Better cache cleanup

src/crypto.cpp

@@ -158,6 +158,9 @@ class Cache
 {
 public:
 	Cache()
+		: derivations(new DerivationsMap())
+		, public_keys(new PublicKeysMap())
+		, tx_keys(new TxKeysMap())
 	{
 		uv_rwlock_init_checked(&derivations_lock);
 		uv_rwlock_init_checked(&public_keys_lock);
@@ -166,6 +169,10 @@ public:
 
 	~Cache()
 	{
+		delete derivations;
+		delete public_keys;
+		delete tx_keys;
+
 		uv_rwlock_destroy(&derivations_lock);
 		uv_rwlock_destroy(&public_keys_lock);
 		uv_rwlock_destroy(&tx_keys_lock);
@@ -180,8 +187,8 @@ public:
 		derivation = {};
 		{
 			ReadLock lock(derivations_lock);
-			auto it = derivations.find(index);
+			auto it = derivations->find(index);
-			if (it != derivations.end()) {
+			if (it != derivations->end()) {
 				const DerivationEntry& entry = it->second;
 				derivation = entry.m_derivation;
 				if (entry.find_view_tag(output_index, view_tag)) {
@@ -210,7 +217,7 @@ public:
 		{
 			WriteLock lock(derivations_lock);
 
-			DerivationEntry& entry = derivations.emplace(index, DerivationEntry{ derivation, {} }).first->second;
+			DerivationEntry& entry = derivations->emplace(index, DerivationEntry{ derivation, {} }).first->second;
 
 			const uint32_t k = static_cast<uint32_t>(output_index << 8) | view_tag;
 			if (std::find(entry.m_viewTags.begin(), entry.m_viewTags.end(), k) == entry.m_viewTags.end()) {
@@ -230,8 +237,8 @@ public:
 
 		{
 			ReadLock lock(public_keys_lock);
-			auto it = public_keys.find(index);
+			auto it = public_keys->find(index);
-			if (it != public_keys.end()) {
+			if (it != public_keys->end()) {
 				derived_key = it->second;
 				return true;
 			}
@@ -257,7 +264,7 @@ public:
 
 		{
 			WriteLock lock(public_keys_lock);
-			public_keys.emplace(index, derived_key);
+			public_keys->emplace(index, derived_key);
 		}
 
 		return true;
@@ -271,8 +278,8 @@ public:
 
 		{
 			ReadLock lock(tx_keys_lock);
-			auto it = tx_keys.find(index);
+			auto it = tx_keys->find(index);
-			if (it != tx_keys.end()) {
+			if (it != tx_keys->end()) {
 				pub = it->second.first;
 				sec = it->second.second;
 				return;
@@ -291,15 +298,27 @@ public:
 
 		{
 			WriteLock lock(tx_keys_lock);
-			tx_keys.emplace(index, std::pair<hash, hash>(pub, sec));
+			tx_keys->emplace(index, std::pair<hash, hash>(pub, sec));
 		}
 	}
 
 	void clear()
 	{
-		{ WriteLock lock(derivations_lock); derivations.clear(); }
+		{
-		{ WriteLock lock(public_keys_lock); public_keys.clear(); }
+			WriteLock lock(derivations_lock);
-		{ WriteLock lock(tx_keys_lock); tx_keys.clear(); }
+			delete derivations;
+			derivations = new DerivationsMap();
+		}
+		{
+			WriteLock lock(public_keys_lock);
+			delete public_keys;
+			public_keys = new PublicKeysMap();
+		}
+		{
+			WriteLock lock(tx_keys_lock);
+			delete tx_keys;
+			tx_keys = new TxKeysMap();
+		}
 	}
 
 private:
@@ -319,14 +338,18 @@ private:
 		}
 	};
 
+	typedef unordered_map<std::array<uint8_t, HASH_SIZE * 2>, DerivationEntry> DerivationsMap;
+	typedef unordered_map<std::array<uint8_t, HASH_SIZE * 2 + sizeof(size_t)>, hash> PublicKeysMap;
+	typedef unordered_map<std::array<uint8_t, HASH_SIZE * 2>, std::pair<hash, hash>> TxKeysMap;
+
 	uv_rwlock_t derivations_lock;
-	unordered_map<std::array<uint8_t, HASH_SIZE * 2>, DerivationEntry> derivations;
+	DerivationsMap* derivations;
 
 	uv_rwlock_t public_keys_lock;
-	unordered_map<std::array<uint8_t, HASH_SIZE * 2 + sizeof(size_t)>, hash> public_keys;
+	PublicKeysMap* public_keys;
 
 	uv_rwlock_t tx_keys_lock;
-	unordered_map<std::array<uint8_t, HASH_SIZE * 2>, std::pair<hash, hash>> tx_keys;
+	TxKeysMap* tx_keys;
 };
 
 static Cache* cache = nullptr;
@@ -379,7 +402,9 @@ void destroy_crypto_cache()
 
 void clear_crypto_cache()
 {
-	cache->clear();
+	if (cache) {
+		cache->clear();
+	}
 }
 
 } // namespace p2pool

src/memory_leak_debug.cpp

@@ -32,15 +32,48 @@ namespace p2pool {
 
 static bool track_memory = false;
 
-constexpr size_t N = 2097152;
+constexpr size_t N = 1 << 22;
-constexpr size_t MAX_FRAMES = 30;
+constexpr size_t MAX_FRAMES = 29;
 
 struct TrackedAllocation
 {
 	void* p;
 	void* stack_trace[MAX_FRAMES];
+	uint64_t allocated_size;
 	uint32_t thread_id;
-	uint32_t allocated_size;
+
+	FORCEINLINE bool operator<(const TrackedAllocation& rhs) { return memcmp(stack_trace, rhs.stack_trace, sizeof(stack_trace)) < 0; }
+	FORCEINLINE bool operator==(const TrackedAllocation& rhs) { return memcmp(stack_trace, rhs.stack_trace, sizeof(stack_trace)) == 0; }
+
+	void print(HANDLE h) const
+	{
+		char buffer[sizeof(SYMBOL_INFO) + MAX_SYM_NAME * sizeof(TCHAR)] = {};
+		PSYMBOL_INFO pSymbol = reinterpret_cast<PSYMBOL_INFO>(buffer);
+
+		pSymbol->SizeOfStruct = sizeof(SYMBOL_INFO);
+		pSymbol->MaxNameLen = MAX_SYM_NAME;
+
+		IMAGEHLP_LINE64 line{};
+		line.SizeOfStruct = sizeof(IMAGEHLP_LINE64);
+
+		for (size_t j = 0; j < MAX_FRAMES; ++j) {
+			const DWORD64 address = reinterpret_cast<DWORD64>(stack_trace[j]);
+			DWORD64 t1 = 0;
+			DWORD t2 = 0;
+			if (SymFromAddr(h, address, &t1, pSymbol) && SymGetLineFromAddr64(h, address, &t2, &line)) {
+				const char* s = line.FileName;
+				const char* file_name = nullptr;
+				while (*s) {
+					if ((*s == '\\') || (*s == '/')) {
+						file_name = s + 1;
+					}
+					++s;
+				}
+				printf("%-25s %s (line %lu)\n", file_name ? file_name : line.FileName, pSymbol->Name, line.LineNumber);
+			}
+		}
+		printf("\n");
+	}
 };
 
 static_assert(sizeof(TrackedAllocation) == 256, "");
@@ -51,15 +84,62 @@ uint32_t first[N];
 uint32_t next[N];
 TrackedAllocation allocations[N];
 uint32_t num_allocations = 0;
+uint64_t total_allocated = 0;
 uint32_t cur_allocation_index = 1;
 
+void show_top_10()
+{
+	TrackedAllocation* buf = reinterpret_cast<TrackedAllocation*>(VirtualAlloc(nullptr, sizeof(TrackedAllocation) * N, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE));
+	if (!buf) {
+		return;
+	}
+
+	const HANDLE h = GetCurrentProcess();
+
+	{
+		p2pool::MutexLock lock(allocation_lock);
+
+		TrackedAllocation* end = buf;
+		for (size_t i = 0; i < N; ++i) {
+			if (allocations[i].allocated_size) {
+				*(end++) = allocations[i];
+			}
+		}
+
+		std::sort(buf, end);
+
+		TrackedAllocation* prev = buf;
+		for (TrackedAllocation* p = buf + 1; p < end; ++p) {
+			if (*p == *prev) {
+				prev->allocated_size += p->allocated_size;
+			}
+			else {
+				++prev;
+				*prev = *p;
+			}
+		}
+		end = prev + 1;
+
+		std::sort(buf, end, [](const auto& a, const auto& b) { return a.allocated_size > b.allocated_size; });
+
+		printf("%I64u total bytes allocated\n", total_allocated);
+
+		for (TrackedAllocation* p = buf; (p < buf + 10) && (p < end); ++p) {
+			printf("%I64u bytes allocated at:\n", p->allocated_size);
+			p->print(h);
+		}
+	}
+
+	VirtualFree(buf, 0, MEM_RELEASE);
+}
+
 FORCEINLINE static void add_alocation(void* p, size_t size)
 {
 	if (!track_memory) {
 		return;
 	}
 
-	void* stack_trace[MAX_FRAMES];
+	void* stack_trace[MAX_FRAMES] = {};
 	DWORD hash;
 	CaptureStackBackTrace(1, MAX_FRAMES, stack_trace, &hash);
 
@@ -74,6 +154,7 @@ FORCEINLINE static void add_alocation(void* p, size_t size)
 		// Make N two times bigger if this triggers
 		__debugbreak();
 	}
+	total_allocated += size;
 
 	for (uint64_t i = cur_allocation_index;; i = (i + 1) & (N - 1)) {
 		if (i && !allocations[i].allocated_size) {
@@ -105,6 +186,7 @@ FORCEINLINE static void remove_allocation(void* p)
 
 	for (uint32_t prev = 0, k = first[index]; k != 0; prev = k, k = next[k]) {
 		if (allocations[k].p == p) {
+			total_allocated -= allocations[k].allocated_size;
 			allocations[k].allocated_size = 0;
 			if (prev) {
 				next[prev] = next[k];
@@ -181,6 +263,8 @@ void* calloc_hook(size_t count, size_t size) noexcept
 
 void memory_tracking_start()
 {
+	SymInitialize(GetCurrentProcess(), NULL, TRUE);
+
 	using namespace p2pool;
 
 	uv_replace_allocator(malloc_hook, realloc_hook, calloc_hook, free_hook);
@@ -196,7 +280,6 @@ void memory_tracking_stop()
 	uv_mutex_destroy(&allocation_lock);
 
 	const HANDLE h = GetCurrentProcess();
-	SymInitialize(h, NULL, TRUE);
 
 	uint64_t total_leaks = 0;
 
@@ -205,39 +288,16 @@ void memory_tracking_stop()
 		if (t.allocated_size) {
 			total_leaks += t.allocated_size;
 
-			char buffer[sizeof(SYMBOL_INFO) + MAX_SYM_NAME * sizeof(TCHAR)] = {};
+			printf("Memory leak detected, %I64u bytes allocated at %p by thread %u:\n", t.allocated_size, t.p, t.thread_id);
-			PSYMBOL_INFO pSymbol = reinterpret_cast<PSYMBOL_INFO>(buffer);
+			t.print(h);
-
-			pSymbol->SizeOfStruct = sizeof(SYMBOL_INFO);
-			pSymbol->MaxNameLen = MAX_SYM_NAME;
-
-			IMAGEHLP_LINE64 line{};
-			line.SizeOfStruct = sizeof(IMAGEHLP_LINE64);
-
-			printf("Memory leak detected, %u bytes allocated at %p by thread %u:\n", t.allocated_size, t.p, t.thread_id);
-			for (size_t j = 0; j < MAX_FRAMES; ++j) {
-				const DWORD64 address = reinterpret_cast<DWORD64>(t.stack_trace[j]);
-				DWORD64 t1 = 0;
-				DWORD t2 = 0;
-				if (SymFromAddr(h, address, &t1, pSymbol) && SymGetLineFromAddr64(h, address, &t2, &line)) {
-					const char* s = line.FileName;
-					const char* file_name = nullptr;
-					while (*s) {
-						if ((*s == '\\') || (*s == '/')) {
-							file_name = s + 1;
-						}
-						++s;
-					}
-					printf("%-25s %s (line %lu)\n", file_name ? file_name : line.FileName, pSymbol->Name, line.LineNumber);
-				}
-			}
-			printf("\n");
 		}
 	}
 
 	if (total_leaks > 0) {
 		printf("%I64u bytes leaked\n\n", total_leaks);
 	}
+
+	SymCleanup(h);
 }
 
 NOINLINE void* operator new(size_t n) { return p2pool::allocate(n); }

src/p2pool.cpp

@@ -302,7 +302,9 @@ void p2pool::handle_miner_data(MinerData& data)
 	);
 
 	// Tx secret keys from all miners change every block, so cache can be cleared here
-	clear_crypto_cache();
+	if (m_sideChain->precalcFinished()) {
+		clear_crypto_cache();
+	}
 
 	if (!is_main_thread()) {
 		update_block_template_async();

src/side_chain.cpp

@@ -170,7 +170,10 @@ SideChain::SideChain(p2pool* pool, NetworkType type, const char* pool_name)
 
 	// Use between 1 and 8 threads
 	if (numThreads < 1) numThreads = 1;
+
+#ifndef _DEBUG
 	if (numThreads > 8) numThreads = 8;
+#endif
 
 	LOGINFO(4, "running " << numThreads << " pre-calculation workers");
 
@@ -2128,6 +2131,9 @@ void SideChain::finish_precalc()
 		LOGERR(1, "exception in finish_precalc(): " << e.what());
 	}
 
+	// Also clear cache because it has data from all old blocks now
+	clear_crypto_cache();
+
 #ifdef DEV_TEST_SYNC
 	if (m_pool) {
 		LOGINFO(0, log::LightGreen() << "[DEV] Synchronization finished successfully, stopping P2Pool now");