Re: [Qemu-devel] [PATCH] test-replication: Lock AioContext around blk_unref()

[Kevin Wolf]

Am 01.10.2018 um 17:40 hat Paolo Bonzini geschrieben:
> On 01/10/2018 16:32, Kevin Wolf wrote:
> > Recently, the test case has started failing because some job related
> > functions want to drop the AioContext lock even though it hasn't been
> > taken:
> > 
> >     (gdb) bt
> >     #0  0x00007f51c067c9fb in raise () from /lib64/libc.so.6
> >     #1  0x00007f51c067e77d in abort () from /lib64/libc.so.6
> >     #2  0x0000558c9d5dde7b in error_exit (err=<optimized out>, 
> > address@hidden <__func__.18373> "qemu_mutex_unlock_impl") at 
> > util/qemu-thread-posix.c:36
> >     #3  0x0000558c9d6b5263 in qemu_mutex_unlock_impl (address@hidden, 
> > address@hidden "util/async.c", address@hidden) at 
> > util/qemu-thread-posix.c:96
> >     #4  0x0000558c9d6b0565 in aio_context_release (address@hidden) at 
> > util/async.c:516
> >     #5  0x0000558c9d5eb3da in job_completed_txn_abort (job=0x558c9f68e640) 
> > at job.c:738
> >     #6  0x0000558c9d5eb227 in job_finish_sync (job=0x558c9f68e640, 
> > address@hidden <job_cancel_err>, address@hidden) at job.c:986
> >     #7  0x0000558c9d5eb8ee in job_cancel_sync (job=<optimized out>) at 
> > job.c:941
> >     #8  0x0000558c9d64d853 in replication_close (bs=<optimized out>) at 
> > block/replication.c:148
> >     #9  0x0000558c9d5e5c9f in bdrv_close (bs=0x558c9f41b020) at block.c:3420
> >     #10 bdrv_delete (bs=0x558c9f41b020) at block.c:3629
> >     #11 bdrv_unref (bs=0x558c9f41b020) at block.c:4685
> >     #12 0x0000558c9d62a3f3 in blk_remove_bs (address@hidden) at 
> > block/block-backend.c:783
> >     #13 0x0000558c9d62a667 in blk_delete (blk=0x558c9f42a7c0) at 
> > block/block-backend.c:402
> >     #14 blk_unref (blk=0x558c9f42a7c0) at block/block-backend.c:457
> >     #15 0x0000558c9d5dfcea in test_secondary_stop () at 
> > tests/test-replication.c:478
> >     #16 0x00007f51c1f13178 in g_test_run_suite_internal () from 
> > /lib64/libglib-2.0.so.0
> >     #17 0x00007f51c1f1337b in g_test_run_suite_internal () from 
> > /lib64/libglib-2.0.so.0
> >     #18 0x00007f51c1f1337b in g_test_run_suite_internal () from 
> > /lib64/libglib-2.0.so.0
> >     #19 0x00007f51c1f13552 in g_test_run_suite () from 
> > /lib64/libglib-2.0.so.0
> >     #20 0x00007f51c1f13571 in g_test_run () from /lib64/libglib-2.0.so.0
> >     #21 0x0000558c9d5de31f in main (argc=<optimized out>, argv=<optimized 
> > out>) at tests/test-replication.c:581
> > 
> > It is yet unclear whether this should really be considered a bug in the
> > test case or whether blk_unref() should work for callers that haven't
> > taken the AioContext lock, but in order to fix the build tests quickly,
> > just take the AioContext lock around blk_unref().
> 
> Given the backtrace, I think bdrv_close should be taking the AioContext
> lock instead of blockdev_close_all_bdrv_states.

Conversely, that would mean that calling bdrv_unref() with the
AioContext lock held is a bug (because close callbacks can involve
AIO_WAIT_WHILE()). I'm not sure if that's very practical.

Of course, there will probably be a lot of callers to fix either way
after we define whether to hold the lock for bdrv_unref() or not. Either
you need to add locking to the places where it's missing or you need to
drop the locks in all other places.

I was leaning towards requiring the lock for bdrv_unref() (and
therefore blk_unref()).

Kevin