about summary refs log tree commit diff
path: root/linuxthreads/spinlock.c
blob: c91a7cfa845fb1113ba30c08cf6d6e94faf19cbc (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
/* Linuxthreads - a simple clone()-based implementation of Posix        */
/* threads for Linux.                                                   */
/* Copyright (C) 1998 Xavier Leroy (Xavier.Leroy@inria.fr)              */
/*                                                                      */
/* This program is free software; you can redistribute it and/or        */
/* modify it under the terms of the GNU Library General Public License  */
/* as published by the Free Software Foundation; either version 2       */
/* of the License, or (at your option) any later version.               */
/*                                                                      */
/* This program is distributed in the hope that it will be useful,      */
/* but WITHOUT ANY WARRANTY; without even the implied warranty of       */
/* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the        */
/* GNU Library General Public License for more details.                 */

/* Internal locks */

#include <errno.h>
#include <sched.h>
#include <time.h>
#include <stdlib.h>
#include <limits.h>
#include "pthread.h"
#include "internals.h"
#include "spinlock.h"
#include "restart.h"

/* The status field of a spinlock has the following meaning:
     0: spinlock is free
     1: spinlock is taken, no thread is waiting on it
  ADDR: psinlock is taken, ADDR is address of thread descriptor for
        first waiting thread, other waiting threads are linked via
        their p_nextlock field.
   The waiting list is not sorted by priority order.
   Actually, we always insert at top of list (sole insertion mode
   that can be performed without locking).
   For __pthread_unlock, we perform a linear search in the list
   to find the highest-priority, oldest waiting thread.
   This is safe because there are no concurrent __pthread_unlock
   operations -- only the thread that locked the mutex can unlock it. */

void internal_function __pthread_lock(struct _pthread_fastlock * lock,
				      pthread_descr self)
{
  long oldstatus, newstatus;
  int spurious_wakeup_count = 0;

  do {
    oldstatus = lock->__status;
    if (oldstatus == 0) {
      newstatus = 1;
    } else {
      if (self == NULL)
	self = thread_self();
      newstatus = (long) self;
    }
    if (self != NULL) {
      THREAD_SETMEM(self, p_nextlock, (pthread_descr) oldstatus);
      /* Make sure the store in p_nextlock completes before performing
         the compare-and-swap */
      MEMORY_BARRIER();
    }
  } while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
                             &lock->__spinlock));

  /* Suspend with guard against spurious wakeup.
     This can happen in pthread_cond_timedwait_relative, when the thread
     wakes up due to timeout and is still on the condvar queue, and then
     locks the queue to remove itself. At that point it may still be on the
     queue, and may be resumed by a condition signal. */

  if (oldstatus != 0) {
    for (;;) {
      suspend(self);
      if (self->p_nextlock != NULL) {
	/* Count resumes that don't belong to us. */
	spurious_wakeup_count++;
	continue;
      }
      break;
    }
  }

  /* Put back any resumes we caught that don't belong to us. */
  while (spurious_wakeup_count--)
    restart(self);
}

int __pthread_unlock(struct _pthread_fastlock * lock)
{
  long oldstatus;
  pthread_descr thr, * ptr, * maxptr;
  int maxprio;

again:
  oldstatus = lock->__status;
  if (oldstatus == 0 || oldstatus == 1) {
    /* No threads are waiting for this lock.  Please note that we also
       enter this case if the lock is not taken at all.  If this wouldn't
       be done here we would crash further down.  */
    if (! compare_and_swap_with_release_semantics (&lock->__status,
						   oldstatus, 0,
						   &lock->__spinlock))
      goto again;
    return 0;
  }
  /* Find thread in waiting queue with maximal priority */
  ptr = (pthread_descr *) &lock->__status;
  thr = (pthread_descr) oldstatus;
  maxprio = 0;
  maxptr = ptr;
  while (thr != (pthread_descr) 1) {
    if (thr->p_priority >= maxprio) {
      maxptr = ptr;
      maxprio = thr->p_priority;
    }
    ptr = &(thr->p_nextlock);
    /* Prevent reordering of the load of lock->__status above and the
       load of *ptr below, as well as reordering of *ptr between
       several iterations of the while loop.  Some processors (e.g.
       multiprocessor Alphas) could perform such reordering even though
       the loads are dependent. */
    READ_MEMORY_BARRIER();
    thr = *ptr;
  }
  /* Prevent reordering of the load of lock->__status above and
     thr->p_nextlock below */
  READ_MEMORY_BARRIER();
  /* Remove max prio thread from waiting list. */
  if (maxptr == (pthread_descr *) &lock->__status) {
    /* If max prio thread is at head, remove it with compare-and-swap
       to guard against concurrent lock operation */
    thr = (pthread_descr) oldstatus;
    if (! compare_and_swap_with_release_semantics
	    (&lock->__status, oldstatus, (long)(thr->p_nextlock),
	     &lock->__spinlock))
      goto again;
  } else {
    /* No risk of concurrent access, remove max prio thread normally */
    thr = *maxptr;
    *maxptr = thr->p_nextlock;
  }
  /* Prevent reordering of store to *maxptr above and store to thr->p_nextlock
     below */
  WRITE_MEMORY_BARRIER();
  /* Wake up the selected waiting thread */
  thr->p_nextlock = NULL;
  restart(thr);

  return 0;
}

/*
 * Alternate fastlocks do not queue threads directly. Instead, they queue
 * these wait queue node structures. When a timed wait wakes up due to
 * a timeout, it can leave its wait node in the queue (because there
 * is no safe way to remove from the quue). Some other thread will
 * deallocate the abandoned node.
 */


struct wait_node {
  struct wait_node *next;	/* Next node in null terminated linked list */
  pthread_descr thr;		/* The thread waiting with this node */
  int abandoned;		/* Atomic flag */
};

static long wait_node_free_list;
static int wait_node_free_list_spinlock;

/* Allocate a new node from the head of the free list using an atomic
   operation, or else using malloc if that list is empty.  A fundamental
   assumption here is that we can safely access wait_node_free_list->next.
   That's because we never free nodes once we allocate them, so a pointer to a
   node remains valid indefinitely. */

static struct wait_node *wait_node_alloc(void)
{
  long oldvalue, newvalue;

  do {
    oldvalue = wait_node_free_list;

    if (oldvalue == 0)
      return malloc(sizeof *wait_node_alloc());

    newvalue = (long) ((struct wait_node *) oldvalue)->next;
    WRITE_MEMORY_BARRIER();
  } while (! compare_and_swap(&wait_node_free_list, oldvalue, newvalue,
                              &wait_node_free_list_spinlock));

  return (struct wait_node *) oldvalue;
}

/* Return a node to the head of the free list using an atomic
   operation. */

static void wait_node_free(struct wait_node *wn)
{
  long oldvalue, newvalue;

  do {
    oldvalue = wait_node_free_list;
    wn->next = (struct wait_node *) oldvalue;
    newvalue = (long) wn;
    WRITE_MEMORY_BARRIER();
  } while (! compare_and_swap(&wait_node_free_list, oldvalue, newvalue,
                              &wait_node_free_list_spinlock));
}

/* Remove a wait node from the specified queue.  It is assumed
   that the removal takes place concurrently with only atomic insertions at the
   head of the queue. */

static void wait_node_dequeue(struct wait_node **pp_head,
			      struct wait_node **pp_node,
			      struct wait_node *p_node,
			      int *spinlock)
{
  long oldvalue, newvalue;

  /* If the node is being deleted from the head of the
     list, it must be deleted using atomic compare-and-swap.
     Otherwise it can be deleted in the straightforward way. */

  if (pp_node == pp_head) {
    oldvalue = (long) p_node;
    newvalue = (long) p_node->next;

    if (compare_and_swap((long *) pp_node, oldvalue, newvalue, spinlock))
      return;

    /* Oops! Compare and swap failed, which means the node is
       no longer first. We delete it using the ordinary method.  But we don't
       know the identity of the node which now holds the pointer to the node
       being deleted, so we must search from the beginning. */

    for (pp_node = pp_head; *pp_node != p_node; pp_node = &(*pp_node)->next)
      ; /* null body */
  }

  *pp_node = p_node->next;
  return;
}

void __pthread_alt_lock(struct _pthread_fastlock * lock,
		        pthread_descr self)
{
  struct wait_node wait_node;
  long oldstatus, newstatus;

  do {
    oldstatus = lock->__status;
    if (oldstatus == 0) {
      newstatus = 1;
    } else {
      if (self == NULL)
	wait_node.thr = self = thread_self();
      newstatus = (long) &wait_node;
    }
    wait_node.abandoned = 0;
    wait_node.next = (struct wait_node *) oldstatus;
    /* Make sure the store in wait_node.next completes before performing
       the compare-and-swap */
    MEMORY_BARRIER();
  } while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
                             &lock->__spinlock));

  /* Suspend. Note that unlike in __pthread_lock, we don't worry
     here about spurious wakeup. That's because this lock is not
     used in situations where that can happen; the restart can
     only come from the previous lock owner. */

  if (oldstatus != 0)
    suspend(self);
}

/* Timed-out lock operation; returns 0 to indicate timeout. */

int __pthread_alt_timedlock(struct _pthread_fastlock * lock,
			    pthread_descr self, const struct timespec *abstime)
{
  struct wait_node *p_wait_node = wait_node_alloc();
  long oldstatus, newstatus;

  /* Out of memory, just give up and do ordinary lock. */
  if (p_wait_node == 0) {
    __pthread_alt_lock(lock, self);
    return 1;
  }

  do {
    oldstatus = lock->__status;
    if (oldstatus == 0) {
      newstatus = 1;
    } else {
      if (self == NULL)
	p_wait_node->thr = self = thread_self();
      newstatus = (long) p_wait_node;
    }
    p_wait_node->abandoned = 0;
    p_wait_node->next = (struct wait_node *) oldstatus;
    /* Make sure the store in wait_node.next completes before performing
       the compare-and-swap */
    MEMORY_BARRIER();
  } while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
                             &lock->__spinlock));

  /* If we did not get the lock, do a timed suspend. If we wake up due
     to a timeout, then there is a race; the old lock owner may try
     to remove us from the queue. This race is resolved by us and the owner
     doing an atomic testandset() to change the state of the wait node from 0
     to 1. If we succeed, then it's a timeout and we abandon the node in the
     queue. If we fail, it means the owner gave us the lock. */

  if (oldstatus != 0) {
    if (timedsuspend(self, abstime) == 0) {
      if (!testandset(&p_wait_node->abandoned))
	return 0; /* Timeout! */

      /* Eat oustanding resume from owner, otherwise wait_node_free() below
	 will race with owner's wait_node_dequeue(). */
      suspend(self);
    }
  }

  wait_node_free(p_wait_node);

  return 1; /* Got the lock! */
}

void __pthread_alt_unlock(struct _pthread_fastlock *lock)
{
  long oldstatus;
  struct wait_node *p_node, **pp_node, *p_max_prio, **pp_max_prio;
  struct wait_node ** const pp_head = (struct wait_node **) &lock->__status;
  int maxprio;

  while (1) {

  /* If no threads are waiting for this lock, try to just
     atomically release it. */

    oldstatus = lock->__status;
    if (oldstatus == 0 || oldstatus == 1) {
      if (compare_and_swap_with_release_semantics (&lock->__status, oldstatus,
						   0, &lock->__spinlock))
	return;
      else
	continue;
    }

    /* Process the entire queue of wait nodes. Remove all abandoned
       wait nodes and put them into the global free queue, and
       remember the one unabandoned node which refers to the thread
       having the highest priority. */

    pp_max_prio = pp_node = pp_head;
    p_max_prio = p_node = *pp_head;
    maxprio = INT_MIN;

    while (p_node != (struct wait_node *) 1) {
      int prio;

      if (p_node->abandoned) {
	/* Remove abandoned node. */
	wait_node_dequeue(pp_head, pp_node, p_node, &lock->__spinlock);
	wait_node_free(p_node);
	READ_MEMORY_BARRIER();
	p_node = *pp_node;
	continue;
      } else if ((prio = p_node->thr->p_priority) >= maxprio) {
	/* Otherwise remember it if its thread has a higher or equal priority
	   compared to that of any node seen thus far. */
	maxprio = prio;
	pp_max_prio = pp_node;
	p_max_prio = p_node;
      }

      pp_node = &p_node->next;
      READ_MEMORY_BARRIER();
      p_node = *pp_node;
    }

    READ_MEMORY_BARRIER();

    /* If all threads abandoned, go back to top */
    if (maxprio == INT_MIN)
      continue;

    ASSERT (p_max_prio != (struct wait_node *) 1);

    /* Now we want to to remove the max priority thread's wait node from
       the list. Before we can do this, we must atomically try to change the
       node's abandon state from zero to nonzero. If we succeed, that means we
       have the node that we will wake up. If we failed, then it means the
       thread timed out and abandoned the node in which case we repeat the
       whole unlock operation. */

    if (!testandset(&p_max_prio->abandoned)) {
      wait_node_dequeue(pp_head, pp_max_prio, p_max_prio, &lock->__spinlock);
      WRITE_MEMORY_BARRIER();
      restart(p_max_prio->thr);
      return;
    }
  }
}


/* Compare-and-swap emulation with a spinlock */

#ifdef TEST_FOR_COMPARE_AND_SWAP
int __pthread_has_cas = 0;
#endif

#if !defined HAS_COMPARE_AND_SWAP || defined TEST_FOR_COMPARE_AND_SWAP

static void __pthread_acquire(int * spinlock);

int __pthread_compare_and_swap(long * ptr, long oldval, long newval,
                               int * spinlock)
{
  int res;
  if (testandset(spinlock)) __pthread_acquire(spinlock);
  if (*ptr == oldval) {
    *ptr = newval; res = 1;
  } else {
    res = 0;
  }
  /* Prevent reordering of store to *ptr above and store to *spinlock below */
  WRITE_MEMORY_BARRIER();
  *spinlock = 0;
  return res;
}

/* This function is called if the inlined test-and-set
   in __pthread_compare_and_swap() failed */

/* The retry strategy is as follows:
   - We test and set the spinlock MAX_SPIN_COUNT times, calling
     sched_yield() each time.  This gives ample opportunity for other
     threads with priority >= our priority to make progress and
     release the spinlock.
   - If a thread with priority < our priority owns the spinlock,
     calling sched_yield() repeatedly is useless, since we're preventing
     the owning thread from making progress and releasing the spinlock.
     So, after MAX_SPIN_LOCK attemps, we suspend the calling thread
     using nanosleep().  This again should give time to the owning thread
     for releasing the spinlock.
     Notice that the nanosleep() interval must not be too small,
     since the kernel does busy-waiting for short intervals in a realtime
     process (!).  The smallest duration that guarantees thread
     suspension is currently 2ms.
   - When nanosleep() returns, we try again, doing MAX_SPIN_COUNT
     sched_yield(), then sleeping again if needed. */

static void __pthread_acquire(int * spinlock)
{
  int cnt = 0;
  struct timespec tm;

  while (testandset(spinlock)) {
    if (cnt < MAX_SPIN_COUNT) {
      sched_yield();
      cnt++;
    } else {
      tm.tv_sec = 0;
      tm.tv_nsec = SPIN_SLEEP_DURATION;
      nanosleep(&tm, NULL);
      cnt = 0;
    }
  }
}

#endif