Skip to contentSkip to conversation

fstests: fix io_uring testing

PATCH Superseded Zorro Lang 4 participants 15 messages Mar 6, 2024 io_uring

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Zorro Lang

author
#0 lore raw

cover fstests: fix io_uring testing

According to the manual of io_uring_queue_init, it doesn't set errno but return the -errno on failure. So we should check the return value of io_uring_queue_init, to make sure if the io_uring is supported by kernel. We've left this problem in xfstests/ltp/fsstress.c long time. (Refer to PATCH 1/3)

And besides kernel build without CONFIG_IO_URING, a system can disable the io_uring supporting manually, by set sysctl kernel.io_uring_disabled=2. The former cause io_uring_queue_init return ENOSYS, but the latter will return EPERM. So I let fsstress to deal with both situations. (Refer to PATCH 2/3)

A question is if we should do "sysctl -w kernel.io_uring_disabled=0 &> /dev/null" at the beginning of each test case (e.g. do that in common/config ?), or leave this decision to the testers (in their test wrapper). Now I only do that in _require_io_uring(). If anyone has any opinions, feel free to reply. (Refer to PATCH 3/3)

Thanks,
Zorro

author
#1 lore raw

PATCH 1/3 fsstress: check io_uring_queue_init errno properly

As the manual of io_uring_queue_init says "io_uring_queue_init(3) returns 0 on success and -errno on failure". We should check if the return value is -ENOSYS, not the errno.

Fixes d15b1721f284 ("ltp/fsstress: don't fail on io_uring ENOSYS")
---
 ltp/fsstress.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
ltp/fsstress.c +2 -2
index 63c75767..482395c4 100644
@@ -763,8 +763,8 @@ int main(int argc, char **argv)
 #ifdef URING
 			have_io_uring = true;
 			/* If ENOSYS, just ignore uring, other errors are fatal. */
-			if (io_uring_queue_init(URING_ENTRIES, &ring, 0)) {
-				if (errno == ENOSYS) {
+			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
+				if (c == -ENOSYS) {
 					have_io_uring = false;
 				} else {
 					fprintf(stderr, "io_uring_queue_init failed\n");
Signature · 1 line
2.43.0
author
#5 lore raw

PATCH 2/3 fsstress: bypass io_uring testing if io_uring_queue_init returns EPERM

In reply to Zorro Lang

I found the io_uring testing still fails as:
io_uring_queue_init failed
even if kernel supports io_uring feature.

That because of the /proc/sys/kernel/io_uring_disabled isn't 0.

Different value means:
0 All processes can create io_uring instances as normal. 1 io_uring creation is disabled (io_uring_setup() will fail with
-EPERM) for unprivileged processes not in the io_uring_group group. Existing io_uring instances can still be used. See the documentation for io_uring_group for more information.
2 io_uring creation is disabled for all processes. io_uring_setup()
always fails with -EPERM. Existing io_uring instances can still be used.

So besides the CONFIG_IO_URING kernel config, there's another switch can on or off the io_uring supporting. And the "2" or "1" might be the default on some systems.

On this situation the io_uring_queue_init returns -EPERM, so I change the fsstress to ignore io_uring testing if io_uring_queue_init returns -ENOSYS or -EPERM. And print different verbose message for debug.

---
 ltp/fsstress.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
ltp/fsstress.c +13 -2
index 482395c4..9c75f27b 100644
@@ -762,12 +762,23 @@ int main(int argc, char **argv)
 #endif
 #ifdef URING
 			have_io_uring = true;
-			/* If ENOSYS, just ignore uring, other errors are fatal. */
+			/*
+			 * If ENOSYS, just ignore uring, due to kernel doesn't support it.
+			 * If EPERM, might due to sysctl kernel.io_uring_disabled isn't 0,
+			 *           or some selinux policies, etc.
+			 * Other errors are fatal.
+			 */
 			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
 				if (c == -ENOSYS) {
 					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't supported by kernel\n");
+				} else if (c == -EPERM) {
+					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't allowed, check io_uring_disabled sysctl or selinux policy\n");
 				} else {
-					fprintf(stderr, "io_uring_queue_init failed\n");
+					fprintf(stderr, "io_uring_queue_init failed, errno=%d\n", c);
 					exit(1);
 				}
 			}
Signature · 1 line
2.43.0
author
#10 lore raw

PATCH 3/3 common/rc: force enable io_uring in _require_io_uring

In reply to Zorro Lang

If kernel supports io_uring, userspace still can/might disable that supporting by set /proc/sys/kernel/io_uring_disabled=2. Let's set it to 0, to always enable io_uring (ignore error if there's not that file).

---
 common/rc | 3 +++
 1 file changed, 3 insertions(+)
common/rc +3 -0
index 50dde313..966c92e3 100644
@@ -2317,6 +2317,9 @@ _require_aiodio()
 # this test requires that the kernel supports IO_URING
 _require_io_uring()
 {
+	# Force enable io_uring if kernel supports it
+	sysctl -w kernel.io_uring_disabled=0 &> /dev/null
+
 	$here/src/feature -R
 	case $? in
 	0)
Signature · 1 line
2.43.0
#11 lore raw

Re: [PATCH 3/3] common/rc: force enable io_uring in _require_io_uring

Quoted 15 lines · Zorro Lang

If kernel supports io_uring, userspace still can/might disable that supporting by set /proc/sys/kernel/io_uring_disabled=2. Let's set it to 0, to always enable io_uring (ignore error if there's not that file).

---
 common/rc | 3 +++
 1 file changed, 3 insertions(+)
common/rc +0 -0
index 50dde313..966c92e3 100644

common/rc _require_aiodio() · line 2321

Quoted 5 lines
common/rc +2 -0
@@ -2317,6 +2317,9 @@ _require_aiodio()
 # this test requires that the kernel supports IO_URING
 _require_io_uring()
 {
+	# Force enable io_uring if kernel supports it
+	sysctl -w kernel.io_uring_disabled=0 &> /dev/null

_require_XXX functions are supposed to be predicates that _notrun the test if XXX isn't configured or available. Shouldn't this be:

	local io_uring_knob="$(sysctl --values kernel.io_uring_disabled)"
	test "$io_uring_knob" -ne 0 && _notrun "io_uring disabled by admin"

Alternately -- if it _is_ ok to turn this knob, then there should be a cleanup method to put it back after the test.

--D

Quoted 8 lines
common/rc +1 -0
+
 	$here/src/feature -R
 	case $? in
 	0)
Signature · 1 line
2.43.0
Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] fsstress: check io_uring_queue_init errno properly

In reply to Zorro Lang
Quoted 3 lines · Zorro Lang

As the manual of io_uring_queue_init says "io_uring_queue_init(3) returns 0 on success and -errno on failure". We should check if the return value is -ENOSYS, not the errno.

/me checks liburing source code and sees that the library returns a negative error code without touching errno (the semi global error code variable) at all. That's an unfortunate quirk of the manpage, but this code here is correct...

Quoted 11 lines
Fixes d15b1721f284 ("ltp/fsstress: don't fail on io_uring ENOSYS")
---
 ltp/fsstress.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
ltp/fsstress.c +0 -0
index 63c75767..482395c4 100644

ltp/fsstress.c int main(int argc, char **argv) · line 770

Quoted 10 lines
ltp/fsstress.c +2 -2
@@ -763,8 +763,8 @@ int main(int argc, char **argv)
 #ifdef URING
 			have_io_uring = true;
 			/* If ENOSYS, just ignore uring, other errors are fatal. */
-			if (io_uring_queue_init(URING_ENTRIES, &ring, 0)) {
-				if (errno == ENOSYS) {
+			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
+				if (c == -ENOSYS) {
 					have_io_uring = false;
 				} else {
 					fprintf(stderr, "io_uring_queue_init failed\n");

But why not:

			c = io_uring_queue_init(...);
			switch (c) {
			case 0:
				have_io_uring = true;
				break;
			case -ENOSYS:
				have_io_uring = false;
				break;
			default:
				fprintf(stderr, "io_uring_queue_init failed\n");
				break;
			}

Especially since you add another case in the next patch?

I'll leave the style nits up to you though:

--D

--D

Quoted 4 lines
Signature · 1 line
2.43.0
Reviewed-by on v1 2/3
#6 lore raw

Re: [PATCH 2/3] fsstress: bypass io_uring testing if io_uring_queue_init returns EPERM

In reply to Zorro Lang
Quoted 34 lines · Zorro Lang

I found the io_uring testing still fails as:
io_uring_queue_init failed
even if kernel supports io_uring feature.

That because of the /proc/sys/kernel/io_uring_disabled isn't 0.

Different value means:
0 All processes can create io_uring instances as normal. 1 io_uring creation is disabled (io_uring_setup() will fail with
-EPERM) for unprivileged processes not in the io_uring_group group. Existing io_uring instances can still be used. See the documentation for io_uring_group for more information.
2 io_uring creation is disabled for all processes. io_uring_setup()
always fails with -EPERM. Existing io_uring instances can still be used.

So besides the CONFIG_IO_URING kernel config, there's another switch can on or off the io_uring supporting. And the "2" or "1" might be the default on some systems.

On this situation the io_uring_queue_init returns -EPERM, so I change the fsstress to ignore io_uring testing if io_uring_queue_init returns -ENOSYS or -EPERM. And print different verbose message for debug.

---
 ltp/fsstress.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
ltp/fsstress.c +0 -0
index 482395c4..9c75f27b 100644

ltp/fsstress.c int main(int argc, char **argv) · line 767

Quoted 7 lines
ltp/fsstress.c +3 -1
@@ -762,12 +762,23 @@ int main(int argc, char **argv)
 #endif
 #ifdef URING
 			have_io_uring = true;
-			/* If ENOSYS, just ignore uring, other errors are fatal. */
+			/*
+			 * If ENOSYS, just ignore uring, due to kernel doesn't support it.
+			 * If EPERM, might due to sysctl kernel.io_uring_disabled isn't 0,

"might be due to..."

With that fixed,

--D

Quoted 22 lines
ltp/fsstress.c +10 -1
+			 *           or some selinux policies, etc.
+			 * Other errors are fatal.
+			 */
 			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
 				if (c == -ENOSYS) {
 					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't supported by kernel\n");
+				} else if (c == -EPERM) {
+					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't allowed, check io_uring_disabled sysctl or selinux policy\n");
 				} else {
-					fprintf(stderr, "io_uring_queue_init failed\n");
+					fprintf(stderr, "io_uring_queue_init failed, errno=%d\n", c);
 					exit(1);
 				}
 			}
Signature · 1 line
2.43.0
Reviewed-by on v1 1/3
#4 lore raw

Re: [PATCH 1/3] fsstress: check io_uring_queue_init errno properly

In reply to Zorro Lang
Quoted 15 lines · Zorro Lang

As the manual of io_uring_queue_init says "io_uring_queue_init(3) returns 0 on success and -errno on failure". We should check if the return value is -ENOSYS, not the errno.

Fixes d15b1721f284 ("ltp/fsstress: don't fail on io_uring ENOSYS")
---
 ltp/fsstress.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
ltp/fsstress.c +0 -0
index 63c75767..482395c4 100644

ltp/fsstress.c int main(int argc, char **argv) · line 770

Quoted 10 lines
ltp/fsstress.c +2 -2
@@ -763,8 +763,8 @@ int main(int argc, char **argv)
 #ifdef URING
 			have_io_uring = true;
 			/* If ENOSYS, just ignore uring, other errors are fatal. */
-			if (io_uring_queue_init(URING_ENTRIES, &ring, 0)) {
-				if (errno == ENOSYS) {
+			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
+				if (c == -ENOSYS) {
 					have_io_uring = false;
 				} else {
 					fprintf(stderr, "io_uring_queue_init failed\n");
Reviewed-by on v1 2/3
#8 lore raw

Re: [PATCH 2/3] fsstress: bypass io_uring testing if io_uring_queue_init returns EPERM

In reply to Zorro Lang
Quoted 46 lines · Zorro Lang

I found the io_uring testing still fails as:
io_uring_queue_init failed
even if kernel supports io_uring feature.

That because of the /proc/sys/kernel/io_uring_disabled isn't 0.

Different value means:
0 All processes can create io_uring instances as normal. 1 io_uring creation is disabled (io_uring_setup() will fail with
-EPERM) for unprivileged processes not in the io_uring_group group. Existing io_uring instances can still be used. See the documentation for io_uring_group for more information.
2 io_uring creation is disabled for all processes. io_uring_setup()
always fails with -EPERM. Existing io_uring instances can still be used.

So besides the CONFIG_IO_URING kernel config, there's another switch can on or off the io_uring supporting. And the "2" or "1" might be the default on some systems.

On this situation the io_uring_queue_init returns -EPERM, so I change the fsstress to ignore io_uring testing if io_uring_queue_init returns -ENOSYS or -EPERM. And print different verbose message for debug.

---
 ltp/fsstress.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
ltp/fsstress.c +6 -1
index 482395c4..9c75f27b 100644
@@ -762,12 +762,23 @@ int main(int argc, char **argv)
 #endif
 #ifdef URING
 			have_io_uring = true;
-			/* If ENOSYS, just ignore uring, other errors are fatal. */
+			/*
+			 * If ENOSYS, just ignore uring, due to kernel doesn't support it.
+			 * If EPERM, might due to sysctl kernel.io_uring_disabled isn't 0,
+			 *           or some selinux policies, etc.
+			 * Other errors are fatal.
+			 */
 			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
 				if (c == -ENOSYS) {

ltp/fsstress.c int main(int argc, char **argv) · line 781

Quoted 10 lines
ltp/fsstress.c +7 -1
 					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't supported by kernel\n");
+				} else if (c == -EPERM) {
+					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't allowed, check io_uring_disabled sysctl or selinux policy\n");
 				} else {
-					fprintf(stderr, "io_uring_queue_init failed\n");
+					fprintf(stderr, "io_uring_queue_init failed, errno=%d\n", c);

I think you want to use -c here, right?

Other than that:

#12 lore raw

Re: [PATCH 3/3] common/rc: force enable io_uring in _require_io_uring

In reply to Darrick J. Wong
Quoted 25 lines · Darrick J. Wong
Quoted 15 lines · Zorro Lang

If kernel supports io_uring, userspace still can/might disable that supporting by set /proc/sys/kernel/io_uring_disabled=2. Let's set it to 0, to always enable io_uring (ignore error if there's not that file).

---
 common/rc | 3 +++
 1 file changed, 3 insertions(+)
common/rc +0 -0
index 50dde313..966c92e3 100644

common/rc _require_aiodio() · line 2321

Quoted 5 lines
common/rc +2 -0
@@ -2317,6 +2317,9 @@ _require_aiodio()
 # this test requires that the kernel supports IO_URING
 _require_io_uring()
 {
+	# Force enable io_uring if kernel supports it
+	sysctl -w kernel.io_uring_disabled=0 &> /dev/null

_require_XXX functions are supposed to be predicates that _notrun the test if XXX isn't configured or available. Shouldn't this be:

Quoted 2 lines
	local io_uring_knob="$(sysctl --values kernel.io_uring_disabled)"
	test "$io_uring_knob" -ne 0 && _notrun "io_uring disabled by admin"

That sounds like a good option to me.

Quoted 2 lines

Alternately -- if it _is_ ok to turn this knob, then there should be a cleanup method to put it back after the test.

I think it would be better not to change the setting, especially if the admin had disabled it.

Cheers,
Jeff

Quoted 11 lines

--D

Quoted 8 lines
common/rc +1 -0
+
 	$here/src/feature -R
 	case $? in
 	0)
Signature · 1 line
2.43.0
#3 lore raw

Re: [PATCH 1/3] fsstress: check io_uring_queue_init errno properly

In reply to Darrick J. Wong
Quoted 9 lines · Darrick J. Wong
Quoted 3 lines · Zorro Lang

As the manual of io_uring_queue_init says "io_uring_queue_init(3) returns 0 on success and -errno on failure". We should check if the return value is -ENOSYS, not the errno.

/me checks liburing source code and sees that the library returns a negative error code without touching errno (the semi global error code variable) at all. That's an unfortunate quirk of the manpage, but this code here is correct...

Yeah, that confuse me too, especially some io_uring functions set errno, some return -errno ...

Quoted 38 lines
Quoted 12 lines
Fixes d15b1721f284 ("ltp/fsstress: don't fail on io_uring ENOSYS")
---
 ltp/fsstress.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
ltp/fsstress.c +0 -0
index 63c75767..482395c4 100644

ltp/fsstress.c int main(int argc, char **argv) · line 770

Quoted 10 lines
ltp/fsstress.c +2 -2
@@ -763,8 +763,8 @@ int main(int argc, char **argv)
 #ifdef URING
 			have_io_uring = true;
 			/* If ENOSYS, just ignore uring, other errors are fatal. */
-			if (io_uring_queue_init(URING_ENTRIES, &ring, 0)) {
-				if (errno == ENOSYS) {
+			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
+				if (c == -ENOSYS) {
 					have_io_uring = false;
 				} else {
 					fprintf(stderr, "io_uring_queue_init failed\n");

But why not:

			c = io_uring_queue_init(...);
			switch (c) {
			case 0:
				have_io_uring = true;
				break;
			case -ENOSYS:
				have_io_uring = false;
				break;
			default:
				fprintf(stderr, "io_uring_queue_init failed\n");
				break;
			}
Quoted 1 line

Especially since you add another case in the next patch?

Sure, that looks more clearly, I'll change this part. Thanks!

Quoted 14 lines

I'll leave the style nits up to you though:

--D

--D

Quoted 4 lines
Signature · 1 line
2.43.0
#7 lore raw

Re: [PATCH 2/3] fsstress: bypass io_uring testing if io_uring_queue_init returns EPERM

In reply to Darrick J. Wong
Quoted 30 lines · Darrick J. Wong
Quoted 29 lines · Zorro Lang

I found the io_uring testing still fails as:
io_uring_queue_init failed
even if kernel supports io_uring feature.

That because of the /proc/sys/kernel/io_uring_disabled isn't 0.

Different value means:
0 All processes can create io_uring instances as normal. 1 io_uring creation is disabled (io_uring_setup() will fail with
-EPERM) for unprivileged processes not in the io_uring_group group. Existing io_uring instances can still be used. See the documentation for io_uring_group for more information.
2 io_uring creation is disabled for all processes. io_uring_setup()
always fails with -EPERM. Existing io_uring instances can still be used.

So besides the CONFIG_IO_URING kernel config, there's another switch can on or off the io_uring supporting. And the "2" or "1" might be the default on some systems.

On this situation the io_uring_queue_init returns -EPERM, so I change the fsstress to ignore io_uring testing if io_uring_queue_init returns -ENOSYS or -EPERM. And print different verbose message for debug.

---
 ltp/fsstress.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
Quoted 14 lines

ltp/fsstress.c int main(int argc, char **argv) · line 767

Quoted 12 lines
ltp/fsstress.c +3 -1
index 482395c4..9c75f27b 100644
@@ -762,12 +762,23 @@ int main(int argc, char **argv)
 #endif
 #ifdef URING
 			have_io_uring = true;
-			/* If ENOSYS, just ignore uring, other errors are fatal. */
+			/*
+			 * If ENOSYS, just ignore uring, due to kernel doesn't support it.
+			 * If EPERM, might due to sysctl kernel.io_uring_disabled isn't 0,

"might be due to..."

Hahaha, as native english speaker so :)

Quoted 3 lines

With that fixed,

Thanks, I'll change below if...else... to switch... as you suggested.

Quoted 26 lines

--D

Quoted 22 lines
ltp/fsstress.c +10 -1
+			 *           or some selinux policies, etc.
+			 * Other errors are fatal.
+			 */
 			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
 				if (c == -ENOSYS) {
 					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't supported by kernel\n");
+				} else if (c == -EPERM) {
+					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't allowed, check io_uring_disabled sysctl or selinux policy\n");
 				} else {
-					fprintf(stderr, "io_uring_queue_init failed\n");
+					fprintf(stderr, "io_uring_queue_init failed, errno=%d\n", c);
 					exit(1);
 				}
 			}
Signature · 1 line
2.43.0
#9 lore raw

Re: [PATCH 2/3] fsstress: bypass io_uring testing if io_uring_queue_init returns EPERM

In reply to Jeff Moyer
Quoted 52 lines · Jeff Moyer
Quoted 50 lines · Zorro Lang

I found the io_uring testing still fails as:
io_uring_queue_init failed
even if kernel supports io_uring feature.

That because of the /proc/sys/kernel/io_uring_disabled isn't 0.

Different value means:
0 All processes can create io_uring instances as normal. 1 io_uring creation is disabled (io_uring_setup() will fail with
-EPERM) for unprivileged processes not in the io_uring_group group. Existing io_uring instances can still be used. See the documentation for io_uring_group for more information.
2 io_uring creation is disabled for all processes. io_uring_setup()
always fails with -EPERM. Existing io_uring instances can still be used.

So besides the CONFIG_IO_URING kernel config, there's another switch can on or off the io_uring supporting. And the "2" or "1" might be the default on some systems.

On this situation the io_uring_queue_init returns -EPERM, so I change the fsstress to ignore io_uring testing if io_uring_queue_init returns -ENOSYS or -EPERM. And print different verbose message for debug.

---
 ltp/fsstress.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
ltp/fsstress.c +9 -1
index 482395c4..9c75f27b 100644
@@ -762,12 +762,23 @@ int main(int argc, char **argv)
 #endif
 #ifdef URING
 			have_io_uring = true;
-			/* If ENOSYS, just ignore uring, other errors are fatal. */
+			/*
+			 * If ENOSYS, just ignore uring, due to kernel doesn't support it.
+			 * If EPERM, might due to sysctl kernel.io_uring_disabled isn't 0,
+			 *           or some selinux policies, etc.
+			 * Other errors are fatal.
+			 */
 			if ((c = io_uring_queue_init(URING_ENTRIES, &ring, 0)) != 0) {
 				if (c == -ENOSYS) {
 					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't supported by kernel\n");
+				} else if (c == -EPERM) {
Quoted 8 lines

ltp/fsstress.c int main(int argc, char **argv) · line 781

Quoted 6 lines
ltp/fsstress.c +4 -1
+					have_io_uring = false;
+					if (verbose)
+						printf("io_uring isn't allowed, check io_uring_disabled sysctl or selinux policy\n");
 				} else {
-					fprintf(stderr, "io_uring_queue_init failed\n");
+					fprintf(stderr, "io_uring_queue_init failed, errno=%d\n", c);

I think you want to use -c here, right?

Sure, will change that, thanks!

Quoted 6 lines

Other than that:

#14 lore raw

Re: [PATCH 3/3] common/rc: force enable io_uring in _require_io_uring

In reply to Darrick J. Wong
Quoted 11 lines · Darrick J. Wong
Quoted 10 lines · Zorro Lang

If kernel supports io_uring, userspace still can/might disable that supporting by set /proc/sys/kernel/io_uring_disabled=2. Let's set it to 0, to always enable io_uring (ignore error if there's not that file).

---
 common/rc | 3 +++
 1 file changed, 3 insertions(+)
Quoted 13 lines

common/rc _require_aiodio() · line 2321

Quoted 10 lines
common/rc +2 -0
index 50dde313..966c92e3 100644
@@ -2317,6 +2317,9 @@ _require_aiodio()
 # this test requires that the kernel supports IO_URING
 _require_io_uring()
 {
+	# Force enable io_uring if kernel supports it
+	sysctl -w kernel.io_uring_disabled=0 &> /dev/null

_require_XXX functions are supposed to be predicates that _notrun the test if XXX isn't configured or available. Shouldn't this be:

Yeah, that makes sense, I tried to find a place to force enable io_uring, forgot it's not proper for _require_XXX function.

Quoted 3 lines
	local io_uring_knob="$(sysctl --values kernel.io_uring_disabled)"
	test "$io_uring_knob" -ne 0 && _notrun "io_uring disabled by admin"

And I think I need to change the src/feature.c:check_uring_support() too, check -EPERM as I did in fsstress.c.

Quoted 3 lines

Alternately -- if it _is_ ok to turn this knob, then there should be a cleanup method to put it back after the test.

Yeah, I'm still wondering if we should let fstests touch/set the io_uring_disabled sysctl. Or we leave this job to the fstests user/wrapper script, fstests just _notrun if io_uring is off.

Thanks,
Zorro

Quoted 12 lines

--D

Quoted 8 lines
common/rc +1 -0
+
 	$here/src/feature -R
 	case $? in
 	0)
Signature · 1 line
2.43.0
#13 lore raw

Re: [PATCH 3/3] common/rc: force enable io_uring in _require_io_uring

In reply to Jeff Moyer
Quoted 30 lines · Jeff Moyer
Quoted 25 lines · Darrick J. Wong
Quoted 15 lines · Zorro Lang

If kernel supports io_uring, userspace still can/might disable that supporting by set /proc/sys/kernel/io_uring_disabled=2. Let's set it to 0, to always enable io_uring (ignore error if there's not that file).

---
 common/rc | 3 +++
 1 file changed, 3 insertions(+)
common/rc +0 -0
index 50dde313..966c92e3 100644

common/rc _require_aiodio() · line 2321

Quoted 5 lines
common/rc +2 -0
@@ -2317,6 +2317,9 @@ _require_aiodio()
 # this test requires that the kernel supports IO_URING
 _require_io_uring()
 {
+	# Force enable io_uring if kernel supports it
+	sysctl -w kernel.io_uring_disabled=0 &> /dev/null

_require_XXX functions are supposed to be predicates that _notrun the test if XXX isn't configured or available. Shouldn't this be:

Quoted 2 lines
	local io_uring_knob="$(sysctl --values kernel.io_uring_disabled)"
	test "$io_uring_knob" -ne 0 && _notrun "io_uring disabled by admin"
Quoted 7 lines

That sounds like a good option to me.

Quoted 2 lines

Alternately -- if it _is_ ok to turn this knob, then there should be a cleanup method to put it back after the test.

I think it would be better not to change the setting, especially if the admin had disabled it.

As a testing environment/system, even if fstests doesn't touch the io_uring_disabled sysctl, the testers should care about that before doing his test. So the fstests users might need to change the io_uring_disabled=0 manually, or do it in their CI test scripts. Maybe we should leave this job to the users.

Thanks,
Zorro

Quoted 17 lines

Cheers,
Jeff

Quoted 11 lines

--D

Quoted 8 lines
common/rc +1 -0
+
 	$here/src/feature -R
 	case $? in
 	0)
Signature · 1 line
2.43.0