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