[LU-6371] sendfile with group lock splits transfer into tiny chunks

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.7.0
• Type: Bug
• Priority: Major
• Reporter: Frank Zago (Inactive)
• Assignee: WC Triage
• Resolution: Duplicate
• Severity: 3

Description

A program using sendfile to copy a file from lustre to lustre, with a group lock on its source will see the transfer split in tiny chunks.

Reproducer attached.

Test setup. Create a 100MiB file.

$ pwd
/mnt/lustre
$ dd if=/dev/zero of=foo1 bs=1M count=100 

Transfer without a group lock:

$ strace ~/lustre-cleanup/lustre/tests/llsendfile4 0 foo1 foo2
...
sendfile(4, 3, NULL, 10485760)          = 10485760
sendfile(4, 3, NULL, 10485760)          = 10485760
sendfile(4, 3, NULL, 10485760)          = 10485760
sendfile(4, 3, NULL, 10485760)          = 10485760
sendfile(4, 3, NULL, 10485760)          = 10485760
sendfile(4, 3, NULL, 10485760)          = 10485760
...

With a group lock

$ strace ~/lustre-cleanup/lustre/tests/llsendfile4 1223 foo1 foo2
...
sendfile(4, 3, NULL, 10485760)          = 28672
sendfile(4, 3, NULL, 10485760)          = 4096
sendfile(4, 3, NULL, 10485760)          = 4096
sendfile(4, 3, NULL, 10485760)          = 8192
sendfile(4, 3, NULL, 10485760)          = 4096
sendfile(4, 3, NULL, 10485760)          = 4096
sendfile(4, 3, NULL, 10485760)          = 4096
...

The second example takes about 10 times longer than the first.

Issue Links

• is related to: LU-6368 ASSERTION( cur->oe_dlmlock == victim-...

Comments

[Patrick Farrell (Inactive)]

Frank reported to me that the patch for https://jira.hpdd.intel.com/browse/LU-6368 fixes this issue.

Honestly, I have no idea why. I'll try to understand why and report back.

[Jinshan Xiong (Inactive)]

It's good for you guys to run test for sendfile cases. Is it possible to generate a few test cases and add them into sanity?

Thanks,

[Frank Zago (Inactive)]

Yes, I'll add some tests too.

[Andreas Dilger]

The patch in question is http://review.whamcloud.com/14093