[PATCH v3] io_uring: limit inflight IO

[PATCH v3] io_uring: limit inflight IO

[Jens Axboe]

With unbounded request times, we can potentially have a lot of IO
inflight. As we provide no real backpressure unless
IORING_SETUP_CQ_NODROP is set, and even there there's quite some delay
between overflows and backpressure being applied, let's put some safety
in place to avoid going way overboard.

This limits the maximum number of inflight IO for any given io_ring_ctx
to twice the CQ ring size. This is a losely managed limit, we only check
for every SQ ring size number of events. That should be good enough to
achieve our goal, which is to prevent massively deep queues. If these
are async requests, they would just be waiting for an execution slot
anyway.

We return -EBUSY if we can't queue anymore IO. The caller should reap
some completions and retry the operation after that. Note that this is
a "should never hit this" kind of condition, as driving the depth into
CQ overflow situations is unreliable.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

Changes since v2:

- Check upfront if we're going over the limit, use the same kind of
  cost amortization logic except something that's appropriate for
  once-per-batch.

- Fold in with the backpressure -EBUSY logic

- Use twice setup CQ ring size as the rough limit

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 81457913e9c9..9dd0f5b5e5b2 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -744,7 +744,7 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
 	struct io_kiocb *req;
 
 	if (!percpu_ref_tryget(&ctx->refs))
-		return NULL;
+		return ERR_PTR(-ENXIO);
 
 	if (!state) {
 		req = kmem_cache_alloc(req_cachep, gfp);
@@ -790,7 +790,7 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
 	if (req)
 		goto got_it;
 	percpu_ref_put(&ctx->refs);
-	return NULL;
+	return ERR_PTR(-EBUSY);
 }
 
 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
@@ -2992,6 +2992,30 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
 	return false;
 }
 
+static bool io_sq_over_limit(struct io_ring_ctx *ctx, unsigned to_submit)
+{
+	unsigned inflight;
+
+	if ((ctx->flags & IORING_SETUP_CQ_NODROP) &&
+	    !list_empty(&ctx->cq_overflow_list))
+		return true;
+
+	/*
+	 * This doesn't need to be super precise, so only check every once
+	 * in a while.
+	 */
+	if ((ctx->cached_sq_head & ctx->sq_mask) !=
+	    ((ctx->cached_sq_head + to_submit) & ctx->sq_mask))
+		return false;
+
+	/*
+	 * Limit us to 2x the CQ ring size
+	 */
+	inflight = ctx->cached_sq_head -
+		  (ctx->cached_cq_tail + atomic_read(&ctx->cached_cq_overflow));
+	return inflight > 2 * ctx->cq_entries;
+}
+
 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 			  struct file *ring_file, int ring_fd,
 			  struct mm_struct **mm, bool async)
@@ -3002,8 +3026,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 	int i, submitted = 0;
 	bool mm_fault = false;
 
-	if ((ctx->flags & IORING_SETUP_CQ_NODROP) &&
-	    !list_empty(&ctx->cq_overflow_list))
+	if (unlikely(io_sq_over_limit(ctx, nr)))
 		return -EBUSY;
 
 	if (nr > IO_PLUG_THRESHOLD) {
@@ -3016,9 +3039,9 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		unsigned int sqe_flags;
 
 		req = io_get_req(ctx, statep);
-		if (unlikely(!req)) {
+		if (unlikely(IS_ERR(req))) {
 			if (!submitted)
-				submitted = -EAGAIN;
+				submitted = PTR_ERR(req);
 			break;
 		}
 		if (!io_get_sqring(ctx, &req->submit)) {
@@ -3039,8 +3062,10 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		if (link && (sqe_flags & IOSQE_IO_DRAIN)) {
 			if (!shadow_req) {
 				shadow_req = io_get_req(ctx, NULL);
-				if (unlikely(!shadow_req))
+				if (unlikely(IS_ERR(shadow_req))) {
+					shadow_req = NULL;
 					goto out;
+				}
 				shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
 				refcount_dec(&shadow_req->refs);
 			}
-- 
Jens Axboe

Re: [PATCH v3] io_uring: limit inflight IO

[Pavel Begunkov]

[-- Attachment #1.1: Type: text/plain, Size: 4589 bytes --]

On 09/11/2019 18:21, Jens Axboe wrote:
> With unbounded request times, we can potentially have a lot of IO
> inflight. As we provide no real backpressure unless
> IORING_SETUP_CQ_NODROP is set, and even there there's quite some delay
> between overflows and backpressure being applied, let's put some safety
> in place to avoid going way overboard.
> 
> This limits the maximum number of inflight IO for any given io_ring_ctx
> to twice the CQ ring size. This is a losely managed limit, we only check
> for every SQ ring size number of events. That should be good enough to
> achieve our goal, which is to prevent massively deep queues. If these
> are async requests, they would just be waiting for an execution slot
> anyway.
> 
> We return -EBUSY if we can't queue anymore IO. The caller should reap
> some completions and retry the operation after that. Note that this is
> a "should never hit this" kind of condition, as driving the depth into
> CQ overflow situations is unreliable.
> 
> Signed-off-by: Jens Axboe <axboe@kernel.dk>
> 
> ---
> 
> Changes since v2:
> 
> - Check upfront if we're going over the limit, use the same kind of
>   cost amortization logic except something that's appropriate for
>   once-per-batch.
> 
> - Fold in with the backpressure -EBUSY logic
> 
> - Use twice setup CQ ring size as the rough limit
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 81457913e9c9..9dd0f5b5e5b2 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -744,7 +744,7 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
>  	struct io_kiocb *req;
>  
>  	if (!percpu_ref_tryget(&ctx->refs))
> -		return NULL;
> +		return ERR_PTR(-ENXIO);
>  
>  	if (!state) {
>  		req = kmem_cache_alloc(req_cachep, gfp);
> @@ -790,7 +790,7 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
>  	if (req)
>  		goto got_it;
>  	percpu_ref_put(&ctx->refs);
> -	return NULL;
> +	return ERR_PTR(-EBUSY);
>  }
>  
>  static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
> @@ -2992,6 +2992,30 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
>  	return false;
>  }
>  
> +static bool io_sq_over_limit(struct io_ring_ctx *ctx, unsigned to_submit)
> +{
> +	unsigned inflight;
> +
> +	if ((ctx->flags & IORING_SETUP_CQ_NODROP) &&
> +	    !list_empty(&ctx->cq_overflow_list))
> +		return true;
> +
> +	/*
> +	 * This doesn't need to be super precise, so only check every once
> +	 * in a while.
> +	 */
> +	if ((ctx->cached_sq_head & ctx->sq_mask) !=
> +	    ((ctx->cached_sq_head + to_submit) & ctx->sq_mask))
> +		return false;

ctx->sq_mask = sq_entries - 1, e.g. 0x0000...ffff.
Thus the code above is modular arithmetic (% sq_entries) and
equivalent to:

if (to_submit != sq_entries)
	return false;
 
I suggest, the intention was:

if ((sq_head & ~mask) == ((sq_head + to_submit) & ~mask))
	return false;


> +
> +	/*
> +	 * Limit us to 2x the CQ ring size
> +	 */
> +	inflight = ctx->cached_sq_head -
> +		  (ctx->cached_cq_tail + atomic_read(&ctx->cached_cq_overflow));
> +	return inflight > 2 * ctx->cq_entries;
> +}
> +
>  static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>  			  struct file *ring_file, int ring_fd,
>  			  struct mm_struct **mm, bool async)
> @@ -3002,8 +3026,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>  	int i, submitted = 0;
>  	bool mm_fault = false;
>  
> -	if ((ctx->flags & IORING_SETUP_CQ_NODROP) &&
> -	    !list_empty(&ctx->cq_overflow_list))
> +	if (unlikely(io_sq_over_limit(ctx, nr)))
>  		return -EBUSY;
>  
>  	if (nr > IO_PLUG_THRESHOLD) {
> @@ -3016,9 +3039,9 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>  		unsigned int sqe_flags;
>  
>  		req = io_get_req(ctx, statep);
> -		if (unlikely(!req)) {
> +		if (unlikely(IS_ERR(req))) {
>  			if (!submitted)
> -				submitted = -EAGAIN;
> +				submitted = PTR_ERR(req);
>  			break;
>  		}
>  		if (!io_get_sqring(ctx, &req->submit)) {
> @@ -3039,8 +3062,10 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>  		if (link && (sqe_flags & IOSQE_IO_DRAIN)) {
>  			if (!shadow_req) {
>  				shadow_req = io_get_req(ctx, NULL);
> -				if (unlikely(!shadow_req))
> +				if (unlikely(IS_ERR(shadow_req))) {
> +					shadow_req = NULL;
>  					goto out;
> +				}
>  				shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
>  				refcount_dec(&shadow_req->refs);
>  			}
> 

-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [PATCH v3] io_uring: limit inflight IO

[Jens Axboe]

On 11/9/19 12:16 PM, Pavel Begunkov wrote:
>> @@ -2992,6 +2992,30 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
>>   	return false;
>>   }
>>   
>> +static bool io_sq_over_limit(struct io_ring_ctx *ctx, unsigned to_submit)
>> +{
>> +	unsigned inflight;
>> +
>> +	if ((ctx->flags & IORING_SETUP_CQ_NODROP) &&
>> +	    !list_empty(&ctx->cq_overflow_list))
>> +		return true;
>> +
>> +	/*
>> +	 * This doesn't need to be super precise, so only check every once
>> +	 * in a while.
>> +	 */
>> +	if ((ctx->cached_sq_head & ctx->sq_mask) !=
>> +	    ((ctx->cached_sq_head + to_submit) & ctx->sq_mask))
>> +		return false;
> 
> ctx->sq_mask = sq_entries - 1, e.g. 0x0000...ffff.
> Thus the code above is modular arithmetic (% sq_entries) and
> equivalent to:
> 
> if (to_submit != sq_entries)
> 	return false;
>   
> I suggest, the intention was:
> 
> if ((sq_head & ~mask) == ((sq_head + to_submit) & ~mask))
> 	return false;

Hah indeed, that was pretty silly. Fixed.

-- 
Jens Axboe