]> git.proxmox.com Git - mirror_qemu.git/commitdiff
util/async: use atomic_mb_set in qemu_bh_cancel
authorSergio Lopez <slp@redhat.com>
Wed, 8 Nov 2017 06:34:47 +0000 (07:34 +0100)
committerStefan Hajnoczi <stefanha@redhat.com>
Wed, 8 Nov 2017 19:09:15 +0000 (19:09 +0000)
Commit b7a745d added a qemu_bh_cancel call to the completion function
as an optimization to prevent it from unnecessarily rescheduling itself.

This completion function is scheduled from worker_thread, after setting
the state of a ThreadPoolElement to THREAD_DONE.

This was considered to be safe, as the completion function restarts the
loop just after the call to qemu_bh_cancel. But, as this loop lacks a HW
memory barrier, the read of req->state may actually happen _before_ the
call, seeing it still as THREAD_QUEUED, and ending the completion
function without having processed a pending TPE linked at pool->head:

         worker thread             |            I/O thread
------------------------------------------------------------------------
                                   | speculatively read req->state
req->state = THREAD_DONE;          |
qemu_bh_schedule(p->completion_bh) |
  bh->scheduled = 1;               |
                                   | qemu_bh_cancel(p->completion_bh)
                                   |   bh->scheduled = 0;
                                   | if (req->state == THREAD_DONE)
                                   |   // sees THREAD_QUEUED

The source of the misunderstanding was that qemu_bh_cancel is now being
used by the _consumer_ rather than the producer, and therefore now needs
to have acquire semantics just like e.g. aio_bh_poll.

In some situations, if there are no other independent requests in the
same aio context that could eventually trigger the scheduling of the
completion function, the omitted TPE and all operations pending on it
will get stuck forever.

[Added Sergio's updated wording about the HW memory barrier.
--Stefan]

Signed-off-by: Sergio Lopez <slp@redhat.com>
Message-id: 20171108063447.2842-1-slp@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
util/async.c

index 355af73ee71b078215e18552fbdcd0ee8c325370..0e1bd8780ae323280bbe4caf5c29d48002b2c3c0 100644 (file)
@@ -174,7 +174,7 @@ void qemu_bh_schedule(QEMUBH *bh)
  */
 void qemu_bh_cancel(QEMUBH *bh)
 {
-    bh->scheduled = 0;
+    atomic_mb_set(&bh->scheduled, 0);
 }
 
 /* This func is async.The bottom half will do the delete action at the finial