• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/base/internal/spinlock.h"
16 
17 #include <algorithm>
18 #include <atomic>
19 #include <limits>
20 
21 #include "absl/base/attributes.h"
22 #include "absl/base/internal/atomic_hook.h"
23 #include "absl/base/internal/cycleclock.h"
24 #include "absl/base/internal/spinlock_wait.h"
25 #include "absl/base/internal/sysinfo.h" /* For NumCPUs() */
26 #include "absl/base/call_once.h"
27 
28 // Description of lock-word:
29 //  31..00: [............................3][2][1][0]
30 //
31 //     [0]: kSpinLockHeld
32 //     [1]: kSpinLockCooperative
33 //     [2]: kSpinLockDisabledScheduling
34 // [31..3]: ONLY kSpinLockSleeper OR
35 //          Wait time in cycles >> PROFILE_TIMESTAMP_SHIFT
36 //
37 // Detailed descriptions:
38 //
39 // Bit [0]: The lock is considered held iff kSpinLockHeld is set.
40 //
41 // Bit [1]: Eligible waiters (e.g. Fibers) may co-operatively reschedule when
42 //          contended iff kSpinLockCooperative is set.
43 //
44 // Bit [2]: This bit is exclusive from bit [1].  It is used only by a
45 //          non-cooperative lock.  When set, indicates that scheduling was
46 //          successfully disabled when the lock was acquired.  May be unset,
47 //          even if non-cooperative, if a ThreadIdentity did not yet exist at
48 //          time of acquisition.
49 //
50 // Bit [3]: If this is the only upper bit ([31..3]) set then this lock was
51 //          acquired without contention, however, at least one waiter exists.
52 //
53 //          Otherwise, bits [31..3] represent the time spent by the current lock
54 //          holder to acquire the lock.  There may be outstanding waiter(s).
55 
56 namespace absl {
57 ABSL_NAMESPACE_BEGIN
58 namespace base_internal {
59 
60 ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES static base_internal::AtomicHook<void (*)(
61     const void *lock, int64_t wait_cycles)>
62     submit_profile_data;
63 
RegisterSpinLockProfiler(void (* fn)(const void * contendedlock,int64_t wait_cycles))64 void RegisterSpinLockProfiler(void (*fn)(const void *contendedlock,
65                                          int64_t wait_cycles)) {
66   submit_profile_data.Store(fn);
67 }
68 
69 // Static member variable definitions.
70 constexpr uint32_t SpinLock::kSpinLockHeld;
71 constexpr uint32_t SpinLock::kSpinLockCooperative;
72 constexpr uint32_t SpinLock::kSpinLockDisabledScheduling;
73 constexpr uint32_t SpinLock::kSpinLockSleeper;
74 constexpr uint32_t SpinLock::kWaitTimeMask;
75 
76 // Uncommon constructors.
SpinLock(base_internal::SchedulingMode mode)77 SpinLock::SpinLock(base_internal::SchedulingMode mode)
78     : lockword_(IsCooperative(mode) ? kSpinLockCooperative : 0) {
79   ABSL_TSAN_MUTEX_CREATE(this, __tsan_mutex_not_static);
80 }
81 
82 // Monitor the lock to see if its value changes within some time period
83 // (adaptive_spin_count loop iterations). The last value read from the lock
84 // is returned from the method.
SpinLoop()85 uint32_t SpinLock::SpinLoop() {
86   // We are already in the slow path of SpinLock, initialize the
87   // adaptive_spin_count here.
88   ABSL_CONST_INIT static absl::once_flag init_adaptive_spin_count;
89   ABSL_CONST_INIT static int adaptive_spin_count = 0;
90   base_internal::LowLevelCallOnce(&init_adaptive_spin_count, []() {
91     adaptive_spin_count = base_internal::NumCPUs() > 1 ? 1000 : 1;
92   });
93 
94   int c = adaptive_spin_count;
95   uint32_t lock_value;
96   do {
97     lock_value = lockword_.load(std::memory_order_relaxed);
98   } while ((lock_value & kSpinLockHeld) != 0 && --c > 0);
99   return lock_value;
100 }
101 
SlowLock()102 void SpinLock::SlowLock() {
103   uint32_t lock_value = SpinLoop();
104   lock_value = TryLockInternal(lock_value, 0);
105   if ((lock_value & kSpinLockHeld) == 0) {
106     return;
107   }
108 
109   base_internal::SchedulingMode scheduling_mode;
110   if ((lock_value & kSpinLockCooperative) != 0) {
111     scheduling_mode = base_internal::SCHEDULE_COOPERATIVE_AND_KERNEL;
112   } else {
113     scheduling_mode = base_internal::SCHEDULE_KERNEL_ONLY;
114   }
115 
116   // The lock was not obtained initially, so this thread needs to wait for
117   // it.  Record the current timestamp in the local variable wait_start_time
118   // so the total wait time can be stored in the lockword once this thread
119   // obtains the lock.
120   int64_t wait_start_time = CycleClock::Now();
121   uint32_t wait_cycles = 0;
122   int lock_wait_call_count = 0;
123   while ((lock_value & kSpinLockHeld) != 0) {
124     // If the lock is currently held, but not marked as having a sleeper, mark
125     // it as having a sleeper.
126     if ((lock_value & kWaitTimeMask) == 0) {
127       // Here, just "mark" that the thread is going to sleep.  Don't store the
128       // lock wait time in the lock as that will cause the current lock
129       // owner to think it experienced contention.
130       if (lockword_.compare_exchange_strong(
131               lock_value, lock_value | kSpinLockSleeper,
132               std::memory_order_relaxed, std::memory_order_relaxed)) {
133         // Successfully transitioned to kSpinLockSleeper.  Pass
134         // kSpinLockSleeper to the SpinLockWait routine to properly indicate
135         // the last lock_value observed.
136         lock_value |= kSpinLockSleeper;
137       } else if ((lock_value & kSpinLockHeld) == 0) {
138         // Lock is free again, so try and acquire it before sleeping.  The
139         // new lock state will be the number of cycles this thread waited if
140         // this thread obtains the lock.
141         lock_value = TryLockInternal(lock_value, wait_cycles);
142         continue;   // Skip the delay at the end of the loop.
143       }
144     }
145 
146     // SpinLockDelay() calls into fiber scheduler, we need to see
147     // synchronization there to avoid false positives.
148     ABSL_TSAN_MUTEX_PRE_DIVERT(this, 0);
149     // Wait for an OS specific delay.
150     base_internal::SpinLockDelay(&lockword_, lock_value, ++lock_wait_call_count,
151                                  scheduling_mode);
152     ABSL_TSAN_MUTEX_POST_DIVERT(this, 0);
153     // Spin again after returning from the wait routine to give this thread
154     // some chance of obtaining the lock.
155     lock_value = SpinLoop();
156     wait_cycles = EncodeWaitCycles(wait_start_time, CycleClock::Now());
157     lock_value = TryLockInternal(lock_value, wait_cycles);
158   }
159 }
160 
SlowUnlock(uint32_t lock_value)161 void SpinLock::SlowUnlock(uint32_t lock_value) {
162   base_internal::SpinLockWake(&lockword_,
163                               false);  // wake waiter if necessary
164 
165   // If our acquisition was contended, collect contentionz profile info.  We
166   // reserve a unitary wait time to represent that a waiter exists without our
167   // own acquisition having been contended.
168   if ((lock_value & kWaitTimeMask) != kSpinLockSleeper) {
169     const uint64_t wait_cycles = DecodeWaitCycles(lock_value);
170     ABSL_TSAN_MUTEX_PRE_DIVERT(this, 0);
171     submit_profile_data(this, wait_cycles);
172     ABSL_TSAN_MUTEX_POST_DIVERT(this, 0);
173   }
174 }
175 
176 // We use the upper 29 bits of the lock word to store the time spent waiting to
177 // acquire this lock.  This is reported by contentionz profiling.  Since the
178 // lower bits of the cycle counter wrap very quickly on high-frequency
179 // processors we divide to reduce the granularity to 2^kProfileTimestampShift
180 // sized units.  On a 4Ghz machine this will lose track of wait times greater
181 // than (2^29/4 Ghz)*128 =~ 17.2 seconds.  Such waits should be extremely rare.
182 static constexpr int kProfileTimestampShift = 7;
183 
184 // We currently reserve the lower 3 bits.
185 static constexpr int kLockwordReservedShift = 3;
186 
EncodeWaitCycles(int64_t wait_start_time,int64_t wait_end_time)187 uint32_t SpinLock::EncodeWaitCycles(int64_t wait_start_time,
188                                     int64_t wait_end_time) {
189   static const int64_t kMaxWaitTime =
190       std::numeric_limits<uint32_t>::max() >> kLockwordReservedShift;
191   int64_t scaled_wait_time =
192       (wait_end_time - wait_start_time) >> kProfileTimestampShift;
193 
194   // Return a representation of the time spent waiting that can be stored in
195   // the lock word's upper bits.
196   uint32_t clamped = static_cast<uint32_t>(
197       std::min(scaled_wait_time, kMaxWaitTime) << kLockwordReservedShift);
198 
199   if (clamped == 0) {
200     return kSpinLockSleeper;  // Just wake waiters, but don't record contention.
201   }
202   // Bump up value if necessary to avoid returning kSpinLockSleeper.
203   const uint32_t kMinWaitTime =
204       kSpinLockSleeper + (1 << kLockwordReservedShift);
205   if (clamped == kSpinLockSleeper) {
206     return kMinWaitTime;
207   }
208   return clamped;
209 }
210 
DecodeWaitCycles(uint32_t lock_value)211 uint64_t SpinLock::DecodeWaitCycles(uint32_t lock_value) {
212   // Cast to uint32_t first to ensure bits [63:32] are cleared.
213   const uint64_t scaled_wait_time =
214       static_cast<uint32_t>(lock_value & kWaitTimeMask);
215   return scaled_wait_time << (kProfileTimestampShift - kLockwordReservedShift);
216 }
217 
218 }  // namespace base_internal
219 ABSL_NAMESPACE_END
220 }  // namespace absl
221