On 1/24/25 04:33, Pavel Begunkov wrote:
Thanks for narrowing it down. Xan, can you try this change please?
Waiters can miss wake ups without it, seems to match the description.
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 9b58ba4616d40..e5a8ee944ef59 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -592,8 +592,10 @@ static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
io_commit_cqring(ctx);
spin_unlock(&ctx->completion_lock);
io_commit_cqring_flush(ctx);
- if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
+ if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
+ smp_mb();
__io_cqring_wake(ctx);
+ }
}
void io_cq_unlock_post(struct io_ring_ctx *ctx)
Thanks Pavel! Early results look very good for this change. I'm now
running 6.1.120 with your added smp_mb() call. The backup process which
had been quickly triggering the issue has been running longer than it
ever did when it would ultimately fail. So that's great!
One sour note: overnight, replication hung on this machine, which is
another failure that started happening with the jump from 6.1.119 to
6.1.123. The machine was running 6.1.124 with the
__io_cq_unlock_post_flush function removed completely. That's the
kernel we had celebrated yesterday for running the backup process
successfully.
So, we might have two separate issues to deal with, unfortunately.
This morning, I found that replication had hung and was behind by some
35,000 seconds. I attached gdb and then detached it, which got things
moving again (which goes the extra mile to prove that this is a very
closely related issue). Then it hung up again at about 25,000 seconds
behind. At that point I rebooted into the new kernel, the 6.1.120
kernel with the added smp_mb() call. The lag is now all the way down to
5,000 seconds without hanging again.
It looks like there are 5 io_uring-related patches in 6.1.122 and
another 1 in 6.1.123. My guess is the replication is hitting a problem
with one of those.
Unfortunately, a replication hang is much harder for me to reproduce
than the issue with the backup procedure, which always failed within 15
minutes. It certainly looks to me like the patched 6.1.120 does not
have the hang (but it's hard to be 100% certain). Perhaps the next step
is to apply the extra smp_mb() call to 6.1.123 and see if I can get
replication to hang.
