Message-ID: <CAJ9a7VhBU4QYWYbzJs2Z91k=NC+xYnmKJ-HH9CKiNdpDxsY1SA@mail.gmail.com>
Date:   Mon, 2 Aug 2021 16:43:59 +0100
From:   Mike Leach <mike.leach@...aro.org>
To:     Leo Yan <leo.yan@...aro.org>
Cc:     James Clark <james.clark@....com>,
        Arnaldo Carvalho de Melo <acme@...nel.org>,
        Mathieu Poirier <mathieu.poirier@...aro.org>,
        Coresight ML <coresight@...ts.linaro.org>,
        Al Grant <al.grant@....com>,
        "Suzuki K. Poulose" <suzuki.poulose@....com>,
        Anshuman Khandual <anshuman.khandual@....com>,
        John Garry <john.garry@...wei.com>,
        Will Deacon <will@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...hat.com>,
        Namhyung Kim <namhyung@...nel.org>,
        linux-arm-kernel <linux-arm-kernel@...ts.infradead.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-perf-users@...r.kernel.org
Subject: Re: [PATCH 2/6] perf cs-etm: Initialise architecture based on TRCIDR1

Hi Leo,

On Mon, 2 Aug 2021 at 16:04, Leo Yan <leo.yan@...aro.org> wrote:
>
> Hi Mike,
>
> On Mon, Aug 02, 2021 at 03:04:14PM +0100, Mike Leach wrote:
>
> [...]
>
> > > > > +#define TRCIDR1_TRCARCHMIN_SHIFT 4
> > > > > +#define TRCIDR1_TRCARCHMIN_MASK  GENMASK(7, 4)
> > > > > +#define TRCIDR1_TRCARCHMIN(x)    (((x) & TRCIDR1_TRCARCHMIN_MASK) >> TRCIDR1_TRCARCHMIN_SHIFT)
> > > > > +static enum _ocsd_arch_version cs_etm_decoder__get_arch_ver(u32 reg_idr1)
> > > > > +{
> > > > > +       /*
> > > > > +        * If the ETM trace minor version is 4 or more then we can assume
> > > > > +        * the architecture is ARCH_AA64 rather than just V8
> > > > > +        */
> > > > > +       return TRCIDR1_TRCARCHMIN(reg_idr1) >= 4 ? ARCH_AA64 : ARCH_V8;
> > > > > +}
> > > >
> > > > This is true for ETM4.x & ETE 1.x (arch 5.x) but not ETM 3.x
> > > > Probably need to beef up this comment or the function name to emphasise this.
> > >
> > > Yeah, I think it's good to change the function name.  Eventually, this
> > > function should only be used for ETM4.x and ETE.
> > >
> > > Another minor comment is: can we refine the arch version number, e.g.
> > > change the OpenCSD's macro "ARCH_AA64" to "ARCH_V8R4", (or
> > > "ARCH_V8R3_AA64"), this can give more clear clue what's the ETM version.
> > >
> >
> > The purpose of these macros is to inform the decoder of the
> > architecture of the PE - not the version of the ETM.
> >
> > These OpenCSD macros are defined by the library headers
> > (ocsd_if_types.h) and not the perf headers.
> > These have been published as the API / ABI for OpenCSD and as such
> > changing them affects all OpenCSD clients, not just perf.
>
> I understand these macros are defined in OpenCSD lib as APIs, since I
> saw these macros have not been widely used in perf tool (e.g.
> ARCH_AA64), so this is why I think it's good to take chance to refine
> the naming conventions.
>

The macros are used in other tools - so changing now affects those
too. Not something I am prepared to do without good reason.

> > This PE architecture version is used along with the core profile to
> > determine which instructions are valid waypoint instructions to
> > associate with atom elements when walking the program image during
> > trace decode.
> >
> > From v8.3  onwards we moved away from filtering on specific
> > architecture versions. This was due to two factors:-
> > 1. The architectural rules now allow architectural features for one
> > increment e.g. Arch 8.4, to be backported into  the previous increment
> > - e,g, 8.3, which made this filtering more difficult to track.
> > 2. After discussion with the PE architects it was clear that
> > instructions in a later architect version would not re-use older
> > opcodes from a previous one and  be nop / invalid in the earlier
> > architectures. (certainly in the scope of AA64). Therefore
> > the policy in the decoder is to check for all the instructions we know
> > about for the latest version of architecture, even if we could be
> > decoding an earlier architecture version. This means we may check for
> > a few more opcodes than necessary for earlier version of the
> > architecture, but the overall decode is more robust and easier to
> > maintain.
> >
> > Therefore for any AA64 core beyond v8.3 - it is safe to use the
> > ARCH_AA64 PE architecture version and the decoder will handle it.
>
> I have no objection for current approach; but two things can cause
> confusions and it might be difficult for maintenance:
>
> - The first thing is now we base on the bit fields TRCIDR1::TRCARCHMIN
>   to decide the PE architecture version.  In the ETMv4 spec,
>   TRCIDR1::TRCARCHMIN is defined as the trace unit minor version,
>   so essentially it's a minor version number for tracer (ETM) but not
>   the PE architecture number.  But now we are using it to decide the
>   PE architecture number (8.3, 8.4, etc...).
>

This is a slight weakness in the implementation of perf. Ideally one
does need to establish the architecture version of the PE - but perf
/cs-etm is using an assumption regarding the profile and version of
the core, according to the ETM / ETE versiom.
That said - the ETM / ETE version numbers do have a strong
relationship with PE architecture version numbers, so this assumption
holds for the current supported devices.

> - The second thing is the macros' naming convention.
>   E.g. "AA64" gives me an impression it is a general naming "Arm Arch 64"
>   for all Arm 64-bit CPUs, it's something like an abbreviation for
>   "aarch64"; so seems to me it doesn't show any meaningful info for PE's
>   architecture version number.  This is why I proposed to use more
>   explict macro definition for architectures (e.g. ARCH_V8R3, ARCH_V8R4,
>   ARCH_V9R0, etc).
>

For modern cores it is sufficient for the decoder to know the profile
and that it is aarch 64 - so yes the macro is simply saying this a
general AA64 core.
The macros for earlier versions are a little more specific as certain
filtering is used according to the version of the PE.

ARCH_V8R4,  ARCH_V9R0 etc would have no significance to the decoder
and would not be useful. If we get to the stage where we need more
specific PE architecture versions - then these can be added as
required.
Using the ARCH_AA64 macro means that we do not have to update the API
for every version update of the architecture, and there are no changes
required to the perf / cs-etm handling.

> If we really want to use ARCH_AA64, it's better to give some comments in
> the code.
>

There are comments in the OpenCSD headers, though additional ones in
the perf  / cs-etm handling soruce code could be added.

Regards

Mike


> Thanks a lot for shared the background info.
>
> Leo



-- 
Mike Leach
Principal Engineer, ARM Ltd.
Manchester Design Centre. UK

Powered by blists - more mailing lists