Re: [uml-devel] Re: [PATCH 07/10] uml: avoid fixing faults while atomic

[Andrew Morton <akpm@osdl.org>]

Blaisorblade <blaisorblade@yahoo.it> wrote:
>
> On Wednesday 21 September 2005 21:49, Andrew Morton wrote:
> > "Paolo 'Blaisorblade' Giarrusso" <blaisorblade@yahoo.it> wrote:
> > > From: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
> 
> > The in_atomic() test in x86's do_page_fault() is in fact a message passed
> > into it from filemap.c's kmap_atomic().
> Ok, this can be ok, but:
> > It has accidental side-effects, 
> > such as making copy_to_user() fail if inside spinlocks when
> > CONFIG_PREEMPT=y.
> Sorry, but should it ever succeed inside spinlocks? I mean, should it ever 
> call down() inside spinlocks? (We never do down_trylock, and ever if we did 
> the x86 trick, that wouldn't make the whole thing safe at all - they still 
> take the spinlock and potentially sleep. And it's legal only if no spinlock 
> is held).

Not sure what you're asking here.

copy_to/from_user() will fail inside spinlock if CONFIG_PREMPT=y and if the
copy happens to cause a fault.  Otherwise it will succeed inside spinlock,
and it won't spew a sleeping-while-atomic warning, because that uses
in_atomic() too.  It might deadlock if we schedule away and try to retake
the same lock.

> Even if spinlocks don't always trigger in_atomic() - which means that we'd 
> need to have a better fix for this.

The patch you have will correctly cause copy_*_user()->pagefault to fail
the copy if the caller has run inc_preempt_count().  It will not cause
copy_*_user()->pagefault to fail inside spinlocks unless UML does
inc_preempt_count() in its spinlock implementation.

> (Btw, I took the above reasoning from something said, as an aside, on LWN.net 
> kernel page, about the FUTEX deadlock on mm->mmap_sem of ~ 2.6.8 - yes, it 
> wasn't the full truth, but not totally dumb).
> 
> > So I think this change is only needed if UML implements kmap_atomic, as in
> > arch/i386/mm/highmem.c, which it surely does not do?
> NACK, see above.

Yup, the patch is needed for the futex code, and for general correct
implementation of inc_preempt_count()'s intended effect.