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/feature_list.h" | |
9 #include "base/files/file.h" | |
10 #include "base/files/file_path.h" | |
11 #include "base/files/memory_mapped_file.h" | |
12 #include "base/logging.h" | |
13 #include "base/memory/ptr_util.h" | |
14 #include "base/metrics/field_trial.h" | |
15 #include "base/metrics/histogram_macros.h" | |
16 #include "base/pending_task.h" | |
17 #include "base/process/process.h" | |
18 #include "base/process/process_handle.h" | |
19 #include "base/stl_util.h" | |
20 #include "base/strings/string_util.h" | |
21 #include "base/threading/platform_thread.h" | |
22 | |
23 namespace base { | |
24 namespace debug { | |
25 | |
26 namespace { | |
27 | |
28 // A number that identifies the memory as having been initialized. It's | |
29 // arbitrary but happens to be the first 8 bytes of SHA1(ThreadActivityTracker). | |
30 // A version number is added on so that major structure changes won't try to | |
31 // read an older version (since the cookie won't match). | |
32 const uint64_t kHeaderCookie = 0xC0029B240D4A3092ULL + 1; // v1 | |
33 | |
34 // The minimum depth a stack should support. | |
35 const int kMinStackDepth = 2; | |
36 | |
37 } // namespace | |
38 | |
39 | |
40 #if !defined(OS_NACL) // NACL doesn't support any kind of file access in build. | |
41 void SetupGlobalActivityTrackerFieldTrial(const FilePath& file) { | |
42 const Feature kActivityTrackerFeature{ | |
Alexei Svitkine (slow)
2016/08/02 21:18:51
Move this outside the function, to the anon namesp
bcwhite
2016/08/04 14:06:20
Done.
| |
43 "ActivityTracking", FEATURE_DISABLED_BY_DEFAULT | |
Alexei Svitkine (slow)
2016/08/02 21:18:51
Wearing my privacy hat, this name be changed? It c
bcwhite
2016/08/04 14:06:20
Done.
| |
44 }; | |
45 | |
46 if (!base::FeatureList::IsEnabled(kActivityTrackerFeature)) | |
47 return; | |
48 | |
49 // TODO(bcwhite): Adjust these numbers once there is real data to show | |
50 // just how much of an arena is necessary. | |
51 const size_t kMemorySize = 1 << 20; // 1 MiB | |
52 const int kStackDepth = 4; | |
53 const uint64_t kAllocatorId = 0; | |
54 const char kAllocatorName[] = "ActivityTracker"; | |
55 | |
56 GlobalActivityTracker::CreateWithFile( | |
57 file.AddExtension(PersistentMemoryAllocator::kFileExtension), | |
58 kMemorySize, kAllocatorId, kAllocatorName, kStackDepth); | |
59 } | |
60 #endif // !defined(OS_NACL) | |
61 | |
62 | |
63 // This information is kept for every thread that is tracked. It is filled | |
64 // the very first time the thread is seen. All fields must be of exact sizes | |
65 // so there is no issue moving between 32 and 64-bit builds. | |
66 struct ThreadActivityTracker::Header { | |
67 // This unique number indicates a valid initialization of the memory. | |
68 uint64_t cookie; | |
69 | |
70 // The process-id and thread-id to which this data belongs. These identifiers | |
71 // are not guaranteed to mean anything but are unique, in combination, among | |
72 // all active trackers. It would be nice to always have the process_id be a | |
73 // 64-bit value but the necessity of having it atomic (for the memory barriers | |
74 // it provides) limits it to the natural word size of the machine. | |
75 #ifdef ARCH_CPU_64_BITS | |
76 std::atomic<int64_t> process_id; | |
77 #else | |
78 std::atomic<int32_t> process_id; | |
79 int32_t process_id_padding; | |
80 #endif | |
81 | |
82 union { | |
83 int64_t as_id; | |
84 #if defined(OS_WIN) | |
85 // On Windows, the handle itself is often a pseudo-handle with a common | |
86 // value meaning "this thread" and so the thread-id is used. The former | |
87 // can be converted to a thread-id with a system call. | |
88 PlatformThreadId as_tid; | |
89 #elif defined(OS_POSIX) | |
90 // On Posix, the handle is always a unique identifier so no conversion | |
91 // needs to be done. However, it's value is officially opaque so there | |
92 // is no one correct way to convert it to a numerical identifier. | |
93 PlatformThreadHandle::Handle as_handle; | |
94 #endif | |
95 } thread_ref; | |
96 | |
97 // The start-time and start-ticks when the data was created. Each activity | |
98 // record has a |time_internal| value that can be converted to a "wall time" | |
99 // with these two values. | |
100 int64_t start_time; | |
101 int64_t start_ticks; | |
102 | |
103 // The number of Activity slots in the data. | |
104 uint32_t stack_slots; | |
105 | |
106 // The current depth of the stack. This may be greater than the number of | |
107 // slots. If the depth exceeds the number of slots, the newest entries | |
108 // won't be recorded. | |
109 std::atomic<uint32_t> current_depth; | |
110 | |
111 // A memory location used to indicate if changes have been made to the stack | |
112 // that would invalidate an in-progress read of its contents. The active | |
113 // tracker will zero the value whenever something gets popped from the | |
114 // stack. A monitoring tracker can write a non-zero value here, copy the | |
115 // stack contents, and read the value to know, if it is still non-zero, that | |
116 // the contents didn't change while being copied. This can handle concurrent | |
117 // snapshot operations only if each snapshot writes a different bit (which | |
118 // is not the current implementation so no parallel snapshots allowed). | |
119 std::atomic<uint32_t> stack_unchanged; | |
120 | |
121 // The name of the thread (up to a maximum length). Dynamic-length names | |
122 // are not practical since the memory has to come from the same persistent | |
123 // allocator that holds this structure and to which this object has no | |
124 // reference. | |
125 char thread_name[32]; | |
126 }; | |
127 | |
128 // It doesn't matter what is contained in this (though it will be all zeros) | |
129 // as only the address of it is important. | |
130 const ThreadActivityTracker::ActivityData | |
131 ThreadActivityTracker::kNullActivityData = {}; | |
132 | |
133 ThreadActivityTracker::ActivityData | |
134 ThreadActivityTracker::ActivityData::ForThread( | |
135 const PlatformThreadHandle& handle) { | |
136 // Header already has a conversion union; reuse that. | |
137 ThreadActivityTracker::Header header; | |
138 header.thread_ref.as_id = 0; // Zero the union in case other is smaller. | |
139 #if defined(OS_WIN) | |
140 header.thread_ref.as_tid = ::GetThreadId(handle.platform_handle()); | |
141 #elif defined(OS_POSIX) | |
142 header.thread_ref.as_handle = handle.platform_handle(); | |
143 #endif | |
144 return ForThread(header.thread_ref.as_id); | |
145 } | |
146 | |
147 ThreadActivityTracker::ActivitySnapshot::ActivitySnapshot() {} | |
148 ThreadActivityTracker::ActivitySnapshot::~ActivitySnapshot() {} | |
149 | |
150 | |
151 ThreadActivityTracker::ThreadActivityTracker(void* base, size_t size) | |
152 : header_(static_cast<Header*>(base)), | |
153 stack_(reinterpret_cast<Activity*>(reinterpret_cast<char*>(base) + | |
154 sizeof(Header))), | |
155 stack_slots_( | |
156 static_cast<uint32_t>((size - sizeof(Header)) / sizeof(Activity))) { | |
157 DCHECK(thread_checker_.CalledOnValidThread()); | |
158 | |
159 // Verify the parameters but fail gracefully if they're not valid so that | |
160 // production code based on external inputs will not crash. IsValid() will | |
161 // return false in this case. | |
162 if (!base || | |
163 // Ensure there is enough space for the header and at least a few records. | |
164 size < sizeof(Header) + kMinStackDepth * sizeof(Activity) || | |
165 // Ensure that the |stack_slots_| calculation didn't overflow. | |
166 (size - sizeof(Header)) / sizeof(Activity) > | |
167 std::numeric_limits<uint32_t>::max()) { | |
168 NOTREACHED(); | |
169 return; | |
170 } | |
171 | |
172 // Ensure that the thread reference doesn't exceed the size of the ID number. | |
173 // This won't compile at the global scope because Header is a private struct. | |
174 static_assert( | |
175 sizeof(header_->thread_ref) == sizeof(header_->thread_ref.as_id), | |
176 "PlatformThreadHandle::Handle is too big to hold in 64-bit ID"); | |
177 | |
178 // Ensure that the alignment of Activity.data is properly aligned to a | |
179 // 64-bit boundary so there are no interoperability-issues across cpu | |
180 // architectures. | |
181 static_assert(offsetof(Activity, data) % sizeof(uint64_t) == 0, | |
182 "ActivityData.data is not 64-bit aligned"); | |
183 | |
184 // Provided memory should either be completely initialized or all zeros. | |
185 if (header_->cookie == 0) { | |
186 // This is a new file. Double-check other fields and then initialize. | |
187 DCHECK_EQ(0, header_->process_id.load(std::memory_order_relaxed)); | |
188 DCHECK_EQ(0, header_->thread_ref.as_id); | |
189 DCHECK_EQ(0, header_->start_time); | |
190 DCHECK_EQ(0, header_->start_ticks); | |
191 DCHECK_EQ(0U, header_->stack_slots); | |
192 DCHECK_EQ(0U, header_->current_depth.load(std::memory_order_relaxed)); | |
193 DCHECK_EQ(0U, header_->stack_unchanged.load(std::memory_order_relaxed)); | |
194 DCHECK_EQ(0, stack_[0].time_internal); | |
195 DCHECK_EQ(0U, stack_[0].origin_address); | |
196 DCHECK_EQ(0U, stack_[0].call_stack[0]); | |
197 DCHECK_EQ(0U, stack_[0].data.task.sequence_id); | |
198 | |
199 #if defined(OS_WIN) | |
200 header_->thread_ref.as_tid = PlatformThread::CurrentId(); | |
201 #elif defined(OS_POSIX) | |
202 header_->thread_ref.as_handle = | |
203 PlatformThread::CurrentHandle().platform_handle(); | |
204 #endif | |
205 header_->start_time = base::Time::Now().ToInternalValue(); | |
206 header_->start_ticks = base::TimeTicks::Now().ToInternalValue(); | |
207 header_->stack_slots = stack_slots_; | |
208 strlcpy(header_->thread_name, PlatformThread::GetName(), | |
209 sizeof(header_->thread_name)); | |
210 header_->cookie = kHeaderCookie; | |
211 | |
212 // This is done last so as to guarantee that everything above is "released" | |
213 // by the time this value gets written. | |
214 header_->process_id.store(GetCurrentProcId(), std::memory_order_release); | |
215 | |
216 valid_ = true; | |
217 DCHECK(IsValid()); | |
218 } else { | |
219 // This is a file with existing data. Perform basic consistency checks. | |
220 valid_ = true; | |
221 valid_ = IsValid(); | |
222 } | |
223 } | |
224 | |
225 ThreadActivityTracker::~ThreadActivityTracker() {} | |
226 | |
227 void ThreadActivityTracker::PushActivity(const void* origin, | |
228 ActivityType type, | |
229 const ActivityData& data) { | |
230 // A thread-checker creates a lock to check the thread-id which means | |
231 // re-entry into this code if lock acquisitions are being tracked. | |
232 DCHECK(type == ACT_LOCK_ACQUIRE || thread_checker_.CalledOnValidThread()); | |
233 | |
234 // Get the current depth of the stack. No access to other memory guarded | |
235 // by this variable is done here so a "relaxed" load is acceptable. | |
236 uint32_t depth = header_->current_depth.load(std::memory_order_relaxed); | |
237 | |
238 // Handle the case where the stack depth has exceeded the storage capacity. | |
239 // Extra entries will be lost leaving only the base of the stack. | |
240 if (depth >= stack_slots_) { | |
241 // Since no other threads modify the data, no compare/exchange is needed. | |
242 // Since no other memory is being modified, a "relaxed" store is acceptable. | |
243 header_->current_depth.store(depth + 1, std::memory_order_relaxed); | |
244 return; | |
245 } | |
246 | |
247 // Get a pointer to the next activity and load it. No atomicity is required | |
248 // here because the memory is known only to this thread. It will be made | |
249 // known to other threads once the depth is incremented. | |
250 Activity* activity = &stack_[depth]; | |
251 activity->time_internal = base::TimeTicks::Now().ToInternalValue(); | |
252 activity->origin_address = reinterpret_cast<uintptr_t>(origin); | |
253 activity->activity_type = type; | |
254 activity->data = data; | |
255 | |
256 #if defined(SYZYASAN) | |
257 // Create a stacktrace from the current location and get the addresses. | |
258 StackTrace stack_trace; | |
259 size_t stack_depth; | |
260 const void* const* stack_addrs = stack_trace.Addresses(&stack_depth); | |
261 // Copy the stack addresses, ignoring the first one (here). | |
262 size_t i; | |
263 for (i = 1; i < stack_depth && i < kActivityCallStackSize; ++i) { | |
264 activity->call_stack[i - 1] = reinterpret_cast<uintptr_t>(stack_addrs[i]); | |
265 } | |
266 activity->call_stack[i - 1] = 0; | |
267 #else | |
268 // Since the memory was initially zero and nothing ever overwrites it in | |
269 // this "else" case, there is no need to write even the null terminator. | |
270 //activity->call_stack[0] = 0; | |
271 #endif | |
272 | |
273 // Save the incremented depth. Because this guards |activity| memory filled | |
274 // above that may be read by another thread once the recorded depth changes, | |
275 // a "release" store is required. | |
276 header_->current_depth.store(depth + 1, std::memory_order_release); | |
277 } | |
278 | |
279 void ThreadActivityTracker::ChangeActivity(ActivityType type, | |
280 const ActivityData& data) { | |
281 DCHECK(thread_checker_.CalledOnValidThread()); | |
282 DCHECK(type != ACT_NULL || &data != &kNullActivityData); | |
283 | |
284 // Get the current depth of the stack and acquire the data held there. | |
285 uint32_t depth = header_->current_depth.load(std::memory_order_acquire); | |
286 DCHECK_LT(0U, depth); | |
287 | |
288 // Update the information if it is being recorded (i.e. within slot limit). | |
289 if (depth <= stack_slots_) { | |
290 Activity* activity = &stack_[depth - 1]; | |
291 | |
292 if (type != ACT_NULL) { | |
293 DCHECK_EQ(activity->activity_type & ACT_CATEGORY_MASK, | |
294 type & ACT_CATEGORY_MASK); | |
295 activity->activity_type = type; | |
296 } | |
297 | |
298 if (&data != &kNullActivityData) | |
299 activity->data = data; | |
300 } | |
301 } | |
302 | |
303 void ThreadActivityTracker::PopActivity() { | |
304 // Do an atomic decrement of the depth. No changes to stack entries guarded | |
305 // by this variable are done here so a "relaxed" operation is acceptable. | |
306 // |depth| will receive the value BEFORE it was modified. | |
307 uint32_t depth = | |
308 header_->current_depth.fetch_sub(1, std::memory_order_relaxed); | |
309 | |
310 // Validate that everything is running correctly. | |
311 DCHECK_LT(0U, depth); | |
312 | |
313 // A thread-checker creates a lock to check the thread-id which means | |
314 // re-entry into this code if lock acquisitions are being tracked. | |
315 DCHECK(stack_[depth - 1].activity_type == ACT_LOCK_ACQUIRE || | |
316 thread_checker_.CalledOnValidThread()); | |
317 | |
318 // The stack has shrunk meaning that some other thread trying to copy the | |
319 // contents for reporting purposes could get bad data. That thread would | |
320 // have written a non-zero value into |stack_unchanged|; clearing it here | |
321 // will let that thread detect that something did change. This needs to | |
322 // happen after the atomic |depth| operation above so a "release" store | |
323 // is required. | |
324 header_->stack_unchanged.store(0, std::memory_order_release); | |
325 } | |
326 | |
327 bool ThreadActivityTracker::IsValid() const { | |
328 if (header_->cookie != kHeaderCookie || | |
329 header_->process_id.load(std::memory_order_relaxed) == 0 || | |
330 header_->thread_ref.as_id == 0 || | |
331 header_->start_time == 0 || | |
332 header_->start_ticks == 0 || | |
333 header_->stack_slots != stack_slots_ || | |
334 header_->thread_name[sizeof(header_->thread_name) - 1] != '\0') { | |
335 return false; | |
336 } | |
337 | |
338 return valid_; | |
339 } | |
340 | |
341 bool ThreadActivityTracker::Snapshot(ActivitySnapshot* output_snapshot) const { | |
342 DCHECK(output_snapshot); | |
343 | |
344 // There is no "called on valid thread" check for this method as it can be | |
345 // called from other threads or even other processes. It is also the reason | |
346 // why atomic operations must be used in certain places above. | |
347 | |
348 // It's possible for the data to change while reading it in such a way that it | |
349 // invalidates the read. Make several attempts but don't try forever. | |
350 const int kMaxAttempts = 10; | |
351 uint32_t depth; | |
352 | |
353 // Stop here if the data isn't valid. | |
354 if (!IsValid()) | |
355 return false; | |
356 | |
357 // Allocate the maximum size for the stack so it doesn't have to be done | |
358 // during the time-sensitive snapshot operation. It is shrunk once the | |
359 // actual size is known. | |
360 output_snapshot->activity_stack.reserve(stack_slots_); | |
361 | |
362 for (int attempt = 0; attempt < kMaxAttempts; ++attempt) { | |
363 // Remember the process and thread IDs to ensure they aren't replaced | |
364 // during the snapshot operation. Use "acquire" to ensure that all the | |
365 // non-atomic fields of the structure are valid (at least at the current | |
366 // moment in time). | |
367 const int64_t starting_process_id = | |
368 header_->process_id.load(std::memory_order_acquire); | |
369 const int64_t starting_thread_id = header_->thread_ref.as_id; | |
370 | |
371 // Write a non-zero value to |stack_unchanged| so it's possible to detect | |
372 // at the end that nothing has changed since copying the data began. A | |
373 // "cst" operation is required to ensure it occurs before everything else. | |
374 // Using "cst" memory ordering is relatively expensive but this is only | |
375 // done during analysis so doesn't directly affect the worker threads. | |
376 header_->stack_unchanged.store(1, std::memory_order_seq_cst); | |
377 | |
378 // Fetching the current depth also "acquires" the contents of the stack. | |
379 depth = header_->current_depth.load(std::memory_order_acquire); | |
380 uint32_t count = std::min(depth, stack_slots_); | |
381 output_snapshot->activity_stack.resize(count); | |
382 if (count > 0) { | |
383 // Copy the existing contents. Memcpy is used for speed. | |
384 memcpy(&output_snapshot->activity_stack[0], stack_, | |
385 count * sizeof(Activity)); | |
386 } | |
387 | |
388 // Retry if something changed during the copy. A "cst" operation ensures | |
389 // it must happen after all the above operations. | |
390 if (!header_->stack_unchanged.load(std::memory_order_seq_cst)) | |
391 continue; | |
392 | |
393 // Stack copied. Record it's full depth. | |
394 output_snapshot->activity_stack_depth = depth; | |
395 | |
396 // TODO(bcwhite): Snapshot other things here. | |
397 | |
398 // Get the general thread information. Loading of "process_id" is guaranteed | |
399 // to be last so that it's possible to detect below if any content has | |
400 // changed while reading it. It's technically possible for a thread to end, | |
401 // have its data cleared, a new thread get created with the same IDs, and | |
402 // it perform an action which starts tracking all in the time since the | |
403 // ID reads above but the chance is so unlikely that it's not worth the | |
404 // effort and complexity of protecting against it (perhaps with an | |
405 // "unchanged" field like is done for the stack). | |
406 output_snapshot->thread_name = | |
407 std::string(header_->thread_name, sizeof(header_->thread_name) - 1); | |
408 output_snapshot->thread_id = header_->thread_ref.as_id; | |
409 output_snapshot->process_id = | |
410 header_->process_id.load(std::memory_order_seq_cst); | |
411 | |
412 // All characters of the thread-name buffer were copied so as to not break | |
413 // if the trailing NUL were missing. Now limit the length if the actual | |
414 // name is shorter. | |
415 output_snapshot->thread_name.resize( | |
416 strlen(output_snapshot->thread_name.c_str())); | |
417 | |
418 // If the process or thread ID has changed then the tracker has exited and | |
419 // the memory reused by a new one. Try again. | |
420 if (output_snapshot->process_id != starting_process_id || | |
421 output_snapshot->thread_id != starting_thread_id) { | |
422 continue; | |
423 } | |
424 | |
425 // Only successful if the data is still valid once everything is done since | |
426 // it's possible for the thread to end somewhere in the middle and all its | |
427 // values become garbage. | |
428 if (!IsValid()) | |
429 return false; | |
430 | |
431 // Change all the timestamps in the activities from "ticks" to "wall" time. | |
432 const Time start_time = Time::FromInternalValue(header_->start_time); | |
433 const int64_t start_ticks = header_->start_ticks; | |
434 for (Activity& activity : output_snapshot->activity_stack) { | |
435 activity.time_internal = | |
436 (start_time + | |
437 TimeDelta::FromInternalValue(activity.time_internal - start_ticks)) | |
438 .ToInternalValue(); | |
439 } | |
440 | |
441 // Success! | |
442 return true; | |
443 } | |
444 | |
445 // Too many attempts. | |
446 return false; | |
447 } | |
448 | |
449 // static | |
450 size_t ThreadActivityTracker::SizeForStackDepth(int stack_depth) { | |
451 return static_cast<size_t>(stack_depth) * sizeof(Activity) + sizeof(Header); | |
452 } | |
453 | |
454 | |
455 GlobalActivityTracker* GlobalActivityTracker::g_tracker_ = nullptr; | |
456 | |
457 GlobalActivityTracker::ManagedActivityTracker::ManagedActivityTracker( | |
458 PersistentMemoryAllocator::Reference mem_reference, | |
459 void* base, | |
460 size_t size) | |
461 : ThreadActivityTracker(base, size), | |
462 mem_reference_(mem_reference), | |
463 mem_base_(base) {} | |
464 | |
465 GlobalActivityTracker::ManagedActivityTracker::~ManagedActivityTracker() { | |
466 // The global |g_tracker_| must point to the owner of this class since all | |
467 // objects of this type must be destructed before |g_tracker_| can be changed | |
468 // (something that only occurs in tests). | |
469 DCHECK(g_tracker_); | |
470 g_tracker_->ReturnTrackerMemory(this); | |
471 } | |
472 | |
473 void GlobalActivityTracker::CreateWithAllocator( | |
474 std::unique_ptr<PersistentMemoryAllocator> allocator, | |
475 int stack_depth) { | |
476 // There's no need to do anything with the result. It is self-managing. | |
477 GlobalActivityTracker* global_tracker = | |
478 new GlobalActivityTracker(std::move(allocator), stack_depth); | |
479 // Create a tracker for this thread since it is known. | |
480 global_tracker->CreateTrackerForCurrentThread(); | |
481 } | |
482 | |
483 #if !defined(OS_NACL) | |
484 // static | |
485 void GlobalActivityTracker::CreateWithFile(const FilePath& file_path, | |
486 size_t size, | |
487 uint64_t id, | |
488 StringPiece name, | |
489 int stack_depth) { | |
490 DCHECK(!file_path.empty()); | |
491 DCHECK_GE(static_cast<uint64_t>(std::numeric_limits<int64_t>::max()), size); | |
492 | |
493 // Create and map the file into memory and make it globally available. | |
494 std::unique_ptr<MemoryMappedFile> mapped_file(new MemoryMappedFile()); | |
495 bool success = | |
496 mapped_file->Initialize(File(file_path, | |
497 File::FLAG_CREATE_ALWAYS | File::FLAG_READ | | |
498 File::FLAG_WRITE | File::FLAG_SHARE_DELETE), | |
499 {0, static_cast<int64_t>(size)}, | |
500 MemoryMappedFile::READ_WRITE_EXTEND); | |
501 DCHECK(success); | |
502 CreateWithAllocator(WrapUnique(new FilePersistentMemoryAllocator( | |
503 std::move(mapped_file), size, id, name, false)), | |
504 stack_depth); | |
505 } | |
506 #endif // !defined(OS_NACL) | |
507 | |
508 // static | |
509 void GlobalActivityTracker::CreateWithLocalMemory(size_t size, | |
510 uint64_t id, | |
511 StringPiece name, | |
512 int stack_depth) { | |
513 CreateWithAllocator( | |
514 WrapUnique(new LocalPersistentMemoryAllocator(size, id, name)), | |
515 stack_depth); | |
516 } | |
517 | |
518 ThreadActivityTracker* GlobalActivityTracker::CreateTrackerForCurrentThread() { | |
519 DCHECK(!this_thread_tracker_.Get()); | |
520 | |
521 PersistentMemoryAllocator::Reference mem_reference = 0; | |
522 void* mem_base = nullptr; | |
523 | |
524 // Get the current count of available memories, acquiring the array values. | |
525 int count = available_memories_count_.load(std::memory_order_acquire); | |
526 while (count > 0) { | |
527 // There is a memory block that was previously released (and zeroed) so | |
528 // just re-use that rather than allocating a new one. Use "relaxed" because | |
529 // the value is guarded by the |count| "acquire". A zero reference replaces | |
530 // the existing value so that it can't be used by another thread that | |
531 // manages to interrupt this one before the count can be decremented. | |
532 // A zero reference is also required for the "push" operation to work | |
533 // once the count finally does get decremented. | |
534 mem_reference = | |
535 available_memories_[count - 1].exchange(0, std::memory_order_relaxed); | |
536 | |
537 // If the reference is zero, it's already been taken but count hasn't yet | |
538 // been decremented. Give that other thread a chance to finish then reload | |
539 // the "count" value and try again. | |
540 if (!mem_reference) { | |
541 PlatformThread::YieldCurrentThread(); | |
542 count = available_memories_count_.load(std::memory_order_acquire); | |
543 continue; | |
544 } | |
545 | |
546 // Decrement the count indicating that the value has been taken. If this | |
547 // fails then another thread has pushed something new and incremented the | |
548 // count. | |
549 // NOTE: |oldcount| will be loaded with the existing value. | |
550 int oldcount = count; | |
551 if (!available_memories_count_.compare_exchange_strong( | |
552 oldcount, count - 1, std::memory_order_acquire, | |
553 std::memory_order_acquire)) { | |
554 DCHECK_LT(count, oldcount); | |
555 | |
556 // Restore the reference that was zeroed above and try again. | |
557 available_memories_[count - 1].store(mem_reference, | |
558 std::memory_order_relaxed); | |
559 count = oldcount; | |
560 continue; | |
561 } | |
562 | |
563 // Turn the reference back into one of the activity-tracker type. | |
564 mem_base = allocator_->GetAsObject<char>(mem_reference, | |
565 kTypeIdActivityTrackerFree); | |
566 DCHECK(mem_base); | |
567 DCHECK_LE(stack_memory_size_, allocator_->GetAllocSize(mem_reference)); | |
568 bool changed = allocator_->ChangeType(mem_reference, kTypeIdActivityTracker, | |
569 kTypeIdActivityTrackerFree); | |
570 DCHECK(changed); | |
571 | |
572 // Success. | |
573 break; | |
574 } | |
575 | |
576 // Handle the case where no previously-used memories are available. | |
577 if (count == 0) { | |
578 // Allocate a block of memory from the persistent segment. | |
579 mem_reference = | |
580 allocator_->Allocate(stack_memory_size_, kTypeIdActivityTracker); | |
581 if (mem_reference) { | |
582 // Success. Convert the reference to an actual memory address. | |
583 mem_base = | |
584 allocator_->GetAsObject<char>(mem_reference, kTypeIdActivityTracker); | |
585 // Make the allocation iterable so it can be found by other processes. | |
586 allocator_->MakeIterable(mem_reference); | |
587 } else { | |
588 // Failure. This shouldn't happen. | |
589 NOTREACHED(); | |
590 // But if it does, probably because the allocator wasn't given enough | |
591 // memory to satisfy all possible requests, handle it gracefully by | |
592 // allocating the required memory from the heap. | |
593 mem_base = new char[stack_memory_size_]; | |
594 memset(mem_base, 0, stack_memory_size_); | |
595 // Report the thread-count at which the allocator was full so that the | |
596 // failure can be seen and underlying memory resized appropriately. | |
597 UMA_HISTOGRAM_COUNTS_1000( | |
598 "UMA.ActivityTracker.ThreadTrackers.MemLimitTrackerCount", | |
599 thread_tracker_count_.load(std::memory_order_relaxed)); | |
600 } | |
601 } | |
602 | |
603 // Create a tracker with the acquired memory and set it as the tracker | |
604 // for this particular thread in thread-local-storage. | |
605 DCHECK(mem_base); | |
606 ManagedActivityTracker* tracker = | |
607 new ManagedActivityTracker(mem_reference, mem_base, stack_memory_size_); | |
608 DCHECK(tracker->IsValid()); | |
609 this_thread_tracker_.Set(tracker); | |
610 int old_count = thread_tracker_count_.fetch_add(1, std::memory_order_relaxed); | |
611 | |
612 UMA_HISTOGRAM_ENUMERATION("UMA.ActivityTracker.ThreadTrackers.Count", | |
613 old_count + 1, kMaxThreadCount); | |
614 return tracker; | |
615 } | |
616 | |
617 void GlobalActivityTracker::ReleaseTrackerForCurrentThreadForTesting() { | |
618 ThreadActivityTracker* tracker = | |
619 reinterpret_cast<ThreadActivityTracker*>(this_thread_tracker_.Get()); | |
620 if (tracker) { | |
621 this_thread_tracker_.Free(); | |
622 delete tracker; | |
623 } | |
624 } | |
625 | |
626 GlobalActivityTracker::GlobalActivityTracker( | |
627 std::unique_ptr<PersistentMemoryAllocator> allocator, | |
628 int stack_depth) | |
629 : allocator_(std::move(allocator)), | |
630 stack_memory_size_(ThreadActivityTracker::SizeForStackDepth(stack_depth)), | |
631 this_thread_tracker_(&OnTLSDestroy), | |
632 thread_tracker_count_(0), | |
633 available_memories_count_(0) { | |
634 // Clear the available-memories array. | |
635 memset(available_memories_, 0, sizeof(available_memories_)); | |
636 | |
637 // Ensure the passed memory is valid and empty (iterator finds nothing). | |
638 uint32_t type; | |
639 DCHECK(!PersistentMemoryAllocator::Iterator(allocator_.get()).GetNext(&type)); | |
640 | |
641 // Ensure that there is no other global object and then make this one such. | |
642 DCHECK(!g_tracker_); | |
643 g_tracker_ = this; | |
644 } | |
645 | |
646 GlobalActivityTracker::~GlobalActivityTracker() { | |
647 DCHECK_EQ(g_tracker_, this); | |
648 DCHECK_EQ(0, thread_tracker_count_.load(std::memory_order_relaxed)); | |
649 g_tracker_ = nullptr; | |
650 } | |
651 | |
652 void GlobalActivityTracker::ReturnTrackerMemory( | |
653 ManagedActivityTracker* tracker) { | |
654 PersistentMemoryAllocator::Reference mem_reference = tracker->mem_reference_; | |
655 void* mem_base = tracker->mem_base_; | |
656 | |
657 // Zero the memory so that it is ready for use if needed again later. It's | |
658 // better to clear the memory now, when a thread is exiting, than to do it | |
659 // when it is first needed by a thread doing actual work. | |
660 memset(mem_base, 0, stack_memory_size_); | |
661 | |
662 // Remove the destructed tracker from the set of known ones. | |
663 DCHECK_LE(1, thread_tracker_count_.load(std::memory_order_relaxed)); | |
664 thread_tracker_count_.fetch_sub(1, std::memory_order_relaxed); | |
665 | |
666 // Deal with the memory that was used by the tracker. | |
667 if (mem_reference) { | |
668 // The memory was within the persistent memory allocator. Change its type | |
669 // so that iteration won't find it. | |
670 allocator_->ChangeType(mem_reference, kTypeIdActivityTrackerFree, | |
671 kTypeIdActivityTracker); | |
672 // There is no way to free memory from a persistent allocator so instead | |
673 // push it on the internal list of available memory blocks. | |
674 while (true) { | |
675 // Get the existing count of available memories and ensure we won't | |
676 // burst the array. Acquire the values in the array. | |
677 int count = available_memories_count_.load(std::memory_order_acquire); | |
678 if (count >= kMaxThreadCount) { | |
679 NOTREACHED(); | |
680 // Storage is full. Just forget about this memory. It won't be re-used | |
681 // but there's no real loss. | |
682 break; | |
683 } | |
684 | |
685 // Write the reference of the memory being returned to this slot in the | |
686 // array. Empty slots have a value of zero so do an atomic compare-and- | |
687 // exchange to ensure that a race condition doesn't exist with another | |
688 // thread doing the same. | |
689 PersistentMemoryAllocator::Reference mem_expected = 0; | |
690 if (!available_memories_[count].compare_exchange_strong( | |
691 mem_expected, mem_reference, std::memory_order_release, | |
692 std::memory_order_relaxed)) { | |
693 PlatformThread::YieldCurrentThread(); | |
694 continue; // Try again. | |
695 } | |
696 | |
697 // Increment the count, releasing the value written to the array. This | |
698 // could fail if a simultaneous "pop" operation decremented the counter. | |
699 // If that happens, clear the array slot and start over. Do a "strong" | |
700 // exchange to avoid spurious retries that can occur with a "weak" one. | |
701 int expected = count; // Updated by compare/exchange. | |
702 if (!available_memories_count_.compare_exchange_strong( | |
703 expected, count + 1, std::memory_order_release, | |
704 std::memory_order_relaxed)) { | |
705 available_memories_[count].store(0, std::memory_order_relaxed); | |
706 continue; | |
707 } | |
708 | |
709 // Count was successfully incremented to reflect the newly added value. | |
710 break; | |
711 } | |
712 } else { | |
713 // The memory was allocated from the process heap. This shouldn't happen | |
714 // because the persistent memory segment should be big enough for all | |
715 // thread stacks but it's better to support falling back to allocation | |
716 // from the heap rather than crash. Everything will work as normal but | |
717 // the data won't be persisted. | |
718 delete[] reinterpret_cast<char*>(mem_base); | |
719 } | |
720 } | |
721 | |
722 // static | |
723 void GlobalActivityTracker::OnTLSDestroy(void* value) { | |
724 delete reinterpret_cast<ManagedActivityTracker*>(value); | |
725 } | |
726 | |
727 | |
728 ScopedActivity::ScopedActivity(const tracked_objects::Location& location, | |
729 uint8_t action, | |
730 uint32_t id, | |
731 int32_t info) | |
732 : GlobalActivityTracker::ScopedThreadActivity( | |
733 location.program_counter(), | |
734 static_cast<ThreadActivityTracker::ActivityType>( | |
735 ThreadActivityTracker::ACT_GENERIC | action), | |
736 ThreadActivityTracker::ActivityData::ForGeneric(id, info), | |
737 /*lock_allowed=*/true), | |
738 id_(id) { | |
739 // The action must not affect the category bits of the activity type. | |
740 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
741 } | |
742 | |
743 void ScopedActivity::ChangeAction(uint8_t action) { | |
744 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
745 ChangeTypeAndData(static_cast<ThreadActivityTracker::ActivityType>( | |
746 ThreadActivityTracker::ACT_GENERIC | action), | |
747 ThreadActivityTracker::kNullActivityData); | |
748 } | |
749 | |
750 void ScopedActivity::ChangeInfo(int32_t info) { | |
751 ChangeTypeAndData(ThreadActivityTracker::ACT_NULL, | |
752 ThreadActivityTracker::ActivityData::ForGeneric(id_, info)); | |
753 } | |
754 | |
755 void ScopedActivity::ChangeActionAndInfo(uint8_t action, int32_t info) { | |
756 DCHECK_EQ(0, action & ThreadActivityTracker::ACT_CATEGORY_MASK); | |
757 ChangeTypeAndData(static_cast<ThreadActivityTracker::ActivityType>( | |
758 ThreadActivityTracker::ACT_GENERIC | action), | |
759 ThreadActivityTracker::ActivityData::ForGeneric(id_, info)); | |
760 } | |
761 | |
762 ScopedTaskRunActivity::ScopedTaskRunActivity(const base::PendingTask& task) | |
763 : GlobalActivityTracker::ScopedThreadActivity( | |
764 task.posted_from.program_counter(), | |
765 ThreadActivityTracker::ACT_TASK_RUN, | |
766 ThreadActivityTracker::ActivityData::ForTask(task.sequence_num), | |
767 /*lock_allowed=*/true) {} | |
768 | |
769 ScopedLockAcquireActivity::ScopedLockAcquireActivity( | |
770 const base::internal::LockImpl* lock) | |
771 : GlobalActivityTracker::ScopedThreadActivity( | |
772 nullptr, | |
773 ThreadActivityTracker::ACT_LOCK_ACQUIRE, | |
774 ThreadActivityTracker::ActivityData::ForLock(lock), | |
775 /*lock_allowed=*/false) {} | |
776 | |
777 ScopedEventWaitActivity::ScopedEventWaitActivity( | |
778 const base::WaitableEvent* event) | |
779 : GlobalActivityTracker::ScopedThreadActivity( | |
780 nullptr, | |
781 ThreadActivityTracker::ACT_EVENT_WAIT, | |
782 ThreadActivityTracker::ActivityData::ForEvent(event), | |
783 /*lock_allowed=*/true) {} | |
784 | |
785 ScopedThreadJoinActivity::ScopedThreadJoinActivity( | |
786 const base::PlatformThreadHandle* thread) | |
787 : GlobalActivityTracker::ScopedThreadActivity( | |
788 nullptr, | |
789 ThreadActivityTracker::ACT_THREAD_JOIN, | |
790 ThreadActivityTracker::ActivityData::ForThread(*thread), | |
791 /*lock_allowed=*/true) {} | |
792 | |
793 #if !defined(OS_NACL) && !defined(OS_IOS) | |
794 ScopedProcessWaitActivity::ScopedProcessWaitActivity( | |
795 const base::Process* process) | |
796 : GlobalActivityTracker::ScopedThreadActivity( | |
797 nullptr, | |
798 ThreadActivityTracker::ACT_PROCESS_WAIT, | |
799 ThreadActivityTracker::ActivityData::ForProcess(process->Pid()), | |
800 /*lock_allowed=*/true) {} | |
801 #endif | |
802 | |
803 } // namespace debug | |
804 } // namespace base | |
OLD | NEW |