-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
This issue was created by maloo for Patrickbot <patrick-bot@mulberrytree.us>
This issue relates to the following test suite run: https://testing.whamcloud.com/test_sets/fb852cf2-0bb6-4701-8ddc-99d6521a3a03
test_48d failed with the following error:
truncate /mnt/lustre/d48d.sanity-sec/f48d.sanity-sec to 8404999 failed
Test session details:
Client
Lustre: 2.17.62.1
Branch: master
Build: https://build.whamcloud.com/job/lustre-reviews/137384
Kernel: 5.14.0-687.60.1.el9_8.aarch64+64k
Server
Lustre: 2.17.62.1
Branch: master
Build: https://build.whamcloud.com/job/lustre-reviews/137384
Kernel: 4.18.0-553.177.1.el8_10_lustre.x86_64
sanity-sec test_48d (added by LU-20964) fails with "truncate ... failed: Operation not supported" when the page being zeroed is still locked by in-flight read ahead.
Failed session: https://testing.whamcloud.com/test_sessions/c31d3160-ba79-4457-9009-9dad79900597 (review-ldiskfs-arm, rocky9.8 aarch64 client, 64K pages, change 70663 ps4 build 137384).
Analysis from the client debug log: ll_setattr_raw() sends the truncate to the MDT, then ll_io_zero_page() runs for page 128. It takes the PW lock (matched) and calls get_folio_nowait() with FGP_NOWAIT, which fails if the page is locked, and the function returns -EOPNOTSUPP (llite_lib.c, "IS_ERR_OR_NULL(folio)" case). ll_setattr_raw() fails the truncate with it. The read ahead RPCs of the test's dd were still completing at that moment (OST_READ replies at ...145.406 and ...145.834, after the truncate at ...145.26), so the read ahead pages were still locked. The test's "sleep 1" is not enough on a slow client, and the kernel returns EOPNOTSUPP to the application for what is only a transient page lock.
Only failure of 48d in 30 days (about 1 of 55 runs on review-ldiskfs-arm, 0 of 9395 suite runs elsewhere), and not caused by change 70663: that change does not touch the setattr or zero page path.
Possible fixes: have ll_io_zero_page() wait for the page lock (or retry) instead of returning -EOPNOTSUPP, and/or wait for read ahead to finish in the test instead of a fixed sleep.
Related: LU-20964 (merged, added ll_io_zero_page() keeping the zero page owned and test 48d).
Raised by Patch Watcher run pw-engineer-70663-ps4-14d32ec52f14 on https://review.whamcloud.com/70663
- is related to
-
LU-20964 Client LBUG/kernel panic: ASSERTION( page->cp_owner != NULL ) in cl_page_list_add via ll_io_zero_page during ftruncate of fscrypt-encrypted file
-
- Resolved
-