Better testing for Stackwalker (#18). r=bryner

 - Test StackwalkerX86 and StackwalkerPPC on the current process, if built
   by a supported compiler (gcc) on a supported (walkable) CPU (x86, ppc).
 - This test is not enabled by default because of certain optimizations
   that interfere with it (stack frame reuse, frame pointer omission).  See
   the comments at the top of stackwalker_selftest.cc.  To enable this
   test in the standard "make check" suite, configure with --enable-selftest.

http://groups.google.com/group/airbag-dev/browse_thread/thread/2ca1c8e72c809bec


git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@35 4c0a9323-5329-0410-9bdc-e9ce6186880e

1 files changed, 252 insertions, 0 deletions

diff --git a/src/processor/stackwalker_selftest.cc b/src/processor/stackwalker_selftest.cc
new file mode 100644
index 00000000..08dea134
--- /dev/null
+++ b/src/processor/stackwalker_selftest.cc
@@ -0,0 +1,252 @@
+// Copyright (C) 2006 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// stackwalker_selftest.cc: Tests StackwalkerX86 or StackwalkerPPC using the
+// running process' stack as test data, if running on an x86 or ppc and
+// compiled with gcc.  This test is not enabled in the "make check" suite
+// by default, because certain optimizations interfere with its proper
+// operation.  To turn it on, configure with --enable-selftest.
+//
+// Optimizations that cause problems:
+//  - stack frame reuse.  The Recursor function here calls itself with
+//    |return Recursor|.  When the caller's frame is reused, it will cause
+//    CountCallerFrames to correctly return the same number of frames
+//    in both the caller and callee.  This is considered an unexpected
+//    condition in the test, which expects a callee to have one more
+//    caller frame in the stack than its caller.
+//  - frame pointer omission.  Even with a stackwalker that understands
+//    this optimization, the code to harness debug information currently
+//    only exists to retrieve it from minidumps, not the current process.
+//
+// This test can also serve as a developmental and debugging aid if
+// PRINT_STACKS is defined.
+//
+// Author: Mark Mentovai
+
+#if defined(__GNUC__) && (defined(__i386__) || defined(__ppc__))
+
+
+#include <cstdio>
+
+#include "google/airbag_types.h"
+#include "google/stack_frame.h"
+#include "processor/memory_region.h"
+#include "processor/minidump_format.h"
+
+using google_airbag::MemoryRegion;
+using google_airbag::StackFrame;
+using google_airbag::StackFrames;
+
+#if defined(__i386__)
+#include "processor/stackwalker_x86.h"
+using google_airbag::StackwalkerX86;
+#elif defined(__ppc__)
+#include "processor/stackwalker_ppc.h"
+using google_airbag::StackwalkerPPC;
+#endif  // __i386__ || __ppc__
+
+#define RECURSION_DEPTH 100
+
+
+// A simple MemoryRegion subclass that provides direct access to this
+// process' memory space by pointer.
+class SelfMemoryRegion : public MemoryRegion {
+ public:
+  virtual u_int64_t GetBase() { return 0; }
+  virtual u_int32_t GetSize() { return 0xffffffff; }
+
+  bool GetMemoryAtAddress(u_int64_t address, u_int8_t*  value) {
+      return GetMemoryAtAddressInternal(address, value); }
+  bool GetMemoryAtAddress(u_int64_t address, u_int16_t* value) {
+      return GetMemoryAtAddressInternal(address, value); }
+  bool GetMemoryAtAddress(u_int64_t address, u_int32_t* value) {
+      return GetMemoryAtAddressInternal(address, value); }
+  bool GetMemoryAtAddress(u_int64_t address, u_int64_t* value) {
+      return GetMemoryAtAddressInternal(address, value); }
+
+ private:
+  template<typename T> bool GetMemoryAtAddressInternal(u_int64_t address,
+                                                       T*        value) {
+    u_int8_t* memory = 0;
+    *value = *reinterpret_cast<const T*>(&memory[address]);
+    return true;
+  }
+};
+
+
+#if defined(__i386__)
+
+// GetEBP returns the current value of the %ebp register.  Because it's
+// implemented as a function, %ebp itself contains GetEBP's frame pointer
+// and not the caller's frame pointer.  Dereference %ebp to obtain the
+// caller's frame pointer, which the compiler-generated preamble stored
+// on the stack (provided frame pointers are not being omitted.)  Because
+// this function depends on the compiler-generated preamble, inlining is
+// disabled.
+static u_int32_t GetEBP() __attribute__((noinline));
+static u_int32_t GetEBP() {
+  u_int32_t ebp;
+  __asm__ __volatile__(
+    "movl (%%ebp), %0"
+    : "=a" (ebp)
+  );
+  return ebp;
+}
+
+
+// GetEIP returns the instruction pointer identifying the next instruction
+// to execute after GetEIP returns.  It obtains this information from the
+// stack, where it was placed by the call instruction that called GetEIP.
+// This function depends on frame pointers not being omitted.  It is possible
+// to write a pure asm version of this routine that has no compiler-generated
+// preamble and uses %esp instead of %ebp; that would function in the
+// absence of frame pointers.  However, the simpler approach is used here
+// because GetEBP and stackwalking necessarily depends on access to frame
+// pointers.  Because this function depends on a call instruction and the
+// compiler-generated preamble, inlining is disabled.
+static u_int32_t GetEIP() __attribute__((noinline));
+static u_int32_t GetEIP() {
+  u_int32_t eip;
+  __asm__ __volatile__(
+    "movl 4(%%ebp), %0"
+    : "=a" (eip)
+  );
+  return eip;
+}
+
+
+#elif defined(__ppc__)
+
+
+// GetSP returns the current value of the %r1 register, which by convention,
+// is the stack pointer on ppc.  Because it's implemented as a function,
+// %r1 itself contains GetSP's own stack pointer and not the caller's stack 
+// pointer.  Dereference %r1 to obtain the caller's stack pointer, which the
+// compiler-generated prolog stored on the stack.  Because this function
+// depends on the compiler-generated prolog, inlining is disabled.
+static u_int32_t GetSP() __attribute__((noinline));
+static u_int32_t GetSP() {
+  u_int32_t sp;
+  __asm__ __volatile__(
+    "lwz %0, 0(r1)"
+    : "=r" (sp)
+  );
+  return sp;
+}
+
+
+// GetPC returns the program counter identifying the next instruction to
+// execute after GetPC returns.  It obtains this information from the
+// link register, where it was placed by the branch instruction that called
+// GetPC.  Because this function depends on the caller's use of a branch
+// instruction, inlining is disabled.
+static u_int32_t GetPC() __attribute__((noinline));
+static u_int32_t GetPC() {
+  u_int32_t lr;
+  __asm__ __volatile__(
+    "mflr %0"
+    : "=r" (lr)
+  );
+  return lr;
+}
+
+
+#endif  // __i386__ || __ppc__
+
+
+// CountCallerFrames returns the number of stack frames beneath the function
+// that called CountCallerFrames.  Because this function's return value
+// is dependent on the size of the stack beneath it, inlining is disabled,
+// and any function that calls this should not be inlined either.
+static unsigned int CountCallerFrames() __attribute__((noinline));
+static unsigned int CountCallerFrames() {
+  SelfMemoryRegion memory;
+
+#if defined(__i386__)
+  MDRawContextX86 context = MDRawContextX86();
+  context.eip = GetEIP();
+  context.ebp = GetEBP();
+
+  StackwalkerX86 stackwalker = StackwalkerX86(&context, &memory, NULL, NULL);
+#elif defined(__ppc__)
+  MDRawContextPPC context = MDRawContextPPC();
+  context.srr0 = GetPC();
+  context.gpr[1] = GetSP();
+
+  StackwalkerPPC stackwalker = StackwalkerPPC(&context, &memory, NULL, NULL);
+#endif  // __i386__ || __ppc__
+
+  StackFrames stack;
+  stackwalker.Walk(&stack);
+
+#ifdef PRINT_STACKS
+  printf("\n");
+  for(unsigned int frame_index = 0;
+      frame_index < stack.size();
+      ++frame_index) {
+    StackFrame *frame = &stack[frame_index];
+    printf("frame %-3d  instruction = 0x%08llx  frame_pointer = 0x%08llx\n",
+           frame_index, frame->instruction, frame->frame_pointer);
+  }
+#endif  // PRINT_STACKS
+
+  // Subtract 1 because the caller wants the number of frames beneath
+  // itself.  Because the caller called us, subract two for our frame and its
+  // frame, which are included in stack->size().
+  return stack.size() - 2;
+}
+
+
+// Recursor verifies that the number stack frames beneath itself is one more
+// than the number of stack frames beneath its parent.  When depth frames
+// have been reached, Recursor stops checking and returns success.  If the
+// frame count check fails at any depth, Recursor will stop and return false.
+// Because this calls CountCallerFrames, inlining is disabled.
+static bool Recursor(unsigned int depth, unsigned int parent_callers)
+    __attribute__((noinline));
+static bool Recursor(unsigned int depth, unsigned int parent_callers) {
+  unsigned int callers = CountCallerFrames();
+  if (callers != parent_callers + 1)
+    return false;
+
+  if (depth)
+    return Recursor(depth - 1, callers);
+
+  // depth == 0
+  return true;
+}
+
+
+// Because this calls CountCallerFrames, inlining is disabled - but because
+// it's main (and nobody calls it other than the entry point), it wouldn't
+// be inlined anyway.
+int main(int argc, char** argv) __attribute__((noinline));
+int main(int argc, char** argv) {
+  return Recursor(RECURSION_DEPTH, CountCallerFrames()) ? 0 : 1;
+}
+
+
+#else  // __GNUC__ && (__i386__ || __ppc__)
+// Not gcc?  We use gcc's __asm__.
+// Not i386 or ppc?  We can only test stacks we know how to walk.
+
+
+int main(int argc, char **argv) {
+  // "make check" interprets an exit status of 77 to mean that the test is
+  // not supported.
+  return 77;
+}
+
+
+#endif  // __GNUC__ && (__i386__ || __ppc__)