source: libcfa/src/concurrency/ready_queue.cfa@ fb15af5

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since fb15af5 was 32c2c5e, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

Removed unused variables
  • Property mode set to 100644
File size: 18.0 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2019 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// ready_queue.cfa --
8//
9// Author : Thierry Delisle
10// Created On : Mon Nov dd 16:29:18 2019
11// Last Modified By :
12// Last Modified On :
13// Update Count :
14//
15
16#define __cforall_thread__
17// #define __CFA_DEBUG_PRINT_READY_QUEUE__
18
19// #define USE_SNZI
20
21#include "bits/defs.hfa"
22#include "kernel_private.hfa"
23
24#define _GNU_SOURCE
25#include "stdlib.hfa"
26#include "math.hfa"
27
28#include <unistd.h>
29
30#include "snzi.hfa"
31#include "ready_subqueue.hfa"
32
33static const size_t cache_line_size = 64;
34
35// No overriden function, no environment variable, no define
36// fall back to a magic number
37#ifndef __CFA_MAX_PROCESSORS__
38 #define __CFA_MAX_PROCESSORS__ 1024
39#endif
40
41#define BIAS 16
42
43// returns the maximum number of processors the RWLock support
44__attribute__((weak)) unsigned __max_processors() {
45 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
46 if(!max_cores_s) {
47 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
48 return __CFA_MAX_PROCESSORS__;
49 }
50
51 char * endptr = 0p;
52 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
53 if(max_cores_l < 1 || max_cores_l > 65535) {
54 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
55 return __CFA_MAX_PROCESSORS__;
56 }
57 if('\0' != *endptr) {
58 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
59 return __CFA_MAX_PROCESSORS__;
60 }
61
62 return max_cores_l;
63}
64
65//=======================================================================
66// Cluster wide reader-writer lock
67//=======================================================================
68void ?{}(__scheduler_RWLock_t & this) {
69 this.max = __max_processors();
70 this.alloc = 0;
71 this.ready = 0;
72 this.lock = false;
73 this.data = alloc(this.max);
74
75 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data )) % 64) );
76 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data + 1)) % 64) );
77 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
78 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
79
80}
81void ^?{}(__scheduler_RWLock_t & this) {
82 free(this.data);
83}
84
85void ?{}( __scheduler_lock_id_t & this, __processor_id_t * proc ) {
86 this.handle = proc;
87 this.lock = false;
88 #ifdef __CFA_WITH_VERIFY__
89 this.owned = false;
90 #endif
91}
92
93//=======================================================================
94// Lock-Free registering/unregistering of threads
95unsigned doregister( struct __processor_id_t * proc ) with(*__scheduler_lock) {
96 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
97
98 // Step - 1 : check if there is already space in the data
99 uint_fast32_t s = ready;
100
101 // Check among all the ready
102 for(uint_fast32_t i = 0; i < s; i++) {
103 __processor_id_t * null = 0p; // Re-write every loop since compare thrashes it
104 if( __atomic_load_n(&data[i].handle, (int)__ATOMIC_RELAXED) == null
105 && __atomic_compare_exchange_n( &data[i].handle, &null, proc, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
106 /*paranoid*/ verify(i < ready);
107 /*paranoid*/ verify(0 == (__alignof__(data[i]) % cache_line_size));
108 /*paranoid*/ verify((((uintptr_t)&data[i]) % cache_line_size) == 0);
109 return i;
110 }
111 }
112
113 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
114
115 // Step - 2 : F&A to get a new spot in the array.
116 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
117 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
118
119 // Step - 3 : Mark space as used and then publish it.
120 __scheduler_lock_id_t * storage = (__scheduler_lock_id_t *)&data[n];
121 (*storage){ proc };
122 while() {
123 unsigned copy = n;
124 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
125 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
126 break;
127 Pause();
128 }
129
130 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
131
132 // Return new spot.
133 /*paranoid*/ verify(n < ready);
134 /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
135 /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
136 return n;
137}
138
139void unregister( struct __processor_id_t * proc ) with(*__scheduler_lock) {
140 unsigned id = proc->id;
141 /*paranoid*/ verify(id < ready);
142 /*paranoid*/ verify(proc == __atomic_load_n(&data[id].handle, __ATOMIC_RELAXED));
143 __atomic_store_n(&data[id].handle, 0p, __ATOMIC_RELEASE);
144
145 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
146}
147
148//-----------------------------------------------------------------------
149// Writer side : acquire when changing the ready queue, e.g. adding more
150// queues or removing them.
151uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
152 /* paranoid */ verify( ! __preemption_enabled() );
153
154 // Step 1 : lock global lock
155 // It is needed to avoid processors that register mid Critical-Section
156 // to simply lock their own lock and enter.
157 __atomic_acquire( &lock );
158
159 // Step 2 : lock per-proc lock
160 // Processors that are currently being registered aren't counted
161 // but can't be in read_lock or in the critical section.
162 // All other processors are counted
163 uint_fast32_t s = ready;
164 for(uint_fast32_t i = 0; i < s; i++) {
165 __atomic_acquire( &data[i].lock );
166 }
167
168 /* paranoid */ verify( ! __preemption_enabled() );
169 return s;
170}
171
172void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
173 /* paranoid */ verify( ! __preemption_enabled() );
174
175 // Step 1 : release local locks
176 // This must be done while the global lock is held to avoid
177 // threads that where created mid critical section
178 // to race to lock their local locks and have the writer
179 // immidiately unlock them
180 // Alternative solution : return s in write_lock and pass it to write_unlock
181 for(uint_fast32_t i = 0; i < last_s; i++) {
182 verify(data[i].lock);
183 __atomic_store_n(&data[i].lock, (bool)false, __ATOMIC_RELEASE);
184 }
185
186 // Step 2 : release global lock
187 /*paranoid*/ assert(true == lock);
188 __atomic_store_n(&lock, (bool)false, __ATOMIC_RELEASE);
189
190 /* paranoid */ verify( ! __preemption_enabled() );
191}
192
193//=======================================================================
194// Cforall Reqdy Queue used for scheduling
195//=======================================================================
196void ?{}(__ready_queue_t & this) with (this) {
197 lanes.data = 0p;
198 lanes.count = 0;
199}
200
201void ^?{}(__ready_queue_t & this) with (this) {
202 verify( 1 == lanes.count );
203 #ifdef USE_SNZI
204 verify( !query( snzi ) );
205 #endif
206 free(lanes.data);
207}
208
209//-----------------------------------------------------------------------
210__attribute__((hot)) bool query(struct cluster * cltr) {
211 #ifdef USE_SNZI
212 return query(cltr->ready_queue.snzi);
213 #endif
214 return true;
215}
216
217static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
218 unsigned i;
219 bool local;
220 #if defined(BIAS)
221 unsigned rlow = r % BIAS;
222 unsigned rhigh = r / BIAS;
223 if((0 != rlow) && preferred >= 0) {
224 // (BIAS - 1) out of BIAS chances
225 // Use perferred queues
226 i = preferred + (rhigh % 4);
227 local = true;
228 }
229 else {
230 // 1 out of BIAS chances
231 // Use all queues
232 i = rhigh;
233 local = false;
234 }
235 #else
236 i = r;
237 local = false;
238 #endif
239 return [i, local];
240}
241
242//-----------------------------------------------------------------------
243__attribute__((hot)) bool push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
244 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
245
246 // write timestamp
247 thrd->link.ts = rdtscl();
248
249 __attribute__((unused)) bool local;
250 __attribute__((unused)) int preferred;
251 #if defined(BIAS)
252 preferred =
253 //*
254 kernelTLS().this_processor ? kernelTLS().this_processor->id * 4 : -1;
255 /*/
256 thrd->link.preferred * 4;
257 //*/
258 #endif
259
260 // Try to pick a lane and lock it
261 unsigned i;
262 do {
263 // Pick the index of a lane
264 // unsigned r = __tls_rand();
265 unsigned r = __tls_rand_fwd();
266 [i, local] = idx_from_r(r, preferred);
267
268 #if !defined(__CFA_NO_STATISTICS__)
269 if(local) {
270 __tls_stats()->ready.pick.push.local++;
271 }
272 #endif
273
274 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
275
276 #if !defined(__CFA_NO_STATISTICS__)
277 __tls_stats()->ready.pick.push.attempt++;
278 #endif
279
280 // If we can't lock it retry
281 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
282
283 bool first = false;
284
285 // Actually push it
286 #ifdef USE_SNZI
287 bool lane_first =
288 #endif
289
290 push(lanes.data[i], thrd);
291
292 #ifdef USE_SNZI
293 // If this lane used to be empty we need to do more
294 if(lane_first) {
295 // Check if the entire queue used to be empty
296 first = !query(snzi);
297
298 // Update the snzi
299 arrive( snzi, i );
300 }
301 #endif
302
303 __tls_rand_advance_bck();
304
305 // Unlock and return
306 __atomic_unlock( &lanes.data[i].lock );
307
308 __cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);
309
310 // Update statistics
311 #if !defined(__CFA_NO_STATISTICS__)
312 #if defined(BIAS)
313 if( local ) __tls_stats()->ready.pick.push.lsuccess++;
314 #endif
315 __tls_stats()->ready.pick.push.success++;
316 #endif
317
318 // return whether or not the list was empty before this push
319 return first;
320}
321
322static struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j);
323static struct $thread * try_pop(struct cluster * cltr, unsigned i);
324
325// Pop from the ready queue from a given cluster
326__attribute__((hot)) $thread * pop(struct cluster * cltr) with (cltr->ready_queue) {
327 /* paranoid */ verify( lanes.count > 0 );
328 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
329 int preferred;
330 #if defined(BIAS)
331 // Don't bother trying locally too much
332 preferred = kernelTLS().this_processor->id * 4;
333 #endif
334
335
336 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
337 #ifdef USE_SNZI
338 while( query(snzi) ) {
339 #else
340 for(25) {
341 #endif
342 // Pick two lists at random
343 // unsigned ri = __tls_rand();
344 // unsigned rj = __tls_rand();
345 unsigned ri = __tls_rand_bck();
346 unsigned rj = __tls_rand_bck();
347
348 unsigned i, j;
349 __attribute__((unused)) bool locali, localj;
350 [i, locali] = idx_from_r(ri, preferred);
351 [j, localj] = idx_from_r(rj, preferred);
352
353 #if !defined(__CFA_NO_STATISTICS__)
354 if(locali) {
355 __tls_stats()->ready.pick.pop.local++;
356 }
357 if(localj) {
358 __tls_stats()->ready.pick.pop.local++;
359 }
360 #endif
361
362 i %= count;
363 j %= count;
364
365 // try popping from the 2 picked lists
366 struct $thread * thrd = try_pop(cltr, i, j);
367 if(thrd) {
368 #if defined(BIAS) && !defined(__CFA_NO_STATISTICS__)
369 if( locali || localj ) __tls_stats()->ready.pick.pop.lsuccess++;
370 #endif
371 return thrd;
372 }
373 }
374
375 // All lanes where empty return 0p
376 return 0p;
377}
378
379__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
380 /* paranoid */ verify( lanes.count > 0 );
381 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
382 unsigned offset = __tls_rand();
383 for(i; count) {
384 unsigned idx = (offset + i) % count;
385 struct $thread * thrd = try_pop(cltr, idx);
386 if(thrd) {
387 return thrd;
388 }
389 }
390
391 // All lanes where empty return 0p
392 return 0p;
393}
394
395
396//-----------------------------------------------------------------------
397// Given 2 indexes, pick the list with the oldest push an try to pop from it
398static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j) with (cltr->ready_queue) {
399 #if !defined(__CFA_NO_STATISTICS__)
400 __tls_stats()->ready.pick.pop.attempt++;
401 #endif
402
403 // Pick the bet list
404 int w = i;
405 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
406 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
407 }
408
409 return try_pop(cltr, w);
410}
411
412static inline struct $thread * try_pop(struct cluster * cltr, unsigned w) with (cltr->ready_queue) {
413 // Get relevant elements locally
414 __intrusive_lane_t & lane = lanes.data[w];
415
416 // If list looks empty retry
417 if( is_empty(lane) ) return 0p;
418
419 // If we can't get the lock retry
420 if( !__atomic_try_acquire(&lane.lock) ) return 0p;
421
422
423 // If list is empty, unlock and retry
424 if( is_empty(lane) ) {
425 __atomic_unlock(&lane.lock);
426 return 0p;
427 }
428
429 // Actually pop the list
430 struct $thread * thrd;
431 thrd = pop(lane);
432
433 /* paranoid */ verify(thrd);
434 /* paranoid */ verify(lane.lock);
435
436 #ifdef USE_SNZI
437 // If this was the last element in the lane
438 if(emptied) {
439 depart( snzi, w );
440 }
441 #endif
442
443 // Unlock and return
444 __atomic_unlock(&lane.lock);
445
446 // Update statistics
447 #if !defined(__CFA_NO_STATISTICS__)
448 __tls_stats()->ready.pick.pop.success++;
449 #endif
450
451 // Update the thread bias
452 thrd->link.preferred = w / 4;
453
454 // return the popped thread
455 return thrd;
456}
457//-----------------------------------------------------------------------
458
459bool remove_head(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
460 for(i; lanes.count) {
461 __intrusive_lane_t & lane = lanes.data[i];
462
463 bool removed = false;
464
465 __atomic_acquire(&lane.lock);
466 if(head(lane)->link.next == thrd) {
467 $thread * pthrd;
468 pthrd = pop(lane);
469
470 /* paranoid */ verify( pthrd == thrd );
471
472 removed = true;
473 #ifdef USE_SNZI
474 if(emptied) {
475 depart( snzi, i );
476 }
477 #endif
478 }
479 __atomic_unlock(&lane.lock);
480
481 if( removed ) return true;
482 }
483 return false;
484}
485
486//-----------------------------------------------------------------------
487
488static void check( __ready_queue_t & q ) with (q) {
489 #if defined(__CFA_WITH_VERIFY__)
490 {
491 for( idx ; lanes.count ) {
492 __intrusive_lane_t & sl = lanes.data[idx];
493 assert(!lanes.data[idx].lock);
494
495 assert(head(sl)->link.prev == 0p );
496 assert(head(sl)->link.next->link.prev == head(sl) );
497 assert(tail(sl)->link.next == 0p );
498 assert(tail(sl)->link.prev->link.next == tail(sl) );
499
500 if(sl.before.link.ts == 0l) {
501 assert(tail(sl)->link.prev == head(sl));
502 assert(head(sl)->link.next == tail(sl));
503 } else {
504 assert(tail(sl)->link.prev != head(sl));
505 assert(head(sl)->link.next != tail(sl));
506 }
507 }
508 }
509 #endif
510}
511
512// Call this function of the intrusive list was moved using memcpy
513// fixes the list so that the pointers back to anchors aren't left dangling
514static inline void fix(__intrusive_lane_t & ll) {
515 // if the list is not empty then follow he pointer and fix its reverse
516 if(!is_empty(ll)) {
517 head(ll)->link.next->link.prev = head(ll);
518 tail(ll)->link.prev->link.next = tail(ll);
519 }
520 // Otherwise just reset the list
521 else {
522 verify(tail(ll)->link.next == 0p);
523 tail(ll)->link.prev = head(ll);
524 head(ll)->link.next = tail(ll);
525 verify(head(ll)->link.prev == 0p);
526 }
527}
528
529// Grow the ready queue
530void ready_queue_grow (struct cluster * cltr, int target) {
531 /* paranoid */ verify( ready_mutate_islocked() );
532 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
533
534 // Make sure that everything is consistent
535 /* paranoid */ check( cltr->ready_queue );
536
537 // grow the ready queue
538 with( cltr->ready_queue ) {
539 #ifdef USE_SNZI
540 ^(snzi){};
541 #endif
542
543 // Find new count
544 // Make sure we always have atleast 1 list
545 size_t ncount = target >= 2 ? target * 4: 1;
546
547 // Allocate new array (uses realloc and memcpies the data)
548 lanes.data = alloc( ncount, lanes.data`realloc );
549
550 // Fix the moved data
551 for( idx; (size_t)lanes.count ) {
552 fix(lanes.data[idx]);
553 }
554
555 // Construct new data
556 for( idx; (size_t)lanes.count ~ ncount) {
557 (lanes.data[idx]){};
558 }
559
560 // Update original
561 lanes.count = ncount;
562
563 #ifdef USE_SNZI
564 // Re-create the snzi
565 snzi{ log2( lanes.count / 8 ) };
566 for( idx; (size_t)lanes.count ) {
567 if( !is_empty(lanes.data[idx]) ) {
568 arrive(snzi, idx);
569 }
570 }
571 #endif
572 }
573
574 // Make sure that everything is consistent
575 /* paranoid */ check( cltr->ready_queue );
576
577 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
578
579 /* paranoid */ verify( ready_mutate_islocked() );
580}
581
582// Shrink the ready queue
583void ready_queue_shrink(struct cluster * cltr, int target) {
584 /* paranoid */ verify( ready_mutate_islocked() );
585 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
586
587 // Make sure that everything is consistent
588 /* paranoid */ check( cltr->ready_queue );
589
590 with( cltr->ready_queue ) {
591 #ifdef USE_SNZI
592 ^(snzi){};
593 #endif
594
595 // Remember old count
596 size_t ocount = lanes.count;
597
598 // Find new count
599 // Make sure we always have atleast 1 list
600 lanes.count = target >= 2 ? target * 4: 1;
601 /* paranoid */ verify( ocount >= lanes.count );
602 /* paranoid */ verify( lanes.count == target * 4 || target < 2 );
603
604 // for printing count the number of displaced threads
605 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
606 __attribute__((unused)) size_t displaced = 0;
607 #endif
608
609 // redistribute old data
610 for( idx; (size_t)lanes.count ~ ocount) {
611 // Lock is not strictly needed but makes checking invariants much easier
612 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
613 verify(locked);
614
615 // As long as we can pop from this lane to push the threads somewhere else in the queue
616 while(!is_empty(lanes.data[idx])) {
617 struct $thread * thrd;
618 thrd = pop(lanes.data[idx]);
619
620 push(cltr, thrd);
621
622 // for printing count the number of displaced threads
623 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
624 displaced++;
625 #endif
626 }
627
628 // Unlock the lane
629 __atomic_unlock(&lanes.data[idx].lock);
630
631 // TODO print the queue statistics here
632
633 ^(lanes.data[idx]){};
634 }
635
636 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
637
638 // Allocate new array (uses realloc and memcpies the data)
639 lanes.data = alloc( lanes.count, lanes.data`realloc );
640
641 // Fix the moved data
642 for( idx; (size_t)lanes.count ) {
643 fix(lanes.data[idx]);
644 }
645
646 #ifdef USE_SNZI
647 // Re-create the snzi
648 snzi{ log2( lanes.count / 8 ) };
649 for( idx; (size_t)lanes.count ) {
650 if( !is_empty(lanes.data[idx]) ) {
651 arrive(snzi, idx);
652 }
653 }
654 #endif
655 }
656
657 // Make sure that everything is consistent
658 /* paranoid */ check( cltr->ready_queue );
659
660 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
661 /* paranoid */ verify( ready_mutate_islocked() );
662}
Note: See TracBrowser for help on using the repository browser.