In the very low memory systems is in the init_bio call
scale parameter set to zero and it leads to creating
zero sized mempool.
This patch prevents pool_entries parameter become zero,
so the created pool have at least 1 entry.
Mempool with 0 entries lead to incorrect behaviour
of mempool_free. (Alloc requests are not waken up
and system stalls in mempool_alloc->ioschedule).
Signed-off-by: Milan Broz <[email protected]>
Index: linux-2.6.18-rc6/fs/bio.c
===================================================================
--- linux-2.6.18-rc6.orig/fs/bio.c
+++ linux-2.6.18-rc6/fs/bio.c
@@ -1142,7 +1142,7 @@ static int biovec_create_pools(struct bi
struct biovec_slab *bp = bvec_slabs + i;
mempool_t **bvp = bs->bvec_pools + i;
- if (i >= scale)
+ if (pool_entries > 1 && i >= scale)
pool_entries >>= 1;
*bvp = mempool_create_slab_pool(pool_entries, bp->slab);
--
VGER BF report: U 0.493705
On Mon, Sep 04 2006, Milan Broz wrote:
> In the very low memory systems is in the init_bio call
> scale parameter set to zero and it leads to creating
> zero sized mempool.
>
> This patch prevents pool_entries parameter become zero,
> so the created pool have at least 1 entry.
>
> Mempool with 0 entries lead to incorrect behaviour
> of mempool_free. (Alloc requests are not waken up
> and system stalls in mempool_alloc->ioschedule).
Good catch, queued. Maybe this is the only such scaling problem,
otherwise it may be a good idea to add a WARN_ON(!min_nr) to the mempool
setup in mm/mempool.c to catch such errors.
--
Jens Axboe