Context Navigation

kernel.cfa @ e8ac228

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since e8ac228 was c33c2af, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Made some of the schedule locking more fine grain.
Property mode set to `100644`
File size: 30.5 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// kernel.c --
8	//
9	// Author : Thierry Delisle
10	// Created On : Tue Jan 17 12:27:26 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 31 07:08:20 2020
13	// Update Count : 71
14	//
15
16	#define __cforall_thread__
17	#define _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 void mark_idle (__cluster_proc_list & idles, processor & proc);
127	static void mark_awake(__cluster_proc_list & idles, processor & proc);
128	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
129
130	extern void __cfa_io_start( processor * );
131	extern bool __cfa_io_drain( processor * );
132	extern void __cfa_io_flush( processor * );
133	extern void __cfa_io_stop ( processor * );
134	static inline bool __maybe_io_drain( processor * );
135
136	extern void __disable_interrupts_hard();
137	extern void __enable_interrupts_hard();
138
139	static inline void __disable_interrupts_checked() {
140	/* paranoid */ verify( __preemption_enabled() );
141	disable_interrupts();
142	/* paranoid */ verify( ! __preemption_enabled() );
143	}
144
145	static 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
155	void 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	ready_schedule_lock();
199	__cfa_io_flush( this );
200	ready_schedule_unlock();
201
202	readyThread = __next_thread_slow( this->cltr );
203	}
204
205	HALT:
206	if( !readyThread ) {
207	// Don't block if we are done
208	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
209
210	#if !defined(__CFA_NO_STATISTICS__)
211	__tls_stats()->ready.sleep.halts++;
212	#endif
213
214	// Push self to idle stack
215	mark_idle(this->cltr->procs, * this);
216
217	// Confirm the ready-queue is empty
218	readyThread = __next_thread_slow( this->cltr );
219	if( readyThread ) {
220	// A thread was found, cancel the halt
221	mark_awake(this->cltr->procs, * this);
222
223	#if !defined(__CFA_NO_STATISTICS__)
224	__tls_stats()->ready.sleep.cancels++;
225	#endif
226
227	// continue the mai loop
228	break HALT;
229	}
230
231	#if !defined(__CFA_NO_STATISTICS__)
232	if(this->print_halts) {
233	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
234	}
235	#endif
236
237	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
238
239	{
240	eventfd_t val;
241	ssize_t ret = read( this->idle, &val, sizeof(val) );
242	if(ret < 0) {
243	switch((int)errno) {
244	case EAGAIN:
245	#if EAGAIN != EWOULDBLOCK
246	case EWOULDBLOCK:
247	#endif
248	case EINTR:
249	// No need to do anything special here, just assume it's a legitimate wake-up
250	break;
251	default:
252	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
253	}
254	}
255	}
256
257	#if !defined(__CFA_NO_STATISTICS__)
258	if(this->print_halts) {
259	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
260	}
261	#endif
262
263	// We were woken up, remove self from idle
264	mark_awake(this->cltr->procs, * this);
265
266	// DON'T just proceed, start looking again
267	continue MAIN_LOOP;
268	}
269
270	/* paranoid */ verify( readyThread );
271
272	// Reset io dirty bit
273	this->io.dirty = false;
274
275	// We found a thread run it
276	__run_thread(this, readyThread);
277
278	// Are we done?
279	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
280
281	if(this->io.pending && !this->io.dirty) {
282	ready_schedule_lock();
283	__cfa_io_flush( this );
284	ready_schedule_unlock();
285	}
286
287	#else
288	#warning new kernel loop
289	SEARCH: {
290	/* paranoid */ verify( ! __preemption_enabled() );
291
292	// First, lock the scheduler since we are searching for a thread
293	ready_schedule_lock();
294
295	// Try to get the next thread
296	readyThread = pop_fast( this->cltr );
297	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
298
299	// If we can't find a thread, might as well flush any outstanding I/O
300	if(this->io.pending) { __cfa_io_flush( this ); }
301
302	// Spin a little on I/O, just in case
303	for(5) {
304	__maybe_io_drain( this );
305	readyThread = pop_fast( this->cltr );
306	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
307	}
308
309	// no luck, try stealing a few times
310	for(5) {
311	if( __maybe_io_drain( this ) ) {
312	readyThread = pop_fast( this->cltr );
313	} else {
314	readyThread = pop_slow( this->cltr );
315	}
316	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
317	}
318
319	// still no luck, search for a thread
320	readyThread = pop_search( this->cltr );
321	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
322
323	// Don't block if we are done
324	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
325	ready_schedule_unlock();
326	break MAIN_LOOP;
327	}
328
329	__STATS( __tls_stats()->ready.sleep.halts++; )
330
331	// Push self to idle stack
332	ready_schedule_unlock();
333	mark_idle(this->cltr->procs, * this);
334	ready_schedule_lock();
335
336	// Confirm the ready-queue is empty
337	__maybe_io_drain( this );
338	readyThread = pop_search( this->cltr );
339	ready_schedule_unlock();
340
341	if( readyThread ) {
342	// A thread was found, cancel the halt
343	mark_awake(this->cltr->procs, * this);
344
345	__STATS( __tls_stats()->ready.sleep.cancels++; )
346
347	// continue the main loop
348	break SEARCH;
349	}
350
351	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
352	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
353
354	{
355	eventfd_t val;
356	ssize_t ret = read( this->idle, &val, sizeof(val) );
357	if(ret < 0) {
358	switch((int)errno) {
359	case EAGAIN:
360	#if EAGAIN != EWOULDBLOCK
361	case EWOULDBLOCK:
362	#endif
363	case EINTR:
364	// No need to do anything special here, just assume it's a legitimate wake-up
365	break;
366	default:
367	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
368	}
369	}
370	}
371
372	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
373
374	// We were woken up, remove self from idle
375	mark_awake(this->cltr->procs, * this);
376
377	// DON'T just proceed, start looking again
378	continue MAIN_LOOP;
379	}
380
381	RUN_THREAD:
382	/* paranoid */ verify( ! __preemption_enabled() );
383	/* paranoid */ verify( readyThread );
384
385	// Reset io dirty bit
386	this->io.dirty = false;
387
388	// We found a thread run it
389	__run_thread(this, readyThread);
390
391	// Are we done?
392	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
393
394	if(this->io.pending && !this->io.dirty) {
395	__cfa_io_flush( this );
396	}
397
398	ready_schedule_lock();
399	__maybe_io_drain( this );
400	ready_schedule_unlock();
401	#endif
402	}
403
404	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
405	}
406
407	__cfa_io_stop( this );
408
409	post( this->terminated );
410
411	if(this == mainProcessor) {
412	// HACK : the coroutine context switch expects this_thread to be set
413	// and it make sense for it to be set in all other cases except here
414	// fake it
415	__cfaabi_tls.this_thread = mainThread;
416	}
417
418	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
419	}
420
421	static int * __volatile_errno() __attribute__((noinline));
422	static int * __volatile_errno() { asm(""); return &errno; }
423
424	// KERNEL ONLY
425	// runThread runs a thread by context switching
426	// from the processor coroutine to the target thread
427	static void __run_thread(processor * this, thread$ * thrd_dst) {
428	/* paranoid */ verify( ! __preemption_enabled() );
429	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
430	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
431	__builtin_prefetch( thrd_dst->context.SP );
432
433	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
434
435	coroutine$ * proc_cor = get_coroutine(this->runner);
436
437	// set state of processor coroutine to inactive
438	verify(proc_cor->state == Active);
439	proc_cor->state = Blocked;
440
441	// Actually run the thread
442	RUNNING: while(true) {
443	thrd_dst->preempted = __NO_PREEMPTION;
444	thrd_dst->state = Active;
445
446	// Update global state
447	kernelTLS().this_thread = thrd_dst;
448
449	/* paranoid */ verify( ! __preemption_enabled() );
450	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
451	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
452	/* paranoid */ verify( thrd_dst->context.SP );
453	/* paranoid */ verify( thrd_dst->state != Halted );
454	/* 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
455	/* 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
456	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
457
458
459
460	// set context switch to the thread that the processor is executing
461	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
462	// when __cfactx_switch returns we are back in the processor coroutine
463
464
465
466	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
467	/* 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 );
468	/* 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 );
469	/* paranoid */ verify( thrd_dst->context.SP );
470	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
471	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
472	/* paranoid */ verify( ! __preemption_enabled() );
473
474	// Reset global state
475	kernelTLS().this_thread = 0p;
476
477	// We just finished running a thread, there are a few things that could have happened.
478	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
479	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
480	// 4 - Preempted
481	// In case 1, we may have won a race so we can't write to the state again.
482	// In case 2, we lost the race so we now own the thread.
483
484	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
485	// The thread was preempted, reschedule it and reset the flag
486	schedule_thread$( thrd_dst, UNPARK_LOCAL );
487	break RUNNING;
488	}
489
490	if(unlikely(thrd_dst->state == Halting)) {
491	// The thread has halted, it should never be scheduled/run again
492	// finish the thread
493	__thread_finish( thrd_dst );
494	break RUNNING;
495	}
496
497	/* paranoid */ verify( thrd_dst->state == Active );
498	thrd_dst->state = Blocked;
499
500	// set state of processor coroutine to active and the thread to inactive
501	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
502	switch(old_ticket) {
503	case TICKET_RUNNING:
504	// This is case 1, the regular case, nothing more is needed
505	break RUNNING;
506	case TICKET_UNBLOCK:
507	#if !defined(__CFA_NO_STATISTICS__)
508	__tls_stats()->ready.threads.threads++;
509	#endif
510	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
511	// In this case, just run it again.
512	continue RUNNING;
513	default:
514	// This makes no sense, something is wrong abort
515	abort();
516	}
517	}
518
519	// Just before returning to the processor, set the processor coroutine to active
520	proc_cor->state = Active;
521
522	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
523
524	#if !defined(__CFA_NO_STATISTICS__)
525	__tls_stats()->ready.threads.threads--;
526	#endif
527
528	/* paranoid */ verify( ! __preemption_enabled() );
529	}
530
531	// KERNEL_ONLY
532	void returnToKernel() {
533	/* paranoid */ verify( ! __preemption_enabled() );
534	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
535	thread$ * thrd_src = kernelTLS().this_thread;
536
537	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
538
539	// Run the thread on this processor
540	{
541	int local_errno = *__volatile_errno();
542	#if defined( __i386 ) \|\| defined( __x86_64 )
543	__x87_store;
544	#endif
545	/* paranoid */ verify( proc_cor->context.SP );
546	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
547	__cfactx_switch( &thrd_src->context, &proc_cor->context );
548	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
549	#if defined( __i386 ) \|\| defined( __x86_64 )
550	__x87_load;
551	#endif
552	*__volatile_errno() = local_errno;
553	}
554
555	#if !defined(__CFA_NO_STATISTICS__)
556	/* paranoid */ verify( thrd_src->last_proc != 0p );
557	if(thrd_src->last_proc != kernelTLS().this_processor) {
558	__tls_stats()->ready.threads.migration++;
559	}
560	#endif
561
562	/* paranoid */ verify( ! __preemption_enabled() );
563	/* 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 );
564	/* 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 );
565	}
566
567	//-----------------------------------------------------------------------------
568	// Scheduler routines
569	// KERNEL ONLY
570	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
571	/* paranoid */ verify( ! __preemption_enabled() );
572	/* paranoid */ verify( ready_schedule_islocked());
573	/* paranoid */ verify( thrd );
574	/* paranoid */ verify( thrd->state != Halted );
575	/* paranoid */ verify( thrd->curr_cluster );
576	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
577	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
578	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
579	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
580	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
581	/* paranoid */ #endif
582	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
583	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
584
585	const bool local = thrd->state != Start;
586	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
587
588	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
589	struct cluster * cl = thrd->curr_cluster;
590	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
591
592	// push the thread to the cluster ready-queue
593	push( cl, thrd, hint );
594
595	// variable thrd is no longer safe to use
596	thrd = 0xdeaddeaddeaddeadp;
597
598	// wake the cluster using the save variable.
599	__wake_one( cl );
600
601	#if !defined(__CFA_NO_STATISTICS__)
602	if( kernelTLS().this_stats ) {
603	__tls_stats()->ready.threads.threads++;
604	if(outside) {
605	__tls_stats()->ready.threads.extunpark++;
606	}
607	}
608	else {
609	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
610	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
611	}
612	#endif
613
614	/* paranoid */ verify( ready_schedule_islocked());
615	/* paranoid */ verify( ! __preemption_enabled() );
616	}
617
618	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
619	ready_schedule_lock();
620	__schedule_thread( thrd, hint );
621	ready_schedule_unlock();
622	}
623
624	// KERNEL ONLY
625	static inline thread$ * __next_thread(cluster * this) with( *this ) {
626	/* paranoid */ verify( ! __preemption_enabled() );
627
628	ready_schedule_lock();
629	thread$ * thrd = pop_fast( this );
630	ready_schedule_unlock();
631
632	/* paranoid */ verify( ! __preemption_enabled() );
633	return thrd;
634	}
635
636	// KERNEL ONLY
637	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
638	/* paranoid */ verify( ! __preemption_enabled() );
639
640	ready_schedule_lock();
641	thread$ * thrd;
642	for(25) {
643	thrd = pop_slow( this );
644	if(thrd) goto RET;
645	}
646	thrd = pop_search( this );
647
648	RET:
649	ready_schedule_unlock();
650
651	/* paranoid */ verify( ! __preemption_enabled() );
652	return thrd;
653	}
654
655	static inline bool __must_unpark( thread$ * thrd ) {
656	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
657	switch(old_ticket) {
658	case TICKET_RUNNING:
659	// Wake won the race, the thread will reschedule/rerun itself
660	return false;
661	case TICKET_BLOCKED:
662	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
663	/* paranoid */ verify( thrd->state == Blocked );
664	return true;
665	default:
666	// This makes no sense, something is wrong abort
667	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
668	}
669	}
670
671	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
672	/* paranoid */ verify( ! __preemption_enabled() );
673	/* paranoid */ verify( ready_schedule_islocked());
674
675	if( !thrd ) return;
676
677	if(__must_unpark(thrd)) {
678	// Wake lost the race,
679	__schedule_thread( thrd, hint );
680	}
681
682	/* paranoid */ verify( ready_schedule_islocked());
683	/* paranoid */ verify( ! __preemption_enabled() );
684	}
685
686	void unpark( thread$ * thrd, unpark_hint hint ) {
687	if( !thrd ) return;
688
689	if(__must_unpark(thrd)) {
690	disable_interrupts();
691	// Wake lost the race,
692	schedule_thread$( thrd, hint );
693	enable_interrupts(false);
694	}
695	}
696
697	void park( void ) {
698	__disable_interrupts_checked();
699	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
700	returnToKernel();
701	__enable_interrupts_checked();
702
703	}
704
705	extern "C" {
706	// Leave the thread monitor
707	// last routine called by a thread.
708	// Should never return
709	void __cfactx_thrd_leave() {
710	thread$ * thrd = active_thread();
711	monitor$ * this = &thrd->self_mon;
712
713	// Lock the monitor now
714	lock( this->lock __cfaabi_dbg_ctx2 );
715
716	disable_interrupts();
717
718	/* paranoid */ verify( ! __preemption_enabled() );
719	/* paranoid */ verify( thrd->state == Active );
720	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
721	/* paranoid */ verify( kernelTLS().this_thread == thrd );
722	/* paranoid */ verify( thrd->context.SP );
723	/* 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 );
724	/* 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 );
725
726	thrd->state = Halting;
727	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
728	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
729	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
730
731	// Leave the thread
732	returnToKernel();
733
734	// Control flow should never reach here!
735	abort();
736	}
737	}
738
739	// KERNEL ONLY
740	bool force_yield( __Preemption_Reason reason ) {
741	__disable_interrupts_checked();
742	thread$ * thrd = kernelTLS().this_thread;
743	/* paranoid */ verify(thrd->state == Active);
744
745	// SKULLDUGGERY: It is possible that we are preempting this thread just before
746	// it was going to park itself. If that is the case and it is already using the
747	// intrusive fields then we can't use them to preempt the thread
748	// If that is the case, abandon the preemption.
749	bool preempted = false;
750	if(thrd->link.next == 0p) {
751	preempted = true;
752	thrd->preempted = reason;
753	returnToKernel();
754	}
755	__enable_interrupts_checked( false );
756	return preempted;
757	}
758
759	//=============================================================================================
760	// Kernel Idle Sleep
761	//=============================================================================================
762	// Wake a thread from the front if there are any
763	static void __wake_one(cluster * this) {
764	/* paranoid */ verify( ! __preemption_enabled() );
765	/* paranoid */ verify( ready_schedule_islocked() );
766
767	// Check if there is a sleeping processor
768	processor * p;
769	unsigned idle;
770	unsigned total;
771	[idle, total, p] = query_idles(this->procs);
772
773	// If no one is sleeping, we are done
774	if( idle == 0 ) return;
775
776	// We found a processor, wake it up
777	eventfd_t val;
778	val = 1;
779	eventfd_write( p->idle, val );
780
781	#if !defined(__CFA_NO_STATISTICS__)
782	if( kernelTLS().this_stats ) {
783	__tls_stats()->ready.sleep.wakes++;
784	}
785	else {
786	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
787	}
788	#endif
789
790	/* paranoid */ verify( ready_schedule_islocked() );
791	/* paranoid */ verify( ! __preemption_enabled() );
792
793	return;
794	}
795
796	// Unconditionnaly wake a thread
797	void __wake_proc(processor * this) {
798	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
799
800	__disable_interrupts_checked();
801	/* paranoid */ verify( ! __preemption_enabled() );
802	eventfd_t val;
803	val = 1;
804	eventfd_write( this->idle, val );
805	__enable_interrupts_checked();
806	}
807
808	static void mark_idle(__cluster_proc_list & this, processor & proc) {
809	/* paranoid */ verify( ! __preemption_enabled() );
810	lock( this );
811	this.idle++;
812	/* paranoid */ verify( this.idle <= this.total );
813	remove(proc);
814	insert_first(this.idles, proc);
815	unlock( this );
816	/* paranoid */ verify( ! __preemption_enabled() );
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	unlock( this );
827	/* paranoid */ verify( ! __preemption_enabled() );
828	}
829
830	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
831	/* paranoid */ verify( ! __preemption_enabled() );
832	/* paranoid */ verify( ready_schedule_islocked() );
833
834	for() {
835	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
836	if( 1 == (l % 2) ) { Pause(); continue; }
837	unsigned idle = this.idle;
838	unsigned total = this.total;
839	processor * proc = &this.idles`first;
840	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
841	asm volatile("": : :"memory");
842	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
843	return [idle, total, proc];
844	}
845
846	/* paranoid */ verify( ready_schedule_islocked() );
847	/* paranoid */ verify( ! __preemption_enabled() );
848	}
849
850	//=============================================================================================
851	// Unexpected Terminating logic
852	//=============================================================================================
853	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
854	thread$ * thrd = __cfaabi_tls.this_thread;
855
856	if(thrd) {
857	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
858	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
859
860	if ( &thrd->self_cor != thrd->curr_cor ) {
861	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
862	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
863	}
864	else {
865	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
866	}
867	}
868	else {
869	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
870	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
871	}
872	}
873
874	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
875	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
876	}
877
878	static __spinlock_t kernel_debug_lock;
879
880	extern "C" {
881	void __cfaabi_bits_acquire() {
882	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
883	}
884
885	void __cfaabi_bits_release() {
886	unlock( kernel_debug_lock );
887	}
888	}
889
890	//=============================================================================================
891	// Kernel Utilities
892	//=============================================================================================
893	#if defined(CFA_HAVE_LINUX_IO_URING_H)
894	#include "io/types.hfa"
895	#endif
896
897	static inline bool __maybe_io_drain( processor * proc ) {
898	bool ret = false;
899	#if defined(CFA_HAVE_LINUX_IO_URING_H)
900	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
901
902	// Check if we should drain the queue
903	$io_context * ctx = proc->io.ctx;
904	unsigned head = *ctx->cq.head;
905	unsigned tail = *ctx->cq.tail;
906	if(head == tail) return false;
907	#if OLD_MAIN
908	ready_schedule_lock();
909	ret = __cfa_io_drain( proc );
910	ready_schedule_unlock();
911	#else
912	ret = __cfa_io_drain( proc );
913	#endif
914	#endif
915	return ret;
916	}
917
918	//-----------------------------------------------------------------------------
919	// Debug
920	__cfaabi_dbg_debug_do(
921	extern "C" {
922	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
923	this.prev_name = prev_name;
924	this.prev_thrd = kernelTLS().this_thread;
925	}
926	}
927	)
928
929	//-----------------------------------------------------------------------------
930	// Debug
931	bool threading_enabled(void) __attribute__((const)) {
932	return true;
933	}
934
935	//-----------------------------------------------------------------------------
936	// Statistics
937	#if !defined(__CFA_NO_STATISTICS__)
938	void print_halts( processor & this ) {
939	this.print_halts = true;
940	}
941
942	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
943	/* paranoid */ verify( cltr->stats );
944
945	processor * it = &list`first;
946	for(unsigned i = 0; i < count; i++) {
947	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
948	/* paranoid */ verify( it->local_data->this_stats );
949	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
950	__tally_stats( cltr->stats, it->local_data->this_stats );
951	it = &(*it)`next;
952	}
953	}
954
955	void crawl_cluster_stats( cluster & this ) {
956	// Stop the world, otherwise stats could get really messed-up
957	// this doesn't solve all problems but does solve many
958	// so it's probably good enough
959	disable_interrupts();
960	uint_fast32_t last_size = ready_mutate_lock();
961
962	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
963	crawl_list(&this, this.procs.idles , this.procs.idle );
964
965	// Unlock the RWlock
966	ready_mutate_unlock( last_size );
967	enable_interrupts();
968	}
969
970
971	void print_stats_now( cluster & this, int flags ) {
972	crawl_cluster_stats( this );
973	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
974	}
975	#endif
976	// Local Variables: //
977	// mode: c //
978	// tab-width: 4 //
979	// End: //

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

Download in other formats:

Original Format