2020-09-17 04:24:21

by Michael Ellerman

[permalink] [raw]
Subject: [PATCH] selftests/harness: Flush stdout before forking

The test harness forks() a child to run each test. Both the parent and
the child print to stdout using libc functions. That can lead to
duplicated (or more) output if the libc buffers are not flushed before
forking.

It's generally not seen when running programs directly, because stdout
will usually be line buffered when it's pointing to a terminal.

This was noticed when running the seccomp_bpf test, eg:

$ ./seccomp_bpf | tee test.log
$ grep -c "TAP version 13" test.log
2

But we only expect the TAP header to appear once.

It can be exacerbated using stdbuf to increase the buffer size:

$ stdbuf -o 1MB ./seccomp_bpf > test.log
$ grep -c "TAP version 13" test.log
13

The fix is simple, we just flush stdout & stderr before fork. Usually
stderr is unbuffered, but that can be changed, so flush it as well
just to be safe.

Signed-off-by: Michael Ellerman <[email protected]>
---
tools/testing/selftests/kselftest_harness.h | 5 +++++
1 file changed, 5 insertions(+)

diff --git a/tools/testing/selftests/kselftest_harness.h b/tools/testing/selftests/kselftest_harness.h
index 4f78e4805633..f19804df244c 100644
--- a/tools/testing/selftests/kselftest_harness.h
+++ b/tools/testing/selftests/kselftest_harness.h
@@ -971,6 +971,11 @@ void __run_test(struct __fixture_metadata *f,

ksft_print_msg(" RUN %s%s%s.%s ...\n",
f->name, variant->name[0] ? "." : "", variant->name, t->name);
+
+ /* Make sure output buffers are flushed before fork */
+ fflush(stdout);
+ fflush(stderr);
+
t->pid = fork();
if (t->pid < 0) {
ksft_print_msg("ERROR SPAWNING TEST CHILD\n");
--
2.25.1


2020-09-17 04:57:39

by Max Filippov

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman <[email protected]> wrote:
>
> The test harness forks() a child to run each test. Both the parent and
> the child print to stdout using libc functions. That can lead to
> duplicated (or more) output if the libc buffers are not flushed before
> forking.
>
> It's generally not seen when running programs directly, because stdout
> will usually be line buffered when it's pointing to a terminal.
>
> This was noticed when running the seccomp_bpf test, eg:
>
> $ ./seccomp_bpf | tee test.log
> $ grep -c "TAP version 13" test.log
> 2
>
> But we only expect the TAP header to appear once.
>
> It can be exacerbated using stdbuf to increase the buffer size:
>
> $ stdbuf -o 1MB ./seccomp_bpf > test.log
> $ grep -c "TAP version 13" test.log
> 13
>
> The fix is simple, we just flush stdout & stderr before fork. Usually
> stderr is unbuffered, but that can be changed, so flush it as well
> just to be safe.
>
> Signed-off-by: Michael Ellerman <[email protected]>
> ---
> tools/testing/selftests/kselftest_harness.h | 5 +++++
> 1 file changed, 5 insertions(+)

Tested-by: Max Filippov <[email protected]>

--
Thanks.
-- Max

2020-09-17 13:00:47

by Shuah Khan

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On 9/16/20 10:53 PM, Max Filippov wrote:
> On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman <[email protected]> wrote:
>>
>> The test harness forks() a child to run each test. Both the parent and
>> the child print to stdout using libc functions. That can lead to
>> duplicated (or more) output if the libc buffers are not flushed before
>> forking.
>>
>> It's generally not seen when running programs directly, because stdout
>> will usually be line buffered when it's pointing to a terminal.
>>
>> This was noticed when running the seccomp_bpf test, eg:
>>
>> $ ./seccomp_bpf | tee test.log
>> $ grep -c "TAP version 13" test.log
>> 2
>>
>> But we only expect the TAP header to appear once.
>>
>> It can be exacerbated using stdbuf to increase the buffer size:
>>
>> $ stdbuf -o 1MB ./seccomp_bpf > test.log
>> $ grep -c "TAP version 13" test.log
>> 13
>>
>> The fix is simple, we just flush stdout & stderr before fork. Usually
>> stderr is unbuffered, but that can be changed, so flush it as well
>> just to be safe.
>>
>> Signed-off-by: Michael Ellerman <[email protected]>
>> ---
>> tools/testing/selftests/kselftest_harness.h | 5 +++++
>> 1 file changed, 5 insertions(+)
>
> Tested-by: Max Filippov <[email protected]>
>

Thank you both. Applying to linux-kselftest fixes for 5.9-rc7

thanks,
-- Shuah

2020-09-17 18:47:37

by Shuah Khan

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On 9/17/20 6:58 AM, Shuah Khan wrote:
> On 9/16/20 10:53 PM, Max Filippov wrote:
>> On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman <[email protected]>
>> wrote:
>>>
>>> The test harness forks() a child to run each test. Both the parent and
>>> the child print to stdout using libc functions. That can lead to
>>> duplicated (or more) output if the libc buffers are not flushed before
>>> forking.
>>>
>>> It's generally not seen when running programs directly, because stdout
>>> will usually be line buffered when it's pointing to a terminal.
>>>
>>> This was noticed when running the seccomp_bpf test, eg:
>>>
>>>    $ ./seccomp_bpf | tee test.log
>>>    $ grep -c "TAP version 13" test.log
>>>    2
>>>
>>> But we only expect the TAP header to appear once.
>>>
>>> It can be exacerbated using stdbuf to increase the buffer size:
>>>
>>>    $ stdbuf -o 1MB ./seccomp_bpf > test.log
>>>    $ grep -c "TAP version 13" test.log
>>>    13
>>>
>>> The fix is simple, we just flush stdout & stderr before fork. Usually
>>> stderr is unbuffered, but that can be changed, so flush it as well
>>> just to be safe.
>>>
>>> Signed-off-by: Michael Ellerman <[email protected]>
>>> ---
>>>   tools/testing/selftests/kselftest_harness.h | 5 +++++
>>>   1 file changed, 5 insertions(+)
>>
>> Tested-by: Max Filippov <[email protected]>
>>
>
> Thank you both. Applying to linux-kselftest fixes for 5.9-rc7
>

Kees,

I haven't pulled this in yet. If you want to take this through seccomp
tree for dependencies:

Acked-by: Shuah Khan <[email protected]>

thanks,
-- Shuah

2020-09-17 22:07:59

by Kees Cook

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On Thu, Sep 17, 2020 at 10:51:32AM -0600, Shuah Khan wrote:
> On 9/17/20 6:58 AM, Shuah Khan wrote:
> > On 9/16/20 10:53 PM, Max Filippov wrote:
> > > On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman
> > > <[email protected]> wrote:
> > > >
> > > > The test harness forks() a child to run each test. Both the parent and
> > > > the child print to stdout using libc functions. That can lead to
> > > > duplicated (or more) output if the libc buffers are not flushed before
> > > > forking.
> > > >
> > > > It's generally not seen when running programs directly, because stdout
> > > > will usually be line buffered when it's pointing to a terminal.
> > > >
> > > > This was noticed when running the seccomp_bpf test, eg:
> > > >
> > > > ?? $ ./seccomp_bpf | tee test.log
> > > > ?? $ grep -c "TAP version 13" test.log
> > > > ?? 2
> > > >
> > > > But we only expect the TAP header to appear once.
> > > >
> > > > It can be exacerbated using stdbuf to increase the buffer size:
> > > >
> > > > ?? $ stdbuf -o 1MB ./seccomp_bpf > test.log
> > > > ?? $ grep -c "TAP version 13" test.log
> > > > ?? 13
> > > >
> > > > The fix is simple, we just flush stdout & stderr before fork. Usually
> > > > stderr is unbuffered, but that can be changed, so flush it as well
> > > > just to be safe.
> > > >
> > > > Signed-off-by: Michael Ellerman <[email protected]>
> > > > ---
> > > > ? tools/testing/selftests/kselftest_harness.h | 5 +++++
> > > > ? 1 file changed, 5 insertions(+)
> > >
> > > Tested-by: Max Filippov <[email protected]>
> > >
> >
> > Thank you both. Applying to linux-kselftest fixes for 5.9-rc7
> >
>
> Kees,
>
> I haven't pulled this in yet. If you want to take this through seccomp
> tree for dependencies:
>
> Acked-by: Shuah Khan <[email protected]>

You can have it -- it's a global fix and more than the seccomp selftest
is likely affected. :)

--
Kees Cook

2020-09-17 22:09:03

by Kees Cook

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On Thu, Sep 17, 2020 at 02:15:19PM +1000, Michael Ellerman wrote:
> The test harness forks() a child to run each test. Both the parent and
> the child print to stdout using libc functions. That can lead to
> duplicated (or more) output if the libc buffers are not flushed before
> forking.
>
> It's generally not seen when running programs directly, because stdout
> will usually be line buffered when it's pointing to a terminal.
>
> This was noticed when running the seccomp_bpf test, eg:
>
> $ ./seccomp_bpf | tee test.log
> $ grep -c "TAP version 13" test.log
> 2

Oh thank you for tracking this down! I thought seccomp wasn't killing a
child or something, and hadn't found it yet. :)

Acked-by: Kees Cook <[email protected]>

--
Kees Cook

2020-09-18 03:39:10

by Michael Ellerman

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

Shuah Khan <[email protected]> writes:
> On 9/16/20 10:53 PM, Max Filippov wrote:
>> On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman <[email protected]> wrote:
>>>
>>> The test harness forks() a child to run each test. Both the parent and
>>> the child print to stdout using libc functions. That can lead to
>>> duplicated (or more) output if the libc buffers are not flushed before
>>> forking.
>>>
>>> It's generally not seen when running programs directly, because stdout
>>> will usually be line buffered when it's pointing to a terminal.
>>>
>>> This was noticed when running the seccomp_bpf test, eg:
>>>
>>> $ ./seccomp_bpf | tee test.log
>>> $ grep -c "TAP version 13" test.log
>>> 2
>>>
>>> But we only expect the TAP header to appear once.
>>>
>>> It can be exacerbated using stdbuf to increase the buffer size:
>>>
>>> $ stdbuf -o 1MB ./seccomp_bpf > test.log
>>> $ grep -c "TAP version 13" test.log
>>> 13
>>>
>>> The fix is simple, we just flush stdout & stderr before fork. Usually
>>> stderr is unbuffered, but that can be changed, so flush it as well
>>> just to be safe.
>>>
>>> Signed-off-by: Michael Ellerman <[email protected]>
>>> ---
>>> tools/testing/selftests/kselftest_harness.h | 5 +++++
>>> 1 file changed, 5 insertions(+)
>>
>> Tested-by: Max Filippov <[email protected]>
>
> Thank you both. Applying to linux-kselftest fixes for 5.9-rc7

It can wait for v5.10 IMHO, but up to you.

cheers

2020-09-18 14:35:45

by Shuah Khan

[permalink] [raw]
Subject: Re: [PATCH] selftests/harness: Flush stdout before forking

On 9/17/20 9:37 PM, Michael Ellerman wrote:
> Shuah Khan <[email protected]> writes:
>> On 9/16/20 10:53 PM, Max Filippov wrote:
>>> On Wed, Sep 16, 2020 at 9:16 PM Michael Ellerman <[email protected]> wrote:
>>>>
>>>> The test harness forks() a child to run each test. Both the parent and
>>>> the child print to stdout using libc functions. That can lead to
>>>> duplicated (or more) output if the libc buffers are not flushed before
>>>> forking.
>>>>
>>>> It's generally not seen when running programs directly, because stdout
>>>> will usually be line buffered when it's pointing to a terminal.
>>>>
>>>> This was noticed when running the seccomp_bpf test, eg:
>>>>
>>>> $ ./seccomp_bpf | tee test.log
>>>> $ grep -c "TAP version 13" test.log
>>>> 2
>>>>
>>>> But we only expect the TAP header to appear once.
>>>>
>>>> It can be exacerbated using stdbuf to increase the buffer size:
>>>>
>>>> $ stdbuf -o 1MB ./seccomp_bpf > test.log
>>>> $ grep -c "TAP version 13" test.log
>>>> 13
>>>>
>>>> The fix is simple, we just flush stdout & stderr before fork. Usually
>>>> stderr is unbuffered, but that can be changed, so flush it as well
>>>> just to be safe.
>>>>
>>>> Signed-off-by: Michael Ellerman <[email protected]>
>>>> ---
>>>> tools/testing/selftests/kselftest_harness.h | 5 +++++
>>>> 1 file changed, 5 insertions(+)
>>>
>>> Tested-by: Max Filippov <[email protected]>
>>
>> Thank you both. Applying to linux-kselftest fixes for 5.9-rc7
>
> It can wait for v5.10 IMHO, but up to you.
>

Good to know. I will send this for 5.10-rc1 then.

thanks,
-- Shuah