On Tue, Aug 09, 2011 at 11:46:09AM +0200, Eric Dumazet wrote:
> Le mardi 09 août 2011 à 18:38 +0900, Akinobu Mita a écrit :
> > 2011/8/9 Eric Dumazet <eric.dumazet@xxxxxxxxx>:
> >
> > >> > diff --git a/mm/slub.c b/mm/slub.c
> > >> > index eb5a8f9..5695f92 100644
> > >> > --- a/mm/slub.c
> > >> > +++ b/mm/slub.c
> > >> > @@ -701,7 +701,7 @@ static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
> > >> > return check_bytes8(start, value, bytes);
> > >> >
> > >> > value64 = value | value << 8 | value << 16 | value << 24;
> > >> > - value64 = value64 | value64 << 32;
> > >> > + value64 = (value64 & 0xffffffff) | value64 << 32;
> > >> > prefix = 8 - ((unsigned long)start) % 8;
> > >> >
> > >> > if (prefix) {
> > >>
> > >> Still buggy I am afraid. Could we use the following ?
> > >>
> > >>
> > >> value64 = value;
> > >> value64 |= value64 << 8;
> > >> value64 |= value64 << 16;
> > >> value64 |= value64 << 32;
> > >>
> > >>
> > >
> > > Well, 'buggy' was not well chosen.
> > >
> > > Another possibility would be to use a multiply if arch has a fast
> > > multiplier...
> > >
> > >
> > > value64 = value;
> > > #if defined(ARCH_HAS_FAST_MULTIPLIER) && BITS_PER_LONG == 64
> > > value64 *= 0x0101010101010101;
> > > #elif defined(ARCH_HAS_FAST_MULTIPLIER)
> > > value64 *= 0x01010101;
> > > value64 |= value64 << 32;
> > > #else
> > > value64 |= value64 << 8;
> > > value64 |= value64 << 16;
> > > value64 |= value64 << 32;
> > > #endif
> >
> > I don't really care about which one should be used. So tell me if I need
> > to resend it with this improvement.
>
> It would be nice to fix all bugs while we review this code.
>
> Lets push your patch and I'll submit a patch for next kernel.
>
> For example, following code is suboptimal :
>
> prefix = 8 - ((unsigned long)start) % 8;
>
> if (prefix) {
> u8 *r = check_bytes8(start, value, prefix);
> if (r)
> return r;
> start += prefix;
> bytes -= prefix;
> }
>
>
> Since we always have prefix = 8 if 'start' is longword aligned, so we
> call check_bytes8() at least once with 8 bytes to compare...
Yeah.
> Also, 32bit arches should be taken into account properly.
At least on x86_32 reading 8 bytes is faster than 4 (I benchmarked it - IIRC
reading 8 bytes speeds up by a factor of ~5 and reading 4 only by ~3.5)
Marcin
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>