linuxthreads/spinlock.c

   1 /* Linuxthreads - a simple clone()-based implementation of Posix        */
   2 /* threads for Linux.                                                   */
   3 /* Copyright (C) 1998 Xavier Leroy (Xavier.Leroy@inria.fr)              */
   4 /*                                                                      */
   5 /* This program is free software; you can redistribute it and/or        */
   6 /* modify it under the terms of the GNU Library General Public License  */
   7 /* as published by the Free Software Foundation; either version 2       */
   8 /* of the License, or (at your option) any later version.               */
   9 /*                                                                      */
  10 /* This program 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        */
  13 /* GNU Library General Public License for more details.                 */
  14
  15 /* Internal locks */
  16
  17 #include <errno.h>
  18 #include <sched.h>
  19 #include <time.h>
  20 #include "pthread.h"
  21 #include "internals.h"
  22 #include "spinlock.h"
  23 #include "restart.h"
  24
  25 /* The status field of a spinlock has the following meaning:
  26      0: spinlock is free
  27      1: spinlock is taken, no thread is waiting on it
  28   ADDR: psinlock is taken, ADDR is address of thread descriptor for
  29         first waiting thread, other waiting threads are linked via
  30         their p_nextlock field.
  31    The waiting list is not sorted by priority order.
  32    Actually, we always insert at top of list (sole insertion mode
  33    that can be performed without locking).
  34    For __pthread_unlock, we perform a linear search in the list
  35    to find the highest-priority, oldest waiting thread.
  36    This is safe because there are no concurrent __pthread_unlock
  37    operations -- only the thread that locked the mutex can unlock it. */
  38
  39 void internal_function __pthread_lock(struct _pthread_fastlock * lock,
  40                                       pthread_descr self)
  41 {
  42   long oldstatus, newstatus;
  43   int spurious_wakeup_count = 0;
  44
  45   do {
  46     oldstatus = lock->__status;
  47     if (oldstatus == 0) {
  48       newstatus = 1;
  49     } else {
  50       if (self == NULL)
  51         self = thread_self();
  52       newstatus = (long) self;
  53     }
  54     if (self != NULL) {
  55       THREAD_SETMEM(self, p_nextlock, (pthread_descr) oldstatus);
  56       /* Make sure the store in p_nextlock completes before performing
  57          the compare-and-swap */
  58       MEMORY_BARRIER();
  59     }
  60   } while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
  61                              &lock->__spinlock));
  62
  63   /* Suspend with guard against spurious wakeup.
  64      This can happen in pthread_cond_timedwait_relative, when the thread
  65      wakes up due to timeout and is still on the condvar queue, and then
  66      locks the queue to remove itself. At that point it may still be on the
  67      queue, and may be resumed by a condition signal. */
  68
  69   if (oldstatus != 0) {
  70     for (;;) {
  71       suspend(self);
  72       if (self->p_nextlock != NULL) {
  73         /* Count resumes that don't belong to us. */
  74         spurious_wakeup_count++;
  75         continue;
  76       }
  77       break;
  78     }
  79   }
  80
  81   /* Put back any resumes we caught that don't belong to us. */
  82   while (spurious_wakeup_count--)
  83     restart(self);
  84 }
  85
  86 int __pthread_unlock(struct _pthread_fastlock * lock)
  87 {
  88   long oldstatus;
  89   pthread_descr thr, * ptr, * maxptr;
  90   int maxprio;
  91
  92 again:
  93   oldstatus = lock->__status;
  94   if (oldstatus == 0 || oldstatus == 1) {
  95     /* No threads are waiting for this lock.  Please note that we also
  96        enter this case if the lock is not taken at all.  If this wouldn't
  97        be done here we would crash further down.  */
  98     if (! compare_and_swap(&lock->__status, oldstatus, 0, &lock->__spinlock))
  99       goto again;
 100     return 0;
 101   }
 102   /* Find thread in waiting queue with maximal priority */
 103   ptr = (pthread_descr *) &lock->__status;
 104   thr = (pthread_descr) oldstatus;
 105   maxprio = 0;
 106   maxptr = ptr;
 107   while (thr != (pthread_descr) 1) {
 108     if (thr->p_priority >= maxprio) {
 109       maxptr = ptr;
 110       maxprio = thr->p_priority;
 111     }
 112     ptr = &(thr->p_nextlock);
 113     /* Prevent reordering of the load of lock->__status above and the
 114        load of *ptr below, as well as reordering of *ptr between
 115        several iterations of the while loop.  Some processors (e.g.
 116        multiprocessor Alphas) could perform such reordering even though
 117        the loads are dependent. */
 118     READ_MEMORY_BARRIER();
 119     thr = *ptr;
 120   }
 121   /* Prevent reordering of the load of lock->__status above and
 122      thr->p_nextlock below */
 123   READ_MEMORY_BARRIER();
 124   /* Remove max prio thread from waiting list. */
 125   if (maxptr == (pthread_descr *) &lock->__status) {
 126     /* If max prio thread is at head, remove it with compare-and-swap
 127        to guard against concurrent lock operation */
 128     thr = (pthread_descr) oldstatus;
 129     if (! compare_and_swap(&lock->__status,
 130                            oldstatus, (long)(thr->p_nextlock),
 131                            &lock->__spinlock))
 132       goto again;
 133   } else {
 134     /* No risk of concurrent access, remove max prio thread normally */
 135     thr = *maxptr;
 136     *maxptr = thr->p_nextlock;
 137   }
 138   /* Prevent reordering of store to *maxptr above and store to thr->p_nextlock
 139      below */
 140   WRITE_MEMORY_BARRIER();
 141   /* Wake up the selected waiting thread */
 142   thr->p_nextlock = NULL;
 143   restart(thr);
 144
 145   return 0;
 146 }
 147
 148
 149 /* Compare-and-swap emulation with a spinlock */
 150
 151 #ifdef TEST_FOR_COMPARE_AND_SWAP
 152 int __pthread_has_cas = 0;
 153 #endif
 154
 155 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 156
 157 static void __pthread_acquire(int * spinlock);
 158
 159 int __pthread_compare_and_swap(long * ptr, long oldval, long newval,
 160                                int * spinlock)
 161 {
 162   int res;
 163   if (testandset(spinlock)) __pthread_acquire(spinlock);
 164   if (*ptr == oldval) {
 165     *ptr = newval; res = 1;
 166   } else {
 167     res = 0;
 168   }
 169   /* Prevent reordering of store to *ptr above and store to *spinlock below */
 170   WRITE_MEMORY_BARRIER();
 171   *spinlock = 0;
 172   return res;
 173 }
 174
 175 /* This function is called if the inlined test-and-set
 176    in __pthread_compare_and_swap() failed */
 177
 178 /* The retry strategy is as follows:
 179    - We test and set the spinlock MAX_SPIN_COUNT times, calling
 180      sched_yield() each time.  This gives ample opportunity for other
 181      threads with priority >= our priority to make progress and
 182      release the spinlock.
 183    - If a thread with priority < our priority owns the spinlock,
 184      calling sched_yield() repeatedly is useless, since we're preventing
 185      the owning thread from making progress and releasing the spinlock.
 186      So, after MAX_SPIN_LOCK attemps, we suspend the calling thread
 187      using nanosleep().  This again should give time to the owning thread
 188      for releasing the spinlock.
 189      Notice that the nanosleep() interval must not be too small,
 190      since the kernel does busy-waiting for short intervals in a realtime
 191      process (!).  The smallest duration that guarantees thread
 192      suspension is currently 2ms.
 193    - When nanosleep() returns, we try again, doing MAX_SPIN_COUNT
 194      sched_yield(), then sleeping again if needed. */
 195
 196 static void __pthread_acquire(int * spinlock)
 197 {
 198   int cnt = 0;
 199   struct timespec tm;
 200
 201   while (testandset(spinlock)) {
 202     if (cnt < MAX_SPIN_COUNT) {
 203       sched_yield();
 204       cnt++;
 205     } else {
 206       tm.tv_sec = 0;
 207       tm.tv_nsec = SPIN_SLEEP_DURATION;
 208       nanosleep(&tm, NULL);
 209       cnt = 0;
 210     }
 211   }
 212 }
 213
 214 #endif