Am 06.05.2020 um 09:02 hat Vladimir Sementsov-Ogievskiy geschrieben:
> 27.04.2020 17:39, Vladimir Sementsov-Ogievskiy wrote:
> > It's safer to expand in_flight request to start before enter to
> > coroutine in synchronous wrappers, due to the following (theoretical)
> > problem:
> >
> > Consider write.
> > It's possible, that qemu_coroutine_enter only schedules execution,
> > assume such case.
> >
> > Then we may possibly have the following:
> >
> > 1. Somehow check that we are not in drained section in outer code.
> >
> > 2. Call bdrv_pwritev(), assuming that it will increase in_flight, which
> > will protect us from starting drained section.
> >
> > 3. It calls bdrv_prwv_co() -> bdrv_coroutine_enter() (not yet increased
> > in_flight).
> >
> > 4. Assume coroutine not yet actually entered, only scheduled, and we go
> > to some code, which starts drained section (as in_flight is zero).
> >
> > 5. Scheduled coroutine starts, and blindly increases in_flight, and we
> > are in drained section with in_flight request.
> >
> > Signed-off-by: Vladimir Sementsov-Ogievskiy <vsement...@virtuozzo.com>
>
> Very interesting: this patch breaks test-replication. It hangs:
>
> (gdb) thr a a bt
>
> Thread 2 (Thread 0x7eff256cd700 (LWP 2843)):
> #0 0x00007eff2f5fd1fd in syscall () from /lib64/libc.so.6
> #1 0x000055af9a9a4f11 in qemu_futex_wait (f=0x55af9aa6f758
> <rcu_call_ready_event>, val=4294967295) at
> /work/src/qemu/up-expand-bdrv-in_flight-bounds/include/qemu/futex.h:29
> #2 0x000055af9a9a50d5 in qemu_event_wait (ev=0x55af9aa6f758
> <rcu_call_ready_event>) at util/qemu-thread-posix.c:459
> #3 0x000055af9a9bd20d in call_rcu_thread (opaque=0x0) at util/rcu.c:260
> #4 0x000055af9a9a5288 in qemu_thread_start (args=0x55af9c4f1b80) at
> util/qemu-thread-posix.c:519
> #5 0x00007eff2f6d44c0 in start_thread () from /lib64/libpthread.so.0
> #6 0x00007eff2f602553 in clone () from /lib64/libc.so.6
>
> Thread 1 (Thread 0x7eff25820a80 (LWP 2842)):
> #0 0x00007eff2f5f7bd6 in ppoll () from /lib64/libc.so.6
> #1 0x000055af9a99e405 in qemu_poll_ns (fds=0x55af9c52a830, nfds=1,
> timeout=-1) at util/qemu-timer.c:335
> #2 0x000055af9a9a1cab in fdmon_poll_wait (ctx=0x55af9c526890,
> ready_list=0x7ffc73e8c5d0, timeout=-1) at util/fdmon-poll.c:79
> #3 0x000055af9a9a160c in aio_poll (ctx=0x55af9c526890, blocking=true) at
> util/aio-posix.c:600
> #4 0x000055af9a8f0bb0 in bdrv_do_drained_begin (bs=0x55af9c52a8d0,
> recursive=false, parent=0x0, ignore_bds_parents=false, poll=true) at
> block/io.c:429
> #5 0x000055af9a8f0c95 in bdrv_drained_begin (bs=0x55af9c52a8d0) at
> block/io.c:435
> #6 0x000055af9a8dc6a8 in blk_drain (blk=0x55af9c542c10) at
> block/block-backend.c:1681
> #7 0x000055af9a8da0b6 in blk_unref (blk=0x55af9c542c10) at
> block/block-backend.c:473
> #8 0x000055af9a8eb5e7 in mirror_exit_common (job=0x55af9c6c45c0) at
> block/mirror.c:667
> #9 0x000055af9a8eb9c1 in mirror_prepare (job=0x55af9c6c45c0) at
> block/mirror.c:765
> #10 0x000055af9a87cd65 in job_prepare (job=0x55af9c6c45c0) at job.c:781
> #11 0x000055af9a87b62a in job_txn_apply (job=0x55af9c6c45c0,
> fn=0x55af9a87cd28 <job_prepare>) at job.c:158
> #12 0x000055af9a87cdee in job_do_finalize (job=0x55af9c6c45c0) at job.c:798
> #13 0x000055af9a87cfb5 in job_completed_txn_success (job=0x55af9c6c45c0) at
> job.c:852
> #14 0x000055af9a87d055 in job_completed (job=0x55af9c6c45c0) at job.c:865
> #15 0x000055af9a87d0a8 in job_exit (opaque=0x55af9c6c45c0) at job.c:885
> #16 0x000055af9a99b981 in aio_bh_call (bh=0x55af9c547440) at util/async.c:136
> #17 0x000055af9a99ba8b in aio_bh_poll (ctx=0x55af9c526890) at util/async.c:164
> #18 0x000055af9a9a17ff in aio_poll (ctx=0x55af9c526890, blocking=true) at
> util/aio-posix.c:650
> #19 0x000055af9a8f7011 in bdrv_flush (bs=0x55af9c53b900) at block/io.c:3019
> #20 0x000055af9a874351 in bdrv_close (bs=0x55af9c53b900) at block.c:4252
> #21 0x000055af9a874ca3 in bdrv_delete (bs=0x55af9c53b900) at block.c:4498
> #22 0x000055af9a877862 in bdrv_unref (bs=0x55af9c53b900) at block.c:5866
> #23 0x000055af9a870837 in bdrv_root_unref_child (child=0x55af9c6c4430) at
> block.c:2684
> #24 0x000055af9a8da9a2 in blk_remove_bs (blk=0x55af9c547bd0) at
> block/block-backend.c:803
> #25 0x000055af9a8d9e54 in blk_delete (blk=0x55af9c547bd0) at
> block/block-backend.c:422
> #26 0x000055af9a8da0f8 in blk_unref (blk=0x55af9c547bd0) at
> block/block-backend.c:477
> #27 0x000055af9a86a6f1 in teardown_secondary () at
> tests/test-replication.c:392
> #28 0x000055af9a86aac1 in test_secondary_stop () at
> tests/test-replication.c:490
> #29 0x00007eff2fd7df7e in g_test_run_suite_internal () from
> /lib64/libglib-2.0.so.0
> #30 0x00007eff2fd7dd24 in g_test_run_suite_internal () from
> /lib64/libglib-2.0.so.0
> #31 0x00007eff2fd7dd24 in g_test_run_suite_internal () from
> /lib64/libglib-2.0.so.0
> #32 0x00007eff2fd7e46a in g_test_run_suite () from /lib64/libglib-2.0.so.0
> #33 0x00007eff2fd7e485 in g_test_run () from /lib64/libglib-2.0.so.0
> #34 0x000055af9a86b19c in main (argc=1, argv=0x7ffc73e8d088) at
> tests/test-replication.c:645
>
>
> (gdb) p ((BlockBackend *)0x55af9c547bd0)->in_flight
> $5 = 0
> (gdb) p ((BlockBackend *)0x55af9c542c10)->in_flight
> $6 = 0
> (gdb) p ((BlockDriverState *)0x55af9c53b900)->in_flight
> $7 = 1
> (gdb) p ((BlockDriverState *)0x55af9c52a8d0)->in_flight
> $8 = 0
> (gdb) fr 20
> #20 0x000055af9a874351 in bdrv_close (bs=0x55af9c53b900) at block.c:4252
> 4252 bdrv_flush(bs);
> (gdb) p bs->node_name
> $9 = "#block5317", '\000' <repeats 21 times>
> (gdb) p bs->drv
> $10 = (BlockDriver *) 0x55af9aa63c40 <bdrv_replication>
> (gdb) p bs->in_flight
> $11 = 1
> (gdb) p bs->tracked_requests
> $12 = {lh_first = 0x0}
>
>
> So, we entered bdrv_flush at frame 19, and increased in_flight. Then
> we go to aio_poll and to nested event loop, and we never return to
> decrease in_flight field.
>
> Hmm. I'm afraid, I don't know what to do with that. Kevin, could you
> take a look? And could similar thing happen with blk layer, because of
> you recent similar patch?
Hmm... You mean blk_prw(), right? Looks like it could have the same
problem, indeed.
Maybe we need to move the blk/bdrv_dec_in_flight to inside the coroutine
(probably to the place where we currently have aio_wait_kick(), which
would already be built in for bdrv_dec_in_flight). This is the last
thing the coroutine does, so presumably it will still be late enough.
Kevin
