Context Navigation

kernel.cfa @ db614d0

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since db614d0 was abcae55, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Removed last_cpu (and small spacing changes)
Property mode set to `100644`
File size: 30.1 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	__cfa_io_flush( this );
199	readyThread = __next_thread_slow( this->cltr );
200	}
201
202	HALT:
203	if( !readyThread ) {
204	// Don't block if we are done
205	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
206
207	#if !defined(__CFA_NO_STATISTICS__)
208	__tls_stats()->ready.sleep.halts++;
209	#endif
210
211	// Push self to idle stack
212	mark_idle(this->cltr->procs, * this);
213
214	// Confirm the ready-queue is empty
215	readyThread = __next_thread_slow( this->cltr );
216	if( readyThread ) {
217	// A thread was found, cancel the halt
218	mark_awake(this->cltr->procs, * this);
219
220	#if !defined(__CFA_NO_STATISTICS__)
221	__tls_stats()->ready.sleep.cancels++;
222	#endif
223
224	// continue the mai loop
225	break HALT;
226	}
227
228	#if !defined(__CFA_NO_STATISTICS__)
229	if(this->print_halts) {
230	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
231	}
232	#endif
233
234	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
235
236	{
237	eventfd_t val;
238	ssize_t ret = read( this->idle, &val, sizeof(val) );
239	if(ret < 0) {
240	switch((int)errno) {
241	case EAGAIN:
242	#if EAGAIN != EWOULDBLOCK
243	case EWOULDBLOCK:
244	#endif
245	case EINTR:
246	// No need to do anything special here, just assume it's a legitimate wake-up
247	break;
248	default:
249	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
250	}
251	}
252	}
253
254	#if !defined(__CFA_NO_STATISTICS__)
255	if(this->print_halts) {
256	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
257	}
258	#endif
259
260	// We were woken up, remove self from idle
261	mark_awake(this->cltr->procs, * this);
262
263	// DON'T just proceed, start looking again
264	continue MAIN_LOOP;
265	}
266
267	/* paranoid */ verify( readyThread );
268
269	// Reset io dirty bit
270	this->io.dirty = false;
271
272	// We found a thread run it
273	__run_thread(this, readyThread);
274
275	// Are we done?
276	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
277
278	if(this->io.pending && !this->io.dirty) {
279	__cfa_io_flush( this );
280	}
281
282	#else
283	#warning new kernel loop
284	SEARCH: {
285	/* paranoid */ verify( ! __preemption_enabled() );
286
287	// First, lock the scheduler since we are searching for a thread
288	ready_schedule_lock();
289
290	// Try to get the next thread
291	readyThread = pop_fast( this->cltr );
292	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
293
294	// If we can't find a thread, might as well flush any outstanding I/O
295	if(this->io.pending) { __cfa_io_flush( this ); }
296
297	// Spin a little on I/O, just in case
298	for(5) {
299	__maybe_io_drain( this );
300	readyThread = pop_fast( this->cltr );
301	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
302	}
303
304	// no luck, try stealing a few times
305	for(5) {
306	if( __maybe_io_drain( this ) ) {
307	readyThread = pop_fast( this->cltr );
308	} else {
309	readyThread = pop_slow( this->cltr );
310	}
311	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
312	}
313
314	// still no luck, search for a thread
315	readyThread = pop_search( this->cltr );
316	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
317
318	// Don't block if we are done
319	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
320
321	__STATS( __tls_stats()->ready.sleep.halts++; )
322
323	// Push self to idle stack
324	ready_schedule_unlock();
325	mark_idle(this->cltr->procs, * this);
326	ready_schedule_lock();
327
328	// Confirm the ready-queue is empty
329	__maybe_io_drain( this );
330	readyThread = pop_search( this->cltr );
331	ready_schedule_unlock();
332
333	if( readyThread ) {
334	// A thread was found, cancel the halt
335	mark_awake(this->cltr->procs, * this);
336
337	__STATS( __tls_stats()->ready.sleep.cancels++; )
338
339	// continue the main loop
340	break SEARCH;
341	}
342
343	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
344	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
345
346	{
347	eventfd_t val;
348	ssize_t ret = read( this->idle, &val, sizeof(val) );
349	if(ret < 0) {
350	switch((int)errno) {
351	case EAGAIN:
352	#if EAGAIN != EWOULDBLOCK
353	case EWOULDBLOCK:
354	#endif
355	case EINTR:
356	// No need to do anything special here, just assume it's a legitimate wake-up
357	break;
358	default:
359	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
360	}
361	}
362	}
363
364	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
365
366	// We were woken up, remove self from idle
367	mark_awake(this->cltr->procs, * this);
368
369	// DON'T just proceed, start looking again
370	continue MAIN_LOOP;
371	}
372
373	RUN_THREAD:
374	/* paranoid */ verify( ! __preemption_enabled() );
375	/* paranoid */ verify( readyThread );
376
377	// Reset io dirty bit
378	this->io.dirty = false;
379
380	// We found a thread run it
381	__run_thread(this, readyThread);
382
383	// Are we done?
384	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
385
386	if(this->io.pending && !this->io.dirty) {
387	__cfa_io_flush( this );
388	}
389
390	ready_schedule_lock();
391	__maybe_io_drain( this );
392	ready_schedule_unlock();
393	#endif
394	}
395
396	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
397	}
398
399	__cfa_io_stop( this );
400
401	post( this->terminated );
402
403	if(this == mainProcessor) {
404	// HACK : the coroutine context switch expects this_thread to be set
405	// and it make sense for it to be set in all other cases except here
406	// fake it
407	__cfaabi_tls.this_thread = mainThread;
408	}
409
410	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
411	}
412
413	static int * __volatile_errno() __attribute__((noinline));
414	static int * __volatile_errno() { asm(""); return &errno; }
415
416	// KERNEL ONLY
417	// runThread runs a thread by context switching
418	// from the processor coroutine to the target thread
419	static void __run_thread(processor * this, thread$ * thrd_dst) {
420	/* paranoid */ verify( ! __preemption_enabled() );
421	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
422	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
423	__builtin_prefetch( thrd_dst->context.SP );
424
425	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
426
427	coroutine$ * proc_cor = get_coroutine(this->runner);
428
429	// set state of processor coroutine to inactive
430	verify(proc_cor->state == Active);
431	proc_cor->state = Blocked;
432
433	// Actually run the thread
434	RUNNING: while(true) {
435	thrd_dst->preempted = __NO_PREEMPTION;
436	thrd_dst->state = Active;
437
438	// Update global state
439	kernelTLS().this_thread = thrd_dst;
440
441	/* paranoid */ verify( ! __preemption_enabled() );
442	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
443	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
444	/* paranoid */ verify( thrd_dst->context.SP );
445	/* paranoid */ verify( thrd_dst->state != Halted );
446	/* 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
447	/* 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
448	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
449
450
451
452	// set context switch to the thread that the processor is executing
453	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
454	// when __cfactx_switch returns we are back in the processor coroutine
455
456
457
458	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
459	/* 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 );
460	/* 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 );
461	/* paranoid */ verify( thrd_dst->context.SP );
462	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
463	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
464	/* paranoid */ verify( ! __preemption_enabled() );
465
466	// Reset global state
467	kernelTLS().this_thread = 0p;
468
469	// We just finished running a thread, there are a few things that could have happened.
470	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
471	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
472	// 4 - Preempted
473	// In case 1, we may have won a race so we can't write to the state again.
474	// In case 2, we lost the race so we now own the thread.
475
476	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
477	// The thread was preempted, reschedule it and reset the flag
478	schedule_thread$( thrd_dst );
479	break RUNNING;
480	}
481
482	if(unlikely(thrd_dst->state == Halting)) {
483	// The thread has halted, it should never be scheduled/run again
484	// finish the thread
485	__thread_finish( thrd_dst );
486	break RUNNING;
487	}
488
489	/* paranoid */ verify( thrd_dst->state == Active );
490	thrd_dst->state = Blocked;
491
492	// set state of processor coroutine to active and the thread to inactive
493	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
494	switch(old_ticket) {
495	case TICKET_RUNNING:
496	// This is case 1, the regular case, nothing more is needed
497	break RUNNING;
498	case TICKET_UNBLOCK:
499	#if !defined(__CFA_NO_STATISTICS__)
500	__tls_stats()->ready.threads.threads++;
501	#endif
502	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
503	// In this case, just run it again.
504	continue RUNNING;
505	default:
506	// This makes no sense, something is wrong abort
507	abort();
508	}
509	}
510
511	// Just before returning to the processor, set the processor coroutine to active
512	proc_cor->state = Active;
513
514	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
515
516	#if !defined(__CFA_NO_STATISTICS__)
517	__tls_stats()->ready.threads.threads--;
518	#endif
519
520	/* paranoid */ verify( ! __preemption_enabled() );
521	}
522
523	// KERNEL_ONLY
524	void returnToKernel() {
525	/* paranoid */ verify( ! __preemption_enabled() );
526	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
527	thread$ * thrd_src = kernelTLS().this_thread;
528
529	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
530
531	// Run the thread on this processor
532	{
533	int local_errno = *__volatile_errno();
534	#if defined( __i386 ) \|\| defined( __x86_64 )
535	__x87_store;
536	#endif
537	/* paranoid */ verify( proc_cor->context.SP );
538	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
539	__cfactx_switch( &thrd_src->context, &proc_cor->context );
540	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
541	#if defined( __i386 ) \|\| defined( __x86_64 )
542	__x87_load;
543	#endif
544	*__volatile_errno() = local_errno;
545	}
546
547	#if !defined(__CFA_NO_STATISTICS__)
548	/* paranoid */ verify( thrd_src->last_proc != 0p );
549	if(thrd_src->last_proc != kernelTLS().this_processor) {
550	__tls_stats()->ready.threads.migration++;
551	}
552	#endif
553
554	/* paranoid */ verify( ! __preemption_enabled() );
555	/* 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 );
556	/* 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 );
557	}
558
559	//-----------------------------------------------------------------------------
560	// Scheduler routines
561	// KERNEL ONLY
562	static void __schedule_thread( thread$ * thrd ) {
563	/* paranoid */ verify( ! __preemption_enabled() );
564	/* paranoid */ verify( ready_schedule_islocked());
565	/* paranoid */ verify( thrd );
566	/* paranoid */ verify( thrd->state != Halted );
567	/* paranoid */ verify( thrd->curr_cluster );
568	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
569	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
570	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
571	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
572	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
573	/* paranoid */ #endif
574	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
575	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
576
577	const bool local = thrd->state != Start;
578	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
579
580	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
581	struct cluster * cl = thrd->curr_cluster;
582	__STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
583
584	// push the thread to the cluster ready-queue
585	push( cl, thrd, local );
586
587	// variable thrd is no longer safe to use
588	thrd = 0xdeaddeaddeaddeadp;
589
590	// wake the cluster using the save variable.
591	__wake_one( cl );
592
593	#if !defined(__CFA_NO_STATISTICS__)
594	if( kernelTLS().this_stats ) {
595	__tls_stats()->ready.threads.threads++;
596	if(outside) {
597	__tls_stats()->ready.threads.extunpark++;
598	}
599	}
600	else {
601	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
602	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
603	}
604	#endif
605
606	/* paranoid */ verify( ready_schedule_islocked());
607	/* paranoid */ verify( ! __preemption_enabled() );
608	}
609
610	void schedule_thread$( thread$ * thrd ) {
611	ready_schedule_lock();
612	__schedule_thread( thrd );
613	ready_schedule_unlock();
614	}
615
616	// KERNEL ONLY
617	static inline thread$ * __next_thread(cluster * this) with( *this ) {
618	/* paranoid */ verify( ! __preemption_enabled() );
619
620	ready_schedule_lock();
621	thread$ * thrd = pop_fast( this );
622	ready_schedule_unlock();
623
624	/* paranoid */ verify( ! __preemption_enabled() );
625	return thrd;
626	}
627
628	// KERNEL ONLY
629	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
630	/* paranoid */ verify( ! __preemption_enabled() );
631
632	ready_schedule_lock();
633	thread$ * thrd;
634	for(25) {
635	thrd = pop_slow( this );
636	if(thrd) goto RET;
637	}
638	thrd = pop_search( this );
639
640	RET:
641	ready_schedule_unlock();
642
643	/* paranoid */ verify( ! __preemption_enabled() );
644	return thrd;
645	}
646
647	static inline bool __must_unpark( thread$ * thrd ) {
648	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
649	switch(old_ticket) {
650	case TICKET_RUNNING:
651	// Wake won the race, the thread will reschedule/rerun itself
652	return false;
653	case TICKET_BLOCKED:
654	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
655	/* paranoid */ verify( thrd->state == Blocked );
656	return true;
657	default:
658	// This makes no sense, something is wrong abort
659	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
660	}
661	}
662
663	void __kernel_unpark( thread$ * thrd ) {
664	/* paranoid */ verify( ! __preemption_enabled() );
665	/* paranoid */ verify( ready_schedule_islocked());
666
667	if( !thrd ) return;
668
669	if(__must_unpark(thrd)) {
670	// Wake lost the race,
671	__schedule_thread( thrd );
672	}
673
674	/* paranoid */ verify( ready_schedule_islocked());
675	/* paranoid */ verify( ! __preemption_enabled() );
676	}
677
678	void unpark( thread$ * thrd ) {
679	if( !thrd ) return;
680
681	if(__must_unpark(thrd)) {
682	disable_interrupts();
683	// Wake lost the race,
684	schedule_thread$( thrd );
685	enable_interrupts(false);
686	}
687	}
688
689	void park( void ) {
690	__disable_interrupts_checked();
691	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
692	returnToKernel();
693	__enable_interrupts_checked();
694
695	}
696
697	extern "C" {
698	// Leave the thread monitor
699	// last routine called by a thread.
700	// Should never return
701	void __cfactx_thrd_leave() {
702	thread$ * thrd = active_thread();
703	monitor$ * this = &thrd->self_mon;
704
705	// Lock the monitor now
706	lock( this->lock __cfaabi_dbg_ctx2 );
707
708	disable_interrupts();
709
710	/* paranoid */ verify( ! __preemption_enabled() );
711	/* paranoid */ verify( thrd->state == Active );
712	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
713	/* paranoid */ verify( kernelTLS().this_thread == thrd );
714	/* paranoid */ verify( thrd->context.SP );
715	/* 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 );
716	/* 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 );
717
718	thrd->state = Halting;
719	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
720	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
721	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
722
723	// Leave the thread
724	returnToKernel();
725
726	// Control flow should never reach here!
727	abort();
728	}
729	}
730
731	// KERNEL ONLY
732	bool force_yield( __Preemption_Reason reason ) {
733	__disable_interrupts_checked();
734	thread$ * thrd = kernelTLS().this_thread;
735	/* paranoid */ verify(thrd->state == Active);
736
737	// SKULLDUGGERY: It is possible that we are preempting this thread just before
738	// it was going to park itself. If that is the case and it is already using the
739	// intrusive fields then we can't use them to preempt the thread
740	// If that is the case, abandon the preemption.
741	bool preempted = false;
742	if(thrd->link.next == 0p) {
743	preempted = true;
744	thrd->preempted = reason;
745	returnToKernel();
746	}
747	__enable_interrupts_checked( false );
748	return preempted;
749	}
750
751	//=============================================================================================
752	// Kernel Idle Sleep
753	//=============================================================================================
754	// Wake a thread from the front if there are any
755	static void __wake_one(cluster * this) {
756	/* paranoid */ verify( ! __preemption_enabled() );
757	/* paranoid */ verify( ready_schedule_islocked() );
758
759	// Check if there is a sleeping processor
760	processor * p;
761	unsigned idle;
762	unsigned total;
763	[idle, total, p] = query_idles(this->procs);
764
765	// If no one is sleeping, we are done
766	if( idle == 0 ) return;
767
768	// We found a processor, wake it up
769	eventfd_t val;
770	val = 1;
771	eventfd_write( p->idle, val );
772
773	#if !defined(__CFA_NO_STATISTICS__)
774	if( kernelTLS().this_stats ) {
775	__tls_stats()->ready.sleep.wakes++;
776	}
777	else {
778	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
779	}
780	#endif
781
782	/* paranoid */ verify( ready_schedule_islocked() );
783	/* paranoid */ verify( ! __preemption_enabled() );
784
785	return;
786	}
787
788	// Unconditionnaly wake a thread
789	void __wake_proc(processor * this) {
790	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
791
792	__disable_interrupts_checked();
793	/* paranoid */ verify( ! __preemption_enabled() );
794	eventfd_t val;
795	val = 1;
796	eventfd_write( this->idle, val );
797	__enable_interrupts_checked();
798	}
799
800	static void mark_idle(__cluster_proc_list & this, processor & proc) {
801	/* paranoid */ verify( ! __preemption_enabled() );
802	lock( this );
803	this.idle++;
804	/* paranoid */ verify( this.idle <= this.total );
805	remove(proc);
806	insert_first(this.idles, proc);
807	unlock( this );
808	/* paranoid */ verify( ! __preemption_enabled() );
809	}
810
811	static void mark_awake(__cluster_proc_list & this, processor & proc) {
812	/* paranoid */ verify( ! __preemption_enabled() );
813	lock( this );
814	this.idle--;
815	/* paranoid */ verify( this.idle >= 0 );
816	remove(proc);
817	insert_last(this.actives, proc);
818	unlock( this );
819	/* paranoid */ verify( ! __preemption_enabled() );
820	}
821
822	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
823	/* paranoid */ verify( ! __preemption_enabled() );
824	/* paranoid */ verify( ready_schedule_islocked() );
825
826	for() {
827	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
828	if( 1 == (l % 2) ) { Pause(); continue; }
829	unsigned idle = this.idle;
830	unsigned total = this.total;
831	processor * proc = &this.idles`first;
832	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
833	asm volatile("": : :"memory");
834	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
835	return [idle, total, proc];
836	}
837
838	/* paranoid */ verify( ready_schedule_islocked() );
839	/* paranoid */ verify( ! __preemption_enabled() );
840	}
841
842	//=============================================================================================
843	// Unexpected Terminating logic
844	//=============================================================================================
845	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
846	thread$ * thrd = __cfaabi_tls.this_thread;
847
848	if(thrd) {
849	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
850	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
851
852	if ( &thrd->self_cor != thrd->curr_cor ) {
853	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
854	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
855	}
856	else {
857	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
858	}
859	}
860	else {
861	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
862	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
863	}
864	}
865
866	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
867	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
868	}
869
870	static __spinlock_t kernel_debug_lock;
871
872	extern "C" {
873	void __cfaabi_bits_acquire() {
874	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
875	}
876
877	void __cfaabi_bits_release() {
878	unlock( kernel_debug_lock );
879	}
880	}
881
882	//=============================================================================================
883	// Kernel Utilities
884	//=============================================================================================
885	#if defined(CFA_HAVE_LINUX_IO_URING_H)
886	#include "io/types.hfa"
887	#endif
888
889	static inline bool __maybe_io_drain( processor * proc ) {
890	bool ret = false;
891	#if defined(CFA_HAVE_LINUX_IO_URING_H)
892	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
893
894	// Check if we should drain the queue
895	$io_context * ctx = proc->io.ctx;
896	unsigned head = *ctx->cq.head;
897	unsigned tail = *ctx->cq.tail;
898	if(head == tail) return false;
899	#if OLD_MAIN
900	ready_schedule_lock();
901	ret = __cfa_io_drain( proc );
902	ready_schedule_unlock();
903	#else
904	ret = __cfa_io_drain( proc );
905	#endif
906	#endif
907	return ret;
908	}
909
910	//-----------------------------------------------------------------------------
911	// Debug
912	__cfaabi_dbg_debug_do(
913	extern "C" {
914	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
915	this.prev_name = prev_name;
916	this.prev_thrd = kernelTLS().this_thread;
917	}
918	}
919	)
920
921	//-----------------------------------------------------------------------------
922	// Debug
923	bool threading_enabled(void) __attribute__((const)) {
924	return true;
925	}
926
927	//-----------------------------------------------------------------------------
928	// Statistics
929	#if !defined(__CFA_NO_STATISTICS__)
930	void print_halts( processor & this ) {
931	this.print_halts = true;
932	}
933
934	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
935	/* paranoid */ verify( cltr->stats );
936
937	processor * it = &list`first;
938	for(unsigned i = 0; i < count; i++) {
939	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
940	/* paranoid */ verify( it->local_data->this_stats );
941	__tally_stats( cltr->stats, it->local_data->this_stats );
942	it = &(*it)`next;
943	}
944	}
945
946	void crawl_cluster_stats( cluster & this ) {
947	// Stop the world, otherwise stats could get really messed-up
948	// this doesn't solve all problems but does solve many
949	// so it's probably good enough
950	uint_fast32_t last_size = ready_mutate_lock();
951
952	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
953	crawl_list(&this, this.procs.idles , this.procs.idle );
954
955	// Unlock the RWlock
956	ready_mutate_unlock( last_size );
957	}
958
959
960	void print_stats_now( cluster & this, int flags ) {
961	crawl_cluster_stats( this );
962	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
963	}
964	#endif
965	// Local Variables: //
966	// mode: c //
967	// tab-width: 4 //
968	// End: //

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

Download in other formats:

Original Format