[Intel-gfx] [patch RFC 00/15] mm/highmem: Provide a preemptible variant of kmap_atomic & friends

[Thomas Gleixner]

On Sun, Sep 20 2020 at 08:41, Thomas Gleixner wrote:
> On Sat, Sep 19 2020 at 10:18, Linus Torvalds wrote:
>> Maybe I've missed something.  Is it because the new interface still
>> does "pagefault_disable()" perhaps?
>>
>> But does it even need the pagefault_disable() at all? Yes, the
>> *atomic* one obviously needed it. But why does this new one need to
>> disable page faults?
>
> It disables pagefaults because it can be called from atomic and
> non-atomic context. That was the point to get rid of
>
>          if (preeemptible())
>          	kmap();
>          else
>                 kmap_atomic();
>
> If it does not disable pagefaults, then it's just a lightweight variant
> of kmap() for short lived mappings.

Actually most usage sites of kmap atomic do not need page faults to be
disabled at all. As Daniel pointed out the implicit pagefault disable
enforces copy_from_user_inatomic() even in places which are clearly
preemptible task context.

As we need to look at each instance anyway we can add the PF disable
explicitely to the very few places which actually need it.

Thanks,

        tglx