Add size threshold to proxy lib to call system allocator

Add a size threshold to proxy lib to call system allocator
when the size is less than the given threshold.

Signed-off-by: Lukasz Dorau <lukasz.dorau@intel.com>

Author: ldorau

.github/workflows/reusable_proxy_lib.yml

@@ -77,6 +77,16 @@ jobs:
         working-directory: ${{env.BUILD_DIR}}
         run: UMF_PROXY="page.disposition=shared-shm" LD_PRELOAD=./lib/libumf_proxy.so /usr/bin/date
 
+      # TODO enable the provider_file_memory_ipc test when the IPC tests with the proxy library are fixed
+      # see the issue: https://github.com/oneapi-src/unified-memory-framework/issues/864
+      - name: Run "ctest --output-on-failure" with proxy library and size.threshold=128
+        working-directory: ${{env.BUILD_DIR}}
+        run: >
+          UMF_PROXY="size.threshold=128"
+          UMF_LOG="level:debug;flush:debug;output:stderr;pid:yes"
+          LD_PRELOAD=./lib/libumf_proxy.so
+          ctest --output-on-failure -E provider_file_memory_ipc
+
       - name: Check coverage
         if:  ${{ matrix.build_type == 'Debug' }}
         working-directory: ${{env.BUILD_DIR}}

src/proxy_lib/proxy_lib.c

@@ -27,6 +27,12 @@
  * - _aligned_offset_recalloc()
  */
 
+#ifndef _WIN32
+#define _GNU_SOURCE
+#include <dlfcn.h>
+#undef _GNU_SOURCE
+#endif /* _WIN32 */
+
 #if (defined PROXY_LIB_USES_JEMALLOC_POOL)
 #include <umf/pools/pool_jemalloc.h>
 #define umfPoolManagerOps umfJemallocPoolOps
@@ -103,38 +109,112 @@ static umf_memory_pool_handle_t Proxy_pool = NULL;
 // it protects us from recursion in umfPool*()
 static __TLS int was_called_from_umfPool = 0;
 
+typedef void *(*system_aligned_alloc_t)(size_t alignment, size_t size);
+typedef void *(*system_calloc_t)(size_t nmemb, size_t size);
+typedef void (*system_free_t)(void *ptr);
+typedef void *(*system_malloc_t)(size_t size);
+typedef size_t (*system_malloc_usable_size_t)(void *ptr);
+typedef void *(*system_realloc_t)(void *ptr, size_t size);
+
+static system_aligned_alloc_t system_aligned_alloc;
+static system_calloc_t system_calloc;
+static system_free_t system_free;
+static system_malloc_t system_malloc;
+static system_malloc_usable_size_t system_malloc_usable_size;
+static system_realloc_t system_realloc;
+static size_t threshold_value = 0;
+
 /*****************************************************************************/
 /*** The constructor and destructor of the proxy library *********************/
 /*****************************************************************************/
 
+#ifndef _WIN32
+static size_t get_size_threshold(void) {
+    char *str_threshold = utils_env_var_get_str("UMF_PROXY", "size.threshold=");
+    if (!str_threshold) {
+        return 0;
+    }
+
+    // move to the beginning of the number
+    str_threshold += strlen("size.threshold=");
+    // find ';' at the end
+    char *end = strstr(str_threshold, ";");
+    if (end) {
+        // replace ';' with '\0' to mark end of the string
+        *end = '\0';
+    }
+
+    size_t int_threshold = (size_t)atoi(str_threshold);
+    LOG_DEBUG("threshold_value = (char *) %s, (int) %zu", str_threshold,
+              int_threshold);
+
+    return int_threshold;
+}
+
+static int dlsym_system_allocator(void) {
+    *((void **)(&system_aligned_alloc)) = dlsym(RTLD_NEXT, "aligned_alloc");
+    *((void **)(&system_calloc)) = dlsym(RTLD_NEXT, "calloc");
+    *((void **)(&system_free)) = dlsym(RTLD_NEXT, "free");
+    *((void **)(&system_malloc)) = dlsym(RTLD_NEXT, "malloc");
+    *((void **)(&system_malloc_usable_size)) =
+        dlsym(RTLD_NEXT, "malloc_usable_size");
+    *((void **)(&system_realloc)) = dlsym(RTLD_NEXT, "realloc");
+
+    if (system_aligned_alloc && system_calloc && system_free && system_malloc &&
+        system_malloc_usable_size && system_realloc) {
+        return 0;
+    }
+
+    *((void **)(&system_aligned_alloc)) = NULL;
+    *((void **)(&system_calloc)) = NULL;
+    *((void **)(&system_free)) = NULL;
+    *((void **)(&system_malloc)) = NULL;
+    *((void **)(&system_malloc_usable_size)) = NULL;
+    *((void **)(&system_realloc)) = NULL;
+
+    return -1;
+}
+#endif /* _WIN32 */
+
 void proxy_lib_create_common(void) {
     utils_log_init();
     umf_os_memory_provider_params_t os_params =
         umfOsMemoryProviderParamsDefault();
     umf_result_t umf_result;
 
 #ifndef _WIN32
-    char shm_name[NAME_MAX];
+    size_t _threshold = get_size_threshold();
+    if (_threshold > 0) {
+        if (dlsym_system_allocator()) {
+            LOG_ERR("initialization of the system allocator failed!");
+            exit(-1);
+        }
+
+        threshold_value = _threshold;
+        LOG_INFO("system allocator initialized, size threshold value = %zu",
+                 threshold_value);
+    }
 
     if (utils_env_var_has_str("UMF_PROXY", "page.disposition=shared-fd")) {
-        LOG_DEBUG("proxy_lib: using the MAP_SHARED visibility mode with the "
-                  "file descriptor duplication");
+        LOG_INFO("proxy_lib: using the MAP_SHARED visibility mode with the "
+                 "file descriptor duplication");
         os_params.visibility = UMF_MEM_MAP_SHARED;
         os_params.shm_name = NULL;
 
     } else if (utils_env_var_has_str("UMF_PROXY",
                                      "page.disposition=shared-shm")) {
         os_params.visibility = UMF_MEM_MAP_SHARED;
 
+        char shm_name[NAME_MAX];
         memset(shm_name, 0, NAME_MAX);
         sprintf(shm_name, "umf_proxy_lib_shm_pid_%i", utils_getpid());
         os_params.shm_name = shm_name;
 
-        LOG_DEBUG("proxy_lib: using the MAP_SHARED visibility mode with the "
-                  "named shared memory: %s",
-                  os_params.shm_name);
+        LOG_INFO("proxy_lib: using the MAP_SHARED visibility mode with the "
+                 "named shared memory: %s",
+                 os_params.shm_name);
     }
-#endif
+#endif /* _WIN32 */
 
     umf_result = umfMemoryProviderCreate(umfOsMemoryProviderOps(), &os_params,
                                          &OS_memory_provider);
@@ -149,16 +229,18 @@ void proxy_lib_create_common(void) {
         LOG_ERR("creating UMF pool manager failed");
         exit(-1);
     }
+
     // The UMF pool has just been created (Proxy_pool != NULL). Stop using
     // the linear allocator and start using the UMF pool allocator from now on.
+    LOG_DEBUG("proxy library initialized");
 }
 
 void proxy_lib_destroy_common(void) {
     if (utils_is_running_in_proxy_lib()) {
         // We cannot destroy 'Base_alloc_leak' nor 'Proxy_pool' nor 'OS_memory_provider',
         // because it could lead to use-after-free in the program's unloader
         // (for example _dl_fini() on Linux).
-        return;
+        goto fini_proxy_lib_destroy_common;
     }
 
     umf_memory_pool_handle_t pool = Proxy_pool;
@@ -168,6 +250,10 @@ void proxy_lib_destroy_common(void) {
     umf_memory_provider_handle_t provider = OS_memory_provider;
     OS_memory_provider = NULL;
     umfMemoryProviderDestroy(provider);
+    LOG_DEBUG("proxy library destroyed");
+
+fini_proxy_lib_destroy_common:
+    LOG_DEBUG("proxy library finalized");
 }
 
 /*****************************************************************************/
@@ -246,6 +332,10 @@ static inline size_t ba_leak_pool_contains_pointer(void *ptr) {
 /*****************************************************************************/
 
 void *malloc(size_t size) {
+    if (size < threshold_value) {
+        return system_malloc(size);
+    }
+
     if (!was_called_from_umfPool && Proxy_pool) {
         was_called_from_umfPool = 1;
         void *ptr = umfPoolMalloc(Proxy_pool, size);
@@ -257,6 +347,10 @@ void *malloc(size_t size) {
 }
 
 void *calloc(size_t nmemb, size_t size) {
+    if ((nmemb * size) < threshold_value) {
+        return system_calloc(nmemb, size);
+    }
+
     if (!was_called_from_umfPool && Proxy_pool) {
         was_called_from_umfPool = 1;
         void *ptr = umfPoolCalloc(Proxy_pool, nmemb, size);
@@ -276,15 +370,20 @@ void free(void *ptr) {
         return;
     }
 
-    if (Proxy_pool) {
+    if (Proxy_pool && (umfPoolByPtr(ptr) == Proxy_pool)) {
         if (umfPoolFree(Proxy_pool, ptr) != UMF_RESULT_SUCCESS) {
             LOG_ERR("umfPoolFree() failed");
-            assert(0);
         }
         return;
     }
 
-    assert(0);
+    if (threshold_value) {
+        system_free(ptr);
+        return;
+    }
+
+    LOG_ERR("free() failed: %p", ptr);
+
     return;
 }
 
@@ -303,18 +402,27 @@ void *realloc(void *ptr, size_t size) {
         return ba_leak_realloc(ptr, size, leak_pool_contains_pointer);
     }
 
-    if (Proxy_pool) {
+    if (Proxy_pool && (umfPoolByPtr(ptr) == Proxy_pool)) {
         was_called_from_umfPool = 1;
         void *new_ptr = umfPoolRealloc(Proxy_pool, ptr, size);
         was_called_from_umfPool = 0;
         return new_ptr;
     }
 
-    assert(0);
+    if (threshold_value) {
+        return system_realloc(ptr, size);
+    }
+
+    LOG_ERR("realloc() failed: %p", ptr);
+
     return NULL;
 }
 
 void *aligned_alloc(size_t alignment, size_t size) {
+    if (size < threshold_value) {
+        return system_aligned_alloc(alignment, size);
+    }
+
     if (!was_called_from_umfPool && Proxy_pool) {
         was_called_from_umfPool = 1;
         void *ptr = umfPoolAlignedMalloc(Proxy_pool, size, alignment);
@@ -330,19 +438,29 @@ size_t _msize(void *ptr) {
 #else
 size_t malloc_usable_size(void *ptr) {
 #endif
-
-    // a check to verify we are running the proxy library
+    // a check to verify if we are running the proxy library
     if (ptr == (void *)0x01) {
         return 0xDEADBEEF;
     }
 
-    if (!was_called_from_umfPool && Proxy_pool) {
+    if (ba_leak_pool_contains_pointer(ptr)) {
+        return 0; // unsupported in case of the ba_leak allocator
+    }
+
+    if (!was_called_from_umfPool && Proxy_pool &&
+        (umfPoolByPtr(ptr) == Proxy_pool)) {
         was_called_from_umfPool = 1;
         size_t size = umfPoolMallocUsableSize(Proxy_pool, ptr);
         was_called_from_umfPool = 0;
         return size;
     }
 
+    if (threshold_value) {
+        return system_malloc_usable_size(ptr);
+    }
+
+    LOG_ERR("malloc_usable_size() failed: %p", ptr);
+
     return 0; // unsupported in this case
 }