Re: [PATCH 1/2] Introduce copy_user_handle_tail routine

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

Andi Kleen <andi@firstfloor.org> writes:

>>> >> Overall you could write it much simpler with a rep ; movs I think,
>>> >> like traditional linux did.
>> >
>> > rep movs can fail.
>
> How? (if it's a byte copy?)

Parameter len is a number of uncopied bytes, it doesn't count bytes
which were loaded into registers before GPF in unrolled
loop. copy_user_handle_tail tries to do a byte copy for, possibly,
remaining bytes, but it can fail at the first read/write, or at the
second, etc. It doesn't know where it will fail.

> The old 2.4 copy_*_user always used to that and it worked just fine AFAIK.

Again, Linus wanted it to be simple plain C routine. rep movs is not in
C language ;)

>>> >> If zerorest is ever 0 then retesting it on every iteration seems
>>> >> somewhat dumb.
>> >
>> > If zerorest is 0, this cycle will never be executed.
>
> Ok but when it's not then it will be executed on each iteration.

Ok, that's matter.

>>> >> I think a simple memset would be actually ok, i don't think we ever zero
>>> >> anything that faults. That would be obviously racy anyways. If the zero
>>> >> are supposed to override something  then a racing user thread could always
>>> >> catch it.
>> >
>> > Linus wanted this routine to be extremely dumb. This is the reason why tail
>> > handling was moved from assembly to C. Yeah, my original patches were in
>> > assembly and on the top of your realization.
>
> My point was that it could be simpler because zeroing should not ever fault
> (copy_in_user is not supposed to zero)

Why do you think that zeroing can never fail, even in userspace?
-- 
wbr, Vitaly