[PATCH 0/3] userfaultfd: selftest: Improve behavior with older kernels

[PATCH 0/3] userfaultfd: selftest: Improve behavior with older kernels

[Thiago Jung Bauermann]

Hello,

A tester ran the upstream selftest on a distro kernel and sounded the alarm when
it reported failures for features which aren't included in that kernel.

This patch set improves the test behavior in that scenario.

Thiago Jung Bauermann (3):
  userfaultfd: selftest: Fix checking of userfaultfd_open() result
  userfaultfd: selftest: Skip test if a feature isn't supported
  userfaultfd: selftest: Report XFAIL if shmem doesn't support zeropage

 tools/testing/selftests/vm/userfaultfd.c | 49 ++++++++++++++++++++++++--------
 1 file changed, 37 insertions(+), 12 deletions(-)

[PATCH 1/3] userfaultfd: selftest: Fix checking of userfaultfd_open() result

[Thiago Jung Bauermann]

If the userfaultfd test is run on a kernel with CONFIG_USERFAULTFD=n, it
will report that the system call is not available yet go ahead and continue
anyway:

  # ./userfaultfd anon 30 1
  nr_pages: 480, nr_pages_per_cpu: 120
  userfaultfd syscall not available in this kernel
  bounces: 0, mode:, register failure

This is because userfaultfd_open() returns 0 on success and 1 on error but
all callers assume that it returns < 0 on error.

Since the convention of the test as a whole is the one used by
userfault_open(), fix its callers instead. Now the test behaves correctly:

  # ./userfaultfd anon 30 1
  nr_pages: 480, nr_pages_per_cpu: 120
  userfaultfd syscall not available in this kernel

Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
---
 tools/testing/selftests/vm/userfaultfd.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index 7b8171e3128a..e4099afe7557 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c
@@ -859,7 +859,7 @@ static int userfaultfd_zeropage_test(void)
 	if (uffd_test_ops->release_pages(area_dst))
 		return 1;
 
-	if (userfaultfd_open(0) < 0)
+	if (userfaultfd_open(0))
 		return 1;
 	uffdio_register.range.start = (unsigned long) area_dst;
 	uffdio_register.range.len = nr_pages * page_size;
@@ -902,7 +902,7 @@ static int userfaultfd_events_test(void)
 
 	features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
 		UFFD_FEATURE_EVENT_REMOVE;
-	if (userfaultfd_open(features) < 0)
+	if (userfaultfd_open(features))
 		return 1;
 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 
@@ -961,7 +961,7 @@ static int userfaultfd_sig_test(void)
 		return 1;
 
 	features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
-	if (userfaultfd_open(features) < 0)
+	if (userfaultfd_open(features))
 		return 1;
 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 
@@ -1027,7 +1027,7 @@ static int userfaultfd_stress(void)
 	if (!area_dst)
 		return 1;
 
-	if (userfaultfd_open(0) < 0)
+	if (userfaultfd_open(0))
 		return 1;
 
 	count_verify = malloc(nr_pages * sizeof(unsigned long long));

Re: [PATCH 1/3] userfaultfd: selftest: Fix checking of userfaultfd_open() result

[Mike Rapoport]

Hi,

On Tue, Jul 24, 2018 at 11:42:07PM -0300, Thiago Jung Bauermann wrote:
> If the userfaultfd test is run on a kernel with CONFIG_USERFAULTFD=n, it
> will report that the system call is not available yet go ahead and continue
> anyway:
> 
>   # ./userfaultfd anon 30 1
>   nr_pages: 480, nr_pages_per_cpu: 120
>   userfaultfd syscall not available in this kernel
>   bounces: 0, mode:, register failure
> 
> This is because userfaultfd_open() returns 0 on success and 1 on error but
> all callers assume that it returns < 0 on error.
> 
> Since the convention of the test as a whole is the one used by
> userfault_open(), fix its callers instead. Now the test behaves correctly:
> 
>   # ./userfaultfd anon 30 1
>   nr_pages: 480, nr_pages_per_cpu: 120
>   userfaultfd syscall not available in this kernel
> 
> Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>

It seems that this patch is superseded by the second patch in this series.

> ---
>  tools/testing/selftests/vm/userfaultfd.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
> index 7b8171e3128a..e4099afe7557 100644
> --- a/tools/testing/selftests/vm/userfaultfd.c
> +++ b/tools/testing/selftests/vm/userfaultfd.c
> @@ -859,7 +859,7 @@ static int userfaultfd_zeropage_test(void)
>  	if (uffd_test_ops->release_pages(area_dst))
>  		return 1;
> 
> -	if (userfaultfd_open(0) < 0)
> +	if (userfaultfd_open(0))
>  		return 1;
>  	uffdio_register.range.start = (unsigned long) area_dst;
>  	uffdio_register.range.len = nr_pages * page_size;
> @@ -902,7 +902,7 @@ static int userfaultfd_events_test(void)
> 
>  	features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
>  		UFFD_FEATURE_EVENT_REMOVE;
> -	if (userfaultfd_open(features) < 0)
> +	if (userfaultfd_open(features))
>  		return 1;
>  	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
> 
> @@ -961,7 +961,7 @@ static int userfaultfd_sig_test(void)
>  		return 1;
> 
>  	features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
> -	if (userfaultfd_open(features) < 0)
> +	if (userfaultfd_open(features))
>  		return 1;
>  	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
> 
> @@ -1027,7 +1027,7 @@ static int userfaultfd_stress(void)
>  	if (!area_dst)
>  		return 1;
> 
> -	if (userfaultfd_open(0) < 0)
> +	if (userfaultfd_open(0))
>  		return 1;
> 
>  	count_verify = malloc(nr_pages * sizeof(unsigned long long));

-- 
Sincerely yours,
Mike.

Re: [PATCH 1/3] userfaultfd: selftest: Fix checking of userfaultfd_open() result

[Thiago Jung Bauermann]

Hello Mike,

Thanks for promptly reviewing the patches.

Mike Rapoport <rppt@linux.vnet.ibm.com> writes:

> Hi,
>
> On Tue, Jul 24, 2018 at 11:42:07PM -0300, Thiago Jung Bauermann wrote:
>> If the userfaultfd test is run on a kernel with CONFIG_USERFAULTFD=n, it
>> will report that the system call is not available yet go ahead and continue
>> anyway:
>>
>>   # ./userfaultfd anon 30 1
>>   nr_pages: 480, nr_pages_per_cpu: 120
>>   userfaultfd syscall not available in this kernel
>>   bounces: 0, mode:, register failure
>>
>> This is because userfaultfd_open() returns 0 on success and 1 on error but
>> all callers assume that it returns < 0 on error.
>>
>> Since the convention of the test as a whole is the one used by
>> userfault_open(), fix its callers instead. Now the test behaves correctly:
>>
>>   # ./userfaultfd anon 30 1
>>   nr_pages: 480, nr_pages_per_cpu: 120
>>   userfaultfd syscall not available in this kernel
>>
>> Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
>
> It seems that this patch is superseded by the second patch in this series.

Yes, but since this is a simple bugfix while the other patch is a
proposed improvement which can be debated, I think it's worthwhile to
keep them separate.

--
Thiago Jung Bauermann
IBM Linux Technology Center

[PATCH 2/3] userfaultfd: selftest: Skip test if a feature isn't supported

[Thiago Jung Bauermann]

If userfaultfd runs on a system that doesn't support some feature it is
trying to test, it currently ends with error code 1 which indicates
test failure:

  # ./userfaultfd anon 10 10
  nr_pages: 160, nr_pages_per_cpu: 80
  bounces: 9, mode: rnd poll, userfaults: 7 59
  bounces: 8, mode: poll, userfaults: 0 0
  bounces: 7, mode: rnd racing ver, userfaults: 45 2
  bounces: 6, mode: racing ver, userfaults: 3 1
  bounces: 5, mode: rnd ver, userfaults: 55 32
  bounces: 4, mode: ver, userfaults: 69 0
  bounces: 3, mode: rnd racing, userfaults: 1 1
  bounces: 2, mode: racing, userfaults: 65 0
  bounces: 1, mode: rnd, userfaults: 44 1
  bounces: 0, mode:, userfaults: 3 2
  testing UFFDIO_ZEROPAGE: done.
  testing signal delivery: UFFDIO_API
  # echo $?
  1

Make the testcase return KSFT_SKIP instead, which is more accurate since it
is not a real test failure:

  # ./userfaultfd anon 10 10
  nr_pages: 160, nr_pages_per_cpu: 80
  bounces: 9, mode: rnd poll, userfaults: 3 0
  bounces: 8, mode: poll, userfaults: 6 1
  bounces: 7, mode: rnd racing ver, userfaults: 3 1
  bounces: 6, mode: racing ver, userfaults: 2 1
  bounces: 5, mode: rnd ver, userfaults: 2 1
  bounces: 4, mode: ver, userfaults: 3 47
  bounces: 3, mode: rnd racing, userfaults: 38 0
  bounces: 2, mode: racing, userfaults: 4 61
  bounces: 1, mode: rnd, userfaults: 16 6
  bounces: 0, mode:, userfaults: 55 25
  testing UFFDIO_ZEROPAGE: done.
  testing signal delivery: UFFDIO_API: Invalid argument
  # echo $?
  4

While at it, also improve the error message of the ioctl(UFFDIO_API) call.

Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
---
 tools/testing/selftests/vm/userfaultfd.c | 41 ++++++++++++++++++++++----------
 1 file changed, 29 insertions(+), 12 deletions(-)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index e4099afe7557..bc9ec38fbc34 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c
@@ -645,8 +645,11 @@ static int userfaultfd_open(int features)
 	uffdio_api.api = UFFD_API;
 	uffdio_api.features = features;
 	if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
-		fprintf(stderr, "UFFDIO_API\n");
-		return 1;
+		int errnum = errno;
+
+		perror("UFFDIO_API");
+
+		return errnum == EINVAL ? KSFT_SKIP : 1;
 	}
 	if (uffdio_api.api != UFFD_API) {
 		fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
@@ -852,6 +855,7 @@ static int userfaultfd_zeropage_test(void)
 {
 	struct uffdio_register uffdio_register;
 	unsigned long expected_ioctls;
+	int err;
 
 	printf("testing UFFDIO_ZEROPAGE: ");
 	fflush(stdout);
@@ -859,8 +863,10 @@ static int userfaultfd_zeropage_test(void)
 	if (uffd_test_ops->release_pages(area_dst))
 		return 1;
 
-	if (userfaultfd_open(0))
-		return 1;
+	err = userfaultfd_open(0);
+	if (err)
+		return err;
+
 	uffdio_register.range.start = (unsigned long) area_dst;
 	uffdio_register.range.len = nr_pages * page_size;
 	uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
@@ -902,8 +908,9 @@ static int userfaultfd_events_test(void)
 
 	features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
 		UFFD_FEATURE_EVENT_REMOVE;
-	if (userfaultfd_open(features))
-		return 1;
+	err = userfaultfd_open(features);
+	if (err)
+		return err;
 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 
 	uffdio_register.range.start = (unsigned long) area_dst;
@@ -961,8 +968,9 @@ static int userfaultfd_sig_test(void)
 		return 1;
 
 	features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
-	if (userfaultfd_open(features))
-		return 1;
+	err = userfaultfd_open(features);
+	if (err)
+		return err;
 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 
 	uffdio_register.range.start = (unsigned long) area_dst;
@@ -1027,8 +1035,9 @@ static int userfaultfd_stress(void)
 	if (!area_dst)
 		return 1;
 
-	if (userfaultfd_open(0))
-		return 1;
+	err = userfaultfd_open(0);
+	if (err)
+		return err;
 
 	count_verify = malloc(nr_pages * sizeof(unsigned long long));
 	if (!count_verify) {
@@ -1199,8 +1208,16 @@ static int userfaultfd_stress(void)
 		return err;
 
 	close(uffd);
-	return userfaultfd_zeropage_test() || userfaultfd_sig_test()
-		|| userfaultfd_events_test();
+
+	err = userfaultfd_zeropage_test();
+	if (err)
+		return err;
+
+	err = userfaultfd_sig_test();
+	if (err)
+		return err;
+
+	return userfaultfd_events_test();
 }
 
 /*

Re: [PATCH 2/3] userfaultfd: selftest: Skip test if a feature isn't supported

[Mike Rapoport]

Hi,

On Tue, Jul 24, 2018 at 11:42:08PM -0300, Thiago Jung Bauermann wrote:
> If userfaultfd runs on a system that doesn't support some feature it is
> trying to test, it currently ends with error code 1 which indicates
> test failure:
> 
>   # ./userfaultfd anon 10 10
>   nr_pages: 160, nr_pages_per_cpu: 80
>   bounces: 9, mode: rnd poll, userfaults: 7 59
>   bounces: 8, mode: poll, userfaults: 0 0
>   bounces: 7, mode: rnd racing ver, userfaults: 45 2
>   bounces: 6, mode: racing ver, userfaults: 3 1
>   bounces: 5, mode: rnd ver, userfaults: 55 32
>   bounces: 4, mode: ver, userfaults: 69 0
>   bounces: 3, mode: rnd racing, userfaults: 1 1
>   bounces: 2, mode: racing, userfaults: 65 0
>   bounces: 1, mode: rnd, userfaults: 44 1
>   bounces: 0, mode:, userfaults: 3 2
>   testing UFFDIO_ZEROPAGE: done.
>   testing signal delivery: UFFDIO_API
>   # echo $?
>   1
> 
> Make the testcase return KSFT_SKIP instead, which is more accurate since it
> is not a real test failure:
> 
>   # ./userfaultfd anon 10 10
>   nr_pages: 160, nr_pages_per_cpu: 80
>   bounces: 9, mode: rnd poll, userfaults: 3 0
>   bounces: 8, mode: poll, userfaults: 6 1
>   bounces: 7, mode: rnd racing ver, userfaults: 3 1
>   bounces: 6, mode: racing ver, userfaults: 2 1
>   bounces: 5, mode: rnd ver, userfaults: 2 1
>   bounces: 4, mode: ver, userfaults: 3 47
>   bounces: 3, mode: rnd racing, userfaults: 38 0
>   bounces: 2, mode: racing, userfaults: 4 61
>   bounces: 1, mode: rnd, userfaults: 16 6
>   bounces: 0, mode:, userfaults: 55 25
>   testing UFFDIO_ZEROPAGE: done.
>   testing signal delivery: UFFDIO_API: Invalid argument
>   # echo $?
>   4
> 
> While at it, also improve the error message of the ioctl(UFFDIO_API) call.

Can you please also make the test return KSFT_SKIP if userfaultfd(2) fails
with ENOSYS?
 
> Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
> ---
>  tools/testing/selftests/vm/userfaultfd.c | 41 ++++++++++++++++++++++----------
>  1 file changed, 29 insertions(+), 12 deletions(-)
> 
> diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
> index e4099afe7557..bc9ec38fbc34 100644
> --- a/tools/testing/selftests/vm/userfaultfd.c
> +++ b/tools/testing/selftests/vm/userfaultfd.c
> @@ -645,8 +645,11 @@ static int userfaultfd_open(int features)
>  	uffdio_api.api = UFFD_API;
>  	uffdio_api.features = features;
>  	if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
> -		fprintf(stderr, "UFFDIO_API\n");
> -		return 1;
> +		int errnum = errno;
> +
> +		perror("UFFDIO_API");
> +
> +		return errnum == EINVAL ? KSFT_SKIP : 1;
>  	}
>  	if (uffdio_api.api != UFFD_API) {
>  		fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
> @@ -852,6 +855,7 @@ static int userfaultfd_zeropage_test(void)
>  {
>  	struct uffdio_register uffdio_register;
>  	unsigned long expected_ioctls;
> +	int err;
> 
>  	printf("testing UFFDIO_ZEROPAGE: ");
>  	fflush(stdout);
> @@ -859,8 +863,10 @@ static int userfaultfd_zeropage_test(void)
>  	if (uffd_test_ops->release_pages(area_dst))
>  		return 1;
> 
> -	if (userfaultfd_open(0))
> -		return 1;
> +	err = userfaultfd_open(0);
> +	if (err)
> +		return err;
> +
>  	uffdio_register.range.start = (unsigned long) area_dst;
>  	uffdio_register.range.len = nr_pages * page_size;
>  	uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
> @@ -902,8 +908,9 @@ static int userfaultfd_events_test(void)
> 
>  	features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
>  		UFFD_FEATURE_EVENT_REMOVE;
> -	if (userfaultfd_open(features))
> -		return 1;
> +	err = userfaultfd_open(features);
> +	if (err)
> +		return err;
>  	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
> 
>  	uffdio_register.range.start = (unsigned long) area_dst;
> @@ -961,8 +968,9 @@ static int userfaultfd_sig_test(void)
>  		return 1;
> 
>  	features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
> -	if (userfaultfd_open(features))
> -		return 1;
> +	err = userfaultfd_open(features);
> +	if (err)
> +		return err;
>  	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
> 
>  	uffdio_register.range.start = (unsigned long) area_dst;
> @@ -1027,8 +1035,9 @@ static int userfaultfd_stress(void)
>  	if (!area_dst)
>  		return 1;
> 
> -	if (userfaultfd_open(0))
> -		return 1;
> +	err = userfaultfd_open(0);
> +	if (err)
> +		return err;
> 
>  	count_verify = malloc(nr_pages * sizeof(unsigned long long));
>  	if (!count_verify) {
> @@ -1199,8 +1208,16 @@ static int userfaultfd_stress(void)
>  		return err;
> 
>  	close(uffd);
> -	return userfaultfd_zeropage_test() || userfaultfd_sig_test()
> -		|| userfaultfd_events_test();
> +
> +	err = userfaultfd_zeropage_test();
> +	if (err)
> +		return err;
> +
> +	err = userfaultfd_sig_test();
> +	if (err)
> +		return err;
> +
> +	return userfaultfd_events_test();
>  }
> 
>  /*

-- 
Sincerely yours,
Mike.

Re: [PATCH 2/3] userfaultfd: selftest: Skip test if a feature isn't supported

[Thiago Jung Bauermann]

Mike Rapoport <rppt@linux.vnet.ibm.com> writes:

> Hi,
>
> On Tue, Jul 24, 2018 at 11:42:08PM -0300, Thiago Jung Bauermann wrote:
>> If userfaultfd runs on a system that doesn't support some feature it is
>> trying to test, it currently ends with error code 1 which indicates
>> test failure:
>> 
>>   # ./userfaultfd anon 10 10
>>   nr_pages: 160, nr_pages_per_cpu: 80
>>   bounces: 9, mode: rnd poll, userfaults: 7 59
>>   bounces: 8, mode: poll, userfaults: 0 0
>>   bounces: 7, mode: rnd racing ver, userfaults: 45 2
>>   bounces: 6, mode: racing ver, userfaults: 3 1
>>   bounces: 5, mode: rnd ver, userfaults: 55 32
>>   bounces: 4, mode: ver, userfaults: 69 0
>>   bounces: 3, mode: rnd racing, userfaults: 1 1
>>   bounces: 2, mode: racing, userfaults: 65 0
>>   bounces: 1, mode: rnd, userfaults: 44 1
>>   bounces: 0, mode:, userfaults: 3 2
>>   testing UFFDIO_ZEROPAGE: done.
>>   testing signal delivery: UFFDIO_API
>>   # echo $?
>>   1
>> 
>> Make the testcase return KSFT_SKIP instead, which is more accurate since it
>> is not a real test failure:
>> 
>>   # ./userfaultfd anon 10 10
>>   nr_pages: 160, nr_pages_per_cpu: 80
>>   bounces: 9, mode: rnd poll, userfaults: 3 0
>>   bounces: 8, mode: poll, userfaults: 6 1
>>   bounces: 7, mode: rnd racing ver, userfaults: 3 1
>>   bounces: 6, mode: racing ver, userfaults: 2 1
>>   bounces: 5, mode: rnd ver, userfaults: 2 1
>>   bounces: 4, mode: ver, userfaults: 3 47
>>   bounces: 3, mode: rnd racing, userfaults: 38 0
>>   bounces: 2, mode: racing, userfaults: 4 61
>>   bounces: 1, mode: rnd, userfaults: 16 6
>>   bounces: 0, mode:, userfaults: 55 25
>>   testing UFFDIO_ZEROPAGE: done.
>>   testing signal delivery: UFFDIO_API: Invalid argument
>>   # echo $?
>>   4
>> 
>> While at it, also improve the error message of the ioctl(UFFDIO_API) call.
>
> Can you please also make the test return KSFT_SKIP if userfaultfd(2) fails
> with ENOSYS?

Sure. v2 will have a separate patch implementing your suggestion.

-- 
Thiago Jung Bauermann
IBM Linux Technology Center

[PATCH 3/3] userfaultfd: selftest: Report XFAIL if shmem doesn't support zeropage

[Thiago Jung Bauermann]

If userfaultfd runs on a system that doesn't support UFFDIO_ZEROPAGE for
shared memory, it currently ends with error code 1 which indicates test
failure:

  # ./userfaultfd shmem 10 10
  nr_pages: 160, nr_pages_per_cpu: 80
  bounces: 9, mode: rnd poll, unexpected missing ioctl for anon memory
  # echo $?
  1

This is a real failure, but expected so signal that to the test harness:

  # ./userfaultfd shmem 10 10
  nr_pages: 160, nr_pages_per_cpu: 80
  bounces: 9, mode: rnd poll, UFFDIO_ZEROPAGE unsupported in shmem VMAs
  # echo $?
  2

Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
---
 tools/testing/selftests/vm/userfaultfd.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index bc9ec38fbc34..686fe96f617f 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c
@@ -1115,6 +1115,14 @@ static int userfaultfd_stress(void)
 		expected_ioctls = uffd_test_ops->expected_ioctls;
 		if ((uffdio_register.ioctls & expected_ioctls) !=
 		    expected_ioctls) {
+			if (test_type == TEST_SHMEM &&
+			    (uffdio_register.ioctls & expected_ioctls) ==
+			    UFFD_API_RANGE_IOCTLS_BASIC) {
+				fprintf(stderr,
+					"UFFDIO_ZEROPAGE unsupported in shmem VMAs\n");
+				return KSFT_XFAIL;
+			}
+
 			fprintf(stderr,
 				"unexpected missing ioctl for anon memory\n");
 			return 1;

Re: [PATCH 3/3] userfaultfd: selftest: Report XFAIL if shmem doesn't support zeropage

[Mike Rapoport]

Hi,

On Tue, Jul 24, 2018 at 11:42:09PM -0300, Thiago Jung Bauermann wrote:
> If userfaultfd runs on a system that doesn't support UFFDIO_ZEROPAGE for
> shared memory, it currently ends with error code 1 which indicates test
> failure:
> 
>   # ./userfaultfd shmem 10 10
>   nr_pages: 160, nr_pages_per_cpu: 80
>   bounces: 9, mode: rnd poll, unexpected missing ioctl for anon memory
>   # echo $?
>   1
> 
> This is a real failure, but expected so signal that to the test harness:

I don't think its a real failure. If the kernel does not support
UFFDIO_ZEROPAGE for shared memory the userfaultfd_zeropage_test can be
simply skipped.
 
>   # ./userfaultfd shmem 10 10
>   nr_pages: 160, nr_pages_per_cpu: 80
>   bounces: 9, mode: rnd poll, UFFDIO_ZEROPAGE unsupported in shmem VMAs
>   # echo $?
>   2
> 
> Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
> ---
>  tools/testing/selftests/vm/userfaultfd.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
> index bc9ec38fbc34..686fe96f617f 100644
> --- a/tools/testing/selftests/vm/userfaultfd.c
> +++ b/tools/testing/selftests/vm/userfaultfd.c
> @@ -1115,6 +1115,14 @@ static int userfaultfd_stress(void)
>  		expected_ioctls = uffd_test_ops->expected_ioctls;
>  		if ((uffdio_register.ioctls & expected_ioctls) !=
>  		    expected_ioctls) {
> +			if (test_type == TEST_SHMEM &&
> +			    (uffdio_register.ioctls & expected_ioctls) ==
> +			    UFFD_API_RANGE_IOCTLS_BASIC) {
> +				fprintf(stderr,
> +					"UFFDIO_ZEROPAGE unsupported in shmem VMAs\n");
> +				return KSFT_XFAIL;
> +			}
> +

By all means, this check should be moved to userfaultfd_zeropage_test().
Ideally, we should call here ksft_test_result_skip() and simply return from
the function.



>  			fprintf(stderr,
>  				"unexpected missing ioctl for anon memory\n");
>  			return 1;

-- 
Sincerely yours,
Mike.

Re: [PATCH 3/3] userfaultfd: selftest: Report XFAIL if shmem doesn't support zeropage

[Thiago Jung Bauermann]

Mike Rapoport <rppt@linux.vnet.ibm.com> writes:

> Hi,
>
> On Tue, Jul 24, 2018 at 11:42:09PM -0300, Thiago Jung Bauermann wrote:
>> If userfaultfd runs on a system that doesn't support UFFDIO_ZEROPAGE for
>> shared memory, it currently ends with error code 1 which indicates test
>> failure:
>>
>>   # ./userfaultfd shmem 10 10
>>   nr_pages: 160, nr_pages_per_cpu: 80
>>   bounces: 9, mode: rnd poll, unexpected missing ioctl for anon memory
>>   # echo $?
>>   1
>>
>> This is a real failure, but expected so signal that to the test harness:
>
> I don't think its a real failure. If the kernel does not support
> UFFDIO_ZEROPAGE for shared memory the userfaultfd_zeropage_test can be
> simply skipped.

Ok, good point. I'll make that change in v2.

>>   # ./userfaultfd shmem 10 10
>>   nr_pages: 160, nr_pages_per_cpu: 80
>>   bounces: 9, mode: rnd poll, UFFDIO_ZEROPAGE unsupported in shmem VMAs
>>   # echo $?
>>   2
>>
>> Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
>> ---
>>  tools/testing/selftests/vm/userfaultfd.c | 8 ++++++++
>>  1 file changed, 8 insertions(+)
>>
>> diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
>> index bc9ec38fbc34..686fe96f617f 100644
>> --- a/tools/testing/selftests/vm/userfaultfd.c
>> +++ b/tools/testing/selftests/vm/userfaultfd.c
>> @@ -1115,6 +1115,14 @@ static int userfaultfd_stress(void)
>>  		expected_ioctls = uffd_test_ops->expected_ioctls;
>>  		if ((uffdio_register.ioctls & expected_ioctls) !=
>>  		    expected_ioctls) {
>> +			if (test_type == TEST_SHMEM &&
>> +			    (uffdio_register.ioctls & expected_ioctls) ==
>> +			    UFFD_API_RANGE_IOCTLS_BASIC) {
>> +				fprintf(stderr,
>> +					"UFFDIO_ZEROPAGE unsupported in shmem VMAs\n");
>> +				return KSFT_XFAIL;
>> +			}
>> +
>
> By all means, this check should be moved to userfaultfd_zeropage_test().

I made that change in v2.

> Ideally, we should call here ksft_test_result_skip() and simply return from
> the function.

In my understanding, calling ksft_test_result_skip() would require
converting the testcase to use the functions that generate TAP output.

Also, returning here isn't actually necessary: from my testing
userfaultfd_stress() doesn't require zeropage support in shmem so if the
only bit missing from uffdio_register.ioctls is the one for
UFFDIO_ZEROPAGE then this error can simply be ignored and the test can
continue. Do you agree?

>>  			fprintf(stderr,
>>  				"unexpected missing ioctl for anon memory\n");
>>  			return 1;


--
Thiago Jung Bauermann
IBM Linux Technology Center