source: libcfa/src/concurrency/ready_queue.cfa@ 9cc3a18

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 9cc3a18 was 9cc3a18, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

Major clean-up before attempting to add new scheduler
  • Property mode set to 100644
File size: 18.6 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 struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j);
55
56
57// returns the maximum number of processors the RWLock support
58__attribute__((weak)) unsigned __max_processors() {
59 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
60 if(!max_cores_s) {
61 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
62 return __CFA_MAX_PROCESSORS__;
63 }
64
65 char * endptr = 0p;
66 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
67 if(max_cores_l < 1 || max_cores_l > 65535) {
68 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
69 return __CFA_MAX_PROCESSORS__;
70 }
71 if('\0' != *endptr) {
72 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
73 return __CFA_MAX_PROCESSORS__;
74 }
75
76 return max_cores_l;
77}
78
79//=======================================================================
80// Cluster wide reader-writer lock
81//=======================================================================
82void ?{}(__scheduler_RWLock_t & this) {
83 this.max = __max_processors();
84 this.alloc = 0;
85 this.ready = 0;
86 this.lock = false;
87 this.data = alloc(this.max);
88
89 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data )) % 64) );
90 /*paranoid*/ verify( 0 == (((uintptr_t)(this.data + 1)) % 64) );
91 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
92 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
93
94}
95void ^?{}(__scheduler_RWLock_t & this) {
96 free(this.data);
97}
98
99void ?{}( __scheduler_lock_id_t & this, __processor_id_t * proc ) {
100 this.handle = proc;
101 this.lock = false;
102 #ifdef __CFA_WITH_VERIFY__
103 this.owned = false;
104 #endif
105}
106
107//=======================================================================
108// Lock-Free registering/unregistering of threads
109void register_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
110 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
111
112 // Step - 1 : check if there is already space in the data
113 uint_fast32_t s = ready;
114
115 // Check among all the ready
116 for(uint_fast32_t i = 0; i < s; i++) {
117 __processor_id_t * null = 0p; // Re-write every loop since compare thrashes it
118 if( __atomic_load_n(&data[i].handle, (int)__ATOMIC_RELAXED) == null
119 && __atomic_compare_exchange_n( &data[i].handle, &null, proc, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
120 /*paranoid*/ verify(i < ready);
121 /*paranoid*/ verify(0 == (__alignof__(data[i]) % cache_line_size));
122 /*paranoid*/ verify((((uintptr_t)&data[i]) % cache_line_size) == 0);
123 proc->id = i;
124 }
125 }
126
127 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
128
129 // Step - 2 : F&A to get a new spot in the array.
130 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
131 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
132
133 // Step - 3 : Mark space as used and then publish it.
134 __scheduler_lock_id_t * storage = (__scheduler_lock_id_t *)&data[n];
135 (*storage){ proc };
136 while() {
137 unsigned copy = n;
138 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
139 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
140 break;
141 Pause();
142 }
143
144 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
145
146 // Return new spot.
147 /*paranoid*/ verify(n < ready);
148 /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
149 /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
150 proc->id = n;
151}
152
153void unregister_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
154 unsigned id = proc->id;
155 /*paranoid*/ verify(id < ready);
156 /*paranoid*/ verify(proc == __atomic_load_n(&data[id].handle, __ATOMIC_RELAXED));
157 __atomic_store_n(&data[id].handle, 0p, __ATOMIC_RELEASE);
158
159 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
160}
161
162//-----------------------------------------------------------------------
163// Writer side : acquire when changing the ready queue, e.g. adding more
164// queues or removing them.
165uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
166 /* paranoid */ verify( ! __preemption_enabled() );
167
168 // Step 1 : lock global lock
169 // It is needed to avoid processors that register mid Critical-Section
170 // to simply lock their own lock and enter.
171 __atomic_acquire( &lock );
172
173 // Step 2 : lock per-proc lock
174 // Processors that are currently being registered aren't counted
175 // but can't be in read_lock or in the critical section.
176 // All other processors are counted
177 uint_fast32_t s = ready;
178 for(uint_fast32_t i = 0; i < s; i++) {
179 __atomic_acquire( &data[i].lock );
180 }
181
182 /* paranoid */ verify( ! __preemption_enabled() );
183 return s;
184}
185
186void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
187 /* paranoid */ verify( ! __preemption_enabled() );
188
189 // Step 1 : release local locks
190 // This must be done while the global lock is held to avoid
191 // threads that where created mid critical section
192 // to race to lock their local locks and have the writer
193 // immidiately unlock them
194 // Alternative solution : return s in write_lock and pass it to write_unlock
195 for(uint_fast32_t i = 0; i < last_s; i++) {
196 verify(data[i].lock);
197 __atomic_store_n(&data[i].lock, (bool)false, __ATOMIC_RELEASE);
198 }
199
200 // Step 2 : release global lock
201 /*paranoid*/ assert(true == lock);
202 __atomic_store_n(&lock, (bool)false, __ATOMIC_RELEASE);
203
204 /* paranoid */ verify( ! __preemption_enabled() );
205}
206
207//=======================================================================
208// Cforall Ready Queue used for scheduling
209//=======================================================================
210void ?{}(__ready_queue_t & this) with (this) {
211 lanes.data = 0p;
212 lanes.tscs = 0p;
213 lanes.count = 0;
214}
215
216void ^?{}(__ready_queue_t & this) with (this) {
217 verify( 1 == lanes.count );
218 free(lanes.data);
219 free(lanes.tscs);
220}
221
222//-----------------------------------------------------------------------
223__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
224 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
225
226 const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
227 /* paranoid */ verify(external || kernelTLS().this_processor->cltr_id < lanes.count );
228
229 // write timestamp
230 thrd->link.ts = rdtscl();
231
232 bool local;
233 int preferred = external ? -1 : kernelTLS().this_processor->cltr_id;
234
235 // Try to pick a lane and lock it
236 unsigned i;
237 do {
238 // Pick the index of a lane
239 unsigned r = __tls_rand_fwd();
240 [i, local] = idx_from_r(r, preferred);
241
242 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
243
244 #if !defined(__CFA_NO_STATISTICS__)
245 if(external) {
246 if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.local, 1, __ATOMIC_RELAXED);
247 __atomic_fetch_add(&cltr->stats->ready.pick.ext.attempt, 1, __ATOMIC_RELAXED);
248 }
249 else {
250 if(local) __tls_stats()->ready.pick.push.local++;
251 __tls_stats()->ready.pick.push.attempt++;
252 }
253 #endif
254
255 #if defined(USE_MPSC)
256 // mpsc always succeeds
257 } while( false );
258 #else
259 // If we can't lock it retry
260 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
261 #endif
262
263 // Actually push it
264 push(lanes.data[i], thrd);
265
266 #if !defined(USE_MPSC)
267 // Unlock and return
268 __atomic_unlock( &lanes.data[i].lock );
269 #endif
270
271 // Mark the current index in the tls rng instance as having an item
272 __tls_rand_advance_bck();
273
274 __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);
275
276 // Update statistics
277 #if !defined(__CFA_NO_STATISTICS__)
278 if(external) {
279 if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.lsuccess, 1, __ATOMIC_RELAXED);
280 __atomic_fetch_add(&cltr->stats->ready.pick.ext.success, 1, __ATOMIC_RELAXED);
281 }
282 else {
283 if(local) __tls_stats()->ready.pick.push.lsuccess++;
284 __tls_stats()->ready.pick.push.success++;
285 }
286 #endif
287}
288
289// Pop from the ready queue from a given cluster
290__attribute__((hot)) $thread * pop(struct cluster * cltr) with (cltr->ready_queue) {
291 /* paranoid */ verify( lanes.count > 0 );
292 /* paranoid */ verify(kernelTLS().this_processor->cltr_id < lanes.count );
293
294 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
295 int preferred = kernelTLS().this_processor->cltr_id;
296
297
298 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
299 for(25) {
300 // Pick two lists at random
301 unsigned ri = __tls_rand_bck();
302 unsigned rj = __tls_rand_bck();
303
304 unsigned i, j;
305 __attribute__((unused)) bool locali, localj;
306 [i, locali] = idx_from_r(ri, preferred);
307 [j, localj] = idx_from_r(rj, preferred);
308
309 #if !defined(__CFA_NO_STATISTICS__)
310 if(locali && localj) {
311 __tls_stats()->ready.pick.pop.local++;
312 }
313 #endif
314
315 i %= count;
316 j %= count;
317
318 // try popping from the 2 picked lists
319 struct $thread * thrd = try_pop(cltr, i, j);
320 if(thrd) {
321 #if !defined(__CFA_NO_STATISTICS__)
322 if( locali || localj ) __tls_stats()->ready.pick.pop.lsuccess++;
323 #endif
324 return thrd;
325 }
326 }
327
328 // All lanes where empty return 0p
329 return 0p;
330}
331
332static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
333 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
334 for(i; lanes.count) {
335 lanes.tscs[i].tv = ts(lanes.data[i]);
336 }
337
338}
339
340//=======================================================================
341// Various Ready Queue utilities
342//=======================================================================
343// these function work the same or almost the same
344// whether they are using work-stealing or relaxed fifo scheduling
345
346//-----------------------------------------------------------------------
347// get index from random number with or without bias towards queues
348static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
349 unsigned i;
350 bool local;
351 unsigned rlow = r % BIAS;
352 unsigned rhigh = r / BIAS;
353 if((0 != rlow) && preferred >= 0) {
354 // (BIAS - 1) out of BIAS chances
355 // Use perferred queues
356 i = preferred + (rhigh % READYQ_SHARD_FACTOR);
357 local = true;
358 }
359 else {
360 // 1 out of BIAS chances
361 // Use all queues
362 i = rhigh;
363 local = false;
364 }
365 return [i, local];
366}
367
368//-----------------------------------------------------------------------
369// try to pop from a lane given by index w
370static inline struct $thread * try_pop(struct cluster * cltr, unsigned w) with (cltr->ready_queue) {
371 // Get relevant elements locally
372 __intrusive_lane_t & lane = lanes.data[w];
373
374 // If list looks empty retry
375 if( is_empty(lane) ) return 0p;
376
377 // If we can't get the lock retry
378 if( !__atomic_try_acquire(&lane.lock) ) return 0p;
379
380 // If list is empty, unlock and retry
381 if( is_empty(lane) ) {
382 __atomic_unlock(&lane.lock);
383 return 0p;
384 }
385
386 // Actually pop the list
387 struct $thread * thrd;
388 thrd = pop(lane);
389
390 /* paranoid */ verify(thrd);
391 /* paranoid */ verify(lane.lock);
392
393 // Unlock and return
394 __atomic_unlock(&lane.lock);
395
396 // Update statistics
397 #if !defined(__CFA_NO_STATISTICS__)
398 __tls_stats()->ready.pick.pop.success++;
399 #endif
400
401 #if defined(USE_WORKSTEALING)
402 lanes.times[i].val = thrd->links.ts;
403 #endif
404
405 // return the popped thread
406 return thrd;
407}
408
409//-----------------------------------------------------------------------
410// try to pop from any lanes making sure you don't miss any threads push
411// before the start of the function
412__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
413 /* paranoid */ verify( lanes.count > 0 );
414 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
415 unsigned offset = __tls_rand();
416 for(i; count) {
417 unsigned idx = (offset + i) % count;
418 struct $thread * thrd = try_pop(cltr, idx);
419 if(thrd) {
420 return thrd;
421 }
422 }
423
424 // All lanes where empty return 0p
425 return 0p;
426}
427
428//-----------------------------------------------------------------------
429// Check that all the intrusive queues in the data structure are still consistent
430static void check( __ready_queue_t & q ) with (q) {
431 #if defined(__CFA_WITH_VERIFY__) && !defined(USE_MPSC)
432 {
433 for( idx ; lanes.count ) {
434 __intrusive_lane_t & sl = lanes.data[idx];
435 assert(!lanes.data[idx].lock);
436
437 assert(head(sl)->link.prev == 0p );
438 assert(head(sl)->link.next->link.prev == head(sl) );
439 assert(tail(sl)->link.next == 0p );
440 assert(tail(sl)->link.prev->link.next == tail(sl) );
441
442 if(is_empty(sl)) {
443 assert(tail(sl)->link.prev == head(sl));
444 assert(head(sl)->link.next == tail(sl));
445 } else {
446 assert(tail(sl)->link.prev != head(sl));
447 assert(head(sl)->link.next != tail(sl));
448 }
449 }
450 }
451 #endif
452}
453
454//-----------------------------------------------------------------------
455// Given 2 indexes, pick the list with the oldest push an try to pop from it
456static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j) with (cltr->ready_queue) {
457 #if !defined(__CFA_NO_STATISTICS__)
458 __tls_stats()->ready.pick.pop.attempt++;
459 #endif
460
461 // Pick the bet list
462 int w = i;
463 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
464 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
465 }
466
467 return try_pop(cltr, w);
468}
469
470// Call this function of the intrusive list was moved using memcpy
471// fixes the list so that the pointers back to anchors aren't left dangling
472static inline void fix(__intrusive_lane_t & ll) {
473 #if !defined(USE_MPSC)
474 // if the list is not empty then follow he pointer and fix its reverse
475 if(!is_empty(ll)) {
476 head(ll)->link.next->link.prev = head(ll);
477 tail(ll)->link.prev->link.next = tail(ll);
478 }
479 // Otherwise just reset the list
480 else {
481 verify(tail(ll)->link.next == 0p);
482 tail(ll)->link.prev = head(ll);
483 head(ll)->link.next = tail(ll);
484 verify(head(ll)->link.prev == 0p);
485 }
486 #endif
487}
488
489static void assign_list(unsigned & value, dlist(processor, processor) & list, unsigned count) {
490 processor * it = &list`first;
491 for(unsigned i = 0; i < count; i++) {
492 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
493 it->cltr_id = value;
494 value += READYQ_SHARD_FACTOR;
495 it = &(*it)`next;
496 }
497}
498
499static void reassign_cltr_id(struct cluster * cltr) {
500 unsigned preferred = 0;
501 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
502 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
503}
504
505// Grow the ready queue
506void ready_queue_grow(struct cluster * cltr) {
507 size_t ncount;
508 int target = cltr->procs.total;
509
510 /* paranoid */ verify( ready_mutate_islocked() );
511 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
512
513 // Make sure that everything is consistent
514 /* paranoid */ check( cltr->ready_queue );
515
516 // grow the ready queue
517 with( cltr->ready_queue ) {
518 // Find new count
519 // Make sure we always have atleast 1 list
520 if(target >= 2) {
521 ncount = target * READYQ_SHARD_FACTOR;
522 } else {
523 ncount = 1;
524 }
525
526 // Allocate new array (uses realloc and memcpies the data)
527 lanes.data = alloc( ncount, lanes.data`realloc );
528
529 // Fix the moved data
530 for( idx; (size_t)lanes.count ) {
531 fix(lanes.data[idx]);
532 }
533
534 // Construct new data
535 for( idx; (size_t)lanes.count ~ ncount) {
536 (lanes.data[idx]){};
537 }
538
539 // Update original
540 lanes.count = ncount;
541 }
542
543 fix_times(cltr);
544
545 reassign_cltr_id(cltr);
546
547 // Make sure that everything is consistent
548 /* paranoid */ check( cltr->ready_queue );
549
550 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
551
552 /* paranoid */ verify( ready_mutate_islocked() );
553}
554
555// Shrink the ready queue
556void ready_queue_shrink(struct cluster * cltr) {
557 /* paranoid */ verify( ready_mutate_islocked() );
558 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
559
560 // Make sure that everything is consistent
561 /* paranoid */ check( cltr->ready_queue );
562
563 int target = cltr->procs.total;
564
565 with( cltr->ready_queue ) {
566 // Remember old count
567 size_t ocount = lanes.count;
568
569 // Find new count
570 // Make sure we always have atleast 1 list
571 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: 1;
572 /* paranoid */ verify( ocount >= lanes.count );
573 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
574
575 // for printing count the number of displaced threads
576 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
577 __attribute__((unused)) size_t displaced = 0;
578 #endif
579
580 // redistribute old data
581 for( idx; (size_t)lanes.count ~ ocount) {
582 // Lock is not strictly needed but makes checking invariants much easier
583 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
584 verify(locked);
585
586 // As long as we can pop from this lane to push the threads somewhere else in the queue
587 while(!is_empty(lanes.data[idx])) {
588 struct $thread * thrd;
589 thrd = pop(lanes.data[idx]);
590
591 push(cltr, thrd);
592
593 // for printing count the number of displaced threads
594 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
595 displaced++;
596 #endif
597 }
598
599 // Unlock the lane
600 __atomic_unlock(&lanes.data[idx].lock);
601
602 // TODO print the queue statistics here
603
604 ^(lanes.data[idx]){};
605 }
606
607 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
608
609 // Allocate new array (uses realloc and memcpies the data)
610 lanes.data = alloc( lanes.count, lanes.data`realloc );
611
612 // Fix the moved data
613 for( idx; (size_t)lanes.count ) {
614 fix(lanes.data[idx]);
615 }
616 }
617
618 fix_times(cltr);
619
620 reassign_cltr_id(cltr);
621
622 // Make sure that everything is consistent
623 /* paranoid */ check( cltr->ready_queue );
624
625 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
626 /* paranoid */ verify( ready_mutate_islocked() );
627}
Note: See TracBrowser for help on using the repository browser.