Hi Karel, Jeffrey, Mizuma-san
Sorry for the late reply.
> From: Masayoshi Mizuma <msys.mizuma@xxxxxxxxx>
> Sent: Thursday, October 1, 2020 12:05 PM
> To: Jeffrey Bastian <jbastian@xxxxxxxxxx>
> Cc: Karel Zak <kzak@xxxxxxxxxx>; Nakamura, Shunsuke
> <nakamura.shun@xxxxxxxxxxx>; util-linux@xxxxxxxxxxxxxxx; Nakamura,
> Shunsuke <nakamura.shun@xxxxxxxxxxx>
> Subject: Re: [PATCH] lscpu: Add FUJITSU aarch64 A64FX cpupart
>
> On Wed, Sep 30, 2020 at 04:53:34PM -0500, Jeffrey Bastian wrote:
> > On Wed, Sep 30, 2020 at 02:49:17PM -0400, Masayoshi Mizuma wrote:
> > > Hi Jeffrey and Karel,
> > >
> > > On Wed, Sep 30, 2020 at 07:40:27PM +0200, Karel Zak wrote:
> > > > On Wed, Sep 30, 2020 at 09:44:09AM -0500, Jeffrey Bastian wrote:
> > > > > On Tue, Sep 29, 2020 at 12:46:27PM +0200, Karel Zak wrote:
> > > > > > On Mon, Sep 28, 2020 at 06:46:14PM +0900, Shunsuke Nakamura
> wrote:
> > > > > > > sys-utils/lscpu-arm.c | 6 ++++++
> > > > > > > 1 file changed, 6 insertions(+)
> > > > > >
> > > > > > Applied, thanks.
> > > > >
> > > > >
> > > > > With the other ARM patch applied ("use SMBIOS tables on ARM for
> > > > > lscpu"), the output changes on an FX700 system:
> > > > >
> > > > > [root@fujitsu-fx700 ~]# ./lscpu | grep -e Vendor -e Model -e Stepping
> > > > > Vendor ID: FUJITSU
> > > > > Model: 0
> > > > > Model name: 461F0010
> > > > > Stepping: 0x1
> > > > >
> > > > >
> > > > > The values here come from the SMBIOS Type 4 record:
> > > > >
> > > > > [root@fujitsu-fx700 ~]# strings
> > > > > /sys/firmware/dmi/entries/4-0/raw
> > > > > CPU1
> > > > > FUJITSU
> > > > > 461F0010
> > > > > 8K083MH14
> > > > >
> > > > >
> > > > > This FX700 system has firmware revision 1.4.0 (Jun 17 2020).
> > > >
> > > > hmm... 461F0010 does not seem user-friendly, "A64FX" seems better.
> > > >
> > > > Shunsuke, what about to update the firmware strings? ;-) (as
> > > > Jeffrey suggested in RH bugzilla)
> > > >
> > > > The another possibility is to exclude some vendors and do not use
> > > > SMBIOS for them (e.g. 0x46 for Fujitsu), but from my point of view
> > > > SMBIOS based solution is more elegant and ready for new CPU types
> > > > than in-lscpu hard coded tables.
> > >
> > > According to SMBIOS specification, the description of Processor
> > > Version (Type4) is "String number describing the Processor". It
> > > shows the version number of the processor, not a model, like as A64FX and
> Cortex-A72 and so on.
> > > So I think we should not use the Processor Version as the modelname...
> >
> > Hello Masa,
> >
> > I think I see your confusion because I was also confused by the
> > wording of the SMBIOS spec. In this case, "String number" means the
> > number of the string appended to the Type 4 record, not a version
> > number of the processor.
> >
> > See section 6.1.3, "Text strings", of the SMBIOS spec (version 3.4.0).
> >
> > For Type 4, there can be up to 6 strings appended to the table:
> > Offset 0x04: Socket Designation
> > Offset 0x07: Processor Manufacturer
> > Offset 0x10: Processor Version
> > Offset 0x20: Serial Number
> > Offset 0x21: Asset Tag
> > Offset 0x22: Part Number
> >
> > Each string is optional. If the byte at one of the above offsets is
> > 0, that means the string is not included. If the byte is non-0, it
> > means refer to the first or second or third (or so on) string appended
> > to the table. It's this non-0 number that the spec calls "String number".
>
> Ohh... thank you for pointing it out! I misunderstood about "String number".
>
> >
> > Example:
> >
> > A Type 4 record that only defined Processor Manufacturer and Processor
> > Version would therefore look like:
> > Offset 0x04: 0x00
> > Offset 0x07: 0x01
> > Offset 0x10: 0x02
> > Offset 0x20: 0x00
> > Offset 0x21: 0x00
> > Offset 0x22: 0x00
> >
> > The text of each string is free form. Most vendors have chosen to use
> > Processor Manufacturer as the CPU vendor, and Processor Version as a
> > text description of the CPU.
> >
> > Some examples from existing servers (copy/paste from the other thread):
> >
> > [root@hp-m400 ~]# ./lscpu | grep -i -e vendor -e model -e stepping
> > Vendor ID: AppliedMicro
> > Model: 1
> > Model name: X-Gene
> > Stepping: 0x0
> >
> > [root@lenovo-hr330a ~]# ./lscpu | grep -i -e vendor -e model -e stepping
> > Vendor ID: Ampere(TM)
> > Model: 2
> > Model name: eMAG
> > Stepping: 0x3
> >
> > [root@hpe-apollo-70 ~]# ./lscpu | grep -i -e vendor -e model -e stepping
> > Vendor ID: Cavium Inc.
> > Model: 1
> > Model name: Cavium ThunderX2(R) CPU CN9980 v2.1
> @ 2.20GHz
> > Stepping: 0x1
> >
> >
> > Therefore, I believe you are free to change the string "461F0010" to
> > "A64FX" in the SMBIOS Type 4 record of the FX700 and FX1000 servers'
> > firmware.
>
> Thanks, I got the most vendors put their CPU model name to Processor Version,
> however it depends on the vendor, right..? That's because the SMBIOS
> specification says just "String number describing the Processor", so vendors can
> put any strings to describe the own processor, like "461F0010"...
>
> But I don't have strong opinion about the Processor Version of A64FX, let's wait
> for Shunsuke's comment.
I checked with the person in charge.
When the A64FX was designed, there was ambiguity in the Processer Version specification.
Therefore we defined it as "461F0010".
Considering the maintenance cost to keep up with the latest CPU, I agree to get information
from the SMBIOS table rather than the hard code.
However, as Mizuma-san said, since "Processer Version" specs depends on the interpretation
of the vendor, the expected value is not output in the processor group existing in the current table
as in the case of the A64FX.
Given these, I propose referring to the SMBIOS table if the processor version does not exist in
the current table. Can you give me your opinion?
Best Regards
>
> Thanks,
> Masa
