[PATCHv4,0/1] xfs: soft lockups in unmapping and reflink remapping path

Message

Ritesh Harjani (IBM) May 7, 2024, 9:36 a.m. UTC

Hello,

In one of the testcases, parallel async dio writes to a file generates
large no. of extents (upto 2M or more), and then this file is cleaned up for
running other I/O tests. In the process of deleting this file, soft lockup
messages are observed. We believe this is happening due to kernel being busy
in unmapping/freeing those extents as part of the transaction processing.
This is similar observation with the same call stack which was also reported
here [1]. I also tried the qemu-img bench testcase shared in [1], and I was
able to reproduce the soft lockup with that on Power.

Similarly another instance was reported where xfs reflink remapping path also
saw a similar softlockup problem while iterating over a million extent entries.

So as I understood from that discussion [1], that kernel is moving towards a new
preemption model, but IIUC, it is still an ongoing work.
Also IMHO, this is still a problem in upstream and in older stable kernels
which we still do support and such a fix might still be necessary for them.

This patch adds the cond_resched() to both xfs_bunmapi_range() and
xfs_reflink_remap_blocks() functions.

v3 -> v4:
Remove cond_resched() from defer finish and add it to both the unmap range and
reflink remap functions individually.

v2 -> v3:
Move cond_resched within xfs_defer_finish_noroll() loop

Ritesh Harjani (IBM) (1):
  xfs: Add cond_resched to block unmap range and reflink remap path

 fs/xfs/libxfs/xfs_bmap.c | 1 +
 fs/xfs/xfs_reflink.c     | 1 +
 2 files changed, 2 insertions(+)

--
2.44.0