Re: [PATCH] Fix copy_user on x86_64

[Vitaly Mayatskikh <v.mayatskih@gmail.com>]

Linus Torvalds <torvalds@linux-foundation.org> writes:

>> This is the patch patch for copy_user routine, you've discussed recently.
>
> I don't think it works right.
>
> Isn't this same routine also used for copy_in_user()? For that case both 
> source _and_ destination can fault, but your fixup routines assume that 
> onle one of them does (ie the fixup for a load-fault does a store for the 
> previously loaded valies, and assumes that it doesn't trap)

Right. I've missed it... :(

> Also, I'd realy rather do this all by handling the "taul" case in C. We 
> already effectively have _half_ that support: the "clear end" flag ends up 
> calling our specialized memset() routine, but it would be much nicer if 
> we:
>
>  - extended the "clear end" flag to be not just "clear end", but also 
>    which direction things are going.
>  - always call a (fixed) fixup-routine that is written in C (because 
>    performance on a cycle basis no longer matters) that gets the remaining 
>    length and the source and destination as arguments, along with the 
>    "clear and direction flag".
>  - make that fixup routine do the byte-exact tests and any necessary 
>    clearing (and return the possibly-fixed-up remaining length).
>
> Notice how this way we still have _optimal_ performance for the case where 
> no fault happens, and we don't need any complex fixups in assembly code at 
> all - the only thing the asm routines need to do is to get the right 
> length (we already have this) and fix up the source/dest pointers (we 
> don't generally have this, although the zero-at-end fixes up the 
> destination pointer in order to zero it, of course).
>
> Hmm?

Seems reasonable. However, we still need specialized memset() routine,
because, again, destination can fail. Thanks for the review, Linus!
-- 
wbr, Vitaly