| 1 | // | 
|---|
| 2 | // Cforall Version 1.0.0 Copyright (C) 2016 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 | // bits/locks.hfa -- Fast internal locks. | 
|---|
| 8 | // | 
|---|
| 9 | // Author           : Thierry Delisle | 
|---|
| 10 | // Created On       : Tue Oct 31 15:14:38 2017 | 
|---|
| 11 | // Last Modified By : Peter A. Buhr | 
|---|
| 12 | // Last Modified On : Wed Aug 12 14:18:07 2020 | 
|---|
| 13 | // Update Count     : 13 | 
|---|
| 14 | // | 
|---|
| 15 |  | 
|---|
| 16 | #pragma once | 
|---|
| 17 |  | 
|---|
| 18 | #include "bits/debug.hfa" | 
|---|
| 19 | #include "bits/defs.hfa" | 
|---|
| 20 | #include <assert.h> | 
|---|
| 21 |  | 
|---|
| 22 | #ifdef __cforall | 
|---|
| 23 | extern "C" { | 
|---|
| 24 | #include <pthread.h> | 
|---|
| 25 | } | 
|---|
| 26 | #endif | 
|---|
| 27 |  | 
|---|
| 28 | // pause to prevent excess processor bus usage | 
|---|
| 29 | #if defined( __i386 ) || defined( __x86_64 ) | 
|---|
| 30 | #define Pause() __asm__ __volatile__ ( "pause" : : : ) | 
|---|
| 31 | #elif defined( __ARM_ARCH ) | 
|---|
| 32 | #define Pause() __asm__ __volatile__ ( "YIELD" : : : ) | 
|---|
| 33 | #else | 
|---|
| 34 | #error unsupported architecture | 
|---|
| 35 | #endif | 
|---|
| 36 |  | 
|---|
| 37 | struct __spinlock_t { | 
|---|
| 38 | // Wrap in struct to prevent false sharing with debug info | 
|---|
| 39 | volatile bool lock; | 
|---|
| 40 | #ifdef __CFA_DEBUG__ | 
|---|
| 41 | // previous function to acquire the lock | 
|---|
| 42 | const char * prev_name; | 
|---|
| 43 | // previous thread to acquire the lock | 
|---|
| 44 | void* prev_thrd; | 
|---|
| 45 | #endif | 
|---|
| 46 | }; | 
|---|
| 47 |  | 
|---|
| 48 | #ifdef __cforall | 
|---|
| 49 | extern "C" { | 
|---|
| 50 | extern void disable_interrupts() OPTIONAL_THREAD; | 
|---|
| 51 | extern void enable_interrupts_noPoll() OPTIONAL_THREAD; | 
|---|
| 52 |  | 
|---|
| 53 | #ifdef __CFA_DEBUG__ | 
|---|
| 54 | void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]); | 
|---|
| 55 | #else | 
|---|
| 56 | #define __cfaabi_dbg_record_lock(x, y) | 
|---|
| 57 | #endif | 
|---|
| 58 | } | 
|---|
| 59 |  | 
|---|
| 60 | static inline void ?{}( __spinlock_t & this ) { | 
|---|
| 61 | this.lock = 0; | 
|---|
| 62 | } | 
|---|
| 63 |  | 
|---|
| 64 | // Lock the spinlock, return false if already acquired | 
|---|
| 65 | static inline bool try_lock  ( __spinlock_t & this __cfaabi_dbg_ctx_param2 ) { | 
|---|
| 66 | disable_interrupts(); | 
|---|
| 67 | bool result = (this.lock == 0) && (__atomic_test_and_set( &this.lock, __ATOMIC_ACQUIRE ) == 0); | 
|---|
| 68 | if( result ) { | 
|---|
| 69 | __cfaabi_dbg_record_lock( this, caller ); | 
|---|
| 70 | } else { | 
|---|
| 71 | enable_interrupts_noPoll(); | 
|---|
| 72 | } | 
|---|
| 73 | return result; | 
|---|
| 74 | } | 
|---|
| 75 |  | 
|---|
| 76 | // Lock the spinlock, spin if already acquired | 
|---|
| 77 | static inline void lock( __spinlock_t & this __cfaabi_dbg_ctx_param2 ) { | 
|---|
| 78 | #ifndef NOEXPBACK | 
|---|
| 79 | enum { SPIN_START = 4, SPIN_END = 64 * 1024, }; | 
|---|
| 80 | unsigned int spin = SPIN_START; | 
|---|
| 81 | #endif | 
|---|
| 82 |  | 
|---|
| 83 | disable_interrupts(); | 
|---|
| 84 | for ( unsigned int i = 1;; i += 1 ) { | 
|---|
| 85 | if ( (this.lock == 0) && (__atomic_test_and_set( &this.lock, __ATOMIC_ACQUIRE ) == 0) ) break; | 
|---|
| 86 | #ifndef NOEXPBACK | 
|---|
| 87 | // exponential spin | 
|---|
| 88 | for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); | 
|---|
| 89 |  | 
|---|
| 90 | // slowly increase by powers of 2 | 
|---|
| 91 | if ( i % 64 == 0 ) spin += spin; | 
|---|
| 92 |  | 
|---|
| 93 | // prevent overflow | 
|---|
| 94 | if ( spin > SPIN_END ) spin = SPIN_START; | 
|---|
| 95 | #else | 
|---|
| 96 | Pause(); | 
|---|
| 97 | #endif | 
|---|
| 98 | } | 
|---|
| 99 | __cfaabi_dbg_record_lock( this, caller ); | 
|---|
| 100 | } | 
|---|
| 101 |  | 
|---|
| 102 | static inline void unlock( __spinlock_t & this ) { | 
|---|
| 103 | __atomic_clear( &this.lock, __ATOMIC_RELEASE ); | 
|---|
| 104 | enable_interrupts_noPoll(); | 
|---|
| 105 | } | 
|---|
| 106 |  | 
|---|
| 107 |  | 
|---|
| 108 | #ifdef __CFA_WITH_VERIFY__ | 
|---|
| 109 | extern bool __cfaabi_dbg_in_kernel(); | 
|---|
| 110 | #endif | 
|---|
| 111 |  | 
|---|
| 112 | extern "C" { | 
|---|
| 113 | char * strerror(int); | 
|---|
| 114 | } | 
|---|
| 115 | #define CHECKED(x) { int err = x; if( err != 0 ) abort("KERNEL ERROR: Operation \"" #x "\" return error %d - %s\n", err, strerror(err)); } | 
|---|
| 116 |  | 
|---|
| 117 | struct __bin_sem_t { | 
|---|
| 118 | pthread_mutex_t         lock; | 
|---|
| 119 | pthread_cond_t          cond; | 
|---|
| 120 | int                     val; | 
|---|
| 121 | }; | 
|---|
| 122 |  | 
|---|
| 123 | static inline void ?{}(__bin_sem_t & this) with( this ) { | 
|---|
| 124 | // Create the mutex with error checking | 
|---|
| 125 | pthread_mutexattr_t mattr; | 
|---|
| 126 | pthread_mutexattr_init( &mattr ); | 
|---|
| 127 | pthread_mutexattr_settype( &mattr, PTHREAD_MUTEX_ERRORCHECK_NP); | 
|---|
| 128 | pthread_mutex_init(&lock, &mattr); | 
|---|
| 129 |  | 
|---|
| 130 | pthread_cond_init (&cond, (const pthread_condattr_t *)0p);  // workaround trac#208: cast should not be required | 
|---|
| 131 | val = 0; | 
|---|
| 132 | } | 
|---|
| 133 |  | 
|---|
| 134 | static inline void ^?{}(__bin_sem_t & this) with( this ) { | 
|---|
| 135 | CHECKED( pthread_mutex_destroy(&lock) ); | 
|---|
| 136 | CHECKED( pthread_cond_destroy (&cond) ); | 
|---|
| 137 | } | 
|---|
| 138 |  | 
|---|
| 139 | static inline void wait(__bin_sem_t & this) with( this ) { | 
|---|
| 140 | verify(__cfaabi_dbg_in_kernel()); | 
|---|
| 141 | CHECKED( pthread_mutex_lock(&lock) ); | 
|---|
| 142 | while(val < 1) { | 
|---|
| 143 | pthread_cond_wait(&cond, &lock); | 
|---|
| 144 | } | 
|---|
| 145 | val -= 1; | 
|---|
| 146 | CHECKED( pthread_mutex_unlock(&lock) ); | 
|---|
| 147 | } | 
|---|
| 148 |  | 
|---|
| 149 | static inline bool post(__bin_sem_t & this) with( this ) { | 
|---|
| 150 | bool needs_signal = false; | 
|---|
| 151 |  | 
|---|
| 152 | CHECKED( pthread_mutex_lock(&lock) ); | 
|---|
| 153 | if(val < 1) { | 
|---|
| 154 | val += 1; | 
|---|
| 155 | pthread_cond_signal(&cond); | 
|---|
| 156 | needs_signal = true; | 
|---|
| 157 | } | 
|---|
| 158 | CHECKED( pthread_mutex_unlock(&lock) ); | 
|---|
| 159 |  | 
|---|
| 160 | return needs_signal; | 
|---|
| 161 | } | 
|---|
| 162 |  | 
|---|
| 163 | #undef CHECKED | 
|---|
| 164 |  | 
|---|
| 165 | struct $thread; | 
|---|
| 166 | extern void park( void ); | 
|---|
| 167 | extern void unpark( struct $thread * this ); | 
|---|
| 168 | static inline struct $thread * active_thread (); | 
|---|
| 169 |  | 
|---|
| 170 | // Semaphore which only supports a single thread | 
|---|
| 171 | struct single_sem { | 
|---|
| 172 | struct $thread * volatile ptr; | 
|---|
| 173 | }; | 
|---|
| 174 |  | 
|---|
| 175 | static inline { | 
|---|
| 176 | void  ?{}(single_sem & this) { | 
|---|
| 177 | this.ptr = 0p; | 
|---|
| 178 | } | 
|---|
| 179 |  | 
|---|
| 180 | void ^?{}(single_sem &) {} | 
|---|
| 181 |  | 
|---|
| 182 | bool wait(single_sem & this) { | 
|---|
| 183 | for() { | 
|---|
| 184 | struct $thread * expected = this.ptr; | 
|---|
| 185 | if(expected == 1p) { | 
|---|
| 186 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 187 | return false; | 
|---|
| 188 | } | 
|---|
| 189 | } | 
|---|
| 190 | else { | 
|---|
| 191 | /* paranoid */ verify( expected == 0p ); | 
|---|
| 192 | if(__atomic_compare_exchange_n(&this.ptr, &expected, active_thread(), false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 193 | park(); | 
|---|
| 194 | return true; | 
|---|
| 195 | } | 
|---|
| 196 | } | 
|---|
| 197 |  | 
|---|
| 198 | } | 
|---|
| 199 | } | 
|---|
| 200 |  | 
|---|
| 201 | bool post(single_sem & this) { | 
|---|
| 202 | for() { | 
|---|
| 203 | struct $thread * expected = this.ptr; | 
|---|
| 204 | if(expected == 1p) return false; | 
|---|
| 205 | if(expected == 0p) { | 
|---|
| 206 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 1p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 207 | return false; | 
|---|
| 208 | } | 
|---|
| 209 | } | 
|---|
| 210 | else { | 
|---|
| 211 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 212 | unpark( expected ); | 
|---|
| 213 | return true; | 
|---|
| 214 | } | 
|---|
| 215 | } | 
|---|
| 216 | } | 
|---|
| 217 | } | 
|---|
| 218 | } | 
|---|
| 219 |  | 
|---|
| 220 | // Synchronozation primitive which only supports a single thread and one post | 
|---|
| 221 | // Similar to a binary semaphore with a 'one shot' semantic | 
|---|
| 222 | // is expected to be discarded after each party call their side | 
|---|
| 223 | struct oneshot { | 
|---|
| 224 | // Internal state : | 
|---|
| 225 | //     0p     : is initial state (wait will block) | 
|---|
| 226 | //     1p     : fulfilled (wait won't block) | 
|---|
| 227 | // any thread : a thread is currently waiting | 
|---|
| 228 | struct $thread * volatile ptr; | 
|---|
| 229 | }; | 
|---|
| 230 |  | 
|---|
| 231 | static inline { | 
|---|
| 232 | void  ?{}(oneshot & this) { | 
|---|
| 233 | this.ptr = 0p; | 
|---|
| 234 | } | 
|---|
| 235 |  | 
|---|
| 236 | void ^?{}(oneshot &) {} | 
|---|
| 237 |  | 
|---|
| 238 | // Wait for the post, return immidiately if it already happened. | 
|---|
| 239 | // return true if the thread was parked | 
|---|
| 240 | bool wait(oneshot & this) { | 
|---|
| 241 | for() { | 
|---|
| 242 | struct $thread * expected = this.ptr; | 
|---|
| 243 | if(expected == 1p) return false; | 
|---|
| 244 | /* paranoid */ verify( expected == 0p ); | 
|---|
| 245 | if(__atomic_compare_exchange_n(&this.ptr, &expected, active_thread(), false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 246 | park(); | 
|---|
| 247 | /* paranoid */ verify( this.ptr == 1p ); | 
|---|
| 248 | return true; | 
|---|
| 249 | } | 
|---|
| 250 | } | 
|---|
| 251 | } | 
|---|
| 252 |  | 
|---|
| 253 | // Mark as fulfilled, wake thread if needed | 
|---|
| 254 | // return true if a thread was unparked | 
|---|
| 255 | bool post(oneshot & this) { | 
|---|
| 256 | struct $thread * got = __atomic_exchange_n( &this.ptr, 1p, __ATOMIC_SEQ_CST); | 
|---|
| 257 | if( got == 0p ) return false; | 
|---|
| 258 | unpark( got ); | 
|---|
| 259 | return true; | 
|---|
| 260 | } | 
|---|
| 261 | } | 
|---|
| 262 |  | 
|---|
| 263 | // base types for future to build upon | 
|---|
| 264 | // It is based on the 'oneshot' type to allow multiple futures | 
|---|
| 265 | // to block on the same instance, permitting users to block a single | 
|---|
| 266 | // thread on "any of" [a given set of] futures. | 
|---|
| 267 | // does not support multiple threads waiting on the same future | 
|---|
| 268 | struct future_t { | 
|---|
| 269 | // Internal state : | 
|---|
| 270 | //     0p      : is initial state (wait will block) | 
|---|
| 271 | //     1p      : fulfilled (wait won't block) | 
|---|
| 272 | //     2p      : in progress () | 
|---|
| 273 | //     3p      : abandoned, server should delete | 
|---|
| 274 | // any oneshot : a context has been setup to wait, a thread could wait on it | 
|---|
| 275 | struct oneshot * volatile ptr; | 
|---|
| 276 | }; | 
|---|
| 277 |  | 
|---|
| 278 | static inline { | 
|---|
| 279 | void  ?{}(future_t & this) { | 
|---|
| 280 | this.ptr = 0p; | 
|---|
| 281 | } | 
|---|
| 282 |  | 
|---|
| 283 | void ^?{}(future_t &) {} | 
|---|
| 284 |  | 
|---|
| 285 | void reset(future_t & this) { | 
|---|
| 286 | // needs to be in 0p or 1p | 
|---|
| 287 | __atomic_exchange_n( &this.ptr, 0p, __ATOMIC_SEQ_CST); | 
|---|
| 288 | } | 
|---|
| 289 |  | 
|---|
| 290 | // check if the future is available | 
|---|
| 291 | bool available( future_t & this ) { | 
|---|
| 292 | return this.ptr == 1p; | 
|---|
| 293 | } | 
|---|
| 294 |  | 
|---|
| 295 | // Prepare the future to be waited on | 
|---|
| 296 | // intented to be use by wait, wait_any, waitfor, etc. rather than used directly | 
|---|
| 297 | bool setup( future_t & this, oneshot & wait_ctx ) { | 
|---|
| 298 | /* paranoid */ verify( wait_ctx.ptr == 0p ); | 
|---|
| 299 | // The future needs to set the wait context | 
|---|
| 300 | for() { | 
|---|
| 301 | struct oneshot * expected = this.ptr; | 
|---|
| 302 | // Is the future already fulfilled? | 
|---|
| 303 | if(expected == 1p) return false; // Yes, just return false (didn't block) | 
|---|
| 304 |  | 
|---|
| 305 | // The future is not fulfilled, try to setup the wait context | 
|---|
| 306 | /* paranoid */ verify( expected == 0p ); | 
|---|
| 307 | if(__atomic_compare_exchange_n(&this.ptr, &expected, &wait_ctx, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 308 | return true; | 
|---|
| 309 | } | 
|---|
| 310 | } | 
|---|
| 311 | } | 
|---|
| 312 |  | 
|---|
| 313 | // Stop waiting on a future | 
|---|
| 314 | // When multiple futures are waited for together in "any of" pattern | 
|---|
| 315 | // futures that weren't fulfilled before the thread woke up | 
|---|
| 316 | // should retract the wait ctx | 
|---|
| 317 | // intented to be use by wait, wait_any, waitfor, etc. rather than used directly | 
|---|
| 318 | void retract( future_t & this, oneshot & wait_ctx ) { | 
|---|
| 319 | // Remove the wait context | 
|---|
| 320 | struct oneshot * got = __atomic_exchange_n( &this.ptr, 0p, __ATOMIC_SEQ_CST); | 
|---|
| 321 |  | 
|---|
| 322 | // got == 0p: future was never actually setup, just return | 
|---|
| 323 | if( got == 0p ) return; | 
|---|
| 324 |  | 
|---|
| 325 | // got == wait_ctx: since fulfil does an atomic_swap, | 
|---|
| 326 | // if we got back the original then no one else saw context | 
|---|
| 327 | // It is safe to delete (which could happen after the return) | 
|---|
| 328 | if( got == &wait_ctx ) return; | 
|---|
| 329 |  | 
|---|
| 330 | // got == 1p: the future is ready and the context was fully consumed | 
|---|
| 331 | // the server won't use the pointer again | 
|---|
| 332 | // It is safe to delete (which could happen after the return) | 
|---|
| 333 | if( got == 1p ) return; | 
|---|
| 334 |  | 
|---|
| 335 | // got == 2p: the future is ready but the context hasn't fully been consumed | 
|---|
| 336 | // spin until it is safe to move on | 
|---|
| 337 | if( got == 2p ) { | 
|---|
| 338 | while( this.ptr != 1p ) Pause(); | 
|---|
| 339 | return; | 
|---|
| 340 | } | 
|---|
| 341 |  | 
|---|
| 342 | // got == any thing else, something wen't wrong here, abort | 
|---|
| 343 | abort("Future in unexpected state"); | 
|---|
| 344 | } | 
|---|
| 345 |  | 
|---|
| 346 | // Mark the future as abandoned, meaning it will be deleted by the server | 
|---|
| 347 | bool abandon( future_t & this ) { | 
|---|
| 348 | /* paranoid */ verify( this.ptr != 3p ); | 
|---|
| 349 |  | 
|---|
| 350 | // Mark the future as abandonned | 
|---|
| 351 | struct oneshot * got = __atomic_exchange_n( &this.ptr, 3p, __ATOMIC_SEQ_CST); | 
|---|
| 352 |  | 
|---|
| 353 | // If the future isn't already fulfilled, let the server delete it | 
|---|
| 354 | if( got == 0p ) return false; | 
|---|
| 355 |  | 
|---|
| 356 | // got == 2p: the future is ready but the context hasn't fully been consumed | 
|---|
| 357 | // spin until it is safe to move on | 
|---|
| 358 | if( got == 2p ) { | 
|---|
| 359 | while( this.ptr != 1p ) Pause(); | 
|---|
| 360 | got = 1p; | 
|---|
| 361 | } | 
|---|
| 362 |  | 
|---|
| 363 | // The future is completed delete it now | 
|---|
| 364 | /* paranoid */ verify( this.ptr != 1p ); | 
|---|
| 365 | free( &this ); | 
|---|
| 366 | return true; | 
|---|
| 367 | } | 
|---|
| 368 |  | 
|---|
| 369 | // from the server side, mark the future as fulfilled | 
|---|
| 370 | // delete it if needed | 
|---|
| 371 | bool fulfil( future_t & this ) { | 
|---|
| 372 | for() { | 
|---|
| 373 | struct oneshot * expected = this.ptr; | 
|---|
| 374 | // was this abandoned? | 
|---|
| 375 | #if defined(__GNUC__) && __GNUC__ >= 7 | 
|---|
| 376 | #pragma GCC diagnostic push | 
|---|
| 377 | #pragma GCC diagnostic ignored "-Wfree-nonheap-object" | 
|---|
| 378 | #endif | 
|---|
| 379 | if( expected == 3p ) { free( &this ); return false; } | 
|---|
| 380 | #if defined(__GNUC__) && __GNUC__ >= 7 | 
|---|
| 381 | #pragma GCC diagnostic pop | 
|---|
| 382 | #endif | 
|---|
| 383 |  | 
|---|
| 384 | /* paranoid */ verify( expected != 1p ); // Future is already fulfilled, should not happen | 
|---|
| 385 | /* paranoid */ verify( expected != 2p ); // Future is bein fulfilled by someone else, this is even less supported then the previous case. | 
|---|
| 386 |  | 
|---|
| 387 | // If there is a wait context, we need to consume it and mark it as consumed after | 
|---|
| 388 | // If there is no context then we can skip the in progress phase | 
|---|
| 389 | struct oneshot * want = expected == 0p ? 1p : 2p; | 
|---|
| 390 | if(__atomic_compare_exchange_n(&this.ptr, &expected, want, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { | 
|---|
| 391 | if( expected == 0p ) { /* paranoid */ verify( this.ptr == 1p); return false; } | 
|---|
| 392 | bool ret = post( *expected ); | 
|---|
| 393 | __atomic_store_n( &this.ptr, 1p, __ATOMIC_SEQ_CST); | 
|---|
| 394 | return ret; | 
|---|
| 395 | } | 
|---|
| 396 | } | 
|---|
| 397 |  | 
|---|
| 398 | } | 
|---|
| 399 |  | 
|---|
| 400 | // Wait for the future to be fulfilled | 
|---|
| 401 | bool wait( future_t & this ) { | 
|---|
| 402 | oneshot temp; | 
|---|
| 403 | if( !setup(this, temp) ) return false; | 
|---|
| 404 |  | 
|---|
| 405 | // Wait context is setup, just wait on it | 
|---|
| 406 | bool ret = wait( temp ); | 
|---|
| 407 |  | 
|---|
| 408 | // Wait for the future to tru | 
|---|
| 409 | while( this.ptr == 2p ) Pause(); | 
|---|
| 410 | // Make sure the state makes sense | 
|---|
| 411 | // Should be fulfilled, could be in progress but it's out of date if so | 
|---|
| 412 | // since if that is the case, the oneshot was fulfilled (unparking this thread) | 
|---|
| 413 | // and the oneshot should not be needed any more | 
|---|
| 414 | __attribute__((unused)) struct oneshot * was = this.ptr; | 
|---|
| 415 | /* paranoid */ verifyf( was == 1p, "Expected this.ptr to be 1p, was %p\n", was ); | 
|---|
| 416 |  | 
|---|
| 417 | // Mark the future as fulfilled, to be consistent | 
|---|
| 418 | // with potential calls to avail | 
|---|
| 419 | // this.ptr = 1p; | 
|---|
| 420 | return ret; | 
|---|
| 421 | } | 
|---|
| 422 | } | 
|---|
| 423 | #endif | 
|---|