[PATCH v2 0/2] improve SQPOLL handling

[PATCH v2 0/2] improve SQPOLL handling

[Xiaoguang Wang]

The first patch tries to improve various issues in current implementation:
  The prepare_to_wait() usage in __io_sq_thread() is weird. If multiple ctxs
share one same poll thread, one ctx will put poll thread in TASK_INTERRUPTIBLE,
but if other ctxs have work to do, we don't need to change task's stat at all.
I think only if all ctxs don't have work to do, we can do it.
  We use round-robin strategy to make multiple ctxs share one same poll thread,
but there are various condition in __io_sq_thread(), which seems complicated and
may affect round-robin strategy.

The second patch adds a IORING_SETUP_SQPOLL_PERCPU flag, for those rings which
have SQPOLL enabled and are willing to be bound to one same cpu, hence share
one same poll thread, add a capability that these rings can share one poll thread
by specifying a new IORING_SETUP_SQPOLL_PERCPU flag. FIO tool can integrate this
feature easily, so we can test multiple rings to share same poll thread easily.


TEST:
  This patch set have passed liburing test cases.

  I also make fio support IORING_SETUP_SQPOLL_PERCPU flag, and make some
io stress tests, no errors or performance regression. See below fio job file:

First in unpatched kernel, I test a fio file which only contains one job
with iodepth being 128, see below:
[global]
ioengine=io_uring
sqthread_poll=1
registerfiles=1
fixedbufs=1
hipri=1
thread=1
bs=4k
direct=1
rw=randread
time_based=1
runtime=120
ramp_time=0
randrepeat=0
group_reporting=1
filename=/dev/nvme0n1
sqthread_poll_cpu=15

[job0]
cpus_allowed=5
iodepth=128
sqthread_poll_cpu=9

performance data: IOPS: 453k, avg lat: 282.37usec


Second in unpatched kernel, I test a fio file which contains 4 jobs
with each iodepth being 32, see below:
[global]
ioengine=io_uring
sqthread_poll=1
registerfiles=1
fixedbufs=1
hipri=1
thread=1
bs=4k
direct=1
rw=randread
time_based=1
runtime=120
ramp_time=0
randrepeat=0
group_reporting=1
filename=/dev/nvme0n1
sqthread_poll_cpu=15

[job0]
cpus_allowed=5
iodepth=32
sqthread_poll_cpu=9

[job1]
cpus_allowed=6
iodepth=32
sqthread_poll_cpu=9

[job2]
cpus_allowed=7
iodepth=32
sqthread_poll_cpu=9

[job3]
cpus_allowed=8
iodepth=32
sqthread_poll_cpu=9
performance data: IOPS: 254k, avg lat: 503.80 usec, obvious performance
drop.

Finally in patched kernel, I test a fio file which contains 4 jobs
with each iodepth being 32, and now we enable sqthread_poll_percpu
flag, see blow:

[global]
ioengine=io_uring
sqthread_poll=1
registerfiles=1
fixedbufs=1
hipri=1
thread=1
bs=4k
direct=1
rw=randread
time_based=1
runtime=120
ramp_time=0
randrepeat=0
group_reporting=1
filename=/dev/nvme0n1
#sqthread_poll_cpu=15
sqthread_poll_percpu=1  # enable percpu feature

[job0]
cpus_allowed=5
iodepth=32
sqthread_poll_cpu=9

[job1]
cpus_allowed=6
iodepth=32
sqthread_poll_cpu=9

[job2]
cpus_allowed=7
iodepth=32
sqthread_poll_cpu=9

performance data: IOPS: 438k, avg lat: 291.69usec


From above teses, we can see that IORING_SETUP_SQPOLL_PERCPU is easy to
use, and no obvious performance regression.
Note I don't test IORING_SETUP_ATTACH_WQ in above three test cases, it's
a little hard to support IORING_SETUP_ATTACH_WQ in fio.

Xiaoguang Wang (2):
  io_uring: refactor io_sq_thread() handling
  io_uring: support multiple rings to share same poll thread by
    specifying same cpu

 fs/io_uring.c                 | 289 +++++++++++++++++++---------------
 include/uapi/linux/io_uring.h |   1 +
 2 files changed, 166 insertions(+), 124 deletions(-)

-- 
2.17.2

[PATCH v2 1/2] io_uring: refactor io_sq_thread() handling

[Xiaoguang Wang]

There are some issues about current io_sq_thread() implementation:
  1. The prepare_to_wait() usage in __io_sq_thread() is weird. If
multiple ctxs share one same poll thread, one ctx will put poll thread
in TASK_INTERRUPTIBLE, but if other ctxs have work to do, we don't
need to change task's stat at all. I think only if all ctxs don't have
work to do, we can do it.
  2. We use round-robin strategy to make multiple ctxs share one same
poll thread, but there are various condition in __io_sq_thread(), which
seems complicated and may affect round-robin strategy.

To improve above issues, I take below actions:
  1. If multiple ctxs share one same poll thread, only if all all ctxs
don't have work to do, we can call prepare_to_wait() and schedule() to
make poll thread enter sleep state.
  2. To make round-robin strategy more straight, I simplify
__io_sq_thread() a bit, it just does io poll and sqes submit work once,
does not check various condition.
  3. For multiple ctxs share one same poll thread, we choose the biggest
sq_thread_idle among these ctxs as timeout condition, and will update
it when ctx is in or out.
  4. Not need to check EBUSY especially, if io_submit_sqes() returns
EBUSY, IORING_SQ_CQ_OVERFLOW should be set, helper in liburing should
be aware of cq overflow and enters kernel to flush work.

Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
---
 fs/io_uring.c | 169 ++++++++++++++++++++------------------------------
 1 file changed, 67 insertions(+), 102 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index badd8e70a10e..e9cde444b34d 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -244,6 +244,8 @@ struct io_sq_data {
 
 	struct task_struct	*thread;
 	struct wait_queue_head	wait;
+
+	unsigned		sq_thread_idle;
 };
 
 struct io_ring_ctx {
@@ -309,7 +311,6 @@ struct io_ring_ctx {
 	struct io_sq_data	*sq_data;	/* if using sq thread polling */
 
 	struct wait_queue_head	sqo_sq_wait;
-	struct wait_queue_entry	sqo_wait_entry;
 	struct list_head	sqd_list;
 
 	/*
@@ -6828,111 +6829,49 @@ static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
 	spin_unlock_irq(&ctx->completion_lock);
 }
 
-static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
-			       int sync, void *key)
-{
-	struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
-	int ret;
-
-	ret = autoremove_wake_function(wqe, mode, sync, key);
-	if (ret) {
-		unsigned long flags;
-
-		spin_lock_irqsave(&ctx->completion_lock, flags);
-		ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
-		spin_unlock_irqrestore(&ctx->completion_lock, flags);
-	}
-	return ret;
-}
-
-enum sq_ret {
-	SQT_IDLE	= 1,
-	SQT_SPIN	= 2,
-	SQT_DID_WORK	= 4,
-};
-
-static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
-				  unsigned long start_jiffies, bool cap_entries)
+static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
 {
-	unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
-	struct io_sq_data *sqd = ctx->sq_data;
 	unsigned int to_submit;
 	int ret = 0;
 
-again:
 	if (!list_empty(&ctx->iopoll_list)) {
 		unsigned nr_events = 0;
 
 		mutex_lock(&ctx->uring_lock);
-		if (!list_empty(&ctx->iopoll_list) && !need_resched())
+		if (!list_empty(&ctx->iopoll_list))
 			io_do_iopoll(ctx, &nr_events, 0);
 		mutex_unlock(&ctx->uring_lock);
 	}
 
 	to_submit = io_sqring_entries(ctx);
-
-	/*
-	 * If submit got -EBUSY, flag us as needing the application
-	 * to enter the kernel to reap and flush events.
-	 */
-	if (!to_submit || ret == -EBUSY || need_resched()) {
-		/*
-		 * Drop cur_mm before scheduling, we can't hold it for
-		 * long periods (or over schedule()). Do this before
-		 * adding ourselves to the waitqueue, as the unuse/drop
-		 * may sleep.
-		 */
-		io_sq_thread_drop_mm_files();
-
-		/*
-		 * We're polling. If we're within the defined idle
-		 * period, then let us spin without work before going
-		 * to sleep. The exception is if we got EBUSY doing
-		 * more IO, we should wait for the application to
-		 * reap events and wake us up.
-		 */
-		if (!list_empty(&ctx->iopoll_list) || need_resched() ||
-		    (!time_after(jiffies, timeout) && ret != -EBUSY &&
-		    !percpu_ref_is_dying(&ctx->refs)))
-			return SQT_SPIN;
-
-		prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
-					TASK_INTERRUPTIBLE);
-
-		/*
-		 * While doing polled IO, before going to sleep, we need
-		 * to check if there are new reqs added to iopoll_list,
-		 * it is because reqs may have been punted to io worker
-		 * and will be added to iopoll_list later, hence check
-		 * the iopoll_list again.
-		 */
-		if ((ctx->flags & IORING_SETUP_IOPOLL) &&
-		    !list_empty_careful(&ctx->iopoll_list)) {
-			finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
-			goto again;
-		}
-
-		to_submit = io_sqring_entries(ctx);
-		if (!to_submit || ret == -EBUSY)
-			return SQT_IDLE;
-	}
-
-	finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
-	io_ring_clear_wakeup_flag(ctx);
-
 	/* if we're handling multiple rings, cap submit size for fairness */
 	if (cap_entries && to_submit > 8)
 		to_submit = 8;
 
-	mutex_lock(&ctx->uring_lock);
-	if (likely(!percpu_ref_is_dying(&ctx->refs)))
-		ret = io_submit_sqes(ctx, to_submit);
-	mutex_unlock(&ctx->uring_lock);
+	if (to_submit) {
+		mutex_lock(&ctx->uring_lock);
+		if (likely(!percpu_ref_is_dying(&ctx->refs)))
+			ret = io_submit_sqes(ctx, to_submit);
+		mutex_unlock(&ctx->uring_lock);
+	}
 
 	if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
 		wake_up(&ctx->sqo_sq_wait);
 
-	return SQT_DID_WORK;
+	return ret;
+}
+
+static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
+{
+	struct io_ring_ctx *ctx;
+	unsigned sq_thread_idle = 0;
+
+	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
+		if (sq_thread_idle < ctx->sq_thread_idle)
+			sq_thread_idle = ctx->sq_thread_idle;
+	}
+
+	sqd->sq_thread_idle = sq_thread_idle;
 }
 
 static void io_sqd_init_new(struct io_sq_data *sqd)
@@ -6941,11 +6880,11 @@ static void io_sqd_init_new(struct io_sq_data *sqd)
 
 	while (!list_empty(&sqd->ctx_new_list)) {
 		ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
-		init_wait(&ctx->sqo_wait_entry);
-		ctx->sqo_wait_entry.func = io_sq_wake_function;
 		list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
 		complete(&ctx->sq_thread_comp);
 	}
+
+	io_sqd_update_thread_idle(sqd);
 }
 
 static int io_sq_thread(void *data)
@@ -6957,7 +6896,8 @@ static int io_sq_thread(void *data)
 	const struct cred *old_cred = NULL;
 	struct io_sq_data *sqd = data;
 	struct io_ring_ctx *ctx;
-	unsigned long start_jiffies;
+	unsigned long timeout;
+	DEFINE_WAIT(wait);
 
 	task_lock(current);
 	current->files = NULL;
@@ -6965,10 +6905,9 @@ static int io_sq_thread(void *data)
 	current->thread_pid = NULL;
 	task_unlock(current);
 
-	start_jiffies = jiffies;
 	while (!kthread_should_stop()) {
-		enum sq_ret ret = 0;
-		bool cap_entries;
+		int ret;
+		bool cap_entries, sqt_spin, needs_sched;
 
 		/*
 		 * Any changes to the sqd lists are synchronized through the
@@ -6978,11 +6917,13 @@ static int io_sq_thread(void *data)
 		if (kthread_should_park())
 			kthread_parkme();
 
-		if (unlikely(!list_empty(&sqd->ctx_new_list)))
+		if (unlikely(!list_empty(&sqd->ctx_new_list))) {
 			io_sqd_init_new(sqd);
+			timeout = jiffies + sqd->sq_thread_idle;
+		}
 
+		sqt_spin = false;
 		cap_entries = !list_is_singular(&sqd->ctx_list);
-
 		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
 			if (current->cred != ctx->creds) {
 				if (old_cred)
@@ -6995,24 +6936,49 @@ static int io_sq_thread(void *data)
 			current->sessionid = ctx->sessionid;
 #endif
 
-			ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
+			ret = __io_sq_thread(ctx, cap_entries);
+			if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
+				sqt_spin = true;
 
 			io_sq_thread_drop_mm_files();
 		}
 
-		if (ret & SQT_SPIN) {
+		if (sqt_spin || !time_after(jiffies, timeout)) {
 			io_run_task_work();
 			cond_resched();
-		} else if (ret == SQT_IDLE) {
-			if (kthread_should_park())
-				continue;
+			if (sqt_spin)
+				timeout = jiffies + sqd->sq_thread_idle;
+			continue;
+		}
+
+		if (kthread_should_park())
+			continue;
+
+		needs_sched = true;
+		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
+		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
+			if ((ctx->flags & IORING_SETUP_IOPOLL) &&
+			    !list_empty_careful(&ctx->iopoll_list)) {
+				needs_sched = false;
+				break;
+			}
+			if (io_sqring_entries(ctx)) {
+				needs_sched = false;
+				break;
+			}
+		}
+
+		if (needs_sched) {
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 				io_ring_set_wakeup_flag(ctx);
+
 			schedule();
-			start_jiffies = jiffies;
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 				io_ring_clear_wakeup_flag(ctx);
 		}
+
+		finish_wait(&sqd->wait, &wait);
+		timeout = jiffies + sqd->sq_thread_idle;
 	}
 
 	io_run_task_work();
@@ -7310,12 +7276,11 @@ static void io_sq_thread_stop(struct io_ring_ctx *ctx)
 
 		mutex_lock(&sqd->ctx_lock);
 		list_del(&ctx->sqd_list);
+		io_sqd_update_thread_idle(sqd);
 		mutex_unlock(&sqd->ctx_lock);
 
-		if (sqd->thread) {
-			finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
+		if (sqd->thread)
 			io_sq_thread_unpark(sqd);
-		}
 
 		io_put_sq_data(sqd);
 		ctx->sq_data = NULL;
-- 
2.17.2

[PATCH v2 2/2] io_uring: support multiple rings to share same poll thread by specifying same cpu

[Xiaoguang Wang]

We have already supported multiple rings to share one same poll thread
by passing IORING_SETUP_ATTACH_WQ, but it's not that convenient to use.
IORING_SETUP_ATTACH_WQ needs users to ensure that a parent ring instance
has already existed, that means it will require app to regulate the
creation oder between uring instances.

Currently we can make this a bit simpler, for those rings which will
have SQPOLL enabled and are willing to be bound to one same cpu, add a
capability that these rings can share one poll thread by specifying
a new IORING_SETUP_SQPOLL_PERCPU flag, then we have 3 cases
  1, IORING_SETUP_ATTACH_WQ: if user specifies this flag, we'll always
try to attach this ring to an existing ring's corresponding poll thread,
no matter whether IORING_SETUP_SQ_AFF or IORING_SETUP_SQPOLL_PERCPU is
set.
  2, IORING_SETUP_SQ_AFF and IORING_SETUP_SQPOLL_PERCPU are both enabled,
for this case, we'll create a single poll thread to be shared by these
rings, and this poll thread is bound to a fixed cpu.
  3, for any other cases, we'll just create one new poll thread for the
corresponding ring.

And for case 2, don't need to regulate creation oder of multiple uring
instances, we use a mutex to synchronize creation, for example, say five
rings which all have IORING_SETUP_SQ_AFF & IORING_SETUP_SQPOLL_PERCPU
enabled, and are willing to be bound same cpu, one ring that gets the
mutex lock will create one poll thread, the other four rings will just
attach themselves the previous created poll thread once they get lock
successfully.

To implement above function, define a percpu io_sq_data array:
    static struct io_sq_data __percpu **percpu_sqd;
When IORING_SETUP_SQ_AFF and IORING_SETUP_SQPOLL_PERCPU are both enabled,
we will use struct io_uring_params' sq_thread_cpu to locate corresponding
sqd, and use this sqd to save poll thread info.

Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
---
 fs/io_uring.c                 | 123 ++++++++++++++++++++++++++++------
 include/uapi/linux/io_uring.h |   1 +
 2 files changed, 102 insertions(+), 22 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index e9cde444b34d..2ff8ed93a400 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -246,8 +246,12 @@ struct io_sq_data {
 	struct wait_queue_head	wait;
 
 	unsigned		sq_thread_idle;
+	unsigned		sq_thread_cpu;
 };
 
+static DEFINE_MUTEX(percpu_sqd_lock);
+static struct io_sq_data __percpu **percpu_sqd;
+
 struct io_ring_ctx {
 	struct {
 		struct percpu_ref	refs;
@@ -7175,8 +7179,17 @@ static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
 	return 0;
 }
 
-static void io_put_sq_data(struct io_sq_data *sqd)
+static void io_put_sq_data(struct io_ring_ctx *ctx, struct io_sq_data *sqd)
 {
+	int percpu = 0;
+
+	if ((ctx->flags & IORING_SETUP_SQ_AFF) &&
+	    (ctx->flags & IORING_SETUP_SQPOLL_PERCPU))
+		percpu = 1;
+
+	if (percpu)
+		mutex_lock(&percpu_sqd_lock);
+
 	if (refcount_dec_and_test(&sqd->refs)) {
 		/*
 		 * The park is a bit of a work-around, without it we get
@@ -7188,8 +7201,13 @@ static void io_put_sq_data(struct io_sq_data *sqd)
 			kthread_stop(sqd->thread);
 		}
 
+		if (percpu)
+			*per_cpu_ptr(percpu_sqd, sqd->sq_thread_cpu) = NULL;
 		kfree(sqd);
 	}
+
+	if (percpu)
+		mutex_unlock(&percpu_sqd_lock);
 }
 
 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
@@ -7218,13 +7236,10 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
 	return sqd;
 }
 
-static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
+static struct io_sq_data *io_alloc_sq_data(struct io_uring_params *p)
 {
 	struct io_sq_data *sqd;
 
-	if (p->flags & IORING_SETUP_ATTACH_WQ)
-		return io_attach_sq_data(p);
-
 	sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
 	if (!sqd)
 		return ERR_PTR(-ENOMEM);
@@ -7256,6 +7271,49 @@ static void io_sq_thread_park(struct io_sq_data *sqd)
 	kthread_park(sqd->thread);
 }
 
+static void io_attach_ctx_to_sqd(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
+{
+	ctx->sq_data = sqd;
+	io_sq_thread_park(sqd);
+	mutex_lock(&sqd->ctx_lock);
+	list_add(&ctx->sqd_list, &sqd->ctx_new_list);
+	mutex_unlock(&sqd->ctx_lock);
+	io_sq_thread_unpark(sqd);
+}
+
+static struct io_sq_data *io_find_or_create_percpu_sq_thread(struct io_ring_ctx *ctx,
+					struct io_uring_params *p)
+{
+	struct io_sq_data *sqd;
+	struct task_struct *tsk;
+	int cpu = p->sq_thread_cpu;
+
+	mutex_lock(&percpu_sqd_lock);
+	sqd = *per_cpu_ptr(percpu_sqd, cpu);
+	if (!sqd) {
+		sqd = io_alloc_sq_data(p);
+		if (IS_ERR(sqd)) {
+			mutex_unlock(&percpu_sqd_lock);
+			return sqd;
+		}
+
+		tsk = kthread_create_on_cpu(io_sq_thread, sqd, cpu, "io_uring-sq");
+		if (IS_ERR(tsk)) {
+			kfree(sqd);
+			sqd = ERR_PTR(PTR_ERR(tsk));
+			mutex_unlock(&percpu_sqd_lock);
+			return sqd;
+		}
+		sqd->sq_thread_cpu = cpu;
+		sqd->thread = tsk;
+		*per_cpu_ptr(percpu_sqd, cpu) = sqd;
+	} else {
+		refcount_inc(&sqd->refs);
+	}
+	mutex_unlock(&percpu_sqd_lock);
+	return sqd;
+}
+
 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
 {
 	struct io_sq_data *sqd = ctx->sq_data;
@@ -7282,7 +7340,7 @@ static void io_sq_thread_stop(struct io_ring_ctx *ctx)
 		if (sqd->thread)
 			io_sq_thread_unpark(sqd);
 
-		io_put_sq_data(sqd);
+		io_put_sq_data(ctx, sqd);
 		ctx->sq_data = NULL;
 	}
 }
@@ -7951,25 +8009,19 @@ static int io_sq_offload_create(struct io_ring_ctx *ctx,
 		if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
 			goto err;
 
-		sqd = io_get_sq_data(p);
-		if (IS_ERR(sqd)) {
-			ret = PTR_ERR(sqd);
-			goto err;
-		}
-
-		ctx->sq_data = sqd;
-		io_sq_thread_park(sqd);
-		mutex_lock(&sqd->ctx_lock);
-		list_add(&ctx->sqd_list, &sqd->ctx_new_list);
-		mutex_unlock(&sqd->ctx_lock);
-		io_sq_thread_unpark(sqd);
-
 		ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
 		if (!ctx->sq_thread_idle)
 			ctx->sq_thread_idle = HZ;
 
-		if (sqd->thread)
+		if (p->flags & IORING_SETUP_ATTACH_WQ) {
+			sqd = io_attach_sq_data(p);
+			if (IS_ERR(sqd)) {
+				ret = PTR_ERR(sqd);
+				goto err;
+			}
+			io_attach_ctx_to_sqd(sqd, ctx);
 			goto done;
+		}
 
 		if (p->flags & IORING_SETUP_SQ_AFF) {
 			int cpu = p->sq_thread_cpu;
@@ -7980,9 +8032,27 @@ static int io_sq_offload_create(struct io_ring_ctx *ctx,
 			if (!cpu_online(cpu))
 				goto err;
 
-			sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
+			if (p->flags & IORING_SETUP_SQPOLL_PERCPU) {
+				sqd = io_find_or_create_percpu_sq_thread(ctx, p);
+				if (IS_ERR(sqd)) {
+					ret = PTR_ERR(sqd);
+					goto err;
+				}
+			} else {
+				sqd = io_alloc_sq_data(p);
+				if (IS_ERR(sqd)) {
+					ret = PTR_ERR(sqd);
+					goto err;
+				}
+				sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
 							cpu, "io_uring-sq");
+			}
 		} else {
+			sqd = io_alloc_sq_data(p);
+			if (IS_ERR(sqd)) {
+				ret = PTR_ERR(sqd);
+				goto err;
+			}
 			sqd->thread = kthread_create(io_sq_thread, sqd,
 							"io_uring-sq");
 		}
@@ -7991,6 +8061,8 @@ static int io_sq_offload_create(struct io_ring_ctx *ctx,
 			sqd->thread = NULL;
 			goto err;
 		}
+		io_attach_ctx_to_sqd(sqd, ctx);
+
 		ret = io_uring_alloc_task_context(sqd->thread);
 		if (ret)
 			goto err;
@@ -9557,7 +9629,7 @@ static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
 	if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
 			IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
 			IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
-			IORING_SETUP_R_DISABLED))
+			IORING_SETUP_R_DISABLED | IORING_SETUP_SQPOLL_PERCPU))
 		return -EINVAL;
 
 	return  io_uring_create(entries, &p, params);
@@ -9910,6 +9982,8 @@ SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
 
 static int __init io_uring_init(void)
 {
+	int cpu;
+
 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
 	BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
 	BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
@@ -9950,6 +10024,11 @@ static int __init io_uring_init(void)
 	BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
 	BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
 	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
+
+	percpu_sqd = alloc_percpu(struct io_sq_data *);
+	for_each_possible_cpu(cpu)
+		*per_cpu_ptr(percpu_sqd, cpu) = NULL;
+
 	return 0;
 };
 __initcall(io_uring_init);
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 557e7eae497f..5bb958359d2f 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -98,6 +98,7 @@ enum {
 #define IORING_SETUP_CLAMP	(1U << 4)	/* clamp SQ/CQ ring sizes */
 #define IORING_SETUP_ATTACH_WQ	(1U << 5)	/* attach to existing wq */
 #define IORING_SETUP_R_DISABLED	(1U << 6)	/* start with ring disabled */
+#define IORING_SETUP_SQPOLL_PERCPU	(1U << 7)	/* use percpu SQ poll thread */
 
 enum {
 	IORING_OP_NOP,
-- 
2.17.2

Re: [PATCH v2 0/2] improve SQPOLL handling

[Xiaoguang Wang]

hi,

A gentle reminder. How does this patch set look now?
I think the first patch looks ok at least.

Regards,
Xiaoguang Wang

> The first patch tries to improve various issues in current implementation:
>    The prepare_to_wait() usage in __io_sq_thread() is weird. If multiple ctxs
> share one same poll thread, one ctx will put poll thread in TASK_INTERRUPTIBLE,
> but if other ctxs have work to do, we don't need to change task's stat at all.
> I think only if all ctxs don't have work to do, we can do it.
>    We use round-robin strategy to make multiple ctxs share one same poll thread,
> but there are various condition in __io_sq_thread(), which seems complicated and
> may affect round-robin strategy.
> 
> The second patch adds a IORING_SETUP_SQPOLL_PERCPU flag, for those rings which
> have SQPOLL enabled and are willing to be bound to one same cpu, hence share
> one same poll thread, add a capability that these rings can share one poll thread
> by specifying a new IORING_SETUP_SQPOLL_PERCPU flag. FIO tool can integrate this
> feature easily, so we can test multiple rings to share same poll thread easily.
> 
> 
> TEST:
>    This patch set have passed liburing test cases.
> 
>    I also make fio support IORING_SETUP_SQPOLL_PERCPU flag, and make some
> io stress tests, no errors or performance regression. See below fio job file:
> 
> First in unpatched kernel, I test a fio file which only contains one job
> with iodepth being 128, see below:
> [global]
> ioengine=io_uring
> sqthread_poll=1
> registerfiles=1
> fixedbufs=1
> hipri=1
> thread=1
> bs=4k
> direct=1
> rw=randread
> time_based=1
> runtime=120
> ramp_time=0
> randrepeat=0
> group_reporting=1
> filename=/dev/nvme0n1
> sqthread_poll_cpu=15
> 
> [job0]
> cpus_allowed=5
> iodepth=128
> sqthread_poll_cpu=9
> 
> performance data: IOPS: 453k, avg lat: 282.37usec
> 
> 
> Second in unpatched kernel, I test a fio file which contains 4 jobs
> with each iodepth being 32, see below:
> [global]
> ioengine=io_uring
> sqthread_poll=1
> registerfiles=1
> fixedbufs=1
> hipri=1
> thread=1
> bs=4k
> direct=1
> rw=randread
> time_based=1
> runtime=120
> ramp_time=0
> randrepeat=0
> group_reporting=1
> filename=/dev/nvme0n1
> sqthread_poll_cpu=15
> 
> [job0]
> cpus_allowed=5
> iodepth=32
> sqthread_poll_cpu=9
> 
> [job1]
> cpus_allowed=6
> iodepth=32
> sqthread_poll_cpu=9
> 
> [job2]
> cpus_allowed=7
> iodepth=32
> sqthread_poll_cpu=9
> 
> [job3]
> cpus_allowed=8
> iodepth=32
> sqthread_poll_cpu=9
> performance data: IOPS: 254k, avg lat: 503.80 usec, obvious performance
> drop.
> 
> Finally in patched kernel, I test a fio file which contains 4 jobs
> with each iodepth being 32, and now we enable sqthread_poll_percpu
> flag, see blow:
> 
> [global]
> ioengine=io_uring
> sqthread_poll=1
> registerfiles=1
> fixedbufs=1
> hipri=1
> thread=1
> bs=4k
> direct=1
> rw=randread
> time_based=1
> runtime=120
> ramp_time=0
> randrepeat=0
> group_reporting=1
> filename=/dev/nvme0n1
> #sqthread_poll_cpu=15
> sqthread_poll_percpu=1  # enable percpu feature
> 
> [job0]
> cpus_allowed=5
> iodepth=32
> sqthread_poll_cpu=9
> 
> [job1]
> cpus_allowed=6
> iodepth=32
> sqthread_poll_cpu=9
> 
> [job2]
> cpus_allowed=7
> iodepth=32
> sqthread_poll_cpu=9
> 
> performance data: IOPS: 438k, avg lat: 291.69usec
> 
> 
>  From above teses, we can see that IORING_SETUP_SQPOLL_PERCPU is easy to
> use, and no obvious performance regression.
> Note I don't test IORING_SETUP_ATTACH_WQ in above three test cases, it's
> a little hard to support IORING_SETUP_ATTACH_WQ in fio.
> 
> Xiaoguang Wang (2):
>    io_uring: refactor io_sq_thread() handling
>    io_uring: support multiple rings to share same poll thread by
>      specifying same cpu
> 
>   fs/io_uring.c                 | 289 +++++++++++++++++++---------------
>   include/uapi/linux/io_uring.h |   1 +
>   2 files changed, 166 insertions(+), 124 deletions(-)
> 

Re: [PATCH v2 0/2] improve SQPOLL handling

[Jens Axboe]

On 11/8/20 7:16 AM, Xiaoguang Wang wrote:
> hi,
> 
> A gentle reminder. How does this patch set look now?
> I think the first patch looks ok at least.

I have applied 1/2 for now, I agree that one looks fine and should get
applied. I'll go over 2/2 soonish.

-- 
Jens Axboe

Re: [PATCH v2 0/2] improve SQPOLL handling

[Xiaoguang Wang]

hi,

> On 11/8/20 7:16 AM, Xiaoguang Wang wrote:
>> hi,
>>
>> A gentle reminder. How does this patch set look now?
>> I think the first patch looks ok at least.
> 
> I have applied 1/2 for now, I agree that one looks fine and should get
> applied. I'll go over 2/2 soonish.
Thanks :)

Regards,
Xiaoguang Wang

>