source: libcfa/src/concurrency/ready_queue.cfa@ 25337e0

ADT ast-experimental enum forall-pointer-decay pthread-emulation qualifiedEnum
Last change on this file since 25337e0 was 25337e0, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Fix some problems with compilation of per-cpu ready queue.
  • Property mode set to 100644
File size: 39.8 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 _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#define USE_AWARE_STEALING
26
27#include "bits/defs.hfa"
28#include "device/cpu.hfa"
29#include "kernel_private.hfa"
30
31#include "stdlib.hfa"
32#include "limits.hfa"
33#include "math.hfa"
34
35#include <errno.h>
36#include <unistd.h>
37
38extern "C" {
39 #include <sys/syscall.h> // __NR_xxx
40}
41
42#include "ready_subqueue.hfa"
43
44static const size_t cache_line_size = 64;
45
46#if !defined(__CFA_NO_STATISTICS__)
47 #define __STATS(...) __VA_ARGS__
48#else
49 #define __STATS(...)
50#endif
51
52// No overriden function, no environment variable, no define
53// fall back to a magic number
54#ifndef __CFA_MAX_PROCESSORS__
55 #define __CFA_MAX_PROCESSORS__ 1024
56#endif
57
58#if defined(USE_AWARE_STEALING)
59 #define READYQ_SHARD_FACTOR 2
60 #define SEQUENTIAL_SHARD 2
61#elif defined(USE_CPU_WORK_STEALING)
62 #define READYQ_SHARD_FACTOR 2
63#elif defined(USE_RELAXED_FIFO)
64 #define BIAS 4
65 #define READYQ_SHARD_FACTOR 4
66 #define SEQUENTIAL_SHARD 1
67#elif defined(USE_WORK_STEALING)
68 #define READYQ_SHARD_FACTOR 2
69 #define SEQUENTIAL_SHARD 2
70#else
71 #error no scheduling strategy selected
72#endif
73
74static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
75static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
76static inline struct thread$ * search(struct cluster * cltr);
77static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
78
79
80// returns the maximum number of processors the RWLock support
81__attribute__((weak)) unsigned __max_processors() {
82 const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
83 if(!max_cores_s) {
84 __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
85 return __CFA_MAX_PROCESSORS__;
86 }
87
88 char * endptr = 0p;
89 long int max_cores_l = strtol(max_cores_s, &endptr, 10);
90 if(max_cores_l < 1 || max_cores_l > 65535) {
91 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
92 return __CFA_MAX_PROCESSORS__;
93 }
94 if('\0' != *endptr) {
95 __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
96 return __CFA_MAX_PROCESSORS__;
97 }
98
99 return max_cores_l;
100}
101
102#if defined(CFA_HAVE_LINUX_LIBRSEQ)
103 // No forward declaration needed
104 #define __kernel_rseq_register rseq_register_current_thread
105 #define __kernel_rseq_unregister rseq_unregister_current_thread
106#elif defined(CFA_HAVE_LINUX_RSEQ_H)
107 static void __kernel_raw_rseq_register (void);
108 static void __kernel_raw_rseq_unregister(void);
109
110 #define __kernel_rseq_register __kernel_raw_rseq_register
111 #define __kernel_rseq_unregister __kernel_raw_rseq_unregister
112#else
113 // No forward declaration needed
114 // No initialization needed
115 static inline void noop(void) {}
116
117 #define __kernel_rseq_register noop
118 #define __kernel_rseq_unregister noop
119#endif
120
121//=======================================================================
122// Cluster wide reader-writer lock
123//=======================================================================
124void ?{}(__scheduler_RWLock_t & this) {
125 this.max = __max_processors();
126 this.alloc = 0;
127 this.ready = 0;
128 this.data = alloc(this.max);
129 this.write_lock = false;
130
131 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
132 /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
133
134}
135void ^?{}(__scheduler_RWLock_t & this) {
136 free(this.data);
137}
138
139
140//=======================================================================
141// Lock-Free registering/unregistering of threads
142unsigned register_proc_id( void ) with(*__scheduler_lock) {
143 __kernel_rseq_register();
144
145 bool * handle = (bool *)&kernelTLS().sched_lock;
146
147 // Step - 1 : check if there is already space in the data
148 uint_fast32_t s = ready;
149
150 // Check among all the ready
151 for(uint_fast32_t i = 0; i < s; i++) {
152 bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
153 /* paranoid */ verify( handle != *cell );
154
155 bool * null = 0p; // Re-write every loop since compare thrashes it
156 if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
157 && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
158 /* paranoid */ verify(i < ready);
159 /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
160 return i;
161 }
162 }
163
164 if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
165
166 // Step - 2 : F&A to get a new spot in the array.
167 uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
168 if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
169
170 // Step - 3 : Mark space as used and then publish it.
171 data[n] = handle;
172 while() {
173 unsigned copy = n;
174 if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
175 && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
176 break;
177 Pause();
178 }
179
180 // Return new spot.
181 /* paranoid */ verify(n < ready);
182 /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
183 return n;
184}
185
186void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
187 /* paranoid */ verify(id < ready);
188 /* paranoid */ verify(id == kernelTLS().sched_id);
189 /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
190
191 bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
192
193 __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
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// caches handling
247
248struct __attribute__((aligned(128))) __ready_queue_caches_t {
249 // Count States:
250 // - 0 : No one is looking after this cache
251 // - 1 : No one is looking after this cache, BUT it's not empty
252 // - 2+ : At least one processor is looking after this cache
253 volatile unsigned count;
254};
255
256void ?{}(__ready_queue_caches_t & this) { this.count = 0; }
257void ^?{}(__ready_queue_caches_t & this) {}
258
259static inline void depart(__ready_queue_caches_t & cache) {
260 /* paranoid */ verify( cache.count > 1);
261 __atomic_fetch_add(&cache.count, -1, __ATOMIC_SEQ_CST);
262 /* paranoid */ verify( cache.count != 0);
263 /* paranoid */ verify( cache.count < 65536 ); // This verify assumes no cluster will have more than 65000 kernel threads mapped to a single cache, which could be correct but is super weird.
264}
265
266static inline void arrive(__ready_queue_caches_t & cache) {
267 // for() {
268 // unsigned expected = cache.count;
269 // unsigned desired = 0 == expected ? 2 : expected + 1;
270 // }
271}
272
273//=======================================================================
274// Cforall Ready Queue used for scheduling
275//=======================================================================
276unsigned long long moving_average(unsigned long long currtsc, unsigned long long instsc, unsigned long long old_avg) {
277 /* paranoid */ verifyf( currtsc < 45000000000000000, "Suspiciously large current time: %'llu (%llx)\n", currtsc, currtsc );
278 /* paranoid */ verifyf( instsc < 45000000000000000, "Suspiciously large insert time: %'llu (%llx)\n", instsc, instsc );
279 /* paranoid */ verifyf( old_avg < 15000000000000, "Suspiciously large previous average: %'llu (%llx)\n", old_avg, old_avg );
280
281 const unsigned long long new_val = currtsc > instsc ? currtsc - instsc : 0;
282 const unsigned long long total_weight = 16;
283 const unsigned long long new_weight = 4;
284 const unsigned long long old_weight = total_weight - new_weight;
285 const unsigned long long ret = ((new_weight * new_val) + (old_weight * old_avg)) / total_weight;
286 return ret;
287}
288
289void ?{}(__ready_queue_t & this) with (this) {
290 #if defined(USE_CPU_WORK_STEALING)
291 lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
292 lanes.data = alloc( lanes.count );
293 lanes.tscs = alloc( lanes.count );
294 lanes.help = alloc( cpu_info.hthrd_count );
295
296 for( idx; (size_t)lanes.count ) {
297 (lanes.data[idx]){};
298 lanes.tscs[idx].tv = rdtscl();
299 lanes.tscs[idx].ma = rdtscl();
300 }
301 for( idx; (size_t)cpu_info.hthrd_count ) {
302 lanes.help[idx].src = 0;
303 lanes.help[idx].dst = 0;
304 lanes.help[idx].tri = 0;
305 }
306 #else
307 lanes.data = 0p;
308 lanes.tscs = 0p;
309 lanes.caches = 0p;
310 lanes.help = 0p;
311 lanes.count = 0;
312 #endif
313}
314
315void ^?{}(__ready_queue_t & this) with (this) {
316 #if !defined(USE_CPU_WORK_STEALING)
317 verify( SEQUENTIAL_SHARD == lanes.count );
318 #endif
319
320 free(lanes.data);
321 free(lanes.tscs);
322 free(lanes.caches);
323 free(lanes.help);
324}
325
326//-----------------------------------------------------------------------
327#if defined(USE_AWARE_STEALING)
328 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
329 processor * const proc = kernelTLS().this_processor;
330 const bool external = (!proc) || (cltr != proc->cltr);
331 const bool remote = hint == UNPARK_REMOTE;
332
333 unsigned i;
334 if( external || remote ) {
335 // Figure out where thread was last time and make sure it's valid
336 /* paranoid */ verify(thrd->preferred >= 0);
337 if(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count) {
338 /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
339 unsigned start = thrd->preferred * READYQ_SHARD_FACTOR;
340 do {
341 unsigned r = __tls_rand();
342 i = start + (r % READYQ_SHARD_FACTOR);
343 /* paranoid */ verify( i < lanes.count );
344 // If we can't lock it retry
345 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
346 } else {
347 do {
348 i = __tls_rand() % lanes.count;
349 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
350 }
351 } else {
352 do {
353 unsigned r = proc->rdq.its++;
354 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
355 /* paranoid */ verify( i < lanes.count );
356 // If we can't lock it retry
357 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
358 }
359
360 // Actually push it
361 push(lanes.data[i], thrd);
362
363 // Unlock and return
364 __atomic_unlock( &lanes.data[i].lock );
365
366 #if !defined(__CFA_NO_STATISTICS__)
367 if(unlikely(external || remote)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
368 else __tls_stats()->ready.push.local.success++;
369 #endif
370 }
371
372 static inline unsigned long long calc_cutoff(const unsigned long long ctsc, const processor * proc, __ready_queue_t & rdq) {
373 unsigned start = proc->rdq.id;
374 unsigned long long max = 0;
375 for(i; READYQ_SHARD_FACTOR) {
376 unsigned long long ptsc = ts(rdq.lanes.data[start + i]);
377 if(ptsc != -1ull) {
378 /* paranoid */ verify( start + i < rdq.lanes.count );
379 unsigned long long tsc = moving_average(ctsc, ptsc, rdq.lanes.tscs[start + i].ma);
380 if(tsc > max) max = tsc;
381 }
382 }
383 return (max + 2 * max) / 2;
384 }
385
386 __attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
387 /* paranoid */ verify( lanes.count > 0 );
388 /* paranoid */ verify( kernelTLS().this_processor );
389 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
390
391 processor * const proc = kernelTLS().this_processor;
392 unsigned this = proc->rdq.id;
393 /* paranoid */ verify( this < lanes.count );
394 __cfadbg_print_safe(ready_queue, "Kernel : pop from %u\n", this);
395
396 // Figure out the current cpu and make sure it is valid
397 const int cpu = __kernel_getcpu();
398 /* paranoid */ verify(cpu >= 0);
399 /* paranoid */ verify(cpu < cpu_info.hthrd_count);
400 unsigned this_cache = cpu_info.llc_map[cpu].cache;
401 __atomic_store_n(&lanes.caches[this / READYQ_SHARD_FACTOR].id, this_cache, __ATOMIC_RELAXED);
402
403 const unsigned long long ctsc = rdtscl();
404
405 if(proc->rdq.target == MAX) {
406 uint64_t chaos = __tls_rand();
407 unsigned ext = chaos & 0xff;
408 unsigned other = (chaos >> 8) % (lanes.count);
409
410 if(ext < 3 || __atomic_load_n(&lanes.caches[other / READYQ_SHARD_FACTOR].id, __ATOMIC_RELAXED) == this_cache) {
411 proc->rdq.target = other;
412 }
413 }
414 else {
415 const unsigned target = proc->rdq.target;
416 __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, lanes.tscs[target].tv);
417 /* paranoid */ verify( lanes.tscs[target].tv != MAX );
418 if(target < lanes.count) {
419 const unsigned long long cutoff = calc_cutoff(ctsc, proc, cltr->ready_queue);
420 const unsigned long long age = moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma);
421 __cfadbg_print_safe(ready_queue, "Kernel : Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no");
422 if(age > cutoff) {
423 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
424 if(t) return t;
425 }
426 }
427 proc->rdq.target = MAX;
428 }
429
430 for(READYQ_SHARD_FACTOR) {
431 unsigned i = this + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
432 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
433 }
434
435 // All lanes where empty return 0p
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 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
444 return search(cltr);
445 }
446#endif
447#if defined(USE_CPU_WORK_STEALING)
448 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
449 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
450
451 processor * const proc = kernelTLS().this_processor;
452 const bool external = (!proc) || (cltr != proc->cltr);
453
454 // Figure out the current cpu and make sure it is valid
455 const int cpu = __kernel_getcpu();
456 /* paranoid */ verify(cpu >= 0);
457 /* paranoid */ verify(cpu < cpu_info.hthrd_count);
458 /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
459
460 // Figure out where thread was last time and make sure it's
461 /* paranoid */ verify(thrd->preferred >= 0);
462 /* paranoid */ verify(thrd->preferred < cpu_info.hthrd_count);
463 /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
464 const int prf = thrd->preferred * READYQ_SHARD_FACTOR;
465
466 const cpu_map_entry_t & map;
467 choose(hint) {
468 case UNPARK_LOCAL : &map = &cpu_info.llc_map[cpu];
469 case UNPARK_REMOTE: &map = &cpu_info.llc_map[prf];
470 }
471 /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
472 /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
473 /* 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);
474
475 const int start = map.self * READYQ_SHARD_FACTOR;
476 unsigned i;
477 do {
478 unsigned r;
479 if(unlikely(external)) { r = __tls_rand(); }
480 else { r = proc->rdq.its++; }
481 choose(hint) {
482 case UNPARK_LOCAL : i = start + (r % READYQ_SHARD_FACTOR);
483 case UNPARK_REMOTE: i = prf + (r % READYQ_SHARD_FACTOR);
484 }
485 // If we can't lock it retry
486 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
487
488 // Actually push it
489 push(lanes.data[i], thrd);
490
491 // Unlock and return
492 __atomic_unlock( &lanes.data[i].lock );
493
494 #if !defined(__CFA_NO_STATISTICS__)
495 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
496 else __tls_stats()->ready.push.local.success++;
497 #endif
498
499 __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);
500
501 }
502
503 // Pop from the ready queue from a given cluster
504 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
505 /* paranoid */ verify( lanes.count > 0 );
506 /* paranoid */ verify( kernelTLS().this_processor );
507
508 processor * const proc = kernelTLS().this_processor;
509 const int cpu = __kernel_getcpu();
510 /* paranoid */ verify(cpu >= 0);
511 /* paranoid */ verify(cpu < cpu_info.hthrd_count);
512 /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
513
514 const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
515 /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
516 /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
517 /* 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);
518
519 const int start = map.self * READYQ_SHARD_FACTOR;
520 const unsigned long long ctsc = rdtscl();
521
522 // Did we already have a help target
523 if(proc->rdq.target == MAX) {
524 unsigned long long max = 0;
525 for(i; READYQ_SHARD_FACTOR) {
526 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
527 if(tsc > max) max = tsc;
528 }
529 // proc->rdq.cutoff = (max + 2 * max) / 2;
530 /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
531 /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.
532
533 if(0 == (__tls_rand() % 100)) {
534 proc->rdq.target = __tls_rand() % lanes.count;
535 } else {
536 unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
537 proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
538 /* paranoid */ verify(proc->rdq.target >= (map.start * READYQ_SHARD_FACTOR));
539 /* paranoid */ verify(proc->rdq.target < ((map.start + map.count) * READYQ_SHARD_FACTOR));
540 }
541
542 /* paranoid */ verify(proc->rdq.target != MAX);
543 }
544 else {
545 unsigned long long max = 0;
546 for(i; READYQ_SHARD_FACTOR) {
547 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
548 if(tsc > max) max = tsc;
549 }
550 const unsigned long long cutoff = (max + 2 * max) / 2;
551 {
552 unsigned target = proc->rdq.target;
553 proc->rdq.target = MAX;
554 lanes.help[target / READYQ_SHARD_FACTOR].tri++;
555 if(moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {
556 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
557 proc->rdq.last = target;
558 if(t) return t;
559 }
560 proc->rdq.target = MAX;
561 }
562
563 unsigned last = proc->rdq.last;
564 if(last != MAX && moving_average(ctsc, lanes.tscs[last].tv, lanes.tscs[last].ma) > cutoff) {
565 thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
566 if(t) return t;
567 }
568 else {
569 proc->rdq.last = MAX;
570 }
571 }
572
573 for(READYQ_SHARD_FACTOR) {
574 unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
575 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
576 }
577
578 // All lanes where empty return 0p
579 return 0p;
580 }
581
582 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
583 processor * const proc = kernelTLS().this_processor;
584 unsigned last = proc->rdq.last;
585 if(last != MAX) {
586 struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal));
587 if(t) return t;
588 proc->rdq.last = MAX;
589 }
590
591 unsigned i = __tls_rand() % lanes.count;
592 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
593 }
594 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
595 return search(cltr);
596 }
597#endif
598#if defined(USE_RELAXED_FIFO)
599 //-----------------------------------------------------------------------
600 // get index from random number with or without bias towards queues
601 static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
602 unsigned i;
603 bool local;
604 unsigned rlow = r % BIAS;
605 unsigned rhigh = r / BIAS;
606 if((0 != rlow) && preferred >= 0) {
607 // (BIAS - 1) out of BIAS chances
608 // Use perferred queues
609 i = preferred + (rhigh % READYQ_SHARD_FACTOR);
610 local = true;
611 }
612 else {
613 // 1 out of BIAS chances
614 // Use all queues
615 i = rhigh;
616 local = false;
617 }
618 return [i, local];
619 }
620
621 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
622 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
623
624 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
625 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
626
627 bool local;
628 int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
629
630 // Try to pick a lane and lock it
631 unsigned i;
632 do {
633 // Pick the index of a lane
634 unsigned r = __tls_rand_fwd();
635 [i, local] = idx_from_r(r, preferred);
636
637 i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
638
639 #if !defined(__CFA_NO_STATISTICS__)
640 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
641 else if(local) __tls_stats()->ready.push.local.attempt++;
642 else __tls_stats()->ready.push.share.attempt++;
643 #endif
644
645 // If we can't lock it retry
646 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
647
648 // Actually push it
649 push(lanes.data[i], thrd);
650
651 // Unlock and return
652 __atomic_unlock( &lanes.data[i].lock );
653
654 // Mark the current index in the tls rng instance as having an item
655 __tls_rand_advance_bck();
656
657 __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);
658
659 // Update statistics
660 #if !defined(__CFA_NO_STATISTICS__)
661 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
662 else if(local) __tls_stats()->ready.push.local.success++;
663 else __tls_stats()->ready.push.share.success++;
664 #endif
665 }
666
667 // Pop from the ready queue from a given cluster
668 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
669 /* paranoid */ verify( lanes.count > 0 );
670 /* paranoid */ verify( kernelTLS().this_processor );
671 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
672
673 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
674 int preferred = kernelTLS().this_processor->rdq.id;
675
676
677 // As long as the list is not empty, try finding a lane that isn't empty and pop from it
678 for(25) {
679 // Pick two lists at random
680 unsigned ri = __tls_rand_bck();
681 unsigned rj = __tls_rand_bck();
682
683 unsigned i, j;
684 __attribute__((unused)) bool locali, localj;
685 [i, locali] = idx_from_r(ri, preferred);
686 [j, localj] = idx_from_r(rj, preferred);
687
688 i %= count;
689 j %= count;
690
691 // try popping from the 2 picked lists
692 struct thread$ * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
693 if(thrd) {
694 return thrd;
695 }
696 }
697
698 // All lanes where empty return 0p
699 return 0p;
700 }
701
702 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
703 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
704 return search(cltr);
705 }
706#endif
707#if defined(USE_WORK_STEALING)
708 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
709 __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
710
711 // #define USE_PREFERRED
712 #if !defined(USE_PREFERRED)
713 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
714 /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
715 #else
716 unsigned preferred = thrd->preferred;
717 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == MAX || thrd->curr_cluster != cltr;
718 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
719
720 unsigned r = preferred % READYQ_SHARD_FACTOR;
721 const unsigned start = preferred - r;
722 #endif
723
724 // Try to pick a lane and lock it
725 unsigned i;
726 do {
727 #if !defined(__CFA_NO_STATISTICS__)
728 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
729 else __tls_stats()->ready.push.local.attempt++;
730 #endif
731
732 if(unlikely(external)) {
733 i = __tls_rand() % lanes.count;
734 }
735 else {
736 #if !defined(USE_PREFERRED)
737 processor * proc = kernelTLS().this_processor;
738 unsigned r = proc->rdq.its++;
739 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
740 #else
741 i = start + (r++ % READYQ_SHARD_FACTOR);
742 #endif
743 }
744 // If we can't lock it retry
745 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
746
747 // Actually push it
748 push(lanes.data[i], thrd);
749
750 // Unlock and return
751 __atomic_unlock( &lanes.data[i].lock );
752
753 #if !defined(__CFA_NO_STATISTICS__)
754 if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
755 else __tls_stats()->ready.push.local.success++;
756 #endif
757
758 __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);
759 }
760
761 // Pop from the ready queue from a given cluster
762 __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
763 /* paranoid */ verify( lanes.count > 0 );
764 /* paranoid */ verify( kernelTLS().this_processor );
765 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
766
767 processor * proc = kernelTLS().this_processor;
768
769 if(proc->rdq.target == MAX) {
770 unsigned long long min = ts(lanes.data[proc->rdq.id]);
771 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
772 unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
773 if(tsc < min) min = tsc;
774 }
775 proc->rdq.cutoff = min;
776 proc->rdq.target = __tls_rand() % lanes.count;
777 }
778 else {
779 unsigned target = proc->rdq.target;
780 proc->rdq.target = MAX;
781 const unsigned long long bias = 0; //2_500_000_000;
782 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
783 if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
784 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
785 if(t) return t;
786 }
787 }
788
789 for(READYQ_SHARD_FACTOR) {
790 unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
791 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
792 }
793 return 0p;
794 }
795
796 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
797 unsigned i = __tls_rand() % lanes.count;
798 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
799 }
800
801 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
802 return search(cltr);
803 }
804#endif
805
806//=======================================================================
807// Various Ready Queue utilities
808//=======================================================================
809// these function work the same or almost the same
810// whether they are using work-stealing or relaxed fifo scheduling
811
812//-----------------------------------------------------------------------
813// try to pop from a lane given by index w
814static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
815 /* paranoid */ verify( w < lanes.count );
816 __STATS( stats.attempt++; )
817
818 // Get relevant elements locally
819 __intrusive_lane_t & lane = lanes.data[w];
820
821 // If list looks empty retry
822 if( is_empty(lane) ) {
823 return 0p;
824 }
825
826 // If we can't get the lock retry
827 if( !__atomic_try_acquire(&lane.lock) ) {
828 return 0p;
829 }
830
831 // If list is empty, unlock and retry
832 if( is_empty(lane) ) {
833 __atomic_unlock(&lane.lock);
834 return 0p;
835 }
836
837 // Actually pop the list
838 struct thread$ * thrd;
839 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
840 unsigned long long tsc_before = ts(lane);
841 #endif
842 unsigned long long tsv;
843 [thrd, tsv] = pop(lane);
844
845 /* paranoid */ verify(thrd);
846 /* paranoid */ verify(tsv);
847 /* paranoid */ verify(lane.lock);
848
849 // Unlock and return
850 __atomic_unlock(&lane.lock);
851
852 // Update statistics
853 __STATS( stats.success++; )
854
855 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
856 if (tsv != MAX) {
857 unsigned long long now = rdtscl();
858 unsigned long long pma = __atomic_load_n(&lanes.tscs[w].ma, __ATOMIC_RELAXED);
859 __atomic_store_n(&lanes.tscs[w].tv, tsv, __ATOMIC_RELAXED);
860 __atomic_store_n(&lanes.tscs[w].ma, moving_average(now, tsc_before, pma), __ATOMIC_RELAXED);
861 }
862 #endif
863
864 #if defined(USE_AWARE_STEALING) || defined(USE_CPU_WORK_STEALING)
865 thrd->preferred = w / READYQ_SHARD_FACTOR;
866 #else
867 thrd->preferred = w;
868 #endif
869
870 // return the popped thread
871 return thrd;
872}
873
874//-----------------------------------------------------------------------
875// try to pop from any lanes making sure you don't miss any threads push
876// before the start of the function
877static inline struct thread$ * search(struct cluster * cltr) with (cltr->ready_queue) {
878 /* paranoid */ verify( lanes.count > 0 );
879 unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
880 unsigned offset = __tls_rand();
881 for(i; count) {
882 unsigned idx = (offset + i) % count;
883 struct thread$ * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
884 if(thrd) {
885 return thrd;
886 }
887 }
888
889 // All lanes where empty return 0p
890 return 0p;
891}
892
893//-----------------------------------------------------------------------
894// get preferred ready for new thread
895unsigned ready_queue_new_preferred() {
896 unsigned pref = 0;
897 if(struct thread$ * thrd = publicTLS_get( this_thread )) {
898 pref = thrd->preferred;
899 }
900 else {
901 #if defined(USE_CPU_WORK_STEALING)
902 pref = __kernel_getcpu();
903 #endif
904 }
905
906 #if defined(USE_CPU_WORK_STEALING)
907 /* paranoid */ verify(pref >= 0);
908 /* paranoid */ verify(pref < cpu_info.hthrd_count);
909 #endif
910
911 return pref;
912}
913
914//-----------------------------------------------------------------------
915// Check that all the intrusive queues in the data structure are still consistent
916static void check( __ready_queue_t & q ) with (q) {
917 #if defined(__CFA_WITH_VERIFY__)
918 {
919 for( idx ; lanes.count ) {
920 __intrusive_lane_t & sl = lanes.data[idx];
921 assert(!lanes.data[idx].lock);
922
923 if(is_empty(sl)) {
924 assert( sl.anchor.next == 0p );
925 assert( sl.anchor.ts == -1llu );
926 assert( mock_head(sl) == sl.prev );
927 } else {
928 assert( sl.anchor.next != 0p );
929 assert( sl.anchor.ts != -1llu );
930 assert( mock_head(sl) != sl.prev );
931 }
932 }
933 }
934 #endif
935}
936
937//-----------------------------------------------------------------------
938// Given 2 indexes, pick the list with the oldest push an try to pop from it
939static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
940 // Pick the bet list
941 int w = i;
942 if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
943 w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
944 }
945
946 return try_pop(cltr, w __STATS(, stats));
947}
948
949// Call this function of the intrusive list was moved using memcpy
950// fixes the list so that the pointers back to anchors aren't left dangling
951static inline void fix(__intrusive_lane_t & ll) {
952 if(is_empty(ll)) {
953 verify(ll.anchor.next == 0p);
954 ll.prev = mock_head(ll);
955 }
956}
957
958static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
959 processor * it = &list`first;
960 for(unsigned i = 0; i < count; i++) {
961 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
962 it->rdq.id = value;
963 it->rdq.target = MAX;
964 value += READYQ_SHARD_FACTOR;
965 it = &(*it)`next;
966 }
967}
968
969static void reassign_cltr_id(struct cluster * cltr) {
970 unsigned preferred = 0;
971 assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
972 assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
973}
974
975static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
976 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING)
977 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
978 for(i; lanes.count) {
979 lanes.tscs[i].tv = rdtscl();
980 lanes.tscs[i].ma = 0;
981 }
982 #endif
983}
984
985#if defined(USE_CPU_WORK_STEALING)
986 // ready_queue size is fixed in this case
987 void ready_queue_grow(struct cluster * cltr) {}
988 void ready_queue_shrink(struct cluster * cltr) {}
989#else
990 // Grow the ready queue
991 void ready_queue_grow(struct cluster * cltr) {
992 size_t ncount;
993 int target = cltr->procs.total;
994
995 /* paranoid */ verify( ready_mutate_islocked() );
996 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
997
998 // Make sure that everything is consistent
999 /* paranoid */ check( cltr->ready_queue );
1000
1001 // grow the ready queue
1002 with( cltr->ready_queue ) {
1003 // Find new count
1004 // Make sure we always have atleast 1 list
1005 if(target >= 2) {
1006 ncount = target * READYQ_SHARD_FACTOR;
1007 } else {
1008 ncount = SEQUENTIAL_SHARD;
1009 }
1010
1011 // Allocate new array (uses realloc and memcpies the data)
1012 lanes.data = alloc( ncount, lanes.data`realloc );
1013
1014 // Fix the moved data
1015 for( idx; (size_t)lanes.count ) {
1016 fix(lanes.data[idx]);
1017 }
1018
1019 // Construct new data
1020 for( idx; (size_t)lanes.count ~ ncount) {
1021 (lanes.data[idx]){};
1022 }
1023
1024 // Update original
1025 lanes.count = ncount;
1026
1027 lanes.caches = alloc( target, lanes.caches`realloc );
1028 }
1029
1030 fix_times(cltr);
1031
1032 reassign_cltr_id(cltr);
1033
1034 // Make sure that everything is consistent
1035 /* paranoid */ check( cltr->ready_queue );
1036
1037 __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
1038
1039 /* paranoid */ verify( ready_mutate_islocked() );
1040 }
1041
1042 // Shrink the ready queue
1043 void ready_queue_shrink(struct cluster * cltr) {
1044 /* paranoid */ verify( ready_mutate_islocked() );
1045 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
1046
1047 // Make sure that everything is consistent
1048 /* paranoid */ check( cltr->ready_queue );
1049
1050 int target = cltr->procs.total;
1051
1052 with( cltr->ready_queue ) {
1053 // Remember old count
1054 size_t ocount = lanes.count;
1055
1056 // Find new count
1057 // Make sure we always have atleast 1 list
1058 lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
1059 /* paranoid */ verify( ocount >= lanes.count );
1060 /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
1061
1062 // for printing count the number of displaced threads
1063 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
1064 __attribute__((unused)) size_t displaced = 0;
1065 #endif
1066
1067 // redistribute old data
1068 for( idx; (size_t)lanes.count ~ ocount) {
1069 // Lock is not strictly needed but makes checking invariants much easier
1070 __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
1071 verify(locked);
1072
1073 // As long as we can pop from this lane to push the threads somewhere else in the queue
1074 while(!is_empty(lanes.data[idx])) {
1075 struct thread$ * thrd;
1076 unsigned long long _;
1077 [thrd, _] = pop(lanes.data[idx]);
1078
1079 push(cltr, thrd, true);
1080
1081 // for printing count the number of displaced threads
1082 #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
1083 displaced++;
1084 #endif
1085 }
1086
1087 // Unlock the lane
1088 __atomic_unlock(&lanes.data[idx].lock);
1089
1090 // TODO print the queue statistics here
1091
1092 ^(lanes.data[idx]){};
1093 }
1094
1095 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
1096
1097 // Allocate new array (uses realloc and memcpies the data)
1098 lanes.data = alloc( lanes.count, lanes.data`realloc );
1099
1100 // Fix the moved data
1101 for( idx; (size_t)lanes.count ) {
1102 fix(lanes.data[idx]);
1103 }
1104
1105 lanes.caches = alloc( target, lanes.caches`realloc );
1106 }
1107
1108 fix_times(cltr);
1109
1110
1111 reassign_cltr_id(cltr);
1112
1113 // Make sure that everything is consistent
1114 /* paranoid */ check( cltr->ready_queue );
1115
1116 __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
1117 /* paranoid */ verify( ready_mutate_islocked() );
1118 }
1119#endif
1120
1121#if !defined(__CFA_NO_STATISTICS__)
1122 unsigned cnt(const __ready_queue_t & this, unsigned idx) {
1123 /* paranoid */ verify(this.lanes.count > idx);
1124 return this.lanes.data[idx].cnt;
1125 }
1126#endif
1127
1128
1129#if defined(CFA_HAVE_LINUX_LIBRSEQ)
1130 // No definition needed
1131#elif defined(CFA_HAVE_LINUX_RSEQ_H)
1132
1133 #if defined( __x86_64 ) || defined( __i386 )
1134 #define RSEQ_SIG 0x53053053
1135 #elif defined( __ARM_ARCH )
1136 #ifdef __ARMEB__
1137 #define RSEQ_SIG 0xf3def5e7 /* udf #24035 ; 0x5de3 (ARMv6+) */
1138 #else
1139 #define RSEQ_SIG 0xe7f5def3 /* udf #24035 ; 0x5de3 */
1140 #endif
1141 #endif
1142
1143 extern void __disable_interrupts_hard();
1144 extern void __enable_interrupts_hard();
1145
1146 static void __kernel_raw_rseq_register (void) {
1147 /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
1148
1149 // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
1150 int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
1151 if(ret != 0) {
1152 int e = errno;
1153 switch(e) {
1154 case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
1155 case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
1156 case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
1157 case EBUSY : abort("KERNEL ERROR: rseq register already registered");
1158 case EPERM : abort("KERNEL ERROR: rseq register sig argument on unregistration does not match the signature received on registration");
1159 default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
1160 }
1161 }
1162 }
1163
1164 static void __kernel_raw_rseq_unregister(void) {
1165 /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
1166
1167 // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
1168 int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
1169 if(ret != 0) {
1170 int e = errno;
1171 switch(e) {
1172 case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
1173 case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
1174 case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
1175 case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
1176 case EPERM : abort("KERNEL ERROR: rseq unregister sig argument on unregistration does not match the signature received on registration");
1177 default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
1178 }
1179 }
1180 }
1181#else
1182 // No definition needed
1183#endif
Note: See TracBrowser for help on using the repository browser.