Re: [PATCH v2 2/2] io_uring: flush timeouts that should already have expired

[Pavel Begunkov <asml.silence@gmail.com>]

On 08/01/2021 15:57, Marcelo Diop-Gonzalez wrote:
> On Sat, Jan 02, 2021 at 08:26:26PM +0000, Pavel Begunkov wrote:
>> On 02/01/2021 19:54, Pavel Begunkov wrote:
>>> On 19/12/2020 19:15, Marcelo Diop-Gonzalez wrote:
>>>> Right now io_flush_timeouts() checks if the current number of events
>>>> is equal to ->timeout.target_seq, but this will miss some timeouts if
>>>> there have been more than 1 event added since the last time they were
>>>> flushed (possible in io_submit_flush_completions(), for example). Fix
>>>> it by recording the starting value of ->cached_cq_overflow -
>>>> ->cq_timeouts instead of the target value, so that we can safely
>>>> (without overflow problems) compare the number of events that have
>>>> happened with the number of events needed to trigger the timeout.
>>
>> https://www.spinics.net/lists/kernel/msg3475160.html
>>
>> The idea was to replace u32 cached_cq_tail with u64 while keeping
>> timeout offsets u32. Assuming that we won't ever hit ~2^62 inflight
>> requests, complete all requests falling into some large enough window
>> behind that u64 cached_cq_tail.
>>
>> simplifying:
>>
>> i64 d = target_off - ctx->u64_cq_tail
>> if (d <= 0 && d > -2^32)
>> 	complete_it()
>>
>> Not fond  of it, but at least worked at that time. You can try out
>> this approach if you want, but would be perfect if you would find
>> something more elegant :)
>>
> 
> What do you think about something like this? I think it's not totally
> correct because it relies on having ->completion_lock in io_timeout() so
> that ->cq_last_tm_flushed is updated, but in case of IORING_SETUP_IOPOLL,
> io_iopoll_complete() doesn't take that lock, and ->uring_lock will not
> be held if io_timeout() is called from io_wq_submit_work(), but maybe
> could still be worth it since that was already possibly a problem?
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index cb57e0360fcb..50984709879c 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -353,6 +353,7 @@ struct io_ring_ctx {
>  		unsigned		cq_entries;
>  		unsigned		cq_mask;
>  		atomic_t		cq_timeouts;
> +		unsigned		cq_last_tm_flush;

It looks like that "last flush" is a good direction.
I think there can be problems at extremes like completing 2^32
requests at once, but should be ok in practice. Anyway better
than it's now.

What about the first patch about overflows and cq_timeouts? I
assume that problem is still there, isn't it?

See comments below, but if it passes liburing tests, please send
a patch.

>  		unsigned long		cq_check_overflow;
>  		struct wait_queue_head	cq_wait;
>  		struct fasync_struct	*cq_fasync;
> @@ -1633,19 +1634,26 @@ static void __io_queue_deferred(struct io_ring_ctx *ctx)
>  
>  static void io_flush_timeouts(struct io_ring_ctx *ctx)
>  {
> +	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
> +

a nit,

if (list_empty()) return; + do {} while();

timeouts can be rare enough

>  	while (!list_empty(&ctx->timeout_list)) {
> +		u32 events_needed, events_got;
>  		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
>  						struct io_kiocb, timeout.list);
>  
>  		if (io_is_timeout_noseq(req))
>  			break;
> -		if (req->timeout.target_seq != ctx->cached_cq_tail
> -					- atomic_read(&ctx->cq_timeouts))
> +

extra new line

> +		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
> +		events_got = seq - ctx->cq_last_tm_flush;
> +		if (events_got < events_needed) 

probably <=

>  			break;

basically it checks that @target is in [last_flush, cur_seq],
it can use such a comment + a note about underflows and using
the modulus arithmetic, like with algebraic rings

>  
>  		list_del_init(&req->timeout.list);
>  		io_kill_timeout(req);
>  	}
> +
> +	ctx->cq_last_tm_flush = seq;
>  }
>  
>  static void io_commit_cqring(struct io_ring_ctx *ctx)
> 

-- 
Pavel Begunkov