nbd: make nbd_export_close_all() synchronous
commit453cc6be0a954ecdb4e0cdbf5b5b87f07b3e1075
authorVladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Tue, 14 Jul 2020 16:22:33 +0000 (14 19:22 +0300)
committerKevin Wolf <kwolf@redhat.com>
Fri, 17 Jul 2020 12:20:57 +0000 (17 14:20 +0200)
treed2e7585e7af3bf1cee1442ef760f490619514b29
parentf21f12936f9db6de4eddfcf9a086b6efd5b093cc
nbd: make nbd_export_close_all() synchronous

Consider nbd_export_close_all(). The call-stack looks like this:
 nbd_export_close_all() -> nbd_export_close -> call client_close() for
each client.

client_close() doesn't guarantee that client is closed: nbd_trip()
keeps reference to it. So, nbd_export_close_all() just reduce
reference counter on export and removes it from the list, but doesn't
guarantee that nbd_trip() finished neither export actually removed.

Let's wait for all exports actually removed.

Without this fix, the following crash is possible:

- export bitmap through internal Qemu NBD server
- connect a client
- shutdown Qemu

On shutdown nbd_export_close_all is called, but it actually don't wait
for nbd_trip() to finish and to release its references. So, export is
not release, and exported bitmap remains busy, and on try to remove the
bitmap (which is part of bdrv_close()) the assertion fails:

bdrv_release_dirty_bitmap_locked: Assertion `!bdrv_dirty_bitmap_busy(bitmap)' failed

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20200714162234.13113-2-vsementsov@virtuozzo.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
nbd/server.c