Context Navigation

kernel.cfa @ fb4ccdf

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

Last change on this file since fb4ccdf was fb4ccdf, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Updated alternate main for processors
Property mode set to `100644`
File size: 29.6 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// kernel.c --
8	//
9	// Author : Thierry Delisle
10	// Created On : Tue Jan 17 12:27:26 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 31 07:08:20 2020
13	// Update Count : 71
14	//
15
16	#define __cforall_thread__
17	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
18
19	//C Includes
20	#include <errno.h>
21	#include <stdio.h>
22	#include <signal.h>
23	#include <unistd.h>
24	extern "C" {
25	#include <sys/eventfd.h>
26	}
27
28	//CFA Includes
29	#include "kernel_private.hfa"
30	#include "preemption.hfa"
31
32	//Private includes
33	#define __CFA_INVOKE_PRIVATE__
34	#include "invoke.h"
35
36	#if !defined(__CFA_NO_STATISTICS__)
37	#define __STATS( ...) __VA_ARGS__
38	#else
39	#define __STATS( ...)
40	#endif
41
42	//-----------------------------------------------------------------------------
43	// Some assembly required
44	#if defined( __i386 )
45	// mxcr : SSE Status and Control bits (control bits are preserved across function calls)
46	// fcw : X87 FPU control word (preserved across function calls)
47	#define __x87_store \
48	uint32_t __mxcr; \
49	uint16_t __fcw; \
50	__asm__ volatile ( \
51	"stmxcsr %0\n" \
52	"fnstcw %1\n" \
53	: "=m" (__mxcr),\
54	"=m" (__fcw) \
55	)
56
57	#define __x87_load \
58	__asm__ volatile ( \
59	"fldcw %1\n" \
60	"ldmxcsr %0\n" \
61	::"m" (__mxcr),\
62	"m" (__fcw) \
63	)
64
65	#elif defined( __x86_64 )
66	#define __x87_store \
67	uint32_t __mxcr; \
68	uint16_t __fcw; \
69	__asm__ volatile ( \
70	"stmxcsr %0\n" \
71	"fnstcw %1\n" \
72	: "=m" (__mxcr),\
73	"=m" (__fcw) \
74	)
75
76	#define __x87_load \
77	__asm__ volatile ( \
78	"fldcw %1\n" \
79	"ldmxcsr %0\n" \
80	:: "m" (__mxcr),\
81	"m" (__fcw) \
82	)
83
84	#elif defined( __arm__ )
85	#define __x87_store
86	#define __x87_load
87
88	#elif defined( __aarch64__ )
89	#define __x87_store \
90	uint32_t __fpcntl[2]; \
91	__asm__ volatile ( \
92	"mrs x9, FPCR\n" \
93	"mrs x10, FPSR\n" \
94	"stp x9, x10, %0\n" \
95	: "=m" (__fpcntl) : : "x9", "x10" \
96	)
97
98	#define __x87_load \
99	__asm__ volatile ( \
100	"ldp x9, x10, %0\n" \
101	"msr FPSR, x10\n" \
102	"msr FPCR, x9\n" \
103	: "=m" (__fpcntl) : : "x9", "x10" \
104	)
105
106	#else
107	#error unsupported hardware architecture
108	#endif
109
110	extern $thread * mainThread;
111	extern processor * mainProcessor;
112
113	//-----------------------------------------------------------------------------
114	// Kernel Scheduling logic
115	static $thread * __next_thread(cluster * this);
116	static $thread * __next_thread_slow(cluster * this);
117	static inline bool __must_unpark( $thread * thrd ) __attribute((nonnull(1)));
118	static void __run_thread(processor * this, $thread * dst);
119	static void __wake_one(cluster * cltr);
120
121	static void mark_idle (__cluster_proc_list & idles, processor & proc);
122	static void mark_awake(__cluster_proc_list & idles, processor & proc);
123	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
124
125	extern void __cfa_io_start( processor * );
126	extern bool __cfa_io_drain( processor * );
127	extern void __cfa_io_flush( processor * );
128	extern void __cfa_io_stop ( processor * );
129	static inline bool __maybe_io_drain( processor * );
130
131	extern void __disable_interrupts_hard();
132	extern void __enable_interrupts_hard();
133
134	static inline void __disable_interrupts_checked() {
135	/* paranoid */ verify( __preemption_enabled() );
136	disable_interrupts();
137	/* paranoid */ verify( ! __preemption_enabled() );
138	}
139
140	static inline void __enable_interrupts_checked( bool poll = true ) {
141	/* paranoid */ verify( ! __preemption_enabled() );
142	enable_interrupts( poll );
143	/* paranoid */ verify( __preemption_enabled() );
144	}
145
146	//=============================================================================================
147	// Kernel Scheduling logic
148	//=============================================================================================
149	//Main of the processor contexts
150	void main(processorCtx_t & runner) {
151	// Because of a bug, we couldn't initialized the seed on construction
152	// Do it here
153	__cfaabi_tls.rand_seed ^= rdtscl();
154	__cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
155	__tls_rand_advance_bck();
156
157	processor * this = runner.proc;
158	verify(this);
159
160	__cfa_io_start( this );
161
162	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
163	#if !defined(__CFA_NO_STATISTICS__)
164	if( this->print_halts ) {
165	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
166	}
167	#endif
168
169	{
170	// Setup preemption data
171	preemption_scope scope = { this };
172
173	// if we need to run some special setup, now is the time to do it.
174	if(this->init.thrd) {
175	this->init.thrd->curr_cluster = this->cltr;
176	__run_thread(this, this->init.thrd);
177	}
178
179	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
180
181	$thread * readyThread = 0p;
182	MAIN_LOOP:
183	for() {
184	#if 1
185	// Check if there is pending io
186	__maybe_io_drain( this );
187
188	// Try to get the next thread
189	readyThread = __next_thread( this->cltr );
190
191	if( !readyThread ) {
192	__cfa_io_flush( this );
193	readyThread = __next_thread_slow( this->cltr );
194	}
195
196	HALT:
197	if( !readyThread ) {
198	// Don't block if we are done
199	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
200
201	#if !defined(__CFA_NO_STATISTICS__)
202	__tls_stats()->ready.sleep.halts++;
203	#endif
204
205	// Push self to idle stack
206	mark_idle(this->cltr->procs, * this);
207
208	// Confirm the ready-queue is empty
209	readyThread = __next_thread_slow( this->cltr );
210	if( readyThread ) {
211	// A thread was found, cancel the halt
212	mark_awake(this->cltr->procs, * this);
213
214	#if !defined(__CFA_NO_STATISTICS__)
215	__tls_stats()->ready.sleep.cancels++;
216	#endif
217
218	// continue the mai loop
219	break HALT;
220	}
221
222	#if !defined(__CFA_NO_STATISTICS__)
223	if(this->print_halts) {
224	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
225	}
226	#endif
227
228	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
229
230	__disable_interrupts_hard();
231	eventfd_t val;
232	eventfd_read( this->idle, &val );
233	__enable_interrupts_hard();
234
235	#if !defined(__CFA_NO_STATISTICS__)
236	if(this->print_halts) {
237	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
238	}
239	#endif
240
241	// We were woken up, remove self from idle
242	mark_awake(this->cltr->procs, * this);
243
244	// DON'T just proceed, start looking again
245	continue MAIN_LOOP;
246	}
247
248	/* paranoid */ verify( readyThread );
249
250	// Reset io dirty bit
251	this->io.dirty = false;
252
253	// We found a thread run it
254	__run_thread(this, readyThread);
255
256	// Are we done?
257	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
258
259	if(this->io.pending && !this->io.dirty) {
260	__cfa_io_flush( this );
261	}
262
263	#else
264
265	SEARCH: {
266	/* paranoid */ verify( ! __preemption_enabled() );
267
268	// First, lock the scheduler since we are searching for a thread
269	ready_schedule_lock();
270
271	// Try to get the next thread
272	readyThread = pop_fast( this->cltr );
273	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
274
275	// If we can't find a thread, might as well flush any outstanding I/O
276	if(this->io.pending) { __cfa_io_flush( this ); }
277
278	// Spin a little on I/O, just in case
279	for(25) {
280	__maybe_io_drain( this );
281	readyThread = pop_fast( this->cltr );
282	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
283	}
284
285	// no luck, try stealing a few times
286	for(25) {
287	if( __maybe_io_drain( this ) ) {
288	readyThread = pop_fast( this->cltr );
289	} else {
290	readyThread = pop_slow( this->cltr );
291	}
292	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
293	}
294
295	// still no luck, search for a thread
296	readyThread = pop_search( this->cltr );
297	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
298
299	// Don't block if we are done
300	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
301
302	__STATS( __tls_stats()->ready.sleep.halts++; )
303
304	// Push self to idle stack
305	ready_schedule_unlock();
306	mark_idle(this->cltr->procs, * this);
307	ready_schedule_lock();
308
309	// Confirm the ready-queue is empty
310	__maybe_io_drain( this );
311	readyThread = pop_search( this->cltr );
312	ready_schedule_unlock();
313
314	if( readyThread ) {
315	// A thread was found, cancel the halt
316	mark_awake(this->cltr->procs, * this);
317
318	__STATS( __tls_stats()->ready.sleep.cancels++; )
319
320	// continue the main loop
321	break SEARCH;
322	}
323
324	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
325	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
326
327	// __disable_interrupts_hard();
328	eventfd_t val;
329	eventfd_read( this->idle, &val );
330	// __enable_interrupts_hard();
331
332	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
333
334	// We were woken up, remove self from idle
335	mark_awake(this->cltr->procs, * this);
336
337	// DON'T just proceed, start looking again
338	continue MAIN_LOOP;
339	}
340
341	RUN_THREAD:
342	/* paranoid */ verify( ! __preemption_enabled() );
343	/* paranoid */ verify( readyThread );
344
345	// Reset io dirty bit
346	this->io.dirty = false;
347
348	// We found a thread run it
349	__run_thread(this, readyThread);
350
351	// Are we done?
352	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
353
354	#if !defined(__CFA_NO_STATISTICS__)
355	unsigned long long curr = rdtscl();
356	if(curr > (last_tally + 500000000)) {
357	__tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
358	last_tally = curr;
359	}
360	#endif
361
362	if(this->io.pending && !this->io.dirty) {
363	__cfa_io_flush( this );
364	}
365
366	ready_schedule_lock();
367	__maybe_io_drain( this );
368	ready_schedule_unlock();
369	#endif
370	}
371
372	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
373	}
374
375	__cfa_io_stop( this );
376
377	post( this->terminated );
378
379	if(this == mainProcessor) {
380	// HACK : the coroutine context switch expects this_thread to be set
381	// and it make sense for it to be set in all other cases except here
382	// fake it
383	__cfaabi_tls.this_thread = mainThread;
384	}
385
386	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
387	}
388
389	static int * __volatile_errno() __attribute__((noinline));
390	static int * __volatile_errno() { asm(""); return &errno; }
391
392	// KERNEL ONLY
393	// runThread runs a thread by context switching
394	// from the processor coroutine to the target thread
395	static void __run_thread(processor * this, $thread * thrd_dst) {
396	/* paranoid */ verify( ! __preemption_enabled() );
397	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
398	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
399	__builtin_prefetch( thrd_dst->context.SP );
400
401	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
402
403	$coroutine * proc_cor = get_coroutine(this->runner);
404
405	// set state of processor coroutine to inactive
406	verify(proc_cor->state == Active);
407	proc_cor->state = Blocked;
408
409	// Actually run the thread
410	RUNNING: while(true) {
411	thrd_dst->preempted = __NO_PREEMPTION;
412	thrd_dst->state = Active;
413
414	// Update global state
415	kernelTLS().this_thread = thrd_dst;
416
417	/* paranoid */ verify( ! __preemption_enabled() );
418	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
419	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
420	/* paranoid */ verify( thrd_dst->context.SP );
421	/* paranoid */ verify( thrd_dst->state != Halted );
422	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) \|\| thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
423	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) \|\| thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
424	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
425
426
427
428	// set context switch to the thread that the processor is executing
429	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
430	// when __cfactx_switch returns we are back in the processor coroutine
431
432	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
433	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
434	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
435	/* paranoid */ verify( thrd_dst->context.SP );
436	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
437	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
438	/* paranoid */ verify( ! __preemption_enabled() );
439
440	// Reset global state
441	kernelTLS().this_thread = 0p;
442
443	// We just finished running a thread, there are a few things that could have happened.
444	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
445	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
446	// 4 - Preempted
447	// In case 1, we may have won a race so we can't write to the state again.
448	// In case 2, we lost the race so we now own the thread.
449
450	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
451	// The thread was preempted, reschedule it and reset the flag
452	schedule_thread$( thrd_dst );
453	break RUNNING;
454	}
455
456	if(unlikely(thrd_dst->state == Halting)) {
457	// The thread has halted, it should never be scheduled/run again
458	// finish the thread
459	__thread_finish( thrd_dst );
460	break RUNNING;
461	}
462
463	/* paranoid */ verify( thrd_dst->state == Active );
464	thrd_dst->state = Blocked;
465
466	// set state of processor coroutine to active and the thread to inactive
467	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
468	switch(old_ticket) {
469	case TICKET_RUNNING:
470	// This is case 1, the regular case, nothing more is needed
471	break RUNNING;
472	case TICKET_UNBLOCK:
473	#if !defined(__CFA_NO_STATISTICS__)
474	__tls_stats()->ready.threads.threads++;
475	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
476	#endif
477	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
478	// In this case, just run it again.
479	continue RUNNING;
480	default:
481	// This makes no sense, something is wrong abort
482	abort();
483	}
484	}
485
486	// Just before returning to the processor, set the processor coroutine to active
487	proc_cor->state = Active;
488
489	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
490
491	#if !defined(__CFA_NO_STATISTICS__)
492	__tls_stats()->ready.threads.threads--;
493	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
494	#endif
495
496	/* paranoid */ verify( ! __preemption_enabled() );
497	}
498
499	// KERNEL_ONLY
500	void returnToKernel() {
501	/* paranoid */ verify( ! __preemption_enabled() );
502	$coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
503	$thread * thrd_src = kernelTLS().this_thread;
504
505	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
506
507	// Run the thread on this processor
508	{
509	int local_errno = *__volatile_errno();
510	#if defined( __i386 ) \|\| defined( __x86_64 )
511	__x87_store;
512	#endif
513	/* paranoid */ verify( proc_cor->context.SP );
514	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
515	__cfactx_switch( &thrd_src->context, &proc_cor->context );
516	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
517	#if defined( __i386 ) \|\| defined( __x86_64 )
518	__x87_load;
519	#endif
520	*__volatile_errno() = local_errno;
521	}
522
523	#if !defined(__CFA_NO_STATISTICS__)
524	/* paranoid */ verify( thrd_src->last_proc != 0p );
525	if(thrd_src->last_proc != kernelTLS().this_processor) {
526	__tls_stats()->ready.threads.migration++;
527	}
528	#endif
529
530	/* paranoid */ verify( ! __preemption_enabled() );
531	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
532	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
533	}
534
535	//-----------------------------------------------------------------------------
536	// Scheduler routines
537	// KERNEL ONLY
538	static void __schedule_thread( $thread * thrd ) {
539	/* paranoid */ verify( ! __preemption_enabled() );
540	/* paranoid */ verify( ready_schedule_islocked());
541	/* paranoid */ verify( thrd );
542	/* paranoid */ verify( thrd->state != Halted );
543	/* paranoid */ verify( thrd->curr_cluster );
544	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
545	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
546	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
547	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
548	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
549	/* paranoid */ #endif
550	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
551	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
552
553
554	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
555
556	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
557	struct cluster * cl = thrd->curr_cluster;
558	__STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
559
560	// push the thread to the cluster ready-queue
561	push( cl, thrd );
562
563	// variable thrd is no longer safe to use
564	thrd = 0xdeaddeaddeaddeadp;
565
566	// wake the cluster using the save variable.
567	__wake_one( cl );
568
569	#if !defined(__CFA_NO_STATISTICS__)
570	if( kernelTLS().this_stats ) {
571	__tls_stats()->ready.threads.threads++;
572	if(outside) {
573	__tls_stats()->ready.threads.extunpark++;
574	}
575	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
576	}
577	else {
578	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
579	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
580	__push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
581	}
582	#endif
583
584	/* paranoid */ verify( ready_schedule_islocked());
585	/* paranoid */ verify( ! __preemption_enabled() );
586	}
587
588	void schedule_thread$( $thread * thrd ) {
589	ready_schedule_lock();
590	__schedule_thread( thrd );
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 ) {
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 );
650	}
651
652	/* paranoid */ verify( ready_schedule_islocked());
653	/* paranoid */ verify( ! __preemption_enabled() );
654	}
655
656	void unpark( $thread * thrd ) {
657	if( !thrd ) return;
658
659	if(__must_unpark(thrd)) {
660	disable_interrupts();
661	// Wake lost the race,
662	schedule_thread$( thrd );
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	processor * p;
739	unsigned idle;
740	unsigned total;
741	[idle, total, p] = query_idles(this->procs);
742
743	// If no one is sleeping, we are done
744	if( idle == 0 ) return;
745
746	// We found a processor, wake it up
747	eventfd_t val;
748	val = 1;
749	eventfd_write( p->idle, val );
750
751	#if !defined(__CFA_NO_STATISTICS__)
752	if( kernelTLS().this_stats ) {
753	__tls_stats()->ready.sleep.wakes++;
754	}
755	else {
756	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
757	}
758	#endif
759
760	/* paranoid */ verify( ready_schedule_islocked() );
761	/* paranoid */ verify( ! __preemption_enabled() );
762
763	return;
764	}
765
766	// Unconditionnaly wake a thread
767	void __wake_proc(processor * this) {
768	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
769
770	__disable_interrupts_checked();
771	/* paranoid */ verify( ! __preemption_enabled() );
772	eventfd_t val;
773	val = 1;
774	eventfd_write( this->idle, val );
775	__enable_interrupts_checked();
776	}
777
778	static void mark_idle(__cluster_proc_list & this, processor & proc) {
779	/* paranoid */ verify( ! __preemption_enabled() );
780	lock( this );
781	this.idle++;
782	/* paranoid */ verify( this.idle <= this.total );
783	remove(proc);
784	insert_first(this.idles, proc);
785	unlock( this );
786	/* paranoid */ verify( ! __preemption_enabled() );
787	}
788
789	static void mark_awake(__cluster_proc_list & this, processor & proc) {
790	/* paranoid */ verify( ! __preemption_enabled() );
791	lock( this );
792	this.idle--;
793	/* paranoid */ verify( this.idle >= 0 );
794	remove(proc);
795	insert_last(this.actives, proc);
796	unlock( this );
797	/* paranoid */ verify( ! __preemption_enabled() );
798	}
799
800	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
801	/* paranoid */ verify( ! __preemption_enabled() );
802	/* paranoid */ verify( ready_schedule_islocked() );
803
804	for() {
805	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
806	if( 1 == (l % 2) ) { Pause(); continue; }
807	unsigned idle = this.idle;
808	unsigned total = this.total;
809	processor * proc = &this.idles`first;
810	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
811	asm volatile("": : :"memory");
812	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
813	return [idle, total, proc];
814	}
815
816	/* paranoid */ verify( ready_schedule_islocked() );
817	/* paranoid */ verify( ! __preemption_enabled() );
818	}
819
820	//=============================================================================================
821	// Unexpected Terminating logic
822	//=============================================================================================
823	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
824	$thread * thrd = __cfaabi_tls.this_thread;
825
826	if(thrd) {
827	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
828	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
829
830	if ( &thrd->self_cor != thrd->curr_cor ) {
831	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
832	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
833	}
834	else {
835	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
836	}
837	}
838	else {
839	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
840	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
841	}
842	}
843
844	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
845	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
846	}
847
848	static __spinlock_t kernel_debug_lock;
849
850	extern "C" {
851	void __cfaabi_bits_acquire() {
852	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
853	}
854
855	void __cfaabi_bits_release() {
856	unlock( kernel_debug_lock );
857	}
858	}
859
860	//=============================================================================================
861	// Kernel Utilities
862	//=============================================================================================
863	#if defined(CFA_HAVE_LINUX_IO_URING_H)
864	#include "io/types.hfa"
865	#endif
866
867	static inline bool __maybe_io_drain( processor * proc ) {
868	bool ret = false;
869	#if defined(CFA_HAVE_LINUX_IO_URING_H)
870	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
871
872	// Check if we should drain the queue
873	$io_context * ctx = proc->io.ctx;
874	unsigned head = *ctx->cq.head;
875	unsigned tail = *ctx->cq.tail;
876	if(head == tail) return false;
877	ready_schedule_lock();
878	ret = __cfa_io_drain( proc );
879	ready_schedule_unlock();
880	#endif
881	return ret;
882	}
883
884	//-----------------------------------------------------------------------------
885	// Debug
886	__cfaabi_dbg_debug_do(
887	extern "C" {
888	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
889	this.prev_name = prev_name;
890	this.prev_thrd = kernelTLS().this_thread;
891	}
892	}
893	)
894
895	//-----------------------------------------------------------------------------
896	// Debug
897	bool threading_enabled(void) __attribute__((const)) {
898	return true;
899	}
900
901	//-----------------------------------------------------------------------------
902	// Statistics
903	#if !defined(__CFA_NO_STATISTICS__)
904	void print_halts( processor & this ) {
905	this.print_halts = true;
906	}
907
908	static void crawl_list( cluster * cltr, dlist(processor, processor) & list, unsigned count ) {
909	/* paranoid */ verify( cltr->stats );
910
911	processor * it = &list`first;
912	for(unsigned i = 0; i < count; i++) {
913	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
914	/* paranoid */ verify( it->local_data->this_stats );
915	__tally_stats( cltr->stats, it->local_data->this_stats );
916	it = &(*it)`next;
917	}
918	}
919
920	void crawl_cluster_stats( cluster & this ) {
921	// Stop the world, otherwise stats could get really messed-up
922	// this doesn't solve all problems but does solve many
923	// so it's probably good enough
924	uint_fast32_t last_size = ready_mutate_lock();
925
926	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
927	crawl_list(&this, this.procs.idles , this.procs.idle );
928
929	// Unlock the RWlock
930	ready_mutate_unlock( last_size );
931	}
932
933
934	void print_stats_now( cluster & this, int flags ) {
935	crawl_cluster_stats( this );
936	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
937	}
938	#endif
939	// Local Variables: //
940	// mode: c //
941	// tab-width: 4 //
942	// End: //

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

Download in other formats:

Original Format