io_uring: add napi busy polling support

This adds the napi busy polling support in io_uring.c. It adds a new napi_list to the io_ring_ctx structure. This list contains the list of napi_id's that are currently enabled for busy polling. This list is used to determine which napi id's enabled busy polling.

To set the new napi busy poll timeout, a new io-uring api has been added. It sets the napi busy poll timeout for the corresponding ring.

There is also a corresponding liburing patch series, which enables this feature. The name of the series is "liburing: add add api for napi busy poll timeout". It also contains two programs to test the this.

Testing has shown that the round-trip times are reduced to 38us from 55us by enabling napi busy polling with a busy poll timeout of 100us.

Changes:
- V3:
- Refreshed to 6.1-rc5
- Added a new io-uring api for the prefer napi busy poll api and wire it to io_napi_busy_loop().
- Removed the unregister (implemented as register)
- Added more performance results to the first commit message.
- V2:
- Add missing defines if CONFIG_NET_RX_BUSY_POLL is not defined
- Changes signature of function io_napi_add_list to static inline if CONFIG_NET_RX_BUSY_POLL is not defined
- define some functions as static

Stefan Roesch (3):
io_uring: add napi busy polling support
io_uring: add api to set napi busy poll timeout.
io_uring: add api to set napi prefer busy poll

 include/linux/io_uring_types.h |   8 +
 include/uapi/linux/io_uring.h  |   4 +
 io_uring/io_uring.c            | 278 +++++++++++++++++++++++++++++++++
 io_uring/napi.h                |  22 +++
 io_uring/poll.c                |   3 +
 io_uring/sqpoll.c              |  10 ++
 6 files changed, 325 insertions(+)
 create mode 100644 io_uring/napi.h

base-commit: 094226ad94f471a9f19e8f8e7140a09c2625abaa

2.30.2

This adds the napi busy polling support in io_uring.c. It adds a new napi_list to the io_ring_ctx structure. This list contains the list of napi_id's that are currently enabled for busy polling. The list is synchronized by the new napi_lock spin lock. The current default napi busy polling time is stored in napi_busy_poll_to. If napi busy polling is not enabled, the value is 0.

The NAPI_TIMEOUT is stored as a timeout to make sure that the time a napi entry is stored in the napi list is limited.

The busy poll timeout is also stored as part of the io_wait_queue. This is necessary as for sq polling the poll interval needs to be adjusted and the napi callback allows only to pass in one value.

This has been tested with two simple programs from the liburing library
repository: the napi client and the napi server program. The client sends a request, which has a timestamp in its payload and the server replies with the same payload. The client calculates the roundtrip time and stores it to calcualte the results.

The client is running on host1 and the server is running on host 2 (in the same rack). The measured times below are roundtrip times. They are average times over 5 runs each. Each run measures 1 million roundtrips.

no rx coal rx coal: frames=88,usecs=33
Default 57us 56us

client_poll=100us 47us 46us

server_poll=100us 51us 46us

client_poll=100us+ 40us 40us
server_poll=100us

client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on client

client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on server

client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on client + server

---
 include/linux/io_uring_types.h |   8 ++
 io_uring/io_uring.c            | 245 +++++++++++++++++++++++++++++++++
 io_uring/napi.h                |  22 +++
 io_uring/poll.c                |   3 +
 io_uring/sqpoll.c              |  10 ++
 5 files changed, 288 insertions(+)
 create mode 100644 io_uring/napi.h

index f5b687a787a3..23993b5d3186 100644
@@ -270,6 +270,14 @@ struct io_ring_ctx {
 	struct xarray		personalities;
 	u32			pers_next;
 
+#ifdef CONFIG_NET_RX_BUSY_POLL
+	struct list_head	napi_list;	/* track busy poll napi_id */
+	spinlock_t		napi_lock;	/* napi_list lock */
+
+	unsigned int		napi_busy_poll_to; /* napi busy poll default timeout */
+	bool			napi_prefer_busy_poll;
+#endif
+
 	struct {
 		/*
 		 * We cache a range of free CQEs we can use, once exhausted it

new file mode 100644
index 000000000000..11602d679629
@@ -0,0 +1,22 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef IOU_NAPI_H
+#define IOU_NAPI_H
+
+#include <linux/kernel.h>
+#include <linux/io_uring.h>
+#include <net/busy_poll.h>
+
+#ifdef CONFIG_NET_RX_BUSY_POLL
+
+void io_napi_add(struct file *file, struct io_ring_ctx *ctx);
+bool io_napi_busy_loop(struct list_head *napi_list, bool prefer_busy_poll);
+
+#else
+
+static inline void io_napi_add(struct file *file, struct io_ring_ctx *ctx)
+{
+}
+
+#endif
+#endif

index f500506984ec..fafce408ee14 100644
@@ -15,6 +15,7 @@
 
 #include "io_uring.h"
 #include "refs.h"
+#include "napi.h"
 #include "opdef.h"
 #include "kbuf.h"
 #include "poll.h"

@@ -250,6 +251,7 @@ static int io_poll_check_events(struct io_kiocb *req, bool *locked)
 				io_req_set_res(req, mask, 0);
 				return IOU_POLL_REMOVE_POLL_USE_RES;
 			}
+			io_napi_add(req->file, ctx);
 		} else {
 			ret = io_poll_issue(req, locked);
 			if (ret == IOU_STOP_MULTISHOT)

@@ -571,6 +573,7 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
 		__io_poll_execute(req, mask);
 		return 0;
 	}
+	io_napi_add(req->file, req->ctx);
 
 	if (ipt->owning) {
 		/*

index 559652380672..6f50bc156038 100644
@@ -15,6 +15,7 @@
 #include <uapi/linux/io_uring.h>
 
 #include "io_uring.h"
+#include "napi.h"
 #include "sqpoll.h"
 
 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8

@@ -193,6 +194,15 @@ static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
 			ret = io_submit_sqes(ctx, to_submit);
 		mutex_unlock(&ctx->uring_lock);
 
+#ifdef CONFIG_NET_RX_BUSY_POLL
+		spin_lock(&ctx->napi_lock);
+		if (!list_empty(&ctx->napi_list) &&
+		    READ_ONCE(ctx->napi_busy_poll_to) > 0 &&
+		    io_napi_busy_loop(&ctx->napi_list, ctx->napi_prefer_busy_poll))
+			++ret;
+		spin_unlock(&ctx->napi_lock);
+#endif
+
 		if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
 			wake_up(&ctx->sqo_sq_wait);
 		if (creds)

2.30.2

This adds an api to register the busy poll timeout from liburing. To be able to use this functionality, the corresponding liburing patch is needed.

---
 include/uapi/linux/io_uring.h |  3 +++
 io_uring/io_uring.c           | 16 ++++++++++++++++
 2 files changed, 19 insertions(+)

index 2df3225b562f..9b5c1df0d1d8 100644
@@ -490,6 +490,9 @@ enum {
 	/* register a range of fixed file slots for automatic slot allocation */
 	IORING_REGISTER_FILE_ALLOC_RANGE	= 25,
 
+	/* set/clear busy poll timeout */
+	IORING_REGISTER_NAPI_BUSY_POLL_TIMEOUT	= 27,
+
 	/* this goes last */
 	IORING_REGISTER_LAST
 };

index 4f432694cbed..f6c9c9cbe0f8 100644
@@ -4122,6 +4122,16 @@ static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
 	return ret;
 }
 
+static int io_register_napi_busy_poll_timeout(struct io_ring_ctx *ctx, unsigned int to)
+{
+#ifdef CONFIG_NET_RX_BUSY_POLL
+	WRITE_ONCE(ctx->napi_busy_poll_to, to);
+	return 0;
+#else
+	return -EINVAL;
+#endif
+}
+
 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 			       void __user *arg, unsigned nr_args)
 	__releases(ctx->uring_lock)

@@ -4282,6 +4292,12 @@ static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 			break;
 		ret = io_register_file_alloc_range(ctx, arg);
 		break;
+	case IORING_REGISTER_NAPI_BUSY_POLL_TIMEOUT:
+		ret = -EINVAL;
+		if (arg || !nr_args)
+			break;
+		ret = io_register_napi_busy_poll_timeout(ctx, nr_args);
+		break;
 	default:
 		ret = -EINVAL;
 		break;

2.30.2

Kind of related to the previous question, but I think we should just add a single REGISTER opcode for this, and define a struct with the delay setting. Add necessary padding for that, check padding for non-zero and EINVAL if set. Then patch 3 can just add that struct field.

The IORING_REGISTER_NAPI can then also return the current settings by just copying it back. That makes it a get/set API as well.

Patch 1 would be unaffected, this just changes patch 2+3 for the io_uring API.

Jens Axboe

This adds an api to register and unregister the napi prefer busy poll setting from liburing. To be able to use this functionality, the corresponding liburing patch is needed.

---
 include/uapi/linux/io_uring.h |  3 ++-
 io_uring/io_uring.c           | 17 +++++++++++++++++
 2 files changed, 19 insertions(+), 1 deletion(-)

index 9b5c1df0d1d8..25b91a4dc103 100644
@@ -490,7 +490,8 @@ enum {
 	/* register a range of fixed file slots for automatic slot allocation */
 	IORING_REGISTER_FILE_ALLOC_RANGE	= 25,
 
-	/* set/clear busy poll timeout */
+	/* set/clear busy poll settings */
+	IORING_REGISTER_NAPI_PREFER_BUSY_POLL	= 26,
 	IORING_REGISTER_NAPI_BUSY_POLL_TIMEOUT	= 27,
 
 	/* this goes last */

index f6c9c9cbe0f8..eca9f7540123 100644
@@ -4122,6 +4122,17 @@ static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
 	return ret;
 }
 
+static int io_register_napi_prefer_busy_poll(struct io_ring_ctx *ctx,
+					     unsigned int prefer_napi)
+{
+#ifdef CONFIG_NET_RX_BUSY_POLL
+	WRITE_ONCE(ctx->napi_prefer_busy_poll, !!prefer_napi);
+	return 0;
+#else
+	return -EINVAL;
+#endif
+}
+
 static int io_register_napi_busy_poll_timeout(struct io_ring_ctx *ctx, unsigned int to)
 {
 #ifdef CONFIG_NET_RX_BUSY_POLL

@@ -4292,6 +4303,12 @@ static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 			break;
 		ret = io_register_file_alloc_range(ctx, arg);
 		break;
+	case IORING_REGISTER_NAPI_PREFER_BUSY_POLL:
+		ret = -EINVAL;
+		if (arg || !nr_args)
+			break;
+		ret = io_register_napi_prefer_busy_poll(ctx, nr_args);
+		break;
 	case IORING_REGISTER_NAPI_BUSY_POLL_TIMEOUT:
 		ret = -EINVAL;
 		if (arg || !nr_args)

2.30.2

Thanks!

Thanks Jakub! Question on the need for patch 3, which I think came about because of comments from you. Can you expand on why we need both an enable and timeout setting? Are there cases where timeout == 0 and enabled == true make sense?

Jens Axboe

The enable is for the "prefer busy poll" mode, rather that just busy polling.

The prefer busy poll mode disables interrupts and arms a (hopefully long enough) fail safe timer, and expects user to come back and busy poll before the timer fires. The timer length is set thru sysfs params for NAPI/queue.

Because the Rx traffic is fully async and not in control of the local app, this gives the local app the ability to postpone the Rx IRQ. No interruptions means lower response latency.
With the expectation that the app will read/"busy poll" next batch of packets once its done servicing the previous batch.

We don't have to implement this bit from the start, "normal" busy poll is already functional with patches 1 and 2.

Gotcha, ok makes sense to me. I'm fine with the patchset, just want a few adjustments on the API side as per previous email. I think Stefan is respinning with that, then we can get it queued up.

Jens Axboe