io_uring: delay sqarray static branch disablement

io_key_has_sqarray static branch can be easily switched on/off by the
user every time patching the kernel. That can be very disruptive as it
might require heavy synchronisation across all CPUs. Use deferred static
keys, which can rate-limit it by deferring, batching and potentially
effectively eliminating dec+inc pairs.

Fixes: 9b296c6 ("io_uring: static_key for !IORING_SETUP_NO_SQARRAY")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Author: isilence

io_uring/io_uring.c

@@ -119,7 +119,7 @@
 static void io_queue_sqe(struct io_kiocb *req, unsigned int extra_flags);
 static void __io_req_caches_free(struct io_ring_ctx *ctx);
 
-static __read_mostly DEFINE_STATIC_KEY_FALSE(io_key_has_sqarray);
+static __read_mostly DEFINE_STATIC_KEY_DEFERRED_FALSE(io_key_has_sqarray, HZ);
 
 struct kmem_cache *req_cachep;
 static struct workqueue_struct *iou_wq __ro_after_init;
@@ -1978,7 +1978,7 @@ static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
 	unsigned mask = ctx->sq_entries - 1;
 	unsigned head = ctx->cached_sq_head++ & mask;
 
-	if (static_branch_unlikely(&io_key_has_sqarray) &&
+	if (static_branch_unlikely(&io_key_has_sqarray.key) &&
 	    (!(ctx->flags & IORING_SETUP_NO_SQARRAY))) {
 		head = READ_ONCE(ctx->sq_array[head]);
 		if (unlikely(head >= ctx->sq_entries)) {
@@ -2173,7 +2173,7 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	io_rings_free(ctx);
 
 	if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
-		static_branch_dec(&io_key_has_sqarray);
+		static_branch_slow_dec_deferred(&io_key_has_sqarray);
 
 	percpu_ref_exit(&ctx->refs);
 	free_uid(ctx->user);
@@ -2946,7 +2946,7 @@ static __cold int io_uring_create(struct io_ctx_config *config)
 	ctx->clock_offset = 0;
 
 	if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
-		static_branch_inc(&io_key_has_sqarray);
+		static_branch_deferred_inc(&io_key_has_sqarray);
 
 	if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
 	    !(ctx->flags & IORING_SETUP_IOPOLL))