[LU-14694] racer timeouts with LU-10948

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.15.0
• Affects Version/s: Upstream
• Severity: 3
• Assignee: Alex Zhuravlev
• Reporter: Alex Zhuravlev

Description

many racer runs timeout, bisected that to patch:
https://review.whamcloud.com/32158/
Subject: LU-10948 llite: Introduce inode open heat counter
Project: fs/lustre-release
Branch: master
Commit: 41d99c4902836b7265db946dfa49cf99381f0db4

Issue Links

• is related to: LU-10948 client cache open lock after N opens

Comments

[Peter Jones]

Landed for 2.15

[Gerrit Updater]

"Oleg Drokin <[e-mail]>" merged in patch https://review.whamcloud.com/43783/
Subject: LU-14694 mdt: do not remove orphans at umount
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 0b94a058fe6732907d7aa3f5de0cbd7888d28ce1

[Gerrit Updater]

Alex Zhuravlev ([e-mail]) uploaded a new patch: https://review.whamcloud.com/43783
Subject: LU-14694 mdt: do not remove orphans at umount
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 9f9e98a5e94e021c68882db5ab3d03484048b810

[Alex Zhuravlev]

I think we need something like above on its own as umount with just evicted client is a legal case.

[Alex Zhuravlev]

probably a bit better option is to re-use existing infrastructure to skip orphan removal:

--- a/lustre/mdt/mdt_handler.c
+++ b/lustre/mdt/mdt_handler.c
@@ -6438,7 +6438,8 @@ static int mdt_export_cleanup(struct obd_export *exp)
                                rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
 
                        /* Don't unlink orphan on failover umount, LU-184 */
-                       if (exp->exp_flags & OBD_OPT_FAILOVER) {
+                       if (exp->exp_flags & OBD_OPT_FAILOVER ||
+                               exp->exp_obd->obd_stopping) {
                                ma->ma_valid = MA_FLAGS;
                                ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
                        }

[Alex Zhuravlev]

basically the problem is that:
1) openlock is taken on directories
2) the client doesn't cancel locks on umount:

                /* obd_force == local only */
                ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
                                       obd->obd_force ? LCF_LOCAL : 0, NULL);

3) thus MDT has to clear those "lost" locks, so close the directories
4) a close may result in directory removal
5) directory can be striped, thus needs all involved MDTs to be healty
6) MDTs are stopped one by one

locally I "solved" the problem disabling opencache for directories:

--- a/lustre/mdt/mdt_open.c
+++ b/lustre/mdt/mdt_open.c
@@ -820,6 +820,9 @@ static int mdt_object_open_lock(struct mdt_thread_info *info,
        if (req_is_replay(mdt_info_req(info)))
                RETURN(0);
 
+       if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
+               open_flags &= ~MDS_OPEN_LOCK;
+
        if (S_ISREG(lu_object_attr(&obj->mot_obj))) {

but it looks like we've got few real problems to solve here. the most serious one, IMO, is to handle such a close on MDT in a better manner, so that umount doesn't get stuck indefinitely.

[Alex Zhuravlev]

--- a/lustre/tests/racer.sh
+++ b/lustre/tests/racer.sh
@@ -180,6 +180,7 @@ test_1() {
                lss_cleanup
        fi
 
+       cancel_lru_locks mdc
        return $rrc
 }
 run_test 1 "racer on clients: ${CLIENTS:-$(hostname)} DURATION=$DURATION"

this helps.. it looks like openhandle locks are lost at umount, causing MDT to close them at export cleanup.

[Alex Zhuravlev]

sanity/29 fails as well:

== sanity test 29: IT_GETATTR regression  ============================================================ 11:52:13 (1621597933)
Lustre: DEBUG MARKER: == sanity test 29: IT_GETATTR regression ============================================================ 11:52:13 (1621597933)
striped dir -i1 -c2 -H all_char /mnt/lustre/d29
first d29
Lustre: DEBUG MARKER: first d29
total 0
-rw-r--r-- 1 root root 0 May 21 11:52 foo
second d29
Lustre: DEBUG MARKER: second d29
total 0
-rw-r--r-- 1 root root 0 May 21 11:52 foo
done
Lustre: DEBUG MARKER: done
 sanity test_29: @@@@@@ FAIL: CURRENT: 10 > 9 
Lustre: DEBUG MARKER: sanity test_29: @@@@@@ FAIL: CURRENT: 10 > 9
  Trace dump:
  = ./../tests/test-framework.sh:6280:error()
  = sanity.sh:3398:test_29()
  = ./../tests/test-framework.sh:6583:run_one()
  = ./../tests/test-framework.sh:6630:run_one_logged()
  = ./../tests/test-framework.sh:6457:run_test()
  = sanity.sh:3410:main()
Dumping lctl log to /tmp/ltest-logs/sanity.test_29.*.1621597935.log
Dumping logs only on local client.
Resetting fail_loc on all nodes...done.
FAIL 29 (2s)

[Alex Zhuravlev]

Also, is this with opencache enabled, or just with the patch applied?

this is the latest clean master

[Alex Zhuravlev]

basically it's MDSCOUNT=2 REFORMAT=yes bash racer.sh