source: libcfa/src/concurrency/kernel.cfa@ ce9ed84

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since ce9ed84 was 34b2796, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Changed comments to ifdef
  • Property mode set to 100644
File size: 29.1 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 __STATS( unsigned long long last_tally = rdtscl(); )
174
175 // if we need to run some special setup, now is the time to do it.
176 if(this->init.thrd) {
177 this->init.thrd->curr_cluster = this->cltr;
178 __run_thread(this, this->init.thrd);
179 }
180
181 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
182
183 $thread * readyThread = 0p;
184 MAIN_LOOP:
185 for() {
186 #if 1
187 // Check if there is pending io
188 __maybe_io_drain( this );
189
190 // Try to get the next thread
191 readyThread = __next_thread( this->cltr );
192
193 if( !readyThread ) {
194 __cfa_io_flush( this );
195 readyThread = __next_thread_slow( this->cltr );
196 }
197
198 HALT:
199 if( !readyThread ) {
200 // Don't block if we are done
201 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
202
203 #if !defined(__CFA_NO_STATISTICS__)
204 __tls_stats()->ready.sleep.halts++;
205 #endif
206
207 // Push self to idle stack
208 mark_idle(this->cltr->procs, * this);
209
210 // Confirm the ready-queue is empty
211 readyThread = __next_thread_slow( this->cltr );
212 if( readyThread ) {
213 // A thread was found, cancel the halt
214 mark_awake(this->cltr->procs, * this);
215
216 #if !defined(__CFA_NO_STATISTICS__)
217 __tls_stats()->ready.sleep.cancels++;
218 #endif
219
220 // continue the mai loop
221 break HALT;
222 }
223
224 #if !defined(__CFA_NO_STATISTICS__)
225 if(this->print_halts) {
226 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
227 }
228 #endif
229
230 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
231
232 __disable_interrupts_hard();
233 eventfd_t val;
234 eventfd_read( this->idle, &val );
235 __enable_interrupts_hard();
236
237 #if !defined(__CFA_NO_STATISTICS__)
238 if(this->print_halts) {
239 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
240 }
241 #endif
242
243 // We were woken up, remove self from idle
244 mark_awake(this->cltr->procs, * this);
245
246 // DON'T just proceed, start looking again
247 continue MAIN_LOOP;
248 }
249
250 /* paranoid */ verify( readyThread );
251
252 // Reset io dirty bit
253 this->io.dirty = false;
254
255 // We found a thread run it
256 __run_thread(this, readyThread);
257
258 // Are we done?
259 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
260
261 #if !defined(__CFA_NO_STATISTICS__)
262 unsigned long long curr = rdtscl();
263 if(curr > (last_tally + 500000000)) {
264 __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
265 last_tally = curr;
266 }
267 #endif
268
269 if(this->io.pending && !this->io.dirty) {
270 __cfa_io_flush( this );
271 }
272
273 #else
274
275 SEARCH: {
276 /* paranoid */ verify( ! __preemption_enabled() );
277 /* paranoid */ verify( kernelTLS().this_proc_id );
278
279 // First, lock the scheduler since we are searching for a thread
280 ready_schedule_lock();
281
282 // Try to get the next thread
283 readyThread = pop_fast( this->cltr );
284 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
285
286 // If we can't find a thread, might as well flush any outstanding I/O
287 if(this->io.pending) { __cfa_io_flush( this ); }
288
289 // Spin a little on I/O, just in case
290 for(25) {
291 __maybe_io_drain( this );
292 readyThread = pop_fast( this->cltr );
293 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
294 }
295
296 // no luck, try stealing a few times
297 for(25) {
298 if( __maybe_io_drain( this ) ) {
299 readyThread = pop_fast( this->cltr );
300 } else {
301 readyThread = pop_slow( this->cltr );
302 }
303 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
304 }
305
306 // still no luck, search for a thread
307 readyThread = pop_search( this->cltr );
308 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
309
310 // Don't block if we are done
311 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
312
313 __STATS( __tls_stats()->ready.sleep.halts++; )
314
315 // Push self to idle stack
316 ready_schedule_unlock();
317 mark_idle(this->cltr->procs, * this);
318 ready_schedule_lock();
319
320 // Confirm the ready-queue is empty
321 __maybe_io_drain( this );
322 readyThread = pop_search( this->cltr );
323 ready_schedule_unlock();
324
325 if( readyThread ) {
326 // A thread was found, cancel the halt
327 mark_awake(this->cltr->procs, * this);
328
329 __STATS( __tls_stats()->ready.sleep.cancels++; )
330
331 // continue the main loop
332 break SEARCH;
333 }
334
335 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl()); )
336 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
337
338 // __disable_interrupts_hard();
339 eventfd_t val;
340 eventfd_read( this->idle, &val );
341 // __enable_interrupts_hard();
342
343 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl()); )
344
345 // We were woken up, remove self from idle
346 mark_awake(this->cltr->procs, * this);
347
348 // DON'T just proceed, start looking again
349 continue MAIN_LOOP;
350 }
351
352 RUN_THREAD:
353 /* paranoid */ verify( kernelTLS().this_proc_id );
354 /* paranoid */ verify( ! __preemption_enabled() );
355 /* paranoid */ verify( readyThread );
356
357 // Reset io dirty bit
358 this->io.dirty = false;
359
360 // We found a thread run it
361 __run_thread(this, readyThread);
362
363 // Are we done?
364 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
365
366 #if !defined(__CFA_NO_STATISTICS__)
367 unsigned long long curr = rdtscl();
368 if(curr > (last_tally + 500000000)) {
369 __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
370 last_tally = curr;
371 }
372 #endif
373
374 if(this->io.pending && !this->io.dirty) {
375 __cfa_io_flush( this );
376 }
377
378 ready_schedule_lock();
379 __maybe_io_drain( this );
380 ready_schedule_unlock();
381 #endif
382 }
383
384 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
385 }
386
387 __cfa_io_stop( this );
388
389 post( this->terminated );
390
391 if(this == mainProcessor) {
392 // HACK : the coroutine context switch expects this_thread to be set
393 // and it make sense for it to be set in all other cases except here
394 // fake it
395 __cfaabi_tls.this_thread = mainThread;
396 }
397
398 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
399}
400
401static int * __volatile_errno() __attribute__((noinline));
402static int * __volatile_errno() { asm(""); return &errno; }
403
404// KERNEL ONLY
405// runThread runs a thread by context switching
406// from the processor coroutine to the target thread
407static void __run_thread(processor * this, $thread * thrd_dst) {
408 /* paranoid */ verify( ! __preemption_enabled() );
409 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
410 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
411 __builtin_prefetch( thrd_dst->context.SP );
412
413 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
414
415 $coroutine * proc_cor = get_coroutine(this->runner);
416
417 // set state of processor coroutine to inactive
418 verify(proc_cor->state == Active);
419 proc_cor->state = Blocked;
420
421 // Actually run the thread
422 RUNNING: while(true) {
423 thrd_dst->preempted = __NO_PREEMPTION;
424 thrd_dst->state = Active;
425
426 // Update global state
427 kernelTLS().this_thread = thrd_dst;
428
429 /* paranoid */ verify( ! __preemption_enabled() );
430 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
431 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
432 /* paranoid */ verify( thrd_dst->context.SP );
433 /* paranoid */ verify( thrd_dst->state != Halted );
434 /* 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
435 /* 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
436 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
437
438
439
440 // set context switch to the thread that the processor is executing
441 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
442 // when __cfactx_switch returns we are back in the processor coroutine
443
444 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
445 /* 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 );
446 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
447 /* paranoid */ verify( thrd_dst->context.SP );
448 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
449 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
450 /* paranoid */ verify( ! __preemption_enabled() );
451
452 // Reset global state
453 kernelTLS().this_thread = 0p;
454
455 // We just finished running a thread, there are a few things that could have happened.
456 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
457 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
458 // 4 - Preempted
459 // In case 1, we may have won a race so we can't write to the state again.
460 // In case 2, we lost the race so we now own the thread.
461
462 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
463 // The thread was preempted, reschedule it and reset the flag
464 schedule_thread$( thrd_dst );
465 break RUNNING;
466 }
467
468 if(unlikely(thrd_dst->state == Halting)) {
469 // The thread has halted, it should never be scheduled/run again
470 // finish the thread
471 __thread_finish( thrd_dst );
472 break RUNNING;
473 }
474
475 /* paranoid */ verify( thrd_dst->state == Active );
476 thrd_dst->state = Blocked;
477
478 // set state of processor coroutine to active and the thread to inactive
479 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
480 switch(old_ticket) {
481 case TICKET_RUNNING:
482 // This is case 1, the regular case, nothing more is needed
483 break RUNNING;
484 case TICKET_UNBLOCK:
485 #if !defined(__CFA_NO_STATISTICS__)
486 __tls_stats()->ready.threads.threads++;
487 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
488 #endif
489 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
490 // In this case, just run it again.
491 continue RUNNING;
492 default:
493 // This makes no sense, something is wrong abort
494 abort();
495 }
496 }
497
498 // Just before returning to the processor, set the processor coroutine to active
499 proc_cor->state = Active;
500
501 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
502
503 #if !defined(__CFA_NO_STATISTICS__)
504 __tls_stats()->ready.threads.threads--;
505 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
506 #endif
507
508 /* paranoid */ verify( ! __preemption_enabled() );
509}
510
511// KERNEL_ONLY
512void returnToKernel() {
513 /* paranoid */ verify( ! __preemption_enabled() );
514 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
515 $thread * thrd_src = kernelTLS().this_thread;
516
517 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
518
519 // Run the thread on this processor
520 {
521 int local_errno = *__volatile_errno();
522 #if defined( __i386 ) || defined( __x86_64 )
523 __x87_store;
524 #endif
525 /* paranoid */ verify( proc_cor->context.SP );
526 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
527 __cfactx_switch( &thrd_src->context, &proc_cor->context );
528 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
529 #if defined( __i386 ) || defined( __x86_64 )
530 __x87_load;
531 #endif
532 *__volatile_errno() = local_errno;
533 }
534
535 #if !defined(__CFA_NO_STATISTICS__)
536 /* paranoid */ verify( thrd_src->last_proc != 0p );
537 if(thrd_src->last_proc != kernelTLS().this_processor) {
538 __tls_stats()->ready.threads.migration++;
539 }
540 #endif
541
542 /* paranoid */ verify( ! __preemption_enabled() );
543 /* 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 );
544 /* 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 );
545}
546
547//-----------------------------------------------------------------------------
548// Scheduler routines
549// KERNEL ONLY
550static void __schedule_thread( $thread * thrd ) {
551 /* paranoid */ verify( ! __preemption_enabled() );
552 /* paranoid */ verify( ready_schedule_islocked());
553 /* paranoid */ verify( thrd );
554 /* paranoid */ verify( thrd->state != Halted );
555 /* paranoid */ verify( thrd->curr_cluster );
556 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
557 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
558 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
559 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
560 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
561 /* paranoid */ #endif
562 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
563 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
564
565
566 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
567
568 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
569 struct cluster * cl = thrd->curr_cluster;
570 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
571
572 // push the thread to the cluster ready-queue
573 push( cl, thrd );
574
575 // variable thrd is no longer safe to use
576 thrd = 0xdeaddeaddeaddeadp;
577
578 // wake the cluster using the save variable.
579 __wake_one( cl );
580
581 #if !defined(__CFA_NO_STATISTICS__)
582 if( kernelTLS().this_stats ) {
583 __tls_stats()->ready.threads.threads++;
584 if(outside) {
585 __tls_stats()->ready.threads.extunpark++;
586 }
587 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
588 }
589 else {
590 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
591 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
592 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
593 }
594 #endif
595
596 /* paranoid */ verify( ready_schedule_islocked());
597 /* paranoid */ verify( ! __preemption_enabled() );
598}
599
600void schedule_thread$( $thread * thrd ) {
601 ready_schedule_lock();
602 __schedule_thread( thrd );
603 ready_schedule_unlock();
604}
605
606// KERNEL ONLY
607static inline $thread * __next_thread(cluster * this) with( *this ) {
608 /* paranoid */ verify( ! __preemption_enabled() );
609
610 ready_schedule_lock();
611 $thread * thrd = pop_fast( this );
612 ready_schedule_unlock();
613
614 /* paranoid */ verify( ! __preemption_enabled() );
615 return thrd;
616}
617
618// KERNEL ONLY
619static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
620 /* paranoid */ verify( ! __preemption_enabled() );
621
622 ready_schedule_lock();
623 $thread * thrd;
624 for(25) {
625 thrd = pop_slow( this );
626 if(thrd) goto RET;
627 }
628 thrd = pop_search( this );
629
630 RET:
631 ready_schedule_unlock();
632
633 /* paranoid */ verify( ! __preemption_enabled() );
634 return thrd;
635}
636
637static inline bool __must_unpark( $thread * thrd ) {
638 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
639 switch(old_ticket) {
640 case TICKET_RUNNING:
641 // Wake won the race, the thread will reschedule/rerun itself
642 return false;
643 case TICKET_BLOCKED:
644 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
645 /* paranoid */ verify( thrd->state == Blocked );
646 return true;
647 default:
648 // This makes no sense, something is wrong abort
649 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
650 }
651}
652
653void __kernel_unpark( $thread * thrd ) {
654 /* paranoid */ verify( ! __preemption_enabled() );
655 /* paranoid */ verify( ready_schedule_islocked());
656
657 if( !thrd ) return;
658
659 if(__must_unpark(thrd)) {
660 // Wake lost the race,
661 __schedule_thread( thrd );
662 }
663
664 /* paranoid */ verify( ready_schedule_islocked());
665 /* paranoid */ verify( ! __preemption_enabled() );
666}
667
668void unpark( $thread * thrd ) {
669 if( !thrd ) return;
670
671 if(__must_unpark(thrd)) {
672 disable_interrupts();
673 // Wake lost the race,
674 schedule_thread$( thrd );
675 enable_interrupts(false);
676 }
677}
678
679void park( void ) {
680 __disable_interrupts_checked();
681 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
682 returnToKernel();
683 __enable_interrupts_checked();
684
685}
686
687extern "C" {
688 // Leave the thread monitor
689 // last routine called by a thread.
690 // Should never return
691 void __cfactx_thrd_leave() {
692 $thread * thrd = active_thread();
693 $monitor * this = &thrd->self_mon;
694
695 // Lock the monitor now
696 lock( this->lock __cfaabi_dbg_ctx2 );
697
698 disable_interrupts();
699
700 /* paranoid */ verify( ! __preemption_enabled() );
701 /* paranoid */ verify( thrd->state == Active );
702 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
703 /* paranoid */ verify( kernelTLS().this_thread == thrd );
704 /* paranoid */ verify( thrd->context.SP );
705 /* 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 );
706 /* 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 );
707
708 thrd->state = Halting;
709 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
710 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
711 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
712
713 // Leave the thread
714 returnToKernel();
715
716 // Control flow should never reach here!
717 abort();
718 }
719}
720
721// KERNEL ONLY
722bool force_yield( __Preemption_Reason reason ) {
723 __disable_interrupts_checked();
724 $thread * thrd = kernelTLS().this_thread;
725 /* paranoid */ verify(thrd->state == Active);
726
727 // SKULLDUGGERY: It is possible that we are preempting this thread just before
728 // it was going to park itself. If that is the case and it is already using the
729 // intrusive fields then we can't use them to preempt the thread
730 // If that is the case, abandon the preemption.
731 bool preempted = false;
732 if(thrd->link.next == 0p) {
733 preempted = true;
734 thrd->preempted = reason;
735 returnToKernel();
736 }
737 __enable_interrupts_checked( false );
738 return preempted;
739}
740
741//=============================================================================================
742// Kernel Idle Sleep
743//=============================================================================================
744// Wake a thread from the front if there are any
745static void __wake_one(cluster * this) {
746 /* paranoid */ verify( ! __preemption_enabled() );
747 /* paranoid */ verify( ready_schedule_islocked() );
748
749 // Check if there is a sleeping processor
750 processor * p;
751 unsigned idle;
752 unsigned total;
753 [idle, total, p] = query_idles(this->procs);
754
755 // If no one is sleeping, we are done
756 if( idle == 0 ) return;
757
758 // We found a processor, wake it up
759 eventfd_t val;
760 val = 1;
761 eventfd_write( p->idle, val );
762
763 #if !defined(__CFA_NO_STATISTICS__)
764 if( kernelTLS().this_stats ) {
765 __tls_stats()->ready.sleep.wakes++;
766 }
767 else {
768 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
769 }
770 #endif
771
772 /* paranoid */ verify( ready_schedule_islocked() );
773 /* paranoid */ verify( ! __preemption_enabled() );
774
775 return;
776}
777
778// Unconditionnaly wake a thread
779void __wake_proc(processor * this) {
780 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
781
782 __disable_interrupts_checked();
783 /* paranoid */ verify( ! __preemption_enabled() );
784 eventfd_t val;
785 val = 1;
786 eventfd_write( this->idle, val );
787 __enable_interrupts_checked();
788}
789
790static void mark_idle(__cluster_proc_list & this, processor & proc) {
791 /* paranoid */ verify( ! __preemption_enabled() );
792 lock( this );
793 this.idle++;
794 /* paranoid */ verify( this.idle <= this.total );
795 remove(proc);
796 insert_first(this.idles, proc);
797 unlock( this );
798 /* paranoid */ verify( ! __preemption_enabled() );
799}
800
801static void mark_awake(__cluster_proc_list & this, processor & proc) {
802 /* paranoid */ verify( ! __preemption_enabled() );
803 lock( this );
804 this.idle--;
805 /* paranoid */ verify( this.idle >= 0 );
806 remove(proc);
807 insert_last(this.actives, proc);
808 unlock( this );
809 /* paranoid */ verify( ! __preemption_enabled() );
810}
811
812static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
813 /* paranoid */ verify( ! __preemption_enabled() );
814 /* paranoid */ verify( ready_schedule_islocked() );
815
816 for() {
817 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
818 if( 1 == (l % 2) ) { Pause(); continue; }
819 unsigned idle = this.idle;
820 unsigned total = this.total;
821 processor * proc = &this.idles`first;
822 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
823 asm volatile("": : :"memory");
824 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
825 return [idle, total, proc];
826 }
827
828 /* paranoid */ verify( ready_schedule_islocked() );
829 /* paranoid */ verify( ! __preemption_enabled() );
830}
831
832//=============================================================================================
833// Unexpected Terminating logic
834//=============================================================================================
835void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
836 $thread * thrd = __cfaabi_tls.this_thread;
837
838 if(thrd) {
839 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
840 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
841
842 if ( &thrd->self_cor != thrd->curr_cor ) {
843 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
844 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
845 }
846 else {
847 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
848 }
849 }
850 else {
851 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
852 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
853 }
854}
855
856int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
857 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
858}
859
860static __spinlock_t kernel_debug_lock;
861
862extern "C" {
863 void __cfaabi_bits_acquire() {
864 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
865 }
866
867 void __cfaabi_bits_release() {
868 unlock( kernel_debug_lock );
869 }
870}
871
872//=============================================================================================
873// Kernel Utilities
874//=============================================================================================
875#if defined(CFA_HAVE_LINUX_IO_URING_H)
876#include "io/types.hfa"
877#endif
878
879static inline bool __maybe_io_drain( processor * proc ) {
880 bool ret = false;
881 #if defined(CFA_HAVE_LINUX_IO_URING_H)
882 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
883
884 // Check if we should drain the queue
885 $io_context * ctx = proc->io.ctx;
886 unsigned head = *ctx->cq.head;
887 unsigned tail = *ctx->cq.tail;
888 if(head == tail) return false;
889 ready_schedule_lock();
890 ret = __cfa_io_drain( proc );
891 ready_schedule_unlock();
892 #endif
893 return ret;
894}
895
896//-----------------------------------------------------------------------------
897// Debug
898__cfaabi_dbg_debug_do(
899 extern "C" {
900 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
901 this.prev_name = prev_name;
902 this.prev_thrd = kernelTLS().this_thread;
903 }
904 }
905)
906
907//-----------------------------------------------------------------------------
908// Debug
909bool threading_enabled(void) __attribute__((const)) {
910 return true;
911}
912
913//-----------------------------------------------------------------------------
914// Statistics
915#if !defined(__CFA_NO_STATISTICS__)
916 void print_halts( processor & this ) {
917 this.print_halts = true;
918 }
919
920 void print_stats_now( cluster & this, int flags ) {
921 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
922 }
923#endif
924// Local Variables: //
925// mode: c //
926// tab-width: 4 //
927// End: //
Note: See TracBrowser for help on using the repository browser.