Hadi Moshayedi 4e35b475a1 Fix learning total_dies in arm64. ...

We used to always determine the number of CPU dies by counting unique
values in `/sys/devices/system/cpu/cpu*/topology/die_id` files. This
method works on x64 systems but is ineffective on ARM64.

The `topology_die_id` function is defined for x64 architectures in
`arch/x86/include/asm/topology.h` but is not implemented for ARM64 in
`arch/arm64/include/asm/topology.h`.

In Linux kernel 5.15 (used in Ubuntu 22.04), the `die_id` attribute is
exposed with a value of -1 if `topology_die_id` is not defined for the
architecture. Therefore, the `die_id` file for ARM64 consistently has
the value -1, causing our method of counting unique values to always
return 1 on Ubuntu 22.04, regardless of the actual number of dies. This
can cause issues if number of sockets is more than 1, since our code
assumes `total_dies` is a multiple of `total_sockets`.

Starting with Linux kernel 5.17, a change was introduced [1] to expose
the `die_id` file only if `topology_die_id` is defined for the
architecture. Consequently, in Linux kernel 6.8 (used in Ubuntu 24.04),
this file is absent for ARM64 systems. As a result, our method of
counting unique values now produces 0 on Ubuntu 24.04.

Given the lack of a straightforward way to determine the number of dies
on ARM64 systems, this patch sets `total_dies` equal to `total_sockets`,
assuming one die per socket.

[1] https://github.com/torvalds/linux/commit/2c4dcd7

2024-11-22 12:08:50 -08:00

1.2 KiB

Raw Permalink Blame History

View Raw