From e5fe80e4e3ae7908e65a42ed5ae01d29e5d37183 Mon Sep 17 00:00:00 2001
From: Dhruba Borthakur <dhruba@dev3936.snc6.facebook.com>
Date: Wed, 29 Aug 2012 09:47:53 -0700
Subject: [PATCH] The sharding of the block cache is limited to 2*20 pieces.

Summary:
The numbers of shards that the block cache is divided into is
configurable. However, if the user specifies that he/she wants
the block cache to be divided into more than 2**20 pieces, then
the system will rey to allocate a huge array of that size) that
could fail.

It is better to limit the sharding of the block cache to an
upper bound. The default sharding is 16 shards (i.e. 2**4)
and the maximum is now 2 million shards (i.e. 2**20).

Also, fixed a bug with the LRUCache where the numShardBits
should be a private member of the LRUCache object rather than
a static variable.

Test Plan:
run db_bench with --cache_numshardbits=64.

Task ID: #

Blame Rev:

Reviewers: heyongqiang

Reviewed By: heyongqiang

Differential Revision: https://reviews.facebook.net/D5013
---
 db/db_bench.cc |  7 ++++++-
 util/cache.cc  | 20 ++++++++++++--------
 2 files changed, 18 insertions(+), 9 deletions(-)

diff --git a/db/db_bench.cc b/db/db_bench.cc
index 4644ed87b..7d3157ba7 100644
--- a/db/db_bench.cc
+++ b/db/db_bench.cc
@@ -1048,7 +1048,12 @@ int main(int argc, char** argv) {
     } else if (sscanf(argv[i], "--cache_size=%ld%c", &n, &junk) == 1) {
       FLAGS_cache_size = n;
     } else if (sscanf(argv[i], "--cache_numshardbits=%d%c", &n, &junk) == 1) {
-      FLAGS_cache_numshardbits = n;
+      if (n < 20) {
+        FLAGS_cache_numshardbits = n;
+      } else {
+        fprintf(stderr, "The cache cannot be sharded into 2**%d pieces\n", n);
+        exit(1);
+      }
     } else if (sscanf(argv[i], "--bloom_bits=%d%c", &n, &junk) == 1) {
       FLAGS_bloom_bits = n;
     } else if (sscanf(argv[i], "--open_files=%d%c", &n, &junk) == 1) {
diff --git a/util/cache.cc b/util/cache.cc
index d76dc15c4..dedb370b3 100644
--- a/util/cache.cc
+++ b/util/cache.cc
@@ -275,21 +275,22 @@ class ShardedLRUCache : public Cache {
   LRUCache* shard_;
   port::Mutex id_mutex_;
   uint64_t last_id_;
+  int numShardBits;
 
   static inline uint32_t HashSlice(const Slice& s) {
     return Hash(s.data(), s.size(), 0);
   }
 
-  static uint32_t Shard(uint32_t hash) {
-    return hash >> (32 - kNumShardBits);
+  uint32_t Shard(uint32_t hash) {
+    return hash >> (32 - numShardBits);
   }
 
-  void init(size_t capacity, int numShardBits) {
-    kNumShardBits = numShardBits;
-    kNumShards = 1 << kNumShardBits;
-    shard_ = new LRUCache[kNumShards];
-    const size_t per_shard = (capacity + (kNumShards - 1)) / kNumShards;
-    for (int s = 0; s < kNumShards; s++) {
+  void init(size_t capacity, int numbits) {
+    numShardBits = numbits;
+    int numShards = 1 << numShardBits;
+    shard_ = new LRUCache[numShards];
+    const size_t per_shard = (capacity + (numShards - 1)) / numShards;
+    for (int s = 0; s < numShards; s++) {
       shard_[s].SetCapacity(per_shard);
     }
   }
@@ -337,6 +338,9 @@ Cache* NewLRUCache(size_t capacity) {
 }
 
 Cache* NewLRUCache(size_t capacity, int numShardBits) {
+  if (numShardBits >= 20) {
+    return NULL;  // the cache cannot be sharded into too many fine pieces
+  }
   return new ShardedLRUCache(capacity, numShardBits);
 }