Il 02/07/2014 13:57, ChenLiang ha scritto:
>>> Hmm, dbs->in_cancel will be true always. Although this will avoid freeing
>>> dbs by dma_comlete.
>>> But it maybe a mistake.
>>
>> This was on purpose; I'm doing the free myself in dma_aio_cancel, so I
>> wanted to avoid the qemu_aio_release from dma_complete. This was in case of
>> a recursive call to dma_complete. But I don't see how that recursive call
>> could happen outside the "if (dbs->acb)"; and inside the "if" the protection
>> is there already.
>>
>> Can you gather the backtraces for _both_ calls to qemu_aio_release, rather
>> than just the second?
>
> (gdb) bt
> #0 qemu_aio_release (p=0x7f44788d1290) at block.c:4260
> #1 0x00007f4477494e5e in dma_complete (dbs=0x7f44788d1290, ret=0) at
> dma-helpers.c:135
> #2 0x00007f44774952c2 in dma_aio_cancel (acb=0x7f44788d1290) at
> dma-helpers.c:195
> #3 0x00007f447744825b in bdrv_aio_cancel (acb=0x7f44788d1290) at block.c:3848
> #4 0x00007f4477513911 in ide_bus_reset (bus=0x7f44785f1bd8) at
> hw/ide/core.c:1957
> #5 0x00007f4477516b3c in piix3_reset (opaque=0x7f44785f1530) at
> hw/ide/piix.c:113
> #6 0x00007f4477647b9f in qemu_devices_reset () at vl.c:2131
> #7 0x00007f4477647c0f in qemu_system_reset (report=true) at vl.c:2140
> #8 0x00007f4477648127 in main_loop_should_exit () at vl.c:2274
> #9 0x00007f447764823a in main_loop () at vl.c:2323
> #10 0x00007f447764f6da in main (argc=57, argv=0x7fff5d194378,
> envp=0x7fff5d194548) at vl.c:4803
And the second is
#7 0x00007f3cb525de5e in dma_complete (dbs=0x7f3cb63f3220, ret=0) at
dma-helpers.c:135
#8 0x00007f3cb525df3d in dma_bdrv_cb (opaque=0x7f3cb63f3220, ret=0) at
dma-helpers.c:152
#9 0x00007f3cb5212102 in bdrv_co_em_bh (opaque=0x7f3cb6398980) at block.c:4127
#10 0x00007f3cb51f6cef in aio_bh_poll (ctx=0x7f3cb622a8f0) at async.c:70
#11 0x00007f3cb51f695a in aio_poll (ctx=0x7f3cb622a8f0, blocking=false) at
aio-posix.c:185
#12 0x00007f3cb51f7056 in aio_ctx_dispatch (source=0x7f3cb622a8f0,
callback=0x0, user_data=0x0)
at async.c:167
#13 0x00007f3cb48b969a in g_main_context_dispatch () from
/usr/lib64/libglib-2.0.so.0
This explains why my patch "fixes" the bug. It turns a double free
into a dangling pointer: the second call now sees in_cancel == true
and skips the free.
The second call should have happened within dma_aio_cancel's call to
bdrv_aio_cancel. This is the real bug.
What is your version of QEMU? I cannot see any where bdrv_co_em_bh is
at line 4127 or bdrv_aio_cancel is at line 3848. Can you reproduce it
with qemu.git master?
Paolo
