source: libcfa/src/concurrency/kernel.cfa@ 93f74c0f

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 93f74c0f was b808625, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added option to ready-queue to push ignoring locality.
  • Property mode set to 100644
File size: 29.5 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, "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, "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 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
426 /* 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 );
427 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
428 /* paranoid */ verify( thrd_dst->context.SP );
429 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
430 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
431 /* paranoid */ verify( ! __preemption_enabled() );
432
433 // Reset global state
434 kernelTLS().this_thread = 0p;
435
436 // We just finished running a thread, there are a few things that could have happened.
437 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
438 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
439 // 4 - Preempted
440 // In case 1, we may have won a race so we can't write to the state again.
441 // In case 2, we lost the race so we now own the thread.
442
443 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
444 // The thread was preempted, reschedule it and reset the flag
445 schedule_thread$( thrd_dst );
446 break RUNNING;
447 }
448
449 if(unlikely(thrd_dst->state == Halting)) {
450 // The thread has halted, it should never be scheduled/run again
451 // finish the thread
452 __thread_finish( thrd_dst );
453 break RUNNING;
454 }
455
456 /* paranoid */ verify( thrd_dst->state == Active );
457 thrd_dst->state = Blocked;
458
459 // set state of processor coroutine to active and the thread to inactive
460 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
461 switch(old_ticket) {
462 case TICKET_RUNNING:
463 // This is case 1, the regular case, nothing more is needed
464 break RUNNING;
465 case TICKET_UNBLOCK:
466 #if !defined(__CFA_NO_STATISTICS__)
467 __tls_stats()->ready.threads.threads++;
468 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
469 #endif
470 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
471 // In this case, just run it again.
472 continue RUNNING;
473 default:
474 // This makes no sense, something is wrong abort
475 abort();
476 }
477 }
478
479 // Just before returning to the processor, set the processor coroutine to active
480 proc_cor->state = Active;
481
482 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
483
484 #if !defined(__CFA_NO_STATISTICS__)
485 __tls_stats()->ready.threads.threads--;
486 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
487 #endif
488
489 /* paranoid */ verify( ! __preemption_enabled() );
490}
491
492// KERNEL_ONLY
493void returnToKernel() {
494 /* paranoid */ verify( ! __preemption_enabled() );
495 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
496 $thread * thrd_src = kernelTLS().this_thread;
497
498 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
499
500 // Run the thread on this processor
501 {
502 int local_errno = *__volatile_errno();
503 #if defined( __i386 ) || defined( __x86_64 )
504 __x87_store;
505 #endif
506 /* paranoid */ verify( proc_cor->context.SP );
507 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
508 __cfactx_switch( &thrd_src->context, &proc_cor->context );
509 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
510 #if defined( __i386 ) || defined( __x86_64 )
511 __x87_load;
512 #endif
513 *__volatile_errno() = local_errno;
514 }
515
516 #if !defined(__CFA_NO_STATISTICS__)
517 /* paranoid */ verify( thrd_src->last_proc != 0p );
518 if(thrd_src->last_proc != kernelTLS().this_processor) {
519 __tls_stats()->ready.threads.migration++;
520 }
521 #endif
522
523 /* paranoid */ verify( ! __preemption_enabled() );
524 /* 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 );
525 /* 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 );
526}
527
528//-----------------------------------------------------------------------------
529// Scheduler routines
530// KERNEL ONLY
531static void __schedule_thread( $thread * thrd ) {
532 /* paranoid */ verify( ! __preemption_enabled() );
533 /* paranoid */ verify( ready_schedule_islocked());
534 /* paranoid */ verify( thrd );
535 /* paranoid */ verify( thrd->state != Halted );
536 /* paranoid */ verify( thrd->curr_cluster );
537 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
538 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
539 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
540 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
541 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
542 /* paranoid */ #endif
543 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
544 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
545
546 const bool local = thrd->state != Start;
547 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
548
549 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
550 struct cluster * cl = thrd->curr_cluster;
551 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
552
553 // push the thread to the cluster ready-queue
554 push( cl, thrd, local );
555
556 // variable thrd is no longer safe to use
557 thrd = 0xdeaddeaddeaddeadp;
558
559 // wake the cluster using the save variable.
560 __wake_one( cl );
561
562 #if !defined(__CFA_NO_STATISTICS__)
563 if( kernelTLS().this_stats ) {
564 __tls_stats()->ready.threads.threads++;
565 if(outside) {
566 __tls_stats()->ready.threads.extunpark++;
567 }
568 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
569 }
570 else {
571 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
572 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
573 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
574 }
575 #endif
576
577 /* paranoid */ verify( ready_schedule_islocked());
578 /* paranoid */ verify( ! __preemption_enabled() );
579}
580
581void schedule_thread$( $thread * thrd ) {
582 ready_schedule_lock();
583 __schedule_thread( thrd );
584 ready_schedule_unlock();
585}
586
587// KERNEL ONLY
588static inline $thread * __next_thread(cluster * this) with( *this ) {
589 /* paranoid */ verify( ! __preemption_enabled() );
590
591 ready_schedule_lock();
592 $thread * thrd = pop_fast( this );
593 ready_schedule_unlock();
594
595 /* paranoid */ verify( ! __preemption_enabled() );
596 return thrd;
597}
598
599// KERNEL ONLY
600static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
601 /* paranoid */ verify( ! __preemption_enabled() );
602
603 ready_schedule_lock();
604 $thread * thrd;
605 for(25) {
606 thrd = pop_slow( this );
607 if(thrd) goto RET;
608 }
609 thrd = pop_search( this );
610
611 RET:
612 ready_schedule_unlock();
613
614 /* paranoid */ verify( ! __preemption_enabled() );
615 return thrd;
616}
617
618static inline bool __must_unpark( $thread * thrd ) {
619 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
620 switch(old_ticket) {
621 case TICKET_RUNNING:
622 // Wake won the race, the thread will reschedule/rerun itself
623 return false;
624 case TICKET_BLOCKED:
625 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
626 /* paranoid */ verify( thrd->state == Blocked );
627 return true;
628 default:
629 // This makes no sense, something is wrong abort
630 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
631 }
632}
633
634void __kernel_unpark( $thread * thrd ) {
635 /* paranoid */ verify( ! __preemption_enabled() );
636 /* paranoid */ verify( ready_schedule_islocked());
637
638 if( !thrd ) return;
639
640 if(__must_unpark(thrd)) {
641 // Wake lost the race,
642 __schedule_thread( thrd );
643 }
644
645 /* paranoid */ verify( ready_schedule_islocked());
646 /* paranoid */ verify( ! __preemption_enabled() );
647}
648
649void unpark( $thread * thrd ) {
650 if( !thrd ) return;
651
652 if(__must_unpark(thrd)) {
653 disable_interrupts();
654 // Wake lost the race,
655 schedule_thread$( thrd );
656 enable_interrupts(false);
657 }
658}
659
660void park( void ) {
661 __disable_interrupts_checked();
662 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
663 returnToKernel();
664 __enable_interrupts_checked();
665
666}
667
668extern "C" {
669 // Leave the thread monitor
670 // last routine called by a thread.
671 // Should never return
672 void __cfactx_thrd_leave() {
673 $thread * thrd = active_thread();
674 $monitor * this = &thrd->self_mon;
675
676 // Lock the monitor now
677 lock( this->lock __cfaabi_dbg_ctx2 );
678
679 disable_interrupts();
680
681 /* paranoid */ verify( ! __preemption_enabled() );
682 /* paranoid */ verify( thrd->state == Active );
683 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
684 /* paranoid */ verify( kernelTLS().this_thread == thrd );
685 /* paranoid */ verify( thrd->context.SP );
686 /* 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 );
687 /* 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 );
688
689 thrd->state = Halting;
690 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
691 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
692 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
693
694 // Leave the thread
695 returnToKernel();
696
697 // Control flow should never reach here!
698 abort();
699 }
700}
701
702// KERNEL ONLY
703bool force_yield( __Preemption_Reason reason ) {
704 __disable_interrupts_checked();
705 $thread * thrd = kernelTLS().this_thread;
706 /* paranoid */ verify(thrd->state == Active);
707
708 // SKULLDUGGERY: It is possible that we are preempting this thread just before
709 // it was going to park itself. If that is the case and it is already using the
710 // intrusive fields then we can't use them to preempt the thread
711 // If that is the case, abandon the preemption.
712 bool preempted = false;
713 if(thrd->link.next == 0p) {
714 preempted = true;
715 thrd->preempted = reason;
716 returnToKernel();
717 }
718 __enable_interrupts_checked( false );
719 return preempted;
720}
721
722//=============================================================================================
723// Kernel Idle Sleep
724//=============================================================================================
725// Wake a thread from the front if there are any
726static void __wake_one(cluster * this) {
727 /* paranoid */ verify( ! __preemption_enabled() );
728 /* paranoid */ verify( ready_schedule_islocked() );
729
730 // Check if there is a sleeping processor
731 processor * p;
732 unsigned idle;
733 unsigned total;
734 [idle, total, p] = query_idles(this->procs);
735
736 // If no one is sleeping, we are done
737 if( idle == 0 ) return;
738
739 // We found a processor, wake it up
740 eventfd_t val;
741 val = 1;
742 eventfd_write( p->idle, val );
743
744 #if !defined(__CFA_NO_STATISTICS__)
745 if( kernelTLS().this_stats ) {
746 __tls_stats()->ready.sleep.wakes++;
747 }
748 else {
749 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
750 }
751 #endif
752
753 /* paranoid */ verify( ready_schedule_islocked() );
754 /* paranoid */ verify( ! __preemption_enabled() );
755
756 return;
757}
758
759// Unconditionnaly wake a thread
760void __wake_proc(processor * this) {
761 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
762
763 __disable_interrupts_checked();
764 /* paranoid */ verify( ! __preemption_enabled() );
765 eventfd_t val;
766 val = 1;
767 eventfd_write( this->idle, val );
768 __enable_interrupts_checked();
769}
770
771static void mark_idle(__cluster_proc_list & this, processor & proc) {
772 /* paranoid */ verify( ! __preemption_enabled() );
773 lock( this );
774 this.idle++;
775 /* paranoid */ verify( this.idle <= this.total );
776 remove(proc);
777 insert_first(this.idles, proc);
778 unlock( this );
779 /* paranoid */ verify( ! __preemption_enabled() );
780}
781
782static void mark_awake(__cluster_proc_list & this, processor & proc) {
783 /* paranoid */ verify( ! __preemption_enabled() );
784 lock( this );
785 this.idle--;
786 /* paranoid */ verify( this.idle >= 0 );
787 remove(proc);
788 insert_last(this.actives, proc);
789 unlock( this );
790 /* paranoid */ verify( ! __preemption_enabled() );
791}
792
793static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
794 /* paranoid */ verify( ! __preemption_enabled() );
795 /* paranoid */ verify( ready_schedule_islocked() );
796
797 for() {
798 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
799 if( 1 == (l % 2) ) { Pause(); continue; }
800 unsigned idle = this.idle;
801 unsigned total = this.total;
802 processor * proc = &this.idles`first;
803 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
804 asm volatile("": : :"memory");
805 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
806 return [idle, total, proc];
807 }
808
809 /* paranoid */ verify( ready_schedule_islocked() );
810 /* paranoid */ verify( ! __preemption_enabled() );
811}
812
813//=============================================================================================
814// Unexpected Terminating logic
815//=============================================================================================
816void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
817 $thread * thrd = __cfaabi_tls.this_thread;
818
819 if(thrd) {
820 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
821 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
822
823 if ( &thrd->self_cor != thrd->curr_cor ) {
824 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
825 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
826 }
827 else {
828 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
829 }
830 }
831 else {
832 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
833 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
834 }
835}
836
837int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
838 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
839}
840
841static __spinlock_t kernel_debug_lock;
842
843extern "C" {
844 void __cfaabi_bits_acquire() {
845 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
846 }
847
848 void __cfaabi_bits_release() {
849 unlock( kernel_debug_lock );
850 }
851}
852
853//=============================================================================================
854// Kernel Utilities
855//=============================================================================================
856#if defined(CFA_HAVE_LINUX_IO_URING_H)
857#include "io/types.hfa"
858#endif
859
860static inline bool __maybe_io_drain( processor * proc ) {
861 bool ret = false;
862 #if defined(CFA_HAVE_LINUX_IO_URING_H)
863 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
864
865 // Check if we should drain the queue
866 $io_context * ctx = proc->io.ctx;
867 unsigned head = *ctx->cq.head;
868 unsigned tail = *ctx->cq.tail;
869 if(head == tail) return false;
870 #if OLD_MAIN
871 ready_schedule_lock();
872 ret = __cfa_io_drain( proc );
873 ready_schedule_unlock();
874 #else
875 ret = __cfa_io_drain( proc );
876 #endif
877 #endif
878 return ret;
879}
880
881//-----------------------------------------------------------------------------
882// Debug
883__cfaabi_dbg_debug_do(
884 extern "C" {
885 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
886 this.prev_name = prev_name;
887 this.prev_thrd = kernelTLS().this_thread;
888 }
889 }
890)
891
892//-----------------------------------------------------------------------------
893// Debug
894bool threading_enabled(void) __attribute__((const)) {
895 return true;
896}
897
898//-----------------------------------------------------------------------------
899// Statistics
900#if !defined(__CFA_NO_STATISTICS__)
901 void print_halts( processor & this ) {
902 this.print_halts = true;
903 }
904
905 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
906 /* paranoid */ verify( cltr->stats );
907
908 processor * it = &list`first;
909 for(unsigned i = 0; i < count; i++) {
910 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
911 /* paranoid */ verify( it->local_data->this_stats );
912 __tally_stats( cltr->stats, it->local_data->this_stats );
913 it = &(*it)`next;
914 }
915 }
916
917 void crawl_cluster_stats( cluster & this ) {
918 // Stop the world, otherwise stats could get really messed-up
919 // this doesn't solve all problems but does solve many
920 // so it's probably good enough
921 uint_fast32_t last_size = ready_mutate_lock();
922
923 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
924 crawl_list(&this, this.procs.idles , this.procs.idle );
925
926 // Unlock the RWlock
927 ready_mutate_unlock( last_size );
928 }
929
930
931 void print_stats_now( cluster & this, int flags ) {
932 crawl_cluster_stats( this );
933 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
934 }
935#endif
936// Local Variables: //
937// mode: c //
938// tab-width: 4 //
939// End: //
Note: See TracBrowser for help on using the repository browser.