Track thread activities in order to diagnose hangs.

base/debug/activity_tracker.cc

+// Copyright 2016 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 "base/debug/activity_tracker.h"
+
+#include <atomic>
+
+#include "base/files/memory_mapped_file.h"
+#include "base/logging.h"
+#include "base/memory/ptr_util.h"
+#include "base/metrics/field_trial.h"
+#include "base/pending_task.h"
+#include "base/stl_util.h"
+
+namespace base {
+namespace debug {
+
+namespace {
+
+// A number that indetifies memory has having been initialized.

[manzagop (departed), 2016/05/20 18:19:29]

nit: typos

[bcwhite, 2016/05/20 19:19:18]

Done.  (I have done zero proof-reading so far.)

+const uint64_t kHeaderCookie = 0x98476A390137E67A + 1;  // v1

[manzagop (departed), 2016/05/20 18:19:29]

Comment on where this value comes from?

[bcwhite, 2016/05/20 19:19:18]

Done.

+
+// The minimum depth a stack should support.
+const int kMinStackDepth = 2;
+
+// Type identifiers used when storing in persistent memory so they can be
+// identified during extraction; the first 4 bytes of the SHA1 of the name
+// is used as a unique integer. A "version number" is added to the base
+// so that, if the structure of that object changes, stored older versions
+// will be safely ignored.
+enum : uint32_t {
+  kTypeIdActivityTracker     = 0x5D7381AF + 1,  // SHA1(ActivityTracker)  v1
+  kTypeIdActivityTrackerFree = 0x3F0272FB,      // SHA1(ActivityTrackerFree)
+};
+
+}  // namespace
+
+const char kActivityTrackingFeatureName[] = "ActivityTracking";
+
+void SetupGlobalActivityTrackerFieldTrial() {
+  const size_t kMemorySize = 1 << 20;  // 1 MiB

[manzagop (departed), 2016/05/20 18:19:30]

Can you comment on how you chose this number? (k threads * depth of 3 * stack
entry size of s?)

[bcwhite, 2016/05/20 19:19:19]

Acknowledged.

+  const int kStackDepth = 3;
+  const uint64_t kAllocatorId = 0;
+  const char kAllocatorName[] = "ActivityTracker";
+
+  const std::string group_name =
+      FieldTrialList::FindFullName(kActivityTrackingFeatureName);
+  if (group_name.empty() || group_name == "Disabled")
+    return;
+
+  if (group_name == "InMemory")

[manzagop (departed), 2016/05/20 18:19:29]

nit: braces when multiline.

[bcwhite, 2016/05/20 19:19:18]

Done.

+    GlobalActivityTracker::CreateWithLocalMemory(kMemorySize, kAllocatorId,
+                                                 kAllocatorName, kStackDepth);
+  else
+    NOTREACHED() << group_name;

[manzagop (departed), 2016/05/20 18:19:30]

Isn't it better to silently disable, and the person looking at the experiment
will notice few users instead of crashes?

[bcwhite, 2016/05/20 19:19:19]

That's what'll happen in a release build.

+}
+
+
+struct ThreadActivityTracker::Header {
+  // This unique number indicates a valid initialization of the memory.
+  uint64_t cookie;
+
+  // The thread-id to which this data belongs. This identifier is not
+  // guaranteed to mean anything, just to be unique among all active
+  // trackers.
+  uint64_t thread_id;
+
+  // The start-time and start-ticks when the data was created. Each activity
+  // record has a |time_ticks| value that can be converted to a "wall time"
+  // with these two values.
+  int64_t start_time;
+  int64_t start_ticks;
+
+  // The number of Activity slots in the data.
+  uint32_t slots;
+
+  // The current depth of the stack. This may be greater than the number of
+  // slots. If the depth exceeds the number of slots, the newest entries
+  // won't be recorded.
+  std::atomic<uint32_t> depth;

[manzagop (departed), 2016/05/20 18:19:30]

Is atomic<POD> considered POD? Can there be version shear issues?

[bcwhite, 2016/05/20 19:19:19]

Yes.  No.  It's exactly the underlying type but with extra access methods.

+
+  // A memory location used to indicate if changes have been made to the stack
+  // that would invalidate an in-progress read of its contents. The active
+  // tracker will zero the value whenever something gets popped from the
+  // stack. A monitoring tracker can write a non-zero value here, copy the
+  // stack contents, and read the value to know, if it is still non-zero, that
+  // the contents didn't change while being copied.
+  std::atomic<int> unchanged;

[manzagop (departed), 2016/05/20 18:19:30]

int vs bool?

[bcwhite, 2016/05/20 19:19:18]

I chose to use a natural word size so there's no need for the CPU to ever deal
with a partial-memory-write. I think x86 is fine with such but other
architectures might not like it.

+};
+
+ThreadActivityTracker::ThreadActivityTracker(void* base, size_t size)
+    : header_(static_cast<Header*>(base)),
+      stack_(reinterpret_cast<StackEntry*>(reinterpret_cast<char*>(base) +
+                                           sizeof(Header))),
+      slots_((size - sizeof(Header)) / sizeof(StackEntry)) {

[manzagop (departed), 2016/05/20 18:19:29]

DCHECK(base)?

[bcwhite, 2016/05/20 19:19:18]

Done.

+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  // Ensure there is enough space for the header and at least a few records.
+  DCHECK_LE(sizeof(Header) + kMinStackDepth * sizeof(StackEntry), size);
+
+  // Ensure that the |slots_| calculation didn't overflow.
+  DCHECK_GE(std::numeric_limits<uint32_t>::max(),
+            (size - sizeof(Header)) / sizeof(StackEntry));
+
+  // Provided memory should either be completely initialized or all zeros.
+  if (header_->cookie == 0) {
+    // This is a new file. Double-check other fields and then initialize.
+    DCHECK_EQ(0U, header_->thread_id);
+    DCHECK_EQ(0, header_->start_time);
+    DCHECK_EQ(0, header_->start_ticks);
+    DCHECK_EQ(0U, header_->slots);
+    DCHECK_EQ(0U, header_->depth.load(std::memory_order_relaxed));
+    DCHECK_EQ(0, header_->unchanged.load(std::memory_order_relaxed));
+    DCHECK_EQ(0, stack_[0].time_ticks);
+    DCHECK_EQ(0, stack_[0].source_address);
+    DCHECK_EQ(0, stack_[0].method_address);
+    DCHECK_EQ(0U, stack_[0].sequence_id);
+
+    header_->cookie = kHeaderCookie;
+    header_->thread_id = static_cast<uint64_t>(PlatformThread::CurrentId());
+    header_->start_time = base::Time::Now().ToInternalValue();
+    header_->start_ticks = base::TimeTicks::Now().ToInternalValue();
+    header_->slots = slots_;
+    valid_ = true;
+  } else {
+    // This is a file with existing data. Perform basic consistency checks.
+    if (header_->cookie != kHeaderCookie ||
+        header_->slots != slots_ ||
+        header_->start_time > base::Time::Now().ToInternalValue() ||
+        stack_[0].time_ticks == 0)
+      return;
+    valid_ = true;
+  }
+}
+
+ThreadActivityTracker::~ThreadActivityTracker() {}
+
+void ThreadActivityTracker::RecordStart(const void* source,
+                                        ActivityType activity,
+                                        intptr_t method,
+                                        uint64_t sequence) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  // Get the current depth of the stack. No access to other memory guarded
+  // by this variable is done here so a "relaxed" load is acceptable.
+  uint32_t depth = header_->depth.load(std::memory_order_relaxed);
+
+  // Handle the case where the stack depth has exceeded the storage capacity.
+  // Extra entries will be lost leaving only the base of the stack.
+  if (depth >= slots_) {
+    // Since no other memory is being modified, a "relaxed" store is acceptable.
+    header_->depth.store(depth + 1, std::memory_order_relaxed);
+    return;
+  }
+
+  // Get a pointer to the next entry and load it. No atomicity is required
+  // here because the memory is known only to this thread. It will be made
+  // known to other threads once the depth is incremented.
+  StackEntry* entry = &stack_[depth];
+  entry->time_ticks = base::TimeTicks::Now().ToInternalValue();
+  entry->activity_type = activity;
+  entry->source_address = reinterpret_cast<intptr_t>(source);
+  entry->method_address = method;
+  entry->sequence_id = sequence;
+
+  // Save the incremented depth. Because this guards |entry| memory filled
+  // above that may be read by another thread once the recorded depth changes,
+  // a "release" store is required.
+  header_->depth.store(depth + 1, std::memory_order_release);
+}
+
+void ThreadActivityTracker::RecordFinish(const void* source) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  // Do an atomic decrement of the depth. No changes to stack entries guarded
+  // by this variable is done here so a "relaxed" operation is acceptable.
+  // |depth| will receive the value before it was modified.
+  uint32_t depth = header_->depth.fetch_sub(1, std::memory_order_relaxed);
+
+  // Validate that everything is running correctly.
+  DCHECK_LT(0U, depth);
+  if (depth <= slots_) {
+    DCHECK_EQ(reinterpret_cast<intptr_t>(source),
+              stack_[depth - 1].source_address);
+  }
+
+  // The stack has shrunk meaning that some other thread trying to copy the
+  // contents for reporting purposes could get bad data. That thread would
+  // have written a non-zero value into |unchanged|; clearing it here will
+  // let that thread detect that something did change. It doesn't matter
+  // when this is done relative to the atomic |depth| operation above so a
+  // "relaxed" access is acceptable.
+  header_->unchanged.store(0, std::memory_order_relaxed);
+}
+
+uint32_t ThreadActivityTracker::CopyStack(std::vector<StackEntry>* stack) {
+  // It's possible for the data to change while reading it. Make several

[manzagop (departed), 2016/05/20 18:19:30]

Nit: add that (only) some of these changes invalidate a read?

[bcwhite, 2016/05/20 19:19:18]

Done.

+  // attempts but don't try forever.
+  const int kMaxAttempts = 10;
+  uint32_t depth;
+
+  // Start with an empty return stack.
+  stack->clear();
+
+  // Stop here if the data isn't valid.
+  if (!valid_)

[manzagop (departed), 2016/05/20 18:19:30]

The return value is ambiguous wrt the function succeeded. Should revise the
interface?

[bcwhite, 2016/05/20 19:19:18]

Caller can always check is_valid() first.  Returning zero isn't unreasonable for
invalid data.

+    return 0;
+
+  for (int attempt = 0; attempt < kMaxAttempts; ++attempt) {
+    // Write a non-zero value to |unchanged| so it's possible to detect at
+    // the end that nothing has changed since copying the data began.
+    header_->unchanged.store(1, std::memory_order_relaxed);
+
+    // Fetching the current depth also "acquires" the contents of the stack.
+    depth = header_->depth.load(std::memory_order_acquire);
+    if (depth == 0)
+      return 0;
+
+    // Copy the existing contents. Memcpy is used for speed.
+    uint32_t count = std::min(depth, slots_);
+    stack->resize(count);
+    memcpy(&(*stack)[0], stack_, count * sizeof(StackEntry));
+
+    // Check to make sure everything was unchanged during the copy.

[manzagop (departed), 2016/05/20 18:19:29]

Is there a constraint there may be at most 1 reader? If so, mention it in the
class doc.

[bcwhite, 2016/05/20 19:19:18]

Good point.  Multiple readers would be possible if each set a different bit of
"unchanged" but it's not worth the complication.

Done.

+    if (header_->unchanged.load(std::memory_order_relaxed))
+      return depth;
+  }
+
+  // If all attempts failed, just return the depth with no content.
+  stack->clear();
+  return depth;
+}
+
+// static
+size_t ThreadActivityTracker::SizeForStackDepth(int stack_depth) {
+  return static_cast<size_t>(stack_depth) * sizeof(StackEntry) + sizeof(Header);
+}
+
+
+GlobalActivityTracker* GlobalActivityTracker::g_tracker_ = nullptr;
+
+GlobalActivityTracker::ManagedActivityTracker::ManagedActivityTracker(
+    PersistentMemoryAllocator::Reference mem_reference,
+    void* base,
+    size_t size)
+    : ThreadActivityTracker(base, size),
+      mem_reference_(mem_reference),
+      mem_base_(base) {}
+
+GlobalActivityTracker::ManagedActivityTracker::~ManagedActivityTracker() {
+  // The global |g_tracker_| must point to the owner of this class since all
+  // objects of this type must be destructed before |g_tracker_| can be changed.
+  DCHECK(g_tracker_);
+  g_tracker_->ReturnTrackerMemory(this, mem_reference_, mem_base_);
+}
+
+void GlobalActivityTracker::CreateWithAllocator(
+    std::unique_ptr<PersistentMemoryAllocator> allocator,
+    int stack_depth) {
+  // There's no need to do anything with the result. It is self-managing.
+  new GlobalActivityTracker(std::move(allocator), stack_depth);
+}
+
+// static
+void GlobalActivityTracker::CreateWithLocalMemory(size_t size,
+                                                  uint64_t id,
+                                                  StringPiece name,
+                                                  int stack_depth) {
+  CreateWithAllocator(
+      WrapUnique(new LocalPersistentMemoryAllocator(size, id, name)),
+      stack_depth);
+}
+
+// static
+void GlobalActivityTracker::CreateWithFile(const FilePath& file_path,
+                                           size_t size,
+                                           uint64_t id,
+                                           StringPiece name,
+                                           int stack_depth) {
+  // Create the file, overwriting anything that was there previously, and set
+  // the length. This will create a space that is zero-filled, a requirement
+  // for operation.
+  File file(file_path,
+            File::FLAG_CREATE_ALWAYS | File::FLAG_READ | File::FLAG_WRITE);
+  file.SetLength(size);
+
+  // Map the file into memory and make it globally available.
+  std::unique_ptr<MemoryMappedFile> mapped_file(new MemoryMappedFile());
+  mapped_file->Initialize(std::move(file), MemoryMappedFile::READ_WRITE);
+  CreateWithAllocator(WrapUnique(new FilePersistentMemoryAllocator(
+                          std::move(mapped_file), id, name)),
+                      stack_depth);
+}
+
+ThreadActivityTracker* GlobalActivityTracker::CreateTrackerForCurrentThread() {
+  DCHECK(!this_thread_tracker_.Get());
+
+  // The lock must be acquired to access the STL data structures.
+  AutoLock auto_lock(lock_);
+
+  PersistentMemoryAllocator::Reference mem_reference;
+  void* mem_base;
+  if (!available_memories_.empty()) {
+    // There is a memory block that was previously released (and zero'd) so
+    // just re-use that rather than allocating a new one.
+    mem_reference = available_memories_.back();
+    available_memories_.pop_back();
+    mem_base = allocator_->GetAsObject<char>(mem_reference,
+                                             kTypeIdActivityTrackerFree);
+    DCHECK(mem_base);
+    DCHECK_LT(stack_memory_, allocator_->GetAllocSize(mem_reference));

[manzagop (departed), 2016/05/20 20:24:21]

Is this not EQ? Or LE?

[bcwhite, 2016/05/20 20:41:18]

The space gets rounded up for alignment reasons so it should be LE.  Odd that it
didn't crash during testing -- will have to investigate that.

+    allocator_->SetType(mem_reference, kTypeIdActivityTracker);
+  } else {
+    // Allocate a block of memory from the persistent segment.
+    mem_reference = allocator_->Allocate(stack_memory_, kTypeIdActivityTracker);
+    if (mem_reference) {
+      // Success. Convert the reference to an actual memory address.
+      mem_base =
+          allocator_->GetAsObject<char>(mem_reference, kTypeIdActivityTracker);
+    } else {
+      // Failure. This should never happen.
+      NOTREACHED();
+      // But if it does, handle it gracefully by allocating the required
+      // memory from the heap.
+      mem_base = new char[stack_memory_];
+      memset(mem_base, 0, stack_memory_);
+    }
+  }
+
+  // Create a tracker with the acquired memory and set it as the tracker
+  // for this particular thread in thread-local-storage.
+  ManagedActivityTracker* tracker =
+      new ManagedActivityTracker(mem_reference, mem_base, stack_memory_);
+  DCHECK(tracker->is_valid());
+  thread_trackers_.insert(tracker);
+  this_thread_tracker_.Set(tracker);
+
+  return tracker;
+}
+
+void GlobalActivityTracker::ReleaseTrackerForCurrentThreadForTesting() {
+  ThreadActivityTracker* tracker =
+      reinterpret_cast<ThreadActivityTracker*>(this_thread_tracker_.Get());
+  if (tracker) {
+    this_thread_tracker_.Free();
+    delete tracker;
+  }
+}
+
+GlobalActivityTracker::GlobalActivityTracker(
+    std::unique_ptr<PersistentMemoryAllocator> allocator,
+    int stack_depth)
+    : allocator_(std::move(allocator)),
+      stack_memory_(ThreadActivityTracker::SizeForStackDepth(stack_depth)),
+      this_thread_tracker_(&OnTLSDestroy) {
+  // Ensure the passed memory is valid and empty (iterator finds nothing).
+  uint32_t type;
+  DCHECK(!PersistentMemoryAllocator::Iterator(allocator_.get()).GetNext(&type));
+
+  // Ensure that there is no other global object and then make this one such.
+  DCHECK(!g_tracker_);
+  g_tracker_ = this;
+
+  // Create a tracker for this thread since it is known.
+  CreateTrackerForCurrentThread();
+}
+
+GlobalActivityTracker::~GlobalActivityTracker() {
+  DCHECK_EQ(g_tracker_, this);
+  DCHECK_EQ(0U, thread_trackers_.size());
+  g_tracker_ = nullptr;
+}
+
+void GlobalActivityTracker::ReturnTrackerMemory(
+    ManagedActivityTracker* tracker,
+    PersistentMemoryAllocator::Reference mem_reference,
+    void* mem_base) {
+  // Zero the memory so that it is ready for use if needed again later. It's
+  // better to clear the memory now, when a thread is exiting, than to do it
+  // when it is first needed by a thread doing actual work.
+  memset(mem_base, 0, stack_memory_);
+
+  // Access to STL structurs requires a lock because this could get called
+  // from any thread.
+  AutoLock auto_lock(lock_);
+
+  // Remove the destructed tracker from the set of known ones.
+  DCHECK(ContainsKey(thread_trackers_, tracker));
+  thread_trackers_.erase(tracker);
+
+  // Deal with the memory that was used by the tracker.
+  if (mem_reference) {
+    // The memory was within the persistent memory allocator. Change its type
+    // so that iteration won't find it.
+    allocator_->SetType(mem_reference, kTypeIdActivityTrackerFree);
+    // There is no way to free memory from a persistent allocator so instead
+    // keep it on the internal list of available memory blocks.
+    DCHECK_LE(stack_memory_, allocator_->GetAllocSize(mem_reference));
+    available_memories_.push_back(mem_reference);
+  } else {
+    // The memory was allocated from the process heap. This shouldn't happen
+    // because the persistent memory segment should be big enough for all
+    // thread stacks but it's better to support falling back to allocation
+    // from the heap rather than crash. Everything will work as normal but
+    // the data won't be persisted.
+    delete[] reinterpret_cast<char*>(mem_base);
+  }
+}
+
+// static
+void GlobalActivityTracker::OnTLSDestroy(void* value) {
+  delete reinterpret_cast<ManagedActivityTracker*>(value);
+}
+
+ScopedTaskActivity::ScopedTaskActivity(const PendingTask& task)
+    : GlobalActivityTracker::ScopedThreadActivity(
+          task.posted_from.program_counter(),
+          ThreadActivityTracker::ACT_TASK,
+          0,
+          task.sequence_num) {}
+
+}  // namespace debug
+}  // namespace base