Context Navigation

kernel.cfa @ cfff639

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

Last change on this file since cfff639 was cfff639, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Merge branch 'master' of plg.uwaterloo.ca:software/cfa/cfa-cc
Property mode set to `100644`
File size: 29.2 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->id, this->name, (void*)this);
166	}
167	#endif
168
169	{
170	// Setup preemption data
171	preemption_scope scope = { this };
172
173	__STATS( unsigned long long last_tally = rdtscl(); )
174
175	// if we need to run some special setup, now is the time to do it.
176	if(this->init.thrd) {
177	this->init.thrd->curr_cluster = this->cltr;
178	__run_thread(this, this->init.thrd);
179	}
180
181	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
182
183	$thread * readyThread = 0p;
184	MAIN_LOOP:
185	for() {
186	// Check if there is pending io
187	__maybe_io_drain( this );
188
189	// Try to get the next thread
190	readyThread = __next_thread( this->cltr );
191
192	if( !readyThread ) {
193	__cfa_io_flush( this );
194	readyThread = __next_thread_slow( this->cltr );
195	}
196
197	HALT:
198	if( !readyThread ) {
199	// Don't block if we are done
200	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
201
202	#if !defined(__CFA_NO_STATISTICS__)
203	__tls_stats()->ready.sleep.halts++;
204	#endif
205
206	// Push self to idle stack
207	mark_idle(this->cltr->procs, * this);
208
209	// Confirm the ready-queue is empty
210	readyThread = __next_thread_slow( this->cltr );
211	if( readyThread ) {
212	// A thread was found, cancel the halt
213	mark_awake(this->cltr->procs, * this);
214
215	#if !defined(__CFA_NO_STATISTICS__)
216	__tls_stats()->ready.sleep.cancels++;
217	#endif
218
219	// continue the mai loop
220	break HALT;
221	}
222
223	#if !defined(__CFA_NO_STATISTICS__)
224	if(this->print_halts) {
225	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
226	}
227	#endif
228
229	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
230
231	__disable_interrupts_hard();
232	eventfd_t val;
233	eventfd_read( this->idle, &val );
234	__enable_interrupts_hard();
235
236	#if !defined(__CFA_NO_STATISTICS__)
237	if(this->print_halts) {
238	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
239	}
240	#endif
241
242	// We were woken up, remove self from idle
243	mark_awake(this->cltr->procs, * this);
244
245	// DON'T just proceed, start looking again
246	continue MAIN_LOOP;
247	}
248
249	/* paranoid */ verify( readyThread );
250
251	// Reset io dirty bit
252	this->io.dirty = false;
253
254	// We found a thread run it
255	__run_thread(this, readyThread);
256
257	// Are we done?
258	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
259
260	#if !defined(__CFA_NO_STATISTICS__)
261	unsigned long long curr = rdtscl();
262	if(curr > (last_tally + 500000000)) {
263	__tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
264	last_tally = curr;
265	}
266	#endif
267
268	if(this->io.pending && !this->io.dirty) {
269	__cfa_io_flush( this );
270	}
271
272	// SEARCH: {
273	// /* paranoid */ verify( ! __preemption_enabled() );
274	// /* paranoid */ verify( kernelTLS().this_proc_id );
275
276	// // First, lock the scheduler since we are searching for a thread
277
278	// // Try to get the next thread
279	// ready_schedule_lock();
280	// readyThread = pop_fast( this->cltr );
281	// ready_schedule_unlock();
282	// if(readyThread) { break SEARCH; }
283
284	// // If we can't find a thread, might as well flush any outstanding I/O
285	// if(this->io.pending) { __cfa_io_flush( this ); }
286
287	// // Spin a little on I/O, just in case
288	// for(25) {
289	// __maybe_io_drain( this );
290	// ready_schedule_lock();
291	// readyThread = pop_fast( this->cltr );
292	// ready_schedule_unlock();
293	// if(readyThread) { break SEARCH; }
294	// }
295
296	// // no luck, try stealing a few times
297	// for(25) {
298	// if( __maybe_io_drain( this ) ) {
299	// ready_schedule_lock();
300	// readyThread = pop_fast( this->cltr );
301	// } else {
302	// ready_schedule_lock();
303	// readyThread = pop_slow( this->cltr );
304	// }
305	// ready_schedule_unlock();
306	// if(readyThread) { break SEARCH; }
307	// }
308
309	// // still no luck, search for a thread
310	// ready_schedule_lock();
311	// readyThread = pop_search( this->cltr );
312	// ready_schedule_unlock();
313	// if(readyThread) { break SEARCH; }
314
315	// // Don't block if we are done
316	// if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
317
318	// __STATS( __tls_stats()->ready.sleep.halts++; )
319
320	// // Push self to idle stack
321	// mark_idle(this->cltr->procs, * this);
322
323	// // Confirm the ready-queue is empty
324	// __maybe_io_drain( this );
325	// ready_schedule_lock();
326	// readyThread = pop_search( this->cltr );
327	// ready_schedule_unlock();
328
329	// if( readyThread ) {
330	// // A thread was found, cancel the halt
331	// mark_awake(this->cltr->procs, * this);
332
333	// __STATS( __tls_stats()->ready.sleep.cancels++; )
334
335	// // continue the main loop
336	// break SEARCH;
337	// }
338
339	// __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl()); )
340	// __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
341
342	// // __disable_interrupts_hard();
343	// eventfd_t val;
344	// eventfd_read( this->idle, &val );
345	// // __enable_interrupts_hard();
346
347	// __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl()); )
348
349	// // We were woken up, remove self from idle
350	// mark_awake(this->cltr->procs, * this);
351
352	// // DON'T just proceed, start looking again
353	// continue MAIN_LOOP;
354	// }
355
356	// RUN_THREAD:
357	// /* paranoid */ verify( kernelTLS().this_proc_id );
358	// /* paranoid */ verify( ! __preemption_enabled() );
359	// /* paranoid */ verify( readyThread );
360
361	// // Reset io dirty bit
362	// this->io.dirty = false;
363
364	// // We found a thread run it
365	// __run_thread(this, readyThread);
366
367	// // Are we done?
368	// if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
369
370	// #if !defined(__CFA_NO_STATISTICS__)
371	// unsigned long long curr = rdtscl();
372	// if(curr > (last_tally + 500000000)) {
373	// __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
374	// last_tally = curr;
375	// }
376	// #endif
377
378	// if(this->io.pending && !this->io.dirty) {
379	// __cfa_io_flush( this );
380	// }
381
382	// // Check if there is pending io
383	// __maybe_io_drain( this );
384	}
385
386	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
387	}
388
389	__cfa_io_stop( this );
390
391	post( this->terminated );
392
393
394	if(this == mainProcessor) {
395	// HACK : the coroutine context switch expects this_thread to be set
396	// and it make sense for it to be set in all other cases except here
397	// fake it
398	__cfaabi_tls.this_thread = mainThread;
399	}
400
401	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
402	}
403
404	static int * __volatile_errno() __attribute__((noinline));
405	static int * __volatile_errno() { asm(""); return &errno; }
406
407	// KERNEL ONLY
408	// runThread runs a thread by context switching
409	// from the processor coroutine to the target thread
410	static void __run_thread(processor * this, $thread * thrd_dst) {
411	/* paranoid */ verify( ! __preemption_enabled() );
412	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
413	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
414	__builtin_prefetch( thrd_dst->context.SP );
415
416	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
417
418	$coroutine * proc_cor = get_coroutine(this->runner);
419
420	// set state of processor coroutine to inactive
421	verify(proc_cor->state == Active);
422	proc_cor->state = Blocked;
423
424	// Actually run the thread
425	RUNNING: while(true) {
426	thrd_dst->preempted = __NO_PREEMPTION;
427	thrd_dst->state = Active;
428
429	// Update global state
430	kernelTLS().this_thread = thrd_dst;
431
432	/* paranoid */ verify( ! __preemption_enabled() );
433	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
434	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
435	/* paranoid */ verify( thrd_dst->context.SP );
436	/* paranoid */ verify( thrd_dst->state != Halted );
437	/* 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
438	/* 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
439	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
440
441
442
443	// set context switch to the thread that the processor is executing
444	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
445	// when __cfactx_switch returns we are back in the processor coroutine
446
447	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
448	/* 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 );
449	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
450	/* paranoid */ verify( thrd_dst->context.SP );
451	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
452	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
453	/* paranoid */ verify( ! __preemption_enabled() );
454
455	// Reset global state
456	kernelTLS().this_thread = 0p;
457
458	// We just finished running a thread, there are a few things that could have happened.
459	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
460	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
461	// 4 - Preempted
462	// In case 1, we may have won a race so we can't write to the state again.
463	// In case 2, we lost the race so we now own the thread.
464
465	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
466	// The thread was preempted, reschedule it and reset the flag
467	schedule_thread$( thrd_dst );
468	break RUNNING;
469	}
470
471	if(unlikely(thrd_dst->state == Halting)) {
472	// The thread has halted, it should never be scheduled/run again
473	// finish the thread
474	__thread_finish( thrd_dst );
475	break RUNNING;
476	}
477
478	/* paranoid */ verify( thrd_dst->state == Active );
479	thrd_dst->state = Blocked;
480
481	// set state of processor coroutine to active and the thread to inactive
482	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
483	switch(old_ticket) {
484	case TICKET_RUNNING:
485	// This is case 1, the regular case, nothing more is needed
486	break RUNNING;
487	case TICKET_UNBLOCK:
488	#if !defined(__CFA_NO_STATISTICS__)
489	__tls_stats()->ready.threads.threads++;
490	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
491	#endif
492	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
493	// In this case, just run it again.
494	continue RUNNING;
495	default:
496	// This makes no sense, something is wrong abort
497	abort();
498	}
499	}
500
501	// Just before returning to the processor, set the processor coroutine to active
502	proc_cor->state = Active;
503
504	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
505
506	#if !defined(__CFA_NO_STATISTICS__)
507	__tls_stats()->ready.threads.threads--;
508	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
509	#endif
510
511	/* paranoid */ verify( ! __preemption_enabled() );
512	}
513
514	// KERNEL_ONLY
515	void returnToKernel() {
516	/* paranoid */ verify( ! __preemption_enabled() );
517	$coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
518	$thread * thrd_src = kernelTLS().this_thread;
519
520	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
521
522	// Run the thread on this processor
523	{
524	int local_errno = *__volatile_errno();
525	#if defined( __i386 ) \|\| defined( __x86_64 )
526	__x87_store;
527	#endif
528	/* paranoid */ verify( proc_cor->context.SP );
529	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
530	__cfactx_switch( &thrd_src->context, &proc_cor->context );
531	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
532	#if defined( __i386 ) \|\| defined( __x86_64 )
533	__x87_load;
534	#endif
535	*__volatile_errno() = local_errno;
536	}
537
538	#if !defined(__CFA_NO_STATISTICS__)
539	/* paranoid */ verify( thrd_src->last_proc != 0p );
540	if(thrd_src->last_proc != kernelTLS().this_processor) {
541	__tls_stats()->ready.threads.migration++;
542	}
543	#endif
544
545	/* paranoid */ verify( ! __preemption_enabled() );
546	/* 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 );
547	/* 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 );
548	}
549
550	//-----------------------------------------------------------------------------
551	// Scheduler routines
552	// KERNEL ONLY
553	void __schedule_thread( $thread * thrd ) {
554	/* paranoid */ verify( ! __preemption_enabled() );
555	/* paranoid */ verify( kernelTLS().this_proc_id );
556	/* paranoid */ verify( ready_schedule_islocked());
557	/* paranoid */ verify( thrd );
558	/* paranoid */ verify( thrd->state != Halted );
559	/* paranoid */ verify( thrd->curr_cluster );
560	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
561	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
562	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
563	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
564	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
565	/* paranoid */ #endif
566	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
567	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
568
569
570	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
571
572	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
573	struct cluster * cl = thrd->curr_cluster;
574	__STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
575
576	// push the thread to the cluster ready-queue
577	push( cl, thrd );
578
579	// variable thrd is no longer safe to use
580	thrd = 0xdeaddeaddeaddeadp;
581
582	// wake the cluster using the save variable.
583	__wake_one( cl );
584
585	#if !defined(__CFA_NO_STATISTICS__)
586	if( kernelTLS().this_stats ) {
587	__tls_stats()->ready.threads.threads++;
588	if(outside) {
589	__tls_stats()->ready.threads.extunpark++;
590	}
591	__push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
592	}
593	else {
594	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
595	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
596	__push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
597	}
598	#endif
599
600	/* paranoid */ verify( ready_schedule_islocked());
601	/* paranoid */ verify( ! __preemption_enabled() );
602	}
603
604	void schedule_thread$( $thread * thrd ) {
605	ready_schedule_lock();
606	__schedule_thread( thrd );
607	ready_schedule_unlock();
608	}
609
610	// KERNEL ONLY
611	static inline $thread * __next_thread(cluster * this) with( *this ) {
612	/* paranoid */ verify( ! __preemption_enabled() );
613	/* paranoid */ verify( kernelTLS().this_proc_id );
614
615	ready_schedule_lock();
616	$thread * thrd = pop_fast( this );
617	ready_schedule_unlock();
618
619	/* paranoid */ verify( kernelTLS().this_proc_id );
620	/* paranoid */ verify( ! __preemption_enabled() );
621	return thrd;
622	}
623
624	// KERNEL ONLY
625	static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
626	/* paranoid */ verify( ! __preemption_enabled() );
627	/* paranoid */ verify( kernelTLS().this_proc_id );
628
629	ready_schedule_lock();
630	$thread * thrd;
631	for(25) {
632	thrd = pop_slow( this );
633	if(thrd) goto RET;
634	}
635	thrd = pop_search( this );
636
637	RET:
638	ready_schedule_unlock();
639
640	/* paranoid */ verify( kernelTLS().this_proc_id );
641	/* paranoid */ verify( ! __preemption_enabled() );
642	return thrd;
643	}
644
645	static inline bool __must_unpark( $thread * thrd ) {
646	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
647	switch(old_ticket) {
648	case TICKET_RUNNING:
649	// Wake won the race, the thread will reschedule/rerun itself
650	return false;
651	case TICKET_BLOCKED:
652	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
653	/* paranoid */ verify( thrd->state == Blocked );
654	return true;
655	default:
656	// This makes no sense, something is wrong abort
657	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
658	}
659	}
660
661	void unpark( $thread * thrd ) {
662	if( !thrd ) return;
663
664	if(__must_unpark(thrd)) {
665	disable_interrupts();
666	// Wake lost the race,
667	schedule_thread$( thrd );
668	enable_interrupts(false);
669	}
670	}
671
672	void park( void ) {
673	__disable_interrupts_checked();
674	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
675	returnToKernel();
676	__enable_interrupts_checked();
677
678	}
679
680	extern "C" {
681	// Leave the thread monitor
682	// last routine called by a thread.
683	// Should never return
684	void __cfactx_thrd_leave() {
685	$thread * thrd = active_thread();
686	$monitor * this = &thrd->self_mon;
687
688	// Lock the monitor now
689	lock( this->lock __cfaabi_dbg_ctx2 );
690
691	disable_interrupts();
692
693	/* paranoid */ verify( ! __preemption_enabled() );
694	/* paranoid */ verify( thrd->state == Active );
695	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
696	/* paranoid */ verify( kernelTLS().this_thread == thrd );
697	/* paranoid */ verify( thrd->context.SP );
698	/* 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 );
699	/* 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 );
700
701	thrd->state = Halting;
702	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
703	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
704	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
705
706	// Leave the thread
707	returnToKernel();
708
709	// Control flow should never reach here!
710	abort();
711	}
712	}
713
714	// KERNEL ONLY
715	bool force_yield( __Preemption_Reason reason ) {
716	__disable_interrupts_checked();
717	$thread * thrd = kernelTLS().this_thread;
718	/* paranoid */ verify(thrd->state == Active);
719
720	// SKULLDUGGERY: It is possible that we are preempting this thread just before
721	// it was going to park itself. If that is the case and it is already using the
722	// intrusive fields then we can't use them to preempt the thread
723	// If that is the case, abandon the preemption.
724	bool preempted = false;
725	if(thrd->link.next == 0p) {
726	preempted = true;
727	thrd->preempted = reason;
728	returnToKernel();
729	}
730	__enable_interrupts_checked( false );
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_checked();
776	/* paranoid */ verify( ! __preemption_enabled() );
777	eventfd_t val;
778	val = 1;
779	eventfd_write( this->idle, val );
780	__enable_interrupts_checked();
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