source: trunk/libs/libpthread/pthread.c@ 581

Last change on this file since 581 was 581, checked in by alain, 8 years ago

1) Improve the busylock debug infrastructure.
2) introduce a non-distributed, but portable implementation for the pthread_barrier.
File size: 16.9 KB
Line 
1/*
2 * pthread.c - User level <pthread> library implementation.
3 *
4 * Author Alain Greiner (2016,2017,2018)
5 *
6 * Copyright (c) UPMC Sorbonne Universites
7 *
8 * This file is part of ALMOS-MKH.
9 *
10 * ALMOS-MKH is free software; you can redistribute it and/or modify it
11 * under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; version 2.0 of the License.
13 *
14 * ALMOS-MKH is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with ALMOS-MKH; if not, write to the Free Software Foundation,
21 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24#include <hal_user.h>
25#include <hal_shared_types.h>
26#include <stdio.h>
27#include <stdlib.h>
28#include <pthread.h>
29#include <shared_pthread.h>
30#include <almosmkh.h>
31#include <syscalls_numbers.h>
32
33
34////////////////////////////////////////////////////////////////////////////////////////////
35// Threads
36////////////////////////////////////////////////////////////////////////////////////////////
37
38/////////////////////////////////////////////////
39int pthread_create( pthread_t * trdid,
40 const pthread_attr_t * attr,
41 void * start_func,
42 void * start_args )
43{
44 return hal_user_syscall( SYS_THREAD_CREATE,
45 (reg_t)trdid,
46 (reg_t)attr,
47 (reg_t)start_func,
48 (reg_t)start_args );
49}
50
51/////////////////////////////////////
52int pthread_join( pthread_t trdid,
53 void ** exit_value )
54{
55 return hal_user_syscall( SYS_THREAD_JOIN,
56 (reg_t)trdid,
57 (reg_t)exit_value, 0, 0 );
58}
59
60///////////////////////////////////////
61int pthread_detach( pthread_t trdid )
62{
63 return hal_user_syscall( SYS_THREAD_DETACH,
64 (reg_t)trdid, 0, 0, 0 );
65}
66
67/////////////////////////////////////
68int pthread_exit( void * exit_value )
69{
70 return hal_user_syscall( SYS_THREAD_EXIT,
71 (reg_t)exit_value, 0, 0, 0 );
72}
73
74/////////////////////////
75int pthread_yield( void )
76{
77 return hal_user_syscall( SYS_THREAD_YIELD, 0, 0, 0, 0 );
78}
79
80////////////////////////////////////////////////////////////////////////////////////////////
81// Mutexes
82////////////////////////////////////////////////////////////////////////////////////////////
83
84//////////////////////////////////////////////////////////
85int pthread_mutex_init( pthread_mutex_t * mutex,
86 const pthread_mutexattr_t * attr )
87{
88 if( attr != NULL )
89 {
90 printf("\n[ERROR] in %s : <attr> argument not supported\n", __FUNCTION__);
91 return -1;
92 }
93
94 return hal_user_syscall( SYS_MUTEX,
95 (reg_t)mutex,
96 MUTEX_INIT,
97 0, 0 );
98}
99
100////////////////////////////////////////////////////
101int pthread_mutex_destroy( pthread_mutex_t * mutex )
102{
103 return hal_user_syscall( SYS_MUTEX,
104 (reg_t)mutex,
105 MUTEX_DESTROY,
106 0, 0 );
107}
108
109/////////////////////////////////////////////////
110int pthread_mutex_lock( pthread_mutex_t * mutex )
111{
112 return hal_user_syscall( SYS_MUTEX,
113 (reg_t)mutex,
114 MUTEX_LOCK,
115 0, 0 );
116}
117
118////////////////////////////////////////////////////
119int pthread_mutex_trylock( pthread_mutex_t * mutex )
120{
121 return hal_user_syscall( SYS_MUTEX,
122 (reg_t)mutex,
123 MUTEX_TRYLOCK,
124 0, 0 );
125}
126
127///////////////////////////////////////////////////
128int pthread_mutex_unlock( pthread_mutex_t * mutex )
129{
130 return hal_user_syscall( SYS_MUTEX,
131 (reg_t)mutex,
132 MUTEX_UNLOCK,
133 0, 0 );
134}
135
136////////////////////////////////////////////////////////////////////////////////////////////
137// Condvars
138////////////////////////////////////////////////////////////////////////////////////////////
139
140///////////////////////////////////////////////
141int pthread_cond_init( pthread_cond_t * cond,
142 pthread_condattr_t * attr )
143{
144 if( attr )
145 {
146 printf("[ERROR] in %s ; <attr> argument must be NULL\n", __FUNCTION__ );
147 return -1;
148 }
149
150 return hal_user_syscall( SYS_CONDVAR,
151 (reg_t)cond,
152 CONDVAR_INIT,
153 0, 0 );
154}
155
156/////////////////////////////////////////////////
157int pthread_cond_destroy( pthread_cond_t * cond )
158{
159 return hal_user_syscall( SYS_CONDVAR,
160 (reg_t)cond,
161 CONDVAR_DESTROY,
162 0, 0 );
163}
164
165//////////////////////////////////////////////
166int pthread_cond_wait( pthread_cond_t * cond,
167 pthread_mutex_t * mutex )
168{
169 return hal_user_syscall( SYS_CONDVAR,
170 (reg_t)cond,
171 CONDVAR_WAIT,
172 (reg_t)mutex,
173 0 );
174}
175
176////////////////////////////////////////////////
177int pthread_cond_signal( pthread_cond_t * cond )
178{
179 return hal_user_syscall( SYS_CONDVAR,
180 (reg_t)cond,
181 CONDVAR_SIGNAL,
182 0, 0 );
183}
184
185///////////////////////////////////////////////////
186int pthread_cond_broadcast( pthread_cond_t * cond )
187{
188 return hal_user_syscall( SYS_CONDVAR,
189 (reg_t)cond,
190 CONDVAR_BROADCAST,
191 0, 0 );
192}
193
194
195////////////////////////////////////////////////////////////////////////////////////////////
196// Barriers
197////////////////////////////////////////////////////////////////////////////////////////////
198
199////////////////////////////////////////////////////////////////
200int pthread_barrier_init( pthread_barrier_t * barrier,
201 const pthread_barrierattr_t * attr,
202 unsigned int count )
203{
204 return hal_user_syscall( SYS_BARRIER,
205 (reg_t)barrier,
206 BARRIER_INIT,
207 (reg_t)count,
208 0 );
209}
210
211//////////////////////////////////////////////////////////
212int pthread_barrier_destroy( pthread_barrier_t * barrier )
213{
214 return hal_user_syscall( SYS_BARRIER,
215 (reg_t)barrier,
216 BARRIER_DESTROY,
217 0, 0 );
218}
219
220///////////////////////////////////////////////////////
221int pthread_barrier_wait( pthread_barrier_t * barrier )
222{
223 return hal_user_syscall( SYS_BARRIER,
224 (reg_t)barrier,
225 BARRIER_WAIT,
226 0, 0 );
227}
228
229/*
230
231////////////////////////////////////////////////////////////////////////////////////////////
232// The following functions define another implementation for the POSX barrier
233// based on a distributed quadtree implemented in user space, and relying
234// on a busy waiting policy.
235////////////////////////////////////////////////////////////////////////////////////////////
236
237
238////////////////////////////////////////////////////////////////////////////////////////////
239// This recursive function initializes the SQT nodes
240// traversing the SQT from root to bottom
241////////////////////////////////////////////////////////////////////////////////////////////
242static void sqt_barrier_build( pthread_barrier_t * barrier,
243 unsigned int x,
244 unsigned int y,
245 unsigned int level,
246 sqt_node_t * parent,
247 unsigned int x_size,
248 unsigned int y_size,
249 unsigned int nthreads )
250{
251 // get target node address
252 sqt_node_t * node = barrier->node[x][y][level];
253
254 if (level == 0 ) // terminal case
255 {
256 // initializes target node
257 node->arity = nthreads;
258 node->count = nthreads;
259 node->sense = 0;
260 node->level = 0;
261 node->parent = parent;
262 node->child[0] = NULL;
263 node->child[1] = NULL;
264 node->child[2] = NULL;
265 node->child[3] = NULL;
266
267#if PTHREAD_BARRIER_DEBUG
268printf("\n[BARRIER] %s : sqt_node[%d][%d][%d] / arity %d / desc %x\n"
269"parent %x / child0 %x / child1 %x / child2 %x / child3 %x\n",
270__FUNCTION__, x, y, level, node->arity, node, node->parent,
271node->child[0], node->child[1], node->child[2], node->child[3] );
272#endif
273
274 }
275 else // non terminal case
276 {
277 unsigned int cx[4]; // x coordinate for children
278 unsigned int cy[4]; // y coordinate for children
279 unsigned int arity = 0;
280 unsigned int i;
281
282 // the child0 coordinates are equal to the parent coordinates
283 // other children coordinates are incremented depending on the level value
284 cx[0] = x;
285 cy[0] = y;
286
287 cx[1] = x;
288 cy[1] = y + (1 << (level-1));
289
290 cx[2] = x + (1 << (level-1));
291 cy[2] = y;
292
293 cx[3] = x + (1 << (level-1));
294 cy[3] = y + (1 << (level-1));
295
296 // initializes parent node taken into account the actual number of childs
297 // child pointer is NULL if coordinates outside the mesh
298 for ( i = 0 ; i < 4 ; i++ )
299 {
300 if ( (cx[i] < x_size) && (cy[i] < y_size) )
301 {
302 node->child[i] = barrier->node[cx[i]][cy[i]][level-1];
303 arity++;
304 }
305 else node->child[i] = NULL;
306 }
307 node->arity = arity;
308 node->count = arity;
309 node->sense = 0;
310 node->level = level;
311 node->parent = parent;
312
313#if PTHREAD_BARRIER_DEBUG
314printf("\n[BARRIER] %s : sqt_node[%d][%d][%d] / arity %d / desc %x\n"
315"parent %x / child0 %x / child1 %x / child2 %x / child3 %x\n",
316__FUNCTION__, x, y, level, node->arity, node, node->parent,
317node->child[0], node->child[1], node->child[2], node->child[3] );
318#endif
319
320 // recursive calls for children nodes
321 for ( i = 0 ; i < 4 ; i++ )
322 {
323 if ( (cx[i] < x_size) && (cy[i] < y_size) )
324 sqt_barrier_build( barrier,
325 cx[i],
326 cy[i],
327 level-1,
328 node,
329 x_size,
330 y_size,
331 nthreads );
332 }
333 }
334} // end sqt_barrier_build()
335
336////////////////////////////////////////////////////////////////
337int pthread_barrier_init( pthread_barrier_t * barrier,
338 const pthread_barrierattr_t * attr,
339 unsigned int count )
340{
341 unsigned int x_size;
342 unsigned int y_size;
343 unsigned int nthreads;
344
345 if( attr != NULL )
346 {
347 x_size = attr->x_size;
348 y_size = attr->y_size;
349 nthreads = attr->nthreads;
350 }
351 else
352 {
353 x_size = 1;
354 y_size = 1;
355 nthreads = count;
356 }
357
358 // check attributes
359 if( (x_size * y_size * nthreads) != count )
360 {
361 printf("\[ERROR] in %s : count != x_size * y_size * nthreads/n", __FUNCTION__);
362 exit( EXIT_FAILURE );
363 }
364
365 // compute SQT levels
366 unsigned int levels;
367 unsigned int z = (x_size > y_size) ? x_size : y_size;
368 levels = (z < 2) ? 1 : (z < 3) ? 2 : (z < 5) ? 3 : (z < 9) ? 4 : 5;
369
370#if PTHREAD_BARRIER_DEBUG
371unsigned int side = (z < 2) ? 1 : (z < 3) ? 2 : (z < 5) ? 4 : (z < 9) ? 8 : 16;
372printf("\n[BARRIER] %s : x_size = %d / y_size = %d / levels = %d / side = %d\n",
373__FUNCTION__ , x_size , y_size , levels , side );
374#endif
375
376 // allocates memory for the SQT nodes and initializes SQT nodes pointers array
377 // the actual number of SQT nodes in a cluster(x,y) depends on (x,y):
378 // At least 1 node / at most 5 nodes
379 unsigned int x; // x coordinate for one SQT node
380 unsigned int y; // y coordinate for one SQT node
381 unsigned int l; // level for one SQT node
382 for ( x = 0 ; x < x_size ; x++ )
383 {
384 for ( y = 0 ; y < y_size ; y++ )
385 {
386 unsigned int cxy = (x<<QDT_YWIDTH) + y;
387
388 for ( l = 0 ; l < levels ; l++ )
389 {
390 if ( ( (l == 0) && ((x&0x00) == 0) && ((y&0x00) == 0) ) ||
391 ( (l == 1) && ((x&0x01) == 0) && ((y&0x01) == 0) ) ||
392 ( (l == 2) && ((x&0x03) == 0) && ((y&0x03) == 0) ) ||
393 ( (l == 3) && ((x&0x07) == 0) && ((y&0x07) == 0) ) ||
394 ( (l == 4) && ((x&0x0F) == 0) && ((y&0x0F) == 0) ) )
395 {
396 sqt_node_t * node = remote_malloc( sizeof(sqt_node_t) , cxy );
397
398 if( node == NULL )
399 {
400 printf("\n[ERROR] in %s : cannot allocate sqt_node in cluster %x\n",
401 __FUNCTION__ , cxy );
402 return -1;
403 }
404
405 barrier->node[x][y][l] = node;
406
407 }
408 }
409 }
410 }
411
412 // recursively initialize all SQT nodes from root to bottom
413 sqt_barrier_build( barrier,
414 0,
415 0,
416 levels-1,
417 NULL,
418 x_size,
419 y_size,
420 nthreads );
421
422 hal_user_fence();
423
424 return 0;
425
426} // end pthread_barrier_init
427
428//////////////////////////////////////////////////////////////////////////////////////////
429// This recursive function decrements the distributed "count" variables,
430// traversing the SQT from bottom to root.
431// The last arrived thread reset the local node before returning.
432//////////////////////////////////////////////////////////////////////////////////////////
433static void sqt_barrier_decrement( sqt_node_t * node )
434{
435
436#if PTHREAD_BARRIER_DEBUG
437unsigned int cxy;
438unsigned int lid;
439get_core( &cxy , &lid );
440printf("\n[BARRIER] %s : core[%x,%d] decrement SQT barrier node %x :\n"
441" level = %d / parent = %x / arity = %d / sense = %d / count = %d\n",
442__FUNCTION__ , cxy , lid , (unsigned int)node ,
443node->level , node->parent, node->arity , node->sense , node->count );
444#endif
445
446 unsigned int expected;
447
448 // compute expected sense value
449 if ( node->sense == 0) expected = 1;
450 else expected = 0;
451
452 // atomically decrement count
453 int count = hal_user_atomic_add( (int *)&node->count , -1 );
454
455 // last arrived thread makes the recursive call
456 if ( count == 1 ) // last thread
457 {
458 // decrement the parent node if the current node is not the root
459 if ( node->parent != NULL ) sqt_barrier_decrement( node->parent );
460
461#if PTHREAD_BARRIER_DEBUG
462printf("\n[BARRIER] %s : core[%x,%d] reset SQT barrier node %x :\n"
463" level = %d / arity = %d / sense = %d / count = %d\n",
464__FUNCTION__ , cxy , lid , (unsigned int)node ,
465node->level , node->arity , node->sense , node->count );
466#endif
467 // reset the current node
468 node->sense = expected;
469 node->count = node->arity;
470
471 return;
472 }
473 else // not the last thread
474 {
475 while( 1 )
476 {
477 // poll sense
478 if( node->sense == expected ) break;
479
480 // deschedule
481 pthread_yield();
482 }
483
484 return;
485 }
486} // end sqt_barrier_decrement()
487
488///////////////////////////////////////////////////////
489int pthread_barrier_wait( pthread_barrier_t * barrier )
490{
491 // get calling core cluster
492 unsigned int cxy;
493 unsigned int lid;
494 get_core( &cxy , &lid );
495
496 // get calling core coordinate
497 unsigned int x = cxy >> QDT_YWIDTH;
498 unsigned int y = cxy & QDT_YMASK;
499
500#if PTHREAD_BARRIER_DEBUG
501printf("\n[BARRIER] %s : core[%x,%d] enter / barrier = %x / node = %x\n",
502__FUNCTION__ , cxy , lid , barrier, barrier->node[x][y][0] );
503#endif
504
505 // recursively decrement count from bottom to root
506 sqt_barrier_decrement( barrier->node[x][y][0] );
507
508 hal_user_fence();
509
510 return 0;
511
512} // end pthread_barrier_wait()
513
514*/
515
516
517
518// Local Variables:
519// tab-width: 4
520// c-basic-offset: 4
521// c-file-offsets:((innamespace . 0)(inline-open . 0))
522// indent-tabs-mode: nil
523// End:
524// vim: filetype=c:expandtab:shiftwidth=4:tabstop=4:softtabstop=4
525
Note: See TracBrowser for help on using the repository browser.