| OLD | NEW |
| (Empty) |
| 1 // Copyright 2016 The Chromium Authors. All rights reserved. | |
| 2 // Use of this source code is governed by a BSD-style license that can be | |
| 3 // found in the LICENSE file. | |
| 4 | |
| 5 #include "base/debug/activity_tracker.h" | |
| 6 | |
| 7 #include "base/debug/stack_trace.h" | |
| 8 #include "base/files/file.h" | |
| 9 #include "base/files/file_path.h" | |
| 10 #include "base/files/memory_mapped_file.h" | |
| 11 #include "base/logging.h" | |
| 12 #include "base/memory/ptr_util.h" | |
| 13 #include "base/metrics/field_trial.h" | |
| 14 #include "base/metrics/histogram_macros.h" | |
| 15 #include "base/pending_task.h" | |
| 16 #include "base/process/process.h" | |
| 17 #include "base/process/process_handle.h" | |
| 18 #include "base/stl_util.h" | |
| 19 #include "base/strings/string_util.h" | |
| 20 #include "base/threading/platform_thread.h" | |
| 21 | |
| 22 namespace base { | |
| 23 namespace debug { | |
| 24 | |
| 25 namespace { | |
| 26 | |
| 27 // A number that identifies the memory as having been initialized. It's | |
| 28 // arbitrary but happens to be the first 8 bytes of SHA1(ThreadActivityTracker). | |
| 29 // A version number is added on so that major structure changes won't try to | |
| 30 // read an older version (since the cookie won't match). | |
| 31 const uint64_t kHeaderCookie = 0xC0029B240D4A3092ULL + 1; // v1 | |
| 32 | |
| 33 // The minimum depth a stack should support. | |
| 34 const int kMinStackDepth = 2; | |
| 35 | |
| 36 } // namespace | |
| 37 | |
| 38 | |
| 39 // This information is kept for every thread that is tracked. It is filled | |
| 40 // the very first time the thread is seen. All fields must be of exact sizes | |
| 41 // so there is no issue moving between 32 and 64-bit builds. | |
| 42 struct ThreadActivityTracker::Header { | |
| 43 // This unique number indicates a valid initialization of the memory. | |
| 44 uint64_t cookie; | |
| 45 | |
| 46 // The process-id and thread-id to which this data belongs. These identifiers | |
| 47 // are not guaranteed to mean anything but are unique, in combination, among | |
| 48 // all active trackers. It would be nice to always have the process_id be a | |
| 49 // 64-bit value but the necessity of having it atomic (for the memory barriers | |
| 50 // it provides) limits it to the natural word size of the machine. | |
| 51 #ifdef ARCH_CPU_64_BITS | |
| 52 std::atomic<int64_t> process_id; | |
| 53 #else | |
| 54 std::atomic<int32_t> process_id; | |
| 55 int32_t process_id_padding; | |
| 56 #endif | |
| 57 | |
| 58 union { | |
| 59 int64_t as_id; | |
| 60 #if defined(OS_WIN) | |
| 61 // On Windows, the handle itself is often a pseudo-handle with a common | |
| 62 // value meaning "this thread" and so the thread-id is used. The former | |
| 63 // can be converted to a thread-id with a system call. | |
| 64 PlatformThreadId as_tid; | |
| 65 #elif defined(OS_POSIX) | |
| 66 // On Posix, the handle is always a unique identifier so no conversion | |
| 67 // needs to be done. However, it's value is officially opaque so there | |
| 68 // is no one correct way to convert it to a numerical identifier. | |
| 69 PlatformThreadHandle::Handle as_handle; | |
| 70 #endif | |
| 71 } thread_ref; | |
| 72 | |
| 73 // The start-time and start-ticks when the data was created. Each activity | |
| 74 // record has a |time_internal| value that can be converted to a "wall time" | |
| 75 // with these two values. | |
| 76 int64_t start_time; | |
| 77 int64_t start_ticks; | |
| 78 | |
| 79 // The number of Activity slots in the data. | |
| 80 uint32_t stack_slots; | |
| 81 | |
| 82 // The current depth of the stack. This may be greater than the number of | |
| 83 // slots. If the depth exceeds the number of slots, the newest entries | |
| 84 // won't be recorded. | |
| 85 std::atomic<uint32_t> current_depth; | |
| 86 | |
| 87 // A memory location used to indicate if changes have been made to the stack | |
| 88 // that would invalidate an in-progress read of its contents. The active | |
| 89 // tracker will zero the value whenever something gets popped from the | |
| 90 // stack. A monitoring tracker can write a non-zero value here, copy the | |
| 91 // stack contents, and read the value to know, if it is still non-zero, that | |
| 92 // the contents didn't change while being copied. This can handle concurrent | |
| 93 // snapshot operations only if each snapshot writes a different bit (which | |
| 94 // is not the current implementation so no parallel snapshots allowed). | |
| 95 std::atomic<uint32_t> stack_unchanged; | |
| 96 | |
| 97 // The name of the thread (up to a maximum length). Dynamic-length names | |
| 98 // are not practical since the memory has to come from the same persistent | |
| 99 // allocator that holds this structure and to which this object has no | |
| 100 // reference. | |
| 101 char thread_name[32]; | |
| 102 }; | |
| 103 | |
| 104 // It doesn't matter what is contained in this (though it will be all zeros) | |
| 105 // as only the address of it is important. | |
| 106 const ThreadActivityTracker::ActivityData | |
| 107 ThreadActivityTracker::kNullActivityData = {}; | |
| 108 | |
| 109 ThreadActivityTracker::ActivityData | |
| 110 ThreadActivityTracker::ActivityData::ForThread( | |
| 111 const PlatformThreadHandle& handle) { | |
| 112 // Header already has a conversion union; reuse that. | |
| 113 ThreadActivityTracker::Header header; | |
| 114 header.thread_ref.as_id = 0; // Zero the union in case other is smaller. | |
| 115 #if defined(OS_WIN) | |
| 116 header.thread_ref.as_tid = ::GetThreadId(handle.platform_handle()); | |
| 117 #elif defined(OS_POSIX) | |
| 118 header.thread_ref.as_handle = handle.platform_handle(); | |
| 119 #endif | |
| 120 return ForThread(header.thread_ref.as_id); | |
| 121 } | |
| 122 | |
| 123 ThreadActivityTracker::ActivitySnapshot::ActivitySnapshot() {} | |
| 124 ThreadActivityTracker::ActivitySnapshot::~ActivitySnapshot() {} | |
| 125 | |
| 126 | |
| 127 ThreadActivityTracker::ThreadActivityTracker(void* base, size_t size) | |
| 128 : header_(static_cast<Header*>(base)), | |
| 129 stack_(reinterpret_cast<Activity*>(reinterpret_cast<char*>(base) + | |
| 130 sizeof(Header))), | |
| 131 stack_slots_( | |
| 132 static_cast<uint32_t>((size - sizeof(Header)) / sizeof(Activity))) { | |
| 133 DCHECK(thread_checker_.CalledOnValidThread()); | |
| 134 | |
| 135 // Verify the parameters but fail gracefully if they're not valid so that | |
| 136 // production code based on external inputs will not crash. IsValid() will | |
| 137 // return false in this case. | |
| 138 if (!base || | |
| 139 // Ensure there is enough space for the header and at least a few records. | |
| 140 size < sizeof(Header) + kMinStackDepth * sizeof(Activity) || | |
| 141 // Ensure that the |stack_slots_| calculation didn't overflow. | |
| 142 (size - sizeof(Header)) / sizeof(Activity) > | |
| 143 std::numeric_limits<uint32_t>::max()) { | |
| 144 NOTREACHED(); | |
| 145 return; | |
| 146 } | |
| 147 | |
| 148 // Ensure that the thread reference doesn't exceed the size of the ID number. | |
| 149 // This won't compile at the global scope because Header is a private struct. | |
| 150 static_assert( | |
| 151 sizeof(header_->thread_ref) == sizeof(header_->thread_ref.as_id), | |
| 152 "PlatformThreadHandle::Handle is too big to hold in 64-bit ID"); | |
| 153 | |
| 154 // Ensure that the alignment of Activity.data is properly aligned to a | |
| 155 // 64-bit boundary so there are no interoperability-issues across cpu | |
| 156 // architectures. | |
| 157 static_assert(offsetof(Activity, data) % sizeof(uint64_t) == 0, | |
| 158 "ActivityData.data is not 64-bit aligned"); | |
| 159 | |
| 160 // Provided memory should either be completely initialized or all zeros. | |
| 161 if (header_->cookie == 0) { | |
| 162 // This is a new file. Double-check other fields and then initialize. | |
| 163 DCHECK_EQ(0, header_->process_id.load(std::memory_order_relaxed)); | |
| 164 DCHECK_EQ(0, header_->thread_ref.as_id); | |
| 165 DCHECK_EQ(0, header_->start_time); | |
| 166 DCHECK_EQ(0, header_->start_ticks); | |
| 167 DCHECK_EQ(0U, header_->stack_slots); | |
| 168 DCHECK_EQ(0U, header_->current_depth.load(std::memory_order_relaxed)); | |
| 169 DCHECK_EQ(0U, header_->stack_unchanged.load(std::memory_order_relaxed)); | |
| 170 DCHECK_EQ(0, stack_[0].time_internal); | |
| 171 DCHECK_EQ(0U, stack_[0].origin_address); | |
| 172 DCHECK_EQ(0U, stack_[0].call_stack[0]); | |
| 173 DCHECK_EQ(0U, stack_[0].data.task.sequence_id); | |
| 174 | |
| 175 #if defined(OS_WIN) | |
| 176 header_->thread_ref.as_tid = PlatformThread::CurrentId(); | |
| 177 #elif defined(OS_POSIX) | |
| 178 header_->thread_ref.as_handle = | |
| 179 PlatformThread::CurrentHandle().platform_handle(); | |
| 180 #endif | |
| 181 header_->start_time = base::Time::Now().ToInternalValue(); | |
| 182 header_->start_ticks = base::TimeTicks::Now().ToInternalValue(); | |
| 183 header_->stack_slots = stack_slots_; | |
| 184 strlcpy(header_->thread_name, PlatformThread::GetName(), | |
| 185 sizeof(header_->thread_name)); | |
| 186 header_->cookie = kHeaderCookie; | |
| 187 | |
| 188 // This is done last so as to guarantee that everything above is "released" | |
| 189 // by the time this value gets written. | |
| 190 header_->process_id.store(GetCurrentProcId(), std::memory_order_release); | |
| 191 | |
| 192 valid_ = true; | |
| 193 DCHECK(IsValid()); | |
| 194 } else { | |
| 195 // This is a file with existing data. Perform basic consistency checks. | |
| 196 valid_ = true; | |
| 197 valid_ = IsValid(); | |
| 198 } | |
| 199 } | |
| 200 | |
| 201 ThreadActivityTracker::~ThreadActivityTracker() {} | |
| 202 | |
| 203 void ThreadActivityTracker::PushActivity(const void* origin, | |
| 204 ActivityType type, | |
| 205 const ActivityData& data) { | |
| 206 // A thread-checker creates a lock to check the thread-id which means | |
| 207 // re-entry into this code if lock acquisitions are being tracked. | |
| 208 DCHECK(type == ACT_LOCK_ACQUIRE || thread_checker_.CalledOnValidThread()); | |
| 209 | |
| 210 // Get the current depth of the stack. No access to other memory guarded | |
| 211 // by this variable is done here so a "relaxed" load is acceptable. | |
| 212 uint32_t depth = header_->current_depth.load(std::memory_order_relaxed); | |
| 213 | |
| 214 // Handle the case where the stack depth has exceeded the storage capacity. | |
| 215 // Extra entries will be lost leaving only the base of the stack. | |
| 216 if (depth >= stack_slots_) { | |
| 217 // Since no other threads modify the data, no compare/exchange is needed. | |
| 218 // Since no other memory is being modified, a "relaxed" store is acceptable. | |
| 219 header_->current_depth.store(depth + 1, std::memory_order_relaxed); | |
| 220 return; | |
| 221 } | |
| 222 | |
| 223 // Get a pointer to the next activity and load it. No atomicity is required | |
| 224 // here because the memory is known only to this thread. It will be made | |
| 225 // known to other threads once the depth is incremented. | |
| 226 Activity* activity = &stack_[depth]; | |
| 227 activity->time_internal = base::TimeTicks::Now().ToInternalValue(); | |
| 228 activity->origin_address = reinterpret_cast<uintptr_t>(origin); | |
| 229 activity->activity_type = type; | |
| 230 activity->data = data; | |
| 231 | |
| 232 #if defined(SYZYASAN) | |
| 233 // Create a stacktrace from the current location and get the addresses. | |
| 234 StackTrace stack_trace; | |
| 235 size_t stack_depth; | |
| 236 const void* const* stack_addrs = stack_trace.Addresses(&stack_depth); | |
| 237 // Copy the stack addresses, ignoring the first one (here). | |
| 238 size_t i; | |
| 239 for (i = 1; i < stack_depth && i < kActivityCallStackSize; ++i) { | |
| 240 activity->call_stack[i - 1] = reinterpret_cast<uintptr_t>(stack_addrs[i]); | |
| 241 } | |
| 242 activity->call_stack[i - 1] = 0; | |
| 243 #else | |
| 244 // Since the memory was initially zero and nothing ever overwrites it in | |
| 245 // this "else" case, there is no need to write even the null terminator. | |
| 246 //activity->call_stack[0] = 0; | |
| 247 #endif | |
| 248 | |
| 249 // Save the incremented depth. Because this guards |activity| memory filled | |
| 250 // above that may be read by another thread once the recorded depth changes, | |
| 251 // a "release" store is required. | |
| 252 header_->current_depth.store(depth + 1, std::memory_order_release); | |
| 253 } | |
| 254 | |
| 255 void ThreadActivityTracker::ChangeActivity(ActivityType type, | |
| 256 const ActivityData& data) { | |
| 257 DCHECK(thread_checker_.CalledOnValidThread()); | |
| 258 DCHECK(type != ACT_NULL || &data != &kNullActivityData); | |
| 259 | |
| 260 // Get the current depth of the stack and acquire the data held there. | |
| 261 uint32_t depth = header_->current_depth.load(std::memory_order_acquire); | |
| 262 DCHECK_LT(0U, depth); | |
| 263 | |
| 264 // Update the information if it is being recorded (i.e. within slot limit). | |
| 265 if (depth <= stack_slots_) { | |
| 266 Activity* activity = &stack_[depth - 1]; | |
| 267 | |
| 268 if (type != ACT_NULL) { | |
| 269 DCHECK_EQ(activity->activity_type & ACT_CATEGORY_MASK, | |
| 270 type & ACT_CATEGORY_MASK); | |
| 271 activity->activity_type = type; | |
| 272 } | |
| 273 | |
| 274 if (&data != &kNullActivityData) | |
| 275 activity->data = data; | |
| 276 } | |
| 277 } | |
| 278 | |
| 279 void ThreadActivityTracker::PopActivity() { | |
| 280 // Do an atomic decrement of the depth. No changes to stack entries guarded | |
| 281 // by this variable are done here so a "relaxed" operation is acceptable. | |
| 282 // |depth| will receive the value BEFORE it was modified. | |
| 283 uint32_t depth = | |
| 284 header_->current_depth.fetch_sub(1, std::memory_order_relaxed); | |
| 285 | |
| 286 // Validate that everything is running correctly. | |
| 287 DCHECK_LT(0U, depth); | |
| 288 | |
| 289 // A thread-checker creates a lock to check the thread-id which means | |
| 290 // re-entry into this code if lock acquisitions are being tracked. | |
| 291 DCHECK(stack_[depth - 1].activity_type == ACT_LOCK_ACQUIRE || | |
| 292 thread_checker_.CalledOnValidThread()); | |
| 293 | |
| 294 // The stack has shrunk meaning that some other thread trying to copy the | |
| 295 // contents for reporting purposes could get bad data. That thread would | |
| 296 // have written a non-zero value into |stack_unchanged|; clearing it here | |
| 297 // will let that thread detect that something did change. This needs to | |
| 298 // happen after the atomic |depth| operation above so a "release" store | |
| 299 // is required. | |
| 300 header_->stack_unchanged.store(0, std::memory_order_release); | |
| 301 } | |
| 302 | |
| 303 bool ThreadActivityTracker::IsValid() const { | |
| 304 if (header_->cookie != kHeaderCookie || | |
| 305 header_->process_id.load(std::memory_order_relaxed) == 0 || | |
| 306 header_->thread_ref.as_id == 0 || | |
| 307 header_->start_time == 0 || | |
| 308 header_->start_ticks == 0 || | |
| 309 header_->stack_slots != stack_slots_ || | |
| 310 header_->thread_name[sizeof(header_->thread_name) - 1] != '\0') { | |
| 311 return false; | |
| 312 } | |
| 313 | |
| 314 return valid_; | |
| 315 } | |
| 316 | |
| 317 bool ThreadActivityTracker::Snapshot(ActivitySnapshot* output_snapshot) const { | |
| 318 DCHECK(output_snapshot); | |
| 319 | |
| 320 // There is no "called on valid thread" check for this method as it can be | |
| 321 // called from other threads or even other processes. It is also the reason | |
| 322 // why atomic operations must be used in certain places above. | |
| 323 | |
| 324 // It's possible for the data to change while reading it in such a way that it | |
| 325 // invalidates the read. Make several attempts but don't try forever. | |
| 326 const int kMaxAttempts = 10; | |
| 327 uint32_t depth; | |
| 328 | |
| 329 // Stop here if the data isn't valid. | |
| 330 if (!IsValid()) | |
| 331 return false; | |
| 332 | |
| 333 // Allocate the maximum size for the stack so it doesn't have to be done | |
| 334 // during the time-sensitive snapshot operation. It is shrunk once the | |
| 335 // actual size is known. | |
| 336 output_snapshot->activity_stack.reserve(stack_slots_); | |
| 337 | |
| 338 for (int attempt = 0; attempt < kMaxAttempts; ++attempt) { | |
| 339 // Remember the process and thread IDs to ensure they aren't replaced | |
| 340 // during the snapshot operation. Use "acquire" to ensure that all the | |
| 341 // non-atomic fields of the structure are valid (at least at the current | |
| 342 // moment in time). | |
| 343 const int64_t starting_process_id = | |
| 344 header_->process_id.load(std::memory_order_acquire); | |
| 345 const int64_t starting_thread_id = header_->thread_ref.as_id; | |
| 346 | |
| 347 // Write a non-zero value to |stack_unchanged| so it's possible to detect | |
| 348 // at the end that nothing has changed since copying the data began. A | |
| 349 // "cst" operation is required to ensure it occurs before everything else. | |
| 350 // Using "cst" memory ordering is relatively expensive but this is only | |
| 351 // done during analysis so doesn't directly affect the worker threads. | |
| 352 header_->stack_unchanged.store(1, std::memory_order_seq_cst); | |
| 353 | |
| 354 // Fetching the current depth also "acquires" the contents of the stack. | |
| 355 depth = header_->current_depth.load(std::memory_order_acquire); | |
| 356 uint32_t count = std::min(depth, stack_slots_); | |
| 357 output_snapshot->activity_stack.resize(count); | |
| 358 if (count > 0) { | |
| 359 // Copy the existing contents. Memcpy is used for speed. | |
| 360 memcpy(&output_snapshot->activity_stack[0], stack_, | |
| 361 count * sizeof(Activity)); | |
| 362 } | |
| 363 | |
| 364 // Retry if something changed during the copy. A "cst" operation ensures | |
| 365 // it must happen after all the above operations. | |
| 366 if (!header_->stack_unchanged.load(std::memory_order_seq_cst)) | |
| 367 continue; | |
| 368 | |
| 369 // Stack copied. Record it's full depth. | |
| 370 output_snapshot->activity_stack_depth = depth; | |
| 371 | |
| 372 // TODO(bcwhite): Snapshot other things here. | |
| 373 | |
| 374 // Get the general thread information. Loading of "process_id" is guaranteed | |
| 375 // to be last so that it's possible to detect below if any content has | |
| 376 // changed while reading it. It's technically possible for a thread to end, | |
| 377 // have its data cleared, a new thread get created with the same IDs, and | |
| 378 // it perform an action which starts tracking all in the time since the | |
| 379 // ID reads above but the chance is so unlikely that it's not worth the | |
| 380 // effort and complexity of protecting against it (perhaps with an | |
| 381 // "unchanged" field like is done for the stack). | |
| 382 output_snapshot->thread_name = | |
| 383 std::string(header_->thread_name, sizeof(header_->thread_name) - 1); | |
| 384 output_snapshot->thread_id = header_->thread_ref.as_id; | |
| 385 output_snapshot->process_id = | |
| 386 header_->process_id.load(std::memory_order_seq_cst); | |
| 387 | |
| 388 // All characters of the thread-name buffer were copied so as to not break | |
| 389 // if the trailing NUL were missing. Now limit the length if the actual | |
| 390 // name is shorter. | |
| 391 output_snapshot->thread_name.resize( | |
| 392 strlen(output_snapshot->thread_name.c_str())); | |
| 393 | |
| 394 // If the process or thread ID has changed then the tracker has exited and | |
| 395 // the memory reused by a new one. Try again. | |
| 396 if (output_snapshot->process_id != starting_process_id || | |
| 397 output_snapshot->thread_id != starting_thread_id) { | |
| 398 continue; | |
| 399 } | |
| 400 | |
| 401 // Only successful if the data is still valid once everything is done since | |
| 402 // it's possible for the thread to end somewhere in the middle and all its | |
| 403 // values become garbage. | |
| 404 if (!IsValid()) | |
| 405 return false; | |
| 406 | |
| 407 // Change all the timestamps in the activities from "ticks" to "wall" time. | |
| 408 const Time start_time = Time::FromInternalValue(header_->start_time); | |
| 409 const int64_t start_ticks = header_->start_ticks; | |
| 410 for (Activity& activity : output_snapshot->activity_stack) { | |
| 411 activity.time_internal = | |
| 412 (start_time + | |
| 413 TimeDelta::FromInternalValue(activity.time_internal - start_ticks)) | |
| 414 .ToInternalValue(); | |
| 415 } | |
| 416 | |
| 417 // Success! | |
| 418 return true; | |
| 419 } | |
| 420 | |
| 421 // Too many attempts. | |
| 422 return false; | |
| 423 } | |
| 424 | |
| 425 // static | |
| 426 size_t ThreadActivityTracker::SizeForStackDepth(int stack_depth) { | |
| 427 return static_cast<size_t>(stack_depth) * sizeof(Activity) + sizeof(Header); | |
| 428 } | |
| 429 | |
| 430 | |
| 431 GlobalActivityTracker* GlobalActivityTracker::g_tracker_ = nullptr; | |
| 432 | |
| 433 GlobalActivityTracker::ManagedActivityTracker::ManagedActivityTracker( | |
| 434 PersistentMemoryAllocator::Reference mem_reference, | |
| 435 void* base, | |
| 436 size_t size) | |
| 437 : ThreadActivityTracker(base, size), | |
| 438 mem_reference_(mem_reference), | |
| 439 mem_base_(base) {} | |
| 440 | |
| 441 GlobalActivityTracker::ManagedActivityTracker::~ManagedActivityTracker() { | |
| 442 // The global |g_tracker_| must point to the owner of this class since all | |
| 443 // objects of this type must be destructed before |g_tracker_| can be changed | |
| 444 // (something that only occurs in tests). | |
| 445 DCHECK(g_tracker_); | |
| 446 g_tracker_->ReturnTrackerMemory(this); | |
| 447 } | |
| 448 | |
| 449 void GlobalActivityTracker::CreateWithAllocator( | |
| 450 std::unique_ptr<PersistentMemoryAllocator> allocator, | |
| 451 int stack_depth) { | |
| 452 // There's no need to do anything with the result. It is self-managing. | |
| 453 GlobalActivityTracker* global_tracker = | |
| 454 new GlobalActivityTracker(std::move(allocator), stack_depth); | |
| 455 // Create a tracker for this thread since it is known. | |
| 456 global_tracker->CreateTrackerForCurrentThread(); | |
| 457 } | |
| 458 | |
| 459 #if !defined(OS_NACL) | |
| 460 // static | |
| 461 void GlobalActivityTracker::CreateWithFile(const FilePath& file_path, | |
| 462 size_t size, | |
| 463 uint64_t id, | |
| 464 StringPiece name, | |
| 465 int stack_depth) { | |
| 466 DCHECK(!file_path.empty()); | |
| 467 DCHECK_GE(static_cast<uint64_t>(std::numeric_limits<int64_t>::max()), size); | |
| 468 | |
| 469 // Create and map the file into memory and make it globally available. | |
| 470 std::unique_ptr<MemoryMappedFile> mapped_file(new MemoryMappedFile()); | |
| 471 bool success = | |
| 472 mapped_file->Initialize(File(file_path, | |
| 473 File::FLAG_CREATE_ALWAYS | File::FLAG_READ | | |
| 474 File::FLAG_WRITE | File::FLAG_SHARE_DELETE), | |
| 475 {0, static_cast<int64_t>(size)}, | |
| 476 MemoryMappedFile::READ_WRITE_EXTEND); | |
| 477 DCHECK(success); | |
| 478 CreateWithAllocator(WrapUnique(new FilePersistentMemoryAllocator( | |
| 479 std::move(mapped_file), size, id, name, false)), | |
| 480 stack_depth); | |
| 481 } | |
| 482 #endif // !defined(OS_NACL) | |
| 483 | |
| 484 // static | |
| 485 void GlobalActivityTracker::CreateWithLocalMemory(size_t size, | |
| 486 uint64_t id, | |
| 487 StringPiece name, | |
| 488 int stack_depth) { | |
| 489 CreateWithAllocator( | |
| 490 WrapUnique(new LocalPersistentMemoryAllocator(size, id, name)), | |
| 491 stack_depth); | |
| 492 } | |
| 493 | |
| 494 ThreadActivityTracker* GlobalActivityTracker::CreateTrackerForCurrentThread() { | |
| 495 DCHECK(!this_thread_tracker_.Get()); | |
| 496 | |
| 497 PersistentMemoryAllocator::Reference mem_reference = 0; | |
| 498 void* mem_base = nullptr; | |
| 499 | |
| 500 // Get the current count of available memories, acquiring the array values. | |
| 501 int count = available_memories_count_.load(std::memory_order_acquire); | |
| 502 while (count > 0) { | |
| 503 // There is a memory block that was previously released (and zeroed) so | |
| 504 // just re-use that rather than allocating a new one. Use "relaxed" because | |
| 505 // the value is guarded by the |count| "acquire". A zero reference replaces | |
| 506 // the existing value so that it can't be used by another thread that | |
| 507 // manages to interrupt this one before the count can be decremented. | |
| 508 // A zero reference is also required for the "push" operation to work | |
| 509 // once the count finally does get decremented. | |
| 510 mem_reference = | |
| 511 available_memories_[count - 1].exchange(0, std::memory_order_relaxed); | |
| 512 | |
| 513 // If the reference is zero, it's already been taken but count hasn't yet | |
| 514 // been decremented. Give that other thread a chance to finish then reload | |
| 515 // the "count" value and try again. | |
| 516 if (!mem_reference) { | |
| 517 PlatformThread::YieldCurrentThread(); | |
| 518 count = available_memories_count_.load(std::memory_order_acquire); | |
| 519 continue; | |
| 520 } | |
| 521 | |
| 522 // Decrement the count indicating that the value has been taken. If this | |
| 523 // fails then another thread has pushed something new and incremented the | |
| 524 // count. | |
| 525 // NOTE: |oldcount| will be loaded with the existing value. | |
| 526 int oldcount = count; | |
| 527 if (!available_memories_count_.compare_exchange_strong( | |
| 528 oldcount, count - 1, std::memory_order_acquire, | |
| 529 std::memory_order_acquire)) { | |
| 530 DCHECK_LT(count, oldcount); | |
| 531 | |
| 532 // Restore the reference that was zeroed above and try again. | |
| 533 available_memories_[count - 1].store(mem_reference, | |
| 534 std::memory_order_relaxed); | |
| 535 count = oldcount; | |
| 536 continue; | |
| 537 } | |
| 538 | |
| 539 // Turn the reference back into one of the activity-tracker type. | |
| 540 mem_base = allocator_->GetAsObject<char>(mem_reference, | |
| 541 kTypeIdActivityTrackerFree); | |
| 542 DCHECK(mem_base); | |
| 543 DCHECK_LE(stack_memory_size_, allocator_->GetAllocSize(mem_reference)); | |
| 544 bool changed = allocator_->ChangeType(mem_reference, kTypeIdActivityTracker, | |
| 545 kTypeIdActivityTrackerFree); | |
| 546 DCHECK(changed); | |
| 547 | |
| 548 // Success. | |
| 549 break; | |
| 550 } | |
| 551 | |
| 552 // Handle the case where no previously-used memories are available. | |
| 553 if (count == 0) { | |
| 554 // Allocate a block of memory from the persistent segment. | |
| 555 mem_reference = | |
| 556 allocator_->Allocate(stack_memory_size_, kTypeIdActivityTracker); | |
| 557 if (mem_reference) { | |
| 558 // Success. Convert the reference to an actual memory address. | |
| 559 mem_base = | |
| 560 allocator_->GetAsObject<char>(mem_reference, kTypeIdActivityTracker); | |
| 561 // Make the allocation iterable so it can be found by other processes. | |
| 562 allocator_->MakeIterable(mem_reference); | |
| 563 } else { | |
| 564 // Failure. This shouldn't happen. | |
| 565 NOTREACHED(); | |
| 566 // But if it does, probably because the allocator wasn't given enough | |
| 567 // memory to satisfy all possible requests, handle it gracefully by | |
| 568 // allocating the required memory from the heap. | |
| 569 mem_base = new char[stack_memory_size_]; | |
| 570 memset(mem_base, 0, stack_memory_size_); | |
| 571 // Report the thread-count at which the allocator was full so that the | |
| 572 // failure can be seen and underlying memory resized appropriately. | |
| 573 UMA_HISTOGRAM_COUNTS_1000( | |
| 574 "ActivityTracker.ThreadTrackers.MemLimitTrackerCount", | |
| 575 thread_tracker_count_.load(std::memory_order_relaxed)); | |
| 576 } | |
| 577 } | |
| 578 | |
| 579 // Create a tracker with the acquired memory and set it as the tracker | |
| 580 // for this particular thread in thread-local-storage. | |
| 581 DCHECK(mem_base); | |
| 582 ManagedActivityTracker* tracker = | |
| 583 new ManagedActivityTracker(mem_reference, mem_base, stack_memory_size_); | |
| 584 DCHECK(tracker->IsValid()); | |
| 585 this_thread_tracker_.Set(tracker); | |
| 586 int old_count = thread_tracker_count_.fetch_add(1, std::memory_order_relaxed); | |
| 587 | |
| 588 UMA_HISTOGRAM_ENUMERATION("ActivityTracker.ThreadTrackers.Count", | |
| 589 old_count + 1, kMaxThreadCount); | |
| 590 return tracker; | |
| 591 } | |
| 592 | |
| 593 void GlobalActivityTracker::ReleaseTrackerForCurrentThreadForTesting() { | |
| 594 ThreadActivityTracker* tracker = | |
| 595 reinterpret_cast<ThreadActivityTracker*>(this_thread_tracker_.Get()); | |
| 596 if (tracker) { | |
| 597 this_thread_tracker_.Free(); | |
| 598 delete tracker; | |
| 599 } | |
| 600 } | |
| 601 | |
| 602 GlobalActivityTracker::GlobalActivityTracker( | |
| 603 std::unique_ptr<PersistentMemoryAllocator> allocator, | |
| 604 int stack_depth) | |
| 605 : allocator_(std::move(allocator)), | |
| 606 stack_memory_size_(ThreadActivityTracker::SizeForStackDepth(stack_depth)), | |
| 607 this_thread_tracker_(&OnTLSDestroy), | |
| 608 thread_tracker_count_(0), | |
| 609 available_memories_count_(0) { | |
| 610 // Clear the available-memories array. | |
| 611 memset(available_memories_, 0, sizeof(available_memories_)); | |
| 612 | |
| 613 // Ensure the passed memory is valid and empty (iterator finds nothing). | |
| 614 uint32_t type; | |
| 615 DCHECK(!PersistentMemoryAllocator::Iterator(allocator_.get()).GetNext(&type)); | |
| 616 | |
| 617 // Ensure that there is no other global object and then make this one such. | |
| 618 DCHECK(!g_tracker_); | |
| 619 g_tracker_ = this; | |
| 620 } | |
| 621 | |
| 622 GlobalActivityTracker::~GlobalActivityTracker() { | |
| 623 DCHECK_EQ(g_tracker_, this); | |
| 624 DCHECK_EQ(0, thread_tracker_count_.load(std::memory_order_relaxed)); | |
| 625 g_tracker_ = nullptr; | |
| 626 } | |
| 627 | |
| 628 void GlobalActivityTracker::ReturnTrackerMemory( | |
| 629 ManagedActivityTracker* tracker) { | |
| 630 PersistentMemoryAllocator::Reference mem_reference = tracker->mem_reference_; | |
| 631 void* mem_base = tracker->mem_base_; | |
| 632 | |
| 633 // Zero the memory so that it is ready for use if needed again later. It's | |
| 634 // better to clear the memory now, when a thread is exiting, than to do it | |
| 635 // when it is first needed by a thread doing actual work. | |
| 636 memset(mem_base, 0, stack_memory_size_); | |
| 637 | |
| 638 // Remove the destructed tracker from the set of known ones. | |
| 639 DCHECK_LE(1, thread_tracker_count_.load(std::memory_order_relaxed)); | |
| 640 thread_tracker_count_.fetch_sub(1, std::memory_order_relaxed); | |
| 641 | |
| 642 // Deal with the memory that was used by the tracker. | |
| 643 if (mem_reference) { | |
| 644 // The memory was within the persistent memory allocator. Change its type | |
| 645 // so that iteration won't find it. | |
| 646 allocator_->ChangeType(mem_reference, kTypeIdActivityTrackerFree, | |
| 647 kTypeIdActivityTracker); | |
| 648 // There is no way to free memory from a persistent allocator so instead | |
| 649 // push it on the internal list of available memory blocks. | |
| 650 while (true) { | |
| 651 // Get the existing count of available memories and ensure we won't | |
| 652 // burst the array. Acquire the values in the array. | |
| 653 int count = available_memories_count_.load(std::memory_order_acquire); | |
| 654 if (count >= kMaxThreadCount) { | |
| 655 NOTREACHED(); | |
| 656 // Storage is full. Just forget about this memory. It won't be re-used | |
| 657 // but there's no real loss. | |
| 658 break; | |
| 659 } | |
| 660 | |
| 661 // Write the reference of the memory being returned to this slot in the | |
| 662 // array. Empty slots have a value of zero so do an atomic compare-and- | |
| 663 // exchange to ensure that a race condition doesn't exist with another | |
| 664 // thread doing the same. | |
| 665 PersistentMemoryAllocator::Reference mem_expected = 0; | |
| 666 if (!available_memories_[count].compare_exchange_strong( | |
| 667 mem_expected, mem_reference, std::memory_order_release, | |
| 668 std::memory_order_relaxed)) { | |
| 669 PlatformThread::YieldCurrentThread(); | |
| 670 continue; // Try again. | |
| 671 } | |
| 672 | |
| 673 // Increment the count, releasing the value written to the array. This | |
| 674 // could fail if a simultaneous "pop" operation decremented the counter. | |
| 675 // If that happens, clear the array slot and start over. Do a "strong" | |
| 676 // exchange to avoid spurious retries that can occur with a "weak" one. | |
| 677 int expected = count; // Updated by compare/exchange. | |
| 678 if (!available_memories_count_.compare_exchange_strong( | |
| 679 expected, count + 1, std::memory_order_release, | |
| 680 std::memory_order_relaxed)) { | |
| 681 available_memories_[count].store(0, std::memory_order_relaxed); | |
| 682 continue; | |
| 683 } | |
| 684 | |
| 685 // Count was successfully incremented to reflect the newly added value. | |
| 686 break; | |
| 687 } | |
| 688 } else { | |
| 689 // The memory was allocated from the process heap. This shouldn't happen | |
| 690 // because the persistent memory segment should be big enough for all | |
| 691 // thread stacks but it's better to support falling back to allocation | |
| 692 // from the heap rather than crash. Everything will work as normal but | |
| 693 // the data won't be persisted. | |
| 694 delete[] reinterpret_cast<char*>(mem_base); | |
| 695 } | |
| 696 } | |
| 697 | |
| 698 // static | |
| 699 void GlobalActivityTracker::OnTLSDestroy(void* value) { | |
| 700 delete reinterpret_cast<ManagedActivityTracker*>(value); | |
| 701 } | |
| 702 | |
| 703 | |
| 704 ScopedActivity::ScopedActivity(const tracked_objects::Location& location, | |
| 705 uint8_t action, | |
| 706 uint32_t id, | |
| 707 int32_t info) | |
| 708 : GlobalActivityTracker::ScopedThreadActivity( | |
| 709 location.program_counter(), | |
| 710 static_cast<ThreadActivityTracker::ActivityType>( | |
| 711 ThreadActivityTracker::ACT_GENERIC | action), | |
| 712 ThreadActivityTracker::ActivityData::ForGeneric(id, info), | |
| 713 /*lock_allowed=*/true), | |
| 714 id_(id) { | |
| 715 // The action must not affect the category bits of the activity type. | |
| 716 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
| 717 } | |
| 718 | |
| 719 void ScopedActivity::ChangeAction(uint8_t action) { | |
| 720 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
| 721 ChangeTypeAndData(static_cast<ThreadActivityTracker::ActivityType>( | |
| 722 ThreadActivityTracker::ACT_GENERIC | action), | |
| 723 ThreadActivityTracker::kNullActivityData); | |
| 724 } | |
| 725 | |
| 726 void ScopedActivity::ChangeInfo(int32_t info) { | |
| 727 ChangeTypeAndData(ThreadActivityTracker::ACT_NULL, | |
| 728 ThreadActivityTracker::ActivityData::ForGeneric(id_, info)); | |
| 729 } | |
| 730 | |
| 731 void ScopedActivity::ChangeActionAndInfo(uint8_t action, int32_t info) { | |
| 732 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
| 733 ChangeTypeAndData(static_cast<ThreadActivityTracker::ActivityType>( | |
| 734 ThreadActivityTracker::ACT_GENERIC | action), | |
| 735 ThreadActivityTracker::ActivityData::ForGeneric(id_, info)); | |
| 736 } | |
| 737 | |
| 738 ScopedTaskRunActivity::ScopedTaskRunActivity(const base::PendingTask& task) | |
| 739 : GlobalActivityTracker::ScopedThreadActivity( | |
| 740 task.posted_from.program_counter(), | |
| 741 ThreadActivityTracker::ACT_TASK_RUN, | |
| 742 ThreadActivityTracker::ActivityData::ForTask(task.sequence_num), | |
| 743 /*lock_allowed=*/true) {} | |
| 744 | |
| 745 ScopedLockAcquireActivity::ScopedLockAcquireActivity( | |
| 746 const base::internal::LockImpl* lock) | |
| 747 : GlobalActivityTracker::ScopedThreadActivity( | |
| 748 nullptr, | |
| 749 ThreadActivityTracker::ACT_LOCK_ACQUIRE, | |
| 750 ThreadActivityTracker::ActivityData::ForLock(lock), | |
| 751 /*lock_allowed=*/false) {} | |
| 752 | |
| 753 ScopedEventWaitActivity::ScopedEventWaitActivity( | |
| 754 const base::WaitableEvent* event) | |
| 755 : GlobalActivityTracker::ScopedThreadActivity( | |
| 756 nullptr, | |
| 757 ThreadActivityTracker::ACT_EVENT_WAIT, | |
| 758 ThreadActivityTracker::ActivityData::ForEvent(event), | |
| 759 /*lock_allowed=*/true) {} | |
| 760 | |
| 761 ScopedThreadJoinActivity::ScopedThreadJoinActivity( | |
| 762 const base::PlatformThreadHandle* thread) | |
| 763 : GlobalActivityTracker::ScopedThreadActivity( | |
| 764 nullptr, | |
| 765 ThreadActivityTracker::ACT_THREAD_JOIN, | |
| 766 ThreadActivityTracker::ActivityData::ForThread(*thread), | |
| 767 /*lock_allowed=*/true) {} | |
| 768 | |
| 769 #if !defined(OS_NACL) && !defined(OS_IOS) | |
| 770 ScopedProcessWaitActivity::ScopedProcessWaitActivity( | |
| 771 const base::Process* process) | |
| 772 : GlobalActivityTracker::ScopedThreadActivity( | |
| 773 nullptr, | |
| 774 ThreadActivityTracker::ACT_PROCESS_WAIT, | |
| 775 ThreadActivityTracker::ActivityData::ForProcess(process->Pid()), | |
| 776 /*lock_allowed=*/true) {} | |
| 777 #endif | |
| 778 | |
| 779 } // namespace debug | |
| 780 } // namespace base | |
| OLD | NEW |