source: libcfa/src/concurrency/ready_queue.cfa@ f558b5f

ADT ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since f558b5f was 43784ac, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Changed libcfathread to consistently define _GNU_SOURCE
  • Property mode set to 100644
File size: 22.3 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 _GNU_SOURCE
18
19// #define __CFA_DEBUG_PRINT_READY_QUEUE__
20
21
22#define USE_RELAXED_FIFO
23// #define USE_WORK_STEALING
24
25#include "bits/defs.hfa"
26#include "kernel_private.hfa"
27
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#if !defined(__CFA_NO_STATISTICS__)
38 #define __STATS(...) __VA_ARGS__
39#else
40 #define __STATS(...)
41#endif
42
43// No overriden function, no environment variable, no define
44// fall back to a magic number
45#ifndef __CFA_MAX_PROCESSORS__
46 #define __CFA_MAX_PROCESSORS__ 1024
47#endif
48
49#if defined(USE_RELAXED_FIFO)
50 #define BIAS 4
51 #define READYQ_SHARD_FACTOR 4
52 #define SEQUENTIAL_SHARD 1
53#elif defined(USE_WORK_STEALING)
54 #define READYQ_SHARD_FACTOR 2
55 #define SEQUENTIAL_SHARD 2
56#else
57 #error no scheduling strategy selected
58#endif
59
60static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
61static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
62static inline struct $thread * search(struct cluster * cltr);
63static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
64
65
66// returns the maximum number of processors the RWLock support
67__attribute__((weak)) unsigned __max_processors() {
68 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
69 if(!max_cores_s) {
70 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
71 return __CFA_MAX_PROCESSORS__;
72 }
73
74 char * endptr = 0p;
75 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
76 if(max_cores_l < 1 || max_cores_l > 65535) {
77 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
78 return __CFA_MAX_PROCESSORS__;
79 }
80 if('\0' != *endptr) {
81 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
82 return __CFA_MAX_PROCESSORS__;
83 }
84
85 return max_cores_l;
86}
87
88//=======================================================================
89// Cluster wide reader-writer lock
90//=======================================================================
91void ?{}(__scheduler_RWLock_t & this) {
92 this.max = __max_processors();
93 this.alloc = 0;
94 this.ready = 0;
95 this.data = alloc(this.max);
96 this.write_lock = false;
97
98 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
99 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
100
101}
102void ^?{}(__scheduler_RWLock_t & this) {
103 free(this.data);
104}
105
106
107//=======================================================================
108// Lock-Free registering/unregistering of threads
109unsigned register_proc_id( void ) with(*__scheduler_lock) {
110 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
111 bool * handle = (bool *)&kernelTLS().sched_lock;
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 bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
119 /* paranoid */ verify( handle != *cell );
120
121 bool * null = 0p; // Re-write every loop since compare thrashes it
122 if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
123 && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
124 /* paranoid */ verify(i < ready);
125 /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
126 return i;
127 }
128 }
129
130 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
131
132 // Step - 2 : F&A to get a new spot in the array.
133 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
134 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
135
136 // Step - 3 : Mark space as used and then publish it.
137 data[n] = handle;
138 while() {
139 unsigned copy = n;
140 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
141 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
142 break;
143 Pause();
144 }
145
146 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
147
148 // Return new spot.
149 /* paranoid */ verify(n < ready);
150 /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
151 return n;
152}
153
154void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
155 /* paranoid */ verify(id < ready);
156 /* paranoid */ verify(id == kernelTLS().sched_id);
157 /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
158
159 bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
160
161 __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
162
163 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
164}
165
166//-----------------------------------------------------------------------
167// Writer side : acquire when changing the ready queue, e.g. adding more
168// queues or removing them.
169uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
170 /* paranoid */ verify( ! __preemption_enabled() );
171 /* paranoid */ verify( ! kernelTLS().sched_lock );
172
173 // Step 1 : lock global lock
174 // It is needed to avoid processors that register mid Critical-Section
175 // to simply lock their own lock and enter.
176 __atomic_acquire( &write_lock );
177
178 // Step 2 : lock per-proc lock
179 // Processors that are currently being registered aren't counted
180 // but can't be in read_lock or in the critical section.
181 // All other processors are counted
182 uint_fast32_t s = ready;
183 for(uint_fast32_t i = 0; i < s; i++) {
184 volatile bool * llock = data[i];
185 if(llock) __atomic_acquire( llock );
186 }
187
188 /* paranoid */ verify( ! __preemption_enabled() );
189 return s;
190}
191
192void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
193 /* paranoid */ verify( ! __preemption_enabled() );
194
195 // Step 1 : release local locks
196 // This must be done while the global lock is held to avoid
197 // threads that where created mid critical section
198 // to race to lock their local locks and have the writer
199 // immidiately unlock them
200 // Alternative solution : return s in write_lock and pass it to write_unlock
201 for(uint_fast32_t i = 0; i < last_s; i++) {
202 volatile bool * llock = data[i];
203 if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
204 }
205
206 // Step 2 : release global lock
207 /*paranoid*/ assert(true == write_lock);
208 __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
209
210 /* paranoid */ verify( ! __preemption_enabled() );
211}
212
213//=======================================================================
214// Cforall Ready Queue used for scheduling
215//=======================================================================
216void ?{}(__ready_queue_t & this) with (this) {
217 lanes.data = 0p;
218 lanes.tscs = 0p;
219 lanes.count = 0;
220}
221
222void ^?{}(__ready_queue_t & this) with (this) {
223 verify( SEQUENTIAL_SHARD == lanes.count );
224 free(lanes.data);
225 free(lanes.tscs);
226}
227
228//-----------------------------------------------------------------------
229#if defined(USE_RELAXED_FIFO)
230 //-----------------------------------------------------------------------
231 // get index from random number with or without bias towards queues
232 static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
233 unsigned i;
234 bool local;
235 unsigned rlow = r % BIAS;
236 unsigned rhigh = r / BIAS;
237 if((0 != rlow) && preferred >= 0) {
238 // (BIAS - 1) out of BIAS chances
239 // Use perferred queues
240 i = preferred + (rhigh % READYQ_SHARD_FACTOR);
241 local = true;
242 }
243 else {
244 // 1 out of BIAS chances
245 // Use all queues
246 i = rhigh;
247 local = false;
248 }
249 return [i, local];
250 }
251
252 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
253 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
254
255 const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
256 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
257
258 bool local;
259 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
260
261 // Try to pick a lane and lock it
262 unsigned i;
263 do {
264 // Pick the index of a lane
265 unsigned r = __tls_rand_fwd();
266 [i, local] = idx_from_r(r, preferred);
267
268 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
269
270 #if !defined(__CFA_NO_STATISTICS__)
271 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
272 else if(local) __tls_stats()->ready.push.local.attempt++;
273 else __tls_stats()->ready.push.share.attempt++;
274 #endif
275
276 // If we can't lock it retry
277 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
278
279 // Actually push it
280 push(lanes.data[i], thrd);
281
282 // Unlock and return
283 __atomic_unlock( &lanes.data[i].lock );
284
285 // Mark the current index in the tls rng instance as having an item
286 __tls_rand_advance_bck();
287
288 __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);
289
290 // Update statistics
291 #if !defined(__CFA_NO_STATISTICS__)
292 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
293 else if(local) __tls_stats()->ready.push.local.success++;
294 else __tls_stats()->ready.push.share.success++;
295 #endif
296 }
297
298 // Pop from the ready queue from a given cluster
299 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
300 /* paranoid */ verify( lanes.count > 0 );
301 /* paranoid */ verify( kernelTLS().this_processor );
302 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
303
304 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
305 int preferred = kernelTLS().this_processor->rdq.id;
306
307
308 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
309 for(25) {
310 // Pick two lists at random
311 unsigned ri = __tls_rand_bck();
312 unsigned rj = __tls_rand_bck();
313
314 unsigned i, j;
315 __attribute__((unused)) bool locali, localj;
316 [i, locali] = idx_from_r(ri, preferred);
317 [j, localj] = idx_from_r(rj, preferred);
318
319 i %= count;
320 j %= count;
321
322 // try popping from the 2 picked lists
323 struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
324 if(thrd) {
325 return thrd;
326 }
327 }
328
329 // All lanes where empty return 0p
330 return 0p;
331 }
332
333 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
334 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
335 return search(cltr);
336 }
337#endif
338#if defined(USE_WORK_STEALING)
339 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
340 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
341
342 // #define USE_PREFERRED
343 #if !defined(USE_PREFERRED)
344 const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
345 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
346 #else
347 unsigned preferred = thrd->preferred;
348 const bool external = push_local || (!kernelTLS().this_processor) || preferred == -1u || thrd->curr_cluster != cltr;
349 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
350
351 unsigned r = preferred % READYQ_SHARD_FACTOR;
352 const unsigned start = preferred - r;
353 #endif
354
355 // Try to pick a lane and lock it
356 unsigned i;
357 do {
358 #if !defined(__CFA_NO_STATISTICS__)
359 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
360 else __tls_stats()->ready.push.local.attempt++;
361 #endif
362
363 if(unlikely(external)) {
364 i = __tls_rand() % lanes.count;
365 }
366 else {
367 #if !defined(USE_PREFERRED)
368 processor * proc = kernelTLS().this_processor;
369 unsigned r = proc->rdq.its++;
370 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
371 #else
372 i = start + (r++ % READYQ_SHARD_FACTOR);
373 #endif
374 }
375 // If we can't lock it retry
376 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
377
378 // Actually push it
379 push(lanes.data[i], thrd);
380
381 // Unlock and return
382 __atomic_unlock( &lanes.data[i].lock );
383
384 #if !defined(__CFA_NO_STATISTICS__)
385 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
386 else __tls_stats()->ready.push.local.success++;
387 #endif
388
389 __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);
390 }
391
392 // Pop from the ready queue from a given cluster
393 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
394 /* paranoid */ verify( lanes.count > 0 );
395 /* paranoid */ verify( kernelTLS().this_processor );
396 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
397
398 processor * proc = kernelTLS().this_processor;
399
400 if(proc->rdq.target == -1u) {
401 unsigned long long min = ts(lanes.data[proc->rdq.id]);
402 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
403 unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
404 if(tsc < min) min = tsc;
405 }
406 proc->rdq.cutoff = min;
407 proc->rdq.target = __tls_rand() % lanes.count;
408 }
409 else {
410 unsigned target = proc->rdq.target;
411 proc->rdq.target = -1u;
412 const unsigned long long bias = 0; //2_500_000_000;
413 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
414 if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
415 $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
416 if(t) return t;
417 }
418 }
419
420 for(READYQ_SHARD_FACTOR) {
421 unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
422 if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
423 }
424 return 0p;
425 }
426
427 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
428 unsigned i = __tls_rand() % lanes.count;
429 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
430 }
431
432 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
433 return search(cltr);
434 }
435#endif
436
437//=======================================================================
438// Various Ready Queue utilities
439//=======================================================================
440// these function work the same or almost the same
441// whether they are using work-stealing or relaxed fifo scheduling
442
443//-----------------------------------------------------------------------
444// try to pop from a lane given by index w
445static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
446 __STATS( stats.attempt++; )
447
448 // Get relevant elements locally
449 __intrusive_lane_t & lane = lanes.data[w];
450
451 // If list looks empty retry
452 if( is_empty(lane) ) {
453 return 0p;
454 }
455
456 // If we can't get the lock retry
457 if( !__atomic_try_acquire(&lane.lock) ) {
458 return 0p;
459 }
460
461 // If list is empty, unlock and retry
462 if( is_empty(lane) ) {
463 __atomic_unlock(&lane.lock);
464 return 0p;
465 }
466
467 // Actually pop the list
468 struct $thread * thrd;
469 unsigned long long tsv;
470 [thrd, tsv] = pop(lane);
471
472 /* paranoid */ verify(thrd);
473 /* paranoid */ verify(tsv);
474 /* paranoid */ verify(lane.lock);
475
476 // Unlock and return
477 __atomic_unlock(&lane.lock);
478
479 // Update statistics
480 __STATS( stats.success++; )
481
482 #if defined(USE_WORK_STEALING)
483 lanes.tscs[w].tv = tsv;
484 #endif
485
486 thrd->preferred = w;
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 __STATS(, __tls_stats()->ready.pop.search));
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__)
515 {
516 for( idx ; lanes.count ) {
517 __intrusive_lane_t & sl = lanes.data[idx];
518 assert(!lanes.data[idx].lock);
519
520 if(is_empty(sl)) {
521 assert( sl.anchor.next == 0p );
522 assert( sl.anchor.ts == 0 );
523 assert( mock_head(sl) == sl.prev );
524 } else {
525 assert( sl.anchor.next != 0p );
526 assert( sl.anchor.ts != 0 );
527 assert( mock_head(sl) != sl.prev );
528 }
529 }
530 }
531 #endif
532}
533
534//-----------------------------------------------------------------------
535// Given 2 indexes, pick the list with the oldest push an try to pop from it
536static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
537 // Pick the bet list
538 int w = i;
539 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
540 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
541 }
542
543 return try_pop(cltr, w __STATS(, stats));
544}
545
546// Call this function of the intrusive list was moved using memcpy
547// fixes the list so that the pointers back to anchors aren't left dangling
548static inline void fix(__intrusive_lane_t & ll) {
549 if(is_empty(ll)) {
550 verify(ll.anchor.next == 0p);
551 ll.prev = mock_head(ll);
552 }
553}
554
555static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
556 processor * it = &list`first;
557 for(unsigned i = 0; i < count; i++) {
558 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
559 it->rdq.id = value;
560 it->rdq.target = -1u;
561 value += READYQ_SHARD_FACTOR;
562 it = &(*it)`next;
563 }
564}
565
566static void reassign_cltr_id(struct cluster * cltr) {
567 unsigned preferred = 0;
568 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
569 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
570}
571
572static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
573 #if defined(USE_WORK_STEALING)
574 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
575 for(i; lanes.count) {
576 unsigned long long tsc = ts(lanes.data[i]);
577 lanes.tscs[i].tv = tsc != 0 ? tsc : rdtscl();
578 }
579 #endif
580}
581
582// Grow the ready queue
583void ready_queue_grow(struct cluster * cltr) {
584 size_t ncount;
585 int target = cltr->procs.total;
586
587 /* paranoid */ verify( ready_mutate_islocked() );
588 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
589
590 // Make sure that everything is consistent
591 /* paranoid */ check( cltr->ready_queue );
592
593 // grow the ready queue
594 with( cltr->ready_queue ) {
595 // Find new count
596 // Make sure we always have atleast 1 list
597 if(target >= 2) {
598 ncount = target * READYQ_SHARD_FACTOR;
599 } else {
600 ncount = SEQUENTIAL_SHARD;
601 }
602
603 // Allocate new array (uses realloc and memcpies the data)
604 lanes.data = alloc( ncount, lanes.data`realloc );
605
606 // Fix the moved data
607 for( idx; (size_t)lanes.count ) {
608 fix(lanes.data[idx]);
609 }
610
611 // Construct new data
612 for( idx; (size_t)lanes.count ~ ncount) {
613 (lanes.data[idx]){};
614 }
615
616 // Update original
617 lanes.count = ncount;
618 }
619
620 fix_times(cltr);
621
622 reassign_cltr_id(cltr);
623
624 // Make sure that everything is consistent
625 /* paranoid */ check( cltr->ready_queue );
626
627 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
628
629 /* paranoid */ verify( ready_mutate_islocked() );
630}
631
632// Shrink the ready queue
633void ready_queue_shrink(struct cluster * cltr) {
634 /* paranoid */ verify( ready_mutate_islocked() );
635 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
636
637 // Make sure that everything is consistent
638 /* paranoid */ check( cltr->ready_queue );
639
640 int target = cltr->procs.total;
641
642 with( cltr->ready_queue ) {
643 // Remember old count
644 size_t ocount = lanes.count;
645
646 // Find new count
647 // Make sure we always have atleast 1 list
648 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
649 /* paranoid */ verify( ocount >= lanes.count );
650 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
651
652 // for printing count the number of displaced threads
653 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
654 __attribute__((unused)) size_t displaced = 0;
655 #endif
656
657 // redistribute old data
658 for( idx; (size_t)lanes.count ~ ocount) {
659 // Lock is not strictly needed but makes checking invariants much easier
660 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
661 verify(locked);
662
663 // As long as we can pop from this lane to push the threads somewhere else in the queue
664 while(!is_empty(lanes.data[idx])) {
665 struct $thread * thrd;
666 unsigned long long _;
667 [thrd, _] = pop(lanes.data[idx]);
668
669 push(cltr, thrd, true);
670
671 // for printing count the number of displaced threads
672 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
673 displaced++;
674 #endif
675 }
676
677 // Unlock the lane
678 __atomic_unlock(&lanes.data[idx].lock);
679
680 // TODO print the queue statistics here
681
682 ^(lanes.data[idx]){};
683 }
684
685 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
686
687 // Allocate new array (uses realloc and memcpies the data)
688 lanes.data = alloc( lanes.count, lanes.data`realloc );
689
690 // Fix the moved data
691 for( idx; (size_t)lanes.count ) {
692 fix(lanes.data[idx]);
693 }
694 }
695
696 fix_times(cltr);
697
698 reassign_cltr_id(cltr);
699
700 // Make sure that everything is consistent
701 /* paranoid */ check( cltr->ready_queue );
702
703 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
704 /* paranoid */ verify( ready_mutate_islocked() );
705}
706
707#if !defined(__CFA_NO_STATISTICS__)
708 unsigned cnt(const __ready_queue_t & this, unsigned idx) {
709 /* paranoid */ verify(this.lanes.count > idx);
710 return this.lanes.data[idx].cnt;
711 }
712#endif
Note: See TracBrowser for help on using the repository browser.