nfsd: convert the lru list into a per-bucket thing
authorTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 6 Aug 2014 17:44:21 +0000 (13:44 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Sun, 17 Aug 2014 16:00:12 +0000 (12:00 -0400)
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
fs/nfsd/nfscache.c

index c74bcd6..c01f44e 100644 (file)
 
 struct nfsd_drc_bucket {
        struct hlist_head cache_hash;
+       struct list_head lru_head;
 };
 
 static struct nfsd_drc_bucket  *drc_hashtbl;
-static struct list_head        lru_head;
 static struct kmem_cache       *drc_slab;
 
 /* max number of entries allowed in the cache */
@@ -40,6 +40,7 @@ static unsigned int           max_drc_entries;
 
 /* number of significant bits in the hash value */
 static unsigned int            maskbits;
+static unsigned int            drc_hashsize;
 
 /*
  * Stats and other tracking of on the duplicate reply cache. All of these and
@@ -167,8 +168,8 @@ nfsd_reply_cache_free(struct svc_cacherep *rp)
 int nfsd_reply_cache_init(void)
 {
        unsigned int hashsize;
+       unsigned int i;
 
-       INIT_LIST_HEAD(&lru_head);
        max_drc_entries = nfsd_cache_size_limit();
        num_drc_entries = 0;
        hashsize = nfsd_hashsize(max_drc_entries);
@@ -183,6 +184,9 @@ int nfsd_reply_cache_init(void)
        drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
        if (!drc_hashtbl)
                goto out_nomem;
+       for (i = 0; i < hashsize; i++)
+               INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
+       drc_hashsize = hashsize;
 
        return 0;
 out_nomem:
@@ -194,17 +198,22 @@ out_nomem:
 void nfsd_reply_cache_shutdown(void)
 {
        struct svc_cacherep     *rp;
+       unsigned int i;
 
        unregister_shrinker(&nfsd_reply_cache_shrinker);
        cancel_delayed_work_sync(&cache_cleaner);
 
-       while (!list_empty(&lru_head)) {
-               rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
-               nfsd_reply_cache_free_locked(rp);
+       for (i = 0; i < drc_hashsize; i++) {
+               struct list_head *head = &drc_hashtbl[i].lru_head;
+               while (!list_empty(head)) {
+                       rp = list_first_entry(head, struct svc_cacherep, c_lru);
+                       nfsd_reply_cache_free_locked(rp);
+               }
        }
 
        kfree (drc_hashtbl);
        drc_hashtbl = NULL;
+       drc_hashsize = 0;
 
        if (drc_slab) {
                kmem_cache_destroy(drc_slab);
@@ -217,10 +226,10 @@ void nfsd_reply_cache_shutdown(void)
  * not already scheduled.
  */
 static void
-lru_put_end(struct svc_cacherep *rp)
+lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
 {
        rp->c_timestamp = jiffies;
-       list_move_tail(&rp->c_lru, &lru_head);
+       list_move_tail(&rp->c_lru, &b->lru_head);
        schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
 }
 
@@ -234,17 +243,13 @@ hash_refile(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
        hlist_add_head(&rp->c_hash, &b->cache_hash);
 }
 
-/*
- * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
- * Also prune the oldest ones when the total exceeds the max number of entries.
- */
 static long
-prune_cache_entries(void)
+prune_bucket(struct nfsd_drc_bucket *b)
 {
        struct svc_cacherep *rp, *tmp;
        long freed = 0;
 
-       list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
+       list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
                /*
                 * Don't free entries attached to calls that are still
                 * in-progress, but do keep scanning the list.
@@ -257,16 +262,33 @@ prune_cache_entries(void)
                nfsd_reply_cache_free_locked(rp);
                freed++;
        }
+       return freed;
+}
+
+/*
+ * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
+ * Also prune the oldest ones when the total exceeds the max number of entries.
+ */
+static long
+prune_cache_entries(void)
+{
+       unsigned int i;
+       long freed = 0;
+       bool cancel = true;
+
+       for (i = 0; i < drc_hashsize; i++) {
+               struct nfsd_drc_bucket *b = &drc_hashtbl[i];
+
+               freed += prune_bucket(b);
+               if (!list_empty(&b->lru_head))
+                       cancel = false;
+       }
 
        /*
-        * Conditionally rearm the job. If we cleaned out the list, then
-        * cancel any pending run (since there won't be any work to do).
-        * Otherwise, we rearm the job or modify the existing one to run in
-        * RC_EXPIRE since we just ran the pruner.
+        * Conditionally rearm the job to run in RC_EXPIRE since we just
+        * ran the pruner.
         */
-       if (list_empty(&lru_head))
-               cancel_delayed_work(&cache_cleaner);
-       else
+       if (!cancel)
                mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
        return freed;
 }
@@ -458,7 +480,7 @@ nfsd_cache_lookup(struct svc_rqst *rqstp)
        rp->c_csum = csum;
 
        hash_refile(b, rp);
-       lru_put_end(rp);
+       lru_put_end(b, rp);
 
        /* release any buffer */
        if (rp->c_type == RC_REPLBUFF) {
@@ -475,7 +497,7 @@ found_entry:
        nfsdstats.rchits++;
        /* We found a matching entry which is either in progress or done. */
        age = jiffies - rp->c_timestamp;
-       lru_put_end(rp);
+       lru_put_end(b, rp);
 
        rtn = RC_DROPIT;
        /* Request being processed or excessive rexmits */
@@ -530,12 +552,17 @@ nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 {
        struct svc_cacherep *rp = rqstp->rq_cacherep;
        struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
+       u32             hash;
+       struct nfsd_drc_bucket *b;
        int             len;
        size_t          bufsize = 0;
 
        if (!rp)
                return;
 
+       hash = nfsd_cache_hash(rp->c_xid);
+       b = &drc_hashtbl[hash];
+
        len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
        len >>= 2;
 
@@ -568,7 +595,7 @@ nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
        }
        spin_lock(&cache_lock);
        drc_mem_usage += bufsize;
-       lru_put_end(rp);
+       lru_put_end(b, rp);
        rp->c_secure = rqstp->rq_secure;
        rp->c_type = cachetype;
        rp->c_state = RC_DONE;