Date:	Fri, 1 Jul 2011 14:23:08 +0200
From:	Stephane Eranian <eranian@...gle.com>
To:	Lin Ming <ming.m.lin@...el.com>
Cc:	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Ingo Molnar <mingo@...e.hu>, Andi Kleen <andi@...stfloor.org>,
	Arnaldo Carvalho de Melo <acme@...stprotocols.net>,
	linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 0/4] perf: Intel uncore pmu counting support

Lin,

Ok, false alarm. Vol3b is correct. I missed the fact that if you want
wrmsrl() to write full width generic counters, you have to use the
alternate MSR alias. I guess that's for backward compatibility.
So for full width wrmsrl on PERFCTR0 -> use 0x41c

uncore counter value: 0x678ffffeeee
core fixed counter value: 0x678ffffeeee
       core generic counter value: 0xffffffffeeee
wide core generic counter value: 0x678ffffeeee

I think full width write to fixed counters has been there
for a long time. Looks like the problem was only present
for core generic counters until SNB.




On Fri, Jul 1, 2011 at 12:49 PM, Stephane Eranian <eranian@...gle.com> wrote:
> On Fri, Jul 1, 2011 at 5:17 AM, Lin Ming <ming.m.lin@...el.com> wrote:
>> On Fri, 2011-07-01 at 00:27 +0800, Stephane Eranian wrote:
>>> On Thu, Jun 30, 2011 at 2:10 PM, Stephane Eranian <eranian@...gle.com> wrote:
>>> > On Thu, Jun 30, 2011 at 10:09 AM, Lin Ming <ming.m.lin@...el.com> wrote:
>>> >> Hi, all
>>> >>
>>> >> I posted uncore patches months ago, but it was pended due to an uncore
>>> >> interrupt problem.
>>> >>
>>> >> This series are cut to support uncore pmu counting only.
>>> >> So uncore interrupt handling is not needed.
>>> >>
>>> > You're making the assumption that when counting, you can never construct
>>> > a measurement that will cause a counter to overflow the 39 bits. If not, then
>>> > you need interrupt handling even when counting.
>>> >
>>> The actual counter width is 48. But wrmsrl() can only write the bottom 32 bits
>>> of a register. I think Intel fixed that only with SandyBridge (see Vol3b). Thus,
>>> the risk of 'silent' wrap around is much higher now as you have only 31 bits
>>> to play with.
>>
>> I just tested wrmsrl on uncore counters and it's surprised to me that it
>> supports full write.
>>
>>        val64.low  = 0xFFFFEEEE;
>>        val64.high = 0x12345678;
>>
>> On Nehalem/Westmere:
>>
>>        msr = 0x3b0; //NHM_MSR_UNCORE_PMC0
>>        wrmsrl(msr, val64.full & 0xfffffffffff); //48 bits counter
>>        rdmsrl(msr, val64.full);
>>        printfk("counter value: 0x%llx\n", val64.full);
>>
>> I got:
>>        counter value: 0x5678ffffeeee
>>
>> On SandyBridge:
>>
>>        msr = 0x716; //SNB_MSR_UNC_CBO_1_PER_CTR0
>>        wrmsrl(msr, val64.full & 0xfffffffffff); //44 bits counter
>>        rdmsrl(msr, val64.full);
>>        printfk("counter value: 0x%llx\n", val64.full);
>>
>> I got:
>>        counter value: 0x678ffffeeee
>>
> I tried on my Nehalem and SandyBridge (model 42) and I get the same results.
> The restriction applies only to core PMU counters then.
>
> But what's so weird is that on SandyBridge the counter counters are still
> limited:
>
> uncore counter value: 0x678ffffeeee
>   core counter value: 0xffffffffeeee
>
> Yet, IA32_PERF_CAPABILITIES.FW_WIDTH (bit13) is set, thus wrmsrl()
> should write the full 48-bit width (see Vol3b 30.2.2.3).
>
> Further testing revealed that you get full width ONLY with the fixed counter
> counters:
>
> FW_WIDTH=1
> uncore counter value: 0x678ffffeeee
> core counter value: 0xffffffffeeee
> fixed counter value: 0x678ffffeeee
>
> I suspect there is a bug in the HW or a bogus description in the SDM.
> I will check on that with Intel.
>
>
>>> But if I read your patch correctly, it seems you are avoiding wrmsrl() on the
>>> counter. Instead, you are reading it when you start (prev_count) and using
>>> that value to compute the delta on stop.
>>>
>>> Am I understanding your workaround correctly?
>>
>> Yes, but I didn't realize that it's a workaround.
>>
>> Lin Ming
>>
>>>
>>>
>>> >
>>> >> The uncore pmu type is allocated dynamically and exported via sysfs.
>>> >> $ cat /sys/bus/event_source/devices/uncore/type
>>> >> 6
>>> >>
>>> >> You can count uncore raw events as below,
>>> >> $ perf stat -e uncore:r0101 ls
>>> >>
>>> >> It reads uncore pmu type id from sysfs to setup perf_event_attr::type.
>>> >>
>>> >> Comments are appreciated.
>>> >>
>>> >> Thanks,
>>> >> Lin Ming
>>> >>
>>> >>
>>> >
>>
>>
>>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists