nptl/pthread_barrier_wait.c

   1 /* Copyright (C) 2003-2020 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
  24
  25 /* Wait on the barrier.
  26
  27    In each round, we wait for a fixed number of threads to enter the barrier
  28    (COUNT).  Once that has happened, exactly these threads are allowed to
  29    leave the barrier.  Note that POSIX does not require that only COUNT
  30    threads can attempt to block using the barrier concurrently.
  31
  32    We count the number of threads that have entered (IN).  Each thread
  33    increments IN when entering, thus getting a position in the sequence of
  34    threads that are or have been waiting (starting with 1, so the position
  35    is the number of threads that have entered so far including the current
  36    thread).
  37    CURRENT_ROUND designates the most recent thread whose round has been
  38    detected as complete.  When a thread detects that enough threads have
  39    entered to make a round complete, it finishes this round by effectively
  40    adding COUNT to CURRENT_ROUND atomically.  Threads that believe that their
  41    round is not complete yet wait until CURRENT_ROUND is not smaller than
  42    their position anymore.
  43
  44    A barrier can be destroyed as soon as no threads are blocked on the
  45    barrier.  This is already the case if just one thread from the last round
  46    has stopped waiting and returned to the caller; the assumption is that
  47    all threads from the round are unblocked atomically, even though they may
  48    return at different times from the respective calls to
  49    pthread_barrier_wait).  Thus, a valid call to pthread_barrier_destroy can
  50    be concurrent with other threads still figuring out that their round has
  51    been completed.  Therefore, threads need to confirm that they have left
  52    the barrier by incrementing OUT, and pthread_barrier_destroy needs to wait
  53    until OUT equals IN.
  54
  55    To avoid an ABA issue for futex_wait on CURRENT_ROUND and for archs with
  56    32b-only atomics, we additionally reset the barrier when IN reaches
  57    a threshold to avoid overflow.  We assume that the total number of threads
  58    is less than UINT_MAX/2, and set the threshold accordingly so that we can
  59    use a simple atomic_fetch_add on IN instead of a CAS when entering.  The
  60    threshold is always set to the end of a round, so all threads that have
  61    entered are either pre-reset threads or post-reset threads (i.e., have a
  62    position larger than the threshold).
  63    Pre-reset threads just run the algorithm explained above.  Post-reset
  64    threads wait until IN is reset to a pre-threshold value.
  65    When the last pre-reset thread leaves the barrier (i.e., OUT equals the
  66    threshold), it resets the barrier to its initial state.  Other (post-reset)
  67    threads wait for the reset to have finished by waiting until IN is less
  68    than the threshold and then restart by trying to enter the barrier again.
  69
  70    We reuse the reset mechanism in pthread_barrier_destroy to get notified
  71    when all threads have left the barrier: We trigger an artificial reset and
  72    wait for the last pre-reset thread to finish reset, thus notifying the
  73    thread that is about to destroy the barrier.
  74
  75    Blocking using futexes is straightforward: pre-reset threads wait for
  76    completion of their round using CURRENT_ROUND as futex word, and post-reset
  77    threads and pthread_barrier_destroy use IN as futex word.
  78
  79    Further notes:
  80    * It is not simple to let some of the post-reset threads help with the
  81      reset because of the ABA issues that arise; therefore, we simply make
  82      the last thread to leave responsible for the reset.
  83    * POSIX leaves it unspecified whether a signal handler running in a thread
  84      that has been unblocked (because its round is complete) can stall all
  85      other threads and prevent them from returning from the barrier.  In this
  86      implementation, other threads will return.  However,
  87      pthread_barrier_destroy will of course wait for the signal handler thread
  88      to confirm that it left the barrier.
  89
  90    TODO We should add spinning with back-off.  Once we do that, we could also
  91    try to avoid the futex_wake syscall when a round is detected as finished.
  92    If we do not spin, it is quite likely that at least some other threads will
  93    have called futex_wait already.  */
  94 int
  95 __pthread_barrier_wait (pthread_barrier_t *barrier)
  96 {
  97   struct pthread_barrier *bar = (struct pthread_barrier *) barrier;
  98
  99   /* How many threads entered so far, including ourself.  */
 100   unsigned int i;
 101
 102  reset_restart:
 103   /* Try to enter the barrier.  We need acquire MO to (1) ensure that if we
 104      observe that our round can be completed (see below for our attempt to do
 105      so), all pre-barrier-entry effects of all threads in our round happen
 106      before us completing the round, and (2) to make our use of the barrier
 107      happen after a potential reset.  We need release MO to make sure that our
 108      pre-barrier-entry effects happen before threads in this round leaving the
 109      barrier.  */
 110   i = atomic_fetch_add_acq_rel (&bar->in, 1) + 1;
 111   /* These loads are after the fetch_add so that we're less likely to first
 112      pull in the cache line as shared.  */
 113   unsigned int count = bar->count;
 114   /* This is the number of threads that can enter before we need to reset.
 115      Always at the end of a round.  */
 116   unsigned int max_in_before_reset = BARRIER_IN_THRESHOLD
 117                                    - BARRIER_IN_THRESHOLD % count;
 118
 119   if (i > max_in_before_reset)
 120     {
 121       /* We're in a reset round.  Just wait for a reset to finish; do not
 122          help finishing previous rounds because this could happen
 123          concurrently with a reset.  */
 124       while (i > max_in_before_reset)
 125         {
 126           futex_wait_simple (&bar->in, i, bar->shared);
 127           /* Relaxed MO is fine here because we just need an indication for
 128              when we should retry to enter (which will use acquire MO, see
 129              above).  */
 130           i = atomic_load_relaxed (&bar->in);
 131         }
 132       goto reset_restart;
 133     }
 134
 135   /* Look at the current round.  At this point, we are just interested in
 136      whether we can complete rounds, based on the information we obtained
 137      through our acquire-MO load of IN.  Nonetheless, if we notice that
 138      our round has been completed using this load, we use the acquire-MO
 139      fence below to make sure that all pre-barrier-entry effects of all
 140      threads in our round happen before us leaving the barrier.  Therefore,
 141      relaxed MO is sufficient.  */
 142   unsigned cr = atomic_load_relaxed (&bar->current_round);
 143
 144   /* Try to finish previous rounds and/or the current round.  We simply
 145      consider just our position here and do not try to do the work of threads
 146      that entered more recently.  */
 147   while (cr + count <= i)
 148     {
 149       /* Calculate the new current round based on how many threads entered.
 150          NEWCR must be larger than CR because CR+COUNT ends a round.  */
 151       unsigned int newcr = i - i % count;
 152       /* Try to complete previous and/or the current round.  We need release
 153          MO to propagate the happens-before that we observed through reading
 154          with acquire MO from IN to other threads.  If the CAS fails, it
 155          is like the relaxed-MO load of CURRENT_ROUND above.  */
 156       if (atomic_compare_exchange_weak_release (&bar->current_round, &cr,
 157                                                 newcr))
 158         {
 159           /* Update CR with the modification we just did.  */
 160           cr = newcr;
 161           /* Wake threads belonging to the rounds we just finished.  We may
 162              wake more threads than necessary if more than COUNT threads try
 163              to block concurrently on the barrier, but this is not a typical
 164              use of barriers.
 165              Note that we can still access SHARED because we haven't yet
 166              confirmed to have left the barrier.  */
 167           futex_wake (&bar->current_round, INT_MAX, bar->shared);
 168           /* We did as much as we could based on our position.  If we advanced
 169              the current round to a round sufficient for us, do not wait for
 170              that to happen and skip the acquire fence (we already
 171              synchronize-with all other threads in our round through the
 172              initial acquire MO fetch_add of IN.  */
 173           if (i <= cr)
 174             goto ready_to_leave;
 175           else
 176             break;
 177         }
 178     }
 179
 180   /* Wait until the current round is more recent than the round we are in.  */
 181   while (i > cr)
 182     {
 183       /* Wait for the current round to finish.  */
 184       futex_wait_simple (&bar->current_round, cr, bar->shared);
 185       /* See the fence below.  */
 186       cr = atomic_load_relaxed (&bar->current_round);
 187     }
 188
 189   /* Our round finished.  Use the acquire MO fence to synchronize-with the
 190      thread that finished the round, either through the initial load of
 191      CURRENT_ROUND above or a failed CAS in the loop above.  */
 192   atomic_thread_fence_acquire ();
 193
 194   /* Now signal that we left.  */
 195   unsigned int o;
 196  ready_to_leave:
 197   /* We need release MO here so that our use of the barrier happens before
 198      reset or memory reuse after pthread_barrier_destroy.  */
 199   o = atomic_fetch_add_release (&bar->out, 1) + 1;
 200   if (o == max_in_before_reset)
 201     {
 202       /* Perform a reset if we are the last pre-reset thread leaving.   All
 203          other threads accessing the barrier are post-reset threads and are
 204          incrementing or spinning on IN.  Thus, resetting IN as the last step
 205          of reset ensures that the reset is not concurrent with actual use of
 206          the barrier.  We need the acquire MO fence so that the reset happens
 207          after use of the barrier by all earlier pre-reset threads.  */
 208       atomic_thread_fence_acquire ();
 209       atomic_store_relaxed (&bar->current_round, 0);
 210       atomic_store_relaxed (&bar->out, 0);
 211       /* When destroying the barrier, we wait for a reset to happen.  Thus,
 212          we must load SHARED now so that this happens before the barrier is
 213          destroyed.  */
 214       int shared = bar->shared;
 215       atomic_store_release (&bar->in, 0);
 216       futex_wake (&bar->in, INT_MAX, shared);
 217
 218     }
 219
 220   /* Return a special value for exactly one thread per round.  */
 221   return i % count == 0 ?  PTHREAD_BARRIER_SERIAL_THREAD : 0;
 222 }
 223 weak_alias (__pthread_barrier_wait, pthread_barrier_wait)