[LU-14399] mount MDT takes very long with hsm enable

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.15.0
• Environment: https://testing.whamcloud.com/test_sets/20fcc789-47a3-44a2-a17e-d6f39fbb3ec0
• Severity: 3
• Assignee: Sergey Cheremencev
• Reporter: Minh Diep

Description

We observed that when mounting MDT with HSM enable, mount command take minutes compare to seconds as before. We saw this in the log

[53618.238941] Lustre: DEBUG MARKER: mkdir -p /mnt/lustre-mds2; mount -t lustre -o localrecov  /dev/mapper/mds2_flakey /mnt/lustre-mds2
[53618.624098] LDISKFS-fs (dm-6): mounted filesystem with ordered data mode. Opts: user_xattr,errors=remount-ro,no_mbcache,nodelalloc
[53720.390690] Lustre: 1722736:0:(mdt_coordinator.c:1114:mdt_hsm_cdt_start()) lustre-MDT0001: trying to init HSM before MDD
[53720.392834] LustreError: 1722736:0:(mdt_coordinator.c:1125:mdt_hsm_cdt_start()) lustre-MDT0001: cannot take the layout locks needed for registered restore: -2
[53720.398049] LustreError: 1722741:0:(mdt_coordinator.c:1090:mdt_hsm_cdt_start()) lustre-MDT0001: Coordinator already started or stopping
[53720.400681] Lustre: lustre-MDT0001: Imperative Recovery not enabled, recovery window 60-180
[53720.424872] Lustre: lustre-MDT0001: in recovery but waiting for the first client to connect
[53720.953893] Lustre: DEBUG MARKER: /usr/sbin/lctl get_param -n health_check
[53722.067555] Lustre: DEBUG MARKER:  

Seems related to LU-13920

Issue Links

• is related to: LU-13920 HSM: hsm_actions are not processed after MDS failover

Comments

[Alex Zhuravlev]

well, I'm running master branch, so clearly LU-14399 is in. this is what helps:\

-       while (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state) && i < obd_timeout) {
+       while (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state) && i < obd_timeout * 2) {

not sure how good it is...

[Sergey Cheremencev]

 I guess "trying to init HSM before MDD" is a hint?

IMO, it hints your build doesn't include "LU-14399 hsm: process hsm_actions in coordinator".

static void cdt_start_pending_restore(struct mdt_device *mdt,
                      struct coordinator *cdt)
{
    struct mdt_thread_info *cdt_mti;
    unsigned int i = 0;
    int rc;    /* wait until MDD initialize hsm actions llog */
    while (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state) && i < obd_timeout) {
        schedule_timeout_interruptible(cfs_time_seconds(1));
        i++;
    }
    if (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state))
        CWARN("%s: trying to init HSM before MDD\n", mdt_obd_name(mdt));
 

"trying to init HSM before MDD" message should be printed by cdt_start_pending_restore, while in your case it is mdt_hsm_cdt_start.

[Alex Zhuravlev]

[ 8878.042110] Lustre: Found index 0 for lustre-MDT0000, updating log
[ 8878.044766] Lustre: Modifying parameter lustre-MDT0000.mdt.identity_upcall in log lustre-MDT0000
[ 8898.880071] Lustre: 497204:0:(mdt_coordinator.c:1145:mdt_hsm_cdt_start()) lustre-MDT0000: trying to init HSM before MDD
[ 8898.888824] LustreError: 497204:0:(mdt_coordinator.c:1156:mdt_hsm_cdt_start()) lustre-MDT0000: cannot take the layout locks needed for registered restore: -2
[ 8899.150565] Lustre: DEBUG MARKER: conf-sanity test_132: @@@@@@ FAIL: Can not take the layout lock

I guess "trying to init HSM before MDD" is a hint?

[Sergey Cheremencev]

Alex, it still doesn't fail on my local VM.
I can you look into the logs if you attach them to the ticket.

[Alex Zhuravlev]

well, essentially just a local VM:

FSTYPE=ldiskfs MDSCOUNT=2 MDSSIZE=300000 OSTSIZE=400000 OSTCOUNT=2 LOGDIR=/tmp/ltest-logs REFORMAT=yes HONOR_EXCEPT=y bash conf-sanity.sh 

[Sergey Cheremencev]

Hello Alex,

what setup did you use? It doesn't fail on my local VM on the latest master(0feec5a3).

[Alex Zhuravlev]

the patch just landed fails every run on my setup:

...
Writing CONFIGS/mountdata
start mds service on tmp.BKaRODHgLn
Starting mds1: -o localrecov /dev/mapper/mds1_flakey /mnt/lustre-mds1
Started lustre-MDT0000
conf-sanity test_132: @@@@@@ FAIL: Can not take the layout lock
Trace dump:
= ./../tests/test-framework.sh:6389:error()
= conf-sanity.sh:9419:test_132()
= ./../tests/test-framework.sh:6693:run_one()
= ./../tests/test-framework.sh:6740:run_one_logged()
= ./../tests/test-framework.sh:6581:run_test()
= conf-sanity.sh:9422:main()
Dumping lctl log to /tmp/ltest-logs/conf-sanity.test_132.*.1642612854.log
Dumping logs only on local client.
FAIL 132 (84s)

[Peter Jones]

Landed for 2.15

[Gerrit Updater]

"Oleg Drokin <[e-mail]>" merged in patch https://review.whamcloud.com/41445/
Subject: LU-14399 hsm: process hsm_actions in coordinator
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: e26d7cc3992252e5fce5a51aee716f933b04c13a

[Gerrit Updater]

Sergey Cheremencev ([e-mail]) uploaded a new patch: https://review.whamcloud.com/42005
Subject: LU-14399 tests: hsm_actions after failover
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 87493ef365d9faaf1f6c1e1a40f65157d37f72dc