1/* Copyright (C) 2003-2021 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Martin Schwidefsky <schwidefsky@de.ibm.com>, 2003.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <https://www.gnu.org/licenses/>. */
18
19#include <errno.h>
20#include <sysdep.h>
21#include <futex-internal.h>
22#include <pthreadP.h>
23#include <shlib-compat.h>
24
25
26/* Wait on the barrier.
27
28 In each round, we wait for a fixed number of threads to enter the barrier
29 (COUNT). Once that has happened, exactly these threads are allowed to
30 leave the barrier. Note that POSIX does not require that only COUNT
31 threads can attempt to block using the barrier concurrently.
32
33 We count the number of threads that have entered (IN). Each thread
34 increments IN when entering, thus getting a position in the sequence of
35 threads that are or have been waiting (starting with 1, so the position
36 is the number of threads that have entered so far including the current
37 thread).
38 CURRENT_ROUND designates the most recent thread whose round has been
39 detected as complete. When a thread detects that enough threads have
40 entered to make a round complete, it finishes this round by effectively
41 adding COUNT to CURRENT_ROUND atomically. Threads that believe that their
42 round is not complete yet wait until CURRENT_ROUND is not smaller than
43 their position anymore.
44
45 A barrier can be destroyed as soon as no threads are blocked on the
46 barrier. This is already the case if just one thread from the last round
47 has stopped waiting and returned to the caller; the assumption is that
48 all threads from the round are unblocked atomically, even though they may
49 return at different times from the respective calls to
50 pthread_barrier_wait). Thus, a valid call to pthread_barrier_destroy can
51 be concurrent with other threads still figuring out that their round has
52 been completed. Therefore, threads need to confirm that they have left
53 the barrier by incrementing OUT, and pthread_barrier_destroy needs to wait
54 until OUT equals IN.
55
56 To avoid an ABA issue for futex_wait on CURRENT_ROUND and for archs with
57 32b-only atomics, we additionally reset the barrier when IN reaches
58 a threshold to avoid overflow. We assume that the total number of threads
59 is less than UINT_MAX/2, and set the threshold accordingly so that we can
60 use a simple atomic_fetch_add on IN instead of a CAS when entering. The
61 threshold is always set to the end of a round, so all threads that have
62 entered are either pre-reset threads or post-reset threads (i.e., have a
63 position larger than the threshold).
64 Pre-reset threads just run the algorithm explained above. Post-reset
65 threads wait until IN is reset to a pre-threshold value.
66 When the last pre-reset thread leaves the barrier (i.e., OUT equals the
67 threshold), it resets the barrier to its initial state. Other (post-reset)
68 threads wait for the reset to have finished by waiting until IN is less
69 than the threshold and then restart by trying to enter the barrier again.
70
71 We reuse the reset mechanism in pthread_barrier_destroy to get notified
72 when all threads have left the barrier: We trigger an artificial reset and
73 wait for the last pre-reset thread to finish reset, thus notifying the
74 thread that is about to destroy the barrier.
75
76 Blocking using futexes is straightforward: pre-reset threads wait for
77 completion of their round using CURRENT_ROUND as futex word, and post-reset
78 threads and pthread_barrier_destroy use IN as futex word.
79
80 Further notes:
81 * It is not simple to let some of the post-reset threads help with the
82 reset because of the ABA issues that arise; therefore, we simply make
83 the last thread to leave responsible for the reset.
84 * POSIX leaves it unspecified whether a signal handler running in a thread
85 that has been unblocked (because its round is complete) can stall all
86 other threads and prevent them from returning from the barrier. In this
87 implementation, other threads will return. However,
88 pthread_barrier_destroy will of course wait for the signal handler thread
89 to confirm that it left the barrier.
90
91 TODO We should add spinning with back-off. Once we do that, we could also
92 try to avoid the futex_wake syscall when a round is detected as finished.
93 If we do not spin, it is quite likely that at least some other threads will
94 have called futex_wait already. */
95int
96___pthread_barrier_wait (pthread_barrier_t *barrier)
97{
98 struct pthread_barrier *bar = (struct pthread_barrier *) barrier;
99
100 /* How many threads entered so far, including ourself. */
101 unsigned int i;
102
103 reset_restart:
104 /* Try to enter the barrier. We need acquire MO to (1) ensure that if we
105 observe that our round can be completed (see below for our attempt to do
106 so), all pre-barrier-entry effects of all threads in our round happen
107 before us completing the round, and (2) to make our use of the barrier
108 happen after a potential reset. We need release MO to make sure that our
109 pre-barrier-entry effects happen before threads in this round leaving the
110 barrier. */
111 i = atomic_fetch_add_acq_rel (&bar->in, 1) + 1;
112 /* These loads are after the fetch_add so that we're less likely to first
113 pull in the cache line as shared. */
114 unsigned int count = bar->count;
115 /* This is the number of threads that can enter before we need to reset.
116 Always at the end of a round. */
117 unsigned int max_in_before_reset = BARRIER_IN_THRESHOLD
118 - BARRIER_IN_THRESHOLD % count;
119
120 if (i > max_in_before_reset)
121 {
122 /* We're in a reset round. Just wait for a reset to finish; do not
123 help finishing previous rounds because this could happen
124 concurrently with a reset. */
125 while (i > max_in_before_reset)
126 {
127 futex_wait_simple (&bar->in, i, bar->shared);
128 /* Relaxed MO is fine here because we just need an indication for
129 when we should retry to enter (which will use acquire MO, see
130 above). */
131 i = atomic_load_relaxed (&bar->in);
132 }
133 goto reset_restart;
134 }
135
136 /* Look at the current round. At this point, we are just interested in
137 whether we can complete rounds, based on the information we obtained
138 through our acquire-MO load of IN. Nonetheless, if we notice that
139 our round has been completed using this load, we use the acquire-MO
140 fence below to make sure that all pre-barrier-entry effects of all
141 threads in our round happen before us leaving the barrier. Therefore,
142 relaxed MO is sufficient. */
143 unsigned cr = atomic_load_relaxed (&bar->current_round);
144
145 /* Try to finish previous rounds and/or the current round. We simply
146 consider just our position here and do not try to do the work of threads
147 that entered more recently. */
148 while (cr + count <= i)
149 {
150 /* Calculate the new current round based on how many threads entered.
151 NEWCR must be larger than CR because CR+COUNT ends a round. */
152 unsigned int newcr = i - i % count;
153 /* Try to complete previous and/or the current round. We need release
154 MO to propagate the happens-before that we observed through reading
155 with acquire MO from IN to other threads. If the CAS fails, it
156 is like the relaxed-MO load of CURRENT_ROUND above. */
157 if (atomic_compare_exchange_weak_release (&bar->current_round, &cr,
158 newcr))
159 {
160 /* Update CR with the modification we just did. */
161 cr = newcr;
162 /* Wake threads belonging to the rounds we just finished. We may
163 wake more threads than necessary if more than COUNT threads try
164 to block concurrently on the barrier, but this is not a typical
165 use of barriers.
166 Note that we can still access SHARED because we haven't yet
167 confirmed to have left the barrier. */
168 futex_wake (&bar->current_round, INT_MAX, bar->shared);
169 /* We did as much as we could based on our position. If we advanced
170 the current round to a round sufficient for us, do not wait for
171 that to happen and skip the acquire fence (we already
172 synchronize-with all other threads in our round through the
173 initial acquire MO fetch_add of IN. */
174 if (i <= cr)
175 goto ready_to_leave;
176 else
177 break;
178 }
179 }
180
181 /* Wait until the current round is more recent than the round we are in. */
182 while (i > cr)
183 {
184 /* Wait for the current round to finish. */
185 futex_wait_simple (&bar->current_round, cr, bar->shared);
186 /* See the fence below. */
187 cr = atomic_load_relaxed (&bar->current_round);
188 }
189
190 /* Our round finished. Use the acquire MO fence to synchronize-with the
191 thread that finished the round, either through the initial load of
192 CURRENT_ROUND above or a failed CAS in the loop above. */
193 atomic_thread_fence_acquire ();
194
195 /* Now signal that we left. */
196 unsigned int o;
197 ready_to_leave:
198 /* We need release MO here so that our use of the barrier happens before
199 reset or memory reuse after pthread_barrier_destroy. */
200 o = atomic_fetch_add_release (&bar->out, 1) + 1;
201 if (o == max_in_before_reset)
202 {
203 /* Perform a reset if we are the last pre-reset thread leaving. All
204 other threads accessing the barrier are post-reset threads and are
205 incrementing or spinning on IN. Thus, resetting IN as the last step
206 of reset ensures that the reset is not concurrent with actual use of
207 the barrier. We need the acquire MO fence so that the reset happens
208 after use of the barrier by all earlier pre-reset threads. */
209 atomic_thread_fence_acquire ();
210 atomic_store_relaxed (&bar->current_round, 0);
211 atomic_store_relaxed (&bar->out, 0);
212 /* When destroying the barrier, we wait for a reset to happen. Thus,
213 we must load SHARED now so that this happens before the barrier is
214 destroyed. */
215 int shared = bar->shared;
216 atomic_store_release (&bar->in, 0);
217 futex_wake (&bar->in, INT_MAX, shared);
218
219 }
220
221 /* Return a special value for exactly one thread per round. */
222 return i % count == 0 ? PTHREAD_BARRIER_SERIAL_THREAD : 0;
223}
224versioned_symbol (libc, ___pthread_barrier_wait, pthread_barrier_wait,
225 GLIBC_2_34);
226libc_hidden_ver (___pthread_barrier_wait, __pthread_barrier_wait)
227#ifndef SHARED
228strong_alias (___pthread_barrier_wait, __pthread_barrier_wait)
229#endif
230
231#if OTHER_SHLIB_COMPAT (libpthread, GLIBC_2_2, GLIBC_2_34)
232compat_symbol (libpthread, ___pthread_barrier_wait, pthread_barrier_wait,
233 GLIBC_2_2);
234#endif
235