Thu Nov 29 07:59:43 2007 Google Inc. <opensource@google.com>

	* google-perftools: version 0.94 release
	* PORTING: MinGW/Msys support -- runs same code as MSVC does (csilvers)
	* PORTING: Add NumCPUs support for Mac OS X (csilvers)
	* Work around a sscanf bug in glibc(?) (waldemar)
	* Fix Windows MSVC bug triggered by thread deletion (csilvers)
	* Fix bug that triggers in MSVC /O2: missing volatile (gpike)
	* March-of-time support: quiet warnings/errors for gcc 4.2, OS X 10.5
	* Modify pprof so it works without nm: useful for windows (csilvers)
	* pprof: Support filtering for CPU profiles (cgd)
	* Bugfix: have realloc report to hooks in all situations (maxim)
	* Speed improvement: replace slow memcpy with std::copy (soren)
	* Speed: better iterator efficiency in RecordRegionRemoval (soren)
	* Speed: minor speed improvements via better bitfield alignment (gpike)
	* Documentation: add documentation of binary profile output (cgd)


git-svn-id: http://gperftools.googlecode.com/svn/trunk@40 6b5cf1ce-ec42-a296-1ba9-69fdba395a50

1 files changed, 105 insertions, 5 deletions

diff --git a/src/stacktrace_x86-inl.h b/src/stacktrace_x86-inl.h
index 5dcbc7a..794aaf5 100644
--- a/src/stacktrace_x86-inl.h
+++ b/src/stacktrace_x86-inl.h
@@ -37,14 +37,29 @@
 #include "google/stacktrace.h"
 
 // Given a pointer to a stack frame, locate and return the calling
-// stackframe, or return NULL if no stackframe can be found. Perform
-// sanity checks to reduce the chance that a bad pointer is returned.
+// stackframe, or return NULL if no stackframe can be found. Perform sanity
+// checks (the strictness of which is controlled by the boolean parameter
+// "STRICT_UNWINDING") to reduce the chance that a bad pointer is returned.
+template<bool STRICT_UNWINDING>
 static void **NextStackFrame(void **old_sp) {
   void **new_sp = (void **) *old_sp;
 
   // Check that the transition from frame pointer old_sp to frame
   // pointer new_sp isn't clearly bogus
-  if (new_sp <= old_sp) return NULL;
+  if (STRICT_UNWINDING) {
+    // With the stack growing downwards, older stack frame must be
+    // at a greater address that the current one.
+    if (new_sp <= old_sp) return NULL;
+    // Assume stack frames larger than 100,000 bytes are bogus.
+    if ((uintptr_t)new_sp - (uintptr_t)old_sp > 100000) return NULL;
+  } else {
+    // In the non-strict mode, allow discontiguous stack frames.
+    // (alternate-signal-stacks for example).
+    if (new_sp == old_sp) return NULL;
+    // And allow frames upto about 1MB.
+    if ((new_sp > old_sp)
+        && ((uintptr_t)new_sp - (uintptr_t)old_sp > 1000000)) return NULL;
+  }
   if ((uintptr_t)new_sp & (sizeof(void *) - 1)) return NULL;
 #ifdef __i386__
   // On 64-bit machines, the stack pointer can be very close to
@@ -52,10 +67,10 @@ static void **NextStackFrame(void **old_sp) {
   // last two pages in the address space
   if ((uintptr_t)new_sp >= 0xffffe000) return NULL;
 #endif
-  if ((uintptr_t)new_sp - (uintptr_t)old_sp > 100000) return NULL;
   return new_sp;
 }
 
+// If you change this function, also change GetStackFrames below.
 int GetStackTrace(void** result, int max_depth, int skip_count) {
   void **sp;
 #ifdef __i386__
@@ -95,7 +110,92 @@ int GetStackTrace(void** result, int max_depth, int skip_count) {
     } else {
       result[n++] = *(sp+1);
     }
-    sp = NextStackFrame(sp);
+    // Use strict unwinding rules.
+    sp = NextStackFrame<true>(sp);
+  }
+  return n;
+}
+
+// If you change this function, also change GetStackTrace above:
+//
+// This GetStackFrames routine shares a lot of code with GetStackTrace
+// above. This code could have been refactored into a common routine,
+// and then both GetStackTrace/GetStackFrames could call that routine.
+// There are two problems with that:
+//
+// (1) The performance of the refactored-code suffers substantially - the
+//     refactored needs to be able to record the stack trace when called
+//     from GetStackTrace, and both the stack trace and stack frame sizes,
+//     when called from GetStackFrames - this introduces enough new
+//     conditionals that GetStackTrace performance can degrade by as much
+//     as 50%.
+//
+// (2) Whether the refactored routine gets inlined into GetStackTrace and
+//     GetStackFrames depends on the compiler, and we can't guarantee the
+//     behavior either-way, even with "__attribute__ ((always_inline))"
+//     or "__attribute__ ((noinline))". But we need this guarantee or the
+//     frame counts may be off by one.
+//
+// Both (1) and (2) can be addressed without this code duplication, by
+// clever use of template functions, and by defining GetStackTrace and
+// GetStackFrames as macros that expand to these template functions.
+// However, this approach comes with its own set of problems - namely,
+// macros and  preprocessor trouble - for example,  if GetStackTrace
+// and/or GetStackFrames is ever defined as a member functions in some
+// class, we are in trouble.
+int GetStackFrames(void** pcs, int* sizes, int max_depth, int skip_count) {
+  void **sp;
+#ifdef __i386__
+  // Stack frame format:
+  //    sp[0]   pointer to previous frame
+  //    sp[1]   caller address
+  //    sp[2]   first argument
+  //    ...
+  sp = (void **)&pcs - 2;
+#endif
+
+#ifdef __x86_64__
+  // __builtin_frame_address(0) can return the wrong address on gcc-4.1.0-k8
+  unsigned long rbp;
+  // Move the value of the register %rbp into the local variable rbp.
+  // We need 'volatile' to prevent this instruction from getting moved
+  // around during optimization to before function prologue is done.
+  // An alternative way to achieve this
+  // would be (before this __asm__ instruction) to call Noop() defined as
+  //   static void Noop() __attribute__ ((noinline));  // prevent inlining
+  //   static void Noop() { asm(""); }  // prevent optimizing-away
+  __asm__ volatile ("mov %%rbp, %0" : "=r" (rbp));
+  // Arguments are passed in registers on x86-64, so we can't just
+  // offset from &pcs
+  sp = (void **) rbp;
+#endif
+
+  int n = 0;
+  while (sp && n < max_depth) {
+    if (*(sp+1) == (void *)0) {
+      // In 64-bit code, we often see a frame that
+      // points to itself and has a return address of 0.
+      break;
+    }
+    // The GetStackFrames routine is called when we are in some
+    // informational context (the failure signal handler for example).
+    // Use the non-strict unwinding rules to produce a stack trace
+    // that is as complete as possible (even if it contains a few bogus
+    // entries in some rare cases).
+    void **next_sp = NextStackFrame<false>(sp);
+    if (skip_count > 0) {
+      skip_count--;
+    } else {
+      pcs[n] = *(sp+1);
+      if (next_sp > sp) {
+        sizes[n] = (uintptr_t)next_sp - (uintptr_t)sp;
+      } else {
+        // A frame-size of 0 is used to indicate unknown frame size.
+        sizes[n] = 0;
+      }
+      n++;
+    }
+    sp = next_sp;
   }
   return n;
 }