Context Navigation

kernel.cfa @ 89eff25

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

Last change on this file since 89eff25 was 89eff25, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Some changes to stats and added back preferred
Property mode set to `100644`
File size: 29.0 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 __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 void __run_thread(processor * this, $thread * dst);
118	static void __wake_one(cluster * cltr);
119
120	static void mark_idle (__cluster_proc_list & idles, processor & proc);
121	static void mark_awake(__cluster_proc_list & idles, processor & proc);
122	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
123
124	extern void __cfa_io_start( processor * );
125	extern bool __cfa_io_drain( processor * );
126	extern void __cfa_io_flush( processor * );
127	extern void __cfa_io_stop ( processor * );
128	static inline bool __maybe_io_drain( processor * );
129
130	extern void __disable_interrupts_hard();
131	extern void __enable_interrupts_hard();
132
133	//=============================================================================================
134	// Kernel Scheduling logic
135	//=============================================================================================
136	//Main of the processor contexts
137	void main(processorCtx_t & runner) {
138	// Because of a bug, we couldn't initialized the seed on construction
139	// Do it here
140	__cfaabi_tls.rand_seed ^= rdtscl();
141	__cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
142	__tls_rand_advance_bck();
143
144	processor * this = runner.proc;
145	verify(this);
146
147	__cfa_io_start( this );
148
149	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
150	#if !defined(__CFA_NO_STATISTICS__)
151	if( this->print_halts ) {
152	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->id, this->name, (void*)this);
153	}
154	#endif
155
156	{
157	// Setup preemption data
158	preemption_scope scope = { this };
159
160	__STATS( unsigned long long last_tally = rdtscl(); )
161
162	// if we need to run some special setup, now is the time to do it.
163	if(this->init.thrd) {
164	this->init.thrd->curr_cluster = this->cltr;
165	__run_thread(this, this->init.thrd);
166	}
167
168	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
169
170	$thread * readyThread = 0p;
171	MAIN_LOOP:
172	for() {
173	// Check if there is pending io
174	__maybe_io_drain( this );
175
176	// Try to get the next thread
177	readyThread = __next_thread( this->cltr );
178
179	if( !readyThread ) {
180	__cfa_io_flush( this );
181	readyThread = __next_thread_slow( this->cltr );
182	}
183
184	HALT:
185	if( !readyThread ) {
186	// Don't block if we are done
187	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
188
189	#if !defined(__CFA_NO_STATISTICS__)
190	__tls_stats()->ready.sleep.halts++;
191	#endif
192
193	// Push self to idle stack
194	mark_idle(this->cltr->procs, * this);
195
196	// Confirm the ready-queue is empty
197	readyThread = __next_thread_slow( this->cltr );
198	if( readyThread ) {
199	// A thread was found, cancel the halt
200	mark_awake(this->cltr->procs, * this);
201
202	#if !defined(__CFA_NO_STATISTICS__)
203	__tls_stats()->ready.sleep.cancels++;
204	#endif
205
206	// continue the mai loop
207	break HALT;
208	}
209
210	#if !defined(__CFA_NO_STATISTICS__)
211	if(this->print_halts) {
212	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
213	}
214	#endif
215
216	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
217
218	__disable_interrupts_hard();
219	eventfd_t val;
220	eventfd_read( this->idle, &val );
221	__enable_interrupts_hard();
222
223	#if !defined(__CFA_NO_STATISTICS__)
224	if(this->print_halts) {
225	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
226	}
227	#endif
228
229	// We were woken up, remove self from idle
230	mark_awake(this->cltr->procs, * this);
231
232	// DON'T just proceed, start looking again
233	continue MAIN_LOOP;
234	}
235
236	/* paranoid */ verify( readyThread );
237
238	// Reset io dirty bit
239	this->io.dirty = false;
240
241	// We found a thread run it
242	__run_thread(this, readyThread);
243
244	// Are we done?
245	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
246
247	#if !defined(__CFA_NO_STATISTICS__)
248	unsigned long long curr = rdtscl();
249	if(curr > (last_tally + 500000000)) {
250	__tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
251	last_tally = curr;
252	}
253	#endif
254
255	if(this->io.pending && !this->io.dirty) {
256	__cfa_io_flush( this );
257	}
258
259	// SEARCH: {
260	// /* paranoid */ verify( ! __preemption_enabled() );
261	// /* paranoid */ verify( kernelTLS().this_proc_id );
262
263	// // First, lock the scheduler since we are searching for a thread
264
265	// // Try to get the next thread
266	// ready_schedule_lock();
267	// readyThread = pop_fast( this->cltr );
268	// ready_schedule_unlock();
269	// if(readyThread) { break SEARCH; }
270
271	// // If we can't find a thread, might as well flush any outstanding I/O
272	// if(this->io.pending) { __cfa_io_flush( this ); }
273
274	// // Spin a little on I/O, just in case
275	// for(25) {
276	// __maybe_io_drain( this );
277	// ready_schedule_lock();
278	// readyThread = pop_fast( this->cltr );
279	// ready_schedule_unlock();
280	// if(readyThread) { break SEARCH; }
281	// }
282
283	// // no luck, try stealing a few times
284	// for(25) {
285	// if( __maybe_io_drain( this ) ) {
286	// ready_schedule_lock();
287	// readyThread = pop_fast( this->cltr );
288	// } else {
289	// ready_schedule_lock();
290	// readyThread = pop_slow( this->cltr );
291	// }
292	// ready_schedule_unlock();
293	// if(readyThread) { break SEARCH; }
294	// }
295
296	// // still no luck, search for a thread
297	// ready_schedule_lock();
298	// readyThread = pop_search( this->cltr );
299	// ready_schedule_unlock();
300	// if(readyThread) { break SEARCH; }
301
302	// // Don't block if we are done
303	// if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
304
305	// __STATS( __tls_stats()->ready.sleep.halts++; )
306
307	// // Push self to idle stack
308	// mark_idle(this->cltr->procs, * this);
309
310	// // Confirm the ready-queue is empty
311	// __maybe_io_drain( this );
312	// ready_schedule_lock();
313	// readyThread = pop_search( this->cltr );
314	// ready_schedule_unlock();
315
316	// if( readyThread ) {
317	// // A thread was found, cancel the halt
318	// mark_awake(this->cltr->procs, * this);
319
320	// __STATS( __tls_stats()->ready.sleep.cancels++; )
321
322	// // continue the main loop
323	// break SEARCH;
324	// }
325
326	// __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl()); )
327	// __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
328
329	// // __disable_interrupts_hard();
330	// eventfd_t val;
331	// eventfd_read( this->idle, &val );
332	// // __enable_interrupts_hard();
333
334	// __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl()); )
335
336	// // We were woken up, remove self from idle
337	// mark_awake(this->cltr->procs, * this);
338
339	// // DON'T just proceed, start looking again
340	// continue MAIN_LOOP;
341	// }
342
343	// RUN_THREAD:
344	// /* paranoid */ verify( kernelTLS().this_proc_id );
345	// /* paranoid */ verify( ! __preemption_enabled() );
346	// /* paranoid */ verify( readyThread );
347
348	// // Reset io dirty bit
349	// this->io.dirty = false;
350
351	// // We found a thread run it
352	// __run_thread(this, readyThread);
353
354	// // Are we done?
355	// if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
356
357	// #if !defined(__CFA_NO_STATISTICS__)
358	// unsigned long long curr = rdtscl();
359	// if(curr > (last_tally + 500000000)) {
360	// __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
361	// last_tally = curr;
362	// }
363	// #endif
364
365	// if(this->io.pending && !this->io.dirty) {
366	// __cfa_io_flush( this );
367	// }
368
369	// // Check if there is pending io
370	// __maybe_io_drain( this );
371	}
372
373	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
374	}
375
376	__cfa_io_stop( this );
377
378	post( this->terminated );
379
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, "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, "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	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
435	/* 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 );
436	/* 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 );
437	/* paranoid */ verify( thrd_dst->context.SP );
438	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
439	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
440	/* paranoid */ verify( ! __preemption_enabled() );
441
442	// Reset global state
443	kernelTLS().this_thread = 0p;
444
445	// We just finished running a thread, there are a few things that could have happened.
446	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
447	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
448	// 4 - Preempted
449	// In case 1, we may have won a race so we can't write to the state again.
450	// In case 2, we lost the race so we now own the thread.
451
452	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
453	// The thread was preempted, reschedule it and reset the flag
454	__schedule_thread( thrd_dst );
455	break RUNNING;
456	}
457
458	if(unlikely(thrd_dst->state == Halting)) {
459	// The thread has halted, it should never be scheduled/run again
460	// finish the thread
461	__thread_finish( thrd_dst );
462	break RUNNING;
463	}
464
465	/* paranoid */ verify( thrd_dst->state == Active );
466	thrd_dst->state = Blocked;
467
468	// set state of processor coroutine to active and the thread to inactive
469	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
470	switch(old_ticket) {
471	case TICKET_RUNNING:
472	// This is case 1, the regular case, nothing more is needed
473	break RUNNING;
474	case TICKET_UNBLOCK:
475	#if !defined(__CFA_NO_STATISTICS__)
476	__tls_stats()->ready.threads.threads++;
477	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
478	#endif
479	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
480	// In this case, just run it again.
481	continue RUNNING;
482	default:
483	// This makes no sense, something is wrong abort
484	abort();
485	}
486	}
487
488	// Just before returning to the processor, set the processor coroutine to active
489	proc_cor->state = Active;
490
491	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
492
493	#if !defined(__CFA_NO_STATISTICS__)
494	__tls_stats()->ready.threads.threads--;
495	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
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 ), "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), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
535	}
536
537	//-----------------------------------------------------------------------------
538	// Scheduler routines
539	// KERNEL ONLY
540	void __schedule_thread( $thread * thrd ) {
541	/* paranoid */ verify( ! __preemption_enabled() );
542	/* paranoid */ verify( kernelTLS().this_proc_id );
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
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 = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
561
562	ready_schedule_lock();
563	// push the thread to the cluster ready-queue
564	push( cl, thrd );
565
566	// variable thrd is no longer safe to use
567
568	// wake the cluster using the save variable.
569	__wake_one( cl );
570	ready_schedule_unlock();
571
572	#if !defined(__CFA_NO_STATISTICS__)
573	if( kernelTLS().this_stats ) {
574	__tls_stats()->ready.threads.threads++;
575	if(outside) {
576	__tls_stats()->ready.threads.extunpark++;
577	}
578	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
579	}
580	else {
581	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
582	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
583	__push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
584	}
585	#endif
586
587	/* paranoid */ verify( ! __preemption_enabled() );
588	}
589
590	// KERNEL ONLY
591	static inline $thread * __next_thread(cluster * this) with( *this ) {
592	/* paranoid */ verify( ! __preemption_enabled() );
593	/* paranoid */ verify( kernelTLS().this_proc_id );
594
595	ready_schedule_lock();
596	$thread * thrd = pop_fast( this );
597	ready_schedule_unlock();
598
599	/* paranoid */ verify( kernelTLS().this_proc_id );
600	/* paranoid */ verify( ! __preemption_enabled() );
601	return thrd;
602	}
603
604	// KERNEL ONLY
605	static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
606	/* paranoid */ verify( ! __preemption_enabled() );
607	/* paranoid */ verify( kernelTLS().this_proc_id );
608
609	ready_schedule_lock();
610	$thread * thrd;
611	for(25) {
612	thrd = pop_slow( this );
613	if(thrd) goto RET;
614	}
615	thrd = pop_search( this );
616
617	RET:
618	ready_schedule_unlock();
619
620	/* paranoid */ verify( kernelTLS().this_proc_id );
621	/* paranoid */ verify( ! __preemption_enabled() );
622	return thrd;
623	}
624
625	void unpark( $thread * thrd ) {
626	if( !thrd ) return;
627
628	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
629	switch(old_ticket) {
630	case TICKET_RUNNING:
631	// Wake won the race, the thread will reschedule/rerun itself
632	break;
633	case TICKET_BLOCKED:
634	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
635	/* paranoid */ verify( thrd->state == Blocked );
636
637	{
638	/* paranoid */ verify( publicTLS_get(this_proc_id) );
639	disable_interrupts();
640
641	/* paranoid */ verify( ! __preemption_enabled() );
642
643	// Wake lost the race,
644	__schedule_thread( thrd );
645
646	/* paranoid */ verify( ! __preemption_enabled() );
647
648	enable_interrupts_noPoll();
649	/* paranoid */ verify( publicTLS_get(this_proc_id) );
650	}
651
652	break;
653	default:
654	// This makes no sense, something is wrong abort
655	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
656	}
657	}
658
659	void park( void ) {
660	/* paranoid */ verify( __preemption_enabled() );
661	disable_interrupts();
662	/* paranoid */ verify( ! __preemption_enabled() );
663	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
664
665	returnToKernel();
666
667	/* paranoid */ verify( ! __preemption_enabled() );
668	enable_interrupts( __cfaabi_dbg_ctx );
669	/* paranoid */ verify( __preemption_enabled() );
670
671	}
672
673	extern "C" {
674	// Leave the thread monitor
675	// last routine called by a thread.
676	// Should never return
677	void __cfactx_thrd_leave() {
678	$thread * thrd = active_thread();
679	$monitor * this = &thrd->self_mon;
680
681	// Lock the monitor now
682	lock( this->lock __cfaabi_dbg_ctx2 );
683
684	disable_interrupts();
685
686	/* paranoid */ verify( ! __preemption_enabled() );
687	/* paranoid */ verify( thrd->state == Active );
688	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
689	/* paranoid */ verify( kernelTLS().this_thread == thrd );
690	/* paranoid */ verify( thrd->context.SP );
691	/* 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 );
692	/* 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 );
693
694	thrd->state = Halting;
695	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
696	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
697	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
698
699	// Leave the thread
700	returnToKernel();
701
702	// Control flow should never reach here!
703	abort();
704	}
705	}
706
707	// KERNEL ONLY
708	bool force_yield( __Preemption_Reason reason ) {
709	/* paranoid */ verify( __preemption_enabled() );
710	disable_interrupts();
711	/* paranoid */ verify( ! __preemption_enabled() );
712
713	$thread * thrd = kernelTLS().this_thread;
714	/* paranoid */ verify(thrd->state == Active);
715
716	// SKULLDUGGERY: It is possible that we are preempting this thread just before
717	// it was going to park itself. If that is the case and it is already using the
718	// intrusive fields then we can't use them to preempt the thread
719	// If that is the case, abandon the preemption.
720	bool preempted = false;
721	if(thrd->link.next == 0p) {
722	preempted = true;
723	thrd->preempted = reason;
724	returnToKernel();
725	}
726
727	/* paranoid */ verify( ! __preemption_enabled() );
728	enable_interrupts_noPoll();
729	/* paranoid */ verify( __preemption_enabled() );
730
731	return preempted;
732	}
733
734	//=============================================================================================
735	// Kernel Idle Sleep
736	//=============================================================================================
737	// Wake a thread from the front if there are any
738	static void __wake_one(cluster * this) {
739	/* paranoid */ verify( ! __preemption_enabled() );
740	/* paranoid */ verify( ready_schedule_islocked() );
741
742	// Check if there is a sleeping processor
743	processor * p;
744	unsigned idle;
745	unsigned total;
746	[idle, total, p] = query_idles(this->procs);
747
748	// If no one is sleeping, we are done
749	if( idle == 0 ) return;
750
751	// We found a processor, wake it up
752	eventfd_t val;
753	val = 1;
754	eventfd_write( p->idle, val );
755
756	#if !defined(__CFA_NO_STATISTICS__)
757	if( kernelTLS().this_stats ) {
758	__tls_stats()->ready.sleep.wakes++;
759	}
760	else {
761	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
762	}
763	#endif
764
765	/* paranoid */ verify( ready_schedule_islocked() );
766	/* paranoid */ verify( ! __preemption_enabled() );
767
768	return;
769	}
770
771	// Unconditionnaly wake a thread
772	void __wake_proc(processor * this) {
773	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
774
775	disable_interrupts();
776	/* paranoid */ verify( ! __preemption_enabled() );
777	eventfd_t val;
778	val = 1;
779	eventfd_write( this->idle, val );
780	enable_interrupts( __cfaabi_dbg_ctx );
781	}
782
783	static void mark_idle(__cluster_proc_list & this, processor & proc) {
784	/* paranoid */ verify( ! __preemption_enabled() );
785	lock( this );
786	this.idle++;
787	/* paranoid */ verify( this.idle <= this.total );
788	remove(proc);
789	insert_first(this.idles, proc);
790	unlock( this );
791	/* paranoid */ verify( ! __preemption_enabled() );
792	}
793
794	static void mark_awake(__cluster_proc_list & this, processor & proc) {
795	/* paranoid */ verify( ! __preemption_enabled() );
796	lock( this );
797	this.idle--;
798	/* paranoid */ verify( this.idle >= 0 );
799	remove(proc);
800	insert_last(this.actives, proc);
801	unlock( this );
802	/* paranoid */ verify( ! __preemption_enabled() );
803	}
804
805	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
806	/* paranoid */ verify( ! __preemption_enabled() );
807	/* paranoid */ verify( ready_schedule_islocked() );
808
809	for() {
810	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
811	if( 1 == (l % 2) ) { Pause(); continue; }
812	unsigned idle = this.idle;
813	unsigned total = this.total;
814	processor * proc = &this.idles`first;
815	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
816	asm volatile("": : :"memory");
817	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
818	return [idle, total, proc];
819	}
820
821	/* paranoid */ verify( ready_schedule_islocked() );
822	/* paranoid */ verify( ! __preemption_enabled() );
823	}
824
825	//=============================================================================================
826	// Unexpected Terminating logic
827	//=============================================================================================
828	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
829	$thread * thrd = __cfaabi_tls.this_thread;
830
831	if(thrd) {
832	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
833	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
834
835	if ( &thrd->self_cor != thrd->curr_cor ) {
836	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
837	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
838	}
839	else {
840	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
841	}
842	}
843	else {
844	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
845	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
846	}
847	}
848
849	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
850	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
851	}
852
853	static __spinlock_t kernel_debug_lock;
854
855	extern "C" {
856	void __cfaabi_bits_acquire() {
857	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
858	}
859
860	void __cfaabi_bits_release() {
861	unlock( kernel_debug_lock );
862	}
863	}
864
865	//=============================================================================================
866	// Kernel Utilities
867	//=============================================================================================
868	#if defined(CFA_HAVE_LINUX_IO_URING_H)
869	#include "io/types.hfa"
870	#endif
871
872	static inline bool __maybe_io_drain( processor * proc ) {
873	#if defined(CFA_HAVE_LINUX_IO_URING_H)
874	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
875
876	// Check if we should drain the queue
877	$io_context * ctx = proc->io.ctx;
878	unsigned head = *ctx->cq.head;
879	unsigned tail = *ctx->cq.tail;
880	if(head == tail) return false;
881	return __cfa_io_drain( proc );
882	#endif
883	}
884
885	//-----------------------------------------------------------------------------
886	// Debug
887	__cfaabi_dbg_debug_do(
888	extern "C" {
889	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
890	this.prev_name = prev_name;
891	this.prev_thrd = kernelTLS().this_thread;
892	}
893	}
894	)
895
896	//-----------------------------------------------------------------------------
897	// Debug
898	bool threading_enabled(void) __attribute__((const)) {
899	return true;
900	}
901
902	//-----------------------------------------------------------------------------
903	// Statistics
904	#if !defined(__CFA_NO_STATISTICS__)
905	void print_halts( processor & this ) {
906	this.print_halts = true;
907	}
908
909	void print_stats_now( cluster & this, int flags ) {
910	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
911	}
912	#endif
913	// Local Variables: //
914	// mode: c //
915	// tab-width: 4 //
916	// End: //

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

Download in other formats:

Original Format