Multi-Process Tracking Support

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 <limits>
 #include <utility>
 
@@ skipping 12 matching lines @@
 #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 = 1 << 10;     // 1 KiB
+const size_t kProcessDataSize = 4 << 10;  // 4 KiB
 const size_t kGlobalDataSize = 16 << 10;  // 16 KiB
 const size_t kMaxUserDataNameLength =
     static_cast<size_t>(std::numeric_limits<uint8_t>::max());
 
 // A constant used to indicate that module information is changing.
 const uint32_t kModuleInformationChanging = 0x80000000;
 
+// An atomically incrementing number, used to check for recreations of objects
+// in the same memory space.
+StaticAtomicSequenceNumber g_next_id;
+
 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;
 #elif defined(OS_POSIX)
   // On Posix, the handle is always a unique identifier so no conversion
   // needs to be done. However, it's value is officially opaque so there
   // is no one correct way to convert it to a numerical identifier.
   PlatformThreadHandle::Handle as_handle;
 #endif
 };
 
+// Get the next non-zero identifier. It is only unique within a process.
+uint32_t GetNextDataId() {
+  uint32_t id;
+  while ((id = g_next_id.GetNext()) == 0)
+    ;
+  return id;
+}
+
+// Finds and reuses a specific allocation or creates a new one.
+PersistentMemoryAllocator::Reference AllocateFrom(
+    PersistentMemoryAllocator* allocator,
+    uint32_t from_type,
+    size_t size,
+    uint32_t to_type) {
+  PersistentMemoryAllocator::Iterator iter(allocator);
+  PersistentMemoryAllocator::Reference ref;
+  while ((ref = iter.GetNextOfType(from_type)) != 0) {
+    DCHECK_LE(size, allocator->GetAllocSize(ref));
+    // This can fail if a another thread has just taken it. It isassumed that
+    // the memory is cleared during the "free" operation.
+    if (allocator->ChangeType(ref, to_type, from_type, /*clear=*/false))
+      return ref;
+  }
+
+  return allocator->Allocate(size, to_type);
+}
+
 // Determines the previous aligned index.
 size_t RoundDownToAlignment(size_t index, size_t alignment) {
   return index & (0 - alignment);
 }
 
 // Determines the next aligned index.
 size_t RoundUpToAlignment(size_t index, size_t alignment) {
   return (index + (alignment - 1)) & (0 - alignment);
 }
 
 }  // namespace
 
+ProcessInfo::ProcessInfo() {}
+ProcessInfo::~ProcessInfo() {}
+
+void ProcessInfo::Release_Initialize() {
+  uint32_t old_id = data_id.load(std::memory_order_acquire);
+  DCHECK_EQ(0U, old_id);
+  process_id = GetCurrentProcId();
+  create_stamp = Time::Now().ToInternalValue();
+  data_id.store(GetNextDataId(), std::memory_order_release);
+}
+
+void ProcessInfo::SetOwningProcessIdForTesting(ProcessId pid, int64_t stamp) {
+  process_id = pid;
+  create_stamp = stamp;

[manzagop (departed), 2017/02/22 20:44:15]

Do you need/want to set data_id so that ProcessInfo::OwningProcessId doesn't hit
the if on l.127?

[bcwhite, 2017/02/22 22:13:02]

All three already have valid, non-zero values.  This just points two of them at
a different process.  The data_id can remain as it is.

[manzagop (departed), 2017/02/24 15:56:35]

Is that because calling Release_Initialize is a precondition? If so, wdyt of
documenting the precondition with a DCHECK on data_id?

[bcwhite, 2017/03/06 16:33:51]

Done.

+}
+
+// static
+bool ProcessInfo::OwningProcessId(const void* memory,
+                                  ProcessId* out_id,
+                                  int64_t* out_stamp) {
+  const ProcessInfo* info = reinterpret_cast<const ProcessInfo*>(memory);
+  uint32_t id = info->data_id.load(std::memory_order_acquire);
+  if (id == 0)
+    return false;
+
+  *out_id = static_cast<ProcessId>(info->process_id);
+  *out_stamp = info->create_stamp;
+  return id == info->data_id.load(std::memory_order_seq_cst);

[manzagop (departed), 2017/02/22 20:44:15]

Can you say more about this operation?

IIUC there are 2 operations on the variable:
1) a release store from 0 to nonzero, in ProcessInfo::Release_Initialize
2) a memset from non-zero to zero, when the PMA reference block is freed

Can there be an issue wrt operation 2? Can you start zero'ing the pid/timestamp
before zero'ing the data_id?

[bcwhite, 2017/02/22 22:13:02]

It could since memset doesn't define the order in which it clears memory.  And
even if it did, the writes are non-atomic so who knows what the compiler and CPU
might do with it.

I think I'll have to fix this inside the PMA to ensure the first word always
gets cleared first.

[manzagop (departed), 2017/02/24 15:56:35]

I'm totally going to forget about this! :) Could you create a bug for it?

[bcwhite, 2017/03/06 16:33:51]

It's already out for review:
https://codereview.chromium.org/2709113003

+}
 
 // It doesn't matter what is contained in this (though it will be all zeros)
 // as only the address of it is important.
 const ActivityData kNullActivityData = {};
 
 ActivityData ActivityData::ForThread(const PlatformThreadHandle& handle) {
   ThreadRef thread_ref;
   thread_ref.as_id = 0;  // Zero the union in case other is smaller.
 #if defined(OS_WIN)
   thread_ref.as_tid = ::GetThreadId(handle.platform_handle());
@@ skipping 150 matching lines @@
 StringPiece ActivityUserData::TypedValue::GetReference() const {
   DCHECK_EQ(RAW_VALUE_REFERENCE, type_);
   return ref_value_;
 }
 
 StringPiece ActivityUserData::TypedValue::GetStringReference() const {
   DCHECK_EQ(STRING_VALUE_REFERENCE, type_);
   return ref_value_;
 }
 
+// These are required because std::atomic is (currently) not a POD type and
+// thus clang requires explicit out-of-line constructors and destructors even
+// when they do nothing.
 ActivityUserData::ValueInfo::ValueInfo() {}
 ActivityUserData::ValueInfo::ValueInfo(ValueInfo&&) = default;
 ActivityUserData::ValueInfo::~ValueInfo() {}
-
-StaticAtomicSequenceNumber ActivityUserData::next_id_;
+ActivityUserData::MemoryHeader::MemoryHeader() {}
+ActivityUserData::MemoryHeader::~MemoryHeader() {}
+ActivityUserData::FieldHeader::FieldHeader() {}
+ActivityUserData::FieldHeader::~FieldHeader() {}
 
 ActivityUserData::ActivityUserData(void* memory, size_t size)
     : memory_(reinterpret_cast<char*>(memory)),
       available_(RoundDownToAlignment(size, kMemoryAlignment)),
-      id_(reinterpret_cast<std::atomic<uint32_t>*>(memory)) {
+      header_(reinterpret_cast<MemoryHeader*>(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_.GetNext()) == 0)
-      ;
-    id_->store(id, std::memory_order_relaxed);
-    DCHECK_NE(0U, id_->load(std::memory_order_relaxed));
-  }
-  memory_ += kMemoryAlignment;
-  available_ -= kMemoryAlignment;
+  static_assert(0 == sizeof(MemoryHeader) % kMemoryAlignment, "invalid header");
+  DCHECK_LT(sizeof(MemoryHeader), available_);
+  if (header_->process_info.data_id.load(std::memory_order_acquire) == 0)
+    header_->process_info.Release_Initialize();
+  memory_ += sizeof(MemoryHeader);
+  available_ -= sizeof(MemoryHeader);
 
   // If there is already data present, load that. This allows the same class
   // to be used for analysis through snapshots.
   ImportExistingData();
 }
 
 ActivityUserData::~ActivityUserData() {}
 
+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;
+}
+
+const void* ActivityUserData::GetBaseAddress() {
+  // The |memory_| pointer advances as elements are written but the |header_|
+  // value is always at the start of the block so just return that.
+  return header_;
+}
+
+void ActivityUserData::SetOwningProcessIdForTesting(ProcessId pid,
+                                                    int64_t stamp) {
+  if (!header_)
+    return;
+  header_->process_info.SetOwningProcessIdForTesting(pid, stamp);
+}
+
+// static
+bool ActivityUserData::OwningProcessId(const void* memory,
+                                       ProcessId* out_id,
+                                       int64_t* out_stamp) {
+  const MemoryHeader* header = reinterpret_cast<const MemoryHeader*>(memory);
+  return ProcessInfo::OwningProcessId(&header->process_info, out_id, out_stamp);
+}
+
 void ActivityUserData::Set(StringPiece name,
                            ValueType type,
                            const void* memory,
                            size_t size) {
   DCHECK_GE(std::numeric_limits<uint8_t>::max(), name.length());
   size = std::min(std::numeric_limits<uint16_t>::max() - (kMemoryAlignment - 1),
                   size);
 
   // It's possible that no user data is being stored.
   if (!memory_)
     return;
 
   // The storage of a name is limited so use that limit during lookup.
   if (name.length() > kMaxUserDataNameLength)
     name.set(name.data(), kMaxUserDataNameLength);
 
   ValueInfo* info;
   auto existing = values_.find(name);
   if (existing != values_.end()) {
     info = &existing->second;
   } else {
     // The name size is limited to what can be held in a single byte but
     // 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);
+        RoundUpToAlignment(sizeof(FieldHeader) + name_size, kMemoryAlignment) -
+        sizeof(FieldHeader);
     size_t value_extent = RoundUpToAlignment(size, kMemoryAlignment);
 
     // The "base size" is the size of the header and (padded) string key. Stop
     // now if there's not room enough for even this.
-    size_t base_size = sizeof(Header) + name_extent;
+    size_t base_size = sizeof(FieldHeader) + name_extent;
     if (base_size > available_)
       return;
 
     // The "full size" is the size for storing the entire value.
     size_t full_size = std::min(base_size + 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 = base_size;
       --name_extent;
       --base_size;
     }
 
     // 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 - base_size, size);
     if (size == 0)
       return;
 
     // Allocate a chunk of memory.
-    Header* header = reinterpret_cast<Header*>(memory_);
+    FieldHeader* header = reinterpret_cast<FieldHeader*>(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));
     header->name_size = static_cast<uint8_t>(name_size);
     header->record_size = full_size;
-    char* name_memory = reinterpret_cast<char*>(header) + sizeof(Header);
+    char* name_memory = reinterpret_cast<char*>(header) + sizeof(FieldHeader);
     void* value_memory =
-        reinterpret_cast<char*>(header) + sizeof(Header) + name_extent;
+        reinterpret_cast<char*>(header) + sizeof(FieldHeader) + name_extent;
     memcpy(name_memory, name.data(), name_size);
     header->type.store(type, std::memory_order_release);
 
     // Create an entry in |values_| so that this field can be found and changed
     // later on without having to allocate new entries.
     StringPiece persistent_name(name_memory, name_size);
     auto inserted =
         values_.insert(std::make_pair(persistent_name, ValueInfo()));
     DCHECK(inserted.second);  // True if inserted, false if existed.
     info = &inserted.first->second;
     info->name = persistent_name;
     info->memory = value_memory;
     info->size_ptr = &header->value_size;
-    info->extent = full_size - sizeof(Header) - name_extent;
+    info->extent = full_size - sizeof(FieldHeader) - name_extent;
     info->type = type;
   }
 
   // Copy the value data to storage. The |size| is written last, atomically, to
   // release the copied data. Until then, a parallel reader will just ignore
   // records with a zero size.
   DCHECK_EQ(type, info->type);
   size = std::min(size, info->extent);
   info->size_ptr->store(0, std::memory_order_seq_cst);
   memcpy(info->memory, memory, size);
   info->size_ptr->store(size, std::memory_order_release);
 }
 
 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_ > sizeof(Header)) {
-    Header* header = reinterpret_cast<Header*>(memory_);
+  while (available_ > sizeof(FieldHeader)) {
+    FieldHeader* header = reinterpret_cast<FieldHeader*>(memory_);
     ValueType type =
         static_cast<ValueType>(header->type.load(std::memory_order_acquire));
     if (type == END_OF_VALUES)
       return;
     if (header->record_size > available_)
       return;
 
-    size_t value_offset = RoundUpToAlignment(sizeof(Header) + header->name_size,
-                                             kMemoryAlignment);
+    size_t value_offset = RoundUpToAlignment(
+        sizeof(FieldHeader) + 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)
       return;
 
     ValueInfo info;
-    info.name = StringPiece(memory_ + sizeof(Header), header->name_size);
+    info.name = StringPiece(memory_ + sizeof(FieldHeader), 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)));
 
     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;
-}
-
-const void* ActivityUserData::GetBaseAddress() {
-  // The |memory_| pointer advances as elements are written but the |id_|
-  // value is always at the start of the block so just return that.
-  return id_;
-}
-
 // 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 {
   // Defined in .h for analyzer access. Increment this if structure changes!
   static constexpr uint32_t kPersistentTypeId =
       GlobalActivityTracker::kTypeIdActivityTracker;
 
   // Expected size for 32/64-bit check.
-  static constexpr size_t kExpectedInstanceSize = 80;
+  static constexpr size_t kExpectedInstanceSize =
+      ProcessInfo::kExpectedInstanceSize + 72;
 
-  // This unique number indicates a valid initialization of the memory.
-  std::atomic<uint32_t> cookie;
+  // This information uniquely identifies a process.
+  ProcessInfo process_info;
 
-  // The number of Activity slots (spaces that can hold an Activity) that
-  // immediately follow this structure in memory.
-  uint32_t stack_slots;
-
-  // The process-id and thread-id (thread_ref.as_id) to which this data belongs.
-  // These identifiers are not guaranteed to mean anything but are unique, in
-  // combination, among all active trackers. It would be nice to always have
-  // the process_id be a 64-bit value but the necessity of having it atomic
-  // (for the memory barriers it provides) limits it to the natural word size
-  // of the machine.
-#ifdef ARCH_CPU_64_BITS
-  std::atomic<int64_t> process_id;
-#else
-  std::atomic<int32_t> process_id;
-  int32_t process_id_padding;
-#endif
+  // The thread-id (thread_ref.as_id) to which this data belongs. This number
+  // is not guaranteed to mean anything but combined with the process-id from
+  // ProcessInfo is unique among all active trackers.
   ThreadRef thread_ref;
 
   // The start-time and start-ticks when the data was created. Each activity
   // record has a |time_internal| 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 (spaces that can hold an Activity) that
+  // immediately follow this structure in memory.
+  uint32_t stack_slots;
+
+  // Some padding to keep everything 64-bit aligned.
+  uint32_t padding;
+
   // 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> current_depth;
 
   // 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
@@ skipping 62 matching lines @@
       sizeof(header_->thread_ref) == sizeof(header_->thread_ref.as_id),
       "PlatformThreadHandle::Handle is too big to hold in 64-bit ID");
 
   // Ensure that the alignment of Activity.data is properly aligned to a
   // 64-bit boundary so there are no interoperability-issues across cpu
   // architectures.
   static_assert(offsetof(Activity, data) % sizeof(uint64_t) == 0,
                 "ActivityData.data is not 64-bit aligned");
 
   // Provided memory should either be completely initialized or all zeros.
-  if (header_->cookie.load(std::memory_order_relaxed) == 0) {
+  if (header_->process_info.data_id.load(std::memory_order_relaxed) == 0) {
     // This is a new file. Double-check other fields and then initialize.
-    DCHECK_EQ(0, header_->process_id.load(std::memory_order_relaxed));
+    DCHECK_EQ(0, header_->process_info.process_id);
+    DCHECK_EQ(0, header_->process_info.create_stamp);
     DCHECK_EQ(0, header_->thread_ref.as_id);
     DCHECK_EQ(0, header_->start_time);
     DCHECK_EQ(0, header_->start_ticks);
     DCHECK_EQ(0U, header_->stack_slots);
     DCHECK_EQ(0U, header_->current_depth.load(std::memory_order_relaxed));
     DCHECK_EQ(0U, header_->stack_unchanged.load(std::memory_order_relaxed));
     DCHECK_EQ(0, stack_[0].time_internal);
     DCHECK_EQ(0U, stack_[0].origin_address);
     DCHECK_EQ(0U, stack_[0].call_stack[0]);
     DCHECK_EQ(0U, stack_[0].data.task.sequence_id);
 
 #if defined(OS_WIN)
     header_->thread_ref.as_tid = PlatformThread::CurrentId();
 #elif defined(OS_POSIX)
     header_->thread_ref.as_handle =
         PlatformThread::CurrentHandle().platform_handle();
 #endif
-    header_->process_id.store(GetCurrentProcId(), std::memory_order_relaxed);
 
     header_->start_time = base::Time::Now().ToInternalValue();
     header_->start_ticks = base::TimeTicks::Now().ToInternalValue();
     header_->stack_slots = stack_slots_;
     strlcpy(header_->thread_name, PlatformThread::GetName(),
             sizeof(header_->thread_name));
 
     // This is done last so as to guarantee that everything above is "released"
     // by the time this value gets written.
-    header_->cookie.store(kHeaderCookie, std::memory_order_release);
+    header_->process_info.Release_Initialize();
 
     valid_ = true;
     DCHECK(IsValid());
   } else {
     // This is a file with existing data. Perform basic consistency checks.
     valid_ = true;
     valid_ = IsValid();
   }
 }
 
@@ skipping 124 matching lines @@
     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 ||
+  if (header_->process_info.data_id.load(std::memory_order_acquire) == 0 ||
+      header_->process_info.process_id == 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::CreateSnapshot(Snapshot* output_snapshot) const {
   DCHECK(output_snapshot);
@@ skipping 10 matching lines @@
   // Stop here if the data isn't valid.
   if (!IsValid())
     return false;
 
   // Allocate the maximum size for the stack so it doesn't have to be done
   // during the time-sensitive snapshot operation. It is shrunk once the
   // actual size is known.
   output_snapshot->activity_stack.reserve(stack_slots_);
 
   for (int attempt = 0; attempt < kMaxAttempts; ++attempt) {
-    // Remember the process and thread IDs to ensure they aren't replaced
-    // during the snapshot operation. Use "acquire" to ensure that all the
-    // non-atomic fields of the structure are valid (at least at the current
-    // moment in time).
-    const int64_t starting_process_id =
-        header_->process_id.load(std::memory_order_acquire);
+    // Remember the data IDs to ensure nothing is replaced during the snapshot
+    // operation. Use "acquire" so that all the non-atomic fields of the
+    // structure are valid (at least at the current moment in time).
+    const uint32_t starting_id =
+        header_->process_info.data_id.load(std::memory_order_acquire);
+    const int64_t starting_process_id = header_->process_info.process_id;
     const int64_t starting_thread_id = header_->thread_ref.as_id;
 
     // Write a non-zero value to |stack_unchanged| so it's possible to detect
     // at the end that nothing has changed since copying the data began. A
     // "cst" operation is required to ensure it occurs before everything else.
     // Using "cst" memory ordering is relatively expensive but this is only
     // done during analysis so doesn't directly affect the worker threads.
     header_->stack_unchanged.store(1, std::memory_order_seq_cst);
 
     // Fetching the current depth also "acquires" the contents of the stack.
@@ skipping 20 matching lines @@
     // to be last so that it's possible to detect below if any content has
     // changed while reading it. It's technically possible for a thread to end,
     // have its data cleared, a new thread get created with the same IDs, and
     // it perform an action which starts tracking all in the time since the
     // ID reads above but the chance is so unlikely that it's not worth the
     // effort and complexity of protecting against it (perhaps with an
     // "unchanged" field like is done for the stack).
     output_snapshot->thread_name =
         std::string(header_->thread_name, sizeof(header_->thread_name) - 1);
     output_snapshot->thread_id = header_->thread_ref.as_id;
-    output_snapshot->process_id =
-        header_->process_id.load(std::memory_order_seq_cst);
+    output_snapshot->process_id = header_->process_info.process_id;
 
     // All characters of the thread-name buffer were copied so as to not break
     // if the trailing NUL were missing. Now limit the length if the actual
     // name is shorter.
     output_snapshot->thread_name.resize(
         strlen(output_snapshot->thread_name.c_str()));
 
-    // If the process or thread ID has changed then the tracker has exited and
-    // the memory reused by a new one. Try again.
-    if (output_snapshot->process_id != starting_process_id ||
+    // If the data ID has changed then the tracker has exited and the memory
+    // reused by a new one. Try again.
+    if (header_->process_info.data_id.load(std::memory_order_seq_cst) !=
+            starting_id ||
+        output_snapshot->process_id != starting_process_id ||
         output_snapshot->thread_id != starting_thread_id) {
       continue;
     }
 
     // Only successful if the data is still valid once everything is done since
     // it's possible for the thread to end somewhere in the middle and all its
     // values become garbage.
     if (!IsValid())
       return false;
 
     // Change all the timestamps in the activities from "ticks" to "wall" time.
     const Time start_time = Time::FromInternalValue(header_->start_time);
     const int64_t start_ticks = header_->start_ticks;
     for (Activity& activity : output_snapshot->activity_stack) {
       activity.time_internal =
           (start_time +
            TimeDelta::FromInternalValue(activity.time_internal - start_ticks))
               .ToInternalValue();
     }
 
     // Success!
     return true;
   }
 
   // Too many attempts.
   return false;
 }
 
+const void* ThreadActivityTracker::GetBaseAddress() {
+  return header_;
+}
+
+void ThreadActivityTracker::SetOwningProcessIdForTesting(ProcessId pid,
+                                                         int64_t stamp) {
+  header_->process_info.SetOwningProcessIdForTesting(pid, stamp);
+}
+
+// static
+bool ThreadActivityTracker::OwningProcessId(const void* memory,
+                                            ProcessId* out_id,
+                                            int64_t* out_stamp) {
+  const Header* header = reinterpret_cast<const Header*>(memory);
+  return ProcessInfo::OwningProcessId(&header->process_info, out_id, out_stamp);
+}
+
 // static
 size_t ThreadActivityTracker::SizeForStackDepth(int stack_depth) {
   return static_cast<size_t>(stack_depth) * sizeof(Activity) + sizeof(Header);
 }
 
 // The instantiation of the GlobalActivityTracker object.
 // The object held here will obviously not be destructed at process exit
 // but that's best since PersistentMemoryAllocator objects (that underlie
 // GlobalActivityTracker objects) are explicitly forbidden from doing anything
 // essential at exit anyway due to the fact that they depend on data managed
@@ skipping 67 matching lines @@
   // These fields never changes and are done before the record is made
   // iterable so no thread protection is necessary.
   size = info.size;
   timestamp = info.timestamp;
   age = info.age;
   memcpy(identifier, info.identifier, sizeof(identifier));
   memcpy(pickle, pickler.data(), pickler.size());
   pickle_size = pickler.size();
   changes.store(0, std::memory_order_relaxed);
 
+  // Initialize the process info.
+  process_info.Release_Initialize();
+
   // Now set those fields that can change.
   return UpdateFrom(info);
 }
 
 bool GlobalActivityTracker::ModuleInfoRecord::UpdateFrom(
     const GlobalActivityTracker::ModuleInfo& info) {
   // Updates can occur after the record is made visible so make changes atomic.
   // A "strong" exchange ensures no false failures.
   uint32_t old_changes = changes.load(std::memory_order_relaxed);
   uint32_t new_changes = old_changes | kModuleInformationChanging;
@@ skipping 54 matching lines @@
       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::GlobalUserData::GlobalUserData(void* memory, size_t size)
+GlobalActivityTracker::ThreadSafeUserData::ThreadSafeUserData(void* memory,
+                                                              size_t size)
     : ActivityUserData(memory, size) {}
 
-GlobalActivityTracker::GlobalUserData::~GlobalUserData() {}
+GlobalActivityTracker::ThreadSafeUserData::~ThreadSafeUserData() {}
 
-void GlobalActivityTracker::GlobalUserData::Set(StringPiece name,
-                                                ValueType type,
-                                                const void* memory,
-                                                size_t size) {
+void GlobalActivityTracker::ThreadSafeUserData::Set(StringPiece name,
+                                                    ValueType type,
+                                                    const void* memory,
+                                                    size_t size) {
   AutoLock lock(data_lock_);
   ActivityUserData::Set(name, type, memory, size);
 }
 
 GlobalActivityTracker::ManagedActivityTracker::ManagedActivityTracker(
     PersistentMemoryAllocator::Reference mem_reference,
     void* base,
     size_t size)
     : ThreadActivityTracker(base, size),
       mem_reference_(mem_reference),
@@ skipping 104 matching lines @@
   return tracker;
 }
 
 void GlobalActivityTracker::ReleaseTrackerForCurrentThreadForTesting() {
   ThreadActivityTracker* tracker =
       reinterpret_cast<ThreadActivityTracker*>(this_thread_tracker_.Get());
   if (tracker)
     delete tracker;
 }
 
+void GlobalActivityTracker::SetBackgroundTaskRunner(
+    const scoped_refptr<TaskRunner>& runner) {
+  AutoLock lock(global_tracker_lock_);
+  background_task_runner_ = runner;
+}
+
+void GlobalActivityTracker::RecordProcessLaunch(ProcessId process_id) {
+  base::AutoLock lock(global_tracker_lock_);
+  DCHECK(!base::ContainsKey(known_processes_, process_id));

[manzagop (departed), 2017/02/22 20:44:15]

This is possible due to pid recycling. The map should contain ProcessInfo?

[bcwhite, 2017/02/22 22:13:02]

It's possible only if there was no corresponding RecordProcessExit for the
previous process of the same ID.  Every Launch should be matched by an Exit
call.  This tells me if I'm missing something.

But I should make the insert safe.

+  known_processes_.insert(process_id);
+}
+
+void GlobalActivityTracker::RecordProcessExit(ProcessId process_id,
+                                              int exit_code) {
+  DCHECK_NE(GetCurrentProcId(), process_id);
+
+  scoped_refptr<TaskRunner> task_runner;
+  {
+    base::AutoLock lock(global_tracker_lock_);
+    task_runner = background_task_runner_;
+    auto found = known_processes_.find(process_id);
+    if (found != known_processes_.end())
+      known_processes_.erase(found);
+    else
+      DLOG(ERROR) << "Recording exit of unknown process #" << process_id;
+  }
+
+  int64_t now_stamp = Time::Now().ToInternalValue();
+
+  // The persistent allocator is thread-safe so run the iteration and
+  // adjustments on a worker thread if one was provided.
+  if (task_runner && !task_runner->RunsTasksOnCurrentThread()) {
+    task_runner->PostTask(FROM_HERE,
+                          Bind(&GlobalActivityTracker::CleanupAfterProcess,
+                               Unretained(this), process_id, now_stamp));
+    return;
+  }
+
+  CleanupAfterProcess(process_id, now_stamp);
+}
+
+void GlobalActivityTracker::CleanupAfterProcess(ProcessId process_id,
+                                                int64_t exit_stamp) {
+  // The process may not have exited cleanly so its necessary to go through
+  // all the data structures it may have allocated in the persistent memory
+  // segment and mark them as "released". This will allow them to be reused
+  // later on. Memory is cleared here, rather than when it's needed, so as to
+  // limit the impact at that critical time.
+  PersistentMemoryAllocator::Iterator iter(allocator_.get());
+  PersistentMemoryAllocator::Reference ref;
+  uint32_t type;
+  while ((ref = iter.GetNext(&type)) != 0) {
+    const void* memory = allocator_->GetAsArray<char>(
+        ref, type, PersistentMemoryAllocator::kSizeAny);
+    ProcessId found_id;
+    int64_t create_stamp;
+
+    switch (type) {
+      case kTypeIdActivityTracker:
+      case kTypeIdUserDataRecord:
+      case kTypeIdProcessDataRecord:
+      case ModuleInfoRecord::kPersistentTypeId:
+        // By convention, the ProcessInfo structure is always the first
+        // field of the structure so there's no need to handle all the
+        // cases separately.
+        if (ProcessInfo::OwningProcessId(memory, &found_id, &create_stamp)) {
+          // Only change the type to be "free" if the process ID matches and
+          // the creation time is before the exit time (so PID re-use doesn't
+          // cause the erasure of something that is in-use).
+          if (found_id == process_id && create_stamp < exit_stamp)
+            allocator_->ChangeType(ref, ~type, type, /*clear=*/true);
+        }
+        break;
+    }
+  }
+}
+
 void GlobalActivityTracker::RecordLogMessage(StringPiece message) {
   // Allocate at least one extra byte so the string is NUL terminated. All
   // memory returned by the allocator is guaranteed to be zeroed.
   PersistentMemoryAllocator::Reference ref =
       allocator_->Allocate(message.size() + 1, kTypeIdGlobalLogMessage);
   char* memory = allocator_->GetAsArray<char>(ref, kTypeIdGlobalLogMessage,
                                               message.size() + 1);
   if (memory) {
     memcpy(memory, message.data(), message.size());
     allocator_->MakeIterable(ref);
@@ skipping 43 matching lines @@
                                 kTypeIdActivityTracker,
                                 kTypeIdActivityTrackerFree,
                                 stack_memory_size_,
                                 kCachedThreadMemories,
                                 /*make_iterable=*/true),
       user_data_allocator_(allocator_.get(),
                            kTypeIdUserDataRecord,
                            kTypeIdUserDataRecordFree,
                            kUserDataSize,
                            kCachedUserDataMemories,
-                           /*make_iterable=*/false),
+                           /*make_iterable=*/true),
+      process_data_(allocator_->GetAsArray<char>(
+                        AllocateFrom(allocator_.get(),
+                                     kTypeIdProcessDataRecordFree,
+                                     kProcessDataSize,
+                                     kTypeIdProcessDataRecord),
+                        kTypeIdProcessDataRecord,
+                        kProcessDataSize),
+                    kProcessDataSize),
       global_data_(
           allocator_->GetAsArray<char>(
               allocator_->Allocate(kGlobalDataSize, kTypeIdGlobalDataRecord),
               kTypeIdGlobalDataRecord,
-              PersistentMemoryAllocator::kSizeAny),
+              kGlobalDataSize),
           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_);
   subtle::Release_Store(&g_tracker_, reinterpret_cast<uintptr_t>(this));
 
-  // The global records must be iterable in order to be found by an analyzer.
+  // The data records must be iterable in order to be found by an analyzer.
+  allocator_->MakeIterable(allocator_->GetAsReference(
+      process_data_.GetBaseAddress(), kTypeIdProcessDataRecord));
   allocator_->MakeIterable(allocator_->GetAsReference(
       global_data_.GetBaseAddress(), kTypeIdGlobalDataRecord));
 
   // Fetch and record all activated field trials.
   FieldTrial::ActiveGroups active_groups;
   FieldTrialList::GetActiveFieldTrialGroups(&active_groups);
   for (auto& group : active_groups)
     RecordFieldTrial(group.trial_name, group.group_name);
 }
 
@@ skipping 102 matching lines @@
     : GlobalActivityTracker::ScopedThreadActivity(
           program_counter,
           nullptr,
           Activity::ACT_PROCESS_WAIT,
           ActivityData::ForProcess(process->Pid()),
           /*lock_allowed=*/true) {}
 #endif
 
 }  // namespace debug
 }  // namespace base