+# собираем модули NFS
+cd linux-*/fs/nfs
+make -C /lib/modules/`uname -r`/build M=$PWD -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD modules_install
+
+# убираем в сторону штатные модули NFS
+mv /lib/modules/`uname -r`/kernel/fs/nfs ~/nfs_orig_`uname -r`
+depmod -a
+
+# выгружаем старые модули и загружаем новые
+rmmod nfsv3 nfs
+modprobe nfsv3
+```
+
+После такой (относительно нехитрой 🙂) манипуляции NFS начинает по умолчанию
+монтироваться с новыми wsize и rsize, и производительность линейной записи в Vitastor-NFS
+исправляется.
+
 ## Горизонтальное масштабирование
 
 Клиент Linux NFS 3.0 не поддерживает встроенное масштабирование или отказоустойчивость.

docs/usage/qemu.en.md

@@ -162,10 +162,12 @@ apt-get install linux-headers-`uname -r`
 apt-get build-dep linux-image-`uname -r`-unsigned
 apt-get source linux-image-`uname -r`-unsigned
 cd linux*/drivers/vdpa
-make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules modules_install
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m modules_install
 cat Module.symvers >> /lib/modules/`uname -r`/build/Module.symvers
 cd ../virtio
-make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules modules_install
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m modules_install
 depmod -a
 ```
 

docs/usage/qemu.ru.md

@@ -165,10 +165,12 @@ apt-get install linux-headers-`uname -r`
 apt-get build-dep linux-image-`uname -r`-unsigned
 apt-get source linux-image-`uname -r`-unsigned
 cd linux*/drivers/vdpa
-make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules modules_install
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m modules_install
 cat Module.symvers >> /lib/modules/`uname -r`/build/Module.symvers
 cd ../virtio
-make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules modules_install
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m -j8 modules
+make -C /lib/modules/`uname -r`/build M=$PWD CONFIG_VDPA=m CONFIG_VDPA_USER=m CONFIG_VIRTIO_VDPA=m modules_install
 depmod -a
 ```
 

mon/lp_optimizer/dsl_pgs.js

@@ -253,7 +253,7 @@ function random_custom_combinations(osd_tree, rules, count, ordered)
         for (let i = 1; i < rules.length; i++)
         {
             const filtered = filter_tree_by_rules(osd_tree, rules[i], selected);
-            const idx = select_murmur3(filtered.length, i => 'p:'+f.id+':'+filtered[i].id);
+            const idx = select_murmur3(filtered.length, i => 'p:'+f.id+':'+(filtered[i].name || filtered[i].id));
             selected.push(idx == null ? { levels: {}, id: null } : filtered[idx]);
         }
         const size = selected.filter(s => s.id !== null).length;
@@ -270,7 +270,7 @@ function random_custom_combinations(osd_tree, rules, count, ordered)
         for (const item_rules of rules)
         {
             const filtered = selected.length ? filter_tree_by_rules(osd_tree, item_rules, selected) : first;
-            const idx = select_murmur3(filtered.length, i => n+':'+filtered[i].id);
+            const idx = select_murmur3(filtered.length, i => n+':'+(filtered[i].name || filtered[i].id));
             selected.push(idx == null ? { levels: {}, id: null } : filtered[idx]);
         }
         const size = selected.filter(s => s.id !== null).length;
@@ -340,9 +340,9 @@ function filter_tree_by_rules(osd_tree, rules, selected)
 }
 
 // Convert from
-// node_list = { id: string|number, level: string, size?: number, parent?: string|number }[]
+// node_list = { id: string|number, name?: string, level: string, size?: number, parent?: string|number }[]
 // to
-// node_tree = { [node_id]: { id, level, size?, parent?, children?: child_node_id[], levels: { [level]: id, ... } } }
+// node_tree = { [node_id]: { id, name?, level, size?, parent?, children?: child_node[], levels: { [level]: id, ... } } }
 function index_tree(node_list)
 {
     const tree = { '': { children: [], levels: {} } };
@@ -357,7 +357,7 @@ function index_tree(node_list)
         tree[parent_id].children = tree[parent_id].children || [];
         tree[parent_id].children.push(tree[node.id]);
     }
-    const cur = tree[''].children;
+    const cur = [ ...tree[''].children ];
     for (let i = 0; i < cur.length; i++)
     {
         cur[i].levels[cur[i].level] = cur[i].id;

mon/lp_optimizer/fold.js

@@ -0,0 +1,244 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+// Extract OSDs from the lowest affected tree level into a separate (flat) map
+// to run PG optimisation on failure domains instead of individual OSDs
+//
+// node_list = same input as for index_tree()
+// rules = [ level, operator, value ][][]
+// returns { nodes: new_node_list, leaves: { new_folded_node_id: [ extracted_leaf_nodes... ] } }
+function fold_failure_domains(node_list, rules)
+{
+    const interest = {};
+    for (const level_rules of rules)
+    {
+        for (const rule of level_rules)
+            interest[rule[0]] = true;
+    }
+    const max_numeric_id = node_list.reduce((a, c) => a < (0|c.id) ? (0|c.id) : a, 0);
+    let next_id = max_numeric_id;
+    const node_map = node_list.reduce((a, c) => { a[c.id||''] = c; return a; }, {});
+    const old_ids_by_new = {};
+    const extracted_nodes = {};
+    let folded = true;
+    while (folded)
+    {
+        const per_parent = {};
+        for (const node_id in node_map)
+        {
+            const node = node_map[node_id];
+            const p = node.parent || '';
+            per_parent[p] = per_parent[p]||[];
+            per_parent[p].push(node);
+        }
+        folded = false;
+        for (const node_id in per_parent)
+        {
+            const fold_node = node_id !== '' && per_parent[node_id].length > 0 && per_parent[node_id].filter(child => per_parent[child.id||''] || interest[child.level]).length == 0;
+            if (fold_node)
+            {
+                const old_node = node_map[node_id];
+                const new_id = ++next_id;
+                node_map[new_id] = {
+                    ...old_node,
+                    id: new_id,
+                    name: node_id, // for use in murmur3 hashes
+                    size: per_parent[node_id].reduce((a, c) => a + (Number(c.size)||0), 0),
+                };
+                delete node_map[node_id];
+                old_ids_by_new[new_id] = node_id;
+                extracted_nodes[new_id] = [];
+                for (const child of per_parent[node_id])
+                {
+                    if (old_ids_by_new[child.id])
+                    {
+                        extracted_nodes[new_id].push(...extracted_nodes[child.id]);
+                        delete extracted_nodes[child.id];
+                    }
+                    else
+                        extracted_nodes[new_id].push(child);
+                    delete node_map[child.id];
+                }
+                folded = true;
+            }
+        }
+    }
+    return { nodes: Object.values(node_map), leaves: extracted_nodes };
+}
+
+// Distribute PGs mapped to "folded" nodes to individual OSDs according to their weights
+// folded_pgs = optimize_result.int_pgs before folding
+// prev_pgs = optional previous PGs from optimize_change() input
+// extracted_nodes = output from fold_failure_domains
+function unfold_failure_domains(folded_pgs, prev_pgs, extracted_nodes)
+{
+    const maps = {};
+    let found = false;
+    for (const new_id in extracted_nodes)
+    {
+        const weights = {};
+        for (const sub_node of extracted_nodes[new_id])
+        {
+            weights[sub_node.id] = sub_node.size;
+        }
+        maps[new_id] = { weights, prev: [], next: [], pos: 0 };
+        found = true;
+    }
+    if (!found)
+    {
+        return folded_pgs;
+    }
+    for (let i = 0; i < folded_pgs.length; i++)
+    {
+        for (let j = 0; j < folded_pgs[i].length; j++)
+        {
+            if (maps[folded_pgs[i][j]])
+            {
+                maps[folded_pgs[i][j]].prev.push(prev_pgs && prev_pgs[i] && prev_pgs[i][j] || 0);
+            }
+        }
+    }
+    for (const new_id in maps)
+    {
+        maps[new_id].next = adjust_distribution(maps[new_id].weights, maps[new_id].prev);
+    }
+    const mapped_pgs = [];
+    for (let i = 0; i < folded_pgs.length; i++)
+    {
+        mapped_pgs.push(folded_pgs[i].map(osd => (maps[osd] ? maps[osd].next[maps[osd].pos++] : osd)));
+    }
+    return mapped_pgs;
+}
+
+// Return the new array of items re-distributed as close as possible to weights in wanted_weights
+// wanted_weights = { [key]: weight }
+// cur_items = key[]
+function adjust_distribution(wanted_weights, cur_items)
+{
+    const item_map = {};
+    for (let i = 0; i < cur_items.length; i++)
+    {
+        const item = cur_items[i];
+        item_map[item] = (item_map[item] || { target: 0, cur: [] });
+        item_map[item].cur.push(i);
+    }
+    let total_weight = 0;
+    for (const item in wanted_weights)
+    {
+        total_weight += Number(wanted_weights[item]) || 0;
+    }
+    for (const item in wanted_weights)
+    {
+        const weight = wanted_weights[item] / total_weight * cur_items.length;
+        if (weight > 0)
+        {
+            item_map[item] = (item_map[item] || { target: 0, cur: [] });
+            item_map[item].target = weight;
+        }
+    }
+    const diff = (item) => (item_map[item].cur.length - item_map[item].target);
+    const most_underweighted = Object.keys(item_map)
+        .filter(item => item_map[item].target > 0)
+        .sort((a, b) => diff(a) - diff(b));
+    // Items with zero target weight MUST never be selected - remove them
+    // and remap each of them to a most underweighted item
+    for (const item in item_map)
+    {
+        if (!item_map[item].target)
+        {
+            const prev = item_map[item];
+            delete item_map[item];
+            for (const idx of prev.cur)
+            {
+                const move_to = most_underweighted[0];
+                item_map[move_to].cur.push(idx);
+                move_leftmost(most_underweighted, diff);
+            }
+        }
+    }
+    // Other over-weighted items are only moved if it improves the distribution
+    while (most_underweighted.length > 1)
+    {
+        const first = most_underweighted[0];
+        const last = most_underweighted[most_underweighted.length-1];
+        const first_diff = diff(first);
+        const last_diff = diff(last);
+        if (Math.abs(first_diff+1)+Math.abs(last_diff-1) < Math.abs(first_diff)+Math.abs(last_diff))
+        {
+            item_map[first].cur.push(item_map[last].cur.pop());
+            move_leftmost(most_underweighted, diff);
+            move_rightmost(most_underweighted, diff);
+        }
+        else
+        {
+            break;
+        }
+    }
+    const new_items = new Array(cur_items.length);
+    for (const item in item_map)
+    {
+        for (const idx of item_map[item].cur)
+        {
+            new_items[idx] = item;
+        }
+    }
+    return new_items;
+}
+
+function move_leftmost(sorted_array, diff)
+{
+    // Re-sort by moving the leftmost item to the right if it changes position
+    const first = sorted_array[0];
+    const new_diff = diff(first);
+    let r = 0;
+    while (r < sorted_array.length-1 && diff(sorted_array[r+1]) <= new_diff)
+        r++;
+    if (r > 0)
+    {
+        for (let i = 0; i < r; i++)
+            sorted_array[i] = sorted_array[i+1];
+        sorted_array[r] = first;
+    }
+}
+
+function move_rightmost(sorted_array, diff)
+{
+    // Re-sort by moving the rightmost item to the left if it changes position
+    const last = sorted_array[sorted_array.length-1];
+    const new_diff = diff(last);
+    let r = sorted_array.length-1;
+    while (r > 0 && diff(sorted_array[r-1]) > new_diff)
+        r--;
+    if (r < sorted_array.length-1)
+    {
+        for (let i = sorted_array.length-1; i > r; i--)
+            sorted_array[i] = sorted_array[i-1];
+        sorted_array[r] = last;
+    }
+}
+
+// map previous PGs to folded nodes
+function fold_prev_pgs(pgs, extracted_nodes)
+{
+    const unmap = {};
+    for (const new_id in extracted_nodes)
+    {
+        for (const sub_node of extracted_nodes[new_id])
+        {
+            unmap[sub_node.id] = new_id;
+        }
+    }
+    const mapped_pgs = [];
+    for (let i = 0; i < pgs.length; i++)
+    {
+        mapped_pgs.push(pgs[i].map(osd => (unmap[osd] || osd)));
+    }
+    return mapped_pgs;
+}
+
+module.exports = {
+    fold_failure_domains,
+    unfold_failure_domains,
+    adjust_distribution,
+    fold_prev_pgs,
+};

mon/lp_optimizer/lp_optimizer.js

@@ -98,6 +98,7 @@ async function optimize_initial({ osd_weights, combinator, pg_count, pg_size = 3
         score: lp_result.score,
         weights: lp_result.vars,
         int_pgs,
+        pg_effsize,
         space: eff * pg_effsize,
         total_space: total_weight,
     };
@@ -409,6 +410,7 @@ async function optimize_change({ prev_pgs: prev_int_pgs, osd_weights, combinator
         int_pgs: new_pgs,
         differs,
         osd_differs,
+        pg_effsize,
         space: pg_effsize * pg_list_space_efficiency(new_pgs, osd_weights, pg_minsize, parity_space),
         total_space: total_weight,
     };

mon/lp_optimizer/test-optimize-fold.js

@@ -0,0 +1,108 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+const assert = require('assert');
+const { fold_failure_domains, unfold_failure_domains, adjust_distribution } = require('./fold.js');
+const DSL = require('./dsl_pgs.js');
+const LPOptimizer = require('./lp_optimizer.js');
+const stableStringify = require('../stable-stringify.js');
+
+async function run()
+{
+    // Test run adjust_distribution
+    console.log('adjust_distribution');
+    const rand = [];
+    for (let i = 0; i < 100; i++)
+    {
+        rand.push(1 + Math.floor(10*Math.random()));
+        // or rand.push(0);
+    }
+    const adj = adjust_distribution({ 1: 1, 2: 1, 3: 1, 4: 1, 5: 1, 6: 1, 7: 1, 8: 1, 9: 1, 10: 1 }, rand);
+    //console.log(rand.join(' '));
+    console.log(rand.reduce((a, c) => { a[c] = (a[c]||0)+1; return a; }, {}));
+    //console.log(adj.join(' '));
+    console.log(adj.reduce((a, c) => { a[c] = (a[c]||0)+1; return a; }, {}));
+    console.log('Movement: '+rand.reduce((a, c, i) => a+(rand[i] != adj[i] ? 1 : 0), 0)+'/'+rand.length);
+
+    console.log('\nfold_failure_domains');
+    console.log(JSON.stringify(fold_failure_domains(
+        [
+            { id: 1,       level: 'osd',  size: 1, parent: 'disk1' },
+            { id: 2,       level: 'osd',  size: 2, parent: 'disk1' },
+            { id: 'disk1', level: 'disk', parent: 'host1' },
+            { id: 'host1', level: 'host', parent: 'dc1' },
+            { id: 'dc1',   level: 'dc' },
+        ],
+        [ [ [ 'dc' ], [ 'host' ] ] ]
+    ), 0, 2));
+
+    console.log('\nfold_failure_domains empty rules');
+    console.log(JSON.stringify(fold_failure_domains(
+        [
+            { id: 1,       level: 'osd',  size: 1, parent: 'disk1' },
+            { id: 2,       level: 'osd',  size: 2, parent: 'disk1' },
+            { id: 'disk1', level: 'disk', parent: 'host1' },
+            { id: 'host1', level: 'host', parent: 'dc1' },
+            { id: 'dc1',   level: 'dc' },
+        ],
+        []
+    ), 0, 2));
+
+    console.log('\noptimize_folded');
+    // 5 DCs, 2 hosts per DC, 10 OSD per host
+    const nodes = [];
+    for (let i = 1; i <= 100; i++)
+    {
+        nodes.push({ id: i, level: 'osd', size: 1, parent: 'host'+(1+(0|((i-1)/10))) });
+    }
+    for (let i = 1; i <= 10; i++)
+    {
+        nodes.push({ id: 'host'+i, level: 'host', parent: 'dc'+(1+(0|((i-1)/2))) });
+    }
+    for (let i = 1; i <= 5; i++)
+    {
+        nodes.push({ id: 'dc'+i, level: 'dc' });
+    }
+
+    // Check rules
+    const rules = DSL.parse_level_indexes({ dc: '112233', host: '123456' }, [ 'dc', 'host', 'osd' ]);
+    assert.deepEqual(rules, [[],[["dc","=",1],["host","!=",[1]]],[["dc","!=",[1]]],[["dc","=",3],["host","!=",[3]]],[["dc","!=",[1,3]]],[["dc","=",5],["host","!=",[5]]]]);
+
+    // Check tree folding
+    const { nodes: folded_nodes, leaves: folded_leaves } = fold_failure_domains(nodes, rules);
+    const expected_folded = [];
+    const expected_leaves = {};
+    for (let i = 1; i <= 10; i++)
+    {
+        expected_folded.push({ id: 100+i, name: 'host'+i, level: 'host', size: 10, parent: 'dc'+(1+(0|((i-1)/2))) });
+        expected_leaves[100+i] = [ ...new Array(10).keys() ].map(k => ({ id: 10*(i-1)+k+1, level: 'osd', size: 1, parent: 'host'+i }));
+    }
+    for (let i = 1; i <= 5; i++)
+    {
+        expected_folded.push({ id: 'dc'+i, level: 'dc' });
+    }
+    assert.equal(stableStringify(folded_nodes), stableStringify(expected_folded));
+    assert.equal(stableStringify(folded_leaves), stableStringify(expected_leaves));
+
+    // Now optimise it
+    console.log('1000 PGs, EC 112233');
+    const leaf_weights = folded_nodes.reduce((a, c) => { if (Number(c.id)) { a[c.id] = c.size; } return a; }, {});
+    let res = await LPOptimizer.optimize_initial({
+        osd_weights: leaf_weights,
+        combinator: new DSL.RuleCombinator(folded_nodes, rules, 10000, false),
+        pg_size: 6,
+        pg_count: 1000,
+        ordered: false,
+    });
+    LPOptimizer.print_change_stats(res, false);
+    assert.equal(res.space, 100, 'Initial distribution');
+
+    const unfolded_res = { ...res };
+    unfolded_res.int_pgs = unfold_failure_domains(res.int_pgs, null, folded_leaves);
+    const osd_weights = nodes.reduce((a, c) => { if (Number(c.id)) { a[c.id] = c.size; } return a; }, {});
+    unfolded_res.space = unfolded_res.pg_effsize * LPOptimizer.pg_list_space_efficiency(unfolded_res.int_pgs, osd_weights, 0, 1);
+    LPOptimizer.print_change_stats(unfolded_res, false);
+    assert.equal(res.space, 100, 'Initial distribution');
+}
+
+run().catch(console.error);

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))
@@ -87,7 +87,7 @@ function make_hier_tree(global_config, tree)
     tree[''] = { children: [] };
     for (const node_id in tree)
     {
-        if (node_id === '' || tree[node_id].level === 'osd' && (!tree[node_id].size || tree[node_id].size <= 0))
+        if (node_id === '' || !(tree[node_id].children||[]).length && (tree[node_id].size||0) <= 0)
         {
             continue;
         }
@@ -107,10 +107,10 @@ function make_hier_tree(global_config, tree)
         deleted = 0;
         for (const node_id in tree)
         {
-            if (tree[node_id].level !== 'osd' && (!tree[node_id].children || !tree[node_id].children.length))
+            if (!(tree[node_id].children||[]).length && (tree[node_id].size||0) <= 0)
             {
                 const parent = tree[node_id].parent;
-                if (parent)
+                if (parent && tree[parent])
                 {
                     tree[parent].children = tree[parent].children.filter(c => c != tree[node_id]);
                 }

mon/pg_gen.js

@@ -3,6 +3,7 @@
 
 const { RuleCombinator } = require('./lp_optimizer/dsl_pgs.js');
 const { SimpleCombinator, flatten_tree } = require('./lp_optimizer/simple_pgs.js');
+const { fold_failure_domains, unfold_failure_domains, fold_prev_pgs } = require('./lp_optimizer/fold.js');
 const { validate_pool_cfg, get_pg_rules } = require('./pool_config.js');
 const LPOptimizer = require('./lp_optimizer/lp_optimizer.js');
 const { scale_pg_count } = require('./pg_utils.js');
@@ -160,7 +161,6 @@ async function generate_pool_pgs(state, global_config, pool_id, osd_tree, levels
         pool_cfg.bitmap_granularity || global_config.bitmap_granularity || 4096,
         pool_cfg.immediate_commit || global_config.immediate_commit || 'all'
     );
-    pool_tree = make_hier_tree(global_config, pool_tree);
     // First try last_clean_pgs to minimize data movement
     let prev_pgs = [];
     for (const pg in ((state.history.last_clean_pgs.items||{})[pool_id]||{}))
@@ -175,14 +175,19 @@ async function generate_pool_pgs(state, global_config, pool_id, osd_tree, levels
             prev_pgs[pg-1] = [ ...state.pg.config.items[pool_id][pg].osd_set ];
         }
     }
+    const use_rules = !global_config.use_old_pg_combinator || pool_cfg.level_placement || pool_cfg.raw_placement;
+    const rules = use_rules ? get_pg_rules(pool_id, pool_cfg, global_config.placement_levels) : null;
+    const folded = fold_failure_domains(Object.values(pool_tree), use_rules ? rules : [ [ [ pool_cfg.failure_domain ] ] ]);
+    // FIXME: Remove/merge make_hier_tree() step somewhere, however it's needed to remove empty nodes
+    const folded_tree = make_hier_tree(global_config, folded.nodes);
     const old_pg_count = prev_pgs.length;
     const optimize_cfg = {
-        osd_weights: Object.values(pool_tree).filter(item => item.level === 'osd').reduce((a, c) => { a[c.id] = c.size; return a; }, {}),
+        osd_weights: folded.nodes.reduce((a, c) => { if (Number(c.id)) { a[c.id] = c.size; } return a; }, {}),
-        combinator: !global_config.use_old_pg_combinator || pool_cfg.level_placement || pool_cfg.raw_placement
+        combinator: use_rules
             // new algorithm:
-            ? new RuleCombinator(pool_tree, get_pg_rules(pool_id, pool_cfg, global_config.placement_levels), pool_cfg.max_osd_combinations)
+            ? new RuleCombinator(folded_tree, rules, pool_cfg.max_osd_combinations)
             // old algorithm:
-            : new SimpleCombinator(flatten_tree(pool_tree[''].children, levels, pool_cfg.failure_domain, 'osd'), pool_cfg.pg_size, pool_cfg.max_osd_combinations),
+            : new SimpleCombinator(flatten_tree(folded_tree[''].children, levels, pool_cfg.failure_domain, 'osd'), pool_cfg.pg_size, pool_cfg.max_osd_combinations),
         pg_count: pool_cfg.pg_count,
         pg_size: pool_cfg.pg_size,
         pg_minsize: pool_cfg.pg_minsize,
@@ -202,12 +207,11 @@ async function generate_pool_pgs(state, global_config, pool_id, osd_tree, levels
         for (const pg of prev_pgs)
         {
             while (pg.length < pool_cfg.pg_size)
-            {
                 pg.push(0);
-            }
         }
+        const folded_prev_pgs = fold_prev_pgs(prev_pgs, folded.leaves);
         optimize_result = await LPOptimizer.optimize_change({
-            prev_pgs,
+            prev_pgs: folded_prev_pgs,
             ...optimize_cfg,
         });
     }
@@ -215,6 +219,10 @@ async function generate_pool_pgs(state, global_config, pool_id, osd_tree, levels
     {
         optimize_result = await LPOptimizer.optimize_initial(optimize_cfg);
     }
+    optimize_result.int_pgs = unfold_failure_domains(optimize_result.int_pgs, prev_pgs, folded.leaves);
+    const osd_weights = Object.values(pool_tree).reduce((a, c) => { if (c.level === 'osd') { a[c.id] = c.size; } return a; }, {});
+    optimize_result.space = optimize_result.pg_effsize * LPOptimizer.pg_list_space_efficiency(optimize_result.int_pgs,
+        osd_weights, optimize_cfg.pg_minsize, 1);
     console.log(`Pool ${pool_id} (${pool_cfg.name || 'unnamed'}):`);
     LPOptimizer.print_change_stats(optimize_result);
     let pg_effsize = pool_cfg.pg_size;

mon/scripts/make-etcd

@@ -40,6 +40,11 @@ async function run()
         console.log("/etc/systemd/system/vitastor-etcd.service already exists");
         process.exit(1);
     }
+    if (!in_docker && fs.existsSync("/etc/systemd/system/etcd.service"))
+    {
+        console.log("/etc/systemd/system/etcd.service already exists");
+        process.exit(1);
+    }
     const config = JSON.parse(fs.readFileSync(config_path, { encoding: 'utf-8' }));
     if (!config.etcd_address)
     {
@@ -66,7 +71,7 @@ async function run()
         console.log('etcd for Vitastor configured. Run `systemctl enable --now vitastor-etcd` to start etcd');
         process.exit(0);
     }
-    await system(`mkdir -p /var/lib/etcd`);
+    await system(`mkdir -p /var/lib/etcd/vitastor`);
     fs.writeFileSync(
         "/etc/systemd/system/vitastor-etcd.service",
 `[Unit]
@@ -77,14 +82,14 @@ Wants=network-online.target local-fs.target time-sync.target
 [Service]
 Restart=always
 Environment=GOGC=50
-ExecStart=etcd --name ${etcd_name} --data-dir /var/lib/etcd \\
+ExecStart=etcd --name ${etcd_name} --data-dir /var/lib/etcd/vitastor \\
     --snapshot-count 10000 --advertise-client-urls http://${etcds[num]}:2379 --listen-client-urls http://${etcds[num]}:2379 \\
     --initial-advertise-peer-urls http://${etcds[num]}:2380 --listen-peer-urls http://${etcds[num]}:2380 \\
     --initial-cluster-token vitastor-etcd-1 --initial-cluster ${etcd_cluster} \\
     --initial-cluster-state new --max-txn-ops=100000 --max-request-bytes=104857600 \\
     --auto-compaction-retention=10 --auto-compaction-mode=revision
-WorkingDirectory=/var/lib/etcd
+WorkingDirectory=/var/lib/etcd/vitastor
-ExecStartPre=+chown -R etcd /var/lib/etcd
+ExecStartPre=+chown -R etcd /var/lib/etcd/vitastor
 User=etcd
 PrivateTmp=false
 TasksMax=infinity
@@ -97,8 +102,9 @@ WantedBy=multi-user.target
 `);
     await system(`useradd etcd`);
     await system(`systemctl daemon-reload`);
-    await system(`systemctl enable etcd`);
+    // Disable distribution etcd unit and enable our one
-    await system(`systemctl start etcd`);
+    await system(`systemctl disable --now etcd`);
+    await system(`systemctl enable --now vitastor-etcd`);
     process.exit(0);
 }
 

mon/stats.js

@@ -87,11 +87,25 @@ function sum_op_stats(all_osd, prev_stats)
                 for (const k in derived[type][op])
                 {
                     sum_diff[type][op] = sum_diff[type][op] || {};
-                    sum_diff[type][op][k] = (sum_diff[type][op][k] || 0n) + derived[type][op][k];
+                    if (k == 'lat')
+                        sum_diff[type][op].lat = (sum_diff[type][op].lat || 0n) + derived[type][op].lat*derived[type][op].iops;
+                    else
+                        sum_diff[type][op][k] = (sum_diff[type][op][k] || 0n) + derived[type][op][k];
                 }
             }
         }
     }
+    // Calculate average (weighted by iops) op latency across all OSDs
+    for (const type in sum_diff)
+    {
+        for (const op in sum_diff[type])
+        {
+            if (sum_diff[type][op].lat)
+            {
+                sum_diff[type][op].lat /= sum_diff[type][op].iops;
+            }
+        }
+    }
     return sum_diff;
 }
 
@@ -271,8 +285,7 @@ function sum_inode_stats(state, prev_stats)
                     const op_st = inode_stats[pool_id][inode_num][op];
                     op_st.bps += op_diff.bps;
                     op_st.iops += op_diff.iops;
-                    op_st.lat += op_diff.lat;
+                    op_st.lat += op_diff.lat*op_diff.iops;
-                    op_st.n_osd = (op_st.n_osd || 0) + 1;
                 }
             }
         }
@@ -285,11 +298,8 @@ function sum_inode_stats(state, prev_stats)
             for (const op of [ 'read', 'write', 'delete' ])
             {
                 const op_st = inode_stats[pool_id][inode_num][op];
-                if (op_st.n_osd)
+                if (op_st.lat)
-                {
+                    op_st.lat /= op_st.iops;
-                    op_st.lat /= BigInt(op_st.n_osd);
-                    delete op_st.n_osd;
-                }
                 if (op_st.bps > 0 || op_st.iops > 0)
                     nonzero = true;
             }

src/blockstore/blockstore_impl.h

@@ -266,6 +266,8 @@ class blockstore_impl_t
     int throttle_threshold_us = 50;
     // Maximum writes between automatically added fsync operations
     uint64_t autosync_writes = 128;
+    // Log level (0-10)
+    int log_level = 0;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;

src/blockstore/blockstore_journal.cpp

@@ -113,10 +113,13 @@ int blockstore_journal_check_t::check_available(blockstore_op_t *op, int entries
     if (!right_dir && next_pos >= bs->journal.used_start-bs->journal.block_size)
     {
         // No space in the journal. Wait until used_start changes.
-        printf(
+        if (bs->log_level > 5)
-            "Ran out of journal space (used_start=%08jx, next_free=%08jx, dirty_start=%08jx)\n",
+        {
-            bs->journal.used_start, bs->journal.next_free, bs->journal.dirty_start
+            printf(
-        );
+                "Ran out of journal space (used_start=%08jx, next_free=%08jx, dirty_start=%08jx)\n",
+                bs->journal.used_start, bs->journal.next_free, bs->journal.dirty_start
+            );
+        }
         PRIV(op)->wait_for = WAIT_JOURNAL;
         bs->flusher->request_trim();
         PRIV(op)->wait_detail = bs->journal.used_start;

src/blockstore/blockstore_open.cpp

@@ -101,6 +101,7 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
         config["journal_no_same_sector_overwrites"] == "1" || config["journal_no_same_sector_overwrites"] == "yes";
     journal.inmemory = config["inmemory_journal"] != "false" && config["inmemory_journal"] != "0" &&
         config["inmemory_journal"] != "no";
+    log_level = strtoull(config["log_level"].c_str(), NULL, 10);
     // Validate
     if (journal.sector_count < 2)
     {

src/client/CMakeLists.txt

@@ -93,7 +93,7 @@ add_executable(test_cluster_client
 	EXCLUDE_FROM_ALL
 	../test/test_cluster_client.cpp
 	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp cluster_client_wb.cpp msgr_op.cpp ../test/mock/messenger.cpp msgr_stop.cpp
-	etcd_state_client.cpp ../util/timerfd_manager.cpp ../util/str_util.cpp ../util/json_util.cpp ../../json11/json11.cpp
+	etcd_state_client.cpp ../util/timerfd_manager.cpp ../util/addr_util.cpp ../util/str_util.cpp ../util/json_util.cpp ../../json11/json11.cpp
 )
 target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)
 target_include_directories(test_cluster_client BEFORE PUBLIC ${CMAKE_SOURCE_DIR}/src/test/mock)

src/client/cluster_client.cpp

@@ -1244,7 +1244,6 @@ int cluster_client_t::try_send(cluster_op_t *op, int i)
                 .req = { .rw = {
                     .header = {
                         .magic = SECONDARY_OSD_OP_MAGIC,
-                        .id = next_op_id(),
                         .opcode = op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP ? OSD_OP_READ : op->opcode,
                     },
                     .inode = op->cur_inode,
@@ -1353,7 +1352,6 @@ void cluster_client_t::send_sync(cluster_op_t *op, cluster_op_part_t *part)
         .req = {
             .hdr = {
                 .magic = SECONDARY_OSD_OP_MAGIC,
-                .id = next_op_id(),
                 .opcode = OSD_OP_SYNC,
             },
         },
@@ -1498,8 +1496,3 @@ void cluster_client_t::copy_part_bitmap(cluster_op_t *op, cluster_op_part_t *par
         part_len--;
     }
 }
-
-uint64_t cluster_client_t::next_op_id()
-{
-    return msgr.next_subop_id++;
-}

src/client/cluster_client.h

@@ -86,8 +86,8 @@ class cluster_client_t
 #ifdef __MOCK__
 public:
 #endif
-    timerfd_manager_t *tfd;
+    timerfd_manager_t *tfd = NULL;
-    ring_loop_t *ringloop;
+    ring_loop_t *ringloop = NULL;
 
     std::map<pool_id_t, uint64_t> pg_counts;
     std::map<pool_pg_num_t, osd_num_t> pg_primary;
@@ -152,7 +152,6 @@ public:
 
     //inline uint32_t get_bs_bitmap_granularity() { return st_cli.global_bitmap_granularity; }
     //inline uint64_t get_bs_block_size() { return st_cli.global_block_size; }
-    uint64_t next_op_id();
 
 #ifndef __MOCK__
 protected:

src/client/cluster_client_list.cpp

@@ -342,7 +342,6 @@ void cluster_client_t::send_list(inode_list_osd_t *cur_list)
         .sec_list = {
             .header = {
                 .magic = SECONDARY_OSD_OP_MAGIC,
-                .id = next_op_id(),
                 .opcode = OSD_OP_SEC_LIST,
             },
             .list_pg = cur_list->pg->pg_num,

src/client/messenger.cpp

@@ -167,6 +167,10 @@ void osd_messenger_t::init()
         }
     }
 #endif
+    if (ringloop)
+    {
+        has_sendmsg_zc = ringloop->has_sendmsg_zc();
+    }
     if (ringloop && iothread_count > 0)
     {
         for (int i = 0; i < iothread_count; i++)
@@ -213,7 +217,6 @@ void osd_messenger_t::init()
                     op->req = (osd_any_op_t){
                         .hdr = {
                             .magic = SECONDARY_OSD_OP_MAGIC,
-                            .id = this->next_subop_id++,
                             .opcode = OSD_OP_PING,
                         },
                     };
@@ -329,6 +332,9 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         this->receive_buffer_size = 65536;
     this->use_sync_send_recv = config["use_sync_send_recv"].bool_value() ||
         config["use_sync_send_recv"].uint64_value();
+    this->min_zerocopy_send_size = config["min_zerocopy_send_size"].is_null()
+        ? DEFAULT_MIN_ZEROCOPY_SEND_SIZE
+        : (int)config["min_zerocopy_send_size"].int64_value();
     this->peer_connect_interval = config["peer_connect_interval"].uint64_value();
     if (!this->peer_connect_interval)
         this->peer_connect_interval = 5;
@@ -622,13 +628,19 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
         .show_conf = {
             .header = {
                 .magic = SECONDARY_OSD_OP_MAGIC,
-                .id = this->next_subop_id++,
                 .opcode = OSD_OP_SHOW_CONFIG,
             },
         },
     };
+    json11::Json::object payload;
+    if (osd_num)
+    {
+        // Inform that we're OSD <osd_num>
+        payload["osd_num"] = osd_num;
+    }
+    payload["features"] = json11::Json::object{ { "check_sequencing", true } };
 #ifdef WITH_RDMA
-    if (rdma_contexts.size())
+    if (!use_rdmacm && rdma_contexts.size())
     {
         // Choose the right context for the selected network
         msgr_rdma_context_t *selected_ctx = choose_rdma_context(cl);
@@ -642,19 +654,20 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
             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 {
+                payload["connect_rdma"] = cl->rdma_conn->addr.to_string();
-                    { "connect_rdma", cl->rdma_conn->addr.to_string() },
+                payload["rdma_max_msg"] = cl->rdma_conn->max_msg;
-                    { "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
+    if (payload.size())
+    {
+        std::string payload_str = json11::Json(payload).dump();
+        op->req.show_conf.json_len = payload_str.size();
+        op->buf = malloc_or_die(payload_str.size());
+        op->iov.push_back(op->buf, payload_str.size());
+        memcpy(op->buf, payload_str.c_str(), payload_str.size());
+    }
     op->callback = [this, cl](osd_op_t *op)
     {
         std::string json_err;
@@ -701,7 +714,7 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
             return;
         }
 #ifdef WITH_RDMA
-        if (cl->rdma_conn && 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) ||
@@ -800,7 +813,8 @@ bool osd_messenger_t::is_rdma_enabled()
 {
     return rdma_contexts.size() > 0;
 }
-
+#endif
+#ifdef WITH_RDMACM
 bool osd_messenger_t::is_use_rdmacm()
 {
     return use_rdmacm;
@@ -896,6 +910,7 @@ static const char* local_only_params[] = {
     "tcp_header_buffer_size",
     "use_rdma",
     "use_sync_send_recv",
+    "min_zerocopy_send_size",
 };
 
 static const char **local_only_end = local_only_params + (sizeof(local_only_params)/sizeof(local_only_params[0]));

src/client/messenger.h

@@ -32,6 +32,8 @@
 
 #define VITASTOR_CONFIG_PATH "/etc/vitastor/vitastor.conf"
 
+#define DEFAULT_MIN_ZEROCOPY_SEND_SIZE 32*1024
+
 #define MSGR_SENDP_HDR 1
 #define MSGR_SENDP_FREE 2
 
@@ -73,12 +75,15 @@ struct osd_client_t
     int read_remaining = 0;
     int read_state = 0;
     osd_op_buf_list_t recv_list;
+    uint64_t read_op_id = 1;
+    bool check_sequencing = false;
 
     // Incoming operations
     std::vector<osd_op_t*> received_ops;
 
     // Outbound operations
     std::map<uint64_t, osd_op_t*> sent_ops;
+    uint64_t send_op_id = 0;
 
     // PGs dirtied by this client's primary-writes
     std::set<pool_pg_num_t> dirty_pgs;
@@ -88,6 +93,7 @@ struct osd_client_t
     int write_state = 0;
     std::vector<iovec> send_list, next_send_list;
     std::vector<msgr_sendp_t> outbox, next_outbox;
+    std::vector<osd_op_t*> zc_free_list;
 
     ~osd_client_t();
 };
@@ -97,6 +103,7 @@ struct osd_wanted_peer_t
     json11::Json raw_address_list;
     json11::Json address_list;
     int port = 0;
+    // FIXME: Remove separate WITH_RDMACM?
 #ifdef WITH_RDMACM
     int rdmacm_port = 0;
 #endif
@@ -175,6 +182,7 @@ protected:
     int osd_ping_timeout = 0;
     int log_level = 0;
     bool use_sync_send_recv = false;
+    int min_zerocopy_send_size = DEFAULT_MIN_ZEROCOPY_SEND_SIZE;
     int iothread_count = 0;
 
 #ifdef WITH_RDMA
@@ -201,11 +209,11 @@ protected:
     std::vector<osd_op_t*> set_immediate_ops;
 
 public:
-    timerfd_manager_t *tfd;
+    timerfd_manager_t *tfd = NULL;
-    ring_loop_t *ringloop;
+    ring_loop_t *ringloop = NULL;
+    bool has_sendmsg_zc = false;
     // osd_num_t is only for logging and asserts
     osd_num_t osd_num;
-    uint64_t next_subop_id = 1;
     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;
@@ -261,7 +269,7 @@ protected:
     void cancel_op(osd_op_t *op);
 
     bool try_send(osd_client_t *cl);
-    void handle_send(int result, osd_client_t *cl);
+    void handle_send(int result, bool prev, bool more, osd_client_t *cl);
 
     bool handle_read(int result, osd_client_t *cl);
     bool handle_read_buffer(osd_client_t *cl, void *curbuf, int remain);
@@ -286,6 +294,7 @@ protected:
     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_set_conn_timeout(rdmacm_connecting_t *conn);
     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);

src/client/msgr_rdma.cpp

@@ -70,6 +70,7 @@ msgr_rdma_context_t::~msgr_rdma_context_t()
 msgr_rdma_connection_t::~msgr_rdma_connection_t()
 {
     ctx->reserve_cqe(-max_send-max_recv);
+#ifdef WITH_RDMACM
     if (qp && !cmid)
         ibv_destroy_qp(qp);
     if (cmid)
@@ -79,6 +80,10 @@ msgr_rdma_connection_t::~msgr_rdma_connection_t()
             rdma_destroy_qp(cmid);
         rdma_destroy_id(cmid);
     }
+#else
+    if (qp)
+        ibv_destroy_qp(qp);
+#endif
     if (recv_buffers.size())
     {
         for (auto b: recv_buffers)
@@ -798,6 +803,9 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
             }
             if (!is_send)
             {
+                // Reset OSD ping state - client is obviously alive
+                cl->ping_time_remaining = 0;
+                cl->idle_time_remaining = osd_idle_timeout;
                 rc->cur_recv--;
                 if (!handle_read_buffer(cl, rc->recv_buffers[rc->next_recv_buf].buf, wc[i].byte_len))
                 {

src/client/msgr_rdmacm.cpp

@@ -70,7 +70,7 @@ void osd_messenger_t::rdmacm_destroy_listener(rdma_cm_id *listener)
 
 void osd_messenger_t::handle_rdmacm_events()
 {
-    // rdma_destroy_id infinitely waits for pthread_cond if called before all events are acked :-(
+    // rdma_destroy_id infinitely waits for pthread_cond if called before all events are acked :-(...
     std::vector<rdma_cm_event> events_copy;
     while (1)
     {
@@ -83,7 +83,15 @@ void osd_messenger_t::handle_rdmacm_events()
             fprintf(stderr, "Failed to get RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
             exit(1);
         }
-        events_copy.push_back(*ev);
+        // ...so we save a copy of all events EXCEPT connection requests, otherwise they sometimes fail with EVENT_DISCONNECT
+        if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
+        {
+            rdmacm_accept(ev);
+        }
+        else
+        {
+            events_copy.push_back(*ev);
+        }
         r = rdma_ack_cm_event(ev);
         if (r != 0)
         {
@@ -96,7 +104,7 @@ void osd_messenger_t::handle_rdmacm_events()
         auto ev = &evl;
         if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
         {
-            rdmacm_accept(ev);
+            // Do nothing, handled above
         }
         else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR ||
             ev->event == RDMA_CM_EVENT_REJECTED ||
@@ -287,29 +295,34 @@ void osd_messenger_t::rdmacm_accept(rdma_cm_event *ev)
         rdma_destroy_id(ev->id);
         return;
     }
-    rdma_context->cm_refs++;
+    // Wait for RDMA_CM_ESTABLISHED, and enable the connection only after it
-    // Wrap into a new msgr_rdma_connection_t
+    auto conn = new rdmacm_connecting_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;
+    conn->peer_fd = fake_fd;
-    auto cl = new osd_client_t();
+    conn->parsed_addr = *(sockaddr_storage*)rdma_get_peer_addr(ev->id);
-    cl->peer_fd = fake_fd;
+    conn->rdma_context = rdma_context;
-    cl->peer_state = PEER_RDMA;
+    rdmacm_set_conn_timeout(conn);
-    cl->peer_addr = *(sockaddr_storage*)rdma_get_peer_addr(ev->id);
+    rdmacm_connecting[ev->id] = conn;
-    cl->in_buf = malloc_or_die(receive_buffer_size);
+    fprintf(stderr, "[OSD %ju] new client %d: connection from %s via RDMA-CM\n", this->osd_num, conn->peer_fd,
-    cl->rdma_conn = conn;
+        addr_to_string(conn->parsed_addr).c_str());
-    clients[fake_fd] = cl;
+}
-    rdmacm_connections[ev->id] = cl;
+
-    // Add initial receive request(s)
+void osd_messenger_t::rdmacm_set_conn_timeout(rdmacm_connecting_t *conn)
-    try_recv_rdma(cl);
+{
-    fprintf(stderr, "[OSD %ju] new client %d: connection from %s via RDMA-CM\n", this->osd_num, fake_fd,
+    conn->timeout_ms = peer_connect_timeout*1000;
-        addr_to_string(cl->peer_addr).c_str());
+    if (peer_connect_timeout > 0)
+    {
+        conn->timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, cmid = conn->cmid](int timer_id)
+        {
+            auto conn = rdmacm_connecting.at(cmid);
+            conn->timeout_id = -1;
+            if (conn->peer_osd)
+                fprintf(stderr, "RDMA-CM connection to %s timed out\n", conn->addr.c_str());
+            else
+                fprintf(stderr, "Incoming RDMA-CM connection from %s timed out\n", addr_to_string(conn->parsed_addr).c_str());
+            rdmacm_on_connect_peer_error(cmid, -EPIPE);
+        });
+    }
 }
 
 void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res)
@@ -332,15 +345,18 @@ void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res)
     }
     rdmacm_connecting.erase(cmid);
     delete conn;
-    if (!disable_tcp)
+    if (peer_osd)
     {
-        // Fall back to TCP instead of just reporting the error to on_connect_peer()
+        if (!disable_tcp)
-        try_connect_peer_tcp(peer_osd, addr.c_str(), tcp_port);
+        {
-    }
+            // Fall back to TCP instead of just reporting the error to on_connect_peer()
-    else
+            try_connect_peer_tcp(peer_osd, addr.c_str(), tcp_port);
-    {
+        }
-        // TCP is disabled
+        else
-        on_connect_peer(peer_osd, res == 0 ? -EINVAL : (res > 0 ? -res : res));
+        {
+            // TCP is disabled
+            on_connect_peer(peer_osd, res == 0 ? -EINVAL : (res > 0 ? -res : res));
+        }
     }
 }
 
@@ -374,6 +390,8 @@ void osd_messenger_t::rdmacm_try_connect_peer(uint64_t peer_osd, const std::stri
         on_connect_peer(peer_osd, res);
         return;
     }
+    if (log_level > 0)
+        fprintf(stderr, "Trying to connect to OSD %ju at %s:%d via RDMA-CM\n", peer_osd, addr.c_str(), rdmacm_port);
     auto conn = new rdmacm_connecting_t;
     rdmacm_connecting[cmid] = conn;
     conn->cmid = cmid;
@@ -383,19 +401,7 @@ void osd_messenger_t::rdmacm_try_connect_peer(uint64_t peer_osd, const std::stri
     conn->parsed_addr = sa;
     conn->rdmacm_port = rdmacm_port;
     conn->tcp_port = fallback_tcp_port;
-    conn->timeout_ms = peer_connect_timeout*1000;
+    rdmacm_set_conn_timeout(conn);
-    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;
@@ -494,7 +500,7 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
     // 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->ctx->cm_refs++; // FIXME now unused, count also connecting_t's when used
     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
@@ -514,14 +520,20 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
     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)
+    if (log_level > 0 && peer_osd)
+    {
         fprintf(stderr, "Successfully connected with OSD %ju using RDMA-CM\n", peer_osd);
+    }
     // Add initial receive request(s)
     try_recv_rdma(cl);
-    osd_peer_fds[peer_osd] = cl->peer_fd;
+    if (peer_osd)
-    on_connect_peer(peer_osd, cl->peer_fd);
+    {
+        check_peer_config(cl);
+    }
 }

src/client/msgr_receive.cpp

@@ -214,6 +214,7 @@ bool osd_messenger_t::handle_read_buffer(osd_client_t *cl, void *curbuf, int rem
 
 bool osd_messenger_t::handle_finished_read(osd_client_t *cl)
 {
+    // Reset OSD ping state
     cl->ping_time_remaining = 0;
     cl->idle_time_remaining = osd_idle_timeout;
     cl->recv_list.reset();
@@ -222,7 +223,19 @@ bool osd_messenger_t::handle_finished_read(osd_client_t *cl)
         if (cl->read_op->req.hdr.magic == SECONDARY_OSD_REPLY_MAGIC)
             return handle_reply_hdr(cl);
         else if (cl->read_op->req.hdr.magic == SECONDARY_OSD_OP_MAGIC)
+        {
+            if (cl->check_sequencing)
+            {
+                if (cl->read_op->req.hdr.id != cl->read_op_id)
+                {
+                    fprintf(stderr, "Warning: operation sequencing is broken on client %d, stopping client\n", cl->peer_fd);
+                    stop_client(cl->peer_fd);
+                    return false;
+                }
+                cl->read_op_id++;
+            }
             handle_op_hdr(cl);
+        }
         else
         {
             fprintf(stderr, "Received garbage: magic=%jx id=%ju opcode=%jx from %d\n", cl->read_op->req.hdr.magic, cl->read_op->req.hdr.id, cl->read_op->req.hdr.opcode, cl->peer_fd);

src/client/msgr_send.cpp

@@ -14,6 +14,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
     if (cur_op->op_type == OSD_OP_OUT)
     {
         clock_gettime(CLOCK_REALTIME, &cur_op->tv_begin);
+        cur_op->req.hdr.id = ++cl->send_op_id;
     }
     else
     {
@@ -203,8 +204,24 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
         cl->refs++;
         ring_data_t* data = ((ring_data_t*)sqe->user_data);
-        data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, cl); };
+        data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, data->prev, data->more, cl); };
-        my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, 0);
+        bool use_zc = has_sendmsg_zc && min_zerocopy_send_size >= 0;
+        if (use_zc && min_zerocopy_send_size > 0)
+        {
+            size_t avg_size = 0;
+            for (size_t i = 0; i < cl->write_msg.msg_iovlen; i++)
+                avg_size += cl->write_msg.msg_iov[i].iov_len;
+            if (avg_size/cl->write_msg.msg_iovlen < min_zerocopy_send_size)
+                use_zc = false;
+        }
+        if (use_zc)
+        {
+            my_uring_prep_sendmsg_zc(sqe, peer_fd, &cl->write_msg, MSG_WAITALL);
+        }
+        else
+        {
+            my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, MSG_WAITALL);
+        }
         if (iothread)
         {
             iothread->add_sqe(sqe_local);
@@ -220,7 +237,7 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
         {
             result = -errno;
         }
-        handle_send(result, cl);
+        handle_send(result, false, false, cl);
     }
     return true;
 }
@@ -240,10 +257,16 @@ void osd_messenger_t::send_replies()
     write_ready_clients.clear();
 }
 
-void osd_messenger_t::handle_send(int result, osd_client_t *cl)
+void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t *cl)
 {
-    cl->write_msg.msg_iovlen = 0;
+    if (!prev)
-    cl->refs--;
+    {
+        cl->write_msg.msg_iovlen = 0;
+    }
+    if (!more)
+    {
+        cl->refs--;
+    }
     if (cl->peer_state == PEER_STOPPED)
     {
         if (cl->refs <= 0)
@@ -261,6 +284,16 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
     }
     if (result >= 0)
     {
+        if (prev)
+        {
+            // Second notification - only free a batch of postponed ops
+            int i = 0;
+            for (; i < cl->zc_free_list.size() && cl->zc_free_list[i]; i++)
+                delete cl->zc_free_list[i];
+            if (i > 0)
+                cl->zc_free_list.erase(cl->zc_free_list.begin(), cl->zc_free_list.begin()+i+1);
+            return;
+        }
         int done = 0;
         while (result > 0 && done < cl->send_list.size())
         {
@@ -270,7 +303,10 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 if (cl->outbox[done].flags & MSGR_SENDP_FREE)
                 {
                     // Reply fully sent
-                    delete cl->outbox[done].op;
+                    if (more)
+                        cl->zc_free_list.push_back(cl->outbox[done].op);
+                    else
+                        delete cl->outbox[done].op;
                 }
                 result -= iov.iov_len;
                 done++;
@@ -282,6 +318,12 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 break;
             }
         }
+        if (more)
+        {
+            auto expected = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+            assert(done == expected);
+            cl->zc_free_list.push_back(NULL); // end marker
+        }
         if (done > 0)
         {
             cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+done);

src/cmd/cli_describe.cpp

@@ -147,7 +147,6 @@ struct cli_describe_t
                 .describe = (osd_op_describe_t){
                     .header = (osd_op_header_t){
                         .magic = SECONDARY_OSD_OP_MAGIC,
-                        .id = parent->cli->next_op_id(),
                         .opcode = OSD_OP_DESCRIBE,
                     },
                     .object_state = object_state,

src/cmd/cli_fix.cpp

@@ -159,7 +159,6 @@ struct cli_fix_t
                 .describe = {
                     .header = {
                         .magic = SECONDARY_OSD_OP_MAGIC,
-                        .id = parent->cli->next_op_id(),
                         .opcode = OSD_OP_DESCRIBE,
                     },
                     .min_inode = obj.inode,
@@ -194,7 +193,6 @@ struct cli_fix_t
                                 .sec_del = {
                                     .header = {
                                         .magic = SECONDARY_OSD_OP_MAGIC,
-                                        .id = parent->cli->next_op_id(),
                                         .opcode = OSD_OP_SEC_DELETE,
                                     },
                                     .oid = {
@@ -242,7 +240,6 @@ struct cli_fix_t
                                         .rw = {
                                             .header = {
                                                 .magic = SECONDARY_OSD_OP_MAGIC,
-                                                .id = parent->cli->next_op_id(),
                                                 .opcode = OSD_OP_SCRUB,
                                             },
                                             .inode = obj.inode,

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_pool_cfg.cpp

@@ -44,10 +44,10 @@ std::string validate_pool_config(json11::Json::object & new_cfg, json11::Json ol
         new_cfg["parity_chunks"] = parity_chunks;
     }
 
-    if (old_cfg.is_null() && new_cfg["scheme"].string_value() == "")
+    if (new_cfg["scheme"].string_value() == "")
     {
         // Default scheme
-        new_cfg["scheme"] = "replicated";
+        new_cfg["scheme"] = old_cfg.is_null() ? "replicated" : old_cfg["scheme"];
     }
     if (new_cfg.find("pg_minsize") == new_cfg.end() && (old_cfg.is_null() || new_cfg.find("pg_size") != new_cfg.end()))
     {

src/cmd/cli_rm_osd.cpp

@@ -5,6 +5,7 @@
 #include "cli.h"
 #include "cluster_client.h"
 #include "str_util.h"
+#include "json_util.h"
 #include "epoll_manager.h"
 
 #include <algorithm>
@@ -22,8 +23,8 @@ struct rm_osd_t
     int state = 0;
     cli_result_t result;
 
-    std::set<uint64_t> to_remove;
+    std::set<osd_num_t> to_remove;
-    std::set<uint64_t> to_restart;
+    std::vector<osd_num_t> still_up;
     json11::Json::array pool_effects;
     json11::Json::array history_updates, history_checks;
     json11::Json new_pgs, new_clean_pgs;
@@ -63,8 +64,17 @@ struct rm_osd_t
             }
             to_remove.insert(osd_id);
         }
-        // Check if OSDs are still used in data distribution
         is_warning = is_dataloss = false;
+        // Check if OSDs are still up
+        for (auto osd_id: to_remove)
+        {
+            if (parent->cli->st_cli.peer_states.find(osd_id) != parent->cli->st_cli.peer_states.end())
+            {
+                is_warning = true;
+                still_up.push_back(osd_id);
+            }
+        }
+        // Check if OSDs are still used in data distribution
         for (auto & pp: parent->cli->st_cli.pool_config)
         {
             // Will OSD deletion make pool incomplete / down / degraded?
@@ -158,6 +168,9 @@ struct rm_osd_t
                                 : strtoupper(e["effect"].string_value())+" PGs"))
                 )+" after deleting OSD(s).\n";
             }
+            if (still_up.size())
+                error += (still_up.size() == 1 ? "OSD " : "OSDs ") + implode(", ", still_up) +
+                    (still_up.size() == 1 ? "is" : "are") + " still up. Use `vitastor-disk purge` to delete them.\n";
             if (is_dataloss && !force_dataloss && !dry_run)
                 error += "OSDs not deleted. Please move data to other OSDs or bypass this check with --allow-data-loss if you know what you are doing.\n";
             else if (is_warning && !force_warning && !dry_run)

src/kv/kv_cli.cpp

@@ -17,6 +17,8 @@
 #include "str_util.h"
 #include "vitastor_kv.h"
 
+#define KV_LIST_BUF_SIZE 65536
+
 const char *exe_name = NULL;
 
 class kv_cli_t
@@ -290,10 +292,26 @@ void kv_cli_t::next_cmd()
 struct kv_cli_list_t
 {
     vitastorkv_dbw_t *db = NULL;
+    std::string buf;
     void *handle = NULL;
     int format = 0;
     int n = 0;
     std::function<void(int)> cb;
+
+    void write(const std::string & str)
+    {
+        if (buf.capacity() < KV_LIST_BUF_SIZE)
+            buf.reserve(KV_LIST_BUF_SIZE);
+        if (buf.size() + str.size() > buf.capacity())
+            flush();
+        buf.append(str.data(), str.size());
+    }
+
+    void flush()
+    {
+        ::write(1, buf.data(), buf.size());
+        buf.resize(0);
+    }
 };
 
 std::vector<std::string> kv_cli_t::parse_cmd(const std::string & str)
@@ -604,11 +622,10 @@ void kv_cli_t::handle_cmd(const std::vector<std::string> & cmd, std::function<vo
             if (res < 0)
             {
                 if (res != -ENOENT)
-                {
                     fprintf(stderr, "Error: %s (code %d)\n", strerror(-res), res);
-                }
                 if (lst->format == 2)
-                    printf("\n}\n");
+                    lst->write("\n}\n");
+                lst->flush();
                 lst->db->list_close(lst->handle);
                 lst->cb(res == -ENOENT ? 0 : res);
                 delete lst;
@@ -616,11 +633,27 @@ void kv_cli_t::handle_cmd(const std::vector<std::string> & cmd, std::function<vo
             else
             {
                 if (lst->format == 2)
-                    printf(lst->n ? ",\n  %s: %s" : "{\n  %s: %s", addslashes(key).c_str(), addslashes(value).c_str());
+                {
+                    lst->write(lst->n ? ",\n  " : "{\n  ");
+                    lst->write(addslashes(key));
+                    lst->write(": ");
+                    lst->write(addslashes(value));
+                }
                 else if (lst->format == 1)
-                    printf("set %s %s\n", auto_addslashes(key).c_str(), value.c_str());
+                {
+                    lst->write("set ");
+                    lst->write(auto_addslashes(key));
+                    lst->write(" ");
+                    lst->write(value);
+                    lst->write("\n");
+                }
                 else
-                    printf("%s = %s\n", key.c_str(), value.c_str());
+                {
+                    lst->write(key);
+                    lst->write(" = ");
+                    lst->write(value);
+                    lst->write("\n");
+                }
                 lst->n++;
                 lst->db->list_next(lst->handle, NULL);
             }

src/kv/kv_db.cpp

@@ -870,7 +870,7 @@ static void get_block(kv_db_t *db, uint64_t offset, int cur_level, int recheck_p
         }
         // Block already in cache, we can proceed
         blk->usage = db->usage_counter;
-        cb(0, BLK_UPDATING);
+        db->cli->msgr.ringloop->set_immediate([=] { cb(0, BLK_UPDATING); });
         return;
     }
     cluster_op_t *op = new cluster_op_t;

src/nfs/nfs_fsstat.cpp

@@ -22,8 +22,8 @@ int nfs3_fsstat_proc(void *opaque, rpc_op_t *rop)
     {
         auto ttb = pst_it->second["total_raw_tb"].number_value();
         auto ftb = (pst_it->second["total_raw_tb"].number_value() - pst_it->second["used_raw_tb"].number_value());
-        tbytes = ttb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
+        tbytes = ttb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)1<<40);
-        fbytes = ftb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
+        fbytes = ftb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)1<<40);
     }
     *reply = (FSSTAT3res){
         .status = NFS3_OK,

src/nfs/nfs_kv_create.cpp

@@ -210,6 +210,7 @@ resume_4:
             st->res = res;
             kv_continue_create(st, 5);
         });
+        return;
 resume_5:
         if (st->res < 0)
         {

src/nfs/nfs_kv_getattr.cpp

@@ -13,6 +13,12 @@ void kv_read_inode(nfs_proxy_t *proxy, uint64_t ino,
     std::function<void(int res, const std::string & value, json11::Json ientry)> cb,
     bool allow_cache)
 {
+    if (!ino)
+    {
+        // Zero value can not exist
+        cb(-ENOENT, "", json11::Json());
+        return;
+    }
     auto key = kv_inode_key(ino);
     proxy->db->get(key, [=](int res, const std::string & value)
     {
@@ -49,7 +55,7 @@ int kv_nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     auto ino = kv_fh_inode(fh);
     if (self->parent->trace)
         fprintf(stderr, "[%d] GETATTR %ju\n", self->nfs_fd, ino);
-    if (!kv_fh_valid(fh))
+    if (!kv_fh_valid(fh) || !ino)
     {
         *reply = (GETATTR3res){ .status = NFS3ERR_INVAL };
         rpc_queue_reply(rop);

src/nfs/nfs_kv_lookup.cpp

@@ -43,9 +43,30 @@ int kv_nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
         uint64_t ino = direntry["ino"].uint64_value();
         kv_read_inode(self->parent, ino, [=](int res, const std::string & value, json11::Json ientry)
         {
-            if (res < 0)
+            if (res == -ENOENT)
             {
-                *reply = (LOOKUP3res){ .status = vitastor_nfs_map_err(res == -ENOENT ? -EIO : res) };
+                *reply = (LOOKUP3res){
+                    .status = NFS3_OK,
+                    .resok = (LOOKUP3resok){
+                        .object = xdr_copy_string(rop->xdrs, kv_fh(ino)),
+                        .obj_attributes = {
+                            .attributes_follow = 1,
+                            .attributes = (fattr3){
+                                .type = (ftype3)0,
+                                .mode = 0666,
+                                .nlink = 1,
+                                .fsid = self->parent->fsid,
+                                .fileid = ino,
+                            },
+                        },
+                    },
+                };
+                rpc_queue_reply(rop);
+                return;
+            }
+            else if (res < 0)
+            {
+                *reply = (LOOKUP3res){ .status = vitastor_nfs_map_err(res) };
                 rpc_queue_reply(rop);
                 return;
             }

src/nfs/nfs_kv_remove.cpp

@@ -89,12 +89,23 @@ resume_1:
 resume_2:
     if (st->res < 0)
     {
-        fprintf(stderr, "error reading inode %s: %s (code %d)\n",
+        if (st->res == -ENOENT)
-            kv_inode_key(st->ino).c_str(), strerror(-st->res), st->res);
+        {
-        auto cb = std::move(st->cb);
+            // Just delete direntry and skip inode
-        cb(st->res);
+            fprintf(stderr, "direntry %s references a non-existing inode %ju, deleting\n",
-        return;
+                kv_direntry_key(st->dir_ino, st->filename).c_str(), st->ino);
+            st->ino = 0;
+        }
+        else
+        {
+            fprintf(stderr, "error reading inode %s: %s (code %d)\n",
+                kv_inode_key(st->ino).c_str(), strerror(-st->res), st->res);
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
     }
+    else
     {
         std::string err;
         st->ientry = json11::Json::parse(st->ientry_text, err);

src/osd/osd_cluster.cpp

@@ -158,18 +158,13 @@ bool osd_t::check_peer_config(osd_client_t *cl, json11::Json conf)
 
 json11::Json osd_t::get_osd_state()
 {
-    std::vector<char> hostname;
-    hostname.resize(1024);
-    while (gethostname(hostname.data(), hostname.size()) < 0 && errno == ENAMETOOLONG)
-        hostname.resize(hostname.size()+1024);
-    hostname.resize(strnlen(hostname.data(), hostname.size()));
     json11::Json::object st;
     st["state"] = "up";
     if (bind_addresses.size() != 1 || bind_addresses[0] != "0.0.0.0")
         st["addresses"] = bind_addresses;
     else
         st["addresses"] = getifaddr_list();
-    st["host"] = std::string(hostname.data(), hostname.size());
+    st["host"] = gethostname_str();
     st["version"] = VITASTOR_VERSION;
     st["port"] = listening_port;
 #ifdef WITH_RDMACM

src/osd/osd_flush.cpp

@@ -209,7 +209,6 @@ bool osd_t::submit_flush_op(pool_id_t pool_id, pg_num_t pg_num, pg_flush_batch_t
             .sec_stab = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = (uint64_t)(rollback ? OSD_OP_SEC_ROLLBACK : OSD_OP_SEC_STABILIZE),
                 },
                 .len = count * sizeof(obj_ver_id),

src/osd/osd_peering.cpp

@@ -391,7 +391,6 @@ void osd_t::submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
             .sec_list = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = OSD_OP_SEC_LIST,
                 },
                 .list_pg = ps->pg_num,

src/osd/osd_primary_chain.cpp

@@ -266,7 +266,6 @@ int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
                     .sec_read_bmp = {
                         .header = {
                             .magic = SECONDARY_OSD_OP_MAGIC,
-                            .id = msgr.next_subop_id++,
                             .opcode = OSD_OP_SEC_READ_BMP,
                         },
                         .len = sizeof(obj_ver_id)*(i+1-prev),

src/osd/osd_primary_subops.cpp

@@ -233,7 +233,6 @@ void osd_t::submit_primary_subop(osd_op_t *cur_op, osd_op_t *subop,
         subop->req.sec_rw = (osd_op_sec_rw_t){
             .header = {
                 .magic = SECONDARY_OSD_OP_MAGIC,
-                .id = msgr.next_subop_id++,
                 .opcode = (uint64_t)(wr ? (cur_op->op_data->pg->scheme == POOL_SCHEME_REPLICATED ? OSD_OP_SEC_WRITE_STABLE : OSD_OP_SEC_WRITE) : OSD_OP_SEC_READ),
             },
             .oid = {
@@ -452,15 +451,16 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
         {
             op_data->errcode = retval;
         }
-        op_data->errors++;
         if (subop->peer_fd >= 0 && retval != -EDOM && retval != -ERANGE &&
             (retval != -ENOSPC || opcode != OSD_OP_SEC_WRITE && opcode != OSD_OP_SEC_WRITE_STABLE) &&
             (retval != -EIO || opcode != OSD_OP_SEC_READ))
         {
             // Drop connection on unexpected errors
-            op_data->drops++;
             msgr.stop_client(subop->peer_fd);
+            op_data->drops++;
         }
+        // Increase op_data->errors after stop_client to prevent >= n_subops running twice
+        op_data->errors++;
     }
     else
     {
@@ -593,7 +593,6 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
             subops[i].req = (osd_any_op_t){ .sec_del = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = OSD_OP_SEC_DELETE,
                 },
                 .oid = chunk.oid,
@@ -653,7 +652,6 @@ int osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
             subops[i].req = (osd_any_op_t){ .sec_sync = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = OSD_OP_SEC_SYNC,
                 },
                 .flags = cur_op->peer_fd == SELF_FD && cur_op->req.hdr.opcode != OSD_OP_SCRUB ? OSD_OP_RECOVERY_RELATED : 0,
@@ -712,7 +710,6 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
             subops[i].req = (osd_any_op_t){ .sec_stab = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = OSD_OP_SEC_STABILIZE,
                 },
                 .len = (uint64_t)(stab_osd.len * sizeof(obj_ver_id)),
@@ -806,7 +803,6 @@ void osd_t::submit_primary_rollback_subops(osd_op_t *cur_op, const uint64_t* osd
                 subop->req = (osd_any_op_t){ .sec_stab = {
                     .header = {
                         .magic = SECONDARY_OSD_OP_MAGIC,
-                        .id = msgr.next_subop_id++,
                         .opcode = OSD_OP_SEC_ROLLBACK,
                     },
                     .len = sizeof(obj_ver_id),

src/osd/osd_rmw.cpp

@@ -162,6 +162,7 @@ struct reed_sol_matrix_t
     int refs = 0;
     int *je_data;
     uint8_t *isal_data;
+    int isal_item_size;
     // 32 bytes = 256/8 = max pg_size/8
     std::map<std::array<uint8_t, 32>, void*> subdata;
     std::map<reed_sol_erased_t, void*> decodings;
@@ -181,20 +182,42 @@ void use_ec(int pg_size, int pg_minsize, bool use)
         }
         int *matrix = reed_sol_vandermonde_coding_matrix(pg_minsize, pg_size-pg_minsize, OSD_JERASURE_W);
         uint8_t *isal_table = NULL;
+        int item_size = 8;
 #ifdef WITH_ISAL
         uint8_t *isal_matrix = (uint8_t*)malloc_or_die(pg_minsize*(pg_size-pg_minsize));
         for (int i = 0; i < pg_minsize*(pg_size-pg_minsize); i++)
         {
             isal_matrix[i] = matrix[i];
         }
-        isal_table = (uint8_t*)malloc_or_die(pg_minsize*(pg_size-pg_minsize)*32);
+        isal_table = (uint8_t*)calloc_or_die(1, pg_minsize*(pg_size-pg_minsize)*32);
         ec_init_tables(pg_minsize, pg_size-pg_minsize, isal_matrix, isal_table);
         free(isal_matrix);
+        for (int i = pg_minsize*(pg_size-pg_minsize)*8; i < pg_minsize*(pg_size-pg_minsize)*32; i++)
+        {
+            if (isal_table[i] != 0)
+            {
+                // ISA-L GF-NI version uses 8-byte table items
+                item_size = 32;
+                break;
+            }
+        }
+        // Sanity check: rows should never consist of all zeroes
+        uint8_t zero_row[pg_minsize*item_size];
+        memset(zero_row, 0, pg_minsize*item_size);
+        for (int i = 0; i < (pg_size-pg_minsize); i++)
+        {
+            if (memcmp(isal_table + i*pg_minsize*item_size, zero_row, pg_minsize*item_size) == 0)
+            {
+                fprintf(stderr, "BUG or ISA-L incompatibility: EC tables shouldn't have all-zero rows\n");
+                abort();
+            }
+        }
 #endif
         matrices[key] = (reed_sol_matrix_t){
             .refs = 0,
             .je_data = matrix,
             .isal_data = isal_table,
+            .isal_item_size = item_size,
         };
         rs_it = matrices.find(key);
     }
@@ -235,7 +258,7 @@ static reed_sol_matrix_t* get_ec_matrix(int pg_size, int pg_minsize)
 // we don't need it. also it makes an extra allocation of int *erased on every call and doesn't cache
 // the decoding matrix.
 // all these flaws are fixed in this function:
-static void* get_jerasure_decoding_matrix(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize)
+static void* get_jerasure_decoding_matrix(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize, int *item_size)
 {
     int edd = 0;
     int erased[pg_size];
@@ -292,6 +315,7 @@ static void* get_jerasure_decoding_matrix(osd_rmw_stripe_t *stripes, int pg_size
         int *erased_copy = (int*)(rectable + 32*smrow*pg_minsize);
         memcpy(erased_copy, erased, pg_size*sizeof(int));
         matrix->decodings.emplace((reed_sol_erased_t){ .data = erased_copy, .size = pg_size }, rectable);
+        *item_size = matrix->isal_item_size;
         return rectable;
 #else
         int *dm_ids = (int*)malloc_or_die(sizeof(int)*(pg_minsize + pg_minsize*pg_minsize + pg_size));
@@ -355,7 +379,8 @@ static void jerasure_matrix_encode_unaligned(int k, int m, int w, int *matrix, c
 #ifdef WITH_ISAL
 void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize, uint32_t bitmap_size)
 {
-    uint8_t *dectable = (uint8_t*)get_jerasure_decoding_matrix(stripes, pg_size, pg_minsize);
+    int item_size = 0;
+    uint8_t *dectable = (uint8_t*)get_jerasure_decoding_matrix(stripes, pg_size, pg_minsize, &item_size);
     if (!dectable)
     {
         return;
@@ -378,7 +403,7 @@ void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsi
                 }
             }
             ec_encode_data(
-                read_end-read_start, pg_minsize, wanted, dectable + wanted_base*32*pg_minsize,
+                read_end-read_start, pg_minsize, wanted, dectable + wanted_base*item_size*pg_minsize,
                 data_ptrs, data_ptrs + pg_minsize
             );
         }
@@ -433,7 +458,7 @@ void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsi
 #else
 void reconstruct_stripes_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize, uint32_t bitmap_size)
 {
-    int *dm_ids = (int*)get_jerasure_decoding_matrix(stripes, pg_size, pg_minsize);
+    int *dm_ids = (int*)get_jerasure_decoding_matrix(stripes, pg_size, pg_minsize, NULL);
     if (!dm_ids)
     {
         return;
@@ -980,7 +1005,7 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
                 {
                     int item_size =
 #ifdef WITH_ISAL
-                        32;
+                        matrix->isal_item_size;
 #else
                         sizeof(int);
 #endif

src/osd/osd_scrub.cpp

@@ -65,7 +65,6 @@ void osd_t::scrub_list(pool_pg_num_t pg_id, osd_num_t role_osd, object_id min_oi
             .sec_list = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
-                    .id = msgr.next_subop_id++,
                     .opcode = OSD_OP_SEC_LIST,
                 },
                 .list_pg = pg_num,

src/osd/osd_secondary.cpp

@@ -198,6 +198,12 @@ void osd_t::exec_show_config(osd_op_t *cur_op)
     json11::Json req_json = cur_op->req.show_conf.json_len > 0
         ? json11::Json::parse(std::string((char *)cur_op->buf), json_err)
         : json11::Json();
+    if (req_json["features"]["check_sequencing"].bool_value())
+    {
+        auto cl = msgr.clients.at(cur_op->peer_fd);
+        cl->check_sequencing = true;
+        cl->read_op_id = cur_op->req.hdr.id + 1;
+    }
     // Expose sensitive configuration values so peers can check them
     json11::Json::object wire_config = json11::Json::object {
         { "osd_num", osd_num },

src/test/mock/messenger.cpp

@@ -21,7 +21,9 @@ osd_messenger_t::~osd_messenger_t()
 
 void osd_messenger_t::outbox_push(osd_op_t *cur_op)
 {
-    clients[cur_op->peer_fd]->sent_ops[cur_op->req.hdr.id] = cur_op;
+    auto cl = clients.at(cur_op->peer_fd);
+    cur_op->req.hdr.id = ++cl->send_op_id;
+    cl->sent_ops[cur_op->req.hdr.id] = cur_op;
 }
 
 void osd_messenger_t::parse_config(const json11::Json & config)

src/util/addr_util.cpp

@@ -245,3 +245,13 @@ int create_and_bind_socket(std::string bind_address, int bind_port, int listen_b
 
     return listen_fd;
 }
+
+std::string gethostname_str()
+{
+    std::string hostname;
+    hostname.resize(1024);
+    while (gethostname((char*)hostname.data(), hostname.size()) < 0 && errno == ENAMETOOLONG)
+        hostname.resize(hostname.size()+1024);
+    hostname.resize(strnlen(hostname.data(), hostname.size()));
+    return hostname;
+}

src/util/addr_util.h

@@ -21,3 +21,4 @@ bool cidr6_match(const in6_addr &address, const in6_addr &network, uint8_t bits)
 bool cidr_sockaddr_match(const sockaddr_storage &addr, const addr_mask_t &mask);
 std::vector<std::string> getifaddr_list(const std::vector<addr_mask_t> & masks = std::vector<addr_mask_t>(), bool include_v6 = false);
 int create_and_bind_socket(std::string bind_address, int bind_port, int listen_backlog, int *listening_port);
+std::string gethostname_str();

src/util/ringloop.cpp

@@ -10,6 +10,10 @@
 
 #include "ringloop.h"
 
+#ifndef IORING_CQE_F_MORE
+#define IORING_CQE_F_MORE (1U << 1)
+#endif
+
 ring_loop_t::ring_loop_t(int qd, bool multithreaded)
 {
     mt = multithreaded;
@@ -30,6 +34,16 @@ ring_loop_t::ring_loop_t(int qd, bool multithreaded)
         free_ring_data[i] = i;
     }
     in_loop = false;
+    auto probe = io_uring_get_probe();
+    if (probe)
+    {
+        support_zc = io_uring_opcode_supported(probe, IORING_OP_SENDMSG_ZC);
+#ifdef IORING_SETUP_R_DISABLED /* liburing 2.0 check */
+        io_uring_free_probe(probe);
+#else
+        free(probe);
+#endif
+    }
 }
 
 ring_loop_t::~ring_loop_t()
@@ -108,7 +122,17 @@ void ring_loop_t::loop()
         if (mt)
             mu.lock();
         struct ring_data_t *d = (struct ring_data_t*)cqe->user_data;
-        if (d->callback)
+        if (cqe->flags & IORING_CQE_F_MORE)
+        {
+            // There will be a second notification
+            d->res = cqe->res;
+            d->more = true;
+            if (d->callback)
+                d->callback(d);
+            d->prev = true;
+            d->more = false;
+        }
+        else if (d->callback)
         {
             // First free ring_data item, then call the callback
             // so it has at least 1 free slot for the next event
@@ -116,7 +140,10 @@ void ring_loop_t::loop()
             struct ring_data_t dl;
             dl.iov = d->iov;
             dl.res = cqe->res;
+            dl.more = false;
+            dl.prev = d->prev;
             dl.callback.swap(d->callback);
+            d->prev = d->more = false;
             free_ring_data[free_ring_data_ptr++] = d - ring_datas;
             if (mt)
                 mu.unlock();

src/util/ringloop.h

@@ -18,6 +18,10 @@
 
 #define RINGLOOP_DEFAULT_SIZE 1024
 
+#ifndef IORING_RECV_MULTISHOT /* liburing-2.3 check */
+#define IORING_OP_SENDMSG_ZC 48
+#endif
+
 static inline void my_uring_prep_rw(int op, struct io_uring_sqe *sqe, int fd, const void *addr, unsigned len, off_t offset)
 {
     // Prepare a read/write operation without clearing user_data
@@ -62,6 +66,12 @@ static inline void my_uring_prep_sendmsg(struct io_uring_sqe *sqe, int fd, const
     sqe->msg_flags = flags;
 }
 
+static inline void my_uring_prep_sendmsg_zc(struct io_uring_sqe *sqe, int fd, const struct msghdr *msg, unsigned flags)
+{
+    my_uring_prep_rw(IORING_OP_SENDMSG_ZC, sqe, fd, msg, 1, 0);
+    sqe->msg_flags = flags;
+}
+
 static inline void my_uring_prep_poll_add(struct io_uring_sqe *sqe, int fd, short poll_mask)
 {
     my_uring_prep_rw(IORING_OP_POLL_ADD, sqe, fd, NULL, 0, 0);
@@ -112,6 +122,8 @@ struct ring_data_t
 {
     struct iovec iov; // for single-entry read/write operations
     int res;
+    bool prev: 1;
+    bool more: 1;
     std::function<void(ring_data_t*)> callback;
 };
 
@@ -133,6 +145,7 @@ class ring_loop_t
     bool loop_again;
     struct io_uring ring;
     int ring_eventfd = -1;
+    bool support_zc = false;
 public:
     ring_loop_t(int qd, bool multithreaded = false);
     ~ring_loop_t();
@@ -144,6 +157,7 @@ public:
     inline void set_immediate(const std::function<void()> cb)
     {
         immediate_queue.push_back(cb);
+        wakeup();
     }
     inline int submit()
     {
@@ -163,6 +177,10 @@ public:
     {
         return loop_again;
     }
+    inline bool has_sendmsg_zc()
+    {
+        return support_zc;
+    }
 
     void loop();
     void wakeup();

tests/common.sh

@@ -87,6 +87,22 @@ wait_etcd()
     done
 }
 
+wait_condition()
+{
+    sec=$1
+    check=$2
+    proc=$3
+    i=0
+    while [[ $i -lt $sec ]]; do
+        eval "$check" && break
+        if [ $i -eq $sec ]; then
+            format_error "$proc couldn't finish in $sec seconds"
+        fi
+        sleep 1
+        i=$((i+1))
+    done
+}
+
 if [[ -n "$ANTIETCD" ]]; then
     ETCDCTL="node mon/node_modules/.bin/anticli -e $ETCD_URL"
     MON_PARAMS="--use_antietcd 1 --antietcd_data_dir ./testdata --antietcd_persist_interval 500 $MON_PARAMS"

tests/run_3osds.sh

@@ -127,22 +127,6 @@ try_reweight()
     sleep 3
 }
 
-wait_condition()
-{
-    sec=$1
-    check=$2
-    proc=$3
-    i=0
-    while [[ $i -lt $sec ]]; do
-        eval "$check" && break
-        if [ $i -eq $sec ]; then
-            format_error "$proc couldn't finish in $sec seconds"
-        fi
-        sleep 1
-        i=$((i+1))
-    done
-}
-
 wait_finish_rebalance()
 {
     sec=$1

tests/test_failure_domain.sh

@@ -31,4 +31,10 @@ sleep 2
 $ETCDCTL get --prefix /vitastor/pg/config --print-value-only | \
     jq -s -e '([ .[0].items["1"] | .[].osd_set | map_values(. | tonumber) | select((.[0] <= 4) != (.[1] <= 4)) ] | length) == 4'
 
+# test pool with size 1
+build/src/cmd/vitastor-cli --etcd_address $ETCD_URL create-pool size1pool -s 1 -n 1 --force
+wait_condition 10 "$ETCDCTL get --prefix /vitastor/pg/config --print-value-only | jq -s -e '.[0].items["'"'"2"'"'"]'"
+
+build/src/cmd/vitastor-cli --etcd_address $ETCD_URL modify-pool size1pool -s 2 --force
+
 format_green OK

tests/test_nfs.sh

@@ -165,4 +165,24 @@ if ls ./testdata/nfs | grep over1; then false; fi
 [[ "`cat ./testdata/nfs/linked1`" = "BABABA" ]]
 format_green "rename over existing file ok"
 
+# check listing and removal of a bad direntry
+sudo umount ./testdata/nfs/
+build/src/kv/vitastor-kv --etcd_address $ETCD_URL fsmeta set d11/settings.jsonLGNmGn '{"ino": 123}'
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs
+ls -l ./testdata/nfs/settings.jsonLGNmGn
+rm ./testdata/nfs/settings.jsonLGNmGn
+build/src/kv/vitastor-kv --etcd_address $ETCD_URL fsmeta get d11/settings.jsonLGNmGn 2>&1 | grep '(code -2)'
+ls -l ./testdata/nfs
+
+# repeat with ino=0
+sudo umount ./testdata/nfs/
+build/src/kv/vitastor-kv --etcd_address $ETCD_URL fsmeta set d11/settings.jsonLGNmGn '{"ino": 0}'
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs
+ls -l ./testdata/nfs/settings.jsonLGNmGn
+rm ./testdata/nfs/settings.jsonLGNmGn
+build/src/kv/vitastor-kv --etcd_address $ETCD_URL fsmeta get d11/settings.jsonLGNmGn 2>&1 | grep '(code -2)'
+ls -l ./testdata/nfs
+
 format_green OK