Re: VERIFY_READ/WRITE in uaccess.h?

[Linus Torvalds <torvalds@linux-foundation.org>]

On Mon, May 11, 2015 at 2:05 PM, H. Peter Anvin <hpa@zytor.com> wrote:
> On 05/10/2015 02:44 AM, Richard Weinberger wrote:
>> Hi!
>>
>> While cleaning up UML's uaccess code I've noticed that not a single architecture
>> is using VERIFY_READ/WRITE in access_ok().
>> One exception is UML, it uses the access type in one check which is in vain anyways.
>> Also asm-generic/uaccess.h drops the type parameter silently.
>>
>> Why do we still carry it around?
>>
>> Is it because we want it for some future architecture which can benefit
>> from it or just because nobody cared enough to do a tree-wide cleanup?
>> I fear it is the latter... ;)
>>
>
> Or, perhaps, nobody noticed?

Not a future architecture. A historical one.

I think the only architecture that needed the VERIFY_READ/WRITE
distinction was the now-dropped i386, where it was used to decide if
we needed to do the manual COW because the 80386 couldn't do COW
correctly in kernel mode.

That one - for the same reasons - also checked the actual accesses,
not just that the range was in user mode. Exactly because it needed to
pre-COW the pages (even if that was then obviously racy in threaded
environments - in practice it worked, and we tried to support the
fundamentally broken i386 hardware protection model for a long time).

                         Linus