Re: might_sleep() in copy_{from,to}_user and friends?

[Andrew Morton]

Arnd Bergmann wrote:
> 
> I have been looking for more places in 2.5 that can be marked
> might_sleep() and noticed that all the functions in asm/uaccess.h
> are not marked although they sleep if the memory they access
> has to be paged in.
> 
> After adding might_sleep() in ten places in asm-i386/uaccess.h
> and arch/i386/lib/usercopy.c, I have been running this kernel
> for about two weeks.

This is an excellent point.  If someone is holding a lock
across a uaccess function and userspace has passed the address
of a valid but not-present page we will hit the "atomic copy_user"
path.  Userspace will be returned an EFAULT and will be left
scratching its head, wondering what it did wrong.

Or the kernel will deadlock, of course.

I don't think we need to add the check to anything other than
ia32.  That will pick up the great bulk of any problems, and
arch-specific code won't be doing these copies much anyway.

So if you could prepare a patch which adds these checks for
ia32 it would be muchly appreciated.

And if you're feeling really keen, Dave Jones has a patch which
makes the might_sleep check a real config option rather than
overloading CONFIG_DEBUG_KERNEL - would be nice to squeeze that
out of him if poss.

Thanks.