io_uring: speedup provided buffer handling

In testing high frequency workloads with provided buffers, we spend a
lot of time in allocating and freeing the buffer units themselves.
Rather than repeatedly free and alloc them, add a recycling cache
instead. There are two caches:

- ctx->io_buffers_cache. This is the one we grab from in the submission
  path, and it's protected by ctx->uring_lock. For inline completions,
  we can recycle straight back to this cache and not need any extra
  locking.

- ctx->io_buffers_comp. If we're not under uring_lock, then we use this
  list to recycle buffers. It's protected by the completion_lock.

On adding a new buffer, check io_buffers_cache. If it's empty, check if
we can splice entries from the io_buffers_comp_cache.

This reduces about 5-10% of overhead from provided buffers, bringing it
pretty close to the non-provided path.

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

1 files changed, 117 insertions, 22 deletions

diff --git a/fs/io_uring.c b/fs/io_uring.c
index daa3609b742f..0f5e999e569f 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -384,6 +384,7 @@ struct io_ring_ctx {
 		struct list_head	ltimeout_list;
 		struct list_head	cq_overflow_list;
 		struct xarray		io_buffers;
+		struct list_head	io_buffers_cache;
 		struct xarray		personalities;
 		u32			pers_next;
 		unsigned		sq_thread_idle;
@@ -426,6 +427,8 @@ struct io_ring_ctx {
 		struct hlist_head	*cancel_hash;
 		unsigned		cancel_hash_bits;
 		bool			poll_multi_queue;
+
+		struct list_head	io_buffers_comp;
 	} ____cacheline_aligned_in_smp;
 
 	struct io_restriction		restrictions;
@@ -441,6 +444,8 @@ struct io_ring_ctx {
 		struct llist_head		rsrc_put_llist;
 		struct list_head		rsrc_ref_list;
 		spinlock_t			rsrc_ref_lock;
+
+		struct list_head	io_buffers_pages;
 	};
 
 	/* Keep this last, we don't need it for the fast path */
@@ -1278,24 +1283,56 @@ static inline void io_req_set_rsrc_node(struct io_kiocb *req,
 	}
 }
 
-static unsigned int __io_put_kbuf(struct io_kiocb *req)
+static unsigned int __io_put_kbuf(struct io_kiocb *req, struct list_head *list)
 {
 	struct io_buffer *kbuf = req->kbuf;
 	unsigned int cflags;
 
-	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
-	cflags |= IORING_CQE_F_BUFFER;
+	cflags = IORING_CQE_F_BUFFER | (kbuf->bid << IORING_CQE_BUFFER_SHIFT);
 	req->flags &= ~REQ_F_BUFFER_SELECTED;
-	kfree(kbuf);
+	list_add(&kbuf->list, list);
 	req->kbuf = NULL;
 	return cflags;
 }
 
-static inline unsigned int io_put_kbuf(struct io_kiocb *req)
+static inline unsigned int io_put_kbuf_comp(struct io_kiocb *req)
 {
 	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
 		return 0;
-	return __io_put_kbuf(req);
+	return __io_put_kbuf(req, &req->ctx->io_buffers_comp);
+}
+
+static inline unsigned int io_put_kbuf(struct io_kiocb *req,
+				       unsigned issue_flags)
+{
+	unsigned int cflags;
+
+	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
+		return 0;
+
+	/*
+	 * We can add this buffer back to two lists:
+	 *
+	 * 1) The io_buffers_cache list. This one is protected by the
+	 *    ctx->uring_lock. If we already hold this lock, add back to this
+	 *    list as we can grab it from issue as well.
+	 * 2) The io_buffers_comp list. This one is protected by the
+	 *    ctx->completion_lock.
+	 *
+	 * We migrate buffers from the comp_list to the issue cache list
+	 * when we need one.
+	 */
+	if (issue_flags & IO_URING_F_UNLOCKED) {
+		struct io_ring_ctx *ctx = req->ctx;
+
+		spin_lock(&ctx->completion_lock);
+		cflags = __io_put_kbuf(req, &ctx->io_buffers_comp);
+		spin_unlock(&ctx->completion_lock);
+	} else {
+		cflags = __io_put_kbuf(req, &req->ctx->io_buffers_cache);
+	}
+
+	return cflags;
 }
 
 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
@@ -1443,6 +1480,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	init_waitqueue_head(&ctx->sqo_sq_wait);
 	INIT_LIST_HEAD(&ctx->sqd_list);
 	INIT_LIST_HEAD(&ctx->cq_overflow_list);
+	INIT_LIST_HEAD(&ctx->io_buffers_cache);
 	init_completion(&ctx->ref_comp);
 	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
 	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
@@ -1451,6 +1489,8 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	spin_lock_init(&ctx->completion_lock);
 	spin_lock_init(&ctx->timeout_lock);
 	INIT_WQ_LIST(&ctx->iopoll_list);
+	INIT_LIST_HEAD(&ctx->io_buffers_pages);
+	INIT_LIST_HEAD(&ctx->io_buffers_comp);
 	INIT_LIST_HEAD(&ctx->defer_list);
 	INIT_LIST_HEAD(&ctx->timeout_list);
 	INIT_LIST_HEAD(&ctx->ltimeout_list);
@@ -2329,7 +2369,8 @@ static void handle_prev_tw_list(struct io_wq_work_node *node,
 		if (likely(*uring_locked))
 			req->io_task_work.func(req, uring_locked);
 		else
-			__io_req_complete_post(req, req->result, io_put_kbuf(req));
+			__io_req_complete_post(req, req->result,
+						io_put_kbuf_comp(req));
 		node = next;
 	} while (node);
 
@@ -2679,7 +2720,7 @@ static int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
 		if (unlikely(req->flags & REQ_F_CQE_SKIP))
 			continue;
 
-		__io_fill_cqe(req, req->result, io_put_kbuf(req));
+		__io_fill_cqe(req, req->result, io_put_kbuf(req, 0));
 		nr_events++;
 	}
 
@@ -2855,14 +2896,14 @@ static bool __io_complete_rw_common(struct io_kiocb *req, long res)
 
 static inline void io_req_task_complete(struct io_kiocb *req, bool *locked)
 {
-	unsigned int cflags = io_put_kbuf(req);
 	int res = req->result;
 
 	if (*locked) {
-		io_req_complete_state(req, res, cflags);
+		io_req_complete_state(req, res, io_put_kbuf(req, 0));
 		io_req_add_compl_list(req);
 	} else {
-		io_req_complete_post(req, res, cflags);
+		io_req_complete_post(req, res,
+					io_put_kbuf(req, IO_URING_F_UNLOCKED));
 	}
 }
 
@@ -2871,7 +2912,8 @@ static void __io_complete_rw(struct io_kiocb *req, long res,
 {
 	if (__io_complete_rw_common(req, res))
 		return;
-	__io_req_complete(req, issue_flags, req->result, io_put_kbuf(req));
+	__io_req_complete(req, issue_flags, req->result,
+				io_put_kbuf(req, issue_flags));
 }
 
 static void io_complete_rw(struct kiocb *kiocb, long res)
@@ -4518,13 +4560,11 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
 
 		nxt = list_first_entry(&buf->list, struct io_buffer, list);
 		list_del(&nxt->list);
-		kfree(nxt);
 		if (++i == nbufs)
 			return i;
 		cond_resched();
 	}
 	i++;
-	kfree(buf);
 	xa_erase(&ctx->io_buffers, bgid);
 
 	return i;
@@ -4590,17 +4630,63 @@ static int io_provide_buffers_prep(struct io_kiocb *req,
 	return 0;
 }
 
-static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
+static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
+{
+	struct io_buffer *buf;
+	struct page *page;
+	int bufs_in_page;
+
+	/*
+	 * Completions that don't happen inline (eg not under uring_lock) will
+	 * add to ->io_buffers_comp. If we don't have any free buffers, check
+	 * the completion list and splice those entries first.
+	 */
+	if (!list_empty_careful(&ctx->io_buffers_comp)) {
+		spin_lock(&ctx->completion_lock);
+		if (!list_empty(&ctx->io_buffers_comp)) {
+			list_splice_init(&ctx->io_buffers_comp,
+						&ctx->io_buffers_cache);
+			spin_unlock(&ctx->completion_lock);
+			return 0;
+		}
+		spin_unlock(&ctx->completion_lock);
+	}
+
+	/*
+	 * No free buffers and no completion entries either. Allocate a new
+	 * page worth of buffer entries and add those to our freelist.
+	 */
+	page = alloc_page(GFP_KERNEL_ACCOUNT);
+	if (!page)
+		return -ENOMEM;
+
+	list_add(&page->lru, &ctx->io_buffers_pages);
+
+	buf = page_address(page);
+	bufs_in_page = PAGE_SIZE / sizeof(*buf);
+	while (bufs_in_page) {
+		list_add_tail(&buf->list, &ctx->io_buffers_cache);
+		buf++;
+		bufs_in_page--;
+	}
+
+	return 0;
+}
+
+static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
+			  struct io_buffer **head)
 {
 	struct io_buffer *buf;
 	u64 addr = pbuf->addr;
 	int i, bid = pbuf->bid;
 
 	for (i = 0; i < pbuf->nbufs; i++) {
-		buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
-		if (!buf)
+		if (list_empty(&ctx->io_buffers_cache) &&
+		    io_refill_buffer_cache(ctx))
 			break;
-
+		buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
+					list);
+		list_del(&buf->list);
 		buf->addr = addr;
 		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
 		buf->bid = bid;
@@ -4632,7 +4718,7 @@ static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
 
 	list = head = xa_load(&ctx->io_buffers, p->bgid);
 
-	ret = io_add_buffers(p, &head);
+	ret = io_add_buffers(ctx, p, &head);
 	if (ret >= 0 && !list) {
 		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
 		if (ret < 0)
@@ -5229,7 +5315,7 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	if (kmsg->free_iov)
 		kfree(kmsg->free_iov);
 	req->flags &= ~REQ_F_NEED_CLEANUP;
-	__io_req_complete(req, issue_flags, ret, io_put_kbuf(req));
+	__io_req_complete(req, issue_flags, ret, io_put_kbuf(req, issue_flags));
 	return 0;
 }
 
@@ -5284,7 +5370,8 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
 out_free:
 		req_set_fail(req);
 	}
-	__io_req_complete(req, issue_flags, ret, io_put_kbuf(req));
+
+	__io_req_complete(req, issue_flags, ret, io_put_kbuf(req, issue_flags));
 	return 0;
 }
 
@@ -6704,7 +6791,7 @@ fail:
 static void io_clean_op(struct io_kiocb *req)
 {
 	if (req->flags & REQ_F_BUFFER_SELECTED)
-		io_put_kbuf(req);
+		io_put_kbuf_comp(req);
 
 	if (req->flags & REQ_F_NEED_CLEANUP) {
 		switch (req->opcode) {
@@ -9475,6 +9562,14 @@ static void io_destroy_buffers(struct io_ring_ctx *ctx)
 
 	xa_for_each(&ctx->io_buffers, index, buf)
 		__io_remove_buffers(ctx, buf, index, -1U);
+
+	while (!list_empty(&ctx->io_buffers_pages)) {
+		struct page *page;
+
+		page = list_first_entry(&ctx->io_buffers_pages, struct page, lru);
+		list_del_init(&page->lru);
+		__free_page(page);
+	}
 }
 
 static void io_req_caches_free(struct io_ring_ctx *ctx)