knfsd: allocate readahead cache in individual chunks
authorJeff Layton <jlayton@redhat.com>
Thu, 14 Aug 2008 02:03:27 +0000 (22:03 -0400)
committerJ. Bruce Fields <bfields@citi.umich.edu>
Mon, 29 Sep 2008 21:56:59 +0000 (17:56 -0400)
I had a report from someone building a large NFS server that they were
unable to start more than 585 nfsd threads. It was reported against an
older kernel using the slab allocator, and I tracked it down to the
large allocation in nfsd_racache_init failing.

It appears that the slub allocator handles large allocations better,
but large contiguous allocations can often be problematic. There
doesn't seem to be any reason that the racache has to be allocated as a
single large chunk. This patch breaks this up so that the racache is
built up from separate allocations.

(Thanks also to Takashi Iwai for a bugfix.)

Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Cc: Takashi Iwai <tiwai@suse.de>
fs/nfsd/vfs.c

index 1319e8027d55a596fe755847df3bd8833387c13f..aa1d0d6489a119a3e8e00aa301cc9a9719006087 100644 (file)
@@ -83,7 +83,6 @@ struct raparm_hbucket {
        spinlock_t              pb_lock;
 } ____cacheline_aligned_in_smp;
 
-static struct raparms *                raparml;
 #define RAPARM_HASH_BITS       4
 #define RAPARM_HASH_SIZE       (1<<RAPARM_HASH_BITS)
 #define RAPARM_HASH_MASK       (RAPARM_HASH_SIZE-1)
@@ -1966,11 +1965,20 @@ nfsd_permission(struct svc_rqst *rqstp, struct svc_export *exp,
 void
 nfsd_racache_shutdown(void)
 {
-       if (!raparml)
-               return;
+       struct raparms *raparm, *last_raparm;
+       unsigned int i;
+
        dprintk("nfsd: freeing readahead buffers.\n");
-       kfree(raparml);
-       raparml = NULL;
+
+       for (i = 0; i < RAPARM_HASH_SIZE; i++) {
+               raparm = raparm_hash[i].pb_head;
+               while(raparm) {
+                       last_raparm = raparm;
+                       raparm = raparm->p_next;
+                       kfree(last_raparm);
+               }
+               raparm_hash[i].pb_head = NULL;
+       }
 }
 /*
  * Initialize readahead param cache
@@ -1981,35 +1989,38 @@ nfsd_racache_init(int cache_size)
        int     i;
        int     j = 0;
        int     nperbucket;
+       struct raparms **raparm = NULL;
 
 
-       if (raparml)
+       if (raparm_hash[0].pb_head)
                return 0;
-       if (cache_size < 2*RAPARM_HASH_SIZE)
-               cache_size = 2*RAPARM_HASH_SIZE;
-       raparml = kcalloc(cache_size, sizeof(struct raparms), GFP_KERNEL);
-
-       if (!raparml) {
-               printk(KERN_WARNING
-                       "nfsd: Could not allocate memory read-ahead cache.\n");
-               return -ENOMEM;
-       }
+       nperbucket = DIV_ROUND_UP(cache_size, RAPARM_HASH_SIZE);
+       if (nperbucket < 2)
+               nperbucket = 2;
+       cache_size = nperbucket * RAPARM_HASH_SIZE;
 
        dprintk("nfsd: allocating %d readahead buffers.\n", cache_size);
-       for (i = 0 ; i < RAPARM_HASH_SIZE ; i++) {
-               raparm_hash[i].pb_head = NULL;
+
+       for (i = 0; i < RAPARM_HASH_SIZE; i++) {
                spin_lock_init(&raparm_hash[i].pb_lock);
-       }
-       nperbucket = DIV_ROUND_UP(cache_size, RAPARM_HASH_SIZE);
-       for (i = 0; i < cache_size - 1; i++) {
-               if (i % nperbucket == 0)
-                       raparm_hash[j++].pb_head = raparml + i;
-               if (i % nperbucket < nperbucket-1)
-                       raparml[i].p_next = raparml + i + 1;
+
+               raparm = &raparm_hash[i].pb_head;
+               for (j = 0; j < nperbucket; j++) {
+                       *raparm = kzalloc(sizeof(struct raparms), GFP_KERNEL);
+                       if (!*raparm)
+                               goto out_nomem;
+                       raparm = &(*raparm)->p_next;
+               }
+               *raparm = NULL;
        }
 
        nfsdstats.ra_size = cache_size;
        return 0;
+
+out_nomem:
+       dprintk("nfsd: kmalloc failed, freeing readahead buffers\n");
+       nfsd_racache_shutdown();
+       return -ENOMEM;
 }
 
 #if defined(CONFIG_NFSD_V2_ACL) || defined(CONFIG_NFSD_V3_ACL)