Release 0.6.0

Warning: upgrading from 0.5.x is currently not supported!
Please create an issue if you really need upgrade capability.

New features:
- Snapshots and Copy-on-Write clones
- Inode (image) names
- Inode I/O and space statistics
- Write throttling for smoothing random write workloads in SSD+HDD configurations

README-ru.md

@@ -22,6 +22,7 @@ Vitastor на данный момент находится в статусе п
 
 Однако следующее уже реализовано:
 
+0.5.x (стабильная версия):
 - Базовая часть - надёжное кластерное блочное хранилище без единой точки отказа
 - Производительность ;-D
 - Несколько схем отказоустойчивости: репликация, XOR n+1 (1 диск чётности), коды коррекции ошибок
@@ -42,20 +43,23 @@ Vitastor на данный момент находится в статусе п
 - NBD-прокси для монтирования образов ядром ("блочное устройство в режиме пользователя")
 - Утилита удаления образов/инодов (vitastor-rm)
 - Пакеты для Debian и CentOS
+
+0.6.x (master-ветка):
 - Статистика операций ввода/вывода и занятого места в разрезе инодов
 - Именование инодов через хранение их метаданных в etcd
 - Снапшоты и copy-on-write клоны
+- Сглаживание производительности случайной записи в SSD+HDD конфигурациях
 
-## Планы разработки
+## Планы развития
 
 - Более корректные скрипты разметки дисков и автоматического запуска OSD
 - Другие инструменты администрирования
 - Плагины для OpenStack, Kubernetes, OpenNebula, Proxmox и других облачных систем
 - iSCSI-прокси
-- Таймауты операций и более быстрое выявление отказов
+- Более быстрое переключение при отказах
 - Фоновая проверка целостности без контрольных сумм (сверка реплик)
 - Контрольные суммы
-- Оптимизации для гибридных SSD+HDD хранилищ
+- Поддержка SSD-кэширования (tiered storage)
 - Поддержка RDMA и NVDIMM
 - Web-интерфейс
 - Возможно, сжатие
@@ -359,9 +363,9 @@ Vitastor с однопоточной NBD прокси на том же стен
   так как в 5.4 есть как минимум 1 известный баг, ведущий к зависанию с io_uring и контроллером HP SmartArray.
 - Установите liburing 0.4 или более новый и его заголовки.
 - Установите lp_solve.
-- Установите etcd. Внимание: вам нужна версия с исправлением отсюда: https://github.com/vitalif/etcd/,
-  из ветки release-3.4, так как в etcd есть баг, который [будет](https://github.com/etcd-io/etcd/pull/12402)
-  исправлен только в 3.4.15. Баг приводит к неспособности Vitastor запустить PG, когда их хотя бы 500 штук.
+- Установите etcd, версии не ниже 3.4.15. Более ранние версии работать не будут из-за различных багов,
+  например [#12402](https://github.com/etcd-io/etcd/pull/12402). Также вы можете взять версию 3.4.13 с
+  этим конкретным исправлением из ветки release-3.4 репозитория https://github.com/vitalif/etcd/.
 - Установите node.js 10 или новее.
 - Установите gcc и g++ 8.x или новее.
 - Склонируйте данный репозиторий с подмодулями: `git clone https://yourcmc.ru/git/vitalif/vitastor/`.

README.md

@@ -16,6 +16,7 @@ with configurable redundancy (replication or erasure codes/XOR).
 Vitastor is currently a pre-release, a lot of features are missing and you can still expect
 breaking changes in the future. However, the following is implemented:
 
+0.5.x (stable):
 - Basic part: highly-available block storage with symmetric clustering and no SPOF
 - Performance ;-D
 - Multiple redundancy schemes: Replication, XOR n+1, Reed-Solomon erasure codes
@@ -36,9 +37,12 @@ breaking changes in the future. However, the following is implemented:
 - NBD proxy for kernel mounts
 - Inode removal tool (vitastor-rm)
 - Packaging for Debian and CentOS
+
+0.6.x (master):
 - Per-inode I/O and space usage statistics
 - Inode metadata storage in etcd
 - Snapshots and copy-on-write image clones
+- Write throttling to smooth random write workloads in SSD+HDD configurations
 
 ## Roadmap
 
@@ -46,10 +50,10 @@ breaking changes in the future. However, the following is implemented:
 - Other administrative tools
 - Plugins for OpenStack, Kubernetes, OpenNebula, Proxmox and other cloud systems
 - iSCSI proxy
-- Operation timeouts and better failure detection
+- Faster failover
 - Scrubbing without checksums (verification of replicas)
 - Checksums
-- SSD+HDD optimizations, possibly including tiered storage and soft journal flushes
+- Tiered storage
 - RDMA and NVDIMM support
 - Web GUI
 - Compression (possibly)
@@ -315,10 +319,9 @@ Vitastor with single-thread NBD on the same hardware:
   there is at least one known io_uring hang with 5.4 and an HP SmartArray controller.
 - Install liburing 0.4 or newer and its headers.
 - Install lp_solve.
-- Install etcd. Attention: you need a fixed version from here: https://github.com/vitalif/etcd/,
-  branch release-3.4, because there is a bug in upstream etcd which makes Vitastor OSDs fail to
-  move PGs out of "starting" state if you have at least around ~500 PGs or so. The custom build
-  will be unnecessary when etcd merges the fix: https://github.com/etcd-io/etcd/pull/12402.
+- Install etcd, at least version 3.4.15. Earlier versions won't work because of various bugs,
+  for example [#12402](https://github.com/etcd-io/etcd/pull/12402). You can also take 3.4.13
+  with this specific fix from here: https://github.com/vitalif/etcd/, branch release-3.4.
 - Install node.js 10 or newer.
 - Install gcc and g++ 8.x or newer.
 - Clone https://yourcmc.ru/git/vitalif/vitastor/ with submodules.

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (0.5.10-1) unstable; urgency=medium
+vitastor (0.6.0-1) unstable; urgency=medium
 
   * Bugfixes
 

debian/vitastor.Dockerfile

@@ -40,10 +40,10 @@ RUN set -e -x; \
     mkdir -p /root/packages/vitastor-$REL; \
     rm -rf /root/packages/vitastor-$REL/*; \
     cd /root/packages/vitastor-$REL; \
-    cp -r /root/vitastor vitastor-0.5.10; \
-    ln -s /root/packages/qemu-$REL/qemu-*/ vitastor-0.5.10/qemu; \
-    ln -s /root/fio-build/fio-*/ vitastor-0.5.10/fio; \
-    cd vitastor-0.5.10; \
+    cp -r /root/vitastor vitastor-0.6.0; \
+    ln -s /root/packages/qemu-$REL/qemu-*/ vitastor-0.6.0/qemu; \
+    ln -s /root/fio-build/fio-*/ vitastor-0.6.0/fio; \
+    cd vitastor-0.6.0; \
     FIO=$(head -n1 fio/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     QEMU=$(head -n1 qemu/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     sh copy-qemu-includes.sh; \
@@ -59,8 +59,8 @@ RUN set -e -x; \
     echo "dep:fio=$FIO" > debian/substvars; \
     echo "dep:qemu=$QEMU" >> debian/substvars; \
     cd /root/packages/vitastor-$REL; \
-    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.5.10.orig.tar.xz vitastor-0.5.10; \
-    cd vitastor-0.5.10; \
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_0.6.0.orig.tar.xz vitastor-0.6.0; \
+    cd vitastor-0.6.0; \
     V=$(head -n1 debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     DEBFULLNAME="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v "$V""$REL" "Rebuild for $REL"; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \

mon/lp-optimizer.js

@@ -104,6 +104,17 @@ async function optimize_initial({ osd_tree, pg_count, pg_size = 3, pg_minsize =
     return res;
 }
 
+function shuffle(array)
+{
+    for (let i = array.length - 1, j, x; i > 0; i--)
+    {
+        j = Math.floor(Math.random() * (i + 1));
+        x = array[i];
+        array[i] = array[j];
+        array[j] = x;
+    }
+}
+
 function make_int_pgs(weights, pg_count)
 {
     const total_weight = Object.values(weights).reduce((a, c) => Number(a) + Number(c), 0);
@@ -120,6 +131,7 @@ function make_int_pgs(weights, pg_count)
         weight_left -= weights[pg_name];
         pg_left -= n;
     }
+    shuffle(int_pgs);
     return int_pgs;
 }
 

mon/make-osd.sh

@@ -53,7 +53,6 @@ ExecStart=/usr/bin/vitastor-osd \\
     --osd_num $OSD_NUM \\
     --disable_data_fsync 1 \\
     --immediate_commit all \\
-    --flusher_count 256 \\
     --disk_alignment 4096 --journal_block_size 4096 --meta_block_size 4096 \\
     --journal_no_same_sector_overwrites true \\
     --journal_sector_buffer_count 1024 \\

mon/make-units.sh

@@ -32,7 +32,8 @@ ExecStart=/usr/local/bin/etcd -name etcd$ETCD_NUM --data-dir /var/lib/etcd$ETCD_
     --advertise-client-urls http://$IP:2379 --listen-client-urls http://$IP:2379 \\
     --initial-advertise-peer-urls http://$IP:2380 --listen-peer-urls http://$IP:2380 \\
     --initial-cluster-token vitastor-etcd-1 --initial-cluster $ETCD_HOSTS \\
-    --initial-cluster-state new --max-txn-ops=100000 --auto-compaction-retention=10 --auto-compaction-mode=revision
+    --initial-cluster-state new --max-txn-ops=100000 --max-request-bytes=104857600 \\
+    --auto-compaction-retention=10 --auto-compaction-mode=revision
 WorkingDirectory=/var/lib/etcd$ETCD_NUM.etcd
 ExecStartPre=+chown -R etcd /var/lib/etcd$ETCD_NUM.etcd
 User=etcd

mon/mon.js

@@ -34,7 +34,7 @@ const etcd_allow = new RegExp('^'+[
     'pg/stats/[1-9]\\d*/[1-9]\\d*',
     'pg/history/[1-9]\\d*/[1-9]\\d*',
     'history/last_clean_pgs',
-    'inode/stats/[1-9]\\d*',
+    'inode/stats/[1-9]\\d*/[1-9]\\d*',
     'stats',
 ].join('$|^')+'$');
 
@@ -96,7 +96,8 @@ const etcd_tree = {
             disable_device_lock,
             // blockstore - configurable
             max_write_iodepth,
-            flusher_count,
+            min_flusher_count: 1,
+            max_flusher_count: 256,
             inmemory_metadata,
             inmemory_journal,
             journal_sector_buffer_count,
@@ -210,7 +211,7 @@ const etcd_tree = {
             /* <pool_id>: {
                 <pg_id>: {
                     primary: osd_num_t,
-                    state: ("starting"|"peering"|"incomplete"|"active"|"stopping"|"offline"|
+                    state: ("starting"|"peering"|"incomplete"|"active"|"repeering"|"stopping"|"offline"|
                         "degraded"|"has_incomplete"|"has_degraded"|"has_misplaced"|"has_unclean"|
                         "has_invalid"|"left_on_dead")[],
                 }
@@ -579,7 +580,7 @@ class Mon
         for (const osd_num of this.all_osds().sort((a, b) => a - b))
         {
             const stat = this.state.osd.stats[osd_num];
-            if (stat.size && (this.state.osd.state[osd_num] || Number(stat.time) >= down_time))
+            if (stat && stat.size && (this.state.osd.state[osd_num] || Number(stat.time) >= down_time))
             {
                 // Numeric IDs are reserved for OSDs
                 const osd_cfg = this.state.config.osd[osd_num];
@@ -730,6 +731,11 @@ class Mon
                 pg_history[i].osd_sets = pg_history[i].osd_sets || [];
                 pg_history[i].osd_sets.push(prev_pgs[i]);
             }
+            if (pg_history[i] && pg_history[i].osd_sets)
+            {
+                pg_history[i].osd_sets = Object.values(pg_history[i].osd_sets
+                    .reduce((a, c) => { a[c.join(' ')] = c; return a; }, {}));
+            }
         });
         for (let i = 0; i < new_pgs.length || i < prev_pgs.length; i++)
         {
@@ -880,7 +886,7 @@ class Mon
     {
         // Take configuration and state, check it against the stored configuration hash
         // Recalculate PGs and save them to etcd if the configuration is changed
-        // FIXME: Also do not change anything if the distribution is good enough and no PGs are degraded
+        // FIXME: Do not change anything if the distribution is good and random enough and no PGs are degraded
         const { up_osds, levels, osd_tree } = this.get_osd_tree();
         const tree_cfg = {
             osd_tree,
@@ -939,7 +945,14 @@ class Mon
                     prev_pgs[pg-1] = this.state.history.last_clean_pgs.items[pool_id][pg].osd_set;
                 }
                 prev_pgs = JSON.parse(JSON.stringify(prev_pgs.length ? prev_pgs : real_prev_pgs));
-                const old_pg_count = prev_pgs.length;
+                const old_pg_count = real_prev_pgs.length;
+                const optimize_cfg = {
+                    osd_tree: pool_tree,
+                    pg_count: pool_cfg.pg_count,
+                    pg_size: pool_cfg.pg_size,
+                    pg_minsize: pool_cfg.pg_minsize,
+                    max_combinations: pool_cfg.max_osd_combinations,
+                };
                 let optimize_result;
                 if (old_pg_count > 0)
                 {
@@ -966,23 +979,22 @@ class Mon
                             pg.pop();
                         }
                     }
-                    optimize_result = await LPOptimizer.optimize_change({
-                        prev_pgs,
-                        osd_tree: pool_tree,
-                        pg_size: pool_cfg.pg_size,
-                        pg_minsize: pool_cfg.pg_minsize,
-                        max_combinations: pool_cfg.max_osd_combinations,
-                    });
+                    if (!this.state.config.pgs.hash)
+                    {
+                        // Re-shuffle PGs
+                        optimize_result = await LPOptimizer.optimize_initial(optimize_cfg);
+                    }
+                    else
+                    {
+                        optimize_result = await LPOptimizer.optimize_change({
+                            prev_pgs,
+                            ...optimize_cfg,
+                        });
+                    }
                 }
                 else
                 {
-                    optimize_result = await LPOptimizer.optimize_initial({
-                        osd_tree: pool_tree,
-                        pg_count: pool_cfg.pg_count,
-                        pg_size: pool_cfg.pg_size,
-                        pg_minsize: pool_cfg.pg_minsize,
-                        max_combinations: pool_cfg.max_osd_combinations,
-                    });
+                    optimize_result = await LPOptimizer.optimize_initial(optimize_cfg);
                 }
                 if (old_pg_count != optimize_result.int_pgs.length)
                 {
@@ -1108,7 +1120,7 @@ class Mon
         const op_stats = {}, subop_stats = {}, recovery_stats = {};
         for (const osd in this.state.osd.stats)
         {
-            const st = this.state.osd.stats[osd];
+            const st = this.state.osd.stats[osd]||{};
             for (const op in st.op_stats||{})
             {
                 op_stats[op] = op_stats[op] || { count: 0n, usec: 0n, bytes: 0n };
@@ -1166,23 +1178,31 @@ class Mon
         });
         for (const osd_num in this.state.osd.space)
         {
-            for (const inode_num in this.state.osd.space[osd_num])
+            for (const pool_id in this.state.osd.space[osd_num])
             {
-                inode_stats[inode_num] = inode_stats[inode_num] || inode_stub();
-                inode_stats[inode_num].raw_used += BigInt(this.state.osd.space[osd_num][inode_num]||0);
+                inode_stats[pool_id] = inode_stats[pool_id] || {};
+                for (const inode_num in this.state.osd.space[osd_num][pool_id])
+                {
+                    inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();
+                    inode_stats[pool_id][inode_num].raw_used += BigInt(this.state.osd.space[osd_num][pool_id][inode_num]||0);
+                }
             }
         }
         for (const osd_num in this.state.osd.inodestats)
         {
             const ist = this.state.osd.inodestats[osd_num];
-            for (const inode_num in ist)
+            for (const pool_id in ist)
             {
-                inode_stats[inode_num] = inode_stats[inode_num] || inode_stub();
-                for (const op of [ 'read', 'write', 'delete' ])
+                inode_stats[pool_id] = inode_stats[pool_id] || {};
+                for (const inode_num in ist[pool_id])
                 {
-                    inode_stats[inode_num][op].count += BigInt(ist[inode_num][op].count||0);
-                    inode_stats[inode_num][op].usec += BigInt(ist[inode_num][op].usec||0);
-                    inode_stats[inode_num][op].bytes += BigInt(ist[inode_num][op].bytes||0);
+                    inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();
+                    for (const op of [ 'read', 'write', 'delete' ])
+                    {
+                        inode_stats[pool_id][inode_num][op].count += BigInt(ist[pool_id][inode_num][op].count||0);
+                        inode_stats[pool_id][inode_num][op].usec += BigInt(ist[pool_id][inode_num][op].usec||0);
+                        inode_stats[pool_id][inode_num][op].bytes += BigInt(ist[pool_id][inode_num][op].bytes||0);
+                    }
                 }
             }
         }
@@ -1248,12 +1268,15 @@ class Mon
         this.serialize_bigints(stats);
         this.serialize_bigints(inode_stats);
         txn.push({ requestPut: { key: b64(this.etcd_prefix+'/stats'), value: b64(JSON.stringify(stats)) } });
-        for (const inode_num in inode_stats)
+        for (const pool_id in inode_stats)
         {
-            txn.push({ requestPut: {
-                key: b64(this.etcd_prefix+'/inode/stats/'+inode_num),
-                value: b64(JSON.stringify(inode_stats[inode_num])),
-            } });
+            for (const inode_num in inode_stats[pool_id])
+            {
+                txn.push({ requestPut: {
+                    key: b64(this.etcd_prefix+'/inode/stats/'+pool_id+'/'+inode_num),
+                    value: b64(JSON.stringify(inode_stats[pool_id][inode_num])),
+                } });
+            }
         }
         if (txn.length)
         {

mon/simple-offsets.js

@@ -51,7 +51,7 @@ async function run()
     const meta_offset = options.journal_offset + Math.ceil(options.journal_size/options.device_block_size)*options.device_block_size;
     const entries_per_block = Math.floor(options.device_block_size / (24 + 2*options.object_size/options.bitmap_granularity/8));
     const object_count = Math.floor((device_size-meta_offset)/options.object_size);
-    const meta_size = Math.ceil(object_count / entries_per_block) * options.device_block_size;
+    const meta_size = Math.ceil(1 + object_count / entries_per_block) * options.device_block_size;
     const data_offset = meta_offset + meta_size;
     const meta_size_fmt = (meta_size > 1024*1024*1024 ? Math.round(meta_size/1024/1024/1024*100)/100+" GB"
         : Math.round(meta_size/1024/1024*100)/100+" MB");
@@ -65,6 +65,9 @@ async function run()
             );
         }
         process.stdout.write(
+            (options.device_block_size != 4096 ?
+                `    --meta_block_size ${options.device}\n`+
+                `    --journal_block-size ${options.device}\n` : '')+
             `    --data_device ${options.device}\n`+
             `    --journal_offset ${options.journal_offset}\n`+
             `    --meta_offset ${meta_offset}\n`+

rpm/build-tarball.sh

@@ -48,4 +48,4 @@ FIO=`rpm -qi fio | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Ve
 QEMU=`rpm -qi qemu qemu-kvm | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Version[\s:]+(\S+)/ && print $1; /^Release[\s:]+(\S+)/ && print "-$1"; }'`
 perl -i -pe 's/(Requires:\s*fio)([^\n]+)?/$1 = '$FIO'/' $VITASTOR/rpm/vitastor-el$EL.spec
 perl -i -pe 's/(Requires:\s*qemu(?:-kvm)?)([^\n]+)?/$1 = '$QEMU'/' $VITASTOR/rpm/vitastor-el$EL.spec
-tar --transform 's#^#vitastor-0.5.10/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.5.10$(rpm --eval '%dist').tar.gz *
+tar --transform 's#^#vitastor-0.6.0/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.6.0$(rpm --eval '%dist').tar.gz *

rpm/vitastor-el7.Dockerfile

@@ -37,7 +37,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.5.10.el7.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.6.0.el7.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el7.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.5.10
+Version:        0.6.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-0.5.10.el7.tar.gz
+Source0:        vitastor-0.6.0.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-0.5.10.el8.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-0.6.0.el8.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el8.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        0.5.10
+Version:        0.6.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-0.5.10.el8.tar.gz
+Source0:        vitastor-0.6.0.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -13,8 +13,8 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVERSION="0.6-dev")
-add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith)
+add_definitions(-DVERSION="0.6.0")
+add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -I ${CMAKE_SOURCE_DIR}/src)
 if (${WITH_ASAN})
 	add_definitions(-fsanitize=address -fno-omit-frame-pointer)
 	add_link_options(-fsanitize=address -fno-omit-frame-pointer)
@@ -66,7 +66,8 @@ target_link_libraries(fio_vitastor_blk
 # vitastor-osd
 add_executable(vitastor-osd
 	osd_main.cpp osd.cpp osd_secondary.cpp msgr_receive.cpp msgr_send.cpp osd_peering.cpp osd_flush.cpp osd_peering_pg.cpp
-	osd_primary.cpp osd_primary_subops.cpp etcd_state_client.cpp messenger.cpp osd_cluster.cpp http_client.cpp osd_ops.cpp pg_states.cpp
+	osd_primary.cpp osd_primary_chain.cpp osd_primary_sync.cpp osd_primary_write.cpp osd_primary_subops.cpp
+	etcd_state_client.cpp messenger.cpp msgr_stop.cpp msgr_op.cpp osd_cluster.cpp http_client.cpp osd_ops.cpp pg_states.cpp
 	osd_rmw.cpp base64.cpp timerfd_manager.cpp epoll_manager.cpp ../json11/json11.cpp
 )
 target_link_libraries(vitastor-osd
@@ -86,7 +87,7 @@ target_link_libraries(fio_vitastor_sec
 # libvitastor_client.so
 add_library(vitastor_client SHARED
 	cluster_client.cpp epoll_manager.cpp etcd_state_client.cpp
-	messenger.cpp msgr_send.cpp msgr_receive.cpp ringloop.cpp ../json11/json11.cpp
+	messenger.cpp msgr_stop.cpp msgr_op.cpp msgr_send.cpp msgr_receive.cpp ringloop.cpp ../json11/json11.cpp
 	http_client.cpp osd_ops.cpp pg_states.cpp timerfd_manager.cpp base64.cpp
 )
 target_link_libraries(vitastor_client
@@ -161,7 +162,8 @@ target_link_libraries(osd_rmw_test Jerasure tcmalloc_minimal)
 
 # stub_uring_osd
 add_executable(stub_uring_osd
-	stub_uring_osd.cpp epoll_manager.cpp messenger.cpp msgr_send.cpp msgr_receive.cpp ringloop.cpp timerfd_manager.cpp ../json11/json11.cpp
+	stub_uring_osd.cpp epoll_manager.cpp messenger.cpp msgr_stop.cpp msgr_op.cpp
+	msgr_send.cpp msgr_receive.cpp ringloop.cpp timerfd_manager.cpp ../json11/json11.cpp
 )
 target_link_libraries(stub_uring_osd
 	${LIBURING_LIBRARIES}
@@ -175,8 +177,17 @@ target_link_libraries(osd_peering_pg_test tcmalloc_minimal)
 # test_allocator
 add_executable(test_allocator test_allocator.cpp allocator.cpp)
 
+# test_cluster_client
+add_executable(test_cluster_client
+	test_cluster_client.cpp
+	pg_states.cpp osd_ops.cpp cluster_client.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
+	etcd_state_client.cpp timerfd_manager.cpp ../json11/json11.cpp
+)
+target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)
+target_include_directories(test_cluster_client PUBLIC ${CMAKE_SOURCE_DIR}/src/mock)
+
 ## test_blockstore, test_shit
-#add_executable(test_blockstore test_blockstore.cpp timerfd_interval.cpp)
+#add_executable(test_blockstore test_blockstore.cpp)
 #target_link_libraries(test_blockstore blockstore)
 #add_executable(test_shit test_shit.cpp osd_peering_pg.cpp)
 #target_link_libraries(test_shit ${LIBURING_LIBRARIES} m)

src/allocator.cpp

@@ -37,6 +37,21 @@ allocator::~allocator()
     delete[] mask;
 }
 
+bool allocator::get(uint64_t addr)
+{
+    if (addr >= size)
+    {
+        return false;
+    }
+    uint64_t p2 = 1, offset = 0;
+    while (p2 * 64 < size)
+    {
+        offset += p2;
+        p2 = p2 * 64;
+    }
+    return ((mask[offset + addr/64] >> (addr % 64)) & 1);
+}
+
 void allocator::set(uint64_t addr, bool value)
 {
     if (addr >= size)

src/allocator.h

@@ -16,6 +16,7 @@ class allocator
 public:
     allocator(uint64_t blocks);
     ~allocator();
+    bool get(uint64_t addr);
     void set(uint64_t addr, bool value);
     uint64_t find_free();
     uint64_t get_free_count();

src/blockstore.cpp

@@ -3,9 +3,9 @@
 
 #include "blockstore_impl.h"
 
-blockstore_t::blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop)
+blockstore_t::blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd)
 {
-    impl = new blockstore_impl_t(config, ringloop);
+    impl = new blockstore_impl_t(config, ringloop, tfd);
 }
 
 blockstore_t::~blockstore_t()
@@ -38,6 +38,11 @@ void blockstore_t::enqueue_op(blockstore_op_t *op)
     impl->enqueue_op(op);
 }
 
+int blockstore_t::read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version)
+{
+    return impl->read_bitmap(oid, target_version, bitmap, result_version);
+}
+
 std::unordered_map<object_id, uint64_t> & blockstore_t::get_unstable_writes()
 {
     return impl->unstable_writes;

src/blockstore.h

@@ -16,6 +16,7 @@
 
 #include "object_id.h"
 #include "ringloop.h"
+#include "timerfd_manager.h"
 
 // Memory alignment for direct I/O (usually 512 bytes)
 // All other alignments must be a multiple of this one
@@ -158,7 +159,7 @@ class blockstore_t
 {
     blockstore_impl_t *impl;
 public:
-    blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop);
+    blockstore_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd);
     ~blockstore_t();
 
     // Event loop
@@ -179,6 +180,9 @@ public:
     // Submission
     void enqueue_op(blockstore_op_t *op);
 
+    // Simplified synchronous operation: get object bitmap & current version
+    int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL);
+
     // Unstable writes are added here (map of object_id -> version)
     std::unordered_map<object_id, uint64_t> & get_unstable_writes();
 

src/blockstore_flush.cpp

@@ -3,12 +3,13 @@
 
 #include "blockstore_impl.h"
 
-journal_flusher_t::journal_flusher_t(int flusher_count, blockstore_impl_t *bs)
+journal_flusher_t::journal_flusher_t(blockstore_impl_t *bs)
 {
     this->bs = bs;
-    this->flusher_count = flusher_count;
-    this->cur_flusher_count = 1;
-    this->target_flusher_count = 1;
+    this->max_flusher_count = bs->max_flusher_count;
+    this->min_flusher_count = bs->min_flusher_count;
+    this->cur_flusher_count = bs->min_flusher_count;
+    this->target_flusher_count = bs->min_flusher_count;
     dequeuing = false;
     trimming = false;
     active_flushers = 0;
@@ -16,11 +17,11 @@ journal_flusher_t::journal_flusher_t(int flusher_count, blockstore_impl_t *bs)
     // FIXME: allow to configure flusher_start_threshold and journal_trim_interval
     flusher_start_threshold = bs->journal_block_size / sizeof(journal_entry_stable);
     journal_trim_interval = 512;
-    journal_trim_counter = 0;
-    trim_wanted = 0;
+    journal_trim_counter = bs->journal.flush_journal ? 1 : 0;
+    trim_wanted = bs->journal.flush_journal ? 1 : 0;
     journal_superblock = bs->journal.inmemory ? bs->journal.buffer : memalign_or_die(MEM_ALIGNMENT, bs->journal_block_size);
-    co = new journal_flusher_co[flusher_count];
-    for (int i = 0; i < flusher_count; i++)
+    co = new journal_flusher_co[max_flusher_count];
+    for (int i = 0; i < max_flusher_count; i++)
     {
         co[i].bs = bs;
         co[i].flusher = this;
@@ -71,10 +72,10 @@ bool journal_flusher_t::is_active()
 void journal_flusher_t::loop()
 {
     target_flusher_count = bs->write_iodepth*2;
-    if (target_flusher_count <= 0)
-        target_flusher_count = 1;
-    else if (target_flusher_count > flusher_count)
-        target_flusher_count = flusher_count;
+    if (target_flusher_count < min_flusher_count)
+        target_flusher_count = min_flusher_count;
+    else if (target_flusher_count > max_flusher_count)
+        target_flusher_count = max_flusher_count;
     if (target_flusher_count > cur_flusher_count)
         cur_flusher_count = target_flusher_count;
     else if (target_flusher_count < cur_flusher_count)
@@ -237,7 +238,8 @@ bool journal_flusher_co::loop()
     else if (wait_state == 21)
         goto resume_21;
 resume_0:
-    if (!flusher->flush_queue.size() || !flusher->dequeuing)
+    if (flusher->flush_queue.size() < flusher->min_flusher_count && !flusher->trim_wanted ||
+        !flusher->flush_queue.size() || !flusher->dequeuing)
     {
 stop_flusher:
         if (flusher->trim_wanted > 0 && flusher->journal_trim_counter > 0)
@@ -483,6 +485,13 @@ resume_1:
         }
         if (has_delete)
         {
+            clean_disk_entry *new_entry = (clean_disk_entry*)(meta_new.buf + meta_new.pos*bs->clean_entry_size);
+            if (new_entry->oid.inode != 0 && new_entry->oid != cur.oid)
+            {
+                printf("Fatal error (metadata corruption or bug): tried to delete metadata entry %lu (%lx:%lx) while deleting %lx:%lx\n",
+                    clean_loc >> bs->block_order, new_entry->oid.inode, new_entry->oid.stripe, cur.oid.inode, cur.oid.stripe);
+                exit(1);
+            }
             // zero out new metadata entry
             memset(meta_new.buf + meta_new.pos*bs->clean_entry_size, 0, bs->clean_entry_size);
         }
@@ -593,6 +602,7 @@ resume_1:
                     .size = sizeof(journal_entry_start),
                     .reserved = 0,
                     .journal_start = new_trim_pos,
+                    .version = JOURNAL_VERSION,
                 };
                 ((journal_entry_start*)flusher->journal_superblock)->crc32 = je_crc32((journal_entry*)flusher->journal_superblock);
                 data->iov = (struct iovec){ flusher->journal_superblock, bs->journal_block_size };
@@ -624,6 +634,12 @@ resume_1:
 #endif
                 flusher->trimming = false;
             }
+            if (bs->journal.flush_journal && !flusher->flush_queue.size())
+            {
+                assert(bs->journal.used_start == bs->journal.next_free);
+                printf("Journal flushed\n");
+                exit(0);
+            }
         }
         // All done
         flusher->active_flushers--;
@@ -654,7 +670,7 @@ bool journal_flusher_co::scan_dirty(int wait_base)
         {
             char err[1024];
             snprintf(
-                err, 1024, "BUG: Unexpected dirty_entry %lx:%lx v%lu unstable state during flush: %d",
+                err, 1024, "BUG: Unexpected dirty_entry %lx:%lx v%lu unstable state during flush: 0x%x",
                 dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version, dirty_it->second.state
             );
             throw std::runtime_error(err);
@@ -783,7 +799,10 @@ void journal_flusher_co::update_clean_db()
     if (old_clean_loc != UINT64_MAX && old_clean_loc != clean_loc)
     {
 #ifdef BLOCKSTORE_DEBUG
-        printf("Free block %lu (new location is %lu)\n", old_clean_loc >> bs->block_order, clean_loc >> bs->block_order);
+        printf("Free block %lu from %lx:%lx v%lu (new location is %lu)\n",
+            old_clean_loc >> bs->block_order,
+            cur.oid.inode, cur.oid.stripe, cur.version,
+            clean_loc >> bs->block_order);
 #endif
         bs->data_alloc->set(old_clean_loc >> bs->block_order, false);
     }
@@ -791,6 +810,11 @@ void journal_flusher_co::update_clean_db()
     {
         auto clean_it = bs->clean_db.find(cur.oid);
         bs->clean_db.erase(clean_it);
+#ifdef BLOCKSTORE_DEBUG
+        printf("Free block %lu from %lx:%lx v%lu (delete)\n",
+            clean_loc >> bs->block_order,
+            cur.oid.inode, cur.oid.stripe, cur.version);
+#endif
         bs->data_alloc->set(clean_loc >> bs->block_order, false);
         clean_loc = UINT64_MAX;
     }
@@ -812,7 +836,7 @@ bool journal_flusher_co::fsync_batch(bool fsync_meta, int wait_base)
         goto resume_1;
     else if (wait_state == wait_base+2)
         goto resume_2;
-    if (!(fsync_meta ? bs->disable_meta_fsync : bs->disable_journal_fsync))
+    if (!(fsync_meta ? bs->disable_meta_fsync : bs->disable_data_fsync))
     {
         cur_sync = flusher->syncs.end();
         while (cur_sync != flusher->syncs.begin())

src/blockstore_flush.h

@@ -79,7 +79,7 @@ class journal_flusher_t
 {
     int trim_wanted = 0;
     bool dequeuing;
-    int flusher_count, cur_flusher_count, target_flusher_count;
+    int min_flusher_count, max_flusher_count, cur_flusher_count, target_flusher_count;
     int flusher_start_threshold;
     journal_flusher_co *co;
     blockstore_impl_t *bs;
@@ -98,7 +98,7 @@ class journal_flusher_t
     std::deque<object_id> flush_queue;
     std::map<object_id, uint64_t> flush_versions;
 public:
-    journal_flusher_t(int flusher_count, blockstore_impl_t *bs);
+    journal_flusher_t(blockstore_impl_t *bs);
     ~journal_flusher_t();
     void loop();
     bool is_active();

src/blockstore_impl.cpp

@@ -3,9 +3,10 @@
 
 #include "blockstore_impl.h"
 
-blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop)
+blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd)
 {
     assert(sizeof(blockstore_op_private_t) <= BS_OP_PRIVATE_DATA_SIZE);
+    this->tfd = tfd;
     this->ringloop = ringloop;
     ring_consumer.loop = [this]() { loop(); };
     ringloop->register_consumer(&ring_consumer);
@@ -31,7 +32,7 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *
             close(journal.fd);
         throw;
     }
-    flusher = new journal_flusher_t(flusher_count, this);
+    flusher = new journal_flusher_t(this);
 }
 
 blockstore_impl_t::~blockstore_impl_t()
@@ -92,10 +93,23 @@ void blockstore_impl_t::loop()
             {
                 delete journal_init_reader;
                 journal_init_reader = NULL;
-                initialized = 10;
+                if (journal.flush_journal)
+                    initialized = 3;
+                else
+                    initialized = 10;
                 ringloop->wakeup();
             }
         }
+        if (initialized == 3)
+        {
+            if (readonly)
+            {
+                printf("Can't flush the journal in readonly mode\n");
+                exit(1);
+            }
+            flusher->loop();
+            ringloop->submit();
+        }
     }
     else
     {
@@ -443,7 +457,7 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
         }
         for (; clean_it != clean_end; clean_it++)
         {
-            if (!pg_count || ((clean_it->first.inode + clean_it->first.stripe / pg_stripe_size) % pg_count) == list_pg)
+            if (!pg_count || ((clean_it->first.stripe / pg_stripe_size) % pg_count) == list_pg) // like map_to_pg()
             {
                 if (stable_count >= stable_alloc)
                 {
@@ -488,7 +502,7 @@ void blockstore_impl_t::process_list(blockstore_op_t *op)
         }
         for (; dirty_it != dirty_end; dirty_it++)
         {
-            if (!pg_count || ((dirty_it->first.oid.inode + dirty_it->first.oid.stripe / pg_stripe_size) % pg_count) == list_pg)
+            if (!pg_count || ((dirty_it->first.oid.stripe / pg_stripe_size) % pg_count) == list_pg) // like map_to_pg()
             {
                 if (IS_DELETE(dirty_it->second.state))
                 {

src/blockstore_impl.h

@@ -9,6 +9,7 @@
 #include <sys/ioctl.h>
 #include <sys/stat.h>
 #include <fcntl.h>
+#include <time.h>
 #include <unistd.h>
 #include <linux/fs.h>
 
@@ -77,6 +78,23 @@
 
 #include "blockstore_journal.h"
 
+// "VITAstor"
+#define BLOCKSTORE_META_MAGIC 0x726F747341544956l
+#define BLOCKSTORE_META_VERSION 1
+
+// metadata header (superblock)
+// FIXME: After adding the OSD superblock, add a key to metadata
+// and journal headers to check if they belong to the same OSD
+struct __attribute__((__packed__)) blockstore_meta_header_t
+{
+    uint64_t zero;
+    uint64_t magic;
+    uint64_t version;
+    uint32_t meta_block_size;
+    uint32_t data_block_size;
+    uint32_t bitmap_granularity;
+};
+
 // 32 bytes = 24 bytes + block bitmap (4 bytes by default) + external attributes (also bitmap, 4 bytes by default)
 // per "clean" entry on disk with fixed metadata tables
 // FIXME: maybe add crc32's to metadata
@@ -158,6 +176,7 @@ struct blockstore_op_private_t
     struct iovec iov_zerofill[3];
     // Warning: must not have a default value here because it's written to before calling constructor in blockstore_write.cpp O_o
     uint64_t real_version;
+    timespec tv_begin;
 
     // Sync
     std::vector<obj_ver_id> sync_big_writes, sync_small_writes;
@@ -199,10 +218,18 @@ class blockstore_impl_t
     // Suitable only for server SSDs with capacitors, requires disabled data and journal fsyncs
     int immediate_commit = IMMEDIATE_NONE;
     bool inmemory_meta = false;
-    // Maximum flusher count
-    unsigned flusher_count;
+    // Maximum and minimum flusher count
+    unsigned max_flusher_count, min_flusher_count;
     // Maximum queue depth
     unsigned max_write_iodepth = 128;
+    // Enable small (journaled) write throttling, useful for the SSD+HDD case
+    bool throttle_small_writes = false;
+    // Target data device iops, bandwidth and parallelism for throttling (100/100/1 is the default for HDD)
+    int throttle_target_iops = 100;
+    int throttle_target_mbs = 100;
+    int throttle_target_parallelism = 1;
+    // Minimum difference in microseconds between target and real execution times to throttle the response
+    int throttle_threshold_us = 50;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;
@@ -212,6 +239,7 @@ class blockstore_impl_t
     blockstore_dirty_db_t dirty_db;
     std::vector<blockstore_op_t*> submit_queue;
     std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
+    int unsynced_big_write_count = 0;
     allocator *data_alloc = NULL;
     uint8_t *zero_object;
 
@@ -232,6 +260,7 @@ class blockstore_impl_t
 
     bool live = false, queue_stall = false;
     ring_loop_t *ringloop;
+    timerfd_manager_t *tfd;
 
     bool stop_sync_submitted;
 
@@ -286,7 +315,7 @@ class blockstore_impl_t
     // Stabilize
     int dequeue_stable(blockstore_op_t *op);
     int continue_stable(blockstore_op_t *op);
-    void mark_stable(const obj_ver_id & ov);
+    void mark_stable(const obj_ver_id & ov, bool forget_dirty = false);
     void handle_stable_event(ring_data_t *data, blockstore_op_t *op);
     void stabilize_object(object_id oid, uint64_t max_ver);
 
@@ -302,7 +331,7 @@ class blockstore_impl_t
 
 public:
 
-    blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop);
+    blockstore_impl_t(blockstore_config_t & config, ring_loop_t *ringloop, timerfd_manager_t *tfd);
     ~blockstore_impl_t();
 
     // Event loop
@@ -323,6 +352,9 @@ public:
     // Submission
     void enqueue_op(blockstore_op_t *op);
 
+    // Simplified synchronous operation: get object bitmap & current version
+    int read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version = NULL);
+
     // Unstable writes are added here (map of object_id -> version)
     std::unordered_map<object_id, uint64_t> unstable_writes;
 

src/blockstore_init.cpp

@@ -3,6 +3,20 @@
 
 #include "blockstore_impl.h"
 
+#define GET_SQE() \
+    sqe = bs->get_sqe();\
+    if (!sqe)\
+        throw std::runtime_error("io_uring is full during initialization");\
+    data = ((ring_data_t*)sqe->user_data)
+
+static bool iszero(uint64_t *buf, int len)
+{
+    for (int i = 0; i < len; i++)
+        if (buf[i] != 0)
+            return false;
+    return true;
+}
+
 blockstore_init_meta::blockstore_init_meta(blockstore_impl_t *bs)
 {
     this->bs = bs;
@@ -10,7 +24,7 @@ blockstore_init_meta::blockstore_init_meta(blockstore_impl_t *bs)
 
 void blockstore_init_meta::handle_event(ring_data_t *data)
 {
-    if (data->res <= 0)
+    if (data->res < 0)
     {
         throw std::runtime_error(
             std::string("read metadata failed at offset ") + std::to_string(metadata_read) +
@@ -28,6 +42,12 @@ int blockstore_init_meta::loop()
 {
     if (wait_state == 1)
         goto resume_1;
+    else if (wait_state == 2)
+        goto resume_2;
+    else if (wait_state == 3)
+        goto resume_3;
+    else if (wait_state == 4)
+        goto resume_4;
     printf("Reading blockstore metadata\n");
     if (bs->inmemory_meta)
         metadata_buffer = bs->metadata_buffer;
@@ -35,22 +55,98 @@ int blockstore_init_meta::loop()
         metadata_buffer = memalign(MEM_ALIGNMENT, 2*bs->metadata_buf_size);
     if (!metadata_buffer)
         throw std::runtime_error("Failed to allocate metadata read buffer");
+    // Read superblock
+    GET_SQE();
+    data->iov = { metadata_buffer, bs->meta_block_size };
+    data->callback = [this](ring_data_t *data) { handle_event(data); };
+    my_uring_prep_readv(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset);
+    bs->ringloop->submit();
+    submitted = 1;
+resume_1:
+    if (submitted)
+    {
+        wait_state = 1;
+        return 1;
+    }
+    if (iszero((uint64_t*)metadata_buffer, bs->meta_block_size / sizeof(uint64_t)))
+    {
+        {
+            blockstore_meta_header_t *hdr = (blockstore_meta_header_t *)metadata_buffer;
+            hdr->zero = 0;
+            hdr->magic = BLOCKSTORE_META_MAGIC;
+            hdr->version = BLOCKSTORE_META_VERSION;
+            hdr->meta_block_size = bs->meta_block_size;
+            hdr->data_block_size = bs->block_size;
+            hdr->bitmap_granularity = bs->bitmap_granularity;
+        }
+        if (bs->readonly)
+        {
+            printf("Skipping metadata initialization because blockstore is readonly\n");
+        }
+        else
+        {
+            printf("Initializing metadata area\n");
+            GET_SQE();
+            data->iov = (struct iovec){ metadata_buffer, bs->meta_block_size };
+            data->callback = [this](ring_data_t *data) { handle_event(data); };
+            my_uring_prep_writev(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset);
+            bs->ringloop->submit();
+            submitted = 1;
+        resume_3:
+            if (submitted > 0)
+            {
+                wait_state = 3;
+                return 1;
+            }
+            zero_on_init = true;
+        }
+    }
+    else
+    {
+        blockstore_meta_header_t *hdr = (blockstore_meta_header_t *)metadata_buffer;
+        if (hdr->zero != 0 ||
+            hdr->magic != BLOCKSTORE_META_MAGIC ||
+            hdr->version != BLOCKSTORE_META_VERSION)
+        {
+            printf(
+                "Metadata is corrupt or old version.\n"
+                " If this is a new OSD please zero out the metadata area before starting it.\n"
+                " If you need to upgrade from 0.5.x please request it via the issue tracker.\n"
+            );
+            exit(1);
+        }
+        if (hdr->meta_block_size != bs->meta_block_size ||
+            hdr->data_block_size != bs->block_size ||
+            hdr->bitmap_granularity != bs->bitmap_granularity)
+        {
+            printf(
+                "Configuration stored in metadata superblock"
+                " (meta_block_size=%u, data_block_size=%u, bitmap_granularity=%u)"
+                " differs from OSD configuration (%lu/%u/%lu).\n",
+                hdr->meta_block_size, hdr->data_block_size, hdr->bitmap_granularity,
+                bs->meta_block_size, bs->block_size, bs->bitmap_granularity
+            );
+            exit(1);
+        }
+    }
+    // Skip superblock
+    bs->meta_offset += bs->meta_block_size;
+    prev_done = 0;
+    done_len = 0;
+    done_pos = 0;
+    metadata_read = 0;
+    // Read the rest of the metadata
     while (1)
     {
-    resume_1:
+    resume_2:
         if (submitted)
         {
-            wait_state = 1;
+            wait_state = 2;
             return 1;
         }
         if (metadata_read < bs->meta_len)
         {
-            sqe = bs->get_sqe();
-            if (!sqe)
-            {
-                throw std::runtime_error("io_uring is full while trying to read metadata");
-            }
-            data = ((ring_data_t*)sqe->user_data);
+            GET_SQE();
             data->iov = {
                 metadata_buffer + (bs->inmemory_meta
                     ? metadata_read
@@ -58,7 +154,14 @@ int blockstore_init_meta::loop()
                 bs->meta_len - metadata_read > bs->metadata_buf_size ? bs->metadata_buf_size : bs->meta_len - metadata_read,
             };
             data->callback = [this](ring_data_t *data) { handle_event(data); };
-            my_uring_prep_readv(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + metadata_read);
+            if (!zero_on_init)
+                my_uring_prep_readv(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + metadata_read);
+            else
+            {
+                // Fill metadata with zeroes
+                memset(data->iov.iov_base, 0, data->iov.iov_len);
+                my_uring_prep_writev(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + metadata_read);
+            }
             bs->ringloop->submit();
             submitted = (prev == 1 ? 2 : 1);
             prev = submitted;
@@ -90,6 +193,21 @@ int blockstore_init_meta::loop()
         free(metadata_buffer);
         metadata_buffer = NULL;
     }
+    if (zero_on_init && !bs->disable_meta_fsync)
+    {
+        GET_SQE();
+        my_uring_prep_fsync(sqe, bs->meta_fd, IORING_FSYNC_DATASYNC);
+        data->iov = { 0 };
+        data->callback = [this](ring_data_t *data) { handle_event(data); };
+        submitted = 1;
+        bs->ringloop->submit();
+    resume_4:
+        if (submitted > 0)
+        {
+            wait_state = 4;
+            return 1;
+        }
+    }
     return 0;
 }
 
@@ -111,7 +229,10 @@ void blockstore_init_meta::handle_entries(void* entries, unsigned count, int blo
                 {
                     // free the previous block
 #ifdef BLOCKSTORE_DEBUG
-                    printf("Free block %lu (new location is %lu)\n", clean_it->second.location >> block_order, done_cnt+i);
+                    printf("Free block %lu from %lx:%lx v%lu (new location is %lu)\n",
+                        clean_it->second.location >> block_order,
+                        clean_it->first.inode, clean_it->first.stripe, clean_it->second.version,
+                        done_cnt+i);
 #endif
                     bs->data_alloc->set(clean_it->second.location >> block_order, false);
                 }
@@ -153,14 +274,6 @@ blockstore_init_journal::blockstore_init_journal(blockstore_impl_t *bs)
     };
 }
 
-bool iszero(uint64_t *buf, int len)
-{
-    for (int i = 0; i < len; i++)
-        if (buf[i] != 0)
-            return false;
-    return true;
-}
-
 void blockstore_init_journal::handle_event(ring_data_t *data1)
 {
     if (data1->res <= 0)
@@ -185,12 +298,6 @@ void blockstore_init_journal::handle_event(ring_data_t *data1)
     submitted_buf = NULL;
 }
 
-#define GET_SQE() \
-    sqe = bs->get_sqe();\
-    if (!sqe)\
-        throw std::runtime_error("io_uring is full while trying to read journal");\
-    data = ((ring_data_t*)sqe->user_data)
-
 int blockstore_init_journal::loop()
 {
     if (wait_state == 1)
@@ -228,7 +335,7 @@ resume_1:
         wait_state = 1;
         return 1;
     }
-    if (iszero((uint64_t*)submitted_buf, 3))
+    if (iszero((uint64_t*)submitted_buf, bs->journal.block_size / sizeof(uint64_t)))
     {
         // Journal is empty
         // FIXME handle this wrapping to journal_block_size better (maybe)
@@ -243,6 +350,7 @@ resume_1:
             .size = sizeof(journal_entry_start),
             .reserved = 0,
             .journal_start = bs->journal.block_size,
+            .version = JOURNAL_VERSION,
         };
         ((journal_entry_start*)submitted_buf)->crc32 = je_crc32((journal_entry*)submitted_buf);
         if (bs->readonly)
@@ -293,11 +401,21 @@ resume_1:
         je_start = (journal_entry_start*)submitted_buf;
         if (je_start->magic != JOURNAL_MAGIC ||
             je_start->type != JE_START ||
-            je_start->size != sizeof(journal_entry_start) ||
-            je_crc32((journal_entry*)je_start) != je_start->crc32)
+            je_crc32((journal_entry*)je_start) != je_start->crc32 ||
+            je_start->size != sizeof(journal_entry_start) && je_start->size != JE_START_LEGACY_SIZE)
         {
             // Entry is corrupt
-            throw std::runtime_error("first entry of the journal is corrupt");
+            fprintf(stderr, "First entry of the journal is corrupt\n");
+            exit(1);
+        }
+        if (je_start->size == JE_START_LEGACY_SIZE || je_start->version != JOURNAL_VERSION)
+        {
+            fprintf(
+                stderr, "The code only supports journal version %d, but it is %lu on disk."
+                    " Please use the previous version to flush the journal before upgrading OSD\n",
+                JOURNAL_VERSION, je_start->size == JE_START_LEGACY_SIZE ? 0 : je_start->version
+            );
+            exit(1);
         }
         next_free = journal_pos = bs->journal.used_start = je_start->journal_start;
         if (!bs->journal.inmemory)
@@ -403,6 +521,18 @@ resume_1:
             }
         }
     }
+    for (auto ov: double_allocs)
+    {
+        auto dirty_it = bs->dirty_db.find(ov);
+        if (dirty_it != bs->dirty_db.end() &&
+            IS_BIG_WRITE(dirty_it->second.state) &&
+            dirty_it->second.location == UINT64_MAX)
+        {
+            printf("Fatal error (bug): %lx:%lx v%lu big_write journal_entry was allocated over another object\n",
+                dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version);
+            exit(1);
+        }
+    }
     bs->flusher->mark_trim_possible();
     bs->journal.dirty_start = bs->journal.next_free;
     printf(
@@ -534,20 +664,20 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         .oid = je->small_write.oid,
                         .version = je->small_write.version,
                     };
-                    void *bmp = (void*)je + sizeof(journal_entry_small_write);
+                    void *bmp = NULL;
+                    void *bmp_from = (void*)je + sizeof(journal_entry_small_write);
                     if (bs->clean_entry_bitmap_size <= sizeof(void*))
                     {
-                        memcpy(&bmp, bmp, bs->clean_entry_bitmap_size);
+                        memcpy(&bmp, bmp_from, bs->clean_entry_bitmap_size);
                     }
-                    else if (!bs->journal.inmemory)
+                    else
                     {
-                        // FIXME Using large blockstore objects and not keeping journal in memory
-                        // will result in a lot of small allocations for entry bitmaps. This can
-                        // only be fixed by using a patched map with dynamic entry size, but not
-                        // the btree_map, because it doesn't keep iterators valid all the time.
-                        void *bmp_cp = malloc_or_die(bs->clean_entry_bitmap_size);
-                        memcpy(bmp_cp, bmp, bs->clean_entry_bitmap_size);
-                        bmp = bmp_cp;
+                        // FIXME Using large blockstore objects will result in a lot of small
+                        // allocations for entry bitmaps. This can only be fixed by using
+                        // a patched map with dynamic entry size, but not the btree_map,
+                        // because it doesn't keep iterators valid all the time.
+                        bmp = malloc_or_die(bs->clean_entry_bitmap_size);
+                        memcpy(bmp, bmp_from, bs->clean_entry_bitmap_size);
                     }
                     bs->dirty_db.emplace(ov, (dirty_entry){
                         .state = (BS_ST_SMALL_WRITE | BS_ST_SYNCED),
@@ -569,7 +699,7 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     unstab = unstab < ov.version ? ov.version : unstab;
                     if (je->type == JE_SMALL_WRITE_INSTANT)
                     {
-                        bs->mark_stable(ov);
+                        bs->mark_stable(ov, true);
                     }
                 }
             }
@@ -599,32 +729,10 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         // its data and metadata are already flushed.
                         // We don't know if newer versions are flushed, but
                         // the previous delete definitely is.
-                        // So we flush previous dirty entries, but retain the clean one.
+                        // So we forget previous dirty entries, but retain the clean one.
                         // This feature is required for writes happening shortly
                         // after deletes.
-                        auto dirty_end = dirty_it;
-                        dirty_end++;
-                        while (1)
-                        {
-                            if (dirty_it == bs->dirty_db.begin())
-                            {
-                                break;
-                            }
-                            dirty_it--;
-                            if (dirty_it->first.oid != je->big_write.oid)
-                            {
-                                dirty_it++;
-                                break;
-                            }
-                        }
-                        auto clean_it = bs->clean_db.find(je->big_write.oid);
-                        bs->erase_dirty(
-                            dirty_it, dirty_end,
-                            clean_it != bs->clean_db.end() ? clean_it->second.location : UINT64_MAX
-                        );
-                        // Remove it from the flusher's queue, too
-                        // Otherwise it may end up referring to a small unstable write after reading the rest of the journal
-                        bs->flusher->remove_flush(je->big_write.oid);
+                        erase_dirty_object(dirty_it);
                     }
                 }
                 auto clean_it = bs->clean_db.find(je->big_write.oid);
@@ -636,22 +744,22 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         .oid = je->big_write.oid,
                         .version = je->big_write.version,
                     };
-                    void *bmp = (void*)je + sizeof(journal_entry_big_write);
+                    void *bmp = NULL;
+                    void *bmp_from = (void*)je + sizeof(journal_entry_big_write);
                     if (bs->clean_entry_bitmap_size <= sizeof(void*))
                     {
-                        memcpy(&bmp, bmp, bs->clean_entry_bitmap_size);
+                        memcpy(&bmp, bmp_from, bs->clean_entry_bitmap_size);
                     }
-                    else if (!bs->journal.inmemory)
+                    else
                     {
-                        // FIXME Using large blockstore objects and not keeping journal in memory
-                        // will result in a lot of small allocations for entry bitmaps. This can
-                        // only be fixed by using a patched map with dynamic entry size, but not
-                        // the btree_map, because it doesn't keep iterators valid all the time.
-                        void *bmp_cp = malloc_or_die(bs->clean_entry_bitmap_size);
-                        memcpy(bmp_cp, bmp, bs->clean_entry_bitmap_size);
-                        bmp = bmp_cp;
+                        // FIXME Using large blockstore objects will result in a lot of small
+                        // allocations for entry bitmaps. This can only be fixed by using
+                        // a patched map with dynamic entry size, but not the btree_map,
+                        // because it doesn't keep iterators valid all the time.
+                        bmp = malloc_or_die(bs->clean_entry_bitmap_size);
+                        memcpy(bmp, bmp_from, bs->clean_entry_bitmap_size);
                     }
-                    bs->dirty_db.emplace(ov, (dirty_entry){
+                    auto dirty_it = bs->dirty_db.emplace(ov, (dirty_entry){
                         .state = (BS_ST_BIG_WRITE | BS_ST_SYNCED),
                         .flags = 0,
                         .location = je->big_write.location,
@@ -659,11 +767,26 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         .len = je->big_write.len,
                         .journal_sector = proc_pos,
                         .bitmap = bmp,
-                    });
+                    }).first;
+                    if (bs->data_alloc->get(je->big_write.location >> bs->block_order))
+                    {
+                        // This is probably a big_write that's already flushed and freed, but it may
+                        // also indicate a bug. So we remember such entries and recheck them afterwards.
+                        // If it's not a bug they won't be present after reading the whole journal.
+                        dirty_it->second.location = UINT64_MAX;
+                        double_allocs.push_back(ov);
+                    }
+                    else
+                    {
 #ifdef BLOCKSTORE_DEBUG
-                    printf("Allocate block %lu\n", je->big_write.location >> bs->block_order);
+                        printf(
+                            "Allocate block (journal) %lu: %lx:%lx v%lu\n",
+                            je->big_write.location >> bs->block_order,
+                            ov.oid.inode, ov.oid.stripe, ov.version
+                        );
 #endif
-                    bs->data_alloc->set(je->big_write.location >> bs->block_order, true);
+                        bs->data_alloc->set(je->big_write.location >> bs->block_order, true);
+                    }
                     bs->journal.used_sectors[proc_pos]++;
 #ifdef BLOCKSTORE_DEBUG
                     printf(
@@ -675,7 +798,7 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     unstab = unstab < ov.version ? ov.version : unstab;
                     if (je->type == JE_BIG_WRITE_INSTANT)
                     {
-                        bs->mark_stable(ov);
+                        bs->mark_stable(ov, true);
                     }
                 }
             }
@@ -689,7 +812,7 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     .oid = je->stable.oid,
                     .version = je->stable.version,
                 };
-                bs->mark_stable(ov);
+                bs->mark_stable(ov, true);
             }
             else if (je->type == JE_ROLLBACK)
             {
@@ -708,9 +831,26 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
 #ifdef BLOCKSTORE_DEBUG
                 printf("je_delete oid=%lx:%lx ver=%lu\n", je->del.oid.inode, je->del.oid.stripe, je->del.version);
 #endif
+                bool dirty_exists = false;
+                auto dirty_it = bs->dirty_db.upper_bound((obj_ver_id){
+                    .oid = je->del.oid,
+                    .version = UINT64_MAX,
+                });
+                if (dirty_it != bs->dirty_db.begin())
+                {
+                    dirty_it--;
+                    dirty_exists = dirty_it->first.oid == je->del.oid;
+                }
                 auto clean_it = bs->clean_db.find(je->del.oid);
-                if (clean_it != bs->clean_db.end() &&
-                    clean_it->second.version < je->del.version)
+                bool clean_exists = (clean_it != bs->clean_db.end() &&
+                    clean_it->second.version < je->del.version);
+                if (!clean_exists && dirty_exists)
+                {
+                    // Clean entry doesn't exist. This means that the delete is already flushed.
+                    // So we must not flush this object anymore.
+                    erase_dirty_object(dirty_it);
+                }
+                else if (clean_exists || dirty_exists)
                 {
                     // oid, version
                     obj_ver_id ov = {
@@ -728,8 +868,9 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                     bs->journal.used_sectors[proc_pos]++;
                     // Deletions are treated as immediately stable, because
                     // "2-phase commit" (write->stabilize) isn't sufficient for them anyway
-                    bs->mark_stable(ov);
+                    bs->mark_stable(ov, true);
                 }
+                // Ignore delete if neither preceding dirty entries nor the clean one are present
             }
             started = true;
             pos += je->size;
@@ -740,3 +881,35 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
     bs->journal.next_free = next_free;
     return 1;
 }
+
+void blockstore_init_journal::erase_dirty_object(blockstore_dirty_db_t::iterator dirty_it)
+{
+    auto oid = dirty_it->first.oid;
+    bool exists = !IS_DELETE(dirty_it->second.state);
+    auto dirty_end = dirty_it;
+    dirty_end++;
+    while (1)
+    {
+        if (dirty_it == bs->dirty_db.begin())
+        {
+            break;
+        }
+        dirty_it--;
+        if (dirty_it->first.oid != oid)
+        {
+            dirty_it++;
+            break;
+        }
+    }
+    auto clean_it = bs->clean_db.find(oid);
+    uint64_t clean_loc = clean_it != bs->clean_db.end()
+        ? clean_it->second.location : UINT64_MAX;
+    if (exists && clean_loc == UINT64_MAX)
+    {
+        bs->inode_space_stats[oid.inode] -= bs->block_size;
+    }
+    bs->erase_dirty(dirty_it, dirty_end, clean_loc);
+    // Remove it from the flusher's queue, too
+    // Otherwise it may end up referring to a small unstable write after reading the rest of the journal
+    bs->flusher->remove_flush(oid);
+}

src/blockstore_init.h

@@ -7,6 +7,7 @@ class blockstore_init_meta
 {
     blockstore_impl_t *bs;
     int wait_state = 0, wait_count = 0;
+    bool zero_on_init = false;
     void *metadata_buffer = NULL;
     uint64_t metadata_read = 0;
     int prev = 0, prev_done = 0, done_len = 0, submitted = 0;
@@ -36,6 +37,7 @@ class blockstore_init_journal
     bool started = false;
     uint64_t next_free;
     std::vector<bs_init_journal_done> done;
+    std::vector<obj_ver_id> double_allocs;
     uint64_t journal_pos = 0;
     uint64_t continue_pos = 0;
     void *init_write_buf = NULL;
@@ -48,6 +50,7 @@ class blockstore_init_journal
     std::function<void(ring_data_t*)> simple_callback;
     int handle_journal_part(void *buf, uint64_t done_pos, uint64_t len);
     void handle_event(ring_data_t *data);
+    void erase_dirty_object(blockstore_dirty_db_t::iterator dirty_it);
 public:
     blockstore_init_journal(blockstore_impl_t* bs);
     int loop();

src/blockstore_journal.h

@@ -7,6 +7,7 @@
 
 #define MIN_JOURNAL_SIZE 4*1024*1024
 #define JOURNAL_MAGIC 0x4A33
+#define JOURNAL_VERSION 1
 #define JOURNAL_BUFFER_SIZE 4*1024*1024
 
 // We reserve some extra space for future stabilize requests during writes
@@ -37,7 +38,9 @@ struct __attribute__((__packed__)) journal_entry_start
     uint32_t size;
     uint32_t reserved;
     uint64_t journal_start;
+    uint64_t version;
 };
+#define JE_START_LEGACY_SIZE 24
 
 struct __attribute__((__packed__)) journal_entry_small_write
 {
@@ -149,6 +152,7 @@ struct journal_t
     int fd;
     uint64_t device_size;
     bool inmemory = false;
+    bool flush_journal = false;
     void *buffer = NULL;
 
     uint64_t block_size;

src/blockstore_open.cpp

@@ -42,6 +42,11 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config)
     {
         disable_flock = true;
     }
+    if (config["flush_journal"] == "true" || config["flush_journal"] == "1" || config["flush_journal"] == "yes")
+    {
+        // Only flush journal and exit
+        journal.flush_journal = true;
+    }
     if (config["immediate_commit"] == "all")
     {
         immediate_commit = IMMEDIATE_ALL;
@@ -69,8 +74,16 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config)
     journal_block_size = strtoull(config["journal_block_size"].c_str(), NULL, 10);
     meta_block_size = strtoull(config["meta_block_size"].c_str(), NULL, 10);
     bitmap_granularity = strtoull(config["bitmap_granularity"].c_str(), NULL, 10);
-    flusher_count = strtoull(config["flusher_count"].c_str(), NULL, 10);
+    max_flusher_count = strtoull(config["max_flusher_count"].c_str(), NULL, 10);
+    if (!max_flusher_count)
+        max_flusher_count = strtoull(config["flusher_count"].c_str(), NULL, 10);
+    min_flusher_count = strtoull(config["min_flusher_count"].c_str(), NULL, 10);
     max_write_iodepth = strtoull(config["max_write_iodepth"].c_str(), NULL, 10);
+    throttle_small_writes = config["throttle_small_writes"] == "true" || config["throttle_small_writes"] == "1" || config["throttle_small_writes"] == "yes";
+    throttle_target_iops = strtoull(config["throttle_target_iops"].c_str(), NULL, 10);
+    throttle_target_mbs = strtoull(config["throttle_target_mbs"].c_str(), NULL, 10);
+    throttle_target_parallelism = strtoull(config["throttle_target_parallelism"].c_str(), NULL, 10);
+    throttle_threshold_us = strtoull(config["throttle_threshold_us"].c_str(), NULL, 10);
     // Validate
     if (!block_size)
     {
@@ -80,9 +93,13 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config)
     {
         throw std::runtime_error("Bad block size");
     }
-    if (!flusher_count)
+    if (!max_flusher_count)
     {
-        flusher_count = 32;
+        max_flusher_count = 256;
+    }
+    if (!min_flusher_count || journal.flush_journal)
+    {
+        min_flusher_count = 1;
     }
     if (!max_write_iodepth)
     {
@@ -168,6 +185,22 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config)
     {
         throw std::runtime_error("immediate_commit=all requires disable_journal_fsync and disable_data_fsync");
     }
+    if (!throttle_target_iops)
+    {
+        throttle_target_iops = 100;
+    }
+    if (!throttle_target_mbs)
+    {
+        throttle_target_mbs = 100;
+    }
+    if (!throttle_target_parallelism)
+    {
+        throttle_target_parallelism = 1;
+    }
+    if (!throttle_threshold_us)
+    {
+        throttle_threshold_us = 50;
+    }
     // init some fields
     clean_entry_bitmap_size = block_size / bitmap_granularity / 8;
     clean_entry_size = sizeof(clean_disk_entry) + 2*clean_entry_bitmap_size;
@@ -224,7 +257,7 @@ void blockstore_impl_t::calc_lengths()
     }
     // required metadata size
     block_count = data_len / block_size;
-    meta_len = ((block_count - 1 + meta_block_size / clean_entry_size) / (meta_block_size / clean_entry_size)) * meta_block_size;
+    meta_len = (1 + (block_count - 1 + meta_block_size / clean_entry_size) / (meta_block_size / clean_entry_size)) * meta_block_size;
     if (meta_area < meta_len)
     {
         throw std::runtime_error("Metadata area is too small, need at least "+std::to_string(meta_len)+" bytes");

src/blockstore_read.cpp

@@ -268,3 +268,50 @@ void blockstore_impl_t::handle_read_event(ring_data_t *data, blockstore_op_t *op
         FINISH_OP(op);
     }
 }
+
+int blockstore_impl_t::read_bitmap(object_id oid, uint64_t target_version, void *bitmap, uint64_t *result_version)
+{
+    auto dirty_it = dirty_db.upper_bound((obj_ver_id){
+        .oid = oid,
+        .version = UINT64_MAX,
+    });
+    if (dirty_it != dirty_db.begin())
+        dirty_it--;
+    if (dirty_it != dirty_db.end())
+    {
+        while (dirty_it->first.oid == oid)
+        {
+            if (target_version >= dirty_it->first.version)
+            {
+                if (result_version)
+                    *result_version = dirty_it->first.version;
+                if (bitmap)
+                {
+                    void *bmp_ptr = (clean_entry_bitmap_size > sizeof(void*) ? dirty_it->second.bitmap : &dirty_it->second.bitmap);
+                    memcpy(bitmap, bmp_ptr, clean_entry_bitmap_size);
+                }
+                return 0;
+            }
+            if (dirty_it == dirty_db.begin())
+                break;
+            dirty_it--;
+        }
+    }
+    auto clean_it = clean_db.find(oid);
+    if (clean_it != clean_db.end())
+    {
+        if (result_version)
+            *result_version = clean_it->second.version;
+        if (bitmap)
+        {
+            void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, clean_entry_bitmap_size);
+            memcpy(bitmap, bmp_ptr, clean_entry_bitmap_size);
+        }
+        return 0;
+    }
+    if (result_version)
+        *result_version = 0;
+    if (bitmap)
+        memset(bitmap, 0, clean_entry_bitmap_size);
+    return -ENOENT;
+}

src/blockstore_rollback.cpp

@@ -248,10 +248,12 @@ void blockstore_impl_t::erase_dirty(blockstore_dirty_db_t::iterator dirty_start,
     }
     while (1)
     {
-        if (IS_BIG_WRITE(dirty_it->second.state) && dirty_it->second.location != clean_loc)
+        if (IS_BIG_WRITE(dirty_it->second.state) && dirty_it->second.location != clean_loc &&
+            dirty_it->second.location != UINT64_MAX)
         {
 #ifdef BLOCKSTORE_DEBUG
-            printf("Free block %lu\n", dirty_it->second.location >> block_order);
+            printf("Free block %lu from %lx:%lx v%lu\n", dirty_it->second.location >> block_order,
+                dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version);
 #endif
             data_alloc->set(dirty_it->second.location >> block_order, false);
         }

src/blockstore_stable.cpp

@@ -168,6 +168,9 @@ resume_5:
     for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
     {
         // Mark all dirty_db entries up to op->version as stable
+#ifdef BLOCKSTORE_DEBUG
+        printf("Stabilize %lx:%lx v%lu\n", v->oid.inode, v->oid.stripe, v->version);
+#endif
         mark_stable(*v);
     }
     // Acknowledge op
@@ -176,31 +179,66 @@ resume_5:
     return 2;
 }
 
-void blockstore_impl_t::mark_stable(const obj_ver_id & v)
+void blockstore_impl_t::mark_stable(const obj_ver_id & v, bool forget_dirty)
 {
     auto dirty_it = dirty_db.find(v);
     if (dirty_it != dirty_db.end())
     {
         while (1)
         {
+            bool was_stable = IS_STABLE(dirty_it->second.state);
             if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_SYNCED)
             {
                 dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_STABLE;
                 // Allocations and deletions are counted when they're stabilized
                 if (IS_BIG_WRITE(dirty_it->second.state))
                 {
-                    inode_space_stats[dirty_it->first.oid.inode] += block_size;
+                    int exists = -1;
+                    if (dirty_it != dirty_db.begin())
+                    {
+                        auto prev_it = dirty_it;
+                        prev_it--;
+                        if (prev_it->first.oid == v.oid)
+                        {
+                            exists = IS_DELETE(prev_it->second.state) ? 0 : 1;
+                        }
+                    }
+                    if (exists == -1)
+                    {
+                        auto clean_it = clean_db.find(v.oid);
+                        exists = clean_it != clean_db.end() ? 1 : 0;
+                    }
+                    if (!exists)
+                    {
+                        inode_space_stats[dirty_it->first.oid.inode] += block_size;
+                    }
                 }
                 else if (IS_DELETE(dirty_it->second.state))
                 {
                     inode_space_stats[dirty_it->first.oid.inode] -= block_size;
                 }
             }
-            else if (IS_STABLE(dirty_it->second.state))
+            if (forget_dirty && (IS_BIG_WRITE(dirty_it->second.state) ||
+                IS_DELETE(dirty_it->second.state)))
             {
+                // Big write overrides all previous dirty entries
+                auto erase_end = dirty_it;
+                while (dirty_it != dirty_db.begin())
+                {
+                    dirty_it--;
+                    if (dirty_it->first.oid != v.oid)
+                    {
+                        dirty_it++;
+                        break;
+                    }
+                }
+                auto clean_it = clean_db.find(v.oid);
+                uint64_t clean_loc = clean_it != clean_db.end()
+                    ? clean_it->second.location : UINT64_MAX;
+                erase_dirty(dirty_it, erase_end, clean_loc);
                 break;
             }
-            if (dirty_it == dirty_db.begin())
+            if (was_stable || dirty_it == dirty_db.begin())
             {
                 break;
             }

src/blockstore_sync.cpp

@@ -24,6 +24,7 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op, bool queue_has_in_prog
     if (PRIV(op)->op_state == 0)
     {
         stop_sync_submitted = false;
+        unsynced_big_write_count -= unsynced_big_writes.size();
         PRIV(op)->sync_big_writes.swap(unsynced_big_writes);
         PRIV(op)->sync_small_writes.swap(unsynced_small_writes);
         PRIV(op)->sync_small_checked = 0;
@@ -79,7 +80,8 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op, bool queue_has_in_prog
         // 2nd step: Data device is synced, prepare & write journal entries
         // Check space in the journal and journal memory buffers
         blockstore_journal_check_t space_check(this);
-        if (!space_check.check_available(op, PRIV(op)->sync_big_writes.size(), sizeof(journal_entry_big_write), JOURNAL_STABILIZE_RESERVATION))
+        if (!space_check.check_available(op, PRIV(op)->sync_big_writes.size(),
+            sizeof(journal_entry_big_write) + clean_entry_bitmap_size, JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
         }
@@ -94,7 +96,7 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op, bool queue_has_in_prog
         int s = 0, cur_sector = -1;
         while (it != PRIV(op)->sync_big_writes.end())
         {
-            if (!journal.entry_fits(sizeof(journal_entry_big_write)) &&
+            if (!journal.entry_fits(sizeof(journal_entry_big_write) + clean_entry_bitmap_size) &&
                 journal.sector_info[journal.cur_sector].dirty)
             {
                 if (cur_sector == -1)
@@ -102,24 +104,27 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op, bool queue_has_in_prog
                 prepare_journal_sector_write(journal, journal.cur_sector, sqe[s++], [this, op](ring_data_t *data) { handle_sync_event(data, op); });
                 cur_sector = journal.cur_sector;
             }
+            auto & dirty_entry = dirty_db.at(*it);
             journal_entry_big_write *je = (journal_entry_big_write*)prefill_single_journal_entry(
-                journal, (dirty_db[*it].state & BS_ST_INSTANT) ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
-                sizeof(journal_entry_big_write)
+                journal, (dirty_entry.state & BS_ST_INSTANT) ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
+                sizeof(journal_entry_big_write) + clean_entry_bitmap_size
             );
-            dirty_db[*it].journal_sector = journal.sector_info[journal.cur_sector].offset;
+            dirty_entry.journal_sector = journal.sector_info[journal.cur_sector].offset;
             journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
 #ifdef BLOCKSTORE_DEBUG
             printf(
                 "journal offset %08lx is used by %lx:%lx v%lu (%lu refs)\n",
-                dirty_db[*it].journal_sector, it->oid.inode, it->oid.stripe, it->version,
+                dirty_entry.journal_sector, it->oid.inode, it->oid.stripe, it->version,
                 journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
             );
 #endif
             je->oid = it->oid;
             je->version = it->version;
-            je->offset = dirty_db[*it].offset;
-            je->len = dirty_db[*it].len;
-            je->location = dirty_db[*it].location;
+            je->offset = dirty_entry.offset;
+            je->len = dirty_entry.len;
+            je->location = dirty_entry.location;
+            memcpy((void*)(je+1), (clean_entry_bitmap_size > sizeof(void*)
+                ? dirty_entry.bitmap : &dirty_entry.bitmap), clean_entry_bitmap_size);
             je->crc32 = je_crc32((journal_entry*)je);
             journal.crc32_last = je->crc32;
             it++;

src/blockstore_write.cpp

@@ -30,10 +30,13 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
             wait_big = (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE
                 ? !IS_SYNCED(dirty_it->second.state)
                 : ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG);
-            if (clean_entry_bitmap_size > sizeof(void*))
-                memcpy(bmp, dirty_it->second.bitmap, clean_entry_bitmap_size);
-            else
-                bmp = dirty_it->second.bitmap;
+            if (!is_del && !deleted)
+            {
+                if (clean_entry_bitmap_size > sizeof(void*))
+                    memcpy(bmp, dirty_it->second.bitmap, clean_entry_bitmap_size);
+                else
+                    bmp = dirty_it->second.bitmap;
+            }
         }
     }
     if (!found)
@@ -42,8 +45,11 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
         if (clean_it != clean_db.end())
         {
             version = clean_it->second.version + 1;
-            void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, clean_entry_bitmap_size);
-            memcpy((clean_entry_bitmap_size > sizeof(void*) ? bmp : &bmp), bmp_ptr, clean_entry_bitmap_size);
+            if (!is_del)
+            {
+                void *bmp_ptr = get_clean_entry_bitmap(clean_it->second.location, clean_entry_bitmap_size);
+                memcpy((clean_entry_bitmap_size > sizeof(void*) ? bmp : &bmp), bmp_ptr, clean_entry_bitmap_size);
+            }
         }
         else
         {
@@ -116,6 +122,8 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
     else
     {
         state = (op->len == block_size || deleted ? BS_ST_BIG_WRITE : BS_ST_SMALL_WRITE);
+        if (state == BS_ST_SMALL_WRITE && throttle_small_writes)
+            clock_gettime(CLOCK_REALTIME, &PRIV(op)->tv_begin);
         if (wait_del)
             state |= BS_ST_WAIT_DEL;
         else if (state == BS_ST_SMALL_WRITE && wait_big)
@@ -241,7 +249,8 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
     if ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE)
     {
         blockstore_journal_check_t space_check(this);
-        if (!space_check.check_available(op, unsynced_big_writes.size() + 1, sizeof(journal_entry_big_write), JOURNAL_STABILIZE_RESERVATION))
+        if (!space_check.check_available(op, unsynced_big_write_count + 1,
+            sizeof(journal_entry_big_write) + clean_entry_bitmap_size, JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
         }
@@ -264,7 +273,10 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         dirty_it->second.location = loc << block_order;
         dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_SUBMITTED;
 #ifdef BLOCKSTORE_DEBUG
-        printf("Allocate block %lu\n", loc);
+        printf(
+            "Allocate block %lu for %lx:%lx v%lu\n",
+            loc, op->oid.inode, op->oid.stripe, op->version
+        );
 #endif
         data_alloc->set(loc, true);
         uint64_t stripe_offset = (op->offset % bitmap_granularity);
@@ -290,11 +302,8 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
         if (immediate_commit != IMMEDIATE_ALL)
         {
-            // Remember big write as unsynced
-            unsynced_big_writes.push_back((obj_ver_id){
-                .oid = op->oid,
-                .version = op->version,
-            });
+            // Increase the counter, but don't save into unsynced_writes yet (can't sync until the write is finished)
+            unsynced_big_write_count++;
             PRIV(op)->op_state = 3;
         }
         else
@@ -307,8 +316,11 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         // Small (journaled) write
         // First check if the journal has sufficient space
         blockstore_journal_check_t space_check(this);
-        if (unsynced_big_writes.size() && !space_check.check_available(op, unsynced_big_writes.size(), sizeof(journal_entry_big_write), 0)
-            || !space_check.check_available(op, 1, sizeof(journal_entry_small_write), op->len + JOURNAL_STABILIZE_RESERVATION))
+        if (unsynced_big_write_count &&
+            !space_check.check_available(op, unsynced_big_write_count,
+                sizeof(journal_entry_big_write) + clean_entry_bitmap_size, 0)
+            || !space_check.check_available(op, 1,
+                sizeof(journal_entry_small_write) + clean_entry_bitmap_size, op->len + JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
         }
@@ -316,8 +328,7 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         // There is sufficient space. Get SQE(s)
         struct io_uring_sqe *sqe1 = NULL;
         if (immediate_commit != IMMEDIATE_NONE ||
-            (journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_small_write) &&
-            journal.sector_info[journal.cur_sector].dirty)
+            !journal.entry_fits(sizeof(journal_entry_small_write) + clean_entry_bitmap_size))
         {
             // Write current journal sector only if it's dirty and full, or in the immediate_commit mode
             BS_SUBMIT_GET_SQE_DECL(sqe1);
@@ -400,14 +411,6 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         {
             journal.next_free = journal_block_size;
         }
-        if (immediate_commit == IMMEDIATE_NONE)
-        {
-            // Remember small write as unsynced
-            unsynced_small_writes.push_back((obj_ver_id){
-                .oid = op->oid,
-                .version = op->version,
-            });
-        }
         if (!PRIV(op)->pending_ops)
         {
             PRIV(op)->op_state = 4;
@@ -423,85 +426,148 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
 
 int blockstore_impl_t::continue_write(blockstore_op_t *op)
 {
-    io_uring_sqe *sqe = NULL;
-    journal_entry_big_write *je;
     int op_state = PRIV(op)->op_state;
-    if (op_state != 2 && op_state != 4)
-    {
-        // In progress
-        return 1;
-    }
-    auto dirty_it = dirty_db.find((obj_ver_id){
-        .oid = op->oid,
-        .version = op->version,
-    });
-    assert(dirty_it != dirty_db.end());
     if (op_state == 2)
         goto resume_2;
     else if (op_state == 4)
         goto resume_4;
+    else if (op_state == 6)
+        goto resume_6;
+    else
+    {
+        // In progress
+        return 1;
+    }
 resume_2:
     // Only for the immediate_commit mode: prepare and submit big_write journal entry
-    BS_SUBMIT_GET_SQE_DECL(sqe);
-    je = (journal_entry_big_write*)prefill_single_journal_entry(
-        journal, op->opcode == BS_OP_WRITE_STABLE ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
-        sizeof(journal_entry_big_write) + clean_entry_bitmap_size
-    );
-    dirty_it->second.journal_sector = journal.sector_info[journal.cur_sector].offset;
-    journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
+    {
+        auto dirty_it = dirty_db.find((obj_ver_id){
+            .oid = op->oid,
+            .version = op->version,
+        });
+        assert(dirty_it != dirty_db.end());
+        io_uring_sqe *sqe = NULL;
+        BS_SUBMIT_GET_SQE_DECL(sqe);
+        journal_entry_big_write *je = (journal_entry_big_write*)prefill_single_journal_entry(
+            journal, op->opcode == BS_OP_WRITE_STABLE ? JE_BIG_WRITE_INSTANT : JE_BIG_WRITE,
+            sizeof(journal_entry_big_write) + clean_entry_bitmap_size
+        );
+        dirty_it->second.journal_sector = journal.sector_info[journal.cur_sector].offset;
+        journal.used_sectors[journal.sector_info[journal.cur_sector].offset]++;
 #ifdef BLOCKSTORE_DEBUG
-    printf(
-        "journal offset %08lx is used by %lx:%lx v%lu (%lu refs)\n",
-        journal.sector_info[journal.cur_sector].offset, op->oid.inode, op->oid.stripe, op->version,
-        journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
-    );
+        printf(
+            "journal offset %08lx is used by %lx:%lx v%lu (%lu refs)\n",
+            journal.sector_info[journal.cur_sector].offset, op->oid.inode, op->oid.stripe, op->version,
+            journal.used_sectors[journal.sector_info[journal.cur_sector].offset]
+        );
 #endif
-    je->oid = op->oid;
-    je->version = op->version;
-    je->offset = op->offset;
-    je->len = op->len;
-    je->location = dirty_it->second.location;
-    memcpy((void*)(je+1), (clean_entry_bitmap_size > sizeof(void*) ? dirty_it->second.bitmap : &dirty_it->second.bitmap), clean_entry_bitmap_size);
-    je->crc32 = je_crc32((journal_entry*)je);
-    journal.crc32_last = je->crc32;
-    prepare_journal_sector_write(journal, journal.cur_sector, sqe,
-        [this, op](ring_data_t *data) { handle_write_event(data, op); });
-    PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 1 + journal.cur_sector;
-    PRIV(op)->pending_ops = 1;
-    PRIV(op)->op_state = 3;
-    return 1;
+        je->oid = op->oid;
+        je->version = op->version;
+        je->offset = op->offset;
+        je->len = op->len;
+        je->location = dirty_it->second.location;
+        memcpy((void*)(je+1), (clean_entry_bitmap_size > sizeof(void*) ? dirty_it->second.bitmap : &dirty_it->second.bitmap), clean_entry_bitmap_size);
+        je->crc32 = je_crc32((journal_entry*)je);
+        journal.crc32_last = je->crc32;
+        prepare_journal_sector_write(journal, journal.cur_sector, sqe,
+            [this, op](ring_data_t *data) { handle_write_event(data, op); });
+        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 1 + journal.cur_sector;
+        PRIV(op)->pending_ops = 1;
+        PRIV(op)->op_state = 3;
+        return 1;
+    }
 resume_4:
     // Switch object state
 #ifdef BLOCKSTORE_DEBUG
-    printf("Ack write %lx:%lx v%lu = state %x\n", op->oid.inode, op->oid.stripe, op->version, dirty_it->second.state);
+    printf("Ack write %lx:%lx v%lu = state 0x%x\n", op->oid.inode, op->oid.stripe, op->version, dirty_it->second.state);
 #endif
-    bool imm = (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE
-        ? (immediate_commit == IMMEDIATE_ALL)
-        : (immediate_commit != IMMEDIATE_NONE);
-    if (imm)
     {
-        auto & unstab = unstable_writes[op->oid];
-        unstab = unstab < op->version ? op->version : unstab;
-    }
-    dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK)
-        | (imm ? BS_ST_SYNCED : BS_ST_WRITTEN);
-    if (imm && ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_DELETE || (dirty_it->second.state & BS_ST_INSTANT)))
-    {
-        // Deletions are treated as immediately stable
-        mark_stable(dirty_it->first);
-    }
-    if (immediate_commit == IMMEDIATE_ALL)
-    {
-        dirty_it++;
-        while (dirty_it != dirty_db.end() && dirty_it->first.oid == op->oid)
+        auto dirty_it = dirty_db.find((obj_ver_id){
+            .oid = op->oid,
+            .version = op->version,
+        });
+        assert(dirty_it != dirty_db.end());
+        bool is_big = (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE;
+        bool imm = is_big ? (immediate_commit == IMMEDIATE_ALL) : (immediate_commit != IMMEDIATE_NONE);
+        if (imm)
         {
-            if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG)
+            auto & unstab = unstable_writes[op->oid];
+            unstab = unstab < op->version ? op->version : unstab;
+        }
+        dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK)
+            | (imm ? BS_ST_SYNCED : BS_ST_WRITTEN);
+        if (imm && ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_DELETE || (dirty_it->second.state & BS_ST_INSTANT)))
+        {
+            // Deletions and 'instant' operations are treated as immediately stable
+            mark_stable(dirty_it->first);
+        }
+        if (!imm)
+        {
+            if (is_big)
             {
-                dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_IN_FLIGHT;
+                // Remember big write as unsynced
+                unsynced_big_writes.push_back((obj_ver_id){
+                    .oid = op->oid,
+                    .version = op->version,
+                });
             }
+            else
+            {
+                // Remember small write as unsynced
+                unsynced_small_writes.push_back((obj_ver_id){
+                    .oid = op->oid,
+                    .version = op->version,
+                });
+            }
+        }
+        if (imm && (dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE)
+        {
+            // Unblock small writes
             dirty_it++;
+            while (dirty_it != dirty_db.end() && dirty_it->first.oid == op->oid)
+            {
+                if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_WAIT_BIG)
+                {
+                    dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_IN_FLIGHT;
+                }
+                dirty_it++;
+            }
+        }
+        // Apply throttling to not fill the journal too fast for the SSD+HDD case
+        if (!is_big && throttle_small_writes)
+        {
+            // Apply throttling
+            timespec tv_end;
+            clock_gettime(CLOCK_REALTIME, &tv_end);
+            uint64_t exec_us =
+                (tv_end.tv_sec - PRIV(op)->tv_begin.tv_sec)*1000000 +
+                (tv_end.tv_nsec - PRIV(op)->tv_begin.tv_nsec)/1000;
+            // Compare with target execution time
+            // 100% free -> target time = 0
+            // 0% free -> target time = iodepth/parallelism * (iops + size/bw) / write per second
+            uint64_t used_start = journal.get_trim_pos();
+            uint64_t journal_free_space = journal.next_free < used_start
+                ? (used_start - journal.next_free)
+                : (journal.len - journal.next_free + used_start - journal.block_size);
+            uint64_t ref_us =
+                (write_iodepth <= throttle_target_parallelism ? 100 : 100*write_iodepth/throttle_target_parallelism)
+                * (1000000/throttle_target_iops + op->len*1000000/throttle_target_mbs/1024/1024)
+                / 100;
+            ref_us -= ref_us * journal_free_space / journal.len;
+            if (ref_us > exec_us + throttle_threshold_us)
+            {
+                // Pause reply
+                tfd->set_timer_us(ref_us-exec_us, false, [this, op](int timer_id)
+                {
+                    PRIV(op)->op_state++;
+                    ringloop->wakeup();
+                });
+                PRIV(op)->op_state = 5;
+                return 1;
+            }
         }
     }
+resume_6:
     // Acknowledge write
     op->retval = op->len;
     write_iodepth--;
@@ -625,14 +691,6 @@ int blockstore_impl_t::dequeue_del(blockstore_op_t *op)
         PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 1 + journal.cur_sector;
         PRIV(op)->pending_ops++;
     }
-    else
-    {
-        // Remember delete as unsynced
-        unsynced_small_writes.push_back((obj_ver_id){
-            .oid = op->oid,
-            .version = op->version,
-        });
-    }
     if (!PRIV(op)->pending_ops)
     {
         PRIV(op)->op_state = 4;

src/cluster_client.h

@@ -8,7 +8,8 @@
 
 #define MIN_BLOCK_SIZE 4*1024
 #define MAX_BLOCK_SIZE 128*1024*1024
-#define DEFAULT_CLIENT_DIRTY_LIMIT 32*1024*1024
+#define DEFAULT_CLIENT_MAX_DIRTY_BYTES 32*1024*1024
+#define DEFAULT_CLIENT_MAX_DIRTY_OPS 1024
 
 struct cluster_op_t;
 
@@ -20,8 +21,7 @@ struct cluster_op_part_t
     pg_num_t pg_num;
     osd_num_t osd_num;
     osd_op_buf_list_t iov;
-    bool sent;
-    bool done;
+    unsigned flags;
     osd_op_t op;
 };
 
@@ -36,19 +36,28 @@ struct cluster_op_t
     std::function<void(cluster_op_t*)> callback;
     ~cluster_op_t();
 protected:
+    int flags = 0;
+    int state = 0;
     uint64_t cur_inode; // for snapshot reads
     void *buf = NULL;
     cluster_op_t *orig_op = NULL;
-    bool is_internal = false;
     bool needs_reslice = false;
     bool up_wait = false;
-    int sent_count = 0, done_count = 0;
+    int inflight_count = 0, done_count = 0;
     std::vector<cluster_op_part_t> parts;
     void *bitmap_buf = NULL, *part_bitmaps = NULL;
     unsigned bitmap_buf_size = 0;
     friend class cluster_client_t;
 };
 
+struct cluster_buffer_t
+{
+    void *buf;
+    uint64_t len;
+    int state;
+};
+
+// FIXME: Split into public and private interfaces
 class cluster_client_t
 {
     timerfd_manager_t *tfd;
@@ -59,28 +68,27 @@ class cluster_client_t
     std::map<pool_id_t, uint64_t> pg_counts;
     bool immediate_commit = false;
     // FIXME: Implement inmemory_commit mode. Note that it requires to return overlapping reads from memory.
-    uint64_t client_dirty_limit = 0;
+    uint64_t client_max_dirty_bytes = 0;
+    uint64_t client_max_dirty_ops = 0;
     int log_level;
     int up_wait_retry_interval = 500; // ms
 
-    uint64_t op_id = 1;
-    ring_consumer_t consumer;
-    // operations currently in progress
-    std::set<cluster_op_t*> cur_ops;
     int retry_timeout_id = 0;
-    // unsynced operations are copied in memory to allow replay when cluster isn't in the immediate_commit mode
-    // unsynced_writes are replayed in any order (because only the SYNC operation guarantees ordering)
-    std::vector<cluster_op_t*> unsynced_writes;
-    std::vector<cluster_op_t*> syncing_writes;
-    cluster_op_t* cur_sync = NULL;
-    std::vector<cluster_op_t*> next_writes;
+    uint64_t op_id = 1;
     std::vector<cluster_op_t*> offline_ops;
-    uint64_t queued_bytes = 0;
+    std::vector<cluster_op_t*> op_queue;
+    std::map<object_id, cluster_buffer_t> dirty_buffers;
+    std::set<osd_num_t> dirty_osds;
+    uint64_t dirty_bytes = 0, dirty_ops = 0;
+
     void *scrap_buffer = NULL;
     unsigned scrap_buffer_size = 0;
 
     bool pgs_loaded = false;
+    ring_consumer_t consumer;
     std::vector<std::function<void(void)>> on_ready_hooks;
+    int continuing_ops = 0;
+    int op_queue_pos = 0;
 
 public:
     etcd_state_client_t st_cli;
@@ -93,19 +101,20 @@ public:
     bool is_ready();
     void on_ready(std::function<void(void)> fn);
 
-protected:
+    static void copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers);
     void continue_ops(bool up_retry = false);
+protected:
+    bool affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd);
+    void flush_buffer(const object_id & oid, cluster_buffer_t *wr);
     void on_load_config_hook(json11::Json::object & config);
     void on_load_pgs_hook(bool success);
-    void on_change_hook(json11::Json::object & changes);
+    void on_change_hook(std::map<std::string, etcd_kv_t> & changes);
     void on_change_osd_state_hook(uint64_t peer_osd);
-    cluster_op_t *copy_write(cluster_op_t *op);
-    void continue_rw(cluster_op_t *op);
+    int continue_rw(cluster_op_t *op);
     void slice_rw(cluster_op_t *op);
     bool try_send(cluster_op_t *op, int i);
-    void execute_sync(cluster_op_t *op);
-    void continue_sync();
-    void finish_sync();
+    int continue_sync(cluster_op_t *op);
     void send_sync(cluster_op_t *op, cluster_op_part_t *part);
     void handle_op_part(cluster_op_part_t *part);
+    void copy_part_bitmap(cluster_op_t *op, cluster_op_part_t *part);
 };

src/etcd_state_client.cpp

@@ -4,8 +4,10 @@
 #include "osd_ops.h"
 #include "pg_states.h"
 #include "etcd_state_client.h"
+#ifndef __MOCK__
 #include "http_client.h"
 #include "base64.h"
+#endif
 
 etcd_state_client_t::~etcd_state_client_t()
 {
@@ -15,16 +17,19 @@ etcd_state_client_t::~etcd_state_client_t()
     }
     watches.clear();
     etcd_watches_initialised = -1;
+#ifndef __MOCK__
     if (etcd_watch_ws)
     {
         etcd_watch_ws->close();
         etcd_watch_ws = NULL;
     }
+#endif
 }
 
-json_kv_t etcd_state_client_t::parse_etcd_kv(const json11::Json & kv_json)
+#ifndef __MOCK__
+etcd_kv_t etcd_state_client_t::parse_etcd_kv(const json11::Json & kv_json)
 {
-    json_kv_t kv;
+    etcd_kv_t kv;
     kv.key = base64_decode(kv_json["key"].string_value());
     std::string json_err, json_text = base64_decode(kv_json["value"].string_value());
     kv.value = json_text == "" ? json11::Json() : json11::Json::parse(json_text, json_err);
@@ -33,6 +38,8 @@ json_kv_t etcd_state_client_t::parse_etcd_kv(const json11::Json & kv_json)
         printf("Bad JSON in etcd key %s: %s (value: %s)\n", kv.key.c_str(), json_err.c_str(), json_text.c_str());
         kv.key = "";
     }
+    else
+        kv.mod_revision = kv_json["mod_revision"].uint64_value();
     return kv;
 }
 
@@ -145,22 +152,22 @@ void etcd_state_client_t::start_etcd_watcher()
                     etcd_watch_revision = data["result"]["header"]["revision"].uint64_value();
                 }
                 // First gather all changes into a hash to remove multiple overwrites
-                json11::Json::object changes;
+                std::map<std::string, etcd_kv_t> changes;
                 for (auto & ev: data["result"]["events"].array_items())
                 {
                     auto kv = parse_etcd_kv(ev["kv"]);
                     if (kv.key != "")
                     {
-                        changes[kv.key] = kv.value;
+                        changes[kv.key] = kv;
                     }
                 }
                 for (auto & kv: changes)
                 {
                     if (this->log_level > 3)
                     {
-                        printf("Incoming event: %s -> %s\n", kv.first.c_str(), kv.second.dump().c_str());
+                        printf("Incoming event: %s -> %s\n", kv.first.c_str(), kv.second.value.dump().c_str());
                     }
-                    parse_state(kv.first, kv.second);
+                    parse_state(kv.second);
                 }
                 // React to changes
                 if (on_change_hook != NULL)
@@ -327,16 +334,33 @@ void etcd_state_client_t::load_pgs()
             for (auto & kv_json: res["response_range"]["kvs"].array_items())
             {
                 auto kv = parse_etcd_kv(kv_json);
-                parse_state(kv.key, kv.value);
+                parse_state(kv);
             }
         }
         on_load_pgs_hook(true);
         start_etcd_watcher();
     });
 }
-
-void etcd_state_client_t::parse_state(const std::string & key, const json11::Json & value)
+#else
+void etcd_state_client_t::parse_config(json11::Json & config)
 {
+}
+
+void etcd_state_client_t::load_global_config()
+{
+    json11::Json::object global_config;
+    on_load_config_hook(global_config);
+}
+
+void etcd_state_client_t::load_pgs()
+{
+}
+#endif
+
+void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
+{
+    const std::string & key = kv.key;
+    const json11::Json & value = kv.value;
     if (key == etcd_prefix+"/config/pools")
     {
         for (auto & pool_item: this->pool_config)
@@ -347,8 +371,10 @@ void etcd_state_client_t::parse_state(const std::string & key, const json11::Jso
         {
             pool_config_t pc;
             // ID
-            pool_id_t pool_id = stoull_full(pool_item.first);
-            if (!pool_id || pool_id >= POOL_ID_MAX)
+            pool_id_t pool_id;
+            char null_byte = 0;
+            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
             {
                 printf("Pool ID %s is invalid (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
@@ -460,16 +486,19 @@ void etcd_state_client_t::parse_state(const std::string & key, const json11::Jso
         }
         for (auto & pool_item: value["items"].object_items())
         {
-            pool_id_t pool_id = stoull_full(pool_item.first);
-            if (!pool_id || pool_id >= POOL_ID_MAX)
+            pool_id_t pool_id;
+            char null_byte = 0;
+            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
             {
                 printf("Pool ID %s is invalid in PG configuration (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
             }
             for (auto & pg_item: pool_item.second.object_items())
             {
-                pg_num_t pg_num = stoull_full(pg_item.first);
-                if (!pg_num)
+                pg_num_t pg_num = 0;
+                sscanf(pg_item.first.c_str(), "%u%c", &pg_num, &null_byte);
+                if (!pg_num || null_byte != 0)
                 {
                     printf("Bad key in pool %u PG configuration: %s (must be a number), skipped\n", pool_id, pg_item.first.c_str());
                     continue;
@@ -682,6 +711,7 @@ void etcd_state_client_t::parse_state(const std::string & key, const json11::Jso
                     .size = value["size"].uint64_value(),
                     .parent_id = parent_inode_num,
                     .readonly = value["readonly"].bool_value(),
+                    .mod_revision = kv.mod_revision,
                 };
                 this->inode_config[inode_num] = cfg;
                 if (cfg.name != "")

src/etcd_state_client.h

@@ -3,8 +3,8 @@
 
 #pragma once
 
+#include "json11/json11.hpp"
 #include "osd_id.h"
-#include "http_client.h"
 #include "timerfd_manager.h"
 
 #define ETCD_CONFIG_WATCH_ID 1
@@ -18,10 +18,11 @@
 
 #define DEFAULT_BLOCK_SIZE 128*1024
 
-struct json_kv_t
+struct etcd_kv_t
 {
     std::string key;
     json11::Json value;
+    uint64_t mod_revision;
 };
 
 struct pg_config_t
@@ -59,6 +60,8 @@ struct inode_config_t
     uint64_t size;
     inode_t parent_id;
     bool readonly;
+    // Change revision of the metadata in etcd
+    uint64_t mod_revision;
 };
 
 struct inode_watch_t
@@ -67,12 +70,14 @@ struct inode_watch_t
     inode_config_t cfg;
 };
 
+struct websocket_t;
+
 struct etcd_state_client_t
 {
 protected:
     std::vector<inode_watch_t*> watches;
     websocket_t *etcd_watch_ws = NULL;
-    uint64_t bs_block_size = 0;
+    uint64_t bs_block_size = DEFAULT_BLOCK_SIZE;
     void add_etcd_url(std::string);
 public:
     std::vector<std::string> etcd_addresses;
@@ -87,20 +92,20 @@ public:
     std::map<inode_t, inode_config_t> inode_config;
     std::map<std::string, inode_t> inode_by_name;
 
-    std::function<void(json11::Json::object &)> on_change_hook;
+    std::function<void(std::map<std::string, etcd_kv_t> &)> on_change_hook;
     std::function<void(json11::Json::object &)> on_load_config_hook;
     std::function<json11::Json()> load_pgs_checks_hook;
     std::function<void(bool)> on_load_pgs_hook;
     std::function<void(pool_id_t, pg_num_t)> on_change_pg_history_hook;
     std::function<void(osd_num_t)> on_change_osd_state_hook;
 
-    json_kv_t parse_etcd_kv(const json11::Json & kv_json);
+    etcd_kv_t parse_etcd_kv(const json11::Json & kv_json);
     void etcd_call(std::string api, json11::Json payload, int timeout, std::function<void(std::string, json11::Json)> callback);
     void etcd_txn(json11::Json txn, int timeout, std::function<void(std::string, json11::Json)> callback);
     void start_etcd_watcher();
     void load_global_config();
     void load_pgs();
-    void parse_state(const std::string & key, const json11::Json & value);
+    void parse_state(const etcd_kv_t & kv);
     void parse_config(json11::Json & config);
     inode_watch_t* watch_inode(std::string name);
     void close_watch(inode_watch_t* watch);

src/fio_engine.cpp

@@ -25,6 +25,7 @@
 //     -bs_config='{"data_device":"./test_data.bin"}' -size=1000M
 
 #include "blockstore.h"
+#include "epoll_manager.h"
 #include "fio_headers.h"
 
 #include "json11/json11.hpp"
@@ -32,6 +33,7 @@
 struct bs_data
 {
     blockstore_t *bs;
+    epoll_manager_t *epmgr;
     ring_loop_t *ringloop;
     /* The list of completed io_u structs. */
     std::vector<io_u*> completed;
@@ -104,6 +106,7 @@ static void bs_cleanup(struct thread_data *td)
         }
     safe:
         delete bsd->bs;
+        delete bsd->epmgr;
         delete bsd->ringloop;
         delete bsd;
     }
@@ -129,7 +132,8 @@ static int bs_init(struct thread_data *td)
         }
     }
     bsd->ringloop = new ring_loop_t(512);
-    bsd->bs = new blockstore_t(config, bsd->ringloop);
+    bsd->epmgr = new epoll_manager_t(bsd->ringloop);
+    bsd->bs = new blockstore_t(config, bsd->ringloop, bsd->epmgr->tfd);
     while (1)
     {
         bsd->ringloop->loop();

src/messenger.cpp

@@ -10,30 +10,16 @@
 
 #include "messenger.h"
 
-osd_op_t::~osd_op_t()
-{
-    assert(!bs_op);
-    assert(!op_data);
-    if (rmw_buf)
-    {
-        free(rmw_buf);
-    }
-    if (buf)
-    {
-        // Note: reusing osd_op_t WILL currently lead to memory leaks
-        // So we don't reuse it, but free it every time
-        free(buf);
-    }
-}
-
 void osd_messenger_t::init()
 {
     keepalive_timer_id = tfd->set_timer(1000, true, [this](int)
     {
-        for (auto cl_it = clients.begin(); cl_it != clients.end();)
+        std::vector<int> to_stop;
+        std::vector<osd_op_t*> to_ping;
+        for (auto cl_it = clients.begin(); cl_it != clients.end(); cl_it++)
         {
-            auto cl = (cl_it++)->second;
-            if (!cl->osd_num)
+            auto cl = cl_it->second;
+            if (!cl->osd_num || cl->peer_state != PEER_CONNECTED)
             {
                 // Do not run keepalive on regular clients
                 continue;
@@ -44,7 +30,8 @@ void osd_messenger_t::init()
                 if (!cl->ping_time_remaining)
                 {
                     // Ping timed out, stop the client
-                    stop_client(cl->peer_fd, true);
+                    printf("Ping timed out for OSD %lu (client %d), disconnecting peer\n", cl->osd_num, cl->peer_fd);
+                    to_stop.push_back(cl->peer_fd);
                 }
             }
             else if (cl->idle_time_remaining > 0)
@@ -70,10 +57,11 @@ void osd_messenger_t::init()
                         delete op;
                         if (fail_fd >= 0)
                         {
+                            printf("Ping failed for OSD %lu (client %d), disconnecting peer\n", cl->osd_num, cl->peer_fd);
                             stop_client(fail_fd, true);
                         }
                     };
-                    outbox_push(op);
+                    to_ping.push_back(op);
                     cl->ping_time_remaining = osd_ping_timeout;
                     cl->idle_time_remaining = osd_idle_timeout;
                 }
@@ -83,6 +71,15 @@ void osd_messenger_t::init()
                 cl->idle_time_remaining = osd_idle_timeout;
             }
         }
+        // Don't stop clients while a 'clients' iterator is still active
+        for (int peer_fd: to_stop)
+        {
+            stop_client(peer_fd, true);
+        }
+        for (auto op: to_ping)
+        {
+            outbox_push(op);
+        }
     });
 }
 
@@ -141,17 +138,14 @@ void osd_messenger_t::connect_peer(uint64_t peer_osd, json11::Json peer_state)
         wanted_peers[peer_osd].port = (int)peer_state["port"].int64_value();
     }
     wanted_peers[peer_osd].address_changed = true;
-    if (!wanted_peers[peer_osd].connecting &&
-        (time(NULL) - wanted_peers[peer_osd].last_connect_attempt) >= peer_connect_interval)
-    {
-        try_connect_peer(peer_osd);
-    }
+    try_connect_peer(peer_osd);
 }
 
 void osd_messenger_t::try_connect_peer(uint64_t peer_osd)
 {
     auto wp_it = wanted_peers.find(peer_osd);
-    if (wp_it == wanted_peers.end())
+    if (wp_it == wanted_peers.end() || wp_it->second.connecting ||
+        (time(NULL) - wp_it->second.last_connect_attempt) < peer_connect_interval)
     {
         return;
     }
@@ -197,10 +191,22 @@ void osd_messenger_t::try_connect_peer_addr(osd_num_t peer_osd, const char *peer
         on_connect_peer(peer_osd, -errno);
         return;
     }
-    int timeout_id = -1;
+    clients[peer_fd] = new osd_client_t();
+    clients[peer_fd]->peer_addr = addr;
+    clients[peer_fd]->peer_port = peer_port;
+    clients[peer_fd]->peer_fd = peer_fd;
+    clients[peer_fd]->peer_state = PEER_CONNECTING;
+    clients[peer_fd]->connect_timeout_id = -1;
+    clients[peer_fd]->osd_num = peer_osd;
+    clients[peer_fd]->in_buf = malloc_or_die(receive_buffer_size);
+    tfd->set_fd_handler(peer_fd, true, [this](int peer_fd, int epoll_events)
+    {
+        // Either OUT (connected) or HUP
+        handle_connect_epoll(peer_fd);
+    });
     if (peer_connect_timeout > 0)
     {
-        timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, peer_fd](int timer_id)
+        clients[peer_fd]->connect_timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, peer_fd](int timer_id)
         {
             osd_num_t peer_osd = clients.at(peer_fd)->osd_num;
             stop_client(peer_fd, true);
@@ -208,20 +214,6 @@ void osd_messenger_t::try_connect_peer_addr(osd_num_t peer_osd, const char *peer
             return;
         });
     }
-    clients[peer_fd] = new osd_client_t((osd_client_t){
-        .peer_addr = addr,
-        .peer_port = peer_port,
-        .peer_fd = peer_fd,
-        .peer_state = PEER_CONNECTING,
-        .connect_timeout_id = timeout_id,
-        .osd_num = peer_osd,
-        .in_buf = malloc_or_die(receive_buffer_size),
-    });
-    tfd->set_fd_handler(peer_fd, true, [this](int peer_fd, int epoll_events)
-    {
-        // Either OUT (connected) or HUP
-        handle_connect_epoll(peer_fd);
-    });
 }
 
 void osd_messenger_t::handle_connect_epoll(int peer_fd)
@@ -357,6 +349,15 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
                 err = true;
                 printf("Connected to OSD %lu instead of OSD %lu, peer state is outdated, disconnecting peer\n", config["osd_num"].uint64_value(), cl->osd_num);
             }
+            else if (config["protocol_version"].uint64_value() != OSD_PROTOCOL_VERSION)
+            {
+                err = true;
+                printf(
+                    "OSD %lu protocol version is %lu, but only version %u is supported.\n"
+                    " If you need to upgrade from 0.5.x please request it via the issue tracker.\n",
+                    cl->osd_num, config["protocol_version"].uint64_value(), OSD_PROTOCOL_VERSION
+                );
+            }
         }
         if (err)
         {
@@ -373,123 +374,6 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
     outbox_push(op);
 }
 
-void osd_messenger_t::cancel_osd_ops(osd_client_t *cl)
-{
-    for (auto p: cl->sent_ops)
-    {
-        cancel_op(p.second);
-    }
-    cl->sent_ops.clear();
-    cl->outbox.clear();
-}
-
-void osd_messenger_t::cancel_op(osd_op_t *op)
-{
-    if (op->op_type == OSD_OP_OUT)
-    {
-        op->reply.hdr.magic = SECONDARY_OSD_REPLY_MAGIC;
-        op->reply.hdr.id = op->req.hdr.id;
-        op->reply.hdr.opcode = op->req.hdr.opcode;
-        op->reply.hdr.retval = -EPIPE;
-        // Copy lambda to be unaffected by `delete op`
-        std::function<void(osd_op_t*)>(op->callback)(op);
-    }
-    else
-    {
-        // This function is only called in stop_client(), so it's fine to destroy the operation
-        delete op;
-    }
-}
-
-void osd_messenger_t::stop_client(int peer_fd, bool force)
-{
-    assert(peer_fd != 0);
-    auto it = clients.find(peer_fd);
-    if (it == clients.end())
-    {
-        return;
-    }
-    uint64_t repeer_osd = 0;
-    osd_client_t *cl = it->second;
-    if (cl->peer_state == PEER_CONNECTED)
-    {
-        if (cl->osd_num)
-        {
-            // Reload configuration from etcd when the connection is dropped
-            if (log_level > 0)
-                printf("[OSD %lu] Stopping client %d (OSD peer %lu)\n", osd_num, peer_fd, cl->osd_num);
-            repeer_osd = cl->osd_num;
-        }
-        else
-        {
-            if (log_level > 0)
-                printf("[OSD %lu] Stopping client %d (regular client)\n", osd_num, peer_fd);
-        }
-    }
-    else if (!force)
-    {
-        return;
-    }
-    cl->peer_state = PEER_STOPPED;
-    clients.erase(it);
-    tfd->set_fd_handler(peer_fd, false, NULL);
-    if (cl->connect_timeout_id >= 0)
-    {
-        tfd->clear_timer(cl->connect_timeout_id);
-        cl->connect_timeout_id = -1;
-    }
-    if (cl->osd_num)
-    {
-        osd_peer_fds.erase(cl->osd_num);
-    }
-    if (cl->read_op)
-    {
-        if (cl->read_op->callback)
-        {
-            cancel_op(cl->read_op);
-        }
-        else
-        {
-            delete cl->read_op;
-        }
-        cl->read_op = NULL;
-    }
-    for (auto rit = read_ready_clients.begin(); rit != read_ready_clients.end(); rit++)
-    {
-        if (*rit == peer_fd)
-        {
-            read_ready_clients.erase(rit);
-            break;
-        }
-    }
-    for (auto wit = write_ready_clients.begin(); wit != write_ready_clients.end(); wit++)
-    {
-        if (*wit == peer_fd)
-        {
-            write_ready_clients.erase(wit);
-            break;
-        }
-    }
-    free(cl->in_buf);
-    cl->in_buf = NULL;
-    close(peer_fd);
-    if (repeer_osd)
-    {
-        // First repeer PGs as canceling OSD ops may push new operations
-        // and we need correct PG states when we do that
-        repeer_pgs(repeer_osd);
-    }
-    if (cl->osd_num)
-    {
-        // Cancel outbound operations
-        cancel_osd_ops(cl);
-    }
-    if (cl->refs <= 0)
-    {
-        delete cl;
-    }
-}
-
 void osd_messenger_t::accept_connections(int listen_fd)
 {
     // Accept new connections
@@ -505,13 +389,12 @@ void osd_messenger_t::accept_connections(int listen_fd)
         fcntl(peer_fd, F_SETFL, fcntl(peer_fd, F_GETFL, 0) | O_NONBLOCK);
         int one = 1;
         setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-        clients[peer_fd] = new osd_client_t((osd_client_t){
-            .peer_addr = addr,
-            .peer_port = ntohs(addr.sin_port),
-            .peer_fd = peer_fd,
-            .peer_state = PEER_CONNECTED,
-            .in_buf = malloc_or_die(receive_buffer_size),
-        });
+        clients[peer_fd] = new osd_client_t();
+        clients[peer_fd]->peer_addr = addr;
+        clients[peer_fd]->peer_port = ntohs(addr.sin_port);
+        clients[peer_fd]->peer_fd = peer_fd;
+        clients[peer_fd]->peer_state = PEER_CONNECTED;
+        clients[peer_fd]->in_buf = malloc_or_die(receive_buffer_size);
         // Add FD to epoll
         tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
         {

src/messenger.h

@@ -14,19 +14,15 @@
 
 #include "malloc_or_die.h"
 #include "json11/json11.hpp"
-#include "osd_ops.h"
+#include "msgr_op.h"
 #include "timerfd_manager.h"
-#include "ringloop.h"
-
-#define OSD_OP_IN 0
-#define OSD_OP_OUT 1
+#include <ringloop.h>
 
 #define CL_READ_HDR 1
 #define CL_READ_DATA 2
 #define CL_READ_REPLY_DATA 3
 #define CL_WRITE_READY 1
 #define CL_WRITE_REPLY 2
-#define OSD_OP_INLINE_BUF_COUNT 16
 
 #define PEER_CONNECTING 1
 #define PEER_CONNECTED 2
@@ -37,164 +33,6 @@
 #define DEFAULT_OSD_PING_TIMEOUT 5
 #define DEFAULT_BITMAP_GRANULARITY 4096
 
-// Kind of a vector with small-list-optimisation
-struct osd_op_buf_list_t
-{
-    int count = 0, alloc = OSD_OP_INLINE_BUF_COUNT, done = 0;
-    iovec *buf = NULL;
-    iovec inline_buf[OSD_OP_INLINE_BUF_COUNT];
-
-    inline osd_op_buf_list_t()
-    {
-        buf = inline_buf;
-    }
-
-    inline osd_op_buf_list_t(const osd_op_buf_list_t & other)
-    {
-        buf = inline_buf;
-        append(other);
-    }
-
-    inline osd_op_buf_list_t & operator = (const osd_op_buf_list_t & other)
-    {
-        reset();
-        append(other);
-        return *this;
-    }
-
-    inline ~osd_op_buf_list_t()
-    {
-        if (buf && buf != inline_buf)
-        {
-            free(buf);
-        }
-    }
-
-    inline void reset()
-    {
-        count = 0;
-        done = 0;
-    }
-
-    inline iovec* get_iovec()
-    {
-        return buf + done;
-    }
-
-    inline int get_size()
-    {
-        return count - done;
-    }
-
-    inline void append(const osd_op_buf_list_t & other)
-    {
-        if (count+other.count > alloc)
-        {
-            if (buf == inline_buf)
-            {
-                int old = alloc;
-                alloc = (((count+other.count+15)/16)*16);
-                buf = (iovec*)malloc(sizeof(iovec) * alloc);
-                if (!buf)
-                {
-                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
-                    exit(1);
-                }
-                memcpy(buf, inline_buf, sizeof(iovec) * old);
-            }
-            else
-            {
-                alloc = (((count+other.count+15)/16)*16);
-                buf = (iovec*)realloc(buf, sizeof(iovec) * alloc);
-                if (!buf)
-                {
-                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
-                    exit(1);
-                }
-            }
-        }
-        for (int i = 0; i < other.count; i++)
-        {
-            buf[count++] = other.buf[i];
-        }
-    }
-
-    inline void push_back(void *nbuf, size_t len)
-    {
-        if (count >= alloc)
-        {
-            if (buf == inline_buf)
-            {
-                int old = alloc;
-                alloc = ((alloc/16)*16 + 1);
-                buf = (iovec*)malloc(sizeof(iovec) * alloc);
-                if (!buf)
-                {
-                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
-                    exit(1);
-                }
-                memcpy(buf, inline_buf, sizeof(iovec)*old);
-            }
-            else
-            {
-                alloc = alloc < 16 ? 16 : (alloc+16);
-                buf = (iovec*)realloc(buf, sizeof(iovec) * alloc);
-                if (!buf)
-                {
-                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
-                    exit(1);
-                }
-            }
-        }
-        buf[count++] = { .iov_base = nbuf, .iov_len = len };
-    }
-
-    inline void eat(int result)
-    {
-        while (result > 0 && done < count)
-        {
-            iovec & iov = buf[done];
-            if (iov.iov_len <= result)
-            {
-                result -= iov.iov_len;
-                done++;
-            }
-            else
-            {
-                iov.iov_len -= result;
-                iov.iov_base += result;
-                break;
-            }
-        }
-    }
-};
-
-struct blockstore_op_t;
-
-struct osd_primary_op_data_t;
-
-struct osd_op_t
-{
-    timespec tv_begin = { 0 }, tv_end = { 0 };
-    uint64_t op_type = OSD_OP_IN;
-    int peer_fd;
-    osd_any_op_t req;
-    osd_any_reply_t reply;
-    blockstore_op_t *bs_op = NULL;
-    void *buf = NULL;
-    // bitmap, bitmap_len, bmp_data are only meaningful for reads
-    void *bitmap = NULL;
-    unsigned bitmap_len = 0;
-    unsigned bmp_data = 0;
-    void *rmw_buf = NULL;
-    osd_primary_op_data_t* op_data = NULL;
-    std::function<void(osd_op_t*)> callback;
-
-    osd_op_buf_list_t iov;
-
-    ~osd_op_t();
-};
-
 struct osd_client_t
 {
     int refs = 0;
@@ -233,6 +71,12 @@ struct osd_client_t
     int write_state = 0;
     std::vector<iovec> send_list, next_send_list;
     std::vector<osd_op_t*> outbox, next_outbox;
+
+    ~osd_client_t()
+    {
+        free(in_buf);
+        in_buf = NULL;
+    }
 };
 
 struct osd_wanted_peer_t
@@ -257,12 +101,9 @@ struct osd_op_stats_t
 
 struct osd_messenger_t
 {
-    timerfd_manager_t *tfd;
-    ring_loop_t *ringloop;
+protected:
     int keepalive_timer_id = -1;
 
-    // osd_num_t is only for logging and asserts
-    osd_num_t osd_num;
     // FIXME: make receive_buffer_size configurable
     int receive_buffer_size = 64*1024;
     int peer_connect_interval = DEFAULT_PEER_CONNECT_INTERVAL;
@@ -272,19 +113,22 @@ struct osd_messenger_t
     int log_level = 0;
     bool use_sync_send_recv = false;
 
-    std::map<osd_num_t, osd_wanted_peer_t> wanted_peers;
-    std::map<uint64_t, int> osd_peer_fds;
-    uint64_t next_subop_id = 1;
-
-    std::map<int, osd_client_t*> clients;
     std::vector<int> read_ready_clients;
     std::vector<int> write_ready_clients;
     std::vector<std::function<void()>> set_immediate;
 
+public:
+    timerfd_manager_t *tfd;
+    ring_loop_t *ringloop;
+    // 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;
     // op statistics
     osd_op_stats_t stats;
 
-public:
     void init();
     void parse_config(const json11::Json & config);
     void connect_peer(uint64_t osd_num, json11::Json peer_state);
@@ -292,7 +136,6 @@ public:
     void outbox_push(osd_op_t *cur_op);
     std::function<void(osd_op_t*)> exec_op;
     std::function<void(osd_num_t)> repeer_pgs;
-    void handle_peer_epoll(int peer_fd, int epoll_events);
     void read_requests();
     void send_replies();
     void accept_connections(int listen_fd);
@@ -301,6 +144,7 @@ public:
 protected:
     void try_connect_peer(uint64_t osd_num);
     void try_connect_peer_addr(osd_num_t peer_osd, const char *peer_host, int peer_port);
+    void handle_peer_epoll(int peer_fd, int epoll_events);
     void handle_connect_epoll(int peer_fd);
     void on_connect_peer(osd_num_t peer_osd, int peer_fd);
     void check_peer_config(osd_client_t *cl);

src/mock/build.sh

@@ -0,0 +1 @@
+g++ -D__MOCK__ -fsanitize=address -g -Wno-pointer-arith pg_states.cpp osd_ops.cpp test_cluster_client.cpp cluster_client.cpp msgr_op.cpp msgr_stop.cpp mock/messenger.cpp etcd_state_client.cpp timerfd_manager.cpp ../json11/json11.cpp -I mock -I . -I ..; ./a.out

src/mock/messenger.cpp

@@ -0,0 +1,44 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <unistd.h>
+#include <stdexcept>
+#include <assert.h>
+
+#include "messenger.h"
+
+void osd_messenger_t::init()
+{
+}
+
+osd_messenger_t::~osd_messenger_t()
+{
+    while (clients.size() > 0)
+    {
+        stop_client(clients.begin()->first, true);
+    }
+}
+
+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;
+}
+
+void osd_messenger_t::parse_config(const json11::Json & config)
+{
+}
+
+void osd_messenger_t::connect_peer(uint64_t peer_osd, json11::Json peer_state)
+{
+    wanted_peers[peer_osd] = (osd_wanted_peer_t){
+        .port = 1,
+    };
+}
+
+void osd_messenger_t::read_requests()
+{
+}
+
+void osd_messenger_t::send_replies()
+{
+}

src/mock/ringloop.h

@@ -0,0 +1,25 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#pragma once
+
+#include <functional>
+
+struct ring_consumer_t
+{
+    std::function<void(void)> loop;
+};
+
+class ring_loop_t
+{
+public:
+    void register_consumer(ring_consumer_t *consumer)
+    {
+    }
+    void unregister_consumer(ring_consumer_t *consumer)
+    {
+    }
+    void submit()
+    {
+    }
+};

src/msgr_op.cpp

@@ -0,0 +1,22 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <assert.h>
+
+#include "msgr_op.h"
+
+osd_op_t::~osd_op_t()
+{
+    assert(!bs_op);
+    assert(!op_data);
+    if (rmw_buf)
+    {
+        free(rmw_buf);
+    }
+    if (buf)
+    {
+        // Note: reusing osd_op_t WILL currently lead to memory leaks
+        // So we don't reuse it, but free it every time
+        free(buf);
+    }
+}

src/msgr_op.h

@@ -0,0 +1,175 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#pragma once
+
+#include <sys/uio.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+
+#include "osd_ops.h"
+
+#define OSD_OP_IN 0
+#define OSD_OP_OUT 1
+
+#define OSD_OP_INLINE_BUF_COUNT 16
+
+// Kind of a vector with small-list-optimisation
+struct osd_op_buf_list_t
+{
+    int count = 0, alloc = OSD_OP_INLINE_BUF_COUNT, done = 0;
+    iovec *buf = NULL;
+    iovec inline_buf[OSD_OP_INLINE_BUF_COUNT];
+
+    inline osd_op_buf_list_t()
+    {
+        buf = inline_buf;
+    }
+
+    inline osd_op_buf_list_t(const osd_op_buf_list_t & other)
+    {
+        buf = inline_buf;
+        append(other);
+    }
+
+    inline osd_op_buf_list_t & operator = (const osd_op_buf_list_t & other)
+    {
+        reset();
+        append(other);
+        return *this;
+    }
+
+    inline ~osd_op_buf_list_t()
+    {
+        if (buf && buf != inline_buf)
+        {
+            free(buf);
+        }
+    }
+
+    inline void reset()
+    {
+        count = 0;
+        done = 0;
+    }
+
+    inline iovec* get_iovec()
+    {
+        return buf + done;
+    }
+
+    inline int get_size()
+    {
+        return count - done;
+    }
+
+    inline void append(const osd_op_buf_list_t & other)
+    {
+        if (count+other.count > alloc)
+        {
+            if (buf == inline_buf)
+            {
+                int old = alloc;
+                alloc = (((count+other.count+15)/16)*16);
+                buf = (iovec*)malloc(sizeof(iovec) * alloc);
+                if (!buf)
+                {
+                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
+                    exit(1);
+                }
+                memcpy(buf, inline_buf, sizeof(iovec) * old);
+            }
+            else
+            {
+                alloc = (((count+other.count+15)/16)*16);
+                buf = (iovec*)realloc(buf, sizeof(iovec) * alloc);
+                if (!buf)
+                {
+                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
+                    exit(1);
+                }
+            }
+        }
+        for (int i = 0; i < other.count; i++)
+        {
+            buf[count++] = other.buf[i];
+        }
+    }
+
+    inline void push_back(void *nbuf, size_t len)
+    {
+        if (count >= alloc)
+        {
+            if (buf == inline_buf)
+            {
+                int old = alloc;
+                alloc = ((alloc/16)*16 + 1);
+                buf = (iovec*)malloc(sizeof(iovec) * alloc);
+                if (!buf)
+                {
+                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
+                    exit(1);
+                }
+                memcpy(buf, inline_buf, sizeof(iovec)*old);
+            }
+            else
+            {
+                alloc = alloc < 16 ? 16 : (alloc+16);
+                buf = (iovec*)realloc(buf, sizeof(iovec) * alloc);
+                if (!buf)
+                {
+                    printf("Failed to allocate %lu bytes\n", sizeof(iovec) * alloc);
+                    exit(1);
+                }
+            }
+        }
+        buf[count++] = { .iov_base = nbuf, .iov_len = len };
+    }
+
+    inline void eat(int result)
+    {
+        while (result > 0 && done < count)
+        {
+            iovec & iov = buf[done];
+            if (iov.iov_len <= result)
+            {
+                result -= iov.iov_len;
+                done++;
+            }
+            else
+            {
+                iov.iov_len -= result;
+                iov.iov_base += result;
+                break;
+            }
+        }
+    }
+};
+
+struct blockstore_op_t;
+
+struct osd_primary_op_data_t;
+
+struct osd_op_t
+{
+    timespec tv_begin = { 0 }, tv_end = { 0 };
+    uint64_t op_type = OSD_OP_IN;
+    int peer_fd;
+    osd_any_op_t req;
+    osd_any_reply_t reply;
+    blockstore_op_t *bs_op = NULL;
+    void *buf = NULL;
+    // bitmap, bitmap_len, bmp_data are only meaningful for reads
+    void *bitmap = NULL;
+    unsigned bitmap_len = 0;
+    unsigned bmp_data = 0;
+    void *rmw_buf = NULL;
+    osd_primary_op_data_t* op_data = NULL;
+    std::function<void(osd_op_t*)> callback;
+
+    osd_op_buf_list_t iov;
+
+    ~osd_op_t();
+};

src/msgr_receive.cpp

@@ -232,6 +232,15 @@ void osd_messenger_t::handle_op_hdr(osd_client_t *cl)
         }
         cl->read_remaining = cur_op->req.sec_stab.len;
     }
+    else if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
+    {
+        if (cur_op->req.sec_read_bmp.len > 0)
+        {
+            cur_op->buf = memalign_or_die(MEM_ALIGNMENT, cur_op->req.sec_read_bmp.len);
+            cl->recv_list.push_back(cur_op->buf, cur_op->req.sec_read_bmp.len);
+        }
+        cl->read_remaining = cur_op->req.sec_read_bmp.len;
+    }
     else if (cur_op->req.hdr.opcode == OSD_OP_READ)
     {
         cl->read_remaining = 0;
@@ -277,17 +286,19 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
     {
         // Read data. In this case we assume that the buffer is preallocated by the caller (!)
         unsigned bmp_len = (op->reply.hdr.opcode == OSD_OP_SEC_READ ? op->reply.sec_rw.attr_len : op->reply.rw.bitmap_len);
-        if (op->reply.hdr.retval != (op->reply.hdr.opcode == OSD_OP_SEC_READ ? op->req.sec_rw.len : op->req.rw.len) ||
-            bmp_len > op->bitmap_len)
+        unsigned expected_size = (op->reply.hdr.opcode == OSD_OP_SEC_READ ? op->req.sec_rw.len : op->req.rw.len);
+        if (op->reply.hdr.retval >= 0 && (op->reply.hdr.retval != expected_size || bmp_len > op->bitmap_len))
         {
             // Check reply length to not overflow the buffer
-            printf("Client %d read reply of different length\n", cl->peer_fd);
+            printf("Client %d read reply of different length: expected %u+%u, got %ld+%u\n",
+                cl->peer_fd, expected_size, op->bitmap_len, op->reply.hdr.retval, bmp_len);
             cl->sent_ops[op->req.hdr.id] = op;
             stop_client(cl->peer_fd);
             return false;
         }
-        if (bmp_len > 0)
+        if (op->reply.hdr.retval >= 0 && bmp_len > 0)
         {
+            assert(op->bitmap);
             cl->recv_list.push_back(op->bitmap, bmp_len);
         }
         if (op->reply.hdr.retval > 0)
@@ -314,6 +325,16 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
         op->buf = memalign_or_die(MEM_ALIGNMENT, cl->read_remaining);
         cl->recv_list.push_back(op->buf, cl->read_remaining);
     }
+    else if (op->reply.hdr.opcode == OSD_OP_SEC_READ_BMP && op->reply.hdr.retval > 0)
+    {
+        assert(!op->iov.count);
+        delete cl->read_op;
+        cl->read_op = op;
+        cl->read_state = CL_READ_REPLY_DATA;
+        cl->read_remaining = op->reply.hdr.retval;
+        op->buf = memalign_or_die(MEM_ALIGNMENT, cl->read_remaining);
+        cl->recv_list.push_back(op->buf, cl->read_remaining);
+    }
     else if (op->reply.hdr.opcode == OSD_OP_SHOW_CONFIG && op->reply.hdr.retval > 0)
     {
         assert(!op->iov.count);

src/msgr_send.cpp

@@ -87,6 +87,14 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             to_outbox.push_back(NULL);
         }
     }
+    if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
+    {
+        if (cur_op->op_type == OSD_OP_IN && cur_op->reply.hdr.retval > 0)
+            to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->reply.hdr.retval });
+        else if (cur_op->op_type == OSD_OP_OUT && cur_op->req.sec_read_bmp.len > 0)
+            to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len });
+        to_outbox.push_back(NULL);
+    }
     if (cur_op->op_type == OSD_OP_IN)
     {
         // To free it later
@@ -203,7 +211,7 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
     cl->refs--;
     if (cl->peer_state == PEER_STOPPED)
     {
-        if (!cl->refs)
+        if (cl->refs <= 0)
         {
             delete cl;
         }

src/msgr_stop.cpp

@@ -0,0 +1,137 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <unistd.h>
+#include <assert.h>
+
+#include "messenger.h"
+
+void osd_messenger_t::cancel_osd_ops(osd_client_t *cl)
+{
+    std::vector<osd_op_t*> cancel_ops;
+    cancel_ops.resize(cl->sent_ops.size());
+    int i = 0;
+    for (auto p: cl->sent_ops)
+    {
+        cancel_ops[i++] = p.second;
+    }
+    cl->sent_ops.clear();
+    cl->outbox.clear();
+    for (auto op: cancel_ops)
+    {
+        cancel_op(op);
+    }
+}
+
+void osd_messenger_t::cancel_op(osd_op_t *op)
+{
+    if (op->op_type == OSD_OP_OUT)
+    {
+        op->reply.hdr.magic = SECONDARY_OSD_REPLY_MAGIC;
+        op->reply.hdr.id = op->req.hdr.id;
+        op->reply.hdr.opcode = op->req.hdr.opcode;
+        op->reply.hdr.retval = -EPIPE;
+        // Copy lambda to be unaffected by `delete op`
+        std::function<void(osd_op_t*)>(op->callback)(op);
+    }
+    else
+    {
+        // This function is only called in stop_client(), so it's fine to destroy the operation
+        delete op;
+    }
+}
+
+void osd_messenger_t::stop_client(int peer_fd, bool force)
+{
+    assert(peer_fd != 0);
+    auto it = clients.find(peer_fd);
+    if (it == clients.end())
+    {
+        return;
+    }
+    osd_client_t *cl = it->second;
+    if (cl->peer_state == PEER_CONNECTING && !force || cl->peer_state == PEER_STOPPED)
+    {
+        return;
+    }
+    if (log_level > 0)
+    {
+        if (cl->osd_num)
+        {
+            printf("[OSD %lu] Stopping client %d (OSD peer %lu)\n", osd_num, peer_fd, cl->osd_num);
+        }
+        else
+        {
+            printf("[OSD %lu] Stopping client %d (regular client)\n", osd_num, peer_fd);
+        }
+    }
+    // First set state to STOPPED so another stop_client() call doesn't try to free it again
+    cl->refs++;
+    cl->peer_state = PEER_STOPPED;
+    if (cl->osd_num)
+    {
+        // ...and forget OSD peer
+        osd_peer_fds.erase(cl->osd_num);
+    }
+#ifndef __MOCK__
+    // Then remove FD from the eventloop so we don't accidentally read something
+    tfd->set_fd_handler(peer_fd, false, NULL);
+    if (cl->connect_timeout_id >= 0)
+    {
+        tfd->clear_timer(cl->connect_timeout_id);
+        cl->connect_timeout_id = -1;
+    }
+    for (auto rit = read_ready_clients.begin(); rit != read_ready_clients.end(); rit++)
+    {
+        if (*rit == peer_fd)
+        {
+            read_ready_clients.erase(rit);
+            break;
+        }
+    }
+    for (auto wit = write_ready_clients.begin(); wit != write_ready_clients.end(); wit++)
+    {
+        if (*wit == peer_fd)
+        {
+            write_ready_clients.erase(wit);
+            break;
+        }
+    }
+#endif
+    if (cl->osd_num)
+    {
+        // Then repeer PGs because cancel_op() callbacks can try to perform
+        // some actions and we need correct PG states to not do something silly
+        repeer_pgs(cl->osd_num);
+    }
+    // Then cancel all operations
+    if (cl->read_op)
+    {
+        if (!cl->read_op->callback)
+        {
+            delete cl->read_op;
+        }
+        cl->read_op = NULL;
+    }
+    if (cl->osd_num)
+    {
+        // Cancel outbound operations
+        cancel_osd_ops(cl);
+    }
+#ifndef __MOCK__
+    // And close the FD only when everything is done
+    // ...because peer_fd number can get reused after close()
+    close(peer_fd);
+#endif
+    // Find the item again because it can be invalidated at this point
+    it = clients.find(peer_fd);
+    if (it != clients.end())
+    {
+        clients.erase(it);
+    }
+    cl->refs--;
+    if (cl->refs <= 0)
+    {
+        delete cl;
+    }
+}

src/osd.cpp

@@ -8,6 +8,7 @@
 #include <arpa/inet.h>
 
 #include "osd.h"
+#include "http_client.h"
 
 osd_t::osd_t(blockstore_config_t & config, ring_loop_t *ringloop)
 {
@@ -19,17 +20,22 @@ osd_t::osd_t(blockstore_config_t & config, ring_loop_t *ringloop)
         bs_bitmap_granularity = DEFAULT_BITMAP_GRANULARITY;
     clean_entry_bitmap_size = bs_block_size / bs_bitmap_granularity / 8;
 
+    zero_buffer_size = 1<<20;
+    zero_buffer = malloc_or_die(zero_buffer_size);
+    memset(zero_buffer, 0, zero_buffer_size);
+
     this->config = config;
     this->ringloop = ringloop;
 
+    epmgr = new epoll_manager_t(ringloop);
+    // FIXME: Use timerfd_interval based directly on io_uring
+    this->tfd = epmgr->tfd;
+
     // FIXME: Create Blockstore from on-disk superblock config and check it against the OSD cluster config
-    this->bs = new blockstore_t(config, ringloop);
+    this->bs = new blockstore_t(config, ringloop, tfd);
 
     parse_config(config);
 
-    epmgr = new epoll_manager_t(ringloop);
-    this->tfd = epmgr->tfd;
-
     this->tfd->set_timer(print_stats_interval*1000, true, [this](int timer_id)
     {
         print_stats();
@@ -57,6 +63,7 @@ osd_t::~osd_t()
     delete epmgr;
     delete bs;
     close(listen_fd);
+    free(zero_buffer);
 }
 
 void osd_t::parse_config(blockstore_config_t & config)
@@ -198,6 +205,8 @@ void osd_t::exec_op(osd_op_t *cur_op)
         delete cur_op;
         return;
     }
+    // Clear the reply buffer
+    memset(cur_op->reply.buf, 0, OSD_PACKET_SIZE);
     inflight_ops++;
     if (cur_op->req.hdr.magic != SECONDARY_OSD_OP_MAGIC ||
         cur_op->req.hdr.opcode < OSD_OP_MIN || cur_op->req.hdr.opcode > OSD_OP_MAX ||
@@ -228,6 +237,7 @@ void osd_t::exec_op(osd_op_t *cur_op)
         cur_op->req.hdr.opcode != OSD_OP_SEC_READ &&
         cur_op->req.hdr.opcode != OSD_OP_SEC_LIST &&
         cur_op->req.hdr.opcode != OSD_OP_READ &&
+        cur_op->req.hdr.opcode != OSD_OP_SEC_READ_BMP &&
         cur_op->req.hdr.opcode != OSD_OP_SHOW_CONFIG)
     {
         // Readonly mode

src/osd.h

@@ -66,6 +66,28 @@ struct inode_stats_t
     uint64_t op_bytes[3] = { 0 };
 };
 
+struct bitmap_request_t
+{
+    osd_num_t osd_num;
+    object_id oid;
+    uint64_t version;
+    void *bmp_buf;
+};
+
+inline bool operator < (const bitmap_request_t & a, const bitmap_request_t & b)
+{
+    return a.osd_num < b.osd_num || a.osd_num == b.osd_num && a.oid < b.oid;
+}
+
+struct osd_chain_read_t
+{
+    int chain_pos;
+    inode_t inode;
+    uint32_t offset, len;
+};
+
+struct osd_rmw_stripe_t;
+
 class osd_t
 {
     // config
@@ -126,6 +148,8 @@ class osd_t
     bool stopping = false;
     int inflight_ops = 0;
     blockstore_t *bs;
+    void *zero_buffer = NULL;
+    uint64_t zero_buffer_size = 0;
     uint32_t bs_block_size, bs_bitmap_granularity, clean_entry_bitmap_size;
     ring_loop_t *ringloop;
     timerfd_manager_t *tfd = NULL;
@@ -147,7 +171,7 @@ class osd_t
     void init_cluster();
     void on_change_osd_state_hook(osd_num_t peer_osd);
     void on_change_pg_history_hook(pool_id_t pool_id, pg_num_t pg_num);
-    void on_change_etcd_state_hook(json11::Json::object & changes);
+    void on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes);
     void on_load_config_hook(json11::Json::object & changes);
     json11::Json on_load_pgs_checks_hook();
     void on_load_pgs_hook(bool success);
@@ -210,23 +234,37 @@ class osd_t
     void continue_primary_del(osd_op_t *cur_op);
     bool check_write_queue(osd_op_t *cur_op, pg_t & pg);
     void remove_object_from_state(object_id & oid, pg_osd_set_state_t *object_state, pg_t &pg);
+    void free_object_state(pg_t & pg, pg_osd_set_state_t **object_state);
     bool remember_unstable_write(osd_op_t *cur_op, pg_t & pg, pg_osd_set_t & loc_set, int base_state);
     void handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op);
     void handle_primary_bs_subop(osd_op_t *subop);
     void add_bs_subop_stats(osd_op_t *subop);
     void pg_cancel_write_queue(pg_t & pg, osd_op_t *first_op, object_id oid, int retval);
-    void submit_primary_subops(int submit_type, uint64_t op_version, int pg_size, const uint64_t* osd_set, osd_op_t *cur_op);
+
+    void submit_primary_subops(int submit_type, uint64_t op_version, const uint64_t* osd_set, osd_op_t *cur_op);
+    int submit_primary_subop_batch(int submit_type, inode_t inode, uint64_t op_version,
+        osd_rmw_stripe_t *stripes, const uint64_t* osd_set, osd_op_t *cur_op, int subop_idx, int zero_read);
     void submit_primary_del_subops(osd_op_t *cur_op, uint64_t *cur_set, uint64_t set_size, pg_osd_set_t & loc_set);
     void submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_delete, int chunks_to_delete_count);
-    void submit_primary_sync_subops(osd_op_t *cur_op);
+    int submit_primary_sync_subops(osd_op_t *cur_op);
     void submit_primary_stab_subops(osd_op_t *cur_op);
 
+    uint64_t* get_object_osd_set(pg_t &pg, object_id &oid, uint64_t *def, pg_osd_set_state_t **object_state);
+
+    void continue_chained_read(osd_op_t *cur_op);
+    int submit_chained_read_requests(pg_t & pg, osd_op_t *cur_op);
+    void send_chained_read_results(pg_t & pg, osd_op_t *cur_op);
+    std::vector<osd_chain_read_t> collect_chained_read_requests(osd_op_t *cur_op);
+    int collect_bitmap_requests(osd_op_t *cur_op, pg_t & pg, std::vector<bitmap_request_t> & bitmap_requests);
+    int submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg);
+    int read_bitmaps(osd_op_t *cur_op, pg_t & pg, int base_state);
+
     inline pg_num_t map_to_pg(object_id oid, uint64_t pg_stripe_size)
     {
         uint64_t pg_count = pg_counts[INODE_POOL(oid.inode)];
         if (!pg_count)
             pg_count = 1;
-        return (oid.inode + oid.stripe / pg_stripe_size) % pg_count + 1;
+        return (oid.stripe / pg_stripe_size) % pg_count + 1;
     }
 
 public:

src/osd_cluster.cpp

@@ -4,6 +4,7 @@
 #include "osd.h"
 #include "base64.h"
 #include "etcd_state_client.h"
+#include "http_client.h"
 #include "osd_rmw.h"
 
 // Startup sequence:
@@ -64,7 +65,7 @@ void osd_t::init_cluster()
         st_cli.log_level = log_level;
         st_cli.on_change_osd_state_hook = [this](osd_num_t peer_osd) { on_change_osd_state_hook(peer_osd); };
         st_cli.on_change_pg_history_hook = [this](pool_id_t pool_id, pg_num_t pg_num) { on_change_pg_history_hook(pool_id, pg_num); };
-        st_cli.on_change_hook = [this](json11::Json::object & changes) { on_change_etcd_state_hook(changes); };
+        st_cli.on_change_hook = [this](std::map<std::string, etcd_kv_t> & changes) { on_change_etcd_state_hook(changes); };
         st_cli.on_load_config_hook = [this](json11::Json::object & cfg) { on_load_config_hook(cfg); };
         st_cli.load_pgs_checks_hook = [this]() { return on_load_pgs_checks_hook(); };
         st_cli.on_load_pgs_hook = [this](bool success) { on_load_pgs_hook(success); };
@@ -182,14 +183,38 @@ void osd_t::report_statistics()
     // Report space usage statistics as a whole
     // Maybe we'll report it using deltas if we tune for a lot of inodes at some point
     json11::Json::object inode_space;
+    json11::Json::object last_stat;
+    pool_id_t last_pool = 0;
     for (auto kv: bs->get_inode_space_stats())
     {
-        inode_space[std::to_string(kv.first)] = kv.second;
+        pool_id_t pool_id = INODE_POOL(kv.first);
+        uint64_t only_inode_num = (kv.first & ((1l << (64-POOL_ID_BITS)) - 1));
+        if (!last_pool || pool_id != last_pool)
+        {
+            if (last_pool)
+                inode_space[std::to_string(last_pool)] = last_stat;
+            last_stat = json11::Json::object();
+            last_pool = pool_id;
+        }
+        last_stat[std::to_string(only_inode_num)] = kv.second;
     }
+    if (last_pool)
+        inode_space[std::to_string(last_pool)] = last_stat;
+    last_stat = json11::Json::object();
+    last_pool = 0;
     json11::Json::object inode_ops;
     for (auto kv: inode_stats)
     {
-        inode_ops[std::to_string(kv.first)] = json11::Json::object {
+        pool_id_t pool_id = INODE_POOL(kv.first);
+        uint64_t only_inode_num = (kv.first & ((1l << (64-POOL_ID_BITS)) - 1));
+        if (!last_pool || pool_id != last_pool)
+        {
+            if (last_pool)
+                inode_ops[std::to_string(last_pool)] = last_stat;
+            last_stat = json11::Json::object();
+            last_pool = pool_id;
+        }
+        last_stat[std::to_string(only_inode_num)] = json11::Json::object {
             { "read", json11::Json::object {
                 { "count", kv.second.op_count[INODE_STATS_READ] },
                 { "usec", kv.second.op_sum[INODE_STATS_READ] },
@@ -207,20 +232,28 @@ void osd_t::report_statistics()
             } },
         };
     }
-    json11::Json::array txn = { json11::Json::object {
-        { "request_put", json11::Json::object {
-            { "key", base64_encode(st_cli.etcd_prefix+"/osd/stats/"+std::to_string(osd_num)) },
-            { "value", base64_encode(get_statistics().dump()) },
-        } },
-        { "request_put", json11::Json::object {
-            { "key", base64_encode(st_cli.etcd_prefix+"/osd/space/"+std::to_string(osd_num)) },
-            { "value", base64_encode(json11::Json(inode_space).dump()) },
-        } },
-        { "request_put", json11::Json::object {
-            { "key", base64_encode(st_cli.etcd_prefix+"/osd/inodestats/"+std::to_string(osd_num)) },
-            { "value", base64_encode(json11::Json(inode_ops).dump()) },
-        } },
-    } };
+    if (last_pool)
+        inode_ops[std::to_string(last_pool)] = last_stat;
+    json11::Json::array txn = {
+        json11::Json::object {
+            { "request_put", json11::Json::object {
+                { "key", base64_encode(st_cli.etcd_prefix+"/osd/stats/"+std::to_string(osd_num)) },
+                { "value", base64_encode(get_statistics().dump()) },
+            } },
+        },
+        json11::Json::object {
+            { "request_put", json11::Json::object {
+                { "key", base64_encode(st_cli.etcd_prefix+"/osd/space/"+std::to_string(osd_num)) },
+                { "value", base64_encode(json11::Json(inode_space).dump()) },
+            } },
+        },
+        json11::Json::object {
+            { "request_put", json11::Json::object {
+                { "key", base64_encode(st_cli.etcd_prefix+"/osd/inodestats/"+std::to_string(osd_num)) },
+                { "value", base64_encode(json11::Json(inode_ops).dump()) },
+            } },
+        },
+    };
     for (auto & p: pgs)
     {
         auto & pg = p.second;
@@ -271,7 +304,7 @@ void osd_t::on_change_osd_state_hook(osd_num_t peer_osd)
     }
 }
 
-void osd_t::on_change_etcd_state_hook(json11::Json::object & changes)
+void osd_t::on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes)
 {
     // FIXME apply config changes in runtime (maybe, some)
     if (run_primary)
@@ -593,7 +626,7 @@ void osd_t::apply_pg_config()
                 }
                 if (currently_taken)
                 {
-                    if (pg_it->second.state & (PG_ACTIVE | PG_INCOMPLETE | PG_PEERING))
+                    if (pg_it->second.state & (PG_ACTIVE | PG_INCOMPLETE | PG_PEERING | PG_REPEERING))
                     {
                         if (pg_it->second.target_set == pg_cfg.target_set)
                         {

src/osd_flush.cpp

@@ -149,10 +149,14 @@ void osd_t::handle_flush_op(bool rollback, pool_id_t pool_id, pg_num_t pg_num, p
         {
             continue_primary_write(op);
         }
-        if (pg.inflight == 0 && (pg.state & PG_STOPPING))
+        if ((pg.state & PG_STOPPING) && pg.inflight == 0 && !pg.flush_batch)
         {
             finish_stop_pg(pg);
         }
+        else if ((pg.state & PG_REPEERING) && pg.inflight == 0 && !pg.flush_batch)
+        {
+            start_pg_peering(pg);
+        }
     }
 }
 
@@ -231,7 +235,8 @@ bool osd_t::pick_next_recovery(osd_recovery_op_t &op)
     {
         for (auto pg_it = pgs.begin(); pg_it != pgs.end(); pg_it++)
         {
-            if ((pg_it->second.state & (PG_ACTIVE | PG_HAS_MISPLACED)) == (PG_ACTIVE | PG_HAS_MISPLACED))
+            // Don't try to "recover" misplaced objects if "recovery" would make them degraded
+            if ((pg_it->second.state & (PG_ACTIVE | PG_DEGRADED | PG_HAS_MISPLACED)) == (PG_ACTIVE | PG_HAS_MISPLACED))
             {
                 for (auto obj_it = pg_it->second.misplaced_objects.begin(); obj_it != pg_it->second.misplaced_objects.end(); obj_it++)
                 {

src/osd_ops.cpp

@@ -20,4 +20,5 @@ const char* osd_op_names[] = {
     "primary_sync",
     "primary_delete",
     "ping",
+    "sec_read_bmp",
 };

src/osd_ops.h

@@ -28,12 +28,14 @@
 #define OSD_OP_SYNC                 13
 #define OSD_OP_DELETE               14
 #define OSD_OP_PING                 15
-#define OSD_OP_MAX                  15
+#define OSD_OP_SEC_READ_BMP         16
+#define OSD_OP_MAX                  16
 // Alignment & limit for read/write operations
 #ifndef MEM_ALIGNMENT
 #define MEM_ALIGNMENT               512
 #endif
 #define OSD_RW_MAX                  64*1024*1024
+#define OSD_PROTOCOL_VERSION        1
 
 // common request and reply headers
 struct __attribute__((__packed__)) osd_op_header_t
@@ -59,7 +61,7 @@ struct __attribute__((__packed__)) osd_reply_header_t
 };
 
 // read or write to the secondary OSD
-struct __attribute__((__packed__)) osd_op_secondary_rw_t
+struct __attribute__((__packed__)) osd_op_sec_rw_t
 {
     osd_op_header_t header;
     // object
@@ -76,7 +78,7 @@ struct __attribute__((__packed__)) osd_op_secondary_rw_t
     uint32_t pad0;
 };
 
-struct __attribute__((__packed__)) osd_reply_secondary_rw_t
+struct __attribute__((__packed__)) osd_reply_sec_rw_t
 {
     osd_reply_header_t header;
     // for reads and writes: assigned or read version number
@@ -87,7 +89,7 @@ struct __attribute__((__packed__)) osd_reply_secondary_rw_t
 };
 
 // delete object on the secondary OSD
-struct __attribute__((__packed__)) osd_op_secondary_del_t
+struct __attribute__((__packed__)) osd_op_sec_del_t
 {
     osd_op_header_t header;
     // object
@@ -96,37 +98,51 @@ struct __attribute__((__packed__)) osd_op_secondary_del_t
     uint64_t version;
 };
 
-struct __attribute__((__packed__)) osd_reply_secondary_del_t
+struct __attribute__((__packed__)) osd_reply_sec_del_t
 {
     osd_reply_header_t header;
     uint64_t version;
 };
 
 // sync to the secondary OSD
-struct __attribute__((__packed__)) osd_op_secondary_sync_t
+struct __attribute__((__packed__)) osd_op_sec_sync_t
 {
     osd_op_header_t header;
 };
 
-struct __attribute__((__packed__)) osd_reply_secondary_sync_t
+struct __attribute__((__packed__)) osd_reply_sec_sync_t
 {
     osd_reply_header_t header;
 };
 
 // stabilize or rollback objects on the secondary OSD
-struct __attribute__((__packed__)) osd_op_secondary_stabilize_t
+struct __attribute__((__packed__)) osd_op_sec_stab_t
 {
     osd_op_header_t header;
     // obj_ver_id array length in bytes
     uint64_t len;
 };
-typedef osd_op_secondary_stabilize_t osd_op_secondary_rollback_t;
+typedef osd_op_sec_stab_t osd_op_sec_rollback_t;
 
-struct __attribute__((__packed__)) osd_reply_secondary_stabilize_t
+struct __attribute__((__packed__)) osd_reply_sec_stab_t
 {
     osd_reply_header_t header;
 };
-typedef osd_reply_secondary_stabilize_t osd_reply_secondary_rollback_t;
+typedef osd_reply_sec_stab_t osd_reply_sec_rollback_t;
+
+// bulk read bitmaps from a secondary OSD
+struct __attribute__((__packed__)) osd_op_sec_read_bmp_t
+{
+    osd_op_header_t header;
+    // obj_ver_id array length in bytes
+    uint64_t len;
+};
+
+struct __attribute__((__packed__)) osd_reply_sec_read_bmp_t
+{
+    // retval is payload length in bytes. payload is {version,bitmap}[]
+    osd_reply_header_t header;
+};
 
 // show configuration
 struct __attribute__((__packed__)) osd_op_show_config_t
@@ -140,7 +156,7 @@ struct __attribute__((__packed__)) osd_reply_show_config_t
 };
 
 // list objects on replica
-struct __attribute__((__packed__)) osd_op_secondary_list_t
+struct __attribute__((__packed__)) osd_op_sec_list_t
 {
     osd_op_header_t header;
     // placement group total number and total count
@@ -151,7 +167,7 @@ struct __attribute__((__packed__)) osd_op_secondary_list_t
     uint64_t min_inode, max_inode;
 };
 
-struct __attribute__((__packed__)) osd_reply_secondary_list_t
+struct __attribute__((__packed__)) osd_reply_sec_list_t
 {
     osd_reply_header_t header;
     // stable object version count. header.retval = total object version count
@@ -169,6 +185,10 @@ struct __attribute__((__packed__)) osd_op_rw_t
     uint64_t offset;
     // length
     uint32_t len;
+    // flags (for future)
+    uint32_t flags;
+    // inode metadata revision
+    uint64_t meta_revision;
 };
 
 struct __attribute__((__packed__)) osd_reply_rw_t
@@ -194,11 +214,12 @@ struct __attribute__((__packed__)) osd_reply_sync_t
 union osd_any_op_t
 {
     osd_op_header_t hdr;
-    osd_op_secondary_rw_t sec_rw;
-    osd_op_secondary_del_t sec_del;
-    osd_op_secondary_sync_t sec_sync;
-    osd_op_secondary_stabilize_t sec_stab;
-    osd_op_secondary_list_t sec_list;
+    osd_op_sec_rw_t sec_rw;
+    osd_op_sec_del_t sec_del;
+    osd_op_sec_sync_t sec_sync;
+    osd_op_sec_stab_t sec_stab;
+    osd_op_sec_read_bmp_t sec_read_bmp;
+    osd_op_sec_list_t sec_list;
     osd_op_show_config_t show_conf;
     osd_op_rw_t rw;
     osd_op_sync_t sync;
@@ -208,11 +229,12 @@ union osd_any_op_t
 union osd_any_reply_t
 {
     osd_reply_header_t hdr;
-    osd_reply_secondary_rw_t sec_rw;
-    osd_reply_secondary_del_t sec_del;
-    osd_reply_secondary_sync_t sec_sync;
-    osd_reply_secondary_stabilize_t sec_stab;
-    osd_reply_secondary_list_t sec_list;
+    osd_reply_sec_rw_t sec_rw;
+    osd_reply_sec_del_t sec_del;
+    osd_reply_sec_sync_t sec_sync;
+    osd_reply_sec_stab_t sec_stab;
+    osd_reply_sec_read_bmp_t sec_read_bmp;
+    osd_reply_sec_list_t sec_list;
     osd_reply_show_config_t show_conf;
     osd_reply_rw_t rw;
     osd_reply_sync_t sync;

src/osd_peering.cpp

@@ -77,10 +77,11 @@ void osd_t::repeer_pgs(osd_num_t peer_osd)
     // Re-peer affected PGs
     for (auto & p: pgs)
     {
+        auto & pg = p.second;
         bool repeer = false;
-        if (p.second.state & (PG_PEERING | PG_ACTIVE | PG_INCOMPLETE))
+        if (pg.state & (PG_PEERING | PG_ACTIVE | PG_INCOMPLETE))
         {
-            for (osd_num_t pg_osd: p.second.all_peers)
+            for (osd_num_t pg_osd: pg.all_peers)
             {
                 if (pg_osd == peer_osd)
                 {
@@ -91,8 +92,17 @@ void osd_t::repeer_pgs(osd_num_t peer_osd)
             if (repeer)
             {
                 // Repeer this pg
-                printf("[PG %u/%u] Repeer because of OSD %lu\n", p.second.pool_id, p.second.pg_num, peer_osd);
-                start_pg_peering(p.second);
+                printf("[PG %u/%u] Repeer because of OSD %lu\n", pg.pool_id, pg.pg_num, peer_osd);
+                if (!(pg.state & (PG_ACTIVE | PG_REPEERING)) || pg.inflight == 0 && !pg.flush_batch)
+                {
+                    start_pg_peering(pg);
+                }
+                else
+                {
+                    // Stop accepting new operations, wait for current ones to finish or fail
+                    pg.state = pg.state & ~PG_ACTIVE | PG_REPEERING;
+                    report_pg_state(pg);
+                }
             }
         }
     }
@@ -334,9 +344,10 @@ void osd_t::submit_sync_and_list_subop(osd_num_t role_osd, pg_peering_state_t *p
             {
                 // FIXME: Mark peer as failed and don't reconnect immediately after dropping the connection
                 printf("Failed to sync OSD %lu: %ld (%s), disconnecting peer\n", role_osd, op->reply.hdr.retval, strerror(-op->reply.hdr.retval));
+                int fail_fd = op->peer_fd;
                 ps->list_ops.erase(role_osd);
-                c_cli.stop_client(op->peer_fd);
                 delete op;
+                c_cli.stop_client(fail_fd);
                 return;
             }
             delete op;
@@ -413,9 +424,10 @@ void osd_t::submit_list_subop(osd_num_t role_osd, pg_peering_state_t *ps)
             if (op->reply.hdr.retval < 0)
             {
                 printf("Failed to get object list from OSD %lu (retval=%ld), disconnecting peer\n", role_osd, op->reply.hdr.retval);
+                int fail_fd = op->peer_fd;
                 ps->list_ops.erase(role_osd);
-                c_cli.stop_client(op->peer_fd);
                 delete op;
+                c_cli.stop_client(fail_fd);
                 return;
             }
             printf(
@@ -484,15 +496,13 @@ bool osd_t::stop_pg(pg_t & pg)
     {
         return false;
     }
-    if (!(pg.state & PG_ACTIVE))
+    if (!(pg.state & (PG_ACTIVE | PG_REPEERING)))
     {
         finish_stop_pg(pg);
         return true;
     }
-    pg.state = pg.state & ~PG_ACTIVE | PG_STOPPING;
-    if (pg.inflight == 0 && !pg.flush_batch &&
-        // We must either forget all PG's unstable writes or wait for it to become clean
-        dirty_pgs.find({ .pool_id = pg.pool_id, .pg_num = pg.pg_num }) == dirty_pgs.end())
+    pg.state = pg.state & ~PG_ACTIVE & ~PG_REPEERING | PG_STOPPING;
+    if (pg.inflight == 0 && !pg.flush_batch)
     {
         finish_stop_pg(pg);
     }

src/osd_peering_pg.cpp

@@ -430,12 +430,13 @@ void pg_t::calc_object_states(int log_level)
 void pg_t::print_state()
 {
     printf(
-        "[PG %u/%u] is %s%s%s%s%s%s%s%s%s%s%s%s%s (%lu objects)\n", pool_id, pg_num,
+        "[PG %u/%u] is %s%s%s%s%s%s%s%s%s%s%s%s%s%s (%lu objects)\n", pool_id, pg_num,
         (state & PG_STARTING) ? "starting" : "",
         (state & PG_OFFLINE) ? "offline" : "",
         (state & PG_PEERING) ? "peering" : "",
         (state & PG_INCOMPLETE) ? "incomplete" : "",
         (state & PG_ACTIVE) ? "active" : "",
+        (state & PG_REPEERING) ? "repeering" : "",
         (state & PG_STOPPING) ? "stopping" : "",
         (state & PG_DEGRADED) ? " + degraded" : "",
         (state & PG_HAS_INCOMPLETE) ? " + has_incomplete" : "",

src/osd_primary.cpp

@@ -19,7 +19,7 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
     // Our EC scheme stores data in fixed chunks equal to (K*block size)
     // K = (pg_size-parity_chunks) in case of EC/XOR, or 1 for replicated pools
     pool_id_t pool_id = INODE_POOL(cur_op->req.rw.inode);
-    // FIXME: We have to access pool config here, so make sure that it doesn't change while its PGs are active...
+    // Note: We read pool config here, so we must NOT change it when PGs are active
     auto pool_cfg_it = st_cli.pool_config.find(pool_id);
     if (pool_cfg_it == st_cli.pool_config.end())
     {
@@ -28,6 +28,7 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
         return false;
     }
     auto & pool_cfg = pool_cfg_it->second;
+    // FIXME: op_data->pg_data_size can probably be removed (there's pg.pg_data_size)
     uint64_t pg_data_size = (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pool_cfg.pg_size-pool_cfg.parity_chunks);
     uint64_t pg_block_size = bs_block_size * pg_data_size;
     object_id oid = {
@@ -35,7 +36,7 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
         // oid.stripe = starting offset of the parity stripe
         .stripe = (cur_op->req.rw.offset/pg_block_size)*pg_block_size,
     };
-    pg_num_t pg_num = (cur_op->req.rw.inode + oid.stripe/pool_cfg.pg_stripe_size) % pg_counts[pool_id] + 1;
+    pg_num_t pg_num = (oid.stripe/pool_cfg.pg_stripe_size) % pg_counts[pool_id] + 1; // like map_to_pg()
     auto pg_it = pgs.find({ .pool_id = pool_id, .pg_num = pg_num });
     if (pg_it == pgs.end() || !(pg_it->second.state & PG_ACTIVE))
     {
@@ -52,26 +53,87 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
         return false;
     }
     int stripe_count = (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pg_it->second.pg_size);
+    int chain_size = 0;
+    if (cur_op->req.hdr.opcode == OSD_OP_READ && cur_op->req.rw.meta_revision > 0)
+    {
+        // Chained read
+        auto inode_it = st_cli.inode_config.find(cur_op->req.rw.inode);
+        if (inode_it->second.mod_revision != cur_op->req.rw.meta_revision)
+        {
+            // Client view of the metadata differs from OSD's view
+            // Operation can't be completed correctly, client should retry later
+            finish_op(cur_op, -EPIPE);
+            return false;
+        }
+        // Find parents from the same pool. Optimized reads only work within pools
+        while (inode_it != st_cli.inode_config.end() && inode_it->second.parent_id &&
+            INODE_POOL(inode_it->second.parent_id) == pg_it->second.pool_id)
+        {
+            chain_size++;
+            inode_it = st_cli.inode_config.find(inode_it->second.parent_id);
+        }
+        if (chain_size)
+        {
+            // Add the original inode
+            chain_size++;
+        }
+    }
     osd_primary_op_data_t *op_data = (osd_primary_op_data_t*)calloc_or_die(
-        1, sizeof(osd_primary_op_data_t) + (clean_entry_bitmap_size + sizeof(osd_rmw_stripe_t)) * stripe_count
+        // Allocate:
+        // - op_data
+        1, sizeof(osd_primary_op_data_t) +
+        // - stripes
+        // - resulting bitmap buffers
+        stripe_count * (clean_entry_bitmap_size + sizeof(osd_rmw_stripe_t)) +
+        chain_size * (
+            // - copy of the chain
+            sizeof(inode_t) +
+            // - bitmap buffers for chained read
+            stripe_count * clean_entry_bitmap_size +
+            // - 'missing' flags for chained reads
+            (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 0 : pg_it->second.pg_size)
+        )
     );
+    void *data_buf = ((void*)op_data) + sizeof(osd_primary_op_data_t);
     op_data->pg_num = pg_num;
     op_data->oid = oid;
-    op_data->stripes = ((osd_rmw_stripe_t*)(op_data+1));
+    op_data->stripes = (osd_rmw_stripe_t*)data_buf;
+    data_buf += sizeof(osd_rmw_stripe_t) * stripe_count;
     op_data->scheme = pool_cfg.scheme;
     op_data->pg_data_size = pg_data_size;
+    op_data->pg_size = pg_it->second.pg_size;
     cur_op->op_data = op_data;
     split_stripes(pg_data_size, bs_block_size, (uint32_t)(cur_op->req.rw.offset - oid.stripe), cur_op->req.rw.len, op_data->stripes);
     // Allocate bitmaps along with stripes to avoid extra allocations and fragmentation
     for (int i = 0; i < stripe_count; i++)
     {
-        op_data->stripes[i].bmp_buf = (void*)(op_data->stripes+stripe_count) + clean_entry_bitmap_size*i;
+        op_data->stripes[i].bmp_buf = data_buf;
+        data_buf += clean_entry_bitmap_size;
+    }
+    op_data->chain_size = chain_size;
+    if (chain_size > 0)
+    {
+        op_data->read_chain = (inode_t*)data_buf;
+        data_buf += sizeof(inode_t) * chain_size;
+        op_data->snapshot_bitmaps = data_buf;
+        data_buf += chain_size * stripe_count * clean_entry_bitmap_size;
+        op_data->missing_flags = (uint8_t*)data_buf;
+        data_buf += chain_size * (pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 0 : pg_it->second.pg_size);
+        // Copy chain
+        int chain_num = 0;
+        op_data->read_chain[chain_num++] = cur_op->req.rw.inode;
+        auto inode_it = st_cli.inode_config.find(cur_op->req.rw.inode);
+        while (inode_it != st_cli.inode_config.end() && inode_it->second.parent_id)
+        {
+            op_data->read_chain[chain_num++] = inode_it->second.parent_id;
+            inode_it = st_cli.inode_config.find(inode_it->second.parent_id);
+        }
     }
     pg_it->second.inflight++;
     return true;
 }
 
-static uint64_t* get_object_osd_set(pg_t &pg, object_id &oid, uint64_t *def, pg_osd_set_state_t **object_state)
+uint64_t* osd_t::get_object_osd_set(pg_t &pg, object_id &oid, uint64_t *def, pg_osd_set_state_t **object_state)
 {
     if (!(pg.state & (PG_HAS_INCOMPLETE | PG_HAS_DEGRADED | PG_HAS_MISPLACED)))
     {
@@ -106,10 +168,17 @@ void osd_t::continue_primary_read(osd_op_t *cur_op)
     {
         return;
     }
-    cur_op->reply.rw.bitmap_len = 0;
     osd_primary_op_data_t *op_data = cur_op->op_data;
-    if (op_data->st == 1)      goto resume_1;
-    else if (op_data->st == 2) goto resume_2;
+    if (op_data->chain_size)
+    {
+        continue_chained_read(cur_op);
+        return;
+    }
+    if (op_data->st == 1)
+        goto resume_1;
+    else if (op_data->st == 2)
+        goto resume_2;
+    cur_op->reply.rw.bitmap_len = 0;
     {
         auto & pg = pgs.at({ .pool_id = INODE_POOL(op_data->oid.inode), .pg_num = op_data->pg_num });
         for (int role = 0; role < op_data->pg_data_size; role++)
@@ -124,8 +193,7 @@ void osd_t::continue_primary_read(osd_op_t *cur_op)
         {
             // Fast happy-path
             cur_op->buf = alloc_read_buffer(op_data->stripes, op_data->pg_data_size, 0);
-            submit_primary_subops(SUBMIT_READ, op_data->target_ver,
-                (op_data->scheme == POOL_SCHEME_REPLICATED ? pg.pg_size : op_data->pg_data_size), pg.cur_set.data(), cur_op);
+            submit_primary_subops(SUBMIT_READ, op_data->target_ver, pg.cur_set.data(), cur_op);
             op_data->st = 1;
         }
         else
@@ -142,7 +210,7 @@ void osd_t::continue_primary_read(osd_op_t *cur_op)
             op_data->scheme = pg.scheme;
             op_data->degraded = 1;
             cur_op->buf = alloc_read_buffer(op_data->stripes, pg.pg_size, 0);
-            submit_primary_subops(SUBMIT_READ, op_data->target_ver, pg.pg_size, cur_set, cur_op);
+            submit_primary_subops(SUBMIT_READ, op_data->target_ver, cur_set, cur_op);
             op_data->st = 1;
         }
     }
@@ -188,612 +256,6 @@ resume_2:
     finish_op(cur_op, cur_op->req.rw.len);
 }
 
-bool osd_t::check_write_queue(osd_op_t *cur_op, pg_t & pg)
-{
-    osd_primary_op_data_t *op_data = cur_op->op_data;
-    // Check if actions are pending for this object
-    auto act_it = pg.flush_actions.lower_bound((obj_piece_id_t){
-        .oid = op_data->oid,
-        .osd_num = 0,
-    });
-    if (act_it != pg.flush_actions.end() &&
-        act_it->first.oid.inode == op_data->oid.inode &&
-        (act_it->first.oid.stripe & ~STRIPE_MASK) == op_data->oid.stripe)
-    {
-        pg.write_queue.emplace(op_data->oid, cur_op);
-        return false;
-    }
-    // Check if there are other write requests to the same object
-    auto vo_it = pg.write_queue.find(op_data->oid);
-    if (vo_it != pg.write_queue.end())
-    {
-        op_data->st = 1;
-        pg.write_queue.emplace(op_data->oid, cur_op);
-        return false;
-    }
-    pg.write_queue.emplace(op_data->oid, cur_op);
-    return true;
-}
-
-void osd_t::continue_primary_write(osd_op_t *cur_op)
-{
-    if (!cur_op->op_data && !prepare_primary_rw(cur_op))
-    {
-        return;
-    }
-    osd_primary_op_data_t *op_data = cur_op->op_data;
-    auto & pg = pgs.at({ .pool_id = INODE_POOL(op_data->oid.inode), .pg_num = op_data->pg_num });
-    if (op_data->st == 1)      goto resume_1;
-    else if (op_data->st == 2) goto resume_2;
-    else if (op_data->st == 3) goto resume_3;
-    else if (op_data->st == 4) goto resume_4;
-    else if (op_data->st == 5) goto resume_5;
-    else if (op_data->st == 6) goto resume_6;
-    else if (op_data->st == 7) goto resume_7;
-    else if (op_data->st == 8) goto resume_8;
-    else if (op_data->st == 9) goto resume_9;
-    else if (op_data->st == 10) goto resume_10;
-    assert(op_data->st == 0);
-    if (!check_write_queue(cur_op, pg))
-    {
-        return;
-    }
-resume_1:
-    // Determine blocks to read and write
-    // Missing chunks are allowed to be overwritten even in incomplete objects
-    // FIXME: Allow to do small writes to the old (degraded/misplaced) OSD set for lower performance impact
-    op_data->prev_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
-    if (op_data->scheme == POOL_SCHEME_REPLICATED)
-    {
-        // Simplified algorithm
-        op_data->stripes[0].write_start = op_data->stripes[0].req_start;
-        op_data->stripes[0].write_end = op_data->stripes[0].req_end;
-        op_data->stripes[0].write_buf = cur_op->buf;
-        op_data->stripes[0].bmp_buf = (void*)(op_data->stripes+1);
-        if (pg.cur_set.data() != op_data->prev_set && (op_data->stripes[0].write_start != 0 ||
-            op_data->stripes[0].write_end != bs_block_size))
-        {
-            // Object is degraded/misplaced and will be moved to <write_osd_set>
-            op_data->stripes[0].read_start = 0;
-            op_data->stripes[0].read_end = bs_block_size;
-            cur_op->rmw_buf = op_data->stripes[0].read_buf = memalign_or_die(MEM_ALIGNMENT, bs_block_size);
-        }
-    }
-    else
-    {
-        cur_op->rmw_buf = calc_rmw(cur_op->buf, op_data->stripes, op_data->prev_set,
-            pg.pg_size, op_data->pg_data_size, pg.pg_cursize, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
-        if (!cur_op->rmw_buf)
-        {
-            // Refuse partial overwrite of an incomplete object
-            cur_op->reply.hdr.retval = -EINVAL;
-            goto continue_others;
-        }
-    }
-    // Read required blocks
-    submit_primary_subops(SUBMIT_RMW_READ, UINT64_MAX, pg.pg_size, op_data->prev_set, cur_op);
-resume_2:
-    op_data->st = 2;
-    return;
-resume_3:
-    if (op_data->errors > 0)
-    {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
-        return;
-    }
-    // Save version override for parallel reads
-    pg.ver_override[op_data->oid] = op_data->fact_ver;
-    if (op_data->scheme == POOL_SCHEME_REPLICATED)
-    {
-        // Set bitmap bits
-        bitmap_set(op_data->stripes[0].bmp_buf, op_data->stripes[0].write_start, op_data->stripes[0].write_end, bs_bitmap_granularity);
-        // Possibly copy new data from the request into the recovery buffer
-        if (pg.cur_set.data() != op_data->prev_set && (op_data->stripes[0].write_start != 0 ||
-            op_data->stripes[0].write_end != bs_block_size))
-        {
-            memcpy(
-                op_data->stripes[0].read_buf + op_data->stripes[0].req_start,
-                op_data->stripes[0].write_buf,
-                op_data->stripes[0].req_end - op_data->stripes[0].req_start
-            );
-            op_data->stripes[0].write_buf = op_data->stripes[0].read_buf;
-            op_data->stripes[0].write_start = 0;
-            op_data->stripes[0].write_end = bs_block_size;
-        }
-    }
-    else
-    {
-        // Recover missing stripes, calculate parity
-        if (pg.scheme == POOL_SCHEME_XOR)
-        {
-            calc_rmw_parity_xor(op_data->stripes, pg.pg_size, op_data->prev_set, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
-        }
-        else if (pg.scheme == POOL_SCHEME_JERASURE)
-        {
-            calc_rmw_parity_jerasure(op_data->stripes, pg.pg_size, op_data->pg_data_size, op_data->prev_set, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
-        }
-    }
-    // Send writes
-    if ((op_data->fact_ver >> (64-PG_EPOCH_BITS)) < pg.epoch)
-    {
-        op_data->target_ver = ((uint64_t)pg.epoch << (64-PG_EPOCH_BITS)) | 1;
-    }
-    else
-    {
-        if ((op_data->fact_ver & (1ul<<(64-PG_EPOCH_BITS) - 1)) == (1ul<<(64-PG_EPOCH_BITS) - 1))
-        {
-            assert(pg.epoch != ((1ul << PG_EPOCH_BITS)-1));
-            pg.epoch++;
-        }
-        op_data->target_ver = op_data->fact_ver + 1;
-    }
-    if (pg.epoch > pg.reported_epoch)
-    {
-        // Report newer epoch before writing
-        // FIXME: We may report only one PG state here...
-        this->pg_state_dirty.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
-        pg.history_changed = true;
-        report_pg_states();
-resume_10:
-        if (pg.epoch > pg.reported_epoch)
-        {
-            op_data->st = 10;
-            return;
-        }
-    }
-    submit_primary_subops(SUBMIT_WRITE, op_data->target_ver, pg.pg_size, pg.cur_set.data(), cur_op);
-resume_4:
-    op_data->st = 4;
-    return;
-resume_5:
-    if (op_data->errors > 0)
-    {
-        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
-        return;
-    }
-resume_6:
-resume_7:
-    if (!remember_unstable_write(cur_op, pg, pg.cur_loc_set, 6))
-    {
-        // FIXME: Check for immediate_commit == IMMEDIATE_SMALL
-        return;
-    }
-    if (op_data->fact_ver == 1)
-    {
-        // Object is created
-        pg.clean_count++;
-        pg.total_count++;
-    }
-    if (op_data->object_state)
-    {
-        {
-            int recovery_type = op_data->object_state->state & (OBJ_DEGRADED|OBJ_INCOMPLETE) ? 0 : 1;
-            recovery_stat_count[0][recovery_type]++;
-            if (!recovery_stat_count[0][recovery_type])
-            {
-                recovery_stat_count[0][recovery_type]++;
-                recovery_stat_bytes[0][recovery_type] = 0;
-            }
-            for (int role = 0; role < (op_data->scheme == POOL_SCHEME_REPLICATED ? 1 : pg.pg_size); role++)
-            {
-                recovery_stat_bytes[0][recovery_type] += op_data->stripes[role].write_end - op_data->stripes[role].write_start;
-            }
-        }
-        // Any kind of a non-clean object can have extra chunks, because we don't record objects
-        // as degraded & misplaced or incomplete & misplaced at the same time. So try to remove extra chunks
-        if (immediate_commit != IMMEDIATE_ALL)
-        {
-            // We can't remove extra chunks yet if fsyncs are explicit, because
-            // new copies may not be committed to stable storage yet
-            // We can only remove extra chunks after a successful SYNC for this PG
-            for (auto & chunk: op_data->object_state->osd_set)
-            {
-                // Check is the same as in submit_primary_del_subops()
-                if (op_data->scheme == POOL_SCHEME_REPLICATED
-                    ? !contains_osd(pg.cur_set.data(), pg.pg_size, chunk.osd_num)
-                    : (chunk.osd_num != pg.cur_set[chunk.role]))
-                {
-                    pg.copies_to_delete_after_sync.push_back((obj_ver_osd_t){
-                        .osd_num = chunk.osd_num,
-                        .oid = {
-                            .inode = op_data->oid.inode,
-                            .stripe = op_data->oid.stripe | (op_data->scheme == POOL_SCHEME_REPLICATED ? 0 : chunk.role),
-                        },
-                        .version = op_data->fact_ver,
-                    });
-                    copies_to_delete_after_sync_count++;
-                }
-            }
-        }
-        else
-        {
-            submit_primary_del_subops(cur_op, pg.cur_set.data(), pg.pg_size, op_data->object_state->osd_set);
-            if (op_data->n_subops > 0)
-            {
-resume_8:
-                op_data->st = 8;
-                return;
-resume_9:
-                if (op_data->errors > 0)
-                {
-                    pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
-                    return;
-                }
-            }
-        }
-        // Clear object state
-        remove_object_from_state(op_data->oid, op_data->object_state, pg);
-        pg.clean_count++;
-    }
-    cur_op->reply.hdr.retval = cur_op->req.rw.len;
-continue_others:
-    // Remove version override
-    pg.ver_override.erase(op_data->oid);
-    object_id oid = op_data->oid;
-    // Remove the operation from queue before calling finish_op so it doesn't see the completed operation in queue
-    auto next_it = pg.write_queue.find(oid);
-    if (next_it != pg.write_queue.end() && next_it->second == cur_op)
-    {
-        pg.write_queue.erase(next_it++);
-    }
-    // finish_op would invalidate next_it if it cleared pg.write_queue, but it doesn't do that :)
-    finish_op(cur_op, cur_op->reply.hdr.retval);
-    // Continue other write operations to the same object
-    if (next_it != pg.write_queue.end() && next_it->first == oid)
-    {
-        osd_op_t *next_op = next_it->second;
-        continue_primary_write(next_op);
-    }
-}
-
-bool osd_t::remember_unstable_write(osd_op_t *cur_op, pg_t & pg, pg_osd_set_t & loc_set, int base_state)
-{
-    osd_primary_op_data_t *op_data = cur_op->op_data;
-    if (op_data->st == base_state)
-    {
-        goto resume_6;
-    }
-    else if (op_data->st == base_state+1)
-    {
-        goto resume_7;
-    }
-    // FIXME: Check for immediate_commit == IMMEDIATE_SMALL
-    if (immediate_commit == IMMEDIATE_ALL)
-    {
-        if (op_data->scheme != POOL_SCHEME_REPLICATED)
-        {
-            // Send STABILIZE ops immediately
-            op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
-            op_data->unstable_writes = new obj_ver_id[loc_set.size()];
-            {
-                int last_start = 0;
-                for (auto & chunk: loc_set)
-                {
-                    op_data->unstable_writes[last_start] = (obj_ver_id){
-                        .oid = {
-                            .inode = op_data->oid.inode,
-                            .stripe = op_data->oid.stripe | chunk.role,
-                        },
-                        .version = op_data->fact_ver,
-                    };
-                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
-                        .osd_num = chunk.osd_num,
-                        .start = last_start,
-                        .len = 1,
-                    });
-                    last_start++;
-                }
-            }
-            submit_primary_stab_subops(cur_op);
-resume_6:
-            op_data->st = 6;
-            return false;
-resume_7:
-            // FIXME: Free those in the destructor?
-            delete op_data->unstable_write_osds;
-            delete[] op_data->unstable_writes;
-            op_data->unstable_writes = NULL;
-            op_data->unstable_write_osds = NULL;
-            if (op_data->errors > 0)
-            {
-                pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
-                return false;
-            }
-        }
-    }
-    else
-    {
-        if (op_data->scheme != POOL_SCHEME_REPLICATED)
-        {
-            // Remember version as unstable for EC/XOR
-            for (auto & chunk: loc_set)
-            {
-                this->dirty_osds.insert(chunk.osd_num);
-                this->unstable_writes[(osd_object_id_t){
-                    .osd_num = chunk.osd_num,
-                    .oid = {
-                        .inode = op_data->oid.inode,
-                        .stripe = op_data->oid.stripe | chunk.role,
-                    },
-                }] = op_data->fact_ver;
-            }
-        }
-        else
-        {
-            // Only remember to sync OSDs for replicated pools
-            for (auto & chunk: loc_set)
-            {
-                this->dirty_osds.insert(chunk.osd_num);
-            }
-        }
-        // Remember PG as dirty to drop the connection when PG goes offline
-        // (this is required because of the "lazy sync")
-        auto cl_it = c_cli.clients.find(cur_op->peer_fd);
-        if (cl_it != c_cli.clients.end())
-        {
-            cl_it->second->dirty_pgs.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
-        }
-        dirty_pgs.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
-    }
-    return true;
-}
-
-// Save and clear unstable_writes -> SYNC all -> STABLE all
-void osd_t::continue_primary_sync(osd_op_t *cur_op)
-{
-    if (!cur_op->op_data)
-    {
-        cur_op->op_data = (osd_primary_op_data_t*)calloc_or_die(1, sizeof(osd_primary_op_data_t));
-    }
-    osd_primary_op_data_t *op_data = cur_op->op_data;
-    if (op_data->st == 1)      goto resume_1;
-    else if (op_data->st == 2) goto resume_2;
-    else if (op_data->st == 3) goto resume_3;
-    else if (op_data->st == 4) goto resume_4;
-    else if (op_data->st == 5) goto resume_5;
-    else if (op_data->st == 6) goto resume_6;
-    else if (op_data->st == 7) goto resume_7;
-    else if (op_data->st == 8) goto resume_8;
-    assert(op_data->st == 0);
-    if (syncs_in_progress.size() > 0)
-    {
-        // Wait for previous syncs, if any
-        // FIXME: We may try to execute the current one in parallel, like in Blockstore, but I'm not sure if it matters at all
-        syncs_in_progress.push_back(cur_op);
-        op_data->st = 1;
-resume_1:
-        return;
-    }
-    else
-    {
-        syncs_in_progress.push_back(cur_op);
-    }
-resume_2:
-    if (dirty_osds.size() == 0)
-    {
-        // Nothing to sync
-        goto finish;
-    }
-    // Save and clear unstable_writes
-    // In theory it is possible to do in on a per-client basis, but this seems to be an unnecessary complication
-    // It would be cool not to copy these here at all, but someone has to deduplicate them by object IDs anyway
-    if (unstable_writes.size() > 0)
-    {
-        op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
-        op_data->unstable_writes = new obj_ver_id[this->unstable_writes.size()];
-        osd_num_t last_osd = 0;
-        int last_start = 0, last_end = 0;
-        for (auto it = this->unstable_writes.begin(); it != this->unstable_writes.end(); it++)
-        {
-            if (last_osd != it->first.osd_num)
-            {
-                if (last_osd != 0)
-                {
-                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
-                        .osd_num = last_osd,
-                        .start = last_start,
-                        .len = last_end - last_start,
-                    });
-                }
-                last_osd = it->first.osd_num;
-                last_start = last_end;
-            }
-            op_data->unstable_writes[last_end] = (obj_ver_id){
-                .oid = it->first.oid,
-                .version = it->second,
-            };
-            last_end++;
-        }
-        if (last_osd != 0)
-        {
-            op_data->unstable_write_osds->push_back((unstable_osd_num_t){
-                .osd_num = last_osd,
-                .start = last_start,
-                .len = last_end - last_start,
-            });
-        }
-        this->unstable_writes.clear();
-    }
-    {
-        void *dirty_buf = malloc_or_die(
-            sizeof(pool_pg_num_t)*dirty_pgs.size() +
-            sizeof(osd_num_t)*dirty_osds.size() +
-            sizeof(obj_ver_osd_t)*this->copies_to_delete_after_sync_count
-        );
-        op_data->dirty_pgs = (pool_pg_num_t*)dirty_buf;
-        op_data->dirty_osds = (osd_num_t*)(dirty_buf + sizeof(pool_pg_num_t)*dirty_pgs.size());
-        op_data->dirty_pg_count = dirty_pgs.size();
-        op_data->dirty_osd_count = dirty_osds.size();
-        if (this->copies_to_delete_after_sync_count)
-        {
-            op_data->copies_to_delete_count = 0;
-            op_data->copies_to_delete = (obj_ver_osd_t*)(op_data->dirty_osds + op_data->dirty_osd_count);
-            for (auto dirty_pg_num: dirty_pgs)
-            {
-                auto & pg = pgs.at(dirty_pg_num);
-                assert(pg.copies_to_delete_after_sync.size() <= this->copies_to_delete_after_sync_count);
-                memcpy(
-                    op_data->copies_to_delete + op_data->copies_to_delete_count,
-                    pg.copies_to_delete_after_sync.data(),
-                    sizeof(obj_ver_osd_t)*pg.copies_to_delete_after_sync.size()
-                );
-                op_data->copies_to_delete_count += pg.copies_to_delete_after_sync.size();
-                this->copies_to_delete_after_sync_count -= pg.copies_to_delete_after_sync.size();
-                pg.copies_to_delete_after_sync.clear();
-            }
-            assert(this->copies_to_delete_after_sync_count == 0);
-        }
-        int dpg = 0;
-        for (auto dirty_pg_num: dirty_pgs)
-        {
-            pgs.at(dirty_pg_num).inflight++;
-            op_data->dirty_pgs[dpg++] = dirty_pg_num;
-        }
-        dirty_pgs.clear();
-        dpg = 0;
-        for (auto osd_num: dirty_osds)
-        {
-            op_data->dirty_osds[dpg++] = osd_num;
-        }
-        dirty_osds.clear();
-    }
-    if (immediate_commit != IMMEDIATE_ALL)
-    {
-        // SYNC
-        submit_primary_sync_subops(cur_op);
-resume_3:
-        op_data->st = 3;
-        return;
-resume_4:
-        if (op_data->errors > 0)
-        {
-            goto resume_6;
-        }
-    }
-    if (op_data->unstable_writes)
-    {
-        // Stabilize version sets, if any
-        submit_primary_stab_subops(cur_op);
-resume_5:
-        op_data->st = 5;
-        return;
-    }
-resume_6:
-    if (op_data->errors > 0)
-    {
-        // Return PGs and OSDs back into their dirty sets
-        for (int i = 0; i < op_data->dirty_pg_count; i++)
-        {
-            dirty_pgs.insert(op_data->dirty_pgs[i]);
-        }
-        for (int i = 0; i < op_data->dirty_osd_count; i++)
-        {
-            dirty_osds.insert(op_data->dirty_osds[i]);
-        }
-        if (op_data->unstable_writes)
-        {
-            // Return objects back into the unstable write set
-            for (auto unstable_osd: *(op_data->unstable_write_osds))
-            {
-                for (int i = 0; i < unstable_osd.len; i++)
-                {
-                    // Except those from peered PGs
-                    auto & w = op_data->unstable_writes[i];
-                    pool_pg_num_t wpg = {
-                        .pool_id = INODE_POOL(w.oid.inode),
-                        .pg_num = map_to_pg(w.oid, st_cli.pool_config.at(INODE_POOL(w.oid.inode)).pg_stripe_size),
-                    };
-                    if (pgs.at(wpg).state & PG_ACTIVE)
-                    {
-                        uint64_t & dest = this->unstable_writes[(osd_object_id_t){
-                            .osd_num = unstable_osd.osd_num,
-                            .oid = w.oid,
-                        }];
-                        dest = dest < w.version ? w.version : dest;
-                        dirty_pgs.insert(wpg);
-                    }
-                }
-            }
-        }
-        if (op_data->copies_to_delete)
-        {
-            // Return 'copies to delete' back into respective PGs
-            for (int i = 0; i < op_data->copies_to_delete_count; i++)
-            {
-                auto & w = op_data->copies_to_delete[i];
-                auto & pg = pgs.at((pool_pg_num_t){
-                    .pool_id = INODE_POOL(w.oid.inode),
-                    .pg_num = map_to_pg(w.oid, st_cli.pool_config.at(INODE_POOL(w.oid.inode)).pg_stripe_size),
-                });
-                if (pg.state & PG_ACTIVE)
-                {
-                    pg.copies_to_delete_after_sync.push_back(w);
-                    copies_to_delete_after_sync_count++;
-                }
-            }
-        }
-    }
-    else if (op_data->copies_to_delete)
-    {
-        // Actually delete copies which we wanted to delete
-        submit_primary_del_batch(cur_op, op_data->copies_to_delete, op_data->copies_to_delete_count);
-resume_7:
-        op_data->st = 7;
-        return;
-resume_8:
-        if (op_data->errors > 0)
-        {
-            goto resume_6;
-        }
-    }
-    for (int i = 0; i < op_data->dirty_pg_count; i++)
-    {
-        auto & pg = pgs.at(op_data->dirty_pgs[i]);
-        pg.inflight--;
-        if ((pg.state & PG_STOPPING) && pg.inflight == 0 && !pg.flush_batch &&
-            // We must either forget all PG's unstable writes or wait for it to become clean
-            dirty_pgs.find({ .pool_id = pg.pool_id, .pg_num = pg.pg_num }) == dirty_pgs.end())
-        {
-            finish_stop_pg(pg);
-        }
-    }
-    // FIXME: Free those in the destructor?
-    free(op_data->dirty_pgs);
-    op_data->dirty_pgs = NULL;
-    op_data->dirty_osds = NULL;
-    if (op_data->unstable_writes)
-    {
-        delete op_data->unstable_write_osds;
-        delete[] op_data->unstable_writes;
-        op_data->unstable_writes = NULL;
-        op_data->unstable_write_osds = NULL;
-    }
-    if (op_data->errors > 0)
-    {
-        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
-    }
-    else
-    {
-finish:
-        if (cur_op->peer_fd)
-        {
-            auto it = c_cli.clients.find(cur_op->peer_fd);
-            if (it != c_cli.clients.end())
-                it->second->dirty_pgs.clear();
-        }
-        finish_op(cur_op, 0);
-    }
-    assert(syncs_in_progress.front() == cur_op);
-    syncs_in_progress.pop_front();
-    if (syncs_in_progress.size() > 0)
-    {
-        cur_op = syncs_in_progress.front();
-        op_data = cur_op->op_data;
-        op_data->st++;
-        goto resume_2;
-    }
-}
-
 // Decrement pg_osd_set_state_t's object_count and change PG state accordingly
 void osd_t::remove_object_from_state(object_id & oid, pg_osd_set_state_t *object_state, pg_t & pg)
 {
@@ -832,10 +294,14 @@ void osd_t::remove_object_from_state(object_id & oid, pg_osd_set_state_t *object
     {
         throw std::runtime_error("BUG: Invalid object state: "+std::to_string(object_state->state));
     }
-    object_state->object_count--;
-    if (!object_state->object_count)
+}
+
+void osd_t::free_object_state(pg_t & pg, pg_osd_set_state_t **object_state)
+{
+    if (*object_state && !(--(*object_state)->object_count))
     {
-        pg.state_dict.erase(object_state->osd_set);
+        pg.state_dict.erase((*object_state)->osd_set);
+        *object_state = NULL;
     }
 }
 
@@ -867,7 +333,7 @@ resume_1:
     // Determine which OSDs contain this object and delete it
     op_data->prev_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
     // Submit 1 read to determine the actual version number
-    submit_primary_subops(SUBMIT_RMW_READ, UINT64_MAX, pg.pg_size, op_data->prev_set, cur_op);
+    submit_primary_subops(SUBMIT_RMW_READ, UINT64_MAX, op_data->prev_set, cur_op);
 resume_2:
     op_data->st = 2;
     return;
@@ -901,22 +367,21 @@ resume_5:
     else
     {
         remove_object_from_state(op_data->oid, op_data->object_state, pg);
+        free_object_state(pg, &op_data->object_state);
     }
     pg.total_count--;
-    object_id oid = op_data->oid;
-    finish_op(cur_op, cur_op->req.rw.len);
-    // Continue other write operations to the same object
-    auto next_it = pg.write_queue.find(oid);
-    auto this_it = next_it;
-    if (this_it != pg.write_queue.end() && this_it->second == cur_op)
+    osd_op_t *next_op = NULL;
+    auto next_it = pg.write_queue.find(op_data->oid);
+    if (next_it != pg.write_queue.end() && next_it->second == cur_op)
     {
-        next_it++;
-        pg.write_queue.erase(this_it);
-        if (next_it != pg.write_queue.end() &&
-            next_it->first == oid)
-        {
-            osd_op_t *next_op = next_it->second;
-            continue_primary_write(next_op);
-        }
+        pg.write_queue.erase(next_it++);
+        if (next_it != pg.write_queue.end() && next_it->first == op_data->oid)
+            next_op = next_it->second;
+    }
+    finish_op(cur_op, cur_op->req.rw.len);
+    if (next_op)
+    {
+        // Continue next write to the same object
+        continue_primary_write(next_op);
     }
 }

src/osd_primary.h

@@ -31,15 +31,31 @@ struct osd_primary_op_data_t
     uint64_t *prev_set = NULL;
     pg_osd_set_state_t *object_state = NULL;
 
-    // for sync. oops, requires freeing
-    std::vector<unstable_osd_num_t> *unstable_write_osds = NULL;
-    pool_pg_num_t *dirty_pgs = NULL;
-    int dirty_pg_count = 0;
-    osd_num_t *dirty_osds = NULL;
-    int dirty_osd_count = 0;
-    obj_ver_id *unstable_writes = NULL;
-    obj_ver_osd_t *copies_to_delete = NULL;
-    int copies_to_delete_count = 0;
+    union
+    {
+        struct
+        {
+            // for sync. oops, requires freeing
+            std::vector<unstable_osd_num_t> *unstable_write_osds;
+            pool_pg_num_t *dirty_pgs;
+            int dirty_pg_count;
+            osd_num_t *dirty_osds;
+            int dirty_osd_count;
+            obj_ver_id *unstable_writes;
+            obj_ver_osd_t *copies_to_delete;
+            int copies_to_delete_count;
+        };
+        struct
+        {
+            // for read_bitmaps
+            void *snapshot_bitmaps;
+            inode_t *read_chain;
+            uint8_t *missing_flags;
+            int chain_size;
+            osd_chain_read_t *chain_reads;
+            int chain_read_count;
+        };
+    };
 };
 
 bool contains_osd(osd_num_t *osd_set, uint64_t size, osd_num_t osd_num);

src/osd_primary_chain.cpp

@@ -0,0 +1,549 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "osd_primary.h"
+#include "allocator.h"
+
+void osd_t::continue_chained_read(osd_op_t *cur_op)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    auto & pg = pgs.at({ .pool_id = INODE_POOL(op_data->oid.inode), .pg_num = op_data->pg_num });
+    if (op_data->st == 1)
+        goto resume_1;
+    else if (op_data->st == 2)
+        goto resume_2;
+    else if (op_data->st == 3)
+        goto resume_3;
+    else if (op_data->st == 4)
+        goto resume_4;
+    cur_op->reply.rw.bitmap_len = 0;
+    for (int role = 0; role < op_data->pg_data_size; role++)
+    {
+        op_data->stripes[role].read_start = op_data->stripes[role].req_start;
+        op_data->stripes[role].read_end = op_data->stripes[role].req_end;
+    }
+resume_1:
+resume_2:
+    // Read bitmaps
+    if (read_bitmaps(cur_op, pg, 1) != 0)
+        return;
+    // Prepare & submit reads
+    if (submit_chained_read_requests(pg, cur_op) != 0)
+        return;
+    if (op_data->n_subops > 0)
+    {
+        // Wait for reads
+        op_data->st = 3;
+resume_3:
+        return;
+    }
+resume_4:
+    if (op_data->errors > 0)
+    {
+        free(op_data->chain_reads);
+        op_data->chain_reads = NULL;
+        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
+        return;
+    }
+    send_chained_read_results(pg, cur_op);
+    finish_op(cur_op, cur_op->req.rw.len);
+}
+
+int osd_t::read_bitmaps(osd_op_t *cur_op, pg_t & pg, int base_state)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    if (op_data->st == base_state)
+        goto resume_0;
+    else if (op_data->st == base_state+1)
+        goto resume_1;
+    if (pg.state == PG_ACTIVE && pg.scheme == POOL_SCHEME_REPLICATED)
+    {
+        // Happy path for clean replicated PGs (all bitmaps are available locally)
+        for (int chain_num = 0; chain_num < op_data->chain_size; chain_num++)
+        {
+            object_id cur_oid = { .inode = op_data->read_chain[chain_num], .stripe = op_data->oid.stripe };
+            auto vo_it = pg.ver_override.find(cur_oid);
+            auto read_version = (vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX);
+            // Read bitmap synchronously from the local database
+            bs->read_bitmap(cur_oid, read_version, op_data->snapshot_bitmaps + chain_num*clean_entry_bitmap_size, NULL);
+        }
+    }
+    else
+    {
+        if (submit_bitmap_subops(cur_op, pg) < 0)
+        {
+            // Failure
+            finish_op(cur_op, -EIO);
+            return -1;
+        }
+resume_0:
+        if (op_data->n_subops > 0)
+        {
+            // Wait for subops
+            op_data->st = base_state;
+            return 1;
+        }
+resume_1:
+        if (pg.scheme != POOL_SCHEME_REPLICATED)
+        {
+            for (int chain_num = 0; chain_num < op_data->chain_size; chain_num++)
+            {
+                // Check if we need to reconstruct any bitmaps
+                for (int i = 0; i < pg.pg_size; i++)
+                {
+                    if (op_data->missing_flags[chain_num*pg.pg_size + i])
+                    {
+                        osd_rmw_stripe_t local_stripes[pg.pg_size] = { 0 };
+                        for (i = 0; i < pg.pg_size; i++)
+                        {
+                            local_stripes[i].missing = op_data->missing_flags[chain_num*pg.pg_size + i] && true;
+                            local_stripes[i].bmp_buf = op_data->snapshot_bitmaps + (chain_num*pg.pg_size + i)*clean_entry_bitmap_size;
+                            local_stripes[i].read_start = local_stripes[i].read_end = 1;
+                        }
+                        if (pg.scheme == POOL_SCHEME_XOR)
+                        {
+                            reconstruct_stripes_xor(local_stripes, pg.pg_size, clean_entry_bitmap_size);
+                        }
+                        else if (pg.scheme == POOL_SCHEME_JERASURE)
+                        {
+                            reconstruct_stripes_jerasure(local_stripes, pg.pg_size, pg.pg_data_size, clean_entry_bitmap_size);
+                        }
+                        break;
+                    }
+                }
+            }
+        }
+    }
+    return 0;
+}
+
+int osd_t::collect_bitmap_requests(osd_op_t *cur_op, pg_t & pg, std::vector<bitmap_request_t> & bitmap_requests)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    for (int chain_num = 0; chain_num < op_data->chain_size; chain_num++)
+    {
+        object_id cur_oid = { .inode = op_data->read_chain[chain_num], .stripe = op_data->oid.stripe };
+        auto vo_it = pg.ver_override.find(cur_oid);
+        uint64_t target_version = vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX;
+        pg_osd_set_state_t *object_state;
+        uint64_t* cur_set = get_object_osd_set(pg, cur_oid, pg.cur_set.data(), &object_state);
+        if (pg.scheme == POOL_SCHEME_REPLICATED)
+        {
+            osd_num_t read_target = 0;
+            for (int i = 0; i < pg.pg_size; i++)
+            {
+                if (cur_set[i] == this->osd_num || cur_set[i] != 0 && read_target == 0)
+                {
+                    // Select local or any other available OSD for reading
+                    read_target = cur_set[i];
+                }
+            }
+            assert(read_target != 0);
+            bitmap_requests.push_back((bitmap_request_t){
+                .osd_num = read_target,
+                .oid = cur_oid,
+                .version = target_version,
+                .bmp_buf = op_data->snapshot_bitmaps + chain_num*clean_entry_bitmap_size,
+            });
+        }
+        else
+        {
+            osd_rmw_stripe_t local_stripes[pg.pg_size];
+            memcpy(local_stripes, op_data->stripes, sizeof(osd_rmw_stripe_t) * pg.pg_size);
+            if (extend_missing_stripes(local_stripes, cur_set, pg.pg_data_size, pg.pg_size) < 0)
+            {
+                free(op_data->snapshot_bitmaps);
+                return -1;
+            }
+            int need_at_least = 0;
+            for (int i = 0; i < pg.pg_size; i++)
+            {
+                if (local_stripes[i].read_end != 0 && cur_set[i] == 0)
+                {
+                    // We need this part of the bitmap, but it's unavailable
+                    need_at_least = pg.pg_data_size;
+                    op_data->missing_flags[chain_num*pg.pg_size + i] = 1;
+                }
+                else
+                {
+                    op_data->missing_flags[chain_num*pg.pg_size + i] = 0;
+                }
+            }
+            int found = 0;
+            for (int i = 0; i < pg.pg_size; i++)
+            {
+                if (cur_set[i] != 0 && (local_stripes[i].read_end != 0 || found < need_at_least))
+                {
+                    // Read part of the bitmap
+                    bitmap_requests.push_back((bitmap_request_t){
+                        .osd_num = cur_set[i],
+                        .oid = {
+                            .inode = cur_oid.inode,
+                            .stripe = cur_oid.stripe | i,
+                        },
+                        .version = target_version,
+                        .bmp_buf = op_data->snapshot_bitmaps + (chain_num*pg.pg_size + i)*clean_entry_bitmap_size,
+                    });
+                    found++;
+                }
+            }
+            // Already checked by extend_missing_stripes, so it's fine to use assert
+            assert(found >= need_at_least);
+        }
+    }
+    std::sort(bitmap_requests.begin(), bitmap_requests.end());
+    return 0;
+}
+
+int osd_t::submit_bitmap_subops(osd_op_t *cur_op, pg_t & pg)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    std::vector<bitmap_request_t> *bitmap_requests = new std::vector<bitmap_request_t>();
+    if (collect_bitmap_requests(cur_op, pg, *bitmap_requests) < 0)
+    {
+        return -1;
+    }
+    op_data->n_subops = 0;
+    for (int i = 0; i < bitmap_requests->size(); i++)
+    {
+        if ((i == bitmap_requests->size()-1 || (*bitmap_requests)[i+1].osd_num != (*bitmap_requests)[i].osd_num) &&
+            (*bitmap_requests)[i].osd_num != this->osd_num)
+        {
+            op_data->n_subops++;
+        }
+    }
+    if (op_data->n_subops)
+    {
+        op_data->fact_ver = 0;
+        op_data->done = op_data->errors = 0;
+        op_data->subops = new osd_op_t[op_data->n_subops];
+    }
+    for (int i = 0, subop_idx = 0, prev = 0; i < bitmap_requests->size(); i++)
+    {
+        if (i == bitmap_requests->size()-1 || (*bitmap_requests)[i+1].osd_num != (*bitmap_requests)[i].osd_num)
+        {
+            osd_num_t subop_osd_num = (*bitmap_requests)[i].osd_num;
+            if (subop_osd_num == this->osd_num)
+            {
+                // Read bitmap synchronously from the local database
+                for (int j = prev; j <= i; j++)
+                {
+                    bs->read_bitmap((*bitmap_requests)[j].oid, (*bitmap_requests)[j].version, (*bitmap_requests)[j].bmp_buf, NULL);
+                }
+            }
+            else
+            {
+                // Send to a remote OSD
+                osd_op_t *subop = op_data->subops+subop_idx;
+                subop->op_type = OSD_OP_OUT;
+                subop->peer_fd = c_cli.osd_peer_fds.at(subop_osd_num);
+                // FIXME: Use the pre-allocated buffer
+                subop->buf = malloc_or_die(sizeof(obj_ver_id)*(i+1-prev));
+                subop->req = (osd_any_op_t){
+                    .sec_read_bmp = {
+                        .header = {
+                            .magic = SECONDARY_OSD_OP_MAGIC,
+                            .id = c_cli.next_subop_id++,
+                            .opcode = OSD_OP_SEC_READ_BMP,
+                        },
+                        .len = sizeof(obj_ver_id)*(i+1-prev),
+                    }
+                };
+                obj_ver_id *ov = (obj_ver_id*)subop->buf;
+                for (int j = prev; j <= i; j++, ov++)
+                {
+                    ov->oid = (*bitmap_requests)[j].oid;
+                    ov->version = (*bitmap_requests)[j].version;
+                }
+                subop->callback = [cur_op, bitmap_requests, prev, i, this](osd_op_t *subop)
+                {
+                    int requested_count = subop->req.sec_read_bmp.len / sizeof(obj_ver_id);
+                    if (subop->reply.hdr.retval == requested_count * (8 + clean_entry_bitmap_size))
+                    {
+                        void *cur_buf = subop->buf + 8;
+                        for (int j = prev; j <= i; j++)
+                        {
+                            memcpy((*bitmap_requests)[j].bmp_buf, cur_buf, clean_entry_bitmap_size);
+                            cur_buf += 8 + clean_entry_bitmap_size;
+                        }
+                    }
+                    if ((cur_op->op_data->errors + cur_op->op_data->done + 1) >= cur_op->op_data->n_subops)
+                    {
+                        delete bitmap_requests;
+                    }
+                    handle_primary_subop(subop, cur_op);
+                };
+                c_cli.outbox_push(subop);
+                subop_idx++;
+            }
+            prev = i+1;
+        }
+    }
+    if (!op_data->n_subops)
+    {
+        delete bitmap_requests;
+    }
+    return 0;
+}
+
+std::vector<osd_chain_read_t> osd_t::collect_chained_read_requests(osd_op_t *cur_op)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    std::vector<osd_chain_read_t> chain_reads;
+    int stripe_count = (op_data->scheme == POOL_SCHEME_REPLICATED ? 1 : op_data->pg_size);
+    memset(op_data->stripes[0].bmp_buf, 0, stripe_count * clean_entry_bitmap_size);
+    uint8_t *global_bitmap = (uint8_t*)op_data->stripes[0].bmp_buf;
+    // We always use at most 1 read request per layer
+    for (int chain_pos = 0; chain_pos < op_data->chain_size; chain_pos++)
+    {
+        uint8_t *part_bitmap = ((uint8_t*)op_data->snapshot_bitmaps) + chain_pos*stripe_count*clean_entry_bitmap_size;
+        int start = (cur_op->req.rw.offset - op_data->oid.stripe)/bs_bitmap_granularity;
+        int end = start + cur_op->req.rw.len/bs_bitmap_granularity;
+        // Skip unneeded part in the beginning
+        while (start < end && (
+            ((global_bitmap[start>>3] >> (start&7)) & 1) ||
+            !((part_bitmap[start>>3] >> (start&7)) & 1)))
+        {
+            start++;
+        }
+        // Skip unneeded part in the end
+        while (start < end && (
+            ((global_bitmap[(end-1)>>3] >> ((end-1)&7)) & 1) ||
+            !((part_bitmap[(end-1)>>3] >> ((end-1)&7)) & 1)))
+        {
+            end--;
+        }
+        if (start < end)
+        {
+            // Copy (OR) bits in between
+            int cur = start;
+            for (; cur < end && (cur & 0x7); cur++)
+            {
+                global_bitmap[cur>>3] = global_bitmap[cur>>3] | (part_bitmap[cur>>3] & (1 << (cur&7)));
+            }
+            for (; cur <= end-8; cur += 8)
+            {
+                global_bitmap[cur>>3] = global_bitmap[cur>>3] | part_bitmap[cur>>3];
+            }
+            for (; cur < end; cur++)
+            {
+                global_bitmap[cur>>3] = global_bitmap[cur>>3] | (part_bitmap[cur>>3] & (1 << (cur&7)));
+            }
+            // Add request
+            chain_reads.push_back((osd_chain_read_t){
+                .chain_pos = chain_pos,
+                .inode = op_data->read_chain[chain_pos],
+                .offset = start*bs_bitmap_granularity,
+                .len = (end-start)*bs_bitmap_granularity,
+            });
+        }
+    }
+    return chain_reads;
+}
+
+int osd_t::submit_chained_read_requests(pg_t & pg, osd_op_t *cur_op)
+{
+    // Decide which parts of which objects we need to read based on bitmaps
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    auto chain_reads = collect_chained_read_requests(cur_op);
+    int stripe_count = (pg.scheme == POOL_SCHEME_REPLICATED ? 1 : pg.pg_size);
+    op_data->chain_read_count = chain_reads.size();
+    op_data->chain_reads = (osd_chain_read_t*)calloc_or_die(
+        1, (sizeof(osd_chain_read_t) + sizeof(osd_rmw_stripe_t)*stripe_count) * chain_reads.size()
+    );
+    osd_rmw_stripe_t *chain_stripes = (osd_rmw_stripe_t*)(((void*)op_data->chain_reads) + sizeof(osd_chain_read_t) * op_data->chain_read_count);
+    // Now process each subrequest as a separate read, including reconstruction if needed
+    // Prepare reads
+    int n_subops = 0;
+    uint64_t read_buffer_size = 0;
+    for (int cri = 0; cri < chain_reads.size(); cri++)
+    {
+        op_data->chain_reads[cri] = chain_reads[cri];
+        object_id cur_oid = { .inode = chain_reads[cri].inode, .stripe = op_data->oid.stripe };
+        // FIXME: maybe introduce split_read_stripes to shorten these lines and to remove read_start=req_start
+        osd_rmw_stripe_t *stripes = chain_stripes + cri*stripe_count;
+        split_stripes(pg.pg_data_size, bs_block_size, chain_reads[cri].offset, chain_reads[cri].len, stripes);
+        if (op_data->scheme == POOL_SCHEME_REPLICATED && !stripes[0].req_end)
+        {
+            continue;
+        }
+        for (int role = 0; role < op_data->pg_data_size; role++)
+        {
+            stripes[role].read_start = stripes[role].req_start;
+            stripes[role].read_end = stripes[role].req_end;
+        }
+        uint64_t *cur_set = pg.cur_set.data();
+        if (pg.state != PG_ACTIVE && op_data->scheme != POOL_SCHEME_REPLICATED)
+        {
+            pg_osd_set_state_t *object_state;
+            cur_set = get_object_osd_set(pg, cur_oid, pg.cur_set.data(), &object_state);
+            if (extend_missing_stripes(stripes, cur_set, pg.pg_data_size, pg.pg_size) < 0)
+            {
+                free(op_data->chain_reads);
+                op_data->chain_reads = NULL;
+                finish_op(cur_op, -EIO);
+                return -1;
+            }
+            op_data->degraded = 1;
+        }
+        if (op_data->scheme == POOL_SCHEME_REPLICATED)
+        {
+            n_subops++;
+            read_buffer_size += stripes[0].read_end - stripes[0].read_start;
+        }
+        else
+        {
+            for (int role = 0; role < pg.pg_size; role++)
+            {
+                if (stripes[role].read_end > 0 && cur_set[role] != 0)
+                    n_subops++;
+                if (stripes[role].read_end > 0)
+                    read_buffer_size += stripes[role].read_end - stripes[role].read_start;
+            }
+        }
+    }
+    cur_op->buf = memalign_or_die(MEM_ALIGNMENT, read_buffer_size);
+    void *cur_buf = cur_op->buf;
+    for (int cri = 0; cri < chain_reads.size(); cri++)
+    {
+        osd_rmw_stripe_t *stripes = chain_stripes + cri*stripe_count;
+        for (int role = 0; role < stripe_count; role++)
+        {
+            if (stripes[role].read_end > 0)
+            {
+                stripes[role].read_buf = cur_buf;
+                stripes[role].bmp_buf = op_data->snapshot_bitmaps + (chain_reads[cri].chain_pos*stripe_count + role)*clean_entry_bitmap_size;
+                cur_buf += stripes[role].read_end - stripes[role].read_start;
+            }
+        }
+    }
+    // Submit all reads
+    op_data->fact_ver = UINT64_MAX;
+    op_data->done = op_data->errors = 0;
+    op_data->n_subops = n_subops;
+    if (!n_subops)
+    {
+        return 0;
+    }
+    op_data->subops = new osd_op_t[n_subops];
+    int cur_subops = 0;
+    for (int cri = 0; cri < chain_reads.size(); cri++)
+    {
+        osd_rmw_stripe_t *stripes = chain_stripes + cri*stripe_count;
+        if (op_data->scheme == POOL_SCHEME_REPLICATED && !stripes[0].req_end)
+        {
+            continue;
+        }
+        object_id cur_oid = { .inode = chain_reads[cri].inode, .stripe = op_data->oid.stripe };
+        auto vo_it = pg.ver_override.find(cur_oid);
+        uint64_t target_ver = vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX;
+        uint64_t *cur_set = pg.cur_set.data();
+        if (pg.state != PG_ACTIVE && op_data->scheme != POOL_SCHEME_REPLICATED)
+        {
+            pg_osd_set_state_t *object_state;
+            cur_set = get_object_osd_set(pg, cur_oid, pg.cur_set.data(), &object_state);
+        }
+        int zero_read = -1;
+        if (op_data->scheme == POOL_SCHEME_REPLICATED)
+        {
+            for (int role = 0; role < op_data->pg_size; role++)
+                if (cur_set[role] == this->osd_num || zero_read == -1)
+                    zero_read = role;
+        }
+        cur_subops += submit_primary_subop_batch(SUBMIT_READ, chain_reads[cri].inode, target_ver, stripes, cur_set, cur_op, cur_subops, zero_read);
+    }
+    assert(cur_subops == n_subops);
+    return 0;
+}
+
+void osd_t::send_chained_read_results(pg_t & pg, osd_op_t *cur_op)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    int stripe_count = (pg.scheme == POOL_SCHEME_REPLICATED ? 1 : pg.pg_size);
+    osd_rmw_stripe_t *chain_stripes = (osd_rmw_stripe_t*)(((void*)op_data->chain_reads) + sizeof(osd_chain_read_t) * op_data->chain_read_count);
+    // Reconstruct parts if needed
+    if (op_data->degraded)
+    {
+        int stripe_count = (pg.scheme == POOL_SCHEME_REPLICATED ? 1 : pg.pg_size);
+        for (int cri = 0; cri < op_data->chain_read_count; cri++)
+        {
+            // Reconstruct missing stripes
+            osd_rmw_stripe_t *stripes = chain_stripes + cri*stripe_count;
+            if (op_data->scheme == POOL_SCHEME_XOR)
+            {
+                reconstruct_stripes_xor(stripes, pg.pg_size, clean_entry_bitmap_size);
+            }
+            else if (op_data->scheme == POOL_SCHEME_JERASURE)
+            {
+                reconstruct_stripes_jerasure(stripes, pg.pg_size, pg.pg_data_size, clean_entry_bitmap_size);
+            }
+        }
+    }
+    // Send bitmap
+    cur_op->reply.rw.bitmap_len = op_data->pg_data_size * clean_entry_bitmap_size;
+    cur_op->iov.push_back(op_data->stripes[0].bmp_buf, cur_op->reply.rw.bitmap_len);
+    // And finally compose the result
+    uint64_t sent = 0;
+    int prev_pos = 0, pos = 0;
+    bool prev_set = false;
+    int prev = (cur_op->req.rw.offset - op_data->oid.stripe) / bs_bitmap_granularity;
+    int end = prev + cur_op->req.rw.len/bs_bitmap_granularity;
+    int cur = prev;
+    while (cur <= end)
+    {
+        bool has_bit = false;
+        if (cur < end)
+        {
+            for (pos = 0; pos < op_data->chain_size; pos++)
+            {
+                has_bit = (((uint8_t*)op_data->snapshot_bitmaps)[pos*stripe_count*clean_entry_bitmap_size + cur/8] >> (cur%8)) & 1;
+                if (has_bit)
+                    break;
+            }
+        }
+        if (has_bit != prev_set || pos != prev_pos || cur == end)
+        {
+            if (cur > prev)
+            {
+                // Send buffer in parts to avoid copying
+                if (!prev_set)
+                {
+                    while ((cur-prev) > zero_buffer_size/bs_bitmap_granularity)
+                    {
+                        cur_op->iov.push_back(zero_buffer, zero_buffer_size);
+                        sent += zero_buffer_size;
+                        prev += zero_buffer_size/bs_bitmap_granularity;
+                    }
+                    cur_op->iov.push_back(zero_buffer, (cur-prev)*bs_bitmap_granularity);
+                    sent += (cur-prev)*bs_bitmap_granularity;
+                }
+                else
+                {
+                    osd_rmw_stripe_t *stripes = chain_stripes + prev_pos*stripe_count;
+                    while (cur > prev)
+                    {
+                        int role = prev*bs_bitmap_granularity/bs_block_size;
+                        int role_start = prev*bs_bitmap_granularity - role*bs_block_size;
+                        int role_end = cur*bs_bitmap_granularity - role*bs_block_size;
+                        if (role_end > bs_block_size)
+                            role_end = bs_block_size;
+                        assert(stripes[role].read_buf);
+                        cur_op->iov.push_back(
+                            stripes[role].read_buf + (role_start - stripes[role].read_start),
+                            role_end - role_start
+                        );
+                        sent += role_end - role_start;
+                        prev += (role_end - role_start)/bs_bitmap_granularity;
+                    }
+                }
+            }
+            prev = cur;
+            prev_pos = pos;
+            prev_set = has_bit;
+        }
+        cur++;
+    }
+    assert(sent == cur_op->req.rw.len);
+    free(op_data->chain_reads);
+    op_data->chain_reads = NULL;
+}

src/osd_primary_subops.cpp

@@ -66,17 +66,16 @@ void osd_t::finish_op(osd_op_t *cur_op, int retval)
             auto & pg = pgs.at({ .pool_id = INODE_POOL(cur_op->op_data->oid.inode), .pg_num = cur_op->op_data->pg_num });
             pg.inflight--;
             assert(pg.inflight >= 0);
-            if ((pg.state & PG_STOPPING) && pg.inflight == 0 && !pg.flush_batch &&
-                // We must either forget all PG's unstable writes or wait for it to become clean
-                dirty_pgs.find({ .pool_id = pg.pool_id, .pg_num = pg.pg_num }) == dirty_pgs.end())
+            if ((pg.state & PG_STOPPING) && pg.inflight == 0 && !pg.flush_batch)
             {
                 finish_stop_pg(pg);
             }
+            else if ((pg.state & PG_REPEERING) && pg.inflight == 0 && !pg.flush_batch)
+            {
+                start_pg_peering(pg);
+            }
         }
         assert(!cur_op->op_data->subops);
-        assert(!cur_op->op_data->unstable_write_osds);
-        assert(!cur_op->op_data->unstable_writes);
-        assert(!cur_op->op_data->dirty_pgs);
         free(cur_op->op_data);
         cur_op->op_data = NULL;
     }
@@ -104,7 +103,7 @@ void osd_t::finish_op(osd_op_t *cur_op, int retval)
     }
 }
 
-void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_size, const uint64_t* osd_set, osd_op_t *cur_op)
+void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, const uint64_t* osd_set, osd_op_t *cur_op)
 {
     bool wr = submit_type == SUBMIT_WRITE;
     osd_primary_op_data_t *op_data = cur_op->op_data;
@@ -112,32 +111,34 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_s
     bool rep = op_data->scheme == POOL_SCHEME_REPLICATED;
     // Allocate subops
     int n_subops = 0, zero_read = -1;
-    for (int role = 0; role < pg_size; role++)
+    for (int role = 0; role < op_data->pg_size; role++)
     {
         if (osd_set[role] == this->osd_num || osd_set[role] != 0 && zero_read == -1)
-        {
             zero_read = role;
-        }
         if (osd_set[role] != 0 && (wr || !rep && stripes[role].read_end != 0))
-        {
             n_subops++;
-        }
     }
     if (!n_subops && (submit_type == SUBMIT_RMW_READ || rep))
-    {
         n_subops = 1;
-    }
     else
-    {
         zero_read = -1;
-    }
     osd_op_t *subops = new osd_op_t[n_subops];
     op_data->fact_ver = 0;
     op_data->done = op_data->errors = 0;
     op_data->n_subops = n_subops;
     op_data->subops = subops;
-    int i = 0;
-    for (int role = 0; role < pg_size; role++)
+    int sent = submit_primary_subop_batch(submit_type, op_data->oid.inode, op_version, op_data->stripes, osd_set, cur_op, 0, zero_read);
+    assert(sent == n_subops);
+}
+
+int osd_t::submit_primary_subop_batch(int submit_type, inode_t inode, uint64_t op_version,
+    osd_rmw_stripe_t *stripes, const uint64_t* osd_set, osd_op_t *cur_op, int subop_idx, int zero_read)
+{
+    bool wr = submit_type == SUBMIT_WRITE;
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    bool rep = op_data->scheme == POOL_SCHEME_REPLICATED;
+    int i = subop_idx;
+    for (int role = 0; role < op_data->pg_size; role++)
     {
         // We always submit zero-length writes to all replicas, even if the stripe is not modified
         if (!(wr || !rep && stripes[role].read_end != 0 || zero_read == role))
@@ -148,20 +149,21 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_s
         if (role_osd_num != 0)
         {
             int stripe_num = rep ? 0 : role;
+            osd_op_t *subop = op_data->subops + i;
             if (role_osd_num == this->osd_num)
             {
-                clock_gettime(CLOCK_REALTIME, &subops[i].tv_begin);
-                subops[i].op_type = (uint64_t)cur_op;
-                subops[i].bitmap = stripes[stripe_num].bmp_buf;
-                subops[i].bitmap_len = clean_entry_bitmap_size;
-                subops[i].bs_op = new blockstore_op_t({
+                clock_gettime(CLOCK_REALTIME, &subop->tv_begin);
+                subop->op_type = (uint64_t)cur_op;
+                subop->bitmap = stripes[stripe_num].bmp_buf;
+                subop->bitmap_len = clean_entry_bitmap_size;
+                subop->bs_op = new blockstore_op_t({
                     .opcode = (uint64_t)(wr ? (rep ? BS_OP_WRITE_STABLE : BS_OP_WRITE) : BS_OP_READ),
-                    .callback = [subop = &subops[i], this](blockstore_op_t *bs_subop)
+                    .callback = [subop, this](blockstore_op_t *bs_subop)
                     {
                         handle_primary_bs_subop(subop);
                     },
                     .oid = {
-                        .inode = op_data->oid.inode,
+                        .inode = inode,
                         .stripe = op_data->oid.stripe | stripe_num,
                     },
                     .version = op_version,
@@ -173,26 +175,26 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_s
 #ifdef OSD_DEBUG
                 printf(
                     "Submit %s to local: %lx:%lx v%lu %u-%u\n", wr ? "write" : "read",
-                    op_data->oid.inode, op_data->oid.stripe | stripe_num, op_version,
-                    subops[i].bs_op->offset, subops[i].bs_op->len
+                    inode, op_data->oid.stripe | stripe_num, op_version,
+                    subop->bs_op->offset, subop->bs_op->len
                 );
 #endif
-                bs->enqueue_op(subops[i].bs_op);
+                bs->enqueue_op(subop->bs_op);
             }
             else
             {
-                subops[i].op_type = OSD_OP_OUT;
-                subops[i].peer_fd = c_cli.osd_peer_fds.at(role_osd_num);
-                subops[i].bitmap = stripes[stripe_num].bmp_buf;
-                subops[i].bitmap_len = clean_entry_bitmap_size;
-                subops[i].req.sec_rw = {
+                subop->op_type = OSD_OP_OUT;
+                subop->peer_fd = c_cli.osd_peer_fds.at(role_osd_num);
+                subop->bitmap = stripes[stripe_num].bmp_buf;
+                subop->bitmap_len = clean_entry_bitmap_size;
+                subop->req.sec_rw = {
                     .header = {
                         .magic = SECONDARY_OSD_OP_MAGIC,
                         .id = c_cli.next_subop_id++,
                         .opcode = (uint64_t)(wr ? (rep ? OSD_OP_SEC_WRITE_STABLE : OSD_OP_SEC_WRITE) : OSD_OP_SEC_READ),
                     },
                     .oid = {
-                        .inode = op_data->oid.inode,
+                        .inode = inode,
                         .stripe = op_data->oid.stripe | stripe_num,
                     },
                     .version = op_version,
@@ -203,40 +205,34 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_s
 #ifdef OSD_DEBUG
                 printf(
                     "Submit %s to osd %lu: %lx:%lx v%lu %u-%u\n", wr ? "write" : "read", role_osd_num,
-                    op_data->oid.inode, op_data->oid.stripe | stripe_num, op_version,
-                    subops[i].req.sec_rw.offset, subops[i].req.sec_rw.len
+                    inode, op_data->oid.stripe | stripe_num, op_version,
+                    subop->req.sec_rw.offset, subop->req.sec_rw.len
                 );
 #endif
                 if (wr)
                 {
                     if (stripes[stripe_num].write_end > stripes[stripe_num].write_start)
                     {
-                        subops[i].iov.push_back(stripes[stripe_num].write_buf, stripes[stripe_num].write_end - stripes[stripe_num].write_start);
+                        subop->iov.push_back(stripes[stripe_num].write_buf, stripes[stripe_num].write_end - stripes[stripe_num].write_start);
                     }
                 }
                 else
                 {
                     if (stripes[stripe_num].read_end > stripes[stripe_num].read_start)
                     {
-                        subops[i].iov.push_back(stripes[stripe_num].read_buf, stripes[stripe_num].read_end - stripes[stripe_num].read_start);
+                        subop->iov.push_back(stripes[stripe_num].read_buf, stripes[stripe_num].read_end - stripes[stripe_num].read_start);
                     }
                 }
-                subops[i].callback = [cur_op, this](osd_op_t *subop)
+                subop->callback = [cur_op, this](osd_op_t *subop)
                 {
-                    int fail_fd = subop->req.hdr.opcode == OSD_OP_SEC_WRITE &&
-                        subop->reply.hdr.retval != subop->req.sec_rw.len ? subop->peer_fd : -1;
                     handle_primary_subop(subop, cur_op);
-                    if (fail_fd >= 0)
-                    {
-                        // write operation failed, drop the connection
-                        c_cli.stop_client(fail_fd);
-                    }
                 };
-                c_cli.outbox_push(&subops[i]);
+                c_cli.outbox_push(subop);
             }
             i++;
         }
     }
+    return i-subop_idx;
 }
 
 static uint64_t bs_op_to_osd_op[] = {
@@ -276,6 +272,7 @@ void osd_t::handle_primary_bs_subop(osd_op_t *subop)
     }
     delete bs_op;
     subop->bs_op = NULL;
+    subop->peer_fd = -1;
     handle_primary_subop(subop, cur_op);
 }
 
@@ -306,8 +303,13 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
 {
     uint64_t opcode = subop->req.hdr.opcode;
     int retval = subop->reply.hdr.retval;
-    int expected = opcode == OSD_OP_SEC_READ || opcode == OSD_OP_SEC_WRITE
-        || opcode == OSD_OP_SEC_WRITE_STABLE ? subop->req.sec_rw.len : 0;
+    int expected;
+    if (opcode == OSD_OP_SEC_READ || opcode == OSD_OP_SEC_WRITE || opcode == OSD_OP_SEC_WRITE_STABLE)
+        expected = subop->req.sec_rw.len;
+    else if (opcode == OSD_OP_SEC_READ_BMP)
+        expected = subop->req.sec_read_bmp.len / sizeof(obj_ver_id) * (8 + clean_entry_bitmap_size);
+    else
+        expected = 0;
     osd_primary_op_data_t *op_data = cur_op->op_data;
     if (retval != expected)
     {
@@ -317,6 +319,11 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
             op_data->epipe++;
         }
         op_data->errors++;
+        if (subop->peer_fd >= 0)
+        {
+            // Drop connection on any error
+            c_cli.stop_client(subop->peer_fd);
+        }
     }
     else
     {
@@ -329,14 +336,17 @@ void osd_t::handle_primary_subop(osd_op_t *subop, osd_op_t *cur_op)
                 ? c_cli.clients[subop->peer_fd]->osd_num : osd_num;
             printf("subop %lu from osd %lu: version = %lu\n", opcode, peer_osd, version);
 #endif
-            if (op_data->fact_ver != 0 && op_data->fact_ver != version)
+            if (op_data->fact_ver != UINT64_MAX)
             {
-                throw std::runtime_error(
-                    "different fact_versions returned from "+std::string(osd_op_names[opcode])+
-                    " subops: "+std::to_string(version)+" vs "+std::to_string(op_data->fact_ver)
-                );
+                if (op_data->fact_ver != 0 && op_data->fact_ver != version)
+                {
+                    throw std::runtime_error(
+                        "different fact_versions returned from "+std::string(osd_op_names[opcode])+
+                        " subops: "+std::to_string(version)+" vs "+std::to_string(op_data->fact_ver)
+                    );
+                }
+                op_data->fact_ver = version;
             }
-            op_data->fact_ver = version;
         }
     }
     if ((op_data->errors + op_data->done) >= op_data->n_subops)
@@ -456,7 +466,7 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
         {
             subops[i].op_type = OSD_OP_OUT;
             subops[i].peer_fd = c_cli.osd_peer_fds.at(chunk.osd_num);
-            subops[i].req.sec_del = {
+            subops[i].req = (osd_any_op_t){ .sec_del = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
                     .id = c_cli.next_subop_id++,
@@ -464,23 +474,17 @@ void osd_t::submit_primary_del_batch(osd_op_t *cur_op, obj_ver_osd_t *chunks_to_
                 },
                 .oid = chunk.oid,
                 .version = chunk.version,
-            };
+            } };
             subops[i].callback = [cur_op, this](osd_op_t *subop)
             {
-                int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : -1;
                 handle_primary_subop(subop, cur_op);
-                if (fail_fd >= 0)
-                {
-                    // delete operation failed, drop the connection
-                    c_cli.stop_client(fail_fd);
-                }
             };
             c_cli.outbox_push(&subops[i]);
         }
     }
 }
 
-void osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
+int osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
 {
     osd_primary_op_data_t *op_data = cur_op->op_data;
     int n_osds = op_data->dirty_osd_count;
@@ -488,6 +492,7 @@ void osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
     op_data->done = op_data->errors = 0;
     op_data->n_subops = n_osds;
     op_data->subops = subops;
+    std::map<uint64_t, int>::iterator peer_it;
     for (int i = 0; i < n_osds; i++)
     {
         osd_num_t sync_osd = op_data->dirty_osds[i];
@@ -504,30 +509,35 @@ void osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
             });
             bs->enqueue_op(subops[i].bs_op);
         }
-        else
+        else if ((peer_it = c_cli.osd_peer_fds.find(sync_osd)) != c_cli.osd_peer_fds.end())
         {
             subops[i].op_type = OSD_OP_OUT;
-            subops[i].peer_fd = c_cli.osd_peer_fds.at(sync_osd);
-            subops[i].req.sec_sync = {
+            subops[i].peer_fd = peer_it->second;
+            subops[i].req = (osd_any_op_t){ .sec_sync = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
                     .id = c_cli.next_subop_id++,
                     .opcode = OSD_OP_SEC_SYNC,
                 },
-            };
+            } };
             subops[i].callback = [cur_op, this](osd_op_t *subop)
             {
-                int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : -1;
                 handle_primary_subop(subop, cur_op);
-                if (fail_fd >= 0)
-                {
-                    // sync operation failed, drop the connection
-                    c_cli.stop_client(fail_fd);
-                }
             };
             c_cli.outbox_push(&subops[i]);
         }
+        else
+        {
+            op_data->done++;
+        }
     }
+    if (op_data->done >= op_data->n_subops)
+    {
+        delete[] op_data->subops;
+        op_data->subops = NULL;
+        return 0;
+    }
+    return 1;
 }
 
 void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
@@ -560,24 +570,18 @@ void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
         {
             subops[i].op_type = OSD_OP_OUT;
             subops[i].peer_fd = c_cli.osd_peer_fds.at(stab_osd.osd_num);
-            subops[i].req.sec_stab = {
+            subops[i].req = (osd_any_op_t){ .sec_stab = {
                 .header = {
                     .magic = SECONDARY_OSD_OP_MAGIC,
                     .id = c_cli.next_subop_id++,
                     .opcode = OSD_OP_SEC_STABILIZE,
                 },
                 .len = (uint64_t)(stab_osd.len * sizeof(obj_ver_id)),
-            };
+            } };
             subops[i].iov.push_back(op_data->unstable_writes + stab_osd.start, stab_osd.len * sizeof(obj_ver_id));
             subops[i].callback = [cur_op, this](osd_op_t *subop)
             {
-                int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : -1;
                 handle_primary_subop(subop, cur_op);
-                if (fail_fd >= 0)
-                {
-                    // sync operation failed, drop the connection
-                    c_cli.stop_client(fail_fd);
-                }
             };
             c_cli.outbox_push(&subops[i]);
         }
@@ -595,7 +599,7 @@ void osd_t::pg_cancel_write_queue(pg_t & pg, osd_op_t *first_op, object_id oid,
         return;
     }
     std::vector<osd_op_t*> cancel_ops;
-    while (it != pg.write_queue.end())
+    while (it != pg.write_queue.end() && it->first == oid)
     {
         cancel_ops.push_back(it->second);
         it++;

src/osd_primary_sync.cpp

@@ -0,0 +1,265 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "osd_primary.h"
+
+// Save and clear unstable_writes -> SYNC all -> STABLE all
+void osd_t::continue_primary_sync(osd_op_t *cur_op)
+{
+    if (!cur_op->op_data)
+    {
+        cur_op->op_data = (osd_primary_op_data_t*)calloc_or_die(1, sizeof(osd_primary_op_data_t));
+    }
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    if (op_data->st == 1)      goto resume_1;
+    else if (op_data->st == 2) goto resume_2;
+    else if (op_data->st == 3) goto resume_3;
+    else if (op_data->st == 4) goto resume_4;
+    else if (op_data->st == 5) goto resume_5;
+    else if (op_data->st == 6) goto resume_6;
+    else if (op_data->st == 7) goto resume_7;
+    else if (op_data->st == 8) goto resume_8;
+    assert(op_data->st == 0);
+    if (syncs_in_progress.size() > 0)
+    {
+        // Wait for previous syncs, if any
+        // FIXME: We may try to execute the current one in parallel, like in Blockstore, but I'm not sure if it matters at all
+        syncs_in_progress.push_back(cur_op);
+        op_data->st = 1;
+resume_1:
+        return;
+    }
+    else
+    {
+        syncs_in_progress.push_back(cur_op);
+    }
+resume_2:
+    if (dirty_osds.size() == 0)
+    {
+        // Nothing to sync
+        goto finish;
+    }
+    // Save and clear unstable_writes
+    // In theory it is possible to do in on a per-client basis, but this seems to be an unnecessary complication
+    // It would be cool not to copy these here at all, but someone has to deduplicate them by object IDs anyway
+    if (unstable_writes.size() > 0)
+    {
+        op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
+        op_data->unstable_writes = new obj_ver_id[this->unstable_writes.size()];
+        osd_num_t last_osd = 0;
+        int last_start = 0, last_end = 0;
+        for (auto it = this->unstable_writes.begin(); it != this->unstable_writes.end(); it++)
+        {
+            if (last_osd != it->first.osd_num)
+            {
+                if (last_osd != 0)
+                {
+                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
+                        .osd_num = last_osd,
+                        .start = last_start,
+                        .len = last_end - last_start,
+                    });
+                }
+                last_osd = it->first.osd_num;
+                last_start = last_end;
+            }
+            op_data->unstable_writes[last_end] = (obj_ver_id){
+                .oid = it->first.oid,
+                .version = it->second,
+            };
+            last_end++;
+        }
+        if (last_osd != 0)
+        {
+            op_data->unstable_write_osds->push_back((unstable_osd_num_t){
+                .osd_num = last_osd,
+                .start = last_start,
+                .len = last_end - last_start,
+            });
+        }
+        this->unstable_writes.clear();
+    }
+    {
+        void *dirty_buf = malloc_or_die(
+            sizeof(pool_pg_num_t)*dirty_pgs.size() +
+            sizeof(osd_num_t)*dirty_osds.size() +
+            sizeof(obj_ver_osd_t)*this->copies_to_delete_after_sync_count
+        );
+        op_data->dirty_pgs = (pool_pg_num_t*)dirty_buf;
+        op_data->dirty_osds = (osd_num_t*)(dirty_buf + sizeof(pool_pg_num_t)*dirty_pgs.size());
+        op_data->dirty_pg_count = dirty_pgs.size();
+        op_data->dirty_osd_count = dirty_osds.size();
+        if (this->copies_to_delete_after_sync_count)
+        {
+            op_data->copies_to_delete_count = 0;
+            op_data->copies_to_delete = (obj_ver_osd_t*)(op_data->dirty_osds + op_data->dirty_osd_count);
+            for (auto dirty_pg_num: dirty_pgs)
+            {
+                auto & pg = pgs.at(dirty_pg_num);
+                assert(pg.copies_to_delete_after_sync.size() <= this->copies_to_delete_after_sync_count);
+                memcpy(
+                    op_data->copies_to_delete + op_data->copies_to_delete_count,
+                    pg.copies_to_delete_after_sync.data(),
+                    sizeof(obj_ver_osd_t)*pg.copies_to_delete_after_sync.size()
+                );
+                op_data->copies_to_delete_count += pg.copies_to_delete_after_sync.size();
+                this->copies_to_delete_after_sync_count -= pg.copies_to_delete_after_sync.size();
+                pg.copies_to_delete_after_sync.clear();
+            }
+            assert(this->copies_to_delete_after_sync_count == 0);
+        }
+        int dpg = 0;
+        for (auto dirty_pg_num: dirty_pgs)
+        {
+            pgs.at(dirty_pg_num).inflight++;
+            op_data->dirty_pgs[dpg++] = dirty_pg_num;
+        }
+        dirty_pgs.clear();
+        dpg = 0;
+        for (auto osd_num: dirty_osds)
+        {
+            op_data->dirty_osds[dpg++] = osd_num;
+        }
+        dirty_osds.clear();
+    }
+    if (immediate_commit != IMMEDIATE_ALL)
+    {
+        // SYNC
+        if (!submit_primary_sync_subops(cur_op))
+        {
+            goto resume_4;
+        }
+resume_3:
+        op_data->st = 3;
+        return;
+resume_4:
+        if (op_data->errors > 0)
+        {
+            goto resume_6;
+        }
+    }
+    if (op_data->unstable_writes)
+    {
+        // Stabilize version sets, if any
+        submit_primary_stab_subops(cur_op);
+resume_5:
+        op_data->st = 5;
+        return;
+    }
+resume_6:
+    if (op_data->errors > 0)
+    {
+        // Return PGs and OSDs back into their dirty sets
+        for (int i = 0; i < op_data->dirty_pg_count; i++)
+        {
+            dirty_pgs.insert(op_data->dirty_pgs[i]);
+        }
+        for (int i = 0; i < op_data->dirty_osd_count; i++)
+        {
+            dirty_osds.insert(op_data->dirty_osds[i]);
+        }
+        if (op_data->unstable_writes)
+        {
+            // Return objects back into the unstable write set
+            for (auto unstable_osd: *(op_data->unstable_write_osds))
+            {
+                for (int i = 0; i < unstable_osd.len; i++)
+                {
+                    // Except those from peered PGs
+                    auto & w = op_data->unstable_writes[i];
+                    pool_pg_num_t wpg = {
+                        .pool_id = INODE_POOL(w.oid.inode),
+                        .pg_num = map_to_pg(w.oid, st_cli.pool_config.at(INODE_POOL(w.oid.inode)).pg_stripe_size),
+                    };
+                    if (pgs.at(wpg).state & PG_ACTIVE)
+                    {
+                        uint64_t & dest = this->unstable_writes[(osd_object_id_t){
+                            .osd_num = unstable_osd.osd_num,
+                            .oid = w.oid,
+                        }];
+                        dest = dest < w.version ? w.version : dest;
+                        dirty_pgs.insert(wpg);
+                    }
+                }
+            }
+        }
+        if (op_data->copies_to_delete)
+        {
+            // Return 'copies to delete' back into respective PGs
+            for (int i = 0; i < op_data->copies_to_delete_count; i++)
+            {
+                auto & w = op_data->copies_to_delete[i];
+                auto & pg = pgs.at((pool_pg_num_t){
+                    .pool_id = INODE_POOL(w.oid.inode),
+                    .pg_num = map_to_pg(w.oid, st_cli.pool_config.at(INODE_POOL(w.oid.inode)).pg_stripe_size),
+                });
+                if (pg.state & PG_ACTIVE)
+                {
+                    pg.copies_to_delete_after_sync.push_back(w);
+                    copies_to_delete_after_sync_count++;
+                }
+            }
+        }
+    }
+    else if (op_data->copies_to_delete)
+    {
+        // Actually delete copies which we wanted to delete
+        submit_primary_del_batch(cur_op, op_data->copies_to_delete, op_data->copies_to_delete_count);
+resume_7:
+        op_data->st = 7;
+        return;
+resume_8:
+        if (op_data->errors > 0)
+        {
+            goto resume_6;
+        }
+    }
+    for (int i = 0; i < op_data->dirty_pg_count; i++)
+    {
+        auto & pg = pgs.at(op_data->dirty_pgs[i]);
+        pg.inflight--;
+        if ((pg.state & PG_STOPPING) && pg.inflight == 0 && !pg.flush_batch)
+        {
+            finish_stop_pg(pg);
+        }
+        else if ((pg.state & PG_REPEERING) && pg.inflight == 0 && !pg.flush_batch)
+        {
+            start_pg_peering(pg);
+        }
+    }
+    // FIXME: Free those in the destructor?
+    free(op_data->dirty_pgs);
+    op_data->dirty_pgs = NULL;
+    op_data->dirty_osds = NULL;
+    if (op_data->unstable_writes)
+    {
+        delete op_data->unstable_write_osds;
+        delete[] op_data->unstable_writes;
+        op_data->unstable_writes = NULL;
+        op_data->unstable_write_osds = NULL;
+    }
+    if (op_data->errors > 0)
+    {
+        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
+    }
+    else
+    {
+finish:
+        if (cur_op->peer_fd)
+        {
+            auto it = c_cli.clients.find(cur_op->peer_fd);
+            if (it != c_cli.clients.end())
+                it->second->dirty_pgs.clear();
+        }
+        finish_op(cur_op, 0);
+    }
+    assert(syncs_in_progress.front() == cur_op);
+    syncs_in_progress.pop_front();
+    if (syncs_in_progress.size() > 0)
+    {
+        cur_op = syncs_in_progress.front();
+        op_data = cur_op->op_data;
+        op_data->st++;
+        goto resume_2;
+    }
+}

src/osd_primary_write.cpp

@@ -0,0 +1,381 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include "osd_primary.h"
+#include "allocator.h"
+
+bool osd_t::check_write_queue(osd_op_t *cur_op, pg_t & pg)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    // Check if actions are pending for this object
+    auto act_it = pg.flush_actions.lower_bound((obj_piece_id_t){
+        .oid = op_data->oid,
+        .osd_num = 0,
+    });
+    if (act_it != pg.flush_actions.end() &&
+        act_it->first.oid.inode == op_data->oid.inode &&
+        (act_it->first.oid.stripe & ~STRIPE_MASK) == op_data->oid.stripe)
+    {
+        pg.write_queue.emplace(op_data->oid, cur_op);
+        return false;
+    }
+    // Check if there are other write requests to the same object
+    auto vo_it = pg.write_queue.find(op_data->oid);
+    if (vo_it != pg.write_queue.end())
+    {
+        op_data->st = 1;
+        pg.write_queue.emplace(op_data->oid, cur_op);
+        return false;
+    }
+    pg.write_queue.emplace(op_data->oid, cur_op);
+    return true;
+}
+
+void osd_t::continue_primary_write(osd_op_t *cur_op)
+{
+    if (!cur_op->op_data && !prepare_primary_rw(cur_op))
+    {
+        return;
+    }
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    auto & pg = pgs.at({ .pool_id = INODE_POOL(op_data->oid.inode), .pg_num = op_data->pg_num });
+    if (op_data->st == 1)      goto resume_1;
+    else if (op_data->st == 2) goto resume_2;
+    else if (op_data->st == 3) goto resume_3;
+    else if (op_data->st == 4) goto resume_4;
+    else if (op_data->st == 5) goto resume_5;
+    else if (op_data->st == 6) goto resume_6;
+    else if (op_data->st == 7) goto resume_7;
+    else if (op_data->st == 8) goto resume_8;
+    else if (op_data->st == 9) goto resume_9;
+    else if (op_data->st == 10) goto resume_10;
+    assert(op_data->st == 0);
+    if (!check_write_queue(cur_op, pg))
+    {
+        return;
+    }
+resume_1:
+    // Determine blocks to read and write
+    // Missing chunks are allowed to be overwritten even in incomplete objects
+    // FIXME: Allow to do small writes to the old (degraded/misplaced) OSD set for lower performance impact
+    op_data->prev_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
+    if (op_data->scheme == POOL_SCHEME_REPLICATED)
+    {
+        // Simplified algorithm
+        op_data->stripes[0].write_start = op_data->stripes[0].req_start;
+        op_data->stripes[0].write_end = op_data->stripes[0].req_end;
+        op_data->stripes[0].write_buf = cur_op->buf;
+        if (pg.cur_set.data() != op_data->prev_set && (op_data->stripes[0].write_start != 0 ||
+            op_data->stripes[0].write_end != bs_block_size))
+        {
+            // Object is degraded/misplaced and will be moved to <write_osd_set>
+            op_data->stripes[0].read_start = 0;
+            op_data->stripes[0].read_end = bs_block_size;
+            cur_op->rmw_buf = op_data->stripes[0].read_buf = memalign_or_die(MEM_ALIGNMENT, bs_block_size);
+        }
+    }
+    else
+    {
+        cur_op->rmw_buf = calc_rmw(cur_op->buf, op_data->stripes, op_data->prev_set,
+            pg.pg_size, op_data->pg_data_size, pg.pg_cursize, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
+        if (!cur_op->rmw_buf)
+        {
+            // Refuse partial overwrite of an incomplete object
+            cur_op->reply.hdr.retval = -EINVAL;
+            goto continue_others;
+        }
+    }
+    // Read required blocks
+    submit_primary_subops(SUBMIT_RMW_READ, UINT64_MAX, op_data->prev_set, cur_op);
+resume_2:
+    op_data->st = 2;
+    return;
+resume_3:
+    if (op_data->errors > 0)
+    {
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        return;
+    }
+    if (op_data->scheme == POOL_SCHEME_REPLICATED)
+    {
+        // Set bitmap bits
+        bitmap_set(op_data->stripes[0].bmp_buf, op_data->stripes[0].write_start,
+            op_data->stripes[0].write_end-op_data->stripes[0].write_start, bs_bitmap_granularity);
+        // Possibly copy new data from the request into the recovery buffer
+        if (pg.cur_set.data() != op_data->prev_set && (op_data->stripes[0].write_start != 0 ||
+            op_data->stripes[0].write_end != bs_block_size))
+        {
+            memcpy(
+                op_data->stripes[0].read_buf + op_data->stripes[0].req_start,
+                op_data->stripes[0].write_buf,
+                op_data->stripes[0].req_end - op_data->stripes[0].req_start
+            );
+            op_data->stripes[0].write_buf = op_data->stripes[0].read_buf;
+            op_data->stripes[0].write_start = 0;
+            op_data->stripes[0].write_end = bs_block_size;
+        }
+    }
+    else
+    {
+        // For EC/XOR pools, save version override to make it impossible
+        // for parallel reads to read different versions of data and parity
+        pg.ver_override[op_data->oid] = op_data->fact_ver;
+        // Recover missing stripes, calculate parity
+        if (pg.scheme == POOL_SCHEME_XOR)
+        {
+            calc_rmw_parity_xor(op_data->stripes, pg.pg_size, op_data->prev_set, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
+        }
+        else if (pg.scheme == POOL_SCHEME_JERASURE)
+        {
+            calc_rmw_parity_jerasure(op_data->stripes, pg.pg_size, op_data->pg_data_size, op_data->prev_set, pg.cur_set.data(), bs_block_size, clean_entry_bitmap_size);
+        }
+    }
+    // Send writes
+    if ((op_data->fact_ver >> (64-PG_EPOCH_BITS)) < pg.epoch)
+    {
+        op_data->target_ver = ((uint64_t)pg.epoch << (64-PG_EPOCH_BITS)) | 1;
+    }
+    else
+    {
+        if ((op_data->fact_ver & (1ul<<(64-PG_EPOCH_BITS) - 1)) == (1ul<<(64-PG_EPOCH_BITS) - 1))
+        {
+            assert(pg.epoch != ((1ul << PG_EPOCH_BITS)-1));
+            pg.epoch++;
+        }
+        op_data->target_ver = op_data->fact_ver + 1;
+    }
+    if (pg.epoch > pg.reported_epoch)
+    {
+        // Report newer epoch before writing
+        // FIXME: We may report only one PG state here...
+        this->pg_state_dirty.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
+        pg.history_changed = true;
+        report_pg_states();
+resume_10:
+        if (pg.epoch > pg.reported_epoch)
+        {
+            op_data->st = 10;
+            return;
+        }
+    }
+    submit_primary_subops(SUBMIT_WRITE, op_data->target_ver, pg.cur_set.data(), cur_op);
+resume_4:
+    op_data->st = 4;
+    return;
+resume_5:
+    if (op_data->scheme != POOL_SCHEME_REPLICATED)
+    {
+        // Remove version override just after the write, but before stabilizing
+        pg.ver_override.erase(op_data->oid);
+    }
+    if (op_data->errors > 0)
+    {
+        pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+        return;
+    }
+    if (op_data->object_state)
+    {
+        // We must forget the unclean state of the object before deleting it
+        // so the next reads don't accidentally read a deleted version
+        // And it should be done at the same time as the removal of the version override
+        remove_object_from_state(op_data->oid, op_data->object_state, pg);
+        pg.clean_count++;
+    }
+resume_6:
+resume_7:
+    if (!remember_unstable_write(cur_op, pg, pg.cur_loc_set, 6))
+    {
+        return;
+    }
+    if (op_data->fact_ver == 1)
+    {
+        // Object is created
+        pg.clean_count++;
+        pg.total_count++;
+    }
+    if (op_data->object_state)
+    {
+        {
+            int recovery_type = op_data->object_state->state & (OBJ_DEGRADED|OBJ_INCOMPLETE) ? 0 : 1;
+            recovery_stat_count[0][recovery_type]++;
+            if (!recovery_stat_count[0][recovery_type])
+            {
+                recovery_stat_count[0][recovery_type]++;
+                recovery_stat_bytes[0][recovery_type] = 0;
+            }
+            for (int role = 0; role < (op_data->scheme == POOL_SCHEME_REPLICATED ? 1 : pg.pg_size); role++)
+            {
+                recovery_stat_bytes[0][recovery_type] += op_data->stripes[role].write_end - op_data->stripes[role].write_start;
+            }
+        }
+        // Any kind of a non-clean object can have extra chunks, because we don't record objects
+        // as degraded & misplaced or incomplete & misplaced at the same time. So try to remove extra chunks
+        if (immediate_commit != IMMEDIATE_ALL)
+        {
+            // We can't remove extra chunks yet if fsyncs are explicit, because
+            // new copies may not be committed to stable storage yet
+            // We can only remove extra chunks after a successful SYNC for this PG
+            for (auto & chunk: op_data->object_state->osd_set)
+            {
+                // Check is the same as in submit_primary_del_subops()
+                if (op_data->scheme == POOL_SCHEME_REPLICATED
+                    ? !contains_osd(pg.cur_set.data(), pg.pg_size, chunk.osd_num)
+                    : (chunk.osd_num != pg.cur_set[chunk.role]))
+                {
+                    pg.copies_to_delete_after_sync.push_back((obj_ver_osd_t){
+                        .osd_num = chunk.osd_num,
+                        .oid = {
+                            .inode = op_data->oid.inode,
+                            .stripe = op_data->oid.stripe | (op_data->scheme == POOL_SCHEME_REPLICATED ? 0 : chunk.role),
+                        },
+                        .version = op_data->fact_ver,
+                    });
+                    copies_to_delete_after_sync_count++;
+                }
+            }
+            free_object_state(pg, &op_data->object_state);
+        }
+        else
+        {
+            submit_primary_del_subops(cur_op, pg.cur_set.data(), pg.pg_size, op_data->object_state->osd_set);
+            free_object_state(pg, &op_data->object_state);
+            if (op_data->n_subops > 0)
+            {
+resume_8:
+                op_data->st = 8;
+                return;
+resume_9:
+                if (op_data->errors > 0)
+                {
+                    pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+                    return;
+                }
+            }
+        }
+    }
+    cur_op->reply.hdr.retval = cur_op->req.rw.len;
+continue_others:
+    osd_op_t *next_op = NULL;
+    auto next_it = pg.write_queue.find(op_data->oid);
+    // Remove the operation from queue before calling finish_op so it doesn't see the completed operation in queue
+    if (next_it != pg.write_queue.end() && next_it->second == cur_op)
+    {
+        pg.write_queue.erase(next_it++);
+        if (next_it != pg.write_queue.end() && next_it->first == op_data->oid)
+            next_op = next_it->second;
+    }
+    // finish_op would invalidate next_it if it cleared pg.write_queue, but it doesn't do that :)
+    finish_op(cur_op, cur_op->req.rw.len);
+    if (next_op)
+    {
+        // Continue next write to the same object
+        continue_primary_write(next_op);
+    }
+}
+
+bool osd_t::remember_unstable_write(osd_op_t *cur_op, pg_t & pg, pg_osd_set_t & loc_set, int base_state)
+{
+    osd_primary_op_data_t *op_data = cur_op->op_data;
+    if (op_data->st == base_state)
+    {
+        goto resume_6;
+    }
+    else if (op_data->st == base_state+1)
+    {
+        goto resume_7;
+    }
+    if (immediate_commit == IMMEDIATE_ALL)
+    {
+immediate:
+        if (op_data->scheme != POOL_SCHEME_REPLICATED)
+        {
+            // Send STABILIZE ops immediately
+            op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
+            op_data->unstable_writes = new obj_ver_id[loc_set.size()];
+            {
+                int last_start = 0;
+                for (auto & chunk: loc_set)
+                {
+                    op_data->unstable_writes[last_start] = (obj_ver_id){
+                        .oid = {
+                            .inode = op_data->oid.inode,
+                            .stripe = op_data->oid.stripe | chunk.role,
+                        },
+                        .version = op_data->fact_ver,
+                    };
+                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
+                        .osd_num = chunk.osd_num,
+                        .start = last_start,
+                        .len = 1,
+                    });
+                    last_start++;
+                }
+            }
+            submit_primary_stab_subops(cur_op);
+resume_6:
+            op_data->st = 6;
+            return false;
+resume_7:
+            // FIXME: Free those in the destructor?
+            delete op_data->unstable_write_osds;
+            delete[] op_data->unstable_writes;
+            op_data->unstable_writes = NULL;
+            op_data->unstable_write_osds = NULL;
+            if (op_data->errors > 0)
+            {
+                pg_cancel_write_queue(pg, cur_op, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
+                return false;
+            }
+        }
+    }
+    else if (immediate_commit == IMMEDIATE_SMALL)
+    {
+        int stripe_count = (op_data->scheme == POOL_SCHEME_REPLICATED ? 1 : op_data->pg_size);
+        for (int role = 0; role < stripe_count; role++)
+        {
+            if (op_data->stripes[role].write_start == 0 &&
+                op_data->stripes[role].write_end == bs_block_size)
+            {
+                // Big write. Treat write as unsynced
+                goto lazy;
+            }
+        }
+        goto immediate;
+    }
+    else
+    {
+lazy:
+        if (op_data->scheme != POOL_SCHEME_REPLICATED)
+        {
+            // Remember version as unstable for EC/XOR
+            for (auto & chunk: loc_set)
+            {
+                this->dirty_osds.insert(chunk.osd_num);
+                this->unstable_writes[(osd_object_id_t){
+                    .osd_num = chunk.osd_num,
+                    .oid = {
+                        .inode = op_data->oid.inode,
+                        .stripe = op_data->oid.stripe | chunk.role,
+                    },
+                }] = op_data->fact_ver;
+            }
+        }
+        else
+        {
+            // Only remember to sync OSDs for replicated pools
+            for (auto & chunk: loc_set)
+            {
+                this->dirty_osds.insert(chunk.osd_num);
+            }
+        }
+        // Remember PG as dirty to drop the connection when PG goes offline
+        // (this is required because of the "lazy sync")
+        auto cl_it = c_cli.clients.find(cur_op->peer_fd);
+        if (cl_it != c_cli.clients.end())
+        {
+            cl_it->second->dirty_pgs.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
+        }
+        dirty_pgs.insert({ .pool_id = pg.pool_id, .pg_num = pg.pg_num });
+    }
+    return true;
+}

src/osd_rmw.cpp

@@ -246,20 +246,23 @@ void reconstruct_stripes_jerasure(osd_rmw_stripe_t *stripes, int pg_size, int pg
     {
         if (stripes[role].read_end != 0 && stripes[role].missing)
         {
-            for (int other = 0; other < pg_size; other++)
+            if (stripes[role].read_end > stripes[role].read_start)
             {
-                if (stripes[other].read_end != 0 && !stripes[other].missing)
+                for (int other = 0; other < pg_size; other++)
                 {
-                    assert(stripes[other].read_start <= stripes[role].read_start);
-                    assert(stripes[other].read_end >= stripes[role].read_end);
-                    data_ptrs[other] = (char*)(stripes[other].read_buf + (stripes[role].read_start - stripes[other].read_start));
+                    if (stripes[other].read_end != 0 && !stripes[other].missing)
+                    {
+                        assert(stripes[other].read_start <= stripes[role].read_start);
+                        assert(stripes[other].read_end >= stripes[role].read_end);
+                        data_ptrs[other] = (char*)(stripes[other].read_buf + (stripes[role].read_start - stripes[other].read_start));
+                    }
                 }
+                data_ptrs[role] = (char*)stripes[role].read_buf;
+                jerasure_matrix_dotprod(
+                    pg_minsize, OSD_JERASURE_W, decoding_matrix+(role*pg_minsize), dm_ids, role,
+                    data_ptrs, data_ptrs+pg_minsize, stripes[role].read_end - stripes[role].read_start
+                );
             }
-            data_ptrs[role] = (char*)stripes[role].read_buf;
-            jerasure_matrix_dotprod(
-                pg_minsize, OSD_JERASURE_W, decoding_matrix+(role*pg_minsize), dm_ids, role,
-                data_ptrs, data_ptrs+pg_minsize, stripes[role].read_end - stripes[role].read_start
-            );
             for (int other = 0; other < pg_size; other++)
             {
                 if (stripes[other].read_end != 0 && !stripes[other].missing)

src/osd_secondary.cpp

@@ -44,6 +44,25 @@ void osd_t::secondary_op_callback(osd_op_t *op)
 
 void osd_t::exec_secondary(osd_op_t *cur_op)
 {
+    if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
+    {
+        int n = cur_op->req.sec_read_bmp.len / sizeof(obj_ver_id);
+        if (n > 0)
+        {
+            obj_ver_id *ov = (obj_ver_id*)cur_op->buf;
+            void *reply_buf = malloc_or_die(n * (8 + clean_entry_bitmap_size));
+            void *cur_buf = reply_buf;
+            for (int i = 0; i < n; i++)
+            {
+                bs->read_bitmap(ov[i].oid, ov[i].version, cur_buf + sizeof(uint64_t), (uint64_t*)cur_buf);
+                cur_buf += (8 + clean_entry_bitmap_size);
+            }
+            free(cur_op->buf);
+            cur_op->buf = reply_buf;
+        }
+        finish_op(cur_op, n * (8 + clean_entry_bitmap_size));
+        return;
+    }
     cur_op->bs_op = new blockstore_op_t();
     cur_op->bs_op->callback = [this, cur_op](blockstore_op_t* bs_op) { secondary_op_callback(cur_op); };
     cur_op->bs_op->opcode = (cur_op->req.hdr.opcode == OSD_OP_SEC_READ ? BS_OP_READ
@@ -126,7 +145,9 @@ void osd_t::exec_secondary(osd_op_t *cur_op)
 void osd_t::exec_show_config(osd_op_t *cur_op)
 {
     // FIXME: Send the real config, not its source
-    std::string cfg_str = json11::Json(config).dump();
+    auto cfg_copy = config;
+    cfg_copy["protocol_version"] = std::to_string(OSD_PROTOCOL_VERSION);
+    std::string cfg_str = json11::Json(cfg_copy).dump();
     cur_op->buf = malloc_or_die(cfg_str.size()+1);
     memcpy(cur_op->buf, cfg_str.c_str(), cfg_str.size()+1);
     cur_op->iov.push_back(cur_op->buf, cfg_str.size()+1);

src/pg_states.cpp

@@ -3,13 +3,14 @@
 
 #include "pg_states.h"
 
-const int pg_state_bit_count = 14;
+const int pg_state_bit_count = 15;
 
-const int pg_state_bits[14] = {
+const int pg_state_bits[15] = {
     PG_STARTING,
     PG_PEERING,
     PG_INCOMPLETE,
     PG_ACTIVE,
+    PG_REPEERING,
     PG_STOPPING,
     PG_OFFLINE,
     PG_DEGRADED,
@@ -21,11 +22,12 @@ const int pg_state_bits[14] = {
     PG_LEFT_ON_DEAD,
 };
 
-const char *pg_state_names[14] = {
+const char *pg_state_names[15] = {
     "starting",
     "peering",
     "incomplete",
     "active",
+    "repeering",
     "stopping",
     "offline",
     "degraded",

src/pg_states.h

@@ -10,16 +10,17 @@
 #define PG_PEERING (1<<1)
 #define PG_INCOMPLETE (1<<2)
 #define PG_ACTIVE (1<<3)
-#define PG_STOPPING (1<<4)
-#define PG_OFFLINE (1<<5)
+#define PG_REPEERING (1<<4)
+#define PG_STOPPING (1<<5)
+#define PG_OFFLINE (1<<6)
 // Plus any of these:
-#define PG_DEGRADED (1<<6)
-#define PG_HAS_INCOMPLETE (1<<7)
-#define PG_HAS_DEGRADED (1<<8)
-#define PG_HAS_MISPLACED (1<<9)
-#define PG_HAS_UNCLEAN (1<<10)
-#define PG_HAS_INVALID (1<<11)
-#define PG_LEFT_ON_DEAD (1<<12)
+#define PG_DEGRADED (1<<7)
+#define PG_HAS_INCOMPLETE (1<<8)
+#define PG_HAS_DEGRADED (1<<9)
+#define PG_HAS_MISPLACED (1<<10)
+#define PG_HAS_UNCLEAN (1<<11)
+#define PG_HAS_INVALID (1<<12)
+#define PG_LEFT_ON_DEAD (1<<13)
 
 // Lower bits that represent object role (EC 0/1/2... or always 0 with replication)
 // 12 bits is a safe default that doesn't depend on pg_stripe_size or pg_block_size

src/test_allocator.cpp

@@ -20,7 +20,15 @@ void alloc_all(int size)
         {
             printf("incorrect block allocated: expected %d, got %lu\n", i, x);
         }
+        if (a->get(x))
+        {
+            printf("not free before set at %d\n", i);
+        }
         a->set(x, true);
+        if (!a->get(x))
+        {
+            printf("free after set at %d\n", i);
+        }
     }
     uint64_t x = a->find_free();
     if (x != UINT64_MAX)

src/test_blockstore.cpp

@@ -2,8 +2,8 @@
 // License: VNPL-1.1 (see README.md for details)
 
 #include <malloc.h>
-#include "timerfd_interval.h"
 #include "blockstore.h"
+#include "epoll_manager.h"
 
 int main(int narg, char *args[])
 {
@@ -12,11 +12,8 @@ int main(int narg, char *args[])
     config["journal_device"] = "./test_journal.bin";
     config["data_device"] = "./test_data.bin";
     ring_loop_t *ringloop = new ring_loop_t(512);
-    blockstore_t *bs = new blockstore_t(config, ringloop);
-    timerfd_interval tick_tfd(ringloop, 1, []()
-    {
-        printf("tick 1s\n");
-    });
+    epoll_manager_t *epmgr = new epoll_manager_t(ringloop);
+    blockstore_t *bs = new blockstore_t(config, ringloop, epmgr->tfd);
 
     blockstore_op_t op;
     int main_state = 0;
@@ -125,6 +122,7 @@ int main(int narg, char *args[])
         ringloop->wait();
     }
     delete bs;
+    delete epmgr;
     delete ringloop;
     return 0;
 }

src/test_cluster_client.cpp

@@ -0,0 +1,407 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+#include "cluster_client.h"
+
+void configure_single_pg_pool(cluster_client_t *cli)
+{
+    cli->st_cli.on_load_pgs_hook(true);
+    cli->st_cli.parse_state((etcd_kv_t){
+        .key = "/config/pools",
+        .value = json11::Json::object {
+            { "1", json11::Json::object {
+                { "name", "hddpool" },
+                { "scheme", "replicated" },
+                { "pg_size", 2 },
+                { "pg_minsize", 1 },
+                { "pg_count", 1 },
+                { "failure_domain", "osd" },
+            } }
+        },
+    });
+    cli->st_cli.parse_state((etcd_kv_t){
+        .key = "/config/pgs",
+        .value = json11::Json::object {
+            { "items", json11::Json::object {
+                { "1", json11::Json::object {
+                    { "1", json11::Json::object {
+                        { "osd_set", json11::Json::array { 1, 2 } },
+                        { "primary", 1 },
+                    } }
+                } }
+            } }
+        },
+    });
+    cli->st_cli.parse_state((etcd_kv_t){
+        .key = "/pg/state/1/1",
+        .value = json11::Json::object {
+            { "peers", json11::Json::array { 1, 2 } },
+            { "primary", 1 },
+            { "state", json11::Json::array { "active" } },
+        },
+    });
+    std::map<std::string, etcd_kv_t> changes;
+    cli->st_cli.on_change_hook(changes);
+}
+
+int *test_write(cluster_client_t *cli, uint64_t offset, uint64_t len, uint8_t c, std::function<void()> cb = NULL)
+{
+    printf("Post write %lx+%lx\n", offset, len);
+    int *r = new int;
+    *r = -1;
+    cluster_op_t *op = new cluster_op_t();
+    op->opcode = OSD_OP_WRITE;
+    op->inode = 0x1000000000001;
+    op->offset = offset;
+    op->len = len;
+    op->iov.push_back(malloc_or_die(len), len);
+    memset(op->iov.buf[0].iov_base, c, len);
+    op->callback = [r, cb](cluster_op_t *op)
+    {
+        if (*r == -1)
+            printf("Error: Not allowed to complete yet\n");
+        assert(*r != -1);
+        *r = op->retval == op->len ? 1 : 0;
+        free(op->iov.buf[0].iov_base);
+        printf("Done write %lx+%lx r=%d\n", op->offset, op->len, op->retval);
+        delete op;
+        if (cb != NULL)
+            cb();
+    };
+    cli->execute(op);
+    return r;
+}
+
+int *test_sync(cluster_client_t *cli)
+{
+    printf("Post sync\n");
+    int *r = new int;
+    *r = -1;
+    cluster_op_t *op = new cluster_op_t();
+    op->opcode = OSD_OP_SYNC;
+    op->callback = [r](cluster_op_t *op)
+    {
+        if (*r == -1)
+            printf("Error: Not allowed to complete yet\n");
+        assert(*r != -1);
+        *r = op->retval == 0 ? 1 : 0;
+        printf("Done sync r=%d\n", op->retval);
+        delete op;
+    };
+    cli->execute(op);
+    return r;
+}
+
+void can_complete(int *r)
+{
+    // Allow the operation to proceed so the test verifies
+    // that it doesn't complete earlier than expected
+    *r = -2;
+}
+
+void check_completed(int *r)
+{
+    assert(*r == 1);
+    delete r;
+}
+
+void pretend_connected(cluster_client_t *cli, osd_num_t osd_num)
+{
+    printf("OSD %lu connected\n", osd_num);
+    int peer_fd = cli->msgr.clients.size() ? std::prev(cli->msgr.clients.end())->first+1 : 10;
+    cli->msgr.osd_peer_fds[osd_num] = peer_fd;
+    cli->msgr.clients[peer_fd] = new osd_client_t();
+    cli->msgr.clients[peer_fd]->osd_num = osd_num;
+    cli->msgr.clients[peer_fd]->peer_state = PEER_CONNECTED;
+    cli->msgr.wanted_peers.erase(osd_num);
+    cli->msgr.repeer_pgs(osd_num);
+}
+
+void pretend_disconnected(cluster_client_t *cli, osd_num_t osd_num)
+{
+    printf("OSD %lu disconnected\n", osd_num);
+    cli->msgr.stop_client(cli->msgr.osd_peer_fds.at(osd_num));
+}
+
+void check_disconnected(cluster_client_t *cli, osd_num_t osd_num)
+{
+    if (cli->msgr.osd_peer_fds.find(osd_num) != cli->msgr.osd_peer_fds.end())
+    {
+        printf("OSD %lu not disconnected as it ought to be\n", osd_num);
+        assert(0);
+    }
+}
+
+void check_op_count(cluster_client_t *cli, osd_num_t osd_num, int ops)
+{
+    int peer_fd = cli->msgr.osd_peer_fds.at(osd_num);
+    int real_ops = cli->msgr.clients[peer_fd]->sent_ops.size();
+    if (real_ops != ops)
+    {
+        printf("error: %d ops expected, but %d queued\n", ops, real_ops);
+        assert(0);
+    }
+}
+
+osd_op_t *find_op(cluster_client_t *cli, osd_num_t osd_num, uint64_t opcode, uint64_t offset, uint64_t len)
+{
+    int peer_fd = cli->msgr.osd_peer_fds.at(osd_num);
+    auto op_it = cli->msgr.clients[peer_fd]->sent_ops.begin();
+    while (op_it != cli->msgr.clients[peer_fd]->sent_ops.end())
+    {
+        auto op = op_it->second;
+        if (op->req.hdr.opcode == opcode && (opcode == OSD_OP_SYNC ||
+            op->req.rw.inode == 0x1000000000001 && op->req.rw.offset == offset && op->req.rw.len == len))
+        {
+            return op;
+        }
+        op_it++;
+    }
+    return NULL;
+}
+
+void pretend_op_completed(cluster_client_t *cli, osd_op_t *op, int64_t retval)
+{
+    assert(op);
+    printf("Pretend completed %s %lx+%x\n", op->req.hdr.opcode == OSD_OP_SYNC
+        ? "sync" : (op->req.hdr.opcode == OSD_OP_WRITE ? "write" : "read"), op->req.rw.offset, op->req.rw.len);
+    uint64_t op_id = op->req.hdr.id;
+    int peer_fd = op->peer_fd;
+    cli->msgr.clients[peer_fd]->sent_ops.erase(op_id);
+    op->reply.hdr.magic = SECONDARY_OSD_REPLY_MAGIC;
+    op->reply.hdr.id = op->req.hdr.id;
+    op->reply.hdr.opcode = op->req.hdr.opcode;
+    op->reply.hdr.retval = retval < 0 ? retval : (op->req.hdr.opcode == OSD_OP_SYNC ? 0 : op->req.rw.len);
+    // Copy lambda to be unaffected by `delete op`
+    std::function<void(osd_op_t*)>(op->callback)(op);
+}
+
+void test1()
+{
+    json11::Json config;
+    timerfd_manager_t *tfd = new timerfd_manager_t([](int fd, bool wr, std::function<void(int, int)> callback){});
+    cluster_client_t *cli = new cluster_client_t(NULL, tfd, config);
+
+    int *r1 = test_write(cli, 0, 4096, 0x55);
+    configure_single_pg_pool(cli);
+    pretend_connected(cli, 1);
+    cli->continue_ops(true);
+    can_complete(r1);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
+    check_completed(r1);
+    pretend_disconnected(cli, 1);
+    int *r2 = test_sync(cli);
+    pretend_connected(cli, 1);
+    check_op_count(cli, 1, 0);
+    cli->continue_ops(true);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 4096), 0);
+    check_op_count(cli, 1, 1);
+    can_complete(r2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_completed(r2);
+    // Check that the client doesn't repeat operations once more
+    pretend_disconnected(cli, 1);
+    pretend_connected(cli, 1);
+    check_op_count(cli, 1, 0);
+
+    // Case:
+    // Write(1) -> Complete Write(1) -> Overwrite(2) -> Complete Write(2)
+    // -> Overwrite(3) -> Drop OSD connection -> Reestablish OSD connection
+    // -> Complete All Posted Writes -> Sync -> Complete Sync
+    // The resulting state of the block must be (3) over (2) over (1).
+    // I.e. the part overwritten by (3) must remain as in (3) and so on.
+
+    // More interesting case:
+    // Same, but both Write(2) and Write(3) must consist of two parts:
+    // one from an OSD 2 that drops connection and other from OSD 1 that doesn't.
+    // The idea is that if the whole Write(2) is repeated when OSD 2 drops connection
+    // then it may also overwrite a part in OSD 1 which shouldn't be overwritten.
+
+    // Another interesting case:
+    // A new operation added during replay (would also break with the previous implementation)
+
+    r1 = test_write(cli, 0, 0x10000, 0x56);
+    can_complete(r1);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x10000), 0);
+    check_completed(r1);
+
+    r1 = test_write(cli, 0xE000, 0x4000, 0x57);
+    can_complete(r1);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0xE000, 0x4000), 0);
+    check_completed(r1);
+
+    r1 = test_write(cli, 0x10000, 0x4000, 0x58);
+
+    pretend_disconnected(cli, 1);
+    pretend_connected(cli, 1);
+    cli->continue_ops(true);
+
+    // Check replay
+    {
+        uint64_t replay_start = UINT64_MAX;
+        uint64_t replay_end = 0;
+        std::vector<osd_op_t*> replay_ops;
+        auto osd_cl = cli->msgr.clients.at(cli->msgr.osd_peer_fds.at(1));
+        for (auto & op_p: osd_cl->sent_ops)
+        {
+            auto op = op_p.second;
+            assert(op->req.hdr.opcode == OSD_OP_WRITE);
+            uint64_t offset = op->req.rw.offset;
+            if (op->req.rw.offset < replay_start)
+                replay_start = op->req.rw.offset;
+            if (op->req.rw.offset+op->req.rw.len > replay_end)
+                replay_end = op->req.rw.offset+op->req.rw.len;
+            for (int buf_idx = 0; buf_idx < op->iov.count; buf_idx++)
+            {
+                for (int i = 0; i < op->iov.buf[buf_idx].iov_len; i++, offset++)
+                {
+                    uint8_t c = offset < 0xE000 ? 0x56 : (offset < 0x10000 ? 0x57 : 0x58);
+                    if (((uint8_t*)op->iov.buf[buf_idx].iov_base)[i] != c)
+                    {
+                        printf("Write replay: mismatch at %lu\n", offset-op->req.rw.offset);
+                        goto fail;
+                    }
+                }
+            }
+        fail:
+            assert(offset == op->req.rw.offset+op->req.rw.len);
+            replay_ops.push_back(op);
+        }
+        if (replay_start != 0 || replay_end != 0x14000)
+        {
+            printf("Write replay: range mismatch: %lx-%lx\n", replay_start, replay_end);
+            assert(0);
+        }
+        for (auto op: replay_ops)
+        {
+            pretend_op_completed(cli, op, 0);
+        }
+    }
+    // Check that the following write finally proceeds
+    check_op_count(cli, 1, 1);
+    can_complete(r1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0x10000, 0x4000), 0);
+    check_completed(r1);
+    check_op_count(cli, 1, 0);
+
+    // Check sync
+    r2 = test_sync(cli);
+    can_complete(r2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_SYNC, 0, 0), 0);
+    check_completed(r2);
+
+    // Check disconnect during write
+    r1 = test_write(cli, 0, 4096, 0x59);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), -EPIPE);
+    check_disconnected(cli, 1);
+    pretend_connected(cli, 1);
+    check_op_count(cli, 1, 0);
+    cli->continue_ops(true);
+    check_op_count(cli, 1, 1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
+    check_op_count(cli, 1, 1);
+    can_complete(r1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
+    check_completed(r1);
+
+    // Check disconnect inside operation callback (reenterability)
+    // Probably doesn't happen too often, but possible in theory
+    r1 = test_write(cli, 0, 0x1000, 0x60, [cli]()
+    {
+        pretend_disconnected(cli, 1);
+    });
+    r2 = test_write(cli, 0x1000, 0x1000, 0x61);
+    check_op_count(cli, 1, 2);
+    can_complete(r1);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
+    check_completed(r1);
+    check_disconnected(cli, 1);
+    pretend_connected(cli, 1);
+    cli->continue_ops(true);
+    check_op_count(cli, 1, 2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0, 0x1000), 0);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0x1000, 0x1000), 0);
+    check_op_count(cli, 1, 1);
+    can_complete(r2);
+    pretend_op_completed(cli, find_op(cli, 1, OSD_OP_WRITE, 0x1000, 0x1000), 0);
+    check_completed(r2);
+
+    // Free client
+    delete cli;
+    delete tfd;
+    printf("[ok] write replay test\n");
+}
+
+void test2()
+{
+    std::map<object_id, cluster_buffer_t> unsynced_writes;
+    cluster_op_t *op = new cluster_op_t();
+    op->opcode = OSD_OP_WRITE;
+    op->inode = 1;
+    op->offset = 0;
+    op->len = 4096;
+    op->iov.push_back(malloc_or_die(4096*1024), 4096);
+    // 0-4k = 0x55
+    memset(op->iov.buf[0].iov_base, 0x55, op->iov.buf[0].iov_len);
+    cluster_client_t::copy_write(op, unsynced_writes);
+    // 8k-12k = 0x66
+    op->offset = 8192;
+    memset(op->iov.buf[0].iov_base, 0x66, op->iov.buf[0].iov_len);
+    cluster_client_t::copy_write(op, unsynced_writes);
+    // 4k-1M+4k = 0x77
+    op->len = op->iov.buf[0].iov_len = 1048576;
+    op->offset = 4096;
+    memset(op->iov.buf[0].iov_base, 0x77, op->iov.buf[0].iov_len);
+    cluster_client_t::copy_write(op, unsynced_writes);
+    // check it
+    assert(unsynced_writes.size() == 4);
+    auto uit = unsynced_writes.begin();
+    int i;
+    assert(uit->first.inode == 1);
+    assert(uit->first.stripe == 0);
+    assert(uit->second.len == 4096);
+    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x55; i++) {}
+    assert(i == uit->second.len);
+    uit++;
+    assert(uit->first.inode == 1);
+    assert(uit->first.stripe == 4096);
+    assert(uit->second.len == 4096);
+    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
+    assert(i == uit->second.len);
+    uit++;
+    assert(uit->first.inode == 1);
+    assert(uit->first.stripe == 8192);
+    assert(uit->second.len == 4096);
+    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
+    assert(i == uit->second.len);
+    uit++;
+    assert(uit->first.inode == 1);
+    assert(uit->first.stripe == 12*1024);
+    assert(uit->second.len == 1016*1024);
+    for (i = 0; i < uit->second.len && ((uint8_t*)uit->second.buf)[i] == 0x77; i++) {}
+    assert(i == uit->second.len);
+    uit++;
+    // free memory
+    free(op->iov.buf[0].iov_base);
+    delete op;
+    for (auto p: unsynced_writes)
+    {
+        free(p.second.buf);
+    }
+    printf("[ok] copy_write test\n");
+}
+
+int main(int narg, char *args[])
+{
+    test1();
+    test2();
+    return 0;
+}

src/timerfd_interval.cpp

@@ -1,64 +0,0 @@
-// Copyright (c) Vitaliy Filippov, 2019+
-// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
-
-#include <sys/timerfd.h>
-#include <sys/poll.h>
-#include <unistd.h>
-#include "timerfd_interval.h"
-
-timerfd_interval::timerfd_interval(ring_loop_t *ringloop, int seconds, std::function<void(void)> cb)
-{
-    wait_state = 0;
-    timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
-    if (timerfd < 0)
-    {
-        throw std::runtime_error(std::string("timerfd_create: ") + strerror(errno));
-    }
-    struct itimerspec exp = {
-        .it_interval = { seconds, 0 },
-        .it_value = { seconds, 0 },
-    };
-    if (timerfd_settime(timerfd, 0, &exp, NULL))
-    {
-        throw std::runtime_error(std::string("timerfd_settime: ") + strerror(errno));
-    }
-    consumer.loop = [this]() { loop(); };
-    ringloop->register_consumer(&consumer);
-    this->ringloop = ringloop;
-    this->callback = cb;
-}
-
-timerfd_interval::~timerfd_interval()
-{
-    ringloop->unregister_consumer(&consumer);
-    close(timerfd);
-}
-
-void timerfd_interval::loop()
-{
-    if (wait_state == 1)
-    {
-        return;
-    }
-    struct io_uring_sqe *sqe = ringloop->get_sqe();
-    if (!sqe)
-    {
-        wait_state = 0;
-        return;
-    }
-    struct ring_data_t *data = ((ring_data_t*)sqe->user_data);
-    my_uring_prep_poll_add(sqe, timerfd, POLLIN);
-    data->callback = [&](ring_data_t *data)
-    {
-        if (data->res < 0)
-        {
-            throw std::runtime_error(std::string("waiting for timer failed: ") + strerror(-data->res));
-        }
-        uint64_t n;
-        read(timerfd, &n, 8);
-        wait_state = 0;
-        callback();
-    };
-    wait_state = 1;
-    ringloop->submit();
-}

src/timerfd_interval.h

@@ -1,19 +0,0 @@
-// Copyright (c) Vitaliy Filippov, 2019+
-// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
-
-#pragma once
-
-#include "ringloop.h"
-
-class timerfd_interval
-{
-    int wait_state;
-    int timerfd;
-    ring_loop_t *ringloop;
-    ring_consumer_t consumer;
-    std::function<void(void)> callback;
-public:
-    timerfd_interval(ring_loop_t *ringloop, int seconds, std::function<void(void)> cb);
-    ~timerfd_interval();
-    void loop();
-};

src/timerfd_manager.cpp

@@ -34,8 +34,8 @@ timerfd_manager_t::~timerfd_manager_t()
 
 void timerfd_manager_t::inc_timer(timerfd_timer_t & t)
 {
-    t.next.tv_sec += t.millis/1000;
-    t.next.tv_nsec += (t.millis%1000)*1000000;
+    t.next.tv_sec += t.micros/1000000;
+    t.next.tv_nsec += (t.micros%1000000)*1000;
     if (t.next.tv_nsec > 1000000000)
     {
         t.next.tv_sec++;
@@ -44,13 +44,18 @@ void timerfd_manager_t::inc_timer(timerfd_timer_t & t)
 }
 
 int timerfd_manager_t::set_timer(uint64_t millis, bool repeat, std::function<void(int)> callback)
+{
+    return set_timer_us(millis*1000, repeat, callback);
+}
+
+int timerfd_manager_t::set_timer_us(uint64_t micros, bool repeat, std::function<void(int)> callback)
 {
     int timer_id = id++;
     timespec start;
     clock_gettime(CLOCK_MONOTONIC, &start);
     timers.push_back({
         .id = timer_id,
-        .millis = millis,
+        .micros = micros,
         .start = start,
         .next = start,
         .repeat = repeat,
@@ -121,7 +126,7 @@ again:
             exp.it_value.tv_sec--;
             exp.it_value.tv_nsec += 1000000000;
         }
-        if (exp.it_value.tv_sec < 0 || !exp.it_value.tv_sec && !exp.it_value.tv_nsec)
+        if (exp.it_value.tv_sec < 0 || exp.it_value.tv_sec == 0 && exp.it_value.tv_nsec <= 0)
         {
             // It already happened
             trigger_nearest();
@@ -159,6 +164,6 @@ void timerfd_manager_t::trigger_nearest()
     {
         timers.erase(timers.begin()+nearest, timers.begin()+nearest+1);
     }
-    cb(nearest_id);
     nearest = -1;
+    cb(nearest_id);
 }

src/timerfd_manager.h

@@ -10,7 +10,7 @@
 struct timerfd_timer_t
 {
     int id;
-    uint64_t millis;
+    uint64_t micros;
     timespec start, next;
     bool repeat;
     std::function<void(int)> callback;
@@ -34,5 +34,6 @@ public:
     timerfd_manager_t(std::function<void(int, bool, std::function<void(int, int)>)> set_fd_handler);
     ~timerfd_manager_t();
     int set_timer(uint64_t millis, bool repeat, std::function<void(int)> callback);
+    int set_timer_us(uint64_t micros, bool repeat, std::function<void(int)> callback);
     void clear_timer(int timer_id);
 };

tests/test_change_pg_count.sh

@@ -2,6 +2,14 @@
 
 . `dirname $0`/common.sh
 
+if [ "$EC" != "" ]; then
+    POOLCFG='"scheme":"xor","pg_size":3,"pg_minsize":2,"parity_chunks":1'
+    NOBJ=512
+else
+    POOLCFG='"scheme":"replicated","pg_size":2,"pg_minsize":2'
+    NOBJ=1024
+fi
+
 dd if=/dev/zero of=./testdata/test_osd1.bin bs=1024 count=1 seek=$((1024*1024-1))
 dd if=/dev/zero of=./testdata/test_osd2.bin bs=1024 count=1 seek=$((1024*1024-1))
 dd if=/dev/zero of=./testdata/test_osd3.bin bs=1024 count=1 seek=$((1024*1024-1))
@@ -28,7 +36,7 @@ cd ..
 node mon/mon-main.js --etcd_url http://$ETCD_URL --etcd_prefix "/vitastor" --verbose 1 &>./testdata/mon.log &
 MON_PID=$!
 
-$ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool","scheme":"replicated","pg_size":2,"pg_minsize":2,"pg_count":16,"failure_domain":"osd"}}'
+$ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool",'$POOLCFG',"pg_count":16,"failure_domain":"osd"}}'
 
 sleep 2
 
@@ -52,7 +60,7 @@ try_change()
         echo --- Change PG count to $n --- >>testdata/osd$i.log
     done
 
-    $ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool","scheme":"replicated","pg_size":2,"pg_minsize":2,"pg_count":'$n',"failure_domain":"osd"}}'
+    $ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool",'$POOLCFG',"pg_count":'$n',"failure_domain":"osd"}}'
 
     for i in {1..10}; do
         ($ETCDCTL get /vitastor/config/pgs --print-value-only | jq -s -e '(.[0].items["1"] | map((.osd_set | select(. > 0)) | length == 2) | length) == '$n) && \
@@ -82,8 +90,8 @@ try_change()
 
     # Check that no objects are lost !
     nobj=`$ETCDCTL get --prefix '/vitastor/pg/stats' --print-value-only | jq -s '[ .[].object_count ] | reduce .[] as $num (0; .+$num)'`
-    if [ "$nobj" -ne 1024 ]; then
-        format_error "Data lost after changing PG count to $n: 1024 objects expected, but got $nobj"
+    if [ "$nobj" -ne $NOBJ ]; then
+        format_error "Data lost after changing PG count to $n: $NOBJ objects expected, but got $nobj"
     fi
 }
 

tests/test_snapshot.sh

@@ -0,0 +1,75 @@
+#!/bin/bash -ex
+
+. `dirname $0`/common.sh
+
+dd if=/dev/zero of=./testdata/test_osd1.bin bs=1024 count=1 seek=$((1024*1024-1))
+dd if=/dev/zero of=./testdata/test_osd2.bin bs=1024 count=1 seek=$((1024*1024-1))
+dd if=/dev/zero of=./testdata/test_osd3.bin bs=1024 count=1 seek=$((1024*1024-1))
+
+build/src/vitastor-osd --osd_num 1 --bind_address 127.0.0.1 --etcd_address $ETCD_URL $(node mon/simple-offsets.js --format options --device ./testdata/test_osd1.bin 2>/dev/null) &>./testdata/osd1.log &
+OSD1_PID=$!
+build/src/vitastor-osd --osd_num 2 --bind_address 127.0.0.1 --etcd_address $ETCD_URL $(node mon/simple-offsets.js --format options --device ./testdata/test_osd2.bin 2>/dev/null) &>./testdata/osd2.log &
+OSD2_PID=$!
+build/src/vitastor-osd --osd_num 3 --bind_address 127.0.0.1 --etcd_address $ETCD_URL $(node mon/simple-offsets.js --format options --device ./testdata/test_osd3.bin 2>/dev/null) &>./testdata/osd3.log &
+OSD3_PID=$!
+
+cd mon
+npm install
+cd ..
+node mon/mon-main.js --etcd_url http://$ETCD_URL --etcd_prefix "/vitastor" &>./testdata/mon.log &
+MON_PID=$!
+
+$ETCDCTL put /vitastor/config/pools '{"1":{"name":"testpool","scheme":"xor","pg_size":3,"pg_minsize":2,"parity_chunks":1,"pg_count":1,"failure_domain":"osd"}}'
+
+sleep 2
+
+if ! ($ETCDCTL get /vitastor/config/pgs --print-value-only | jq -s -e '(. | length) != 0 and (.[0].items["1"]["1"].osd_set | sort) == ["1","2","3"]'); then
+    format_error "FAILED: 1 PG NOT CONFIGURED"
+fi
+
+if ! ($ETCDCTL get /vitastor/pg/state/1/1 --print-value-only | jq -s -e '(. | length) != 0 and .[0].state == ["active"]'); then
+    format_error "FAILED: 1 PG NOT UP"
+fi
+
+if ! cmp build/src/block-vitastor.so /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so; then
+    sudo rm -f /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so
+    sudo ln -s "$(realpath .)/build/src/block-vitastor.so" /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so
+fi
+
+# Test basic write and snapshot
+
+$ETCDCTL put /vitastor/config/inode/1/2 '{"name":"testimg","size":'$((32*1024*1024))'}'
+
+LD_PRELOAD=libasan.so.5 \
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -fsync=1 -rw=write \
+        -etcd=$ETCD_URL -pool=1 -inode=2 -size=32M -cluster_log_level=10
+
+$ETCDCTL put /vitastor/config/inode/1/2 '{"name":"testimg@0","size":'$((32*1024*1024))'}'
+$ETCDCTL put /vitastor/config/inode/1/3 '{"parent_id":2,"name":"testimg","size":'$((32*1024*1024))'}'
+
+LD_PRELOAD=libasan.so.5 \
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4k -direct=1 -iodepth=1 -fsync=32 -buffer_pattern=0xdeadface \
+        -rw=randwrite -etcd=$ETCD_URL -image=testimg -number_ios=1024
+
+LD_PRELOAD=libasan.so.5 \
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -rw=read -etcd=$ETCD_URL -pool=1 -inode=3 -size=32M
+
+qemu-img convert -S 4096 -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=3:size=$((32*1024*1024))" \
+    -O raw ./testdata/merged.bin
+
+qemu-img convert -S 4096 -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:image=testimg@0" \
+    -O raw ./testdata/layer0.bin
+
+$ETCDCTL put /vitastor/config/inode/1/3 '{"name":"testimg","size":'$((32*1024*1024))'}'
+
+qemu-img convert -S 4096 -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:image=testimg" \
+    -O raw ./testdata/layer1.bin
+
+node mon/merge.js ./testdata/layer0.bin ./testdata/layer1.bin ./testdata/check.bin
+
+cmp ./testdata/merged.bin ./testdata/check.bin
+
+format_green OK

tests/test_write.sh

@@ -34,40 +34,24 @@ fi
 #LD_PRELOAD=libasan.so.5 \
 #    fio -thread -name=test -ioengine=build/src/libfio_vitastor_sec.so -bs=4k -fsync=128 `$ETCDCTL get /vitastor/osd/state/1 --print-value-only | jq -r '"-host="+.addresses[0]+" -port="+(.port|tostring)'` -rw=write -size=32M
 
-# Test basic write and snapshot
-
-$ETCDCTL put /vitastor/config/inode/1/2 '{"name":"testimg","size":'$((32*1024*1024))'}'
+if ! cmp build/src/block-vitastor.so /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so; then
+    sudo rm -f /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so
+    sudo ln -s "$(realpath .)/build/src/block-vitastor.so" /usr/lib/x86_64-linux-gnu/qemu/block-vitastor.so
+fi
 
 LD_PRELOAD=libasan.so.5 \
-fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -fsync=1 -rw=write \
-    -etcd=$ETCD_URL -pool=1 -inode=2 -size=32M -cluster_log_level=10
-
-$ETCDCTL put /vitastor/config/inode/1/2 '{"name":"testimg@0","size":'$((32*1024*1024))'}'
-$ETCDCTL put /vitastor/config/inode/1/3 '{"parent_id":2,"name":"testimg","size":'$((32*1024*1024))'}'
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -fsync=1 -rw=write -etcd=$ETCD_URL -pool=1 -inode=1 -size=128M -cluster_log_level=10
 
 LD_PRELOAD=libasan.so.5 \
-fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4k -direct=1 -iodepth=1 -fsync=32 -buffer_pattern=0xdeadface \
-    -rw=randwrite -etcd=$ETCD_URL -image=testimg -number_ios=1024
-
-LD_PRELOAD=libasan.so.5 \
-fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -rw=read -etcd=$ETCD_URL -pool=1 -inode=3 -size=32M
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4k -direct=1 -iodepth=1 -fsync=32 -buffer_pattern=0xdeadface \
+        -rw=randwrite -etcd=$ETCD_URL -pool=1 -inode=1 -size=128M -number_ios=1024
 
 qemu-img convert -S 4096 -p \
-  -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=3:size=$((32*1024*1024))" \
-  -O raw ./testdata/merged.bin
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=1:size=$((128*1024*1024))" \
+    -O raw ./testdata/read.bin
 
 qemu-img convert -S 4096 -p \
-  -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:image=testimg@0" \
-  -O raw ./testdata/layer0.bin
-
-$ETCDCTL put /vitastor/config/inode/1/3 '{"name":"testimg","size":'$((32*1024*1024))'}'
-
-qemu-img convert -S 4096 -p \
-  -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:image=testimg" \
-  -O raw ./testdata/layer1.bin
-
-node mon/merge.js ./testdata/layer0.bin ./testdata/layer1.bin ./testdata/check.bin
-
-cmp ./testdata/merged.bin ./testdata/check.bin
+    -f raw ./testdata/read.bin \
+    -O raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=1:size=$((128*1024*1024))"
 
 format_green OK