From: Igor Fedotov <ifedotov@mirantis.com>
Date: Thu, 7 Jul 2016 17:50:56 +0000 (+0300)
Subject: os/bluestore: collect 'dirty' blobs on per-transaction basis to avoid onode::blob_map... 
X-Git-Tag: ses5-milestone5~146^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=refs%2Fpull%2F10215%2Fhead;p=ceph.git

os/bluestore: collect 'dirty' blobs on per-transaction basis to avoid onode::blob_map enumeration

Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
---

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 2ffbbf189cb0..6e7c137e7063 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -4727,20 +4727,10 @@ void BlueStore::_txc_state_proc(TransContext *txc)
       //assert(txc->osr->qlock.is_locked());  // see _txc_finish_io
       txc->log_state_latency(logger, l_bluestore_state_io_done_lat);
       txc->state = TransContext::STATE_KV_QUEUED;
-      // FIXME: use a per-txc dirty blob list?
-      
-      if (txc->first_collection) {
-        (txc->first_collection)->lock.get_read();
+      for (auto& b : txc->blobs) {
+        b->bc.finish_write(txc->seq);
       }
-      for (auto& o : txc->onodes) {
-        for (auto& p : o->blob_map.blob_map) {
-	    p.bc.finish_write(txc->seq);
-	}
-      }
-      if (txc->first_collection) {
-        (txc->first_collection)->lock.put_read();
-      }
-      
+      txc->blobs.clear();
       if (!g_conf->bluestore_sync_transaction) {
 	if (g_conf->bluestore_sync_submit_transaction) {
 	  _txc_finalize_kv(txc, txc->t);
@@ -5939,8 +5929,8 @@ void BlueStore::_do_write_small(
 	       << " pad 0x" << head_pad << " + 0x" << tail_pad
 	       << std::dec << " of mutable " << blob << ": " << b << dendl;
       assert(b->blob.is_unreferenced(b_off, b_len));
-      b->bc.write(txc->seq, b_off, padded,
-		  wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+      _buffer_cache_write(txc, b, b_off, padded, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+
       b->blob.map_bl(
 	b_off, padded,
 	[&](uint64_t offset, uint64_t length, bufferlist& t) {
@@ -6007,8 +5997,8 @@ void BlueStore::_do_write_small(
 	b->blob.is_allocated(b_off, b_len)) {
       bluestore_wal_op_t *op = _get_wal_op(txc, o);
       op->op = bluestore_wal_op_t::OP_WRITE;
-      b->bc.write(txc->seq, b_off, padded,
-		  wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+      _buffer_cache_write(txc, b, b_off, padded, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+
       b->blob.map(
 	b_off, b_len,
 	[&](uint64_t offset, uint64_t length) {
@@ -6038,7 +6028,7 @@ void BlueStore::_do_write_small(
   b = o->blob_map.new_blob(c->cache);
   unsigned alloc_len = min_alloc_size;
   uint64_t b_off = P2PHASE(offset, alloc_len);
-  b->bc.write(txc->seq, b_off, bl, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+  _buffer_cache_write(txc, b, b_off, bl, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
   _pad_zeros(&bl, &b_off, block_size);
   bluestore_lextent_t lex(b->id, P2PHASE(offset, alloc_len), length);
   o->onode.set_lextent(offset, lex, &b->blob, &wctx->lex_old);
@@ -6071,7 +6061,7 @@ void BlueStore::_do_write_big(
     auto l = MIN(max_blob_len, length);
     bufferlist t;
     blp.copy(l, t);
-    b->bc.write(txc->seq, 0, t, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
+    _buffer_cache_write(txc, b, 0, t, wctx->buffered ? 0 : Buffer::FLAG_NOCACHE);
     wctx->write(b, l, 0, t, false);
     bluestore_lextent_t lex(b->id, 0, l);
     o->onode.set_lextent(offset, lex, &b->blob, &wctx->lex_old);
diff --git a/src/os/bluestore/BlueStore.h b/src/os/bluestore/BlueStore.h
index c51f29b6d42d..91457ceeef95 100644
--- a/src/os/bluestore/BlueStore.h
+++ b/src/os/bluestore/BlueStore.h
@@ -820,6 +820,8 @@ public:
 
     set<OnodeRef> onodes;     ///< these onodes need to be updated/written
     set<BnodeRef> bnodes;     ///< these bnodes need to be updated/written
+    set<BlobRef> blobs;       ///< these blobs need to be updated on io completion
+
     KeyValueDB::Transaction t; ///< then we will commit this
     Context *oncommit;         ///< signal on commit
     Context *onreadable;         ///< signal on readable
@@ -1272,6 +1274,15 @@ private:
     boost::dynamic_bitset<> &used_blocks,
     store_statfs_t& expected_statfs);
 
+  void _buffer_cache_write(
+    TransContext *txc,
+    BlobRef b,
+    uint64_t offset,
+    bufferlist& bl,
+    unsigned flags) {
+    b->bc.write(txc->seq, offset, bl, flags);
+    txc->blobs.insert(b);
+  }
 public:
   BlueStore(CephContext *cct, const string& path);
   ~BlueStore();