/src/abseil-cpp/absl/time/clock.cc
Line | Count | Source |
1 | | // Copyright 2017 The Abseil Authors. |
2 | | // |
3 | | // Licensed under the Apache License, Version 2.0 (the "License"); |
4 | | // you may not use this file except in compliance with the License. |
5 | | // You may obtain a copy of the License at |
6 | | // |
7 | | // https://www.apache.org/licenses/LICENSE-2.0 |
8 | | // |
9 | | // Unless required by applicable law or agreed to in writing, software |
10 | | // distributed under the License is distributed on an "AS IS" BASIS, |
11 | | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | | // See the License for the specific language governing permissions and |
13 | | // limitations under the License. |
14 | | |
15 | | #include "absl/time/clock.h" |
16 | | |
17 | | #include <algorithm> |
18 | | #include <atomic> |
19 | | #include <cerrno> |
20 | | #include <cstdint> |
21 | | #include <ctime> |
22 | | #include <limits> |
23 | | |
24 | | #include "absl/base/attributes.h" |
25 | | #include "absl/base/config.h" |
26 | | #include "absl/base/internal/spinlock.h" |
27 | | #include "absl/base/internal/unscaledcycleclock.h" |
28 | | #include "absl/base/internal/unscaledcycleclock_config.h" |
29 | | #include "absl/base/macros.h" |
30 | | #include "absl/base/optimization.h" |
31 | | #include "absl/base/port.h" |
32 | | #include "absl/base/thread_annotations.h" |
33 | | #include "absl/time/time.h" |
34 | | |
35 | | #ifdef _WIN32 |
36 | | #include <windows.h> |
37 | | #endif |
38 | | |
39 | | namespace absl { |
40 | | ABSL_NAMESPACE_BEGIN |
41 | 0 | Time Now() { |
42 | | // TODO(bww): Get a timespec instead so we don't have to divide. |
43 | 0 | int64_t n = absl::GetCurrentTimeNanos(); |
44 | 0 | if (n >= 0) { |
45 | 0 | return time_internal::FromUnixDuration( |
46 | 0 | time_internal::MakeDuration(n / 1000000000, n % 1000000000 * 4)); |
47 | 0 | } |
48 | 0 | return time_internal::FromUnixDuration(absl::Nanoseconds(n)); |
49 | 0 | } |
50 | | ABSL_NAMESPACE_END |
51 | | } // namespace absl |
52 | | |
53 | | // Decide if we should use the fast GetCurrentTimeNanos() algorithm based on the |
54 | | // cyclecounter, otherwise just get the time directly from the OS on every call. |
55 | | // By default, the fast algorithm based on the cyclecount is disabled because in |
56 | | // certain situations, for example, if the OS enters a "sleep" mode, it may |
57 | | // produce incorrect values immediately upon waking. |
58 | | // This can be chosen at compile-time via |
59 | | // -DABSL_USE_CYCLECLOCK_FOR_GET_CURRENT_TIME_NANOS=[0|1] |
60 | | #ifndef ABSL_USE_CYCLECLOCK_FOR_GET_CURRENT_TIME_NANOS |
61 | | #define ABSL_USE_CYCLECLOCK_FOR_GET_CURRENT_TIME_NANOS 0 |
62 | | #endif |
63 | | |
64 | | #if defined(__APPLE__) || defined(_WIN32) |
65 | | #include "absl/time/internal/get_current_time_chrono.inc" |
66 | | #else |
67 | | #include "absl/time/internal/get_current_time_posix.inc" |
68 | | #endif |
69 | | |
70 | | // Allows override by test. |
71 | | #ifndef GET_CURRENT_TIME_NANOS_FROM_SYSTEM |
72 | | #define GET_CURRENT_TIME_NANOS_FROM_SYSTEM() \ |
73 | 0 | ::absl::time_internal::GetCurrentTimeNanosFromSystem() |
74 | | #endif |
75 | | |
76 | | #if !ABSL_USE_CYCLECLOCK_FOR_GET_CURRENT_TIME_NANOS |
77 | | namespace absl { |
78 | | ABSL_NAMESPACE_BEGIN |
79 | 0 | int64_t GetCurrentTimeNanos() { return GET_CURRENT_TIME_NANOS_FROM_SYSTEM(); } |
80 | | ABSL_NAMESPACE_END |
81 | | } // namespace absl |
82 | | #else // Use the cyclecounter-based implementation below. |
83 | | |
84 | | // Allows override by test. |
85 | | #ifndef GET_CURRENT_TIME_NANOS_CYCLECLOCK_NOW |
86 | | #define GET_CURRENT_TIME_NANOS_CYCLECLOCK_NOW() \ |
87 | | ::absl::time_internal::UnscaledCycleClockWrapperForGetCurrentTime::Now() |
88 | | #endif |
89 | | |
90 | | namespace absl { |
91 | | ABSL_NAMESPACE_BEGIN |
92 | | namespace time_internal { |
93 | | |
94 | | // On some processors, consecutive reads of the cycle counter may yield the |
95 | | // same value (weakly-increasing). In debug mode, clear the least significant |
96 | | // bits to discourage depending on a strictly-increasing Now() value. |
97 | | // In x86-64's debug mode, discourage depending on a strictly-increasing Now() |
98 | | // value. |
99 | | #if !defined(NDEBUG) && defined(__x86_64__) |
100 | | constexpr int64_t kCycleClockNowMask = ~int64_t{0xff}; |
101 | | #else |
102 | | constexpr int64_t kCycleClockNowMask = ~int64_t{0}; |
103 | | #endif |
104 | | |
105 | | // This is a friend wrapper around UnscaledCycleClock::Now() |
106 | | // (needed to access UnscaledCycleClock). |
107 | | class UnscaledCycleClockWrapperForGetCurrentTime { |
108 | | public: |
109 | | static int64_t Now() { |
110 | | return base_internal::UnscaledCycleClock::Now() & kCycleClockNowMask; |
111 | | } |
112 | | }; |
113 | | } // namespace time_internal |
114 | | |
115 | | // uint64_t is used in this module to provide an extra bit in multiplications |
116 | | |
117 | | // --------------------------------------------------------------------- |
118 | | // An implementation of reader-write locks that use no atomic ops in the read |
119 | | // case. This is a generalization of Lamport's method for reading a multiword |
120 | | // clock. Increment a word on each write acquisition, using the low-order bit |
121 | | // as a spinlock; the word is the high word of the "clock". Readers read the |
122 | | // high word, then all other data, then the high word again, and repeat the |
123 | | // read if the reads of the high words yields different answers, or an odd |
124 | | // value (either case suggests possible interference from a writer). |
125 | | // Here we use a spinlock to ensure only one writer at a time, rather than |
126 | | // spinning on the bottom bit of the word to benefit from SpinLock |
127 | | // spin-delay tuning. |
128 | | |
129 | | // Acquire seqlock (*seq) and return the value to be written to unlock. |
130 | | static inline uint64_t SeqAcquire(std::atomic<uint64_t>* seq) { |
131 | | uint64_t x = seq->fetch_add(1, std::memory_order_relaxed); |
132 | | |
133 | | // We put a release fence between update to *seq and writes to shared data. |
134 | | // Thus all stores to shared data are effectively release operations and |
135 | | // update to *seq above cannot be re-ordered past any of them. Note that |
136 | | // this barrier is not for the fetch_add above. A release barrier for the |
137 | | // fetch_add would be before it, not after. |
138 | | std::atomic_thread_fence(std::memory_order_release); |
139 | | |
140 | | return x + 2; // original word plus 2 |
141 | | } |
142 | | |
143 | | // Release seqlock (*seq) by writing x to it---a value previously returned by |
144 | | // SeqAcquire. |
145 | | static inline void SeqRelease(std::atomic<uint64_t>* seq, uint64_t x) { |
146 | | // The unlock store to *seq must have release ordering so that all |
147 | | // updates to shared data must finish before this store. |
148 | | seq->store(x, std::memory_order_release); // release lock for readers |
149 | | } |
150 | | |
151 | | // --------------------------------------------------------------------- |
152 | | |
153 | | // "nsscaled" is unit of time equal to a (2**kScale)th of a nanosecond. |
154 | | enum { kScale = 30 }; |
155 | | |
156 | | // The minimum interval between samples of the time base. |
157 | | // We pick enough time to amortize the cost of the sample, |
158 | | // to get a reasonably accurate cycle counter rate reading, |
159 | | // and not so much that calculations will overflow 64-bits. |
160 | | static const uint64_t kMinNSBetweenSamples = 2000 << 20; |
161 | | |
162 | | // We require that kMinNSBetweenSamples shifted by kScale |
163 | | // have at least a bit left over for 64-bit calculations. |
164 | | static_assert(((kMinNSBetweenSamples << (kScale + 1)) >> (kScale + 1)) == |
165 | | kMinNSBetweenSamples, |
166 | | "cannot represent kMaxBetweenSamplesNSScaled"); |
167 | | |
168 | | // data from a sample of the kernel's time value |
169 | | struct TimeSampleAtomic { |
170 | | std::atomic<uint64_t> raw_ns{0}; // raw kernel time |
171 | | std::atomic<uint64_t> base_ns{0}; // our estimate of time |
172 | | std::atomic<uint64_t> base_cycles{0}; // cycle counter reading |
173 | | std::atomic<uint64_t> nsscaled_per_cycle{0}; // cycle period |
174 | | // cycles before we'll sample again (a scaled reciprocal of the period, |
175 | | // to avoid a division on the fast path). |
176 | | std::atomic<uint64_t> min_cycles_per_sample{0}; |
177 | | }; |
178 | | // Same again, but with non-atomic types |
179 | | struct TimeSample { |
180 | | uint64_t raw_ns = 0; // raw kernel time |
181 | | uint64_t base_ns = 0; // our estimate of time |
182 | | uint64_t base_cycles = 0; // cycle counter reading |
183 | | uint64_t nsscaled_per_cycle = 0; // cycle period |
184 | | uint64_t min_cycles_per_sample = 0; // approx cycles before next sample |
185 | | }; |
186 | | |
187 | | struct ABSL_CACHELINE_ALIGNED TimeState { |
188 | | std::atomic<uint64_t> seq{0}; |
189 | | TimeSampleAtomic last_sample; // the last sample; under seq |
190 | | |
191 | | // The following counters are used only by the test code. |
192 | | int64_t stats_initializations{0}; |
193 | | int64_t stats_reinitializations{0}; |
194 | | int64_t stats_calibrations{0}; |
195 | | int64_t stats_slow_paths{0}; |
196 | | int64_t stats_fast_slow_paths{0}; |
197 | | |
198 | | uint64_t last_now_cycles ABSL_GUARDED_BY(lock){0}; |
199 | | |
200 | | // Used by GetCurrentTimeNanosFromKernel(). |
201 | | // We try to read clock values at about the same time as the kernel clock. |
202 | | // This value gets adjusted up or down as estimate of how long that should |
203 | | // take, so we can reject attempts that take unusually long. |
204 | | std::atomic<uint64_t> approx_syscall_time_in_cycles{10 * 1000}; |
205 | | // Number of times in a row we've seen a kernel time call take substantially |
206 | | // less than approx_syscall_time_in_cycles. |
207 | | std::atomic<uint32_t> kernel_time_seen_smaller{0}; |
208 | | |
209 | | // A reader-writer lock protecting the static locations below. |
210 | | // See SeqAcquire() and SeqRelease() above. |
211 | | absl::base_internal::SpinLock lock{base_internal::SCHEDULE_KERNEL_ONLY}; |
212 | | }; |
213 | | ABSL_CONST_INIT static TimeState time_state; |
214 | | |
215 | | // Return the time in ns as told by the kernel interface. Place in *cycleclock |
216 | | // the value of the cycleclock at about the time of the syscall. |
217 | | // This call represents the time base that this module synchronizes to. |
218 | | // Ensures that *cycleclock does not step back by up to (1 << 16) from |
219 | | // last_cycleclock, to discard small backward counter steps. (Larger steps are |
220 | | // assumed to be complete resyncs, which shouldn't happen. If they do, a full |
221 | | // reinitialization of the outer algorithm should occur.) |
222 | | static int64_t GetCurrentTimeNanosFromKernel(uint64_t last_cycleclock, |
223 | | uint64_t* cycleclock) |
224 | | ABSL_EXCLUSIVE_LOCKS_REQUIRED(time_state.lock) { |
225 | | uint64_t local_approx_syscall_time_in_cycles = // local copy |
226 | | time_state.approx_syscall_time_in_cycles.load(std::memory_order_relaxed); |
227 | | |
228 | | int64_t current_time_nanos_from_system; |
229 | | uint64_t before_cycles; |
230 | | uint64_t after_cycles; |
231 | | uint64_t elapsed_cycles; |
232 | | int loops = 0; |
233 | | do { |
234 | | before_cycles = |
235 | | static_cast<uint64_t>(GET_CURRENT_TIME_NANOS_CYCLECLOCK_NOW()); |
236 | | current_time_nanos_from_system = GET_CURRENT_TIME_NANOS_FROM_SYSTEM(); |
237 | | after_cycles = |
238 | | static_cast<uint64_t>(GET_CURRENT_TIME_NANOS_CYCLECLOCK_NOW()); |
239 | | // elapsed_cycles is unsigned, so is large on overflow |
240 | | elapsed_cycles = after_cycles - before_cycles; |
241 | | if (elapsed_cycles >= local_approx_syscall_time_in_cycles && |
242 | | ++loops == 20) { // clock changed frequencies? Back off. |
243 | | loops = 0; |
244 | | if (local_approx_syscall_time_in_cycles < 1000 * 1000) { |
245 | | local_approx_syscall_time_in_cycles = |
246 | | (local_approx_syscall_time_in_cycles + 1) << 1; |
247 | | } |
248 | | time_state.approx_syscall_time_in_cycles.store( |
249 | | local_approx_syscall_time_in_cycles, std::memory_order_relaxed); |
250 | | } |
251 | | } while (elapsed_cycles >= local_approx_syscall_time_in_cycles || |
252 | | last_cycleclock - after_cycles < (static_cast<uint64_t>(1) << 16)); |
253 | | |
254 | | // Adjust approx_syscall_time_in_cycles to be within a factor of 2 |
255 | | // of the typical time to execute one iteration of the loop above. |
256 | | if ((local_approx_syscall_time_in_cycles >> 1) < elapsed_cycles) { |
257 | | // measured time is no smaller than half current approximation |
258 | | time_state.kernel_time_seen_smaller.store(0, std::memory_order_relaxed); |
259 | | } else if (time_state.kernel_time_seen_smaller.fetch_add( |
260 | | 1, std::memory_order_relaxed) >= 3) { |
261 | | // smaller delays several times in a row; reduce approximation by 12.5% |
262 | | const uint64_t new_approximation = |
263 | | local_approx_syscall_time_in_cycles - |
264 | | (local_approx_syscall_time_in_cycles >> 3); |
265 | | time_state.approx_syscall_time_in_cycles.store(new_approximation, |
266 | | std::memory_order_relaxed); |
267 | | time_state.kernel_time_seen_smaller.store(0, std::memory_order_relaxed); |
268 | | } |
269 | | |
270 | | *cycleclock = after_cycles; |
271 | | return current_time_nanos_from_system; |
272 | | } |
273 | | |
274 | | static int64_t GetCurrentTimeNanosSlowPath() ABSL_ATTRIBUTE_COLD; |
275 | | |
276 | | // Read the contents of *atomic into *sample. |
277 | | // Each field is read atomically, but to maintain atomicity between fields, |
278 | | // the access must be done under a lock. |
279 | | static void ReadTimeSampleAtomic(const struct TimeSampleAtomic* atomic, |
280 | | struct TimeSample* sample) { |
281 | | sample->base_ns = atomic->base_ns.load(std::memory_order_relaxed); |
282 | | sample->base_cycles = atomic->base_cycles.load(std::memory_order_relaxed); |
283 | | sample->nsscaled_per_cycle = |
284 | | atomic->nsscaled_per_cycle.load(std::memory_order_relaxed); |
285 | | sample->min_cycles_per_sample = |
286 | | atomic->min_cycles_per_sample.load(std::memory_order_relaxed); |
287 | | sample->raw_ns = atomic->raw_ns.load(std::memory_order_relaxed); |
288 | | } |
289 | | |
290 | | // Public routine. |
291 | | // Algorithm: We wish to compute real time from a cycle counter. In normal |
292 | | // operation, we construct a piecewise linear approximation to the kernel time |
293 | | // source, using the cycle counter value. The start of each line segment is at |
294 | | // the same point as the end of the last, but may have a different slope (that |
295 | | // is, a different idea of the cycle counter frequency). Every couple of |
296 | | // seconds, the kernel time source is sampled and compared with the current |
297 | | // approximation. A new slope is chosen that, if followed for another couple |
298 | | // of seconds, will correct the error at the current position. The information |
299 | | // for a sample is in the "last_sample" struct. The linear approximation is |
300 | | // estimated_time = last_sample.base_ns + |
301 | | // last_sample.ns_per_cycle * (counter_reading - last_sample.base_cycles) |
302 | | // (ns_per_cycle is actually stored in different units and scaled, to avoid |
303 | | // overflow). The base_ns of the next linear approximation is the |
304 | | // estimated_time using the last approximation; the base_cycles is the cycle |
305 | | // counter value at that time; the ns_per_cycle is the number of ns per cycle |
306 | | // measured since the last sample, but adjusted so that most of the difference |
307 | | // between the estimated_time and the kernel time will be corrected by the |
308 | | // estimated time to the next sample. In normal operation, this algorithm |
309 | | // relies on: |
310 | | // - the cycle counter and kernel time rates not changing a lot in a few |
311 | | // seconds. |
312 | | // - the client calling into the code often compared to a couple of seconds, so |
313 | | // the time to the next correction can be estimated. |
314 | | // Any time ns_per_cycle is not known, a major error is detected, or the |
315 | | // assumption about frequent calls is violated, the implementation returns the |
316 | | // kernel time. It records sufficient data that a linear approximation can |
317 | | // resume a little later. |
318 | | |
319 | | int64_t GetCurrentTimeNanos() { |
320 | | // read the data from the "last_sample" struct (but don't need raw_ns yet) |
321 | | // The reads of "seq" and test of the values emulate a reader lock. |
322 | | uint64_t base_ns; |
323 | | uint64_t base_cycles; |
324 | | uint64_t nsscaled_per_cycle; |
325 | | uint64_t min_cycles_per_sample; |
326 | | uint64_t seq_read0; |
327 | | uint64_t seq_read1; |
328 | | |
329 | | // If we have enough information to interpolate, the value returned will be |
330 | | // derived from this cycleclock-derived time estimate. On some platforms |
331 | | // (POWER) the function to retrieve this value has enough complexity to |
332 | | // contribute to register pressure - reading it early before initializing |
333 | | // the other pieces of the calculation minimizes spill/restore instructions, |
334 | | // minimizing icache cost. |
335 | | uint64_t now_cycles = |
336 | | static_cast<uint64_t>(GET_CURRENT_TIME_NANOS_CYCLECLOCK_NOW()); |
337 | | |
338 | | // Acquire pairs with the barrier in SeqRelease - if this load sees that |
339 | | // store, the shared-data reads necessarily see that SeqRelease's updates |
340 | | // to the same shared data. |
341 | | seq_read0 = time_state.seq.load(std::memory_order_acquire); |
342 | | |
343 | | // The algorithm does not require that the following four loads be ordered |
344 | | // with respect to one another; it requires only that they precede the load of |
345 | | // time_state.seq below them. Nevertheless, we mark each of them as an |
346 | | // acquire-load, rather than using a barrier immediately before the |
347 | | // time_state.seq load, because the former is likely faster on most CPUs of |
348 | | // interest. Architectures that may see a regression because of this approach |
349 | | // include PowerPC and MIPS. |
350 | | base_ns = time_state.last_sample.base_ns.load(std::memory_order_acquire); |
351 | | base_cycles = |
352 | | time_state.last_sample.base_cycles.load(std::memory_order_acquire); |
353 | | nsscaled_per_cycle = |
354 | | time_state.last_sample.nsscaled_per_cycle.load(std::memory_order_acquire); |
355 | | min_cycles_per_sample = time_state.last_sample.min_cycles_per_sample.load( |
356 | | std::memory_order_acquire); |
357 | | |
358 | | // The shared-data reads are effectively acquire ordered, and the |
359 | | // shared-data writes are effectively release ordered. Therefore if our |
360 | | // shared-data reads see any of a particular update's shared-data writes, |
361 | | // seq_read1 is guaranteed to see that update's SeqAcquire. |
362 | | seq_read1 = time_state.seq.load(std::memory_order_relaxed); |
363 | | |
364 | | // Fast path. Return if min_cycles_per_sample has not yet elapsed since the |
365 | | // last sample, and we read a consistent sample. The fast path activates |
366 | | // only when min_cycles_per_sample is non-zero, which happens when we get an |
367 | | // estimate for the cycle time. The predicate will fail if now_cycles < |
368 | | // base_cycles, or if some other thread is in the slow path. |
369 | | // |
370 | | // Since we now read now_cycles before base_ns, it is possible for now_cycles |
371 | | // to be less than base_cycles (if we were interrupted between those loads and |
372 | | // last_sample was updated). This is harmless, because delta_cycles will wrap |
373 | | // and report a time much much bigger than min_cycles_per_sample. In that case |
374 | | // we will take the slow path. |
375 | | uint64_t delta_cycles; |
376 | | if (seq_read0 == seq_read1 && (seq_read0 & 1) == 0 && |
377 | | (delta_cycles = now_cycles - base_cycles) < min_cycles_per_sample) { |
378 | | return static_cast<int64_t>( |
379 | | base_ns + ((delta_cycles * nsscaled_per_cycle) >> kScale)); |
380 | | } |
381 | | return GetCurrentTimeNanosSlowPath(); |
382 | | } |
383 | | |
384 | | // Return (a << kScale)/b. |
385 | | // Zero is returned if b==0. Scaling is performed internally to |
386 | | // preserve precision without overflow. |
387 | | static uint64_t SafeDivideAndScale(uint64_t a, uint64_t b) { |
388 | | // Find maximum safe_shift so that |
389 | | // 0 <= safe_shift <= kScale and (a << safe_shift) does not overflow. |
390 | | int safe_shift = kScale; |
391 | | while (((a << safe_shift) >> safe_shift) != a) { |
392 | | safe_shift--; |
393 | | } |
394 | | uint64_t scaled_b = b >> (kScale - safe_shift); |
395 | | uint64_t quotient = 0; |
396 | | if (scaled_b != 0) { |
397 | | quotient = (a << safe_shift) / scaled_b; |
398 | | } |
399 | | return quotient; |
400 | | } |
401 | | |
402 | | static uint64_t UpdateLastSample( |
403 | | uint64_t now_cycles, uint64_t now_ns, uint64_t delta_cycles, |
404 | | const struct TimeSample* sample) ABSL_ATTRIBUTE_COLD; |
405 | | |
406 | | // The slow path of GetCurrentTimeNanos(). This is taken while gathering |
407 | | // initial samples, when enough time has elapsed since the last sample, and if |
408 | | // any other thread is writing to last_sample. |
409 | | // |
410 | | // Manually mark this 'noinline' to minimize stack frame size of the fast |
411 | | // path. Without this, sometimes a compiler may inline this big block of code |
412 | | // into the fast path. That causes lots of register spills and reloads that |
413 | | // are unnecessary unless the slow path is taken. |
414 | | // |
415 | | // TODO(absl-team): Remove this attribute when our compiler is smart enough |
416 | | // to do the right thing. |
417 | | ABSL_ATTRIBUTE_NOINLINE |
418 | | static int64_t GetCurrentTimeNanosSlowPath() |
419 | | ABSL_LOCKS_EXCLUDED(time_state.lock) { |
420 | | // Serialize access to slow-path. Fast-path readers are not blocked yet, and |
421 | | // code below must not modify last_sample until the seqlock is acquired. |
422 | | base_internal::SpinLockHolder l(time_state.lock); |
423 | | |
424 | | // Sample the kernel time base. This is the definition of |
425 | | // "now" if we take the slow path. |
426 | | uint64_t now_cycles; |
427 | | uint64_t now_ns = static_cast<uint64_t>( |
428 | | GetCurrentTimeNanosFromKernel(time_state.last_now_cycles, &now_cycles)); |
429 | | time_state.last_now_cycles = now_cycles; |
430 | | |
431 | | uint64_t estimated_base_ns; |
432 | | |
433 | | // ---------- |
434 | | // Read the "last_sample" values again; this time holding the write lock. |
435 | | struct TimeSample sample; |
436 | | ReadTimeSampleAtomic(&time_state.last_sample, &sample); |
437 | | |
438 | | // ---------- |
439 | | // Try running the fast path again; another thread may have updated the |
440 | | // sample between our run of the fast path and the sample we just read. |
441 | | uint64_t delta_cycles = now_cycles - sample.base_cycles; |
442 | | if (delta_cycles < sample.min_cycles_per_sample) { |
443 | | // Another thread updated the sample. This path does not take the seqlock |
444 | | // so that blocked readers can make progress without blocking new readers. |
445 | | estimated_base_ns = |
446 | | sample.base_ns + ((delta_cycles * sample.nsscaled_per_cycle) >> kScale); |
447 | | time_state.stats_fast_slow_paths++; |
448 | | } else { |
449 | | estimated_base_ns = |
450 | | UpdateLastSample(now_cycles, now_ns, delta_cycles, &sample); |
451 | | } |
452 | | |
453 | | return static_cast<int64_t>(estimated_base_ns); |
454 | | } |
455 | | |
456 | | // Main part of the algorithm. Locks out readers, updates the approximation |
457 | | // using the new sample from the kernel, and stores the result in last_sample |
458 | | // for readers. Returns the new estimated time. |
459 | | static uint64_t UpdateLastSample(uint64_t now_cycles, uint64_t now_ns, |
460 | | uint64_t delta_cycles, |
461 | | const struct TimeSample* sample) |
462 | | ABSL_EXCLUSIVE_LOCKS_REQUIRED(time_state.lock) { |
463 | | uint64_t estimated_base_ns = now_ns; |
464 | | uint64_t lock_value = |
465 | | SeqAcquire(&time_state.seq); // acquire seqlock to block readers |
466 | | |
467 | | // The 5s in the next if-statement limits the time for which we will trust |
468 | | // the cycle counter and our last sample to give a reasonable result. |
469 | | // Errors in the rate of the source clock can be multiplied by the ratio |
470 | | // between this limit and kMinNSBetweenSamples. |
471 | | if (sample->raw_ns == 0 || // no recent sample, or clock went backwards |
472 | | sample->raw_ns + static_cast<uint64_t>(5) * 1000 * 1000 * 1000 < now_ns || |
473 | | now_ns < sample->raw_ns || now_cycles < sample->base_cycles) { |
474 | | // record this sample, and forget any previously known slope. |
475 | | time_state.last_sample.raw_ns.store(now_ns, std::memory_order_relaxed); |
476 | | time_state.last_sample.base_ns.store(estimated_base_ns, |
477 | | std::memory_order_relaxed); |
478 | | time_state.last_sample.base_cycles.store(now_cycles, |
479 | | std::memory_order_relaxed); |
480 | | time_state.last_sample.nsscaled_per_cycle.store(0, |
481 | | std::memory_order_relaxed); |
482 | | time_state.last_sample.min_cycles_per_sample.store( |
483 | | 0, std::memory_order_relaxed); |
484 | | time_state.stats_initializations++; |
485 | | } else if (sample->raw_ns + 500 * 1000 * 1000 < now_ns && |
486 | | sample->base_cycles + 50 < now_cycles) { |
487 | | // Enough time has passed to compute the cycle time. |
488 | | if (sample->nsscaled_per_cycle != 0) { // Have a cycle time estimate. |
489 | | // Compute time from counter reading, but avoiding overflow |
490 | | // delta_cycles may be larger than on the fast path. |
491 | | uint64_t estimated_scaled_ns; |
492 | | int s = -1; |
493 | | do { |
494 | | s++; |
495 | | estimated_scaled_ns = (delta_cycles >> s) * sample->nsscaled_per_cycle; |
496 | | } while (estimated_scaled_ns / sample->nsscaled_per_cycle != |
497 | | (delta_cycles >> s)); |
498 | | estimated_base_ns = |
499 | | sample->base_ns + (estimated_scaled_ns >> (kScale - s)); |
500 | | } |
501 | | |
502 | | // Compute the assumed cycle time kMinNSBetweenSamples ns into the future |
503 | | // assuming the cycle counter rate stays the same as the last interval. |
504 | | uint64_t ns = now_ns - sample->raw_ns; |
505 | | uint64_t measured_nsscaled_per_cycle = SafeDivideAndScale(ns, delta_cycles); |
506 | | |
507 | | uint64_t assumed_next_sample_delta_cycles = |
508 | | SafeDivideAndScale(kMinNSBetweenSamples, measured_nsscaled_per_cycle); |
509 | | |
510 | | // Estimate low by this much. |
511 | | int64_t diff_ns = static_cast<int64_t>(now_ns - estimated_base_ns); |
512 | | |
513 | | // We want to set nsscaled_per_cycle so that our estimate of the ns time |
514 | | // at the assumed cycle time is the assumed ns time. |
515 | | // That is, we want to set nsscaled_per_cycle so: |
516 | | // kMinNSBetweenSamples + diff_ns == |
517 | | // (assumed_next_sample_delta_cycles * nsscaled_per_cycle) >> kScale |
518 | | // But we wish to damp oscillations, so instead correct only most |
519 | | // of our current error, by solving: |
520 | | // kMinNSBetweenSamples + diff_ns - (diff_ns / 16) == |
521 | | // (assumed_next_sample_delta_cycles * nsscaled_per_cycle) >> kScale |
522 | | ns = static_cast<uint64_t>(static_cast<int64_t>(kMinNSBetweenSamples) + |
523 | | diff_ns - (diff_ns / 16)); |
524 | | uint64_t new_nsscaled_per_cycle = |
525 | | SafeDivideAndScale(ns, assumed_next_sample_delta_cycles); |
526 | | if (new_nsscaled_per_cycle != 0 && diff_ns < 100 * 1000 * 1000 && |
527 | | -diff_ns < 100 * 1000 * 1000) { |
528 | | // record the cycle time measurement |
529 | | time_state.last_sample.nsscaled_per_cycle.store( |
530 | | new_nsscaled_per_cycle, std::memory_order_relaxed); |
531 | | uint64_t new_min_cycles_per_sample = |
532 | | SafeDivideAndScale(kMinNSBetweenSamples, new_nsscaled_per_cycle); |
533 | | time_state.last_sample.min_cycles_per_sample.store( |
534 | | new_min_cycles_per_sample, std::memory_order_relaxed); |
535 | | time_state.stats_calibrations++; |
536 | | } else { // something went wrong; forget the slope |
537 | | time_state.last_sample.nsscaled_per_cycle.store( |
538 | | 0, std::memory_order_relaxed); |
539 | | time_state.last_sample.min_cycles_per_sample.store( |
540 | | 0, std::memory_order_relaxed); |
541 | | estimated_base_ns = now_ns; |
542 | | time_state.stats_reinitializations++; |
543 | | } |
544 | | time_state.last_sample.raw_ns.store(now_ns, std::memory_order_relaxed); |
545 | | time_state.last_sample.base_ns.store(estimated_base_ns, |
546 | | std::memory_order_relaxed); |
547 | | time_state.last_sample.base_cycles.store(now_cycles, |
548 | | std::memory_order_relaxed); |
549 | | } else { |
550 | | // have a sample, but no slope; waiting for enough time for a calibration |
551 | | time_state.stats_slow_paths++; |
552 | | } |
553 | | |
554 | | SeqRelease(&time_state.seq, lock_value); // release the readers |
555 | | |
556 | | return estimated_base_ns; |
557 | | } |
558 | | ABSL_NAMESPACE_END |
559 | | } // namespace absl |
560 | | #endif // ABSL_USE_CYCLECLOCK_FOR_GET_CURRENT_TIME_NANOS |
561 | | |
562 | | namespace absl { |
563 | | ABSL_NAMESPACE_BEGIN |
564 | | namespace { |
565 | | |
566 | | // Returns the maximum duration that SleepOnce() can sleep for. |
567 | 0 | constexpr absl::Duration MaxSleep() { |
568 | | #ifdef _WIN32 |
569 | | // Windows Sleep() takes unsigned long argument in milliseconds. |
570 | | return absl::Milliseconds( |
571 | | std::numeric_limits<unsigned long>::max()); // NOLINT(runtime/int) |
572 | | #else |
573 | 0 | return absl::Seconds(std::numeric_limits<time_t>::max()); |
574 | 0 | #endif |
575 | 0 | } |
576 | | |
577 | | // Sleeps for the given duration. |
578 | | // REQUIRES: to_sleep <= MaxSleep(). |
579 | 0 | void SleepOnce(absl::Duration to_sleep) { |
580 | | #ifdef _WIN32 |
581 | | Sleep(static_cast<DWORD>(to_sleep / absl::Milliseconds(1))); |
582 | | #else |
583 | 0 | struct timespec sleep_time = absl::ToTimespec(to_sleep); |
584 | 0 | while (nanosleep(&sleep_time, &sleep_time) != 0 && errno == EINTR) { |
585 | | // Ignore signals and wait for the full interval to elapse. |
586 | 0 | } |
587 | 0 | #endif |
588 | 0 | } |
589 | | |
590 | | } // namespace |
591 | | ABSL_NAMESPACE_END |
592 | | } // namespace absl |
593 | | |
594 | | extern "C" { |
595 | | |
596 | | ABSL_ATTRIBUTE_WEAK void ABSL_INTERNAL_C_SYMBOL(AbslInternalSleepFor)( |
597 | 0 | absl::Duration duration) { |
598 | 0 | while (duration > absl::ZeroDuration()) { |
599 | 0 | absl::Duration to_sleep = std::min(duration, absl::MaxSleep()); |
600 | 0 | absl::SleepOnce(to_sleep); |
601 | 0 | duration -= to_sleep; |
602 | 0 | } |
603 | 0 | } |
604 | | |
605 | | } // extern "C" |