WIP NFS RDMA support

.gitea/workflows/test.yml

@@ -288,6 +288,24 @@ jobs:
           echo ""
         done
 
+  test_create_halfhost:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_create_halfhost.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_failure_domain:
     runs-on: ubuntu-latest
     needs: build

docs/config/network.en.md

@@ -68,11 +68,17 @@ but they are not connected to the cluster.
 - Type: string
 
 RDMA device name to use for Vitastor OSD communications (for example,
-"rocep5s0f0"). Now Vitastor supports all adapters, even ones without
-ODP support, like Mellanox ConnectX-3 and non-Mellanox cards.
+"rocep5s0f0"). If not specified, Vitastor will try to find an RoCE
+device matching [osd_network](osd.en.md#osd_network), preferring RoCEv2,
+or choose the first available RDMA device if no RoCE devices are
+found or if `osd_network` is not specified. Auto-selection is also
+unsupported with old libibverbs < v32, like in Debian 10 Buster or
+CentOS 7.
 
-Versions up to Vitastor 1.2.0 required ODP which is only present in
-Mellanox ConnectX >= 4. See also [rdma_odp](#rdma_odp).
+Vitastor supports all adapters, even ones without ODP support, like
+Mellanox ConnectX-3 and non-Mellanox cards. Versions up to Vitastor
+1.2.0 required ODP which is only present in Mellanox ConnectX >= 4.
+See also [rdma_odp](#rdma_odp).
 
 Run `ibv_devinfo -v` as root to list available RDMA devices and their
 features.
@@ -95,15 +101,17 @@ your device has.
 ## rdma_gid_index
 
 - Type: integer
-- Default: 0
 
 Global address identifier index of the RDMA device to use. Different GID
 indexes may correspond to different protocols like RoCEv1, RoCEv2 and iWARP.
 Search for "GID" in `ibv_devinfo -v` output to determine which GID index
 you need.
 
-**IMPORTANT:** If you want to use RoCEv2 (as recommended) then the correct
-rdma_gid_index is usually 1 (IPv6) or 3 (IPv4).
+If not specified, Vitastor will try to auto-select a RoCEv2 IPv4 GID, then
+RoCEv2 IPv6 GID, then RoCEv1 IPv4 GID, then RoCEv1 IPv6 GID, then IB GID.
+GID auto-selection is unsupported with libibverbs < v32.
+
+A correct rdma_gid_index for RoCEv2 is usually 1 (IPv6) or 3 (IPv4).
 
 ## rdma_mtu
 

docs/config/network.ru.md

@@ -71,12 +71,17 @@ RDMA может быть нужно только если у клиентов е
 - Тип: строка
 
 Название RDMA-устройства для связи с Vitastor OSD (например, "rocep5s0f0").
-Сейчас Vitastor поддерживает все модели адаптеров, включая те, у которых
-нет поддержки ODP, то есть вы можете использовать RDMA с ConnectX-3 и
-картами производства не Mellanox.
+Если не указано, Vitastor попробует найти RoCE-устройство, соответствующее
+[osd_network](osd.en.md#osd_network), предпочитая RoCEv2, или выбрать первое
+попавшееся RDMA-устройство, если RoCE-устройств нет или если сеть `osd_network`
+не задана. Также автовыбор не поддерживается со старыми версиями библиотеки
+libibverbs < v32, например в Debian 10 Buster или CentOS 7.
 
-Версии Vitastor до 1.2.0 включительно требовали ODP, который есть только
-на Mellanox ConnectX 4 и более новых. См. также [rdma_odp](#rdma_odp).
+Vitastor поддерживает все модели адаптеров, включая те, у которых
+нет поддержки ODP, то есть вы можете использовать RDMA с ConnectX-3 и
+картами производства не Mellanox. Версии Vitastor до 1.2.0 включительно
+требовали ODP, который есть только на Mellanox ConnectX 4 и более новых.
+См. также [rdma_odp](#rdma_odp).
 
 Запустите `ibv_devinfo -v` от имени суперпользователя, чтобы посмотреть
 список доступных RDMA-устройств, их параметры и возможности.
@@ -101,15 +106,18 @@ Control) и ECN (Explicit Congestion Notification).
 ## rdma_gid_index
 
 - Тип: целое число
-- Значение по умолчанию: 0
 
 Номер глобального идентификатора адреса RDMA-устройства, который следует
 использовать. Разным gid_index могут соответствовать разные протоколы связи:
 RoCEv1, RoCEv2, iWARP. Чтобы понять, какой нужен вам - смотрите строчки со
 словом "GID" в выводе команды `ibv_devinfo -v`.
 
-**ВАЖНО:** Если вы хотите использовать RoCEv2 (как мы и рекомендуем), то
-правильный rdma_gid_index, как правило, 1 (IPv6) или 3 (IPv4).
+Если не указан, Vitastor попробует автоматически выбрать сначала GID,
+соответствующий RoCEv2 IPv4, потом RoCEv2 IPv6, потом RoCEv1 IPv4, потом
+RoCEv1 IPv6, потом IB. Авто-выбор GID не поддерживается со старыми версиями
+libibverbs < v32.
+
+Правильный rdma_gid_index для RoCEv2, как правило, 1 (IPv6) или 3 (IPv4).
 
 ## rdma_mtu
 

docs/config/src/network.yml

@@ -48,11 +48,17 @@
   type: string
   info: |
     RDMA device name to use for Vitastor OSD communications (for example,
-    "rocep5s0f0"). Now Vitastor supports all adapters, even ones without
-    ODP support, like Mellanox ConnectX-3 and non-Mellanox cards.
+    "rocep5s0f0"). If not specified, Vitastor will try to find an RoCE
+    device matching [osd_network](osd.en.md#osd_network), preferring RoCEv2,
+    or choose the first available RDMA device if no RoCE devices are
+    found or if `osd_network` is not specified. Auto-selection is also
+    unsupported with old libibverbs < v32, like in Debian 10 Buster or
+    CentOS 7.
 
-    Versions up to Vitastor 1.2.0 required ODP which is only present in
-    Mellanox ConnectX >= 4. See also [rdma_odp](#rdma_odp).
+    Vitastor supports all adapters, even ones without ODP support, like
+    Mellanox ConnectX-3 and non-Mellanox cards. Versions up to Vitastor
+    1.2.0 required ODP which is only present in Mellanox ConnectX >= 4.
+    See also [rdma_odp](#rdma_odp).
 
     Run `ibv_devinfo -v` as root to list available RDMA devices and their
     features.
@@ -64,12 +70,17 @@
     PFC (Priority Flow Control) and ECN (Explicit Congestion Notification).
   info_ru: |
     Название RDMA-устройства для связи с Vitastor OSD (например, "rocep5s0f0").
-    Сейчас Vitastor поддерживает все модели адаптеров, включая те, у которых
-    нет поддержки ODP, то есть вы можете использовать RDMA с ConnectX-3 и
-    картами производства не Mellanox.
+    Если не указано, Vitastor попробует найти RoCE-устройство, соответствующее
+    [osd_network](osd.en.md#osd_network), предпочитая RoCEv2, или выбрать первое
+    попавшееся RDMA-устройство, если RoCE-устройств нет или если сеть `osd_network`
+    не задана. Также автовыбор не поддерживается со старыми версиями библиотеки
+    libibverbs < v32, например в Debian 10 Buster или CentOS 7.
 
-    Версии Vitastor до 1.2.0 включительно требовали ODP, который есть только
-    на Mellanox ConnectX 4 и более новых. См. также [rdma_odp](#rdma_odp).
+    Vitastor поддерживает все модели адаптеров, включая те, у которых
+    нет поддержки ODP, то есть вы можете использовать RDMA с ConnectX-3 и
+    картами производства не Mellanox. Версии Vitastor до 1.2.0 включительно
+    требовали ODP, который есть только на Mellanox ConnectX 4 и более новых.
+    См. также [rdma_odp](#rdma_odp).
 
     Запустите `ibv_devinfo -v` от имени суперпользователя, чтобы посмотреть
     список доступных RDMA-устройств, их параметры и возможности.
@@ -94,23 +105,29 @@
     `ibv_devinfo -v`.
 - name: rdma_gid_index
   type: int
-  default: 0
   info: |
     Global address identifier index of the RDMA device to use. Different GID
     indexes may correspond to different protocols like RoCEv1, RoCEv2 and iWARP.
     Search for "GID" in `ibv_devinfo -v` output to determine which GID index
     you need.
 
-    **IMPORTANT:** If you want to use RoCEv2 (as recommended) then the correct
-    rdma_gid_index is usually 1 (IPv6) or 3 (IPv4).
+    If not specified, Vitastor will try to auto-select a RoCEv2 IPv4 GID, then
+    RoCEv2 IPv6 GID, then RoCEv1 IPv4 GID, then RoCEv1 IPv6 GID, then IB GID.
+    GID auto-selection is unsupported with libibverbs < v32.
+
+    A correct rdma_gid_index for RoCEv2 is usually 1 (IPv6) or 3 (IPv4).
   info_ru: |
     Номер глобального идентификатора адреса RDMA-устройства, который следует
     использовать. Разным gid_index могут соответствовать разные протоколы связи:
     RoCEv1, RoCEv2, iWARP. Чтобы понять, какой нужен вам - смотрите строчки со
     словом "GID" в выводе команды `ibv_devinfo -v`.
 
-    **ВАЖНО:** Если вы хотите использовать RoCEv2 (как мы и рекомендуем), то
-    правильный rdma_gid_index, как правило, 1 (IPv6) или 3 (IPv4).
+    Если не указан, Vitastor попробует автоматически выбрать сначала GID,
+    соответствующий RoCEv2 IPv4, потом RoCEv2 IPv6, потом RoCEv1 IPv4, потом
+    RoCEv1 IPv6, потом IB. Авто-выбор GID не поддерживается со старыми версиями
+    libibverbs < v32.
+
+    Правильный rdma_gid_index для RoCEv2, как правило, 1 (IPv6) или 3 (IPv4).
 - name: rdma_mtu
   type: int
   default: 4096

mon/etcd_adapter.js

@@ -232,6 +232,7 @@ class EtcdAdapter
     async become_master()
     {
         const state = { ...this.mon.get_mon_state(), id: ''+this.mon.etcd_lease_id };
+        console.log('Waiting to become master');
         // eslint-disable-next-line no-constant-condition
         while (1)
         {
@@ -243,7 +244,6 @@ class EtcdAdapter
             {
                 break;
             }
-            console.log('Waiting to become master');
             await new Promise(ok => setTimeout(ok, this.mon.config.etcd_start_timeout));
         }
         console.log('Became master');

mon/etcd_schema.js

@@ -56,6 +56,7 @@ const etcd_tree = {
             osd_out_time: 600, // seconds. min: 0
             placement_levels: { datacenter: 1, rack: 2, host: 3, osd: 4, ... },
             use_old_pg_combinator: false,
+            osd_backfillfull_ratio: 0.99,
             // client and osd
             tcp_header_buffer_size: 65536,
             use_sync_send_recv: false,

mon/mon.js

@@ -74,6 +74,7 @@ class Mon
         this.state = JSON.parse(JSON.stringify(etcd_tree));
         this.prev_stats = { osd_stats: {}, osd_diff: {} };
         this.recheck_pgs_active = false;
+        this.updating_total_stats = false;
         this.watcher_active = false;
         this.old_pg_config = false;
         this.old_pg_stats_seen = false;
@@ -658,7 +659,19 @@ class Mon
                 this.etcd_watch_revision, pool_id, up_osds, osd_tree, real_prev_pgs, pool_res.pgs, pg_history);
         }
         new_pg_config.hash = tree_hash;
-        return await this.save_pg_config(new_pg_config, etcd_request);
+        const { backfillfull_pools, backfillfull_osds } = sum_object_counts(
+            { ...this.state, pg: { ...this.state.pg, config: new_pg_config } }, this.config
+        );
+        if (backfillfull_pools.join(',') != ((this.state.pg.config||{}).backfillfull_pools||[]).join(','))
+        {
+            this.log_backfillfull(backfillfull_osds, backfillfull_pools);
+        }
+        new_pg_config.backfillfull_pools = backfillfull_pools.length ? backfillfull_pools : undefined;
+        if (!await this.save_pg_config(new_pg_config, etcd_request))
+        {
+            return false;
+        }
+        return true;
     }
 
     async save_pg_config(new_pg_config, etcd_request = { compare: [], success: [] })
@@ -730,7 +743,7 @@ class Mon
     async update_total_stats()
     {
         const txn = [];
-        const { object_counts, object_bytes } = sum_object_counts(this.state, this.config);
+        const { object_counts, object_bytes, backfillfull_pools, backfillfull_osds } = sum_object_counts(this.state, this.config);
         let stats = sum_op_stats(this.state.osd, this.prev_stats);
         let { inode_stats, seen_pools } = sum_inode_stats(this.state, this.prev_stats);
         stats.object_counts = object_counts;
@@ -783,6 +796,27 @@ class Mon
         {
             await this.etcd.etcd_call('/kv/txn', { success: txn }, this.config.etcd_mon_timeout, 0);
         }
+        if (!this.recheck_pgs_active &&
+            backfillfull_pools.join(',') != ((this.state.pg.config||{}).backfillfull_pools||[]).join(','))
+        {
+            this.log_backfillfull(backfillfull_osds, backfillfull_pools);
+            const new_pg_config = { ...this.state.pg.config, backfillfull_pools: backfillfull_pools.length ? backfillfull_pools : undefined };
+            await this.save_pg_config(new_pg_config);
+        }
+    }
+
+    log_backfillfull(osds, pools)
+    {
+        for (const osd in osds)
+        {
+            const bf = osds[osd];
+            console.log('OSD '+osd+' may fill up during rebalance: capacity '+(bf.cap/1024n/1024n)+
+                ' MB, target user data '+(bf.clean/1024n/1024n)+' MB');
+        }
+        console.log(
+            (pools.length ? 'Pool(s) '+pools.join(', ') : 'No pools')+
+            ' are backfillfull now, applying rebalance configuration'
+        );
     }
 
     schedule_update_stats()
@@ -794,7 +828,21 @@ class Mon
         this.stats_timer = setTimeout(() =>
         {
             this.stats_timer = null;
-            this.update_total_stats().catch(console.error);
+            if (this.updating_total_stats)
+            {
+                this.schedule_update_stats();
+                return;
+            }
+            this.updating_total_stats = true;
+            try
+            {
+                this.update_total_stats().catch(console.error);
+            }
+            catch (e)
+            {
+                console.error(e);
+            }
+            this.updating_total_stats = false;
         }, this.config.mon_stats_timeout);
     }
 

mon/stats.js

@@ -109,6 +109,8 @@ function sum_object_counts(state, global_config)
             pgstats[pool_id] = { ...(state.pg.stats[pool_id] || {}), ...(pgstats[pool_id] || {}) };
         }
     }
+    const pool_per_osd = {};
+    const clean_per_osd = {};
     for (const pool_id in pgstats)
     {
         let object_size = 0;
@@ -143,10 +145,45 @@ function sum_object_counts(state, global_config)
                         object_bytes[k] += BigInt(st[k+'_count']) * object_size;
                     }
                 }
+                if (st.object_count)
+                {
+                    for (const pg_osd of (((state.pg.config.items||{})[pool_id]||{})[pg_num]||{}).osd_set||[])
+                    {
+                        if (!(pg_osd in clean_per_osd))
+                        {
+                            clean_per_osd[pg_osd] = 0n;
+                        }
+                        clean_per_osd[pg_osd] += BigInt(st.object_count);
+                        pool_per_osd[pg_osd] = pool_per_osd[pg_osd]||{};
+                        pool_per_osd[pg_osd][pool_id] = true;
+                    }
+                }
             }
         }
     }
-    return { object_counts, object_bytes };
+    // If clean_per_osd[osd] is larger than osd capacity then it will fill up during rebalance
+    let backfillfull_pools = {};
+    let backfillfull_osds = {};
+    for (const osd in clean_per_osd)
+    {
+        const st = state.osd.stats[osd];
+        if (!st || !st.size || !st.data_block_size)
+        {
+            continue;
+        }
+        let cap = BigInt(st.size)/BigInt(st.data_block_size);
+        cap = cap * BigInt((global_config.osd_backfillfull_ratio||0.99)*1000000) / 1000000n;
+        if (cap < clean_per_osd[osd])
+        {
+            backfillfull_osds[osd] = { cap: BigInt(st.size), clean: clean_per_osd[osd]*BigInt(st.data_block_size) };
+            for (const pool_id in pool_per_osd[osd])
+            {
+                backfillfull_pools[pool_id] = true;
+            }
+        }
+    }
+    backfillfull_pools = Object.keys(backfillfull_pools).sort();
+    return { object_counts, object_bytes, backfillfull_pools, backfillfull_osds };
 }
 
 // sum_inode_stats(this.state, this.prev_stats)

src/client/etcd_state_client.cpp

@@ -785,7 +785,7 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         }
         for (auto & pool_item: value.object_items())
         {
-            pool_config_t pc;
+            pool_config_t pc = {};
             // ID
             pool_id_t pool_id;
             char null_byte = 0;
@@ -931,12 +931,28 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             // Ignore old key if the new one is present
             return;
         }
+        for (auto & pool_id_json: value["backfillfull_pools"].array_items())
+        {
+            auto pool_id = pool_id_json.uint64_value();
+            auto pool_it = this->pool_config.find(pool_id);
+            if (pool_it != this->pool_config.end())
+            {
+                pool_it->second.backfillfull |= 2;
+            }
+        }
         for (auto & pool_item: this->pool_config)
         {
             for (auto & pg_item: pool_item.second.pg_config)
             {
                 pg_item.second.config_exists = false;
             }
+            // 3 = was 1 and became 1, 0 = was 0 and became 0
+            if (pool_item.second.backfillfull == 2 || pool_item.second.backfillfull == 1)
+            {
+                if (on_change_backfillfull_hook)
+                    on_change_backfillfull_hook(pool_item.first);
+            }
+            pool_item.second.backfillfull = pool_item.second.backfillfull >> 1;
         }
         for (auto & pool_item: value["items"].object_items())
         {

src/client/etcd_state_client.h

@@ -62,6 +62,7 @@ struct pool_config_t
     std::map<pg_num_t, pg_config_t> pg_config;
     uint64_t scrub_interval;
     std::string used_for_fs;
+    int backfillfull;
 };
 
 struct inode_config_t
@@ -131,6 +132,7 @@ public:
     std::function<json11::Json()> load_pgs_checks_hook;
     std::function<void(bool)> on_load_pgs_hook;
     std::function<void()> on_change_pool_config_hook;
+    std::function<void(pool_id_t)> on_change_backfillfull_hook;
     std::function<void(pool_id_t, pg_num_t, osd_num_t)> on_change_pg_state_hook;
     std::function<void(pool_id_t, pg_num_t)> on_change_pg_history_hook;
     std::function<void(osd_num_t)> on_change_osd_state_hook;

src/client/msgr_rdma.cpp

@@ -70,6 +70,7 @@ msgr_rdma_connection_t::~msgr_rdma_connection_t()
     send_out_size = 0;
 }
 
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
 static bool is_ipv4_gid(ibv_gid_entry *gidx)
 {
     return (((uint64_t*)gidx->gid.raw)[0] == 0 &&
@@ -131,6 +132,7 @@ static matched_dev match_device(ibv_device **dev_list, addr_mask_t *networks, in
     ibv_port_attr portinfo;
     ibv_gid_entry best_gidx;
     int res;
+    bool have_non_roce = false, have_roce = false;
     for (int i = 0; dev_list[i]; ++i)
     {
         auto dev = dev_list[i];
@@ -161,6 +163,11 @@ static matched_dev match_device(ibv_device **dev_list, addr_mask_t *networks, in
                     else
                         break;
                 }
+                if (gidx.gid_type != IBV_GID_TYPE_ROCE_V1 &&
+                    gidx.gid_type != IBV_GID_TYPE_ROCE_V2)
+                    have_non_roce = true;
+                else
+                    have_roce = true;
                 if (match_gid(&gidx, networks, nnet))
                 {
                     // Prefer RoCEv2
@@ -186,8 +193,13 @@ cleanup:
     {
         log_rdma_dev_port_gid(dev_list[best.dev], best.port, best.gid, best_gidx);
     }
+    if (best.dev < 0 && have_non_roce && !have_roce)
+    {
+        best.dev = -2;
+    }
     return best;
 }
+#endif
 
 msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_networks, const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level)
 {
@@ -229,6 +241,7 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
             goto cleanup;
         }
     }
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
     else if (osd_networks.size())
     {
         std::vector<addr_mask_t> nets;
@@ -237,11 +250,17 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
             nets.push_back(cidr_parse(netstr));
         }
         auto best = match_device(dev_list, nets.data(), nets.size(), log_level);
-        if (best.dev < 0)
+        if (best.dev == -2)
+        {
+            best.dev = 0;
+            if (log_level > 0)
+                fprintf(stderr, "No RoCE devices found, using first available RDMA device %s\n", ibv_get_device_name(*dev_list));
+        }
+        else if (best.dev < 0)
         {
             if (log_level > 0)
-                fprintf(stderr, "RDMA device matching osd_network is not found, using first available device\n");
-            best.dev = 0;
+                fprintf(stderr, "RDMA device matching osd_network is not found, disabling RDMA\n");
+            goto cleanup;
         }
         else
         {
@@ -250,6 +269,7 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
         }
         ctx->dev = dev_list[best.dev];
     }
+#endif
     else
     {
         ctx->dev = *dev_list;
@@ -275,18 +295,22 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
         goto cleanup;
     }
 
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
     if (gid_index != -1)
+#endif
     {
-        ctx->gid_index = gid_index;
-        if (ibv_query_gid_ex(ctx->context, ib_port, gid_index, &ctx->my_gid, 0))
+        ctx->gid_index = gid_index < 0 ? 0 : gid_index;
+        if (ibv_query_gid(ctx->context, ib_port, gid_index, &ctx->my_gid))
         {
             fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), gid_index);
             goto cleanup;
         }
     }
+#ifdef IBV_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT
     else
     {
         // Auto-guess GID
+        ibv_gid_entry best_gidx;
         for (int k = 0; k < ctx->portinfo.gid_tbl_len; k++)
         {
             ibv_gid_entry gidx;
@@ -301,21 +325,24 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(std::vector<std::string> osd_ne
             {
                 continue;
             }
-            // Prefer IPv4 RoCEv2 GID by default
+            // Prefer IPv4 RoCEv2 -> IPv6 RoCEv2 -> IPv4 RoCEv1 -> IPv6 RoCEv1 -> IB
             if (gid_index == -1 ||
-                gidx.gid_type == IBV_GID_TYPE_ROCE_V2 &&
-                (ctx->my_gid.gid_type != IBV_GID_TYPE_ROCE_V2 || is_ipv4_gid(&gidx)))
+                gidx.gid_type == IBV_GID_TYPE_ROCE_V2 && best_gidx.gid_type != IBV_GID_TYPE_ROCE_V2 ||
+                gidx.gid_type == IBV_GID_TYPE_ROCE_V1 && best_gidx.gid_type == IBV_GID_TYPE_IB ||
+                gidx.gid_type == best_gidx.gid_type && is_ipv4_gid(&gidx))
             {
                 gid_index = k;
-                ctx->my_gid = gidx;
+                best_gidx = gidx;
             }
         }
         ctx->gid_index = gid_index = (gid_index == -1 ? 0 : gid_index);
         if (log_level > 0)
         {
-            log_rdma_dev_port_gid(ctx->dev, ctx->ib_port, ctx->gid_index, ctx->my_gid);
+            log_rdma_dev_port_gid(ctx->dev, ctx->ib_port, ctx->gid_index, best_gidx);
         }
+        ctx->my_gid = best_gidx.gid;
     }
+#endif
 
     ctx->pd = ibv_alloc_pd(ctx->context);
     if (!ctx->pd)
@@ -431,7 +458,7 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
     }
 
     conn->addr.lid = ctx->my_lid;
-    conn->addr.gid = ctx->my_gid.gid;
+    conn->addr.gid = ctx->my_gid;
     conn->addr.qpn = conn->qp->qp_num;
     conn->addr.psn = lrand48() & 0xffffff;
 

src/client/msgr_rdma.h

@@ -31,7 +31,7 @@ struct msgr_rdma_context_t
     uint8_t ib_port;
     uint8_t gid_index;
     uint16_t my_lid;
-    ibv_gid_entry my_gid;
+    ibv_gid my_gid;
     uint32_t mtu;
     int max_cqe = 0;
     int used_max_cqe = 0;

src/client/nbd_proxy.cpp

@@ -64,7 +64,7 @@ static void netlink_sock_alloc(struct netlink_ctx *ctx)
     if (nl_driver_id < 0)
     {
         nl_socket_free(sk);
-        fail("Couldn't resolve the nbd netlink family\n");
+        fail("Couldn't resolve the nbd netlink family: %s (code %d)\n", nl_geterror(nl_driver_id), nl_driver_id);
     }
 
     ctx->driver_id = nl_driver_id;
@@ -555,7 +555,12 @@ help:
         fcntl(sockfd[0], F_SETFL, fcntl(sockfd[0], F_GETFL, 0) | O_NONBLOCK);
         nbd_fd = sockfd[0];
         load_module();
+
         bool bg = cfg["foreground"].is_null();
+        if (cfg["logfile"].string_value() != "")
+        {
+            logfile = cfg["logfile"].string_value();
+        }
 
         if (netlink)
         {
@@ -588,6 +593,10 @@ help:
             }
             close(sockfd[1]);
             printf("/dev/nbd%d\n", err);
+            if (bg)
+            {
+                daemonize_reopen_stdio();
+            }
 #else
             fprintf(stderr, "netlink support is disabled in this build\n");
             exit(1);
@@ -631,14 +640,10 @@ help:
                     }
                 }
             }
-        }
-        if (cfg["logfile"].string_value() != "")
-        {
-            logfile = cfg["logfile"].string_value();
-        }
-        if (bg)
-        {
-            daemonize();
+            if (bg)
+            {
+                daemonize();
+            }
         }
         // Initialize read state
         read_state = CL_READ_HDR;
@@ -716,13 +721,17 @@ help:
         }
     }
 
-    void daemonize()
+    void daemonize_fork()
     {
         if (fork())
             exit(0);
         setsid();
         if (fork())
             exit(0);
+    }
+
+    void daemonize_reopen_stdio()
+    {
         close(0);
         close(1);
         close(2);
@@ -733,6 +742,12 @@ help:
             fprintf(stderr, "Warning: Failed to chdir into /\n");
     }
 
+    void daemonize()
+    {
+        daemonize_fork();
+        daemonize_reopen_stdio();
+    }
+
     json11::Json::object list_mapped()
     {
         const char *self_filename = exe_name;
@@ -783,8 +798,9 @@ help:
                 if (!strcmp(pid_filename, self_filename))
                 {
                     json11::Json::object cfg = nbd_proxy::parse_args(argv.size(), argv.data());
-                    if (cfg["command"] == "map")
+                    if (cfg["command"] == "map" || cfg["command"] == "netlink-map")
                     {
+                        cfg["interface"] = (cfg["command"] == "netlink-map") ? "netlink" : "nbd";
                         cfg.erase("command");
                         cfg["pid"] = pid;
                         mapped["/dev/nbd"+std::to_string(dev_num)] = cfg;

src/cmd/cli_dd.cpp

@@ -261,6 +261,7 @@ struct dd_out_info_t
             else
             {
                 // ok
+                out_size = owatch->cfg.size;
                 return true;
             }
             // Wait for sub-command
@@ -882,6 +883,8 @@ resume_2:
         oinfo.end_fsync = oinfo.end_fsync && oinfo.out_seekable;
         read_offset = 0;
         read_end = iinfo.in_seekable ? iinfo.in_size-iseek : 0;
+        if (oinfo.out_size && (!read_end || read_end > oinfo.out_size-oseek))
+            read_end = oinfo.out_size-oseek;
         if (bytelimit && (!read_end || read_end > bytelimit))
             read_end = bytelimit;
         clock_gettime(CLOCK_REALTIME, &tv_begin);

src/cmd/cli_pool_create.cpp

@@ -35,6 +35,7 @@ struct pool_creator_t
     uint64_t new_pools_mod_rev;
     json11::Json state_node_tree;
     json11::Json new_pools;
+    std::map<osd_num_t, json11::Json> osd_stats;
 
     bool is_done() { return state == 100; }
 
@@ -46,8 +47,6 @@ struct pool_creator_t
             goto resume_2;
         else if (state == 3)
             goto resume_3;
-        else if (state == 4)
-            goto resume_4;
         else if (state == 5)
             goto resume_5;
         else if (state == 6)
@@ -90,13 +89,19 @@ resume_1:
         // If not forced, check that we have enough osds for pg_size
         if (!force)
         {
-            // Get node_placement configuration from etcd
+            // Get node_placement configuration from etcd and OSD stats
             parent->etcd_txn(json11::Json::object {
                 { "success", json11::Json::array {
                     json11::Json::object {
                         { "request_range", json11::Json::object {
                             { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/config/node_placement") },
-                        } }
+                        } },
+                    },
+                    json11::Json::object {
+                        { "request_range", json11::Json::object {
+                            { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/osd/stats/") },
+                            { "range_end", base64_encode(parent->cli->st_cli.etcd_prefix+"/osd/stats0") },
+                        } },
                     },
                 } },
             });
@@ -112,10 +117,21 @@ resume_2:
                 return;
             }
 
-            // Get state_node_tree based on node_placement and osd peer states
+            // Get state_node_tree based on node_placement and osd stats
             {
-                auto kv = parent->cli->st_cli.parse_etcd_kv(parent->etcd_result["responses"][0]["response_range"]["kvs"][0]);
-                state_node_tree = get_state_node_tree(kv.value.object_items());
+                auto node_placement_kv = parent->cli->st_cli.parse_etcd_kv(parent->etcd_result["responses"][0]["response_range"]["kvs"][0]);
+                timespec tv_now;
+                clock_gettime(CLOCK_REALTIME, &tv_now);
+                uint64_t osd_out_time = parent->cli->config["osd_out_time"].uint64_value();
+                if (!osd_out_time)
+                    osd_out_time = 600;
+                osd_stats.clear();
+                parent->iterate_kvs_1(parent->etcd_result["responses"][1]["response_range"]["kvs"], "/osd/stats/", [&](uint64_t cur_osd, json11::Json value)
+                {
+                    if ((uint64_t)value["time"].number_value()+osd_out_time >= tv_now.tv_sec)
+                        osd_stats[cur_osd] = value;
+                });
+                state_node_tree = get_state_node_tree(node_placement_kv.value.object_items(), osd_stats);
             }
 
             // Skip tag checks, if pool has none
@@ -158,42 +174,18 @@ resume_3:
                 }
             }
 
-            // Get stats (for block_size, bitmap_granularity, ...) of osds in state_node_tree
-            {
-                json11::Json::array osd_stats;
-
-                for (auto osd_num: state_node_tree["osds"].array_items())
-                {
-                    osd_stats.push_back(json11::Json::object {
-                        { "request_range", json11::Json::object {
-                            { "key", base64_encode(parent->cli->st_cli.etcd_prefix+"/osd/stats/"+osd_num.as_string()) },
-                        } }
-                    });
-                }
-
-                parent->etcd_txn(json11::Json::object{ { "success", osd_stats } });
-            }
-
-            state = 4;
-resume_4:
-            if (parent->waiting > 0)
-                return;
-            if (parent->etcd_err.err)
-            {
-                result = parent->etcd_err;
-                state = 100;
-                return;
-            }
-
             // Filter osds from state_node_tree based on pool parameters and osd stats
             {
-                std::vector<json11::Json> osd_stats;
-                for (auto & ocr: parent->etcd_result["responses"].array_items())
+                std::vector<json11::Json> filtered_osd_stats;
+                for (auto & osd_num: state_node_tree["osds"].array_items())
                 {
-                    auto kv = parent->cli->st_cli.parse_etcd_kv(ocr["response_range"]["kvs"][0]);
-                    osd_stats.push_back(kv.value);
+                    auto st_it = osd_stats.find(osd_num.uint64_value());
+                    if (st_it != osd_stats.end())
+                    {
+                        filtered_osd_stats.push_back(st_it->second);
+                    }
                 }
-                guess_block_size(osd_stats);
+                guess_block_size(filtered_osd_stats);
                 state_node_tree = filter_state_node_tree_by_stats(state_node_tree, osd_stats);
             }
 
@@ -201,8 +193,7 @@ resume_4:
             {
                 auto failure_domain = cfg["failure_domain"].string_value() == ""
                     ? "host" : cfg["failure_domain"].string_value();
-                uint64_t max_pg_size = get_max_pg_size(state_node_tree["nodes"].object_items(),
-                    failure_domain, cfg["root_node"].string_value());
+                uint64_t max_pg_size = get_max_pg_size(state_node_tree, failure_domain, cfg["root_node"].string_value());
 
                 if (cfg["pg_size"].uint64_value() > max_pg_size)
                 {
@@ -358,56 +349,50 @@ resume_8:
 
     // Returns a JSON object of form {"nodes": {...}, "osds": [...]} that
     // contains: all nodes (osds, hosts, ...) based on node_placement config
-    // and current peer state, and a list of active peer osds.
-    json11::Json get_state_node_tree(json11::Json::object node_placement)
+    // and current osd stats.
+    json11::Json get_state_node_tree(json11::Json::object node_placement, std::map<osd_num_t, json11::Json> & osd_stats)
     {
-        // Erase non-peer osd nodes from node_placement
+        // Erase non-existing osd nodes from node_placement
         for (auto np_it = node_placement.begin(); np_it != node_placement.end();)
         {
             // Numeric nodes are osds
             osd_num_t osd_num = stoull_full(np_it->first);
 
-            // If node is osd and it is not in peer states, erase it
-            if (osd_num > 0 &&
-                parent->cli->st_cli.peer_states.find(osd_num) == parent->cli->st_cli.peer_states.end())
-            {
+            // If node is osd and its stats do not exist, erase it
+            if (osd_num > 0 && osd_stats.find(osd_num) == osd_stats.end())
                 node_placement.erase(np_it++);
-            }
             else
                 np_it++;
         }
 
-        // List of peer osds
-        std::vector<std::string> peer_osds;
+        // List of osds
+        std::vector<std::string> existing_osds;
 
-        // Record peer osds and add missing osds/hosts to np
-        for (auto & ps: parent->cli->st_cli.peer_states)
+        // Record osds and add missing osds/hosts to np
+        for (auto & ps: osd_stats)
         {
             std::string osd_num = std::to_string(ps.first);
 
-            // Record peer osd
-            peer_osds.push_back(osd_num);
+            // Record osd
+            existing_osds.push_back(osd_num);
 
-            // Add osd, if necessary
-            if (node_placement.find(osd_num) == node_placement.end())
+            // Add host if necessary
+            std::string osd_host = ps.second["host"].as_string();
+            if (node_placement.find(osd_host) == node_placement.end())
             {
-                std::string osd_host = ps.second["host"].as_string();
-
-                // Add host, if necessary
-                if (node_placement.find(osd_host) == node_placement.end())
-                {
-                    node_placement[osd_host] = json11::Json::object {
-                        { "level", "host" }
-                    };
-                }
-
-                node_placement[osd_num] = json11::Json::object {
-                    { "parent", osd_host }
+                node_placement[osd_host] = json11::Json::object {
+                    { "level", "host" }
                 };
             }
+
+            // Add osd
+            node_placement[osd_num] = json11::Json::object {
+                { "parent", node_placement[osd_num]["parent"].is_null() ? osd_host : node_placement[osd_num]["parent"] },
+                { "level", "osd" },
+            };
         }
 
-        return json11::Json::object { { "osds", peer_osds }, { "nodes", node_placement } };
+        return json11::Json::object { { "osds", existing_osds }, { "nodes", node_placement } };
     }
 
     // Returns new state_node_tree based on given state_node_tree with osds
@@ -534,15 +519,13 @@ resume_8:
     // filtered out by stats parameters (block_size, bitmap_granularity) in
     // given osd_stats and current pool config.
     // Requires: state_node_tree["osds"] must match osd_stats 1-1
-    json11::Json filter_state_node_tree_by_stats(const json11::Json & state_node_tree, std::vector<json11::Json> & osd_stats)
+    json11::Json filter_state_node_tree_by_stats(const json11::Json & state_node_tree, std::map<osd_num_t, json11::Json> & osd_stats)
     {
-        auto & osds = state_node_tree["osds"].array_items();
-
         // Accepted state_node_tree nodes
         auto accepted_nodes = state_node_tree["nodes"].object_items();
 
         // List of accepted osds
-        std::vector<std::string> accepted_osds;
+        json11::Json::array accepted_osds;
 
         block_size = cfg["block_size"].uint64_value()
             ? cfg["block_size"].uint64_value()
@@ -554,21 +537,25 @@ resume_8:
             ? etcd_state_client_t::parse_immediate_commit(cfg["immediate_commit"].string_value(), IMMEDIATE_ALL)
             : parent->cli->st_cli.global_immediate_commit;
 
-        for (size_t i = 0; i < osd_stats.size(); i++)
+        for (auto osd_num_json: state_node_tree["osds"].array_items())
         {
-            auto & os = osd_stats[i];
-            // Get osd number
-            auto osd_num = osds[i].as_string();
+            auto osd_num = osd_num_json.uint64_value();
+            auto os_it = osd_stats.find(osd_num);
+            if (os_it == osd_stats.end())
+            {
+                continue;
+            }
+            auto & os = os_it->second;
             if (!os["data_block_size"].is_null() && os["data_block_size"] != block_size ||
                 !os["bitmap_granularity"].is_null() && os["bitmap_granularity"] != bitmap_granularity ||
                 !os["immediate_commit"].is_null() &&
                 etcd_state_client_t::parse_immediate_commit(os["immediate_commit"].string_value(), IMMEDIATE_NONE) < immediate_commit)
             {
-                accepted_nodes.erase(osd_num);
+                accepted_nodes.erase(osd_num_json.as_string());
             }
             else
             {
-                accepted_osds.push_back(osd_num);
+                accepted_osds.push_back(osd_num_json);
             }
         }
 
@@ -576,87 +563,28 @@ resume_8:
     }
 
     // Returns maximum pg_size possible for given node_tree and failure_domain, starting at parent_node
-    uint64_t get_max_pg_size(json11::Json::object node_tree, const std::string & level, const std::string & parent_node)
+    uint64_t get_max_pg_size(json11::Json state_node_tree, const std::string & level, const std::string & root_node)
     {
-        uint64_t max_pg_sz = 0;
-
-        std::vector<std::string> nodes;
-
-        // Check if parent node is an osd (numeric)
-        if (parent_node != "" && stoull_full(parent_node))
+        std::set<std::string> level_seen;
+        for (auto & osd: state_node_tree["osds"].array_items())
         {
-            // Add it to node list if osd is in node tree
-            if (node_tree.find(parent_node) != node_tree.end())
-                nodes.push_back(parent_node);
-        }
-        // If parent node given, ...
-        else if (parent_node != "")
-        {
-            // ... look for children nodes of this parent
-            for (auto & sn: node_tree)
+            // find OSD parent at <level>, but stop at <root_node>
+            auto cur_id = osd.string_value();
+            auto cur = state_node_tree["nodes"][cur_id];
+            while (!cur.is_null())
             {
-                auto & props = sn.second.object_items();
-
-                auto parent_prop = props.find("parent");
-                if (parent_prop != props.end() && (parent_prop->second.as_string() == parent_node))
+                if (cur["level"] == level)
                 {
-                    nodes.push_back(sn.first);
-
-                    // If we're not looking for all osds, we only need a single
-                    // child osd node
-                    if (level != "osd" && stoull_full(sn.first))
-                        break;
+                    level_seen.insert(cur_id);
+                    break;
                 }
+                if (cur_id == root_node)
+                    break;
+                cur_id = cur["parent"].string_value();
+                cur = state_node_tree["nodes"][cur_id];
             }
         }
-        // No parent node given, and we're not looking for all osds
-        else if (level != "osd")
-        {
-            // ... look for all level nodes
-            for (auto & sn: node_tree)
-            {
-                auto & props = sn.second.object_items();
-
-                auto level_prop = props.find("level");
-                if (level_prop != props.end() && (level_prop->second.as_string() == level))
-                {
-                    nodes.push_back(sn.first);
-                }
-            }
-        }
-        // Otherwise, ...
-        else
-        {
-            // ... we're looking for osd nodes only
-            for (auto & sn: node_tree)
-            {
-                if (stoull_full(sn.first))
-                {
-                    nodes.push_back(sn.first);
-                }
-            }
-        }
-
-        // Process gathered nodes
-        for (auto & node: nodes)
-        {
-            // Check for osd node, return constant max size
-            if (stoull_full(node))
-            {
-                max_pg_sz += 1;
-            }
-            // Otherwise, ...
-            else
-            {
-                // ... exclude parent node from tree, and ...
-                node_tree.erase(parent_node);
-
-                // ... descend onto the resulting tree
-                max_pg_sz += get_max_pg_size(node_tree, level, node);
-            }
-        }
-
-        return max_pg_sz;
+        return level_seen.size();
     }
 
     json11::Json create_pool(const etcd_kv_t & kv)

src/nfs/nfs_proxy.cpp

@@ -499,6 +499,19 @@ void nfs_proxy_t::check_default_pool()
     }
 }
 
+void nfs_proxy_t::create_client()
+{
+    auto cli = new nfs_client_t();
+    cli->parent = this->proxy;
+    if (kvfs)
+        nfs_kv_procs(cli);
+    else
+        nfs_block_procs(cli);
+    for (auto & fn: pmap.proc_table)
+        cli->proc_table.insert(fn);
+    rpc_clients.insert(cli);
+}
+
 void nfs_proxy_t::do_accept(int listen_fd)
 {
     struct sockaddr_storage addr;
@@ -512,18 +525,8 @@ void nfs_proxy_t::do_accept(int listen_fd)
         fcntl(nfs_fd, F_SETFL, fcntl(nfs_fd, F_GETFL, 0) | O_NONBLOCK);
         int one = 1;
         setsockopt(nfs_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-        auto cli = new nfs_client_t();
-        if (kvfs)
-            nfs_kv_procs(cli);
-        else
-            nfs_block_procs(cli);
-        cli->parent = this;
+        auto cli = this->create_client();
         cli->nfs_fd = nfs_fd;
-        for (auto & fn: pmap.proc_table)
-        {
-            cli->proc_table.insert(fn);
-        }
-        rpc_clients[nfs_fd] = cli;
         epmgr->tfd->set_fd_handler(nfs_fd, true, [cli](int nfs_fd, int epoll_events)
         {
             // Handle incoming event
@@ -781,7 +784,7 @@ void nfs_client_t::stop()
     if (refs <= 0)
     {
         auto parent = this->parent;
-        parent->rpc_clients.erase(nfs_fd);
+        parent->rpc_clients.erase(this);
         parent->active_connections--;
         parent->epmgr->tfd->set_fd_handler(nfs_fd, true, NULL);
         close(nfs_fd);
@@ -1085,8 +1088,8 @@ void nfs_proxy_t::daemonize()
     // Stop all clients because client I/O sometimes breaks during daemonize
     // I.e. the new process stops receiving events on the old FD
     // It doesn't happen if we call sleep(1) here, but we don't want to call sleep(1)...
-    for (auto & clp: rpc_clients)
-        clp.second->stop();
+    for (auto & cli: rpc_clients)
+        cli->stop();
     if (fork())
         exit(0);
     setsid();

src/nfs/nfs_proxy.h

@@ -55,7 +55,7 @@ public:
     vitastorkv_dbw_t *db = NULL;
     kv_fs_state_t *kvfs = NULL;
     block_fs_state_t *blockfs = NULL;
-    std::map<int, nfs_client_t*> rpc_clients;
+    std::set<nfs_client_t*> rpc_clients;
 
     std::vector<XDR*> xdr_pool;
 
@@ -72,6 +72,7 @@ public:
     void watch_stats();
     void parse_stats(etcd_kv_t & kv);
     void check_default_pool();
+    nfs_client_t* create_client();
     void do_accept(int listen_fd);
     void daemonize();
     void write_pid();
@@ -101,6 +102,8 @@ struct rpc_free_buffer_t
     unsigned size;
 };
 
+struct nfs_rdma_conn_t;
+
 class nfs_client_t
 {
 public:
@@ -108,6 +111,7 @@ public:
     int refs = 0;
     bool stopped = false;
     std::set<rpc_service_proc_t> proc_table;
+    nfs_rdma_conn_t *rdma_conn = NULL;
 
     // <TCP>
     int nfs_fd;

src/nfs/nfs_proxy_rdma.cpp

@@ -45,6 +45,7 @@ struct nfs_rdma_context_t
     int max_send = 8, max_recv = 8; --- FIXME max_send and max_recv should probably be equal
     uint64_t max_send_size = 256*1024, max_recv_size = 256*1024;
 
+    nfs_proxy_t *proxy = NULL;
     epoll_manager_t *epmgr = NULL;
 
     int max_cqe = 0, used_max_cqe = 0;
@@ -76,6 +77,7 @@ struct nfs_rdma_conn_t
 nfs_rdma_context_t* nfs_proxy_t::create_rdma(const std::string & bind_address, int rdmacm_port)
 {
     nfs_rdma_context_t* self = new nfs_rdma_context_t;
+    self->proxy = this;
     self->epmgr = epmgr;
     self->bind_address = bind_address;
     self->rdmacm_port = rdmacm_port;
@@ -301,6 +303,8 @@ void nfs_rdma_context_t::rdmacm_accept(rdma_cm_event *ev)
         conn->id = ctx->id;
         rdma_connections[ctx->id] = conn;
         rdma_connections_by_qp[conn->id->qp->qp_num];
+        auto cli = this->proxy->create_client();
+        cli->rdma_conn = conn;
     }
 }
 

src/osd/osd.h

@@ -226,6 +226,7 @@ class osd_t
     void parse_config(bool init);
     void init_cluster();
     void on_change_osd_state_hook(osd_num_t peer_osd);
+    void on_change_backfillfull_hook(pool_id_t pool_id);
     void on_change_pg_history_hook(pool_id_t pool_id, pg_num_t pg_num);
     void on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes);
     void on_load_config_hook(json11::Json::object & changes);

src/osd/osd_cluster.cpp

@@ -65,6 +65,7 @@ void osd_t::init_cluster()
         st_cli.tfd = tfd;
         st_cli.log_level = log_level;
         st_cli.on_change_osd_state_hook = [this](osd_num_t peer_osd) { on_change_osd_state_hook(peer_osd); };
+        st_cli.on_change_backfillfull_hook = [this](pool_id_t pool_id) { on_change_backfillfull_hook(pool_id); };
         st_cli.on_change_pg_history_hook = [this](pool_id_t pool_id, pg_num_t pg_num) { on_change_pg_history_hook(pool_id, pg_num); };
         st_cli.on_change_hook = [this](std::map<std::string, etcd_kv_t> & changes) { on_change_etcd_state_hook(changes); };
         st_cli.on_load_config_hook = [this](json11::Json::object & cfg) { on_load_config_hook(cfg); };
@@ -414,6 +415,14 @@ void osd_t::on_change_osd_state_hook(osd_num_t peer_osd)
     }
 }
 
+void osd_t::on_change_backfillfull_hook(pool_id_t pool_id)
+{
+    if (!(peering_state & (OSD_RECOVERING | OSD_FLUSHING_PGS)))
+    {
+        peering_state = peering_state | OSD_RECOVERING;
+    }
+}
+
 void osd_t::on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes)
 {
     if (changes.find(st_cli.etcd_prefix+"/config/global") != changes.end())

src/osd/osd_flush.cpp

@@ -252,10 +252,18 @@ bool osd_t::pick_next_recovery(osd_recovery_op_t &op)
             auto mask = recovery_last_degraded ? (PG_ACTIVE | PG_HAS_DEGRADED) : (PG_ACTIVE | PG_DEGRADED | PG_HAS_MISPLACED);
             auto check = recovery_last_degraded ? (PG_ACTIVE | PG_HAS_DEGRADED) : (PG_ACTIVE | PG_HAS_MISPLACED);
             // Restart scanning from the same PG as the last time
+        restart:
             for (auto pg_it = pgs.lower_bound(recovery_last_pg); pg_it != pgs.end(); pg_it++)
             {
                 if ((pg_it->second.state & mask) == check)
                 {
+                    auto pool_it = st_cli.pool_config.find(pg_it->first.pool_id);
+                    if (pool_it != st_cli.pool_config.end() && pool_it->second.backfillfull)
+                    {
+                        // Skip the pool
+                        recovery_last_pg.pool_id++;
+                        goto restart;
+                    }
                     auto & src = recovery_last_degraded ? pg_it->second.degraded_objects : pg_it->second.misplaced_objects;
                     assert(src.size() > 0);
                     // Restart scanning from the next object

tests/run_tests.sh

@@ -19,9 +19,12 @@ ANTIETCD=1 ./test_etcd_fail.sh
 
 ./test_interrupted_rebalance.sh
 IMMEDIATE_COMMIT=1 ./test_interrupted_rebalance.sh
+
 SCHEME=ec ./test_interrupted_rebalance.sh
 SCHEME=ec IMMEDIATE_COMMIT=1 ./test_interrupted_rebalance.sh
 
+./test_create_halfhost.sh
+
 ./test_failure_domain.sh
 
 ./test_snapshot.sh

tests/test_create_halfhost.sh

@@ -0,0 +1,35 @@
+#!/bin/bash -ex
+
+. `dirname $0`/common.sh
+
+node mon/mon-main.js $MON_PARAMS --etcd_address $ETCD_URL --etcd_prefix "/vitastor" >>./testdata/mon.log 2>&1 &
+MON_PID=$!
+wait_etcd
+
+TIME=$(date '+%s')
+$ETCDCTL put /vitastor/config/global '{"placement_levels":{"dc":10,"host":100,"half":105,"osd":110}}'
+$ETCDCTL put /vitastor/config/node_placement '{
+  "h11":{"level":"half","parent":"host1"},
+  "h12":{"level":"half","parent":"host1"},
+  "h21":{"level":"half","parent":"host2"},
+  "h22":{"level":"half","parent":"host2"},
+  "h31":{"level":"half","parent":"host3"},
+  "h32":{"level":"half","parent":"host3"},
+  "1":{"parent":"h11"},
+  "2":{"parent":"h12"},
+  "3":{"parent":"h21"},
+  "4":{"parent":"h22"},
+  "5":{"parent":"h31"},
+  "6":{"parent":"h32"}
+}'
+$ETCDCTL put /vitastor/osd/stats/1 '{"host":"host1","size":1073741824,"time":"'$TIME'"}'
+$ETCDCTL put /vitastor/osd/stats/2 '{"host":"host1","size":1073741824,"time":"'$TIME'"}'
+$ETCDCTL put /vitastor/osd/stats/3 '{"host":"host2","size":1073741824,"time":"'$TIME'"}'
+$ETCDCTL put /vitastor/osd/stats/4 '{"host":"host2","size":1073741824,"time":"'$TIME'"}'
+$ETCDCTL put /vitastor/osd/stats/5 '{"host":"host3","size":1073741824,"time":"'$TIME'"}'
+$ETCDCTL put /vitastor/osd/stats/6 '{"host":"host3","size":1073741824,"time":"'$TIME'"}'
+build/src/cmd/vitastor-cli --etcd_address $ETCD_URL osd-tree
+# check that it doesn't fail
+build/src/cmd/vitastor-cli --etcd_address $ETCD_URL create-pool testpool --ec 2+1 -n 32
+
+format_green OK