• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 
2 //          Copyright Oliver Kowalke 2016.
3 // Distributed under the Boost Software License, Version 1.0.
4 //    (See accompanying file LICENSE_1_0.txt or copy at
5 //          http://www.boost.org/LICENSE_1_0.txt)
6 
7 #ifndef BOOST_FIBERS_SPINLOCK_TTAS_H
8 #define BOOST_FIBERS_SPINLOCK_TTAS_H
9 
10 #include <atomic>
11 #include <chrono>
12 #include <cmath>
13 #include <random>
14 #include <thread>
15 
16 #include <boost/fiber/detail/config.hpp>
17 #include <boost/fiber/detail/cpu_relax.hpp>
18 #include <boost/fiber/detail/spinlock_status.hpp>
19 
20 // based on informations from:
21 // https://software.intel.com/en-us/articles/benefitting-power-and-performance-sleep-loops
22 // https://software.intel.com/en-us/articles/long-duration-spin-wait-loops-on-hyper-threading-technology-enabled-intel-processors
23 
24 namespace boost {
25 namespace fibers {
26 namespace detail {
27 
28 class spinlock_ttas {
29 private:
30     template< typename FBSplk >
31     friend class spinlock_rtm;
32 
33     std::atomic< spinlock_status >              state_{ spinlock_status::unlocked };
34 
35 public:
36     spinlock_ttas() = default;
37 
38     spinlock_ttas( spinlock_ttas const&) = delete;
39     spinlock_ttas & operator=( spinlock_ttas const&) = delete;
40 
lock()41     void lock() noexcept {
42         static thread_local std::minstd_rand generator{ std::random_device{}() };
43         std::size_t collisions = 0 ;
44         for (;;) {
45             // avoid using multiple pause instructions for a delay of a specific cycle count
46             // the delay of cpu_relax() (pause on Intel) depends on the processor family
47             // the cycle count can not guaranteed from one system to the next
48             // -> check the shared variable 'state_' in between each cpu_relax() to prevent
49             //    unnecessarily long delays on some systems
50             std::size_t retries = 0;
51             // test shared variable 'status_'
52             // first access to 'state_' -> chache miss
53             // sucessive acccess to 'state_' -> cache hit
54             // if 'state_' was released by other fiber
55             // cached 'state_' is invalidated -> cache miss
56             while ( spinlock_status::locked == state_.load( std::memory_order_relaxed) ) {
57 #if !defined(BOOST_FIBERS_SPIN_SINGLE_CORE)
58                 if ( BOOST_FIBERS_SPIN_BEFORE_SLEEP0 > retries) {
59                     ++retries;
60                     // give CPU a hint that this thread is in a "spin-wait" loop
61                     // delays the next instruction's execution for a finite period of time (depends on processor family)
62                     // the CPU is not under demand, parts of the pipeline are no longer being used
63                     // -> reduces the power consumed by the CPU
64                     // -> prevent pipeline stalls
65                     cpu_relax();
66                 } else if ( BOOST_FIBERS_SPIN_BEFORE_YIELD > retries) {
67                     // std::this_thread::sleep_for( 0us) has a fairly long instruction path length,
68                     // combined with an expensive ring3 to ring 0 transition costing about 1000 cycles
69                     // std::this_thread::sleep_for( 0us) lets give up this_thread the remaining part of its time slice
70                     // if and only if a thread of equal or greater priority is ready to run
71                     static constexpr std::chrono::microseconds us0{ 0 };
72                     std::this_thread::sleep_for( us0);
73                 } else {
74                     // std::this_thread::yield() allows this_thread to give up the remaining part of its time slice,
75                     // but only to another thread on the same processor
76                     // instead of constant checking, a thread only checks if no other useful work is pending
77                     std::this_thread::yield();
78                 }
79 #else
80                 std::this_thread::yield();
81 #endif
82             }
83             // test-and-set shared variable 'status_'
84             // everytime 'status_' is signaled over the bus, even if the test failes
85             if ( spinlock_status::locked == state_.exchange( spinlock_status::locked, std::memory_order_acquire) ) {
86                 // spinlock now contended
87                 // utilize 'Binary Exponential Backoff' algorithm
88                 // linear_congruential_engine is a random number engine based on Linear congruential generator (LCG)
89                 std::uniform_int_distribution< std::size_t > distribution{
90                     0, static_cast< std::size_t >( 1) << (std::min)(collisions, static_cast< std::size_t >( BOOST_FIBERS_CONTENTION_WINDOW_THRESHOLD)) };
91                 const std::size_t z = distribution( generator);
92                 ++collisions;
93                 for ( std::size_t i = 0; i < z; ++i) {
94                     // -> reduces the power consumed by the CPU
95                     // -> prevent pipeline stalls
96                     cpu_relax();
97                 }
98             } else {
99                 // success, thread has acquired the lock
100                 break;
101             }
102         }
103     }
104 
try_lock()105     bool try_lock() noexcept {
106         return spinlock_status::unlocked == state_.exchange( spinlock_status::locked, std::memory_order_acquire);
107     }
108 
unlock()109     void unlock() noexcept {
110         state_.store( spinlock_status::unlocked, std::memory_order_release);
111     }
112 };
113 
114 }}}
115 
116 #endif // BOOST_FIBERS_SPINLOCK_TTAS_H
117