-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
After a PCC file is mmapped, pcc_mmap_mapping_set() points the Lustre inode mapping a_ops at the PCC backend (e.g. ext4) aops, until detach. On kernels v6.10+ those aops have no >direct_IO. If a PCC-RO read then fails with an error pcc_io_tolerate() does not tolerate (any backend -EIO, or OBD_FAIL_LLITE_PCC_FAKE_ERROR), do_file_read_iter() falls back to the Lustre read path with IOCB_DIRECT still set: ll_file_io_generic()>vvp_io_read_start()>generic_file_read_iter() then calls mapping>a_ops->direct_IO(), which is NULL in the backend aops.
It can be hit by a file opened O_DIRECT before mmap, or by hybrid IO (large buffered reads get IOCB_DIRECT, on by default). Change 67003 (LU-20388, PCC mmap fix for mmap_prepare) additionally sets FMODE_CAN_ODIRECT unconditionally in ll_file_open(), which lets explicit O_DIRECT opens succeed in this state on newer kernels.
Suggested fix: keep IOCB_DIRECT off the Lustre fallback (or skip the fallback) while inode->i_mapping->a_ops != &ll_aops, with a regression test. Found reviewing https://review.whamcloud.com/c/fs/lustre-release/+/67003 (AI review comment on lustre/llite/file.c); Oleg Drokin asked for a separate ticket and a separate patch.
- is related to
-
LU-20388 Ubuntu 26.04 sanity-pcc test_2b: FAIL: /mnt/lustre/d2b.sanity-pcc/multiop expected pcc state: none, but got: readwrite
-
- Open
-