i#2566 anon text: Use the app path instead of maps comments

For the application module, we use the application path obtained from
early injection or /proc/self/exe on Linux, rather than
/proc/self/maps comments.  The maps comments can be unreliable in the
face of anonymous or deleted-file mremaps used for hugepage backing
and other features.

Adds a test case to the existing "burst_maps" test.

Having the right module full_path helps many cases, including the
forthcoming restartable sequences ("rseq") support for #2350.

Issue: #2566, #2350

clients/drcachesim/tests/burst_maps.cpp

@@ -1,5 +1,5 @@
 /* **********************************************************
- * Copyright (c) 2017 Google, Inc.  All rights reserved.
+ * Copyright (c) 2017-2019 Google, Inc.  All rights reserved.
  * **********************************************************/
 
 /*
@@ -57,6 +57,8 @@
 #define MAPS_LINE_MAX8 73 /* sum of 16  1  16  1 4 1 16 1 5 1 10 1 */
 #define MAPS_LINE_MAX MAPS_LINE_MAX8
 
+char exe_path[MAPS_LINE_LENGTH];
+
 void *
 find_exe_base()
 {
@@ -82,6 +84,8 @@ find_exe_base()
         if (len < 4)
             comment_buffer[0] = '\0';
         if (strstr(comment_buffer, "burst_maps") != 0) {
+            strncpy(exe_path, comment_buffer, BUFFER_SIZE_ELEMENTS(exe_path));
+            NULL_TERMINATE_BUFFER(exe_path);
             fclose(maps);
             return vm_start;
         }
@@ -137,6 +141,17 @@ clobber_mapping()
     copy_and_remap(exe, 8 * clobber_size, clobber_size);
 }
 
+DR_EXPORT void
+dr_client_main(client_id_t id, int argc, const char *argv[])
+{
+    /* Test the full_path used by DR when the maps file comments can't be used. */
+    module_data_t *exe = dr_get_main_module();
+    assert(exe != nullptr);
+    assert(exe->full_path != nullptr);
+    assert(strcmp(exe->full_path, exe_path) == 0);
+    dr_free_module_data(exe);
+}
+
 int
 main(int argc, const char *argv[])
 {

core/unix/os.c

@@ -9118,15 +9118,25 @@ os_walk_address_space(memquery_iter_t *iter, bool add_modules)
                 iter->vm_start, iter->vm_start + image_size, iter->inode, iter->comment);
 
             if (add_modules) {
+                const char *modpath = iter->comment;
                 /* look for executable */
 #    ifdef LINUX
                 exec_match = get_application_name();
                 if (exec_match != NULL && exec_match[0] != '\0')
                     found_exec = (strcmp(iter->comment, exec_match) == 0);
                 /* Handle an anon region for the header (i#2566) */
                 if (!found_exec && executable_start != NULL &&
-                    executable_start == iter->vm_start)
+                    executable_start == iter->vm_start) {
                     found_exec = true;
+                    /* The maps file's first entry may not have the path, in the
+                     * presence of mremapping for hugepages (i#2566; i#3387) (this
+                     * could happen for libraries too, but we don't have alternatives
+                     * there).  Or, it may have an incorrect path.  Prefer the path
+                     * we recorded in early injection or obtained from
+                     * /proc/self/exe.
+                     */
+                    modpath = get_application_name();
+                }
 #    else
                 /* We don't have a nice normalized name: it can have ./ or ../ inside
                  * it.  But, we can distinguish an exe from a lib here, even for PIE,
@@ -9156,8 +9166,8 @@ os_walk_address_space(memquery_iter_t *iter, bool add_modules)
                 /* We don't yet know whether contiguous so we have to settle for the
                  * first segment's size.  We'll update it in module_list_add().
                  */
-                module_list_add(iter->vm_start, mod_first_end - mod_base, false,
-                                iter->comment, iter->inode);
+                module_list_add(iter->vm_start, mod_first_end - mod_base, false, modpath,
+                                iter->inode);
 
 #    ifdef MACOS
                 /* look for dyld */