block/mirror: fix active mirror dead-lock in mirror_wait_on_conflicts

It's possible that requests start to wait each other in
mirror_wait_on_conflicts(). To avoid it let's use same technique as in
block/io.c in bdrv_wait_serialising_requests_locked() /
bdrv_find_conflicting_request(): don't wait on intersecting request if
it is already waiting for some other request.

For details of the dead-lock look at testIntersectingActiveIO()
test-case which we actually fixing now.

Fixes: d06107ade0
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210702211636.228981-4-vsementsov@virtuozzo.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This commit is contained in:
Vladimir Sementsov-Ogievskiy 2021-07-03 00:16:36 +03:00 committed by Kevin Wolf
parent e0f69d83d5
commit d44dae1a7c
2 changed files with 17 additions and 13 deletions

View File

@ -107,6 +107,7 @@ struct MirrorOp {
bool is_in_flight; bool is_in_flight;
CoQueue waiting_requests; CoQueue waiting_requests;
Coroutine *co; Coroutine *co;
MirrorOp *waiting_for_op;
QTAILQ_ENTRY(MirrorOp) next; QTAILQ_ENTRY(MirrorOp) next;
}; };
@ -159,7 +160,18 @@ static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
if (ranges_overlap(self_start_chunk, self_nb_chunks, if (ranges_overlap(self_start_chunk, self_nb_chunks,
op_start_chunk, op_nb_chunks)) op_start_chunk, op_nb_chunks))
{ {
/*
* If the operation is already (indirectly) waiting for us, or
* will wait for us as soon as it wakes up, then just go on
* (instead of producing a deadlock in the former case).
*/
if (op->waiting_for_op) {
continue;
}
self->waiting_for_op = op;
qemu_co_queue_wait(&op->waiting_requests, NULL); qemu_co_queue_wait(&op->waiting_requests, NULL);
self->waiting_for_op = NULL;
break; break;
} }
} }

View File

@ -143,10 +143,6 @@ class TestActiveMirror(iotests.QMPTestCase):
self.potential_writes_in_flight = False self.potential_writes_in_flight = False
def testIntersectingActiveIO(self): def testIntersectingActiveIO(self):
# FIXME: test-case is dead-locking. To reproduce dead-lock just drop
# this return statement
return
# Fill the source image # Fill the source image
result = self.vm.hmp_qemu_io('source', 'write -P 1 0 2M') result = self.vm.hmp_qemu_io('source', 'write -P 1 0 2M')
@ -180,18 +176,14 @@ class TestActiveMirror(iotests.QMPTestCase):
# Now we resumed 1, so 2 and 3 goes to the next iteration of while loop # Now we resumed 1, so 2 and 3 goes to the next iteration of while loop
# in mirror_wait_on_conflicts(). They don't exit, as bitmap is dirty # in mirror_wait_on_conflicts(). They don't exit, as bitmap is dirty
# due to request 4. And they start to wait: 2 wait for 3, 3 wait for 2 # due to request 4.
# - DEAD LOCK. # In the past at that point 2 and 3 would wait for each other producing
# Note that it's important that we add request 4 at last: requests are # a dead-lock. Now this is fixed and they will wait for request 4.
# appended to the list, so we are sure that 4 is last in the list, so 2
# and 3 now waits for each other, not for 4.
self.vm.hmp_qemu_io('source', 'resume B') self.vm.hmp_qemu_io('source', 'resume B')
# Resuming 4 doesn't help, 2 and 3 already dead-locked # After resuming 4, one of 2 and 3 goes first and set in_flight_bitmap,
# To check the dead-lock run: # so the other will wait for it.
# gdb -p $(pidof qemu-system-x86_64) -ex 'set $job=(MirrorBlockJob *)jobs.lh_first' -ex 'p *$job->ops_in_flight.tqh_first' -ex 'p *$job->ops_in_flight.tqh_first->next.tqe_next'
# You'll see two MirrorOp objects waiting on each other
result = self.vm.qmp('block-job-set-speed', device='mirror', speed=0) result = self.vm.qmp('block-job-set-speed', device='mirror', speed=0)
self.assert_qmp(result, 'return', {}) self.assert_qmp(result, 'return', {})