source: libcfa/src/concurrency/kernel.cfa@ b51e389c

ADT ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since b51e389c was 733fd3d, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added missing corctx_flag in assertion
  • Property mode set to 100644
File size: 29.7 KB
Line 
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>
24extern "C" {
25 #include <sys/eventfd.h>
26}
27
28//CFA Includes
29#include "kernel_private.hfa"
30#include "preemption.hfa"
31
32//Private includes
33#define __CFA_INVOKE_PRIVATE__
34#include "invoke.h"
35
36#if !defined(__CFA_NO_STATISTICS__)
37 #define __STATS( ...) __VA_ARGS__
38#else
39 #define __STATS( ...)
40#endif
41
42//-----------------------------------------------------------------------------
43// Some assembly required
44#if defined( __i386 )
45 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
46 // fcw : X87 FPU control word (preserved across function calls)
47 #define __x87_store \
48 uint32_t __mxcr; \
49 uint16_t __fcw; \
50 __asm__ volatile ( \
51 "stmxcsr %0\n" \
52 "fnstcw %1\n" \
53 : "=m" (__mxcr),\
54 "=m" (__fcw) \
55 )
56
57 #define __x87_load \
58 __asm__ volatile ( \
59 "fldcw %1\n" \
60 "ldmxcsr %0\n" \
61 ::"m" (__mxcr),\
62 "m" (__fcw) \
63 )
64
65#elif defined( __x86_64 )
66 #define __x87_store \
67 uint32_t __mxcr; \
68 uint16_t __fcw; \
69 __asm__ volatile ( \
70 "stmxcsr %0\n" \
71 "fnstcw %1\n" \
72 : "=m" (__mxcr),\
73 "=m" (__fcw) \
74 )
75
76 #define __x87_load \
77 __asm__ volatile ( \
78 "fldcw %1\n" \
79 "ldmxcsr %0\n" \
80 :: "m" (__mxcr),\
81 "m" (__fcw) \
82 )
83
84#elif defined( __arm__ )
85 #define __x87_store
86 #define __x87_load
87
88#elif defined( __aarch64__ )
89 #define __x87_store \
90 uint32_t __fpcntl[2]; \
91 __asm__ volatile ( \
92 "mrs x9, FPCR\n" \
93 "mrs x10, FPSR\n" \
94 "stp x9, x10, %0\n" \
95 : "=m" (__fpcntl) : : "x9", "x10" \
96 )
97
98 #define __x87_load \
99 __asm__ volatile ( \
100 "ldp x9, x10, %0\n" \
101 "msr FPSR, x10\n" \
102 "msr FPCR, x9\n" \
103 : "=m" (__fpcntl) : : "x9", "x10" \
104 )
105
106#else
107 #error unsupported hardware architecture
108#endif
109
110extern $thread * mainThread;
111extern processor * mainProcessor;
112
113//-----------------------------------------------------------------------------
114// Kernel Scheduling logic
115static $thread * __next_thread(cluster * this);
116static $thread * __next_thread_slow(cluster * this);
117static inline bool __must_unpark( $thread * thrd ) __attribute((nonnull(1)));
118static void __run_thread(processor * this, $thread * dst);
119static void __wake_one(cluster * cltr);
120
121static void mark_idle (__cluster_proc_list & idles, processor & proc);
122static void mark_awake(__cluster_proc_list & idles, processor & proc);
123static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
124
125extern void __cfa_io_start( processor * );
126extern bool __cfa_io_drain( processor * );
127extern void __cfa_io_flush( processor * );
128extern void __cfa_io_stop ( processor * );
129static inline bool __maybe_io_drain( processor * );
130
131extern void __disable_interrupts_hard();
132extern void __enable_interrupts_hard();
133
134static inline void __disable_interrupts_checked() {
135 /* paranoid */ verify( __preemption_enabled() );
136 disable_interrupts();
137 /* paranoid */ verify( ! __preemption_enabled() );
138}
139
140static inline void __enable_interrupts_checked( bool poll = true ) {
141 /* paranoid */ verify( ! __preemption_enabled() );
142 enable_interrupts( poll );
143 /* paranoid */ verify( __preemption_enabled() );
144}
145
146//=============================================================================================
147// Kernel Scheduling logic
148//=============================================================================================
149//Main of the processor contexts
150void main(processorCtx_t & runner) {
151 // Because of a bug, we couldn't initialized the seed on construction
152 // Do it here
153 __cfaabi_tls.rand_seed ^= rdtscl();
154 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
155 __tls_rand_advance_bck();
156
157 processor * this = runner.proc;
158 verify(this);
159
160 __cfa_io_start( this );
161
162 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
163 #if !defined(__CFA_NO_STATISTICS__)
164 if( this->print_halts ) {
165 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
166 }
167 #endif
168
169 {
170 // Setup preemption data
171 preemption_scope scope = { this };
172
173 // if we need to run some special setup, now is the time to do it.
174 if(this->init.thrd) {
175 this->init.thrd->curr_cluster = this->cltr;
176 __run_thread(this, this->init.thrd);
177 }
178
179 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
180
181 $thread * readyThread = 0p;
182 MAIN_LOOP:
183 for() {
184 #define OLD_MAIN 1
185 #if OLD_MAIN
186 // Check if there is pending io
187 __maybe_io_drain( this );
188
189 // Try to get the next thread
190 readyThread = __next_thread( this->cltr );
191
192 if( !readyThread ) {
193 __cfa_io_flush( this );
194 readyThread = __next_thread_slow( this->cltr );
195 }
196
197 HALT:
198 if( !readyThread ) {
199 // Don't block if we are done
200 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
201
202 #if !defined(__CFA_NO_STATISTICS__)
203 __tls_stats()->ready.sleep.halts++;
204 #endif
205
206 // Push self to idle stack
207 mark_idle(this->cltr->procs, * this);
208
209 // Confirm the ready-queue is empty
210 readyThread = __next_thread_slow( this->cltr );
211 if( readyThread ) {
212 // A thread was found, cancel the halt
213 mark_awake(this->cltr->procs, * this);
214
215 #if !defined(__CFA_NO_STATISTICS__)
216 __tls_stats()->ready.sleep.cancels++;
217 #endif
218
219 // continue the mai loop
220 break HALT;
221 }
222
223 #if !defined(__CFA_NO_STATISTICS__)
224 if(this->print_halts) {
225 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
226 }
227 #endif
228
229 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
230
231 __disable_interrupts_hard();
232 eventfd_t val;
233 eventfd_read( this->idle, &val );
234 __enable_interrupts_hard();
235
236 #if !defined(__CFA_NO_STATISTICS__)
237 if(this->print_halts) {
238 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
239 }
240 #endif
241
242 // We were woken up, remove self from idle
243 mark_awake(this->cltr->procs, * this);
244
245 // DON'T just proceed, start looking again
246 continue MAIN_LOOP;
247 }
248
249 /* paranoid */ verify( readyThread );
250
251 // Reset io dirty bit
252 this->io.dirty = false;
253
254 // We found a thread run it
255 __run_thread(this, readyThread);
256
257 // Are we done?
258 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
259
260 if(this->io.pending && !this->io.dirty) {
261 __cfa_io_flush( this );
262 }
263
264 #else
265 #warning new kernel loop
266 SEARCH: {
267 /* paranoid */ verify( ! __preemption_enabled() );
268
269 // First, lock the scheduler since we are searching for a thread
270 ready_schedule_lock();
271
272 // Try to get the next thread
273 readyThread = pop_fast( this->cltr );
274 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
275
276 // If we can't find a thread, might as well flush any outstanding I/O
277 if(this->io.pending) { __cfa_io_flush( this ); }
278
279 // Spin a little on I/O, just in case
280 for(5) {
281 __maybe_io_drain( this );
282 readyThread = pop_fast( this->cltr );
283 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
284 }
285
286 // no luck, try stealing a few times
287 for(5) {
288 if( __maybe_io_drain( this ) ) {
289 readyThread = pop_fast( this->cltr );
290 } else {
291 readyThread = pop_slow( this->cltr );
292 }
293 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
294 }
295
296 // still no luck, search for a thread
297 readyThread = pop_search( this->cltr );
298 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
299
300 // Don't block if we are done
301 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
302
303 __STATS( __tls_stats()->ready.sleep.halts++; )
304
305 // Push self to idle stack
306 ready_schedule_unlock();
307 mark_idle(this->cltr->procs, * this);
308 ready_schedule_lock();
309
310 // Confirm the ready-queue is empty
311 __maybe_io_drain( this );
312 readyThread = pop_search( this->cltr );
313 ready_schedule_unlock();
314
315 if( readyThread ) {
316 // A thread was found, cancel the halt
317 mark_awake(this->cltr->procs, * this);
318
319 __STATS( __tls_stats()->ready.sleep.cancels++; )
320
321 // continue the main loop
322 break SEARCH;
323 }
324
325 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
326 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
327
328 // __disable_interrupts_hard();
329 eventfd_t val;
330 eventfd_read( this->idle, &val );
331 // __enable_interrupts_hard();
332
333 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
334
335 // We were woken up, remove self from idle
336 mark_awake(this->cltr->procs, * this);
337
338 // DON'T just proceed, start looking again
339 continue MAIN_LOOP;
340 }
341
342 RUN_THREAD:
343 /* paranoid */ verify( ! __preemption_enabled() );
344 /* paranoid */ verify( readyThread );
345
346 // Reset io dirty bit
347 this->io.dirty = false;
348
349 // We found a thread run it
350 __run_thread(this, readyThread);
351
352 // Are we done?
353 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
354
355 if(this->io.pending && !this->io.dirty) {
356 __cfa_io_flush( this );
357 }
358
359 ready_schedule_lock();
360 __maybe_io_drain( this );
361 ready_schedule_unlock();
362 #endif
363 }
364
365 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
366 }
367
368 __cfa_io_stop( this );
369
370 post( this->terminated );
371
372 if(this == mainProcessor) {
373 // HACK : the coroutine context switch expects this_thread to be set
374 // and it make sense for it to be set in all other cases except here
375 // fake it
376 __cfaabi_tls.this_thread = mainThread;
377 }
378
379 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
380}
381
382static int * __volatile_errno() __attribute__((noinline));
383static int * __volatile_errno() { asm(""); return &errno; }
384
385// KERNEL ONLY
386// runThread runs a thread by context switching
387// from the processor coroutine to the target thread
388static void __run_thread(processor * this, $thread * thrd_dst) {
389 /* paranoid */ verify( ! __preemption_enabled() );
390 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
391 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
392 __builtin_prefetch( thrd_dst->context.SP );
393
394 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
395
396 $coroutine * proc_cor = get_coroutine(this->runner);
397
398 // set state of processor coroutine to inactive
399 verify(proc_cor->state == Active);
400 proc_cor->state = Blocked;
401
402 // Actually run the thread
403 RUNNING: while(true) {
404 thrd_dst->preempted = __NO_PREEMPTION;
405 thrd_dst->state = Active;
406
407 // Update global state
408 kernelTLS().this_thread = thrd_dst;
409
410 /* paranoid */ verify( ! __preemption_enabled() );
411 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
412 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
413 /* paranoid */ verify( thrd_dst->context.SP );
414 /* paranoid */ verify( thrd_dst->state != Halted );
415 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor || thrd_dst->corctx_flag, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
416 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor || thrd_dst->corctx_flag, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
417 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
418
419
420
421 // set context switch to the thread that the processor is executing
422 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
423 // when __cfactx_switch returns we are back in the processor coroutine
424
425
426
427 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
428 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->corctx_flag, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
429 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->corctx_flag, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
430 /* paranoid */ verify( thrd_dst->context.SP );
431 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
432 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
433 /* paranoid */ verify( ! __preemption_enabled() );
434
435 // Reset global state
436 kernelTLS().this_thread = 0p;
437
438 // We just finished running a thread, there are a few things that could have happened.
439 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
440 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
441 // 4 - Preempted
442 // In case 1, we may have won a race so we can't write to the state again.
443 // In case 2, we lost the race so we now own the thread.
444
445 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
446 // The thread was preempted, reschedule it and reset the flag
447 schedule_thread$( thrd_dst );
448 break RUNNING;
449 }
450
451 if(unlikely(thrd_dst->state == Halting)) {
452 // The thread has halted, it should never be scheduled/run again
453 // finish the thread
454 __thread_finish( thrd_dst );
455 break RUNNING;
456 }
457
458 /* paranoid */ verify( thrd_dst->state == Active );
459 thrd_dst->state = Blocked;
460
461 // set state of processor coroutine to active and the thread to inactive
462 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
463 switch(old_ticket) {
464 case TICKET_RUNNING:
465 // This is case 1, the regular case, nothing more is needed
466 break RUNNING;
467 case TICKET_UNBLOCK:
468 #if !defined(__CFA_NO_STATISTICS__)
469 __tls_stats()->ready.threads.threads++;
470 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
471 #endif
472 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
473 // In this case, just run it again.
474 continue RUNNING;
475 default:
476 // This makes no sense, something is wrong abort
477 abort();
478 }
479 }
480
481 // Just before returning to the processor, set the processor coroutine to active
482 proc_cor->state = Active;
483
484 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
485
486 #if !defined(__CFA_NO_STATISTICS__)
487 __tls_stats()->ready.threads.threads--;
488 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
489 #endif
490
491 /* paranoid */ verify( ! __preemption_enabled() );
492}
493
494// KERNEL_ONLY
495void returnToKernel() {
496 /* paranoid */ verify( ! __preemption_enabled() );
497 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
498 $thread * thrd_src = kernelTLS().this_thread;
499
500 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
501
502 // Run the thread on this processor
503 {
504 int local_errno = *__volatile_errno();
505 #if defined( __i386 ) || defined( __x86_64 )
506 __x87_store;
507 #endif
508 /* paranoid */ verify( proc_cor->context.SP );
509 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
510 __cfactx_switch( &thrd_src->context, &proc_cor->context );
511 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
512 #if defined( __i386 ) || defined( __x86_64 )
513 __x87_load;
514 #endif
515 *__volatile_errno() = local_errno;
516 }
517
518 #if !defined(__CFA_NO_STATISTICS__)
519 /* paranoid */ verify( thrd_src->last_proc != 0p );
520 if(thrd_src->last_proc != kernelTLS().this_processor) {
521 __tls_stats()->ready.threads.migration++;
522 }
523 #endif
524
525 /* paranoid */ verify( ! __preemption_enabled() );
526 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ) || thrd_src->corctx_flag, "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
527 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit) || thrd_src->corctx_flag, "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
528}
529
530//-----------------------------------------------------------------------------
531// Scheduler routines
532// KERNEL ONLY
533static void __schedule_thread( $thread * thrd ) {
534 /* paranoid */ verify( ! __preemption_enabled() );
535 /* paranoid */ verify( ready_schedule_islocked());
536 /* paranoid */ verify( thrd );
537 /* paranoid */ verify( thrd->state != Halted );
538 /* paranoid */ verify( thrd->curr_cluster );
539 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
540 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
541 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
542 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
543 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
544 /* paranoid */ #endif
545 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
546 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
547
548 const bool local = thrd->state != Start;
549 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
550
551 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
552 struct cluster * cl = thrd->curr_cluster;
553 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
554
555 // push the thread to the cluster ready-queue
556 push( cl, thrd, local );
557
558 // variable thrd is no longer safe to use
559 thrd = 0xdeaddeaddeaddeadp;
560
561 // wake the cluster using the save variable.
562 __wake_one( cl );
563
564 #if !defined(__CFA_NO_STATISTICS__)
565 if( kernelTLS().this_stats ) {
566 __tls_stats()->ready.threads.threads++;
567 if(outside) {
568 __tls_stats()->ready.threads.extunpark++;
569 }
570 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
571 }
572 else {
573 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
574 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
575 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
576 }
577 #endif
578
579 /* paranoid */ verify( ready_schedule_islocked());
580 /* paranoid */ verify( ! __preemption_enabled() );
581}
582
583void schedule_thread$( $thread * thrd ) {
584 ready_schedule_lock();
585 __schedule_thread( thrd );
586 ready_schedule_unlock();
587}
588
589// KERNEL ONLY
590static inline $thread * __next_thread(cluster * this) with( *this ) {
591 /* paranoid */ verify( ! __preemption_enabled() );
592
593 ready_schedule_lock();
594 $thread * thrd = pop_fast( this );
595 ready_schedule_unlock();
596
597 /* paranoid */ verify( ! __preemption_enabled() );
598 return thrd;
599}
600
601// KERNEL ONLY
602static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
603 /* paranoid */ verify( ! __preemption_enabled() );
604
605 ready_schedule_lock();
606 $thread * thrd;
607 for(25) {
608 thrd = pop_slow( this );
609 if(thrd) goto RET;
610 }
611 thrd = pop_search( this );
612
613 RET:
614 ready_schedule_unlock();
615
616 /* paranoid */ verify( ! __preemption_enabled() );
617 return thrd;
618}
619
620static inline bool __must_unpark( $thread * thrd ) {
621 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
622 switch(old_ticket) {
623 case TICKET_RUNNING:
624 // Wake won the race, the thread will reschedule/rerun itself
625 return false;
626 case TICKET_BLOCKED:
627 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
628 /* paranoid */ verify( thrd->state == Blocked );
629 return true;
630 default:
631 // This makes no sense, something is wrong abort
632 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
633 }
634}
635
636void __kernel_unpark( $thread * thrd ) {
637 /* paranoid */ verify( ! __preemption_enabled() );
638 /* paranoid */ verify( ready_schedule_islocked());
639
640 if( !thrd ) return;
641
642 if(__must_unpark(thrd)) {
643 // Wake lost the race,
644 __schedule_thread( thrd );
645 }
646
647 /* paranoid */ verify( ready_schedule_islocked());
648 /* paranoid */ verify( ! __preemption_enabled() );
649}
650
651void unpark( $thread * thrd ) {
652 if( !thrd ) return;
653
654 if(__must_unpark(thrd)) {
655 disable_interrupts();
656 // Wake lost the race,
657 schedule_thread$( thrd );
658 enable_interrupts(false);
659 }
660}
661
662void park( void ) {
663 __disable_interrupts_checked();
664 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
665 returnToKernel();
666 __enable_interrupts_checked();
667
668}
669
670extern "C" {
671 // Leave the thread monitor
672 // last routine called by a thread.
673 // Should never return
674 void __cfactx_thrd_leave() {
675 $thread * thrd = active_thread();
676 $monitor * this = &thrd->self_mon;
677
678 // Lock the monitor now
679 lock( this->lock __cfaabi_dbg_ctx2 );
680
681 disable_interrupts();
682
683 /* paranoid */ verify( ! __preemption_enabled() );
684 /* paranoid */ verify( thrd->state == Active );
685 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
686 /* paranoid */ verify( kernelTLS().this_thread == thrd );
687 /* paranoid */ verify( thrd->context.SP );
688 /* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) > ((uintptr_t)__get_stack(thrd->curr_cor)->limit), "ERROR : $thread %p has been corrupted.\n StackPointer too large.\n", thrd );
689 /* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) < ((uintptr_t)__get_stack(thrd->curr_cor)->base ), "ERROR : $thread %p has been corrupted.\n StackPointer too small.\n", thrd );
690
691 thrd->state = Halting;
692 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
693 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
694 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
695
696 // Leave the thread
697 returnToKernel();
698
699 // Control flow should never reach here!
700 abort();
701 }
702}
703
704// KERNEL ONLY
705bool force_yield( __Preemption_Reason reason ) {
706 __disable_interrupts_checked();
707 $thread * thrd = kernelTLS().this_thread;
708 /* paranoid */ verify(thrd->state == Active);
709
710 // SKULLDUGGERY: It is possible that we are preempting this thread just before
711 // it was going to park itself. If that is the case and it is already using the
712 // intrusive fields then we can't use them to preempt the thread
713 // If that is the case, abandon the preemption.
714 bool preempted = false;
715 if(thrd->link.next == 0p) {
716 preempted = true;
717 thrd->preempted = reason;
718 returnToKernel();
719 }
720 __enable_interrupts_checked( false );
721 return preempted;
722}
723
724//=============================================================================================
725// Kernel Idle Sleep
726//=============================================================================================
727// Wake a thread from the front if there are any
728static void __wake_one(cluster * this) {
729 /* paranoid */ verify( ! __preemption_enabled() );
730 /* paranoid */ verify( ready_schedule_islocked() );
731
732 // Check if there is a sleeping processor
733 processor * p;
734 unsigned idle;
735 unsigned total;
736 [idle, total, p] = query_idles(this->procs);
737
738 // If no one is sleeping, we are done
739 if( idle == 0 ) return;
740
741 // We found a processor, wake it up
742 eventfd_t val;
743 val = 1;
744 eventfd_write( p->idle, val );
745
746 #if !defined(__CFA_NO_STATISTICS__)
747 if( kernelTLS().this_stats ) {
748 __tls_stats()->ready.sleep.wakes++;
749 }
750 else {
751 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
752 }
753 #endif
754
755 /* paranoid */ verify( ready_schedule_islocked() );
756 /* paranoid */ verify( ! __preemption_enabled() );
757
758 return;
759}
760
761// Unconditionnaly wake a thread
762void __wake_proc(processor * this) {
763 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
764
765 __disable_interrupts_checked();
766 /* paranoid */ verify( ! __preemption_enabled() );
767 eventfd_t val;
768 val = 1;
769 eventfd_write( this->idle, val );
770 __enable_interrupts_checked();
771}
772
773static void mark_idle(__cluster_proc_list & this, processor & proc) {
774 /* paranoid */ verify( ! __preemption_enabled() );
775 lock( this );
776 this.idle++;
777 /* paranoid */ verify( this.idle <= this.total );
778 remove(proc);
779 insert_first(this.idles, proc);
780 unlock( this );
781 /* paranoid */ verify( ! __preemption_enabled() );
782}
783
784static void mark_awake(__cluster_proc_list & this, processor & proc) {
785 /* paranoid */ verify( ! __preemption_enabled() );
786 lock( this );
787 this.idle--;
788 /* paranoid */ verify( this.idle >= 0 );
789 remove(proc);
790 insert_last(this.actives, proc);
791 unlock( this );
792 /* paranoid */ verify( ! __preemption_enabled() );
793}
794
795static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
796 /* paranoid */ verify( ! __preemption_enabled() );
797 /* paranoid */ verify( ready_schedule_islocked() );
798
799 for() {
800 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
801 if( 1 == (l % 2) ) { Pause(); continue; }
802 unsigned idle = this.idle;
803 unsigned total = this.total;
804 processor * proc = &this.idles`first;
805 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
806 asm volatile("": : :"memory");
807 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
808 return [idle, total, proc];
809 }
810
811 /* paranoid */ verify( ready_schedule_islocked() );
812 /* paranoid */ verify( ! __preemption_enabled() );
813}
814
815//=============================================================================================
816// Unexpected Terminating logic
817//=============================================================================================
818void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
819 $thread * thrd = __cfaabi_tls.this_thread;
820
821 if(thrd) {
822 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
823 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
824
825 if ( &thrd->self_cor != thrd->curr_cor ) {
826 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
827 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
828 }
829 else {
830 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
831 }
832 }
833 else {
834 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
835 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
836 }
837}
838
839int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
840 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
841}
842
843static __spinlock_t kernel_debug_lock;
844
845extern "C" {
846 void __cfaabi_bits_acquire() {
847 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
848 }
849
850 void __cfaabi_bits_release() {
851 unlock( kernel_debug_lock );
852 }
853}
854
855//=============================================================================================
856// Kernel Utilities
857//=============================================================================================
858#if defined(CFA_HAVE_LINUX_IO_URING_H)
859#include "io/types.hfa"
860#endif
861
862static inline bool __maybe_io_drain( processor * proc ) {
863 bool ret = false;
864 #if defined(CFA_HAVE_LINUX_IO_URING_H)
865 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
866
867 // Check if we should drain the queue
868 $io_context * ctx = proc->io.ctx;
869 unsigned head = *ctx->cq.head;
870 unsigned tail = *ctx->cq.tail;
871 if(head == tail) return false;
872 #if OLD_MAIN
873 ready_schedule_lock();
874 ret = __cfa_io_drain( proc );
875 ready_schedule_unlock();
876 #else
877 ret = __cfa_io_drain( proc );
878 #endif
879 #endif
880 return ret;
881}
882
883//-----------------------------------------------------------------------------
884// Debug
885__cfaabi_dbg_debug_do(
886 extern "C" {
887 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
888 this.prev_name = prev_name;
889 this.prev_thrd = kernelTLS().this_thread;
890 }
891 }
892)
893
894//-----------------------------------------------------------------------------
895// Debug
896bool threading_enabled(void) __attribute__((const)) {
897 return true;
898}
899
900//-----------------------------------------------------------------------------
901// Statistics
902#if !defined(__CFA_NO_STATISTICS__)
903 void print_halts( processor & this ) {
904 this.print_halts = true;
905 }
906
907 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
908 /* paranoid */ verify( cltr->stats );
909
910 processor * it = &list`first;
911 for(unsigned i = 0; i < count; i++) {
912 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
913 /* paranoid */ verify( it->local_data->this_stats );
914 __tally_stats( cltr->stats, it->local_data->this_stats );
915 it = &(*it)`next;
916 }
917 }
918
919 void crawl_cluster_stats( cluster & this ) {
920 // Stop the world, otherwise stats could get really messed-up
921 // this doesn't solve all problems but does solve many
922 // so it's probably good enough
923 uint_fast32_t last_size = ready_mutate_lock();
924
925 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
926 crawl_list(&this, this.procs.idles , this.procs.idle );
927
928 // Unlock the RWlock
929 ready_mutate_unlock( last_size );
930 }
931
932
933 void print_stats_now( cluster & this, int flags ) {
934 crawl_cluster_stats( this );
935 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
936 }
937#endif
938// Local Variables: //
939// mode: c //
940// tab-width: 4 //
941// End: //
Note: See TracBrowser for help on using the repository browser.