Context Navigation

kernel.cfa @ 8464edf

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum

Last change on this file since 8464edf was 8464edf, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
'print_stats_now' now pulls stats from processors.
Property mode set to `100644`
File size: 29.7 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// kernel.c --
8	//
9	// Author : Thierry Delisle
10	// Created On : Tue Jan 17 12:27:26 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 31 07:08:20 2020
13	// Update Count : 71
14	//
15
16	#define __cforall_thread__
17	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
18
19	//C Includes
20	#include <errno.h>
21	#include <stdio.h>
22	#include <signal.h>
23	#include <unistd.h>
24	extern "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
110	extern $thread * mainThread;
111	extern processor * mainProcessor;
112
113	//-----------------------------------------------------------------------------
114	// Kernel Scheduling logic
115	static $thread * __next_thread(cluster * this);
116	static $thread * __next_thread_slow(cluster * this);
117	static inline bool __must_unpark( $thread * thrd ) __attribute((nonnull(1)));
118	static void __run_thread(processor * this, $thread * dst);
119	static void __wake_one(cluster * cltr);
120
121	static void mark_idle (__cluster_proc_list & idles, processor & proc);
122	static void mark_awake(__cluster_proc_list & idles, processor & proc);
123	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
124
125	extern void __cfa_io_start( processor * );
126	extern bool __cfa_io_drain( processor * );
127	extern void __cfa_io_flush( processor * );
128	extern void __cfa_io_stop ( processor * );
129	static inline bool __maybe_io_drain( processor * );
130
131	extern void __disable_interrupts_hard();
132	extern void __enable_interrupts_hard();
133
134	static inline void __disable_interrupts_checked() {
135	/* paranoid */ verify( __preemption_enabled() );
136	disable_interrupts();
137	/* paranoid */ verify( ! __preemption_enabled() );
138	}
139
140	static 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
150	void 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	#if 1
185	// Check if there is pending io
186	__maybe_io_drain( this );
187
188	// Try to get the next thread
189	readyThread = __next_thread( this->cltr );
190
191	if( !readyThread ) {
192	__cfa_io_flush( this );
193	readyThread = __next_thread_slow( this->cltr );
194	}
195
196	HALT:
197	if( !readyThread ) {
198	// Don't block if we are done
199	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
200
201	#if !defined(__CFA_NO_STATISTICS__)
202	__tls_stats()->ready.sleep.halts++;
203	#endif
204
205	// Push self to idle stack
206	mark_idle(this->cltr->procs, * this);
207
208	// Confirm the ready-queue is empty
209	readyThread = __next_thread_slow( this->cltr );
210	if( readyThread ) {
211	// A thread was found, cancel the halt
212	mark_awake(this->cltr->procs, * this);
213
214	#if !defined(__CFA_NO_STATISTICS__)
215	__tls_stats()->ready.sleep.cancels++;
216	#endif
217
218	// continue the mai loop
219	break HALT;
220	}
221
222	#if !defined(__CFA_NO_STATISTICS__)
223	if(this->print_halts) {
224	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
225	}
226	#endif
227
228	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
229
230	__disable_interrupts_hard();
231	eventfd_t val;
232	eventfd_read( this->idle, &val );
233	__enable_interrupts_hard();
234
235	#if !defined(__CFA_NO_STATISTICS__)
236	if(this->print_halts) {
237	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
238	}
239	#endif
240
241	// We were woken up, remove self from idle
242	mark_awake(this->cltr->procs, * this);
243
244	// DON'T just proceed, start looking again
245	continue MAIN_LOOP;
246	}
247
248	/* paranoid */ verify( readyThread );
249
250	// Reset io dirty bit
251	this->io.dirty = false;
252
253	// We found a thread run it
254	__run_thread(this, readyThread);
255
256	// Are we done?
257	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
258
259	if(this->io.pending && !this->io.dirty) {
260	__cfa_io_flush( this );
261	}
262
263	#else
264
265	SEARCH: {
266	/* paranoid */ verify( ! __preemption_enabled() );
267	/* paranoid */ verify( kernelTLS().this_proc_id );
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(25) {
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(25) {
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->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->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( kernelTLS().this_proc_id );
344	/* paranoid */ verify( ! __preemption_enabled() );
345	/* paranoid */ verify( readyThread );
346
347	// Reset io dirty bit
348	this->io.dirty = false;
349
350	// We found a thread run it
351	__run_thread(this, readyThread);
352
353	// Are we done?
354	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
355
356	#if !defined(__CFA_NO_STATISTICS__)
357	unsigned long long curr = rdtscl();
358	if(curr > (last_tally + 500000000)) {
359	__tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
360	last_tally = curr;
361	}
362	#endif
363
364	if(this->io.pending && !this->io.dirty) {
365	__cfa_io_flush( this );
366	}
367
368	ready_schedule_lock();
369	__maybe_io_drain( this );
370	ready_schedule_unlock();
371	#endif
372	}
373
374	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
375	}
376
377	__cfa_io_stop( this );
378
379	post( this->terminated );
380
381	if(this == mainProcessor) {
382	// HACK : the coroutine context switch expects this_thread to be set
383	// and it make sense for it to be set in all other cases except here
384	// fake it
385	__cfaabi_tls.this_thread = mainThread;
386	}
387
388	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
389	}
390
391	static int * __volatile_errno() __attribute__((noinline));
392	static int * __volatile_errno() { asm(""); return &errno; }
393
394	// KERNEL ONLY
395	// runThread runs a thread by context switching
396	// from the processor coroutine to the target thread
397	static void __run_thread(processor * this, $thread * thrd_dst) {
398	/* paranoid */ verify( ! __preemption_enabled() );
399	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
400	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
401	__builtin_prefetch( thrd_dst->context.SP );
402
403	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
404
405	$coroutine * proc_cor = get_coroutine(this->runner);
406
407	// set state of processor coroutine to inactive
408	verify(proc_cor->state == Active);
409	proc_cor->state = Blocked;
410
411	// Actually run the thread
412	RUNNING: while(true) {
413	thrd_dst->preempted = __NO_PREEMPTION;
414	thrd_dst->state = Active;
415
416	// Update global state
417	kernelTLS().this_thread = thrd_dst;
418
419	/* paranoid */ verify( ! __preemption_enabled() );
420	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
421	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
422	/* paranoid */ verify( thrd_dst->context.SP );
423	/* paranoid */ verify( thrd_dst->state != Halted );
424	/* 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
425	/* 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
426	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
427
428
429
430	// set context switch to the thread that the processor is executing
431	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
432	// when __cfactx_switch returns we are back in the processor coroutine
433
434	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
435	/* 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 );
436	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
437	/* paranoid */ verify( thrd_dst->context.SP );
438	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
439	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
440	/* paranoid */ verify( ! __preemption_enabled() );
441
442	// Reset global state
443	kernelTLS().this_thread = 0p;
444
445	// We just finished running a thread, there are a few things that could have happened.
446	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
447	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
448	// 4 - Preempted
449	// In case 1, we may have won a race so we can't write to the state again.
450	// In case 2, we lost the race so we now own the thread.
451
452	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
453	// The thread was preempted, reschedule it and reset the flag
454	schedule_thread$( thrd_dst );
455	break RUNNING;
456	}
457
458	if(unlikely(thrd_dst->state == Halting)) {
459	// The thread has halted, it should never be scheduled/run again
460	// finish the thread
461	__thread_finish( thrd_dst );
462	break RUNNING;
463	}
464
465	/* paranoid */ verify( thrd_dst->state == Active );
466	thrd_dst->state = Blocked;
467
468	// set state of processor coroutine to active and the thread to inactive
469	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
470	switch(old_ticket) {
471	case TICKET_RUNNING:
472	// This is case 1, the regular case, nothing more is needed
473	break RUNNING;
474	case TICKET_UNBLOCK:
475	#if !defined(__CFA_NO_STATISTICS__)
476	__tls_stats()->ready.threads.threads++;
477	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
478	#endif
479	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
480	// In this case, just run it again.
481	continue RUNNING;
482	default:
483	// This makes no sense, something is wrong abort
484	abort();
485	}
486	}
487
488	// Just before returning to the processor, set the processor coroutine to active
489	proc_cor->state = Active;
490
491	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
492
493	#if !defined(__CFA_NO_STATISTICS__)
494	__tls_stats()->ready.threads.threads--;
495	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
496	#endif
497
498	/* paranoid */ verify( ! __preemption_enabled() );
499	}
500
501	// KERNEL_ONLY
502	void returnToKernel() {
503	/* paranoid */ verify( ! __preemption_enabled() );
504	$coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
505	$thread * thrd_src = kernelTLS().this_thread;
506
507	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
508
509	// Run the thread on this processor
510	{
511	int local_errno = *__volatile_errno();
512	#if defined( __i386 ) \|\| defined( __x86_64 )
513	__x87_store;
514	#endif
515	/* paranoid */ verify( proc_cor->context.SP );
516	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
517	__cfactx_switch( &thrd_src->context, &proc_cor->context );
518	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
519	#if defined( __i386 ) \|\| defined( __x86_64 )
520	__x87_load;
521	#endif
522	*__volatile_errno() = local_errno;
523	}
524
525	#if !defined(__CFA_NO_STATISTICS__)
526	/* paranoid */ verify( thrd_src->last_proc != 0p );
527	if(thrd_src->last_proc != kernelTLS().this_processor) {
528	__tls_stats()->ready.threads.migration++;
529	}
530	#endif
531
532	/* paranoid */ verify( ! __preemption_enabled() );
533	/* 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 );
534	/* 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 );
535	}
536
537	//-----------------------------------------------------------------------------
538	// Scheduler routines
539	// KERNEL ONLY
540	static void __schedule_thread( $thread * thrd ) {
541	/* paranoid */ verify( ! __preemption_enabled() );
542	/* paranoid */ verify( ready_schedule_islocked());
543	/* paranoid */ verify( thrd );
544	/* paranoid */ verify( thrd->state != Halted );
545	/* paranoid */ verify( thrd->curr_cluster );
546	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
547	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
548	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
549	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
550	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
551	/* paranoid */ #endif
552	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
553	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
554
555
556	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
557
558	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
559	struct cluster * cl = thrd->curr_cluster;
560	__STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
561
562	// push the thread to the cluster ready-queue
563	push( cl, thrd );
564
565	// variable thrd is no longer safe to use
566	thrd = 0xdeaddeaddeaddeadp;
567
568	// wake the cluster using the save variable.
569	__wake_one( cl );
570
571	#if !defined(__CFA_NO_STATISTICS__)
572	if( kernelTLS().this_stats ) {
573	__tls_stats()->ready.threads.threads++;
574	if(outside) {
575	__tls_stats()->ready.threads.extunpark++;
576	}
577	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
578	}
579	else {
580	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
581	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
582	__push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
583	}
584	#endif
585
586	/* paranoid */ verify( ready_schedule_islocked());
587	/* paranoid */ verify( ! __preemption_enabled() );
588	}
589
590	void schedule_thread$( $thread * thrd ) {
591	ready_schedule_lock();
592	__schedule_thread( thrd );
593	ready_schedule_unlock();
594	}
595
596	// KERNEL ONLY
597	static inline $thread * __next_thread(cluster * this) with( *this ) {
598	/* paranoid */ verify( ! __preemption_enabled() );
599
600	ready_schedule_lock();
601	$thread * thrd = pop_fast( this );
602	ready_schedule_unlock();
603
604	/* paranoid */ verify( ! __preemption_enabled() );
605	return thrd;
606	}
607
608	// KERNEL ONLY
609	static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
610	/* paranoid */ verify( ! __preemption_enabled() );
611
612	ready_schedule_lock();
613	$thread * thrd;
614	for(25) {
615	thrd = pop_slow( this );
616	if(thrd) goto RET;
617	}
618	thrd = pop_search( this );
619
620	RET:
621	ready_schedule_unlock();
622
623	/* paranoid */ verify( ! __preemption_enabled() );
624	return thrd;
625	}
626
627	static inline bool __must_unpark( $thread * thrd ) {
628	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
629	switch(old_ticket) {
630	case TICKET_RUNNING:
631	// Wake won the race, the thread will reschedule/rerun itself
632	return false;
633	case TICKET_BLOCKED:
634	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
635	/* paranoid */ verify( thrd->state == Blocked );
636	return true;
637	default:
638	// This makes no sense, something is wrong abort
639	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
640	}
641	}
642
643	void __kernel_unpark( $thread * thrd ) {
644	/* paranoid */ verify( ! __preemption_enabled() );
645	/* paranoid */ verify( ready_schedule_islocked());
646
647	if( !thrd ) return;
648
649	if(__must_unpark(thrd)) {
650	// Wake lost the race,
651	__schedule_thread( thrd );
652	}
653
654	/* paranoid */ verify( ready_schedule_islocked());
655	/* paranoid */ verify( ! __preemption_enabled() );
656	}
657
658	void unpark( $thread * thrd ) {
659	if( !thrd ) return;
660
661	if(__must_unpark(thrd)) {
662	disable_interrupts();
663	// Wake lost the race,
664	schedule_thread$( thrd );
665	enable_interrupts(false);
666	}
667	}
668
669	void park( void ) {
670	__disable_interrupts_checked();
671	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
672	returnToKernel();
673	__enable_interrupts_checked();
674
675	}
676
677	extern "C" {
678	// Leave the thread monitor
679	// last routine called by a thread.
680	// Should never return
681	void __cfactx_thrd_leave() {
682	$thread * thrd = active_thread();
683	$monitor * this = &thrd->self_mon;
684
685	// Lock the monitor now
686	lock( this->lock __cfaabi_dbg_ctx2 );
687
688	disable_interrupts();
689
690	/* paranoid */ verify( ! __preemption_enabled() );
691	/* paranoid */ verify( thrd->state == Active );
692	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
693	/* paranoid */ verify( kernelTLS().this_thread == thrd );
694	/* paranoid */ verify( thrd->context.SP );
695	/* 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 );
696	/* 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 );
697
698	thrd->state = Halting;
699	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
700	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
701	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
702
703	// Leave the thread
704	returnToKernel();
705
706	// Control flow should never reach here!
707	abort();
708	}
709	}
710
711	// KERNEL ONLY
712	bool force_yield( __Preemption_Reason reason ) {
713	__disable_interrupts_checked();
714	$thread * thrd = kernelTLS().this_thread;
715	/* paranoid */ verify(thrd->state == Active);
716
717	// SKULLDUGGERY: It is possible that we are preempting this thread just before
718	// it was going to park itself. If that is the case and it is already using the
719	// intrusive fields then we can't use them to preempt the thread
720	// If that is the case, abandon the preemption.
721	bool preempted = false;
722	if(thrd->link.next == 0p) {
723	preempted = true;
724	thrd->preempted = reason;
725	returnToKernel();
726	}
727	__enable_interrupts_checked( false );
728	return preempted;
729	}
730
731	//=============================================================================================
732	// Kernel Idle Sleep
733	//=============================================================================================
734	// Wake a thread from the front if there are any
735	static void __wake_one(cluster * this) {
736	/* paranoid */ verify( ! __preemption_enabled() );
737	/* paranoid */ verify( ready_schedule_islocked() );
738
739	// Check if there is a sleeping processor
740	processor * p;
741	unsigned idle;
742	unsigned total;
743	[idle, total, p] = query_idles(this->procs);
744
745	// If no one is sleeping, we are done
746	if( idle == 0 ) return;
747
748	// We found a processor, wake it up
749	eventfd_t val;
750	val = 1;
751	eventfd_write( p->idle, val );
752
753	#if !defined(__CFA_NO_STATISTICS__)
754	if( kernelTLS().this_stats ) {
755	__tls_stats()->ready.sleep.wakes++;
756	}
757	else {
758	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
759	}
760	#endif
761
762	/* paranoid */ verify( ready_schedule_islocked() );
763	/* paranoid */ verify( ! __preemption_enabled() );
764
765	return;
766	}
767
768	// Unconditionnaly wake a thread
769	void __wake_proc(processor * this) {
770	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
771
772	__disable_interrupts_checked();
773	/* paranoid */ verify( ! __preemption_enabled() );
774	eventfd_t val;
775	val = 1;
776	eventfd_write( this->idle, val );
777	__enable_interrupts_checked();
778	}
779
780	static void mark_idle(__cluster_proc_list & this, processor & proc) {
781	/* paranoid */ verify( ! __preemption_enabled() );
782	lock( this );
783	this.idle++;
784	/* paranoid */ verify( this.idle <= this.total );
785	remove(proc);
786	insert_first(this.idles, proc);
787	unlock( this );
788	/* paranoid */ verify( ! __preemption_enabled() );
789	}
790
791	static void mark_awake(__cluster_proc_list & this, processor & proc) {
792	/* paranoid */ verify( ! __preemption_enabled() );
793	lock( this );
794	this.idle--;
795	/* paranoid */ verify( this.idle >= 0 );
796	remove(proc);
797	insert_last(this.actives, proc);
798	unlock( this );
799	/* paranoid */ verify( ! __preemption_enabled() );
800	}
801
802	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
803	/* paranoid */ verify( ! __preemption_enabled() );
804	/* paranoid */ verify( ready_schedule_islocked() );
805
806	for() {
807	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
808	if( 1 == (l % 2) ) { Pause(); continue; }
809	unsigned idle = this.idle;
810	unsigned total = this.total;
811	processor * proc = &this.idles`first;
812	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
813	asm volatile("": : :"memory");
814	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
815	return [idle, total, proc];
816	}
817
818	/* paranoid */ verify( ready_schedule_islocked() );
819	/* paranoid */ verify( ! __preemption_enabled() );
820	}
821
822	//=============================================================================================
823	// Unexpected Terminating logic
824	//=============================================================================================
825	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
826	$thread * thrd = __cfaabi_tls.this_thread;
827
828	if(thrd) {
829	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
830	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
831
832	if ( &thrd->self_cor != thrd->curr_cor ) {
833	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
834	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
835	}
836	else {
837	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
838	}
839	}
840	else {
841	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
842	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
843	}
844	}
845
846	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
847	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
848	}
849
850	static __spinlock_t kernel_debug_lock;
851
852	extern "C" {
853	void __cfaabi_bits_acquire() {
854	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
855	}
856
857	void __cfaabi_bits_release() {
858	unlock( kernel_debug_lock );
859	}
860	}
861
862	//=============================================================================================
863	// Kernel Utilities
864	//=============================================================================================
865	#if defined(CFA_HAVE_LINUX_IO_URING_H)
866	#include "io/types.hfa"
867	#endif
868
869	static inline bool __maybe_io_drain( processor * proc ) {
870	bool ret = false;
871	#if defined(CFA_HAVE_LINUX_IO_URING_H)
872	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
873
874	// Check if we should drain the queue
875	$io_context * ctx = proc->io.ctx;
876	unsigned head = *ctx->cq.head;
877	unsigned tail = *ctx->cq.tail;
878	if(head == tail) return false;
879	ready_schedule_lock();
880	ret = __cfa_io_drain( proc );
881	ready_schedule_unlock();
882	#endif
883	return ret;
884	}
885
886	//-----------------------------------------------------------------------------
887	// Debug
888	__cfaabi_dbg_debug_do(
889	extern "C" {
890	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
891	this.prev_name = prev_name;
892	this.prev_thrd = kernelTLS().this_thread;
893	}
894	}
895	)
896
897	//-----------------------------------------------------------------------------
898	// Debug
899	bool threading_enabled(void) __attribute__((const)) {
900	return true;
901	}
902
903	//-----------------------------------------------------------------------------
904	// Statistics
905	#if !defined(__CFA_NO_STATISTICS__)
906	void print_halts( processor & this ) {
907	this.print_halts = true;
908	}
909
910	static void crawl_list( cluster * cltr, dlist(processor, processor) & list, unsigned count ) {
911	/* paranoid */ verify( cltr->stats );
912
913	processor * it = &list`first;
914	for(unsigned i = 0; i < count; i++) {
915	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
916	/* paranoid */ verify( it->local_data->this_stats );
917	__tally_stats( cltr->stats, it->local_data->this_stats );
918	it = &(*it)`next;
919	}
920	}
921
922	void crawl_cluster_stats( cluster & this ) {
923	// Stop the world, otherwise stats could get really messed-up
924	// this doesn't solve all problems but does solve many
925	// so it's probably good enough
926	uint_fast32_t last_size = ready_mutate_lock();
927
928	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
929	crawl_list(&this, this.procs.idles , this.procs.idle );
930
931	// Unlock the RWlock
932	ready_mutate_unlock( last_size );
933	}
934
935
936	void print_stats_now( cluster & this, int flags ) {
937	crawl_cluster_stats( this );
938	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
939	}
940	#endif
941	// Local Variables: //
942	// mode: c //
943	// tab-width: 4 //
944	// End: //

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format