[RFC PATCH v1 34/50] mm/slub.c: Use cheaper prandom source in shuffle_freelist

From: George Spelvin
Date: Sat Mar 28 2020 - 12:46:15 EST


The pre-generated permutation which we're choosing a random starting
offset into was itself generated with prandom (in freelist_randomize()),
so there's not much point using a crypto-quality generator here.

Also, prandom_u32_max() uses a multiplicative algorithm for generating
random integers in a range, which is faster than modulo.

TODO: Figure out a better algorithm for the whole thing. A single
permutation with a random starting point is a bit limiting.
Can we make a full shuffle fast enough. or do we need to stick
with something less general?

We could easily add an outer offset: off2 + random_seq[off1 + i]

Perhaps instead of just a random starting point, a random start
and step? Unfortunately, the step must be relatively prime to the
count, and the latter is not chosen to make things convenient.
But it's easy enough to precompute a list of possible steps.
Or we could at least allow steps of +1 and -1.

Should random_seq be changed periodically? It's a separately
allocated structure, so it's easy to allocate a new one and swap
it out atomically.

or even an Enigma-style double rotor:
page[i] = off3 + random_seq2[off2 + random_seq1[off1 + i]]

Signed-off-by: George Spelvin <lkml@xxxxxxx>
Cc: Thomas Garnier <thgarnie@xxxxxxxxxxxx>
Cc: Yu Zhao <yuzhao@xxxxxxxxxx>
Cc: Sean Rees <sean@xxxxxxxxxx>
---
mm/slub.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/slub.c b/mm/slub.c
index 8eafccf759409..94d765666cff0 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1580,7 +1580,7 @@ static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
return false;

freelist_count = oo_objects(s->oo);
- pos = get_random_int() % freelist_count;
+ pos = prandom_u32_max(freelist_count);

page_limit = page->objects * s->size;
start = fixup_red_left(s, page_address(page));
--
2.26.0