fs: Remove the buffer_head kmem_cache constructor to reduce memory usage
under slub.
When using slub, having a kmem_cache constructor forces slub to add a
free pointer to the size of the cached object, which can have a
significant impact to the number of small objects that can fit into a
slab.
As buffer_head is relatively small and we can have large numbers of
them, removing the constructor is a definite win.
On x86_64 removing the constructor gives me 39 objects/slab, 3 more than
without the patch. And on x86_32 73 objects/slab, which is 9 more.
As alloc_buffer_head() already initializes each new object there is very
little difference in actual code run.
Signed-off-by: Richard Kennedy <[email protected]>
----
This patch against 2.6.33-rc7
I've been running this patch for over a week on both a x86_64 desktop &
a x86_32 laptop with no problems, only having fewer pages in the
buffer_head cache :)
regards
Richard
diff --git a/fs/buffer.c b/fs/buffer.c
index 6fa5302..bc3212e 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -3265,7 +3265,7 @@ static void recalc_bh_state(void)
struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
{
- struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
+ struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
if (ret) {
INIT_LIST_HEAD(&ret->b_assoc_buffers);
get_cpu_var(bh_accounting).nr++;
@@ -3352,15 +3352,6 @@ int bh_submit_read(struct buffer_head *bh)
}
EXPORT_SYMBOL(bh_submit_read);
-static void
-init_buffer_head(void *data)
-{
- struct buffer_head *bh = data;
-
- memset(bh, 0, sizeof(*bh));
- INIT_LIST_HEAD(&bh->b_assoc_buffers);
-}
-
void __init buffer_init(void)
{
int nrpages;
@@ -3369,7 +3360,7 @@ void __init buffer_init(void)
sizeof(struct buffer_head), 0,
(SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
SLAB_MEM_SPREAD),
- init_buffer_head);
+ NULL);
/*
* Limit the bh occupancy to 10% of ZONE_NORMAL
On Tue, Feb 09, 2010 at 01:29:51PM +0000, Richard Kennedy wrote:
> fs: Remove the buffer_head kmem_cache constructor to reduce memory usage
> under slub.
>
> When using slub, having a kmem_cache constructor forces slub to add a
> free pointer to the size of the cached object, which can have a
> significant impact to the number of small objects that can fit into a
> slab.
>
> As buffer_head is relatively small and we can have large numbers of
> them, removing the constructor is a definite win.
>
> On x86_64 removing the constructor gives me 39 objects/slab, 3 more than
> without the patch. And on x86_32 73 objects/slab, which is 9 more.
>
> As alloc_buffer_head() already initializes each new object there is very
> little difference in actual code run.
>
> Signed-off-by: Richard Kennedy <[email protected]>
Looks fine to me. Seems like it should reduce temporal cache footprint
too by just touching the objects as they are used.
Acked-by: Nick Piggin <[email protected]>
> ----
> This patch against 2.6.33-rc7
>
> I've been running this patch for over a week on both a x86_64 desktop &
> a x86_32 laptop with no problems, only having fewer pages in the
> buffer_head cache :)
>
> regards
> Richard
>
>
>
>
> diff --git a/fs/buffer.c b/fs/buffer.c
> index 6fa5302..bc3212e 100644
> --- a/fs/buffer.c
> +++ b/fs/buffer.c
> @@ -3265,7 +3265,7 @@ static void recalc_bh_state(void)
>
> struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
> {
> - struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
> + struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
> if (ret) {
> INIT_LIST_HEAD(&ret->b_assoc_buffers);
> get_cpu_var(bh_accounting).nr++;
> @@ -3352,15 +3352,6 @@ int bh_submit_read(struct buffer_head *bh)
> }
> EXPORT_SYMBOL(bh_submit_read);
>
> -static void
> -init_buffer_head(void *data)
> -{
> - struct buffer_head *bh = data;
> -
> - memset(bh, 0, sizeof(*bh));
> - INIT_LIST_HEAD(&bh->b_assoc_buffers);
> -}
> -
> void __init buffer_init(void)
> {
> int nrpages;
> @@ -3369,7 +3360,7 @@ void __init buffer_init(void)
> sizeof(struct buffer_head), 0,
> (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
> SLAB_MEM_SPREAD),
> - init_buffer_head);
> + NULL);
>
> /*
> * Limit the bh occupancy to 10% of ZONE_NORMAL
>
On 02/09/2010 08:29 AM, Richard Kennedy wrote:
> fs: Remove the buffer_head kmem_cache constructor to reduce memory usage
> under slub.
>
> When using slub, having a kmem_cache constructor forces slub to add a
> free pointer to the size of the cached object, which can have a
> significant impact to the number of small objects that can fit into a
> slab.
>
> As buffer_head is relatively small and we can have large numbers of
> them, removing the constructor is a definite win.
>
> On x86_64 removing the constructor gives me 39 objects/slab, 3 more than
> without the patch. And on x86_32 73 objects/slab, which is 9 more.
>
> As alloc_buffer_head() already initializes each new object there is very
> little difference in actual code run.
>
> Signed-off-by: Richard Kennedy<[email protected]>
Reviewed-by: Rik van Riel <[email protected]>
--
All rights reversed.