From: Jeff Layton <jlayton@redhat.com>
Date: Wed, 14 Jan 2009 08:25:53 -0500
Subject: [nfs] knfsd: alloc readahead cache in individual chunks
Message-id: 1231939553-19087-5-git-send-email-jlayton@redhat.com
O-Subject: [RHEL5.4 PATCH 4/4] BZ#459397: knfsd: allocate readahead cache in individual chunks
Bugzilla: 459397

Upstream: 54a66e548079f12a6f54c3cae96812a9ed9b54ae

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>

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 18e11f3..d531106 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -90,7 +90,6 @@ 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)
@@ -1854,11 +1853,20 @@ 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
@@ -1869,35 +1877,38 @@ 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)