From: Adam Kupczyk <akupczyk@redhat.com>
Date: Sun, 27 Nov 2022 11:44:10 +0000 (+0000)
Subject: os/bluestore: Refactor, add _dup_writing when cloning
X-Git-Tag: v19.0.0~486^2~24
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=6213a94a83899bdc90d8aecaab410ba03b10b6eb;p=ceph.git

os/bluestore: Refactor, add _dup_writing when cloning

With BufferSpace now attached to Blob (was SharedBlob inside it),
on-the-fly 'writing' buffers must be copied to clones.
Otherwise those objects will read data from disk before it is written there.

Signed-off-by: Adam Kupczyk <akupczyk@redhat.com>
---

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 58fd3233ba29..0378dbf555f0 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -1890,6 +1890,27 @@ void BlueStore::BufferSpace::_finish_write(BufferCacheShard* cache, uint64_t seq
   cache->_audit("finish_write end");
 }
 
+/*
+  copy Buffers that are in writing queue
+  returns:
+  true  if something copied
+  false if nothing copied
+*/
+bool BlueStore::BufferSpace::_dup_writing(BufferCacheShard* cache, BufferSpace* to)
+{
+  bool copied = false;
+  if (!writing.empty()) {
+    copied = true;
+    for (auto it = writing.begin(); it != writing.end(); ++it) {
+      Buffer& b = *it;
+      Buffer* to_b = new Buffer(to, b.state, b.seq, b.offset, b.data, b.flags);
+      ceph_assert(to_b->is_writing());
+      to->_add_buffer(cache, to_b, 0, nullptr);
+    }
+  }
+  return copied;
+}
+
 void BlueStore::BufferSpace::split(BufferCacheShard* cache, size_t pos, BlueStore::BufferSpace &r)
 {
   std::lock_guard lk(cache->lock);
@@ -2580,6 +2601,16 @@ void BlueStore::ExtentMap::dup(BlueStore* b, TransContext* txc,
       e.blob->last_encoded_id = n;
       id_to_blob[n] = cb;
       e.blob->dup(*cb);
+      // By default do not copy buffers to clones, and let them read data by themselves.
+      // The exception are 'writing' buffers, which are not yet stable on device.
+      bool some_copied = e.blob->bc._dup_writing(cb->shared_blob->get_cache(), &cb->bc);
+      if (some_copied) {
+	// Pretend we just wrote those buffers;
+	// we need to get _finish_write called, so we can clear then from writing list.
+	// Otherwise it will be stuck until someone does write-op on clone.
+	txc->blobs_written.insert(cb);
+      }
+
       // bump the extent refs on the copied blob's extents
       for (auto p : blob.get_extents()) {
         if (p.is_valid()) {
diff --git a/src/os/bluestore/BlueStore.h b/src/os/bluestore/BlueStore.h
index b0449451c15f..24b5cd143b86 100644
--- a/src/os/bluestore/BlueStore.h
+++ b/src/os/bluestore/BlueStore.h
@@ -464,6 +464,7 @@ public:
       discard(cache, offset, (uint32_t)-1 - offset);
     }
 
+    bool _dup_writing(BufferCacheShard* cache, BufferSpace* bc);
     void split(BufferCacheShard* cache, size_t pos, BufferSpace &r);
 
     void dump(BufferCacheShard* cache, ceph::Formatter *f) const {