Skip to contentSkip to conversation

coretemp driver change on dual-die/package systems

DISCUSSION Zhang Rui 2 participants 3 messages Feb 11, 2019 – Feb 12, 2019 Hardware monitoring

lore
Share & tools
lore

Zhang Rui

author
#0 lore raw

Hi, Jean and Guenter,

Per the Intel Software Developer's Manual, which you can found at www.i ntel.com/sdm, when CPUID.1F is present, it is preferred over CPUID.B.

New dual-die/package systems will see this difference:
   CPUID.0B enumerates the CPUs on each die as if they were in different packages.
   CPUID.1F enumerates the CPUs on each die within the same package.

This will manifest in the sysfs physical_package_id attribute. ie. In the example above, CPUID.B would cause lscpu to show 2 packages, and CPUID.1F will cause lscpu to show 1 package.

Also, with CPUID.B the concept of a package-scope MSR and a die-scope MSR are synonymous.  With CPUID.1F, it is possible for some MSRs to have die-scope, and other MSRs to have package-scope.

MSR_IA32_PACKAGE_THERM_STATUS is a die-scope MSR, thus we need to update the coretemp driver to become multi-die-aware when we support CPUID.1F.

Previously, we create one hwmon device for each package, now we need to create one hwmon device for each die.
But there is one problem left. For each coretemp hwmon device, the "temp1_input" attribute represents the temperature got from MSR_IA32_PACKAGE_THERM_STATUS, and "temp1_label" is "Package id X", where X is the logical package id.
Now we create one hwmon device for each die, thus temp1_label can not use logical package id any more, because there are two dies in the same package.
To me, there are two choices,
1. changing "temp1_label" from "Package id X" to "Package id Y", where Y is just a unique number instead of logical package id, say, using ida.
2. changing "temp1_label" from "Package id X" to "Package id X Die id Y", where Y is the die id.

Question is I'm not sure how temp1_label is used and if this change will break any userspace, like lm-sensors?

thanks,
rui

maintainer
#1 lore raw
Quoted 38 lines · Zhang Rui

Hi, Jean and Guenter,

Per the Intel Software Developer's Manual, which you can found at www.i ntel.com/sdm, when CPUID.1F is present, it is preferred over CPUID.B.

New dual-die/package systems will see this difference:
   CPUID.0B enumerates the CPUs on each die as if they were in
different packages.
   CPUID.1F enumerates the CPUs on each die within the same package.

This will manifest in the sysfs physical_package_id attribute. ie. In the example above, CPUID.B would cause lscpu to show 2 packages, and CPUID.1F will cause lscpu to show 1 package.

Also, with CPUID.B the concept of a package-scope MSR and a die-scope MSR are synonymous.  With CPUID.1F, it is possible for some MSRs to have die-scope, and other MSRs to have package-scope.

MSR_IA32_PACKAGE_THERM_STATUS is a die-scope MSR, thus we need to update the coretemp driver to become multi-die-aware when we support CPUID.1F.

Previously, we create one hwmon device for each package, now we need to create one hwmon device for each die.
But there is one problem left. For each coretemp hwmon device, the "temp1_input" attribute represents the temperature got from MSR_IA32_PACKAGE_THERM_STATUS, and "temp1_label" is "Package id X", where X is the logical package id.
Now we create one hwmon device for each die, thus temp1_label can not use logical package id any more, because there are two dies in the same package.
To me, there are two choices,
1. changing "temp1_label" from "Package id X" to "Package id Y", where Y is just a unique number instead of logical package id, say, using ida.
2. changing "temp1_label" from "Package id X" to "Package id X Die id Y", where Y is the die id.

Quoted 3 lines

Question is I'm not sure how temp1_label is used and if this change will break any userspace, like lm-sensors?

Please feel free to change the label as it makes sense. I would suggest option 2 to avoid confusion. The string it reports is not part of the ABI and can be changed. It can be overwritten with sensors3.conf anyway, so nothing should depend on it.

Guenter

author
#2 lore raw
Quoted 55 lines · Guenter Roeck
Quoted 51 lines · Zhang Rui

Hi, Jean and Guenter,

Per the Intel Software Developer's Manual, which you can found at www.i
ntel.com/sdm, when CPUID.1F is present, it is preferred over CPUID.B.

New dual-die/package systems will see this difference:
    CPUID.0B enumerates the CPUs on each die as if they were in different packages.
    CPUID.1F enumerates the CPUs on each die within the same package.

This will manifest in the sysfs physical_package_id attribute. ie. In
the example above, CPUID.B would cause lscpu to show 2 packages, and
CPUID.1F will cause lscpu to show 1 package.

Also, with CPUID.B the concept of a package-scope MSR and a die- scope
MSR are synonymous.  With CPUID.1F, it is possible for some MSRs to have die-scope, and other MSRs to have package-scope.

MSR_IA32_PACKAGE_THERM_STATUS is a die-scope MSR, thus we need to update the coretemp driver to become multi-die-aware when we support
CPUID.1F.

Previously, we create one hwmon device for each package, now we need to
create one hwmon device for each die.
But there is one problem left. For each coretemp hwmon device, the "temp1_input" attribute represents the temperature got from MSR_IA32_PACKAGE_THERM_STATUS, and "temp1_label" is "Package id X", where X is the logical package id.
Now we create one hwmon device for each die, thus temp1_label can not
use logical package id any more, because there are two dies in the same
package.
To me, there are two choices,
1. changing "temp1_label" from "Package id X" to "Package id Y", where
Y is just a unique number instead of logical package id, say, using ida.
2. changing "temp1_label" from "Package id X" to "Package id X Die id
Y", where Y is the die id.

Quoted 3 lines

Question is I'm not sure how temp1_label is used and if this change will break any userspace, like lm-sensors?

Quoted 7 lines

Please feel free to change the label as it makes sense. I would suggest option
2 to avoid confusion. The string it reports is not part of the ABI and
can be changed. It can be overwritten with sensors3.conf anyway, so nothing
should depend on it.

Good to know. Thanks for the clarification, Guenter!

-rui

Quoted 2 lines

Guenter