Fix and improve parallel allocation

- Do not try to allocate more DB blocks in an inode block until it's "confirmed" and "locked" by the first write
- Do not recheck for new zero DB blocks on first write into an inode block - a CAS failure means someone else is already writing into it
- Throw new allocation blocks away regardless of whether the known_version is 0 on a CAS failure

src/kv_db.cpp

@@ -123,6 +123,13 @@ struct kv_continue_write_t
     std::function<void(int)> cb;
 };
 
+struct kv_alloc_block_t
+{
+    uint64_t offset;
+    bool writing;
+    bool confirmed;
+};
+
 struct kv_db_t
 {
     cluster_client_t *cli = NULL;
@@ -137,7 +144,7 @@ struct kv_db_t
     uint64_t evict_unused_age = 1000;
     uint64_t evict_max_misses = 10;
     uint64_t evict_attempts_per_level = 3;
-    uint64_t allocate_blocks = 4;
+    uint64_t max_allocate_blocks = 4;
     uint64_t log_level = 1;
 
     // state
@@ -146,7 +153,7 @@ struct kv_db_t
     int base_block_level = 0;
     int usage_counter = 1;
     int allocating_block_pos = 0;
-    std::vector<uint64_t> allocating_blocks;
+    std::vector<kv_alloc_block_t> allocating_blocks;
     std::set<uint64_t> block_levels;
     std::map<uint64_t, kv_block_t> block_cache;
     std::map<uint64_t, uint64_t> known_versions;
@@ -160,6 +167,8 @@ struct kv_db_t
 
     uint64_t alloc_block();
     void clear_allocation_block(uint64_t offset);
+    void confirm_allocation_block(uint64_t offset);
+    void stop_writing_new(uint64_t offset);
 
     void open(inode_t inode_id, json11::Json cfg, std::function<void(int)> cb);
     void set_config(json11::Json cfg);
@@ -539,7 +548,7 @@ void kv_db_t::set_config(json11::Json cfg)
     this->evict_attempts_per_level = cfg["kv_evict_attempts_per_level"].is_null() ? 3 : cfg["kv_evict_attempts_per_level"].uint64_value();
     this->evict_unused_age = cfg["kv_evict_unused_age"].is_null() ? 1000 : cfg["kv_evict_unused_age"].uint64_value();
     this->cache_max_blocks = this->memory_limit / this->kv_block_size;
-    this->allocate_blocks = cfg["kv_allocate_blocks"].uint64_value() ? cfg["kv_allocate_blocks"].uint64_value() : 4;
+    this->max_allocate_blocks = cfg["kv_allocate_blocks"].uint64_value() ? cfg["kv_allocate_blocks"].uint64_value() : 4;
     this->log_level = !cfg["kv_log_level"].is_null() ? cfg["kv_log_level"].uint64_value() : 1;
 }
 
@@ -567,49 +576,96 @@ void kv_db_t::close(std::function<void()> cb)
 
 uint64_t kv_db_t::alloc_block()
 {
-    // round robin between <allocate_blocks> blocks
+    // select from at least <max_allocate_blocks> allocation blocks
-    while (allocating_blocks.size() < allocate_blocks)
+    while (allocating_blocks.size() < max_allocate_blocks)
     {
-        allocating_blocks.push_back(UINT64_MAX);
+        allocating_blocks.push_back({ .offset = UINT64_MAX });
     }
-    if (allocating_blocks[allocating_block_pos] == UINT64_MAX)
+    bool found = false;
+    for (int i = 0; i < allocating_blocks.size(); i++)
     {
-        // try v0 and only v0 of a new inode block
+        int next = (allocating_block_pos+i) % allocating_blocks.size();
-        while (known_versions[next_free/ino_block_size] != 0)
+        if (allocating_blocks[allocating_block_pos].offset != UINT64_MAX &&
+            allocating_blocks[next].confirmed && !allocating_blocks[next].writing)
         {
-            next_free += ino_block_size;
+            allocating_block_pos = next;
+            found = true;
+            break;
         }
-        allocating_blocks[allocating_block_pos] = next_free;
+    }
+    if (!found)
+    {
+        // Allow to allocate more new blocks in parallel than <max_allocate_blocks>
+        allocating_blocks.push_back({ .offset = UINT64_MAX });
+        allocating_block_pos = allocating_blocks.size()-1;
+    }
+    if (allocating_blocks[allocating_block_pos].offset == UINT64_MAX)
+    {
+        // allocate new blocks in the end
+        auto known_max = known_versions.end();
+        while (known_max != known_versions.begin())
+        {
+            known_max--;
+            // try v0 and only v0 of a new inode block
+            if (known_max->second != 0)
+            {
+                auto probably_unused = (known_max->first+1)*ino_block_size;
+                if (next_free < probably_unused)
+                    next_free = probably_unused;
+                break;
+            }
+        }
+        allocating_blocks[allocating_block_pos] = {
+            .offset = next_free,
+            .writing = false,
+            .confirmed = false,
+        };
         next_free += ino_block_size;
     }
-    auto pos = allocating_blocks[allocating_block_pos];
+    auto pos = allocating_blocks[allocating_block_pos].offset;
-    allocating_blocks[allocating_block_pos] += kv_block_size;
+    allocating_blocks[allocating_block_pos].writing = true;
-    if (!(allocating_blocks[allocating_block_pos] % ino_block_size))
+    allocating_blocks[allocating_block_pos].offset += kv_block_size;
+    if (!(allocating_blocks[allocating_block_pos].offset % ino_block_size))
     {
-        // Allow online reconfiguration
+        // Allow to reconfigure <max_allocate_blocks> online
-        if (allocating_blocks.size() > allocate_blocks)
+        if (allocating_blocks.size() > max_allocate_blocks)
             allocating_blocks.erase(allocating_blocks.begin()+allocating_block_pos, allocating_blocks.begin()+allocating_block_pos+1);
         else
-            allocating_blocks[allocating_block_pos] = UINT64_MAX;
+            allocating_blocks[allocating_block_pos].offset = UINT64_MAX;
     }
-    allocating_block_pos = (allocating_block_pos+1) % allocating_blocks.size();
     assert(block_cache.find(pos) == block_cache.end());
     return pos;
 }
 
 void kv_db_t::clear_allocation_block(uint64_t offset)
 {
-    // We want to be first when writing the first KV block into a new inode block
+    for (int i = 0; i < allocating_blocks.size(); i++)
-    // After it other writers may modify already placed blocks and the version
-    // will increase, but they won't use free space in it for new blocks
-    if (known_versions[offset/ino_block_size] == 0)
     {
-        for (int i = 0; i < allocating_blocks.size(); i++)
+        if (allocating_blocks[i].offset/ino_block_size == offset/ino_block_size)
         {
-            if (allocating_blocks[i]/ino_block_size == offset/ino_block_size)
+            allocating_blocks[i].offset = UINT64_MAX;
-            {
+        }
-                allocating_blocks[i] = UINT64_MAX;
+    }
-            }
+}
+
+void kv_db_t::confirm_allocation_block(uint64_t offset)
+{
+    for (int i = 0; i < allocating_blocks.size(); i++)
+    {
+        if (allocating_blocks[i].offset/ino_block_size == offset/ino_block_size)
+        {
+            allocating_blocks[i].confirmed = true;
+        }
+    }
+}
+
+void kv_db_t::stop_writing_new(uint64_t offset)
+{
+    for (int i = 0; i < allocating_blocks.size(); i++)
+    {
+        if (allocating_blocks[i].offset/ino_block_size == offset/ino_block_size)
+        {
+            allocating_blocks[i].writing = false;
         }
     }
 }
@@ -711,7 +767,10 @@ static void invalidate(kv_db_t *db, uint64_t offset, uint64_t version)
 {
     if (db->known_versions[offset/db->ino_block_size] < version)
     {
-        db->clear_allocation_block(offset);
+        if (db->known_versions[offset/db->ino_block_size] == 0)
+        {
+            db->clear_allocation_block(offset);
+        }
         auto b_it = db->block_cache.lower_bound(offset/db->ino_block_size * db->ino_block_size);
         while (b_it != db->block_cache.end() && b_it->first/db->ino_block_size == offset/db->ino_block_size)
         {
@@ -1230,14 +1289,41 @@ static kv_block_t *create_new_block(kv_db_t *db, kv_block_t *old_blk, const std:
     return blk;
 }
 
+static void write_new_block(kv_db_t *db, kv_block_t *blk, std::function<void(int, kv_block_t *)> cb);
+
+static void place_again(kv_db_t *db, kv_block_t *blk, std::function<void(int, kv_block_t *)> cb)
+{
+    auto old_offset = blk->offset;
+    auto new_offset = db->alloc_block();
+    del_block_level(db, blk);
+    std::swap(db->block_cache[new_offset], db->block_cache[old_offset]);
+    db->block_cache.erase(old_offset);
+    auto new_blk = &db->block_cache[new_offset];
+    new_blk->offset = new_offset;
+    new_blk->invalidated = false;
+    add_block_level(db, new_blk);
+    write_new_block(db, new_blk, cb);
+}
+
 static void write_new_block(kv_db_t *db, kv_block_t *blk, std::function<void(int, kv_block_t *)> cb)
 {
     write_block(db, blk, [=](int res)
     {
+        db->stop_writing_new(blk->offset);
         if (res == -EINTR)
         {
             // CAS failure => re-read, then, if not zero, find position again and retry
-            db->clear_allocation_block(blk->offset);
+            if (!(blk->offset % db->ino_block_size))
+            {
+                // Any failed write of the first block within an inode block
+                // means that someone else is already allocating blocks in it,
+                // so we HAVE to move to the next inode block immediately
+                db->clear_allocation_block(blk->offset);
+                place_again(db, blk, cb);
+                return;
+            }
+            // On the other hand, if the block is already "ours", then live parts
+            // of it may change and we MAY recheck if the block is still zero on CAS failure
             cluster_op_t *op = new cluster_op_t;
             op->opcode = OSD_OP_READ;
             op->inode = db->inode_id;
@@ -1264,16 +1350,7 @@ static void write_new_block(kv_db_t *db, kv_block_t *blk, std::function<void(int
                 else
                 {
                     // Block is already occupied => place again
-                    auto old_offset = blk->offset;
+                    place_again(db, blk, cb);
-                    auto new_offset = db->alloc_block();
-                    del_block_level(db, blk);
-                    std::swap(db->block_cache[new_offset], db->block_cache[old_offset]);
-                    db->block_cache.erase(old_offset);
-                    auto new_blk = &db->block_cache[new_offset];
-                    new_blk->offset = new_offset;
-                    new_blk->invalidated = false;
-                    add_block_level(db, new_blk);
-                    write_new_block(db, new_blk, cb);
                 }
             };
             db->cli->execute(op);
@@ -1281,6 +1358,7 @@ static void write_new_block(kv_db_t *db, kv_block_t *blk, std::function<void(int
         else if (res != 0)
         {
             // Other failure => free the new unreferenced block and die
+            db->clear_allocation_block(blk->offset);
             del_block_level(db, blk);
             db->block_cache.erase(blk->offset);
             cb(res > 0 ? -EIO : res, NULL);
@@ -1288,6 +1366,12 @@ static void write_new_block(kv_db_t *db, kv_block_t *blk, std::function<void(int
         else
         {
             // OK
+            if (!(blk->offset % db->ino_block_size))
+            {
+                // A successful first write into a new allocation block
+                // confirms that it's now "locked" by us
+                db->confirm_allocation_block(blk->offset);
+            }
             cb(0, blk);
         }
     });
@@ -1405,6 +1489,7 @@ void kv_op_t::create_root()
     {
         if (res == -EINTR)
         {
+            db->clear_allocation_block(blk->offset);
             auto blk_offset = blk->offset;
             del_block_level(db, blk);
             db->block_cache.erase(blk_offset);
@@ -1413,6 +1498,8 @@ void kv_op_t::create_root()
         }
         else
         {
+            db->stop_writing_new(blk->offset);
+            db->confirm_allocation_block(blk->offset);
             db->stop_updating(blk);
             finish(res);
         }