nfsd: keep stats on worst hash balancing seen so far
authorJeff Layton <jlayton@redhat.com>
Wed, 27 Mar 2013 14:15:39 +0000 (10:15 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Wed, 3 Apr 2013 15:47:25 +0000 (11:47 -0400)
The typical case with the DRC is a cache miss, so if we keep track of
the max number of entries that we've ever walked over in a search, then
we should have a reasonable estimate of the longest hash chain that
we've ever seen.

With that, we'll also keep track of the total size of the cache when we
see the longest chain. In the case of a tie, we prefer to track the
smallest total cache size in order to properly gauge the worst-case
ratio of max vs. avg chain length.

Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
fs/nfsd/nfscache.c

index fd81ca79a00205a8bfc91b62257bfacf0ee9cc33..17cb0d6b9944eff727d390185febf31c4579bdfb 100644 (file)
@@ -41,6 +41,12 @@ static unsigned int          payload_misses;
 /* amount of memory (in bytes) currently consumed by the DRC */
 static unsigned int            drc_mem_usage;
 
+/* longest hash chain seen */
+static unsigned int            longest_chain;
+
+/* size of cache when we saw the longest hash chain */
+static unsigned int            longest_chain_cachesize;
+
 /*
  * Calculate the hash index from an XID.
  */
@@ -319,15 +325,30 @@ nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
 static struct svc_cacherep *
 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
 {
-       struct svc_cacherep     *rp;
+       struct svc_cacherep     *rp, *ret = NULL;
        struct hlist_head       *rh;
+       unsigned int            entries = 0;
 
        rh = &cache_hash[request_hash(rqstp->rq_xid)];
        hlist_for_each_entry(rp, rh, c_hash) {
-               if (nfsd_cache_match(rqstp, csum, rp))
-                       return rp;
+               ++entries;
+               if (nfsd_cache_match(rqstp, csum, rp)) {
+                       ret = rp;
+                       break;
+               }
+       }
+
+       /* tally hash chain length stats */
+       if (entries > longest_chain) {
+               longest_chain = entries;
+               longest_chain_cachesize = num_drc_entries;
+       } else if (entries == longest_chain) {
+               /* prefer to keep the smallest cachesize possible here */
+               longest_chain_cachesize = min(longest_chain_cachesize,
+                                               num_drc_entries);
        }
-       return NULL;
+
+       return ret;
 }
 
 /*
@@ -573,6 +594,8 @@ static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
        seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
        seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
        seq_printf(m, "payload misses:        %u\n", payload_misses);
+       seq_printf(m, "longest chain len:     %u\n", longest_chain);
+       seq_printf(m, "cachesize at longest:  %u\n", longest_chain_cachesize);
        spin_unlock(&cache_lock);
        return 0;
 }