Re: [PATCH] x86: Optimize tail handling for copy_user

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

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

> On Mon, 28 Jul 2008, Vitaly Mayatskikh wrote:
>>
>> Reduce protection faults count in copy_user_handle_tail routine by
>> limiting clear length to the end of page as was suggested by Linus.
>
> No, you did it wrong.
>
> The page _clearing_ has to be to the end of the copy. Yes, in practice, I 
> don't think the kernel actually very often does "copy_from_user()" calls 
> that are page crossers in the kernel, but they do happen.
>
> It's the _copy_ part that needn't cross pages, because we know that 
> something faulted, and we _know_ that it wasn't the clearing.
>
> Really: the clearing should be just a "memset()" with no limitations. It 
> can never fault without it being a serious bug, and then we should not fix 
> it up, but we should oops!

Right. I will take your code discussed previously. The only question is:
what will be a good place for *_IS_USERSPACE and CLEAR_REMAINDER flags
definitions and assembly macro ALIGN_DESTINATION (common for
copy_user_64.S and copy_user_nocache_64.S)? I have a new file
include/asm-x86/copy_user.h at the moment, but not sure if you will
think it's ok to put that stuff here:

include/asm-x86/copy_user.h:
#ifndef _COPY_USER_H
#define _COPY_USER_H

/* Flags for copy_user_handle_tail */
#define CLEAR_REMAINDER         1
#define DEST_IS_USERSPACE       2
#define SOURCE_IS_USERSPACE     4

/* Align destination */
#define FIX_ALIGNMENT 1

#define BYTES_LEFT_IN_PAGE(ptr) \
        (unsigned long)((PAGE_MASK & ((long)(ptr) + PAGE_SIZE)) - (long)(ptr))

#ifdef __ASSEMBLY__
        .macro ALIGN_DESTINATION
...
#endif /* FIX_ALIGNMENT */
        .endm
#endif /* __ASSEMBLY__ */
#endif /* _COPY_USER_H */

-- 
wbr, Vitaly