Le 14/09/2021 à 14:10, Ard Biesheuvel a écrit :
Commit c65eacbe290b ("sched/core: Allow putting thread_info into
task_struct") mentions that, along with moving thread_info into
task_struct, the cpu field is moved out of the former into the latter,
but does not explain why.
I think it does explain why (init/Kconfig): "an arch will need to remove
all thread_info fields except flags".
IIUC initially the intention with THREAD_INFO_IN_TASK was to remove
everything from thread_info, but at the end it didn't happen it seems.
While collaborating with Keith on adding THREAD_INFO_IN_TASK support to
ARM, we noticed that keeping CPU in task_struct is problematic for
architectures that define raw_smp_processor_id() in terms of this field,
as it requires linux/sched.h to be included, which causes a lot of pain
in terms of circular dependencies (or 'header soup', as the original
commit refers to it).
For examples of how existing architectures work around this, please
refer to patches #6 or #7. In the former case, it uses an awful
asm-offsets hack to index thread_info/current without using its type
definition. The latter approach simply keeps a copy of the task_struct
CPU field in thread_info, and keeps it in sync at context switch time.
It was a pain when implementing that on powerpc, so I really like your
idea, the series looks good to me.
Patch #8 reverts this latter approach for ARM, but this code is still
under review so it does not currently apply to mainline.
We also discussed introducing yet another Kconfig symbol to indicate
that the arch has THREAD_INFO_IN_TASK enabled but still prefers to keep
its CPU field in thread_info, but simply keeping it in thread_info in
all cases seems to be the cleanest approach here.
Yes, if we can avoid yet another config, that's better. We already have
so many configs that are supposed to be temporary and have lasted for
years if not decades.
Christophe
