cleanup atomics

Author: bratpiorka

cmake/helpers.cmake

@@ -273,7 +273,14 @@ function(add_umf_target_compile_options name)
                     # disable warning 6326: Potential comparison of a constant
                     # with another constant
                     /wd6326
-                    # disable 4200 warning: nonstandard extension used:
+                    # disable warning 28112: a variable (ptr) which is accessed
+                    # via an Interlocked function must always be accessed via an
+                    # Interlocked function
+                    /wd28112
+                    # disable warning 4324: structure was padded due to
+                    # alignment specifier
+                    /wd4324
+                    # disable warning 4200: nonstandard extension used:
                     # zero-sized array in struct/union
                     /wd4200)
         if(UMF_DEVELOPER_MODE)

src/critnib/critnib.c

@@ -1,6 +1,6 @@
 /*
  *
- * Copyright (C) 2023-2024 Intel Corporation
+ * Copyright (C) 2023-2025 Intel Corporation
  *
  * Under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
@@ -119,6 +119,7 @@ struct critnib_leaf {
 
 struct critnib {
     struct critnib_node *root;
+    uint64_t remove_count;
 
     /* pool of freed nodes: singly linked list, next at child[0] */
     struct critnib_node *deleted_node;
@@ -128,29 +129,9 @@ struct critnib {
     struct critnib_node *pending_del_nodes[DELETED_LIFE];
     struct critnib_leaf *pending_del_leaves[DELETED_LIFE];
 
-    uint64_t remove_count;
-
     struct utils_mutex_t mutex; /* writes/removes */
 };
 
-/*
- * atomic load
- */
-static void load(void *src, void *dst) {
-    utils_atomic_load_acquire((word *)src, (word *)dst);
-}
-
-static void load64(uint64_t *src, uint64_t *dst) {
-    utils_atomic_load_acquire(src, dst);
-}
-
-/*
- * atomic store
- */
-static void store(void *dst, void *src) {
-    utils_atomic_store_release((word *)dst, (word)src);
-}
-
 /*
  * internal: is_leaf -- check tagged pointer for leafness
  */
@@ -180,7 +161,7 @@ static inline unsigned slice_index(word key, sh_t shift) {
  * critnib_new -- allocates a new critnib structure
  */
 struct critnib *critnib_new(void) {
-    struct critnib *c = umf_ba_global_alloc(sizeof(struct critnib));
+    struct critnib *c = umf_ba_global_aligned_alloc(sizeof(struct critnib), 8);
     if (!c) {
         return NULL;
     }
@@ -272,7 +253,7 @@ static void free_node(struct critnib *__restrict c,
  */
 static struct critnib_node *alloc_node(struct critnib *__restrict c) {
     if (!c->deleted_node) {
-        return umf_ba_global_alloc(sizeof(struct critnib_node));
+        return umf_ba_global_aligned_alloc(sizeof(struct critnib_node), 8);
     }
 
     struct critnib_node *n = c->deleted_node;
@@ -303,7 +284,7 @@ static void free_leaf(struct critnib *__restrict c,
  */
 static struct critnib_leaf *alloc_leaf(struct critnib *__restrict c) {
     if (!c->deleted_leaf) {
-        return umf_ba_global_alloc(sizeof(struct critnib_leaf));
+        return umf_ba_global_aligned_alloc(sizeof(struct critnib_leaf), 8);
     }
 
     struct critnib_leaf *k = c->deleted_leaf;
@@ -343,10 +324,8 @@ int critnib_insert(struct critnib *c, word key, void *value, int update) {
 
     struct critnib_node *n = c->root;
     if (!n) {
-        store(&c->root, kn);
-
+        utils_atomic_store_release_ptr((void **)&c->root, kn);
         utils_mutex_unlock(&c->mutex);
-
         return 0;
     }
 
@@ -361,7 +340,8 @@ int critnib_insert(struct critnib *c, word key, void *value, int update) {
 
     if (!n) {
         n = prev;
-        store(&n->child[slice_index(key, n->shift)], kn);
+        utils_atomic_store_release_ptr(
+            (void **)&n->child[slice_index(key, n->shift)], kn);
 
         utils_mutex_unlock(&c->mutex);
 
@@ -406,7 +386,7 @@ int critnib_insert(struct critnib *c, word key, void *value, int update) {
     m->child[slice_index(path, sh)] = n;
     m->shift = sh;
     m->path = key & path_mask(sh);
-    store(parent, m);
+    utils_atomic_store_release_ptr((void **)parent, m);
 
     utils_mutex_unlock(&c->mutex);
 
@@ -427,7 +407,8 @@ void *critnib_remove(struct critnib *c, word key) {
         goto not_found;
     }
 
-    word del = (utils_atomic_increment(&c->remove_count) - 1) % DELETED_LIFE;
+    word del =
+        (utils_atomic_increment_u64(&c->remove_count) - 1) % DELETED_LIFE;
     free_node(c, c->pending_del_nodes[del]);
     free_leaf(c, c->pending_del_leaves[del]);
     c->pending_del_nodes[del] = NULL;
@@ -436,7 +417,7 @@ void *critnib_remove(struct critnib *c, word key) {
     if (is_leaf(n)) {
         k = to_leaf(n);
         if (k->key == key) {
-            store(&c->root, NULL);
+            utils_atomic_store_release_ptr((void **)&c->root, NULL);
             goto del_leaf;
         }
 
@@ -466,7 +447,8 @@ void *critnib_remove(struct critnib *c, word key) {
         goto not_found;
     }
 
-    store(&n->child[slice_index(key, n->shift)], NULL);
+    utils_atomic_store_release_ptr(
+        (void **)&n->child[slice_index(key, n->shift)], NULL);
 
     /* Remove the node if there's only one remaining child. */
     int ochild = -1;
@@ -482,7 +464,7 @@ void *critnib_remove(struct critnib *c, word key) {
 
     ASSERTne(ochild, -1);
 
-    store(n_parent, n->child[ochild]);
+    utils_atomic_store_release_ptr((void **)n_parent, n->child[ochild]);
     c->pending_del_nodes[del] = n;
 
 del_leaf:
@@ -511,22 +493,23 @@ void *critnib_get(struct critnib *c, word key) {
     do {
         struct critnib_node *n;
 
-        load64(&c->remove_count, &wrs1);
-        load(&c->root, &n);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs1);
+        utils_atomic_load_acquire_ptr((void **)&c->root, (void **)&n);
 
         /*
 		 * critbit algorithm: dive into the tree, looking at nothing but
 		 * each node's critical bit^H^H^Hnibble.  This means we risk
 		 * going wrong way if our path is missing, but that's ok...
 		 */
         while (n && !is_leaf(n)) {
-            load(&n->child[slice_index(key, n->shift)], &n);
+            utils_atomic_load_acquire_ptr(
+                (void **)&n->child[slice_index(key, n->shift)], (void **)&n);
         }
 
         /* ... as we check it at the end. */
         struct critnib_leaf *k = to_leaf(n);
         res = (n && k->key == key) ? k->value : NULL;
-        load64(&c->remove_count, &wrs2);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs2);
     } while (wrs1 + DELETED_LIFE <= wrs2);
 
     return res;
@@ -597,7 +580,7 @@ static struct critnib_leaf *find_le(struct critnib_node *__restrict n,
     /* recursive call: follow the path */
     {
         struct critnib_node *m;
-        load(&n->child[nib], &m);
+        utils_atomic_load_acquire_ptr((void **)&n->child[nib], (void **)&m);
         struct critnib_leaf *k = find_le(m, key);
         if (k) {
             return k;
@@ -611,7 +594,7 @@ static struct critnib_leaf *find_le(struct critnib_node *__restrict n,
 	 */
     for (; nib > 0; nib--) {
         struct critnib_node *m;
-        load(&n->child[nib - 1], &m);
+        utils_atomic_load_acquire_ptr((void **)&n->child[nib - 1], (void **)&m);
         if (m) {
             n = m;
             if (is_leaf(n)) {
@@ -635,12 +618,12 @@ void *critnib_find_le(struct critnib *c, word key) {
     void *res;
 
     do {
-        load64(&c->remove_count, &wrs1);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs1);
         struct critnib_node *n; /* avoid a subtle TOCTOU */
-        load(&c->root, &n);
+        utils_atomic_load_acquire_ptr((void **)&c->root, (void **)&n);
         struct critnib_leaf *k = n ? find_le(n, key) : NULL;
         res = k ? k->value : NULL;
-        load64(&c->remove_count, &wrs2);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs2);
     } while (wrs1 + DELETED_LIFE <= wrs2);
 
     return res;
@@ -694,7 +677,7 @@ static struct critnib_leaf *find_ge(struct critnib_node *__restrict n,
     unsigned nib = slice_index(key, n->shift);
     {
         struct critnib_node *m;
-        load(&n->child[nib], &m);
+        utils_atomic_load_acquire_ptr((void **)&n->child[nib], (void **)&m);
         struct critnib_leaf *k = find_ge(m, key);
         if (k) {
             return k;
@@ -703,7 +686,7 @@ static struct critnib_leaf *find_ge(struct critnib_node *__restrict n,
 
     for (; nib < NIB; nib++) {
         struct critnib_node *m;
-        load(&n->child[nib + 1], &m);
+        utils_atomic_load_acquire_ptr((void **)&n->child[nib + 1], (void **)&m);
         if (m) {
             n = m;
             if (is_leaf(n)) {
@@ -741,17 +724,19 @@ int critnib_find(struct critnib *c, uintptr_t key, enum find_dir_t dir,
     }
 
     do {
-        load64(&c->remove_count, &wrs1);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs1);
         struct critnib_node *n;
-        load(&c->root, &n);
+        utils_atomic_load_acquire_ptr((void **)&c->root, (void **)&n);
 
         if (dir < 0) {
             k = find_le(n, key);
         } else if (dir > 0) {
             k = find_ge(n, key);
         } else {
             while (n && !is_leaf(n)) {
-                load(&n->child[slice_index(key, n->shift)], &n);
+                utils_atomic_load_acquire_ptr(
+                    (void **)&n->child[slice_index(key, n->shift)],
+                    (void **)&n);
             }
 
             struct critnib_leaf *kk = to_leaf(n);
@@ -761,7 +746,7 @@ int critnib_find(struct critnib *c, uintptr_t key, enum find_dir_t dir,
             _rkey = k->key;
             _rvalue = k->value;
         }
-        load64(&c->remove_count, &wrs2);
+        utils_atomic_load_acquire_u64(&c->remove_count, &wrs2);
     } while (wrs1 + DELETED_LIFE <= wrs2);
 
     if (k) {

src/ipc_cache.c

@@ -232,7 +232,7 @@ umf_result_t umfIpcOpenedCacheGet(ipc_opened_cache_handle_t cache,
 
 exit:
     if (ret == UMF_RESULT_SUCCESS) {
-        utils_atomic_increment(&entry->ref_count);
+        utils_atomic_increment_u64(&entry->ref_count);
         *retEntry = &entry->value;
     }
 

src/libumf.c

@@ -24,10 +24,10 @@
 
 umf_memory_tracker_handle_t TRACKER = NULL;
 
-static unsigned long long umfRefCount = 0;
+static uint64_t umfRefCount = 0;
 
 int umfInit(void) {
-    if (utils_fetch_and_add64(&umfRefCount, 1) == 0) {
+    if (utils_fetch_and_add_u64(&umfRefCount, 1) == 0) {
         utils_log_init();
         TRACKER = umfMemoryTrackerCreate();
         if (!TRACKER) {
@@ -54,7 +54,7 @@ int umfInit(void) {
 }
 
 void umfTearDown(void) {
-    if (utils_fetch_and_add64(&umfRefCount, -1) == 1) {
+    if (utils_fetch_and_sub_u64(&umfRefCount, 1) == 1) {
 #if !defined(_WIN32) && !defined(UMF_NO_HWLOC)
         umfMemspaceHostAllDestroy();
         umfMemspaceHighestCapacityDestroy();

src/provider/provider_os_memory.c

@@ -587,7 +587,7 @@ static umf_result_t os_initialize(void *params, void **provider) {
     }
 
     os_memory_provider_t *os_provider =
-        umf_ba_global_alloc(sizeof(os_memory_provider_t));
+        umf_ba_global_aligned_alloc(sizeof(os_memory_provider_t), 8);
     if (!os_provider) {
         return UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
     }
@@ -934,7 +934,7 @@ static membind_t membindFirst(os_memory_provider_t *provider, void *addr,
 
     if (provider->mode == UMF_NUMA_MODE_INTERLEAVE) {
         assert(provider->part_size != 0);
-        size_t s = utils_fetch_and_add64(&provider->alloc_sum, size);
+        size_t s = utils_fetch_and_add_u64(&provider->alloc_sum, size);
         membind.node = (s / provider->part_size) % provider->nodeset_len;
         membind.bitmap = provider->nodeset[membind.node];
         membind.bind_size = ALIGN_UP(provider->part_size, membind.page_size);

src/provider/provider_os_memory_internal.h

@@ -58,7 +58,7 @@ typedef struct os_memory_provider_t {
     int numa_flags; // combination of hwloc flags
 
     size_t part_size;
-    size_t alloc_sum; // sum of all allocations - used for manual interleaving
+    uint64_t alloc_sum; // sum of all allocations - used for manual interleaving
 
     struct {
         unsigned weight;

src/provider/provider_tracking.c

@@ -565,7 +565,7 @@ static umf_result_t trackingGetIpcHandle(void *provider, const void *ptr,
                 return ret;
             }
 
-            cache_value->handle_id = utils_atomic_increment(&IPC_HANDLE_ID);
+            cache_value->handle_id = utils_atomic_increment_u64(&IPC_HANDLE_ID);
             cache_value->ipcDataSize = ipcDataSize;
 
             int insRes = critnib_insert(p->ipcCache, (uintptr_t)ptr,
@@ -703,18 +703,20 @@ static umf_result_t trackingOpenIpcHandle(void *provider, void *providerIpcData,
     assert(cache_entry != NULL);
 
     void *mapped_ptr = NULL;
-    utils_atomic_load_acquire(&(cache_entry->mapped_base_ptr), &mapped_ptr);
+    utils_atomic_load_acquire_ptr(&(cache_entry->mapped_base_ptr),
+                                  (void **)&mapped_ptr);
     if (mapped_ptr == NULL) {
         utils_mutex_lock(&(cache_entry->mmap_lock));
-        utils_atomic_load_acquire(&(cache_entry->mapped_base_ptr), &mapped_ptr);
+        utils_atomic_load_acquire_ptr(&(cache_entry->mapped_base_ptr),
+                                      (void **)&mapped_ptr);
         if (mapped_ptr == NULL) {
             ret = upstreamOpenIPCHandle(p, providerIpcData,
                                         ipcUmfData->baseSize, &mapped_ptr);
             if (ret == UMF_RESULT_SUCCESS) {
                 // Put to the cache
                 cache_entry->mapped_size = ipcUmfData->baseSize;
-                utils_atomic_store_release(&(cache_entry->mapped_base_ptr),
-                                           mapped_ptr);
+                utils_atomic_store_release_ptr(&(cache_entry->mapped_base_ptr),
+                                               mapped_ptr);
             }
         }
         utils_mutex_unlock(&(cache_entry->mmap_lock));