Use PoolAllocator for CCoinsMap

In my benchmarks, using this pool allocator for CCoinsMap gives about
20% faster `-reindex-chainstate` with -dbcache=5000 with practically the
same memory usage. The change in max RSS changed was 0.3%.

The `validation_flush_tests` tests need to be updated because
memory allocation is now done in large pools instead of one node at a
time, so the limits need to be updated accordingly.

Author: martinus

src/coins.cpp

@@ -34,7 +34,7 @@ size_t CCoinsViewBacked::EstimateSize() const { return base->EstimateSize(); }
 
 CCoinsViewCache::CCoinsViewCache(CCoinsView* baseIn, bool deterministic) :
     CCoinsViewBacked(baseIn), m_deterministic(deterministic),
-    cacheCoins(0, SaltedOutpointHasher(/*deterministic=*/deterministic))
+    cacheCoins(0, SaltedOutpointHasher(/*deterministic=*/deterministic), CCoinsMap::key_equal{}, &m_cache_coins_memory_resource)
 {}
 
 size_t CCoinsViewCache::DynamicMemoryUsage() const {
@@ -317,7 +317,9 @@ void CCoinsViewCache::ReallocateCache()
     // Cache should be empty when we're calling this.
     assert(cacheCoins.size() == 0);
     cacheCoins.~CCoinsMap();
-    ::new (&cacheCoins) CCoinsMap(0, SaltedOutpointHasher(/*deterministic=*/m_deterministic));
+    m_cache_coins_memory_resource.~CCoinsMapMemoryResource();
+    ::new (&m_cache_coins_memory_resource) CCoinsMapMemoryResource{};
+    ::new (&cacheCoins) CCoinsMap{0, SaltedOutpointHasher{/*deterministic=*/m_deterministic}, CCoinsMap::key_equal{}, &m_cache_coins_memory_resource};
 }
 
 void CCoinsViewCache::SanityCheck() const

src/coins.h

@@ -11,6 +11,7 @@
 #include <memusage.h>
 #include <primitives/transaction.h>
 #include <serialize.h>
+#include <support/allocators/pool.h>
 #include <uint256.h>
 #include <util/hasher.h>
 
@@ -131,7 +132,23 @@ struct CCoinsCacheEntry
     CCoinsCacheEntry(Coin&& coin_, unsigned char flag) : coin(std::move(coin_)), flags(flag) {}
 };
 
-typedef std::unordered_map<COutPoint, CCoinsCacheEntry, SaltedOutpointHasher> CCoinsMap;
+/**
+ * PoolAllocator's MAX_BLOCK_SIZE_BYTES parameter here uses sizeof the data, and adds the size
+ * of 4 pointers. We do not know the exact node size used in the std::unordered_node implementation
+ * because it is implementation defined. Most implementations have an overhead of 1 or 2 pointers,
+ * so nodes can be connected in a linked list, and in some cases the hash value is stored as well.
+ * Using an additional sizeof(void*)*4 for MAX_BLOCK_SIZE_BYTES should thus be sufficient so that
+ * all implementations can allocate the nodes from the PoolAllocator.
+ */
+using CCoinsMap = std::unordered_map<COutPoint,
+                                     CCoinsCacheEntry,
+                                     SaltedOutpointHasher,
+                                     std::equal_to<COutPoint>,
+                                     PoolAllocator<std::pair<const COutPoint, CCoinsCacheEntry>,
+                                                   sizeof(std::pair<const COutPoint, CCoinsCacheEntry>) + sizeof(void*) * 4,
+                                                   alignof(void*)>>;
+
+using CCoinsMapMemoryResource = CCoinsMap::allocator_type::ResourceType;
 
 /** Cursor for iterating over CoinsView state */
 class CCoinsViewCursor
@@ -220,6 +237,7 @@ class CCoinsViewCache : public CCoinsViewBacked
      * declared as "const".
      */
     mutable uint256 hashBlock;
+    mutable CCoinsMapMemoryResource m_cache_coins_memory_resource{};
     mutable CCoinsMap cacheCoins;
 
     /* Cached dynamic memory usage for the inner Coin objects. */

src/test/coins_tests.cpp

@@ -6,6 +6,7 @@
 #include <coins.h>
 #include <script/standard.h>
 #include <streams.h>
+#include <test/util/poolresourcetester.h>
 #include <test/util/random.h>
 #include <test/util/setup_common.h>
 #include <txdb.h>
@@ -612,7 +613,8 @@ void GetCoinsMapEntry(const CCoinsMap& map, CAmount& value, char& flags, const C
 
 void WriteCoinsViewEntry(CCoinsView& view, CAmount value, char flags)
 {
-    CCoinsMap map;
+    CCoinsMapMemoryResource resource;
+    CCoinsMap map{0, CCoinsMap::hasher{}, CCoinsMap::key_equal{}, &resource};
     InsertCoinsMapEntry(map, value, flags);
     BOOST_CHECK(view.BatchWrite(map, {}));
 }
@@ -911,6 +913,7 @@ void TestFlushBehavior(
     CAmount value;
     char flags;
     size_t cache_usage;
+    size_t cache_size;
 
     auto flush_all = [&all_caches](bool erase) {
         // Flush in reverse order to ensure that flushes happen from children up.
@@ -935,6 +938,8 @@ void TestFlushBehavior(
     view->AddCoin(outp, Coin(coin), false);
 
     cache_usage = view->DynamicMemoryUsage();
+    cache_size = view->map().size();
+
     // `base` shouldn't have coin (no flush yet) but `view` should have cached it.
     BOOST_CHECK(!base.HaveCoin(outp));
     BOOST_CHECK(view->HaveCoin(outp));
@@ -949,6 +954,7 @@ void TestFlushBehavior(
 
     // CoinsMap usage should be unchanged since we didn't erase anything.
     BOOST_CHECK_EQUAL(cache_usage, view->DynamicMemoryUsage());
+    BOOST_CHECK_EQUAL(cache_size, view->map().size());
 
     // --- 3. Ensuring the entry still exists in the cache and has been written to parent
     //
@@ -965,8 +971,10 @@ void TestFlushBehavior(
         //
         flush_all(/*erase=*/ true);
 
-        // Memory usage should have gone down.
-        BOOST_CHECK(view->DynamicMemoryUsage() < cache_usage);
+        // Memory does not necessarily go down due to the map using a memory pool
+        BOOST_TEST(view->DynamicMemoryUsage() <= cache_usage);
+        // Size of the cache must go down though
+        BOOST_TEST(view->map().size() < cache_size);
 
         // --- 5. Ensuring the entry is no longer in the cache
         //
@@ -1076,4 +1084,29 @@ BOOST_AUTO_TEST_CASE(ccoins_flush_behavior)
     }
 }
 
+BOOST_AUTO_TEST_CASE(coins_resource_is_used)
+{
+    CCoinsMapMemoryResource resource;
+    PoolResourceTester::CheckAllDataAccountedFor(resource);
+
+    {
+        CCoinsMap map{0, CCoinsMap::hasher{}, CCoinsMap::key_equal{}, &resource};
+        BOOST_TEST(memusage::DynamicUsage(map) >= resource.ChunkSizeBytes());
+
+        map.reserve(1000);
+
+        // The resource has preallocated a chunk, so we should have space for at several nodes without the need to allocate anything else.
+        const auto usage_before = memusage::DynamicUsage(map);
+
+        COutPoint out_point{};
+        for (size_t i = 0; i < 1000; ++i) {
+            out_point.n = i;
+            map[out_point];
+        }
+        BOOST_TEST(usage_before == memusage::DynamicUsage(map));
+    }
+
+    PoolResourceTester::CheckAllDataAccountedFor(resource);
+}
+
 BOOST_AUTO_TEST_SUITE_END()

src/test/fuzz/coins_view.cpp

@@ -115,7 +115,8 @@ FUZZ_TARGET_INIT(coins_view, initialize_coins_view)
                 random_mutable_transaction = *opt_mutable_transaction;
             },
             [&] {
-                CCoinsMap coins_map;
+                CCoinsMapMemoryResource resource;
+                CCoinsMap coins_map{0, SaltedOutpointHasher{/*deterministic=*/true}, CCoinsMap::key_equal{}, &resource};
                 LIMITED_WHILE(fuzzed_data_provider.ConsumeBool(), 10000) {
                     CCoinsCacheEntry coins_cache_entry;
                     coins_cache_entry.flags = fuzzed_data_provider.ConsumeIntegral<unsigned char>();

src/test/validation_flush_tests.cpp

@@ -36,12 +36,12 @@ BOOST_AUTO_TEST_CASE(getcoinscachesizestate)
         BOOST_TEST_MESSAGE("CCoinsViewCache memory usage: " << view.DynamicMemoryUsage());
     };
 
-    constexpr size_t MAX_COINS_CACHE_BYTES = 1024;
+    // PoolResource defaults to 256 KiB that will be allocated, so we'll take that and make it a bit larger.
+    constexpr size_t MAX_COINS_CACHE_BYTES = 262144 + 512;
 
     // Without any coins in the cache, we shouldn't need to flush.
-    BOOST_CHECK_EQUAL(
-        chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/0),
-        CoinsCacheSizeState::OK);
+    BOOST_TEST(
+        chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/ 0) != CoinsCacheSizeState::CRITICAL);
 
     // If the initial memory allocations of cacheCoins don't match these common
     // cases, we can't really continue to make assertions about memory usage.
@@ -71,13 +71,21 @@ BOOST_AUTO_TEST_CASE(getcoinscachesizestate)
     // cacheCoins (unordered_map) preallocates.
     constexpr int COINS_UNTIL_CRITICAL{3};
 
+    // no coin added, so we have plenty of space left.
+    BOOST_CHECK_EQUAL(
+        chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes*/ 0),
+        CoinsCacheSizeState::OK);
+
     for (int i{0}; i < COINS_UNTIL_CRITICAL; ++i) {
         const COutPoint res = AddTestCoin(view);
         print_view_mem_usage(view);
         BOOST_CHECK_EQUAL(view.AccessCoin(res).DynamicMemoryUsage(), COIN_SIZE);
+
+        // adding first coin causes the MemoryResource to allocate one 256 KiB chunk of memory,
+        // pushing us immediately over to LARGE
         BOOST_CHECK_EQUAL(
-            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/0),
-            CoinsCacheSizeState::OK);
+            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/ 0),
+            CoinsCacheSizeState::LARGE);
     }
 
     // Adding some additional coins will push us over the edge to CRITICAL.
@@ -94,16 +102,16 @@ BOOST_AUTO_TEST_CASE(getcoinscachesizestate)
         chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/0),
         CoinsCacheSizeState::CRITICAL);
 
-    // Passing non-zero max mempool usage should allow us more headroom.
+    // Passing non-zero max mempool usage (512 KiB) should allow us more headroom.
     BOOST_CHECK_EQUAL(
-        chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/1 << 10),
+        chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/ 1 << 19),
         CoinsCacheSizeState::OK);
 
     for (int i{0}; i < 3; ++i) {
         AddTestCoin(view);
         print_view_mem_usage(view);
         BOOST_CHECK_EQUAL(
-            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/1 << 10),
+            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes=*/ 1 << 19),
             CoinsCacheSizeState::OK);
     }
 
@@ -119,7 +127,7 @@ BOOST_AUTO_TEST_CASE(getcoinscachesizestate)
         BOOST_CHECK(usage_percentage >= 0.9);
         BOOST_CHECK(usage_percentage < 1);
         BOOST_CHECK_EQUAL(
-            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, 1 << 10),
+            chainstate.GetCoinsCacheSizeState(MAX_COINS_CACHE_BYTES, /*max_mempool_size_bytes*/ 1 << 10), // 1024
             CoinsCacheSizeState::LARGE);
     }