Merge pull request #8678

f698f2b Refactored rx-slow-hash.c (SChernykh)

CMakeLists.txt

@@ -794,7 +794,7 @@ else()
     set(USE_LTO_DEFAULT false)
     set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--stack,10485760")
     if(NOT BUILD_64)
-      add_definitions(-DWINVER=0x0501 -D_WIN32_WINNT=0x0501)
+      add_definitions(-DWINVER=0x0600 -D_WIN32_WINNT=0x0600)
     endif()
   endif()
   set(C_WARNINGS "-Waggregate-return -Wnested-externs -Wold-style-definition -Wstrict-prototypes")

src/crypto/c_threads.h

@@ -30,29 +30,39 @@
 #pragma once
 
 #ifdef _WIN32
+
 #include <windows.h>
-#define CTHR_MUTEX_TYPE	HANDLE
-#define CTHR_MUTEX_INIT	NULL
-#define CTHR_MUTEX_LOCK(x)	do { if (x == NULL) { \
-    HANDLE p = CreateMutex(NULL, FALSE, NULL); \
-    if (InterlockedCompareExchangePointer((PVOID*)&x, (PVOID)p, NULL) != NULL) \
-      CloseHandle(p); \
-  } WaitForSingleObject(x, INFINITE); } while(0)
-#define CTHR_MUTEX_UNLOCK(x)	ReleaseMutex(x)
+
+#define CTHR_RWLOCK_TYPE	SRWLOCK
+#define CTHR_RWLOCK_INIT	SRWLOCK_INIT
+#define CTHR_RWLOCK_LOCK_WRITE(x)	AcquireSRWLockExclusive(&x)
+#define CTHR_RWLOCK_UNLOCK_WRITE(x)	ReleaseSRWLockExclusive(&x)
+#define CTHR_RWLOCK_LOCK_READ(x)	AcquireSRWLockShared(&x)
+#define CTHR_RWLOCK_UNLOCK_READ(x)	ReleaseSRWLockShared(&x)
+#define CTHR_RWLOCK_TRYLOCK_READ(x)	TryAcquireSRWLockShared(&x)
+
 #define CTHR_THREAD_TYPE	HANDLE
 #define CTHR_THREAD_RTYPE	void
 #define CTHR_THREAD_RETURN	return
-#define CTHR_THREAD_CREATE(thr, func, arg)	thr = (HANDLE)_beginthread(func, 0, arg)
-#define CTHR_THREAD_JOIN(thr)			WaitForSingleObject(thr, INFINITE)
+#define CTHR_THREAD_CREATE(thr, func, arg)	((thr = (HANDLE)_beginthread(func, 0, arg)) != -1L)
+#define CTHR_THREAD_JOIN(thr)			WaitForSingleObject((HANDLE)thr, INFINITE)
+
 #else
+
 #include <pthread.h>
-#define CTHR_MUTEX_TYPE pthread_mutex_t
-#define CTHR_MUTEX_INIT	PTHREAD_MUTEX_INITIALIZER
-#define CTHR_MUTEX_LOCK(x)	pthread_mutex_lock(&x)
-#define CTHR_MUTEX_UNLOCK(x)	pthread_mutex_unlock(&x)
+
+#define CTHR_RWLOCK_TYPE	pthread_rwlock_t
+#define CTHR_RWLOCK_INIT	PTHREAD_RWLOCK_INITIALIZER
+#define CTHR_RWLOCK_LOCK_WRITE(x)	pthread_rwlock_wrlock(&x)
+#define CTHR_RWLOCK_UNLOCK_WRITE(x)	pthread_rwlock_unlock(&x)
+#define CTHR_RWLOCK_LOCK_READ(x)	pthread_rwlock_rdlock(&x)
+#define CTHR_RWLOCK_UNLOCK_READ(x)	pthread_rwlock_unlock(&x)
+#define CTHR_RWLOCK_TRYLOCK_READ(x)	(pthread_rwlock_tryrdlock(&x) == 0)
+
 #define CTHR_THREAD_TYPE pthread_t
 #define CTHR_THREAD_RTYPE	void *
 #define CTHR_THREAD_RETURN	return NULL
-#define CTHR_THREAD_CREATE(thr, func, arg)	pthread_create(&thr, NULL, func, arg)
+#define CTHR_THREAD_CREATE(thr, func, arg)	(pthread_create(&thr, NULL, func, arg) == 0)
 #define CTHR_THREAD_JOIN(thr)			pthread_join(thr, NULL)
+
 #endif

src/crypto/hash-ops.h

@@ -97,5 +97,9 @@ void rx_slow_hash_allocate_state(void);
 void rx_slow_hash_free_state(void);
 uint64_t rx_seedheight(const uint64_t height);
 void rx_seedheights(const uint64_t height, uint64_t *seed_height, uint64_t *next_height);
-void rx_slow_hash(const uint64_t mainheight, const uint64_t seedheight, const char *seedhash, const void *data, size_t length, char *hash, int miners, int is_alt);
-void rx_reorg(const uint64_t split_height);
+
+void rx_set_main_seedhash(const char *seedhash, size_t max_dataset_init_threads);
+void rx_slow_hash(const char *seedhash, const void *data, size_t length, char *result_hash);
+
+void rx_set_miner_thread(uint32_t value, size_t max_dataset_init_threads);
+uint32_t rx_get_miner_thread(void);

src/crypto/rx-slow-hash.c

@@ -43,32 +43,38 @@
 
 #define RX_LOGCAT	"randomx"
 
+static CTHR_RWLOCK_TYPE main_dataset_lock = CTHR_RWLOCK_INIT;
+static CTHR_RWLOCK_TYPE main_cache_lock = CTHR_RWLOCK_INIT;
+
+static randomx_dataset *main_dataset = NULL;
+static randomx_cache *main_cache = NULL;
+static char main_seedhash[HASH_SIZE];
+static int main_seedhash_set = 0;
+
+static CTHR_RWLOCK_TYPE secondary_cache_lock = CTHR_RWLOCK_INIT;
+
+static randomx_cache *secondary_cache = NULL;
+static char secondary_seedhash[HASH_SIZE];
+static int secondary_seedhash_set = 0;
+
 #if defined(_MSC_VER)
 #define THREADV __declspec(thread)
 #else
 #define THREADV __thread
 #endif
 
-typedef struct rx_state {
-  CTHR_MUTEX_TYPE rs_mutex;
-  char rs_hash[HASH_SIZE];
-  uint64_t  rs_height;
-  randomx_cache *rs_cache;
-} rx_state;
+static THREADV randomx_vm *main_vm_full = NULL;
+static THREADV randomx_vm *main_vm_light = NULL;
+static THREADV randomx_vm *secondary_vm_light = NULL;
 
-static CTHR_MUTEX_TYPE rx_mutex = CTHR_MUTEX_INIT;
-static CTHR_MUTEX_TYPE rx_dataset_mutex = CTHR_MUTEX_INIT;
+static THREADV uint32_t miner_thread = 0;
 
-static rx_state rx_s[2] = {{CTHR_MUTEX_INIT,{0},0,0},{CTHR_MUTEX_INIT,{0},0,0}};
-
-static randomx_dataset *rx_dataset;
-static int rx_dataset_nomem;
-static int rx_dataset_nolp;
-static uint64_t rx_dataset_height;
-static THREADV randomx_vm *rx_vm = NULL;
+static bool is_main(const char* seedhash) { return main_seedhash_set && (memcmp(seedhash, main_seedhash, HASH_SIZE) == 0); }
+static bool is_secondary(const char* seedhash) { return secondary_seedhash_set && (memcmp(seedhash, secondary_seedhash, HASH_SIZE) == 0); }
 
 static void local_abort(const char *msg)
 {
+  merror(RX_LOGCAT, "%s", msg);
   fprintf(stderr, "%s\n", msg);
 #ifdef NDEBUG
   _exit(1);
@@ -77,6 +83,16 @@ static void local_abort(const char *msg)
 #endif
 }
 
+static void hash2hex(const char* hash, char* hex) {
+  const char* d = "0123456789abcdef";
+  for (int i = 0; i < HASH_SIZE; ++i) {
+    const uint8_t b = hash[i];
+    hex[i * 2 + 0] = d[b >> 4];
+    hex[i * 2 + 1] = d[b & 15];
+  }
+  hex[HASH_SIZE * 2] = '\0';
+}
+
 static inline int disabled_flags(void) {
   static int flags = -1;
 
@@ -157,19 +173,6 @@ static unsigned int get_seedhash_epoch_blocks(void)
   return blocks;
 }
 
-void rx_reorg(const uint64_t split_height) {
-  int i;
-  CTHR_MUTEX_LOCK(rx_mutex);
-  for (i=0; i<2; i++) {
-    if (split_height <= rx_s[i].rs_height) {
-      if (rx_s[i].rs_height == rx_dataset_height)
-        rx_dataset_height = 1;
-      rx_s[i].rs_height = 1;	/* set to an invalid seed height */
-    }
-  }
-  CTHR_MUTEX_UNLOCK(rx_mutex);
-}
-
 uint64_t rx_seedheight(const uint64_t height) {
   const uint64_t seedhash_epoch_lag = get_seedhash_epoch_lag();
   const uint64_t seedhash_epoch_blocks = get_seedhash_epoch_blocks();
@@ -183,6 +186,95 @@ void rx_seedheights(const uint64_t height, uint64_t *seedheight, uint64_t *nexth
   *nextheight = rx_seedheight(height + get_seedhash_epoch_lag());
 }
 
+static void rx_alloc_dataset(randomx_flags flags, randomx_dataset** dataset, int ignore_env)
+{
+  if (*dataset) {
+    return;
+  }
+
+  if (disabled_flags() & RANDOMX_FLAG_FULL_MEM) {
+    static int shown = 0;
+    if (!shown) {
+      shown = 1;
+      minfo(RX_LOGCAT, "RandomX dataset is disabled by MONERO_RANDOMX_UMASK environment variable.");
+    }
+    return;
+  }
+
+  if (!ignore_env && !getenv("MONERO_RANDOMX_FULL_MEM")) {
+    static int shown = 0;
+    if (!shown) {
+      shown = 1;
+      minfo(RX_LOGCAT, "RandomX dataset is not enabled by default. Use MONERO_RANDOMX_FULL_MEM environment variable to enable it.");
+    }
+    return;
+  }
+
+  *dataset = randomx_alloc_dataset((flags | RANDOMX_FLAG_LARGE_PAGES) & ~disabled_flags());
+  if (!*dataset) {
+    mwarning(RX_LOGCAT, "Couldn't allocate RandomX dataset using large pages");
+    *dataset = randomx_alloc_dataset(flags & ~disabled_flags());
+    if (!*dataset) {
+      merror(RX_LOGCAT, "Couldn't allocate RandomX dataset");
+    }
+  }
+}
+
+static void rx_alloc_cache(randomx_flags flags, randomx_cache** cache)
+{
+  if (*cache) {
+    return;
+  }
+
+  *cache = randomx_alloc_cache((flags | RANDOMX_FLAG_LARGE_PAGES) & ~disabled_flags());
+  if (!*cache) {
+    mwarning(RX_LOGCAT, "Couldn't allocate RandomX cache using large pages");
+    *cache = randomx_alloc_cache(flags & ~disabled_flags());
+    if (!*cache) local_abort("Couldn't allocate RandomX cache");
+  }
+}
+
+static void rx_init_full_vm(randomx_flags flags, randomx_vm** vm)
+{
+  if (*vm || !main_dataset || (disabled_flags() & RANDOMX_FLAG_FULL_MEM)) {
+    return;
+  }
+
+  if ((flags & RANDOMX_FLAG_JIT) && !miner_thread) {
+    flags |= RANDOMX_FLAG_SECURE;
+  }
+
+  *vm = randomx_create_vm((flags | RANDOMX_FLAG_LARGE_PAGES | RANDOMX_FLAG_FULL_MEM) & ~disabled_flags(), NULL, main_dataset);
+  if (!*vm) {
+    mwarning(RX_LOGCAT, "Couldn't allocate RandomX full VM using large pages");
+    *vm = randomx_create_vm((flags | RANDOMX_FLAG_FULL_MEM) & ~disabled_flags(), NULL, main_dataset);
+    if (!*vm) {
+      merror(RX_LOGCAT, "Couldn't allocate RandomX full VM");
+    }
+  }
+}
+
+static void rx_init_light_vm(randomx_flags flags, randomx_vm** vm, randomx_cache* cache)
+{
+  if (*vm) {
+    randomx_vm_set_cache(*vm, cache);
+    return;
+  }
+
+  if ((flags & RANDOMX_FLAG_JIT) && !miner_thread) {
+    flags |= RANDOMX_FLAG_SECURE;
+  }
+
+  flags &= ~RANDOMX_FLAG_FULL_MEM;
+
+  *vm = randomx_create_vm((flags | RANDOMX_FLAG_LARGE_PAGES) & ~disabled_flags(), cache, NULL);
+  if (!*vm) {
+    mwarning(RX_LOGCAT, "Couldn't allocate RandomX light VM using large pages");
+    *vm = randomx_create_vm(flags & ~disabled_flags(), cache, NULL);
+    if (!*vm) local_abort("Couldn't allocate RandomX light VM");
+  }
+}
+
 typedef struct seedinfo {
   randomx_cache *si_cache;
   unsigned long si_start;
@@ -191,187 +283,230 @@ typedef struct seedinfo {
 
 static CTHR_THREAD_RTYPE rx_seedthread(void *arg) {
   seedinfo *si = arg;
-  randomx_init_dataset(rx_dataset, si->si_cache, si->si_start, si->si_count);
+  randomx_init_dataset(main_dataset, si->si_cache, si->si_start, si->si_count);
   CTHR_THREAD_RETURN;
 }
 
-static void rx_initdata(randomx_cache *rs_cache, const int miners, const uint64_t seedheight) {
-  if (miners > 1) {
-    unsigned long delta = randomx_dataset_item_count() / miners;
-    unsigned long start = 0;
-    int i;
-    seedinfo *si;
-    CTHR_THREAD_TYPE *st;
-    si = malloc(miners * sizeof(seedinfo));
-    if (si == NULL)
-      local_abort("Couldn't allocate RandomX mining threadinfo");
-    st = malloc(miners * sizeof(CTHR_THREAD_TYPE));
-    if (st == NULL) {
-      free(si);
-      local_abort("Couldn't allocate RandomX mining threadlist");
-    }
-    for (i=0; i<miners-1; i++) {
-      si[i].si_cache = rs_cache;
-      si[i].si_start = start;
-      si[i].si_count = delta;
-      start += delta;
-    }
-    si[i].si_cache = rs_cache;
+static void rx_init_dataset(size_t max_threads) {
+  if (!main_dataset) {
+    return;
+  }
+
+  // leave 2 CPU cores for other tasks
+  const size_t num_threads = (max_threads < 4) ? 1 : (max_threads - 2);
+  seedinfo* si = malloc(num_threads * sizeof(seedinfo));
+  if (!si) local_abort("Couldn't allocate RandomX mining threadinfo");
+
+  const uint32_t delta = randomx_dataset_item_count() / num_threads;
+  uint32_t start = 0;
+
+  const size_t n1 = num_threads - 1;
+  for (size_t i = 0; i < n1; ++i) {
+    si[i].si_cache = main_cache;
     si[i].si_start = start;
-    si[i].si_count = randomx_dataset_item_count() - start;
-    for (i=1; i<miners; i++) {
-      CTHR_THREAD_CREATE(st[i], rx_seedthread, &si[i]);
-    }
-    randomx_init_dataset(rx_dataset, rs_cache, 0, si[0].si_count);
-    for (i=1; i<miners; i++) {
-      CTHR_THREAD_JOIN(st[i]);
+    si[i].si_count = delta;
+    start += delta;
+  }
+
+  si[n1].si_cache = main_cache;
+  si[n1].si_start = start;
+  si[n1].si_count = randomx_dataset_item_count() - start;
+
+  CTHR_THREAD_TYPE *st = malloc(num_threads * sizeof(CTHR_THREAD_TYPE));
+  if (!st) local_abort("Couldn't allocate RandomX mining threadlist");
+
+  CTHR_RWLOCK_LOCK_READ(main_cache_lock);
+  for (size_t i = 0; i < n1; ++i) {
+    if (!CTHR_THREAD_CREATE(st[i], rx_seedthread, &si[i])) {
+      local_abort("Couldn't start RandomX seed thread");
     }
-    free(st);
-    free(si);
-  } else {
-    randomx_init_dataset(rx_dataset, rs_cache, 0, randomx_dataset_item_count());
   }
-  rx_dataset_height = seedheight;
+  rx_seedthread(&si[n1]);
+  for (size_t i = 0; i < n1; ++i) CTHR_THREAD_JOIN(st[i]);
+  CTHR_RWLOCK_UNLOCK_READ(main_cache_lock);
+
+  free(st);
+  free(si);
+
+  minfo(RX_LOGCAT, "RandomX dataset initialized");
 }
 
-void rx_slow_hash(const uint64_t mainheight, const uint64_t seedheight, const char *seedhash, const void *data, size_t length,
-  char *hash, int miners, int is_alt) {
-  uint64_t s_height = rx_seedheight(mainheight);
-  int toggle = (s_height & get_seedhash_epoch_blocks()) != 0;
-  randomx_flags flags = enabled_flags() & ~disabled_flags();
-  rx_state *rx_sp;
-  randomx_cache *cache;
-
-  CTHR_MUTEX_LOCK(rx_mutex);
-
-  /* if alt block but with same seed as mainchain, no need for alt cache */
-  if (is_alt) {
-    if (s_height == seedheight && !memcmp(rx_s[toggle].rs_hash, seedhash, HASH_SIZE))
-      is_alt = 0;
-  } else {
-  /* RPC could request an earlier block on mainchain */
-    if (s_height > seedheight)
-      is_alt = 1;
-    /* miner can be ahead of mainchain */
-    else if (s_height < seedheight)
-      toggle ^= 1;
-  }
-
-  toggle ^= (is_alt != 0);
-
-  rx_sp = &rx_s[toggle];
-  CTHR_MUTEX_LOCK(rx_sp->rs_mutex);
-  CTHR_MUTEX_UNLOCK(rx_mutex);
-
-  cache = rx_sp->rs_cache;
-  if (cache == NULL) {
-    if (!(disabled_flags() & RANDOMX_FLAG_LARGE_PAGES)) {
-      cache = randomx_alloc_cache(flags | RANDOMX_FLAG_LARGE_PAGES);
-      if (cache == NULL) {
-        mdebug(RX_LOGCAT, "Couldn't use largePages for RandomX cache");
-      }
-    }
-    if (cache == NULL) {
-      cache = randomx_alloc_cache(flags);
-      if (cache == NULL)
-        local_abort("Couldn't allocate RandomX cache");
-    }
+typedef struct thread_info {
+  char seedhash[HASH_SIZE];
+  size_t max_threads;
+} thread_info;
+
+static CTHR_THREAD_RTYPE rx_set_main_seedhash_thread(void *arg) {
+  thread_info* info = arg;
+
+  CTHR_RWLOCK_LOCK_WRITE(main_dataset_lock);
+  CTHR_RWLOCK_LOCK_WRITE(main_cache_lock);
+
+  // Double check that seedhash wasn't already updated
+  if (is_main(info->seedhash)) {
+    CTHR_RWLOCK_UNLOCK_WRITE(main_cache_lock);
+    CTHR_RWLOCK_UNLOCK_WRITE(main_dataset_lock);
+    free(info);
+    CTHR_THREAD_RETURN;
   }
-  if (rx_sp->rs_height != seedheight || rx_sp->rs_cache == NULL || memcmp(seedhash, rx_sp->rs_hash, HASH_SIZE)) {
-    randomx_init_cache(cache, seedhash, HASH_SIZE);
-    rx_sp->rs_cache = cache;
-    rx_sp->rs_height = seedheight;
-    memcpy(rx_sp->rs_hash, seedhash, HASH_SIZE);
+  memcpy(main_seedhash, info->seedhash, HASH_SIZE);
+  main_seedhash_set = 1;
+
+  char buf[HASH_SIZE * 2 + 1];
+  hash2hex(main_seedhash, buf);
+  minfo(RX_LOGCAT, "RandomX new main seed hash is %s", buf);
+
+  const randomx_flags flags = enabled_flags() & ~disabled_flags();
+  rx_alloc_dataset(flags, &main_dataset, 0);
+  rx_alloc_cache(flags, &main_cache);
+
+  randomx_init_cache(main_cache, info->seedhash, HASH_SIZE);
+  minfo(RX_LOGCAT, "RandomX main cache initialized");
+
+  CTHR_RWLOCK_UNLOCK_WRITE(main_cache_lock);
+
+  // From this point, rx_slow_hash can calculate hashes in light mode, but dataset is not initialized yet
+  rx_init_dataset(info->max_threads);
+
+  CTHR_RWLOCK_UNLOCK_WRITE(main_dataset_lock);
+
+  free(info);
+  CTHR_THREAD_RETURN;
+}
+
+void rx_set_main_seedhash(const char *seedhash, size_t max_dataset_init_threads) {
+  // Early out if seedhash didn't change
+  if (is_main(seedhash)) {
+    return;
   }
-  if (rx_vm == NULL) {
-    if ((flags & RANDOMX_FLAG_JIT) && !miners) {
-        flags |= RANDOMX_FLAG_SECURE & ~disabled_flags();
-    }
-    if (miners && (disabled_flags() & RANDOMX_FLAG_FULL_MEM)) {
-      miners = 0;
-    }
-    if (miners) {
-      CTHR_MUTEX_LOCK(rx_dataset_mutex);
-      if (!rx_dataset_nomem) {
-        if (rx_dataset == NULL) {
-          if (!(disabled_flags() & RANDOMX_FLAG_LARGE_PAGES)) {
-            rx_dataset = randomx_alloc_dataset(RANDOMX_FLAG_LARGE_PAGES);
-            if (rx_dataset == NULL) {
-              mdebug(RX_LOGCAT, "Couldn't use largePages for RandomX dataset");
-            }
-          }
-          if (rx_dataset == NULL)
-            rx_dataset = randomx_alloc_dataset(RANDOMX_FLAG_DEFAULT);
-          if (rx_dataset != NULL)
-            rx_initdata(rx_sp->rs_cache, miners, seedheight);
-        }
-      }
-      if (rx_dataset != NULL)
-        flags |= RANDOMX_FLAG_FULL_MEM;
-      else {
-        miners = 0;
-        if (!rx_dataset_nomem) {
-          rx_dataset_nomem = 1;
-          mwarning(RX_LOGCAT, "Couldn't allocate RandomX dataset for miner");
+
+  // Update main cache and dataset in the background
+  thread_info* info = malloc(sizeof(thread_info));
+  if (!info) local_abort("Couldn't allocate RandomX mining threadinfo");
+
+  memcpy(info->seedhash, seedhash, HASH_SIZE);
+  info->max_threads = max_dataset_init_threads;
+
+  CTHR_THREAD_TYPE t;
+  if (!CTHR_THREAD_CREATE(t, rx_set_main_seedhash_thread, info)) {
+    local_abort("Couldn't start RandomX seed thread");
+  }
+}
+
+void rx_slow_hash(const char *seedhash, const void *data, size_t length, char *result_hash) {
+  const randomx_flags flags = enabled_flags() & ~disabled_flags();
+  int success = 0;
+
+  // Fast path (seedhash == main_seedhash)
+  // Multiple threads can run in parallel in fast or light mode, 1-2 ms or 10-15 ms per hash per thread
+  if (is_main(seedhash)) {
+    // If CTHR_RWLOCK_TRYLOCK_READ fails it means dataset is being initialized now, so use the light mode
+    if (main_dataset && CTHR_RWLOCK_TRYLOCK_READ(main_dataset_lock)) {
+      // Double check that main_seedhash didn't change
+      if (is_main(seedhash)) {
+        rx_init_full_vm(flags, &main_vm_full);
+        if (main_vm_full) {
+          randomx_calculate_hash(main_vm_full, data, length, result_hash);
+          success = 1;
         }
       }
-      CTHR_MUTEX_UNLOCK(rx_dataset_mutex);
-    }
-    if (!(disabled_flags() & RANDOMX_FLAG_LARGE_PAGES) && !rx_dataset_nolp) {
-      rx_vm = randomx_create_vm(flags | RANDOMX_FLAG_LARGE_PAGES, rx_sp->rs_cache, rx_dataset);
-      if(rx_vm == NULL) { //large pages failed
-        mdebug(RX_LOGCAT, "Couldn't use largePages for RandomX VM");
-        rx_dataset_nolp = 1;
+      CTHR_RWLOCK_UNLOCK_READ(main_dataset_lock);
+    } else {
+      CTHR_RWLOCK_LOCK_READ(main_cache_lock);
+      // Double check that main_seedhash didn't change
+      if (is_main(seedhash)) {
+        rx_init_light_vm(flags, &main_vm_light, main_cache);
+        randomx_calculate_hash(main_vm_light, data, length, result_hash);
+        success = 1;
       }
+      CTHR_RWLOCK_UNLOCK_READ(main_cache_lock);
     }
-    if (rx_vm == NULL)
-      rx_vm = randomx_create_vm(flags, rx_sp->rs_cache, rx_dataset);
-    if(rx_vm == NULL) {//fallback if everything fails
-      flags = RANDOMX_FLAG_DEFAULT | (miners ? RANDOMX_FLAG_FULL_MEM : 0);
-      rx_vm = randomx_create_vm(flags, rx_sp->rs_cache, rx_dataset);
-    }
-    if (rx_vm == NULL)
-      local_abort("Couldn't allocate RandomX VM");
-  } else if (miners) {
-    CTHR_MUTEX_LOCK(rx_dataset_mutex);
-    if (rx_dataset != NULL && rx_dataset_height != seedheight)
-      rx_initdata(cache, miners, seedheight);
-    else if (rx_dataset == NULL) {
-      /* this is a no-op if the cache hasn't changed */
-      randomx_vm_set_cache(rx_vm, rx_sp->rs_cache);
+  }
+
+  if (success) {
+    return;
+  }
+
+  char buf[HASH_SIZE * 2 + 1];
+
+  // Slow path (seedhash != main_seedhash, but seedhash == secondary_seedhash)
+  // Multiple threads can run in parallel in light mode, 10-15 ms per hash per thread
+  if (!secondary_cache) {
+    CTHR_RWLOCK_LOCK_WRITE(secondary_cache_lock);
+    if (!secondary_cache) {
+      hash2hex(seedhash, buf);
+      minfo(RX_LOGCAT, "RandomX new secondary seed hash is %s", buf);
+
+      rx_alloc_cache(flags, &secondary_cache);
+      randomx_init_cache(secondary_cache, seedhash, HASH_SIZE);
+      minfo(RX_LOGCAT, "RandomX secondary cache updated");
+      memcpy(secondary_seedhash, seedhash, HASH_SIZE);
+      secondary_seedhash_set = 1;
     }
-    CTHR_MUTEX_UNLOCK(rx_dataset_mutex);
-  } else {
-    /* this is a no-op if the cache hasn't changed */
-    randomx_vm_set_cache(rx_vm, rx_sp->rs_cache);
-  }
-  /* mainchain users can run in parallel */
-  if (!is_alt)
-    CTHR_MUTEX_UNLOCK(rx_sp->rs_mutex);
-  randomx_calculate_hash(rx_vm, data, length, hash);
-  /* altchain slot users always get fully serialized */
-  if (is_alt)
-    CTHR_MUTEX_UNLOCK(rx_sp->rs_mutex);
-}
+    CTHR_RWLOCK_UNLOCK_WRITE(secondary_cache_lock);
+  }
 
-void rx_slow_hash_allocate_state(void) {
+  CTHR_RWLOCK_LOCK_READ(secondary_cache_lock);
+  if (is_secondary(seedhash)) {
+    rx_init_light_vm(flags, &secondary_vm_light, secondary_cache);
+    randomx_calculate_hash(secondary_vm_light, data, length, result_hash);
+    success = 1;
+  }
+  CTHR_RWLOCK_UNLOCK_READ(secondary_cache_lock);
+
+  if (success) {
+    return;
+  }
+
+  // Slowest path (seedhash != main_seedhash, seedhash != secondary_seedhash)
+  // Only one thread runs at a time and updates secondary_seedhash if needed, up to 200-500 ms per hash
+  CTHR_RWLOCK_LOCK_WRITE(secondary_cache_lock);
+  if (!is_secondary(seedhash)) {
+    hash2hex(seedhash, buf);
+    minfo(RX_LOGCAT, "RandomX new secondary seed hash is %s", buf);
+
+    randomx_init_cache(secondary_cache, seedhash, HASH_SIZE);
+    minfo(RX_LOGCAT, "RandomX secondary cache updated");
+    memcpy(secondary_seedhash, seedhash, HASH_SIZE);
+    secondary_seedhash_set = 1;
+  }
+  rx_init_light_vm(flags, &secondary_vm_light, secondary_cache);
+  randomx_calculate_hash(secondary_vm_light, data, length, result_hash);
+  CTHR_RWLOCK_UNLOCK_WRITE(secondary_cache_lock);
 }
 
-void rx_slow_hash_free_state(void) {
-  if (rx_vm != NULL) {
-    randomx_destroy_vm(rx_vm);
-    rx_vm = NULL;
+void rx_set_miner_thread(uint32_t value, size_t max_dataset_init_threads) {
+  miner_thread = value;
+
+  // If dataset is not allocated yet, try to allocate and initialize it
+  CTHR_RWLOCK_LOCK_WRITE(main_dataset_lock);
+  if (main_dataset) {
+    CTHR_RWLOCK_UNLOCK_WRITE(main_dataset_lock);
+    return;
   }
+
+  const randomx_flags flags = enabled_flags() & ~disabled_flags();
+  rx_alloc_dataset(flags, &main_dataset, 1);
+  rx_init_dataset(max_dataset_init_threads);
+
+  CTHR_RWLOCK_UNLOCK_WRITE(main_dataset_lock);
+}
+
+uint32_t rx_get_miner_thread() {
+  return miner_thread;
 }
 
-void rx_stop_mining(void) {
-  CTHR_MUTEX_LOCK(rx_dataset_mutex);
-  if (rx_dataset != NULL) {
-    randomx_dataset *rd = rx_dataset;
-    rx_dataset = NULL;
-    randomx_release_dataset(rd);
+void rx_slow_hash_allocate_state() {}
+
+static void rx_destroy_vm(randomx_vm** vm) {
+  if (*vm) {
+    randomx_destroy_vm(*vm);
+    *vm = NULL;
   }
-  rx_dataset_nomem = 0;
-  rx_dataset_nolp = 0;
-  CTHR_MUTEX_UNLOCK(rx_dataset_mutex);
+}
+
+void rx_slow_hash_free_state() {
+  rx_destroy_vm(&main_vm_full);
+  rx_destroy_vm(&main_vm_light);
+  rx_destroy_vm(&secondary_vm_light);
 }

src/cryptonote_basic/miner.cpp

@@ -82,6 +82,7 @@
 using namespace epee;
 
 #include "miner.h"
+#include "crypto/hash.h"
 
 
 extern "C" void slow_hash_allocate_state();
@@ -436,7 +437,6 @@ namespace cryptonote
   {
     m_stop = true;
   }
-  extern "C" void rx_stop_mining(void);
   //-----------------------------------------------------------------------------------------------------
   bool miner::stop()
   {
@@ -469,7 +469,6 @@ namespace cryptonote
     MINFO("Mining has been stopped, " << m_threads.size() << " finished" );
     m_threads.clear();
     m_threads_autodetect.clear();
-    rx_stop_mining();
     return true;
   }
   //-----------------------------------------------------------------------------------------------------
@@ -524,6 +523,8 @@ namespace cryptonote
   bool miner::worker_thread()
   {
     const uint32_t th_local_index = m_thread_index++; // atomically increment, getting value before increment
+    crypto::rx_set_miner_thread(th_local_index, tools::get_max_concurrency());
+
     MLOG_SET_THREAD_NAME(std::string("[miner ") + std::to_string(th_local_index) + "]");
     MGINFO("Miner thread was started ["<< th_local_index << "]");
     uint32_t nonce = m_starter_nonce + th_local_index;

src/cryptonote_core/blockchain.cpp

@@ -456,6 +456,14 @@ bool Blockchain::init(BlockchainDB* db, const network_type nettype, bool offline
     if (!update_next_cumulative_weight_limit())
       return false;
   }
+
+  if (m_hardfork->get_current_version() >= RX_BLOCK_VERSION)
+  {
+    const crypto::hash seedhash = get_block_id_by_height(crypto::rx_seedheight(m_db->height()));
+    if (seedhash != crypto::null_hash)
+      rx_set_main_seedhash(seedhash.data, tools::get_max_concurrency());
+  }
+
   return true;
 }
 //------------------------------------------------------------------
@@ -570,6 +578,12 @@ void Blockchain::pop_blocks(uint64_t nblocks)
 
   if (stop_batch)
     m_db->batch_stop();
+
+  if (m_hardfork->get_current_version() >= RX_BLOCK_VERSION)
+  {
+    const crypto::hash seedhash = get_block_id_by_height(crypto::rx_seedheight(m_db->height()));
+    rx_set_main_seedhash(seedhash.data, tools::get_max_concurrency());
+  }
 }
 //------------------------------------------------------------------
 // This function tells BlockchainDB to remove the top block from the
@@ -1239,18 +1253,20 @@ bool Blockchain::switch_to_alternative_blockchain(std::list<block_extended_info>
   }
 
   m_hardfork->reorganize_from_chain_height(split_height);
-  get_block_longhash_reorg(split_height);
 
   std::shared_ptr<tools::Notify> reorg_notify = m_reorg_notify;
   if (reorg_notify)
     reorg_notify->notify("%s", std::to_string(split_height).c_str(), "%h", std::to_string(m_db->height()).c_str(),
         "%n", std::to_string(m_db->height() - split_height).c_str(), "%d", std::to_string(discarded_blocks).c_str(), NULL);
 
+  const uint64_t new_height = m_db->height();
+  const crypto::hash seedhash = get_block_id_by_height(crypto::rx_seedheight(new_height));
+
   crypto::hash prev_id;
   if (!get_block_hash(alt_chain.back().bl, prev_id))
     MERROR("Failed to get block hash of an alternative chain's tip");
   else
-    send_miner_notifications(prev_id, alt_chain.back().already_generated_coins);
+    send_miner_notifications(new_height, seedhash, prev_id, alt_chain.back().already_generated_coins);
 
   for (const auto& notifier : m_block_notifiers)
   {
@@ -1262,6 +1278,9 @@ bool Blockchain::switch_to_alternative_blockchain(std::list<block_extended_info>
     }
   }
 
+  if (m_hardfork->get_current_version() >= RX_BLOCK_VERSION)
+    rx_set_main_seedhash(seedhash.data, tools::get_max_concurrency());
+
   MGINFO_GREEN("REORGANIZE SUCCESS! on height: " << split_height << ", new blockchain size: " << m_db->height());
   return true;
 }
@@ -2001,7 +2020,7 @@ bool Blockchain::handle_alternative_block(const block& b, const crypto::hash& id
       {
         seedhash = get_block_id_by_height(seedheight);
       }
-      get_altblock_longhash(bei.bl, proof_of_work, get_current_blockchain_height(), bei.height, seedheight, seedhash);
+      get_altblock_longhash(bei.bl, proof_of_work, seedhash);
     } else
     {
       get_block_longhash(this, bei.bl, proof_of_work, bei.height, 0);
@@ -4552,11 +4571,15 @@ leave:
     }
   }
 
-  send_miner_notifications(id, already_generated_coins);
+  const crypto::hash seedhash = get_block_id_by_height(crypto::rx_seedheight(new_height));
+  send_miner_notifications(new_height, seedhash, id, already_generated_coins);
 
   for (const auto& notifier: m_block_notifiers)
     notifier(new_height - 1, {std::addressof(bl), 1});
 
+  if (m_hardfork->get_current_version() >= RX_BLOCK_VERSION)
+    rx_set_main_seedhash(seedhash.data, tools::get_max_concurrency());
+
   return true;
 }
 //------------------------------------------------------------------
@@ -5761,24 +5784,15 @@ void Blockchain::cache_block_template(const block &b, const cryptonote::account_
   m_btc_valid = true;
 }
 
-void Blockchain::send_miner_notifications(const crypto::hash &prev_id, uint64_t already_generated_coins)
+void Blockchain::send_miner_notifications(uint64_t height, const crypto::hash &seed_hash, const crypto::hash &prev_id, uint64_t already_generated_coins)
 {
   if (m_miner_notifiers.empty())
     return;
 
-  const uint64_t height = m_db->height();
   const uint8_t major_version = m_hardfork->get_ideal_version(height);
   const difficulty_type diff = get_difficulty_for_next_block();
   const uint64_t median_weight = m_current_block_cumul_weight_median;
 
-  crypto::hash seed_hash = crypto::null_hash;
-  if (m_hardfork->get_current_version() >= RX_BLOCK_VERSION)
-  {
-    uint64_t seed_height, next_height;
-    crypto::rx_seedheights(height, &seed_height, &next_height);
-    seed_hash = get_block_id_by_height(seed_height);
-  }
-
   std::vector<tx_block_template_backlog_entry> tx_backlog;
   m_tx_pool.get_block_template_backlog(tx_backlog);
 

src/cryptonote_core/blockchain.h

@@ -1598,9 +1598,11 @@ namespace cryptonote
     /**
      * @brief sends new block notifications to ZMQ `miner_data` subscribers
      *
+     * @param height current blockchain height
+     * @param seed_hash seed hash to use for mining
      * @param prev_id hash of new blockchain tip
      * @param already_generated_coins total coins mined by the network so far
      */
-    void send_miner_notifications(const crypto::hash &prev_id, uint64_t already_generated_coins);
+    void send_miner_notifications(uint64_t height, const crypto::hash &seed_hash, const crypto::hash &prev_id, uint64_t already_generated_coins);
   };
 }  // namespace cryptonote

src/cryptonote_core/cryptonote_tx_utils.cpp

@@ -669,10 +669,10 @@ namespace cryptonote
     return true;
   }
   //---------------------------------------------------------------
-  void get_altblock_longhash(const block& b, crypto::hash& res, const uint64_t main_height, const uint64_t height, const uint64_t seed_height, const crypto::hash& seed_hash)
+  void get_altblock_longhash(const block& b, crypto::hash& res, const crypto::hash& seed_hash)
   {
     blobdata bd = get_block_hashing_blob(b);
-    rx_slow_hash(main_height, seed_height, seed_hash.data, bd.data(), bd.size(), res.data, 0, 1);
+    rx_slow_hash(seed_hash.data, bd.data(), bd.size(), res.data);
   }
 
   bool get_block_longhash(const Blockchain *pbc, const blobdata& bd, crypto::hash& res, const uint64_t height, const int major_version, const crypto::hash *seed_hash, const int miners)
@@ -686,20 +686,16 @@ namespace cryptonote
     }
     if (major_version >= RX_BLOCK_VERSION)
     {
-      uint64_t seed_height, main_height;
       crypto::hash hash;
       if (pbc != NULL)
       {
-        seed_height = rx_seedheight(height);
+        const uint64_t seed_height = rx_seedheight(height);
         hash = seed_hash ? *seed_hash : pbc->get_pending_block_id_by_height(seed_height);
-        main_height = pbc->get_current_blockchain_height();
       } else
       {
         memset(&hash, 0, sizeof(hash));  // only happens when generating genesis block
-        seed_height = 0;
-        main_height = 0;
       }
-      rx_slow_hash(main_height, seed_height, hash.data, bd.data(), bd.size(), res.data, seed_hash ? 0 : miners, !!seed_hash);
+      rx_slow_hash(hash.data, bd.data(), bd.size(), res.data);
     } else {
       const int pow_variant = major_version >= 7 ? major_version - 6 : 0;
       crypto::cn_slow_hash(bd.data(), bd.size(), res, pow_variant, height);
@@ -713,20 +709,10 @@ namespace cryptonote
 	return get_block_longhash(pbc, bd, res, height, b.major_version, seed_hash, miners);
   }
 
-  bool get_block_longhash(const Blockchain *pbc, const block& b, crypto::hash& res, const uint64_t height, const int miners)
-  {
-    return get_block_longhash(pbc, b, res, height, NULL, miners);
-  }
-
-  crypto::hash get_block_longhash(const Blockchain *pbc, const block& b, const uint64_t height, const int miners)
+  crypto::hash get_block_longhash(const Blockchain *pbc, const block& b, const uint64_t height, const crypto::hash *seed_hash, const int miners)
   {
     crypto::hash p = crypto::null_hash;
-    get_block_longhash(pbc, b, p, height, miners);
+    get_block_longhash(pbc, b, p, height, seed_hash, miners);
     return p;
   }
-
-  void get_block_longhash_reorg(const uint64_t split_height)
-  {
-    rx_reorg(split_height);
-  }
 }

src/cryptonote_core/cryptonote_tx_utils.h

@@ -144,14 +144,10 @@ namespace cryptonote
     );
 
   class Blockchain;
-  bool get_block_longhash(const Blockchain *pb, const blobdata& bd, crypto::hash& res, const uint64_t height,
-    const int major_version, const crypto::hash *seed_hash, const int miners);
-  bool get_block_longhash(const Blockchain *pb, const block& b, crypto::hash& res, const uint64_t height, const int miners);
-  bool get_block_longhash(const Blockchain *pb, const block& b, crypto::hash& res, const uint64_t height, const crypto::hash *seed_hash, const int miners);
-  void get_altblock_longhash(const block& b, crypto::hash& res, const uint64_t main_height, const uint64_t height,
-    const uint64_t seed_height, const crypto::hash& seed_hash);
-  crypto::hash get_block_longhash(const Blockchain *pb, const block& b, const uint64_t height, const int miners);
-  void get_block_longhash_reorg(const uint64_t split_height);
+  bool get_block_longhash(const Blockchain *pb, const blobdata& bd, crypto::hash& res, const uint64_t height, const int major_version, const crypto::hash *seed_hash, const int miners = 0);
+  bool get_block_longhash(const Blockchain *pb, const block& b, crypto::hash& res, const uint64_t height, const crypto::hash *seed_hash = nullptr, const int miners = 0);
+  crypto::hash get_block_longhash(const Blockchain *pb, const block& b, const uint64_t height, const crypto::hash *seed_hash = nullptr, const int miners = 0);
+  void get_altblock_longhash(const block& b, crypto::hash& res, const crypto::hash& seed_hash);
 
 }
 

src/rpc/rpc_payment.cpp

@@ -236,10 +236,8 @@ namespace cryptonote
     *(uint32_t*)(hashing_blob.data() + 39) = SWAP32LE(nonce);
     if (block.major_version >= RX_BLOCK_VERSION)
     {
-      const uint64_t seed_height = is_current ? info.seed_height : info.previous_seed_height;
       const crypto::hash &seed_hash = is_current ? info.seed_hash : info.previous_seed_hash;
-      const uint64_t height = cryptonote::get_block_height(block);
-      crypto::rx_slow_hash(height, seed_height, seed_hash.data, hashing_blob.data(), hashing_blob.size(), hash.data, 0, 0);
+      crypto::rx_slow_hash(seed_hash.data, hashing_blob.data(), hashing_blob.size(), hash.data);
     }
     else
     {

src/wallet/wallet_rpc_payments.cpp

@@ -155,8 +155,7 @@ bool wallet2::search_for_rpc_payment(uint64_t credits_target, uint32_t n_threads
         const uint8_t major_version = hashing_blob[0];
         if (major_version >= RX_BLOCK_VERSION)
         {
-          const int miners = 1;
-          crypto::rx_slow_hash(height, seed_height, seed_hash.data, hashing_blob.data(), hashing_blob.size(), hash[i].data, miners, 0);
+          crypto::rx_slow_hash(seed_hash.data, hashing_blob.data(), hashing_blob.size(), hash[i].data);
         }
         else
         {