[LU-10048] osd-ldiskfs to truncate outside of main transaction

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.13.0
• Affects Version/s: Lustre 2.12.0, Lustre 2.10.4, Lustre 2.10.5
• Assignee: Alex Zhuravlev
• Reporter: Alex Zhuravlev

Description

this is needed to implement (transaction first; locking next) order to unify locking among MDT/OST/OUT

Issue Links

• is blocked by: LU-11764 osd_declare_qid() should not rely on constant id
• is duplicated by: LU-11613 MDS and OSS locked up wait_transaction_locked+0x85/0xd0 [jbd2]
• is duplicated by: LU-8786 Terrible i/o performance of a test application doing repeatable writes and truncates
• is related to: LU-11685 removing file names and freeing inodes are not atomic on the OST
• is related to: LU-11465 OSS/MDS deadlock in 2.10.5
• is related to: LU-5994 DT transaction start and object lock ordering
• is related to: LU-13234 ldiskfs/namei.c:3310 ldiskfs_orphan_add+0x11e/0x2a0 [ldiskfs]
• is related to: LU-12977 fix i_mutex for ldiskfs_truncate() in osd_execute_truncate()
• is related to: LU-5152 Can't enforce block quota when unprivileged user change group
• is related to: LU-8806 LFSCK hangs on MDT - osp_precreate_cleanup_orphans - cannot cleanup orphans: rc = -107

Comments

[Etienne Aujames]

Hello,

We have applied the https://review.whamcloud.com/43277 ("LU-10048 ofd: take local locks within transaction") + https://review.whamcloud.com/43278 ("LU-13093 osd: fix osd_attr_set race") on the problematic filesystem. The issue never occurred after.

We were able to reproduce this issue in 5/10 min with many creations of small files and several threads doing file migrations ("lfs migrate" between OSTs).

[Etienne Aujames]

Hello Alex,

Did you have the time to look to our backtrace?

If you need more data from the crashdump I can get you some.

[Etienne Aujames]

I have added the our backtrace to this tickets: crash_bt_jbd2_locked_20210412.log

[Etienne Aujames]

The issue seems to occurs relatively often with a lot of migrate threads (today 4 times on different OSTs).

We will test the backports quickly on a production environment (maybe this week).

I will try to get backtrace from the crashdump (manually triggered) tomorrow.

[Alex Zhuravlev]

please, generate full backtrace (

echo t >/proc/sysrq-trigger

) and attach to the ticket

[Etienne Aujames]

The "LU-10048 osd: async truncate" has already landed on b2_12.

[Alex Zhuravlev]

as for fast journal - we've got number of nodes running all the tests on RAM-backed devices 24h a day.

[Alex Zhuravlev]

there is yet another patch under LU-10048 - LU-10048 osd: async truncate, cf29a5e7bfa254ccfcea023028fe7da80503c512 in master

[Etienne Aujames]

Hello,

I have backported the https://review.whamcloud.com/43277 ("LU-10048 ofd: take local locks within transaction") along with the https://review.whamcloud.com/43278 ("LU-13093 osd: fix osd_attr_set race").

It seems we trigger that bug in 2.12.6 with external journal (on flash dev) for rotational disk while several migrate thread is running on a robinhood client (512 thread).

jbd2 journal seems to be locked when the transaction is in T_LOCKED.

transaction_t.t_state = T_LOCKED
transaction_t.t_handle_count = 40
transaction_t.t_updates = 1
nbr of task in j_wait_transaction_locked = 324
nbr of task in j_wait_update = 0

[Gerrit Updater]

Etienne AUJAMES ([e-mail]) uploaded a new patch: https://review.whamcloud.com/43277
Subject: LU-10048 ofd: take local locks within transaction
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: 4dc5634dcdf3381a7ea698605a3c7a345f689a4c