Re: [PATCH] Tweak test-*exit-race tests to honor the maximum number of processes

[Carlos O'Donell <carlos at redhat dot com>]

On 01/30/2018 03:36 PM, Dmitry V. Levin wrote:
> On Tue, Jan 30, 2018 at 12:04:06PM -0800, Carlos O'Donell wrote:
>> On 01/30/2018 08:54 AM, Dmitry V. Levin wrote:
>>> When the number of threads created by test-*exit-race tests is close
>>> to the maximum number of simultaneous processes per uid, these tests
>>> fail with the following diagnostics:
>>>
>>> error: xpthread_check_return.c:32: pthread_create: Resource temporarily unavailable
>>>
>>> Workaround this issue by limiting the number of threads created
>>> by these tests to sysconf (_SC_CHILD_MAX) / 2.
>>>
>>> Tested that these tests still fail reliably on x86_64 with glibc-2.26
>>> and sysconf (_SC_CHILD_MAX) == 512.
>>>
>>> *  stdlib/test-atexit-race-common.c: Include <unistd.h>.
>>> (kNumThreads): Remove const qualifier.
>>> (do_test): Limit kNumThreads to sysconf (_SC_CHILD_MAX) / 2.
>>
>> The test should return UNSUPPORTED if an expected minimum number of threads
>> required to reliably trigger the test cannot be created e.g. 512.
> 
> It depends on what do you call reliably.  For example, on an x86_64 box
> with `getconf _NPROCESSORS_ONLN` == 32 the chance of test-atexit-race
> *not* to trigger a segfault with glibc-2.26 was about 0.08% with current
> value of kNumThreads == 1024, and this chance increased roughly 4 times
> to about 0.34% with kNumThreads == 256.

Reliably is > 50% of the time in my opinion. This means that if you sample
the test 3-5 times you are almost guaranteed a failure (87.5-97%, multiplicative
statistics).

-- 
Cheers,
Carlos.