rhashtable: kill ht->shift atomic operations
authorDaniel Borkmann <daniel@iogearbox.net>
Thu, 12 Mar 2015 14:28:40 +0000 (15:28 +0100)
committerDavid S. Miller <davem@davemloft.net>
Fri, 13 Mar 2015 03:02:30 +0000 (23:02 -0400)
Commit c0c09bfdc415 ("rhashtable: avoid unnecessary wakeup for worker
queue") changed ht->shift to be atomic, which is actually unnecessary.

Instead of leaving the current shift in the core rhashtable structure,
it can be cached inside the individual bucket tables.

There, it will only be initialized once during a new table allocation
in the shrink/expansion slow path, and from then onward it stays immutable
for the rest of the bucket table liftime.

That allows shift to be non-atomic. The patch also moves hash_rnd
management into the table setup. The rhashtable structure now consumes
3 instead of 4 cachelines.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Ying Xue <ying.xue@windriver.com>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/rhashtable.h
lib/rhashtable.c

index 5ef8ea551556f1bba6ed3e0e5f52a9b1022770d6..c93ff8ac474a14a877830e879820dc849573179c 100644 (file)
@@ -50,6 +50,7 @@ struct rhash_head {
  * struct bucket_table - Table of hash buckets
  * @size: Number of hash buckets
  * @hash_rnd: Random seed to fold into hash
+ * @shift: Current size (1 << shift)
  * @locks_mask: Mask to apply before accessing locks[]
  * @locks: Array of spinlocks protecting individual buckets
  * @buckets: size * hash buckets
@@ -57,6 +58,7 @@ struct rhash_head {
 struct bucket_table {
        size_t                  size;
        u32                     hash_rnd;
+       u32                     shift;
        unsigned int            locks_mask;
        spinlock_t              *locks;
 
@@ -99,7 +101,6 @@ struct rhashtable_params {
  * @tbl: Bucket table
  * @future_tbl: Table under construction during expansion/shrinking
  * @nelems: Number of elements in table
- * @shift: Current size (1 << shift)
  * @p: Configuration parameters
  * @run_work: Deferred worker to expand/shrink asynchronously
  * @mutex: Mutex to protect current/future table swapping
@@ -110,12 +111,11 @@ struct rhashtable {
        struct bucket_table __rcu       *tbl;
        struct bucket_table __rcu       *future_tbl;
        atomic_t                        nelems;
-       atomic_t                        shift;
+       bool                            being_destroyed;
        struct rhashtable_params        p;
        struct work_struct              run_work;
        struct mutex                    mutex;
        struct list_head                walkers;
-       bool                            being_destroyed;
 };
 
 /**
index 68210cc2bab8e53c35edd1a791378935d83c4808..adea791ea3ab85d8ea582256e374d9d692579b4a 100644 (file)
@@ -147,7 +147,7 @@ static void bucket_table_free(const struct bucket_table *tbl)
 }
 
 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
-                                              size_t nbuckets)
+                                              size_t nbuckets, u32 hash_rnd)
 {
        struct bucket_table *tbl = NULL;
        size_t size;
@@ -162,6 +162,8 @@ static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
                return NULL;
 
        tbl->size = nbuckets;
+       tbl->shift = ilog2(nbuckets);
+       tbl->hash_rnd = hash_rnd;
 
        if (alloc_bucket_locks(ht, tbl) < 0) {
                bucket_table_free(tbl);
@@ -177,25 +179,27 @@ static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 /**
  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
  * @ht:                hash table
- * @new_size:  new table size
+ * @tbl:       current table
  */
-static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
+static bool rht_grow_above_75(const struct rhashtable *ht,
+                             const struct bucket_table *tbl)
 {
        /* Expand table when exceeding 75% load */
-       return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
-              (!ht->p.max_shift || atomic_read(&ht->shift) < ht->p.max_shift);
+       return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) &&
+              (!ht->p.max_shift || tbl->shift < ht->p.max_shift);
 }
 
 /**
  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
  * @ht:                hash table
- * @new_size:  new table size
+ * @tbl:       current table
  */
-static bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
+static bool rht_shrink_below_30(const struct rhashtable *ht,
+                               const struct bucket_table *tbl)
 {
        /* Shrink table beneath 30% load */
-       return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
-              (atomic_read(&ht->shift) > ht->p.min_shift);
+       return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) &&
+              tbl->shift > ht->p.min_shift;
 }
 
 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash)
@@ -310,16 +314,11 @@ int rhashtable_expand(struct rhashtable *ht)
 
        ASSERT_RHT_MUTEX(ht);
 
-       new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
+       new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, old_tbl->hash_rnd);
        if (new_tbl == NULL)
                return -ENOMEM;
 
-       new_tbl->hash_rnd = old_tbl->hash_rnd;
-
-       atomic_inc(&ht->shift);
-
        rhashtable_rehash(ht, new_tbl);
-
        return 0;
 }
 EXPORT_SYMBOL_GPL(rhashtable_expand);
@@ -342,20 +341,15 @@ EXPORT_SYMBOL_GPL(rhashtable_expand);
  */
 int rhashtable_shrink(struct rhashtable *ht)
 {
-       struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
+       struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 
        ASSERT_RHT_MUTEX(ht);
 
-       new_tbl = bucket_table_alloc(ht, tbl->size / 2);
+       new_tbl = bucket_table_alloc(ht, old_tbl->size / 2, old_tbl->hash_rnd);
        if (new_tbl == NULL)
                return -ENOMEM;
 
-       new_tbl->hash_rnd = tbl->hash_rnd;
-
-       atomic_dec(&ht->shift);
-
        rhashtable_rehash(ht, new_tbl);
-
        return 0;
 }
 EXPORT_SYMBOL_GPL(rhashtable_shrink);
@@ -376,9 +370,9 @@ static void rht_deferred_worker(struct work_struct *work)
        list_for_each_entry(walker, &ht->walkers, list)
                walker->resize = true;
 
-       if (rht_grow_above_75(ht, tbl->size))
+       if (rht_grow_above_75(ht, tbl))
                rhashtable_expand(ht);
-       else if (rht_shrink_below_30(ht, tbl->size))
+       else if (rht_shrink_below_30(ht, tbl))
                rhashtable_shrink(ht);
 unlock:
        mutex_unlock(&ht->mutex);
@@ -431,7 +425,7 @@ static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
        rcu_assign_pointer(tbl->buckets[hash], obj);
 
        atomic_inc(&ht->nelems);
-       if (no_resize_running && rht_grow_above_75(ht, tbl->size))
+       if (no_resize_running && rht_grow_above_75(ht, tbl))
                schedule_work(&ht->run_work);
 
 exit:
@@ -539,7 +533,7 @@ bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
                bool no_resize_running = tbl == old_tbl;
 
                atomic_dec(&ht->nelems);
-               if (no_resize_running && rht_shrink_below_30(ht, tbl->size))
+               if (no_resize_running && rht_shrink_below_30(ht, tbl))
                        schedule_work(&ht->run_work);
        }
 
@@ -913,6 +907,7 @@ int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
 {
        struct bucket_table *tbl;
        size_t size;
+       u32 hash_rnd;
 
        size = HASH_DEFAULT_SIZE;
 
@@ -939,14 +934,14 @@ int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
        else
                ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 
-       tbl = bucket_table_alloc(ht, size);
+       get_random_bytes(&hash_rnd, sizeof(hash_rnd));
+
+       tbl = bucket_table_alloc(ht, size, hash_rnd);
        if (tbl == NULL)
                return -ENOMEM;
 
-       get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
-
        atomic_set(&ht->nelems, 0);
-       atomic_set(&ht->shift, ilog2(tbl->size));
+
        RCU_INIT_POINTER(ht->tbl, tbl);
        RCU_INIT_POINTER(ht->future_tbl, tbl);