io_uring: use percpu counters to track inflight requests

Even though we place the req_issued and req_complete in separate
cachelines, there's considerable overhead in doing the atomics
particularly on the completion side.

Get rid of having the two counters, and just use a percpu_counter for
this. That's what it was made for, after all. This considerably
reduces the overhead in __io_free_req().

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

fs/io_uring.c

@@ -1810,7 +1810,7 @@ static void __io_free_req(struct io_kiocb *req)
 
 	io_dismantle_req(req);
 
-	atomic_long_inc(&tctx->req_complete);
+	percpu_counter_dec(&tctx->inflight);
 	if (tctx->in_idle)
 		wake_up(&tctx->wait);
 	put_task_struct(req->task);
@@ -2089,7 +2089,9 @@ static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
 	if (rb->to_free)
 		__io_req_free_batch_flush(ctx, rb);
 	if (rb->task) {
-		atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
+		struct io_uring_task *tctx = rb->task->io_uring;
+
+		percpu_counter_sub(&tctx->inflight, rb->task_refs);
 		put_task_struct_many(rb->task, rb->task_refs);
 		rb->task = NULL;
 	}
@@ -2106,7 +2108,9 @@ static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
 
 	if (req->task != rb->task) {
 		if (rb->task) {
-			atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
+			struct io_uring_task *tctx = rb->task->io_uring;
+
+			percpu_counter_sub(&tctx->inflight, rb->task_refs);
 			put_task_struct_many(rb->task, rb->task_refs);
 		}
 		rb->task = req->task;
@@ -6524,7 +6528,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
 	if (!percpu_ref_tryget_many(&ctx->refs, nr))
 		return -EAGAIN;
 
-	atomic_long_add(nr, &current->io_uring->req_issue);
+	percpu_counter_add(&current->io_uring->inflight, nr);
 	refcount_add(nr, &current->usage);
 
 	io_submit_state_start(&state, ctx, nr);
@@ -6566,10 +6570,12 @@ fail_req:
 
 	if (unlikely(submitted != nr)) {
 		int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
+		struct io_uring_task *tctx = current->io_uring;
+		int unused = nr - ref_used;
 
-		percpu_ref_put_many(&ctx->refs, nr - ref_used);
-		atomic_long_sub(nr - ref_used, &current->io_uring->req_issue);
-		put_task_struct_many(current, nr - ref_used);
+		percpu_ref_put_many(&ctx->refs, unused);
+		percpu_counter_sub(&tctx->inflight, unused);
+		put_task_struct_many(current, unused);
 	}
 	if (link)
 		io_queue_link_head(link, &state.comp);
@@ -7687,17 +7693,22 @@ out_fput:
 static int io_uring_alloc_task_context(struct task_struct *task)
 {
 	struct io_uring_task *tctx;
+	int ret;
 
 	tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
 	if (unlikely(!tctx))
 		return -ENOMEM;
 
+	ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
+	if (unlikely(ret)) {
+		kfree(tctx);
+		return ret;
+	}
+
 	xa_init(&tctx->xa);
 	init_waitqueue_head(&tctx->wait);
 	tctx->last = NULL;
 	tctx->in_idle = 0;
-	atomic_long_set(&tctx->req_issue, 0);
-	atomic_long_set(&tctx->req_complete, 0);
 	io_init_identity(&tctx->__identity);
 	tctx->identity = &tctx->__identity;
 	task->io_uring = tctx;
@@ -7712,6 +7723,7 @@ void __io_uring_free(struct task_struct *tsk)
 	WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
 	if (tctx->identity != &tctx->__identity)
 		kfree(tctx->identity);
+	percpu_counter_destroy(&tctx->inflight);
 	kfree(tctx);
 	tsk->io_uring = NULL;
 }
@@ -8696,12 +8708,6 @@ void __io_uring_files_cancel(struct files_struct *files)
 	}
 }
 
-static inline bool io_uring_task_idle(struct io_uring_task *tctx)
-{
-	return atomic_long_read(&tctx->req_issue) ==
-		atomic_long_read(&tctx->req_complete);
-}
-
 /*
  * Find any io_uring fd that this task has registered or done IO on, and cancel
  * requests.
@@ -8710,14 +8716,16 @@ void __io_uring_task_cancel(void)
 {
 	struct io_uring_task *tctx = current->io_uring;
 	DEFINE_WAIT(wait);
-	long completions;
+	s64 inflight;
 
 	/* make sure overflow events are dropped */
 	tctx->in_idle = true;
 
-	while (!io_uring_task_idle(tctx)) {
+	do {
 		/* read completions before cancelations */
-		completions = atomic_long_read(&tctx->req_complete);
+		inflight = percpu_counter_sum(&tctx->inflight);
+		if (!inflight)
+			break;
 		__io_uring_files_cancel(NULL);
 
 		prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
@@ -8726,12 +8734,10 @@ void __io_uring_task_cancel(void)
 		 * If we've seen completions, retry. This avoids a race where
 		 * a completion comes in before we did prepare_to_wait().
 		 */
-		if (completions != atomic_long_read(&tctx->req_complete))
+		if (inflight != percpu_counter_sum(&tctx->inflight))
 			continue;
-		if (io_uring_task_idle(tctx))
-			break;
 		schedule();
-	}
+	} while (1);
 
 	finish_wait(&tctx->wait, &wait);
 	tctx->in_idle = false;

include/linux/io_uring.h

@@ -23,13 +23,10 @@ struct io_uring_task {
 	struct xarray		xa;
 	struct wait_queue_head	wait;
 	struct file		*last;
-	atomic_long_t		req_issue;
+	struct percpu_counter	inflight;
 	struct io_identity	__identity;
 	struct io_identity	*identity;
-
-	/* completion side */
-	bool			in_idle ____cacheline_aligned_in_smp;
-	atomic_long_t		req_complete;
+	bool			in_idle;
 };
 
 #if defined(CONFIG_IO_URING)