[PATCH] (repost) kmem_cache_zalloc

[PATCH] (repost) kmem_cache_zalloc

[Eric Sandeen]

(reposting)

There was a brief thread on this patch a while ago, please see 
http://www.uwsg.iu.edu/hypermail/linux/kernel/0203.3/0601.html

In short, XFS is using a kmem_cache_zalloc() function which just
does kmem_cache_alloc + memset.

We'd like to incorporate this into the kernel proper, and several others
chimed in that it would be useful, so here's the patch.  If it's a no-go
with Linus, we can roll this functionality back under fs/xfs to reduce
our changes in the mainline kernel.

Brian Gerst suggested adding a flag to the cache to tell
kmem_cache_zalloc() to zero the object, and this also sounds like a
reasonable way to go, but there was no discussion after that.

Thanks,

-Eric

--- linux-orig/include/linux/slab.h	Mon Mar 18 14:37:14 2002
+++ linux/include/linux/slab.h	Wed Apr  3 14:58:40 2002
@@ -56,6 +56,7 @@
 extern int kmem_cache_destroy(kmem_cache_t *);
 extern int kmem_cache_shrink(kmem_cache_t *);
 extern void *kmem_cache_alloc(kmem_cache_t *, int);
+extern void *kmem_cache_zalloc(kmem_cache_t *, int);
 extern void kmem_cache_free(kmem_cache_t *, void *);
 
 extern void *kmalloc(size_t, int);
--- linux-orig/mm/slab.c	Mon Mar 18 14:37:18 2002
+++ linux/mm/slab.c	Tue Apr  2 12:56:38 2002
@@ -1611,6 +1611,23 @@
 	local_irq_restore(flags);
 }
 
+void *
+kmem_cache_zalloc(kmem_cache_t *cachep, int flags)
+{
+	void    *ptr;
+	ptr = __kmem_cache_alloc(cachep, flags);
+	if (ptr)
+#if DEBUG
+		memset(ptr, 0, cachep->objsize -
+			(cachep->flags & SLAB_RED_ZONE ? 2*BYTES_PER_WORD : 0));
+#else
+		memset(ptr, 0, cachep->objsize);
+#endif
+
+	return ptr;
+}
+
+
 /**
  * kfree - free previously allocated memory
  * @objp: pointer returned by kmalloc.
--- linux-orig/kernel/ksyms.c	Mon Mar 18 14:37:03 2002
+++ linux/kernel/ksyms.c	Tue Apr  2 12:56:38 2002
@@ -102,6 +115,7 @@
 EXPORT_SYMBOL(kmem_cache_destroy);
 EXPORT_SYMBOL(kmem_cache_shrink);
 EXPORT_SYMBOL(kmem_cache_alloc);
+EXPORT_SYMBOL(kmem_cache_zalloc);
 EXPORT_SYMBOL(kmem_cache_free);
 EXPORT_SYMBOL(kmalloc);
 EXPORT_SYMBOL(kfree);
-- 
Eric Sandeen      XFS for Linux     http://oss.sgi.com/projects/xfs
sandeen@sgi.com   SGI, Inc.

Re: [PATCH] (repost) kmem_cache_zalloc

[arjan]

In article <1019682472.15455.33.camel@stout.americas.sgi.com> you wrote:
> We'd like to incorporate this into the kernel proper, and several others
> chimed in that it would be useful, so here's the patch.  If it's a no-go
> with Linus, we can roll this functionality back under fs/xfs to reduce
> our changes in the mainline kernel.

personally I liked the kcalloc suggesition more; that would fix a lot of
multiplication exploitable bugs at the same time

Re: [PATCH] (repost) kmem_cache_zalloc

[Christoph Hellwig]

On Wed, Apr 24, 2002 at 04:07:52PM -0500, Eric Sandeen wrote:
> (reposting)
> 
> There was a brief thread on this patch a while ago, please see 
> http://www.uwsg.iu.edu/hypermail/linux/kernel/0203.3/0601.html
> 
> In short, XFS is using a kmem_cache_zalloc() function which just
> does kmem_cache_alloc + memset.

Hi Eric,

I don't think kmem_cache_zalloc is a good idea. The idea behind the slab
cache is to allow object reuse by storing constructed objects in the
caches, and a memset directly after the alloc destroys the object state.
A kmen_zalloc/kzalloc might make more sense.

Re: [PATCH] (repost) kmem_cache_zalloc

[Eric Sandeen]

Hi Christoph - 

On Thu, 2002-04-25 at 03:41, Christoph Hellwig wrote:

> I don't think kmem_cache_zalloc is a good idea. The idea behind the slab
> cache is to allow object reuse by storing constructed objects in the
> caches, and a memset directly after the alloc destroys the object state.
> A kmen_zalloc/kzalloc might make more sense.

The constructor is one part of it, but what about more efficient memory
use?  If we let things fall into the default power-of-two caches, we'd
waste quite a lot of memory.

See http://oss.sgi.com/~sandeen/slabinfo.html for example.

On this machine we'd use 30% more memory by using kmem_alloc vs
kmem_cache_alloc.

-Eric

-- 
Eric Sandeen      XFS for Linux     http://oss.sgi.com/projects/xfs
sandeen@sgi.com   SGI, Inc.

Re: [PATCH] (repost) kmem_cache_zalloc

[Alan Cox]

> caches, and a memset directly after the alloc destroys the object state.
> A kmen_zalloc/kzalloc might make more sense.

s/kzalloc/kcalloc/

Then our api continues to remind people of the C one so is easier to 
remember (and we need calloc stuff anyway in multiple places)