i#140 pcprof: add client pc sampling support (#2883)

Added dr_where_am_i() to better support client self-profiling via sampling. This also provides the fragment tag, by refactoring the pcprofile code into a new helper fcache_refine_whereami(). Renamed the whereami types to better avoid name conflicts: s/WHERE_xxx/DR_WHERE_xxx/ and s/where_am_i_t/dr_where_am_i_t/. Exported the dr_where_am_i_t enum. Fixed a bug where an uninitialized mcontext was passed to a client timer callback: we only filled in the mcontext for a DR-internal callback. Added a test of client sampling to client.timer Issue: #140
DynamoRIO · Mar 15, 2018 · ce3a9f4 · ce3a9f4
1 parent 3f2dd1d
commit ce3a9f4
Show file tree

Hide file tree

Showing 26 changed files with 312 additions and 211 deletions.
diff --git a/api/docs/release.dox b/api/docs/release.dox
@@ -132,7 +132,7 @@ Dr. Memory Framework (DRMF) in the same package as DynamoRIO.  DRMF
 provides the umbra, drsyscall, and drsymcache Extensions for use by
 clients.
 
-The changes between version \DR_VERSION and 6.2.0 include the following minor
+The changes between version \DR_VERSION and 7.0.0 include the following minor
 compatibility changes:
 
  - The drltrace tool has been moved to the Dr.Memory Framework.
@@ -223,6 +223,7 @@ Further non-compatibility-affecting changes include:
    drmgr_unregister_pre_syscall_event_user_data() to enable passing of user data.
  - Added drmgr_register_post_syscall_event_user_data() and
    drmgr_unregister_post_syscall_event_user_data() to enable passing of user data.
+ - Added dr_where_am_i() to better support client self-profiling via sampling.
 
 **************************************************
 <hr>

diff --git a/core/arch/emit_utils_shared.c b/core/arch/emit_utils_shared.c
@@ -5638,11 +5638,11 @@ insert_entering_native(dcontext_t *dcontext, instrlist_t *ilist, instr_t *where,
      */
 
     /* C equivalent:
-     *   whereami = WHERE_APP
+     *   whereami = DR_WHERE_APP
      */
     PRE(ilist, where,
         instr_create_save_immed_to_dc_via_reg(dcontext, reg_dc, WHEREAMI_OFFSET,
-                                              (ptr_int_t) WHERE_APP, OPSZ_4));
+                                              (ptr_int_t) DR_WHERE_APP, OPSZ_4));
 
     /* skip C equivalent:
      *   STATS_INC(num_native_module_enter)
@@ -5703,11 +5703,11 @@ insert_entering_non_native(dcontext_t *dcontext, instrlist_t *ilist, instr_t *wh
                          ilist, where, reg_dc);
 
     /* C equivalent:
-     *   whereami = WHERE_FCACHE
+     *   whereami = DR_WHERE_FCACHE
      */
     PRE(ilist, where,
         instr_create_save_immed_to_dc_via_reg(dcontext, reg_dc, WHEREAMI_OFFSET,
-                                              (ptr_int_t) WHERE_FCACHE, OPSZ_4));
+                                              (ptr_int_t) DR_WHERE_FCACHE, OPSZ_4));
 }
 
 /* Emit code to transfer execution from native module to code cache of non-native

diff --git a/core/arch/interp.c b/core/arch/interp.c
@@ -5150,10 +5150,10 @@ build_basic_block_fragment(dcontext_t *dcontext, app_pc start, uint initial_flag
 {
     fragment_t *f;
     build_bb_t bb;
-    where_am_i_t wherewasi = dcontext->whereami;
+    dr_where_am_i_t wherewasi = dcontext->whereami;
     bool image_entry;
     KSTART(bb_building);
-    dcontext->whereami = WHERE_INTERP;
+    dcontext->whereami = DR_WHERE_INTERP;
 
     /* Neither thin_client nor hotp_only should be building any bbs. */
     ASSERT(!RUNNING_WITHOUT_CODE_CACHE());

diff --git a/core/arch/mangle_shared.c b/core/arch/mangle_shared.c
@@ -1,5 +1,5 @@
 /* ******************************************************************************
- * Copyright (c) 2010-2017 Google, Inc.  All rights reserved.
+ * Copyright (c) 2010-2018 Google, Inc.  All rights reserved.
  * Copyright (c) 2010 Massachusetts Institute of Technology  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
  * ******************************************************************************/
@@ -437,7 +437,7 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
                 XINST_CREATE_store(dcontext,
                                    OPND_CREATE_MEMPTR(SCRATCH_REG0, 0),
                                    opnd_create_reg(SCRATCH_REG1)));
-            instrlist_insert_mov_immed_ptrsz(dcontext, (ptr_int_t)WHERE_CLEAN_CALLEE,
+            instrlist_insert_mov_immed_ptrsz(dcontext, (ptr_int_t)DR_WHERE_CLEAN_CALLEE,
                                              opnd_create_reg(SCRATCH_REG1),
                                              ilist, instr, NULL, NULL);
             PRE(ilist, instr,
@@ -449,14 +449,14 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
                                   OPND_CREATE_MEMPTR(SCRATCH_REG0, 0)));
 # else
             PRE(ilist, instr,
-                instr_create_save_immed_to_dc_via_reg(dcontext, SCRATCH_REG0,
-                                                      WHEREAMI_OFFSET,
-                                                      (uint) WHERE_CLEAN_CALLEE, OPSZ_4));
+                instr_create_save_immed_to_dc_via_reg
+                (dcontext, SCRATCH_REG0, WHEREAMI_OFFSET,
+                 (uint) DR_WHERE_CLEAN_CALLEE, OPSZ_4));
 # endif
         } else {
             PRE(ilist, instr, XINST_CREATE_store(dcontext,
                 opnd_create_dcontext_field(dcontext, WHEREAMI_OFFSET),
-                OPND_CREATE_INT32(WHERE_CLEAN_CALLEE)));
+                OPND_CREATE_INT32(DR_WHERE_CLEAN_CALLEE)));
         }
     }
 #endif
@@ -480,9 +480,9 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
         uint whereami;
 
         if (TEST(META_CALL_RETURNS_TO_NATIVE, flags))
-            whereami = (uint) WHERE_APP;
+            whereami = (uint) DR_WHERE_APP;
         else
-            whereami = (uint) WHERE_FCACHE;
+            whereami = (uint) DR_WHERE_FCACHE;
 
         if (SCRATCH_ALWAYS_TLS()) {
             /* SCRATCH_REG0 is dead here: restore of the app stack will clobber xax */

diff --git a/core/arch/sideline.c b/core/arch/sideline.c
@@ -591,7 +591,7 @@ sideline_optimize(fragment_t *f,
     pause_for_sideline = dcontext->owning_thread;
     ASSERT(is_thread_known(pause_for_sideline));
 
-    if (dcontext->whereami != WHERE_FCACHE) {
+    if (dcontext->whereami != DR_WHERE_FCACHE) {
         /* wait for thread to reach waiting point in dispatch */
         LOG(logfile, LOG_SIDELINE, VERB_3,
             "\nsideline_optimize: waiting for target thread "TIDFMT"\n",

diff --git a/core/arch/x86_code.c b/core/arch/x86_code.c
@@ -410,7 +410,7 @@ nt_continue_setup(priv_mcontext_t *mc)
     dcontext->next_tag = next_pc;
     ASSERT(dcontext->next_tag != NULL);
     set_last_exit(dcontext, (linkstub_t *) get_asynch_linkstub());
-    dcontext->whereami = WHERE_TRAMPOLINE;
+    dcontext->whereami = DR_WHERE_TRAMPOLINE;
 
     *get_mcontext(dcontext) = *mc;
     /* clear pc */

diff --git a/core/dispatch.c b/core/dispatch.c
@@ -147,7 +147,7 @@ dispatch(dcontext_t *dcontext)
 # endif
     ASSERT(dcontext == get_thread_private_dcontext() ||
            /* i#813: the app hit our post-sysenter hook while native */
-           (dcontext->whereami == WHERE_APP &&
+           (dcontext->whereami == DR_WHERE_APP &&
             dcontext->last_exit == get_syscall_linkstub()));
 #else
 # ifdef UNIX
@@ -517,7 +517,7 @@ dispatch_enter_fcache(dcontext_t *dcontext, fragment_t *targetf)
          * to be safe for unlinking.
          */
         KSTOP_NOT_MATCHING(fcache_default);
-        dcontext->whereami = WHERE_DISPATCH;
+        dcontext->whereami = DR_WHERE_DISPATCH;
         enter_couldbelinking(dcontext, NULL, true);
         dcontext->next_tag = dcontext->asynch_target;
         LOG(THREAD, LOG_DISPATCH, 2,
@@ -569,7 +569,7 @@ enter_fcache(dcontext_t *dcontext, fcache_enter_func_t entry, cache_pc pc)
     }
 #endif
 
-    dcontext->whereami = WHERE_FCACHE;
+    dcontext->whereami = DR_WHERE_FCACHE;
     (*entry)(dcontext);
     IF_WINDOWS(ASSERT_NOT_REACHED()); /* returns for signals on unix */
 }
@@ -711,7 +711,7 @@ dispatch_enter_native(dcontext_t *dcontext)
 #endif
     }
     set_fcache_target(dcontext, dcontext->next_tag);
-    dcontext->whereami = WHERE_APP;
+    dcontext->whereami = DR_WHERE_APP;
 #ifdef UNIX
     do {
         (*go_native)(dcontext);
@@ -731,14 +731,14 @@ static void
 dispatch_enter_dynamorio(dcontext_t *dcontext)
 {
     /* We're transitioning to DynamoRIO from somewhere: either the fcache,
-     * the kernel (WHERE_TRAMPOLINE), or the app itself via our start/stop API.
-     * N.B.: set whereami to WHERE_APP iff this is the first dispatch() entry
+     * the kernel (DR_WHERE_TRAMPOLINE), or the app itself via our start/stop API.
+     * N.B.: set whereami to DR_WHERE_APP iff this is the first dispatch() entry
      * for this thread!
      */
-    where_am_i_t wherewasi = dcontext->whereami;
+    dr_where_am_i_t wherewasi = dcontext->whereami;
 #ifdef UNIX
-    if (!(wherewasi == WHERE_FCACHE || wherewasi == WHERE_TRAMPOLINE ||
-          wherewasi == WHERE_APP)) {
+    if (!(wherewasi == DR_WHERE_FCACHE || wherewasi == DR_WHERE_TRAMPOLINE ||
+          wherewasi == DR_WHERE_APP)) {
         /* This is probably our own syscalls hitting our own sysenter
          * hook (PR 212570), since we're not completely user library
          * independent (PR 206369).
@@ -756,9 +756,9 @@ dispatch_enter_dynamorio(dcontext_t *dcontext)
                        "DR's own syscall (via user library) hit the sysenter hook");
     }
 #endif
-    ASSERT(wherewasi == WHERE_FCACHE || wherewasi == WHERE_TRAMPOLINE ||
-           wherewasi == WHERE_APP);
-    dcontext->whereami = WHERE_DISPATCH;
+    ASSERT(wherewasi == DR_WHERE_FCACHE || wherewasi == DR_WHERE_TRAMPOLINE ||
+           wherewasi == DR_WHERE_APP);
+    dcontext->whereami = DR_WHERE_DISPATCH;
     ASSERT_LOCAL_HEAP_UNPROTECTED(dcontext);
     ASSERT(check_should_be_protected(DATASEC_RARELY_PROT));
     /* CANNOT hold any locks across cache execution, as our thread synch
@@ -775,7 +775,7 @@ dispatch_enter_dynamorio(dcontext_t *dcontext)
 #endif
 
     DOLOG(2, LOG_INTERP, {
-        if (wherewasi == WHERE_APP) {
+        if (wherewasi == DR_WHERE_APP) {
             LOG(THREAD, LOG_INTERP, 2, "\ninitial dispatch: target = "PFX"\n",
                 dcontext->next_tag);
             dump_mcontext_callstack(dcontext);
@@ -791,7 +791,7 @@ dispatch_enter_dynamorio(dcontext_t *dcontext)
      * messy that we're violating assumption of no ptrs...
      */
 
-    if (wherewasi == WHERE_APP) { /* first entrance */
+    if (wherewasi == DR_WHERE_APP) { /* first entrance */
         if (dcontext->last_exit == get_syscall_linkstub()) {
             /* i#813: the app hit our post-sysenter hook while native.
              * XXX: should we try to process ni syscalls here?  But we're only
@@ -870,7 +870,7 @@ dispatch_enter_dynamorio(dcontext_t *dcontext)
     /* KSWITCHed next time around for a better explanation */
     KSTART_DC(dcontext, dispatch_num_exits);
 
-    if (wherewasi != WHERE_APP) { /* if not first entrance */
+    if (wherewasi != DR_WHERE_APP) { /* if not first entrance */
         if (get_at_syscall(dcontext))
             handle_post_system_call(dcontext);
 
@@ -967,7 +967,7 @@ dispatch_enter_dynamorio(dcontext_t *dcontext)
         ASSERT(LINKSTUB_FAKE(dcontext->last_exit));
     }
 
-    if (wherewasi != WHERE_APP) { /* if not first entrance */
+    if (wherewasi != DR_WHERE_APP) { /* if not first entrance */
         /* now fully process the last cache exit as couldbelinking */
         dispatch_exit_fcache(dcontext);
     }
@@ -2047,7 +2047,7 @@ handle_system_call(dcontext_t *dcontext)
             /* avoid synch errors with dispatch -- since enter_fcache will set
              * whereami for prev dcontext, not real one!
              */
-            tmp_dcontext->whereami = WHERE_FCACHE;
+            tmp_dcontext->whereami = DR_WHERE_FCACHE;
         }
 #endif
 
@@ -2085,7 +2085,7 @@ handle_system_call(dcontext_t *dcontext)
         if (dcontext->signals_pending) {
             /* i#2019: see comments in dispatch_enter_fcache() */
             KSTOP(syscall_fcache);
-            dcontext->whereami = WHERE_DISPATCH;
+            dcontext->whereami = DR_WHERE_DISPATCH;
             set_at_syscall(dcontext, false);
             /* We need to remember both the post-syscall resumption point and
              * the fact that we need to execute a syscall, but we only have
@@ -2237,7 +2237,7 @@ handle_callback_return(dcontext_t *dcontext)
     SELF_PROTECT_LOCAL(dcontext, READONLY);
 
     /* obey flushing protocol, plus set whereami (both using real dcontext) */
-    dcontext->whereami = WHERE_FCACHE;
+    dcontext->whereami = DR_WHERE_FCACHE;
     set_at_syscall(dcontext, true); /* will be set to false on other end's post-syscall */
     ASSERT(!is_couldbelinking(dcontext));
 

diff --git a/core/dynamo.c b/core/dynamo.c
@@ -1704,7 +1704,7 @@ initialize_dynamo_context(dcontext_t *dcontext)
      */
     memset(dcontext->upcontext_ptr, 0, sizeof(unprotected_context_t));
     dcontext->initialized = true;
-    dcontext->whereami = WHERE_APP;
+    dcontext->whereami = DR_WHERE_APP;
     dcontext->next_tag = NULL;
     dcontext->native_exec_postsyscall = NULL;
     memset(dcontext->native_retstack, 0, sizeof(dcontext->native_retstack));

diff --git a/core/fcache.c b/core/fcache.c
@@ -1,5 +1,5 @@
 /* **********************************************************
- * Copyright (c) 2011-2017 Google, Inc.  All rights reserved.
+ * Copyright (c) 2011-2018 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
  * **********************************************************/
 
@@ -1176,6 +1176,45 @@ fcache_fragment_pclookup(dcontext_t *dcontext, cache_pc lookup_pc, fragment_t *w
     return found;
 }
 
+/* This is safe to call from a signal handler. */
+dr_where_am_i_t
+fcache_refine_whereami(dcontext_t *dcontext, dr_where_am_i_t whereami, app_pc pc,
+                       OUT fragment_t **containing_fragment)
+{
+    if (whereami != DR_WHERE_FCACHE) {
+        if (containing_fragment != NULL)
+            *containing_fragment = NULL;
+        return whereami;
+    }
+    fragment_t wrapper;
+    fragment_t *fragment = fragment_pclookup(dcontext, pc, &wrapper);
+    if (fragment == NULL) {
+        /* Since we're DR_WHERE_FCACHE, our locks shouldn't be held.
+         * XXX: we could double-check fcache_unit_areas.lock before
+         * calling (case 1317) and assert on it.
+         */
+        if (in_fcache(pc)) {
+            whereami = DR_WHERE_UNKNOWN;
+        } else {
+            /* Identify parts of our assembly code now.
+             * It's all generated and post-process can't identify.
+             * Assume code order is as follows:
+             */
+            if (in_context_switch_code(dcontext, (cache_pc)pc)) {
+                whereami = DR_WHERE_CONTEXT_SWITCH;
+            } else if (in_indirect_branch_lookup_code(dcontext,
+                                                      (cache_pc)pc)) {
+                whereami = DR_WHERE_IBL;
+            } else {
+                whereami = DR_WHERE_UNKNOWN;
+            }
+        }
+    }
+    if (containing_fragment != NULL)
+        *containing_fragment = fragment;
+    return whereami;
+}
+
 #ifdef DEBUG
 static bool
 fcache_pc_in_live_unit(fcache_t *cache, cache_pc pc)

diff --git a/core/fcache.h b/core/fcache.h
@@ -1,4 +1,5 @@
 /* **********************************************************
+ * Copyright (c) 2018 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2008 VMware, Inc.  All rights reserved.
  * **********************************************************/
 
@@ -141,6 +142,11 @@ bool fragment_lookup_deleted(dcontext_t *dcontext, app_pc tag);
 fragment_t *
 fcache_fragment_pclookup(dcontext_t *dcontext, cache_pc lookup_pc, fragment_t *wrapper);
 
+/* This is safe to call from a signal handler. */
+dr_where_am_i_t
+fcache_refine_whereami(dcontext_t *dcontext, dr_where_am_i_t whereami, app_pc pc,
+                       OUT fragment_t **containing_fragment);
+
 void
 fcache_coarse_cache_delete(dcontext_t *dcontext, coarse_info_t *info);
 

diff --git a/core/globals.h b/core/globals.h
@@ -1,5 +1,5 @@
 /* **********************************************************
- * Copyright (c) 2011-2017 Google, Inc.  All rights reserved.
+ * Copyright (c) 2011-2018 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
  * **********************************************************/
 
@@ -598,25 +598,30 @@ extern mutex_t bb_building_lock;
 extern volatile bool bb_lock_start;
 extern recursive_lock_t change_linking_lock;
 
-/* where the current app thread's control is */
+/* DR_API EXPORT BEGIN */
+/**
+ * Identifies where a thread's control is at any one point.
+ * Used with client PC sampling using dr_set_itimer().
+ */
 typedef enum {
-    WHERE_APP=0,
-    WHERE_INTERP,
-    WHERE_DISPATCH,
-    WHERE_MONITOR,
-    WHERE_SYSCALL_HANDLER,
-    WHERE_SIGNAL_HANDLER,
-    WHERE_TRAMPOLINE,
-    WHERE_CONTEXT_SWITCH,
-    WHERE_IBL,
-    WHERE_FCACHE,
-    WHERE_CLEAN_CALLEE,
-    WHERE_UNKNOWN,
+    DR_WHERE_APP=0,            /**< Control is in native application code. */
+    DR_WHERE_INTERP,           /**< Control is in basic block building. */
+    DR_WHERE_DISPATCH,         /**< Control is in dispatch. */
+    DR_WHERE_MONITOR,          /**< Control is in trace building. */
+    DR_WHERE_SYSCALL_HANDLER,  /**< Control is in system call handling. */
+    DR_WHERE_SIGNAL_HANDLER,   /**< Control is in signal handling. */
+    DR_WHERE_TRAMPOLINE,       /**< Control is in trampoline hooks. */
+    DR_WHERE_CONTEXT_SWITCH,   /**< Control is in context switching. */
+    DR_WHERE_IBL,              /**< Control is in inlined indirect branch lookup. */
+    DR_WHERE_FCACHE,           /**< Control is in the code cache. */
+    DR_WHERE_CLEAN_CALLEE,     /**< Control is in a clean call. */
+    DR_WHERE_UNKNOWN,          /**< Control is in an unknown location. */
 #ifdef HOT_PATCHING_INTERFACE
-    WHERE_HOTPATCH,
+    DR_WHERE_HOTPATCH,         /**< Control is in hotpatching. */
 #endif
-    WHERE_LAST
-} where_am_i_t;
+    DR_WHERE_LAST              /**< Equals the count of DR_WHERE_xxx locations. */
+} dr_where_am_i_t;
+/* DR_API EXPORT END */
 
 /* make args easier to read for protection change calls
  * since only two possibilities not using new type
@@ -786,7 +791,7 @@ struct _dcontext_t {
         coarse_info_t *dir_exit;
     } coarse_exit;
 
-    where_am_i_t   whereami;        /* where control is at the moment */
+    dr_where_am_i_t   whereami;        /* where control is at the moment */
 #ifdef UNIX
     char           signals_pending; /* != 0: pending; < 0: currently handling one */
 #endif