Context Navigation

kernel.cfa @ 813dfd86

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since 813dfd86 was a633f6f, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Fix the new FD change.
Property mode set to `100644`
File size: 29.9 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>
27	extern "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
115	extern thread$ * mainThread;
116	extern processor * mainProcessor;
117
118	//-----------------------------------------------------------------------------
119	// Kernel Scheduling logic
120	static thread$ * __next_thread(cluster * this);
121	static thread$ * __next_thread_slow(cluster * this);
122	static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
123	static void __run_thread(processor * this, thread$ * dst);
124	static void __wake_one(cluster * cltr);
125
126	static bool mark_idle (__cluster_proc_list & idles, processor & proc);
127	static void mark_awake(__cluster_proc_list & idles, processor & proc);
128
129	extern void __cfa_io_start( processor * );
130	extern bool __cfa_io_drain( processor * );
131	extern void __cfa_io_flush( processor * );
132	extern void __cfa_io_stop ( processor * );
133	static inline bool __maybe_io_drain( processor * );
134
135	extern void __disable_interrupts_hard();
136	extern void __enable_interrupts_hard();
137
138	static inline void __disable_interrupts_checked() {
139	/* paranoid */ verify( __preemption_enabled() );
140	disable_interrupts();
141	/* paranoid */ verify( ! __preemption_enabled() );
142	}
143
144	static inline void __enable_interrupts_checked( bool poll = true ) {
145	/* paranoid */ verify( ! __preemption_enabled() );
146	enable_interrupts( poll );
147	/* paranoid */ verify( __preemption_enabled() );
148	}
149
150	//=============================================================================================
151	// Kernel Scheduling logic
152	//=============================================================================================
153	//Main of the processor contexts
154	void main(processorCtx_t & runner) {
155	// Because of a bug, we couldn't initialized the seed on construction
156	// Do it here
157	__cfaabi_tls.rand_seed ^= rdtscl();
158	__cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
159	__tls_rand_advance_bck();
160
161	processor * this = runner.proc;
162	verify(this);
163
164	__cfa_io_start( this );
165
166	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
167	#if !defined(__CFA_NO_STATISTICS__)
168	if( this->print_halts ) {
169	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
170	}
171	#endif
172
173	{
174	// Setup preemption data
175	preemption_scope scope = { this };
176
177	// if we need to run some special setup, now is the time to do it.
178	if(this->init.thrd) {
179	this->init.thrd->curr_cluster = this->cltr;
180	__run_thread(this, this->init.thrd);
181	}
182
183	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
184
185	thread$ * readyThread = 0p;
186	MAIN_LOOP:
187	for() {
188	#define OLD_MAIN 1
189	#if OLD_MAIN
190	// Check if there is pending io
191	__maybe_io_drain( this );
192
193	// Try to get the next thread
194	readyThread = __next_thread( this->cltr );
195
196	if( !readyThread ) {
197	ready_schedule_lock();
198	__cfa_io_flush( this );
199	ready_schedule_unlock();
200
201	readyThread = __next_thread_slow( this->cltr );
202	}
203
204	HALT:
205	if( !readyThread ) {
206	// Don't block if we are done
207	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
208
209	#if !defined(__CFA_NO_STATISTICS__)
210	__tls_stats()->ready.sleep.halts++;
211	#endif
212
213	// Push self to idle stack
214	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
215
216	// Confirm the ready-queue is empty
217	readyThread = __next_thread_slow( this->cltr );
218	if( readyThread ) {
219	// A thread was found, cancel the halt
220	mark_awake(this->cltr->procs, * this);
221
222	#if !defined(__CFA_NO_STATISTICS__)
223	__tls_stats()->ready.sleep.cancels++;
224	#endif
225
226	// continue the mai loop
227	break HALT;
228	}
229
230	#if !defined(__CFA_NO_STATISTICS__)
231	if(this->print_halts) {
232	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
233	}
234	#endif
235
236	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
237
238	{
239	eventfd_t val;
240	ssize_t ret = read( this->idle, &val, sizeof(val) );
241	if(ret < 0) {
242	switch((int)errno) {
243	case EAGAIN:
244	#if EAGAIN != EWOULDBLOCK
245	case EWOULDBLOCK:
246	#endif
247	case EINTR:
248	// No need to do anything special here, just assume it's a legitimate wake-up
249	break;
250	default:
251	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
252	}
253	}
254	}
255
256	#if !defined(__CFA_NO_STATISTICS__)
257	if(this->print_halts) {
258	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
259	}
260	#endif
261
262	// We were woken up, remove self from idle
263	mark_awake(this->cltr->procs, * this);
264
265	// DON'T just proceed, start looking again
266	continue MAIN_LOOP;
267	}
268
269	/* paranoid */ verify( readyThread );
270
271	// Reset io dirty bit
272	this->io.dirty = false;
273
274	// We found a thread run it
275	__run_thread(this, readyThread);
276
277	// Are we done?
278	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
279
280	if(this->io.pending && !this->io.dirty) {
281	ready_schedule_lock();
282	__cfa_io_flush( this );
283	ready_schedule_unlock();
284	}
285
286	#else
287	#warning new kernel loop
288	SEARCH: {
289	/* paranoid */ verify( ! __preemption_enabled() );
290
291	// First, lock the scheduler since we are searching for a thread
292	ready_schedule_lock();
293
294	// Try to get the next thread
295	readyThread = pop_fast( this->cltr );
296	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
297
298	// If we can't find a thread, might as well flush any outstanding I/O
299	if(this->io.pending) { __cfa_io_flush( this ); }
300
301	// Spin a little on I/O, just in case
302	for(5) {
303	__maybe_io_drain( this );
304	readyThread = pop_fast( this->cltr );
305	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
306	}
307
308	// no luck, try stealing a few times
309	for(5) {
310	if( __maybe_io_drain( this ) ) {
311	readyThread = pop_fast( this->cltr );
312	} else {
313	readyThread = pop_slow( this->cltr );
314	}
315	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
316	}
317
318	// still no luck, search for a thread
319	readyThread = pop_search( this->cltr );
320	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
321
322	// Don't block if we are done
323	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
324	ready_schedule_unlock();
325	break MAIN_LOOP;
326	}
327
328	__STATS( __tls_stats()->ready.sleep.halts++; )
329
330	// Push self to idle stack
331	ready_schedule_unlock();
332	if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
333	ready_schedule_lock();
334
335	// Confirm the ready-queue is empty
336	__maybe_io_drain( this );
337	readyThread = pop_search( this->cltr );
338	ready_schedule_unlock();
339
340	if( readyThread ) {
341	// A thread was found, cancel the halt
342	mark_awake(this->cltr->procs, * this);
343
344	__STATS( __tls_stats()->ready.sleep.cancels++; )
345
346	// continue the main loop
347	break SEARCH;
348	}
349
350	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
351	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
352
353	{
354	eventfd_t val;
355	ssize_t ret = read( this->idle, &val, sizeof(val) );
356	if(ret < 0) {
357	switch((int)errno) {
358	case EAGAIN:
359	#if EAGAIN != EWOULDBLOCK
360	case EWOULDBLOCK:
361	#endif
362	case EINTR:
363	// No need to do anything special here, just assume it's a legitimate wake-up
364	break;
365	default:
366	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
367	}
368	}
369	}
370
371	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
372
373	// We were woken up, remove self from idle
374	mark_awake(this->cltr->procs, * this);
375
376	// DON'T just proceed, start looking again
377	continue MAIN_LOOP;
378	}
379
380	RUN_THREAD:
381	/* paranoid */ verify( ! __preemption_enabled() );
382	/* paranoid */ verify( readyThread );
383
384	// Reset io dirty bit
385	this->io.dirty = false;
386
387	// We found a thread run it
388	__run_thread(this, readyThread);
389
390	// Are we done?
391	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
392
393	if(this->io.pending && !this->io.dirty) {
394	__cfa_io_flush( this );
395	}
396
397	ready_schedule_lock();
398	__maybe_io_drain( this );
399	ready_schedule_unlock();
400	#endif
401	}
402
403	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
404	}
405
406	__cfa_io_stop( this );
407
408	post( this->terminated );
409
410	if(this == mainProcessor) {
411	// HACK : the coroutine context switch expects this_thread to be set
412	// and it make sense for it to be set in all other cases except here
413	// fake it
414	__cfaabi_tls.this_thread = mainThread;
415	}
416
417	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
418	}
419
420	static int * __volatile_errno() __attribute__((noinline));
421	static int * __volatile_errno() { asm(""); return &errno; }
422
423	// KERNEL ONLY
424	// runThread runs a thread by context switching
425	// from the processor coroutine to the target thread
426	static void __run_thread(processor * this, thread$ * thrd_dst) {
427	/* paranoid */ verify( ! __preemption_enabled() );
428	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
429	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
430	__builtin_prefetch( thrd_dst->context.SP );
431
432	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
433
434	coroutine$ * proc_cor = get_coroutine(this->runner);
435
436	// set state of processor coroutine to inactive
437	verify(proc_cor->state == Active);
438	proc_cor->state = Blocked;
439
440	// Actually run the thread
441	RUNNING: while(true) {
442	thrd_dst->preempted = __NO_PREEMPTION;
443	thrd_dst->state = Active;
444
445	// Update global state
446	kernelTLS().this_thread = thrd_dst;
447
448	/* paranoid */ verify( ! __preemption_enabled() );
449	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
450	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
451	/* paranoid */ verify( thrd_dst->context.SP );
452	/* paranoid */ verify( thrd_dst->state != Halted );
453	/* 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
454	/* 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
455	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
456
457
458
459	// set context switch to the thread that the processor is executing
460	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
461	// when __cfactx_switch returns we are back in the processor coroutine
462
463
464
465	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
466	/* 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 );
467	/* 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 );
468	/* paranoid */ verify( thrd_dst->context.SP );
469	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
470	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
471	/* paranoid */ verify( ! __preemption_enabled() );
472
473	// Reset global state
474	kernelTLS().this_thread = 0p;
475
476	// We just finished running a thread, there are a few things that could have happened.
477	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
478	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
479	// 4 - Preempted
480	// In case 1, we may have won a race so we can't write to the state again.
481	// In case 2, we lost the race so we now own the thread.
482
483	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
484	// The thread was preempted, reschedule it and reset the flag
485	schedule_thread$( thrd_dst, UNPARK_LOCAL );
486	break RUNNING;
487	}
488
489	if(unlikely(thrd_dst->state == Halting)) {
490	// The thread has halted, it should never be scheduled/run again
491	// finish the thread
492	__thread_finish( thrd_dst );
493	break RUNNING;
494	}
495
496	/* paranoid */ verify( thrd_dst->state == Active );
497	thrd_dst->state = Blocked;
498
499	// set state of processor coroutine to active and the thread to inactive
500	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
501	switch(old_ticket) {
502	case TICKET_RUNNING:
503	// This is case 1, the regular case, nothing more is needed
504	break RUNNING;
505	case TICKET_UNBLOCK:
506	#if !defined(__CFA_NO_STATISTICS__)
507	__tls_stats()->ready.threads.threads++;
508	#endif
509	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
510	// In this case, just run it again.
511	continue RUNNING;
512	default:
513	// This makes no sense, something is wrong abort
514	abort();
515	}
516	}
517
518	// Just before returning to the processor, set the processor coroutine to active
519	proc_cor->state = Active;
520
521	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
522
523	#if !defined(__CFA_NO_STATISTICS__)
524	__tls_stats()->ready.threads.threads--;
525	#endif
526
527	/* paranoid */ verify( ! __preemption_enabled() );
528	}
529
530	// KERNEL_ONLY
531	void returnToKernel() {
532	/* paranoid */ verify( ! __preemption_enabled() );
533	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
534	thread$ * thrd_src = kernelTLS().this_thread;
535
536	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
537
538	// Run the thread on this processor
539	{
540	int local_errno = *__volatile_errno();
541	#if defined( __i386 ) \|\| defined( __x86_64 )
542	__x87_store;
543	#endif
544	/* paranoid */ verify( proc_cor->context.SP );
545	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
546	__cfactx_switch( &thrd_src->context, &proc_cor->context );
547	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
548	#if defined( __i386 ) \|\| defined( __x86_64 )
549	__x87_load;
550	#endif
551	*__volatile_errno() = local_errno;
552	}
553
554	#if !defined(__CFA_NO_STATISTICS__)
555	/* paranoid */ verify( thrd_src->last_proc != 0p );
556	if(thrd_src->last_proc != kernelTLS().this_processor) {
557	__tls_stats()->ready.threads.migration++;
558	}
559	#endif
560
561	/* paranoid */ verify( ! __preemption_enabled() );
562	/* 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 );
563	/* 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 );
564	}
565
566	//-----------------------------------------------------------------------------
567	// Scheduler routines
568	// KERNEL ONLY
569	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
570	/* paranoid */ verify( ! __preemption_enabled() );
571	/* paranoid */ verify( ready_schedule_islocked());
572	/* paranoid */ verify( thrd );
573	/* paranoid */ verify( thrd->state != Halted );
574	/* paranoid */ verify( thrd->curr_cluster );
575	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
576	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
577	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
578	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
579	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
580	/* paranoid */ #endif
581	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
582	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
583
584	const bool local = thrd->state != Start;
585	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
586
587	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
588	struct cluster * cl = thrd->curr_cluster;
589	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
590
591	// push the thread to the cluster ready-queue
592	push( cl, thrd, hint );
593
594	// variable thrd is no longer safe to use
595	thrd = 0xdeaddeaddeaddeadp;
596
597	// wake the cluster using the save variable.
598	__wake_one( cl );
599
600	#if !defined(__CFA_NO_STATISTICS__)
601	if( kernelTLS().this_stats ) {
602	__tls_stats()->ready.threads.threads++;
603	if(outside) {
604	__tls_stats()->ready.threads.extunpark++;
605	}
606	}
607	else {
608	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
609	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
610	}
611	#endif
612
613	/* paranoid */ verify( ready_schedule_islocked());
614	/* paranoid */ verify( ! __preemption_enabled() );
615	}
616
617	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
618	ready_schedule_lock();
619	__schedule_thread( thrd, hint );
620	ready_schedule_unlock();
621	}
622
623	// KERNEL ONLY
624	static inline thread$ * __next_thread(cluster * this) with( *this ) {
625	/* paranoid */ verify( ! __preemption_enabled() );
626
627	ready_schedule_lock();
628	thread$ * thrd = pop_fast( this );
629	ready_schedule_unlock();
630
631	/* paranoid */ verify( ! __preemption_enabled() );
632	return thrd;
633	}
634
635	// KERNEL ONLY
636	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
637	/* paranoid */ verify( ! __preemption_enabled() );
638
639	ready_schedule_lock();
640	thread$ * thrd;
641	for(25) {
642	thrd = pop_slow( this );
643	if(thrd) goto RET;
644	}
645	thrd = pop_search( this );
646
647	RET:
648	ready_schedule_unlock();
649
650	/* paranoid */ verify( ! __preemption_enabled() );
651	return thrd;
652	}
653
654	static inline bool __must_unpark( thread$ * thrd ) {
655	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
656	switch(old_ticket) {
657	case TICKET_RUNNING:
658	// Wake won the race, the thread will reschedule/rerun itself
659	return false;
660	case TICKET_BLOCKED:
661	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
662	/* paranoid */ verify( thrd->state == Blocked );
663	return true;
664	default:
665	// This makes no sense, something is wrong abort
666	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
667	}
668	}
669
670	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
671	/* paranoid */ verify( ! __preemption_enabled() );
672	/* paranoid */ verify( ready_schedule_islocked());
673
674	if( !thrd ) return;
675
676	if(__must_unpark(thrd)) {
677	// Wake lost the race,
678	__schedule_thread( thrd, hint );
679	}
680
681	/* paranoid */ verify( ready_schedule_islocked());
682	/* paranoid */ verify( ! __preemption_enabled() );
683	}
684
685	void unpark( thread$ * thrd, unpark_hint hint ) {
686	if( !thrd ) return;
687
688	if(__must_unpark(thrd)) {
689	disable_interrupts();
690	// Wake lost the race,
691	schedule_thread$( thrd, hint );
692	enable_interrupts(false);
693	}
694	}
695
696	void park( void ) {
697	__disable_interrupts_checked();
698	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
699	returnToKernel();
700	__enable_interrupts_checked();
701
702	}
703
704	extern "C" {
705	// Leave the thread monitor
706	// last routine called by a thread.
707	// Should never return
708	void __cfactx_thrd_leave() {
709	thread$ * thrd = active_thread();
710	monitor$ * this = &thrd->self_mon;
711
712	// Lock the monitor now
713	lock( this->lock __cfaabi_dbg_ctx2 );
714
715	disable_interrupts();
716
717	/* paranoid */ verify( ! __preemption_enabled() );
718	/* paranoid */ verify( thrd->state == Active );
719	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
720	/* paranoid */ verify( kernelTLS().this_thread == thrd );
721	/* paranoid */ verify( thrd->context.SP );
722	/* 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 );
723	/* 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 );
724
725	thrd->state = Halting;
726	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
727	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
728	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
729
730	// Leave the thread
731	returnToKernel();
732
733	// Control flow should never reach here!
734	abort();
735	}
736	}
737
738	// KERNEL ONLY
739	bool force_yield( __Preemption_Reason reason ) {
740	__disable_interrupts_checked();
741	thread$ * thrd = kernelTLS().this_thread;
742	/* paranoid */ verify(thrd->state == Active);
743
744	// SKULLDUGGERY: It is possible that we are preempting this thread just before
745	// it was going to park itself. If that is the case and it is already using the
746	// intrusive fields then we can't use them to preempt the thread
747	// If that is the case, abandon the preemption.
748	bool preempted = false;
749	if(thrd->link.next == 0p) {
750	preempted = true;
751	thrd->preempted = reason;
752	returnToKernel();
753	}
754	__enable_interrupts_checked( false );
755	return preempted;
756	}
757
758	//=============================================================================================
759	// Kernel Idle Sleep
760	//=============================================================================================
761	// Wake a thread from the front if there are any
762	static void __wake_one(cluster * this) {
763	/* paranoid */ verify( ! __preemption_enabled() );
764	/* paranoid */ verify( ready_schedule_islocked() );
765
766	// Check if there is a sleeping processor
767	int fd = __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST);
768
769	// If no one is sleeping, we are done
770	if( fd == 0 ) return;
771
772	// We found a processor, wake it up
773	eventfd_t val;
774	val = 1;
775	eventfd_write( fd, val );
776
777	#if !defined(__CFA_NO_STATISTICS__)
778	if( kernelTLS().this_stats ) {
779	__tls_stats()->ready.sleep.wakes++;
780	}
781	else {
782	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
783	}
784	#endif
785
786	/* paranoid */ verify( ready_schedule_islocked() );
787	/* paranoid */ verify( ! __preemption_enabled() );
788
789	return;
790	}
791
792	// Unconditionnaly wake a thread
793	void __wake_proc(processor * this) {
794	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
795
796	__disable_interrupts_checked();
797	/* paranoid */ verify( ! __preemption_enabled() );
798	eventfd_t val;
799	val = 1;
800	eventfd_write( this->idle, val );
801	__enable_interrupts_checked();
802	}
803
804	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
805	/* paranoid */ verify( ! __preemption_enabled() );
806	if(!try_lock( this )) return false;
807	this.idle++;
808	/* paranoid */ verify( this.idle <= this.total );
809	remove(proc);
810	insert_first(this.idles, proc);
811
812	__atomic_store_n(&this.fd, proc.idle, __ATOMIC_SEQ_CST);
813	unlock( this );
814	/* paranoid */ verify( ! __preemption_enabled() );
815
816	return true;
817	}
818
819	static void mark_awake(__cluster_proc_list & this, processor & proc) {
820	/* paranoid */ verify( ! __preemption_enabled() );
821	lock( this );
822	this.idle--;
823	/* paranoid */ verify( this.idle >= 0 );
824	remove(proc);
825	insert_last(this.actives, proc);
826
827	{
828	int fd = 0;
829	if(!this.idles`isEmpty) fd = this.idles`first.idle;
830	__atomic_store_n(&this.fd, fd, __ATOMIC_SEQ_CST);
831	}
832
833	unlock( this );
834	/* paranoid */ verify( ! __preemption_enabled() );
835	}
836
837	//=============================================================================================
838	// Unexpected Terminating logic
839	//=============================================================================================
840	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
841	thread$ * thrd = __cfaabi_tls.this_thread;
842
843	if(thrd) {
844	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
845	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
846
847	if ( &thrd->self_cor != thrd->curr_cor ) {
848	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
849	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
850	}
851	else {
852	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
853	}
854	}
855	else {
856	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
857	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
858	}
859	}
860
861	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
862	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
863	}
864
865	static __spinlock_t kernel_debug_lock;
866
867	extern "C" {
868	void __cfaabi_bits_acquire() {
869	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
870	}
871
872	void __cfaabi_bits_release() {
873	unlock( kernel_debug_lock );
874	}
875	}
876
877	//=============================================================================================
878	// Kernel Utilities
879	//=============================================================================================
880	#if defined(CFA_HAVE_LINUX_IO_URING_H)
881	#include "io/types.hfa"
882	#endif
883
884	static inline bool __maybe_io_drain( processor * proc ) {
885	bool ret = false;
886	#if defined(CFA_HAVE_LINUX_IO_URING_H)
887	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
888
889	// Check if we should drain the queue
890	$io_context * ctx = proc->io.ctx;
891	unsigned head = *ctx->cq.head;
892	unsigned tail = *ctx->cq.tail;
893	if(head == tail) return false;
894	#if OLD_MAIN
895	ready_schedule_lock();
896	ret = __cfa_io_drain( proc );
897	ready_schedule_unlock();
898	#else
899	ret = __cfa_io_drain( proc );
900	#endif
901	#endif
902	return ret;
903	}
904
905	//-----------------------------------------------------------------------------
906	// Debug
907	__cfaabi_dbg_debug_do(
908	extern "C" {
909	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
910	this.prev_name = prev_name;
911	this.prev_thrd = kernelTLS().this_thread;
912	}
913	}
914	)
915
916	//-----------------------------------------------------------------------------
917	// Debug
918	bool threading_enabled(void) __attribute__((const)) {
919	return true;
920	}
921
922	//-----------------------------------------------------------------------------
923	// Statistics
924	#if !defined(__CFA_NO_STATISTICS__)
925	void print_halts( processor & this ) {
926	this.print_halts = true;
927	}
928
929	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
930	/* paranoid */ verify( cltr->stats );
931
932	processor * it = &list`first;
933	for(unsigned i = 0; i < count; i++) {
934	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
935	/* paranoid */ verify( it->local_data->this_stats );
936	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
937	__tally_stats( cltr->stats, it->local_data->this_stats );
938	it = &(*it)`next;
939	}
940	}
941
942	void crawl_cluster_stats( cluster & this ) {
943	// Stop the world, otherwise stats could get really messed-up
944	// this doesn't solve all problems but does solve many
945	// so it's probably good enough
946	disable_interrupts();
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	enable_interrupts();
955	}
956
957
958	void print_stats_now( cluster & this, int flags ) {
959	crawl_cluster_stats( this );
960	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
961	}
962	#endif
963	// Local Variables: //
964	// mode: c //
965	// tab-width: 4 //
966	// End: //

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

Download in other formats:

Original Format