File drbd-fix-termination-of-verify-with-stop-sector.patch of Package drbd.28199
From 263ac567d1b4bb285c13dadaa7cb59b6a2dc8d48 Mon Sep 17 00:00:00 2001 From: Joel Colledge <joel.colledge@linbit.com> Date: Fri, 28 May 2021 10:41:48 +0200 Subject: [PATCH] drbd: fix termination of verify with stop sector Using the verify stop sector functionality could result in DRBD getting stuck in VerifyS/VerifyT states. It was caused by a race condition, but such that it was reliably reproducible in some environments. When this occurred, lines such as the following were repeatedly logged on the VerifyS node: drbd test/0 drbd1000 node: Retrying drbd_rs_del_all() later. refcnt=1 The root of the issue was a bug in make_ov_request which caused it to always issue one more request even if the stop sector had been reached. When the "last" OV request was done, make_ov_request was scheduled, as well as drbd_resync_finished. These two functions raced. The function make_ov_request locking a resync extent, and the function drbd_resync_finished clearing the resync extents. When make_ov_request won the race, the resync extents could not be cleared and the verify operation could not be finished. With corking enabled and no other activity causing anything to be sent, the extra resync extent remained locked indefinitely because the OV request was never actually sent. With corking disabled or sender activity (of a sort which is not blocked by the locked resync extent, e.g. writes to a different disk region), the verify operation did continue to make slow progress until it eventually reached the end of the disk. Fix this by only ensuring that one OV request is sent for the whole verify operation rather than one every time make_ov_request is called. --- drbd/drbd_sender.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drbd/drbd_sender.c b/drbd/drbd_sender.c index fac63d75..2dc1ba41 100644 --- a/drbd/drbd_sender.c +++ b/drbd/drbd_sender.c @@ -947,7 +947,7 @@ static int make_ov_request(struct drbd_peer_device *peer_device, unsigned int se /* We check for "finished" only in the reply path: * w_e_end_ov_reply(). * We need to send at least one request out. */ - stop_sector_reached = i > 0 + stop_sector_reached = sector > peer_device->ov_start_sector && verify_can_do_stop_sector(peer_device) && sector >= peer_device->ov_stop_sector; if (stop_sector_reached) -- 2.16.4
