NativeStackSampler implementation for Mac.

base/profiler/native_stack_sampler_mac.cc

+// Copyright 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include <dlfcn.h>
+#include <mach/mach.h>
+#include <mach/thread_status.h>
+#include <mach-o/swap.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <map>
+
+#include "base/logging.h"
+#include "base/md5.h"
+#include "base/profiler/native_stack_sampler.h"
+#include "base/strings/string_number_conversions.h"
+#include "base/time/time.h"
+
+namespace base {
+
+namespace {
+

[robliao, 2015/09/14 21:58:17]

Nit: Remove extra line.

[sydli, 2015/09/14 23:00:13]

Done.

+
+// Copy of x86_64 thread context structure from x86_thread_state64_t type.
+// Copied struct since fields can have different names on different versions of
+// Darwin.
+struct ThreadContext {
+  uint64_t rax;
+  uint64_t rbx;
+  uint64_t rcx;
+  uint64_t rdx;
+  uint64_t rdi;
+  uint64_t rsi;
+  uint64_t rbp;
+  uint64_t rsp;
+  uint64_t r8;
+  uint64_t r9;
+  uint64_t r10;
+  uint64_t r11;
+  uint64_t r12;
+  uint64_t r13;
+  uint64_t r14;
+  uint64_t r15;
+  uint64_t rip;
+  uint64_t rflags;
+  uint64_t cs;
+  uint64_t fs;
+  uint64_t gs;
+};
+
+// Struct for stack walking (represents stack state on any function call that
+// pushes a frame pointer).
+struct StackFrame {
+  // Pointer to caller's frame (rbp).
+  struct StackFrame* prev;
+  // Address in caller for callee to return to.
+  uintptr_t return_addr;
+};
+
+// Overwrites |dst| with StackFrame referenced in |src|. Returns true if stack
+// jump was successful. Failure implies address in |src| is corrupt.
+bool SafeJump(const StackFrame* src, StackFrame* dst) {

[robliao, 2015/09/14 21:58:17]

Can src be a const StackFrame& ?

[sydli, 2015/09/14 23:00:13]

Don't think so; I pass in frame.prev (which is a StackFrame*) as src, which I
don't want to dereference until vm_read_overwrite.

+  vm_size_t bytesCopied = 0;

[robliao, 2015/09/14 21:58:16]

ignored_bytes_copied

[sydli, 2015/09/14 23:00:13]

Done.

+  return vm_read_overwrite(mach_task_self(), (vm_address_t)src,
+                           (vm_size_t)sizeof(StackFrame), (vm_address_t)dst,
+                           &bytesCopied) == KERN_SUCCESS;
+}
+
+// Functions related to retrieving Mach-O Identifer ----------------------------

[robliao, 2015/09/14 21:58:17]

Remove trailing -'s

[sydli, 2015/09/14 23:00:13]

Done.

+//
+// These functions were cannibalized from Mach-O Identifier procedures found in
+// breakpad/src/common/mac. Support for non-X86_64 architectures and MD5 IDs
+// were removed to simplify the code.
+
+// Returns offset in bytes where the x86_64 header is located in binary
+// loaded at |module_addr|. Returns 0 if |module_addr| is not a valid FAT
+// Mach-O binary or has not been built for x86_64.
+off_t GetMach64HeaderOffset(const void* module_addr) {
+  fat_header hdr;
+  memcpy(&hdr, module_addr, sizeof(hdr));

[robliao, 2015/09/14 21:58:17]

Safety: Where do we get the assurance that module_addr is at least
sizeof(fat_header)?

[sydli, 2015/09/14 23:00:13]

This is an assumption we make since |module_addr| is supposedly where a module
is loaded into memory. Unsure what would be the best way to check this besides
checking the results of dladdr (where |module_addr| comes from); any
recommendations?

[robliao, 2015/09/14 23:28:05]

The sections don't appear to have size info, so I guess we'll have to stick with
this.

+  if (hdr.magic != FAT_MAGIC && hdr.magic != FAT_CIGAM)
+    return 0;
+
+  if (hdr.magic == FAT_CIGAM)
+    swap_fat_header(&hdr, NXHostByteOrder());
+
+  // Search all FAT architectures for x86_64.
+  off_t offset = sizeof(hdr);
+  fat_arch arch;
+  for (uint32_t i = 0; i < hdr.nfat_arch; ++i) {
+    memcpy(&arch, (const char*)module_addr + offset, sizeof(arch));

[robliao, 2015/09/14 21:58:17]

Would casting module_addr+offset to a fat_arch[] work?
At minimum, use reintrepret_cast<const char*>() from here and forwards.

[sydli, 2015/09/14 23:00:13]

module_addr has to be a pointer of a byte-wide type so we can offset exactly
|offset| bytes. Afterwards we can cast to fat_arch[] if needed. What is the
standard pointer type to use for arithmetic?

[robliao, 2015/09/14 23:28:05]

const char* is likely what you want since you're operating off of the assumption
that you're working with bytes here.

You could also use uint8_t*

+
+    if (NXHostByteOrder() != NX_BigEndian)
+      swap_fat_arch(&arch, 1, NXHostByteOrder());
+
+    if (arch.cputype == CPU_TYPE_X86_64)
+      return arch.offset;
+    offset += sizeof(arch);
+  }
+  return 0;
+}
+
+// Returns true if Mach-O binary at |module_addr| was built specifically for
+// x86_64 cpu architecture.
+bool IsX64Header(const void* module_addr) {
+  mach_header_64 hdr;
+  memcpy(&hdr, module_addr, sizeof(hdr));
+  if (hdr.magic != MH_MAGIC_64 && hdr.magic != MH_CIGAM_64)
+    return false;
+  if (hdr.magic == MH_CIGAM_64)
+    swap_mach_header_64(&hdr, NXHostByteOrder());
+
+  if (hdr.cputype != CPU_TYPE_X86_64)

[robliao, 2015/09/14 21:58:16]

return hdr.cputype == CPU_TYPE_X86_64;

[sydli, 2015/09/14 23:00:13]

Done.

+    return false;
+  return true;
+}
+
+// Fills |id| with UUID of x86_64 Mach-O binary loaded at |module_addr|.

[robliao, 2015/09/14 21:58:17]

Nit: {the} UUID of {the} x86_64

[sydli, 2015/09/14 23:00:13]

Done.

+// |offset| is the offset in bytes into |module_addr| where the x86_64 header

[robliao, 2015/09/14 21:58:17]

It might be instructive to do some ASCII Art towards the top, something like

+ <-- module_addr
|
| <-- + Offset, where the headers start.
.
.
.

[sydli, 2015/09/14 23:00:13]

This is only accurate for FAT Mach-O headers. Updated to comment that offset is
only relevant if binary is FAT. In this case, should I diagram both layouts? I
linked the relevant spec as well.

[robliao, 2015/09/14 23:28:05]

The link is good, but the diagram is more helpful in case the link goes bad.

[sydli, 2015/09/15 00:20:13]

Ok.

+// is located. |offset| should be non-zero if binary is FAT and contains
+// multiple architecture headers. Returns false if header is malformed or the
+// header does not specify the UUID load command.
+bool GetX64UUIDAt(const void* module_addr, unsigned char id[16], off_t offset) {
+  mach_header_64 hdr;
+  if (!memcpy(&hdr, (const char*)module_addr + offset, sizeof(hdr)) ||
+      (hdr.magic != MH_MAGIC_64 && hdr.magic != MH_CIGAM_64))
+    return false;

[robliao, 2015/09/14 21:58:17]

When a conditional spans multiple lines, use braces.

[sydli, 2015/09/14 23:00:13]

Done.

[robliao, 2015/09/14 23:28:05]

I assume this is the one you forgot?

[sydli, 2015/09/15 00:20:13]

Yep! Added it in new patchset.

+
+  bool swap = hdr.magic == MH_CIGAM_64;
+  if (swap)
+    swap_mach_header_64(&hdr, NXHostByteOrder());
+
+  // Search all load commands for UUID command.
+  offset += sizeof(hdr);
+  for (uint32_t i = 0; i < hdr.ncmds; ++i) {
+    load_command cmd;

[robliao, 2015/09/14 21:58:17]

Move cmd outside the loop to avoid constructing this hdr.ncmds times.

[sydli, 2015/09/14 23:00:13]

Done.

+    memcpy(&cmd, (const char*)module_addr + offset, sizeof(cmd));

[robliao, 2015/09/14 21:58:16]

Could represent module_addr + offset as a temp value so that you don't have to
copy it again below.

[sydli, 2015/09/14 23:00:13]

Done.

+
+    if (swap)
+      swap_load_command(&cmd, NXHostByteOrder());
+    if (cmd.cmd == LC_UUID) {
+      uuid_command uuid_cmd;
+      memcpy(&uuid_cmd, (const char*)module_addr + offset, sizeof(uuid_cmd));
+      if (swap)
+        swap_uuid_command(&uuid_cmd, NXHostByteOrder());
+      memcpy(id, &uuid_cmd, sizeof(uuid_cmd));

[robliao, 2015/09/14 21:58:17]

Safety: This copy probably should be no longer than sizeof(id).
Might be useful to have a static assert that sizeof(uuid_command) == sizeof(id)

[sydli, 2015/09/14 23:00:13]

Done.

+      return true;
+    }
+    offset += cmd.cmdsize;
+  }
+  return false;
+}
+
+// Fills |id| with Mach-O UUID retrieved from Mach-O binary loaded at
+// |module_addr|. This function returns false if the binary was not built for
+// X86_64 or if UUID cannot be found.
+bool GetUUID(const void* module_addr, unsigned char id[16]) {
+  off_t offset = 0;
+  if (!IsX64Header(module_addr) &&
+      !(offset = GetMach64HeaderOffset(module_addr)))
+    return false;
+  return GetX64UUIDAt(module_addr, id, offset);
+}
+
+// Returns hex encoding of a 16-byte ID for binary loaded at |module_addr|.
+// Returns empty string if UUID cannot be found at |module_addr|.
+std::string GetUniqueId(const void* module_addr) {
+  unsigned char id[16];
+  if (!GetUUID(module_addr, id))
+    return "";
+  return HexEncode(id, 16);
+}
+
+// Functions related to grabbing a stack trace --------------------------------
+
+// Fills |state| with |target_thread|'s context.
+bool GetThreadState(thread_act_t target_thread, ThreadContext* state) {
+  mach_msg_type_number_t count =
+      static_cast<mach_msg_type_number_t>(MACHINE_THREAD_STATE_COUNT);
+  return thread_get_state(target_thread, x86_THREAD_STATE64,
+                          reinterpret_cast<thread_state_t>(state),
+                          &count) == KERN_SUCCESS;
+}
+
+// Walks |thread_handle|'s stack and fills |instruction_pointers|.
+// Returns number of frames in stack, unless there's a corrupt ebp, in which

[robliao, 2015/09/14 21:58:17]

corrupt ebp or missing ebp or both?

[sydli, 2015/09/14 23:00:13]

A corrupt base pointer implies that it is missing (unless something is horribly
wrong). Updated comment.

+// this function will return the number of frames up until the frame with
+// a corrupt ebp. This procedure occurs while thread is suspended, so it should
+// take as little time as possible.
+int RecordStack(mach_port_t thread_handle,
+                int max_stack_size,

[robliao, 2015/09/14 21:58:17]

Maybe instruction_pointers_size to make it obvious that this describes the size
of the instruction_pointers array?

[sydli, 2015/09/14 23:00:14]

Done.

+                const void* instruction_pointers[]) {

[robliao, 2015/09/14 21:58:17]

Should this be const uintptr_t?

[sydli, 2015/09/14 23:00:13]

Done. Changed where appropriate.

+  ThreadContext state;
+  if (!GetThreadState(thread_handle, &state))
+    return 0;
+
+  StackFrame frame;
+  frame.prev = reinterpret_cast<StackFrame*>(state.rbp);
+  int i = 0;
+  for (; i < max_stack_size; i++) {
+    // Three cases for end-of-stack condition:
+    //   1) We tried to jump off the stack.
+    //   2) We jumped to a lower address.
+    //   3) We reached a return (instruction) address of 0.
+    StackFrame* old_esp = frame.prev;
+    if (!SafeJump(frame.prev, &frame) || frame.prev < old_esp ||
+        !frame.return_addr)
+      return i;

[robliao, 2015/09/14 21:58:17]

When a conditional spans multiple lines, add braces around the single statement.

[sydli, 2015/09/14 23:00:13]

Done.

+    instruction_pointers[i] = reinterpret_cast<const void*>(frame.return_addr);
+  }
+  return i;
+}
+
+// Adds library referenced by |instruction_pointer| to |sample| and |modules|.
+// Records library's (1) filepath, (2) base address, and (3) unique ID.
+void AddModule(StackSamplingProfiler::Sample* sample,
+               std::vector<StackSamplingProfiler::Module>* modules,
+               std::map<void*, size_t>* module_to_index,
+               const void* instruction_pointer) {
+  char filepath[PATH_MAX + 1];
+  Dl_info inf;
+  dladdr(instruction_pointer, &inf);
+  auto loc = module_to_index->find(inf.dli_fbase);

[robliao, 2015/09/14 21:58:16]

What is loc?

[sydli, 2015/09/14 23:00:13]

Pair result from module -> index map. Updated name to module_index.

+  if (loc == module_to_index->end()) {
+    realpath(inf.dli_fname, filepath);
+    StackSamplingProfiler::Module module(
+        inf.dli_fbase, GetUniqueId(inf.dli_fbase), base::FilePath(filepath));
+    modules->push_back(module);
+
+    loc = module_to_index->insert(

[robliao, 2015/09/14 21:58:17]

Linebreak after the -> and reflow.

[sydli, 2015/09/14 23:00:13]

Done.

+                             std::make_pair(inf.dli_fbase, modules->size() - 1))
+              .first;
+  }
+  sample->push_back(
+      StackSamplingProfiler::Frame(instruction_pointer, loc->second));
+}
+
+// Fills |sample| with Frames and Modules referenced by |instruction_pointers|.
+void FillSample(const void* instruction_pointers[],
+                std::vector<StackSamplingProfiler::Module>* modules,
+                int stack_depth,
+                StackSamplingProfiler::Sample* sample) {
+  std::map<void*, size_t> module_index;
+  for (int i = 0; i < stack_depth; i++) {
+    AddModule(sample, modules, &module_index, instruction_pointers[i]);
+  }
+}
+
+// NativeStackSamplerMac ------------------------------------------------------
+
+class NativeStackSamplerMac : public NativeStackSampler {
+ public:
+  explicit NativeStackSamplerMac(pid_t thread_handle);
+  ~NativeStackSamplerMac() override;
+
+  // StackSamplingProfiler::NativeStackSampler:
+  void ProfileRecordingStarting(
+      std::vector<StackSamplingProfiler::Module>* modules) override;
+  void RecordStackSample(StackSamplingProfiler::Sample* sample) override;
+  void ProfileRecordingStopped() override;
+
+ private:
+  mach_port_t thread_handle_;
+
+  // Weak. Points to the modules associated with the profile being recorded
+  // between ProfileRecordingStarting() and ProfileRecordingStopped().
+  std::vector<StackSamplingProfiler::Module>* current_modules_;
+
+  DISALLOW_COPY_AND_ASSIGN(NativeStackSamplerMac);
+};
+
+// The PlatformThreadId of the given thread is actually a typedef of
+// mach_port_t.
+// (base/threading/platform_thread_posix.cc:128)
+NativeStackSamplerMac::NativeStackSamplerMac(pid_t thread_handle)
+    : thread_handle_(static_cast<mach_port_t>(thread_handle)) {}
+
+NativeStackSamplerMac::~NativeStackSamplerMac() {}
+
+void NativeStackSamplerMac::ProfileRecordingStarting(
+    std::vector<StackSamplingProfiler::Module>* modules) {
+  current_modules_ = modules;
+}
+
+void NativeStackSamplerMac::RecordStackSample(
+    StackSamplingProfiler::Sample* sample) {
+  DCHECK(current_modules_);
+  const int max_stack_size = 64;

[robliao, 2015/09/14 21:58:17]

MAX_STACK_SIZE. Consider moving this to the top of the namespace.

[sydli, 2015/09/14 23:00:13]

Done. Made this a static const in NativeStackSamplerMac.

+  const void* instruction_pointers[max_stack_size] = {0};
+
+  thread_suspend(thread_handle_);
+  int stack_depth =
+      RecordStack(thread_handle_, max_stack_size, instruction_pointers);
+  thread_resume(thread_handle_);
+  FillSample(instruction_pointers, current_modules_, stack_depth, sample);
+}
+
+void NativeStackSamplerMac::ProfileRecordingStopped() {
+  current_modules_ = nullptr;
+}
+
+}  // namespace
+
+scoped_ptr<NativeStackSampler> NativeStackSampler::Create(
+    PlatformThreadId thread_id) {
+#if defined(OS_MACOSX) && !defined(OS_NACL)
+  return scoped_ptr<NativeStackSampler>(new NativeStackSamplerMac(thread_id));
+#endif
+  return scoped_ptr<NativeStackSampler>();
+}
+
+}  // namespace base