Support for extracting user-data from activity tracking.

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 <algorithm>
 
 #include "base/debug/stack_trace.h"
 #include "base/files/file.h"
 #include "base/files/file_path.h"
 #include "base/files/memory_mapped_file.h"
 #include "base/logging.h"
 #include "base/memory/ptr_util.h"
+#include "base/memory/ptr_util.h"
 #include "base/metrics/field_trial.h"
 #include "base/metrics/histogram_macros.h"
 #include "base/pending_task.h"
 #include "base/process/process.h"
 #include "base/process/process_handle.h"
 #include "base/stl_util.h"
 #include "base/strings/string_util.h"
 #include "base/threading/platform_thread.h"
 
 namespace base {
 namespace debug {
 
 namespace {
 
 // A number that identifies the memory as having been initialized. It's
 // arbitrary but happens to be the first 4 bytes of SHA1(ThreadActivityTracker).
 // A version number is added on so that major structure changes won't try to
 // read an older version (since the cookie won't match).
 const uint32_t kHeaderCookie = 0xC0029B24UL + 2;  // v2
 
 // The minimum depth a stack should support.
 const int kMinStackDepth = 2;
 
 // The amount of memory set aside for holding arbitrary user data (key/value
 // pairs) globally or associated with ActivityData entries.
 const size_t kUserDataSize = 1024;    // bytes
-const size_t kGlobalDataSize = 1024;  // bytes
+const size_t kGlobalDataSize = 4096;  // bytes
 const size_t kMaxUserDataNameLength =
     static_cast<size_t>(std::numeric_limits<uint8_t>::max());
 
 union ThreadRef {
   int64_t as_id;
 #if defined(OS_WIN)
   // On Windows, the handle itself is often a pseudo-handle with a common
   // value meaning "this thread" and so the thread-id is used. The former
   // can be converted to a thread-id with a system call.
   PlatformThreadId as_tid;
@@ skipping 135 matching lines @@
   size_t i;
   for (i = 1; i < stack_depth && i < kActivityCallStackSize; ++i) {
     activity->call_stack[i - 1] = reinterpret_cast<uintptr_t>(stack_addrs[i]);
   }
   activity->call_stack[i - 1] = 0;
 #else
   activity->call_stack[0] = 0;
 #endif
 }
 
-ActivitySnapshot::ActivitySnapshot() {}
-ActivitySnapshot::~ActivitySnapshot() {}
+ActivityUserData::TypedValue::TypedValue() {}
+ActivityUserData::TypedValue::TypedValue(const TypedValue& other) = default;
+ActivityUserData::TypedValue::~TypedValue() {}
+
+StringPiece ActivityUserData::TypedValue::Get() const {
+  DCHECK_EQ(RAW_VALUE, type);
+  return long_value;
+}
+
+StringPiece ActivityUserData::TypedValue::GetReference() const {
+  DCHECK_EQ(RAW_VALUE_REFERENCE, type);
+  return ref_value;
+}
+
+StringPiece ActivityUserData::TypedValue::GetString() const {
+  DCHECK_EQ(STRING_VALUE, type);
+  return long_value;
+}
+
+StringPiece ActivityUserData::TypedValue::GetStringReference() const {
+  DCHECK_EQ(STRING_VALUE_REFERENCE, type);
+  return ref_value;
+}
+
+bool ActivityUserData::TypedValue::GetBool() const {
+  DCHECK_EQ(BOOL_VALUE, type);
+  return short_value != 0;
+}
+
+char ActivityUserData::TypedValue::GetChar() const {
+  DCHECK_EQ(CHAR_VALUE, type);
+  return static_cast<char>(short_value);
+}
+
+int64_t ActivityUserData::TypedValue::GetInt() const {
+  DCHECK_EQ(SIGNED_VALUE, type);
+  return static_cast<int64_t>(short_value);
+}
+
+uint64_t ActivityUserData::TypedValue::GetUint() const {
+  DCHECK_EQ(UNSIGNED_VALUE, type);
+  return static_cast<uint64_t>(short_value);
+}
 
 ActivityUserData::ValueInfo::ValueInfo() {}
 ActivityUserData::ValueInfo::ValueInfo(ValueInfo&&) = default;
 ActivityUserData::ValueInfo::~ValueInfo() {}
 
+std::atomic<uint32_t> ActivityUserData::next_id_;

[manzagop (departed), 2016/11/29 22:42:58]

Did you mean to set a value?

[bcwhite, 2016/12/01 16:29:39]

Adding "=1" caused a compile error on some platforms.  All static values are
always initialized to zero if there is no explicit initializer.

+
 ActivityUserData::ActivityUserData(void* memory, size_t size)
-    : memory_(static_cast<char*>(memory)), available_(size) {}
+    : memory_(reinterpret_cast<char*>(memory)),
+      available_(size),
+      id_(reinterpret_cast<std::atomic<uint32_t>*>(memory)) {
+  // It's possible that no user data is being stored.
+  if (!memory_)
+    return;
+
+  DCHECK_LT(kMemoryAlignment, available_);
+  if (id_->load(std::memory_order_relaxed) == 0) {
+    // Generate a new ID and store it in the first 32-bit word of memory_.
+    // |id_| must be non-zero for non-sink instances.
+    uint32_t id;
+    while ((id = next_id_.fetch_add(1, std::memory_order_relaxed)) == 0)
+      ;
+    id_->store(id, std::memory_order_relaxed);
+    DCHECK_NE(0U, id_->load(std::memory_order_relaxed));
+  }
+  memory_ += kMemoryAlignment;
+  available_ -= kMemoryAlignment;
+
+  // If there is already data present, load that. This allows the same class
+  // to be used for analysis through snapshots.
+  ImportExistingData();

[manzagop (departed), 2016/11/29 22:42:58]

Worth avoiding the import when the id was initially 0? Not stricly necessary
because we clear the memory on reference release, but may make things locally
clearer.

[bcwhite, 2016/12/01 16:29:39]

It would require making a copy of the initial ID value and adding a condition
here.  That seems less obvious.

+}
 
 ActivityUserData::~ActivityUserData() {}
 
 void ActivityUserData::Set(StringPiece name,
                            ValueType type,
                            const void* memory,
                            size_t size) {
   DCHECK(thread_checker_.CalledOnValidThread());
   DCHECK_GE(std::numeric_limits<uint8_t>::max(), name.length());
   size = std::min(std::numeric_limits<uint16_t>::max() - (kMemoryAlignment - 1),
@@ skipping 16 matching lines @@
     // because there are not alignment constraints on strings, it's set tight
     // against the header. Its extent (the reserved space, even if it's not
     // all used) is calculated so that, when pressed against the header, the
     // following field will be aligned properly.
     size_t name_size = name.length();
     size_t name_extent =
         RoundUpToAlignment(sizeof(Header) + name_size, kMemoryAlignment) -
         sizeof(Header);
     size_t value_extent = RoundUpToAlignment(size, kMemoryAlignment);
 
-    // The "basic size" is the minimum size of the record. It's possible that
-    // lengthy values will get truncated but there must be at least some bytes
-    // available.
-    size_t basic_size = sizeof(Header) + name_extent + kMemoryAlignment;
-    if (basic_size > available_)
-      return;  // No space to store even the smallest value.
-
-    // The "full size" is the size for storing the entire value, truncated
-    // to the amount of available memory.
+    // The "full size" is the size for storing the entire value.
     size_t full_size =
         std::min(sizeof(Header) + name_extent + value_extent, available_);
+
+    // If the value is actually a single byte, see if it can be stuffed at the
+    // end of the name extent rather than wasting kMemoryAlignment bytes.
+    if (size == 1 && name_extent > name_size) {
+      full_size = sizeof(Header) + name_extent;

[manzagop (departed), 2016/11/29 22:42:58]

std::min(sizeof(Header) + name_extent, available_)?

[bcwhite, 2016/12/01 16:29:39]

Done, but with a separate check placed above.

+      --name_extent;
+    }
+
+    // Truncate the stored size to the amount of available memory. Stop now if
+    // there's not any room for even part of the value.
     size = std::min(full_size - sizeof(Header) - name_extent, size);

[manzagop (departed), 2016/11/29 22:42:58]

I guess full size could be equal to available_ which could be less than
(sizeof(Header) + name_extent), e.g. if available_ is 0), so there could be an
underflow?

[bcwhite, 2016/12/01 16:29:39]

Done as part of the above.

+    if (size == 0)
+      return;
 
     // Allocate a chunk of memory.
     Header* header = reinterpret_cast<Header*>(memory_);
     memory_ += full_size;
     available_ -= full_size;
 
     // Datafill the header and name records. Memory must be zeroed. The |type|
     // is written last, atomically, to release all the other values.
     DCHECK_EQ(END_OF_VALUES, header->type.load(std::memory_order_relaxed));
     DCHECK_EQ(0, header->value_size.load(std::memory_order_relaxed));
@@ skipping 32 matching lines @@
 void ActivityUserData::SetReference(StringPiece name,
                                     ValueType type,
                                     const void* memory,
                                     size_t size) {
   ReferenceRecord rec;
   rec.address = reinterpret_cast<uintptr_t>(memory);
   rec.size = size;
   Set(name, type, &rec, sizeof(rec));
 }
 
+void ActivityUserData::ImportExistingData() const {
+  while (available_ > 0) {

[manzagop (departed), 2016/11/29 22:42:58]

0 -> sizeof(Header)?

[bcwhite, 2016/12/01 16:29:39]

Done.

+    Header* header = reinterpret_cast<Header*>(memory_);
+    ValueType type =
+        static_cast<ValueType>(header->type.load(std::memory_order_acquire));
+    if (type == END_OF_VALUES)
+      return;
+    if (header->record_size > available_)
+      NOTREACHED(); // return;
+
+    size_t value_offset = RoundUpToAlignment(sizeof(Header) + header->name_size,
+                                             kMemoryAlignment);
+    if (header->record_size == value_offset &&
+        header->value_size.load(std::memory_order_relaxed) == 1) {
+      value_offset -= 1;
+    }
+    if (value_offset + header->value_size > header->record_size)
+      NOTREACHED(); // return;
+
+    ValueInfo info;
+    info.name = StringPiece(memory_ + sizeof(Header), header->name_size);
+    info.type = type;
+    info.memory = memory_ + value_offset;
+    info.size_ptr = &header->value_size;
+    info.extent = header->record_size - value_offset;
+
+    StringPiece key(info.name);
+    values_.insert(std::make_pair(key, std::move(info)));

[manzagop (departed), 2016/11/29 22:42:58]

Do you want to support successive calls of ImportExistingData (ie updates to
keys already inserted by a previous call)? If not, then mention it at
declaration.

[bcwhite, 2016/12/01 16:29:39]

It's called as part of every snapshot operation in case the system is live and
new things have been added.

[manzagop (departed), 2016/12/02 22:13:32]

Hm my comment was totally off! You scan from available_ and there's no need to
rescan what's been scanned as keys are immutable and you have a pointer to the
value's size in case it changed. Nothing to see here! :)

[bcwhite, 2016/12/08 21:30:56]

Ah.  The |memory_| pointer always points to the end so no rescan occurs.  It
just continues where it left off, which is important on a live system where
something else is updating the memory.

+
+    memory_ += header->record_size;
+    available_ -= header->record_size;
+  }
+}
+
+bool ActivityUserData::CreateSnapshot(Snapshot* output_snapshot) const {
+  DCHECK(output_snapshot);
+  DCHECK(output_snapshot->empty());
+
+  // Find any new data that may have been added by an active instance of this
+  // class that is adding records.
+  ImportExistingData();
+
+  for (const auto& entry : values_) {
+    TypedValue value;
+    value.type = entry.second.type;
+    DCHECK_GE(entry.second.extent,
+              entry.second.size_ptr->load(std::memory_order_relaxed));
+
+    switch (entry.second.type) {
+      case RAW_VALUE:
+      case STRING_VALUE:
+        value.long_value =
+            std::string(reinterpret_cast<char*>(entry.second.memory),
+                        entry.second.size_ptr->load(std::memory_order_relaxed));
+        break;
+      case RAW_VALUE_REFERENCE:
+      case STRING_VALUE_REFERENCE: {
+        ReferenceRecord* ref =
+            reinterpret_cast<ReferenceRecord*>(entry.second.memory);
+        value.ref_value = StringPiece(
+            reinterpret_cast<char*>(static_cast<uintptr_t>(ref->address)),
+            static_cast<size_t>(ref->size));
+      } break;
+      case BOOL_VALUE:
+      case CHAR_VALUE:
+        value.short_value = *reinterpret_cast<char*>(entry.second.memory);
+        break;
+      case SIGNED_VALUE:
+      case UNSIGNED_VALUE:
+        value.short_value = *reinterpret_cast<uint64_t*>(entry.second.memory);
+        break;
+      case END_OF_VALUES:  // Included for completeness purposes.
+        NOTREACHED();
+    }
+    auto inserted = output_snapshot->insert(
+        std::make_pair(entry.second.name.as_string(), std::move(value)));
+    DCHECK(inserted.second);  // True if inserted, false if existed.
+  }
+
+  return true;
+}
+
+ActivitySnapshot::ActivitySnapshot() {}
+ActivitySnapshot::~ActivitySnapshot() {}
+
 // This information is kept for every thread that is tracked. It is filled
 // the very first time the thread is seen. All fields must be of exact sizes
 // so there is no issue moving between 32 and 64-bit builds.
 struct ThreadActivityTracker::Header {
   // Expected size for 32/64-bit check.
   static constexpr size_t kExpectedInstanceSize = 80;
 
   // This unique number indicates a valid initialization of the memory.
   std::atomic<uint32_t> cookie;
 
@@ skipping 59 matching lines @@
     tracker_->PopActivity(activity_id_);
 }
 
 void ThreadActivityTracker::ScopedActivity::ChangeTypeAndData(
     Activity::Type type,
     const ActivityData& data) {
   if (tracker_)
     tracker_->ChangeActivity(activity_id_, type, data);
 }
 
-ActivityUserData& ThreadActivityTracker::ScopedActivity::user_data() {
-  if (!user_data_) {
-    if (tracker_)
-      user_data_ = tracker_->GetUserData(activity_id_);
-    else
-      user_data_ = MakeUnique<ActivityUserData>(nullptr, 0);
-  }
-  return *user_data_;
-}
-
 ThreadActivityTracker::ThreadActivityTracker(void* base, size_t size)
     : header_(static_cast<Header*>(base)),
       stack_(reinterpret_cast<Activity*>(reinterpret_cast<char*>(base) +
                                          sizeof(Header))),
       stack_slots_(
           static_cast<uint32_t>((size - sizeof(Header)) / sizeof(Activity))) {
   DCHECK(thread_checker_.CalledOnValidThread());
 
   // Verify the parameters but fail gracefully if they're not valid so that
   // production code based on external inputs will not crash.  IsValid() will
@@ skipping 135 matching lines @@
       header_->current_depth.fetch_sub(1, std::memory_order_relaxed) - 1;
 
   // Validate that everything is running correctly.
   DCHECK_EQ(id, depth);
 
   // A thread-checker creates a lock to check the thread-id which means
   // re-entry into this code if lock acquisitions are being tracked.
   DCHECK(stack_[depth].activity_type == Activity::ACT_LOCK_ACQUIRE ||
          thread_checker_.CalledOnValidThread());
 
-  // Check if there was any user-data memory. It isn't free'd until later
-  // because the call to release it can push something on the stack.
-  PersistentMemoryAllocator::Reference user_data = stack_[depth].user_data;
-  stack_[depth].user_data = 0;
-
   // 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 |stack_unchanged|; clearing it here
   // will let that thread detect that something did change. This needs to
   // happen after the atomic |depth| operation above so a "release" store
   // is required.
   header_->stack_unchanged.store(0, std::memory_order_release);
-
-  // Release resources located above. All stack processing is done so it's
-  // safe if some outside code does another push.
-  if (user_data)
-    GlobalActivityTracker::Get()->ReleaseUserDataMemory(&user_data);
 }
 
 std::unique_ptr<ActivityUserData> ThreadActivityTracker::GetUserData(
-    ActivityId id) {
+    ActivityId id,
+    ActivityTrackerMemoryAllocator* allocator) {
   // User-data is only stored for activities actually held in the stack.
   if (id < stack_slots_) {
+    // Don't allow user data for lock acquisition as recursion may occur.
+    if (stack_[id].activity_type == Activity::ACT_LOCK_ACQUIRE) {
+      NOTREACHED();
+      return MakeUnique<ActivityUserData>(nullptr, 0);
+    }
+
+    // Get (or reuse) a block of memory and create a real UserData object
+    // on it.
+    PersistentMemoryAllocator::Reference ref = allocator->GetObjectReference();
     void* memory =
-        GlobalActivityTracker::Get()->GetUserDataMemory(&stack_[id].user_data);
-    if (memory)
-      return MakeUnique<ActivityUserData>(memory, kUserDataSize);
+        allocator->GetAsArray<char>(ref, PersistentMemoryAllocator::kSizeAny);
+    if (memory) {
+      std::unique_ptr<ActivityUserData> user_data =
+          MakeUnique<ActivityUserData>(memory, kUserDataSize);
+      stack_[id].user_data_ref = ref;
+      stack_[id].user_data_id = user_data->id();
+      return user_data;
+    }
   }
 
   // Return a dummy object that will still accept (but ignore) Set() calls.
   return MakeUnique<ActivityUserData>(nullptr, 0);
 }
 
+bool ThreadActivityTracker::HasUserData(ActivityId id) {
+  // User-data is only stored for activities actually held in the stack.
+  return (id < stack_slots_ && stack_[id].user_data_ref);
+}
+
+void ThreadActivityTracker::ReleaseUserData(
+    ActivityId id,
+    ActivityTrackerMemoryAllocator* allocator) {
+  // User-data is only stored for activities actually held in the stack.
+  if (id < stack_slots_ && stack_[id].user_data_ref) {
+    allocator->ReleaseObjectReference(stack_[id].user_data_ref);
+    stack_[id].user_data_ref = 0;
+  }
+}
+
 bool ThreadActivityTracker::IsValid() const {
   if (header_->cookie.load(std::memory_order_acquire) != kHeaderCookie ||
       header_->process_id.load(std::memory_order_relaxed) == 0 ||
       header_->thread_ref.as_id == 0 ||
       header_->start_time == 0 ||
       header_->start_ticks == 0 ||
       header_->stack_slots != stack_slots_ ||
       header_->thread_name[sizeof(header_->thread_name) - 1] != '\0') {
     return false;
   }
 
   return valid_;
 }
 
-bool ThreadActivityTracker::Snapshot(ActivitySnapshot* output_snapshot) const {
+bool ThreadActivityTracker::CreateSnapshot(
+    ActivitySnapshot* output_snapshot) const {
   DCHECK(output_snapshot);
 
   // There is no "called on valid thread" check for this method as it can be
   // called from other threads or even other processes. It is also the reason
   // why atomic operations must be used in certain places above.
 
   // It's possible for the data to change while reading it in such a way that it
   // invalidates the read. Make several attempts but don't try forever.
   const int kMaxAttempts = 10;
   uint32_t depth;
@@ skipping 95 matching lines @@
 }
 
 // static
 size_t ThreadActivityTracker::SizeForStackDepth(int stack_depth) {
   return static_cast<size_t>(stack_depth) * sizeof(Activity) + sizeof(Header);
 }
 
 
 GlobalActivityTracker* GlobalActivityTracker::g_tracker_ = nullptr;
 
+GlobalActivityTracker::ScopedThreadActivity::ScopedThreadActivity(
+    const void* program_counter,
+    const void* origin,
+    Activity::Type type,
+    const ActivityData& data,
+    bool lock_allowed)
+    : ThreadActivityTracker::ScopedActivity(
+          GetOrCreateTracker(lock_allowed),
+          program_counter,
+          origin,
+          type,
+          data) {}
+
+GlobalActivityTracker::ScopedThreadActivity::~ScopedThreadActivity() {
+  if (tracker_ && tracker_->HasUserData(activity_id_)) {
+    GlobalActivityTracker* global = GlobalActivityTracker::Get();
+    AutoLock lock(global->user_data_allocator_lock_);
+    tracker_->ReleaseUserData(activity_id_, &global->user_data_allocator_);
+  }
+}
+
+ActivityUserData& GlobalActivityTracker::ScopedThreadActivity::user_data() {
+  if (!user_data_) {
+    if (tracker_) {
+      GlobalActivityTracker* global = GlobalActivityTracker::Get();
+      AutoLock lock(global->user_data_allocator_lock_);
+      user_data_ =
+          tracker_->GetUserData(activity_id_, &global->user_data_allocator_);
+    } else {
+      user_data_ = MakeUnique<ActivityUserData>(nullptr, 0);
+    }
+  }
+  return *user_data_;
+}
+
 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
@@ skipping 105 matching lines @@
   return tracker;
 }
 
 void GlobalActivityTracker::ReleaseTrackerForCurrentThreadForTesting() {
   ThreadActivityTracker* tracker =
       reinterpret_cast<ThreadActivityTracker*>(this_thread_tracker_.Get());
   if (tracker)
     delete tracker;
 }
 
-void* GlobalActivityTracker::GetUserDataMemory(
-    PersistentMemoryAllocator::Reference* reference) {
-  if (!*reference) {
-    base::AutoLock autolock(user_data_allocator_lock_);
-    *reference = user_data_allocator_.GetObjectReference();
-    if (!*reference)
-      return nullptr;
-  }
-
-  void* memory = allocator_->GetAsArray<char>(
-      *reference, kTypeIdUserDataRecord, PersistentMemoryAllocator::kSizeAny);
-  DCHECK(memory);
-  return memory;
-}
-
-void GlobalActivityTracker::ReleaseUserDataMemory(
-    PersistentMemoryAllocator::Reference* reference) {
-  DCHECK(*reference);
-  base::AutoLock autolock(user_data_allocator_lock_);
-  user_data_allocator_.ReleaseObjectReference(*reference);
-  *reference = PersistentMemoryAllocator::kReferenceNull;
-}
-
 GlobalActivityTracker::GlobalActivityTracker(
     std::unique_ptr<PersistentMemoryAllocator> allocator,
     int stack_depth)
     : allocator_(std::move(allocator)),
       stack_memory_size_(ThreadActivityTracker::SizeForStackDepth(stack_depth)),
       this_thread_tracker_(&OnTLSDestroy),
       thread_tracker_count_(0),
       thread_tracker_allocator_(allocator_.get(),
                                 kTypeIdActivityTracker,
                                 kTypeIdActivityTrackerFree,
@@ skipping 12 matching lines @@
               kTypeIdGlobalDataRecord,
               PersistentMemoryAllocator::kSizeAny),
           kGlobalDataSize) {
   // 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;
+
+  // The global user-data record must be iterable in order to be found by an
+  // analyzer.
+  // allocator_->MakeIterable(...);

[manzagop (departed), 2016/11/29 22:42:58]

Uncomment?

[bcwhite, 2016/12/01 16:29:39]

Done.  I had to wait for a separate CL to land.

 }
 
 GlobalActivityTracker::~GlobalActivityTracker() {
   DCHECK_EQ(g_tracker_, this);
   DCHECK_EQ(0, thread_tracker_count_.load(std::memory_order_relaxed));
   g_tracker_ = nullptr;
 }
 
 void GlobalActivityTracker::ReturnTrackerMemory(
     ManagedActivityTracker* tracker) {
@@ skipping 94 matching lines @@
     : GlobalActivityTracker::ScopedThreadActivity(
           program_counter,
           nullptr,
           Activity::ACT_PROCESS_WAIT,
           ActivityData::ForProcess(process->Pid()),
           /*lock_allowed=*/true) {}
 #endif
 
 }  // namespace debug
 }  // namespace base