2005-01-27 03:36:26

by Anton Blanchard

Subject: [PATCH] Fix large allocation in nfsd init

Hi,

I just added an NFS mount to a ppc64 box that had been up for a while.
This required inserting the nfsd module. Unfortunately it failed:

modprobe: page allocation failure. order:5, mode:0xd0
Trace:
[c0000000000ba0f8] alloc_pages_current+0xc0/0xe4
[c0000000000941fc] __get_free_pages+0x54/0x1e0
[d00000000046f7d8] nfsd_cache_init+0x54/0x1a4 [nfsd]
[d0000000004782cc] init_nfsd+0x30/0x2564 [nfsd]
[c000000000084bec] sys_init_module+0x23c/0x5ac
[c00000000001045c] ret_from_syscall_1+0x0/0xa4
nfsd: cannot allocate 98304 bytes for reply cache

An order 5 allocation. Replace it with a vmalloc.

Anton

Signed-off-by: Anton Blanchard <[email protected]>

diff -puN fs/nfsd/nfscache.c~fix_nfsd_allocation fs/nfsd/nfscache.c
--- foobar2/fs/nfsd/nfscache.c~fix_nfsd_allocation 2005-01-27 13:56:14.248445552 +1100
+++ foobar2-anton/fs/nfsd/nfscache.c 2005-01-27 14:22:52.764642869 +1100
@@ -15,6 +15,7 @@
#include <linux/slab.h>
#include <linux/string.h>
#include <linux/spinlock.h>
+#include <linux/vmalloc.h>

#include <linux/sunrpc/svc.h>
#include <linux/nfsd/nfsd.h>
@@ -56,14 +57,10 @@ nfsd_cache_init(void)
struct svc_cacherep *rp;
struct nfscache_head *rh;
size_t i;
- unsigned long order;
-

i = CACHESIZE * sizeof (struct svc_cacherep);
- for (order = 0; (PAGE_SIZE << order) < i; order++)
- ;
- nfscache = (struct svc_cacherep *)
- __get_free_pages(GFP_KERNEL, order);
+
+ nfscache = vmalloc(i);
if (!nfscache) {
printk (KERN_ERR "nfsd: cannot allocate %Zd bytes for reply cache\n", i);
return;
@@ -73,7 +70,7 @@ nfsd_cache_init(void)
i = HASHSIZE * sizeof (struct nfscache_head);
hash_list = kmalloc (i, GFP_KERNEL);
if (!hash_list) {
- free_pages ((unsigned long)nfscache, order);
+ vfree(nfscache);
nfscache = NULL;
printk (KERN_ERR "nfsd: cannot allocate %Zd bytes for hash list\n", i);
return;
@@ -102,8 +99,6 @@ void
nfsd_cache_shutdown(void)
{
struct svc_cacherep *rp;
- size_t i;
- unsigned long order;

for (rp = lru_head; rp; rp = rp->c_lru_next) {
if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
@@ -112,10 +107,7 @@ nfsd_cache_shutdown(void)

cache_disabled = 1;

- i = CACHESIZE * sizeof (struct svc_cacherep);
- for (order = 0; (PAGE_SIZE << order) < i; order++)
- ;
- free_pages ((unsigned long)nfscache, order);
+ vfree(nfscache);
nfscache = NULL;
kfree (hash_list);
hash_list = NULL;
_

2005-01-27 04:03:04

by NeilBrown

Subject: Re: [PATCH] Fix large allocation in nfsd init

On Thursday January 27, [email protected] wrote:
>
> Hi,
>
> I just added an NFS mount to a ppc64 box that had been up for a while.
> This required inserting the nfsd module. Unfortunately it failed:
>
> modprobe: page allocation failure. order:5, mode:0xd0
> Trace:
> [c0000000000ba0f8] alloc_pages_current+0xc0/0xe4
> [c0000000000941fc] __get_free_pages+0x54/0x1e0
> [d00000000046f7d8] nfsd_cache_init+0x54/0x1a4 [nfsd]
> [d0000000004782cc] init_nfsd+0x30/0x2564 [nfsd]
> [c000000000084bec] sys_init_module+0x23c/0x5ac
> [c00000000001045c] ret_from_syscall_1+0x0/0xa4
> nfsd: cannot allocate 98304 bytes for reply cache
>
> An order 5 allocation. Replace it with a vmalloc.

Given that the purpose of this order-5 allocation is to provide
storage for 1024 "struct svc_cacherep" structs, it would seem that a
better approach would be to just do 1024 kmallocs.

I'll try to knock a patch together in next week sometime.

Thanks,

NeilBrown

2005-01-29 11:23:57

by Anton Blanchard

Subject: Re: [PATCH] Fix large allocation in nfsd init

Hi Neil,

> Given that the purpose of this order-5 allocation is to provide
> storage for 1024 "struct svc_cacherep" structs, it would seem that a
> better approach would be to just do 1024 kmallocs.
>
> I'll try to knock a patch together in next week sometime.

That works for me.

Anton