Context Navigation

kernel.cfa @ 1757f98

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since 1757f98 was 1757f98, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Refactoring idle sleep to try and help the change from idle on read to idle on io_uring_enter.
Property mode set to `100644`
File size: 30.0 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// kernel.c --
8	//
9	// Author : Thierry Delisle
10	// Created On : Tue Jan 17 12:27:26 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 31 07:08:20 2020
13	// Update Count : 71
14	//
15
16	#define __cforall_thread__
17	#define _GNU_SOURCE
18
19	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
20
21	//C Includes
22	#include <errno.h>
23	#include <stdio.h>
24	#include <string.h>
25	#include <signal.h>
26	#include <unistd.h>
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 idle_sleep(processor * proc);
127	static bool mark_idle (__cluster_proc_list & idles, processor & proc);
128	static void mark_awake(__cluster_proc_list & idles, processor & proc);
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	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
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	idle_sleep( this );
232
233	// We were woken up, remove self from idle
234	mark_awake(this->cltr->procs, * this);
235
236	// DON'T just proceed, start looking again
237	continue MAIN_LOOP;
238	}
239
240	/* paranoid */ verify( readyThread );
241
242	// Reset io dirty bit
243	this->io.dirty = false;
244
245	// We found a thread run it
246	__run_thread(this, readyThread);
247
248	// Are we done?
249	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
250
251	if(this->io.pending && !this->io.dirty) {
252	ready_schedule_lock();
253	__cfa_io_flush( this );
254	ready_schedule_unlock();
255	}
256
257	#else
258	#warning new kernel loop
259	SEARCH: {
260	/* paranoid */ verify( ! __preemption_enabled() );
261
262	// First, lock the scheduler since we are searching for a thread
263	ready_schedule_lock();
264
265	// Try to get the next thread
266	readyThread = pop_fast( this->cltr );
267	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
268
269	// If we can't find a thread, might as well flush any outstanding I/O
270	if(this->io.pending) { __cfa_io_flush( this ); }
271
272	// Spin a little on I/O, just in case
273	for(5) {
274	__maybe_io_drain( this );
275	readyThread = pop_fast( this->cltr );
276	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
277	}
278
279	// no luck, try stealing a few times
280	for(5) {
281	if( __maybe_io_drain( this ) ) {
282	readyThread = pop_fast( this->cltr );
283	} else {
284	readyThread = pop_slow( this->cltr );
285	}
286	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
287	}
288
289	// still no luck, search for a thread
290	readyThread = pop_search( this->cltr );
291	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
292
293	// Don't block if we are done
294	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
295	ready_schedule_unlock();
296	break MAIN_LOOP;
297	}
298
299	__STATS( __tls_stats()->ready.sleep.halts++; )
300
301	// Push self to idle stack
302	ready_schedule_unlock();
303	if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
304	ready_schedule_lock();
305
306	// Confirm the ready-queue is empty
307	__maybe_io_drain( this );
308	readyThread = pop_search( this->cltr );
309	ready_schedule_unlock();
310
311	if( readyThread ) {
312	// A thread was found, cancel the halt
313	mark_awake(this->cltr->procs, * this);
314
315	__STATS( __tls_stats()->ready.sleep.cancels++; )
316
317	// continue the main loop
318	break SEARCH;
319	}
320
321	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
322	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
323
324	{
325	eventfd_t val;
326	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
327	if(ret < 0) {
328	switch((int)errno) {
329	case EAGAIN:
330	#if EAGAIN != EWOULDBLOCK
331	case EWOULDBLOCK:
332	#endif
333	case EINTR:
334	// No need to do anything special here, just assume it's a legitimate wake-up
335	break;
336	default:
337	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
338	}
339	}
340	}
341
342	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
343
344	// We were woken up, remove self from idle
345	mark_awake(this->cltr->procs, * this);
346
347	// DON'T just proceed, start looking again
348	continue MAIN_LOOP;
349	}
350
351	RUN_THREAD:
352	/* paranoid */ verify( ! __preemption_enabled() );
353	/* paranoid */ verify( readyThread );
354
355	// Reset io dirty bit
356	this->io.dirty = false;
357
358	// We found a thread run it
359	__run_thread(this, readyThread);
360
361	// Are we done?
362	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
363
364	if(this->io.pending && !this->io.dirty) {
365	__cfa_io_flush( this );
366	}
367
368	ready_schedule_lock();
369	__maybe_io_drain( this );
370	ready_schedule_unlock();
371	#endif
372	}
373
374	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
375	}
376
377	__cfa_io_stop( this );
378
379	post( this->terminated );
380
381	if(this == mainProcessor) {
382	// HACK : the coroutine context switch expects this_thread to be set
383	// and it make sense for it to be set in all other cases except here
384	// fake it
385	__cfaabi_tls.this_thread = mainThread;
386	}
387
388	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
389	}
390
391	static int * __volatile_errno() __attribute__((noinline));
392	static int * __volatile_errno() { asm(""); return &errno; }
393
394	// KERNEL ONLY
395	// runThread runs a thread by context switching
396	// from the processor coroutine to the target thread
397	static void __run_thread(processor * this, thread$ * thrd_dst) {
398	/* paranoid */ verify( ! __preemption_enabled() );
399	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
400	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
401	__builtin_prefetch( thrd_dst->context.SP );
402
403	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
404
405	coroutine$ * proc_cor = get_coroutine(this->runner);
406
407	// set state of processor coroutine to inactive
408	verify(proc_cor->state == Active);
409	proc_cor->state = Blocked;
410
411	// Actually run the thread
412	RUNNING: while(true) {
413	thrd_dst->preempted = __NO_PREEMPTION;
414	thrd_dst->state = Active;
415
416	// Update global state
417	kernelTLS().this_thread = thrd_dst;
418
419	/* paranoid */ verify( ! __preemption_enabled() );
420	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
421	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
422	/* paranoid */ verify( thrd_dst->context.SP );
423	/* paranoid */ verify( thrd_dst->state != Halted );
424	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) \|\| thrd_dst->curr_cor == proc_cor \|\| 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
425	/* 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
426	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
427
428
429
430	// set context switch to the thread that the processor is executing
431	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
432	// when __cfactx_switch returns we are back in the processor coroutine
433
434
435
436	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
437	/* 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 );
438	/* 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 );
439	/* paranoid */ verify( thrd_dst->context.SP );
440	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
441	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
442	/* paranoid */ verify( ! __preemption_enabled() );
443
444	// Reset global state
445	kernelTLS().this_thread = 0p;
446
447	// We just finished running a thread, there are a few things that could have happened.
448	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
449	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
450	// 4 - Preempted
451	// In case 1, we may have won a race so we can't write to the state again.
452	// In case 2, we lost the race so we now own the thread.
453
454	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
455	// The thread was preempted, reschedule it and reset the flag
456	schedule_thread$( thrd_dst, UNPARK_LOCAL );
457	break RUNNING;
458	}
459
460	if(unlikely(thrd_dst->state == Halting)) {
461	// The thread has halted, it should never be scheduled/run again
462	// finish the thread
463	__thread_finish( thrd_dst );
464	break RUNNING;
465	}
466
467	/* paranoid */ verify( thrd_dst->state == Active );
468	thrd_dst->state = Blocked;
469
470	// set state of processor coroutine to active and the thread to inactive
471	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
472	switch(old_ticket) {
473	case TICKET_RUNNING:
474	// This is case 1, the regular case, nothing more is needed
475	break RUNNING;
476	case TICKET_UNBLOCK:
477	#if !defined(__CFA_NO_STATISTICS__)
478	__tls_stats()->ready.threads.threads++;
479	#endif
480	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
481	// In this case, just run it again.
482	continue RUNNING;
483	default:
484	// This makes no sense, something is wrong abort
485	abort();
486	}
487	}
488
489	// Just before returning to the processor, set the processor coroutine to active
490	proc_cor->state = Active;
491
492	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
493
494	#if !defined(__CFA_NO_STATISTICS__)
495	__tls_stats()->ready.threads.threads--;
496	#endif
497
498	/* paranoid */ verify( ! __preemption_enabled() );
499	}
500
501	// KERNEL_ONLY
502	void returnToKernel() {
503	/* paranoid */ verify( ! __preemption_enabled() );
504	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
505	thread$ * thrd_src = kernelTLS().this_thread;
506
507	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
508
509	// Run the thread on this processor
510	{
511	int local_errno = *__volatile_errno();
512	#if defined( __i386 ) \|\| defined( __x86_64 )
513	__x87_store;
514	#endif
515	/* paranoid */ verify( proc_cor->context.SP );
516	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
517	__cfactx_switch( &thrd_src->context, &proc_cor->context );
518	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
519	#if defined( __i386 ) \|\| defined( __x86_64 )
520	__x87_load;
521	#endif
522	*__volatile_errno() = local_errno;
523	}
524
525	#if !defined(__CFA_NO_STATISTICS__)
526	/* paranoid */ verify( thrd_src->last_proc != 0p );
527	if(thrd_src->last_proc != kernelTLS().this_processor) {
528	__tls_stats()->ready.threads.migration++;
529	}
530	#endif
531
532	/* paranoid */ verify( ! __preemption_enabled() );
533	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ) \|\| thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_src );
534	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit) \|\| thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_src );
535	}
536
537	//-----------------------------------------------------------------------------
538	// Scheduler routines
539	// KERNEL ONLY
540	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
541	/* paranoid */ verify( ! __preemption_enabled() );
542	/* paranoid */ verify( ready_schedule_islocked());
543	/* paranoid */ verify( thrd );
544	/* paranoid */ verify( thrd->state != Halted );
545	/* paranoid */ verify( thrd->curr_cluster );
546	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
547	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
548	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
549	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
550	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
551	/* paranoid */ #endif
552	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
553	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
554
555	const bool local = thrd->state != Start;
556	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
557
558	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
559	struct cluster * cl = thrd->curr_cluster;
560	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
561
562	// push the thread to the cluster ready-queue
563	push( cl, thrd, hint );
564
565	// variable thrd is no longer safe to use
566	thrd = 0xdeaddeaddeaddeadp;
567
568	// wake the cluster using the save variable.
569	__wake_one( cl );
570
571	#if !defined(__CFA_NO_STATISTICS__)
572	if( kernelTLS().this_stats ) {
573	__tls_stats()->ready.threads.threads++;
574	if(outside) {
575	__tls_stats()->ready.threads.extunpark++;
576	}
577	}
578	else {
579	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
580	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
581	}
582	#endif
583
584	/* paranoid */ verify( ready_schedule_islocked());
585	/* paranoid */ verify( ! __preemption_enabled() );
586	}
587
588	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
589	ready_schedule_lock();
590	__schedule_thread( thrd, hint );
591	ready_schedule_unlock();
592	}
593
594	// KERNEL ONLY
595	static inline thread$ * __next_thread(cluster * this) with( *this ) {
596	/* paranoid */ verify( ! __preemption_enabled() );
597
598	ready_schedule_lock();
599	thread$ * thrd = pop_fast( this );
600	ready_schedule_unlock();
601
602	/* paranoid */ verify( ! __preemption_enabled() );
603	return thrd;
604	}
605
606	// KERNEL ONLY
607	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
608	/* paranoid */ verify( ! __preemption_enabled() );
609
610	ready_schedule_lock();
611	thread$ * thrd;
612	for(25) {
613	thrd = pop_slow( this );
614	if(thrd) goto RET;
615	}
616	thrd = pop_search( this );
617
618	RET:
619	ready_schedule_unlock();
620
621	/* paranoid */ verify( ! __preemption_enabled() );
622	return thrd;
623	}
624
625	static inline bool __must_unpark( thread$ * thrd ) {
626	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
627	switch(old_ticket) {
628	case TICKET_RUNNING:
629	// Wake won the race, the thread will reschedule/rerun itself
630	return false;
631	case TICKET_BLOCKED:
632	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
633	/* paranoid */ verify( thrd->state == Blocked );
634	return true;
635	default:
636	// This makes no sense, something is wrong abort
637	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
638	}
639	}
640
641	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
642	/* paranoid */ verify( ! __preemption_enabled() );
643	/* paranoid */ verify( ready_schedule_islocked());
644
645	if( !thrd ) return;
646
647	if(__must_unpark(thrd)) {
648	// Wake lost the race,
649	__schedule_thread( thrd, hint );
650	}
651
652	/* paranoid */ verify( ready_schedule_islocked());
653	/* paranoid */ verify( ! __preemption_enabled() );
654	}
655
656	void unpark( thread$ * thrd, unpark_hint hint ) {
657	if( !thrd ) return;
658
659	if(__must_unpark(thrd)) {
660	disable_interrupts();
661	// Wake lost the race,
662	schedule_thread$( thrd, hint );
663	enable_interrupts(false);
664	}
665	}
666
667	void park( void ) {
668	__disable_interrupts_checked();
669	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
670	returnToKernel();
671	__enable_interrupts_checked();
672
673	}
674
675	extern "C" {
676	// Leave the thread monitor
677	// last routine called by a thread.
678	// Should never return
679	void __cfactx_thrd_leave() {
680	thread$ * thrd = active_thread();
681	monitor$ * this = &thrd->self_mon;
682
683	// Lock the monitor now
684	lock( this->lock __cfaabi_dbg_ctx2 );
685
686	disable_interrupts();
687
688	/* paranoid */ verify( ! __preemption_enabled() );
689	/* paranoid */ verify( thrd->state == Active );
690	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
691	/* paranoid */ verify( kernelTLS().this_thread == thrd );
692	/* paranoid */ verify( thrd->context.SP );
693	/* 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 );
694	/* 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 );
695
696	thrd->state = Halting;
697	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
698	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
699	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
700
701	// Leave the thread
702	returnToKernel();
703
704	// Control flow should never reach here!
705	abort();
706	}
707	}
708
709	// KERNEL ONLY
710	bool force_yield( __Preemption_Reason reason ) {
711	__disable_interrupts_checked();
712	thread$ * thrd = kernelTLS().this_thread;
713	/* paranoid */ verify(thrd->state == Active);
714
715	// SKULLDUGGERY: It is possible that we are preempting this thread just before
716	// it was going to park itself. If that is the case and it is already using the
717	// intrusive fields then we can't use them to preempt the thread
718	// If that is the case, abandon the preemption.
719	bool preempted = false;
720	if(thrd->link.next == 0p) {
721	preempted = true;
722	thrd->preempted = reason;
723	returnToKernel();
724	}
725	__enable_interrupts_checked( false );
726	return preempted;
727	}
728
729	//=============================================================================================
730	// Kernel Idle Sleep
731	//=============================================================================================
732	// Wake a thread from the front if there are any
733	static void __wake_one(cluster * this) {
734	/* paranoid */ verify( ! __preemption_enabled() );
735	/* paranoid */ verify( ready_schedule_islocked() );
736
737	// Check if there is a sleeping processor
738	int fd = __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST);
739
740	// If no one is sleeping, we are done
741	if( fd == 0 ) return;
742
743	// We found a processor, wake it up
744	eventfd_t val;
745	val = 1;
746	eventfd_write( fd, val );
747
748	#if !defined(__CFA_NO_STATISTICS__)
749	if( kernelTLS().this_stats ) {
750	__tls_stats()->ready.sleep.wakes++;
751	}
752	else {
753	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
754	}
755	#endif
756
757	/* paranoid */ verify( ready_schedule_islocked() );
758	/* paranoid */ verify( ! __preemption_enabled() );
759
760	return;
761	}
762
763	// Unconditionnaly wake a thread
764	void __wake_proc(processor * this) {
765	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
766
767	__disable_interrupts_checked();
768	/* paranoid */ verify( ! __preemption_enabled() );
769	eventfd_t val;
770	val = 1;
771	eventfd_write( this->idle_fd, val );
772	__enable_interrupts_checked();
773	}
774
775	static void idle_sleep(processor * this) {
776	#if !defined(__CFA_NO_STATISTICS__)
777	if(this->print_halts) {
778	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
779	}
780	#endif
781
782	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
783
784	{
785	eventfd_t val;
786	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
787	if(ret < 0) {
788	switch((int)errno) {
789	case EAGAIN:
790	#if EAGAIN != EWOULDBLOCK
791	case EWOULDBLOCK:
792	#endif
793	case EINTR:
794	// No need to do anything special here, just assume it's a legitimate wake-up
795	break;
796	default:
797	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
798	}
799	}
800	}
801
802	#if !defined(__CFA_NO_STATISTICS__)
803	if(this->print_halts) {
804	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
805	}
806	#endif
807	}
808
809	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
810	/* paranoid */ verify( ! __preemption_enabled() );
811	if(!try_lock( this )) return false;
812	this.idle++;
813	/* paranoid */ verify( this.idle <= this.total );
814	remove(proc);
815	insert_first(this.idles, proc);
816
817	__atomic_store_n(&this.fd, proc.idle_fd, __ATOMIC_SEQ_CST);
818	unlock( this );
819	/* paranoid */ verify( ! __preemption_enabled() );
820
821	return true;
822	}
823
824	static void mark_awake(__cluster_proc_list & this, processor & proc) {
825	/* paranoid */ verify( ! __preemption_enabled() );
826	lock( this );
827	this.idle--;
828	/* paranoid */ verify( this.idle >= 0 );
829	remove(proc);
830	insert_last(this.actives, proc);
831
832	{
833	int fd = 0;
834	if(!this.idles`isEmpty) fd = this.idles`first.idle_fd;
835	__atomic_store_n(&this.fd, fd, __ATOMIC_SEQ_CST);
836	}
837
838	unlock( this );
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	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
942	__tally_stats( cltr->stats, it->local_data->this_stats );
943	it = &(*it)`next;
944	}
945	}
946
947	void crawl_cluster_stats( cluster & this ) {
948	// Stop the world, otherwise stats could get really messed-up
949	// this doesn't solve all problems but does solve many
950	// so it's probably good enough
951	disable_interrupts();
952	uint_fast32_t last_size = ready_mutate_lock();
953
954	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
955	crawl_list(&this, this.procs.idles , this.procs.idle );
956
957	// Unlock the RWlock
958	ready_mutate_unlock( last_size );
959	enable_interrupts();
960	}
961
962
963	void print_stats_now( cluster & this, int flags ) {
964	crawl_cluster_stats( this );
965	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
966	}
967	#endif
968	// Local Variables: //
969	// mode: c //
970	// tab-width: 4 //
971	// End: //

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

Download in other formats:

Original Format