source: libcfa/src/concurrency/ready_queue.cfa@ ec1c674

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 ec1c674 was c993b15, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Changed RW lock to avoid hitting the global array on schedule.

  • 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 __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#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) with (cltr->ready_queue) {
253 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
254
255 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
256 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
257
258 // write timestamp
259 thrd->link.ts = rdtscl();
260
261 bool local;
262 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
263
264 // Try to pick a lane and lock it
265 unsigned i;
266 do {
267 // Pick the index of a lane
268 unsigned r = __tls_rand_fwd();
269 [i, local] = idx_from_r(r, preferred);
270
271 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
272
273 #if !defined(__CFA_NO_STATISTICS__)
274 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
275 else if(local) __tls_stats()->ready.push.local.attempt++;
276 else __tls_stats()->ready.push.share.attempt++;
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(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
303 else if(local) __tls_stats()->ready.push.local.success++;
304 else __tls_stats()->ready.push.share.success++;
305 #endif
306 }
307
308 // Pop from the ready queue from a given cluster
309 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
310 /* paranoid */ verify( lanes.count > 0 );
311 /* paranoid */ verify( kernelTLS().this_processor );
312 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
313
314 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
315 int preferred = kernelTLS().this_processor->rdq.id;
316
317
318 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
319 for(25) {
320 // Pick two lists at random
321 unsigned ri = __tls_rand_bck();
322 unsigned rj = __tls_rand_bck();
323
324 unsigned i, j;
325 __attribute__((unused)) bool locali, localj;
326 [i, locali] = idx_from_r(ri, preferred);
327 [j, localj] = idx_from_r(rj, preferred);
328
329 i %= count;
330 j %= count;
331
332 // try popping from the 2 picked lists
333 struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
334 if(thrd) {
335 return thrd;
336 }
337 }
338
339 // All lanes where empty return 0p
340 return 0p;
341 }
342
343 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
344 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
345 return search(cltr);
346 }
347#endif
348#if defined(USE_WORK_STEALING)
349 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
350 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
351
352 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
353 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
354
355 // write timestamp
356 thrd->link.ts = rdtscl();
357
358 // Try to pick a lane and lock it
359 unsigned i;
360 do {
361 #if !defined(__CFA_NO_STATISTICS__)
362 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
363 else __tls_stats()->ready.push.local.attempt++;
364 #endif
365
366 if(unlikely(external)) {
367 i = __tls_rand() % lanes.count;
368 }
369 else {
370 processor * proc = kernelTLS().this_processor;
371 unsigned r = proc->rdq.its++;
372 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
373 }
374
375
376 #if defined(USE_MPSC)
377 // mpsc always succeeds
378 } while( false );
379 #else
380 // If we can't lock it retry
381 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
382 #endif
383
384 // Actually push it
385 push(lanes.data[i], thrd);
386
387 #if !defined(USE_MPSC)
388 // Unlock and return
389 __atomic_unlock( &lanes.data[i].lock );
390 #endif
391
392 #if !defined(__CFA_NO_STATISTICS__)
393 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
394 else __tls_stats()->ready.push.local.success++;
395 #endif
396
397 __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);
398 }
399
400 // Pop from the ready queue from a given cluster
401 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
402 /* paranoid */ verify( lanes.count > 0 );
403 /* paranoid */ verify( kernelTLS().this_processor );
404 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
405
406 processor * proc = kernelTLS().this_processor;
407
408 if(proc->rdq.target == -1u) {
409 proc->rdq.target = __tls_rand() % lanes.count;
410 unsigned it1 = proc->rdq.itr;
411 unsigned it2 = proc->rdq.itr + 1;
412 unsigned idx1 = proc->rdq.id + (it1 % READYQ_SHARD_FACTOR);
413 unsigned idx2 = proc->rdq.id + (it2 % READYQ_SHARD_FACTOR);
414 unsigned long long tsc1 = ts(lanes.data[idx1]);
415 unsigned long long tsc2 = ts(lanes.data[idx2]);
416 proc->rdq.cutoff = min(tsc1, tsc2);
417 if(proc->rdq.cutoff == 0) proc->rdq.cutoff = -1ull;
418 }
419 else {
420 unsigned target = proc->rdq.target;
421 proc->rdq.target = -1u;
422 if(lanes.tscs[target].tv < proc->rdq.cutoff) {
423 $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
424 if(t) return t;
425 }
426 }
427
428 for(READYQ_SHARD_FACTOR) {
429 unsigned i = proc->rdq.id + (--proc->rdq.itr % READYQ_SHARD_FACTOR);
430 if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
431 }
432 return 0p;
433 }
434
435 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
436 unsigned i = __tls_rand() % lanes.count;
437 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
438 }
439
440 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
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 __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
454 __STATS( stats.attempt++; )
455
456 // Get relevant elements locally
457 __intrusive_lane_t & lane = lanes.data[w];
458
459 // If list looks empty retry
460 if( is_empty(lane) ) {
461 __STATS( stats.espec++; )
462 return 0p;
463 }
464
465 // If we can't get the lock retry
466 if( !__atomic_try_acquire(&lane.lock) ) {
467 __STATS( stats.elock++; )
468 return 0p;
469 }
470
471 // If list is empty, unlock and retry
472 if( is_empty(lane) ) {
473 __atomic_unlock(&lane.lock);
474 __STATS( stats.eempty++; )
475 return 0p;
476 }
477
478 // Actually pop the list
479 struct $thread * thrd;
480 thrd = pop(lane);
481
482 /* paranoid */ verify(thrd);
483 /* paranoid */ verify(lane.lock);
484
485 // Unlock and return
486 __atomic_unlock(&lane.lock);
487
488 // Update statistics
489 __STATS( stats.success++; )
490
491 #if defined(USE_WORK_STEALING)
492 lanes.tscs[w].tv = thrd->link.ts;
493 #endif
494
495 // return the popped thread
496 return thrd;
497}
498
499//-----------------------------------------------------------------------
500// try to pop from any lanes making sure you don't miss any threads push
501// before the start of the function
502static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
503 /* paranoid */ verify( lanes.count > 0 );
504 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
505 unsigned offset = __tls_rand();
506 for(i; count) {
507 unsigned idx = (offset + i) % count;
508 struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
509 if(thrd) {
510 return thrd;
511 }
512 }
513
514 // All lanes where empty return 0p
515 return 0p;
516}
517
518//-----------------------------------------------------------------------
519// Check that all the intrusive queues in the data structure are still consistent
520static void check( __ready_queue_t & q ) with (q) {
521 #if defined(__CFA_WITH_VERIFY__) && !defined(USE_MPSC)
522 {
523 for( idx ; lanes.count ) {
524 __intrusive_lane_t & sl = lanes.data[idx];
525 assert(!lanes.data[idx].lock);
526
527 assert(head(sl)->link.prev == 0p );
528 assert(head(sl)->link.next->link.prev == head(sl) );
529 assert(tail(sl)->link.next == 0p );
530 assert(tail(sl)->link.prev->link.next == tail(sl) );
531
532 if(is_empty(sl)) {
533 assert(tail(sl)->link.prev == head(sl));
534 assert(head(sl)->link.next == tail(sl));
535 } else {
536 assert(tail(sl)->link.prev != head(sl));
537 assert(head(sl)->link.next != tail(sl));
538 }
539 }
540 }
541 #endif
542}
543
544//-----------------------------------------------------------------------
545// Given 2 indexes, pick the list with the oldest push an try to pop from it
546static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
547 // Pick the bet list
548 int w = i;
549 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
550 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
551 }
552
553 return try_pop(cltr, w __STATS(, stats));
554}
555
556// Call this function of the intrusive list was moved using memcpy
557// fixes the list so that the pointers back to anchors aren't left dangling
558static inline void fix(__intrusive_lane_t & ll) {
559 #if !defined(USE_MPSC)
560 // if the list is not empty then follow he pointer and fix its reverse
561 if(!is_empty(ll)) {
562 head(ll)->link.next->link.prev = head(ll);
563 tail(ll)->link.prev->link.next = tail(ll);
564 }
565 // Otherwise just reset the list
566 else {
567 verify(tail(ll)->link.next == 0p);
568 tail(ll)->link.prev = head(ll);
569 head(ll)->link.next = tail(ll);
570 verify(head(ll)->link.prev == 0p);
571 }
572 #endif
573}
574
575static void assign_list(unsigned & value, dlist(processor, processor) & list, unsigned count) {
576 processor * it = &list`first;
577 for(unsigned i = 0; i < count; i++) {
578 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
579 it->rdq.id = value;
580 it->rdq.target = -1u;
581 value += READYQ_SHARD_FACTOR;
582 it = &(*it)`next;
583 }
584}
585
586static void reassign_cltr_id(struct cluster * cltr) {
587 unsigned preferred = 0;
588 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
589 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
590}
591
592static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
593 #if defined(USE_WORK_STEALING)
594 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
595 for(i; lanes.count) {
596 lanes.tscs[i].tv = ts(lanes.data[i]);
597 }
598 #endif
599}
600
601// Grow the ready queue
602void ready_queue_grow(struct cluster * cltr) {
603 size_t ncount;
604 int target = cltr->procs.total;
605
606 /* paranoid */ verify( ready_mutate_islocked() );
607 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
608
609 // Make sure that everything is consistent
610 /* paranoid */ check( cltr->ready_queue );
611
612 // grow the ready queue
613 with( cltr->ready_queue ) {
614 // Find new count
615 // Make sure we always have atleast 1 list
616 if(target >= 2) {
617 ncount = target * READYQ_SHARD_FACTOR;
618 } else {
619 ncount = SEQUENTIAL_SHARD;
620 }
621
622 // Allocate new array (uses realloc and memcpies the data)
623 lanes.data = alloc( ncount, lanes.data`realloc );
624
625 // Fix the moved data
626 for( idx; (size_t)lanes.count ) {
627 fix(lanes.data[idx]);
628 }
629
630 // Construct new data
631 for( idx; (size_t)lanes.count ~ ncount) {
632 (lanes.data[idx]){};
633 }
634
635 // Update original
636 lanes.count = ncount;
637 }
638
639 fix_times(cltr);
640
641 reassign_cltr_id(cltr);
642
643 // Make sure that everything is consistent
644 /* paranoid */ check( cltr->ready_queue );
645
646 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
647
648 /* paranoid */ verify( ready_mutate_islocked() );
649}
650
651// Shrink the ready queue
652void ready_queue_shrink(struct cluster * cltr) {
653 /* paranoid */ verify( ready_mutate_islocked() );
654 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
655
656 // Make sure that everything is consistent
657 /* paranoid */ check( cltr->ready_queue );
658
659 int target = cltr->procs.total;
660
661 with( cltr->ready_queue ) {
662 // Remember old count
663 size_t ocount = lanes.count;
664
665 // Find new count
666 // Make sure we always have atleast 1 list
667 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
668 /* paranoid */ verify( ocount >= lanes.count );
669 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
670
671 // for printing count the number of displaced threads
672 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
673 __attribute__((unused)) size_t displaced = 0;
674 #endif
675
676 // redistribute old data
677 for( idx; (size_t)lanes.count ~ ocount) {
678 // Lock is not strictly needed but makes checking invariants much easier
679 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
680 verify(locked);
681
682 // As long as we can pop from this lane to push the threads somewhere else in the queue
683 while(!is_empty(lanes.data[idx])) {
684 struct $thread * thrd;
685 thrd = pop(lanes.data[idx]);
686
687 push(cltr, thrd);
688
689 // for printing count the number of displaced threads
690 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
691 displaced++;
692 #endif
693 }
694
695 // Unlock the lane
696 __atomic_unlock(&lanes.data[idx].lock);
697
698 // TODO print the queue statistics here
699
700 ^(lanes.data[idx]){};
701 }
702
703 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
704
705 // Allocate new array (uses realloc and memcpies the data)
706 lanes.data = alloc( lanes.count, lanes.data`realloc );
707
708 // Fix the moved data
709 for( idx; (size_t)lanes.count ) {
710 fix(lanes.data[idx]);
711 }
712 }
713
714 fix_times(cltr);
715
716 reassign_cltr_id(cltr);
717
718 // Make sure that everything is consistent
719 /* paranoid */ check( cltr->ready_queue );
720
721 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
722 /* paranoid */ verify( ready_mutate_islocked() );
723}
Note: See TracBrowser for help on using the repository browser.