> When the owner is killed, vhost_worker_killed() detaches the worker > from the vqs. A bio completing after that fails to queue the > completion work, so its request stays on llhead and is never removed > from req_inflight. vhost_blk_flush() on release then waits forever > and the process is stuck in D state. Any kill of a VM with I/O in > flight can hit this. > > If the work cannot be queued, drop the completed requests right in > the bio end_io: forget them without writing a status, as there is no > guest left to see it. llist_del_all() makes sure a request is dropped > once even if it races with the last run of the host kick handler. > forget_request() can now run in irq context, so take flush_lock with > interrupts disabled. The end_io may run in atomic context and the > request buffers may be vmalloc'ed, so free them on release instead: > a dropped request is the one left with ->inflight set. > > Once a request is on llhead, a concurrent consumer may forget it and > let release free the request array and blk while the end_io is still > using blk_vq. Keep the end_io tail in an RCU read-side section and > wait for a grace period before freeing the requests. The same window > existed before with the host kick handler as the consumer. > > https://virtuozzo.atlassian.net/browse/VSTOR-152175 > > Fixes: c57a98d65796 ("drivers/vhost: vhost-blk accelerator for virtio-blk > guests") > Feature: vhost-blk: in-kernel accelerator for virtio-blk guests > Signed-off-by: Alexander Ivanov <[email protected]>
Sashiko has reviewed this patch and found no issues. -- Sashiko AI review ยท https://sashiko.vzint.dev/#/patchset/[email protected]?part=9 _______________________________________________ Devel mailing list [email protected] https://lists.openvz.org/mailman/listinfo/devel
