Re: [RFC 3/4] perf: Allow per PMU access control

[Tvrtko Ursulin <tursulin@ursulin.net>]

On 27/06/18 10:47, Alexey Budankov wrote:
> 
> 
> On 27.06.2018 12:15, Tvrtko Ursulin wrote:
>>
>> On 26/06/18 18:25, Alexey Budankov wrote:
>>> Hi,
>>>
>>> On 26.06.2018 18:36, Tvrtko Ursulin wrote:
>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>>>
>>>> For situations where sysadmins might want to allow different level of
>>>> access control for different PMUs, we start creating per-PMU
>>>> perf_event_paranoid controls in sysfs.
>>>>
>>>> These work in equivalent fashion as the existing perf_event_paranoid
>>>> sysctl, which now becomes the parent control for each PMU.
>>>>
>>>> On PMU registration the global/parent value will be inherited by each PMU,
>>>> as it will be propagated to all registered PMUs when the sysctl is
>>>> updated.
>>>>
>>>> At any later point individual PMU access controls, located in
>>>> <sysfs>/device/<pmu-name>/perf_event_paranoid, can be adjusted to achieve
>>>> fine grained access control.
>>>>
>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>>> Cc: Thomas Gleixner <tglx@linutronix.de>
>>>> Cc: Peter Zijlstra <peterz@infradead.org>
>>>> Cc: Ingo Molnar <mingo@redhat.com>
>>>> Cc: "H. Peter Anvin" <hpa@zytor.com>
>>>> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
>>>> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
>>>> Cc: Jiri Olsa <jolsa@redhat.com>
>>>> Cc: Namhyung Kim <namhyung@kernel.org>
>>>> Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
>>>> Cc: Andi Kleen <ak@linux.intel.com>
>>>> Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
>>>> Cc: linux-kernel@vger.kernel.org
>>>> Cc: x86@kernel.org
>>>> ---
>>>>    include/linux/perf_event.h | 12 ++++++--
>>>>    kernel/events/core.c       | 59 ++++++++++++++++++++++++++++++++++++++
>>>>    kernel/sysctl.c            |  4 ++-
>>>>    3 files changed, 71 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
>>>> index d7938d88c028..22e91cc2d9e1 100644
>>>> --- a/include/linux/perf_event.h
>>>> +++ b/include/linux/perf_event.h
>>>> @@ -271,6 +271,9 @@ struct pmu {
>>>>        /* number of address filters this PMU can do */
>>>>        unsigned int            nr_addr_filters;
>>>>    +    /* per PMU access control */
>>>> +    int                perf_event_paranoid;
>>>
>>> It looks like it needs to be declared as atomic and atomic_read/atomic_write
>>> operations need to be explicitly used below in the patch as far this
>>> variable may be manipulated by different threads at the same time
>>> without explicit locking.
>>
>> It is just a write of an integer from either sysfs access or sysctl. As such I don't think going atomic would change anything. There is no RMW or increment or anything on it.
>>
>> Unless there are architectures where int stores are not atomic? But then the existing sysctl would have the same issue. So I suspect we can indeed rely on int store being atomic.
> 
> Yep, aligned word read/write is atomic on Intel and there is no runtime issue
> currently, but the implementation itself is multithread and implicitly relies
> on that atomicity so my suggestion is just explicitly code that assumption :).
> Also, as you mentioned, that makes the arch independent part of code more portable.

Perhaps we are not on the same page, but my argument was that the 
current sysctl (or sysctl via proc) has the same issue with multiple 
threads potentially writing to it. As long as the end result is a valid 
value it is not a problem. So I don't see what this patch changes in 
that respect. Different tasks writing either of the sysctl/sysfs values 
race, but end up with valid values everywhere. If we can rely on int 
stores to be atomic on all architectures I don't see an effective 
difference after changing to atomic_t, or even taking the pmu mutex over 
the per PMU sysfs writes. So I don't see value in complicating the code 
with either approach but am happy to be proved wrong if that is the case.

Regards,

Tvrtko