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 <stdlib.h>
  21 #include <limits.h>
  22 #include "pthread.h"
  23 #include "internals.h"
  24 #include "spinlock.h"
  25 #include "restart.h"
  26
  27 static void __pthread_acquire(int * spinlock);
  28
  29 static inline void __pthread_release(int * spinlock)
  30 {
  31   WRITE_MEMORY_BARRIER();
  32   *spinlock = __LT_SPINLOCK_INIT;
  33   __asm __volatile ("" : "=m" (*spinlock) : "0" (*spinlock));
  34 }
  35
  36
  37 /* The status field of a spinlock is a pointer whose least significant
  38    bit is a locked flag.
  39
  40    Thus the field values have the following meanings:
  41
  42    status == 0:       spinlock is free
  43    status == 1:       spinlock is taken; no thread is waiting on it
  44
  45    (status & 1) == 1: spinlock is taken and (status & ~1L) is a
  46                       pointer to the first waiting thread; other
  47                       waiting threads are linked via the p_nextlock
  48                       field.
  49    (status & 1) == 0: same as above, but spinlock is not taken.
  50
  51    The waiting list is not sorted by priority order.
  52    Actually, we always insert at top of list (sole insertion mode
  53    that can be performed without locking).
  54    For __pthread_unlock, we perform a linear search in the list
  55    to find the highest-priority, oldest waiting thread.
  56    This is safe because there are no concurrent __pthread_unlock
  57    operations -- only the thread that locked the mutex can unlock it. */
  58
  59
  60 void internal_function __pthread_lock(struct _pthread_fastlock * lock,
  61                                       pthread_descr self)
  62 {
  63 #if defined HAS_COMPARE_AND_SWAP
  64   long oldstatus, newstatus;
  65   int successful_seizure, spurious_wakeup_count;
  66   int spin_count;
  67 #endif
  68
  69 #if defined TEST_FOR_COMPARE_AND_SWAP
  70   if (!__pthread_has_cas)
  71 #endif
  72 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
  73   {
  74     __pthread_acquire(&lock->__spinlock);
  75     return;
  76   }
  77 #endif
  78
  79 #if defined HAS_COMPARE_AND_SWAP
  80   /* First try it without preparation.  Maybe it's a completely
  81      uncontested lock.  */
  82   if (lock->__status == 0 && __compare_and_swap (&lock->__status, 0, 1))
  83     return;
  84
  85   spurious_wakeup_count = 0;
  86   spin_count = 0;
  87
  88 again:
  89
  90   /* On SMP, try spinning to get the lock. */
  91
  92   if (__pthread_smp_kernel) {
  93     int max_count = lock->__spinlock * 2 + 10;
  94
  95     if (max_count > MAX_ADAPTIVE_SPIN_COUNT)
  96       max_count = MAX_ADAPTIVE_SPIN_COUNT;
  97
  98     for (spin_count = 0; spin_count < max_count; spin_count++) {
  99       if (((oldstatus = lock->__status) & 1) == 0) {
 100         if(__compare_and_swap(&lock->__status, oldstatus, oldstatus | 1))
 101         {
 102           if (spin_count)
 103             lock->__spinlock += (spin_count - lock->__spinlock) / 8;
 104           READ_MEMORY_BARRIER();
 105           return;
 106         }
 107       }
 108 #ifdef BUSY_WAIT_NOP
 109       BUSY_WAIT_NOP;
 110 #endif
 111       __asm __volatile ("" : "=m" (lock->__status) : "0" (lock->__status));
 112     }
 113
 114     lock->__spinlock += (spin_count - lock->__spinlock) / 8;
 115   }
 116
 117   /* No luck, try once more or suspend. */
 118
 119   do {
 120     oldstatus = lock->__status;
 121     successful_seizure = 0;
 122
 123     if ((oldstatus & 1) == 0) {
 124       newstatus = oldstatus | 1;
 125       successful_seizure = 1;
 126     } else {
 127       if (self == NULL)
 128         self = thread_self();
 129       newstatus = (long) self | 1;
 130     }
 131
 132     if (self != NULL) {
 133       THREAD_SETMEM(self, p_nextlock, (pthread_descr) (oldstatus & ~1L));
 134       /* Make sure the store in p_nextlock completes before performing
 135          the compare-and-swap */
 136       MEMORY_BARRIER();
 137     }
 138   } while(! __compare_and_swap(&lock->__status, oldstatus, newstatus));
 139
 140   /* Suspend with guard against spurious wakeup.
 141      This can happen in pthread_cond_timedwait_relative, when the thread
 142      wakes up due to timeout and is still on the condvar queue, and then
 143      locks the queue to remove itself. At that point it may still be on the
 144      queue, and may be resumed by a condition signal. */
 145
 146   if (!successful_seizure) {
 147     for (;;) {
 148       suspend(self);
 149       if (self->p_nextlock != NULL) {
 150         /* Count resumes that don't belong to us. */
 151         spurious_wakeup_count++;
 152         continue;
 153       }
 154       break;
 155     }
 156     goto again;
 157   }
 158
 159   /* Put back any resumes we caught that don't belong to us. */
 160   while (spurious_wakeup_count--)
 161     restart(self);
 162
 163   READ_MEMORY_BARRIER();
 164 #endif
 165 }
 166
 167 int __pthread_unlock(struct _pthread_fastlock * lock)
 168 {
 169 #if defined HAS_COMPARE_AND_SWAP
 170   long oldstatus;
 171   pthread_descr thr, * ptr, * maxptr;
 172   int maxprio;
 173 #endif
 174
 175 #if defined TEST_FOR_COMPARE_AND_SWAP
 176   if (!__pthread_has_cas)
 177 #endif
 178 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 179   {
 180     __pthread_release(&lock->__spinlock);
 181     return 0;
 182   }
 183 #endif
 184
 185 #if defined HAS_COMPARE_AND_SWAP
 186   WRITE_MEMORY_BARRIER();
 187
 188 again:
 189   while ((oldstatus = lock->__status) == 1) {
 190     if (__compare_and_swap_with_release_semantics(&lock->__status,
 191         oldstatus, 0))
 192       return 0;
 193   }
 194
 195   /* Find thread in waiting queue with maximal priority */
 196   ptr = (pthread_descr *) &lock->__status;
 197   thr = (pthread_descr) (oldstatus & ~1L);
 198   maxprio = 0;
 199   maxptr = ptr;
 200
 201   /* Before we iterate over the wait queue, we need to execute
 202      a read barrier, otherwise we may read stale contents of nodes that may
 203      just have been inserted by other processors. One read barrier is enough to
 204      ensure we have a stable list; we don't need one for each pointer chase
 205      through the list, because we are the owner of the lock; other threads
 206      can only add nodes at the front; if a front node is consistent,
 207      the ones behind it must also be. */
 208
 209   READ_MEMORY_BARRIER();
 210
 211   while (thr != 0) {
 212     if (thr->p_priority >= maxprio) {
 213       maxptr = ptr;
 214       maxprio = thr->p_priority;
 215     }
 216     ptr = &(thr->p_nextlock);
 217     thr = *ptr;
 218   }
 219
 220   /* Remove max prio thread from waiting list. */
 221   if (maxptr == (pthread_descr *) &lock->__status) {
 222     /* If max prio thread is at head, remove it with compare-and-swap
 223        to guard against concurrent lock operation. This removal
 224        also has the side effect of marking the lock as released
 225        because the new status comes from thr->p_nextlock whose
 226        least significant bit is clear. */
 227     thr = (pthread_descr) (oldstatus & ~1L);
 228     if (! __compare_and_swap_with_release_semantics
 229             (&lock->__status, oldstatus, (long)(thr->p_nextlock)))
 230       goto again;
 231   } else {
 232     /* No risk of concurrent access, remove max prio thread normally.
 233        But in this case we must also flip the least significant bit
 234        of the status to mark the lock as released. */
 235     thr = *maxptr;
 236     *maxptr = thr->p_nextlock;
 237
 238     /* Ensure deletion from linked list completes before we
 239        release the lock. */
 240     WRITE_MEMORY_BARRIER();
 241
 242     do {
 243       oldstatus = lock->__status;
 244     } while (!__compare_and_swap_with_release_semantics(&lock->__status,
 245              oldstatus, oldstatus & ~1L));
 246   }
 247
 248   /* Wake up the selected waiting thread. Woken thread can check
 249      its own p_nextlock field for NULL to detect that it has been removed. No
 250      barrier is needed here, since restart() and suspend() take
 251      care of memory synchronization. */
 252
 253   thr->p_nextlock = NULL;
 254   restart(thr);
 255
 256   return 0;
 257 #endif
 258 }
 259
 260 /*
 261  * Alternate fastlocks do not queue threads directly. Instead, they queue
 262  * these wait queue node structures. When a timed wait wakes up due to
 263  * a timeout, it can leave its wait node in the queue (because there
 264  * is no safe way to remove from the quue). Some other thread will
 265  * deallocate the abandoned node.
 266  */
 267
 268
 269 struct wait_node {
 270   struct wait_node *next;       /* Next node in null terminated linked list */
 271   pthread_descr thr;            /* The thread waiting with this node */
 272   int abandoned;                /* Atomic flag */
 273 };
 274
 275 static long wait_node_free_list;
 276 static int wait_node_free_list_spinlock;
 277
 278 /* Allocate a new node from the head of the free list using an atomic
 279    operation, or else using malloc if that list is empty.  A fundamental
 280    assumption here is that we can safely access wait_node_free_list->next.
 281    That's because we never free nodes once we allocate them, so a pointer to a
 282    node remains valid indefinitely. */
 283
 284 static struct wait_node *wait_node_alloc(void)
 285 {
 286     struct wait_node *new_node = 0;
 287
 288     __pthread_acquire(&wait_node_free_list_spinlock);
 289     if (wait_node_free_list != 0) {
 290       new_node = (struct wait_node *) wait_node_free_list;
 291       wait_node_free_list = (long) new_node->next;
 292     }
 293     WRITE_MEMORY_BARRIER();
 294     __pthread_release(&wait_node_free_list_spinlock);
 295
 296     if (new_node == 0)
 297       return malloc(sizeof *wait_node_alloc());
 298
 299     return new_node;
 300 }
 301
 302 /* Return a node to the head of the free list using an atomic
 303    operation. */
 304
 305 static void wait_node_free(struct wait_node *wn)
 306 {
 307     __pthread_acquire(&wait_node_free_list_spinlock);
 308     wn->next = (struct wait_node *) wait_node_free_list;
 309     wait_node_free_list = (long) wn;
 310     WRITE_MEMORY_BARRIER();
 311     __pthread_release(&wait_node_free_list_spinlock);
 312     return;
 313 }
 314
 315 #if defined HAS_COMPARE_AND_SWAP
 316
 317 /* Remove a wait node from the specified queue.  It is assumed
 318    that the removal takes place concurrently with only atomic insertions at the
 319    head of the queue. */
 320
 321 static void wait_node_dequeue(struct wait_node **pp_head,
 322                               struct wait_node **pp_node,
 323                               struct wait_node *p_node)
 324 {
 325   /* If the node is being deleted from the head of the
 326      list, it must be deleted using atomic compare-and-swap.
 327      Otherwise it can be deleted in the straightforward way. */
 328
 329   if (pp_node == pp_head) {
 330     /* We don't need a read barrier between these next two loads,
 331        because it is assumed that the caller has already ensured
 332        the stability of *p_node with respect to p_node. */
 333
 334     long oldvalue = (long) p_node;
 335     long newvalue = (long) p_node->next;
 336
 337     if (__compare_and_swap((long *) pp_node, oldvalue, newvalue))
 338       return;
 339
 340     /* Oops! Compare and swap failed, which means the node is
 341        no longer first. We delete it using the ordinary method.  But we don't
 342        know the identity of the node which now holds the pointer to the node
 343        being deleted, so we must search from the beginning. */
 344
 345     for (pp_node = pp_head; p_node != *pp_node; ) {
 346       pp_node = &(*pp_node)->next;
 347       READ_MEMORY_BARRIER(); /* Stabilize *pp_node for next iteration. */
 348     }
 349   }
 350
 351   *pp_node = p_node->next;
 352   return;
 353 }
 354
 355 #endif
 356
 357 void __pthread_alt_lock(struct _pthread_fastlock * lock,
 358                         pthread_descr self)
 359 {
 360 #if defined HAS_COMPARE_AND_SWAP
 361   long oldstatus, newstatus;
 362 #endif
 363   struct wait_node wait_node;
 364
 365 #if defined TEST_FOR_COMPARE_AND_SWAP
 366   if (!__pthread_has_cas)
 367 #endif
 368 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 369   {
 370     int suspend_needed = 0;
 371     __pthread_acquire(&lock->__spinlock);
 372
 373     if (lock->__status == 0)
 374       lock->__status = 1;
 375     else {
 376       if (self == NULL)
 377         self = thread_self();
 378
 379       wait_node.abandoned = 0;
 380       wait_node.next = (struct wait_node *) lock->__status;
 381       wait_node.thr = self;
 382       lock->__status = (long) &wait_node;
 383       suspend_needed = 1;
 384     }
 385
 386     __pthread_release(&lock->__spinlock);
 387
 388     if (suspend_needed)
 389       suspend (self);
 390     return;
 391   }
 392 #endif
 393
 394 #if defined HAS_COMPARE_AND_SWAP
 395   do {
 396     oldstatus = lock->__status;
 397     if (oldstatus == 0) {
 398       newstatus = 1;
 399     } else {
 400       if (self == NULL)
 401         self = thread_self();
 402       wait_node.thr = self;
 403       newstatus = (long) &wait_node;
 404     }
 405     wait_node.abandoned = 0;
 406     wait_node.next = (struct wait_node *) oldstatus;
 407     /* Make sure the store in wait_node.next completes before performing
 408        the compare-and-swap */
 409     MEMORY_BARRIER();
 410   } while(! __compare_and_swap(&lock->__status, oldstatus, newstatus));
 411
 412   /* Suspend. Note that unlike in __pthread_lock, we don't worry
 413      here about spurious wakeup. That's because this lock is not
 414      used in situations where that can happen; the restart can
 415      only come from the previous lock owner. */
 416
 417   if (oldstatus != 0)
 418     suspend(self);
 419
 420   READ_MEMORY_BARRIER();
 421 #endif
 422 }
 423
 424 /* Timed-out lock operation; returns 0 to indicate timeout. */
 425
 426 int __pthread_alt_timedlock(struct _pthread_fastlock * lock,
 427                             pthread_descr self, const struct timespec *abstime)
 428 {
 429   long oldstatus = 0;
 430 #if defined HAS_COMPARE_AND_SWAP
 431   long newstatus;
 432 #endif
 433   struct wait_node *p_wait_node = wait_node_alloc();
 434
 435   /* Out of memory, just give up and do ordinary lock. */
 436   if (p_wait_node == 0) {
 437     __pthread_alt_lock(lock, self);
 438     return 1;
 439   }
 440
 441 #if defined TEST_FOR_COMPARE_AND_SWAP
 442   if (!__pthread_has_cas)
 443 #endif
 444 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 445   {
 446     __pthread_acquire(&lock->__spinlock);
 447
 448     if (lock->__status == 0)
 449       lock->__status = 1;
 450     else {
 451       if (self == NULL)
 452         self = thread_self();
 453
 454       p_wait_node->abandoned = 0;
 455       p_wait_node->next = (struct wait_node *) lock->__status;
 456       p_wait_node->thr = self;
 457       lock->__status = (long) p_wait_node;
 458       oldstatus = 1; /* force suspend */
 459     }
 460
 461     __pthread_release(&lock->__spinlock);
 462     goto suspend;
 463   }
 464 #endif
 465
 466 #if defined HAS_COMPARE_AND_SWAP
 467   do {
 468     oldstatus = lock->__status;
 469     if (oldstatus == 0) {
 470       newstatus = 1;
 471     } else {
 472       if (self == NULL)
 473         self = thread_self();
 474       p_wait_node->thr = self;
 475       newstatus = (long) p_wait_node;
 476     }
 477     p_wait_node->abandoned = 0;
 478     p_wait_node->next = (struct wait_node *) oldstatus;
 479     /* Make sure the store in wait_node.next completes before performing
 480        the compare-and-swap */
 481     MEMORY_BARRIER();
 482   } while(! __compare_and_swap(&lock->__status, oldstatus, newstatus));
 483 #endif
 484
 485 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 486   suspend:
 487 #endif
 488
 489   /* If we did not get the lock, do a timed suspend. If we wake up due
 490      to a timeout, then there is a race; the old lock owner may try
 491      to remove us from the queue. This race is resolved by us and the owner
 492      doing an atomic testandset() to change the state of the wait node from 0
 493      to 1. If we succeed, then it's a timeout and we abandon the node in the
 494      queue. If we fail, it means the owner gave us the lock. */
 495
 496   if (oldstatus != 0) {
 497     if (timedsuspend(self, abstime) == 0) {
 498       if (!testandset(&p_wait_node->abandoned))
 499         return 0; /* Timeout! */
 500
 501       /* Eat oustanding resume from owner, otherwise wait_node_free() below
 502          will race with owner's wait_node_dequeue(). */
 503       suspend(self);
 504     }
 505   }
 506
 507   wait_node_free(p_wait_node);
 508
 509   READ_MEMORY_BARRIER();
 510
 511   return 1; /* Got the lock! */
 512 }
 513
 514 void __pthread_alt_unlock(struct _pthread_fastlock *lock)
 515 {
 516   struct wait_node *p_node, **pp_node, *p_max_prio, **pp_max_prio;
 517   struct wait_node ** const pp_head = (struct wait_node **) &lock->__status;
 518   int maxprio;
 519
 520   WRITE_MEMORY_BARRIER();
 521
 522 #if defined TEST_FOR_COMPARE_AND_SWAP
 523   if (!__pthread_has_cas)
 524 #endif
 525 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 526   {
 527     __pthread_acquire(&lock->__spinlock);
 528   }
 529 #endif
 530
 531   while (1) {
 532
 533   /* If no threads are waiting for this lock, try to just
 534      atomically release it. */
 535 #if defined TEST_FOR_COMPARE_AND_SWAP
 536     if (!__pthread_has_cas)
 537 #endif
 538 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 539     {
 540       if (lock->__status == 0 || lock->__status == 1) {
 541         lock->__status = 0;
 542         break;
 543       }
 544     }
 545 #endif
 546
 547 #if defined TEST_FOR_COMPARE_AND_SWAP
 548     else
 549 #endif
 550
 551 #if defined HAS_COMPARE_AND_SWAP
 552     {
 553       long oldstatus = lock->__status;
 554       if (oldstatus == 0 || oldstatus == 1) {
 555         if (__compare_and_swap_with_release_semantics (&lock->__status, oldstatus, 0))
 556           break;
 557         else
 558           continue;
 559       }
 560     }
 561 #endif
 562
 563     /* Process the entire queue of wait nodes. Remove all abandoned
 564        wait nodes and put them into the global free queue, and
 565        remember the one unabandoned node which refers to the thread
 566        having the highest priority. */
 567
 568     pp_max_prio = pp_node = pp_head;
 569     p_max_prio = p_node = *pp_head;
 570     maxprio = INT_MIN;
 571
 572     READ_MEMORY_BARRIER(); /* Prevent access to stale data through p_node */
 573
 574     while (p_node != (struct wait_node *) 1) {
 575       int prio;
 576
 577       if (p_node->abandoned) {
 578         /* Remove abandoned node. */
 579 #if defined TEST_FOR_COMPARE_AND_SWAP
 580         if (!__pthread_has_cas)
 581 #endif
 582 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 583           *pp_node = p_node->next;
 584 #endif
 585 #if defined TEST_FOR_COMPARE_AND_SWAP
 586         else
 587 #endif
 588 #if defined HAS_COMPARE_AND_SWAP
 589           wait_node_dequeue(pp_head, pp_node, p_node);
 590 #endif
 591         wait_node_free(p_node);
 592         /* Note that the next assignment may take us to the beginning
 593            of the queue, to newly inserted nodes, if pp_node == pp_head.
 594            In that case we need a memory barrier to stabilize the first of
 595            these new nodes. */
 596         p_node = *pp_node;
 597         if (pp_node == pp_head)
 598           READ_MEMORY_BARRIER(); /* No stale reads through p_node */
 599         continue;
 600       } else if ((prio = p_node->thr->p_priority) >= maxprio) {
 601         /* Otherwise remember it if its thread has a higher or equal priority
 602            compared to that of any node seen thus far. */
 603         maxprio = prio;
 604         pp_max_prio = pp_node;
 605         p_max_prio = p_node;
 606       }
 607
 608       /* This canno6 jump backward in the list, so no further read
 609          barrier is needed. */
 610       pp_node = &p_node->next;
 611       p_node = *pp_node;
 612     }
 613
 614     /* If all threads abandoned, go back to top */
 615     if (maxprio == INT_MIN)
 616       continue;
 617
 618     ASSERT (p_max_prio != (struct wait_node *) 1);
 619
 620     /* Now we want to to remove the max priority thread's wait node from
 621        the list. Before we can do this, we must atomically try to change the
 622        node's abandon state from zero to nonzero. If we succeed, that means we
 623        have the node that we will wake up. If we failed, then it means the
 624        thread timed out and abandoned the node in which case we repeat the
 625        whole unlock operation. */
 626
 627     if (!testandset(&p_max_prio->abandoned)) {
 628 #if defined TEST_FOR_COMPARE_AND_SWAP
 629       if (!__pthread_has_cas)
 630 #endif
 631 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 632         *pp_max_prio = p_max_prio->next;
 633 #endif
 634 #if defined TEST_FOR_COMPARE_AND_SWAP
 635       else
 636 #endif
 637 #if defined HAS_COMPARE_AND_SWAP
 638         wait_node_dequeue(pp_head, pp_max_prio, p_max_prio);
 639 #endif
 640       restart(p_max_prio->thr);
 641       break;
 642     }
 643   }
 644
 645 #if defined TEST_FOR_COMPARE_AND_SWAP
 646   if (!__pthread_has_cas)
 647 #endif
 648 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 649   {
 650     __pthread_release(&lock->__spinlock);
 651   }
 652 #endif
 653 }
 654
 655
 656 /* Compare-and-swap emulation with a spinlock */
 657
 658 #ifdef TEST_FOR_COMPARE_AND_SWAP
 659 int __pthread_has_cas = 0;
 660 #endif
 661
 662 #if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP
 663
 664 int __pthread_compare_and_swap(long * ptr, long oldval, long newval,
 665                                int * spinlock)
 666 {
 667   int res;
 668
 669   __pthread_acquire(spinlock);
 670
 671   if (*ptr == oldval) {
 672     *ptr = newval; res = 1;
 673   } else {
 674     res = 0;
 675   }
 676
 677   __pthread_release(spinlock);
 678
 679   return res;
 680 }
 681
 682 #endif
 683
 684 /* The retry strategy is as follows:
 685    - We test and set the spinlock MAX_SPIN_COUNT times, calling
 686      sched_yield() each time.  This gives ample opportunity for other
 687      threads with priority >= our priority to make progress and
 688      release the spinlock.
 689    - If a thread with priority < our priority owns the spinlock,
 690      calling sched_yield() repeatedly is useless, since we're preventing
 691      the owning thread from making progress and releasing the spinlock.
 692      So, after MAX_SPIN_LOCK attemps, we suspend the calling thread
 693      using nanosleep().  This again should give time to the owning thread
 694      for releasing the spinlock.
 695      Notice that the nanosleep() interval must not be too small,
 696      since the kernel does busy-waiting for short intervals in a realtime
 697      process (!).  The smallest duration that guarantees thread
 698      suspension is currently 2ms.
 699    - When nanosleep() returns, we try again, doing MAX_SPIN_COUNT
 700      sched_yield(), then sleeping again if needed. */
 701
 702 static void __pthread_acquire(int * spinlock)
 703 {
 704   int cnt = 0;
 705   struct timespec tm;
 706
 707   READ_MEMORY_BARRIER();
 708
 709   while (testandset(spinlock)) {
 710     if (cnt < MAX_SPIN_COUNT) {
 711       sched_yield();
 712       cnt++;
 713     } else {
 714       tm.tv_sec = 0;
 715       tm.tv_nsec = SPIN_SLEEP_DURATION;
 716       nanosleep(&tm, NULL);
 717       cnt = 0;
 718     }
 719   }
 720 }