Re: soft lockup with io_uring

[Jens Axboe <axboe@kernel.dk>]

On 8/21/19 4:48 PM, Sagi Grimberg wrote:
> Hey,
> 
> Just ran io-uring-bench on my VM to /dev/nullb0 and got the following
> soft lockup [1], the reproducer is as simple as:
> 
> modprobe null_blk
> tools/io_uring/io_uring-bench /dev/nullb0
> 
> It looks like io_iopoll_getevents() can hog the cpu, however I don't
> yet really know what is preventing it from quickly exceeding min and
> punting back...
> 
> Adding this makes the problem go away:
> --
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 8b9dbf3b2298..aba03eee5c81 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -779,6 +779,7 @@ static int io_iopoll_getevents(struct io_ring_ctx
> *ctx, unsigned int *nr_events,
>                           return ret;
>                   if (!min || *nr_events >= min)
>                           return 0;
> +               cond_resched();
>           }
> 
>           return 1;
> --
> 
> But I do not know if this is the correct way to fix this, or what
> exactly is the issue, but thought I send it out given its so
> easy to reproduce.

I wonder what your .config is, can you attach it?

Also, please try my for-linus branch, it's got a few tweaks for how
we handle polling (and when we back off). Doesn't affect the inner
loop, so might not change anything for you.

If not, might be better to have a need_resched() terminator in there,
like we have in the outer loop.

-- 
Jens Axboe