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

From: Thomas Gleixner
Date: Mon Sep 21 2020 - 03:39:53 EST

On Sun, Sep 20 2020 at 10:42, Linus Torvalds wrote:
> On Sun, Sep 20, 2020 at 10:40 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>> I think the more obvious solution is to split the whole exercise:
>> schedule()
>> prepare_switch()
>> unmap()
>> switch_to()
>> finish_switch()
>> map()
> Yeah, that looks much easier to explain. Ack.

So far so good, but Peter Z. just pointed out to me that I completely
missed the fact that this cannot work.

If a task is migrated to a different CPU then the mapping address will
change which will explode in colourful ways.

On RT kernels this works because we ping the task to the CPU via
migrate_disable(). On a !RT kernel migrate_disable() maps to
preempt_disable() which brings us back to square one.

/me goes back to the drawing board.