[PATCH] io_uring: releasing CPU resources when polling

[PATCH] io_uring: releasing CPU resources when polling

[Xue]

From: hexue <xue01.he@samsung.com>

This patch is intended to reduce the CPU usage of io_uring in polling
mode. When io_uring is configured to use polling but the
underlying device is not configured with poll queue, this patch is
trying to optimaize the CPU when io_uring is polling for completions
of IOs that will actually be completed via interrupts.

The patch is implemented as follows:
        - Get the poll queue information of the underling device
        in io_uring.

        - If there is no poll queue, IO will be arrive as
        interruption, then io_uring's polling detection will keep
        spinning empty, resulting in a waste of CPU resources,
        so let the process release CPU before IO completes.

        - Record the running time and context switching time of each
        IO, and use these time to determine whether a process continue
        to schedule.
        In specific implementations, applying to consecutive IOs, each
        IO is judged based on the recording time of the previous IO,
        and when the idle time is greater than 1us, it will schedule to
        ensure that the performance is not compromised.

        - Adaptive adjustment to different devices. Due to the real-time
        nature of time recording, each device's IO processing speed is
        different, so the CPU optimization effect will vary, but could
        ensure that the performance will not be degraded.

        - Set a interface enables application to choose whether or not
        to require this feature.
        Add a tag of flag, when the application adds this tag, the
        optimization provided by this patch will only be adopted if
        the above condition is met, i.e., if io_uring is polling and
        the underlying device is not configured with a poll queue,
        which means that the user chooses the CPU low-loss prioritized
        approach for IO. If the application does not set this tag, then
        even if the conditions are met, it will remain the same as
        native io_uring, with no changes or impacts.

The CPU optimization of patch is tested as follows:
        In the case of reaching the peak of the disk's spec, the
        performance is not significantly reduced from native kernel,
        but the CPU usage of per CPU is reduced by about 50% for
        sequential R&W, and by ~80% for Randomwrite.

        This optimization does not affect IO performance in other cases,
        e.g., when both the underlying device and io_uring allow
        polling, performance is not affected by this patch.

                - test tool: Fio 3.35, 8 core VM
                - test method: Run performance tests on bare disks
                and use the htop tool to observe CPU utilization
                - Fio peak workload command:
                  [global]
                  ioengine=io_uring
                  norandommap=1
                  randrepeat=0
                  refill_buffers
                  group_reporting
                  ramp_time=30s
                  time_based
                  runtime=1m
                  filename=/dev/nvme0n1
                  hipri=1
                  direct=1
                  iodepth=64

                  [disk0]                 |     [disk0]
                  bs=4k                   |     bs=128k
                  numjobs=16              |     numjobs=1
                  rw=randread/randwrite   |     rw=read/write

                - Detailed test results
                  |-------------|---------|--------------------|
                  |  rw=read    |  BW/s   |per CPU utilization |
                  |  Native     |  10.9G  |       100%         |
                  |Optimization |  10.8G  |       44%          |

                  |-------------|---------|--------------------|
                  |  rw=write   |  BW/s   |per CPU utilization |
                  |  Native     |  6175   |       100%         |
                  |Optimization |  6150   |       32%          |

                  |-------------|---------|--------------------|
                  | rw=randread |  KIOPS  |per CPU utilization |
                  |  Native     |  1680   |       100%         |
                  |Optimization |  1608   |       100%         |

                  |-------------|---------|--------------------|
                  | rw=randwrite|  KIOPS  |per CPU utilization |
                  |  Native     |   225   |       100%         |
                  |Optimization |   225   |       15%          |

Signed-off-by: hexue <xue01.he@samsung.com>
---
 include/linux/io_uring_types.h | 12 ++++++++
 include/uapi/linux/io_uring.h  |  1 +
 io_uring/io_uring.c            | 47 ++++++++++++++++++++++++++++-
 io_uring/io_uring.h            |  2 ++
 io_uring/rw.c                  | 55 ++++++++++++++++++++++++++++++++++
 5 files changed, 116 insertions(+), 1 deletion(-)

diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index 854ad67a5f70..55d22f6e1eb0 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -224,6 +224,12 @@ struct io_alloc_cache {
 	size_t			elem_size;
 };
 
+struct iopoll_info {
+	bool		poll_state;
+	long		last_runtime;
+	long		last_irqtime;
+};
+
 struct io_ring_ctx {
 	/* const or read-mostly hot data */
 	struct {
@@ -421,6 +427,7 @@ struct io_ring_ctx {
 	unsigned short			n_sqe_pages;
 	struct page			**ring_pages;
 	struct page			**sqe_pages;
+	struct xarray		poll_array;
 };
 
 struct io_tw_state {
@@ -641,6 +648,11 @@ struct io_kiocb {
 		u64			extra1;
 		u64			extra2;
 	} big_cqe;
+	/* for adaptive iopoll */
+	int				poll_flag;
+	bool			poll_state;
+	struct timespec64		iopoll_start;
+	struct timespec64		iopoll_end;
 };
 
 struct io_overflow_cqe {
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 7a673b52827b..cd11a7786c51 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -198,6 +198,7 @@ enum {
  * Removes indirection through the SQ index array.
  */
 #define IORING_SETUP_NO_SQARRAY		(1U << 16)
+#define IORING_SETUP_NO_POLLQUEUE	(1U << 17)
 
 enum io_uring_op {
 	IORING_OP_NOP,
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index cd9a137ad6ce..9609acc60868 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -79,6 +79,8 @@
 
 #include <uapi/linux/io_uring.h>
 
+#include <linux/time.h>
+#include <linux/timekeeping.h>
 #include "io-wq.h"
 
 #include "io_uring.h"
@@ -122,6 +124,9 @@
 #define IO_COMPL_BATCH			32
 #define IO_REQ_ALLOC_BATCH		8
 
+#define IO_POLL_QUEUE 1
+#define IO_NO_POLL_QUEUE 0
+
 enum {
 	IO_CHECK_CQ_OVERFLOW_BIT,
 	IO_CHECK_CQ_DROPPED_BIT,
@@ -311,6 +316,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 		goto err;
 
 	ctx->flags = p->flags;
+	xa_init(&ctx->poll_array);
 	atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
 	init_waitqueue_head(&ctx->sqo_sq_wait);
 	INIT_LIST_HEAD(&ctx->sqd_list);
@@ -1875,11 +1881,32 @@ static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
 	return !!req->file;
 }
 
+/* Get poll queue information of the device */
+int get_poll_queue_state(struct io_kiocb *req)
+{
+	struct block_device *bdev;
+	struct request_queue *q;
+	struct inode *inode;
+
+	inode = req->file->f_inode;
+	if (!inode->i_rdev)
+		return 1;
+	bdev = blkdev_get_no_open(inode->i_rdev);
+	q = bdev->bd_queue;
+	if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) {
+		return IO_NO_POLL_QUEUE;
+	} else {
+		return IO_POLL_QUEUE;
+	}
+}
+
 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
 {
 	const struct io_issue_def *def = &io_issue_defs[req->opcode];
 	const struct cred *creds = NULL;
+	struct io_ring_ctx *ctx = req->ctx;
 	int ret;
+	u32 index;
 
 	if (unlikely(!io_assign_file(req, def, issue_flags)))
 		return -EBADF;
@@ -1890,6 +1917,21 @@ static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
 	if (!def->audit_skip)
 		audit_uring_entry(req->opcode);
 
+	if (ctx->flags & IORING_SETUP_NO_POLLQUEUE) {
+		index = req->file->f_inode->i_rdev;
+		struct iopoll_info *entry = xa_load(&ctx->poll_array, index);
+
+		if (!entry) {
+			entry = kmalloc(sizeof(struct iopoll_info), GFP_KERNEL);
+			entry->poll_state = get_poll_queue_state(req);
+			entry->last_runtime = 0;
+			entry->last_irqtime = 0;
+			xa_store(&ctx->poll_array, index, entry, GFP_KERNEL);
+			}
+		req->poll_state = entry->poll_state;
+		ktime_get_ts64(&req->iopoll_start);
+	}
+
 	ret = def->issue(req, issue_flags);
 
 	if (!def->audit_skip)
@@ -2176,6 +2218,8 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	req->file = NULL;
 	req->rsrc_node = NULL;
 	req->task = current;
+	req->poll_flag = 0;
+	req->poll_state = 1;
 
 	if (unlikely(opcode >= IORING_OP_LAST)) {
 		req->opcode = 0;
@@ -2921,6 +2965,7 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	kfree(ctx->cancel_table_locked.hbs);
 	kfree(ctx->io_bl);
 	xa_destroy(&ctx->io_bl_xa);
+	xa_destroy(&ctx->poll_array);
 	kfree(ctx);
 }
 
@@ -4050,7 +4095,7 @@ static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
 			IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
 			IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
 			IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
-			IORING_SETUP_NO_SQARRAY))
+			IORING_SETUP_NO_SQARRAY | IORING_SETUP_NO_POLLQUEUE))
 		return -EINVAL;
 
 	return io_uring_create(entries, &p, params);
diff --git a/io_uring/io_uring.h b/io_uring/io_uring.h
index d5495710c178..4281a0bb7ed9 100644
--- a/io_uring/io_uring.h
+++ b/io_uring/io_uring.h
@@ -125,6 +125,8 @@ static inline void io_req_task_work_add(struct io_kiocb *req)
 	__io_req_task_work_add(req, 0);
 }
 
+#define LEFT_TIME 3000
+
 #define io_for_each_link(pos, head) \
 	for (pos = (head); pos; pos = pos->link)
 
diff --git a/io_uring/rw.c b/io_uring/rw.c
index d5e79d9bdc71..db589a2cf659 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c
@@ -1118,6 +1118,44 @@ void io_rw_fail(struct io_kiocb *req)
 	io_req_set_res(req, res, req->cqe.flags);
 }
 
+void io_delay(struct io_kiocb *req, struct iopoll_info *entry)
+{
+	struct hrtimer_sleeper timer;
+	ktime_t kt;
+	struct timespec64 tc, oldtc;
+	enum hrtimer_mode mode;
+	long sleep_ti;
+
+	if (req->poll_flag == 1)
+		return;
+
+	if (entry->last_runtime <= entry->last_irqtime || (entry->last_runtime - entry->last_irqtime) < LEFT_TIME)
+		return;
+
+	req->poll_flag = 1;
+	ktime_get_ts64(&oldtc);
+	sleep_ti = (entry->last_runtime - entry->last_irqtime) / 2;
+	kt = ktime_set(0, sleep_ti);
+
+	mode = HRTIMER_MODE_REL;
+	hrtimer_init_sleeper_on_stack(&timer, CLOCK_MONOTONIC, mode);
+	hrtimer_set_expires(&timer.timer, kt);
+
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	hrtimer_sleeper_start_expires(&timer, mode);
+	if (timer.task) {
+		io_schedule();
+	}
+	hrtimer_cancel(&timer.timer);
+	mode = HRTIMER_MODE_ABS;
+
+	__set_current_state(TASK_RUNNING);
+	destroy_hrtimer_on_stack(&timer.timer);
+
+	ktime_get_ts64(&tc);
+	entry->last_irqtime = tc.tv_nsec - oldtc.tv_nsec - sleep_ti;
+}
+
 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
 {
 	struct io_wq_work_node *pos, *start, *prev;
@@ -1136,12 +1174,28 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
 		struct file *file = req->file;
 		int ret;
+		u32 index = file->f_inode->i_rdev;
 
 		/*
 		 * Move completed and retryable entries to our local lists.
 		 * If we find a request that requires polling, break out
 		 * and complete those lists first, if we have entries there.
 		 */
+
+		if ((ctx->flags & IORING_SETUP_NO_POLLQUEUE) && !req->poll_state) {
+			struct iopoll_info *entry = xa_load(&ctx->poll_array, index);
+
+			do {
+				if (READ_ONCE(req->iopoll_completed)) {
+					ktime_get_ts64(&req->iopoll_end);
+					entry->last_runtime = req->iopoll_end.tv_nsec - req->iopoll_start.tv_nsec;
+					break;
+				}
+				io_delay(req, entry);
+			} while (1);
+			goto complete;
+		}
+
 		if (READ_ONCE(req->iopoll_completed))
 			break;
 
@@ -1172,6 +1226,7 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
 	else if (!pos)
 		return 0;
 
+complete:
 	prev = start;
 	wq_list_for_each_resume(pos, prev) {
 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
-- 
2.34.1

Re:io_uring: releasing CPU resources when polling

[Xue]

Hi,

I hope this message finds you well.

I'm waiting to follow up on the patch I submitted on 3.18,
titled "io_uring: releasing CPU resources when polling".

I haven't received feedback yet and wondering if you had
a chance to look at it. Any guidance or suggestions you could
provide would be greatly appreciated.

Thanks,
Xue He

Re: io_uring: releasing CPU resources when polling

[Jens Axboe]

On 3/25/24 9:23 PM, Xue wrote:
> Hi,
> 
> I hope this message finds you well.
> 
> I'm waiting to follow up on the patch I submitted on 3.18,
> titled "io_uring: releasing CPU resources when polling".
> 
> I haven't received feedback yet and wondering if you had
> a chance to look at it. Any guidance or suggestions you could
> provide would be greatly appreciated.

I did take a look at it, and I have to be honest - I don't like it at
all. It's a lot of expensive code in the fast path, for a problem that
should not really exist. The system is misconfigured if you're doing
polled IO for devices that don't have a poll queue. At some point the
block layer returned -EOPNOTSUPP for that, and honestly I think that's a
MUCH better solution than adding expensive code in the fast path for
something that is really a badly configured setup.

-- 
Jens Axboe

Re: Re: io_uring: releasing CPU resources when polling.

[hexue]

On 3/26/24  3:39, Jens Axboe wrote:
>On 3/25/24 9:23 PM, Xue wrote:
>> Hi,
>> 
>> I hope this message finds you well.
>> 
>> I'm waiting to follow up on the patch I submitted on 3.18,
>> titled "io_uring: releasing CPU resources when polling".
>> 
>> I haven't received feedback yet and wondering if you had
>> a chance to look at it. Any guidance or suggestions you could
>> provide would be greatly appreciated.
>
>I did take a look at it, and I have to be honest - I don't like it at
>all. It's a lot of expensive code in the fast path, for a problem that
>should not really exist. The system is misconfigured if you're doing
>polled IO for devices that don't have a poll queue. At some point the
>block layer returned -EOPNOTSUPP for that, and honestly I think that's a
>MUCH better solution than adding expensive code in the fast path for
>something that is really a badly configured setup.

Sorry for my late reply, if you think that the scenario where if you're 
doing polled IO for devices that don't have a poll queue is just a 
misconfigured and does not need to be changed too much, then I'm inclined
to extend this scenario to all devices, I think it's an effective way to
release CPU resources, and I verified this and found that it does have a
very good benefit. At the same time I have reduce the code in the fast
path. I will release the v2 version of the code with my test results,
and please reconsider the feasibility of this solution.

--
Xue

Re: io_uring: releasing CPU resources when polling

[hexue]

Sorry, please ignore this patch, I will resend one later, I'm
sorry for improper operation.

--
hexue