2006-03-15 17:37:04

by John Hawkes

[permalink] [raw]
Subject: [PATCH] fix alloc_large_system_hash roundup

The "rounded up to nearest power of 2 in size" algorithm in
alloc_large_system_hash is not correct. As coded, it takes an otherwise
acceptable power-of-2 value and doubles it. For example, we see the
error if we boot with thash_entries=2097152 which produces a hash
table with 4194304 entries.

Signed-off-by: John Hawkes <[email protected]>

Index: linux/mm/page_alloc.c
===================================================================
--- linux.orig/mm/page_alloc.c 2006-03-14 15:25:40.000000000 -0800
+++ linux/mm/page_alloc.c 2006-03-14 16:06:48.000000000 -0800
@@ -2686,7 +2686,7 @@ void *__init alloc_large_system_hash(con
numentries <<= (PAGE_SHIFT - scale);
}
/* rounded up to nearest power of 2 in size */
- numentries = 1UL << (long_log2(numentries) + 1);
+ numentries = 1UL << (long_log2(2*numentries - 1));

/* limit allocation size to 1/16 total memory by default */
if (max == 0) {


2006-03-15 17:40:57

by Roland Dreier

[permalink] [raw]
Subject: Re: [PATCH] fix alloc_large_system_hash roundup

> /* rounded up to nearest power of 2 in size */
> - numentries = 1UL << (long_log2(numentries) + 1);
> + numentries = 1UL << (long_log2(2*numentries - 1));

How about just using roundup_pow_of_two()? You could kill the comment
too then.

- R.

2006-03-15 19:06:40

by Roland Dreier

[permalink] [raw]
Subject: Re: [PATCH] fix alloc_large_system_hash roundup

Kenneth> roundup_pow_of_two uses fls, but fls takes an "int"
Kenneth> argument. I think that function is buggy on 64-bit
Kenneth> arch. Is it an oversight or something?

Huh, looks like you're right. I never looked inside fls() before.
Yes, roundup_pow_of_two() should probably be fixed, since a naive
person (like me) would look at it and think it works on longs.

- R.

2006-03-15 18:29:14

by Chen, Kenneth W

[permalink] [raw]
Subject: RE: [PATCH] fix alloc_large_system_hash roundup

Roland Dreier wrote on Wednesday, March 15, 2006 9:41 AM
> > /* rounded up to nearest power of 2 in size */
> > - numentries = 1UL << (long_log2(numentries) + 1);
> > + numentries = 1UL << (long_log2(2*numentries - 1));
>
> How about just using roundup_pow_of_two()? You could kill the comment
> too then.

roundup_pow_of_two uses fls, but fls takes an "int" argument. I think
that function is buggy on 64-bit arch. Is it an oversight or something?


static inline unsigned long roundup_pow_of_two(unsigned long x)
{
return (1UL << fls(x - 1));
}

- Ken