archive mirror
 help / color / mirror / Atom feed
From: John Garry <>
To: Jeremy Linton <>, <>,
Cc: <>, <>, <>,
	<>, <>,
Subject: Re: [PATCH RFC 0/2] Add basic generic ACPI soc driver
Date: Tue, 28 Jan 2020 20:07:04 +0000	[thread overview]
Message-ID: <> (raw)
In-Reply-To: <>

Hi Jeremy,

>> I did consider DMI, but I want something more generic, i.e. could 
>> cover embedded/DT systems also.
>> And I need to double check if DMI really has the info I require. Last 
>> time I checked, it didn't for my dev board, but I know that some 
>> fields are simply not filled in.
> Well the info is probably there, but that doesn't mean it should be used 
> programmatically. As your board shows, its not that reliable. And 
> looking at the linked patch I see you mention that.

Right, I am trying to stay away from that.

>>> But, quickly looking at the use case, I can't help but think you 
>>> don't really want any of the above, or the PPTT id. It seems the 
>>> mapping should actually be tied directly to the uncore PMU 
>>> definition, rather than a soc/machine/whatever identifier. Which 
>>> would imply keying off one of the ACPI object identifiers for the PMU 
>>> itself.
>> So a PMU device (/sys/bus/event_source/devices) does not have a link 
>> to the ACPI object identifiers or uncore PMU platform device etc.
>> And even if it did, there is no clear link between that ACPI object 
>> and the events it supports for that implementation.
> Having a direct link isn't ideal either. It seems you do mention the pmu 
> naming conventions, which can be controlled based on ACPI object 
> identifiers.

Not necessarily.

  Something like "uncore_dmc_hsi1" where the appended bits
> could for example vary on _CID+_UID or DT name.

We already do include some naming from ACPI tables in naming (see, 
but this is not good enough. I'll explain below.

> Not sure that is a particularly good suggestion either, but I do think 
> its a better idea to tie the mapping to the pmu type/man/version concept 
> than the SOC it appears in. The sysfs-bus-event_source-devices-* ABI 
> docs are noticeably silent on the format of the pmu name (is that 
> somewhere else?).

I would say that there is a lack of PMU naming convention, which I did 
note in my referenced patchset.

Apart from that, I think that this problem can be better explained with 
the SMMUv3 PMCG example.

So this PMU has support for a number of IMP DEF events. The SMMUv3 PMCG 
has no method to identify the implementation, so we cannot know which 
IMP DEF events are supported for a specific implementation.

The PMCG PMU naming is fixed, and is in the form smmuv3_pmcg_XXXX - so 
we cannot use some special naming. And the XXXX does not tell us 
anything about the implementation to help know the IMP DEF events.

Now the perf tool has support to know which CPU+uncore events are 
supported for a particular CPU through pmu-events feature - see 

The perf tool includes a number of per-CPU event tables.

The matching of per-CPU event table the perf tool uses is based on 
knowing the host CPUID - this is easy to retrieve this via some special 
arch-specific CPU reg, etc. So once it knows the CPUID, "perf list" 
command can show all the events for that CPU.

Now we can extend this idea for the PMCG PMU to support the IMP DEF 
events. For this, we add support for a table of "system" PMU events per 
SoC - similar to the CPU tables - containing the PMCG events. We cannot 
use the CPUID to match the event table for SoC, as a CPUID is not always 
specific to a SoC - that's definite for ARM world and definite for 
SMMUv3 PMCG. So then perf tool needs to know some SoC identifier to 
match the per-SoC events table. That's why I want the SoC id in readable 
form in sysfs.

To add a final note on uncore PMUs, for ARM this is bit of grey area. So 
currently we match uncore PMUs on CPUID. However I figure some SoC 
implementer could take, for example, an A72, and add some uncore PMUs. 
As such, we cannot always match on CPUID, so being able to match on a 
SoC identifier would be better also.

Hope it explains.


      reply	other threads:[~2020-01-28 20:07 UTC|newest]

Thread overview: 30+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-01-28 11:14 John Garry
2020-01-28 11:14 ` [PATCH RFC 1/2] ACPI/PPTT: Add acpi_pptt_get_package_info() API John Garry
2020-01-28 12:34   ` Sudeep Holla
2020-01-28 14:04     ` John Garry
2020-01-28 14:54       ` Sudeep Holla
2020-01-29 11:03         ` John Garry
2020-01-30 11:23     ` Sudeep Holla
2020-01-30 16:12       ` John Garry
2020-01-30 17:41         ` Sudeep Holla
2020-01-31 10:58           ` John Garry
2020-01-28 11:14 ` [PATCH RFC 2/2] soc: Add a basic ACPI generic driver John Garry
2020-01-28 11:56   ` Greg KH
2020-01-28 13:33     ` John Garry
2020-01-28 12:50   ` Arnd Bergmann
2020-01-28 14:46     ` John Garry
2020-01-28 15:20   ` Sudeep Holla
2020-01-28 15:59     ` John Garry
2020-01-28 16:17       ` Sudeep Holla
2020-01-28 17:51   ` Olof Johansson
2020-01-28 18:22     ` John Garry
2020-01-28 19:11       ` Rafael J. Wysocki
2020-01-28 19:28         ` John Garry
2020-01-28 22:30           ` Rafael J. Wysocki
2020-01-29 10:27             ` John Garry
2020-01-28 20:06       ` Olof Johansson
2020-01-29  9:58         ` John Garry
2020-01-28 16:56 ` [PATCH RFC 0/2] Add basic generic ACPI soc driver Jeremy Linton
2020-01-28 17:28   ` John Garry
2020-01-28 19:04     ` Jeremy Linton
2020-01-28 20:07       ` John Garry [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \ \ \ \ \ \ \ \ \ \ \ \
    --subject='Re: [PATCH RFC 0/2] Add basic generic ACPI soc driver' \

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).