Re: [PATCH v6 2/2] drivers/perf: Add CCPI2 PMU support in ThunderX2 UNCORE driver.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 18/10/2019 05:21, Ganapatrao Kulkarni wrote:
Hi Will,

On Thu, Oct 17, 2019 at 9:17 PM Will Deacon <will@xxxxxxxxxx> wrote:

On Thu, Oct 17, 2019 at 12:38:51PM +0530, Ganapatrao Kulkarni wrote:
On Wed, Oct 16, 2019 at 7:01 PM John Garry <john.garry@xxxxxxxxxx> wrote:
+TX2_EVENT_ATTR(req_pktsent, CCPI2_EVENT_REQ_PKT_SENT);
+TX2_EVENT_ATTR(snoop_pktsent, CCPI2_EVENT_SNOOP_PKT_SENT);
+TX2_EVENT_ATTR(data_pktsent, CCPI2_EVENT_DATA_PKT_SENT);
+TX2_EVENT_ATTR(gic_pktsent, CCPI2_EVENT_GIC_PKT_SENT);
+
+static struct attribute *ccpi2_pmu_events_attrs[] = {
+     &tx2_pmu_event_attr_req_pktsent.attr.attr,
+     &tx2_pmu_event_attr_snoop_pktsent.attr.attr,
+     &tx2_pmu_event_attr_data_pktsent.attr.attr,
+     &tx2_pmu_event_attr_gic_pktsent.attr.attr,
+     NULL,
+};

Hi Ganapatrao,

Have you considered adding these as uncore pmu-events in the perf tool?

At the moment no, since the number of events exposed/listed are very few.

Then sounds like a perfect time to nip it in the bud before the list grows
;)

I had internal discussion with architecture team, they have confirmed
that, these are the only published events and no plan to add new.
However, If any such request comes from HW team in future, i will add
them to JSON files.

Don't you find perf list is swamped with all the uncore events?

For Huawei platform, I find this:
./perf list pmu | grep "Kernel PMU event" | grep hisi | wc -l
648

That's because we have so many instances of the same PMUs, not because there are many events per PMU.

TBH, I would like to delete all the events from the hisi uncore kernel drivers, now that they're supported in the perf tool, but I think that would constitute an ABI breakage.

Maybe there is a way to hide them, but I couldn't find it.

John


I have incorporate all your previous comments, Can you please Ack and
queue it to 5.5?


If you can manage with these things in userspace, then I agree with John
that it would be preferential to do it that way. It also offers more
flexibility if we get the metricgroup stuff working properly (I think it's
buggered for big/little atm).

Will

Thanks,
Ganapat

.






[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux