Hi Kent,
Firstly, I confirmed that today's master seems to avoid the splat I sent
previously (re: your comment about a reverse journal replay patch or
some such).
I still reproduce the stall issue on this system. After peeling away at
it, I was eventually able to reproduce without the drop writes
(dm-flakey) behavior from the test, and with fio using either the libaio
or sync I/O engine options. The sync I/O mode fortunately provides a
more useful stack trace:
# cat /proc/177747/stack
[<0>] bch2_dio_write_flush+0x122/0x160 [bcachefs]
[<0>] bch2_direct_write+0xb53/0xce0 [bcachefs]
[<0>] bch2_write_iter+0x142/0xc70 [bcachefs]
[<0>] vfs_write+0x29b/0x470
[<0>] ksys_write+0x6f/0xf0
[<0>] do_syscall_64+0x86/0x170
[<0>] entry_SYSCALL_64_after_hwframe+0x6e/0x76
... which resolves down to the closure_sync() call in
bch2_dio_write_flush(). The problem seems to go away if I remove the
preceding journal flush from that function. This seems to rule out
io_uring/aio and instead suggest that we're getting stuck somehow
waiting on a journal flush.
Based on that I went back to the first commit before 746a33c96b7a0
("bcachefs: better journal pipelining"). With that, I can run hundreds
of iterations of generic/703 without a problem, so this appears to be a
regression associated with the journal pipeline improvements. I'm
currently re-running on the last known good commit with my test tweaks
backed out (i.e. so back to io_uring and drop writes) just to
corroborate that it's the same problem, but so far it's running as
expected...
Brian
