source: libcfa/src/concurrency/kernel.cfa@ 0a79fd9

ADT ast-experimental enum forall-pointer-decay pthread-emulation qualifiedEnum
Last change on this file since 0a79fd9 was 445f984, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Forgot to commit changes to include.
  • Property mode set to 100644
File size: 30.0 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 _GNU_SOURCE
18
19// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
20
21//C Includes
22#include <errno.h>
23#include <stdio.h>
24#include <string.h>
25#include <signal.h>
26#include <unistd.h>
27extern "C" {
28 #include <sys/eventfd.h>
29}
30
31//CFA Includes
32#include "kernel_private.hfa"
33#include "preemption.hfa"
34#include "strstream.hfa"
35#include "device/cpu.hfa"
36
37//Private includes
38#define __CFA_INVOKE_PRIVATE__
39#include "invoke.h"
40
41#if !defined(__CFA_NO_STATISTICS__)
42 #define __STATS( ...) __VA_ARGS__
43#else
44 #define __STATS( ...)
45#endif
46
47//-----------------------------------------------------------------------------
48// Some assembly required
49#if defined( __i386 )
50 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
51 // fcw : X87 FPU control word (preserved across function calls)
52 #define __x87_store \
53 uint32_t __mxcr; \
54 uint16_t __fcw; \
55 __asm__ volatile ( \
56 "stmxcsr %0\n" \
57 "fnstcw %1\n" \
58 : "=m" (__mxcr),\
59 "=m" (__fcw) \
60 )
61
62 #define __x87_load \
63 __asm__ volatile ( \
64 "fldcw %1\n" \
65 "ldmxcsr %0\n" \
66 ::"m" (__mxcr),\
67 "m" (__fcw) \
68 )
69
70#elif defined( __x86_64 )
71 #define __x87_store \
72 uint32_t __mxcr; \
73 uint16_t __fcw; \
74 __asm__ volatile ( \
75 "stmxcsr %0\n" \
76 "fnstcw %1\n" \
77 : "=m" (__mxcr),\
78 "=m" (__fcw) \
79 )
80
81 #define __x87_load \
82 __asm__ volatile ( \
83 "fldcw %1\n" \
84 "ldmxcsr %0\n" \
85 :: "m" (__mxcr),\
86 "m" (__fcw) \
87 )
88
89#elif defined( __arm__ )
90 #define __x87_store
91 #define __x87_load
92
93#elif defined( __aarch64__ )
94 #define __x87_store \
95 uint32_t __fpcntl[2]; \
96 __asm__ volatile ( \
97 "mrs x9, FPCR\n" \
98 "mrs x10, FPSR\n" \
99 "stp x9, x10, %0\n" \
100 : "=m" (__fpcntl) : : "x9", "x10" \
101 )
102
103 #define __x87_load \
104 __asm__ volatile ( \
105 "ldp x9, x10, %0\n" \
106 "msr FPSR, x10\n" \
107 "msr FPCR, x9\n" \
108 : "=m" (__fpcntl) : : "x9", "x10" \
109 )
110
111#else
112 #error unsupported hardware architecture
113#endif
114
115extern thread$ * mainThread;
116extern processor * mainProcessor;
117
118//-----------------------------------------------------------------------------
119// Kernel Scheduling logic
120static thread$ * __next_thread(cluster * this);
121static thread$ * __next_thread_slow(cluster * this);
122static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
123static void __run_thread(processor * this, thread$ * dst);
124static void __wake_one(cluster * cltr);
125
126static void mark_idle (__cluster_proc_list & idles, processor & proc);
127static void mark_awake(__cluster_proc_list & idles, processor & proc);
128static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
129
130extern void __cfa_io_start( processor * );
131extern bool __cfa_io_drain( processor * );
132extern void __cfa_io_flush( processor * );
133extern void __cfa_io_stop ( processor * );
134static inline bool __maybe_io_drain( processor * );
135
136extern void __disable_interrupts_hard();
137extern void __enable_interrupts_hard();
138
139static inline void __disable_interrupts_checked() {
140 /* paranoid */ verify( __preemption_enabled() );
141 disable_interrupts();
142 /* paranoid */ verify( ! __preemption_enabled() );
143}
144
145static inline void __enable_interrupts_checked( bool poll = true ) {
146 /* paranoid */ verify( ! __preemption_enabled() );
147 enable_interrupts( poll );
148 /* paranoid */ verify( __preemption_enabled() );
149}
150
151//=============================================================================================
152// Kernel Scheduling logic
153//=============================================================================================
154//Main of the processor contexts
155void main(processorCtx_t & runner) {
156 // Because of a bug, we couldn't initialized the seed on construction
157 // Do it here
158 __cfaabi_tls.rand_seed ^= rdtscl();
159 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
160 __tls_rand_advance_bck();
161
162 processor * this = runner.proc;
163 verify(this);
164
165 __cfa_io_start( this );
166
167 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
168 #if !defined(__CFA_NO_STATISTICS__)
169 if( this->print_halts ) {
170 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
171 }
172 #endif
173
174 {
175 // Setup preemption data
176 preemption_scope scope = { this };
177
178 // if we need to run some special setup, now is the time to do it.
179 if(this->init.thrd) {
180 this->init.thrd->curr_cluster = this->cltr;
181 __run_thread(this, this->init.thrd);
182 }
183
184 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
185
186 thread$ * readyThread = 0p;
187 MAIN_LOOP:
188 for() {
189 #define OLD_MAIN 1
190 #if OLD_MAIN
191 // Check if there is pending io
192 __maybe_io_drain( this );
193
194 // Try to get the next thread
195 readyThread = __next_thread( this->cltr );
196
197 if( !readyThread ) {
198 __cfa_io_flush( this );
199 readyThread = __next_thread_slow( this->cltr );
200 }
201
202 HALT:
203 if( !readyThread ) {
204 // Don't block if we are done
205 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
206
207 #if !defined(__CFA_NO_STATISTICS__)
208 __tls_stats()->ready.sleep.halts++;
209 #endif
210
211 // Push self to idle stack
212 mark_idle(this->cltr->procs, * this);
213
214 // Confirm the ready-queue is empty
215 readyThread = __next_thread_slow( this->cltr );
216 if( readyThread ) {
217 // A thread was found, cancel the halt
218 mark_awake(this->cltr->procs, * this);
219
220 #if !defined(__CFA_NO_STATISTICS__)
221 __tls_stats()->ready.sleep.cancels++;
222 #endif
223
224 // continue the mai loop
225 break HALT;
226 }
227
228 #if !defined(__CFA_NO_STATISTICS__)
229 if(this->print_halts) {
230 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
231 }
232 #endif
233
234 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
235
236 {
237 eventfd_t val;
238 ssize_t ret = read( this->idle, &val, sizeof(val) );
239 if(ret < 0) {
240 switch((int)errno) {
241 case EAGAIN:
242 #if EAGAIN != EWOULDBLOCK
243 case EWOULDBLOCK:
244 #endif
245 case EINTR:
246 // No need to do anything special here, just assume it's a legitimate wake-up
247 break;
248 default:
249 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
250 }
251 }
252 }
253
254 #if !defined(__CFA_NO_STATISTICS__)
255 if(this->print_halts) {
256 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
257 }
258 #endif
259
260 // We were woken up, remove self from idle
261 mark_awake(this->cltr->procs, * this);
262
263 // DON'T just proceed, start looking again
264 continue MAIN_LOOP;
265 }
266
267 /* paranoid */ verify( readyThread );
268
269 // Reset io dirty bit
270 this->io.dirty = false;
271
272 // We found a thread run it
273 __run_thread(this, readyThread);
274
275 // Are we done?
276 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
277
278 if(this->io.pending && !this->io.dirty) {
279 __cfa_io_flush( this );
280 }
281
282 #else
283 #warning new kernel loop
284 SEARCH: {
285 /* paranoid */ verify( ! __preemption_enabled() );
286
287 // First, lock the scheduler since we are searching for a thread
288 ready_schedule_lock();
289
290 // Try to get the next thread
291 readyThread = pop_fast( this->cltr );
292 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
293
294 // If we can't find a thread, might as well flush any outstanding I/O
295 if(this->io.pending) { __cfa_io_flush( this ); }
296
297 // Spin a little on I/O, just in case
298 for(5) {
299 __maybe_io_drain( this );
300 readyThread = pop_fast( this->cltr );
301 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
302 }
303
304 // no luck, try stealing a few times
305 for(5) {
306 if( __maybe_io_drain( this ) ) {
307 readyThread = pop_fast( this->cltr );
308 } else {
309 readyThread = pop_slow( this->cltr );
310 }
311 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
312 }
313
314 // still no luck, search for a thread
315 readyThread = pop_search( this->cltr );
316 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
317
318 // Don't block if we are done
319 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
320
321 __STATS( __tls_stats()->ready.sleep.halts++; )
322
323 // Push self to idle stack
324 ready_schedule_unlock();
325 mark_idle(this->cltr->procs, * this);
326 ready_schedule_lock();
327
328 // Confirm the ready-queue is empty
329 __maybe_io_drain( this );
330 readyThread = pop_search( this->cltr );
331 ready_schedule_unlock();
332
333 if( readyThread ) {
334 // A thread was found, cancel the halt
335 mark_awake(this->cltr->procs, * this);
336
337 __STATS( __tls_stats()->ready.sleep.cancels++; )
338
339 // continue the main loop
340 break SEARCH;
341 }
342
343 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
344 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
345
346 // __disable_interrupts_hard();
347 eventfd_t val;
348 eventfd_read( this->idle, &val );
349 // __enable_interrupts_hard();
350
351 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
352
353 // We were woken up, remove self from idle
354 mark_awake(this->cltr->procs, * this);
355
356 // DON'T just proceed, start looking again
357 continue MAIN_LOOP;
358 }
359
360 RUN_THREAD:
361 /* paranoid */ verify( ! __preemption_enabled() );
362 /* paranoid */ verify( readyThread );
363
364 // Reset io dirty bit
365 this->io.dirty = false;
366
367 // We found a thread run it
368 __run_thread(this, readyThread);
369
370 // Are we done?
371 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
372
373 if(this->io.pending && !this->io.dirty) {
374 __cfa_io_flush( this );
375 }
376
377 ready_schedule_lock();
378 __maybe_io_drain( this );
379 ready_schedule_unlock();
380 #endif
381 }
382
383 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
384 }
385
386 __cfa_io_stop( this );
387
388 post( this->terminated );
389
390 if(this == mainProcessor) {
391 // HACK : the coroutine context switch expects this_thread to be set
392 // and it make sense for it to be set in all other cases except here
393 // fake it
394 __cfaabi_tls.this_thread = mainThread;
395 }
396
397 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
398}
399
400static int * __volatile_errno() __attribute__((noinline));
401static int * __volatile_errno() { asm(""); return &errno; }
402
403// KERNEL ONLY
404// runThread runs a thread by context switching
405// from the processor coroutine to the target thread
406static void __run_thread(processor * this, thread$ * thrd_dst) {
407 /* paranoid */ verify( ! __preemption_enabled() );
408 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
409 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
410 __builtin_prefetch( thrd_dst->context.SP );
411
412 int curr = __kernel_getcpu();
413 if(thrd_dst->last_cpu != curr) {
414 int64_t l = thrd_dst->last_cpu;
415 int64_t c = curr;
416 int64_t v = (l << 32) | c;
417 __push_stat( __tls_stats(), v, false, "Processor", this );
418 }
419
420 thrd_dst->last_cpu = curr;
421
422 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
423
424 coroutine$ * proc_cor = get_coroutine(this->runner);
425
426 // set state of processor coroutine to inactive
427 verify(proc_cor->state == Active);
428 proc_cor->state = Blocked;
429
430 // Actually run the thread
431 RUNNING: while(true) {
432 thrd_dst->preempted = __NO_PREEMPTION;
433 thrd_dst->state = Active;
434
435 // Update global state
436 kernelTLS().this_thread = thrd_dst;
437
438 /* paranoid */ verify( ! __preemption_enabled() );
439 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
440 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
441 /* paranoid */ verify( thrd_dst->context.SP );
442 /* paranoid */ verify( thrd_dst->state != Halted );
443 /* 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
444 /* 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
445 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
446
447
448
449 // set context switch to the thread that the processor is executing
450 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
451 // when __cfactx_switch returns we are back in the processor coroutine
452
453
454
455 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
456 /* 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 );
457 /* 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 );
458 /* paranoid */ verify( thrd_dst->context.SP );
459 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
460 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
461 /* paranoid */ verify( ! __preemption_enabled() );
462
463 // Reset global state
464 kernelTLS().this_thread = 0p;
465
466 // We just finished running a thread, there are a few things that could have happened.
467 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
468 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
469 // 4 - Preempted
470 // In case 1, we may have won a race so we can't write to the state again.
471 // In case 2, we lost the race so we now own the thread.
472
473 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
474 // The thread was preempted, reschedule it and reset the flag
475 schedule_thread$( thrd_dst );
476 break RUNNING;
477 }
478
479 if(unlikely(thrd_dst->state == Halting)) {
480 // The thread has halted, it should never be scheduled/run again
481 // finish the thread
482 __thread_finish( thrd_dst );
483 break RUNNING;
484 }
485
486 /* paranoid */ verify( thrd_dst->state == Active );
487 thrd_dst->state = Blocked;
488
489 // set state of processor coroutine to active and the thread to inactive
490 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
491 switch(old_ticket) {
492 case TICKET_RUNNING:
493 // This is case 1, the regular case, nothing more is needed
494 break RUNNING;
495 case TICKET_UNBLOCK:
496 #if !defined(__CFA_NO_STATISTICS__)
497 __tls_stats()->ready.threads.threads++;
498 #endif
499 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
500 // In this case, just run it again.
501 continue RUNNING;
502 default:
503 // This makes no sense, something is wrong abort
504 abort();
505 }
506 }
507
508 // Just before returning to the processor, set the processor coroutine to active
509 proc_cor->state = Active;
510
511 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
512
513 #if !defined(__CFA_NO_STATISTICS__)
514 __tls_stats()->ready.threads.threads--;
515 #endif
516
517 /* paranoid */ verify( ! __preemption_enabled() );
518}
519
520// KERNEL_ONLY
521void returnToKernel() {
522 /* paranoid */ verify( ! __preemption_enabled() );
523 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
524 thread$ * thrd_src = kernelTLS().this_thread;
525
526 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
527
528 // Run the thread on this processor
529 {
530 int local_errno = *__volatile_errno();
531 #if defined( __i386 ) || defined( __x86_64 )
532 __x87_store;
533 #endif
534 /* paranoid */ verify( proc_cor->context.SP );
535 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
536 __cfactx_switch( &thrd_src->context, &proc_cor->context );
537 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
538 #if defined( __i386 ) || defined( __x86_64 )
539 __x87_load;
540 #endif
541 *__volatile_errno() = local_errno;
542 }
543
544 #if !defined(__CFA_NO_STATISTICS__)
545 /* paranoid */ verify( thrd_src->last_proc != 0p );
546 if(thrd_src->last_proc != kernelTLS().this_processor) {
547 __tls_stats()->ready.threads.migration++;
548 }
549 #endif
550
551 /* paranoid */ verify( ! __preemption_enabled() );
552 /* 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 );
553 /* 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 );
554}
555
556//-----------------------------------------------------------------------------
557// Scheduler routines
558// KERNEL ONLY
559static void __schedule_thread( thread$ * thrd ) {
560 /* paranoid */ verify( ! __preemption_enabled() );
561 /* paranoid */ verify( ready_schedule_islocked());
562 /* paranoid */ verify( thrd );
563 /* paranoid */ verify( thrd->state != Halted );
564 /* paranoid */ verify( thrd->curr_cluster );
565 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
566 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
567 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
568 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
569 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
570 /* paranoid */ #endif
571 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
572 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
573
574 const bool local = thrd->state != Start;
575 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
576
577 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
578 struct cluster * cl = thrd->curr_cluster;
579 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
580
581 // push the thread to the cluster ready-queue
582 push( cl, thrd, local );
583
584 // variable thrd is no longer safe to use
585 thrd = 0xdeaddeaddeaddeadp;
586
587 // wake the cluster using the save variable.
588 __wake_one( cl );
589
590 #if !defined(__CFA_NO_STATISTICS__)
591 if( kernelTLS().this_stats ) {
592 __tls_stats()->ready.threads.threads++;
593 if(outside) {
594 __tls_stats()->ready.threads.extunpark++;
595 }
596 }
597 else {
598 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
599 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
600 }
601 #endif
602
603 /* paranoid */ verify( ready_schedule_islocked());
604 /* paranoid */ verify( ! __preemption_enabled() );
605}
606
607void schedule_thread$( thread$ * thrd ) {
608 ready_schedule_lock();
609 __schedule_thread( thrd );
610 ready_schedule_unlock();
611}
612
613// KERNEL ONLY
614static inline thread$ * __next_thread(cluster * this) with( *this ) {
615 /* paranoid */ verify( ! __preemption_enabled() );
616
617 ready_schedule_lock();
618 thread$ * thrd = pop_fast( this );
619 ready_schedule_unlock();
620
621 /* paranoid */ verify( ! __preemption_enabled() );
622 return thrd;
623}
624
625// KERNEL ONLY
626static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
627 /* paranoid */ verify( ! __preemption_enabled() );
628
629 ready_schedule_lock();
630 thread$ * thrd;
631 for(25) {
632 thrd = pop_slow( this );
633 if(thrd) goto RET;
634 }
635 thrd = pop_search( this );
636
637 RET:
638 ready_schedule_unlock();
639
640 /* paranoid */ verify( ! __preemption_enabled() );
641 return thrd;
642}
643
644static inline bool __must_unpark( thread$ * thrd ) {
645 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
646 switch(old_ticket) {
647 case TICKET_RUNNING:
648 // Wake won the race, the thread will reschedule/rerun itself
649 return false;
650 case TICKET_BLOCKED:
651 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
652 /* paranoid */ verify( thrd->state == Blocked );
653 return true;
654 default:
655 // This makes no sense, something is wrong abort
656 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
657 }
658}
659
660void __kernel_unpark( thread$ * thrd ) {
661 /* paranoid */ verify( ! __preemption_enabled() );
662 /* paranoid */ verify( ready_schedule_islocked());
663
664 if( !thrd ) return;
665
666 if(__must_unpark(thrd)) {
667 // Wake lost the race,
668 __schedule_thread( thrd );
669 }
670
671 /* paranoid */ verify( ready_schedule_islocked());
672 /* paranoid */ verify( ! __preemption_enabled() );
673}
674
675void unpark( thread$ * thrd ) {
676 if( !thrd ) return;
677
678 if(__must_unpark(thrd)) {
679 disable_interrupts();
680 // Wake lost the race,
681 schedule_thread$( thrd );
682 enable_interrupts(false);
683 }
684}
685
686void park( void ) {
687 __disable_interrupts_checked();
688 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
689 returnToKernel();
690 __enable_interrupts_checked();
691
692}
693
694extern "C" {
695 // Leave the thread monitor
696 // last routine called by a thread.
697 // Should never return
698 void __cfactx_thrd_leave() {
699 thread$ * thrd = active_thread();
700 monitor$ * this = &thrd->self_mon;
701
702 // Lock the monitor now
703 lock( this->lock __cfaabi_dbg_ctx2 );
704
705 disable_interrupts();
706
707 /* paranoid */ verify( ! __preemption_enabled() );
708 /* paranoid */ verify( thrd->state == Active );
709 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
710 /* paranoid */ verify( kernelTLS().this_thread == thrd );
711 /* paranoid */ verify( thrd->context.SP );
712 /* 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 );
713 /* 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 );
714
715 thrd->state = Halting;
716 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
717 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
718 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
719
720 // Leave the thread
721 returnToKernel();
722
723 // Control flow should never reach here!
724 abort();
725 }
726}
727
728// KERNEL ONLY
729bool force_yield( __Preemption_Reason reason ) {
730 __disable_interrupts_checked();
731 thread$ * thrd = kernelTLS().this_thread;
732 /* paranoid */ verify(thrd->state == Active);
733
734 // SKULLDUGGERY: It is possible that we are preempting this thread just before
735 // it was going to park itself. If that is the case and it is already using the
736 // intrusive fields then we can't use them to preempt the thread
737 // If that is the case, abandon the preemption.
738 bool preempted = false;
739 if(thrd->link.next == 0p) {
740 preempted = true;
741 thrd->preempted = reason;
742 returnToKernel();
743 }
744 __enable_interrupts_checked( false );
745 return preempted;
746}
747
748//=============================================================================================
749// Kernel Idle Sleep
750//=============================================================================================
751// Wake a thread from the front if there are any
752static void __wake_one(cluster * this) {
753 /* paranoid */ verify( ! __preemption_enabled() );
754 /* paranoid */ verify( ready_schedule_islocked() );
755
756 // Check if there is a sleeping processor
757 processor * p;
758 unsigned idle;
759 unsigned total;
760 [idle, total, p] = query_idles(this->procs);
761
762 // If no one is sleeping, we are done
763 if( idle == 0 ) return;
764
765 // We found a processor, wake it up
766 eventfd_t val;
767 val = 1;
768 eventfd_write( p->idle, val );
769
770 #if !defined(__CFA_NO_STATISTICS__)
771 if( kernelTLS().this_stats ) {
772 __tls_stats()->ready.sleep.wakes++;
773 }
774 else {
775 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
776 }
777 #endif
778
779 /* paranoid */ verify( ready_schedule_islocked() );
780 /* paranoid */ verify( ! __preemption_enabled() );
781
782 return;
783}
784
785// Unconditionnaly wake a thread
786void __wake_proc(processor * this) {
787 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
788
789 __disable_interrupts_checked();
790 /* paranoid */ verify( ! __preemption_enabled() );
791 eventfd_t val;
792 val = 1;
793 eventfd_write( this->idle, val );
794 __enable_interrupts_checked();
795}
796
797static void mark_idle(__cluster_proc_list & this, processor & proc) {
798 /* paranoid */ verify( ! __preemption_enabled() );
799 lock( this );
800 this.idle++;
801 /* paranoid */ verify( this.idle <= this.total );
802 remove(proc);
803 insert_first(this.idles, proc);
804 unlock( this );
805 /* paranoid */ verify( ! __preemption_enabled() );
806}
807
808static void mark_awake(__cluster_proc_list & this, processor & proc) {
809 /* paranoid */ verify( ! __preemption_enabled() );
810 lock( this );
811 this.idle--;
812 /* paranoid */ verify( this.idle >= 0 );
813 remove(proc);
814 insert_last(this.actives, proc);
815 unlock( this );
816 /* paranoid */ verify( ! __preemption_enabled() );
817}
818
819static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
820 /* paranoid */ verify( ! __preemption_enabled() );
821 /* paranoid */ verify( ready_schedule_islocked() );
822
823 for() {
824 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
825 if( 1 == (l % 2) ) { Pause(); continue; }
826 unsigned idle = this.idle;
827 unsigned total = this.total;
828 processor * proc = &this.idles`first;
829 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
830 asm volatile("": : :"memory");
831 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
832 return [idle, total, proc];
833 }
834
835 /* paranoid */ verify( ready_schedule_islocked() );
836 /* paranoid */ verify( ! __preemption_enabled() );
837}
838
839//=============================================================================================
840// Unexpected Terminating logic
841//=============================================================================================
842void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
843 thread$ * thrd = __cfaabi_tls.this_thread;
844
845 if(thrd) {
846 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
847 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
848
849 if ( &thrd->self_cor != thrd->curr_cor ) {
850 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
851 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
852 }
853 else {
854 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
855 }
856 }
857 else {
858 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
859 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
860 }
861}
862
863int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
864 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
865}
866
867static __spinlock_t kernel_debug_lock;
868
869extern "C" {
870 void __cfaabi_bits_acquire() {
871 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
872 }
873
874 void __cfaabi_bits_release() {
875 unlock( kernel_debug_lock );
876 }
877}
878
879//=============================================================================================
880// Kernel Utilities
881//=============================================================================================
882#if defined(CFA_HAVE_LINUX_IO_URING_H)
883#include "io/types.hfa"
884#endif
885
886static inline bool __maybe_io_drain( processor * proc ) {
887 bool ret = false;
888 #if defined(CFA_HAVE_LINUX_IO_URING_H)
889 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
890
891 // Check if we should drain the queue
892 $io_context * ctx = proc->io.ctx;
893 unsigned head = *ctx->cq.head;
894 unsigned tail = *ctx->cq.tail;
895 if(head == tail) return false;
896 #if OLD_MAIN
897 ready_schedule_lock();
898 ret = __cfa_io_drain( proc );
899 ready_schedule_unlock();
900 #else
901 ret = __cfa_io_drain( proc );
902 #endif
903 #endif
904 return ret;
905}
906
907//-----------------------------------------------------------------------------
908// Debug
909__cfaabi_dbg_debug_do(
910 extern "C" {
911 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
912 this.prev_name = prev_name;
913 this.prev_thrd = kernelTLS().this_thread;
914 }
915 }
916)
917
918//-----------------------------------------------------------------------------
919// Debug
920bool threading_enabled(void) __attribute__((const)) {
921 return true;
922}
923
924//-----------------------------------------------------------------------------
925// Statistics
926#if !defined(__CFA_NO_STATISTICS__)
927 void print_halts( processor & this ) {
928 this.print_halts = true;
929 }
930
931 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
932 /* paranoid */ verify( cltr->stats );
933
934 processor * it = &list`first;
935 for(unsigned i = 0; i < count; i++) {
936 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
937 /* paranoid */ verify( it->local_data->this_stats );
938 __tally_stats( cltr->stats, it->local_data->this_stats );
939 it = &(*it)`next;
940 }
941 }
942
943 void crawl_cluster_stats( cluster & this ) {
944 // Stop the world, otherwise stats could get really messed-up
945 // this doesn't solve all problems but does solve many
946 // so it's probably good enough
947 uint_fast32_t last_size = ready_mutate_lock();
948
949 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
950 crawl_list(&this, this.procs.idles , this.procs.idle );
951
952 // Unlock the RWlock
953 ready_mutate_unlock( last_size );
954 }
955
956
957 void print_stats_now( cluster & this, int flags ) {
958 crawl_cluster_stats( this );
959 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
960 }
961#endif
962// Local Variables: //
963// mode: c //
964// tab-width: 4 //
965// End: //
Note: See TracBrowser for help on using the repository browser.