lkml.org 
[lkml]   [2010]   [Nov]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [rfc 1/3] perf, x86: P4 PMU - describe config format
From
On Fri, Nov 26, 2010 at 2:50 PM, Cyrill Gorcunov <gorcunov@openvz.org> wrote:
> On Fri, Nov 26, 2010 at 4:10 PM, Stephane Eranian <eranian@google.com> wrote:
>> On Fri, Nov 26, 2010 at 2:04 PM, Cyrill Gorcunov <gorcunov@openvz.org> wrote:
>>> On 11/26/10, Stephane Eranian <eranian@google.com> wrote:
>>>> On Fri, Nov 26, 2010 at 12:58 PM, Cyrill Gorcunov <gorcunov@openvz.org>
>>>> wrote:
>>>>> On 11/26/10, Stephane Eranian <eranian@google.com> wrote:
>>>>>> On Fri, Nov 26, 2010 at 12:32 PM, Cyrill Gorcunov <gorcunov@openvz.org>
>>>>>> wrote:
>>>>>>> On Fri, Nov 26, 2010 at 2:14 PM, Cyrill Gorcunov <gorcunov@openvz.org>
>>>>>>> wrote:
>>>>>>>> Stephane, this is a misprint, I'll update this comments on format
>>>>>>>> (giod catch btw!). in real low 32 bits are considered as cccr in ht
>>>>>>>> mode. wait a bit, i'll post update.
>>>>>>>>
>>>>>>>> On 11/26/10, Stephane Eranian <eranian@google.com> wrote:
>>>>>>> ...
>>>>>>>>>> + *    Low 32 bits
>>>>>>>>>> + *    -----------
>>>>>>>>>> + *      0-6: P4_PEBS_METRIC enum
>>>>>>>>>> + *     7-11:                    reserved
>>>>>>>>>> + *       12: Active thread
>>>>>>>>>
>>>>>>>>> I don't understand bit 12. In the actual register, it
>>>>>>>>> corresponds to the enable bit. Seems you're overriding
>>>>>>>>> its usage. Do I interpret this as saying: 0 = enable when
>>>>>>>>> running on thread0, 1=monitoring when running on thread1?
>>>>>>>>> And if I don't care?
>>>>>>> ...
>>>>>>> I believe it simply escaped quilt refresh somehow. Here is the
>>>>>>> 'refreshed'
>>>>>>> copy (note the low bits 12-19 updated here).
>>>>>>> ---
>>>>>>> perf, x86: P4 PMU - describe config format v2
>>>>>>>
>>>>>>> Add description of .config in a sake of RAW events.
>>>>>>> At least this should bring some light to those who
>>>>>>> will be reading this code.
>>>>>>>
>>>>>>> Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
>>>>>>> CC: Lin Ming <ming.m.lin@intel.com>
>>>>>>> CC: Stephane Eranian <eranian@google.com>
>>>>>>> CC: Peter Zijlstra <peterz@infradead.org>
>>>>>>> ---
>>>>>>>  arch/x86/include/asm/perf_event_p4.h |   63
>>>>>>> ++++++++++++++++++++++++++++++-----
>>>>>>>  1 file changed, 55 insertions(+), 8 deletions(-)
>>>>>>>
>>>>>>> Index: linux-2.6.tip/arch/x86/include/asm/perf_event_p4.h
>>>>>>> ===================================================================
>>>>>>> --- linux-2.6.tip.orig/arch/x86/include/asm/perf_event_p4.h
>>>>>>> +++ linux-2.6.tip/arch/x86/include/asm/perf_event_p4.h
>>>>>>> @@ -744,14 +744,6 @@ enum P4_ESCR_EMASKS {
>>>>>>>  };
>>>>>>>
>>>>>>>  /*
>>>>>>> - * P4 PEBS specifics (Replay Event only)
>>>>>>> - *
>>>>>>> - * Format (bits):
>>>>>>> - *   0-6: metric from P4_PEBS_METRIC enum
>>>>>>> - *    7 : reserved
>>>>>>> - *    8 : reserved
>>>>>>> - * 9-11 : reserved
>>>>>>> - *
>>>>>>>  * Note we have UOP and PEBS bits reserved for now
>>>>>>>  * just in case if we will need them once
>>>>>>>  */
>>>>>>> @@ -788,5 +780,60 @@ enum P4_PEBS_METRIC {
>>>>>>>        P4_PEBS_METRIC__max
>>>>>>>  };
>>>>>>>
>>>>>>> +/*
>>>>>>> + * Notes on internal configuration of ESCR+CCCR tuples
>>>>>>> + *
>>>>>>> + * Since P4 has quite the different architecture of
>>>>>>> + * performance registers in compare with "architectural"
>>>>>>> + * once and we have on 64 bits to keep configuration
>>>>>>> + * of performance event, the following trick is used.
>>>>>>> + *
>>>>>>> + * 1) Since both ESCR and CCCR registers have only low
>>>>>>> + *    32 bits valuable, we pack them into a single 64 bit
>>>>>>> + *    configuration. Low 32 bits of such config correspond
>>>>>>> + *    to low 32 bits of CCCR register and high 32 bits
>>>>>>> + *    correspond to low 32 bits of ESCR register.
>>>>>>> + *
>>>>>>> + * 2) The meaning of every bit of such config field can
>>>>>>> + *    be found in Intel SDM but it should be noted that
>>>>>>> + *    we "borrow" some reserved bits for own usage and
>>>>>>> + *    clean them or set to a proper value when we do
>>>>>>> + *    a real write to hardware registers.
>>>>>>> + *
>>>>>>> + * 3) The format of bits of config is the following
>>>>>>> + *    and should be either 0 or set to some predefined
>>>>>>> + *    values:
>>>>>>> + *
>>>>>>> + *    Low 32 bits
>>>>>>> + *    -----------
>>>>>>> + *      0-6: P4_PEBS_METRIC enum
>>>>>>> + *     7-11:                    reserved
>>>>>>> + *       12:                    reserved (Enable)
>>>>>>> + *    13-15:                    reserved (ESCR select)
>>>>>>> + *    16-17: Active Thread
>>>>>>
>>>>>> HW has the active thread bits reserved to 0x3.
>>>>>> what about you? If not, then explain what they
>>>>>> mean.
>>>>>>
>>>>> hm, not sure i follow, hw allows you to pass any of 4 values for that
>>>>> field, so i simply pass it to kernel and then propagate to real cccr
>>>>> register. if machine is not ht capable it might be a problem, but i
>>>>> left it to a caller to set proper thread value here. i believe that
>>>>> you read cccr spec for non-ht machine while ht machine has a bit more
>>>>> flags to set.
>>>>>
>>>> You're right, I missed Figure-30.29. So you honor the field. The counter
>>>> won't count anything if the task is not running on the corresponding
>>>> HT thread, then.
>>>
>>> no ;) p4 treat this field in different meaning in compare to others
>>> perf archs. the condition is not 'where' but 'when'. ie you can
>>> specify 'when' to run counter: if none therad is active, single any
>>> thread is active, both threads are active and finally any thread is
>>> active (which is exactly 0b11 in single thread machine). i know it's
>>> weird ;)
>>>
>> That is weird indeed.
>>
>> I can see this being somewhat useful in per-cpu mode. But this is
>> weird in per-thread mode. That means you will count in your thread
>> (OS and HW) only when the other HW thread runs. I am wondering
>> about the security implications of this. I suspect you'd have to
>> disallow this mode in per-thread mode.
>
> No, single thread mode means _any_ single thread is running,
> Stephane I'll describe some more a bit later (as only reach home),
> ok?

From the manual:

00 — None. Count only when neither logical processor is active.
01 — Single. Count only when one logical processor is active (either 0 or 1).
10 — Both. Count only when both logical processors are active.
11 — Any. Count when either logical processor is active.

In per-thread mode, you won't hit 00. I suspect you want to
disallow 01, 10 (or CAP_SYS_ADMIN). Otherwise, you want
to force 11, i.e., can't figure out what's going on in the other
HT thread.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2010-11-26 14:57    [W:0.081 / U:1.784 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site