Release 0.5.13

Another followup to 0.5.11

debian/changelog

@@ -1,4 +1,4 @@
-vitastor (0.5.10-1) unstable; urgency=medium
+vitastor (0.5.13-1) 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-0.5.13; \
+    ln -s /root/packages/qemu-$REL/qemu-*/ vitastor-0.5.13/qemu; \
+    ln -s /root/fio-build/fio-*/ vitastor-0.5.13/fio; \
+    cd vitastor-0.5.13; \
     FIO=$(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_0.5.13.orig.tar.xz vitastor-0.5.13; \
+    cd vitastor-0.5.13; \
     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

@@ -92,7 +92,8 @@ 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,
@@ -182,7 +183,7 @@ 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")[],
                 }
@@ -541,7 +542,7 @@ 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];
@@ -692,6 +693,11 @@ 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++)
         {
@@ -842,7 +848,7 @@ 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,
@@ -901,7 +907,14 @@ 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)
                 {
@@ -928,24 +941,23 @@ 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_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_change({
+                            prev_pgs,
+                            ...optimize_cfg,
                         });
                     }
+                }
+                else
+                {
+                    optimize_result = await LPOptimizer.optimize_initial(optimize_cfg);
+                }
                 if (old_pg_count != optimize_result.int_pgs.length)
                 {
                     console.log(
@@ -1072,7 +1084,7 @@ 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 };

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-0.5.13/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-0.5.13$(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-0.5.13.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:        0.5.13
 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-0.5.13.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-0.5.13.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:        0.5.13
 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-0.5.13.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -14,7 +14,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 endif()
 
 add_definitions(-DVERSION="0.6-dev")
-add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith)
+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_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}
@@ -175,6 +177,15 @@ 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)
 #target_link_libraries(test_blockstore blockstore)

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

@@ -58,7 +58,7 @@ uint64_t blockstore_t::get_free_block_count()
     return impl->get_free_block_count();
 }
 
-uint32_t blockstore_t::get_disk_alignment()
+uint32_t blockstore_t::get_bitmap_granularity()
 {
-    return impl->get_disk_alignment();
+    return impl->get_bitmap_granularity();
 }

src/blockstore.h

@@ -183,5 +183,5 @@ public:
     uint64_t get_block_count();
     uint64_t get_free_block_count();
 
-    uint32_t get_disk_alignment();
+    uint32_t get_bitmap_granularity();
 };

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;
@@ -19,8 +20,8 @@ journal_flusher_t::journal_flusher_t(int flusher_count, blockstore_impl_t *bs)
     journal_trim_counter = 0;
     trim_wanted = 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)
@@ -482,6 +484,14 @@ 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);
         }
         else
@@ -646,7 +656,7 @@ 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);
@@ -775,7 +785,10 @@ 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);
     }
@@ -783,6 +796,11 @@ 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;
     }
@@ -804,7 +822,7 @@ 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

@@ -80,7 +80,7 @@ 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;
@@ -99,7 +99,7 @@ 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

@@ -10,9 +10,9 @@ blockstore_impl_t::blockstore_impl_t(blockstore_config_t & config, ring_loop_t *
     ring_consumer.loop = [this]() { loop(); };
     ringloop->register_consumer(&ring_consumer);
     initialized = 0;
-    zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, block_size);
     data_fd = meta_fd = journal.fd = -1;
     parse_config(config);
+    zero_object = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, block_size);
     try
     {
         open_data();
@@ -31,7 +31,7 @@ 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()

src/blockstore_impl.h

@@ -197,8 +197,8 @@ 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;
     /******* END OF OPTIONS *******/
@@ -210,6 +210,7 @@ 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;
 
@@ -283,7 +284,7 @@ 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);
 
@@ -326,5 +327,5 @@ public:
     inline uint32_t get_block_size() { return block_size; }
     inline uint64_t get_block_count() { return block_count; }
     inline uint64_t get_free_block_count() { return data_alloc->get_free_count(); }
-    inline uint32_t get_disk_alignment() { return disk_alignment; }
+    inline uint32_t get_bitmap_granularity() { return disk_alignment; }
 };

src/blockstore_init.cpp

@@ -111,7 +111,10 @@ 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);
                 }
@@ -399,6 +402,18 @@ 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(
@@ -549,7 +564,7 @@ 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);
                     }
                 }
             }
@@ -579,32 +594,10 @@ 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);
@@ -616,18 +609,33 @@ int blockstore_init_journal::handle_journal_part(void *buf, uint64_t done_pos, u
                         .oid = je->big_write.oid,
                         .version = je->big_write.version,
                     };
-                    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,
                         .offset = je->big_write.offset,
                         .len = je->big_write.len,
                         .journal_sector = proc_pos,
-                    });
+                    }).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->journal.used_sectors[proc_pos]++;
 #ifdef BLOCKSTORE_DEBUG
                     printf(
@@ -639,7 +647,7 @@ 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);
                     }
                 }
             }
@@ -653,7 +661,7 @@ 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)
             {
@@ -672,9 +680,26 @@ 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 = {
@@ -692,8 +717,9 @@ 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;
@@ -704,3 +730,30 @@ 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;
+    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;
+    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

@@ -36,6 +36,7 @@ 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 +49,7 @@ 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_open.cpp

@@ -69,7 +69,10 @@ 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);
     // Validate
     if (!block_size)
@@ -80,9 +83,13 @@ 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)
+    {
+        min_flusher_count = 1;
     }
     if (!max_write_iodepth)
     {

src/blockstore_rollback.cpp

@@ -163,10 +163,7 @@ void blockstore_impl_t::mark_rolled_back(const obj_ver_id & ov)
         auto rm_start = it;
         auto rm_end = it;
         it--;
-        while (it->first.oid == ov.oid &&
-            it->first.version > ov.version &&
-            !IS_IN_FLIGHT(it->second.state) &&
-            !IS_STABLE(it->second.state))
+        while (1)
         {
             if (it->first.oid != ov.oid)
                 break;
@@ -176,7 +173,7 @@ void blockstore_impl_t::mark_rolled_back(const obj_ver_id & ov)
                     max_unstable = it->first.version;
                 break;
             }
-            else if (IS_STABLE(it->second.state))
+            else if (IS_IN_FLIGHT(it->second.state) || IS_STABLE(it->second.state))
                 break;
             // Remove entry
             rm_start = it;
@@ -187,7 +184,6 @@ void blockstore_impl_t::mark_rolled_back(const obj_ver_id & ov)
         if (rm_start != rm_end)
         {
             erase_dirty(rm_start, rm_end, UINT64_MAX);
-        }
             auto unstab_it = unstable_writes.find(ov.oid);
             if (unstab_it != unstable_writes.end())
             {
@@ -197,6 +193,7 @@ void blockstore_impl_t::mark_rolled_back(const obj_ver_id & ov)
                     unstab_it->second = max_unstable;
             }
         }
+    }
 }
 
 void blockstore_impl_t::handle_rollback_event(ring_data_t *data, blockstore_op_t *op)
@@ -251,10 +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
@@ -176,22 +179,39 @@ 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;
             }
-            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;
                     }
-            if (dirty_it == dirty_db.begin())
+                }
+                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 (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;

src/blockstore_write.cpp

@@ -201,7 +201,7 @@ 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), JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
         }
@@ -224,7 +224,10 @@ 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);
@@ -250,11 +253,8 @@ 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
@@ -267,7 +267,7 @@ 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)
+        if (unsynced_big_write_count && !space_check.check_available(op, unsynced_big_write_count, sizeof(journal_entry_big_write), 0)
             || !space_check.check_available(op, 1, sizeof(journal_entry_small_write), op->len + JOURNAL_STABILIZE_RESERVATION))
         {
             return 0;
@@ -359,14 +359,6 @@ 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;
@@ -431,7 +423,7 @@ resume_2:
 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)
@@ -445,11 +437,31 @@ resume_4:
         | (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
+        // Deletions and 'instant' operations are treated as immediately stable
         mark_stable(dirty_it->first);
     }
-    if (immediate_commit == IMMEDIATE_ALL)
+    if (!imm)
     {
+        if ((dirty_it->second.state & BS_ST_TYPE_MASK) == BS_ST_BIG_WRITE)
+        {
+            // 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)
         {
@@ -583,14 +595,6 @@ 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.cpp

@@ -2,8 +2,17 @@
 // License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
 
 #include <stdexcept>
+#include <assert.h>
 #include "cluster_client.h"
 
+#define PART_SENT 1
+#define PART_DONE 2
+#define PART_ERROR 4
+#define CACHE_DIRTY 1
+#define CACHE_FLUSHING 2
+#define CACHE_REPEATING 3
+#define OP_FLUSH_BUFFER 2
+
 cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config)
 {
     this->ringloop = ringloop;
@@ -20,39 +29,17 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
             // peer_osd just connected
             continue_ops();
         }
-        else if (unsynced_writes.size())
+        else if (dirty_buffers.size())
         {
             // peer_osd just dropped connection
-            for (auto op: syncing_writes)
+            // determine WHICH dirty_buffers are now obsolete and repeat them
+            for (auto & wr: dirty_buffers)
             {
-                for (auto & part: op->parts)
+                if (affects_osd(wr.first.inode, wr.first.stripe, wr.second.len, peer_osd) &&
+                    wr.second.state != CACHE_REPEATING)
                 {
-                    if (part.osd_num == peer_osd && part.done)
-                    {
-                        // repeat this operation
-                        part.osd_num = 0;
-                        part.done = false;
-                        assert(!part.sent);
-                        op->done_count--;
-                    }
-                }
-            }
-            for (auto op: unsynced_writes)
-            {
-                for (auto & part: op->parts)
-                {
-                    if (part.osd_num == peer_osd && part.done)
-                    {
-                        // repeat this operation
-                        part.osd_num = 0;
-                        part.done = false;
-                        assert(!part.sent);
-                        op->done_count--;
-                    }
-                }
-                if (op->done_count < op->parts.size())
-                {
-                    cur_ops.insert(op);
+                    // FIXME: Flush in larger parts
+                    flush_buffer(wr.first, &wr.second);
                 }
             }
             continue_ops();
@@ -90,37 +77,87 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
 
 cluster_client_t::~cluster_client_t()
 {
+    for (auto bp: dirty_buffers)
+    {
+        free(bp.second.buf);
+    }
+    dirty_buffers.clear();
     if (ringloop)
     {
         ringloop->unregister_consumer(&consumer);
     }
 }
 
-void cluster_client_t::stop()
-{
-    while (msgr.clients.size() > 0)
-    {
-        msgr.stop_client(msgr.clients.begin()->first);
-    }
-}
-
 void cluster_client_t::continue_ops(bool up_retry)
 {
-    for (auto op_it = cur_ops.begin(); op_it != cur_ops.end(); )
+    if (!pgs_loaded)
     {
-        if ((*op_it)->up_wait)
+        // We're offline
+        return;
+    }
+    if (continuing_ops)
     {
-            if (up_retry)
+        // Attempt to reenter the function
+        continuing_ops = 2;
+        return;
+    }
+restart:
+    continuing_ops = 1;
+    op_queue_pos = 0;
+    bool has_flushes = false, has_writes = false;
+    while (op_queue_pos < op_queue.size())
     {
-                (*op_it)->up_wait = false;
-                continue_rw(*op_it++);
+        auto op = op_queue[op_queue_pos];
+        bool rm = false, is_flush = op->flags & OP_FLUSH_BUFFER;
+        auto opcode = op->opcode;
+        if (!op->up_wait || up_retry)
+        {
+            op->up_wait = false;
+            if (opcode == OSD_OP_READ || opcode == OSD_OP_WRITE)
+            {
+                if (is_flush || !has_flushes)
+                {
+                    // Regular writes can't proceed before buffer flushes
+                    rm = continue_rw(op);
+                }
+            }
+            else if (opcode == OSD_OP_SYNC)
+            {
+                if (!has_writes)
+                {
+                    // SYNC can't proceed before previous writes
+                    rm = continue_sync(op);
+                }
+            }
+        }
+        if (opcode == OSD_OP_WRITE)
+        {
+            has_writes = has_writes || !rm;
+            if (is_flush)
+            {
+                has_flushes = has_writes || !rm;
+            }
+        }
+        else if (opcode == OSD_OP_SYNC)
+        {
+            // Postpone writes until previous SYNC completes
+            // ...so dirty_writes can't contain anything newer than SYNC
+            has_flushes = has_writes || !rm;
+        }
+        if (rm)
+        {
+            op_queue.erase(op_queue.begin()+op_queue_pos, op_queue.begin()+op_queue_pos+1);
         }
         else
-                op_it++;
+        {
+            op_queue_pos++;
         }
-        else
-            continue_rw(*op_it++);
+        if (continuing_ops == 2)
+        {
+            goto restart;
         }
+    }
+    continuing_ops = 0;
 }
 
 static uint32_t is_power_of_two(uint64_t value)
@@ -141,16 +178,11 @@ static uint32_t is_power_of_two(uint64_t value)
 void cluster_client_t::on_load_config_hook(json11::Json::object & config)
 {
     bs_block_size = config["block_size"].uint64_value();
-    bs_disk_alignment = config["disk_alignment"].uint64_value();
     bs_bitmap_granularity = config["bitmap_granularity"].uint64_value();
     if (!bs_block_size)
     {
         bs_block_size = DEFAULT_BLOCK_SIZE;
     }
-    if (!bs_disk_alignment)
-    {
-        bs_disk_alignment = DEFAULT_DISK_ALIGNMENT;
-    }
     if (!bs_bitmap_granularity)
     {
         bs_bitmap_granularity = DEFAULT_BITMAP_GRANULARITY;
@@ -165,13 +197,26 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & config)
         // Cluster-wide immediate_commit mode
         immediate_commit = true;
     }
+    if (config.find("client_max_dirty_bytes") != config.end())
+    {
+        client_max_dirty_bytes = config["client_max_dirty_bytes"].uint64_value();
+    }
     else if (config.find("client_dirty_limit") != config.end())
     {
-        client_dirty_limit = config["client_dirty_limit"].uint64_value();
+        // Old name
+        client_max_dirty_bytes = config["client_dirty_limit"].uint64_value();
     }
-    if (!client_dirty_limit)
+    if (config.find("client_max_dirty_ops") != config.end())
     {
-        client_dirty_limit = DEFAULT_CLIENT_DIRTY_LIMIT;
+        client_max_dirty_ops = config["client_max_dirty_ops"].uint64_value();
+    }
+    if (!client_max_dirty_bytes)
+    {
+        client_max_dirty_bytes = DEFAULT_CLIENT_MAX_DIRTY_BYTES;
+    }
+    if (!client_max_dirty_ops)
+    {
+        client_max_dirty_ops = DEFAULT_CLIENT_MAX_DIRTY_OPS;
     }
     up_wait_retry_interval = config["up_wait_retry_interval"].uint64_value();
     if (!up_wait_retry_interval)
@@ -215,23 +260,10 @@ void cluster_client_t::on_change_hook(json11::Json::object & changes)
         {
             // At this point, all pool operations should have been suspended
             // And now they have to be resliced!
-            for (auto op: cur_ops)
+            for (auto op: op_queue)
             {
-                if (INODE_POOL(op->inode) == pool_item.first)
-                {
-                    op->needs_reslice = true;
-                }
-            }
-            for (auto op: unsynced_writes)
-            {
-                if (INODE_POOL(op->inode) == pool_item.first)
-                {
-                    op->needs_reslice = true;
-                }
-            }
-            for (auto op: syncing_writes)
-            {
-                if (INODE_POOL(op->inode) == pool_item.first)
+                if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_READ) &&
+                    INODE_POOL(op->inode) == pool_item.first)
                 {
                     op->needs_reslice = true;
                 }
@@ -250,6 +282,11 @@ void cluster_client_t::on_change_osd_state_hook(uint64_t peer_osd)
     }
 }
 
+bool cluster_client_t::is_ready()
+{
+    return pgs_loaded;
+}
+
 void cluster_client_t::on_ready(std::function<void(void)> fn)
 {
     if (pgs_loaded)
@@ -265,21 +302,15 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
 /**
  * How writes are synced when immediate_commit is false
  *
- * 1) accept up to <client_dirty_limit> write operations for execution,
- *    queue all subsequent writes into <next_writes>
- * 2) accept exactly one SYNC, queue all subsequent SYNCs into <next_writes>, too
- * 3) "continue" all accepted writes
- *
  * "Continue" WRITE:
- * 1) if the operation is not a copy yet - copy it (required for replay)
- * 2) if the operation is not sliced yet - slice it
- * 3) if the operation doesn't require reslice - try to connect & send all remaining parts
- * 4) if any of them fail due to disconnected peers or PGs not up, repeat after reconnecting or small timeout
- * 5) if any of them fail due to other errors, fail the operation and forget it from the current "unsynced batch"
- * 6) if PG count changes before all parts are done, wait for all in-progress parts to finish,
+ * 1) if the operation is not sliced yet - slice it
+ * 2) if the operation doesn't require reslice - try to connect & send all remaining parts
+ * 3) if any of them fail due to disconnected peers or PGs not up, repeat after reconnecting or small timeout
+ * 4) if any of them fail due to other errors, fail the operation and forget it from the current "unsynced batch"
+ * 5) if PG count changes before all parts are done, wait for all in-progress parts to finish,
  *    throw all results away, reslice and resubmit op
- * 7) when all parts are done, try to "continue" the current SYNC
- * 8) if the operation succeeds, but then some OSDs drop their connections, repeat
+ * 6) when all parts are done, try to "continue" the current SYNC
+ * 7) if the operation succeeds, but then some OSDs drop their connections, repeat
  *    parts from the current "unsynced batch" previously sent to those OSDs in any order
  *
  * "Continue" current SYNC:
@@ -289,181 +320,277 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
  * 4) if any of them fail due to disconnected peers, repeat SYNC after repeating all writes
  * 5) if any of them fail due to other errors, fail the SYNC operation
  */
-
 void cluster_client_t::execute(cluster_op_t *op)
 {
-    if (!pgs_loaded)
-    {
-        // We're offline
-        offline_ops.push_back(op);
-        return;
-    }
-    op->retval = 0;
-    if (op->opcode != OSD_OP_SYNC && op->opcode != OSD_OP_READ && op->opcode != OSD_OP_WRITE ||
-        (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_WRITE) && (!op->inode || !op->len ||
-        op->offset % bs_disk_alignment || op->len % bs_disk_alignment))
+    if (op->opcode != OSD_OP_SYNC && op->opcode != OSD_OP_READ && op->opcode != OSD_OP_WRITE)
     {
         op->retval = -EINVAL;
         std::function<void(cluster_op_t*)>(op->callback)(op);
         return;
     }
-    if (op->opcode == OSD_OP_SYNC)
-    {
-        execute_sync(op);
-        return;
-    }
+    op->retval = 0;
     if (op->opcode == OSD_OP_WRITE && !immediate_commit)
     {
-        if (next_writes.size() > 0)
-        {
-            assert(cur_sync);
-            next_writes.push_back(op);
-            return;
-        }
-        if (queued_bytes >= client_dirty_limit)
+        if (dirty_bytes >= client_max_dirty_bytes || dirty_ops >= client_max_dirty_ops)
         {
             // Push an extra SYNC operation to flush previous writes
-            next_writes.push_back(op);
             cluster_op_t *sync_op = new cluster_op_t;
-            sync_op->is_internal = true;
             sync_op->opcode = OSD_OP_SYNC;
-            sync_op->callback = [](cluster_op_t* sync_op) {};
-            execute_sync(sync_op);
-            return;
+            sync_op->callback = [](cluster_op_t* sync_op)
+            {
+                delete sync_op;
+            };
+            op_queue.push_back(sync_op);
+            dirty_bytes = 0;
+            dirty_ops = 0;
         }
-        queued_bytes += op->len;
+        dirty_bytes += op->len;
+        dirty_ops++;
     }
-    cur_ops.insert(op);
-    continue_rw(op);
+    else if (op->opcode == OSD_OP_SYNC)
+    {
+        dirty_bytes = 0;
+        dirty_ops = 0;
+    }
+    op_queue.push_back(op);
+    continue_ops();
 }
 
-void cluster_client_t::continue_rw(cluster_op_t *op)
+void cluster_client_t::copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers)
 {
+    // Save operation for replay when one of PGs goes out of sync
+    // (primary OSD drops our connection in this case)
+    auto dirty_it = dirty_buffers.lower_bound((object_id){
+        .inode = op->inode,
+        .stripe = op->offset,
+    });
+    while (dirty_it != dirty_buffers.begin())
+    {
+        dirty_it--;
+        if (dirty_it->first.inode != op->inode ||
+            (dirty_it->first.stripe + dirty_it->second.len) <= op->offset)
+        {
+            dirty_it++;
+            break;
+        }
+    }
+    uint64_t pos = op->offset, len = op->len, iov_idx = 0, iov_pos = 0;
+    while (len > 0)
+    {
+        uint64_t new_len = 0;
+        if (dirty_it == dirty_buffers.end())
+        {
+            new_len = len;
+        }
+        else if (dirty_it->first.inode != op->inode || dirty_it->first.stripe > pos)
+        {
+            new_len = dirty_it->first.stripe - pos;
+            if (new_len > len)
+            {
+                new_len = len;
+            }
+        }
+        if (new_len > 0)
+        {
+            dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+                .inode = op->inode,
+                .stripe = pos,
+            }, (cluster_buffer_t){
+                .buf = malloc_or_die(new_len),
+                .len = new_len,
+            });
+        }
+        // FIXME: Split big buffers into smaller ones on overwrites. But this will require refcounting
+        dirty_it->second.state = CACHE_DIRTY;
+        uint64_t cur_len = (dirty_it->first.stripe + dirty_it->second.len - pos);
+        if (cur_len > len)
+        {
+            cur_len = len;
+        }
+        while (cur_len > 0 && iov_idx < op->iov.count)
+        {
+            unsigned iov_len = (op->iov.buf[iov_idx].iov_len - iov_pos);
+            if (iov_len <= cur_len)
+            {
+                memcpy(dirty_it->second.buf + pos - dirty_it->first.stripe,
+                    op->iov.buf[iov_idx].iov_base + iov_pos, iov_len);
+                pos += iov_len;
+                len -= iov_len;
+                cur_len -= iov_len;
+                iov_pos = 0;
+                iov_idx++;
+            }
+            else
+            {
+                memcpy(dirty_it->second.buf + pos - dirty_it->first.stripe,
+                    op->iov.buf[iov_idx].iov_base + iov_pos, cur_len);
+                pos += cur_len;
+                len -= cur_len;
+                iov_pos += cur_len;
+                cur_len = 0;
+            }
+        }
+        dirty_it++;
+    }
+}
+
+void cluster_client_t::flush_buffer(const object_id & oid, cluster_buffer_t *wr)
+{
+    wr->state = CACHE_REPEATING;
+    cluster_op_t *op = new cluster_op_t;
+    op->flags = OP_FLUSH_BUFFER;
+    op->opcode = OSD_OP_WRITE;
+    op->inode = oid.inode;
+    op->offset = oid.stripe;
+    op->len = wr->len;
+    op->iov.push_back(wr->buf, wr->len);
+    op->callback = [wr](cluster_op_t* op)
+    {
+        if (wr->state == CACHE_REPEATING)
+        {
+            wr->state = CACHE_DIRTY;
+        }
+        delete op;
+    };
+    op_queue.insert(op_queue.begin(), op);
+    if (continuing_ops)
+    {
+        continuing_ops = 2;
+        op_queue_pos++;
+    }
+}
+
+int cluster_client_t::continue_rw(cluster_op_t *op)
+{
+    if (op->state == 0)
+        goto resume_0;
+    else if (op->state == 1)
+        goto resume_1;
+    else if (op->state == 2)
+        goto resume_2;
+    else if (op->state == 3)
+        goto resume_3;
+resume_0:
+    if (!op->len || op->offset % bs_bitmap_granularity || op->len % bs_bitmap_granularity)
+    {
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return 1;
+    }
+    {
         pool_id_t pool_id = INODE_POOL(op->inode);
         if (!pool_id)
         {
             op->retval = -EINVAL;
             std::function<void(cluster_op_t*)>(op->callback)(op);
-        return;
+            return 1;
         }
         if (st_cli.pool_config.find(pool_id) == st_cli.pool_config.end() ||
             st_cli.pool_config[pool_id].real_pg_count == 0)
         {
             // Postpone operations to unknown pools
-        return;
+            return 0;
         }
-    if (op->opcode == OSD_OP_WRITE && !immediate_commit && !op->is_internal)
-    {
-        // Save operation for replay when PG goes out of sync
-        // (primary OSD drops our connection in this case)
-        cluster_op_t *op_copy = new cluster_op_t();
-        op_copy->is_internal = true;
-        op_copy->orig_op = op;
-        op_copy->opcode = op->opcode;
-        op_copy->inode = op->inode;
-        op_copy->offset = op->offset;
-        op_copy->len = op->len;
-        op_copy->buf = malloc_or_die(op->len);
-        op_copy->iov.push_back(op_copy->buf, op->len);
-        op_copy->callback = [](cluster_op_t* op_copy)
-        {
-            if (op_copy->orig_op)
-            {
-                // Acknowledge write and forget the original pointer
-                op_copy->orig_op->retval = op_copy->retval;
-                std::function<void(cluster_op_t*)>(op_copy->orig_op->callback)(op_copy->orig_op);
-                op_copy->orig_op = NULL;
     }
-        };
-        void *cur_buf = op_copy->buf;
-        for (int i = 0; i < op->iov.count; i++)
+    if (op->opcode == OSD_OP_WRITE)
     {
-            memcpy(cur_buf, op->iov.buf[i].iov_base, op->iov.buf[i].iov_len);
-            cur_buf += op->iov.buf[i].iov_len;
-        }
-        unsynced_writes.push_back(op_copy);
-        cur_ops.erase(op);
-        cur_ops.insert(op_copy);
-        op = op_copy;
-    }
-    if (!op->parts.size())
+        if (!immediate_commit && !(op->flags & OP_FLUSH_BUFFER))
         {
+            copy_write(op, dirty_buffers);
+        }
+    }
+resume_1:
     // Slice the operation into parts
     slice_rw(op);
-    }
-    if (!op->needs_reslice)
-    {
+    op->needs_reslice = false;
+resume_2:
     // Send unsent parts, if they're not subject to change
-        for (auto & op_part: op->parts)
+    op->state = 3;
+    if (op->needs_reslice)
     {
-            if (!op_part.sent && !op_part.done)
+        for (int i = 0; i < op->parts.size(); i++)
         {
-                try_send(op, &op_part);
+            if (!(op->parts[i].flags & PART_SENT) && op->retval)
+            {
+                op->retval = -EPIPE;
+            }
+        }
+        goto resume_3;
+    }
+    for (int i = 0; i < op->parts.size(); i++)
+    {
+        if (!(op->parts[i].flags & PART_SENT))
+        {
+            if (!try_send(op, i))
+            {
+                // We'll need to retry again
+                op->up_wait = true;
+                if (!retry_timeout_id)
+                {
+                    retry_timeout_id = tfd->set_timer(up_wait_retry_interval, false, [this](int)
+                    {
+                        retry_timeout_id = 0;
+                        continue_ops(true);
+                    });
+                }
+                op->state = 2;
             }
         }
     }
-    if (!op->sent_count)
+    if (op->state == 2)
     {
+        return 0;
+    }
+resume_3:
+    if (op->inflight_count > 0)
+    {
+        op->state = 3;
+        return 0;
+    }
     if (op->done_count >= op->parts.size())
     {
         // Finished successfully
         // Even if the PG count has changed in meanwhile we treat it as success
         // because if some operations were invalid for the new PG count we'd get errors
-            cur_ops.erase(op);
         op->retval = op->len;
         std::function<void(cluster_op_t*)>(op->callback)(op);
-            continue_sync();
-            return;
+        return 1;
     }
     else if (op->retval != 0 && op->retval != -EPIPE)
     {
         // Fatal error (not -EPIPE)
-            cur_ops.erase(op);
-            if (!immediate_commit && op->opcode == OSD_OP_WRITE)
-            {
-                for (int i = 0; i < unsynced_writes.size(); i++)
-                {
-                    if (unsynced_writes[i] == op)
-                    {
-                        unsynced_writes.erase(unsynced_writes.begin()+i, unsynced_writes.begin()+i+1);
-                        break;
-                    }
-                }
-            }
-            bool del = op->is_internal;
         std::function<void(cluster_op_t*)>(op->callback)(op);
-            if (del)
-            {
-                if (op->buf)
-                    free(op->buf);
-                delete op;
-            }
-            continue_sync();
-            return;
+        return 1;
     }
     else
     {
-            // -EPIPE or no error - clear the error
+        // -EPIPE - clear the error and retry
         op->retval = 0;
         if (op->needs_reslice)
         {
             op->parts.clear();
             op->done_count = 0;
-                op->needs_reslice = false;
-                continue_rw(op);
-            }
+            goto resume_1;
+        }
+        else
+        {
+            for (int i = 0; i < op->parts.size(); i++)
+            {
+                op->parts[i].flags = 0;
+            }
+            goto resume_2;
         }
     }
+    return 0;
 }
 
 void cluster_client_t::slice_rw(cluster_op_t *op)
 {
     // Slice the request into individual object stripe requests
     // Primary OSDs still operate individual stripes, but their size is multiplied by PG minsize in case of EC
-    auto & pool_cfg = st_cli.pool_config[INODE_POOL(op->inode)];
-    uint64_t pg_block_size = bs_block_size * (
-        pool_cfg.scheme == POOL_SCHEME_REPLICATED ? 1 : pool_cfg.pg_size-pool_cfg.parity_chunks
-    );
+    auto & pool_cfg = st_cli.pool_config.at(INODE_POOL(op->inode));
+    uint32_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;
     uint64_t first_stripe = (op->offset / pg_block_size) * pg_block_size;
     uint64_t last_stripe = ((op->offset + op->len + pg_block_size - 1) / pg_block_size - 1) * pg_block_size;
     op->retval = 0;
@@ -482,8 +609,7 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             .offset = begin,
             .len = (uint32_t)(end - begin),
             .pg_num = pg_num,
-            .sent = false,
-            .done = false,
+            .flags = 0,
         };
         int left = end-begin;
         while (left > 0 && iov_idx < op->iov.count)
@@ -507,8 +633,28 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
     }
 }
 
-bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
+bool cluster_client_t::affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd)
 {
+    auto & pool_cfg = st_cli.pool_config.at(INODE_POOL(inode));
+    uint32_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;
+    uint64_t first_stripe = (offset / pg_block_size) * pg_block_size;
+    uint64_t last_stripe = ((offset + len + pg_block_size - 1) / pg_block_size - 1) * pg_block_size;
+    for (uint64_t stripe = first_stripe; stripe <= last_stripe; stripe += pg_block_size)
+    {
+        pg_num_t pg_num = (stripe/pool_cfg.pg_stripe_size) % pool_cfg.real_pg_count + 1; // like map_to_pg()
+        auto pg_it = pool_cfg.pg_config.find(pg_num);
+        if (pg_it != pool_cfg.pg_config.end() && pg_it->second.cur_primary == osd)
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+bool cluster_client_t::try_send(cluster_op_t *op, int i)
+{
+    auto part = &op->parts[i];
     auto & pool_cfg = st_cli.pool_config[INODE_POOL(op->inode)];
     auto pg_it = pool_cfg.pg_config.find(part->pg_num);
     if (pg_it != pool_cfg.pg_config.end() &&
@@ -520,8 +666,8 @@ bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
         {
             int peer_fd = peer_it->second;
             part->osd_num = primary_osd;
-            part->sent = true;
-            op->sent_count++;
+            part->flags |= PART_SENT;
+            op->inflight_count++;
             part->op = (osd_op_t){
                 .op_type = OSD_OP_OUT,
                 .peer_fd = peer_fd,
@@ -552,137 +698,99 @@ bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
     return false;
 }
 
-void cluster_client_t::execute_sync(cluster_op_t *op)
+int cluster_client_t::continue_sync(cluster_op_t *op)
 {
-    if (immediate_commit)
+    if (op->state == 1)
+        goto resume_1;
+    if (immediate_commit || !dirty_osds.size())
     {
-        // Syncs are not required in the immediate_commit mode
+        // Sync is not required in the immediate_commit mode or if there are no dirty_osds
         op->retval = 0;
         std::function<void(cluster_op_t*)>(op->callback)(op);
-    }
-    else if (cur_sync != NULL)
-    {
-        next_writes.push_back(op);
-    }
-    else
-    {
-        cur_sync = op;
-        continue_sync();
-    }
-}
-
-void cluster_client_t::continue_sync()
-{
-    if (!cur_sync || cur_sync->parts.size() > 0)
-    {
-        // Already submitted
-        return;
-    }
-    cur_sync->retval = 0;
-    std::set<osd_num_t> sync_osds;
-    for (auto prev_op: unsynced_writes)
-    {
-        if (prev_op->done_count < prev_op->parts.size())
-        {
-            // Writes not finished yet
-            return;
-        }
-        for (auto & part: prev_op->parts)
-        {
-            if (part.osd_num)
-            {
-                sync_osds.insert(part.osd_num);
-            }
-        }
-    }
-    if (!sync_osds.size())
-    {
-        // No dirty writes
-        finish_sync();
-        return;
+        return 1;
     }
     // Check that all OSD connections are still alive
-    for (auto sync_osd: sync_osds)
+    for (auto sync_osd: dirty_osds)
     {
         auto peer_it = msgr.osd_peer_fds.find(sync_osd);
         if (peer_it == msgr.osd_peer_fds.end())
         {
-            // SYNC is pointless to send to a non connected OSD
-            return;
+            return 0;
         }
     }
-    syncing_writes.swap(unsynced_writes);
     // Post sync to affected OSDs
-    cur_sync->parts.resize(sync_osds.size());
-    int i = 0;
-    for (auto sync_osd: sync_osds)
+    for (auto & prev_op: dirty_buffers)
     {
-        cur_sync->parts[i] = {
-            .parent = cur_sync,
+        if (prev_op.second.state == CACHE_DIRTY)
+        {
+            prev_op.second.state = CACHE_FLUSHING;
+        }
+    }
+    op->parts.resize(dirty_osds.size());
+    op->retval = 0;
+    {
+        int i = 0;
+        for (auto sync_osd: dirty_osds)
+        {
+            op->parts[i] = {
+                .parent = op,
                 .osd_num = sync_osd,
-            .sent = false,
-            .done = false,
+                .flags = 0,
             };
-        send_sync(cur_sync, &cur_sync->parts[i]);
+            send_sync(op, &op->parts[i]);
             i++;
         }
-}
-
-void cluster_client_t::finish_sync()
-{
-    int retval = cur_sync->retval;
-    if (retval != 0)
+    }
+    dirty_osds.clear();
+resume_1:
+    if (op->inflight_count > 0)
     {
-        for (auto op: syncing_writes)
+        op->state = 1;
+        return 0;
+    }
+    if (op->retval != 0)
     {
-            if (op->done_count < op->parts.size())
+        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); uw_it++)
         {
-                cur_ops.insert(op);
+            if (uw_it->second.state == CACHE_FLUSHING)
+            {
+                uw_it->second.state = CACHE_DIRTY;
             }
         }
-        unsynced_writes.insert(unsynced_writes.begin(), syncing_writes.begin(), syncing_writes.end());
-        syncing_writes.clear();
-    }
-    if (retval == -EPIPE)
+        if (op->retval == -EPIPE)
         {
             // Retry later
-        cur_sync->parts.clear();
-        cur_sync->retval = 0;
-        cur_sync->sent_count = 0;
-        cur_sync->done_count = 0;
-        return;
+            op->parts.clear();
+            op->retval = 0;
+            op->inflight_count = 0;
+            op->done_count = 0;
+            op->state = 0;
+            return 0;
         }
-    std::function<void(cluster_op_t*)>(cur_sync->callback)(cur_sync);
-    if (!retval)
+    }
+    else
     {
-        for (auto op: syncing_writes)
+        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); )
         {
-            assert(op->sent_count == 0);
-            if (op->is_internal)
+            if (uw_it->second.state == CACHE_FLUSHING)
             {
-                if (op->buf)
-                    free(op->buf);
-                delete op;
+                free(uw_it->second.buf);
+                dirty_buffers.erase(uw_it++);
+            }
+            else
+                uw_it++;
         }
     }
-        syncing_writes.clear();
-    }
-    cur_sync = NULL;
-    queued_bytes = 0;
-    std::vector<cluster_op_t*> next_wr_copy;
-    next_wr_copy.swap(next_writes);
-    for (auto next_op: next_wr_copy)
-    {
-        execute(next_op);
-    }
+    std::function<void(cluster_op_t*)>(op->callback)(op);
+    return 1;
 }
 
 void cluster_client_t::send_sync(cluster_op_t *op, cluster_op_part_t *part)
 {
     auto peer_it = msgr.osd_peer_fds.find(part->osd_num);
     assert(peer_it != msgr.osd_peer_fds.end());
-    part->sent = true;
-    op->sent_count++;
+    part->flags |= PART_SENT;
+    op->inflight_count++;
     part->op = (osd_op_t){
         .op_type = OSD_OP_OUT,
         .peer_fd = peer_it->second,
@@ -704,19 +812,18 @@ void cluster_client_t::send_sync(cluster_op_t *op, cluster_op_part_t *part)
 void cluster_client_t::handle_op_part(cluster_op_part_t *part)
 {
     cluster_op_t *op = part->parent;
-    part->sent = false;
-    op->sent_count--;
+    op->inflight_count--;
     int expected = part->op.req.hdr.opcode == OSD_OP_SYNC ? 0 : part->op.req.rw.len;
     if (part->op.reply.hdr.retval != expected)
     {
         // Operation failed, retry
         printf(
-            "Operation failed on OSD %lu: retval=%ld (expected %d), dropping connection\n",
-            part->osd_num, part->op.reply.hdr.retval, expected
+            "%s operation failed on OSD %lu: retval=%ld (expected %d), dropping connection\n",
+            osd_op_names[part->op.req.hdr.opcode], part->osd_num, part->op.reply.hdr.retval, expected
         );
-        msgr.stop_client(part->op.peer_fd);
         if (part->op.reply.hdr.retval == -EPIPE)
         {
+            // Mark op->up_wait = true before stopping the client
             op->up_wait = true;
             if (!retry_timeout_id)
             {
@@ -732,23 +839,18 @@ void cluster_client_t::handle_op_part(cluster_op_part_t *part)
             // Don't overwrite other errors with -EPIPE
             op->retval = part->op.reply.hdr.retval;
         }
+        msgr.stop_client(part->op.peer_fd);
+        part->flags |= PART_ERROR;
     }
     else
     {
         // OK
-        part->done = true;
+        dirty_osds.insert(part->osd_num);
+        part->flags |= PART_DONE;
         op->done_count++;
     }
-    if (op->sent_count == 0)
+    if (op->inflight_count == 0)
     {
-        if (op->opcode == OSD_OP_SYNC)
-        {
-            assert(op == cur_sync);
-            finish_sync();
-        }
-        else if (!op->up_wait)
-        {
-            continue_rw(op);
-        }
+        continue_ops();
     }
 }

src/cluster_client.h

@@ -10,7 +10,8 @@
 #define MAX_BLOCK_SIZE 128*1024*1024
 #define DEFAULT_DISK_ALIGNMENT 4096
 #define DEFAULT_BITMAP_GRANULARITY 4096
-#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;
 
@@ -22,8 +23,7 @@ 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;
 };
 
@@ -37,47 +37,53 @@ struct cluster_op_t
     osd_op_buf_list_t iov;
     std::function<void(cluster_op_t*)> callback;
 protected:
+    int flags = 0;
+    int state = 0;
     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;
     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;
     ring_loop_t *ringloop;
 
     uint64_t bs_block_size = 0;
-    uint64_t bs_disk_alignment = 0;
     uint64_t bs_bitmap_granularity = 0;
     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;
 
     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;
@@ -87,21 +93,22 @@ public:
     cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config);
     ~cluster_client_t();
     void execute(cluster_op_t *op);
+    bool is_ready();
     void on_ready(std::function<void(void)> fn);
-    void stop();
 
-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_osd_state_hook(uint64_t peer_osd);
-    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, cluster_op_part_t *part);
-    void execute_sync(cluster_op_t *op);
-    void continue_sync();
-    void finish_sync();
+    bool try_send(cluster_op_t *op, int i);
+    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);
 };

src/etcd_state_client.cpp

@@ -4,19 +4,24 @@
 #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()
 {
     etcd_watches_initialised = -1;
+#ifndef __MOCK__
     if (etcd_watch_ws)
     {
         etcd_watch_ws->close();
         etcd_watch_ws = NULL;
     }
+#endif
 }
 
+#ifndef __MOCK__
 json_kv_t etcd_state_client_t::parse_etcd_kv(const json11::Json & kv_json)
 {
     json_kv_t kv;
@@ -323,6 +328,26 @@ void etcd_state_client_t::load_pgs()
         start_etcd_watcher();
     });
 }
+#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 json_kv_t & kv)
+{
+    parse_state(kv.key, kv.value);
+}
 
 void etcd_state_client_t::parse_state(const std::string & key, const json11::Json & value)
 {
@@ -336,8 +361,10 @@ 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;
@@ -449,16 +476,19 @@ 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;

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
@@ -52,9 +52,13 @@ struct pool_config_t
     std::map<pg_num_t, pg_config_t> pg_config;
 };
 
+struct websocket_t;
+
 struct etcd_state_client_t
 {
 protected:
+    websocket_t *etcd_watch_ws = NULL;
+    uint64_t bs_block_size = DEFAULT_BLOCK_SIZE;
     void add_etcd_url(std::string);
 public:
     std::vector<std::string> etcd_addresses;
@@ -64,8 +68,6 @@ public:
 
     int etcd_watches_initialised = 0;
     uint64_t etcd_watch_revision = 0;
-    websocket_t *etcd_watch_ws = NULL;
-    uint64_t bs_block_size = 0;
     std::map<pool_id_t, pool_config_t> pool_config;
     std::map<osd_num_t, json11::Json> peer_states;
 
@@ -82,6 +84,7 @@ public:
     void start_etcd_watcher();
     void load_global_config();
     void load_pgs();
+    void parse_state(const json_kv_t & kv);
     void parse_state(const std::string & key, const json11::Json & value);
     void parse_config(json11::Json & config);
     ~etcd_state_client_t();

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);
-    }
 }
 
 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)
@@ -373,123 +365,6 @@ 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 +380,12 @@ 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
@@ -36,160 +32,6 @@
 #define DEFAULT_PEER_CONNECT_TIMEOUT 5
 #define DEFAULT_OSD_PING_TIMEOUT 5
 
-// 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;
-    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;
-    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;
@@ -228,6 +70,12 @@ 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
@@ -252,12 +100,9 @@ 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;
@@ -267,19 +112,22 @@ 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);
@@ -287,7 +135,6 @@ 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);
@@ -296,6 +143,7 @@ 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 +1,171 @@
+// 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;
+    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;
+    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_send.cpp

@@ -180,7 +180,7 @@ 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

@@ -8,16 +8,20 @@
 #include <arpa/inet.h>
 
 #include "osd.h"
+#include "http_client.h"
 
-osd_t::osd_t(blockstore_config_t & config, blockstore_t *bs, ring_loop_t *ringloop)
+osd_t::osd_t(blockstore_config_t & config, ring_loop_t *ringloop)
 {
+    config["entry_attr_size"] = "0";
+
     this->config = config;
-    this->bs = bs;
     this->ringloop = ringloop;
 
+    // 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_block_size = bs->get_block_size();
-    // FIXME: use bitmap granularity instead
-    this->bs_disk_alignment = bs->get_disk_alignment();
+    this->bs_bitmap_granularity = bs->get_bitmap_granularity();
 
     parse_config(config);
 
@@ -49,6 +53,7 @@ osd_t::~osd_t()
 {
     ringloop->unregister_consumer(&consumer);
     delete epmgr;
+    delete bs;
     close(listen_fd);
 }
 
@@ -171,7 +176,7 @@ bool osd_t::shutdown()
     {
         return false;
     }
-    return bs->is_safe_to_stop();
+    return !bs || bs->is_safe_to_stop();
 }
 
 void osd_t::loop()
@@ -191,6 +196,8 @@ 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 ||
@@ -198,14 +205,14 @@ void osd_t::exec_op(osd_op_t *cur_op)
             cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
             cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE) &&
             (cur_op->req.sec_rw.len > OSD_RW_MAX ||
-            cur_op->req.sec_rw.len % bs_disk_alignment ||
-            cur_op->req.sec_rw.offset % bs_disk_alignment)) ||
+            cur_op->req.sec_rw.len % bs_bitmap_granularity ||
+            cur_op->req.sec_rw.offset % bs_bitmap_granularity)) ||
         ((cur_op->req.hdr.opcode == OSD_OP_READ ||
             cur_op->req.hdr.opcode == OSD_OP_WRITE ||
             cur_op->req.hdr.opcode == OSD_OP_DELETE) &&
             (cur_op->req.rw.len > OSD_RW_MAX ||
-            cur_op->req.rw.len % bs_disk_alignment ||
-            cur_op->req.rw.offset % bs_disk_alignment)))
+            cur_op->req.rw.len % bs_bitmap_granularity ||
+            cur_op->req.rw.offset % bs_bitmap_granularity)))
     {
         // Bad command
         finish_op(cur_op, -EINVAL);

src/osd.h

@@ -115,7 +115,7 @@ class osd_t
     bool stopping = false;
     int inflight_ops = 0;
     blockstore_t *bs;
-    uint32_t bs_block_size, bs_disk_alignment;
+    uint32_t bs_block_size, bs_bitmap_granularity;
     ring_loop_t *ringloop;
     timerfd_manager_t *tfd = NULL;
     epoll_manager_t *epmgr = NULL;
@@ -198,6 +198,7 @@ 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);
@@ -206,9 +207,11 @@ class osd_t
     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_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);
+
     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)];
@@ -218,7 +221,7 @@ class osd_t
     }
 
 public:
-    osd_t(blockstore_config_t & config, blockstore_t *bs, ring_loop_t *ringloop);
+    osd_t(blockstore_config_t & config, ring_loop_t *ringloop);
     ~osd_t();
     void force_stop(int exitcode);
     bool shutdown();

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:
@@ -557,7 +558,7 @@ 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_main.cpp

@@ -41,16 +41,13 @@ int main(int narg, char *args[])
     signal(SIGINT, handle_sigint);
     signal(SIGTERM, handle_sigint);
     ring_loop_t *ringloop = new ring_loop_t(512);
-    // FIXME: Create Blockstore from on-disk superblock config and check it against the OSD cluster config
-    blockstore_t *bs = new blockstore_t(config, ringloop);
-    osd = new osd_t(config, bs, ringloop);
+    osd = new osd_t(config, ringloop);
     while (1)
     {
         ringloop->loop();
         ringloop->wait();
     }
     delete osd;
-    delete bs;
     delete ringloop;
     return 0;
 }

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

@@ -18,7 +18,7 @@ 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())
     {
@@ -44,8 +44,8 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
         return false;
     }
     if ((cur_op->req.rw.offset + cur_op->req.rw.len) > (oid.stripe + pg_block_size) ||
-        (cur_op->req.rw.offset % bs_disk_alignment) != 0 ||
-        (cur_op->req.rw.len % bs_disk_alignment) != 0)
+        (cur_op->req.rw.offset % bs_bitmap_granularity) != 0 ||
+        (cur_op->req.rw.len % bs_bitmap_granularity) != 0)
     {
         finish_op(cur_op, -EINVAL);
         return false;
@@ -64,7 +64,7 @@ bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
     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)))
     {
@@ -177,609 +177,6 @@ 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;
-        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);
-        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)
-    {
-        // Only (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);
-        }
-        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);
-        }
-    }
-    // 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)
 {
@@ -818,10 +215,14 @@ 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;
     }
 }
 
@@ -887,22 +288,21 @@ 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;
+    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)
+    {
+        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);
-    // 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)
+    if (next_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 next write to the same object
         continue_primary_write(next_op);
     }
-    }
 }

src/osd_primary_subops.cpp

@@ -43,12 +43,14 @@ 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);
@@ -194,14 +196,7 @@ void osd_t::submit_primary_subops(int submit_type, uint64_t op_version, int pg_s
                 }
                 subops[i].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]);
             }
@@ -247,6 +242,7 @@ 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);
 }
 
@@ -288,6 +284,11 @@ 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
     {
@@ -427,7 +428,7 @@ 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++,
@@ -435,23 +436,17 @@ 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;
@@ -459,6 +454,7 @@ 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];
@@ -475,30 +471,35 @@ 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)
@@ -531,24 +532,18 @@ 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]);
         }
@@ -566,7 +561,7 @@ 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 +1,378 @@
+// 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);
+        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;
+    }
+    if (op_data->scheme == POOL_SCHEME_REPLICATED)
+    {
+        // Only (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);
+        }
+        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);
+        }
+    }
+    // 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->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/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/qemu_proxy.cpp

@@ -47,7 +47,6 @@ public:
 
     ~QemuProxy()
     {
-        cli->stop();
         delete cli;
         delete tfd;
     }

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_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((json_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((json_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((json_kv_t){
+        .key = "/pg/state/1/1",
+        .value = json11::Json::object {
+            { "peers", json11::Json::array { 1, 2 } },
+            { "primary", 1 },
+            { "state", json11::Json::array { "active" } },
+        },
+    });
+    json11::Json::object 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_manager.cpp

@@ -121,7 +121,7 @@ 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 +159,6 @@ void timerfd_manager_t::trigger_nearest()
     {
         timers.erase(timers.begin()+nearest, timers.begin()+nearest+1);
     }
-    cb(nearest_id);
     nearest = -1;
+    cb(nearest_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_write.sh

@@ -35,6 +35,18 @@ fi
 #    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
 
 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=$ETCD_URL -pool=1 -inode=1 -size=1G -cluster_log_level=10
+    fio -thread -name=test -ioengine=build/src/libfio_vitastor.so -bs=4M -direct=1 -iodepth=1 -fsync=1 -rw=write -etcd=$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=$ETCD_URL -pool=1 -inode=1 -size=128M -number_ios=1024
+
+qemu-img convert -S 4096 -p \
+    -f raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=1:size=$((128*1024*1024))" \
+    -O raw ./testdata/read.bin
+
+qemu-img convert -S 4096 -p \
+    -f raw ./testdata/read.bin \
+    -O raw "vitastor:etcd_host=127.0.0.1\:$ETCD_PORT/v3:pool=1:inode=1:size=$((128*1024*1024))"
 
 format_green OK