Skip to contentSkip to conversation

engines/io_uring: add sqthread_poll_percpu option

PATCH New Xiaoguang Wang 1 participant 1 message Nov 3, 2020 io_uring

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 engines/io_uring: add sqthread_poll_percpu option

This option is only meaningful when sqthread_poll and sqthread_poll_cpu are both set. If this option is effective, for multiple io_uring instances which are all bound to one same cpu, only a kernel thread is created for this cpu to perform these io_uring instances' submission queue polling.

---
 engines/io_uring.c  | 13 +++++++++++++
 fio.1               |  6 ++++++
 os/linux/io_uring.h |  1 +
 3 files changed, 20 insertions(+)
engines/io_uring.c +13 -0
index ec8cb18a..8a447b62 100644
@@ -77,6 +77,7 @@ struct ioring_options {
 	unsigned int fixedbufs;
 	unsigned int registerfiles;
 	unsigned int sqpoll_thread;
+	unsigned int sqpoll_thread_percpu;
 	unsigned int sqpoll_set;
 	unsigned int sqpoll_cpu;
 	unsigned int nonvectored;
@@ -160,6 +161,16 @@ static struct fio_option options[] = {
 		.category = FIO_OPT_C_ENGINE,
 		.group	= FIO_OPT_G_IOURING,
 	},
+	{
+		.name	= "sqthread_poll_percpu",
+		.lname	= "Kernel percpu SQ thread polling",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct ioring_options, sqpoll_thread_percpu),
+		.help	= "Offload submission/completion to kernel thread, use percpu thread",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_IOURING,
+	},
+
 	{
 		.name	= "sqthread_poll_cpu",
 		.lname	= "SQ Thread Poll CPU",
@@ -596,6 +607,8 @@ static int fio_ioring_queue_init(struct thread_data *td)
 		p.flags |= IORING_SETUP_IOPOLL;
 	if (o->sqpoll_thread) {
 		p.flags |= IORING_SETUP_SQPOLL;
+		if (o->sqpoll_thread_percpu)
+			p.flags |= IORING_SETUP_SQPOLL_PERCPU;
 		if (o->sqpoll_set) {
 			p.flags |= IORING_SETUP_SQ_AFF;
 			p.sq_thread_cpu = o->sqpoll_cpu;
fio.1 +6 -0
index 1c90e4a5..d367134f 100644
@@ -1851,6 +1851,12 @@ the cost of using more CPU in the system.
 When `sqthread_poll` is set, this option provides a way to define which CPU
 should be used for the polling thread.
 .TP
+.BI (io_uring)sqthread_poll_percpu
+This option is only meaningful when `sqthread_poll` and `sqthread_poll_cpu` are
+both set. If this option is effective, for multiple io_uring instances which are all
+bound to one same cpu, only a kernel thread is created for this cpu to perform these
+io_uring instances' submission queue polling.
+.TP
 .BI (libaio)userspace_reap
 Normally, with the libaio engine in use, fio will use the
 \fBio_getevents\fR\|(3) system call to reap newly returned events. With
os/linux/io_uring.h +1 -0
index d39b45fd..0a4c8a35 100644
@@ -99,6 +99,7 @@ enum {
 #define IORING_SETUP_CQSIZE	(1U << 3)	/* app defines CQ size */
 #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_SQPOLL_PERCPU (1U << 7)       /* sq_thread_cpu is valid */
 
 enum {
 	IORING_OP_NOP,
Signature · 1 line
2.17.2