Fix data corruption in MXM by registering to OPAL memory release hooks and removing any mappings created by mxm

This commit was SVN r28489.

ompi/mca/mtl/mxm/mtl_mxm.c

@@ -15,6 +15,7 @@
 #include "ompi/mca/mtl/base/mtl_base_datatype.h"
 #include "ompi/proc/proc.h"
 #include "ompi/communicator/communicator.h"
+#include "opal/memoryhooks/memory.h"
 
 #include "mtl_mxm.h"
 #include "mtl_mxm_types.h"
@@ -39,7 +40,11 @@ mca_mtl_mxm_module_t ompi_mtl_mxm = {
        ompi_mtl_mxm_cancel,
        ompi_mtl_mxm_add_comm,
        ompi_mtl_mxm_del_comm
-    }
+    },
+    0,
+    0,
+    NULL,
+    NULL
 };
 
 
@@ -78,6 +83,10 @@ static uint32_t ompi_mtl_mxm_get_job_id(void)
 }
 
 int ompi_mtl_mxm_progress(void);
+#if MXM_API >= MXM_VERSION(2,0)
+static void ompi_mtl_mxm_mem_release_cb(void *buf, size_t length,
+                                        void *cbdata, bool from_alloc);
+#endif
 
 #if MXM_API < MXM_VERSION(2, 0)
 static int ompi_mtl_mxm_get_ep_address(ompi_mtl_mxm_ep_conn_info_t *ep_info, mxm_ptl_id_t ptlid)
@@ -412,11 +421,22 @@ int ompi_mtl_mxm_module_init(void)
      
     /* Register the MXM progress function */
     opal_progress_register(ompi_mtl_mxm_progress);
+
+#if MXM_API >= MXM_VERSION(2,0)
+    if (ompi_mtl_mxm.using_mem_hooks) {
+        opal_mem_hooks_register_release(ompi_mtl_mxm_mem_release_cb, NULL);
+    }
+#endif
     return OMPI_SUCCESS;
 }
 
 int ompi_mtl_mxm_finalize(struct mca_mtl_base_module_t* mtl)
 {
+#if MXM_API >= MXM_VERSION(2,0)
+    if (ompi_mtl_mxm.using_mem_hooks) {
+        opal_mem_hooks_unregister_release(ompi_mtl_mxm_mem_release_cb);
+    }
+#endif
     opal_progress_unregister(ompi_mtl_mxm_progress);
     mxm_ep_destroy(ompi_mtl_mxm.ep);
     return OMPI_SUCCESS;
@@ -566,6 +586,15 @@ int ompi_mtl_mxm_progress(void)
     return 1;
 }
 
+#if MXM_API >= MXM_VERSION(2,0)
+static void ompi_mtl_mxm_mem_release_cb(void *buf, size_t length,
+                                        void *cbdata, bool from_alloc)
+{
+    mxm_mem_unmap(ompi_mtl_mxm.mxm_context, buf, length,
+                  from_alloc ? MXM_MEM_UNMAP_MARK_INVALID : 0);
+}
+#endif
+
 #if MXM_API >= MXM_VERSION(1,5)
 OBJ_CLASS_INSTANCE(
         ompi_mtl_mxm_message_t,

ompi/mca/mtl/mxm/mtl_mxm_component.c

@@ -11,6 +11,7 @@
 
 #include "opal/util/output.h"
 #include "ompi/proc/proc.h"
+#include "opal/memoryhooks/memory.h"
 
 #include "mtl_mxm.h"
 #include "mtl_mxm_types.h"
@@ -103,20 +104,35 @@ static int ompi_mtl_mxm_component_open(void)
         return OMPI_ERR_NOT_AVAILABLE;
     }
 #if MXM_API < MXM_VERSION(1,5)
-        mxm_fill_context_opts(&ompi_mtl_mxm.mxm_opts);
+    mxm_fill_context_opts(&ompi_mtl_mxm.mxm_opts);
-        err = mxm_init(&ompi_mtl_mxm.mxm_opts, &ompi_mtl_mxm.mxm_context);
+    err = mxm_init(&ompi_mtl_mxm.mxm_opts, &ompi_mtl_mxm.mxm_context);
-        MXM_VERBOSE(1, "mxm component open");
+    MXM_VERBOSE(1, "mxm component open");
 #else
-        err = mxm_config_read_context_opts(&ompi_mtl_mxm.mxm_opts);
-        if (err != MXM_OK) {
-           MXM_ERROR("Failed to parse MXM configuration");
-            return OPAL_ERR_BAD_PARAM;
-        }
 
-        err = mxm_init(ompi_mtl_mxm.mxm_opts, &ompi_mtl_mxm.mxm_context);
+#if MXM_API >= MXM_VERSION(2,0)
-        MXM_VERBOSE(1, "mxm component open");
+    /* Register memory hooks */
+    if ((OPAL_MEMORY_FREE_SUPPORT | OPAL_MEMORY_MUNMAP_SUPPORT) ==
+        ((OPAL_MEMORY_FREE_SUPPORT | OPAL_MEMORY_MUNMAP_SUPPORT) &
+         opal_mem_hooks_support_level()))
+    {
+        setenv("MXM_MEM_ON_DEMAND_MAP", "y", 0);
+        MXM_VERBOSE(1, "Enabling on-demand memory mapping");
+        ompi_mtl_mxm.using_mem_hooks = 1;
+    } else {
+        MXM_VERBOSE(1, "Disabling on-demand memory mapping");
+        ompi_mtl_mxm.using_mem_hooks = 0;
+    }
 #endif
 
+    err = mxm_config_read_context_opts(&ompi_mtl_mxm.mxm_opts);
+    if (err != MXM_OK) {
+        MXM_ERROR("Failed to parse MXM configuration");
+        return OPAL_ERR_BAD_PARAM;
+    }
+
+    err = mxm_init(ompi_mtl_mxm.mxm_opts, &ompi_mtl_mxm.mxm_context);
+    MXM_VERBOSE(1, "mxm component open");
+#endif
 
     if (MXM_OK != err) {
         if (MXM_ERR_NO_DEVICE == err) {
@@ -158,8 +174,7 @@ static int ompi_mtl_mxm_component_close(void)
     unsigned long cur_ver;
 
     cur_ver = mxm_get_version();
-
+    if ((cur_ver == MXM_API) && (ompi_mtl_mxm.mxm_context != NULL)) {
-    if (cur_ver == MXM_API) {
         mxm_cleanup(ompi_mtl_mxm.mxm_context);
         ompi_mtl_mxm.mxm_context = NULL;
 
@@ -189,4 +204,3 @@ ompi_mtl_mxm_component_init(bool enable_progress_threads,
             sizeof(mca_mtl_mxm_request_t) - sizeof(struct mca_mtl_request_t);
     return &ompi_mtl_mxm.super;
 }
-

ompi/mca/mtl/mxm/mtl_mxm_types.h

@@ -29,12 +29,15 @@ typedef struct mca_mtl_mxm_module_t {
     int                   verbose;
     int                   mxm_np;
     mxm_h                 mxm_context;
+    mxm_ep_h              ep;
 #if MXM_API < MXM_VERSION(1,5)
     mxm_context_opts_t    mxm_opts;
 #else
     mxm_context_opts_t   *mxm_opts;
 #endif
-    mxm_ep_h              ep;
+#if MXM_API >= MXM_VERSION(2,0)
+    int                   using_mem_hooks;
+#endif
 } mca_mtl_mxm_module_t;