| 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 | // kernel.c --
|
|---|
| 8 | //
|
|---|
| 9 | // Author : Thierry Delisle
|
|---|
| 10 | // Created On : Tue Jan 17 12:27:26 2017
|
|---|
| 11 | // Last Modified By : Peter A. Buhr
|
|---|
| 12 | // Last Modified On : Mon Aug 31 07:08:20 2020
|
|---|
| 13 | // Update Count : 71
|
|---|
| 14 | //
|
|---|
| 15 |
|
|---|
| 16 | #define __cforall_thread__
|
|---|
| 17 | // #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
|
|---|
| 18 |
|
|---|
| 19 | //C Includes
|
|---|
| 20 | #include <errno.h>
|
|---|
| 21 | #include <stdio.h>
|
|---|
| 22 | #include <signal.h>
|
|---|
| 23 | #include <unistd.h>
|
|---|
| 24 |
|
|---|
| 25 | //CFA Includes
|
|---|
| 26 | #include "kernel_private.hfa"
|
|---|
| 27 | #include "preemption.hfa"
|
|---|
| 28 |
|
|---|
| 29 | //Private includes
|
|---|
| 30 | #define __CFA_INVOKE_PRIVATE__
|
|---|
| 31 | #include "invoke.h"
|
|---|
| 32 |
|
|---|
| 33 |
|
|---|
| 34 | //-----------------------------------------------------------------------------
|
|---|
| 35 | // Some assembly required
|
|---|
| 36 | #if defined( __i386 )
|
|---|
| 37 | // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
|
|---|
| 38 | // fcw : X87 FPU control word (preserved across function calls)
|
|---|
| 39 | #define __x87_store \
|
|---|
| 40 | uint32_t __mxcr; \
|
|---|
| 41 | uint16_t __fcw; \
|
|---|
| 42 | __asm__ volatile ( \
|
|---|
| 43 | "stmxcsr %0\n" \
|
|---|
| 44 | "fnstcw %1\n" \
|
|---|
| 45 | : "=m" (__mxcr),\
|
|---|
| 46 | "=m" (__fcw) \
|
|---|
| 47 | )
|
|---|
| 48 |
|
|---|
| 49 | #define __x87_load \
|
|---|
| 50 | __asm__ volatile ( \
|
|---|
| 51 | "fldcw %1\n" \
|
|---|
| 52 | "ldmxcsr %0\n" \
|
|---|
| 53 | ::"m" (__mxcr),\
|
|---|
| 54 | "m" (__fcw) \
|
|---|
| 55 | )
|
|---|
| 56 |
|
|---|
| 57 | #elif defined( __x86_64 )
|
|---|
| 58 | #define __x87_store \
|
|---|
| 59 | uint32_t __mxcr; \
|
|---|
| 60 | uint16_t __fcw; \
|
|---|
| 61 | __asm__ volatile ( \
|
|---|
| 62 | "stmxcsr %0\n" \
|
|---|
| 63 | "fnstcw %1\n" \
|
|---|
| 64 | : "=m" (__mxcr),\
|
|---|
| 65 | "=m" (__fcw) \
|
|---|
| 66 | )
|
|---|
| 67 |
|
|---|
| 68 | #define __x87_load \
|
|---|
| 69 | __asm__ volatile ( \
|
|---|
| 70 | "fldcw %1\n" \
|
|---|
| 71 | "ldmxcsr %0\n" \
|
|---|
| 72 | :: "m" (__mxcr),\
|
|---|
| 73 | "m" (__fcw) \
|
|---|
| 74 | )
|
|---|
| 75 |
|
|---|
| 76 | #elif defined( __arm__ )
|
|---|
| 77 | #define __x87_store
|
|---|
| 78 | #define __x87_load
|
|---|
| 79 |
|
|---|
| 80 | #elif defined( __aarch64__ )
|
|---|
| 81 | #define __x87_store \
|
|---|
| 82 | uint32_t __fpcntl[2]; \
|
|---|
| 83 | __asm__ volatile ( \
|
|---|
| 84 | "mrs x9, FPCR\n" \
|
|---|
| 85 | "mrs x10, FPSR\n" \
|
|---|
| 86 | "stp x9, x10, %0\n" \
|
|---|
| 87 | : "=m" (__fpcntl) : : "x9", "x10" \
|
|---|
| 88 | )
|
|---|
| 89 |
|
|---|
| 90 | #define __x87_load \
|
|---|
| 91 | __asm__ volatile ( \
|
|---|
| 92 | "ldp x9, x10, %0\n" \
|
|---|
| 93 | "msr FPSR, x10\n" \
|
|---|
| 94 | "msr FPCR, x9\n" \
|
|---|
| 95 | : "=m" (__fpcntl) : : "x9", "x10" \
|
|---|
| 96 | )
|
|---|
| 97 |
|
|---|
| 98 | #else
|
|---|
| 99 | #error unsupported hardware architecture
|
|---|
| 100 | #endif
|
|---|
| 101 |
|
|---|
| 102 | extern $thread * mainThread;
|
|---|
| 103 | extern processor * mainProcessor;
|
|---|
| 104 |
|
|---|
| 105 | //-----------------------------------------------------------------------------
|
|---|
| 106 | // Kernel Scheduling logic
|
|---|
| 107 | static $thread * __next_thread(cluster * this);
|
|---|
| 108 | static $thread * __next_thread_slow(cluster * this);
|
|---|
| 109 | static void __run_thread(processor * this, $thread * dst);
|
|---|
| 110 | static void __wake_one(struct __processor_id_t * id, cluster * cltr);
|
|---|
| 111 |
|
|---|
| 112 | static void push (__cluster_idles & idles, processor & proc);
|
|---|
| 113 | static void remove(__cluster_idles & idles, processor & proc);
|
|---|
| 114 | static [unsigned idle, unsigned total, * processor] query( & __cluster_idles idles );
|
|---|
| 115 |
|
|---|
| 116 |
|
|---|
| 117 | //=============================================================================================
|
|---|
| 118 | // Kernel Scheduling logic
|
|---|
| 119 | //=============================================================================================
|
|---|
| 120 | //Main of the processor contexts
|
|---|
| 121 | void main(processorCtx_t & runner) {
|
|---|
| 122 | // Because of a bug, we couldn't initialized the seed on construction
|
|---|
| 123 | // Do it here
|
|---|
| 124 | kernelTLS.rand_seed ^= rdtscl();
|
|---|
| 125 | kernelTLS.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
|
|---|
| 126 | __tls_rand_advance_bck();
|
|---|
| 127 |
|
|---|
| 128 | processor * this = runner.proc;
|
|---|
| 129 | verify(this);
|
|---|
| 130 |
|
|---|
| 131 | __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
|
|---|
| 132 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 133 | if( this->print_halts ) {
|
|---|
| 134 | __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->id, this->name, (void*)this);
|
|---|
| 135 | }
|
|---|
| 136 | #endif
|
|---|
| 137 |
|
|---|
| 138 | {
|
|---|
| 139 | // Setup preemption data
|
|---|
| 140 | preemption_scope scope = { this };
|
|---|
| 141 |
|
|---|
| 142 | __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
|
|---|
| 143 |
|
|---|
| 144 | $thread * readyThread = 0p;
|
|---|
| 145 | MAIN_LOOP:
|
|---|
| 146 | for() {
|
|---|
| 147 | // Try to get the next thread
|
|---|
| 148 | readyThread = __next_thread( this->cltr );
|
|---|
| 149 |
|
|---|
| 150 | if( !readyThread ) {
|
|---|
| 151 | readyThread = __next_thread_slow( this->cltr );
|
|---|
| 152 | }
|
|---|
| 153 |
|
|---|
| 154 | HALT:
|
|---|
| 155 | if( !readyThread ) {
|
|---|
| 156 | // Don't block if we are done
|
|---|
| 157 | if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
|
|---|
| 158 |
|
|---|
| 159 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 160 | __tls_stats()->ready.sleep.halts++;
|
|---|
| 161 | #endif
|
|---|
| 162 |
|
|---|
| 163 | // Push self to idle stack
|
|---|
| 164 | push(this->cltr->idles, * this);
|
|---|
| 165 |
|
|---|
| 166 | // Confirm the ready-queue is empty
|
|---|
| 167 | readyThread = __next_thread_slow( this->cltr );
|
|---|
| 168 | if( readyThread ) {
|
|---|
| 169 | // A thread was found, cancel the halt
|
|---|
| 170 | remove(this->cltr->idles, * this);
|
|---|
| 171 |
|
|---|
| 172 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 173 | __tls_stats()->ready.sleep.cancels++;
|
|---|
| 174 | #endif
|
|---|
| 175 |
|
|---|
| 176 | // continue the mai loop
|
|---|
| 177 | break HALT;
|
|---|
| 178 | }
|
|---|
| 179 |
|
|---|
| 180 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 181 | if(this->print_halts) {
|
|---|
| 182 | __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
|
|---|
| 183 | }
|
|---|
| 184 | #endif
|
|---|
| 185 |
|
|---|
| 186 | wait( this->idle );
|
|---|
| 187 |
|
|---|
| 188 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 189 | if(this->print_halts) {
|
|---|
| 190 | __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
|
|---|
| 191 | }
|
|---|
| 192 | #endif
|
|---|
| 193 |
|
|---|
| 194 | // We were woken up, remove self from idle
|
|---|
| 195 | remove(this->cltr->idles, * this);
|
|---|
| 196 |
|
|---|
| 197 | // DON'T just proceed, start looking again
|
|---|
| 198 | continue MAIN_LOOP;
|
|---|
| 199 | }
|
|---|
| 200 |
|
|---|
| 201 | /* paranoid */ verify( readyThread );
|
|---|
| 202 |
|
|---|
| 203 | // We found a thread run it
|
|---|
| 204 | __run_thread(this, readyThread);
|
|---|
| 205 |
|
|---|
| 206 | // Are we done?
|
|---|
| 207 | if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
|
|---|
| 208 | }
|
|---|
| 209 |
|
|---|
| 210 | __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
|
|---|
| 211 | }
|
|---|
| 212 |
|
|---|
| 213 | V( this->terminated );
|
|---|
| 214 |
|
|---|
| 215 | if(this == mainProcessor) {
|
|---|
| 216 | // HACK : the coroutine context switch expects this_thread to be set
|
|---|
| 217 | // and it make sense for it to be set in all other cases except here
|
|---|
| 218 | // fake it
|
|---|
| 219 | kernelTLS.this_thread = mainThread;
|
|---|
| 220 | }
|
|---|
| 221 |
|
|---|
| 222 | __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
|
|---|
| 223 | }
|
|---|
| 224 |
|
|---|
| 225 | static int * __volatile_errno() __attribute__((noinline));
|
|---|
| 226 | static int * __volatile_errno() { asm(""); return &errno; }
|
|---|
| 227 |
|
|---|
| 228 | // KERNEL ONLY
|
|---|
| 229 | // runThread runs a thread by context switching
|
|---|
| 230 | // from the processor coroutine to the target thread
|
|---|
| 231 | static void __run_thread(processor * this, $thread * thrd_dst) {
|
|---|
| 232 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 233 | /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
|
|---|
| 234 | /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
|
|---|
| 235 | __builtin_prefetch( thrd_dst->context.SP );
|
|---|
| 236 |
|
|---|
| 237 | $coroutine * proc_cor = get_coroutine(this->runner);
|
|---|
| 238 |
|
|---|
| 239 | // Update global state
|
|---|
| 240 | kernelTLS.this_thread = thrd_dst;
|
|---|
| 241 |
|
|---|
| 242 | // set state of processor coroutine to inactive
|
|---|
| 243 | verify(proc_cor->state == Active);
|
|---|
| 244 | proc_cor->state = Blocked;
|
|---|
| 245 |
|
|---|
| 246 | // Actually run the thread
|
|---|
| 247 | RUNNING: while(true) {
|
|---|
| 248 | thrd_dst->preempted = __NO_PREEMPTION;
|
|---|
| 249 | thrd_dst->state = Active;
|
|---|
| 250 |
|
|---|
| 251 | __cfaabi_dbg_debug_do(
|
|---|
| 252 | thrd_dst->park_stale = true;
|
|---|
| 253 | thrd_dst->unpark_stale = true;
|
|---|
| 254 | )
|
|---|
| 255 |
|
|---|
| 256 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 257 | /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
|
|---|
| 258 | /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
|
|---|
| 259 | /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
|
|---|
| 260 |
|
|---|
| 261 | // set context switch to the thread that the processor is executing
|
|---|
| 262 | verify( thrd_dst->context.SP );
|
|---|
| 263 | __cfactx_switch( &proc_cor->context, &thrd_dst->context );
|
|---|
| 264 | // when __cfactx_switch returns we are back in the processor coroutine
|
|---|
| 265 |
|
|---|
| 266 | /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
|
|---|
| 267 | /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
|
|---|
| 268 | /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
|
|---|
| 269 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 270 |
|
|---|
| 271 |
|
|---|
| 272 | // We just finished running a thread, there are a few things that could have happened.
|
|---|
| 273 | // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
|
|---|
| 274 | // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
|
|---|
| 275 | // 4 - Preempted
|
|---|
| 276 | // In case 1, we may have won a race so we can't write to the state again.
|
|---|
| 277 | // In case 2, we lost the race so we now own the thread.
|
|---|
| 278 |
|
|---|
| 279 | if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
|
|---|
| 280 | // The thread was preempted, reschedule it and reset the flag
|
|---|
| 281 | __schedule_thread( (__processor_id_t*)this, thrd_dst );
|
|---|
| 282 | break RUNNING;
|
|---|
| 283 | }
|
|---|
| 284 |
|
|---|
| 285 | if(unlikely(thrd_dst->state == Halted)) {
|
|---|
| 286 | // The thread has halted, it should never be scheduled/run again
|
|---|
| 287 | // We may need to wake someone up here since
|
|---|
| 288 | unpark( this->destroyer __cfaabi_dbg_ctx2 );
|
|---|
| 289 | this->destroyer = 0p;
|
|---|
| 290 | break RUNNING;
|
|---|
| 291 | }
|
|---|
| 292 |
|
|---|
| 293 | /* paranoid */ verify( thrd_dst->state == Active );
|
|---|
| 294 | thrd_dst->state = Blocked;
|
|---|
| 295 |
|
|---|
| 296 | // set state of processor coroutine to active and the thread to inactive
|
|---|
| 297 | int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
|
|---|
| 298 | __cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
|
|---|
| 299 | switch(old_ticket) {
|
|---|
| 300 | case 1:
|
|---|
| 301 | // This is case 1, the regular case, nothing more is needed
|
|---|
| 302 | break RUNNING;
|
|---|
| 303 | case 2:
|
|---|
| 304 | // This is case 2, the racy case, someone tried to run this thread before it finished blocking
|
|---|
| 305 | // In this case, just run it again.
|
|---|
| 306 | continue RUNNING;
|
|---|
| 307 | default:
|
|---|
| 308 | // This makes no sense, something is wrong abort
|
|---|
| 309 | abort();
|
|---|
| 310 | }
|
|---|
| 311 | }
|
|---|
| 312 |
|
|---|
| 313 | // Just before returning to the processor, set the processor coroutine to active
|
|---|
| 314 | proc_cor->state = Active;
|
|---|
| 315 | kernelTLS.this_thread = 0p;
|
|---|
| 316 |
|
|---|
| 317 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 318 | }
|
|---|
| 319 |
|
|---|
| 320 | // KERNEL_ONLY
|
|---|
| 321 | void returnToKernel() {
|
|---|
| 322 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 323 | $coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
|
|---|
| 324 | $thread * thrd_src = kernelTLS.this_thread;
|
|---|
| 325 |
|
|---|
| 326 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 327 | struct processor * last_proc = kernelTLS.this_processor;
|
|---|
| 328 | #endif
|
|---|
| 329 |
|
|---|
| 330 | // Run the thread on this processor
|
|---|
| 331 | {
|
|---|
| 332 | int local_errno = *__volatile_errno();
|
|---|
| 333 | #if defined( __i386 ) || defined( __x86_64 )
|
|---|
| 334 | __x87_store;
|
|---|
| 335 | #endif
|
|---|
| 336 | verify( proc_cor->context.SP );
|
|---|
| 337 | __cfactx_switch( &thrd_src->context, &proc_cor->context );
|
|---|
| 338 | #if defined( __i386 ) || defined( __x86_64 )
|
|---|
| 339 | __x87_load;
|
|---|
| 340 | #endif
|
|---|
| 341 | *__volatile_errno() = local_errno;
|
|---|
| 342 | }
|
|---|
| 343 |
|
|---|
| 344 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 345 | if(last_proc != kernelTLS.this_processor) {
|
|---|
| 346 | __tls_stats()->ready.threads.migration++;
|
|---|
| 347 | }
|
|---|
| 348 | #endif
|
|---|
| 349 |
|
|---|
| 350 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 351 | /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
|
|---|
| 352 | /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
|
|---|
| 353 | }
|
|---|
| 354 |
|
|---|
| 355 | //-----------------------------------------------------------------------------
|
|---|
| 356 | // Scheduler routines
|
|---|
| 357 | // KERNEL ONLY
|
|---|
| 358 | void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
|
|---|
| 359 | /* paranoid */ verify( thrd );
|
|---|
| 360 | /* paranoid */ verify( thrd->state != Halted );
|
|---|
| 361 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 362 | /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
|
|---|
| 363 | /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
|
|---|
| 364 | "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
|
|---|
| 365 | /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
|
|---|
| 366 | "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
|
|---|
| 367 | /* paranoid */ #endif
|
|---|
| 368 | /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
|
|---|
| 369 |
|
|---|
| 370 | if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
|
|---|
| 371 |
|
|---|
| 372 | ready_schedule_lock ( id );
|
|---|
| 373 | push( thrd->curr_cluster, thrd );
|
|---|
| 374 | __wake_one(id, thrd->curr_cluster);
|
|---|
| 375 | ready_schedule_unlock( id );
|
|---|
| 376 |
|
|---|
| 377 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 378 | }
|
|---|
| 379 |
|
|---|
| 380 | // KERNEL ONLY
|
|---|
| 381 | static inline $thread * __next_thread(cluster * this) with( *this ) {
|
|---|
| 382 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 383 |
|
|---|
| 384 | ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
|
|---|
| 385 | $thread * thrd = pop( this );
|
|---|
| 386 | ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
|
|---|
| 387 |
|
|---|
| 388 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 389 | return thrd;
|
|---|
| 390 | }
|
|---|
| 391 |
|
|---|
| 392 | // KERNEL ONLY
|
|---|
| 393 | static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
|
|---|
| 394 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 395 |
|
|---|
| 396 | ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
|
|---|
| 397 | $thread * thrd = pop_slow( this );
|
|---|
| 398 | ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
|
|---|
| 399 |
|
|---|
| 400 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 401 | return thrd;
|
|---|
| 402 | }
|
|---|
| 403 |
|
|---|
| 404 | // KERNEL ONLY unpark with out disabling interrupts
|
|---|
| 405 | void __unpark( struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
|
|---|
| 406 | // record activity
|
|---|
| 407 | __cfaabi_dbg_record_thrd( *thrd, false, caller );
|
|---|
| 408 |
|
|---|
| 409 | int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
|
|---|
| 410 | __cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
|
|---|
| 411 | switch(old_ticket) {
|
|---|
| 412 | case 1:
|
|---|
| 413 | // Wake won the race, the thread will reschedule/rerun itself
|
|---|
| 414 | break;
|
|---|
| 415 | case 0:
|
|---|
| 416 | /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
|
|---|
| 417 | /* paranoid */ verify( thrd->state == Blocked );
|
|---|
| 418 |
|
|---|
| 419 | // Wake lost the race,
|
|---|
| 420 | __schedule_thread( id, thrd );
|
|---|
| 421 | break;
|
|---|
| 422 | default:
|
|---|
| 423 | // This makes no sense, something is wrong abort
|
|---|
| 424 | abort();
|
|---|
| 425 | }
|
|---|
| 426 | }
|
|---|
| 427 |
|
|---|
| 428 | void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
|
|---|
| 429 | if( !thrd ) return;
|
|---|
| 430 |
|
|---|
| 431 | disable_interrupts();
|
|---|
| 432 | __unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
|
|---|
| 433 | enable_interrupts( __cfaabi_dbg_ctx );
|
|---|
| 434 | }
|
|---|
| 435 |
|
|---|
| 436 | void park( __cfaabi_dbg_ctx_param ) {
|
|---|
| 437 | /* paranoid */ verify( kernelTLS.preemption_state.enabled );
|
|---|
| 438 | disable_interrupts();
|
|---|
| 439 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 440 | /* paranoid */ verify( kernelTLS.this_thread->preempted == __NO_PREEMPTION );
|
|---|
| 441 |
|
|---|
| 442 | // record activity
|
|---|
| 443 | __cfaabi_dbg_record_thrd( *kernelTLS.this_thread, true, caller );
|
|---|
| 444 |
|
|---|
| 445 | returnToKernel();
|
|---|
| 446 |
|
|---|
| 447 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 448 | enable_interrupts( __cfaabi_dbg_ctx );
|
|---|
| 449 | /* paranoid */ verify( kernelTLS.preemption_state.enabled );
|
|---|
| 450 |
|
|---|
| 451 | }
|
|---|
| 452 |
|
|---|
| 453 | // KERNEL ONLY
|
|---|
| 454 | void __leave_thread() {
|
|---|
| 455 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 456 | returnToKernel();
|
|---|
| 457 | abort();
|
|---|
| 458 | }
|
|---|
| 459 |
|
|---|
| 460 | // KERNEL ONLY
|
|---|
| 461 | bool force_yield( __Preemption_Reason reason ) {
|
|---|
| 462 | /* paranoid */ verify( kernelTLS.preemption_state.enabled );
|
|---|
| 463 | disable_interrupts();
|
|---|
| 464 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 465 |
|
|---|
| 466 | $thread * thrd = kernelTLS.this_thread;
|
|---|
| 467 | /* paranoid */ verify(thrd->state == Active);
|
|---|
| 468 |
|
|---|
| 469 | // SKULLDUGGERY: It is possible that we are preempting this thread just before
|
|---|
| 470 | // it was going to park itself. If that is the case and it is already using the
|
|---|
| 471 | // intrusive fields then we can't use them to preempt the thread
|
|---|
| 472 | // If that is the case, abandon the preemption.
|
|---|
| 473 | bool preempted = false;
|
|---|
| 474 | if(thrd->link.next == 0p) {
|
|---|
| 475 | preempted = true;
|
|---|
| 476 | thrd->preempted = reason;
|
|---|
| 477 | returnToKernel();
|
|---|
| 478 | }
|
|---|
| 479 |
|
|---|
| 480 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 481 | enable_interrupts_noPoll();
|
|---|
| 482 | /* paranoid */ verify( kernelTLS.preemption_state.enabled );
|
|---|
| 483 |
|
|---|
| 484 | return preempted;
|
|---|
| 485 | }
|
|---|
| 486 |
|
|---|
| 487 | //=============================================================================================
|
|---|
| 488 | // Kernel Idle Sleep
|
|---|
| 489 | //=============================================================================================
|
|---|
| 490 | // Wake a thread from the front if there are any
|
|---|
| 491 | static void __wake_one(struct __processor_id_t * id, cluster * this) {
|
|---|
| 492 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 493 | /* paranoid */ verify( ready_schedule_islocked( id ) );
|
|---|
| 494 |
|
|---|
| 495 | // Check if there is a sleeping processor
|
|---|
| 496 | processor * p;
|
|---|
| 497 | unsigned idle;
|
|---|
| 498 | unsigned total;
|
|---|
| 499 | [idle, total, p] = query(this->idles);
|
|---|
| 500 |
|
|---|
| 501 | // If no one is sleeping, we are done
|
|---|
| 502 | if( idle == 0 ) return;
|
|---|
| 503 |
|
|---|
| 504 | // We found a processor, wake it up
|
|---|
| 505 | post( p->idle );
|
|---|
| 506 |
|
|---|
| 507 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 508 | __tls_stats()->ready.sleep.wakes++;
|
|---|
| 509 | #endif
|
|---|
| 510 |
|
|---|
| 511 | /* paranoid */ verify( ready_schedule_islocked( id ) );
|
|---|
| 512 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 513 |
|
|---|
| 514 | return;
|
|---|
| 515 | }
|
|---|
| 516 |
|
|---|
| 517 | // Unconditionnaly wake a thread
|
|---|
| 518 | void __wake_proc(processor * this) {
|
|---|
| 519 | __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
|
|---|
| 520 |
|
|---|
| 521 | disable_interrupts();
|
|---|
| 522 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 523 | bool ret = post( this->idle );
|
|---|
| 524 | enable_interrupts( __cfaabi_dbg_ctx );
|
|---|
| 525 | }
|
|---|
| 526 |
|
|---|
| 527 | static void push (__cluster_idles & this, processor & proc) {
|
|---|
| 528 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 529 | lock( this );
|
|---|
| 530 | this.idle++;
|
|---|
| 531 | /* paranoid */ verify( this.idle <= this.total );
|
|---|
| 532 |
|
|---|
| 533 | insert_first(this.list, proc);
|
|---|
| 534 | unlock( this );
|
|---|
| 535 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 536 | }
|
|---|
| 537 |
|
|---|
| 538 | static void remove(__cluster_idles & this, processor & proc) {
|
|---|
| 539 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 540 | lock( this );
|
|---|
| 541 | this.idle--;
|
|---|
| 542 | /* paranoid */ verify( this.idle >= 0 );
|
|---|
| 543 |
|
|---|
| 544 | remove(proc);
|
|---|
| 545 | unlock( this );
|
|---|
| 546 | /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
|
|---|
| 547 | }
|
|---|
| 548 |
|
|---|
| 549 | static [unsigned idle, unsigned total, * processor] query( & __cluster_idles this ) {
|
|---|
| 550 | for() {
|
|---|
| 551 | uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
|
|---|
| 552 | if( 1 == (l % 2) ) { Pause(); continue; }
|
|---|
| 553 | unsigned idle = this.idle;
|
|---|
| 554 | unsigned total = this.total;
|
|---|
| 555 | processor * proc = &this.list`first;
|
|---|
| 556 | // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
|
|---|
| 557 | asm volatile("": : :"memory");
|
|---|
| 558 | if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
|
|---|
| 559 | return [idle, total, proc];
|
|---|
| 560 | }
|
|---|
| 561 | }
|
|---|
| 562 |
|
|---|
| 563 | //=============================================================================================
|
|---|
| 564 | // Unexpected Terminating logic
|
|---|
| 565 | //=============================================================================================
|
|---|
| 566 | static __spinlock_t kernel_abort_lock;
|
|---|
| 567 | static bool kernel_abort_called = false;
|
|---|
| 568 |
|
|---|
| 569 | void * kernel_abort(void) __attribute__ ((__nothrow__)) {
|
|---|
| 570 | // abort cannot be recursively entered by the same or different processors because all signal handlers return when
|
|---|
| 571 | // the globalAbort flag is true.
|
|---|
| 572 | lock( kernel_abort_lock __cfaabi_dbg_ctx2 );
|
|---|
| 573 |
|
|---|
| 574 | // first task to abort ?
|
|---|
| 575 | if ( kernel_abort_called ) { // not first task to abort ?
|
|---|
| 576 | unlock( kernel_abort_lock );
|
|---|
| 577 |
|
|---|
| 578 | sigset_t mask;
|
|---|
| 579 | sigemptyset( &mask );
|
|---|
| 580 | sigaddset( &mask, SIGALRM ); // block SIGALRM signals
|
|---|
| 581 | sigaddset( &mask, SIGUSR1 ); // block SIGALRM signals
|
|---|
| 582 | sigsuspend( &mask ); // block the processor to prevent further damage during abort
|
|---|
| 583 | _exit( EXIT_FAILURE ); // if processor unblocks before it is killed, terminate it
|
|---|
| 584 | }
|
|---|
| 585 | else {
|
|---|
| 586 | kernel_abort_called = true;
|
|---|
| 587 | unlock( kernel_abort_lock );
|
|---|
| 588 | }
|
|---|
| 589 |
|
|---|
| 590 | return kernelTLS.this_thread;
|
|---|
| 591 | }
|
|---|
| 592 |
|
|---|
| 593 | void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
|
|---|
| 594 | $thread * thrd = ( $thread * ) kernel_data;
|
|---|
| 595 |
|
|---|
| 596 | if(thrd) {
|
|---|
| 597 | int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
|
|---|
| 598 | __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
|
|---|
| 599 |
|
|---|
| 600 | if ( &thrd->self_cor != thrd->curr_cor ) {
|
|---|
| 601 | len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
|
|---|
| 602 | __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
|
|---|
| 603 | }
|
|---|
| 604 | else {
|
|---|
| 605 | __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
|
|---|
| 606 | }
|
|---|
| 607 | }
|
|---|
| 608 | else {
|
|---|
| 609 | int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
|
|---|
| 610 | __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
|
|---|
| 611 | }
|
|---|
| 612 | }
|
|---|
| 613 |
|
|---|
| 614 | int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
|
|---|
| 615 | return get_coroutine(kernelTLS.this_thread) == get_coroutine(mainThread) ? 4 : 2;
|
|---|
| 616 | }
|
|---|
| 617 |
|
|---|
| 618 | static __spinlock_t kernel_debug_lock;
|
|---|
| 619 |
|
|---|
| 620 | extern "C" {
|
|---|
| 621 | void __cfaabi_bits_acquire() {
|
|---|
| 622 | lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
|
|---|
| 623 | }
|
|---|
| 624 |
|
|---|
| 625 | void __cfaabi_bits_release() {
|
|---|
| 626 | unlock( kernel_debug_lock );
|
|---|
| 627 | }
|
|---|
| 628 | }
|
|---|
| 629 |
|
|---|
| 630 | //=============================================================================================
|
|---|
| 631 | // Kernel Utilities
|
|---|
| 632 | //=============================================================================================
|
|---|
| 633 | //-----------------------------------------------------------------------------
|
|---|
| 634 | // Locks
|
|---|
| 635 | void ?{}( semaphore & this, int count = 1 ) {
|
|---|
| 636 | (this.lock){};
|
|---|
| 637 | this.count = count;
|
|---|
| 638 | (this.waiting){};
|
|---|
| 639 | }
|
|---|
| 640 | void ^?{}(semaphore & this) {}
|
|---|
| 641 |
|
|---|
| 642 | bool P(semaphore & this) with( this ){
|
|---|
| 643 | lock( lock __cfaabi_dbg_ctx2 );
|
|---|
| 644 | count -= 1;
|
|---|
| 645 | if ( count < 0 ) {
|
|---|
| 646 | // queue current task
|
|---|
| 647 | append( waiting, kernelTLS.this_thread );
|
|---|
| 648 |
|
|---|
| 649 | // atomically release spin lock and block
|
|---|
| 650 | unlock( lock );
|
|---|
| 651 | park( __cfaabi_dbg_ctx );
|
|---|
| 652 | return true;
|
|---|
| 653 | }
|
|---|
| 654 | else {
|
|---|
| 655 | unlock( lock );
|
|---|
| 656 | return false;
|
|---|
| 657 | }
|
|---|
| 658 | }
|
|---|
| 659 |
|
|---|
| 660 | bool V(semaphore & this) with( this ) {
|
|---|
| 661 | $thread * thrd = 0p;
|
|---|
| 662 | lock( lock __cfaabi_dbg_ctx2 );
|
|---|
| 663 | count += 1;
|
|---|
| 664 | if ( count <= 0 ) {
|
|---|
| 665 | // remove task at head of waiting list
|
|---|
| 666 | thrd = pop_head( waiting );
|
|---|
| 667 | }
|
|---|
| 668 |
|
|---|
| 669 | unlock( lock );
|
|---|
| 670 |
|
|---|
| 671 | // make new owner
|
|---|
| 672 | unpark( thrd __cfaabi_dbg_ctx2 );
|
|---|
| 673 |
|
|---|
| 674 | return thrd != 0p;
|
|---|
| 675 | }
|
|---|
| 676 |
|
|---|
| 677 | bool V(semaphore & this, unsigned diff) with( this ) {
|
|---|
| 678 | $thread * thrd = 0p;
|
|---|
| 679 | lock( lock __cfaabi_dbg_ctx2 );
|
|---|
| 680 | int release = max(-count, (int)diff);
|
|---|
| 681 | count += diff;
|
|---|
| 682 | for(release) {
|
|---|
| 683 | unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 );
|
|---|
| 684 | }
|
|---|
| 685 |
|
|---|
| 686 | unlock( lock );
|
|---|
| 687 |
|
|---|
| 688 | return thrd != 0p;
|
|---|
| 689 | }
|
|---|
| 690 |
|
|---|
| 691 | //-----------------------------------------------------------------------------
|
|---|
| 692 | // Debug
|
|---|
| 693 | __cfaabi_dbg_debug_do(
|
|---|
| 694 | extern "C" {
|
|---|
| 695 | void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
|
|---|
| 696 | this.prev_name = prev_name;
|
|---|
| 697 | this.prev_thrd = kernelTLS.this_thread;
|
|---|
| 698 | }
|
|---|
| 699 |
|
|---|
| 700 | void __cfaabi_dbg_record_thrd($thread & this, bool park, const char prev_name[]) {
|
|---|
| 701 | if(park) {
|
|---|
| 702 | this.park_caller = prev_name;
|
|---|
| 703 | this.park_stale = false;
|
|---|
| 704 | }
|
|---|
| 705 | else {
|
|---|
| 706 | this.unpark_caller = prev_name;
|
|---|
| 707 | this.unpark_stale = false;
|
|---|
| 708 | }
|
|---|
| 709 | }
|
|---|
| 710 | }
|
|---|
| 711 | )
|
|---|
| 712 |
|
|---|
| 713 | //-----------------------------------------------------------------------------
|
|---|
| 714 | // Debug
|
|---|
| 715 | bool threading_enabled(void) __attribute__((const)) {
|
|---|
| 716 | return true;
|
|---|
| 717 | }
|
|---|
| 718 |
|
|---|
| 719 | //-----------------------------------------------------------------------------
|
|---|
| 720 | // Statistics
|
|---|
| 721 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 722 | void print_halts( processor & this ) {
|
|---|
| 723 | this.print_halts = true;
|
|---|
| 724 | }
|
|---|
| 725 | #endif
|
|---|
| 726 | // Local Variables: //
|
|---|
| 727 | // mode: c //
|
|---|
| 728 | // tab-width: 4 //
|
|---|
| 729 | // End: //
|
|---|