Re: [Qemu-devel] [PATCH] aio: strengthen memory barriers for bottom half scheduling

[Leveille, Paul]

Paolo,

I've applied your patch in place of my prototype patch and, as expected, it's 
working fine. Thanks!

-----Original Message-----
From: Paolo Bonzini [mailto:address@hidden 
Sent: Tuesday, April 07, 2015 11:16 AM
To: address@hidden
Cc: address@hidden; Leveille, Paul; address@hidden
Subject: [PATCH] aio: strengthen memory barriers for bottom half scheduling

There are two problems with memory barriers in async.c.  The fix is to use 
atomic_xchg in order to achieve sequential consistency between the scheduling 
of a bottom half and the corresponding execution.

First, if bh->scheduled is already 1 in qemu_bh_schedule, QEMU does not execute 
a memory barrier to order any writes needed by the callback before the read of 
bh->scheduled.  If the other side sees req->state as THREAD_ACTIVE, the 
callback is not invoked and you get deadlock.

Second, the memory barrier in aio_bh_poll is too weak.  Without this patch, it 
is possible that bh->scheduled = 0 is not "published" until after the callback 
has returned.  Another thread wants to schedule the bottom half, but it sees 
bh->scheduled = 1 and does nothing.  This causes a lost wakeup.  The memory 
barrier should have been changed to smp_mb() in commit 924fe12 (aio: fix 
qemu_bh_schedule() bh->ctx race condition,
2014-06-03) together with qemu_bh_schedule()'s.  Guess who reviewed that patch?

Both of these involve a store and a load, so they are reproducible on
x86_64 as well.  It is however much easier on aarch64, where the libguestfs 
test suite triggers the bug fairly easily.  Even there the failure can go away 
or appear depending on compiler optimization level, tracing options, or even 
kernel debugging options.

Paul Leveille however reported how to trigger the problem within 15 minutes on 
x86_64 as well.  His (untested) recipe, reproduced here for reference, is the 
following:

   1) Qcow2 (or 3) is critical – raw files alone seem to avoid the problem.

   2) Use “cache=directsync” rather than the default of
   “cache=none” to make it happen easier.

   3) Use a server with a write-back RAID controller to allow for rapid
   IO rates.

   4) Run a random-access load that (mostly) writes chunks to various
   files on the virtual block device.

      a. I use ‘diskload.exe c:25’, a Microsoft HCT load
         generator, on Windows VMs.

      b. Iometer can probably be configured to generate a similar load.

   5) Run multiple VMs in parallel, against the same storage device,
   to shake the failure out sooner.

   6) IvyBridge and Haswell processors for certain; not sure about others.

A similar patch survived over 12 hours of testing, where an unpatched QEMU 
would fail within 15 minutes.

This bug is, most likely, also the cause of failures in the libguestfs 
testsuite on AArch64.

Thanks to Laszlo Ersek for initially reporting this bug, to Stefan Hajnoczi for 
suggesting closer examination of qemu_bh_schedule, and to Paul for providing 
test input and a prototype patch.

Reported-by: Laszlo Ersek <address@hidden>
Reported-by: Paul Leveille <address@hidden>
Reported-by: John Snow <address@hidden>
Suggested-by: Paul Leveille <address@hidden>
Suggested-by: Stefan Hajnoczi <address@hidden>
Signed-off-by: Paolo Bonzini <address@hidden>
---
        Not yet tested on AArch64, will do it tomorrow.  Paul, it would
        be great if you could test this patch too!

 async.c | 28 ++++++++++++----------------
 1 file changed, 12 insertions(+), 16 deletions(-)

diff --git a/async.c b/async.c
index 2be88cc..2b51e87 100644
--- a/async.c
+++ b/async.c
@@ -72,12 +72,13 @@ int aio_bh_poll(AioContext *ctx)
         /* Make sure that fetching bh happens before accessing its members */
         smp_read_barrier_depends();
         next = bh->next;
-        if (!bh->deleted && bh->scheduled) {
-            bh->scheduled = 0;
-            /* Paired with write barrier in bh schedule to ensure reading for
-             * idle & callbacks coming after bh's scheduling.
-             */
-            smp_rmb();
+        /* The atomic_xchg is paired with the one in qemu_bh_schedule.  The
+         * implicit memory barrier ensures that the callback sees all writes
+         * done by the scheduling thread.  It also ensures that the scheduling
+         * thread sees the zero before bh->cb has run, and thus will call
+         * aio_notify again if necessary.
+         */
+        if (!bh->deleted && atomic_xchg(&bh->scheduled, 0)) {
             if (!bh->idle)
                 ret = 1;
             bh->idle = 0;
@@ -108,33 +109,28 @@ int aio_bh_poll(AioContext *ctx)
 
 void qemu_bh_schedule_idle(QEMUBH *bh)
 {
-    if (bh->scheduled)
-        return;
     bh->idle = 1;
     /* Make sure that idle & any writes needed by the callback are done
      * before the locations are read in the aio_bh_poll.
      */
-    smp_wmb();
-    bh->scheduled = 1;
+    atomic_mb_set(&bh->scheduled, 1);
 }
 
 void qemu_bh_schedule(QEMUBH *bh)
 {
     AioContext *ctx;
 
-    if (bh->scheduled)
-        return;
     ctx = bh->ctx;
     bh->idle = 0;
-    /* Make sure that:
+    /* The memory barrier implicit in atomic_xchg makes sure that:
      * 1. idle & any writes needed by the callback are done before the
      *    locations are read in the aio_bh_poll.
      * 2. ctx is loaded before scheduled is set and the callback has a chance
      *    to execute.
      */
-    smp_mb();
-    bh->scheduled = 1;
-    aio_notify(ctx);
+    if (atomic_xchg(&bh->scheduled, 1) == 0) {
+        aio_notify(ctx);
+    }
 }
 
 
--
2.3.4