source: libcfa/src/concurrency/ready_queue.cfa@ 431cd4f

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum stuck-waitfor-destruct
Last change on this file since 431cd4f was 431cd4f, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

Added alternative to relaxed-fifo scheduler.
Disabled by default
  • Property mode set to 100644
File size: 21.6 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_MPSC
20
21#define USE_RELAXED_FIFO
22// #define USE_WORK_STEALING
23
24#include "bits/defs.hfa"
25#include "kernel_private.hfa"
26
27#define _GNU_SOURCE
28#include "stdlib.hfa"
29#include "math.hfa"
30
31#include <unistd.h>
32
33#include "ready_subqueue.hfa"
34
35static const size_t cache_line_size = 64;
36
37// No overriden function, no environment variable, no define
38// fall back to a magic number
39#ifndef __CFA_MAX_PROCESSORS__
40 #define __CFA_MAX_PROCESSORS__ 1024
41#endif
42
43#if defined(USE_RELAXED_FIFO)
44 #define BIAS 4
45 #define READYQ_SHARD_FACTOR 4
46#elif defined(USE_WORK_STEALING)
47 #define READYQ_SHARD_FACTOR 2
48#else
49 #error no scheduling strategy selected
50#endif
51
52static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
53static inline struct $thread * try_pop(struct cluster * cltr, unsigned w);
54static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j);
55static inline struct $thread * search(struct cluster * cltr);
56
57
58// returns the maximum number of processors the RWLock support
59__attribute__((weak)) unsigned __max_processors() {
60 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
61 if(!max_cores_s) {
62 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
63 return __CFA_MAX_PROCESSORS__;
64 }
65
66 char * endptr = 0p;
67 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
68 if(max_cores_l < 1 || max_cores_l > 65535) {
69 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
70 return __CFA_MAX_PROCESSORS__;
71 }
72 if('\0' != *endptr) {
73 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
74 return __CFA_MAX_PROCESSORS__;
75 }
76
77 return max_cores_l;
78}
79
80//=======================================================================
81// Cluster wide reader-writer lock
82//=======================================================================
83void ?{}(__scheduler_RWLock_t & this) {
84 this.max = __max_processors();
85 this.alloc = 0;
86 this.ready = 0;
87 this.lock = false;
88 this.data = alloc(this.max);
89
90 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data )) % 64) );
91 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data + 1)) % 64) );
92 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
93 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
94
95}
96void ^?{}(__scheduler_RWLock_t & this) {
97 free(this.data);
98}
99
100void ?{}( __scheduler_lock_id_t & this, __processor_id_t * proc ) {
101 this.handle = proc;
102 this.lock = false;
103 #ifdef __CFA_WITH_VERIFY__
104 this.owned = false;
105 #endif
106}
107
108//=======================================================================
109// Lock-Free registering/unregistering of threads
110void register_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
111 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
112
113 // Step - 1 : check if there is already space in the data
114 uint_fast32_t s = ready;
115
116 // Check among all the ready
117 for(uint_fast32_t i = 0; i < s; i++) {
118 __processor_id_t * null = 0p; // Re-write every loop since compare thrashes it
119 if( __atomic_load_n(&data[i].handle, (int)__ATOMIC_RELAXED) == null
120 && __atomic_compare_exchange_n( &data[i].handle, &null, proc, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
121 /*paranoid*/ verify(i < ready);
122 /*paranoid*/ verify(0 == (__alignof__(data[i]) % cache_line_size));
123 /*paranoid*/ verify((((uintptr_t)&data[i]) % cache_line_size) == 0);
124 proc->id = i;
125 }
126 }
127
128 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
129
130 // Step - 2 : F&A to get a new spot in the array.
131 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
132 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
133
134 // Step - 3 : Mark space as used and then publish it.
135 __scheduler_lock_id_t * storage = (__scheduler_lock_id_t *)&data[n];
136 (*storage){ proc };
137 while() {
138 unsigned copy = n;
139 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
140 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
141 break;
142 Pause();
143 }
144
145 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
146
147 // Return new spot.
148 /*paranoid*/ verify(n < ready);
149 /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
150 /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
151 proc->id = n;
152}
153
154void unregister_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
155 unsigned id = proc->id;
156 /*paranoid*/ verify(id < ready);
157 /*paranoid*/ verify(proc == __atomic_load_n(&data[id].handle, __ATOMIC_RELAXED));
158 __atomic_store_n(&data[id].handle, 0p, __ATOMIC_RELEASE);
159
160 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
161}
162
163//-----------------------------------------------------------------------
164// Writer side : acquire when changing the ready queue, e.g. adding more
165// queues or removing them.
166uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
167 /* paranoid */ verify( ! __preemption_enabled() );
168
169 // Step 1 : lock global lock
170 // It is needed to avoid processors that register mid Critical-Section
171 // to simply lock their own lock and enter.
172 __atomic_acquire( &lock );
173
174 // Step 2 : lock per-proc lock
175 // Processors that are currently being registered aren't counted
176 // but can't be in read_lock or in the critical section.
177 // All other processors are counted
178 uint_fast32_t s = ready;
179 for(uint_fast32_t i = 0; i < s; i++) {
180 __atomic_acquire( &data[i].lock );
181 }
182
183 /* paranoid */ verify( ! __preemption_enabled() );
184 return s;
185}
186
187void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
188 /* paranoid */ verify( ! __preemption_enabled() );
189
190 // Step 1 : release local locks
191 // This must be done while the global lock is held to avoid
192 // threads that where created mid critical section
193 // to race to lock their local locks and have the writer
194 // immidiately unlock them
195 // Alternative solution : return s in write_lock and pass it to write_unlock
196 for(uint_fast32_t i = 0; i < last_s; i++) {
197 verify(data[i].lock);
198 __atomic_store_n(&data[i].lock, (bool)false, __ATOMIC_RELEASE);
199 }
200
201 // Step 2 : release global lock
202 /*paranoid*/ assert(true == lock);
203 __atomic_store_n(&lock, (bool)false, __ATOMIC_RELEASE);
204
205 /* paranoid */ verify( ! __preemption_enabled() );
206}
207
208//=======================================================================
209// Cforall Ready Queue used for scheduling
210//=======================================================================
211void ?{}(__ready_queue_t & this) with (this) {
212 lanes.data = 0p;
213 lanes.tscs = 0p;
214 lanes.count = 0;
215}
216
217void ^?{}(__ready_queue_t & this) with (this) {
218 verify( 1 == lanes.count );
219 free(lanes.data);
220 free(lanes.tscs);
221}
222
223//-----------------------------------------------------------------------
224#if defined(USE_RELAXED_FIFO)
225 //-----------------------------------------------------------------------
226 // get index from random number with or without bias towards queues
227 static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
228 unsigned i;
229 bool local;
230 unsigned rlow = r % BIAS;
231 unsigned rhigh = r / BIAS;
232 if((0 != rlow) && preferred >= 0) {
233 // (BIAS - 1) out of BIAS chances
234 // Use perferred queues
235 i = preferred + (rhigh % READYQ_SHARD_FACTOR);
236 local = true;
237 }
238 else {
239 // 1 out of BIAS chances
240 // Use all queues
241 i = rhigh;
242 local = false;
243 }
244 return [i, local];
245 }
246
247 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
248 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
249
250 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
251 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
252
253 // write timestamp
254 thrd->link.ts = rdtscl();
255
256 bool local;
257 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
258
259 // Try to pick a lane and lock it
260 unsigned i;
261 do {
262 // Pick the index of a lane
263 unsigned r = __tls_rand_fwd();
264 [i, local] = idx_from_r(r, preferred);
265
266 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
267
268 #if !defined(__CFA_NO_STATISTICS__)
269 if(external) {
270 if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.local, 1, __ATOMIC_RELAXED);
271 __atomic_fetch_add(&cltr->stats->ready.pick.ext.attempt, 1, __ATOMIC_RELAXED);
272 }
273 else {
274 if(local) __tls_stats()->ready.pick.push.local++;
275 __tls_stats()->ready.pick.push.attempt++;
276 }
277 #endif
278
279 #if defined(USE_MPSC)
280 // mpsc always succeeds
281 } while( false );
282 #else
283 // If we can't lock it retry
284 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
285 #endif
286
287 // Actually push it
288 push(lanes.data[i], thrd);
289
290 #if !defined(USE_MPSC)
291 // Unlock and return
292 __atomic_unlock( &lanes.data[i].lock );
293 #endif
294
295 // Mark the current index in the tls rng instance as having an item
296 __tls_rand_advance_bck();
297
298 __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);
299
300 // Update statistics
301 #if !defined(__CFA_NO_STATISTICS__)
302 if(external) {
303 if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.lsuccess, 1, __ATOMIC_RELAXED);
304 __atomic_fetch_add(&cltr->stats->ready.pick.ext.success, 1, __ATOMIC_RELAXED);
305 }
306 else {
307 if(local) __tls_stats()->ready.pick.push.lsuccess++;
308 __tls_stats()->ready.pick.push.success++;
309 }
310 #endif
311 }
312
313 // Pop from the ready queue from a given cluster
314 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
315 /* paranoid */ verify( lanes.count > 0 );
316 /* paranoid */ verify( kernelTLS().this_processor );
317 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
318
319 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
320 int preferred = kernelTLS().this_processor->rdq.id;
321
322
323 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
324 for(25) {
325 // Pick two lists at random
326 unsigned ri = __tls_rand_bck();
327 unsigned rj = __tls_rand_bck();
328
329 unsigned i, j;
330 __attribute__((unused)) bool locali, localj;
331 [i, locali] = idx_from_r(ri, preferred);
332 [j, localj] = idx_from_r(rj, preferred);
333
334 #if !defined(__CFA_NO_STATISTICS__)
335 if(locali && localj) {
336 __tls_stats()->ready.pick.pop.local++;
337 }
338 #endif
339
340 i %= count;
341 j %= count;
342
343 // try popping from the 2 picked lists
344 struct $thread * thrd = try_pop(cltr, i, j);
345 if(thrd) {
346 #if !defined(__CFA_NO_STATISTICS__)
347 if( locali || localj ) __tls_stats()->ready.pick.pop.lsuccess++;
348 #endif
349 return thrd;
350 }
351 }
352
353 // All lanes where empty return 0p
354 return 0p;
355 }
356
357 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) {
358 return search(cltr);
359 }
360#endif
361#if defined(USE_WORK_STEALING)
362 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
363 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
364
365 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
366 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
367
368 // write timestamp
369 thrd->link.ts = rdtscl();
370
371 // Try to pick a lane and lock it
372 unsigned i;
373 do {
374 if(unlikely(external)) {
375 i = __tls_rand() % lanes.count;
376 }
377 else {
378 processor * proc = kernelTLS().this_processor;
379 unsigned r = proc->rdq.its++;
380 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
381 }
382
383
384 #if defined(USE_MPSC)
385 // mpsc always succeeds
386 } while( false );
387 #else
388 // If we can't lock it retry
389 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
390 #endif
391
392 // Actually push it
393 push(lanes.data[i], thrd);
394
395 #if !defined(USE_MPSC)
396 // Unlock and return
397 __atomic_unlock( &lanes.data[i].lock );
398 #endif
399
400 __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);
401 }
402
403 // Pop from the ready queue from a given cluster
404 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
405 /* paranoid */ verify( lanes.count > 0 );
406 /* paranoid */ verify( kernelTLS().this_processor );
407 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
408
409 processor * proc = kernelTLS().this_processor;
410
411 if(proc->rdq.target == -1u) {
412 proc->rdq.target = __tls_rand() % lanes.count;
413 unsigned it1 = proc->rdq.itr;
414 unsigned it2 = proc->rdq.itr + 1;
415 unsigned idx1 = proc->rdq.id + (it1 % READYQ_SHARD_FACTOR);
416 unsigned idx2 = proc->rdq.id + (it1 % READYQ_SHARD_FACTOR);
417 unsigned long long tsc1 = ts(lanes.data[idx1]);
418 unsigned long long tsc2 = ts(lanes.data[idx2]);
419 proc->rdq.cutoff = min(tsc1, tsc2);
420 }
421 else if(lanes.tscs[proc->rdq.target].tv < proc->rdq.cutoff) {
422 $thread * t = try_pop(cltr, proc->rdq.target);
423 proc->rdq.target = -1u;
424 if(t) return t;
425 }
426
427 for(READYQ_SHARD_FACTOR) {
428 unsigned i = proc->rdq.id + (--proc->rdq.itr % READYQ_SHARD_FACTOR);
429 if($thread * t = try_pop(cltr, i)) return t;
430 }
431 return 0p;
432 }
433
434 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
435 for(25) {
436 unsigned i = __tls_rand() % lanes.count;
437 $thread * t = try_pop(cltr, i);
438 if(t) return t;
439 }
440
441 return search(cltr);
442 }
443#endif
444
445//=======================================================================
446// Various Ready Queue utilities
447//=======================================================================
448// these function work the same or almost the same
449// whether they are using work-stealing or relaxed fifo scheduling
450
451//-----------------------------------------------------------------------
452// try to pop from a lane given by index w
453static inline struct $thread * try_pop(struct cluster * cltr, unsigned w) with (cltr->ready_queue) {
454 // Get relevant elements locally
455 __intrusive_lane_t & lane = lanes.data[w];
456
457 // If list looks empty retry
458 if( is_empty(lane) ) return 0p;
459
460 // If we can't get the lock retry
461 if( !__atomic_try_acquire(&lane.lock) ) return 0p;
462
463 // If list is empty, unlock and retry
464 if( is_empty(lane) ) {
465 __atomic_unlock(&lane.lock);
466 return 0p;
467 }
468
469 // Actually pop the list
470 struct $thread * thrd;
471 thrd = pop(lane);
472
473 /* paranoid */ verify(thrd);
474 /* paranoid */ verify(lane.lock);
475
476 // Unlock and return
477 __atomic_unlock(&lane.lock);
478
479 // Update statistics
480 #if !defined(__CFA_NO_STATISTICS__)
481 __tls_stats()->ready.pick.pop.success++;
482 #endif
483
484 #if defined(USE_WORK_STEALING)
485 lanes.tscs[w].tv = thrd->link.ts;
486 #endif
487
488 // return the popped thread
489 return thrd;
490}
491
492//-----------------------------------------------------------------------
493// try to pop from any lanes making sure you don't miss any threads push
494// before the start of the function
495static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
496 /* paranoid */ verify( lanes.count > 0 );
497 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
498 unsigned offset = __tls_rand();
499 for(i; count) {
500 unsigned idx = (offset + i) % count;
501 struct $thread * thrd = try_pop(cltr, idx);
502 if(thrd) {
503 return thrd;
504 }
505 }
506
507 // All lanes where empty return 0p
508 return 0p;
509}
510
511//-----------------------------------------------------------------------
512// Check that all the intrusive queues in the data structure are still consistent
513static void check( __ready_queue_t & q ) with (q) {
514 #if defined(__CFA_WITH_VERIFY__) && !defined(USE_MPSC)
515 {
516 for( idx ; lanes.count ) {
517 __intrusive_lane_t & sl = lanes.data[idx];
518 assert(!lanes.data[idx].lock);
519
520 assert(head(sl)->link.prev == 0p );
521 assert(head(sl)->link.next->link.prev == head(sl) );
522 assert(tail(sl)->link.next == 0p );
523 assert(tail(sl)->link.prev->link.next == tail(sl) );
524
525 if(is_empty(sl)) {
526 assert(tail(sl)->link.prev == head(sl));
527 assert(head(sl)->link.next == tail(sl));
528 } else {
529 assert(tail(sl)->link.prev != head(sl));
530 assert(head(sl)->link.next != tail(sl));
531 }
532 }
533 }
534 #endif
535}
536
537//-----------------------------------------------------------------------
538// Given 2 indexes, pick the list with the oldest push an try to pop from it
539static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j) with (cltr->ready_queue) {
540 #if !defined(__CFA_NO_STATISTICS__)
541 __tls_stats()->ready.pick.pop.attempt++;
542 #endif
543
544 // Pick the bet list
545 int w = i;
546 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
547 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
548 }
549
550 return try_pop(cltr, w);
551}
552
553// Call this function of the intrusive list was moved using memcpy
554// fixes the list so that the pointers back to anchors aren't left dangling
555static inline void fix(__intrusive_lane_t & ll) {
556 #if !defined(USE_MPSC)
557 // if the list is not empty then follow he pointer and fix its reverse
558 if(!is_empty(ll)) {
559 head(ll)->link.next->link.prev = head(ll);
560 tail(ll)->link.prev->link.next = tail(ll);
561 }
562 // Otherwise just reset the list
563 else {
564 verify(tail(ll)->link.next == 0p);
565 tail(ll)->link.prev = head(ll);
566 head(ll)->link.next = tail(ll);
567 verify(head(ll)->link.prev == 0p);
568 }
569 #endif
570}
571
572static void assign_list(unsigned & value, dlist(processor, processor) & list, unsigned count) {
573 processor * it = &list`first;
574 for(unsigned i = 0; i < count; i++) {
575 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
576 it->rdq.id = value;
577 it->rdq.target = -1u;
578 value += READYQ_SHARD_FACTOR;
579 it = &(*it)`next;
580 }
581}
582
583static void reassign_cltr_id(struct cluster * cltr) {
584 unsigned preferred = 0;
585 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
586 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
587}
588
589static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
590 #if defined(USE_WORK_STEALING)
591 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
592 for(i; lanes.count) {
593 lanes.tscs[i].tv = ts(lanes.data[i]);
594 }
595 #endif
596}
597
598// Grow the ready queue
599void ready_queue_grow(struct cluster * cltr) {
600 size_t ncount;
601 int target = cltr->procs.total;
602
603 /* paranoid */ verify( ready_mutate_islocked() );
604 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
605
606 // Make sure that everything is consistent
607 /* paranoid */ check( cltr->ready_queue );
608
609 // grow the ready queue
610 with( cltr->ready_queue ) {
611 // Find new count
612 // Make sure we always have atleast 1 list
613 if(target >= 2) {
614 ncount = target * READYQ_SHARD_FACTOR;
615 } else {
616 ncount = 1;
617 }
618
619 // Allocate new array (uses realloc and memcpies the data)
620 lanes.data = alloc( ncount, lanes.data`realloc );
621
622 // Fix the moved data
623 for( idx; (size_t)lanes.count ) {
624 fix(lanes.data[idx]);
625 }
626
627 // Construct new data
628 for( idx; (size_t)lanes.count ~ ncount) {
629 (lanes.data[idx]){};
630 }
631
632 // Update original
633 lanes.count = ncount;
634 }
635
636 fix_times(cltr);
637
638 reassign_cltr_id(cltr);
639
640 // Make sure that everything is consistent
641 /* paranoid */ check( cltr->ready_queue );
642
643 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
644
645 /* paranoid */ verify( ready_mutate_islocked() );
646}
647
648// Shrink the ready queue
649void ready_queue_shrink(struct cluster * cltr) {
650 /* paranoid */ verify( ready_mutate_islocked() );
651 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
652
653 // Make sure that everything is consistent
654 /* paranoid */ check( cltr->ready_queue );
655
656 int target = cltr->procs.total;
657
658 with( cltr->ready_queue ) {
659 // Remember old count
660 size_t ocount = lanes.count;
661
662 // Find new count
663 // Make sure we always have atleast 1 list
664 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: 1;
665 /* paranoid */ verify( ocount >= lanes.count );
666 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
667
668 // for printing count the number of displaced threads
669 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
670 __attribute__((unused)) size_t displaced = 0;
671 #endif
672
673 // redistribute old data
674 for( idx; (size_t)lanes.count ~ ocount) {
675 // Lock is not strictly needed but makes checking invariants much easier
676 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
677 verify(locked);
678
679 // As long as we can pop from this lane to push the threads somewhere else in the queue
680 while(!is_empty(lanes.data[idx])) {
681 struct $thread * thrd;
682 thrd = pop(lanes.data[idx]);
683
684 push(cltr, thrd);
685
686 // for printing count the number of displaced threads
687 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
688 displaced++;
689 #endif
690 }
691
692 // Unlock the lane
693 __atomic_unlock(&lanes.data[idx].lock);
694
695 // TODO print the queue statistics here
696
697 ^(lanes.data[idx]){};
698 }
699
700 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
701
702 // Allocate new array (uses realloc and memcpies the data)
703 lanes.data = alloc( lanes.count, lanes.data`realloc );
704
705 // Fix the moved data
706 for( idx; (size_t)lanes.count ) {
707 fix(lanes.data[idx]);
708 }
709 }
710
711 fix_times(cltr);
712
713 reassign_cltr_id(cltr);
714
715 // Make sure that everything is consistent
716 /* paranoid */ check( cltr->ready_queue );
717
718 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
719 /* paranoid */ verify( ready_mutate_islocked() );
720}
Note: See TracBrowser for help on using the repository browser.