On Fri, 2021-12-31 at 01:42 +0000, Matthew Wilcox wrote:

On Thu, Dec 30, 2021 at 02:35:22PM -0500, trondmy@kernel.org wrote:

quoted

 Workqueue: xfs-conv/md127 xfs_end_io [xfs]
 RIP: 0010:_raw_spin_unlock_irqrestore+0x11/0x20
 Code: 7c ff 48 29 e8 4c 39 e0 76 cf 80 0b 08 eb 8c 90 90 90 90 90
90 90 90 90 90 0f 1f 44 00 00 e8 e6 db 7e ff 66 90 48 89 f7 57 9d
<0f> 1f 44 00 00 c3 66 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 8b 07
 RSP: 0018:ffffac51d26dfd18 EFLAGS: 00000202 ORIG_RAX:
ffffffffffffff12
 RAX: 0000000000000001 RBX: ffffffff980085a0 RCX: dead000000000200
 RDX: ffffac51d3893c40 RSI: 0000000000000202 RDI: 0000000000000202
 RBP: 0000000000000202 R08: ffffac51d3893c40 R09: 0000000000000000
 R10: 00000000000000b9 R11: 00000000000004b3 R12: 0000000000000a20
 R13: ffffd228f3e5a200 R14: ffff963cf7f58d10 R15: ffffd228f3e5a200
 FS:  0000000000000000(0000) GS:ffff9625bfb00000(0000)
knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007f5035487500 CR3: 0000000432810004 CR4: 00000000003706e0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 Call Trace:
  wake_up_page_bit+0x8a/0x110
  iomap_finish_ioend+0xd7/0x1c0
  iomap_finish_ioends+0x7f/0xb0

quoted

+++ b/fs/iomap/buffered-io.c

@@ -1052,9 +1052,11 @@ iomap_finish_ioend(struct iomap_ioend

*ioend, int error)
                        next = bio->bi_private;
 
                /* walk each page on bio, ending page IO on them */
-               bio_for_each_segment_all(bv, bio, iter_all)
+               bio_for_each_segment_all(bv, bio, iter_all) {
                        iomap_finish_page_writeback(inode, bv-

quoted

bv_page, error,

                                        bv->bv_len);
+                       cond_resched();
+               }
                bio_put(bio);
        }
        /* The ioend has been freed by bio_put() */

As I recall, iomap_finish_ioend() can be called in softirq (or even
hardirq?) context currently.  I think we've seen similar things
before,
and the solution suggested at the time was to aggregate fewer
writeback
pages into a single bio.

I haven't seen any evidence that iomap_finish_ioend() is being called
from anything other than a regular task context. Where can it be called
from softirq/hardirq and why is that a requirement?

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com