Skip to contentSkip to conversation

fixed worker

PATCH In review Hao Xu 1 participant 13 messages Jun 27, 2022 – Jul 12, 2022 io_uring

lore
Share & tools
lore

Hao Xu

author
#0 lore raw

v5 cover fixed worker

From: Hao Xu <[email protected]>

This is the second version of fixed worker implementation. Wrote a nop test program to test it, 3 fixed-workers VS 3 normal workers. normal workers:
./run_nop_wqe.sh nop_wqe_normal 200000 100 3 1-3

        time spent: 17314274 usecs      IOPS: 1155116
        time spent: 17016942 usecs      IOPS: 1175299
        time spent: 17908684 usecs      IOPS: 1116776

fixed workers:
./run_nop_wqe.sh nop_wqe_fixed 200000 100 3 1-3

        time spent: 10464397 usecs      IOPS: 1911242
        time spent: 9610976 usecs       IOPS: 2080954
        time spent: 9807361 usecs       IOPS: 2039284

About 2x improvement. From perf result, almost no acct->lock contension. Test program: https://github.com/HowHsu/liburing/tree/fixed_worker
liburing/test/nop_wqe.c

v3->v4:
- make work in fixed worker's private worfixed worker
- tweak the io_wqe_acct struct to make it clearer

v4->v5:
- code clean
- rebase against the for-5.20/io_uring

Hao Xu (11):
io-wq: add a worker flag for individual exit
io-wq: change argument of create_io_worker() for convienence
io-wq: add infra data structure for fixed workers
io-wq: tweak io_get_acct()
io-wq: fixed worker initialization
io-wq: fixed worker exit
io-wq: implement fixed worker logic
io-wq: batch the handling of fixed worker private works
io_uring: add register fixed worker interface
io-wq: add an work list for fixed worker
io_uring: cancel works in exec work list for fixed worker

 include/uapi/linux/io_uring.h |  11 +
 io_uring/io-wq.c              | 495 ++++++++++++++++++++++++++++++----
 io_uring/io-wq.h              |   8 +
 io_uring/io_uring.c           |  71 +++++
 4 files changed, 537 insertions(+), 48 deletions(-)

base-commit: 094abe8fbccb0d79bef982c67eb7372e92452c0e

Signature · 1 line
2.25.1
author
#1 lore raw

PATCH v5 1/11 io-wq: add a worker flag for individual exit

From: Hao Xu <[email protected]>

Add a worker flag to control exit of an individual worker, this is needed for fixed worker in the next patches but also as a generic functionality.

---
 io_uring/io-wq.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
io_uring/io-wq.c +10 -1
index 824623bcf1a5..0c26805ca6de 100644
@@ -26,6 +26,7 @@ enum {
 	IO_WORKER_F_RUNNING	= 2,	/* account as running */
 	IO_WORKER_F_FREE	= 4,	/* worker on free list */
 	IO_WORKER_F_BOUND	= 8,	/* is doing bounded work */
+	IO_WORKER_F_EXIT	= 32,	/* worker is exiting */
 };
 
 enum {
@@ -639,8 +640,12 @@ static int io_wqe_worker(void *data)
 	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 		long ret;
 
+		if (worker->flags & IO_WORKER_F_EXIT)
+			break;
+
 		set_current_state(TASK_INTERRUPTIBLE);
-		while (io_acct_run_queue(acct))
+		while (!(worker->flags & IO_WORKER_F_EXIT) &&
+		       io_acct_run_queue(acct))
 			io_worker_handle_work(worker);
 
 		raw_spin_lock(&wqe->lock);
@@ -656,6 +661,10 @@ static int io_wqe_worker(void *data)
 		raw_spin_unlock(&wqe->lock);
 		if (io_flush_signals())
 			continue;
+		if (worker->flags & IO_WORKER_F_EXIT) {
+			__set_current_state(TASK_RUNNING);
+			break;
+		}
 		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 		if (signal_pending(current)) {
 			struct ksignal ksig;
Signature · 1 line
2.25.1
author
#2 lore raw

PATCH v5 2/11 io-wq: change argument of create_io_worker() for convienence

In reply to Hao Xu

From: Hao Xu <[email protected]>

Change index to acct itself for create_io_worker() for convienence in the next patches.

---
 io_uring/io-wq.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)
io_uring/io-wq.c +7 -5
index 0c26805ca6de..35ce622f77ba 100644
@@ -139,7 +139,8 @@ struct io_cb_cancel_data {
 	bool cancel_all;
 };
 
-static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
+static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe,
+			     struct io_wqe_acct *acct);
 static void io_wqe_dec_running(struct io_worker *worker);
 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 					struct io_wqe_acct *acct,
@@ -306,7 +307,7 @@ static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 	raw_spin_unlock(&wqe->lock);
 	atomic_inc(&acct->nr_running);
 	atomic_inc(&wqe->wq->worker_refs);
-	return create_io_worker(wqe->wq, wqe, acct->index);
+	return create_io_worker(wqe->wq, wqe, acct);
 }
 
 static void io_wqe_inc_running(struct io_worker *worker)
@@ -335,7 +336,7 @@ static void create_worker_cb(struct callback_head *cb)
 	}
 	raw_spin_unlock(&wqe->lock);
 	if (do_create) {
-		create_io_worker(wq, wqe, worker->create_index);
+		create_io_worker(wq, wqe, acct);
 	} else {
 		atomic_dec(&acct->nr_running);
 		io_worker_ref_put(wq);
@@ -812,9 +813,10 @@ static void io_workqueue_create(struct work_struct *work)
 		kfree(worker);
 }
 
-static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
+static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe,
+			     struct io_wqe_acct *acct)
 {
-	struct io_wqe_acct *acct = &wqe->acct[index];
+	int index = acct->index;
 	struct io_worker *worker;
 	struct task_struct *tsk;
 
Signature · 1 line
2.25.1
author
#3 lore raw

PATCH v5 3/11 io-wq: add infra data structure for fixed workers

In reply to Hao Xu

From: Hao Xu <[email protected]>

Add data sttructure and basic initialization for fixed worker.

---
 io_uring/io-wq.c | 103 ++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 92 insertions(+), 11 deletions(-)
io_uring/io-wq.c +92 -11
index 35ce622f77ba..d9b3aeea2c6a 100644
@@ -26,6 +26,7 @@ enum {
 	IO_WORKER_F_RUNNING	= 2,	/* account as running */
 	IO_WORKER_F_FREE	= 4,	/* worker on free list */
 	IO_WORKER_F_BOUND	= 8,	/* is doing bounded work */
+	IO_WORKER_F_FIXED	= 16,	/* is a fixed worker */
 	IO_WORKER_F_EXIT	= 32,	/* worker is exiting */
 };
 
@@ -37,6 +38,66 @@ enum {
 	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
 };
 
+struct io_wqe_acct {
+/*
+ *	union {
+ *		// (1) for wqe->acct (normal worker)
+ *		struct {
+ *			unsigned nr_workers;
+ *			unsigned max_workers;
+ *			struct io_wq_work_list work_list;
+ *		};
+ *		// (2) for wqe->fixed_acct (fixed worker)
+ *		struct {
+ *			unsigned nr_workers;
+ *			unsigned max_workers;
+ *			unsigned nr_fixed;
+ *			unsigned max_works;
+ *			struct io_worker **fixed_workers;
+ *		};
+ *		// (3) for fixed worker's private acct
+ *		struct {
+ *			unsigned nr_works;
+ *			unsigned max_works;
+ *			struct io_wq_work_list work_list;
+ *		};
+ *	};
+ */
+	union {
+		struct {
+			unsigned nr_workers;
+			unsigned max_workers;
+		};
+		unsigned nr_works;
+	};
+	unsigned max_works;
+	union {
+		struct io_wq_work_list work_list;
+		struct {
+			unsigned nr_fixed;
+			struct io_worker **fixed_workers;
+		};
+	};
+
+	/*
+	 * nr_running is not meaningful for fixed worker
+	 * but still keep the same logic for it for the
+	 * convinence for now. So do nr_workers and
+	 * max_workers.
+	 */
+	atomic_t nr_running;
+	/*
+	 * For 1), it protects the work_list, the other two member nr_workers
+	 * and max_workers are protected by wqe->lock.
+	 * For 2), it protects nr_fixed, max_works, fixed_workers
+	 * For 3), it protects nr_works, max_works and work_list.
+	 */
+	raw_spinlock_t lock;
+	int index;
+	unsigned long flags;
+	bool fixed_worker_registered;
+};
+
 /*
  * One for each thread in a wqe pool
  */
@@ -62,6 +123,8 @@ struct io_worker {
 		struct rcu_head rcu;
 		struct work_struct work;
 	};
+	int index;
+	struct io_wqe_acct acct;
 };
 
 #if BITS_PER_LONG == 64
@@ -72,16 +135,6 @@ struct io_worker {
 
 #define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
 
-struct io_wqe_acct {
-	unsigned nr_workers;
-	unsigned max_workers;
-	int index;
-	atomic_t nr_running;
-	raw_spinlock_t lock;
-	struct io_wq_work_list work_list;
-	unsigned long flags;
-};
-
 enum {
 	IO_WQ_ACCT_BOUND,
 	IO_WQ_ACCT_UNBOUND,
@@ -94,6 +147,7 @@ enum {
 struct io_wqe {
 	raw_spinlock_t lock;
 	struct io_wqe_acct acct[IO_WQ_ACCT_NR];
+	struct io_wqe_acct fixed_acct[IO_WQ_ACCT_NR];
 
 	int node;
 
@@ -1205,6 +1259,31 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 			atomic_set(&acct->nr_running, 0);
 			INIT_WQ_LIST(&acct->work_list);
 			raw_spin_lock_init(&acct->lock);
+
+			acct = &wqe->fixed_acct[i];
+			acct->index = i;
+			INIT_WQ_LIST(&acct->work_list);
+			raw_spin_lock_init(&acct->lock);
+			/*
+			 * nr_running for a fixed worker is meaningless
+			 * for now, init it to 1 to wround around the
+			 * io_wqe_dec_running logic
+			 */
+			atomic_set(&acct->nr_running, 1);
+			/*
+			 * max_workers for a fixed worker is meaningless
+			 * for now, init it so since number of fixed workers
+			 * should be controlled by users.
+			 */
+			acct->max_workers = task_rlimit(current, RLIMIT_NPROC);
+			raw_spin_lock_init(&acct->lock);
+			/*
+			 * For fixed worker, not necessary
+			 * but do it explicitly for clearity
+			 */
+			acct->nr_fixed = 0;
+			acct->max_works = 0;
+			acct->fixed_workers = NULL;
 		}
 		wqe->wq = wq;
 		raw_spin_lock_init(&wqe->lock);
@@ -1287,7 +1366,7 @@ static void io_wq_exit_workers(struct io_wq *wq)
 
 static void io_wq_destroy(struct io_wq *wq)
 {
-	int node;
+	int i, node;
 
 	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
 
@@ -1299,6 +1378,8 @@ static void io_wq_destroy(struct io_wq *wq)
 		};
 		io_wqe_cancel_pending_work(wqe, &match);
 		free_cpumask_var(wqe->cpu_mask);
+		for (i = 0; i < IO_WQ_ACCT_NR; i++)
+			kfree(wqe->fixed_acct[i].fixed_workers);
 		kfree(wqe);
 	}
 	io_wq_put_hash(wq->hash);
Signature · 1 line
2.25.1
author
#4 lore raw

PATCH v5 4/11 io-wq: tweak io_get_acct()

In reply to Hao Xu

From: Hao Xu <[email protected]>

Add an argument for io_get_acct() to indicate fixed or normal worker

---
 io_uring/io-wq.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
io_uring/io-wq.c +9 -5
index d9b3aeea2c6a..7775ba5fddba 100644
@@ -213,20 +213,24 @@ static void io_worker_release(struct io_worker *worker)
 		complete(&worker->ref_done);
 }
 
-static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
+static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound,
+					      bool fixed)
 {
-	return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
+	unsigned index = bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND;
+
+	return fixed ? &wqe->fixed_acct[index] : &wqe->acct[index];
 }
 
 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 						   struct io_wq_work *work)
 {
-	return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
+	return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND), false);
 }
 
 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
 {
-	return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
+	return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND,
+			   worker->flags & IO_WORKER_F_FIXED);
 }
 
 static void io_worker_ref_put(struct io_wq *wq)
@@ -1129,7 +1133,7 @@ static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 	int i;
 retry:
 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
-		struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
+		struct io_wqe_acct *acct = io_get_acct(wqe, i == 0, false);
 
 		if (io_acct_cancel_pending_work(wqe, acct, match)) {
 			if (match->cancel_all)
Signature · 1 line
2.25.1
author
#5 lore raw

PATCH v5 5/11 io-wq: fixed worker initialization

In reply to Hao Xu

From: Hao Xu <[email protected]>

Implementation of the fixed worker initialization.

---
 io_uring/io-wq.c | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)
io_uring/io-wq.c +24 -0
index 7775ba5fddba..70240b847112 100644
@@ -779,6 +779,26 @@ void io_wq_worker_sleeping(struct task_struct *tsk)
 	io_wqe_dec_running(worker);
 }
 
+static void io_init_new_fixed_worker(struct io_wqe *wqe,
+				     struct io_worker *worker)
+{
+	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
+	struct io_wqe_acct *iw_acct = &worker->acct;
+	unsigned index = acct->index;
+	unsigned *nr_fixed;
+
+	raw_spin_lock(&acct->lock);
+	nr_fixed = &acct->nr_fixed;
+	acct->fixed_workers[*nr_fixed] = worker;
+	worker->index = (*nr_fixed)++;
+	iw_acct->nr_works = 0;
+	iw_acct->max_works = acct->max_works;
+	iw_acct->index = index;
+	INIT_WQ_LIST(&iw_acct->work_list);
+	raw_spin_lock_init(&iw_acct->lock);
+	raw_spin_unlock(&acct->lock);
+}
+
 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
 			       struct task_struct *tsk)
 {
@@ -792,6 +812,8 @@ static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
 	list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 	worker->flags |= IO_WORKER_F_FREE;
 	raw_spin_unlock(&wqe->lock);
+	if (worker->flags & IO_WORKER_F_FIXED)
+		io_init_new_fixed_worker(wqe, worker);
 	wake_up_new_task(tsk);
 }
 
@@ -898,6 +920,8 @@ static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe,
 
 	if (index == IO_WQ_ACCT_BOUND)
 		worker->flags |= IO_WORKER_F_BOUND;
+	if (&wqe->fixed_acct[index] == acct)
+		worker->flags |= IO_WORKER_F_FIXED;
 
 	tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
 	if (!IS_ERR(tsk)) {
Signature · 1 line
2.25.1
author
#6 lore raw

PATCH v5 6/11 io-wq: fixed worker exit

In reply to Hao Xu

From: Hao Xu <[email protected]>

Implement the fixed worker exit

---
 io_uring/io-wq.c | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)
io_uring/io-wq.c +26 -0
index 70240b847112..11fe5787c409 100644
@@ -264,6 +264,29 @@ static bool io_task_worker_match(struct callback_head *cb, void *data)
 	return worker == data;
 }
 
+static void io_fixed_worker_exit(struct io_worker *worker)
+{
+	int *nr_fixed;
+	int index = worker->acct.index;
+	struct io_wqe *wqe = worker->wqe;
+	struct io_wqe_acct *acct = io_get_acct(wqe, index == 0, true);
+	struct io_worker **fixed_workers;
+
+	raw_spin_lock(&acct->lock);
+	fixed_workers = acct->fixed_workers;
+	if (!fixed_workers || worker->index == -1) {
+		raw_spin_unlock(&acct->lock);
+		return;
+	}
+	nr_fixed = &acct->nr_fixed;
+	/* reuse variable index to represent fixed worker index in its array */
+	index = worker->index;
+	fixed_workers[index] = fixed_workers[*nr_fixed - 1];
+	(*nr_fixed)--;
+	fixed_workers[index]->index = index;
+	raw_spin_unlock(&acct->lock);
+}
+
 static void io_worker_exit(struct io_worker *worker)
 {
 	struct io_wqe *wqe = worker->wqe;
@@ -687,6 +710,7 @@ static int io_wqe_worker(void *data)
 	struct io_wqe *wqe = worker->wqe;
 	struct io_wq *wq = wqe->wq;
 	bool last_timeout = false;
+	bool fixed = worker->flags & IO_WORKER_F_FIXED;
 	char buf[TASK_COMM_LEN];
 
 	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
@@ -737,6 +761,8 @@ static int io_wqe_worker(void *data)
 
 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
 		io_worker_handle_work(worker);
+	if (fixed)
+		io_fixed_worker_exit(worker);
 
 	audit_free(current);
 	io_worker_exit(worker);
Signature · 1 line
2.25.1
author
#7 lore raw

PATCH v5 7/11 io-wq: implement fixed worker logic

In reply to Hao Xu

From: Hao Xu <[email protected]>

The current implementation of io-wq has big spinlock contension. The main reason is the single work list model. All producers(who insert works) and consumers(io-workers) have to grap wqe->lock to move ahead. Set max_worker to 3 or 4, do a fio read test, we can see 40%~50% lock contension.
Introduce fixed io-workers which sticks there to handle works and have their own work list.

previous:

producer0 ---insert---> work_list ---get---> io-worker0,1,2

now:

                     ---> private work_list0 --get--> fixed-worker0
                    /
  producer0 --insert----> private work_list1 --get--> fixed-worker1
      |             \
      |              ---> private work_list2 --get--> fixed-worker2
      |
      |---insert---> public work_list --get--> (normal)io-worker

Since each fixed-worker has a private work list, the contension will be limited to a smaller range(the private work list).
Logic of fixed-worker: first handle private works then public ones. Logic of normal io-worker: only handle public works.
Logic of producer: 1) randomly pick a private work list and check if it
is full, insert the work if it's not
2) insert the work to the public work list if 1) fails.
The get logic of a private list: fixed-worker grab all the works in its private work list(like what tctx_task_work() does) rather than one by one.(this code is in the next patches as a optimization)

To achieve this, we need to add an io_wqe_acct for each fixed-worker struct, and though this we can leverage the old code as much as possible, which makes the new design clean and compatible.

Good things of this feature:
1) bound and unbound work lists now have different spinlocks.
2) much smaller contension between work producers and consumers.
3) fixed workers are friendly for users to control: binding cpus, reset priority etc.

Wrote a nop test program to test it, 3 fixed-workers VS 3 normal workers. normal workers:
./run_nop_wqe.sh nop_wqe_normal 200000 100 3 1-3

        time spent: 17314274 usecs      IOPS: 1155116
        time spent: 17016942 usecs      IOPS: 1175299
        time spent: 17908684 usecs      IOPS: 1116776

fixed workers:
./run_nop_wqe.sh nop_wqe_fixed 200000 100 3 1-3

        time spent: 10464397 usecs      IOPS: 1911242
        time spent: 9610976 usecs       IOPS: 2080954
        time spent: 9807361 usecs       IOPS: 2039284

About 2x improvement. From perf result, almost no acct->lock contension. Test program: https://github.com/HowHsu/liburing/tree/fixed_worker

---
 io_uring/io-wq.c | 148 ++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 122 insertions(+), 26 deletions(-)
io_uring/io-wq.c +122 -26
index 11fe5787c409..c87ba38f27b1 100644
@@ -71,6 +71,7 @@ struct io_wqe_acct {
 		unsigned nr_works;
 	};
 	unsigned max_works;
+	unsigned work_seq;
 	union {
 		struct io_wq_work_list work_list;
 		struct {
@@ -629,9 +630,9 @@ static void io_assign_current_work(struct io_worker *worker,
 
 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 
-static void io_worker_handle_work(struct io_worker *worker)
+static void io_worker_handle_work(struct io_worker *worker,
+				  struct io_wqe_acct *acct)
 {
-	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 	struct io_wqe *wqe = worker->wqe;
 	struct io_wq *wq = wqe->wq;
 	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
@@ -703,19 +704,31 @@ static void io_worker_handle_work(struct io_worker *worker)
 	} while (1);
 }
 
+static inline void io_worker_handle_private_work(struct io_worker *worker)
+{
+	io_worker_handle_work(worker, &worker->acct);
+}
+
+static inline void io_worker_handle_public_work(struct io_worker *worker)
+{
+	io_worker_handle_work(worker, io_wqe_get_acct(worker));
+}
+
 static int io_wqe_worker(void *data)
 {
 	struct io_worker *worker = data;
-	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 	struct io_wqe *wqe = worker->wqe;
 	struct io_wq *wq = wqe->wq;
-	bool last_timeout = false;
+	struct io_wqe_acct *acct =
+		io_get_acct(wqe, worker->flags & IO_WORKER_F_BOUND, false);
 	bool fixed = worker->flags & IO_WORKER_F_FIXED;
+	bool last_timeout = false;
 	char buf[TASK_COMM_LEN];
 
 	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 
-	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
+	snprintf(buf, sizeof(buf), fixed ? "iou-fix-%d" : "iou-wrk-%d",
+		 wq->task->pid);
 	set_task_comm(current, buf);
 
 	audit_alloc_kernel(current);
@@ -727,13 +740,24 @@ static int io_wqe_worker(void *data)
 			break;
 
 		set_current_state(TASK_INTERRUPTIBLE);
-		while (!(worker->flags & IO_WORKER_F_EXIT) &&
-		       io_acct_run_queue(acct))
-			io_worker_handle_work(worker);
-
+		if (fixed) {
+			while (io_acct_run_queue(&worker->acct))
+				io_worker_handle_private_work(worker);
+			if (io_acct_run_queue(acct))
+				io_worker_handle_public_work(worker);
+		} else {
+			while (io_acct_run_queue(acct))
+				io_worker_handle_public_work(worker);
+		}
 		raw_spin_lock(&wqe->lock);
-		/* timed out, exit unless we're the last worker */
-		if (last_timeout && acct->nr_workers > 1) {
+		/* timed out, a worker will exit only if:
+		 * - not a fixed worker
+		 * - not the last non-fixed worker
+		 *
+		 * the second condition is due to we need at least one worker to
+		 * handle the public work list.
+		 */
+		if (last_timeout && !fixed && acct->nr_workers > 1) {
 			acct->nr_workers--;
 			raw_spin_unlock(&wqe->lock);
 			__set_current_state(TASK_RUNNING);
@@ -759,10 +783,18 @@ static int io_wqe_worker(void *data)
 		last_timeout = !ret;
 	}
 
-	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
-		io_worker_handle_work(worker);
-	if (fixed)
+	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && !fixed)
+		io_worker_handle_public_work(worker);
+	if (fixed) {
 		io_fixed_worker_exit(worker);
+		/*
+		 * Check and handle private work list again
+		 * to avoid race with private work insertion
+		 * TODO: an alternative way is to deliver
+		 * works to the public work list
+		 */
+		io_worker_handle_private_work(worker);
+	}
 
 	audit_free(current);
 	io_worker_exit(worker);
@@ -1006,9 +1038,9 @@ static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 	} while (work);
 }
 
-static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
+static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work,
+			       struct io_wqe_acct *acct)
 {
-	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 	unsigned int hash;
 	struct io_wq_work *tail;
 
@@ -1027,6 +1059,45 @@ static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
 }
 
+static bool io_wqe_insert_private_work(struct io_wqe *wqe,
+				       struct io_wq_work *work,
+				       struct io_wqe_acct *acct)
+{
+	unsigned int nr_fixed;
+	struct io_worker *fixed_worker;
+	struct io_wqe_acct *iw_acct;
+	unsigned int fixed_worker_index;
+
+	raw_spin_lock(&acct->lock);
+	nr_fixed = acct->nr_fixed;
+	if (!nr_fixed) {
+		raw_spin_unlock(&acct->lock);
+		return false;
+	}
+
+	fixed_worker_index = (acct->work_seq++) % nr_fixed;
+	fixed_worker = acct->fixed_workers[fixed_worker_index];
+	if (!fixed_worker || fixed_worker->flags & IO_WORKER_F_EXIT) {
+		raw_spin_unlock(&acct->lock);
+		return false;
+	}
+	iw_acct = &fixed_worker->acct;
+
+	raw_spin_lock(&iw_acct->lock);
+	if (iw_acct->nr_works < iw_acct->max_works) {
+		io_wqe_insert_work(wqe, work, iw_acct);
+		iw_acct->nr_works++;
+		raw_spin_unlock(&iw_acct->lock);
+		wake_up_process(fixed_worker->task);
+		raw_spin_unlock(&acct->lock);
+		return true;
+	}
+	raw_spin_unlock(&iw_acct->lock);
+	raw_spin_unlock(&acct->lock);
+
+	return false;
+}
+
 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 {
 	return work == data;
@@ -1035,6 +1106,7 @@ static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 {
 	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
+	struct io_wqe_acct *fixed_acct;
 	struct io_cb_cancel_data match;
 	unsigned work_flags = work->flags;
 	bool do_create;
@@ -1049,8 +1121,14 @@ static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 		return;
 	}
 
+	fixed_acct = io_get_acct(wqe, !acct->index, true);
+	if (fixed_acct->fixed_worker_registered && !io_wq_is_hashed(work)) {
+		if (io_wqe_insert_private_work(wqe, work, fixed_acct))
+			return;
+	}
+
 	raw_spin_lock(&acct->lock);
-	io_wqe_insert_work(wqe, work);
+	io_wqe_insert_work(wqe, work, acct);
 	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 	raw_spin_unlock(&acct->lock);
 
@@ -1136,9 +1214,9 @@ static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 
 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 					 struct io_wq_work *work,
-					 struct io_wq_work_node *prev)
+					 struct io_wq_work_node *prev,
+					 struct io_wqe_acct *acct)
 {
-	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 	unsigned int hash = io_get_work_hash(work);
 	struct io_wq_work *prev_work = NULL;
 
@@ -1165,7 +1243,7 @@ static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 		work = container_of(node, struct io_wq_work, list);
 		if (!match->fn(work, match->data))
 			continue;
-		io_wqe_remove_pending(wqe, work, prev);
+		io_wqe_remove_pending(wqe, work, prev, acct);
 		raw_spin_unlock(&acct->lock);
 		io_run_cancel(work, wqe);
 		match->nr_pending++;
@@ -1180,17 +1258,35 @@ static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 				       struct io_cb_cancel_data *match)
 {
-	int i;
-retry:
-	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
-		struct io_wqe_acct *acct = io_get_acct(wqe, i == 0, false);
+	int i, j;
+	struct io_wqe_acct *acct, *iw_acct;
 
+retry_public:
+	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
+		acct = io_get_acct(wqe, i == 0, false);
 		if (io_acct_cancel_pending_work(wqe, acct, match)) {
 			if (match->cancel_all)
-				goto retry;
-			break;
+				goto retry_public;
+			return;
 		}
 	}
+
+retry_private:
+	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
+		acct = io_get_acct(wqe, i == 0, true);
+		raw_spin_lock(&acct->lock);
+		for (j = 0; j < acct->nr_fixed; j++) {
+			iw_acct = &acct->fixed_workers[j]->acct;
+			if (io_acct_cancel_pending_work(wqe, iw_acct, match)) {
+				if (match->cancel_all) {
+					raw_spin_unlock(&acct->lock);
+					goto retry_private;
+				}
+				break;
+			}
+		}
+		raw_spin_unlock(&acct->lock);
+	}
 }
 
 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
Signature · 1 line
2.25.1
author
#8 lore raw

PATCH v5 8/11 io-wq: batch the handling of fixed worker private works

In reply to Hao Xu

From: Hao Xu <[email protected]>

Reduce acct->lock contension by batching the handling of private work list for fixed_workers.

---
 io_uring/io-wq.c | 42 +++++++++++++++++++++++++++++++++---------
 io_uring/io-wq.h |  5 +++++
 2 files changed, 38 insertions(+), 9 deletions(-)
io_uring/io-wq.c +33 -9
index c87ba38f27b1..ce754c78ecac 100644
@@ -545,8 +545,23 @@ static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
 	return ret;
 }
 
+static inline void conditional_acct_lock(struct io_wqe_acct *acct,
+					 bool needs_lock)
+{
+	if (needs_lock)
+		raw_spin_lock(&acct->lock);
+}
+
+static inline void conditional_acct_unlock(struct io_wqe_acct *acct,
+					   bool needs_lock)
+{
+	if (needs_lock)
+		raw_spin_unlock(&acct->lock);
+}
+
 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
-					   struct io_worker *worker)
+					   struct io_worker *worker,
+					   bool needs_lock)
 	__must_hold(acct->lock)
 {
 	struct io_wq_work_node *node, *prev;
@@ -554,6 +569,7 @@ static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
 	unsigned int stall_hash = -1U;
 	struct io_wqe *wqe = worker->wqe;
 
+	conditional_acct_lock(acct, needs_lock);
 	wq_list_for_each(node, prev, &acct->work_list) {
 		unsigned int hash;
 
@@ -562,6 +578,7 @@ static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
 		/* not hashed, can run anytime */
 		if (!io_wq_is_hashed(work)) {
 			wq_list_del(&acct->work_list, node, prev);
+			conditional_acct_unlock(acct, needs_lock);
 			return work;
 		}
 
@@ -573,6 +590,7 @@ static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
 		if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
 			wqe->hash_tail[hash] = NULL;
 			wq_list_cut(&acct->work_list, &tail->list, prev);
+			conditional_acct_unlock(acct, needs_lock);
 			return work;
 		}
 		if (stall_hash == -1U)
@@ -589,15 +607,16 @@ static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
 		 * work being added and clearing the stalled bit.
 		 */
 		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
-		raw_spin_unlock(&acct->lock);
+		conditional_acct_unlock(acct, needs_lock);
 		unstalled = io_wait_on_hash(wqe, stall_hash);
-		raw_spin_lock(&acct->lock);
+		conditional_acct_lock(acct, needs_lock);
 		if (unstalled) {
 			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 			if (wq_has_sleeper(&wqe->wq->hash->wait))
 				wake_up(&wqe->wq->hash->wait);
 		}
 	}
+	conditional_acct_unlock(acct, needs_lock);
 
 	return NULL;
 }
@@ -631,7 +650,7 @@ static void io_assign_current_work(struct io_worker *worker,
 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 
 static void io_worker_handle_work(struct io_worker *worker,
-				  struct io_wqe_acct *acct)
+				  struct io_wqe_acct *acct, bool needs_lock)
 {
 	struct io_wqe *wqe = worker->wqe;
 	struct io_wq *wq = wqe->wq;
@@ -647,9 +666,7 @@ static void io_worker_handle_work(struct io_worker *worker,
 		 * can't make progress, any work completion or insertion will
 		 * clear the stalled flag.
 		 */
-		raw_spin_lock(&acct->lock);
-		work = io_get_next_work(acct, worker);
-		raw_spin_unlock(&acct->lock);
+		work = io_get_next_work(acct, worker, needs_lock);
 		if (work) {
 			__io_worker_busy(wqe, worker);
 
@@ -706,12 +723,19 @@ static void io_worker_handle_work(struct io_worker *worker,
 
 static inline void io_worker_handle_private_work(struct io_worker *worker)
 {
-	io_worker_handle_work(worker, &worker->acct);
+	struct io_wqe_acct acct;
+
+	raw_spin_lock(&worker->acct.lock);
+	acct = worker->acct;
+	wq_list_clean(&worker->acct.work_list);
+	worker->acct.nr_works = 0;
+	raw_spin_unlock(&worker->acct.lock);
+	io_worker_handle_work(worker, &acct, false);
 }
 
 static inline void io_worker_handle_public_work(struct io_worker *worker)
 {
-	io_worker_handle_work(worker, io_wqe_get_acct(worker));
+	io_worker_handle_work(worker, io_wqe_get_acct(worker), true);
 }
 
 static int io_wqe_worker(void *data)
io_uring/io-wq.h +5 -0
index 10b80ef78bb8..78efbb8c53f0 100644
@@ -32,6 +32,11 @@ enum io_wq_cancel {
 	(list)->first = NULL;					\
 } while (0)
 
+static inline void wq_list_clean(struct io_wq_work_list *list)
+{
+	list->first = list->last = NULL;
+}
+
 static inline void wq_list_add_after(struct io_wq_work_node *node,
 				     struct io_wq_work_node *pos,
 				     struct io_wq_work_list *list)
Signature · 1 line
2.25.1
author
#9 lore raw

PATCH v5 9/11 io_uring: add register fixed worker interface

In reply to Hao Xu

From: Hao Xu <[email protected]>

Add an io_uring_register() interface to register fixed workers and indicate its work capacity.
The argument is an array of two elements each is

    struct {
    	__s32 nr_workers;
    	__s32 max_works;
    }

(nr_workers, max_works) meaning

nr_workers or max_works < -1 invalid
nr_workers or max_works == -1 get the old value back
nr_workers or max_works >= 0 get the old value and set to the
value

---
 include/uapi/linux/io_uring.h |  11 ++++
 io_uring/io-wq.c              | 101 ++++++++++++++++++++++++++++++++++
 io_uring/io-wq.h              |   3 +
 io_uring/io_uring.c           |  71 ++++++++++++++++++++++++
 4 files changed, 186 insertions(+)
include/uapi/linux/io_uring.h +11 -0
index 8715f0942ec2..5480829d07c0 100644
@@ -423,6 +423,12 @@ enum {
 	IORING_REGISTER_PBUF_RING		= 22,
 	IORING_UNREGISTER_PBUF_RING		= 23,
 
+	/* set number of fixed workers and number
+	 * of works in a private work list which
+	 * belongs to a fixed worker
+	 */
+	IORING_REGISTER_IOWQ_FIXED_WORKERS	= 24,
+
 	/* this goes last */
 	IORING_REGISTER_LAST
 };
@@ -558,4 +564,9 @@ struct io_uring_getevents_arg {
 	__u64	ts;
 };
 
+struct io_uring_fixed_worker_arg {
+	__s32	nr_workers;
+	__s32	max_works;
+};
+
 #endif
io_uring/io-wq.c +101 -0

107 lines not on this page · raw message

io_uring/io-wq.h +3 -0
index 78efbb8c53f0..fbbe13d75595 100644
@@ -3,6 +3,7 @@
 
 #include <linux/refcount.h>
 #include <linux/io_uring_types.h>
+#include <uapi/linux/io_uring.h>
 
 struct io_wq;
 
@@ -188,6 +189,8 @@ void io_wq_hash_work(struct io_wq_work *work, void *val);
 
 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask);
 int io_wq_max_workers(struct io_wq *wq, int *new_count);
+int io_wq_fixed_workers(struct io_wq *wq,
+			struct io_uring_fixed_worker_arg *new_count);
 
 static inline bool io_wq_is_hashed(struct io_wq_work *work)
 {
io_uring/io_uring.c +71 -0

83 lines not on this page · raw message

Signature · 1 line
2.25.1
author
#10 lore raw

PATCH v5 10/11 io-wq: add an work list for fixed worker

In reply to Hao Xu

From: Hao Xu <[email protected]>

Previously when a fixed worker handles its private works, it get all of them from worker->acct.work_list to a temporary acct->work_list. This prevents work cancellation since the cancellation process cannot find works from this temporary work list. Thus add a new acct so to address it.

---
 io_uring/io-wq.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
io_uring/io-wq.c +13 -7
index d54056b98e2b..38d88da70a40 100644
@@ -126,6 +126,7 @@ struct io_worker {
 	};
 	int index;
 	struct io_wqe_acct acct;
+	struct io_wqe_acct exec_acct;
 };
 
 #if BITS_PER_LONG == 64
@@ -723,14 +724,17 @@ static void io_worker_handle_work(struct io_worker *worker,
 
 static inline void io_worker_handle_private_work(struct io_worker *worker)
 {
-	struct io_wqe_acct acct;
+	struct io_wqe_acct *acct = &worker->acct;
+	struct io_wqe_acct *exec_acct = &worker->exec_acct;
 
-	raw_spin_lock(&worker->acct.lock);
-	acct = worker->acct;
-	wq_list_clean(&worker->acct.work_list);
-	worker->acct.nr_works = 0;
-	raw_spin_unlock(&worker->acct.lock);
-	io_worker_handle_work(worker, &acct, false);
+	raw_spin_lock(&acct->lock);
+	exec_acct->nr_works = acct->nr_works;
+	exec_acct->max_works = acct->max_works;
+	exec_acct->work_list = acct->work_list;
+	wq_list_clean(&acct->work_list);
+	acct->nr_works = 0;
+	raw_spin_unlock(&acct->lock);
+	io_worker_handle_work(worker, exec_acct, false);
 }
 
 static inline void io_worker_handle_public_work(struct io_worker *worker)
@@ -866,6 +870,7 @@ static void io_init_new_fixed_worker(struct io_wqe *wqe,
 {
 	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 	struct io_wqe_acct *iw_acct = &worker->acct;
+	struct io_wqe_acct *exec_acct = &worker->exec_acct;
 	unsigned index = acct->index;
 	unsigned *nr_fixed;
 
@@ -878,6 +883,7 @@ static void io_init_new_fixed_worker(struct io_wqe *wqe,
 	iw_acct->index = index;
 	INIT_WQ_LIST(&iw_acct->work_list);
 	raw_spin_lock_init(&iw_acct->lock);
+	raw_spin_lock_init(&exec_acct->lock);
 	raw_spin_unlock(&acct->lock);
 }
 
Signature · 1 line
2.25.1
author
#11 lore raw

PATCH v5 11/11 io_uring: cancel works in exec work list for fixed worker

In reply to Hao Xu

From: Hao Xu <[email protected]>

When users want to cancel a request, look into the exec work list of fixed worker as well. It's not sane to ignore it.

---
 io_uring/io-wq.c | 44 +++++++++++++++++++++++++++++++++++---------
 1 file changed, 35 insertions(+), 9 deletions(-)
io_uring/io-wq.c +35 -9
index 38d88da70a40..9e8ad7455373 100644
@@ -1285,32 +1285,44 @@ static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 	return false;
 }
 
-static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
-				       struct io_cb_cancel_data *match)
+static void io_wqe_cancel_pending_work_normal(struct io_wqe *wqe,
+					      struct io_cb_cancel_data *match)
 {
-	int i, j;
-	struct io_wqe_acct *acct, *iw_acct;
+	int i;
+	struct io_wqe_acct *acct;
 
-retry_public:
+retry:
 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
 		acct = io_get_acct(wqe, i == 0, false);
 		if (io_acct_cancel_pending_work(wqe, acct, match)) {
 			if (match->cancel_all)
-				goto retry_public;
+				goto retry;
 			return;
 		}
 	}
+}
+
+static void __io_wqe_cancel_pending_work_fixed(struct io_wqe *wqe,
+					       struct io_cb_cancel_data *match,
+					       bool exec)
+{
+	int i, j;
+	struct io_wqe_acct *acct, *iw_acct;
 
-retry_private:
+retry:
 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
 		acct = io_get_acct(wqe, i == 0, true);
 		raw_spin_lock(&acct->lock);
 		for (j = 0; j < acct->nr_fixed; j++) {
-			iw_acct = &acct->fixed_workers[j]->acct;
+			if (exec)
+				iw_acct = &acct->fixed_workers[j]->acct;
+			else
+				iw_acct = &acct->fixed_workers[j]->exec_acct;
+
 			if (io_acct_cancel_pending_work(wqe, iw_acct, match)) {
 				if (match->cancel_all) {
 					raw_spin_unlock(&acct->lock);
-					goto retry_private;
+					goto retry;
 				}
 				break;
 			}
@@ -1319,6 +1331,20 @@ static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 	}
 }
 
+static void io_wqe_cancel_pending_work_fixed(struct io_wqe *wqe,
+					     struct io_cb_cancel_data *match)
+{
+	__io_wqe_cancel_pending_work_fixed(wqe, match, false);
+	__io_wqe_cancel_pending_work_fixed(wqe, match, true);
+}
+
+static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
+				       struct io_cb_cancel_data *match)
+{
+	io_wqe_cancel_pending_work_normal(wqe, match);
+	io_wqe_cancel_pending_work_fixed(wqe, match);
+}
+
 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
 				       struct io_cb_cancel_data *match)
 {
Signature · 1 line
2.25.1
author
#12 lore raw
In reply to Hao Xu
Quoted 42 lines · Hao Xu

From: Hao Xu <[email protected]>

This is the second version of fixed worker implementation. Wrote a nop test program to test it, 3 fixed-workers VS 3 normal workers. normal workers:
./run_nop_wqe.sh nop_wqe_normal 200000 100 3 1-3

         time spent: 17314274 usecs      IOPS: 1155116
         time spent: 17016942 usecs      IOPS: 1175299
         time spent: 17908684 usecs      IOPS: 1116776

fixed workers:
./run_nop_wqe.sh nop_wqe_fixed 200000 100 3 1-3

         time spent: 10464397 usecs      IOPS: 1911242
         time spent: 9610976 usecs       IOPS: 2080954
         time spent: 9807361 usecs       IOPS: 2039284

About 2x improvement. From perf result, almost no acct->lock contension. Test program: https://github.com/HowHsu/liburing/tree/fixed_worker
liburing/test/nop_wqe.c

v3->v4:
- make work in fixed worker's private worfixed worker
- tweak the io_wqe_acct struct to make it clearer

v4->v5:
- code clean
- rebase against the for-5.20/io_uring

Hao Xu (11):
io-wq: add a worker flag for individual exit
io-wq: change argument of create_io_worker() for convienence
io-wq: add infra data structure for fixed workers
io-wq: tweak io_get_acct()
io-wq: fixed worker initialization
io-wq: fixed worker exit
io-wq: implement fixed worker logic
io-wq: batch the handling of fixed worker private works
io_uring: add register fixed worker interface
io-wq: add an work list for fixed worker
io_uring: cancel works in exec work list for fixed worker

Quoted 8 lines
  include/uapi/linux/io_uring.h |  11 +
  io_uring/io-wq.c              | 495 ++++++++++++++++++++++++++++++----
  io_uring/io-wq.h              |   8 +
  io_uring/io_uring.c           |  71 +++++
  4 files changed, 537 insertions(+), 48 deletions(-)

base-commit: 094abe8fbccb0d79bef982c67eb7372e92452c0e

ping...