[PATCHSET for-next] Add alloc cache for sendmsg/recvmsg

[PATCHSET for-next] Add alloc cache for sendmsg/recvmsg

[Jens Axboe]

Hi,

This abstracts out the alloc cache we have for apoll entries, and extends
it to be usable for recv/sendmsg as well. During that abstraction, impose
an upper limit for cached entries as well.

This yields a 4-5% performance increase running netbench using
sendmsg/recvmsg rather than plan send and recv.

Post 5.20, I suspect we can get rid of using io_async_msghdr for single
vector sendmsg/recvmsg, which will make this less relevant. But as this
work isn't done yet, and the support for eg ITER_UBUF isn't upstream yet
either, this can help fill the gap.

-- 
Jens Axboe

[PATCH 1/4] io_uring: move apoll cache to poll.c

[Jens Axboe]

This is where it's used, move the flush handler in there.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 io_uring/io_uring.c | 12 ------------
 io_uring/poll.c     | 12 ++++++++++++
 io_uring/poll.h     |  2 ++
 3 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index caf979cd4327..4d1ce58b015e 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -2445,18 +2445,6 @@ static void io_req_caches_free(struct io_ring_ctx *ctx)
 	mutex_unlock(&ctx->uring_lock);
 }
 
-static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
-{
-	struct async_poll *apoll;
-
-	while (!list_empty(&ctx->apoll_cache)) {
-		apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
-						poll.wait.entry);
-		list_del(&apoll->poll.wait.entry);
-		kfree(apoll);
-	}
-}
-
 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
 {
 	io_sq_thread_finish(ctx);
diff --git a/io_uring/poll.c b/io_uring/poll.c
index 57747d92bba4..f0fe209490d8 100644
--- a/io_uring/poll.c
+++ b/io_uring/poll.c
@@ -958,3 +958,15 @@ int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
 	io_req_set_res(req, ret, 0);
 	return IOU_OK;
 }
+
+void io_flush_apoll_cache(struct io_ring_ctx *ctx)
+{
+	struct async_poll *apoll;
+
+	while (!list_empty(&ctx->apoll_cache)) {
+		apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
+						poll.wait.entry);
+		list_del(&apoll->poll.wait.entry);
+		kfree(apoll);
+	}
+}
diff --git a/io_uring/poll.h b/io_uring/poll.h
index c40673d7da01..95f192c7babb 100644
--- a/io_uring/poll.h
+++ b/io_uring/poll.h
@@ -30,3 +30,5 @@ int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags);
 bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
 			bool cancel_all);
+
+void io_flush_apoll_cache(struct io_ring_ctx *ctx);
-- 
2.35.1

[PATCH 2/4] io_uring: add abstraction around apoll cache

[Jens Axboe]

In preparation for adding limits, and one more user, abstract out the
core bits of the allocation+free cache.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/io_uring_types.h |  6 +++++-
 io_uring/alloc_cache.h         |  4 ++++
 io_uring/io_uring.c            |  7 ++++---
 io_uring/poll.c                | 16 ++++++++--------
 io_uring/poll.h                |  5 ++++-
 5 files changed, 25 insertions(+), 13 deletions(-)
 create mode 100644 io_uring/alloc_cache.h

diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index 26ef11e978d4..b548da03b563 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -158,6 +158,10 @@ struct io_ev_fd {
 	struct rcu_head		rcu;
 };
 
+struct io_alloc_cache {
+	struct hlist_head	list;
+};
+
 struct io_ring_ctx {
 	/* const or read-mostly hot data */
 	struct {
@@ -216,7 +220,7 @@ struct io_ring_ctx {
 
 		struct io_hash_table	cancel_table_locked;
 		struct list_head	cq_overflow_list;
-		struct list_head	apoll_cache;
+		struct io_alloc_cache	apoll_cache;
 		struct xarray		personalities;
 		u32			pers_next;
 	} ____cacheline_aligned_in_smp;
diff --git a/io_uring/alloc_cache.h b/io_uring/alloc_cache.h
new file mode 100644
index 000000000000..49ac6ae237ef
--- /dev/null
+++ b/io_uring/alloc_cache.h
@@ -0,0 +1,4 @@
+static inline void io_alloc_cache_init(struct io_alloc_cache *cache)
+{
+	INIT_HLIST_HEAD(&cache->list);
+}
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 4d1ce58b015e..3b9033c401bf 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -92,6 +92,7 @@
 
 #include "timeout.h"
 #include "poll.h"
+#include "alloc_cache.h"
 
 #define IORING_MAX_ENTRIES	32768
 #define IORING_MAX_CQ_ENTRIES	(2 * IORING_MAX_ENTRIES)
@@ -295,7 +296,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	INIT_LIST_HEAD(&ctx->sqd_list);
 	INIT_LIST_HEAD(&ctx->cq_overflow_list);
 	INIT_LIST_HEAD(&ctx->io_buffers_cache);
-	INIT_LIST_HEAD(&ctx->apoll_cache);
+	io_alloc_cache_init(&ctx->apoll_cache);
 	init_completion(&ctx->ref_comp);
 	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
 	mutex_init(&ctx->uring_lock);
@@ -1180,8 +1181,8 @@ void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
 
 				if (apoll->double_poll)
 					kfree(apoll->double_poll);
-				list_add(&apoll->poll.wait.entry,
-						&ctx->apoll_cache);
+				hlist_add_head(&apoll->cache_list,
+						&ctx->apoll_cache.list);
 				req->flags &= ~REQ_F_POLLED;
 			}
 			if (req->flags & IO_REQ_LINK_FLAGS)
diff --git a/io_uring/poll.c b/io_uring/poll.c
index f0fe209490d8..f3aae3cc6501 100644
--- a/io_uring/poll.c
+++ b/io_uring/poll.c
@@ -589,10 +589,10 @@ static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
 		apoll = req->apoll;
 		kfree(apoll->double_poll);
 	} else if (!(issue_flags & IO_URING_F_UNLOCKED) &&
-		   !list_empty(&ctx->apoll_cache)) {
-		apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
-						poll.wait.entry);
-		list_del_init(&apoll->poll.wait.entry);
+		   !hlist_empty(&ctx->apoll_cache.list)) {
+		apoll = hlist_entry(ctx->apoll_cache.list.first,
+						struct async_poll, cache_list);
+		hlist_del(&apoll->cache_list);
 	} else {
 		apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
 		if (unlikely(!apoll))
@@ -963,10 +963,10 @@ void io_flush_apoll_cache(struct io_ring_ctx *ctx)
 {
 	struct async_poll *apoll;
 
-	while (!list_empty(&ctx->apoll_cache)) {
-		apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
-						poll.wait.entry);
-		list_del(&apoll->poll.wait.entry);
+	while (!hlist_empty(&ctx->apoll_cache.list)) {
+		apoll = hlist_entry(ctx->apoll_cache.list.first,
+						struct async_poll, cache_list);
+		hlist_del(&apoll->cache_list);
 		kfree(apoll);
 	}
 }
diff --git a/io_uring/poll.h b/io_uring/poll.h
index 95f192c7babb..cb528f8ef203 100644
--- a/io_uring/poll.h
+++ b/io_uring/poll.h
@@ -14,7 +14,10 @@ struct io_poll {
 };
 
 struct async_poll {
-	struct io_poll		poll;
+	union {
+		struct io_poll		poll;
+		struct hlist_node	cache_list;
+	};
 	struct io_poll		*double_poll;
 };
 
-- 
2.35.1

[PATCH 3/4] io_uring: impose max limit on apoll cache

[Jens Axboe]

Caches like this tend to grow to the peak size, and then never get any
smaller. Impose a max limit on the size, to prevent it from growing too
big.

A somewhat randomly chosen 512 is the max size we'll allow the cache
to get. If a batch of frees come in and would bring it over that, we
simply start kfree'ing the surplus.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/io_uring_types.h |  1 +
 io_uring/alloc_cache.h         | 15 +++++++++++++++
 io_uring/io_uring.c            |  8 ++++++--
 io_uring/poll.c                |  2 ++
 4 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index b548da03b563..bf8f95332eda 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -160,6 +160,7 @@ struct io_ev_fd {
 
 struct io_alloc_cache {
 	struct hlist_head	list;
+	unsigned int		nr_cached;
 };
 
 struct io_ring_ctx {
diff --git a/io_uring/alloc_cache.h b/io_uring/alloc_cache.h
index 49ac6ae237ef..0e64030f1ae0 100644
--- a/io_uring/alloc_cache.h
+++ b/io_uring/alloc_cache.h
@@ -1,4 +1,19 @@
+/*
+ * Don't allow the cache to grow beyond this size.
+ */
+#define IO_ALLOC_CACHE_MAX	512
+
+static inline bool io_alloc_cache_store(struct io_alloc_cache *cache)
+{
+	if (cache->nr_cached < IO_ALLOC_CACHE_MAX) {
+		cache->nr_cached++;
+		return true;
+	}
+	return false;
+}
+
 static inline void io_alloc_cache_init(struct io_alloc_cache *cache)
 {
 	INIT_HLIST_HEAD(&cache->list);
+	cache->nr_cached = 0;
 }
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 3b9033c401bf..b5098773d924 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -1181,8 +1181,12 @@ void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
 
 				if (apoll->double_poll)
 					kfree(apoll->double_poll);
-				hlist_add_head(&apoll->cache_list,
-						&ctx->apoll_cache.list);
+				if (io_alloc_cache_store(&ctx->apoll_cache)) {
+					hlist_add_head(&apoll->cache_list,
+							&ctx->apoll_cache.list);
+				} else {
+					kfree(apoll);
+				}
 				req->flags &= ~REQ_F_POLLED;
 			}
 			if (req->flags & IO_REQ_LINK_FLAGS)
diff --git a/io_uring/poll.c b/io_uring/poll.c
index f3aae3cc6501..cc49160975cb 100644
--- a/io_uring/poll.c
+++ b/io_uring/poll.c
@@ -593,6 +593,7 @@ static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
 		apoll = hlist_entry(ctx->apoll_cache.list.first,
 						struct async_poll, cache_list);
 		hlist_del(&apoll->cache_list);
+		ctx->apoll_cache.nr_cached--;
 	} else {
 		apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
 		if (unlikely(!apoll))
@@ -969,4 +970,5 @@ void io_flush_apoll_cache(struct io_ring_ctx *ctx)
 		hlist_del(&apoll->cache_list);
 		kfree(apoll);
 	}
+	ctx->apoll_cache.nr_cached = 0;
 }
-- 
2.35.1

[PATCH 4/4] io_uring: add netmsg cache

[Jens Axboe]

For recvmsg/sendmsg, if they don't complete inline, we currently need
to allocate a struct io_async_msghdr for each request. This is a
somewhat large struct.

Hook up sendmsg/recvmsg to use the io_alloc_cache. This reduces the
alloc + free overhead considerably, yielding 4-5% of extra performance
running netbench.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/io_uring_types.h |  6 ++-
 io_uring/io_uring.c            |  3 ++
 io_uring/net.c                 | 73 +++++++++++++++++++++++++++++-----
 io_uring/net.h                 | 11 ++++-
 4 files changed, 81 insertions(+), 12 deletions(-)

diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index bf8f95332eda..d54b8b7e0746 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -222,8 +222,7 @@ struct io_ring_ctx {
 		struct io_hash_table	cancel_table_locked;
 		struct list_head	cq_overflow_list;
 		struct io_alloc_cache	apoll_cache;
-		struct xarray		personalities;
-		u32			pers_next;
+		struct io_alloc_cache	netmsg_cache;
 	} ____cacheline_aligned_in_smp;
 
 	/* IRQ completion list, under ->completion_lock */
@@ -241,6 +240,9 @@ struct io_ring_ctx {
 	unsigned int		file_alloc_start;
 	unsigned int		file_alloc_end;
 
+	struct xarray		personalities;
+	u32			pers_next;
+
 	struct {
 		/*
 		 * We cache a range of free CQEs we can use, once exhausted it
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index b5098773d924..32110c5b4059 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -89,6 +89,7 @@
 #include "kbuf.h"
 #include "rsrc.h"
 #include "cancel.h"
+#include "net.h"
 
 #include "timeout.h"
 #include "poll.h"
@@ -297,6 +298,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	INIT_LIST_HEAD(&ctx->cq_overflow_list);
 	INIT_LIST_HEAD(&ctx->io_buffers_cache);
 	io_alloc_cache_init(&ctx->apoll_cache);
+	io_alloc_cache_init(&ctx->netmsg_cache);
 	init_completion(&ctx->ref_comp);
 	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
 	mutex_init(&ctx->uring_lock);
@@ -2473,6 +2475,7 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
 		__io_cqring_overflow_flush(ctx, true);
 	io_eventfd_unregister(ctx);
 	io_flush_apoll_cache(ctx);
+	io_flush_netmsg_cache(ctx);
 	mutex_unlock(&ctx->uring_lock);
 	io_destroy_buffers(ctx);
 	if (ctx->sq_creds)
diff --git a/io_uring/net.c b/io_uring/net.c
index 6679069eeef1..ba7e94ff287c 100644
--- a/io_uring/net.c
+++ b/io_uring/net.c
@@ -12,6 +12,7 @@
 
 #include "io_uring.h"
 #include "kbuf.h"
+#include "alloc_cache.h"
 #include "net.h"
 
 #if defined(CONFIG_NET)
@@ -97,18 +98,57 @@ static bool io_net_retry(struct socket *sock, int flags)
 	return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
 }
 
+static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
+{
+	struct io_async_msghdr *hdr = req->async_data;
+
+	if (!hdr || issue_flags & IO_URING_F_UNLOCKED)
+		return;
+
+	if (io_alloc_cache_store(&req->ctx->netmsg_cache)) {
+		hlist_add_head(&hdr->cache_list, &req->ctx->netmsg_cache.list);
+		req->async_data = NULL;
+		req->flags &= ~REQ_F_ASYNC_DATA;
+	}
+}
+
+static struct io_async_msghdr *io_recvmsg_alloc_async(struct io_kiocb *req,
+						      unsigned int issue_flags)
+{
+	struct io_ring_ctx *ctx = req->ctx;
+
+	if (!(issue_flags & IO_URING_F_UNLOCKED) &&
+	    !hlist_empty(&ctx->netmsg_cache.list)) {
+		struct io_async_msghdr *hdr;
+
+		hdr = hlist_entry(ctx->netmsg_cache.list.first,
+					struct io_async_msghdr, cache_list);
+		ctx->netmsg_cache.nr_cached--;
+		hlist_del(&hdr->cache_list);
+		req->flags |= REQ_F_ASYNC_DATA;
+		req->async_data = hdr;
+		return hdr;
+	}
+
+	if (!io_alloc_async_data(req))
+		return req->async_data;
+
+	return NULL;
+}
+
 static int io_setup_async_msg(struct io_kiocb *req,
-			      struct io_async_msghdr *kmsg)
+			      struct io_async_msghdr *kmsg,
+			      unsigned int issue_flags)
 {
 	struct io_async_msghdr *async_msg = req->async_data;
 
 	if (async_msg)
 		return -EAGAIN;
-	if (io_alloc_async_data(req)) {
+	async_msg = io_recvmsg_alloc_async(req, issue_flags);
+	if (!async_msg) {
 		kfree(kmsg->free_iov);
 		return -ENOMEM;
 	}
-	async_msg = req->async_data;
 	req->flags |= REQ_F_NEED_CLEANUP;
 	memcpy(async_msg, kmsg, sizeof(*kmsg));
 	async_msg->msg.msg_name = &async_msg->addr;
@@ -195,7 +235,7 @@ int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
 
 	if (!(req->flags & REQ_F_POLLED) &&
 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
-		return io_setup_async_msg(req, kmsg);
+		return io_setup_async_msg(req, kmsg, issue_flags);
 
 	flags = sr->msg_flags;
 	if (issue_flags & IO_URING_F_NONBLOCK)
@@ -207,13 +247,13 @@ int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
 
 	if (ret < min_ret) {
 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
-			return io_setup_async_msg(req, kmsg);
+			return io_setup_async_msg(req, kmsg, issue_flags);
 		if (ret == -ERESTARTSYS)
 			ret = -EINTR;
 		if (ret > 0 && io_net_retry(sock, flags)) {
 			sr->done_io += ret;
 			req->flags |= REQ_F_PARTIAL_IO;
-			return io_setup_async_msg(req, kmsg);
+			return io_setup_async_msg(req, kmsg, issue_flags);
 		}
 		req_set_fail(req);
 	}
@@ -221,6 +261,7 @@ int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
 	if (kmsg->free_iov)
 		kfree(kmsg->free_iov);
 	req->flags &= ~REQ_F_NEED_CLEANUP;
+	io_netmsg_recycle(req, issue_flags);
 	if (ret >= 0)
 		ret += sr->done_io;
 	else if (sr->done_io)
@@ -495,7 +536,7 @@ int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 
 	if (!(req->flags & REQ_F_POLLED) &&
 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
-		return io_setup_async_msg(req, kmsg);
+		return io_setup_async_msg(req, kmsg, issue_flags);
 
 	if (io_do_buffer_select(req)) {
 		void __user *buf;
@@ -519,13 +560,13 @@ int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg, kmsg->uaddr, flags);
 	if (ret < min_ret) {
 		if (ret == -EAGAIN && force_nonblock)
-			return io_setup_async_msg(req, kmsg);
+			return io_setup_async_msg(req, kmsg, issue_flags);
 		if (ret == -ERESTARTSYS)
 			ret = -EINTR;
 		if (ret > 0 && io_net_retry(sock, flags)) {
 			sr->done_io += ret;
 			req->flags |= REQ_F_PARTIAL_IO;
-			return io_setup_async_msg(req, kmsg);
+			return io_setup_async_msg(req, kmsg, issue_flags);
 		}
 		req_set_fail(req);
 	} else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
@@ -535,6 +576,7 @@ int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	/* fast path, check for non-NULL to avoid function call */
 	if (kmsg->free_iov)
 		kfree(kmsg->free_iov);
+	io_netmsg_recycle(req, issue_flags);
 	req->flags &= ~REQ_F_NEED_CLEANUP;
 	if (ret > 0)
 		ret += sr->done_io;
@@ -848,4 +890,17 @@ int io_connect(struct io_kiocb *req, unsigned int issue_flags)
 	io_req_set_res(req, ret, 0);
 	return IOU_OK;
 }
+
+void io_flush_netmsg_cache(struct io_ring_ctx *ctx)
+{
+	while (!hlist_empty(&ctx->netmsg_cache.list)) {
+		struct io_async_msghdr *hdr;
+
+		hdr = hlist_entry(ctx->netmsg_cache.list.first,
+					struct io_async_msghdr, cache_list);
+		hlist_del(&hdr->cache_list);
+		kfree(hdr);
+	}
+	ctx->netmsg_cache.nr_cached = 0;
+}
 #endif
diff --git a/io_uring/net.h b/io_uring/net.h
index 81d71d164770..576efb602c7f 100644
--- a/io_uring/net.h
+++ b/io_uring/net.h
@@ -5,7 +5,10 @@
 
 #if defined(CONFIG_NET)
 struct io_async_msghdr {
-	struct iovec			fast_iov[UIO_FASTIOV];
+	union {
+		struct iovec		fast_iov[UIO_FASTIOV];
+		struct hlist_node	cache_list;
+	};
 	/* points to an allocated iov, if NULL we use fast_iov instead */
 	struct iovec			*free_iov;
 	struct sockaddr __user		*uaddr;
@@ -40,4 +43,10 @@ int io_socket(struct io_kiocb *req, unsigned int issue_flags);
 int io_connect_prep_async(struct io_kiocb *req);
 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
 int io_connect(struct io_kiocb *req, unsigned int issue_flags);
+
+void io_flush_netmsg_cache(struct io_ring_ctx *ctx);
+#else
+static inline void io_flush_netmsg_cache(struct io_ring_ctx *ctx)
+{
+}
 #endif
-- 
2.35.1

Re: [PATCH 4/4] io_uring: add netmsg cache

[Dylan Yudaken]

On Thu, 2022-07-07 at 17:23 -0600, Jens Axboe wrote:
> For recvmsg/sendmsg, if they don't complete inline, we currently need
> to allocate a struct io_async_msghdr for each request. This is a
> somewhat large struct.
> 
> Hook up sendmsg/recvmsg to use the io_alloc_cache. This reduces the
> alloc + free overhead considerably, yielding 4-5% of extra
> performance
> running netbench.
> 
> Signed-off-by: Jens Axboe <axboe@kernel.dk>
> ---
>  include/linux/io_uring_types.h |  6 ++-
>  io_uring/io_uring.c            |  3 ++
>  io_uring/net.c                 | 73 +++++++++++++++++++++++++++++---
> --
>  io_uring/net.h                 | 11 ++++-
>  4 files changed, 81 insertions(+), 12 deletions(-)
> 
> diff --git a/include/linux/io_uring_types.h
> b/include/linux/io_uring_types.h
> index bf8f95332eda..d54b8b7e0746 100644
> --- a/include/linux/io_uring_types.h
> +++ b/include/linux/io_uring_types.h
> @@ -222,8 +222,7 @@ struct io_ring_ctx {
>                 struct io_hash_table    cancel_table_locked;
>                 struct list_head        cq_overflow_list;
>                 struct io_alloc_cache   apoll_cache;
> -               struct xarray           personalities;
> -               u32                     pers_next;
> +               struct io_alloc_cache   netmsg_cache;
>         } ____cacheline_aligned_in_smp;
>  
>         /* IRQ completion list, under ->completion_lock */
> @@ -241,6 +240,9 @@ struct io_ring_ctx {
>         unsigned int            file_alloc_start;
>         unsigned int            file_alloc_end;
>  
> +       struct xarray           personalities;
> +       u32                     pers_next;
> +
>         struct {
>                 /*
>                  * We cache a range of free CQEs we can use, once
> exhausted it
> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
> index b5098773d924..32110c5b4059 100644
> --- a/io_uring/io_uring.c
> +++ b/io_uring/io_uring.c
> @@ -89,6 +89,7 @@
>  #include "kbuf.h"
>  #include "rsrc.h"
>  #include "cancel.h"
> +#include "net.h"
>  
>  #include "timeout.h"
>  #include "poll.h"
> @@ -297,6 +298,7 @@ static __cold struct io_ring_ctx
> *io_ring_ctx_alloc(struct io_uring_params *p)
>         INIT_LIST_HEAD(&ctx->cq_overflow_list);
>         INIT_LIST_HEAD(&ctx->io_buffers_cache);
>         io_alloc_cache_init(&ctx->apoll_cache);
> +       io_alloc_cache_init(&ctx->netmsg_cache);
>         init_completion(&ctx->ref_comp);
>         xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
>         mutex_init(&ctx->uring_lock);
> @@ -2473,6 +2475,7 @@ static __cold void io_ring_ctx_free(struct
> io_ring_ctx *ctx)
>                 __io_cqring_overflow_flush(ctx, true);
>         io_eventfd_unregister(ctx);
>         io_flush_apoll_cache(ctx);
> +       io_flush_netmsg_cache(ctx);
>         mutex_unlock(&ctx->uring_lock);
>         io_destroy_buffers(ctx);
>         if (ctx->sq_creds)
> diff --git a/io_uring/net.c b/io_uring/net.c
> index 6679069eeef1..ba7e94ff287c 100644
> --- a/io_uring/net.c
> +++ b/io_uring/net.c
> @@ -12,6 +12,7 @@
>  
>  #include "io_uring.h"
>  #include "kbuf.h"
> +#include "alloc_cache.h"
>  #include "net.h"
>  
>  #if defined(CONFIG_NET)
> @@ -97,18 +98,57 @@ static bool io_net_retry(struct socket *sock, int
> flags)
>         return sock->type == SOCK_STREAM || sock->type ==
> SOCK_SEQPACKET;
>  }
>  
> +static void io_netmsg_recycle(struct io_kiocb *req, unsigned int
> issue_flags)
> +{
> +       struct io_async_msghdr *hdr = req->async_data;
> +
> +       if (!hdr || issue_flags & IO_URING_F_UNLOCKED)
> +               return;
> +
> +       if (io_alloc_cache_store(&req->ctx->netmsg_cache)) {
> +               hlist_add_head(&hdr->cache_list, &req->ctx-
> >netmsg_cache.list);

can io_alloc_cache_store just do the store?
would be nicer to have cache::list be generally unused outside of the
cache code.

> +               req->async_data = NULL;
> +               req->flags &= ~REQ_F_ASYNC_DATA;
> +       }
> +}
> +
> +static struct io_async_msghdr *io_recvmsg_alloc_async(struct
> io_kiocb *req,
> +                                                     unsigned int
> issue_flags)
> +{
> +       struct io_ring_ctx *ctx = req->ctx;
> +
> +       if (!(issue_flags & IO_URING_F_UNLOCKED) &&
> +           !hlist_empty(&ctx->netmsg_cache.list)) {
> +               struct io_async_msghdr *hdr;
> +
> +               hdr = hlist_entry(ctx->netmsg_cache.list.first,
> +                                       struct io_async_msghdr,
> cache_list);
> +               ctx->netmsg_cache.nr_cached--;
> +               hlist_del(&hdr->cache_list);

ditto here. I think all the hlist stuff and the nr_cached manipulation
can be wrapped up

> +               req->flags |= REQ_F_ASYNC_DATA;
> +               req->async_data = hdr;
> +               return hdr;
> +       }
> +
> +       if (!io_alloc_async_data(req))
> +               return req->async_data;
> +
> +       return NULL;
> +}
> +
>  static int io_setup_async_msg(struct io_kiocb *req,
> -                             struct io_async_msghdr *kmsg)
> +                             struct io_async_msghdr *kmsg,
> +                             unsigned int issue_flags)
>  {
>         struct io_async_msghdr *async_msg = req->async_data;
>  
>         if (async_msg)
>                 return -EAGAIN;
> -       if (io_alloc_async_data(req)) {
> +       async_msg = io_recvmsg_alloc_async(req, issue_flags);
> +       if (!async_msg) {
>                 kfree(kmsg->free_iov);
>                 return -ENOMEM;
>         }
> -       async_msg = req->async_data;
>         req->flags |= REQ_F_NEED_CLEANUP;
>         memcpy(async_msg, kmsg, sizeof(*kmsg));
>         async_msg->msg.msg_name = &async_msg->addr;
> @@ -195,7 +235,7 @@ int io_sendmsg(struct io_kiocb *req, unsigned int
> issue_flags)
>  
>         if (!(req->flags & REQ_F_POLLED) &&
>             (sr->flags & IORING_RECVSEND_POLL_FIRST))
> -               return io_setup_async_msg(req, kmsg);
> +               return io_setup_async_msg(req, kmsg, issue_flags);
>  
>         flags = sr->msg_flags;
>         if (issue_flags & IO_URING_F_NONBLOCK)
> @@ -207,13 +247,13 @@ int io_sendmsg(struct io_kiocb *req, unsigned
> int issue_flags)
>  
>         if (ret < min_ret) {
>                 if (ret == -EAGAIN && (issue_flags &
> IO_URING_F_NONBLOCK))
> -                       return io_setup_async_msg(req, kmsg);
> +                       return io_setup_async_msg(req, kmsg,
> issue_flags);
>                 if (ret == -ERESTARTSYS)
>                         ret = -EINTR;
>                 if (ret > 0 && io_net_retry(sock, flags)) {
>                         sr->done_io += ret;
>                         req->flags |= REQ_F_PARTIAL_IO;
> -                       return io_setup_async_msg(req, kmsg);
> +                       return io_setup_async_msg(req, kmsg,
> issue_flags);
>                 }
>                 req_set_fail(req);
>         }
> @@ -221,6 +261,7 @@ int io_sendmsg(struct io_kiocb *req, unsigned int
> issue_flags)
>         if (kmsg->free_iov)
>                 kfree(kmsg->free_iov);
>         req->flags &= ~REQ_F_NEED_CLEANUP;
> +       io_netmsg_recycle(req, issue_flags);
>         if (ret >= 0)
>                 ret += sr->done_io;
>         else if (sr->done_io)
> @@ -495,7 +536,7 @@ int io_recvmsg(struct io_kiocb *req, unsigned int
> issue_flags)
>  
>         if (!(req->flags & REQ_F_POLLED) &&
>             (sr->flags & IORING_RECVSEND_POLL_FIRST))
> -               return io_setup_async_msg(req, kmsg);
> +               return io_setup_async_msg(req, kmsg, issue_flags);
>  
>         if (io_do_buffer_select(req)) {
>                 void __user *buf;
> @@ -519,13 +560,13 @@ int io_recvmsg(struct io_kiocb *req, unsigned
> int issue_flags)
>         ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg, kmsg-
> >uaddr, flags);
>         if (ret < min_ret) {
>                 if (ret == -EAGAIN && force_nonblock)
> -                       return io_setup_async_msg(req, kmsg);
> +                       return io_setup_async_msg(req, kmsg,
> issue_flags);
>                 if (ret == -ERESTARTSYS)
>                         ret = -EINTR;
>                 if (ret > 0 && io_net_retry(sock, flags)) {
>                         sr->done_io += ret;
>                         req->flags |= REQ_F_PARTIAL_IO;
> -                       return io_setup_async_msg(req, kmsg);
> +                       return io_setup_async_msg(req, kmsg,
> issue_flags);
>                 }
>                 req_set_fail(req);
>         } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags &
> (MSG_TRUNC | MSG_CTRUNC))) {
> @@ -535,6 +576,7 @@ int io_recvmsg(struct io_kiocb *req, unsigned int
> issue_flags)
>         /* fast path, check for non-NULL to avoid function call */
>         if (kmsg->free_iov)
>                 kfree(kmsg->free_iov);
> +       io_netmsg_recycle(req, issue_flags);
>         req->flags &= ~REQ_F_NEED_CLEANUP;
>         if (ret > 0)
>                 ret += sr->done_io;
> @@ -848,4 +890,17 @@ int io_connect(struct io_kiocb *req, unsigned
> int issue_flags)
>         io_req_set_res(req, ret, 0);
>         return IOU_OK;
>  }
> +
> +void io_flush_netmsg_cache(struct io_ring_ctx *ctx)
> +{
> +       while (!hlist_empty(&ctx->netmsg_cache.list)) {
> +               struct io_async_msghdr *hdr;
> +
> +               hdr = hlist_entry(ctx->netmsg_cache.list.first,
> +                                       struct io_async_msghdr,
> cache_list);
> +               hlist_del(&hdr->cache_list);
> +               kfree(hdr);
> +       }
> +       ctx->netmsg_cache.nr_cached = 0;
> +}

again - could be put somewhere common. I assume there will not be much
more cleanup than simple kfree

>  #endif
> diff --git a/io_uring/net.h b/io_uring/net.h
> index 81d71d164770..576efb602c7f 100644
> --- a/io_uring/net.h
> +++ b/io_uring/net.h
> @@ -5,7 +5,10 @@
>  
>  #if defined(CONFIG_NET)
>  struct io_async_msghdr {
> -       struct iovec                    fast_iov[UIO_FASTIOV];
> +       union {
> +               struct iovec            fast_iov[UIO_FASTIOV];
> +               struct hlist_node       cache_list;
> +       };
>         /* points to an allocated iov, if NULL we use fast_iov
> instead */
>         struct iovec                    *free_iov;
>         struct sockaddr __user          *uaddr;
> @@ -40,4 +43,10 @@ int io_socket(struct io_kiocb *req, unsigned int
> issue_flags);
>  int io_connect_prep_async(struct io_kiocb *req);
>  int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe
> *sqe);
>  int io_connect(struct io_kiocb *req, unsigned int issue_flags);
> +
> +void io_flush_netmsg_cache(struct io_ring_ctx *ctx);
> +#else
> +static inline void io_flush_netmsg_cache(struct io_ring_ctx *ctx)
> +{
> +}
>  #endif

Re: [PATCH 4/4] io_uring: add netmsg cache

[Jens Axboe]

On 7/8/22 1:33 AM, Dylan Yudaken wrote:
> On Thu, 2022-07-07 at 17:23 -0600, Jens Axboe wrote:
>> For recvmsg/sendmsg, if they don't complete inline, we currently need
>> to allocate a struct io_async_msghdr for each request. This is a
>> somewhat large struct.
>>
>> Hook up sendmsg/recvmsg to use the io_alloc_cache. This reduces the
>> alloc + free overhead considerably, yielding 4-5% of extra
>> performance
>> running netbench.
>>
>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>> ---
>>  include/linux/io_uring_types.h |  6 ++-
>>  io_uring/io_uring.c            |  3 ++
>>  io_uring/net.c                 | 73 +++++++++++++++++++++++++++++---
>> --
>>  io_uring/net.h                 | 11 ++++-
>>  4 files changed, 81 insertions(+), 12 deletions(-)
>>
>> diff --git a/include/linux/io_uring_types.h
>> b/include/linux/io_uring_types.h
>> index bf8f95332eda..d54b8b7e0746 100644
>> --- a/include/linux/io_uring_types.h
>> +++ b/include/linux/io_uring_types.h
>> @@ -222,8 +222,7 @@ struct io_ring_ctx {
>>                 struct io_hash_table    cancel_table_locked;
>>                 struct list_head        cq_overflow_list;
>>                 struct io_alloc_cache   apoll_cache;
>> -               struct xarray           personalities;
>> -               u32                     pers_next;
>> +               struct io_alloc_cache   netmsg_cache;
>>         } ____cacheline_aligned_in_smp;
>>  
>>         /* IRQ completion list, under ->completion_lock */
>> @@ -241,6 +240,9 @@ struct io_ring_ctx {
>>         unsigned int            file_alloc_start;
>>         unsigned int            file_alloc_end;
>>  
>> +       struct xarray           personalities;
>> +       u32                     pers_next;
>> +
>>         struct {
>>                 /*
>>                  * We cache a range of free CQEs we can use, once
>> exhausted it
>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
>> index b5098773d924..32110c5b4059 100644
>> --- a/io_uring/io_uring.c
>> +++ b/io_uring/io_uring.c
>> @@ -89,6 +89,7 @@
>>  #include "kbuf.h"
>>  #include "rsrc.h"
>>  #include "cancel.h"
>> +#include "net.h"
>>  
>>  #include "timeout.h"
>>  #include "poll.h"
>> @@ -297,6 +298,7 @@ static __cold struct io_ring_ctx
>> *io_ring_ctx_alloc(struct io_uring_params *p)
>>         INIT_LIST_HEAD(&ctx->cq_overflow_list);
>>         INIT_LIST_HEAD(&ctx->io_buffers_cache);
>>         io_alloc_cache_init(&ctx->apoll_cache);
>> +       io_alloc_cache_init(&ctx->netmsg_cache);
>>         init_completion(&ctx->ref_comp);
>>         xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
>>         mutex_init(&ctx->uring_lock);
>> @@ -2473,6 +2475,7 @@ static __cold void io_ring_ctx_free(struct
>> io_ring_ctx *ctx)
>>                 __io_cqring_overflow_flush(ctx, true);
>>         io_eventfd_unregister(ctx);
>>         io_flush_apoll_cache(ctx);
>> +       io_flush_netmsg_cache(ctx);
>>         mutex_unlock(&ctx->uring_lock);
>>         io_destroy_buffers(ctx);
>>         if (ctx->sq_creds)
>> diff --git a/io_uring/net.c b/io_uring/net.c
>> index 6679069eeef1..ba7e94ff287c 100644
>> --- a/io_uring/net.c
>> +++ b/io_uring/net.c
>> @@ -12,6 +12,7 @@
>>  
>>  #include "io_uring.h"
>>  #include "kbuf.h"
>> +#include "alloc_cache.h"
>>  #include "net.h"
>>  
>>  #if defined(CONFIG_NET)
>> @@ -97,18 +98,57 @@ static bool io_net_retry(struct socket *sock, int
>> flags)
>>         return sock->type == SOCK_STREAM || sock->type ==
>> SOCK_SEQPACKET;
>>  }
>>  
>> +static void io_netmsg_recycle(struct io_kiocb *req, unsigned int
>> issue_flags)
>> +{
>> +       struct io_async_msghdr *hdr = req->async_data;
>> +
>> +       if (!hdr || issue_flags & IO_URING_F_UNLOCKED)
>> +               return;
>> +
>> +       if (io_alloc_cache_store(&req->ctx->netmsg_cache)) {
>> +               hlist_add_head(&hdr->cache_list, &req->ctx-
>>> netmsg_cache.list);
> 
> can io_alloc_cache_store just do the store?
> would be nicer to have cache::list be generally unused outside of the
> cache code.

We could do that if we just make the hlist_node be inside a struct.
Would probably allow cleaning up the get-entry etc too, let me give that
a whirl.

-- 
Jens Axboe

[PATCH 3/4] io_uring: impose max limit on apoll cache

[Jens Axboe]

Caches like this tend to grow to the peak size, and then never get any
smaller. Impose a max limit on the size, to prevent it from growing too
big.

A somewhat randomly chosen 512 is the max size we'll allow the cache
to get. If a batch of frees come in and would bring it over that, we
simply start kfree'ing the surplus.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/io_uring_types.h |  1 +
 io_uring/alloc_cache.h         | 16 ++++++++++++++--
 io_uring/io_uring.c            |  3 ++-
 3 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index b548da03b563..bf8f95332eda 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -160,6 +160,7 @@ struct io_ev_fd {
 
 struct io_alloc_cache {
 	struct hlist_head	list;
+	unsigned int		nr_cached;
 };
 
 struct io_ring_ctx {
diff --git a/io_uring/alloc_cache.h b/io_uring/alloc_cache.h
index 98f2374c37c7..729793ae9712 100644
--- a/io_uring/alloc_cache.h
+++ b/io_uring/alloc_cache.h
@@ -1,14 +1,24 @@
 #ifndef IOU_ALLOC_CACHE_H
 #define IOU_ALLOC_CACHE_H
 
+/*
+ * Don't allow the cache to grow beyond this size.
+ */
+#define IO_ALLOC_CACHE_MAX	512
+
 struct io_cache_entry {
 	struct hlist_node	node;
 };
 
-static inline void io_alloc_cache_put(struct io_alloc_cache *cache,
+static inline bool io_alloc_cache_put(struct io_alloc_cache *cache,
 				      struct io_cache_entry *entry)
 {
-	hlist_add_head(&entry->node, &cache->list);
+	if (cache->nr_cached < IO_ALLOC_CACHE_MAX) {
+		cache->nr_cached++;
+		hlist_add_head(&entry->node, &cache->list);
+		return true;
+	}
+	return false;
 }
 
 static inline struct io_cache_entry *io_alloc_cache_get(struct io_alloc_cache *cache)
@@ -26,6 +36,7 @@ static inline struct io_cache_entry *io_alloc_cache_get(struct io_alloc_cache *c
 static inline void io_alloc_cache_init(struct io_alloc_cache *cache)
 {
 	INIT_HLIST_HEAD(&cache->list);
+	cache->nr_cached = 0;
 }
 
 static inline void io_alloc_cache_free(struct io_alloc_cache *cache,
@@ -37,5 +48,6 @@ static inline void io_alloc_cache_free(struct io_alloc_cache *cache,
 		hlist_del(node);
 		free(container_of(node, struct io_cache_entry, node));
 	}
+	cache->nr_cached = 0;
 }
 #endif
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index a360a3d390c6..c9c23e459766 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -1181,7 +1181,8 @@ void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
 
 				if (apoll->double_poll)
 					kfree(apoll->double_poll);
-				io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache);
+				if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
+					kfree(apoll);
 				req->flags &= ~REQ_F_POLLED;
 			}
 			if (req->flags & IO_REQ_LINK_FLAGS)
-- 
2.35.1