[LU-17963] sometime touch command cannot change mtime

• Type: Bug
• Resolution: Incomplete
• Priority: Minor
• Severity: 3
• Assignee: Feng Lei
• Reporter: Feng Lei

Description

First create a file, then change its mtime immediately (touch command, or utimes/utimensat system call). Sometime mtime is not changed.

The bug can be reproduced in sanity/39_d easily in the development of nanosecond timestamps, but rarely/never be reproduced in current version with second granularity timestamps.

Issue Links

• is related to: LU-1158 nanosecond timestamp support for Lustre

Comments

[Andreas Dilger]

We mark tickets Resolved instead of Closed, since it is still allows changing some fields of the ticket (eg. Labels, versions, etc).

[Feng Lei]

In https://review.whamcloud.com/c/fs/lustre-release/+/53313, which enable nanosecond field for timestamps, the bug is fixed.

If timestamps are still in second, the bug is never reproduced.

This bug can be re-opened if sanity/test_39d failed.

[Feng Lei]

utimes() is dropped only when new ctime is smaller than existing ctime. In the end of function mdd_fix_attr():

    if (la->la_valid & LA_CTIME) {
        /**
         * The pure setattr, it has the priority over what is
         * already set, do not drop it if ctime is equal.
         */
        if (la->la_ctime < oattr->la_ctime)
            la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
    }
 

This bug is actually found during the development of nanosecond timestamps. If the time granularity is nanosecond, it's easy to detect the difference of new ctime and old ctime. But if the time granularity is second, new ctime is almost always equal to old ctime. That's the reason it's not found in current lustre.

[Andreas Dilger]

flei, I haven't checked, but is it possible that the utimes() update is dropped because the ctime is the same?

[Gerrit Updater]

"Feng Lei <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/55497
Subject: LU-17963 llite: get timestamps in an accurate way
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: fb383e763dc7e960024ce105141f4f41c915052b

[Feng Lei]

MDS has an optimization that: when setting file's timestamps, it compares the new ctime to be set with the existing ctime. If the new ctime is before the existing one, it will ignore the setting operation.

In a tight operation series open/close/utimes, file's timestamps are firstly set by close(), then by utimes(). The utimes() operation gets ctime with current_time() function, which may get a lagged time than the real time. So there is a tiny window that: close() gets ctime at N seconds, but utimes() later get ctime at (N-1) second. Then the utimes() operation is ignored.