]> git-server-git.apps.pok.os.sepia.ceph.com Git - ceph.git/commitdiff
librbd/cache/pwl/ssd: avoid corrupting m_first_free_entry
authorIlya Dryomov <idryomov@gmail.com>
Fri, 21 May 2021 13:27:31 +0000 (15:27 +0200)
committerDeepika Upadhyay <dupadhya@redhat.com>
Fri, 5 Nov 2021 09:22:02 +0000 (14:52 +0530)
In append_op_log_entries(), new_first_free_entry is read after
append_ops() returns.  This can result in accessing freed memory
because all I/Os may complete and append_ctx callback may run
by the time new_first_free_entry is read.  Garbage value gets
written to m_first_free_entry and depending on the circumstances
it may allow AbstractWriteLog code to accept more dirty user data
than we have space for.  Luckily we usually crash before then.

Fixes: https://tracker.ceph.com/issues/50832
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
(cherry picked from commit d83a0f6db8ff26eeb2c817b1bd192fb357f715df)

src/librbd/cache/pwl/ssd/WriteLog.cc

index 34a724b62ed9c8bf763fced8ae67056117064d60..56f73941f36c204cbe1424e0b4b9469cff1be21b 100644 (file)
@@ -845,6 +845,12 @@ void WriteLog<I>::append_ops(GenericLogOperations &ops, Context *ctx,
     write_log_entries(log_entries, aio, new_first_free_entry);
   }
 
+  {
+    std::lock_guard locker1(m_lock);
+    m_first_free_entry = *new_first_free_entry;
+    m_bytes_allocated -= bytes_to_free;
+  }
+
   bdev->aio_submit(&aio->ioc);
 }