knfsd: nfsd4: vary maximum delegation limit based on RAM size

Our original NFSv4 delegation policy was to give out a read delegation on any
open when it was possible to.

Since the lifetime of a delegation isn't limited to that of an open, a client
may quite reasonably hang on to a delegation as long as it has the inode
cached.  This becomes an obvious problem the first time a client's inode cache
approaches the size of the server's total memory.

Our first quick solution was to add a hard-coded limit.  This patch makes a
mild incremental improvement by varying that limit according to the server's
total memory size, allowing at most 4 delegations per megabyte of RAM.

My quick back-of-the-envelope calculation finds that in the worst case (where
every delegation is for a different inode), a delegation could take about
1.5K, which would make the worst case usage about 6% of memory.  The new limit
works out to be about the same as the old on a 1-gig server.

[akpm@linux-foundation.org: Don't needlessly bloat vmlinux]
[akpm@linux-foundation.org: Make it right for highmem machines]
Signed-off-by: "J. Bruce Fields" <bfields@citi.umich.edu>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

3 files changed, 27 insertions, 1 deletions

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 9cc31eaf3857..46249886ea86 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -49,6 +49,7 @@
 #include <linux/nfsd/state.h>
 #include <linux/nfsd/xdr4.h>
 #include <linux/namei.h>
+#include <linux/swap.h>
 #include <linux/mutex.h>
 #include <linux/lockd/bind.h>
 #include <linux/module.h>
@@ -150,6 +151,7 @@ get_nfs4_file(struct nfs4_file *fi)
 }
 
 static int num_delegations;
+unsigned int max_delegations;
 
 /*
  * Open owner state (share locks)
@@ -193,7 +195,7 @@ alloc_init_deleg(struct nfs4_client *clp, struct nfs4_stateid *stp, struct svc_f
         struct nfs4_callback *cb = &stp->st_stateowner->so_client->cl_callback;
 
         dprintk("NFSD alloc_init_deleg\n");
-        if (num_delegations > STATEID_HASH_SIZE * 4)
+        if (num_delegations > max_delegations)
                 return NULL;
         dp = kmem_cache_alloc(deleg_slab, GFP_KERNEL);
         if (dp == NULL)
@@ -3197,6 +3199,27 @@ get_nfs4_grace_period(void)
         return max(user_lease_time, lease_time) * HZ;
 }
 
+/*
+ * Since the lifetime of a delegation isn't limited to that of an open, a
+ * client may quite reasonably hang on to a delegation as long as it has
+ * the inode cached.  This becomes an obvious problem the first time a
+ * client's inode cache approaches the size of the server's total memory.
+ *
+ * For now we avoid this problem by imposing a hard limit on the number
+ * of delegations, which varies according to the server's memory size.
+ */
+static void
+set_max_delegations(void)
+{
+        /*
+         * Allow at most 4 delegations per megabyte of RAM.  Quick
+         * estimates suggest that in the worst case (where every delegation
+         * is for a different inode), a delegation could take about 1.5K,
+         * giving a worst case usage of about 6% of memory.
+         */
+        max_delegations = nr_free_buffer_pages() >> (20 - 2 - PAGE_SHIFT);
+}
+
 /* initialization to perform when the nfsd service is started: */
 
 static void
@@ -3212,6 +3235,7 @@ __nfs4_state_start(void)
                grace_time/HZ);
         laundry_wq = create_singlethread_workqueue("nfsd4");
         queue_delayed_work(laundry_wq, &laundromat_work, grace_time);
+        set_max_delegations();
 }
 
 int
diff --git a/include/linux/nfsd/nfsd.h b/include/linux/nfsd/nfsd.h
index 0d8420497765..ce5e345a9bce 100644
--- a/include/linux/nfsd/nfsd.h
+++ b/include/linux/nfsd/nfsd.h
@@ -148,6 +148,7 @@ extern int nfsd_max_blksize;
  * NFSv4 State
  */
 #ifdef CONFIG_NFSD_V4
+extern unsigned int max_delegations;
 void nfs4_state_init(void);
 int nfs4_state_start(void);
 void nfs4_state_shutdown(void);
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 1a889c3fec59..e2a10b957f23 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1484,6 +1484,7 @@ unsigned int nr_free_buffer_pages(void)
 {
         return nr_free_zone_pages(gfp_zone(GFP_USER));
 }
+EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
 
 /*
  * Amount of free RAM allocatable within all zones