hi,
You can use this patch to test IORING_SETUP_SQPOLL_PERCPU flag introduced in
patch " io_uring: support multiple rings to share same poll thread by specifying same cpu".
Below is a example fio job:
[global]
ioengine=io_uring
sqthread_poll=1
sqthread_poll_percpu=1
registerfiles=1
fixedbufs=1
hipri=0
thread=1
bs=4k
direct=1
rw=randread
time_based=1
runtime=60
randrepeat=0
filename=/dev/vdb
[job1]
cpus_allowed=2
sqthread_poll_cpu=17
iodepth=128
[job2]
cpus_allowed=3
sqthread_poll_cpu=17
iodepth=128
[job3]
cpus_allowed=4
sqthread_poll_cpu=17
iodepth=128
[job4]
cpus_allowed=5
sqthread_poll_cpu=17
iodepth=128
Regards,
Xiaoguang Wang
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.
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@xxxxxxxxxxxxxxxxx>
---
engines/io_uring.c | 13 +++++++++++++
fio.1 | 6 ++++++
os/linux/io_uring.h | 1 +
3 files changed, 20 insertions(+)
diff --git a/engines/io_uring.c b/engines/io_uring.c
index ec8cb18a..8a447b62 100644
--- a/engines/io_uring.c
+++ b/engines/io_uring.c
@@ -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;
diff --git a/fio.1 b/fio.1
index 1c90e4a5..d367134f 100644
--- a/fio.1
+++ b/fio.1
@@ -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
diff --git a/os/linux/io_uring.h b/os/linux/io_uring.h
index d39b45fd..0a4c8a35 100644
--- a/os/linux/io_uring.h
+++ b/os/linux/io_uring.h
@@ -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,
