Re: perf: bug in rdpmc/mmap accounting after exec
From: Andy Lutomirski
Date: Thu Aug 03 2017 - 20:37:55 EST
On Wed, Aug 2, 2017 at 10:39 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Wed, Jul 26, 2017 at 03:39:01PM -0400, Vince Weaver wrote:
>> Hello
>>
>> so one last bug found by the PAPI testsuite.
>>
>> This one involves the rdpmc auto-disable on last unmap of an event
>> feature.
>>
>> Failing test case:
>>
>> fd=perf_event_open();
>> addr=mmap(fd);
>> exec() // without closing or unmapping the event
>> fd=perf_event_open();
>> addr=mmap(fd);
>> rdpmc() // GPFs due to rdpmc being disabled
>>
>> I won't pretend to be able to follow the rdpmc disabling code, but if I
>> add some printks it looks like
>> current->mm->context.perf_rdpmc_allowed
>> isn't properly being reset on exec?
>>
>> In fact, current->mm->context.perf_rdpmc_allowed goes negative which seems
>> like it shouldn't happen?
>>
>> Anyway, a test case for this can be found in the perf_event_tests,
>> tests/rdpmc/rdpmc_exec_papi
>
> Good find that...
>
> The below seems to fix that for me.
>
...because execve plays funny games with non-current mms. Whoops.
Reviewed-by: Andy Lutomirski <luto@xxxxxxxxxx>
Can you send it upstream and tag it for stable?