Re: [Qemu-devel] question: a dead loop in qemu when do blockJobAbort and vm suspend coinstantaneously

[Fam Zheng <famz@redhat.com>]

On Sat, 06/09 17:10, l00284672 wrote:
> Hi, I found a dead loop in qemu when do blockJobAbort and vm suspend
> coinstantaneously.
> 
> The qemu bt is below:
> 
> #0  0x00007ff58b53af1f in ppoll () from /lib64/libc.so.6
> #1  0x00000000007fdbd9 in ppoll (__ss=0x0, __timeout=0x7ffcf7055390,
> __nfds=<optimized out>, __fds=<optimized out>) at
> /usr/include/bits/poll2.h:77
> #2  qemu_poll_ns (fds=<optimized out>, nfds=<optimized out>,
> timeout=timeout@entry=0) at util/qemu-timer.c:334
> #3  0x00000000007ff83a in aio_poll (ctx=ctx@entry=0x269d800,
> blocking=blocking@entry=false) at util/aio-posix.c:629
> #4  0x0000000000776e91 in bdrv_drain_recurse (bs=bs@entry=0x26d9cb0) at
> block/io.c:198
> #5  0x0000000000776ef2 in bdrv_drain_recurse (bs=bs@entry=0x3665990) at
> block/io.c:215
> #6  0x00000000007774b8 in bdrv_do_drained_begin (bs=0x3665990,
> recursive=<optimized out>, parent=0x0) at block/io.c:291
> #7  0x000000000076a79e in blk_drain (blk=0x2780fc0) at
> block/block-backend.c:1586
> #8  0x000000000072d2a9 in block_job_drain (job=0x29df040) at blockjob.c:123
> #9  0x000000000072d228 in block_job_detach_aio_context (opaque=0x29df040) at
> blockjob.c:139
> #10 0x00000000007298b1 in bdrv_detach_aio_context (bs=bs@entry=0x3665990) at
> block.c:4885
> #11 0x0000000000729a46 in bdrv_set_aio_context (bs=0x3665990,
> new_context=0x268e140) at block.c:4946
> #12 0x0000000000499743 in virtio_blk_data_plane_stop (vdev=<optimized out>)
> at
> /mnt/sdb/lzg/code/shequ_code/5_29/qemu/hw/block/dataplane/virtio-blk.c:285
> #13 0x00000000006bce30 in virtio_bus_stop_ioeventfd (bus=0x3de5378) at
> hw/virtio/virtio-bus.c:246
> #14 0x00000000004c654d in virtio_vmstate_change (opaque=0x3de53f0,
> running=<optimized out>, state=<optimized out>)
>     at /mnt/sdb/lzg/code/shequ_code/5_29/qemu/hw/virtio/virtio.c:2222
> #15 0x0000000000561b52 in vm_state_notify (running=running@entry=0,
> state=state@entry=RUN_STATE_PAUSED) at vl.c:1514
> #16 0x000000000045d67a in do_vm_stop (state=state@entry=RUN_STATE_PAUSED,
> send_stop=send_stop@entry=true)
>     at /mnt/sdb/lzg/code/shequ_code/5_29/qemu/cpus.c:1012
> #17 0x000000000045dafd in vm_stop (state=state@entry=RUN_STATE_PAUSED) at
> /mnt/sdb/lzg/code/shequ_code/5_29/qemu/cpus.c:2035
> #18 0x000000000057301b in qmp_stop (errp=errp@entry=0x7ffcf70556f0) at
> qmp.c:106
> #19 0x000000000056bf7a in qmp_marshal_stop (args=<optimized out>,
> ret=<optimized out>, errp=0x7ffcf7055738) at qapi/qapi-commands-misc.c:784
> #20 0x00000000007f2d27 in do_qmp_dispatch (errp=0x7ffcf7055730,
> request=0x3e121e0, cmds=<optimized out>) at qapi/qmp-dispatch.c:119
> #21 qmp_dispatch (cmds=<optimized out>, request=request@entry=0x26f2800) at
> qapi/qmp-dispatch.c:168
> #22 0x00000000004655be in monitor_qmp_dispatch_one
> (req_obj=req_obj@entry=0x39abff0) at
> /mnt/sdb/lzg/code/shequ_code/5_29/qemu/monitor.c:4088
> #23 0x0000000000465894 in monitor_qmp_bh_dispatcher (data=<optimized out>)
> at /mnt/sdb/lzg/code/shequ_code/5_29/qemu/monitor.c:4146
> #24 0x00000000007fc571 in aio_bh_call (bh=0x26de7e0) at util/async.c:90
> #25 aio_bh_poll (ctx=ctx@entry=0x268dd50) at util/async.c:118
> #26 0x00000000007ff6f0 in aio_dispatch (ctx=0x268dd50) at
> util/aio-posix.c:436
> #27 0x00000000007fc44e in aio_ctx_dispatch (source=<optimized out>,
> callback=<optimized out>, user_data=<optimized out>) at util/async.c:261
> #28 0x00007ff58bc7c99a in g_main_context_dispatch () from
> /lib64/libglib-2.0.so.0
> #29 0x00000000007fea3a in glib_pollfds_poll () at util/main-loop.c:215
> #30 os_host_main_loop_wait (timeout=<optimized out>) at util/main-loop.c:238
> #31 main_loop_wait (nonblocking=nonblocking@entry=0) at util/main-loop.c:497
> #32 0x0000000000561cad in main_loop () at vl.c:1848
> #33 0x000000000041995c in main (argc=<optimized out>, argv=<optimized out>,
> envp=<optimized out>) at vl.c:4605
> 
> The disk is a virtio-blk dataplane disk with a mirror job running.  The dead
> loop is here:
> 
> static void block_job_detach_aio_context(void *opaque)
> {
>     BlockJob *job = opaque;
> 
>     /* In case the job terminates during aio_poll()... */
>     job_ref(&job->job);
> 
>     job_pause(&job->job);
> 
>   while (!job->job.paused && !job_is_completed(&job->job)) {
>         job_drain(&job->job);
>     }
> 
>     job->job.aio_context = NULL;
>     job_unref(&job->job);
> }
> 
> The job is deferred to main loop now,  but the  job_drain only processes the
> AIO context of bs which has no more work to do,
> 
> while the main loop BH is scheduled for setting the job->completed flag is
> never processed.

In that case, main loop's AioContext should be driven like in job_finish_sync().
Could you try this patch?

diff --git a/blockjob.c b/blockjob.c
index 0306533a2e..72aa82ac2d 100644
--- a/blockjob.c
+++ b/blockjob.c
@@ -135,7 +135,15 @@ static void block_job_detach_aio_context(void *opaque)

     job_pause(&job->job);

-    while (!job->job.paused && !job_is_completed(&job->job)) {
+
+    while (true) {
+        if (job->job.paused || job_is_completed(&job->job)) {
+            break;
+        }
+        if (job->deferred_to_main_loop) {
+            aio_poll(qemu_get_aio_context(), true);
+            continue;
+        }
         job_drain(&job->job);
     }