source: libcfa/src/concurrency/ready_queue.cfa@ 64a7146

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast new-ast-unique-expr pthread-emulation qualifiedEnum stuck-waitfor-destruct
Last change on this file since 64a7146 was 64a7146, checked in by Thierry Delisle <tdelisle@…>, 6 years ago

Fixed idle sleep to no-longer use a spinlock, broke registration and gdbtools in the process
  • Property mode set to 100644
File size: 25.4 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#include "bits/defs.hfa"
20#include "kernel_private.hfa"
21
22#define _GNU_SOURCE
23#include "stdlib.hfa"
24#include "math.hfa"
25
26#include <unistd.h>
27
28static const size_t cache_line_size = 64;
29
30// No overriden function, no environment variable, no define
31// fall back to a magic number
32#ifndef __CFA_MAX_PROCESSORS__
33 #define __CFA_MAX_PROCESSORS__ 1024
34#endif
35
36#define BIAS 64
37
38// returns the maximum number of processors the RWLock support
39__attribute__((weak)) unsigned __max_processors() {
40 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
41 if(!max_cores_s) {
42 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
43 return __CFA_MAX_PROCESSORS__;
44 }
45
46 char * endptr = 0p;
47 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
48 if(max_cores_l < 1 || max_cores_l > 65535) {
49 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
50 return __CFA_MAX_PROCESSORS__;
51 }
52 if('\0' != *endptr) {
53 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
54 return __CFA_MAX_PROCESSORS__;
55 }
56
57 return max_cores_l;
58}
59
60//=======================================================================
61// Cluster wide reader-writer lock
62//=======================================================================
63void ?{}(__scheduler_RWLock_t & this) {
64 this.max = __max_processors();
65 this.alloc = 0;
66 this.ready = 0;
67 this.lock = false;
68 this.data = alloc(this.max);
69
70 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data )) % 64) );
71 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data + 1)) % 64) );
72 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
73 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
74
75}
76void ^?{}(__scheduler_RWLock_t & this) {
77 free(this.data);
78}
79
80void ?{}( __scheduler_lock_id_t & this, __processor_id_t * proc ) {
81 this.handle = proc;
82 this.lock = false;
83 #ifdef __CFA_WITH_VERIFY__
84 this.owned = false;
85 #endif
86}
87
88//=======================================================================
89// Lock-Free registering/unregistering of threads
90unsigned doregister( struct __processor_id_t * proc ) with(*__scheduler_lock) {
91 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
92
93 // Step - 1 : check if there is already space in the data
94 uint_fast32_t s = ready;
95
96 // Check among all the ready
97 for(uint_fast32_t i = 0; i < s; i++) {
98 __processor_id_t * null = 0p; // Re-write every loop since compare thrashes it
99 if( __atomic_load_n(&data[i].handle, (int)__ATOMIC_RELAXED) == null
100 && __atomic_compare_exchange_n( &data[i].handle, &null, proc, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
101 /*paranoid*/ verify(i < ready);
102 /*paranoid*/ verify(0 == (__alignof__(data[i]) % cache_line_size));
103 /*paranoid*/ verify((((uintptr_t)&data[i]) % cache_line_size) == 0);
104 return i;
105 }
106 }
107
108 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
109
110 // Step - 2 : F&A to get a new spot in the array.
111 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
112 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
113
114 // Step - 3 : Mark space as used and then publish it.
115 __scheduler_lock_id_t * storage = (__scheduler_lock_id_t *)&data[n];
116 (*storage){ proc };
117 while(true) {
118 unsigned copy = n;
119 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
120 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
121 break;
122 asm volatile("pause");
123 }
124
125 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
126
127 // Return new spot.
128 /*paranoid*/ verify(n < ready);
129 /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
130 /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
131 return n;
132}
133
134void unregister( struct __processor_id_t * proc ) with(*__scheduler_lock) {
135 unsigned id = proc->id;
136 /*paranoid*/ verify(id < ready);
137 /*paranoid*/ verify(proc == __atomic_load_n(&data[id].handle, __ATOMIC_RELAXED));
138 __atomic_store_n(&data[id].handle, 0p, __ATOMIC_RELEASE);
139
140 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
141}
142
143//-----------------------------------------------------------------------
144// Writer side : acquire when changing the ready queue, e.g. adding more
145// queues or removing them.
146uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
147 // Step 1 : lock global lock
148 // It is needed to avoid processors that register mid Critical-Section
149 // to simply lock their own lock and enter.
150 __atomic_acquire( &lock );
151
152 // Step 2 : lock per-proc lock
153 // Processors that are currently being registered aren't counted
154 // but can't be in read_lock or in the critical section.
155 // All other processors are counted
156 uint_fast32_t s = ready;
157 for(uint_fast32_t i = 0; i < s; i++) {
158 __atomic_acquire( &data[i].lock );
159 }
160
161 return s;
162}
163
164void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
165 // Step 1 : release local locks
166 // This must be done while the global lock is held to avoid
167 // threads that where created mid critical section
168 // to race to lock their local locks and have the writer
169 // immidiately unlock them
170 // Alternative solution : return s in write_lock and pass it to write_unlock
171 for(uint_fast32_t i = 0; i < last_s; i++) {
172 verify(data[i].lock);
173 __atomic_store_n(&data[i].lock, (bool)false, __ATOMIC_RELEASE);
174 }
175
176 // Step 2 : release global lock
177 /*paranoid*/ assert(true == lock);
178 __atomic_store_n(&lock, (bool)false, __ATOMIC_RELEASE);
179}
180
181//=======================================================================
182// Intrusive Queue used by ready queue
183//=======================================================================
184// Intrusives lanes which are used by the relaxed ready queue
185struct __attribute__((aligned(128))) __intrusive_lane_t {
186 // spin lock protecting the queue
187 volatile bool lock;
188
189 // anchor for the head and the tail of the queue
190 struct __sentinel_t {
191 // Link lists fields
192 // instrusive link field for threads
193 // must be exactly as in $thread
194 __thread_desc_link link;
195 } before, after;
196
197 // Optional statistic counters
198 #if !defined(__CFA_NO_SCHED_STATS__)
199 struct __attribute__((aligned(64))) {
200 // difference between number of push and pops
201 ssize_t diff;
202
203 // total number of pushes and pops
204 size_t push;
205 size_t pop ;
206 } stat;
207 #endif
208};
209
210void ?{}(__intrusive_lane_t & this);
211void ^?{}(__intrusive_lane_t & this);
212
213// Get the head pointer (one before the first element) from the anchor
214static inline $thread * head(const __intrusive_lane_t & this) {
215 $thread * rhead = ($thread *)(
216 (uintptr_t)( &this.before ) - offsetof( $thread, link )
217 );
218 /* paranoid */ verify(rhead);
219 return rhead;
220}
221
222// Get the tail pointer (one after the last element) from the anchor
223static inline $thread * tail(const __intrusive_lane_t & this) {
224 $thread * rtail = ($thread *)(
225 (uintptr_t)( &this.after ) - offsetof( $thread, link )
226 );
227 /* paranoid */ verify(rtail);
228 return rtail;
229}
230
231// Ctor
232void ?{}( __intrusive_lane_t & this ) {
233 this.lock = false;
234
235 this.before.link.prev = 0p;
236 this.before.link.next = tail(this);
237 this.before.link.ts = 0;
238
239 this.after .link.prev = head(this);
240 this.after .link.next = 0p;
241 this.after .link.ts = 0;
242
243 #if !defined(__CFA_NO_SCHED_STATS__)
244 this.stat.diff = 0;
245 this.stat.push = 0;
246 this.stat.pop = 0;
247 #endif
248
249 // We add a boat-load of assertions here because the anchor code is very fragile
250 /* paranoid */ verify(((uintptr_t)( head(this) ) + offsetof( $thread, link )) == (uintptr_t)(&this.before));
251 /* paranoid */ verify(((uintptr_t)( tail(this) ) + offsetof( $thread, link )) == (uintptr_t)(&this.after ));
252 /* paranoid */ verify(head(this)->link.prev == 0p );
253 /* paranoid */ verify(head(this)->link.next == tail(this) );
254 /* paranoid */ verify(tail(this)->link.next == 0p );
255 /* paranoid */ verify(tail(this)->link.prev == head(this) );
256 /* paranoid */ verify(&head(this)->link.prev == &this.before.link.prev );
257 /* paranoid */ verify(&head(this)->link.next == &this.before.link.next );
258 /* paranoid */ verify(&tail(this)->link.prev == &this.after .link.prev );
259 /* paranoid */ verify(&tail(this)->link.next == &this.after .link.next );
260 /* paranoid */ verify(sizeof(__intrusive_lane_t) == 128);
261 /* paranoid */ verify(sizeof(this) == 128);
262 /* paranoid */ verify(__alignof__(__intrusive_lane_t) == 128);
263 /* paranoid */ verify(__alignof__(this) == 128);
264 /* paranoid */ verifyf(((intptr_t)(&this) % 128) == 0, "Expected address to be aligned %p %% 128 == %zd", &this, ((intptr_t)(&this) % 128));
265}
266
267// Dtor is trivial
268void ^?{}( __intrusive_lane_t & this ) {
269 // Make sure the list is empty
270 /* paranoid */ verify(head(this)->link.prev == 0p );
271 /* paranoid */ verify(head(this)->link.next == tail(this) );
272 /* paranoid */ verify(tail(this)->link.next == 0p );
273 /* paranoid */ verify(tail(this)->link.prev == head(this) );
274}
275
276// Push a thread onto this lane
277// returns true of lane was empty before push, false otherwise
278bool push(__intrusive_lane_t & this, $thread * node) {
279 #if defined(__CFA_WITH_VERIFY__)
280 /* paranoid */ verify(this.lock);
281 /* paranoid */ verify(node->link.ts != 0);
282 /* paranoid */ verify(node->link.next == 0p);
283 /* paranoid */ verify(node->link.prev == 0p);
284 /* paranoid */ verify(tail(this)->link.next == 0p);
285 /* paranoid */ verify(head(this)->link.prev == 0p);
286
287 if(this.before.link.ts == 0l) {
288 /* paranoid */ verify(tail(this)->link.prev == head(this));
289 /* paranoid */ verify(head(this)->link.next == tail(this));
290 } else {
291 /* paranoid */ verify(tail(this)->link.prev != head(this));
292 /* paranoid */ verify(head(this)->link.next != tail(this));
293 }
294 #endif
295
296 // Get the relevant nodes locally
297 $thread * tail = tail(this);
298 $thread * prev = tail->link.prev;
299
300 // Do the push
301 node->link.next = tail;
302 node->link.prev = prev;
303 prev->link.next = node;
304 tail->link.prev = node;
305
306 // Update stats
307 #if !defined(__CFA_NO_SCHED_STATS__)
308 this.stat.diff++;
309 this.stat.push++;
310 #endif
311
312 verify(node->link.next == tail(this));
313
314 // Check if the queue used to be empty
315 if(this.before.link.ts == 0l) {
316 this.before.link.ts = node->link.ts;
317 /* paranoid */ verify(node->link.prev == head(this));
318 return true;
319 }
320 return false;
321}
322
323// Pop a thread from this lane (must be non-empty)
324// returns popped
325// returns true of lane was empty before push, false otherwise
326[$thread *, bool] pop(__intrusive_lane_t & this) {
327 /* paranoid */ verify(this.lock);
328 /* paranoid */ verify(this.before.link.ts != 0ul);
329
330 // Get anchors locally
331 $thread * head = head(this);
332 $thread * tail = tail(this);
333
334 // Get the relevant nodes locally
335 $thread * node = head->link.next;
336 $thread * next = node->link.next;
337
338 /* paranoid */ verify(node != tail);
339 /* paranoid */ verify(node);
340
341 // Do the pop
342 head->link.next = next;
343 next->link.prev = head;
344 node->link.[next, prev] = 0p;
345
346 // Update head time stamp
347 this.before.link.ts = next->link.ts;
348
349 // Update stats
350 #ifndef __CFA_NO_SCHED_STATS__
351 this.stat.diff--;
352 this.stat.pop ++;
353 #endif
354
355 // Check if we emptied list and return accordingly
356 /* paranoid */ verify(tail(this)->link.next == 0p);
357 /* paranoid */ verify(head(this)->link.prev == 0p);
358 if(next == tail) {
359 /* paranoid */ verify(this.before.link.ts == 0);
360 /* paranoid */ verify(tail(this)->link.prev == head(this));
361 /* paranoid */ verify(head(this)->link.next == tail(this));
362 return [node, true];
363 }
364 else {
365 /* paranoid */ verify(next->link.ts != 0);
366 /* paranoid */ verify(tail(this)->link.prev != head(this));
367 /* paranoid */ verify(head(this)->link.next != tail(this));
368 /* paranoid */ verify(this.before.link.ts != 0);
369 return [node, false];
370 }
371}
372
373// Check whether or not list is empty
374static inline bool is_empty(__intrusive_lane_t & this) {
375 // Cannot verify here since it may not be locked
376 return this.before.link.ts == 0;
377}
378
379// Return the timestamp
380static inline unsigned long long ts(__intrusive_lane_t & this) {
381 // Cannot verify here since it may not be locked
382 return this.before.link.ts;
383}
384
385//=======================================================================
386// Scalable Non-Zero counter
387//=======================================================================
388
389union __snzi_val_t {
390 uint64_t _all;
391 struct __attribute__((packed)) {
392 char cnt;
393 uint64_t ver:56;
394 };
395};
396
397bool cas(volatile __snzi_val_t & self, __snzi_val_t & exp, char _cnt, uint64_t _ver) {
398 __snzi_val_t t;
399 t.ver = _ver;
400 t.cnt = _cnt;
401 /* paranoid */ verify(t._all == ((_ver << 8) | ((unsigned char)_cnt)));
402 return __atomic_compare_exchange_n(&self._all, &exp._all, t._all, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
403}
404
405bool cas(volatile __snzi_val_t & self, __snzi_val_t & exp, const __snzi_val_t & tar) {
406 return __atomic_compare_exchange_n(&self._all, &exp._all, tar._all, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
407}
408
409void ?{}( __snzi_val_t & this ) { this._all = 0; }
410void ?{}( __snzi_val_t & this, const volatile __snzi_val_t & o) { this._all = o._all; }
411
412struct __attribute__((aligned(128))) __snzi_node_t {
413 volatile __snzi_val_t value;
414 struct __snzi_node_t * parent;
415 bool is_root;
416};
417
418static inline void arrive( __snzi_node_t & );
419static inline void depart( __snzi_node_t & );
420
421#define __snzi_half -1
422
423//--------------------------------------------------
424// Root node
425static void arrive_r( __snzi_node_t & this ) {
426 /* paranoid */ verify( this.is_root );
427 __atomic_fetch_add(&this.value._all, 1, __ATOMIC_SEQ_CST);
428}
429
430static void depart_r( __snzi_node_t & this ) {
431 /* paranoid */ verify( this.is_root );
432 __atomic_fetch_sub(&this.value._all, 1, __ATOMIC_SEQ_CST);
433}
434
435//--------------------------------------------------
436// Hierarchical node
437static void arrive_h( __snzi_node_t & this ) {
438 int undoArr = 0;
439 bool success = false;
440 while(!success) {
441 __snzi_val_t x = { this.value };
442 /* paranoid */ verify(x.cnt <= 120);
443 if( x.cnt >= 1 ) {
444 if( cas( this.value, x, x.cnt + 1, x.ver ) ) {
445 success = true;
446 }
447 }
448 /* paranoid */ verify(x.cnt <= 120);
449 if( x.cnt == 0 ) {
450 if( cas( this.value, x, __snzi_half, x.ver + 1) ) {
451 success = true;
452 x.cnt = __snzi_half;
453 x.ver = x.ver + 1;
454 }
455 }
456 /* paranoid */ verify(x.cnt <= 120);
457 if( x.cnt == __snzi_half ) {
458 /* paranoid */ verify( this.parent);
459 arrive( *this.parent );
460 if( !cas( this.value, x, 1, x.ver) ) {
461 undoArr = undoArr + 1;
462 }
463 }
464 }
465
466 for(int i = 0; i < undoArr; i++) {
467 /* paranoid */ verify( this.parent );
468 depart( *this.parent );
469 }
470}
471
472static void depart_h( __snzi_node_t & this ) {
473 while(true) {
474 const __snzi_val_t x = { this.value };
475 /* paranoid */ verifyf(x.cnt >= 1, "%d", x.cnt);
476 if( cas( this.value, x, x.cnt - 1, x.ver ) ) {
477 if( x.cnt == 1 ) {
478 /* paranoid */ verify( this.parent );
479 depart( *this.parent );
480 }
481 return;
482 }
483 }
484}
485
486//--------------------------------------------------
487// All nodes
488static inline void arrive( __snzi_node_t & this ) {
489 if(this.is_root) arrive_r( this );
490 else arrive_h( this );
491}
492
493static inline void depart( __snzi_node_t & this ) {
494 if(this.is_root) depart_r( this );
495 else depart_h( this );
496}
497
498static inline bool query( __snzi_node_t & this ) {
499 /* paranoid */ verify( this.is_root );
500 return this.value._all > 0;
501}
502
503//--------------------------------------------------
504// SNZI object
505void ?{}( __snzi_t & this, unsigned depth ) with( this ) {
506 mask = (1 << depth) - 1;
507 root = (1 << (depth + 1)) - 2;
508 nodes = alloc( root + 1 );
509
510 int width = 1 << depth;
511 for(int i = 0; i < root; i++) {
512 nodes[i].value._all = 0;
513 nodes[i].parent = &nodes[(i / 2) + width ];
514 nodes[i].is_root = false;
515 }
516
517 nodes[ root ].value._all = 0;
518 nodes[ root ].parent = 0p;
519 nodes[ root ].is_root = true;
520}
521
522void ^?{}( __snzi_t & this ) {
523 free( this.nodes );
524}
525
526static inline void arrive( __snzi_t & this, int idx) {
527 idx &= this.mask;
528 arrive( this.nodes[idx] );
529}
530
531static inline void depart( __snzi_t & this, int idx) {
532 idx &= this.mask;
533 depart( this.nodes[idx] );
534}
535
536static inline bool query( const __snzi_t & this ) {
537 return query( this.nodes[ this.root ] );
538}
539
540//=======================================================================
541// Cforall Reqdy Queue used by ready queue
542//=======================================================================
543
544void ?{}(__ready_queue_t & this) with (this) {
545
546 lanes.data = alloc(4);
547 for( i; 4 ) {
548 (lanes.data[i]){};
549 }
550 lanes.count = 4;
551 snzi{ log2( lanes.count / 8 ) };
552}
553
554void ^?{}(__ready_queue_t & this) with (this) {
555 verify( 4 == lanes.count );
556 verify( !query( snzi ) );
557
558 ^(snzi){};
559
560 for( i; 4 ) {
561 ^(lanes.data[i]){};
562 }
563 free(lanes.data);
564}
565
566//-----------------------------------------------------------------------
567__attribute__((hot)) bool query(struct cluster * cltr) {
568 return query(cltr->ready_queue.snzi);
569}
570
571//-----------------------------------------------------------------------
572__attribute__((hot)) bool push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
573 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
574
575 // write timestamp
576 thrd->link.ts = rdtscl();
577
578 // Try to pick a lane and lock it
579 unsigned i;
580 do {
581 // Pick the index of a lane
582 #if defined(BIAS)
583 unsigned r = __tls_rand();
584 unsigned rlow = r % BIAS;
585 unsigned rhigh = r / BIAS;
586 if(0 != (rlow % BIAS) && kernelTLS.this_processor) {
587 // (BIAS - 1) out of BIAS chances
588 // Use perferred queues
589 i = (kernelTLS.this_processor->id * 4) + (rhigh % 4);
590 }
591 else {
592 // 1 out of BIAS chances
593 // Use all queues
594 i = rhigh;
595 }
596 #else
597 i = __tls_rand();
598 #endif
599
600 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
601
602 #if !defined(__CFA_NO_STATISTICS__)
603 __tls_stats()->ready.pick.push.attempt++;
604 #endif
605
606 // If we can't lock it retry
607 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
608
609 bool first = false;
610
611 // Actually push it
612 bool lane_first = push(lanes.data[i], thrd);
613
614 // If this lane used to be empty we need to do more
615 if(lane_first) {
616 // Check if the entire queue used to be empty
617 first = !query(snzi);
618
619 // Update the snzi
620 arrive( snzi, i );
621 }
622
623 // Unlock and return
624 __atomic_unlock( &lanes.data[i].lock );
625
626 __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);
627
628 // Update statistics
629 #if !defined(__CFA_NO_STATISTICS__)
630 __tls_stats()->ready.pick.push.success++;
631 #endif
632
633 // return whether or not the list was empty before this push
634 return first;
635}
636
637//-----------------------------------------------------------------------
638// Given 2 indexes, pick the list with the oldest push an try to pop from it
639static struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j) with (cltr->ready_queue) {
640 #if !defined(__CFA_NO_STATISTICS__)
641 __tls_stats()->ready.pick.pop.attempt++;
642 #endif
643
644 // Pick the bet list
645 int w = i;
646 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
647 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
648 }
649
650 // Get relevant elements locally
651 __intrusive_lane_t & lane = lanes.data[w];
652
653 // If list looks empty retry
654 if( is_empty(lane) ) return 0p;
655
656 // If we can't get the lock retry
657 if( !__atomic_try_acquire(&lane.lock) ) return 0p;
658
659
660 // If list is empty, unlock and retry
661 if( is_empty(lane) ) {
662 __atomic_unlock(&lane.lock);
663 return 0p;
664 }
665
666 // Actually pop the list
667 struct $thread * thrd;
668 bool emptied;
669 [thrd, emptied] = pop(lane);
670
671 /* paranoid */ verify(thrd);
672 /* paranoid */ verify(lane.lock);
673
674 // If this was the last element in the lane
675 if(emptied) {
676 depart( snzi, w );
677 }
678
679 // Unlock and return
680 __atomic_unlock(&lane.lock);
681
682 // Update statistics
683 #if !defined(__CFA_NO_STATISTICS__)
684 __tls_stats()->ready.pick.pop.success++;
685 #endif
686
687 // return the popped thread
688 return thrd;
689}
690
691// Pop from the ready queue from a given cluster
692__attribute__((hot)) $thread * pop(struct cluster * cltr) with (cltr->ready_queue) {
693 /* paranoid */ verify( lanes.count > 0 );
694
695 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
696 while( query(snzi) ) {
697 // Pick two lists at random
698 #if defined(BIAS)
699 unsigned i = __tls_rand();
700 unsigned j = __tls_rand();
701
702 if(0 == (i % BIAS)) {
703 i = i / BIAS;
704 }
705 else {
706 i = ((kernelTLS.this_processor->id * 4) + ((i / BIAS) % 4));
707 j = ((kernelTLS.this_processor->id * 4) + ((j / BIAS) % 4));
708 }
709 #else
710 unsigned i = __tls_rand();
711 unsigned j = __tls_rand();
712 #endif
713
714 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
715 j %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
716
717 // try popping from the 2 picked lists
718 struct $thread * thrd = try_pop(cltr, i, j);
719 if(thrd) return thrd;
720 }
721
722 // All lanes where empty return 0p
723 return 0p;
724}
725
726//-----------------------------------------------------------------------
727
728static void check( __ready_queue_t & q ) with (q) {
729 #if defined(__CFA_WITH_VERIFY__)
730 {
731 for( idx ; lanes.count ) {
732 __intrusive_lane_t & sl = lanes.data[idx];
733 assert(!lanes.data[idx].lock);
734
735 assert(head(sl)->link.prev == 0p );
736 assert(head(sl)->link.next->link.prev == head(sl) );
737 assert(tail(sl)->link.next == 0p );
738 assert(tail(sl)->link.prev->link.next == tail(sl) );
739
740 if(sl.before.link.ts == 0l) {
741 assert(tail(sl)->link.prev == head(sl));
742 assert(head(sl)->link.next == tail(sl));
743 } else {
744 assert(tail(sl)->link.prev != head(sl));
745 assert(head(sl)->link.next != tail(sl));
746 }
747 }
748 }
749 #endif
750}
751
752// Call this function of the intrusive list was moved using memcpy
753// fixes the list so that the pointers back to anchors aren't left dangling
754static inline void fix(__intrusive_lane_t & ll) {
755 // if the list is not empty then follow he pointer and fix its reverse
756 if(!is_empty(ll)) {
757 head(ll)->link.next->link.prev = head(ll);
758 tail(ll)->link.prev->link.next = tail(ll);
759 }
760 // Otherwise just reset the list
761 else {
762 verify(tail(ll)->link.next == 0p);
763 tail(ll)->link.prev = head(ll);
764 head(ll)->link.next = tail(ll);
765 verify(head(ll)->link.prev == 0p);
766 }
767}
768
769// Grow the ready queue
770void ready_queue_grow (struct cluster * cltr) {
771 /* paranoid */ verify( ready_mutate_islocked() );
772 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
773
774 // Make sure that everything is consistent
775 /* paranoid */ check( cltr->ready_queue );
776
777 // grow the ready queue
778 with( cltr->ready_queue ) {
779 ^(snzi){};
780
781 size_t ncount = lanes.count;
782
783 // increase count
784 ncount += 4;
785
786 // Allocate new array (uses realloc and memcpies the data)
787 lanes.data = alloc(lanes.data, ncount);
788
789 // Fix the moved data
790 for( idx; (size_t)lanes.count ) {
791 fix(lanes.data[idx]);
792 }
793
794 // Construct new data
795 for( idx; (size_t)lanes.count ~ ncount) {
796 (lanes.data[idx]){};
797 }
798
799 // Update original
800 lanes.count = ncount;
801
802 // Re-create the snzi
803 snzi{ log2( lanes.count / 8 ) };
804 for( idx; (size_t)lanes.count ) {
805 if( !is_empty(lanes.data[idx]) ) {
806 arrive(snzi, idx);
807 }
808 }
809 }
810
811 // Make sure that everything is consistent
812 /* paranoid */ check( cltr->ready_queue );
813
814 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
815
816 /* paranoid */ verify( ready_mutate_islocked() );
817}
818
819// Shrink the ready queue
820void ready_queue_shrink(struct cluster * cltr) {
821 /* paranoid */ verify( ready_mutate_islocked() );
822 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
823
824 // Make sure that everything is consistent
825 /* paranoid */ check( cltr->ready_queue );
826
827 with( cltr->ready_queue ) {
828 ^(snzi){};
829
830 size_t ocount = lanes.count;
831 // Check that we have some space left
832 if(ocount < 8) abort("Program attempted to destroy more Ready Queues than were created");
833
834 // reduce the actual count so push doesn't use the old queues
835 lanes.count -= 4;
836 verify(ocount > lanes.count);
837
838 // for printing count the number of displaced threads
839 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
840 __attribute__((unused)) size_t displaced = 0;
841 #endif
842
843 // redistribute old data
844 for( idx; (size_t)lanes.count ~ ocount) {
845 // Lock is not strictly needed but makes checking invariants much easier
846 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
847 verify(locked);
848
849 // As long as we can pop from this lane to push the threads somewhere else in the queue
850 while(!is_empty(lanes.data[idx])) {
851 struct $thread * thrd;
852 __attribute__((unused)) bool _;
853 [thrd, _] = pop(lanes.data[idx]);
854
855 push(cltr, thrd);
856
857 // for printing count the number of displaced threads
858 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
859 displaced++;
860 #endif
861 }
862
863 // Unlock the lane
864 __atomic_unlock(&lanes.data[idx].lock);
865
866 // TODO print the queue statistics here
867
868 ^(lanes.data[idx]){};
869 }
870
871 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
872
873 // Allocate new array (uses realloc and memcpies the data)
874 lanes.data = alloc(lanes.data, lanes.count);
875
876 // Fix the moved data
877 for( idx; (size_t)lanes.count ) {
878 fix(lanes.data[idx]);
879 }
880
881 // Re-create the snzi
882 snzi{ log2( lanes.count / 8 ) };
883 for( idx; (size_t)lanes.count ) {
884 if( !is_empty(lanes.data[idx]) ) {
885 arrive(snzi, idx);
886 }
887 }
888 }
889
890 // Make sure that everything is consistent
891 /* paranoid */ check( cltr->ready_queue );
892
893 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
894 /* paranoid */ verify( ready_mutate_islocked() );
895}
Note: See TracBrowser for help on using the repository browser.