Changeset 175f9f4 for libcfa/src/concurrency
- Timestamp:
- Jan 18, 2022, 1:16:23 PM (2 years ago)
- Branches:
- ADT, ast-experimental, enum, forall-pointer-decay, master, pthread-emulation, qualifiedEnum
- Children:
- 1e8b4b49, adfd125
- Parents:
- 21a5bfb7 (diff), 91a72ef (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the(diff)links above to see all the changes relative to each parent. - Location:
- libcfa/src/concurrency
- Files:
-
- 10 edited
-
clib/cfathread.cfa (modified) (2 diffs)
-
invoke.h (modified) (2 diffs)
-
io.cfa (modified) (1 diff)
-
kernel.cfa (modified) (1 diff)
-
kernel.hfa (modified) (4 diffs)
-
kernel/fwd.hfa (modified) (3 diffs)
-
kernel/startup.cfa (modified) (6 diffs)
-
locks.hfa (modified) (2 diffs)
-
ready_queue.cfa (modified) (24 diffs)
-
thread.cfa (modified) (4 diffs)
Legend:
- Unmodified
- Added
- Removed
-
libcfa/src/concurrency/clib/cfathread.cfa
r21a5bfb7 r175f9f4 22 22 #include "thread.hfa" 23 23 #include "time.hfa" 24 #include "stdlib.hfa" 24 25 25 26 #include "cfathread.h" … … 195 196 eevent.data.u64 = (uint64_t)active_thread(); 196 197 197 int id = thread_rand() % poller_cnt;198 int id = prng() % poller_cnt; 198 199 if(0 != epoll_ctl(poller_fds[id], EPOLL_CTL_ADD, fd, &eevent)) 199 200 { -
libcfa/src/concurrency/invoke.h
r21a5bfb7 r175f9f4 10 10 // Created On : Tue Jan 17 12:27:26 2016 11 11 // Last Modified By : Peter A. Buhr 12 // Last Modified On : Thu Jan 6 16:37:40202213 // Update Count : 4 712 // Last Modified On : Sun Jan 9 19:06:45 2022 13 // Update Count : 48 14 14 // 15 15 … … 211 211 struct processor * last_proc; 212 212 213 uint32_t random_state; // fast random numbers 214 213 215 #if defined( __CFA_WITH_VERIFY__ ) 214 216 void * canary; -
libcfa/src/concurrency/io.cfa
r21a5bfb7 r175f9f4 552 552 /* paranoid */ verify( proc == __cfaabi_tls.this_processor ); 553 553 /* paranoid */ verify( ! __preemption_enabled() ); 554 555 return true; 554 556 } 555 557 #endif -
libcfa/src/concurrency/kernel.cfa
r21a5bfb7 r175f9f4 554 554 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary ); 555 555 556 const bool local = thrd->state != Start;557 556 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready; 558 557 -
libcfa/src/concurrency/kernel.hfa
r21a5bfb7 r175f9f4 67 67 unsigned target; 68 68 unsigned last; 69 unsigned cnt; 70 unsigned long long int cutoff; 69 signed cpu; 71 70 } rdq; 72 71 … … 152 151 volatile unsigned long long tv; 153 152 volatile unsigned long long ma; 153 }; 154 155 struct __attribute__((aligned(16))) __cache_id_t { 156 volatile unsigned id; 154 157 }; 155 158 … … 164 167 static inline void ^?{}(__timestamp_t & this) {} 165 168 169 struct __attribute__((aligned(128))) __ready_queue_caches_t; 170 void ?{}(__ready_queue_caches_t & this); 171 void ^?{}(__ready_queue_caches_t & this); 172 166 173 //TODO adjust cache size to ARCHITECTURE 167 // Structure holding the re laxed ready queue174 // Structure holding the ready queue 168 175 struct __ready_queue_t { 169 176 // Data tracking the actual lanes … … 177 184 // Array of times 178 185 __timestamp_t * volatile tscs; 186 187 __cache_id_t * volatile caches; 179 188 180 189 // Array of stats -
libcfa/src/concurrency/kernel/fwd.hfa
r21a5bfb7 r175f9f4 77 77 78 78 static inline uint64_t __tls_rand() { 79 return 79 80 #if defined(__SIZEOF_INT128__) 80 return __lehmer64( kernelTLS().rand_seed );81 lehmer64( kernelTLS().rand_seed ); 81 82 #else 82 return __xorshift64( kernelTLS().rand_seed );83 xorshift_13_7_17( kernelTLS().rand_seed ); 83 84 #endif 84 85 } 85 86 86 #define M (1_l64u << 48_l64u)87 #define A (25214903917_l64u)88 #define AI (18446708753438544741_l64u)89 #define C (11_l64u)90 #define D (16_l64u)91 92 87 static inline unsigned __tls_rand_fwd() { 93 94 kernelTLS().ready_rng.fwd_seed = (A * kernelTLS().ready_rng.fwd_seed + C) & (M - 1); 95 return kernelTLS().ready_rng.fwd_seed >> D; 88 return LCGBI_fwd( kernelTLS().ready_rng.fwd_seed ); 96 89 } 97 90 98 91 static inline unsigned __tls_rand_bck() { 99 unsigned int r = kernelTLS().ready_rng.bck_seed >> D; 100 kernelTLS().ready_rng.bck_seed = AI * (kernelTLS().ready_rng.bck_seed - C) & (M - 1); 101 return r; 102 } 103 104 #undef M 105 #undef A 106 #undef AI 107 #undef C 108 #undef D 92 return LCGBI_bck( kernelTLS().ready_rng.bck_seed ); 93 } 109 94 110 95 static inline void __tls_rand_advance_bck(void) { … … 112 97 } 113 98 } 114 115 116 99 117 100 extern void disable_interrupts(); … … 142 125 } 143 126 } 144 145 extern uint64_t thread_rand();146 127 147 128 // Semaphore which only supports a single thread -
libcfa/src/concurrency/kernel/startup.cfa
r21a5bfb7 r175f9f4 34 34 #include "kernel_private.hfa" 35 35 #include "startup.hfa" // STARTUP_PRIORITY_XXX 36 #include "limits.hfa" 36 37 #include "math.hfa" 37 38 … … 101 102 extern void __wake_proc(processor *); 102 103 extern int cfa_main_returned; // from interpose.cfa 104 extern uint32_t __global_random_seed; 103 105 104 106 //----------------------------------------------------------------------------- … … 174 176 this.context = &storage_mainThreadCtx; 175 177 } 176 177 178 178 179 … … 489 490 preferred = ready_queue_new_preferred(); 490 491 last_proc = 0p; 492 random_state = __global_random_seed; 491 493 #if defined( __CFA_WITH_VERIFY__ ) 492 494 canary = 0x0D15EA5E0D15EA5Ep; … … 513 515 this.rdq.its = 0; 514 516 this.rdq.itr = 0; 515 this.rdq.id = -1u; 516 this.rdq.target = -1u; 517 this.rdq.last = -1u; 518 this.rdq.cutoff = 0ull; 517 this.rdq.id = MAX; 518 this.rdq.target = MAX; 519 this.rdq.last = MAX; 520 this.rdq.cpu = 0; 521 // this.rdq.cutoff = 0ull; 519 522 do_terminate = false; 520 523 preemption_alarm = 0p; … … 684 687 uint_fast32_t last_size; 685 688 [this->unique_id, last_size] = ready_mutate_register(); 689 690 this->rdq.cpu = __kernel_getcpu(); 686 691 687 692 this->cltr->procs.total += 1u; -
libcfa/src/concurrency/locks.hfa
r21a5bfb7 r175f9f4 29 29 #include "time_t.hfa" 30 30 #include "time.hfa" 31 32 //-----------------------------------------------------------------------------33 // Semaphores34 35 // '0-nary' semaphore36 // Similar to a counting semaphore except the value of one is never reached37 // as a consequence, a V() that would bring the value to 1 *spins* until38 // a P consumes it39 struct Semaphore0nary {40 __spinlock_t lock; // needed to protect41 mpsc_queue(thread$) queue;42 };43 44 static inline bool P(Semaphore0nary & this, thread$ * thrd) {45 /* paranoid */ verify(!thrd`next);46 /* paranoid */ verify(!(&(*thrd)`next));47 48 push(this.queue, thrd);49 return true;50 }51 52 static inline bool P(Semaphore0nary & this) {53 thread$ * thrd = active_thread();54 P(this, thrd);55 park();56 return true;57 }58 59 static inline thread$ * V(Semaphore0nary & this, bool doUnpark = true) {60 thread$ * next;61 lock(this.lock __cfaabi_dbg_ctx2);62 for (;;) {63 next = pop(this.queue);64 if (next) break;65 Pause();66 }67 unlock(this.lock);68 69 if (doUnpark) unpark(next);70 return next;71 }72 73 // Wrapper used on top of any sempahore to avoid potential locking74 struct BinaryBenaphore {75 volatile ssize_t counter;76 };77 78 static inline {79 void ?{}(BinaryBenaphore & this) { this.counter = 0; }80 void ?{}(BinaryBenaphore & this, zero_t) { this.counter = 0; }81 void ?{}(BinaryBenaphore & this, one_t ) { this.counter = 1; }82 83 // returns true if no blocking needed84 bool P(BinaryBenaphore & this) {85 return __atomic_fetch_sub(&this.counter, 1, __ATOMIC_SEQ_CST) > 0;86 }87 88 bool tryP(BinaryBenaphore & this) {89 ssize_t c = this.counter;90 /* paranoid */ verify( c > MIN );91 return (c >= 1) && __atomic_compare_exchange_n(&this.counter, &c, c-1, false, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);92 }93 94 // returns true if notify needed95 bool V(BinaryBenaphore & this) {96 ssize_t c = 0;97 for () {98 /* paranoid */ verify( this.counter < MAX );99 if (__atomic_compare_exchange_n(&this.counter, &c, c+1, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {100 if (c == 0) return true;101 /* paranoid */ verify(c < 0);102 return false;103 } else {104 if (c == 1) return true;105 /* paranoid */ verify(c < 1);106 Pause();107 }108 }109 }110 }111 112 // Binary Semaphore based on the BinaryBenaphore on top of the 0-nary Semaphore113 struct ThreadBenaphore {114 BinaryBenaphore ben;115 Semaphore0nary sem;116 };117 118 static inline void ?{}(ThreadBenaphore & this) {}119 static inline void ?{}(ThreadBenaphore & this, zero_t) { (this.ben){ 0 }; }120 static inline void ?{}(ThreadBenaphore & this, one_t ) { (this.ben){ 1 }; }121 122 static inline bool P(ThreadBenaphore & this) { return P(this.ben) ? false : P(this.sem); }123 static inline bool tryP(ThreadBenaphore & this) { return tryP(this.ben); }124 static inline bool P(ThreadBenaphore & this, bool wait) { return wait ? P(this) : tryP(this); }125 126 static inline thread$ * V(ThreadBenaphore & this, bool doUnpark = true) {127 if (V(this.ben)) return 0p;128 return V(this.sem, doUnpark);129 }130 31 131 32 //----------------------------------------------------------------------------- … … 171 72 static inline void on_wakeup( owner_lock & this, size_t v ) { on_wakeup ( (blocking_lock &)this, v ); } 172 73 static inline void on_notify( owner_lock & this, struct thread$ * t ) { on_notify( (blocking_lock &)this, t ); } 173 174 struct fast_lock {175 thread$ * volatile owner;176 ThreadBenaphore sem;177 };178 179 static inline void ?{}(fast_lock & this) __attribute__((deprecated("use linear_backoff_then_block_lock instead")));180 static inline void ?{}(fast_lock & this) { this.owner = 0p; }181 182 static inline bool $try_lock(fast_lock & this, thread$ * thrd) {183 thread$ * exp = 0p;184 return __atomic_compare_exchange_n(&this.owner, &exp, thrd, false, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);185 }186 187 static inline void lock( fast_lock & this ) __attribute__((deprecated("use linear_backoff_then_block_lock instead"), artificial));188 static inline void lock( fast_lock & this ) {189 thread$ * thrd = active_thread();190 /* paranoid */verify(thrd != this.owner);191 192 for (;;) {193 if ($try_lock(this, thrd)) return;194 P(this.sem);195 }196 }197 198 static inline bool try_lock( fast_lock & this ) __attribute__((deprecated("use linear_backoff_then_block_lock instead"), artificial));199 static inline bool try_lock ( fast_lock & this ) {200 thread$ * thrd = active_thread();201 /* paranoid */ verify(thrd != this.owner);202 return $try_lock(this, thrd);203 }204 205 static inline thread$ * unlock( fast_lock & this ) __attribute__((deprecated("use linear_backoff_then_block_lock instead"), artificial));206 static inline thread$ * unlock( fast_lock & this ) {207 /* paranoid */ verify(active_thread() == this.owner);208 209 // open 'owner' before unlocking anyone210 // so new and unlocked threads don't park incorrectly.211 // This may require additional fencing on ARM.212 this.owner = 0p;213 214 return V(this.sem);215 }216 217 static inline size_t on_wait( fast_lock & this ) { unlock(this); return 0; }218 static inline void on_wakeup( fast_lock & this, size_t ) { lock(this); }219 static inline void on_notify( fast_lock &, struct thread$ * t ) { unpark(t); }220 74 221 75 struct mcs_node { -
libcfa/src/concurrency/ready_queue.cfa
r21a5bfb7 r175f9f4 20 20 21 21 22 #define USE_RELAXED_FIFO22 // #define USE_RELAXED_FIFO 23 23 // #define USE_WORK_STEALING 24 24 // #define USE_CPU_WORK_STEALING 25 #define USE_AWARE_STEALING 25 26 26 27 #include "bits/defs.hfa" … … 29 30 30 31 #include "stdlib.hfa" 32 #include "limits.hfa" 31 33 #include "math.hfa" 32 34 … … 54 56 #endif 55 57 56 #if defined(USE_CPU_WORK_STEALING) 58 #if defined(USE_AWARE_STEALING) 59 #define READYQ_SHARD_FACTOR 2 60 #define SEQUENTIAL_SHARD 2 61 #elif defined(USE_CPU_WORK_STEALING) 57 62 #define READYQ_SHARD_FACTOR 2 58 63 #elif defined(USE_RELAXED_FIFO) … … 138 143 __kernel_rseq_register(); 139 144 140 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);141 145 bool * handle = (bool *)&kernelTLS().sched_lock; 142 146 … … 174 178 } 175 179 176 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);177 178 180 // Return new spot. 179 181 /* paranoid */ verify(n < ready); … … 190 192 191 193 __atomic_store_n(cell, 0p, __ATOMIC_RELEASE); 192 193 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);194 194 195 195 __kernel_rseq_unregister(); … … 244 244 245 245 //======================================================================= 246 // caches handling 247 248 struct __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 }; 255 256 void ?{}(__ready_queue_caches_t & this) { this.count = 0; } 257 void ^?{}(__ready_queue_caches_t & this) {} 258 259 static 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 } 265 266 static 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 //======================================================================= 246 274 // Cforall Ready Queue used for scheduling 247 275 //======================================================================= 248 unsigned 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; 276 unsigned 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; 253 287 } 254 288 … … 271 305 } 272 306 #else 273 lanes.data = 0p; 274 lanes.tscs = 0p; 275 lanes.help = 0p; 276 lanes.count = 0; 307 lanes.data = 0p; 308 lanes.tscs = 0p; 309 lanes.caches = 0p; 310 lanes.help = 0p; 311 lanes.count = 0; 277 312 #endif 278 313 } … … 285 320 free(lanes.data); 286 321 free(lanes.tscs); 322 free(lanes.caches); 287 323 free(lanes.help); 288 324 } 289 325 290 326 //----------------------------------------------------------------------- 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; 401 402 // Super important: don't write the same value over and over again 403 // We want to maximise our chances that his particular values stays in cache 404 if(lanes.caches[this / READYQ_SHARD_FACTOR].id != this_cache) 405 __atomic_store_n(&lanes.caches[this / READYQ_SHARD_FACTOR].id, this_cache, __ATOMIC_RELAXED); 406 407 const unsigned long long ctsc = rdtscl(); 408 409 if(proc->rdq.target == MAX) { 410 uint64_t chaos = __tls_rand(); 411 unsigned ext = chaos & 0xff; 412 unsigned other = (chaos >> 8) % (lanes.count); 413 414 if(ext < 3 || __atomic_load_n(&lanes.caches[other / READYQ_SHARD_FACTOR].id, __ATOMIC_RELAXED) == this_cache) { 415 proc->rdq.target = other; 416 } 417 } 418 else { 419 const unsigned target = proc->rdq.target; 420 __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, lanes.tscs[target].tv); 421 /* paranoid */ verify( lanes.tscs[target].tv != MAX ); 422 if(target < lanes.count) { 423 const unsigned long long cutoff = calc_cutoff(ctsc, proc, cltr->ready_queue); 424 const unsigned long long age = moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma); 425 __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"); 426 if(age > cutoff) { 427 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 428 if(t) return t; 429 } 430 } 431 proc->rdq.target = MAX; 432 } 433 434 for(READYQ_SHARD_FACTOR) { 435 unsigned i = this + (proc->rdq.itr++ % READYQ_SHARD_FACTOR); 436 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t; 437 } 438 439 // All lanes where empty return 0p 440 return 0p; 441 442 } 443 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) { 444 unsigned i = __tls_rand() % lanes.count; 445 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal)); 446 } 447 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) { 448 return search(cltr); 449 } 450 #endif 291 451 #if defined(USE_CPU_WORK_STEALING) 292 452 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) { … … 350 510 /* paranoid */ verify( kernelTLS().this_processor ); 351 511 512 processor * const proc = kernelTLS().this_processor; 352 513 const int cpu = __kernel_getcpu(); 353 514 /* paranoid */ verify(cpu >= 0); … … 360 521 /* 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); 361 522 362 processor * const proc = kernelTLS().this_processor;363 523 const int start = map.self * READYQ_SHARD_FACTOR; 364 524 const unsigned long long ctsc = rdtscl(); 365 525 366 526 // Did we already have a help target 367 if(proc->rdq.target == -1u) {527 if(proc->rdq.target == MAX) { 368 528 unsigned long long max = 0; 369 529 for(i; READYQ_SHARD_FACTOR) { 370 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);530 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 371 531 if(tsc > max) max = tsc; 372 532 } 373 proc->rdq.cutoff = (max + 2 * max) / 2;533 // proc->rdq.cutoff = (max + 2 * max) / 2; 374 534 /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores. 375 535 /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores. … … 384 544 } 385 545 386 /* paranoid */ verify(proc->rdq.target != -1u);546 /* paranoid */ verify(proc->rdq.target != MAX); 387 547 } 388 548 else { 389 549 unsigned long long max = 0; 390 550 for(i; READYQ_SHARD_FACTOR) { 391 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);551 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 392 552 if(tsc > max) max = tsc; 393 553 } … … 395 555 { 396 556 unsigned target = proc->rdq.target; 397 proc->rdq.target = -1u;557 proc->rdq.target = MAX; 398 558 lanes.help[target / READYQ_SHARD_FACTOR].tri++; 399 if(moving_average(ctsc -lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {559 if(moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) { 400 560 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 401 561 proc->rdq.last = target; 402 562 if(t) return t; 403 else proc->rdq.target = -1u;404 563 } 405 else proc->rdq.target = -1u;564 proc->rdq.target = MAX; 406 565 } 407 566 408 567 unsigned last = proc->rdq.last; 409 if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) <cutoff) {568 if(last != MAX && moving_average(ctsc, lanes.tscs[last].tv, lanes.tscs[last].ma) > cutoff) { 410 569 thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help)); 411 570 if(t) return t; 412 571 } 413 572 else { 414 proc->rdq.last = -1u;573 proc->rdq.last = MAX; 415 574 } 416 575 } … … 428 587 processor * const proc = kernelTLS().this_processor; 429 588 unsigned last = proc->rdq.last; 430 if(last != -1u) {589 if(last != MAX) { 431 590 struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal)); 432 591 if(t) return t; 433 proc->rdq.last = -1u;592 proc->rdq.last = MAX; 434 593 } 435 594 … … 560 719 #else 561 720 unsigned preferred = thrd->preferred; 562 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == -1u|| thrd->curr_cluster != cltr;721 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == MAX || thrd->curr_cluster != cltr; 563 722 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count ); 564 723 … … 612 771 processor * proc = kernelTLS().this_processor; 613 772 614 if(proc->rdq.target == -1u) {773 if(proc->rdq.target == MAX) { 615 774 unsigned long long min = ts(lanes.data[proc->rdq.id]); 616 775 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) { … … 623 782 else { 624 783 unsigned target = proc->rdq.target; 625 proc->rdq.target = -1u;784 proc->rdq.target = MAX; 626 785 const unsigned long long bias = 0; //2_500_000_000; 627 786 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff; … … 658 817 // try to pop from a lane given by index w 659 818 static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) { 819 /* paranoid */ verify( w < lanes.count ); 660 820 __STATS( stats.attempt++; ) 661 821 … … 681 841 // Actually pop the list 682 842 struct thread$ * thrd; 683 unsigned long long tsc_before = ts(lane); 843 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 844 unsigned long long tsc_before = ts(lane); 845 #endif 684 846 unsigned long long tsv; 685 847 [thrd, tsv] = pop(lane); … … 695 857 __STATS( stats.success++; ) 696 858 697 #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); 859 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 860 if (tsv != MAX) { 861 unsigned long long now = rdtscl(); 862 unsigned long long pma = __atomic_load_n(&lanes.tscs[w].ma, __ATOMIC_RELAXED); 863 __atomic_store_n(&lanes.tscs[w].tv, tsv, __ATOMIC_RELAXED); 864 __atomic_store_n(&lanes.tscs[w].ma, moving_average(now, tsc_before, pma), __ATOMIC_RELAXED); 865 } 701 866 #endif 702 867 703 #if defined(USE_ CPU_WORK_STEALING)868 #if defined(USE_AWARE_STEALING) || defined(USE_CPU_WORK_STEALING) 704 869 thrd->preferred = w / READYQ_SHARD_FACTOR; 705 870 #else … … 800 965 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count); 801 966 it->rdq.id = value; 802 it->rdq.target = -1u;967 it->rdq.target = MAX; 803 968 value += READYQ_SHARD_FACTOR; 804 969 it = &(*it)`next; … … 813 978 814 979 static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) { 815 #if defined(USE_ WORK_STEALING)980 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) 816 981 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc); 817 982 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); 983 lanes.tscs[i].tv = rdtscl(); 984 lanes.tscs[i].ma = 0; 821 985 } 822 986 #endif … … 864 1028 // Update original 865 1029 lanes.count = ncount; 1030 1031 lanes.caches = alloc( target, lanes.caches`realloc ); 866 1032 } 867 1033 … … 940 1106 fix(lanes.data[idx]); 941 1107 } 1108 1109 lanes.caches = alloc( target, lanes.caches`realloc ); 942 1110 } 943 1111 944 1112 fix_times(cltr); 1113 945 1114 946 1115 reassign_cltr_id(cltr); -
libcfa/src/concurrency/thread.cfa
r21a5bfb7 r175f9f4 10 10 // Created On : Tue Jan 17 12:27:26 2017 11 11 // Last Modified By : Peter A. Buhr 12 // Last Modified On : Wed Dec 4 09:17:49 201913 // Update Count : 912 // Last Modified On : Thu Jan 13 20:11:55 2022 13 // Update Count : 42 14 14 // 15 15 … … 25 25 #include "invoke.h" 26 26 27 uint64_t thread_rand();27 extern uint32_t __global_random_seed; 28 28 29 29 //----------------------------------------------------------------------------- 30 30 // Thread ctors and dtors 31 void ?{}( thread$ & this, const char * const name, cluster & cl, void * storage, size_t storageSize ) with( this ) {31 void ?{}( thread$ & this, const char * const name, cluster & cl, void * storage, size_t storageSize ) with( this ) { 32 32 context{ 0p, 0p }; 33 33 self_cor{ name, storage, storageSize }; … … 45 45 preferred = ready_queue_new_preferred(); 46 46 last_proc = 0p; 47 random_state = __global_random_seed; 47 48 #if defined( __CFA_WITH_VERIFY__ ) 48 49 canary = 0x0D15EA5E0D15EA5Ep; … … 171 172 } 172 173 173 uint64_t thread_rand() { 174 disable_interrupts(); 175 uint64_t ret = __tls_rand(); 176 enable_interrupts(); 177 return ret; 178 } 174 //----------------------------------------------------------------------------- 175 #define GENERATOR LCG 176 177 void set_seed( uint32_t seed ) { 178 active_thread()->random_state = __global_random_seed = seed; 179 GENERATOR( active_thread()->random_state ); 180 } // set_seed 181 uint32_t prng( void ) { return GENERATOR( active_thread()->random_state ); } // [0,UINT_MAX] 179 182 180 183 // Local Variables: //
Note: See TracChangeset
for help on using the changeset viewer.
