source: libcfa/src/concurrency/ready_queue.cfa@ f04a3df6

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 f04a3df6 was 2b96031, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added new subqueue implementation.
Seems faster will test on another machine before full replacement.
  • Property mode set to 100644
File size: 22.8 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 #if !defined(USE_NEW_SUBQUEUE)
260 thrd->link.ts = rdtscl();
261 #endif
262
263 bool local;
264 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
265
266 // Try to pick a lane and lock it
267 unsigned i;
268 do {
269 // Pick the index of a lane
270 unsigned r = __tls_rand_fwd();
271 [i, local] = idx_from_r(r, preferred);
272
273 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
274
275 #if !defined(__CFA_NO_STATISTICS__)
276 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
277 else if(local) __tls_stats()->ready.push.local.attempt++;
278 else __tls_stats()->ready.push.share.attempt++;
279 #endif
280
281 #if defined(USE_MPSC)
282 // mpsc always succeeds
283 } while( false );
284 #else
285 // If we can't lock it retry
286 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
287 #endif
288
289 // Actually push it
290 push(lanes.data[i], thrd);
291
292 #if !defined(USE_MPSC)
293 // Unlock and return
294 __atomic_unlock( &lanes.data[i].lock );
295 #endif
296
297 // Mark the current index in the tls rng instance as having an item
298 __tls_rand_advance_bck();
299
300 __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);
301
302 // Update statistics
303 #if !defined(__CFA_NO_STATISTICS__)
304 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
305 else if(local) __tls_stats()->ready.push.local.success++;
306 else __tls_stats()->ready.push.share.success++;
307 #endif
308 }
309
310 // Pop from the ready queue from a given cluster
311 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
312 /* paranoid */ verify( lanes.count > 0 );
313 /* paranoid */ verify( kernelTLS().this_processor );
314 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
315
316 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
317 int preferred = kernelTLS().this_processor->rdq.id;
318
319
320 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
321 for(25) {
322 // Pick two lists at random
323 unsigned ri = __tls_rand_bck();
324 unsigned rj = __tls_rand_bck();
325
326 unsigned i, j;
327 __attribute__((unused)) bool locali, localj;
328 [i, locali] = idx_from_r(ri, preferred);
329 [j, localj] = idx_from_r(rj, preferred);
330
331 i %= count;
332 j %= count;
333
334 // try popping from the 2 picked lists
335 struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
336 if(thrd) {
337 return thrd;
338 }
339 }
340
341 // All lanes where empty return 0p
342 return 0p;
343 }
344
345 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
346 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
347 return search(cltr);
348 }
349#endif
350#if defined(USE_WORK_STEALING)
351 __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
352 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
353
354 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
355 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
356
357 // write timestamp
358 #if !defined(USE_NEW_SUBQUEUE)
359 thrd->link.ts = rdtscl();
360 #endif
361
362 // Try to pick a lane and lock it
363 unsigned i;
364 do {
365 #if !defined(__CFA_NO_STATISTICS__)
366 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
367 else __tls_stats()->ready.push.local.attempt++;
368 #endif
369
370 if(unlikely(external)) {
371 i = __tls_rand() % lanes.count;
372 }
373 else {
374 processor * proc = kernelTLS().this_processor;
375 unsigned r = proc->rdq.its++;
376 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
377 }
378
379
380 #if defined(USE_MPSC)
381 // mpsc always succeeds
382 } while( false );
383 #else
384 // If we can't lock it retry
385 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
386 #endif
387
388 // Actually push it
389 push(lanes.data[i], thrd);
390
391 #if !defined(USE_MPSC)
392 // Unlock and return
393 __atomic_unlock( &lanes.data[i].lock );
394 #endif
395
396 #if !defined(__CFA_NO_STATISTICS__)
397 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
398 else __tls_stats()->ready.push.local.success++;
399 #endif
400
401 __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);
402 }
403
404 // Pop from the ready queue from a given cluster
405 __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
406 /* paranoid */ verify( lanes.count > 0 );
407 /* paranoid */ verify( kernelTLS().this_processor );
408 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
409
410 processor * proc = kernelTLS().this_processor;
411
412 if(proc->rdq.target == -1u) {
413 proc->rdq.target = __tls_rand() % lanes.count;
414 unsigned it1 = proc->rdq.itr;
415 unsigned it2 = proc->rdq.itr + 1;
416 unsigned idx1 = proc->rdq.id + (it1 % READYQ_SHARD_FACTOR);
417 unsigned idx2 = proc->rdq.id + (it2 % READYQ_SHARD_FACTOR);
418 unsigned long long tsc1 = ts(lanes.data[idx1]);
419 unsigned long long tsc2 = ts(lanes.data[idx2]);
420 proc->rdq.cutoff = min(tsc1, tsc2);
421 if(proc->rdq.cutoff == 0) proc->rdq.cutoff = -1ull;
422 }
423 else {
424 unsigned target = proc->rdq.target;
425 proc->rdq.target = -1u;
426 if(lanes.tscs[target].tv < proc->rdq.cutoff) {
427 $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
428 if(t) return t;
429 }
430 }
431
432 for(READYQ_SHARD_FACTOR) {
433 unsigned i = proc->rdq.id + (--proc->rdq.itr % READYQ_SHARD_FACTOR);
434 if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
435 }
436 return 0p;
437 }
438
439 __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
440 unsigned i = __tls_rand() % lanes.count;
441 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
442 }
443
444 __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
445 return search(cltr);
446 }
447#endif
448
449//=======================================================================
450// Various Ready Queue utilities
451//=======================================================================
452// these function work the same or almost the same
453// whether they are using work-stealing or relaxed fifo scheduling
454
455//-----------------------------------------------------------------------
456// try to pop from a lane given by index w
457static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
458 __STATS( stats.attempt++; )
459
460 // Get relevant elements locally
461 __intrusive_lane_t & lane = lanes.data[w];
462
463 // If list looks empty retry
464 if( is_empty(lane) ) {
465 __STATS( stats.espec++; )
466 return 0p;
467 }
468
469 // If we can't get the lock retry
470 if( !__atomic_try_acquire(&lane.lock) ) {
471 __STATS( stats.elock++; )
472 return 0p;
473 }
474
475 // If list is empty, unlock and retry
476 if( is_empty(lane) ) {
477 __atomic_unlock(&lane.lock);
478 __STATS( stats.eempty++; )
479 return 0p;
480 }
481
482 // Actually pop the list
483 struct $thread * thrd;
484 thrd = pop(lane);
485
486 /* paranoid */ verify(thrd);
487 /* paranoid */ verify(lane.lock);
488
489 // Unlock and return
490 __atomic_unlock(&lane.lock);
491
492 // Update statistics
493 __STATS( stats.success++; )
494
495 #if defined(USE_WORK_STEALING)
496 lanes.tscs[w].tv = thrd->link.ts;
497 #endif
498
499 // return the popped thread
500 return thrd;
501}
502
503//-----------------------------------------------------------------------
504// try to pop from any lanes making sure you don't miss any threads push
505// before the start of the function
506static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
507 /* paranoid */ verify( lanes.count > 0 );
508 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
509 unsigned offset = __tls_rand();
510 for(i; count) {
511 unsigned idx = (offset + i) % count;
512 struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
513 if(thrd) {
514 return thrd;
515 }
516 }
517
518 // All lanes where empty return 0p
519 return 0p;
520}
521
522//-----------------------------------------------------------------------
523// Check that all the intrusive queues in the data structure are still consistent
524static void check( __ready_queue_t & q ) with (q) {
525 #if defined(__CFA_WITH_VERIFY__) && !defined(USE_MPSC)
526 {
527 for( idx ; lanes.count ) {
528 __intrusive_lane_t & sl = lanes.data[idx];
529 assert(!lanes.data[idx].lock);
530
531 #if defined(USE_NEW_SUBQUEUE)
532 if(is_empty(sl)) {
533 assert( sl.anchor.next == 0p );
534 assert( sl.anchor.ts == 0 );
535 assert( mock_head(sl) == sl.prev );
536 } else {
537 assert( sl.anchor.next != 0p );
538 assert( sl.anchor.ts != 0 );
539 assert( mock_head(sl) != sl.prev );
540 }
541 #else
542 assert(head(sl)->link.prev == 0p );
543 assert(head(sl)->link.next->link.prev == head(sl) );
544 assert(tail(sl)->link.next == 0p );
545 assert(tail(sl)->link.prev->link.next == tail(sl) );
546
547 if(is_empty(sl)) {
548 assert(tail(sl)->link.prev == head(sl));
549 assert(head(sl)->link.next == tail(sl));
550 } else {
551 assert(tail(sl)->link.prev != head(sl));
552 assert(head(sl)->link.next != tail(sl));
553 }
554 #endif
555 }
556 }
557 #endif
558}
559
560//-----------------------------------------------------------------------
561// Given 2 indexes, pick the list with the oldest push an try to pop from it
562static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
563 // Pick the bet list
564 int w = i;
565 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
566 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
567 }
568
569 return try_pop(cltr, w __STATS(, stats));
570}
571
572// Call this function of the intrusive list was moved using memcpy
573// fixes the list so that the pointers back to anchors aren't left dangling
574static inline void fix(__intrusive_lane_t & ll) {
575 #if !defined(USE_MPSC)
576 #if defined(USE_NEW_SUBQUEUE)
577 if(is_empty(ll)) {
578 verify(ll.anchor.next == 0p);
579 ll.prev = mock_head(ll);
580 }
581 #else
582 // if the list is not empty then follow he pointer and fix its reverse
583 if(!is_empty(ll)) {
584 head(ll)->link.next->link.prev = head(ll);
585 tail(ll)->link.prev->link.next = tail(ll);
586 }
587 // Otherwise just reset the list
588 else {
589 verify(tail(ll)->link.next == 0p);
590 tail(ll)->link.prev = head(ll);
591 head(ll)->link.next = tail(ll);
592 verify(head(ll)->link.prev == 0p);
593 }
594 #endif
595 #endif
596}
597
598static void assign_list(unsigned & value, dlist(processor, processor) & list, unsigned count) {
599 processor * it = &list`first;
600 for(unsigned i = 0; i < count; i++) {
601 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
602 it->rdq.id = value;
603 it->rdq.target = -1u;
604 value += READYQ_SHARD_FACTOR;
605 it = &(*it)`next;
606 }
607}
608
609static void reassign_cltr_id(struct cluster * cltr) {
610 unsigned preferred = 0;
611 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
612 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
613}
614
615static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
616 #if defined(USE_WORK_STEALING)
617 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
618 for(i; lanes.count) {
619 lanes.tscs[i].tv = ts(lanes.data[i]);
620 }
621 #endif
622}
623
624// Grow the ready queue
625void ready_queue_grow(struct cluster * cltr) {
626 size_t ncount;
627 int target = cltr->procs.total;
628
629 /* paranoid */ verify( ready_mutate_islocked() );
630 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
631
632 // Make sure that everything is consistent
633 /* paranoid */ check( cltr->ready_queue );
634
635 // grow the ready queue
636 with( cltr->ready_queue ) {
637 // Find new count
638 // Make sure we always have atleast 1 list
639 if(target >= 2) {
640 ncount = target * READYQ_SHARD_FACTOR;
641 } else {
642 ncount = SEQUENTIAL_SHARD;
643 }
644
645 // Allocate new array (uses realloc and memcpies the data)
646 lanes.data = alloc( ncount, lanes.data`realloc );
647
648 // Fix the moved data
649 for( idx; (size_t)lanes.count ) {
650 fix(lanes.data[idx]);
651 }
652
653 // Construct new data
654 for( idx; (size_t)lanes.count ~ ncount) {
655 (lanes.data[idx]){};
656 }
657
658 // Update original
659 lanes.count = ncount;
660 }
661
662 fix_times(cltr);
663
664 reassign_cltr_id(cltr);
665
666 // Make sure that everything is consistent
667 /* paranoid */ check( cltr->ready_queue );
668
669 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
670
671 /* paranoid */ verify( ready_mutate_islocked() );
672}
673
674// Shrink the ready queue
675void ready_queue_shrink(struct cluster * cltr) {
676 /* paranoid */ verify( ready_mutate_islocked() );
677 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
678
679 // Make sure that everything is consistent
680 /* paranoid */ check( cltr->ready_queue );
681
682 int target = cltr->procs.total;
683
684 with( cltr->ready_queue ) {
685 // Remember old count
686 size_t ocount = lanes.count;
687
688 // Find new count
689 // Make sure we always have atleast 1 list
690 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
691 /* paranoid */ verify( ocount >= lanes.count );
692 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
693
694 // for printing count the number of displaced threads
695 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
696 __attribute__((unused)) size_t displaced = 0;
697 #endif
698
699 // redistribute old data
700 for( idx; (size_t)lanes.count ~ ocount) {
701 // Lock is not strictly needed but makes checking invariants much easier
702 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
703 verify(locked);
704
705 // As long as we can pop from this lane to push the threads somewhere else in the queue
706 while(!is_empty(lanes.data[idx])) {
707 struct $thread * thrd;
708 thrd = pop(lanes.data[idx]);
709
710 push(cltr, thrd);
711
712 // for printing count the number of displaced threads
713 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
714 displaced++;
715 #endif
716 }
717
718 // Unlock the lane
719 __atomic_unlock(&lanes.data[idx].lock);
720
721 // TODO print the queue statistics here
722
723 ^(lanes.data[idx]){};
724 }
725
726 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
727
728 // Allocate new array (uses realloc and memcpies the data)
729 lanes.data = alloc( lanes.count, lanes.data`realloc );
730
731 // Fix the moved data
732 for( idx; (size_t)lanes.count ) {
733 fix(lanes.data[idx]);
734 }
735 }
736
737 fix_times(cltr);
738
739 reassign_cltr_id(cltr);
740
741 // Make sure that everything is consistent
742 /* paranoid */ check( cltr->ready_queue );
743
744 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
745 /* paranoid */ verify( ready_mutate_islocked() );
746}
Note: See TracBrowser for help on using the repository browser.