bihash table size perf/scale improvements

Directly allocate and carve cache-line-aligned chunks of virtual
memory. To a first approximation, bihash wasn't using
clib_mem_free(...).

We eliminate mheap object header/trailers, which improves space
efficiency. We also eliminate the 4gb bihash table size limit. An 8_8
bihash w/ 100 million random entries uses 3.8 Gbytes.

Change-Id: Icf925fdf99bce7d6ac407ac4edd30560b8f04808
Signed-off-by: Dave Barach <dave@barachs.net>

4 files changed, 84 insertions, 44 deletions

diff --git a/src/vnet/fib/ip6_fib.c b/src/vnet/fib/ip6_fib.c
index bb092cea1bf..9f58e4b6a9e 100644
--- a/src/vnet/fib/ip6_fib.c
+++ b/src/vnet/fib/ip6_fib.c
@@ -600,12 +600,20 @@ ip6_fib_table_show_one (ip6_fib_t *fib,
 u8 *
 format_ip6_fib_table_memory (u8 * s, va_list * args)
 {
+    uword bytes_inuse;
+
+    bytes_inuse = 
+        ip6_main.ip6_table[IP6_FIB_TABLE_NON_FWDING].ip6_hash.alloc_arena_next
+        - ip6_main.ip6_table[IP6_FIB_TABLE_NON_FWDING].ip6_hash.alloc_arena;
+
+    bytes_inuse += 
+        ip6_main.ip6_table[IP6_FIB_TABLE_FWDING].ip6_hash.alloc_arena_next
+        - ip6_main.ip6_table[IP6_FIB_TABLE_FWDING].ip6_hash.alloc_arena;
+
     s = format(s, "%=30s %=6d %=8ld\n",
                "IPv6 unicast",
                pool_elts(ip6_main.fibs),
-               mheap_bytes(ip6_main.ip6_table[IP6_FIB_TABLE_NON_FWDING].ip6_hash.mheap) +
-               mheap_bytes(ip6_main.ip6_table[IP6_FIB_TABLE_FWDING].ip6_hash.mheap));
-
+               bytes_inuse);
     return (s);
 }
 
diff --git a/src/vppinfra/bihash_template.c b/src/vppinfra/bihash_template.c
index 2b40af31d6f..89ae847c036 100644
--- a/src/vppinfra/bihash_template.c
+++ b/src/vppinfra/bihash_template.c
@@ -15,10 +15,28 @@
 
 /** @cond DOCUMENTATION_IS_IN_BIHASH_DOC_H */
 
+static inline void *BV (alloc_aligned) (BVT (clib_bihash) * h, uword nbytes)
+{
+  uword rv;
+
+  /* Round to an even number of cache lines */
+  nbytes += CLIB_CACHE_LINE_BYTES - 1;
+  nbytes &= ~(CLIB_CACHE_LINE_BYTES - 1);
+
+  rv = h->alloc_arena_next;
+  h->alloc_arena_next += nbytes;
+
+  if (rv >= (h->alloc_arena + h->alloc_arena_size))
+    os_out_of_memory ();
+
+  return (void *) rv;
+}
+
+
 void BV (clib_bihash_init)
   (BVT (clib_bihash) * h, char *name, u32 nbuckets, uword memory_size)
 {
-  void *oldheap;
+  uword bucket_size;
   int i;
 
   nbuckets = 1 << (max_log2 (nbuckets));
@@ -29,19 +47,19 @@ void BV (clib_bihash_init)
   h->cache_hits = 0;
   h->cache_misses = 0;
 
-  h->mheap = mheap_alloc (0 /* use VM */ , memory_size);
+  h->alloc_arena = (uword) clib_mem_vm_alloc (memory_size);
+  h->alloc_arena_next = h->alloc_arena;
+  h->alloc_arena_size = memory_size;
 
-  oldheap = clib_mem_set_heap (h->mheap);
-  vec_validate_aligned (h->buckets, nbuckets - 1, CLIB_CACHE_LINE_BYTES);
-  h->writer_lock = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES,
-					   CLIB_CACHE_LINE_BYTES);
+  bucket_size = nbuckets * sizeof (h->buckets[0]);
+  h->buckets = BV (alloc_aligned) (h, bucket_size);
+
+  h->writer_lock = BV (alloc_aligned) (h, CLIB_CACHE_LINE_BYTES);
   h->writer_lock[0] = 0;
 
   for (i = 0; i < nbuckets; i++)
     BV (clib_bihash_reset_cache) (h->buckets + i);
 
-  clib_mem_set_heap (oldheap);
-
   h->fmt_fn = NULL;
 }
 
@@ -53,7 +71,9 @@ void BV (clib_bihash_set_kvp_format_fn) (BVT (clib_bihash) * h,
 
 void BV (clib_bihash_free) (BVT (clib_bihash) * h)
 {
-  mheap_free (h->mheap);
+  vec_free (h->working_copies);
+  vec_free (h->freelists);
+  clib_mem_vm_free ((void *) (h->alloc_arena), h->alloc_arena_size);
   memset (h, 0, sizeof (*h));
 }
 
@@ -62,17 +82,12 @@ BVT (clib_bihash_value) *
 BV (value_alloc) (BVT (clib_bihash) * h, u32 log2_pages)
 {
   BVT (clib_bihash_value) * rv = 0;
-  void *oldheap;
 
   ASSERT (h->writer_lock[0]);
   if (log2_pages >= vec_len (h->freelists) || h->freelists[log2_pages] == 0)
     {
-      oldheap = clib_mem_set_heap (h->mheap);
-
-      vec_validate (h->freelists, log2_pages);
-      rv = clib_mem_alloc_aligned ((sizeof (*rv) * (1 << log2_pages)),
-				   CLIB_CACHE_LINE_BYTES);
-      clib_mem_set_heap (oldheap);
+      vec_validate_init_empty (h->freelists, log2_pages, 0);
+      rv = BV (alloc_aligned) (h, (sizeof (*rv) * (1 << log2_pages)));
       goto initialize;
     }
   rv = h->freelists[log2_pages];
@@ -106,17 +121,14 @@ BV (make_working_copy) (BVT (clib_bihash) * h, BVT (clib_bihash_bucket) * b)
 {
   BVT (clib_bihash_value) * v;
   BVT (clib_bihash_bucket) working_bucket __attribute__ ((aligned (8)));
-  void *oldheap;
   BVT (clib_bihash_value) * working_copy;
   u32 thread_index = os_get_thread_index ();
   int log2_working_copy_length;
 
   if (thread_index >= vec_len (h->working_copies))
     {
-      oldheap = clib_mem_set_heap (h->mheap);
       vec_validate (h->working_copies, thread_index);
       vec_validate_init_empty (h->working_copy_lengths, thread_index, ~0);
-      clib_mem_set_heap (oldheap);
     }
 
   /*
@@ -128,22 +140,20 @@ BV (make_working_copy) (BVT (clib_bihash) * h, BVT (clib_bihash_bucket) * b)
   log2_working_copy_length = h->working_copy_lengths[thread_index];
 
   h->saved_bucket.as_u64 = b->as_u64;
-  oldheap = clib_mem_set_heap (h->mheap);
 
   if (b->log2_pages > log2_working_copy_length)
     {
-      if (working_copy)
-	clib_mem_free (working_copy);
-
-      working_copy = clib_mem_alloc_aligned
-	(sizeof (working_copy[0]) * (1 << b->log2_pages),
-	 CLIB_CACHE_LINE_BYTES);
+      /*
+       * It's not worth the bookkeeping to free working copies
+       *   if (working_copy)
+       *     clib_mem_free (working_copy);
+       */
+      working_copy = BV (alloc_aligned)
+	(h, sizeof (working_copy[0]) * (1 << b->log2_pages));
       h->working_copy_lengths[thread_index] = b->log2_pages;
       h->working_copies[thread_index] = working_copy;
     }
 
-  clib_mem_set_heap (oldheap);
-
   /* Lock the bucket... */
   while (BV (clib_bihash_lock_bucket) (b) == 0)
     ;
@@ -554,6 +564,7 @@ u8 *BV (format_bihash) (u8 * s, va_list * args)
   u64 active_elements = 0;
   u64 active_buckets = 0;
   u64 linear_buckets = 0;
+  u64 used_bytes;
 
   s = format (s, "Hash table %s\n", h->name ? h->name : (u8 *) "(unnamed)");
 
@@ -633,8 +644,13 @@ u8 *BV (format_bihash) (u8 * s, va_list * args)
   s = format (s, "    %lld linear search buckets\n", linear_buckets);
   s = format (s, "    %lld cache hits, %lld cache misses\n",
 	      h->cache_hits, h->cache_misses);
-  if (h->mheap)
-    s = format (s, "    mheap: %U", format_mheap, h->mheap, 0 /* verbose */ );
+  used_bytes = h->alloc_arena_next - h->alloc_arena;
+  s = format (s,
+	      "    arena: base %llx, next %llx\n"
+	      "           used %lld b (%lld Mbytes) of %lld b (%lld Mbytes)\n",
+	      h->alloc_arena, h->alloc_arena_next,
+	      used_bytes, used_bytes >> 20,
+	      h->alloc_arena_size, h->alloc_arena_size >> 20);
   return s;
 }
 
diff --git a/src/vppinfra/bihash_template.h b/src/vppinfra/bihash_template.h
index 4e5d995cd9f..81d9ffad41e 100644
--- a/src/vppinfra/bihash_template.h
+++ b/src/vppinfra/bihash_template.h
@@ -89,7 +89,14 @@ typedef struct
   u64 cache_misses;
 
     BVT (clib_bihash_value) ** freelists;
-  void *mheap;
+
+  /*
+   * Backing store allocation. Since bihash mananges its own
+   * freelists, we simple dole out memory at alloc_arena_next.
+   */
+  uword alloc_arena;
+  uword alloc_arena_next;
+  uword alloc_arena_size;
 
   /**
     * A custom format function to print the Key and Value of bihash_key instead of default hexdump
@@ -224,7 +231,7 @@ static inline void BV (clib_bihash_unlock_bucket)
 static inline void *BV (clib_bihash_get_value) (BVT (clib_bihash) * h,
 						uword offset)
 {
-  u8 *hp = h->mheap;
+  u8 *hp = (u8 *) h->alloc_arena;
   u8 *vp = hp + offset;
 
   return (void *) vp;
@@ -235,10 +242,9 @@ static inline uword BV (clib_bihash_get_offset) (BVT (clib_bihash) * h,
 {
   u8 *hp, *vp;
 
-  hp = (u8 *) h->mheap;
+  hp = (u8 *) h->alloc_arena;
   vp = (u8 *) v;
 
-  ASSERT ((vp - hp) < 0x100000000ULL);
   return vp - hp;
 }
 
diff --git a/src/vppinfra/test_bihash_template.c b/src/vppinfra/test_bihash_template.c
index 2d4b553d259..bdcf2cd6a81 100644
--- a/src/vppinfra/test_bihash_template.c
+++ b/src/vppinfra/test_bihash_template.c
@@ -36,6 +36,7 @@ typedef struct
   int non_random_keys;
   uword *key_hash;
   u64 *keys;
+  uword hash_memory_size;
     BVT (clib_bihash) hash;
   clib_time_t clib_time;
 
@@ -101,8 +102,7 @@ test_bihash (test_main_t * tm)
 
   h = &tm->hash;
 
-  BV (clib_bihash_init) (h, "test", tm->nbuckets, 3ULL << 30);
-
+  BV (clib_bihash_init) (h, "test", tm->nbuckets, tm->hash_memory_size);
 
   for (acycle = 0; acycle < tm->ncycles; acycle++)
     {
@@ -269,10 +269,11 @@ test_bihash (test_main_t * tm)
 	}
 
       /* Clean up side-bet hash table and random key vector */
-      for (i = 0; i < tm->nitems; i++)
-	hash_unset (tm->key_hash, tm->keys[i]);
-
+      hash_free (tm->key_hash);
       vec_reset_length (tm->keys);
+      /* Recreate hash table if we're going to need it again */
+      if (acycle != (tm->ncycles - 1))
+	tm->key_hash = hash_create (tm->nitems, sizeof (uword));
     }
 
   fformat (stdout, "End of run, should be empty...\n");
@@ -322,6 +323,7 @@ test_bihash_main (test_main_t * tm)
   int which = 0;
 
   tm->report_every_n = 1;
+  tm->hash_memory_size = 4095ULL << 20;
 
   while (unformat_check_input (i) != UNFORMAT_END_OF_INPUT)
     {
@@ -344,6 +346,9 @@ test_bihash_main (test_main_t * tm)
 	;
       else if (unformat (i, "report-every %d", &tm->report_every_n))
 	;
+      else if (unformat (i, "memory-size %U",
+			 unformat_memory_size, &tm->hash_memory_size))
+	;
       else if (unformat (i, "vec64"))
 	which = 1;
       else if (unformat (i, "cache"))
@@ -356,6 +361,12 @@ test_bihash_main (test_main_t * tm)
 				  format_unformat_error, i);
     }
 
+  /* Preallocate hash table, key vector */
+  tm->key_hash = hash_create (tm->nitems, sizeof (uword));
+  vec_validate (tm->keys, tm->nitems - 1);
+  _vec_len (tm->keys) = 0;
+
+
   switch (which)
     {
     case 0:
@@ -385,7 +396,7 @@ main (int argc, char *argv[])
   clib_error_t *error;
   test_main_t *tm = &test_main;
 
-  clib_mem_init (0, 3ULL << 30);
+  clib_mem_init (0, 4095ULL << 20);
 
   tm->input = &i;
   tm->seed = 0xdeaddabe;
@@ -396,7 +407,6 @@ main (int argc, char *argv[])
   tm->verbose = 1;
   tm->search_iter = 1;
   tm->careful_delete_tests = 0;
-  tm->key_hash = hash_create (0, sizeof (uword));
   clib_time_init (&tm->clib_time);
 
   unformat_init_command_line (&i, argv);