Merge pull request #7909 from yosefe/topic/memtype-cache-fixes-v1.12.x

Memtype cache fixes - v1.12.x

NEWS

@@ -12,6 +12,7 @@
 * Fixed memory hooks for Cuda 11.5
 * Fixed memory type cache merge
 * Fixed continuously triggering wakeup fd when keepalive is used
+* Fixed memtype cache fallback when memory hooks are not installed
 
 ## 1.12.0 (January 12, 2022)
 ### Features:

config/m4/ucm.m4

@@ -46,6 +46,13 @@ AC_CHECK_DECLS([MADV_FREE,
                [#include <sys/mman.h>])
 
 
+#
+# getauxval()
+#
+AC_CHECK_DECLS([getauxval], [], [],
+               [#include <sys/auxv.h>])
+
+
 # BISTRO hooks infrastructure
 #
 # SYS_xxx macro

src/ucm/cuda/cudamem.c

@@ -194,75 +194,66 @@ static ucm_cuda_func_t ucm_cuda_runtime_funcs[] = {
     {{NULL}, NULL}
 };
 
-static int ucm_cuda_allow_hook_mode(ucm_mmap_hook_mode_t mode)
-{
-    return (ucm_global_opts.cuda_hook_modes & UCS_BIT(mode)) &&
-           (ucm_get_hook_mode(mode) == mode);
-}
-
 static ucs_status_t
-ucm_cuda_install_hooks(ucm_cuda_func_t *funcs, int *used_reloc,
-                       const char *name)
+ucm_cuda_install_hooks(ucm_cuda_func_t *funcs, const char *name,
+                       ucm_mmap_hook_mode_t mode, int *installed_hooks_p)
 {
-    const char UCS_V_UNUSED *hook_mode;
-    unsigned num_bistro, num_reloc;
     ucm_cuda_func_t *func;
     ucs_status_t status;
     void *func_ptr;
+    int count;
+
+    if (*installed_hooks_p & UCS_BIT(mode)) {
+        return UCS_OK;
+    }
 
-    num_bistro  = 0;
-    num_reloc   = 0;
+    if (!(ucm_global_opts.cuda_hook_modes & UCS_BIT(mode))) {
+        /* Disabled by configuration */
+        ucm_debug("cuda memory hooks mode %s is disabled for %s API",
+                  ucm_mmap_hook_modes[mode], name);
+        return UCS_OK;
+    }
+
+    count = 0;
     for (func = funcs; func->patch.symbol != NULL; ++func) {
         func_ptr = ucm_reloc_get_orig(func->patch.symbol, func->patch.value);
         if (func_ptr == NULL) {
             continue;
         }
 
-        status = UCS_ERR_UNSUPPORTED;
-
-        if (ucm_cuda_allow_hook_mode(UCM_MMAP_HOOK_BISTRO)) {
+        if (mode == UCM_MMAP_HOOK_BISTRO) {
             status = ucm_bistro_patch(func_ptr, func->patch.value,
                                       func->patch.symbol, func->orig_func_ptr,
                                       NULL);
-            if (status == UCS_OK) {
-                ucm_trace("installed bistro hook for '%s': %s",
-                          func->patch.symbol, ucs_status_string(status));
-                ++num_bistro;
-                continue;
-            }
-
-            ucm_debug("failed to install bistro hook for '%s'",
-                      func->patch.symbol);
+        } else if (mode == UCM_MMAP_HOOK_RELOC) {
+            status = ucm_reloc_modify(&func->patch);
+        } else {
+            break;
         }
 
-        if (ucm_cuda_allow_hook_mode(UCM_MMAP_HOOK_RELOC)) {
-            status = ucm_reloc_modify(&func->patch);
-            if (status == UCS_OK) {
-                ++num_reloc;
-                ucm_trace("installed reloc hook on '%s'", func->patch.symbol);
-                continue;
-            }
-
-            ucm_debug("failed to install relocation table hook for '%s'",
-                      func->patch.symbol);
+        if (status != UCS_OK) {
+            ucm_diag("failed to install %s hook for '%s'",
+                     ucm_mmap_hook_modes[mode], func->patch.symbol);
+            return status;
         }
 
-        ucm_diag("failed to install hook for '%s'", func->patch.symbol);
-        return status;
+        ucm_debug("installed %s hook for '%s'", ucm_mmap_hook_modes[mode],
+                  func->patch.symbol);
+        ++count;
     }
 
-    *used_reloc = num_reloc > 0;
-    ucm_info("cuda memory hooks on %s API: installed %u bistro and %u reloc",
-             name, num_bistro, num_reloc);
+    *installed_hooks_p |= UCS_BIT(mode);
+    ucm_info("cuda memory hooks mode %s: installed %d on %s API",
+             ucm_mmap_hook_modes[mode], count, name);
     return UCS_OK;
 }
 
 static ucs_status_t ucm_cudamem_install(int events)
 {
-    static int ucm_cudamem_installed     = 0;
     static pthread_mutex_t install_mutex = PTHREAD_MUTEX_INITIALIZER;
+    static int driver_api_hooks          = 0;
+    static int runtime_api_hooks         = 0;
     ucs_status_t status                  = UCS_OK;
-    int used_reloc;
 
     if (!(events & (UCM_EVENT_MEM_TYPE_ALLOC | UCM_EVENT_MEM_TYPE_FREE))) {
         goto out;
@@ -276,26 +267,22 @@ static ucs_status_t ucm_cudamem_install(int events)
 
     pthread_mutex_lock(&install_mutex);
 
-    if (ucm_cudamem_installed) {
+    status = ucm_cuda_install_hooks(ucm_cuda_driver_funcs, "driver",
+                                    UCM_MMAP_HOOK_BISTRO, &driver_api_hooks);
+    if (status != UCS_OK) {
         goto out_unlock;
     }
 
-    status = ucm_cuda_install_hooks(ucm_cuda_driver_funcs, &used_reloc,
-                                    "driver");
+    status = ucm_cuda_install_hooks(ucm_cuda_driver_funcs, "driver",
+                                    UCM_MMAP_HOOK_RELOC, &driver_api_hooks);
     if (status != UCS_OK) {
-        ucm_warn("failed to install cuda memory hooks on driver API");
-    } else if (!used_reloc) {
-        ucm_cudamem_installed = 1;
-    } else if (status == UCS_OK) {
-        /* Failed to install bistro hooks on all driver APIs, so need to install
-           hooks on runtime APIs. */
-        status = ucm_cuda_install_hooks(ucm_cuda_runtime_funcs, &used_reloc,
-                                        "runtime");
-        if (status == UCS_OK) {
-            ucm_cudamem_installed = 1;
-        } else {
-            ucm_warn("failed to install cuda memory hooks on runtime API")
-        }
+        goto out_unlock;
+    }
+
+    status = ucm_cuda_install_hooks(ucm_cuda_runtime_funcs, "runtime",
+                                    UCM_MMAP_HOOK_RELOC, &runtime_api_hooks);
+    if (status != UCS_OK) {
+        goto out_unlock;
     }
 
 out_unlock:

src/ucm/mmap/install.c

@@ -58,7 +58,6 @@
     ((UCM_MMAP_MAX_EVENT_NAME_LEN + 2) * \
     ucs_static_array_size(ucm_mmap_event_name))
 
-extern const char *ucm_mmap_hook_modes[];
 
 typedef struct ucm_mmap_func {
     ucm_reloc_patch_t    patch;
@@ -89,6 +88,15 @@ static ucm_mmap_func_t ucm_mmap_funcs[] = {
 static pthread_mutex_t ucm_mmap_install_mutex = PTHREAD_MUTEX_INITIALIZER;
 static int ucm_mmap_installed_events = 0; /* events that were reported as installed */
 
+const char *ucm_mmap_hook_modes[] = {
+    [UCM_MMAP_HOOK_NONE]   = "none",
+    [UCM_MMAP_HOOK_RELOC]  = UCM_MMAP_HOOK_RELOC_STR,
+#if UCM_BISTRO_HOOKS
+    [UCM_MMAP_HOOK_BISTRO] = UCM_MMAP_HOOK_BISTRO_STR,
+#endif
+    [UCM_MMAP_HOOK_LAST]   = NULL
+};
+
 static const char *ucm_mmap_event_name[] = {
     /* Native events */
     UCM_MMAP_EVENT_NAME_ENTRY(MMAP),

src/ucm/mmap/mmap.h

@@ -39,6 +39,13 @@ ucs_status_t ucm_mmap_test_installed_events(int events);
 ucs_status_t ucm_mmap_test_events(int events, const char *event_type);
 void ucm_mmap_init();
 
+
+/**
+ * Memory hooks mode names.
+ */
+extern const char *ucm_mmap_hook_modes[];
+
+
 static UCS_F_ALWAYS_INLINE ucm_mmap_hook_mode_t ucm_mmap_hook_mode(void)
 {
     return ucm_get_hook_mode(ucm_global_opts.mmap_hook_mode);

src/ucm/util/reloc.c

@@ -29,6 +29,11 @@
 #include <link.h>
 #include <limits.h>
 
+#ifdef HAVE_DECL_GETAUXVAL
+#include <sys/auxv.h>
+#endif
+
+
 /* Ensure this macro is defined (from <link.h>) - otherwise, cppcheck might
    fail with an "unknown macro" warning */
 #ifndef ElfW
@@ -116,6 +121,14 @@ static ucs_status_t ucm_reloc_get_aux_phsize(int *phsize_p)
         return UCS_OK;
     }
 
+#ifdef HAVE_DECL_GETAUXVAL
+    phsize = getauxval(AT_PHENT);
+    if (phsize > 0) {
+        *phsize_p = phsize;
+        return UCS_OK;
+    }
+#endif
+
     fd = open(proc_auxv_filename, O_RDONLY);
     if (fd < 0) {
         ucm_error("failed to open '%s' for reading: %m", proc_auxv_filename);

src/ucp/core/ucp_context.h

@@ -505,25 +505,19 @@ ucp_memory_detect_internal(ucp_context_h context, const void *address,
     }
 
     status = ucs_memtype_cache_lookup(address, length, mem_info);
-    if (status != UCS_ERR_NO_ELEM) {
-        if (ucs_likely(status != UCS_OK)) {
-            ucs_assert(status == UCS_ERR_NO_ELEM);
-            goto out_host_mem;
-        }
-
-        if ((mem_info->type != UCS_MEMORY_TYPE_UNKNOWN) &&
-            ((mem_info->sys_dev != UCS_SYS_DEVICE_ID_UNKNOWN))) {
-            return;
-        }
-
-        /* Fall thru to slow-path memory type and system device detection by UCT
-         * memory domains. In any case, the memory type cache is not expected to
-         * return HOST memory type.
-         */
-        ucs_assert(mem_info->type != UCS_MEMORY_TYPE_HOST);
+    if (ucs_likely(status == UCS_ERR_NO_ELEM)) {
+        goto out_host_mem;
+    } else if ((status == UCS_ERR_UNSUPPORTED) ||
+               ((status == UCS_OK) &&
+                ((mem_info->type == UCS_MEMORY_TYPE_UNKNOWN) ||
+                 (mem_info->sys_dev == UCS_SYS_DEVICE_ID_UNKNOWN)))) {
+        ucp_memory_detect_slowpath(context, address, length, mem_info);
+    } else {
+        ucs_assertv(status == UCS_OK, "%s (%d)", ucs_status_string(status),
+                    status);
     }
 
-    ucp_memory_detect_slowpath(context, address, length, mem_info);
+    /* Memory type and system device was detected successfully */
     return;
 
 out_host_mem:

src/ucs/config/global_opts.c

@@ -37,7 +37,7 @@ ucs_global_opts_t ucs_global_opts = {
     .debug_signo           = SIGHUP,
     .log_level_trigger     = UCS_LOG_LEVEL_FATAL,
     .warn_unused_env_vars  = 1,
-    .enable_memtype_cache  = 1,
+    .enable_memtype_cache  = UCS_TRY,
     .async_max_events      = 64,
     .async_signo           = SIGALRM,
     .stats_dest            = "",
@@ -145,9 +145,9 @@ static ucs_config_field_t ucs_global_opts_table[] = {
   "configuration parser.",
   ucs_offsetof(ucs_global_opts_t, warn_unused_env_vars), UCS_CONFIG_TYPE_BOOL},
 
-  {"MEMTYPE_CACHE", "y",
+  {"MEMTYPE_CACHE", "try",
    "Enable memory type (cuda/rocm) cache",
-   ucs_offsetof(ucs_global_opts_t, enable_memtype_cache), UCS_CONFIG_TYPE_BOOL},
+   ucs_offsetof(ucs_global_opts_t, enable_memtype_cache), UCS_CONFIG_TYPE_TERNARY},
 
  {"ASYNC_MAX_EVENTS", "1024", /* TODO remove this; resize mpmc */
   "Maximal number of events which can be handled from one context",

src/ucs/config/global_opts.h

@@ -80,7 +80,7 @@ typedef struct {
     unsigned                   async_max_events;
 
     /** Memtype cache */
-    int                        enable_memtype_cache;
+    ucs_ternary_auto_value_t   enable_memtype_cache;
 
     /* Destination for statistics: udp:host:port / file:path / stdout
      */

src/ucs/config/ucm_opts.c

@@ -12,21 +12,13 @@
 
 #include <ucm/api/ucm.h>
 #include <ucm/util/log.h>
+#include <ucm/util/sys.h>
 #include <ucm/mmap/mmap.h>
 #include <ucs/sys/compiler.h>
 
 
 #define UCM_CONFIG_PREFIX   "MEM_"
 
-static const char *ucm_mmap_hook_modes[] = {
-    [UCM_MMAP_HOOK_NONE]   = "none",
-    [UCM_MMAP_HOOK_RELOC]  = UCM_MMAP_HOOK_RELOC_STR,
-#if UCM_BISTRO_HOOKS
-    [UCM_MMAP_HOOK_BISTRO] = UCM_MMAP_HOOK_BISTRO_STR,
-#endif
-    [UCM_MMAP_HOOK_LAST]   = NULL
-};
-
 static const char *ucm_module_unload_prevent_modes[] = {
     [UCM_UNLOAD_PREVENT_MODE_LAZY] = "lazy",
     [UCM_UNLOAD_PREVENT_MODE_NOW]  = "now",
@@ -71,9 +63,10 @@ static ucs_config_field_t ucm_global_config_table[] = {
 
   {"CUDA_HOOK_MODE",
 #if UCM_BISTRO_HOOKS
-    UCM_MMAP_HOOK_BISTRO_STR ","
-#endif
+   UCM_MMAP_HOOK_BISTRO_STR,
+#else
    UCM_MMAP_HOOK_RELOC_STR,
+#endif
    "Cuda memory hook modes. A combination of:\n"
    " none   - Don't set Cuda hooks.\n"
    " reloc  - Use ELF relocation table to set hooks. In this mode, if any\n"