| 1 | //===-- xray_fdr_logging.cpp -----------------------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file is a part of XRay, a dynamic runtime instrumentation system. |
| 10 | // |
| 11 | // Here we implement the Flight Data Recorder mode for XRay, where we use |
| 12 | // compact structures to store records in memory as well as when writing out the |
| 13 | // data to files. |
| 14 | // |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | #include "xray_fdr_logging.h" |
| 17 | #include <cassert> |
| 18 | #include <cstddef> |
| 19 | #include <errno.h> |
| 20 | #include <limits> |
| 21 | #include <memory> |
| 22 | #include <pthread.h> |
| 23 | #include <sys/time.h> |
| 24 | #include <time.h> |
| 25 | #include <unistd.h> |
| 26 | |
| 27 | #include "sanitizer_common/sanitizer_allocator_internal.h" |
| 28 | #include "sanitizer_common/sanitizer_atomic.h" |
| 29 | #include "sanitizer_common/sanitizer_common.h" |
| 30 | #include "xray/xray_interface.h" |
| 31 | #include "xray/xray_records.h" |
| 32 | #include "xray_allocator.h" |
| 33 | #include "xray_buffer_queue.h" |
| 34 | #include "xray_defs.h" |
| 35 | #include "xray_fdr_controller.h" |
| 36 | #include "xray_fdr_flags.h" |
| 37 | #include "xray_fdr_log_writer.h" |
| 38 | #include "xray_flags.h" |
| 39 | #include "xray_recursion_guard.h" |
| 40 | #include "xray_tsc.h" |
| 41 | #include "xray_utils.h" |
| 42 | |
| 43 | namespace __xray { |
| 44 | |
| 45 | static atomic_sint32_t LoggingStatus = { |
| 46 | .val_dont_use: XRayLogInitStatus::XRAY_LOG_UNINITIALIZED}; |
| 47 | |
| 48 | namespace { |
| 49 | |
| 50 | // Group together thread-local-data in a struct, then hide it behind a function |
| 51 | // call so that it can be initialized on first use instead of as a global. We |
| 52 | // force the alignment to 64-bytes for x86 cache line alignment, as this |
| 53 | // structure is used in the hot path of implementation. |
| 54 | struct XRAY_TLS_ALIGNAS(64) ThreadLocalData { |
| 55 | BufferQueue::Buffer Buffer{}; |
| 56 | BufferQueue *BQ = nullptr; |
| 57 | |
| 58 | using LogWriterStorage = std::byte[sizeof(FDRLogWriter)]; |
| 59 | alignas(FDRLogWriter) LogWriterStorage LWStorage; |
| 60 | FDRLogWriter *Writer = nullptr; |
| 61 | |
| 62 | using ControllerStorage = std::byte[sizeof(FDRController<>)]; |
| 63 | alignas(FDRController<>) ControllerStorage CStorage; |
| 64 | FDRController<> *Controller = nullptr; |
| 65 | }; |
| 66 | |
| 67 | } // namespace |
| 68 | |
| 69 | static_assert(std::is_trivially_destructible<ThreadLocalData>::value, |
| 70 | "ThreadLocalData must be trivially destructible" ); |
| 71 | |
| 72 | // Use a global pthread key to identify thread-local data for logging. |
| 73 | static pthread_key_t Key; |
| 74 | |
| 75 | // Global BufferQueue. |
| 76 | static std::byte BufferQueueStorage[sizeof(BufferQueue)]; |
| 77 | static BufferQueue *BQ = nullptr; |
| 78 | |
| 79 | // Global thresholds for function durations. |
| 80 | static atomic_uint64_t ThresholdTicks{.val_dont_use: 0}; |
| 81 | |
| 82 | // Global for ticks per second. |
| 83 | static atomic_uint64_t TicksPerSec{.val_dont_use: 0}; |
| 84 | |
| 85 | static atomic_sint32_t LogFlushStatus = { |
| 86 | .val_dont_use: XRayLogFlushStatus::XRAY_LOG_NOT_FLUSHING}; |
| 87 | |
| 88 | // This function will initialize the thread-local data structure used by the FDR |
| 89 | // logging implementation and return a reference to it. The implementation |
| 90 | // details require a bit of care to maintain. |
| 91 | // |
| 92 | // First, some requirements on the implementation in general: |
| 93 | // |
| 94 | // - XRay handlers should not call any memory allocation routines that may |
| 95 | // delegate to an instrumented implementation. This means functions like |
| 96 | // malloc() and free() should not be called while instrumenting. |
| 97 | // |
| 98 | // - We would like to use some thread-local data initialized on first-use of |
| 99 | // the XRay instrumentation. These allow us to implement unsynchronized |
| 100 | // routines that access resources associated with the thread. |
| 101 | // |
| 102 | // The implementation here uses a few mechanisms that allow us to provide both |
| 103 | // the requirements listed above. We do this by: |
| 104 | // |
| 105 | // 1. Using a thread-local aligned storage buffer for representing the |
| 106 | // ThreadLocalData struct. This data will be uninitialized memory by |
| 107 | // design. |
| 108 | // |
| 109 | // 2. Not requiring a thread exit handler/implementation, keeping the |
| 110 | // thread-local as purely a collection of references/data that do not |
| 111 | // require cleanup. |
| 112 | // |
| 113 | // We're doing this to avoid using a `thread_local` object that has a |
| 114 | // non-trivial destructor, because the C++ runtime might call std::malloc(...) |
| 115 | // to register calls to destructors. Deadlocks may arise when, for example, an |
| 116 | // externally provided malloc implementation is XRay instrumented, and |
| 117 | // initializing the thread-locals involves calling into malloc. A malloc |
| 118 | // implementation that does global synchronization might be holding a lock for a |
| 119 | // critical section, calling a function that might be XRay instrumented (and |
| 120 | // thus in turn calling into malloc by virtue of registration of the |
| 121 | // thread_local's destructor). |
| 122 | #if XRAY_HAS_TLS_ALIGNAS |
| 123 | static_assert(alignof(ThreadLocalData) >= 64, |
| 124 | "ThreadLocalData must be cache line aligned." ); |
| 125 | #endif |
| 126 | static ThreadLocalData &getThreadLocalData() { |
| 127 | alignas(ThreadLocalData) thread_local std::byte |
| 128 | TLDStorage[sizeof(ThreadLocalData)]; |
| 129 | |
| 130 | if (pthread_getspecific(key: Key) == NULL) { |
| 131 | new (reinterpret_cast<ThreadLocalData *>(&TLDStorage)) ThreadLocalData{}; |
| 132 | pthread_setspecific(key: Key, pointer: &TLDStorage); |
| 133 | } |
| 134 | |
| 135 | return *reinterpret_cast<ThreadLocalData *>(&TLDStorage); |
| 136 | } |
| 137 | |
| 138 | static XRayFileHeader &() { |
| 139 | alignas(XRayFileHeader) static std::byte HStorage[sizeof(XRayFileHeader)]; |
| 140 | static pthread_once_t OnceInit = PTHREAD_ONCE_INIT; |
| 141 | static bool TSCSupported = true; |
| 142 | static uint64_t CycleFrequency = NanosecondsPerSecond; |
| 143 | pthread_once( |
| 144 | once_control: &OnceInit, init_routine: +[] { |
| 145 | XRayFileHeader &H = reinterpret_cast<XRayFileHeader &>(HStorage); |
| 146 | // Version 2 of the log writes the extents of the buffer, instead of |
| 147 | // relying on an end-of-buffer record. |
| 148 | // Version 3 includes PID metadata record. |
| 149 | // Version 4 includes CPU data in the custom event records. |
| 150 | // Version 5 uses relative deltas for custom and typed event records, |
| 151 | // and removes the CPU data in custom event records (similar to how |
| 152 | // function records use deltas instead of full TSCs and rely on other |
| 153 | // metadata records for TSC wraparound and CPU migration). |
| 154 | H.Version = 5; |
| 155 | H.Type = FileTypes::FDR_LOG; |
| 156 | |
| 157 | // Test for required CPU features and cache the cycle frequency |
| 158 | TSCSupported = probeRequiredCPUFeatures(); |
| 159 | if (TSCSupported) |
| 160 | CycleFrequency = getTSCFrequency(); |
| 161 | H.CycleFrequency = CycleFrequency; |
| 162 | |
| 163 | // FIXME: Actually check whether we have 'constant_tsc' and |
| 164 | // 'nonstop_tsc' before setting the values in the header. |
| 165 | H.ConstantTSC = 1; |
| 166 | H.NonstopTSC = 1; |
| 167 | }); |
| 168 | return reinterpret_cast<XRayFileHeader &>(HStorage); |
| 169 | } |
| 170 | |
| 171 | // This is the iterator implementation, which knows how to handle FDR-mode |
| 172 | // specific buffers. This is used as an implementation of the iterator function |
| 173 | // needed by __xray_set_buffer_iterator(...). It maintains a global state of the |
| 174 | // buffer iteration for the currently installed FDR mode buffers. In particular: |
| 175 | // |
| 176 | // - If the argument represents the initial state of XRayBuffer ({nullptr, 0}) |
| 177 | // then the iterator returns the header information. |
| 178 | // - If the argument represents the header information ({address of header |
| 179 | // info, size of the header info}) then it returns the first FDR buffer's |
| 180 | // address and extents. |
| 181 | // - It will keep returning the next buffer and extents as there are more |
| 182 | // buffers to process. When the input represents the last buffer, it will |
| 183 | // return the initial state to signal completion ({nullptr, 0}). |
| 184 | // |
| 185 | // See xray/xray_log_interface.h for more details on the requirements for the |
| 186 | // implementations of __xray_set_buffer_iterator(...) and |
| 187 | // __xray_log_process_buffers(...). |
| 188 | XRayBuffer fdrIterator(const XRayBuffer B) { |
| 189 | DCHECK(internal_strcmp(__xray_log_get_current_mode(), "xray-fdr" ) == 0); |
| 190 | DCHECK(BQ->finalizing()); |
| 191 | |
| 192 | if (BQ == nullptr || !BQ->finalizing()) { |
| 193 | if (Verbosity()) |
| 194 | Report( |
| 195 | format: "XRay FDR: Failed global buffer queue is null or not finalizing!\n" ); |
| 196 | return {.Data: nullptr, .Size: 0}; |
| 197 | } |
| 198 | |
| 199 | // We use a global scratch-pad for the header information, which only gets |
| 200 | // initialized the first time this function is called. We'll update one part |
| 201 | // of this information with some relevant data (in particular the number of |
| 202 | // buffers to expect). |
| 203 | alignas( |
| 204 | XRayFileHeader) static std::byte [sizeof(XRayFileHeader)]; |
| 205 | static pthread_once_t = PTHREAD_ONCE_INIT; |
| 206 | pthread_once( |
| 207 | once_control: &HeaderOnce, init_routine: +[] { |
| 208 | reinterpret_cast<XRayFileHeader &>(HeaderStorage) = |
| 209 | fdrCommonHeaderInfo(); |
| 210 | }); |
| 211 | |
| 212 | // We use a convenience alias for code referring to Header from here on out. |
| 213 | auto & = reinterpret_cast<XRayFileHeader &>(HeaderStorage); |
| 214 | if (B.Data == nullptr && B.Size == 0) { |
| 215 | Header.FdrData = FdrAdditionalHeaderData{.ThreadBufferSize: BQ->ConfiguredBufferSize()}; |
| 216 | return XRayBuffer{.Data: static_cast<void *>(&Header), .Size: sizeof(Header)}; |
| 217 | } |
| 218 | |
| 219 | static BufferQueue::const_iterator It{}; |
| 220 | static BufferQueue::const_iterator End{}; |
| 221 | static uint8_t *CurrentBuffer{nullptr}; |
| 222 | static size_t SerializedBufferSize = 0; |
| 223 | if (B.Data == static_cast<void *>(&Header) && B.Size == sizeof(Header)) { |
| 224 | // From this point on, we provide raw access to the raw buffer we're getting |
| 225 | // from the BufferQueue. We're relying on the iterators from the current |
| 226 | // Buffer queue. |
| 227 | It = BQ->cbegin(); |
| 228 | End = BQ->cend(); |
| 229 | } |
| 230 | |
| 231 | if (CurrentBuffer != nullptr) { |
| 232 | deallocateBuffer(B: CurrentBuffer, S: SerializedBufferSize); |
| 233 | CurrentBuffer = nullptr; |
| 234 | } |
| 235 | |
| 236 | if (It == End) |
| 237 | return {.Data: nullptr, .Size: 0}; |
| 238 | |
| 239 | // Set up the current buffer to contain the extents like we would when writing |
| 240 | // out to disk. The difference here would be that we still write "empty" |
| 241 | // buffers, or at least go through the iterators faithfully to let the |
| 242 | // handlers see the empty buffers in the queue. |
| 243 | // |
| 244 | // We need this atomic fence here to ensure that writes happening to the |
| 245 | // buffer have been committed before we load the extents atomically. Because |
| 246 | // the buffer is not explicitly synchronised across threads, we rely on the |
| 247 | // fence ordering to ensure that writes we expect to have been completed |
| 248 | // before the fence are fully committed before we read the extents. |
| 249 | atomic_thread_fence(mo: memory_order_acquire); |
| 250 | auto BufferSize = atomic_load(a: It->Extents, mo: memory_order_acquire); |
| 251 | SerializedBufferSize = BufferSize + sizeof(MetadataRecord); |
| 252 | CurrentBuffer = allocateBuffer(S: SerializedBufferSize); |
| 253 | if (CurrentBuffer == nullptr) |
| 254 | return {.Data: nullptr, .Size: 0}; |
| 255 | |
| 256 | // Write out the extents as a Metadata Record into the CurrentBuffer. |
| 257 | MetadataRecord ExtentsRecord; |
| 258 | ExtentsRecord.Type = uint8_t(RecordType::Metadata); |
| 259 | ExtentsRecord.RecordKind = |
| 260 | uint8_t(MetadataRecord::RecordKinds::BufferExtents); |
| 261 | internal_memcpy(dest: ExtentsRecord.Data, src: &BufferSize, n: sizeof(BufferSize)); |
| 262 | auto AfterExtents = |
| 263 | static_cast<char *>(internal_memcpy(dest: CurrentBuffer, src: &ExtentsRecord, |
| 264 | n: sizeof(MetadataRecord))) + |
| 265 | sizeof(MetadataRecord); |
| 266 | internal_memcpy(dest: AfterExtents, src: It->Data, n: BufferSize); |
| 267 | |
| 268 | XRayBuffer Result; |
| 269 | Result.Data = CurrentBuffer; |
| 270 | Result.Size = SerializedBufferSize; |
| 271 | ++It; |
| 272 | return Result; |
| 273 | } |
| 274 | |
| 275 | // Must finalize before flushing. |
| 276 | XRayLogFlushStatus fdrLoggingFlush() XRAY_NEVER_INSTRUMENT { |
| 277 | if (atomic_load(a: &LoggingStatus, mo: memory_order_acquire) != |
| 278 | XRayLogInitStatus::XRAY_LOG_FINALIZED) { |
| 279 | if (Verbosity()) |
| 280 | Report(format: "Not flushing log, implementation is not finalized.\n" ); |
| 281 | return XRayLogFlushStatus::XRAY_LOG_NOT_FLUSHING; |
| 282 | } |
| 283 | |
| 284 | if (atomic_exchange(a: &LogFlushStatus, v: XRayLogFlushStatus::XRAY_LOG_FLUSHING, |
| 285 | mo: memory_order_release) == |
| 286 | XRayLogFlushStatus::XRAY_LOG_FLUSHING) { |
| 287 | if (Verbosity()) |
| 288 | Report(format: "Not flushing log, implementation is still flushing.\n" ); |
| 289 | return XRayLogFlushStatus::XRAY_LOG_NOT_FLUSHING; |
| 290 | } |
| 291 | |
| 292 | if (BQ == nullptr) { |
| 293 | if (Verbosity()) |
| 294 | Report(format: "Cannot flush when global buffer queue is null.\n" ); |
| 295 | return XRayLogFlushStatus::XRAY_LOG_NOT_FLUSHING; |
| 296 | } |
| 297 | |
| 298 | // We wait a number of milliseconds to allow threads to see that we've |
| 299 | // finalised before attempting to flush the log. |
| 300 | SleepForMillis(millis: fdrFlags()->grace_period_ms); |
| 301 | |
| 302 | // At this point, we're going to uninstall the iterator implementation, before |
| 303 | // we decide to do anything further with the global buffer queue. |
| 304 | __xray_log_remove_buffer_iterator(); |
| 305 | |
| 306 | // Once flushed, we should set the global status of the logging implementation |
| 307 | // to "uninitialized" to allow for FDR-logging multiple runs. |
| 308 | auto ResetToUnitialized = at_scope_exit(fn: [] { |
| 309 | atomic_store(a: &LoggingStatus, v: XRayLogInitStatus::XRAY_LOG_UNINITIALIZED, |
| 310 | mo: memory_order_release); |
| 311 | }); |
| 312 | |
| 313 | auto CleanupBuffers = at_scope_exit(fn: [] { |
| 314 | auto &TLD = getThreadLocalData(); |
| 315 | if (TLD.Controller != nullptr) |
| 316 | TLD.Controller->flush(); |
| 317 | }); |
| 318 | |
| 319 | if (fdrFlags()->no_file_flush) { |
| 320 | if (Verbosity()) |
| 321 | Report(format: "XRay FDR: Not flushing to file, 'no_file_flush=true'.\n" ); |
| 322 | |
| 323 | atomic_store(a: &LogFlushStatus, v: XRayLogFlushStatus::XRAY_LOG_FLUSHED, |
| 324 | mo: memory_order_release); |
| 325 | return XRayLogFlushStatus::XRAY_LOG_FLUSHED; |
| 326 | } |
| 327 | |
| 328 | // We write out the file in the following format: |
| 329 | // |
| 330 | // 1) We write down the XRay file header with version 1, type FDR_LOG. |
| 331 | // 2) Then we use the 'apply' member of the BufferQueue that's live, to |
| 332 | // ensure that at this point in time we write down the buffers that have |
| 333 | // been released (and marked "used") -- we dump the full buffer for now |
| 334 | // (fixed-sized) and let the tools reading the buffers deal with the data |
| 335 | // afterwards. |
| 336 | // |
| 337 | LogWriter *LW = LogWriter::Open(); |
| 338 | if (LW == nullptr) { |
| 339 | auto Result = XRayLogFlushStatus::XRAY_LOG_NOT_FLUSHING; |
| 340 | atomic_store(a: &LogFlushStatus, v: Result, mo: memory_order_release); |
| 341 | return Result; |
| 342 | } |
| 343 | |
| 344 | XRayFileHeader = fdrCommonHeaderInfo(); |
| 345 | Header.FdrData = FdrAdditionalHeaderData{.ThreadBufferSize: BQ->ConfiguredBufferSize()}; |
| 346 | LW->WriteAll(Begin: reinterpret_cast<char *>(&Header), |
| 347 | End: reinterpret_cast<char *>(&Header) + sizeof(Header)); |
| 348 | |
| 349 | // Release the current thread's buffer before we attempt to write out all the |
| 350 | // buffers. This ensures that in case we had only a single thread going, that |
| 351 | // we are able to capture the data nonetheless. |
| 352 | auto &TLD = getThreadLocalData(); |
| 353 | if (TLD.Controller != nullptr) |
| 354 | TLD.Controller->flush(); |
| 355 | |
| 356 | BQ->apply(Fn: [&](const BufferQueue::Buffer &B) { |
| 357 | // Starting at version 2 of the FDR logging implementation, we only write |
| 358 | // the records identified by the extents of the buffer. We use the Extents |
| 359 | // from the Buffer and write that out as the first record in the buffer. We |
| 360 | // still use a Metadata record, but fill in the extents instead for the |
| 361 | // data. |
| 362 | MetadataRecord ExtentsRecord; |
| 363 | auto BufferExtents = atomic_load(a: B.Extents, mo: memory_order_acquire); |
| 364 | DCHECK(BufferExtents <= B.Size); |
| 365 | ExtentsRecord.Type = uint8_t(RecordType::Metadata); |
| 366 | ExtentsRecord.RecordKind = |
| 367 | uint8_t(MetadataRecord::RecordKinds::BufferExtents); |
| 368 | internal_memcpy(dest: ExtentsRecord.Data, src: &BufferExtents, n: sizeof(BufferExtents)); |
| 369 | if (BufferExtents > 0) { |
| 370 | LW->WriteAll(Begin: reinterpret_cast<char *>(&ExtentsRecord), |
| 371 | End: reinterpret_cast<char *>(&ExtentsRecord) + |
| 372 | sizeof(MetadataRecord)); |
| 373 | LW->WriteAll(Begin: reinterpret_cast<char *>(B.Data), |
| 374 | End: reinterpret_cast<char *>(B.Data) + BufferExtents); |
| 375 | } |
| 376 | }); |
| 377 | |
| 378 | atomic_store(a: &LogFlushStatus, v: XRayLogFlushStatus::XRAY_LOG_FLUSHED, |
| 379 | mo: memory_order_release); |
| 380 | return XRayLogFlushStatus::XRAY_LOG_FLUSHED; |
| 381 | } |
| 382 | |
| 383 | XRayLogInitStatus fdrLoggingFinalize() XRAY_NEVER_INSTRUMENT { |
| 384 | s32 CurrentStatus = XRayLogInitStatus::XRAY_LOG_INITIALIZED; |
| 385 | if (!atomic_compare_exchange_strong(a: &LoggingStatus, cmp: &CurrentStatus, |
| 386 | xchg: XRayLogInitStatus::XRAY_LOG_FINALIZING, |
| 387 | mo: memory_order_release)) { |
| 388 | if (Verbosity()) |
| 389 | Report(format: "Cannot finalize log, implementation not initialized.\n" ); |
| 390 | return static_cast<XRayLogInitStatus>(CurrentStatus); |
| 391 | } |
| 392 | |
| 393 | // Do special things to make the log finalize itself, and not allow any more |
| 394 | // operations to be performed until re-initialized. |
| 395 | if (BQ == nullptr) { |
| 396 | if (Verbosity()) |
| 397 | Report(format: "Attempting to finalize an uninitialized global buffer!\n" ); |
| 398 | } else { |
| 399 | BQ->finalize(); |
| 400 | } |
| 401 | |
| 402 | atomic_store(a: &LoggingStatus, v: XRayLogInitStatus::XRAY_LOG_FINALIZED, |
| 403 | mo: memory_order_release); |
| 404 | return XRayLogInitStatus::XRAY_LOG_FINALIZED; |
| 405 | } |
| 406 | |
| 407 | struct TSCAndCPU { |
| 408 | uint64_t TSC = 0; |
| 409 | unsigned char CPU = 0; |
| 410 | }; |
| 411 | |
| 412 | static TSCAndCPU getTimestamp() XRAY_NEVER_INSTRUMENT { |
| 413 | // We want to get the TSC as early as possible, so that we can check whether |
| 414 | // we've seen this CPU before. We also do it before we load anything else, |
| 415 | // to allow for forward progress with the scheduling. |
| 416 | TSCAndCPU Result; |
| 417 | |
| 418 | // Test once for required CPU features |
| 419 | static pthread_once_t OnceProbe = PTHREAD_ONCE_INIT; |
| 420 | static bool TSCSupported = true; |
| 421 | pthread_once( |
| 422 | once_control: &OnceProbe, init_routine: +[] { TSCSupported = probeRequiredCPUFeatures(); }); |
| 423 | |
| 424 | if (TSCSupported) { |
| 425 | Result.TSC = __xray::readTSC(CPU&: Result.CPU); |
| 426 | } else { |
| 427 | // FIXME: This code needs refactoring as it appears in multiple locations |
| 428 | timespec TS; |
| 429 | int result = clock_gettime(CLOCK_REALTIME, tp: &TS); |
| 430 | if (result != 0) { |
| 431 | Report(format: "clock_gettime(2) return %d, errno=%d" , result, int(errno)); |
| 432 | TS = {.tv_sec: 0, .tv_nsec: 0}; |
| 433 | } |
| 434 | Result.CPU = 0; |
| 435 | Result.TSC = TS.tv_sec * __xray::NanosecondsPerSecond + TS.tv_nsec; |
| 436 | } |
| 437 | return Result; |
| 438 | } |
| 439 | |
| 440 | thread_local atomic_uint8_t Running{.val_dont_use: 0}; |
| 441 | |
| 442 | static bool setupTLD(ThreadLocalData &TLD) XRAY_NEVER_INSTRUMENT { |
| 443 | // Check if we're finalizing, before proceeding. |
| 444 | { |
| 445 | auto Status = atomic_load(a: &LoggingStatus, mo: memory_order_acquire); |
| 446 | if (Status == XRayLogInitStatus::XRAY_LOG_FINALIZING || |
| 447 | Status == XRayLogInitStatus::XRAY_LOG_FINALIZED) { |
| 448 | if (TLD.Controller != nullptr) { |
| 449 | TLD.Controller->flush(); |
| 450 | TLD.Controller = nullptr; |
| 451 | } |
| 452 | return false; |
| 453 | } |
| 454 | } |
| 455 | |
| 456 | if (UNLIKELY(TLD.Controller == nullptr)) { |
| 457 | // Set up the TLD buffer queue. |
| 458 | if (UNLIKELY(BQ == nullptr)) |
| 459 | return false; |
| 460 | TLD.BQ = BQ; |
| 461 | |
| 462 | // Check that we have a valid buffer. |
| 463 | if (TLD.Buffer.Generation != BQ->generation() && |
| 464 | TLD.BQ->releaseBuffer(Buf&: TLD.Buffer) != BufferQueue::ErrorCode::Ok) |
| 465 | return false; |
| 466 | |
| 467 | // Set up a buffer, before setting up the log writer. Bail out on failure. |
| 468 | if (TLD.BQ->getBuffer(Buf&: TLD.Buffer) != BufferQueue::ErrorCode::Ok) |
| 469 | return false; |
| 470 | |
| 471 | // Set up the Log Writer for this thread. |
| 472 | if (UNLIKELY(TLD.Writer == nullptr)) { |
| 473 | auto *LWStorage = reinterpret_cast<FDRLogWriter *>(&TLD.LWStorage); |
| 474 | new (LWStorage) FDRLogWriter(TLD.Buffer); |
| 475 | TLD.Writer = LWStorage; |
| 476 | } else { |
| 477 | TLD.Writer->resetRecord(); |
| 478 | } |
| 479 | |
| 480 | auto *CStorage = reinterpret_cast<FDRController<> *>(&TLD.CStorage); |
| 481 | new (CStorage) |
| 482 | FDRController<>(TLD.BQ, TLD.Buffer, *TLD.Writer, clock_gettime, |
| 483 | atomic_load_relaxed(a: &ThresholdTicks)); |
| 484 | TLD.Controller = CStorage; |
| 485 | } |
| 486 | |
| 487 | DCHECK_NE(TLD.Controller, nullptr); |
| 488 | return true; |
| 489 | } |
| 490 | |
| 491 | void fdrLoggingHandleArg0(int32_t FuncId, |
| 492 | XRayEntryType Entry) XRAY_NEVER_INSTRUMENT { |
| 493 | auto TC = getTimestamp(); |
| 494 | auto &TSC = TC.TSC; |
| 495 | auto &CPU = TC.CPU; |
| 496 | RecursionGuard Guard{Running}; |
| 497 | if (!Guard) |
| 498 | return; |
| 499 | |
| 500 | auto &TLD = getThreadLocalData(); |
| 501 | if (!setupTLD(TLD)) |
| 502 | return; |
| 503 | |
| 504 | switch (Entry) { |
| 505 | case XRayEntryType::ENTRY: |
| 506 | case XRayEntryType::LOG_ARGS_ENTRY: |
| 507 | TLD.Controller->functionEnter(FuncId, TSC, CPU); |
| 508 | return; |
| 509 | case XRayEntryType::EXIT: |
| 510 | TLD.Controller->functionExit(FuncId, TSC, CPU); |
| 511 | return; |
| 512 | case XRayEntryType::TAIL: |
| 513 | TLD.Controller->functionTailExit(FuncId, TSC, CPU); |
| 514 | return; |
| 515 | case XRayEntryType::CUSTOM_EVENT: |
| 516 | case XRayEntryType::TYPED_EVENT: |
| 517 | break; |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | void fdrLoggingHandleArg1(int32_t FuncId, XRayEntryType Entry, |
| 522 | uint64_t Arg) XRAY_NEVER_INSTRUMENT { |
| 523 | auto TC = getTimestamp(); |
| 524 | auto &TSC = TC.TSC; |
| 525 | auto &CPU = TC.CPU; |
| 526 | RecursionGuard Guard{Running}; |
| 527 | if (!Guard) |
| 528 | return; |
| 529 | |
| 530 | auto &TLD = getThreadLocalData(); |
| 531 | if (!setupTLD(TLD)) |
| 532 | return; |
| 533 | |
| 534 | switch (Entry) { |
| 535 | case XRayEntryType::ENTRY: |
| 536 | case XRayEntryType::LOG_ARGS_ENTRY: |
| 537 | TLD.Controller->functionEnterArg(FuncId, TSC, CPU, Arg); |
| 538 | return; |
| 539 | case XRayEntryType::EXIT: |
| 540 | TLD.Controller->functionExit(FuncId, TSC, CPU); |
| 541 | return; |
| 542 | case XRayEntryType::TAIL: |
| 543 | TLD.Controller->functionTailExit(FuncId, TSC, CPU); |
| 544 | return; |
| 545 | case XRayEntryType::CUSTOM_EVENT: |
| 546 | case XRayEntryType::TYPED_EVENT: |
| 547 | break; |
| 548 | } |
| 549 | } |
| 550 | |
| 551 | void fdrLoggingHandleCustomEvent(void *Event, |
| 552 | std::size_t EventSize) XRAY_NEVER_INSTRUMENT { |
| 553 | auto TC = getTimestamp(); |
| 554 | auto &TSC = TC.TSC; |
| 555 | auto &CPU = TC.CPU; |
| 556 | RecursionGuard Guard{Running}; |
| 557 | if (!Guard) |
| 558 | return; |
| 559 | |
| 560 | // Complain when we ever get at least one custom event that's larger than what |
| 561 | // we can possibly support. |
| 562 | if (EventSize > |
| 563 | static_cast<std::size_t>(std::numeric_limits<int32_t>::max())) { |
| 564 | static pthread_once_t Once = PTHREAD_ONCE_INIT; |
| 565 | pthread_once( |
| 566 | once_control: &Once, init_routine: +[] { |
| 567 | Report(format: "Custom event size too large; truncating to %d.\n" , |
| 568 | std::numeric_limits<int32_t>::max()); |
| 569 | }); |
| 570 | } |
| 571 | |
| 572 | auto &TLD = getThreadLocalData(); |
| 573 | if (!setupTLD(TLD)) |
| 574 | return; |
| 575 | |
| 576 | int32_t ReducedEventSize = static_cast<int32_t>(EventSize); |
| 577 | TLD.Controller->customEvent(TSC, CPU, Event, EventSize: ReducedEventSize); |
| 578 | } |
| 579 | |
| 580 | void fdrLoggingHandleTypedEvent(size_t EventType, const void *Event, |
| 581 | size_t EventSize) noexcept |
| 582 | XRAY_NEVER_INSTRUMENT { |
| 583 | auto TC = getTimestamp(); |
| 584 | auto &TSC = TC.TSC; |
| 585 | auto &CPU = TC.CPU; |
| 586 | RecursionGuard Guard{Running}; |
| 587 | if (!Guard) |
| 588 | return; |
| 589 | |
| 590 | // Complain when we ever get at least one typed event that's larger than what |
| 591 | // we can possibly support. |
| 592 | if (EventSize > |
| 593 | static_cast<std::size_t>(std::numeric_limits<int32_t>::max())) { |
| 594 | static pthread_once_t Once = PTHREAD_ONCE_INIT; |
| 595 | pthread_once( |
| 596 | once_control: &Once, init_routine: +[] { |
| 597 | Report(format: "Typed event size too large; truncating to %d.\n" , |
| 598 | std::numeric_limits<int32_t>::max()); |
| 599 | }); |
| 600 | } |
| 601 | |
| 602 | auto &TLD = getThreadLocalData(); |
| 603 | if (!setupTLD(TLD)) |
| 604 | return; |
| 605 | |
| 606 | int32_t ReducedEventSize = static_cast<int32_t>(EventSize); |
| 607 | TLD.Controller->typedEvent(TSC, CPU, EventType: static_cast<uint16_t>(EventType), Event, |
| 608 | EventSize: ReducedEventSize); |
| 609 | } |
| 610 | |
| 611 | XRayLogInitStatus fdrLoggingInit(size_t, size_t, void *Options, |
| 612 | size_t OptionsSize) XRAY_NEVER_INSTRUMENT { |
| 613 | if (Options == nullptr) |
| 614 | return XRayLogInitStatus::XRAY_LOG_UNINITIALIZED; |
| 615 | |
| 616 | s32 CurrentStatus = XRayLogInitStatus::XRAY_LOG_UNINITIALIZED; |
| 617 | if (!atomic_compare_exchange_strong(a: &LoggingStatus, cmp: &CurrentStatus, |
| 618 | xchg: XRayLogInitStatus::XRAY_LOG_INITIALIZING, |
| 619 | mo: memory_order_release)) { |
| 620 | if (Verbosity()) |
| 621 | Report(format: "Cannot initialize already initialized implementation.\n" ); |
| 622 | return static_cast<XRayLogInitStatus>(CurrentStatus); |
| 623 | } |
| 624 | |
| 625 | if (Verbosity()) |
| 626 | Report(format: "Initializing FDR mode with options: %s\n" , |
| 627 | static_cast<const char *>(Options)); |
| 628 | |
| 629 | // TODO: Factor out the flags specific to the FDR mode implementation. For |
| 630 | // now, use the global/single definition of the flags, since the FDR mode |
| 631 | // flags are already defined there. |
| 632 | FlagParser FDRParser; |
| 633 | FDRFlags FDRFlags; |
| 634 | registerXRayFDRFlags(P: &FDRParser, F: &FDRFlags); |
| 635 | FDRFlags.setDefaults(); |
| 636 | |
| 637 | // Override first from the general XRAY_DEFAULT_OPTIONS compiler-provided |
| 638 | // options until we migrate everyone to use the XRAY_FDR_OPTIONS |
| 639 | // compiler-provided options. |
| 640 | FDRParser.ParseString(s: useCompilerDefinedFlags()); |
| 641 | FDRParser.ParseString(s: useCompilerDefinedFDRFlags()); |
| 642 | auto *EnvOpts = GetEnv(name: "XRAY_FDR_OPTIONS" ); |
| 643 | if (EnvOpts == nullptr) |
| 644 | EnvOpts = "" ; |
| 645 | FDRParser.ParseString(s: EnvOpts); |
| 646 | |
| 647 | // FIXME: Remove this when we fully remove the deprecated flags. |
| 648 | if (internal_strlen(s: EnvOpts) == 0) { |
| 649 | FDRFlags.func_duration_threshold_us = |
| 650 | flags()->xray_fdr_log_func_duration_threshold_us; |
| 651 | FDRFlags.grace_period_ms = flags()->xray_fdr_log_grace_period_ms; |
| 652 | } |
| 653 | |
| 654 | // The provided options should always override the compiler-provided and |
| 655 | // environment-variable defined options. |
| 656 | FDRParser.ParseString(s: static_cast<const char *>(Options)); |
| 657 | *fdrFlags() = FDRFlags; |
| 658 | auto BufferSize = FDRFlags.buffer_size; |
| 659 | auto BufferMax = FDRFlags.buffer_max; |
| 660 | |
| 661 | if (BQ == nullptr) { |
| 662 | bool Success = false; |
| 663 | BQ = reinterpret_cast<BufferQueue *>(&BufferQueueStorage); |
| 664 | new (BQ) BufferQueue(BufferSize, BufferMax, Success); |
| 665 | if (!Success) { |
| 666 | Report(format: "BufferQueue init failed.\n" ); |
| 667 | return XRayLogInitStatus::XRAY_LOG_UNINITIALIZED; |
| 668 | } |
| 669 | } else { |
| 670 | if (BQ->init(BS: BufferSize, BC: BufferMax) != BufferQueue::ErrorCode::Ok) { |
| 671 | if (Verbosity()) |
| 672 | Report(format: "Failed to re-initialize global buffer queue. Init failed.\n" ); |
| 673 | return XRayLogInitStatus::XRAY_LOG_UNINITIALIZED; |
| 674 | } |
| 675 | } |
| 676 | |
| 677 | static pthread_once_t OnceInit = PTHREAD_ONCE_INIT; |
| 678 | pthread_once( |
| 679 | once_control: &OnceInit, init_routine: +[] { |
| 680 | atomic_store(a: &TicksPerSec, |
| 681 | v: probeRequiredCPUFeatures() ? getTSCFrequency() |
| 682 | : __xray::NanosecondsPerSecond, |
| 683 | mo: memory_order_release); |
| 684 | pthread_key_create( |
| 685 | key: &Key, destr_function: +[](void *TLDPtr) { |
| 686 | if (TLDPtr == nullptr) |
| 687 | return; |
| 688 | auto &TLD = *reinterpret_cast<ThreadLocalData *>(TLDPtr); |
| 689 | if (TLD.BQ == nullptr) |
| 690 | return; |
| 691 | if (TLD.Buffer.Data == nullptr) |
| 692 | return; |
| 693 | auto EC = TLD.BQ->releaseBuffer(Buf&: TLD.Buffer); |
| 694 | if (EC != BufferQueue::ErrorCode::Ok) |
| 695 | Report(format: "At thread exit, failed to release buffer at %p; " |
| 696 | "error=%s\n" , |
| 697 | TLD.Buffer.Data, BufferQueue::getErrorString(E: EC)); |
| 698 | }); |
| 699 | }); |
| 700 | |
| 701 | atomic_store(a: &ThresholdTicks, |
| 702 | v: atomic_load_relaxed(a: &TicksPerSec) * |
| 703 | fdrFlags()->func_duration_threshold_us / 1000000, |
| 704 | mo: memory_order_release); |
| 705 | // Arg1 handler should go in first to avoid concurrent code accidentally |
| 706 | // falling back to arg0 when it should have ran arg1. |
| 707 | __xray_set_handler_arg1(entry: fdrLoggingHandleArg1); |
| 708 | // Install the actual handleArg0 handler after initialising the buffers. |
| 709 | __xray_set_handler(entry: fdrLoggingHandleArg0); |
| 710 | __xray_set_customevent_handler(entry: fdrLoggingHandleCustomEvent); |
| 711 | __xray_set_typedevent_handler(entry: fdrLoggingHandleTypedEvent); |
| 712 | |
| 713 | // Install the buffer iterator implementation. |
| 714 | __xray_log_set_buffer_iterator(Iterator: fdrIterator); |
| 715 | |
| 716 | atomic_store(a: &LoggingStatus, v: XRayLogInitStatus::XRAY_LOG_INITIALIZED, |
| 717 | mo: memory_order_release); |
| 718 | |
| 719 | if (Verbosity()) |
| 720 | Report(format: "XRay FDR init successful.\n" ); |
| 721 | return XRayLogInitStatus::XRAY_LOG_INITIALIZED; |
| 722 | } |
| 723 | |
| 724 | bool fdrLogDynamicInitializer() XRAY_NEVER_INSTRUMENT { |
| 725 | XRayLogImpl Impl{ |
| 726 | .log_init: fdrLoggingInit, |
| 727 | .log_finalize: fdrLoggingFinalize, |
| 728 | .handle_arg0: fdrLoggingHandleArg0, |
| 729 | .flush_log: fdrLoggingFlush, |
| 730 | }; |
| 731 | auto RegistrationResult = __xray_log_register_mode(Mode: "xray-fdr" , Impl); |
| 732 | if (RegistrationResult != XRayLogRegisterStatus::XRAY_REGISTRATION_OK && |
| 733 | Verbosity()) { |
| 734 | Report(format: "Cannot register XRay FDR mode to 'xray-fdr'; error = %d\n" , |
| 735 | RegistrationResult); |
| 736 | return false; |
| 737 | } |
| 738 | |
| 739 | if (flags()->xray_fdr_log || |
| 740 | !internal_strcmp(s1: flags()->xray_mode, s2: "xray-fdr" )) { |
| 741 | auto SelectResult = __xray_log_select_mode(Mode: "xray-fdr" ); |
| 742 | if (SelectResult != XRayLogRegisterStatus::XRAY_REGISTRATION_OK && |
| 743 | Verbosity()) { |
| 744 | Report(format: "Cannot select XRay FDR mode as 'xray-fdr'; error = %d\n" , |
| 745 | SelectResult); |
| 746 | return false; |
| 747 | } |
| 748 | } |
| 749 | return true; |
| 750 | } |
| 751 | |
| 752 | } // namespace __xray |
| 753 | |
| 754 | static auto UNUSED Unused = __xray::fdrLogDynamicInitializer(); |
| 755 | |