source: libcfa/src/concurrency/ready_queue.cfa@ c33c2af

ADT ast-experimental enum forall-pointer-decay pthread-emulation qualifiedEnum
Last change on this file since c33c2af was 75c7252, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

CPU workstealing now has correct remote unpark.
  • Property mode set to 100644
File size: 33.3 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// #define USE_CPU_WORK_STEALING
25
26#include "bits/defs.hfa"
27#include "device/cpu.hfa"
28#include "kernel_private.hfa"
29
30#include "stdlib.hfa"
31#include "math.hfa"
32
33#include <errno.h>
34#include <unistd.h>
35
36extern "C" {
37 #include <sys/syscall.h> // __NR_xxx
38}
39
40#include "ready_subqueue.hfa"
41
42static const size_t cache_line_size = 64;
43
44#if !defined(__CFA_NO_STATISTICS__)
45 #define __STATS(...) __VA_ARGS__
46#else
47 #define __STATS(...)
48#endif
49
50// No overriden function, no environment variable, no define
51// fall back to a magic number
52#ifndef __CFA_MAX_PROCESSORS__
53 #define __CFA_MAX_PROCESSORS__ 1024
54#endif
55
56#if defined(USE_CPU_WORK_STEALING)
57 #define READYQ_SHARD_FACTOR 2
58#elif defined(USE_RELAXED_FIFO)
59 #define BIAS 4
60 #define READYQ_SHARD_FACTOR 4
61 #define SEQUENTIAL_SHARD 1
62#elif defined(USE_WORK_STEALING)
63 #define READYQ_SHARD_FACTOR 2
64 #define SEQUENTIAL_SHARD 2
65#else
66 #error no scheduling strategy selected
67#endif
68
69static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
70static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
71static inline struct thread$ * search(struct cluster * cltr);
72static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
73
74
75// returns the maximum number of processors the RWLock support
76__attribute__((weak)) unsigned __max_processors() {
77 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
78 if(!max_cores_s) {
79 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
80 return __CFA_MAX_PROCESSORS__;
81 }
82
83 char * endptr = 0p;
84 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
85 if(max_cores_l < 1 || max_cores_l > 65535) {
86 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
87 return __CFA_MAX_PROCESSORS__;
88 }
89 if('\0' != *endptr) {
90 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
91 return __CFA_MAX_PROCESSORS__;
92 }
93
94 return max_cores_l;
95}
96
97#if defined(CFA_HAVE_LINUX_LIBRSEQ)
98 // No forward declaration needed
99 #define __kernel_rseq_register rseq_register_current_thread
100 #define __kernel_rseq_unregister rseq_unregister_current_thread
101#elif defined(CFA_HAVE_LINUX_RSEQ_H)
102 static void __kernel_raw_rseq_register (void);
103 static void __kernel_raw_rseq_unregister(void);
104
105 #define __kernel_rseq_register __kernel_raw_rseq_register
106 #define __kernel_rseq_unregister __kernel_raw_rseq_unregister
107#else
108 // No forward declaration needed
109 // No initialization needed
110 static inline void noop(void) {}
111
112 #define __kernel_rseq_register noop
113 #define __kernel_rseq_unregister noop
114#endif
115
116//=======================================================================
117// Cluster wide reader-writer lock
118//=======================================================================
119void ?{}(__scheduler_RWLock_t & this) {
120 this.max = __max_processors();
121 this.alloc = 0;
122 this.ready = 0;
123 this.data = alloc(this.max);
124 this.write_lock = false;
125
126 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
127 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
128
129}
130void ^?{}(__scheduler_RWLock_t & this) {
131 free(this.data);
132}
133
134
135//=======================================================================
136// Lock-Free registering/unregistering of threads
137unsigned register_proc_id( void ) with(*__scheduler_lock) {
138 __kernel_rseq_register();
139
140 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
141 bool * handle = (bool *)&kernelTLS().sched_lock;
142
143 // Step - 1 : check if there is already space in the data
144 uint_fast32_t s = ready;
145
146 // Check among all the ready
147 for(uint_fast32_t i = 0; i < s; i++) {
148 bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
149 /* paranoid */ verify( handle != *cell );
150
151 bool * null = 0p; // Re-write every loop since compare thrashes it
152 if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
153 && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
154 /* paranoid */ verify(i < ready);
155 /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
156 return i;
157 }
158 }
159
160 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
161
162 // Step - 2 : F&A to get a new spot in the array.
163 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
164 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
165
166 // Step - 3 : Mark space as used and then publish it.
167 data[n] = handle;
168 while() {
169 unsigned copy = n;
170 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
171 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
172 break;
173 Pause();
174 }
175
176 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
177
178 // Return new spot.
179 /* paranoid */ verify(n < ready);
180 /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
181 return n;
182}
183
184void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
185 /* paranoid */ verify(id < ready);
186 /* paranoid */ verify(id == kernelTLS().sched_id);
187 /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
188
189 bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
190
191 __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
192
193 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
194
195 __kernel_rseq_unregister();
196}
197
198//-----------------------------------------------------------------------
199// Writer side : acquire when changing the ready queue, e.g. adding more
200// queues or removing them.
201uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
202 /* paranoid */ verify( ! __preemption_enabled() );
203 /* paranoid */ verify( ! kernelTLS().sched_lock );
204
205 // Step 1 : lock global lock
206 // It is needed to avoid processors that register mid Critical-Section
207 // to simply lock their own lock and enter.
208 __atomic_acquire( &write_lock );
209
210 // Step 2 : lock per-proc lock
211 // Processors that are currently being registered aren't counted
212 // but can't be in read_lock or in the critical section.
213 // All other processors are counted
214 uint_fast32_t s = ready;
215 for(uint_fast32_t i = 0; i < s; i++) {
216 volatile bool * llock = data[i];
217 if(llock) __atomic_acquire( llock );
218 }
219
220 /* paranoid */ verify( ! __preemption_enabled() );
221 return s;
222}
223
224void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
225 /* paranoid */ verify( ! __preemption_enabled() );
226
227 // Step 1 : release local locks
228 // This must be done while the global lock is held to avoid
229 // threads that where created mid critical section
230 // to race to lock their local locks and have the writer
231 // immidiately unlock them
232 // Alternative solution : return s in write_lock and pass it to write_unlock
233 for(uint_fast32_t i = 0; i < last_s; i++) {
234 volatile bool * llock = data[i];
235 if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
236 }
237
238 // Step 2 : release global lock
239 /*paranoid*/ assert(true == write_lock);
240 __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
241
242 /* paranoid */ verify( ! __preemption_enabled() );
243}
244
245//=======================================================================
246// Cforall Ready Queue used for scheduling
247//=======================================================================
248unsigned long long moving_average(unsigned long long nval, unsigned long long oval) {
249 const unsigned long long tw = 16;
250 const unsigned long long nw = 4;
251 const unsigned long long ow = tw - nw;
252 return ((nw * nval) + (ow * oval)) / tw;
253}
254
255void ?{}(__ready_queue_t & this) with (this) {
256 #if defined(USE_CPU_WORK_STEALING)
257 lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
258 lanes.data = alloc( lanes.count );
259 lanes.tscs = alloc( lanes.count );
260 lanes.help = alloc( cpu_info.hthrd_count );
261
262 for( idx; (size_t)lanes.count ) {
263 (lanes.data[idx]){};
264 lanes.tscs[idx].tv = rdtscl();
265 lanes.tscs[idx].ma = rdtscl();
266 }
267 for( idx; (size_t)cpu_info.hthrd_count ) {
268 lanes.help[idx].src = 0;
269 lanes.help[idx].dst = 0;
270 lanes.help[idx].tri = 0;
271 }
272 #else
273 lanes.data = 0p;
274 lanes.tscs = 0p;
275 lanes.help = 0p;
276 lanes.count = 0;
277 #endif
278}
279
280void ^?{}(__ready_queue_t & this) with (this) {
281 #if !defined(USE_CPU_WORK_STEALING)
282 verify( SEQUENTIAL_SHARD == lanes.count );
283 #endif
284
285 free(lanes.data);
286 free(lanes.tscs);
287 free(lanes.help);
288}
289
290//-----------------------------------------------------------------------
291#if defined(USE_CPU_WORK_STEALING)
292 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
293 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
294
295 processor * const proc = kernelTLS().this_processor;
296 const bool external = (!proc) || (cltr != proc->cltr);
297
298 // Figure out the current cpu and make sure it is valid
299 const int cpu = __kernel_getcpu();
300 /* paranoid */ verify(cpu >= 0);
301 /* paranoid */ verify(cpu < cpu_info.hthrd_count);
302 /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
303
304 // Figure out where thread was last time and make sure it's
305 /* paranoid */ verify(thrd->preferred >= 0);
306 /* paranoid */ verify(thrd->preferred < cpu_info.hthrd_count);
307 /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
308 const int prf = thrd->preferred * READYQ_SHARD_FACTOR;
309
310 const cpu_map_entry_t & map;
311 choose(hint) {
312 case UNPARK_LOCAL : &map = &cpu_info.llc_map[cpu];
313 case UNPARK_REMOTE: &map = &cpu_info.llc_map[prf];
314 }
315 /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
316 /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
317 /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
318
319 const int start = map.self * READYQ_SHARD_FACTOR;
320 unsigned i;
321 do {
322 unsigned r;
323 if(unlikely(external)) { r = __tls_rand(); }
324 else { r = proc->rdq.its++; }
325 choose(hint) {
326 case UNPARK_LOCAL : i = start + (r % READYQ_SHARD_FACTOR);
327 case UNPARK_REMOTE: i = prf + (r % READYQ_SHARD_FACTOR);
328 }
329 // If we can't lock it retry
330 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
331
332 // Actually push it
333 push(lanes.data[i], thrd);
334
335 // Unlock and return
336 __atomic_unlock( &lanes.data[i].lock );
337
338 #if !defined(__CFA_NO_STATISTICS__)
339 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
340 else __tls_stats()->ready.push.local.success++;
341 #endif
342
343 __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);
344
345 }
346
347 // Pop from the ready queue from a given cluster
348 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
349 /* paranoid */ verify( lanes.count > 0 );
350 /* paranoid */ verify( kernelTLS().this_processor );
351
352 const int cpu = __kernel_getcpu();
353 /* paranoid */ verify(cpu >= 0);
354 /* paranoid */ verify(cpu < cpu_info.hthrd_count);
355 /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
356
357 const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
358 /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
359 /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
360 /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
361
362 processor * const proc = kernelTLS().this_processor;
363 const int start = map.self * READYQ_SHARD_FACTOR;
364 const unsigned long long ctsc = rdtscl();
365
366 // Did we already have a help target
367 if(proc->rdq.target == -1u) {
368 unsigned long long max = 0;
369 for(i; READYQ_SHARD_FACTOR) {
370 unsigned long long tsc = moving_average(ctsc - ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
371 if(tsc > max) max = tsc;
372 }
373 proc->rdq.cutoff = (max + 2 * max) / 2;
374 /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
375 /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.
376
377 if(0 == (__tls_rand() % 100)) {
378 proc->rdq.target = __tls_rand() % lanes.count;
379 } else {
380 unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
381 proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
382 /* paranoid */ verify(proc->rdq.target >= (map.start * READYQ_SHARD_FACTOR));
383 /* paranoid */ verify(proc->rdq.target < ((map.start + map.count) * READYQ_SHARD_FACTOR));
384 }
385
386 /* paranoid */ verify(proc->rdq.target != -1u);
387 }
388 else {
389 unsigned long long max = 0;
390 for(i; READYQ_SHARD_FACTOR) {
391 unsigned long long tsc = moving_average(ctsc - ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
392 if(tsc > max) max = tsc;
393 }
394 const unsigned long long cutoff = (max + 2 * max) / 2;
395 {
396 unsigned target = proc->rdq.target;
397 proc->rdq.target = -1u;
398 lanes.help[target / READYQ_SHARD_FACTOR].tri++;
399 if(moving_average(ctsc - lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {
400 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
401 proc->rdq.last = target;
402 if(t) return t;
403 else proc->rdq.target = -1u;
404 }
405 else proc->rdq.target = -1u;
406 }
407
408 unsigned last = proc->rdq.last;
409 if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
410 thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
411 if(t) return t;
412 }
413 else {
414 proc->rdq.last = -1u;
415 }
416 }
417
418 for(READYQ_SHARD_FACTOR) {
419 unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
420 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
421 }
422
423 // All lanes where empty return 0p
424 return 0p;
425 }
426
427 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
428 processor * const proc = kernelTLS().this_processor;
429 unsigned last = proc->rdq.last;
430 if(last != -1u) {
431 struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal));
432 if(t) return t;
433 proc->rdq.last = -1u;
434 }
435
436 unsigned i = __tls_rand() % lanes.count;
437 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
438 }
439 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
440 return search(cltr);
441 }
442#endif
443#if defined(USE_RELAXED_FIFO)
444 //-----------------------------------------------------------------------
445 // get index from random number with or without bias towards queues
446 static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
447 unsigned i;
448 bool local;
449 unsigned rlow = r % BIAS;
450 unsigned rhigh = r / BIAS;
451 if((0 != rlow) && preferred >= 0) {
452 // (BIAS - 1) out of BIAS chances
453 // Use perferred queues
454 i = preferred + (rhigh % READYQ_SHARD_FACTOR);
455 local = true;
456 }
457 else {
458 // 1 out of BIAS chances
459 // Use all queues
460 i = rhigh;
461 local = false;
462 }
463 return [i, local];
464 }
465
466 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
467 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
468
469 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
470 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
471
472 bool local;
473 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
474
475 // Try to pick a lane and lock it
476 unsigned i;
477 do {
478 // Pick the index of a lane
479 unsigned r = __tls_rand_fwd();
480 [i, local] = idx_from_r(r, preferred);
481
482 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
483
484 #if !defined(__CFA_NO_STATISTICS__)
485 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
486 else if(local) __tls_stats()->ready.push.local.attempt++;
487 else __tls_stats()->ready.push.share.attempt++;
488 #endif
489
490 // If we can't lock it retry
491 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
492
493 // Actually push it
494 push(lanes.data[i], thrd);
495
496 // Unlock and return
497 __atomic_unlock( &lanes.data[i].lock );
498
499 // Mark the current index in the tls rng instance as having an item
500 __tls_rand_advance_bck();
501
502 __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);
503
504 // Update statistics
505 #if !defined(__CFA_NO_STATISTICS__)
506 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
507 else if(local) __tls_stats()->ready.push.local.success++;
508 else __tls_stats()->ready.push.share.success++;
509 #endif
510 }
511
512 // Pop from the ready queue from a given cluster
513 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
514 /* paranoid */ verify( lanes.count > 0 );
515 /* paranoid */ verify( kernelTLS().this_processor );
516 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
517
518 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
519 int preferred = kernelTLS().this_processor->rdq.id;
520
521
522 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
523 for(25) {
524 // Pick two lists at random
525 unsigned ri = __tls_rand_bck();
526 unsigned rj = __tls_rand_bck();
527
528 unsigned i, j;
529 __attribute__((unused)) bool locali, localj;
530 [i, locali] = idx_from_r(ri, preferred);
531 [j, localj] = idx_from_r(rj, preferred);
532
533 i %= count;
534 j %= count;
535
536 // try popping from the 2 picked lists
537 struct thread$ * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
538 if(thrd) {
539 return thrd;
540 }
541 }
542
543 // All lanes where empty return 0p
544 return 0p;
545 }
546
547 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
548 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
549 return search(cltr);
550 }
551#endif
552#if defined(USE_WORK_STEALING)
553 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
554 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
555
556 // #define USE_PREFERRED
557 #if !defined(USE_PREFERRED)
558 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
559 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
560 #else
561 unsigned preferred = thrd->preferred;
562 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == -1u || thrd->curr_cluster != cltr;
563 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
564
565 unsigned r = preferred % READYQ_SHARD_FACTOR;
566 const unsigned start = preferred - r;
567 #endif
568
569 // Try to pick a lane and lock it
570 unsigned i;
571 do {
572 #if !defined(__CFA_NO_STATISTICS__)
573 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
574 else __tls_stats()->ready.push.local.attempt++;
575 #endif
576
577 if(unlikely(external)) {
578 i = __tls_rand() % lanes.count;
579 }
580 else {
581 #if !defined(USE_PREFERRED)
582 processor * proc = kernelTLS().this_processor;
583 unsigned r = proc->rdq.its++;
584 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
585 #else
586 i = start + (r++ % READYQ_SHARD_FACTOR);
587 #endif
588 }
589 // If we can't lock it retry
590 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
591
592 // Actually push it
593 push(lanes.data[i], thrd);
594
595 // Unlock and return
596 __atomic_unlock( &lanes.data[i].lock );
597
598 #if !defined(__CFA_NO_STATISTICS__)
599 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
600 else __tls_stats()->ready.push.local.success++;
601 #endif
602
603 __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);
604 }
605
606 // Pop from the ready queue from a given cluster
607 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
608 /* paranoid */ verify( lanes.count > 0 );
609 /* paranoid */ verify( kernelTLS().this_processor );
610 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
611
612 processor * proc = kernelTLS().this_processor;
613
614 if(proc->rdq.target == -1u) {
615 unsigned long long min = ts(lanes.data[proc->rdq.id]);
616 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
617 unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
618 if(tsc < min) min = tsc;
619 }
620 proc->rdq.cutoff = min;
621 proc->rdq.target = __tls_rand() % lanes.count;
622 }
623 else {
624 unsigned target = proc->rdq.target;
625 proc->rdq.target = -1u;
626 const unsigned long long bias = 0; //2_500_000_000;
627 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
628 if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
629 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
630 if(t) return t;
631 }
632 }
633
634 for(READYQ_SHARD_FACTOR) {
635 unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
636 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
637 }
638 return 0p;
639 }
640
641 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
642 unsigned i = __tls_rand() % lanes.count;
643 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
644 }
645
646 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
647 return search(cltr);
648 }
649#endif
650
651//=======================================================================
652// Various Ready Queue utilities
653//=======================================================================
654// these function work the same or almost the same
655// whether they are using work-stealing or relaxed fifo scheduling
656
657//-----------------------------------------------------------------------
658// try to pop from a lane given by index w
659static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
660 __STATS( stats.attempt++; )
661
662 // Get relevant elements locally
663 __intrusive_lane_t & lane = lanes.data[w];
664
665 // If list looks empty retry
666 if( is_empty(lane) ) {
667 return 0p;
668 }
669
670 // If we can't get the lock retry
671 if( !__atomic_try_acquire(&lane.lock) ) {
672 return 0p;
673 }
674
675 // If list is empty, unlock and retry
676 if( is_empty(lane) ) {
677 __atomic_unlock(&lane.lock);
678 return 0p;
679 }
680
681 // Actually pop the list
682 struct thread$ * thrd;
683 unsigned long long tsc_before = ts(lane);
684 unsigned long long tsv;
685 [thrd, tsv] = pop(lane);
686
687 /* paranoid */ verify(thrd);
688 /* paranoid */ verify(tsv);
689 /* paranoid */ verify(lane.lock);
690
691 // Unlock and return
692 __atomic_unlock(&lane.lock);
693
694 // Update statistics
695 __STATS( stats.success++; )
696
697 #if defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
698 unsigned long long now = rdtscl();
699 lanes.tscs[w].tv = tsv;
700 lanes.tscs[w].ma = moving_average(now > tsc_before ? now - tsc_before : 0, lanes.tscs[w].ma);
701 #endif
702
703 #if defined(USE_CPU_WORK_STEALING)
704 thrd->preferred = w / READYQ_SHARD_FACTOR;
705 #else
706 thrd->preferred = w;
707 #endif
708
709 // return the popped thread
710 return thrd;
711}
712
713//-----------------------------------------------------------------------
714// try to pop from any lanes making sure you don't miss any threads push
715// before the start of the function
716static inline struct thread$ * search(struct cluster * cltr) with (cltr->ready_queue) {
717 /* paranoid */ verify( lanes.count > 0 );
718 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
719 unsigned offset = __tls_rand();
720 for(i; count) {
721 unsigned idx = (offset + i) % count;
722 struct thread$ * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
723 if(thrd) {
724 return thrd;
725 }
726 }
727
728 // All lanes where empty return 0p
729 return 0p;
730}
731
732//-----------------------------------------------------------------------
733// get preferred ready for new thread
734unsigned ready_queue_new_preferred() {
735 unsigned pref = 0;
736 if(struct thread$ * thrd = publicTLS_get( this_thread )) {
737 pref = thrd->preferred;
738 }
739 else {
740 #if defined(USE_CPU_WORK_STEALING)
741 pref = __kernel_getcpu();
742 #endif
743 }
744
745 #if defined(USE_CPU_WORK_STEALING)
746 /* paranoid */ verify(pref >= 0);
747 /* paranoid */ verify(pref < cpu_info.hthrd_count);
748 #endif
749
750 return pref;
751}
752
753//-----------------------------------------------------------------------
754// Check that all the intrusive queues in the data structure are still consistent
755static void check( __ready_queue_t & q ) with (q) {
756 #if defined(__CFA_WITH_VERIFY__)
757 {
758 for( idx ; lanes.count ) {
759 __intrusive_lane_t & sl = lanes.data[idx];
760 assert(!lanes.data[idx].lock);
761
762 if(is_empty(sl)) {
763 assert( sl.anchor.next == 0p );
764 assert( sl.anchor.ts == -1llu );
765 assert( mock_head(sl) == sl.prev );
766 } else {
767 assert( sl.anchor.next != 0p );
768 assert( sl.anchor.ts != -1llu );
769 assert( mock_head(sl) != sl.prev );
770 }
771 }
772 }
773 #endif
774}
775
776//-----------------------------------------------------------------------
777// Given 2 indexes, pick the list with the oldest push an try to pop from it
778static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
779 // Pick the bet list
780 int w = i;
781 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
782 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
783 }
784
785 return try_pop(cltr, w __STATS(, stats));
786}
787
788// Call this function of the intrusive list was moved using memcpy
789// fixes the list so that the pointers back to anchors aren't left dangling
790static inline void fix(__intrusive_lane_t & ll) {
791 if(is_empty(ll)) {
792 verify(ll.anchor.next == 0p);
793 ll.prev = mock_head(ll);
794 }
795}
796
797static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
798 processor * it = &list`first;
799 for(unsigned i = 0; i < count; i++) {
800 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
801 it->rdq.id = value;
802 it->rdq.target = -1u;
803 value += READYQ_SHARD_FACTOR;
804 it = &(*it)`next;
805 }
806}
807
808static void reassign_cltr_id(struct cluster * cltr) {
809 unsigned preferred = 0;
810 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
811 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
812}
813
814static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
815 #if defined(USE_WORK_STEALING)
816 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
817 for(i; lanes.count) {
818 unsigned long long tsc1 = ts(lanes.data[i]);
819 unsigned long long tsc2 = rdtscl();
820 lanes.tscs[i].tv = min(tsc1, tsc2);
821 }
822 #endif
823}
824
825#if defined(USE_CPU_WORK_STEALING)
826 // ready_queue size is fixed in this case
827 void ready_queue_grow(struct cluster * cltr) {}
828 void ready_queue_shrink(struct cluster * cltr) {}
829#else
830 // Grow the ready queue
831 void ready_queue_grow(struct cluster * cltr) {
832 size_t ncount;
833 int target = cltr->procs.total;
834
835 /* paranoid */ verify( ready_mutate_islocked() );
836 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
837
838 // Make sure that everything is consistent
839 /* paranoid */ check( cltr->ready_queue );
840
841 // grow the ready queue
842 with( cltr->ready_queue ) {
843 // Find new count
844 // Make sure we always have atleast 1 list
845 if(target >= 2) {
846 ncount = target * READYQ_SHARD_FACTOR;
847 } else {
848 ncount = SEQUENTIAL_SHARD;
849 }
850
851 // Allocate new array (uses realloc and memcpies the data)
852 lanes.data = alloc( ncount, lanes.data`realloc );
853
854 // Fix the moved data
855 for( idx; (size_t)lanes.count ) {
856 fix(lanes.data[idx]);
857 }
858
859 // Construct new data
860 for( idx; (size_t)lanes.count ~ ncount) {
861 (lanes.data[idx]){};
862 }
863
864 // Update original
865 lanes.count = ncount;
866 }
867
868 fix_times(cltr);
869
870 reassign_cltr_id(cltr);
871
872 // Make sure that everything is consistent
873 /* paranoid */ check( cltr->ready_queue );
874
875 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
876
877 /* paranoid */ verify( ready_mutate_islocked() );
878 }
879
880 // Shrink the ready queue
881 void ready_queue_shrink(struct cluster * cltr) {
882 /* paranoid */ verify( ready_mutate_islocked() );
883 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
884
885 // Make sure that everything is consistent
886 /* paranoid */ check( cltr->ready_queue );
887
888 int target = cltr->procs.total;
889
890 with( cltr->ready_queue ) {
891 // Remember old count
892 size_t ocount = lanes.count;
893
894 // Find new count
895 // Make sure we always have atleast 1 list
896 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
897 /* paranoid */ verify( ocount >= lanes.count );
898 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
899
900 // for printing count the number of displaced threads
901 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
902 __attribute__((unused)) size_t displaced = 0;
903 #endif
904
905 // redistribute old data
906 for( idx; (size_t)lanes.count ~ ocount) {
907 // Lock is not strictly needed but makes checking invariants much easier
908 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
909 verify(locked);
910
911 // As long as we can pop from this lane to push the threads somewhere else in the queue
912 while(!is_empty(lanes.data[idx])) {
913 struct thread$ * thrd;
914 unsigned long long _;
915 [thrd, _] = pop(lanes.data[idx]);
916
917 push(cltr, thrd, true);
918
919 // for printing count the number of displaced threads
920 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
921 displaced++;
922 #endif
923 }
924
925 // Unlock the lane
926 __atomic_unlock(&lanes.data[idx].lock);
927
928 // TODO print the queue statistics here
929
930 ^(lanes.data[idx]){};
931 }
932
933 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
934
935 // Allocate new array (uses realloc and memcpies the data)
936 lanes.data = alloc( lanes.count, lanes.data`realloc );
937
938 // Fix the moved data
939 for( idx; (size_t)lanes.count ) {
940 fix(lanes.data[idx]);
941 }
942 }
943
944 fix_times(cltr);
945
946 reassign_cltr_id(cltr);
947
948 // Make sure that everything is consistent
949 /* paranoid */ check( cltr->ready_queue );
950
951 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
952 /* paranoid */ verify( ready_mutate_islocked() );
953 }
954#endif
955
956#if !defined(__CFA_NO_STATISTICS__)
957 unsigned cnt(const __ready_queue_t & this, unsigned idx) {
958 /* paranoid */ verify(this.lanes.count > idx);
959 return this.lanes.data[idx].cnt;
960 }
961#endif
962
963
964#if defined(CFA_HAVE_LINUX_LIBRSEQ)
965 // No definition needed
966#elif defined(CFA_HAVE_LINUX_RSEQ_H)
967
968 #if defined( __x86_64 ) || defined( __i386 )
969 #define RSEQ_SIG 0x53053053
970 #elif defined( __ARM_ARCH )
971 #ifdef __ARMEB__
972 #define RSEQ_SIG 0xf3def5e7 /* udf #24035 ; 0x5de3 (ARMv6+) */
973 #else
974 #define RSEQ_SIG 0xe7f5def3 /* udf #24035 ; 0x5de3 */
975 #endif
976 #endif
977
978 extern void __disable_interrupts_hard();
979 extern void __enable_interrupts_hard();
980
981 static void __kernel_raw_rseq_register (void) {
982 /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
983
984 // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
985 int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
986 if(ret != 0) {
987 int e = errno;
988 switch(e) {
989 case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
990 case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
991 case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
992 case EBUSY : abort("KERNEL ERROR: rseq register already registered");
993 case EPERM : abort("KERNEL ERROR: rseq register sig argument on unregistration does not match the signature received on registration");
994 default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
995 }
996 }
997 }
998
999 static void __kernel_raw_rseq_unregister(void) {
1000 /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
1001
1002 // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
1003 int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
1004 if(ret != 0) {
1005 int e = errno;
1006 switch(e) {
1007 case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
1008 case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
1009 case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
1010 case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
1011 case EPERM : abort("KERNEL ERROR: rseq unregister sig argument on unregistration does not match the signature received on registration");
1012 default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
1013 }
1014 }
1015 }
1016#else
1017 // No definition needed
1018#endif
Note: See TracBrowser for help on using the repository browser.