Context Navigation

kernel.cfa @ 9cdfa5fb

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since 9cdfa5fb was c86ee4c, 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: 29.6 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 _GNU_SOURCE
18
19	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
20
21	//C Includes
22	#include <errno.h>
23	#include <stdio.h>
24	#include <signal.h>
25	#include <unistd.h>
26	extern "C" {
27	#include <sys/eventfd.h>
28	}
29
30	//CFA Includes
31	#include "kernel_private.hfa"
32	#include "preemption.hfa"
33
34	//Private includes
35	#define __CFA_INVOKE_PRIVATE__
36	#include "invoke.h"
37
38	#if !defined(__CFA_NO_STATISTICS__)
39	#define __STATS( ...) __VA_ARGS__
40	#else
41	#define __STATS( ...)
42	#endif
43
44	//-----------------------------------------------------------------------------
45	// Some assembly required
46	#if defined( __i386 )
47	// mxcr : SSE Status and Control bits (control bits are preserved across function calls)
48	// fcw : X87 FPU control word (preserved across function calls)
49	#define __x87_store \
50	uint32_t __mxcr; \
51	uint16_t __fcw; \
52	__asm__ volatile ( \
53	"stmxcsr %0\n" \
54	"fnstcw %1\n" \
55	: "=m" (__mxcr),\
56	"=m" (__fcw) \
57	)
58
59	#define __x87_load \
60	__asm__ volatile ( \
61	"fldcw %1\n" \
62	"ldmxcsr %0\n" \
63	::"m" (__mxcr),\
64	"m" (__fcw) \
65	)
66
67	#elif defined( __x86_64 )
68	#define __x87_store \
69	uint32_t __mxcr; \
70	uint16_t __fcw; \
71	__asm__ volatile ( \
72	"stmxcsr %0\n" \
73	"fnstcw %1\n" \
74	: "=m" (__mxcr),\
75	"=m" (__fcw) \
76	)
77
78	#define __x87_load \
79	__asm__ volatile ( \
80	"fldcw %1\n" \
81	"ldmxcsr %0\n" \
82	:: "m" (__mxcr),\
83	"m" (__fcw) \
84	)
85
86	#elif defined( __arm__ )
87	#define __x87_store
88	#define __x87_load
89
90	#elif defined( __aarch64__ )
91	#define __x87_store \
92	uint32_t __fpcntl[2]; \
93	__asm__ volatile ( \
94	"mrs x9, FPCR\n" \
95	"mrs x10, FPSR\n" \
96	"stp x9, x10, %0\n" \
97	: "=m" (__fpcntl) : : "x9", "x10" \
98	)
99
100	#define __x87_load \
101	__asm__ volatile ( \
102	"ldp x9, x10, %0\n" \
103	"msr FPSR, x10\n" \
104	"msr FPCR, x9\n" \
105	: "=m" (__fpcntl) : : "x9", "x10" \
106	)
107
108	#else
109	#error unsupported hardware architecture
110	#endif
111
112	extern thread$ * mainThread;
113	extern processor * mainProcessor;
114
115	//-----------------------------------------------------------------------------
116	// Kernel Scheduling logic
117	static thread$ * __next_thread(cluster * this);
118	static thread$ * __next_thread_slow(cluster * this);
119	static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
120	static void __run_thread(processor * this, thread$ * dst);
121	static void __wake_one(cluster * cltr);
122
123	static void mark_idle (__cluster_proc_list & idles, processor & proc);
124	static void mark_awake(__cluster_proc_list & idles, processor & proc);
125	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
126
127	extern void __cfa_io_start( processor * );
128	extern bool __cfa_io_drain( processor * );
129	extern void __cfa_io_flush( processor * );
130	extern void __cfa_io_stop ( processor * );
131	static inline bool __maybe_io_drain( processor * );
132
133	extern void __disable_interrupts_hard();
134	extern void __enable_interrupts_hard();
135
136	static inline void __disable_interrupts_checked() {
137	/* paranoid */ verify( __preemption_enabled() );
138	disable_interrupts();
139	/* paranoid */ verify( ! __preemption_enabled() );
140	}
141
142	static inline void __enable_interrupts_checked( bool poll = true ) {
143	/* paranoid */ verify( ! __preemption_enabled() );
144	enable_interrupts( poll );
145	/* paranoid */ verify( __preemption_enabled() );
146	}
147
148	//=============================================================================================
149	// Kernel Scheduling logic
150	//=============================================================================================
151	//Main of the processor contexts
152	void main(processorCtx_t & runner) {
153	// Because of a bug, we couldn't initialized the seed on construction
154	// Do it here
155	__cfaabi_tls.rand_seed ^= rdtscl();
156	__cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
157	__tls_rand_advance_bck();
158
159	processor * this = runner.proc;
160	verify(this);
161
162	__cfa_io_start( this );
163
164	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
165	#if !defined(__CFA_NO_STATISTICS__)
166	if( this->print_halts ) {
167	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
168	}
169	#endif
170
171	{
172	// Setup preemption data
173	preemption_scope scope = { this };
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	#define OLD_MAIN 1
187	#if OLD_MAIN
188	// Check if there is pending io
189	__maybe_io_drain( this );
190
191	// Try to get the next thread
192	readyThread = __next_thread( this->cltr );
193
194	if( !readyThread ) {
195	__cfa_io_flush( this );
196	readyThread = __next_thread_slow( this->cltr );
197	}
198
199	HALT:
200	if( !readyThread ) {
201	// Don't block if we are done
202	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
203
204	#if !defined(__CFA_NO_STATISTICS__)
205	__tls_stats()->ready.sleep.halts++;
206	#endif
207
208	// Push self to idle stack
209	mark_idle(this->cltr->procs, * this);
210
211	// Confirm the ready-queue is empty
212	readyThread = __next_thread_slow( this->cltr );
213	if( readyThread ) {
214	// A thread was found, cancel the halt
215	mark_awake(this->cltr->procs, * this);
216
217	#if !defined(__CFA_NO_STATISTICS__)
218	__tls_stats()->ready.sleep.cancels++;
219	#endif
220
221	// continue the mai loop
222	break HALT;
223	}
224
225	#if !defined(__CFA_NO_STATISTICS__)
226	if(this->print_halts) {
227	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
228	}
229	#endif
230
231	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
232
233	__disable_interrupts_hard();
234	eventfd_t val;
235	eventfd_read( this->idle, &val );
236	__enable_interrupts_hard();
237
238	#if !defined(__CFA_NO_STATISTICS__)
239	if(this->print_halts) {
240	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
241	}
242	#endif
243
244	// We were woken up, remove self from idle
245	mark_awake(this->cltr->procs, * this);
246
247	// DON'T just proceed, start looking again
248	continue MAIN_LOOP;
249	}
250
251	/* paranoid */ verify( readyThread );
252
253	// Reset io dirty bit
254	this->io.dirty = false;
255
256	// We found a thread run it
257	__run_thread(this, readyThread);
258
259	// Are we done?
260	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
261
262	if(this->io.pending && !this->io.dirty) {
263	__cfa_io_flush( this );
264	}
265
266	#else
267	#warning new kernel loop
268	SEARCH: {
269	/* paranoid */ verify( ! __preemption_enabled() );
270
271	// First, lock the scheduler since we are searching for a thread
272	ready_schedule_lock();
273
274	// Try to get the next thread
275	readyThread = pop_fast( this->cltr );
276	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
277
278	// If we can't find a thread, might as well flush any outstanding I/O
279	if(this->io.pending) { __cfa_io_flush( this ); }
280
281	// Spin a little on I/O, just in case
282	for(5) {
283	__maybe_io_drain( this );
284	readyThread = pop_fast( this->cltr );
285	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
286	}
287
288	// no luck, try stealing a few times
289	for(5) {
290	if( __maybe_io_drain( this ) ) {
291	readyThread = pop_fast( this->cltr );
292	} else {
293	readyThread = pop_slow( this->cltr );
294	}
295	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
296	}
297
298	// still no luck, search for a thread
299	readyThread = pop_search( this->cltr );
300	if(readyThread) { ready_schedule_unlock(); 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	ready_schedule_unlock();
309	mark_idle(this->cltr->procs, * this);
310	ready_schedule_lock();
311
312	// Confirm the ready-queue is empty
313	__maybe_io_drain( this );
314	readyThread = pop_search( this->cltr );
315	ready_schedule_unlock();
316
317	if( readyThread ) {
318	// A thread was found, cancel the halt
319	mark_awake(this->cltr->procs, * this);
320
321	__STATS( __tls_stats()->ready.sleep.cancels++; )
322
323	// continue the main loop
324	break SEARCH;
325	}
326
327	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
328	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
329
330	// __disable_interrupts_hard();
331	eventfd_t val;
332	eventfd_read( this->idle, &val );
333	// __enable_interrupts_hard();
334
335	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
336
337	// We were woken up, remove self from idle
338	mark_awake(this->cltr->procs, * this);
339
340	// DON'T just proceed, start looking again
341	continue MAIN_LOOP;
342	}
343
344	RUN_THREAD:
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(this->io.pending && !this->io.dirty) {
358	__cfa_io_flush( this );
359	}
360
361	ready_schedule_lock();
362	__maybe_io_drain( this );
363	ready_schedule_unlock();
364	#endif
365	}
366
367	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
368	}
369
370	__cfa_io_stop( this );
371
372	post( this->terminated );
373
374	if(this == mainProcessor) {
375	// HACK : the coroutine context switch expects this_thread to be set
376	// and it make sense for it to be set in all other cases except here
377	// fake it
378	__cfaabi_tls.this_thread = mainThread;
379	}
380
381	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
382	}
383
384	static int * __volatile_errno() __attribute__((noinline));
385	static int * __volatile_errno() { asm(""); return &errno; }
386
387	// KERNEL ONLY
388	// runThread runs a thread by context switching
389	// from the processor coroutine to the target thread
390	static void __run_thread(processor * this, thread$ * thrd_dst) {
391	/* paranoid */ verify( ! __preemption_enabled() );
392	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
393	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
394	__builtin_prefetch( thrd_dst->context.SP );
395
396	int curr = __kernel_getcpu();
397	if(thrd_dst->last_cpu != curr) {
398	int64_t l = thrd_dst->last_cpu;
399	int64_t c = curr;
400	int64_t v = (l << 32) \| c;
401	__push_stat( __tls_stats(), v, false, "Processor", this );
402	}
403
404	thrd_dst->last_cpu = curr;
405
406	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
407
408	coroutine$ * proc_cor = get_coroutine(this->runner);
409
410	// set state of processor coroutine to inactive
411	verify(proc_cor->state == Active);
412	proc_cor->state = Blocked;
413
414	// Actually run the thread
415	RUNNING: while(true) {
416	thrd_dst->preempted = __NO_PREEMPTION;
417	thrd_dst->state = Active;
418
419	// Update global state
420	kernelTLS().this_thread = thrd_dst;
421
422	/* paranoid */ verify( ! __preemption_enabled() );
423	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
424	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
425	/* paranoid */ verify( thrd_dst->context.SP );
426	/* paranoid */ verify( thrd_dst->state != Halted );
427	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) \|\| thrd_dst->curr_cor == proc_cor \|\| thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
428	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) \|\| thrd_dst->curr_cor == proc_cor \|\| thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
429	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
430
431
432
433	// set context switch to the thread that the processor is executing
434	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
435	// when __cfactx_switch returns we are back in the processor coroutine
436
437
438
439	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
440	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) \|\| thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
441	/* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) \|\| thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
442	/* paranoid */ verify( thrd_dst->context.SP );
443	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
444	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
445	/* paranoid */ verify( ! __preemption_enabled() );
446
447	// Reset global state
448	kernelTLS().this_thread = 0p;
449
450	// We just finished running a thread, there are a few things that could have happened.
451	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
452	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
453	// 4 - Preempted
454	// In case 1, we may have won a race so we can't write to the state again.
455	// In case 2, we lost the race so we now own the thread.
456
457	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
458	// The thread was preempted, reschedule it and reset the flag
459	schedule_thread$( thrd_dst );
460	break RUNNING;
461	}
462
463	if(unlikely(thrd_dst->state == Halting)) {
464	// The thread has halted, it should never be scheduled/run again
465	// finish the thread
466	__thread_finish( thrd_dst );
467	break RUNNING;
468	}
469
470	/* paranoid */ verify( thrd_dst->state == Active );
471	thrd_dst->state = Blocked;
472
473	// set state of processor coroutine to active and the thread to inactive
474	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
475	switch(old_ticket) {
476	case TICKET_RUNNING:
477	// This is case 1, the regular case, nothing more is needed
478	break RUNNING;
479	case TICKET_UNBLOCK:
480	#if !defined(__CFA_NO_STATISTICS__)
481	__tls_stats()->ready.threads.threads++;
482	#endif
483	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
484	// In this case, just run it again.
485	continue RUNNING;
486	default:
487	// This makes no sense, something is wrong abort
488	abort();
489	}
490	}
491
492	// Just before returning to the processor, set the processor coroutine to active
493	proc_cor->state = Active;
494
495	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
496
497	#if !defined(__CFA_NO_STATISTICS__)
498	__tls_stats()->ready.threads.threads--;
499	#endif
500
501	/* paranoid */ verify( ! __preemption_enabled() );
502	}
503
504	// KERNEL_ONLY
505	void returnToKernel() {
506	/* paranoid */ verify( ! __preemption_enabled() );
507	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
508	thread$ * thrd_src = kernelTLS().this_thread;
509
510	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
511
512	// Run the thread on this processor
513	{
514	int local_errno = *__volatile_errno();
515	#if defined( __i386 ) \|\| defined( __x86_64 )
516	__x87_store;
517	#endif
518	/* paranoid */ verify( proc_cor->context.SP );
519	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
520	__cfactx_switch( &thrd_src->context, &proc_cor->context );
521	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
522	#if defined( __i386 ) \|\| defined( __x86_64 )
523	__x87_load;
524	#endif
525	*__volatile_errno() = local_errno;
526	}
527
528	#if !defined(__CFA_NO_STATISTICS__)
529	/* paranoid */ verify( thrd_src->last_proc != 0p );
530	if(thrd_src->last_proc != kernelTLS().this_processor) {
531	__tls_stats()->ready.threads.migration++;
532	}
533	#endif
534
535	/* paranoid */ verify( ! __preemption_enabled() );
536	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ) \|\| thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_src );
537	/* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit) \|\| thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_src );
538	}
539
540	//-----------------------------------------------------------------------------
541	// Scheduler routines
542	// KERNEL ONLY
543	static void __schedule_thread( thread$ * thrd ) {
544	/* paranoid */ verify( ! __preemption_enabled() );
545	/* paranoid */ verify( ready_schedule_islocked());
546	/* paranoid */ verify( thrd );
547	/* paranoid */ verify( thrd->state != Halted );
548	/* paranoid */ verify( thrd->curr_cluster );
549	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
550	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
551	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
552	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
553	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
554	/* paranoid */ #endif
555	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
556	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
557
558	const bool local = thrd->state != Start;
559	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
560
561	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
562	struct cluster * cl = thrd->curr_cluster;
563	__STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
564
565	// push the thread to the cluster ready-queue
566	push( cl, thrd, local );
567
568	// variable thrd is no longer safe to use
569	thrd = 0xdeaddeaddeaddeadp;
570
571	// wake the cluster using the save variable.
572	__wake_one( cl );
573
574	#if !defined(__CFA_NO_STATISTICS__)
575	if( kernelTLS().this_stats ) {
576	__tls_stats()->ready.threads.threads++;
577	if(outside) {
578	__tls_stats()->ready.threads.extunpark++;
579	}
580	}
581	else {
582	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
583	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
584	}
585	#endif
586
587	/* paranoid */ verify( ready_schedule_islocked());
588	/* paranoid */ verify( ! __preemption_enabled() );
589	}
590
591	void schedule_thread$( thread$ * thrd ) {
592	ready_schedule_lock();
593	__schedule_thread( thrd );
594	ready_schedule_unlock();
595	}
596
597	// KERNEL ONLY
598	static inline thread$ * __next_thread(cluster * this) with( *this ) {
599	/* paranoid */ verify( ! __preemption_enabled() );
600
601	ready_schedule_lock();
602	thread$ * thrd = pop_fast( this );
603	ready_schedule_unlock();
604
605	/* paranoid */ verify( ! __preemption_enabled() );
606	return thrd;
607	}
608
609	// KERNEL ONLY
610	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
611	/* paranoid */ verify( ! __preemption_enabled() );
612
613	ready_schedule_lock();
614	thread$ * thrd;
615	for(25) {
616	thrd = pop_slow( this );
617	if(thrd) goto RET;
618	}
619	thrd = pop_search( this );
620
621	RET:
622	ready_schedule_unlock();
623
624	/* paranoid */ verify( ! __preemption_enabled() );
625	return thrd;
626	}
627
628	static inline bool __must_unpark( thread$ * thrd ) {
629	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
630	switch(old_ticket) {
631	case TICKET_RUNNING:
632	// Wake won the race, the thread will reschedule/rerun itself
633	return false;
634	case TICKET_BLOCKED:
635	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
636	/* paranoid */ verify( thrd->state == Blocked );
637	return true;
638	default:
639	// This makes no sense, something is wrong abort
640	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
641	}
642	}
643
644	void __kernel_unpark( thread$ * thrd ) {
645	/* paranoid */ verify( ! __preemption_enabled() );
646	/* paranoid */ verify( ready_schedule_islocked());
647
648	if( !thrd ) return;
649
650	if(__must_unpark(thrd)) {
651	// Wake lost the race,
652	__schedule_thread( thrd );
653	}
654
655	/* paranoid */ verify( ready_schedule_islocked());
656	/* paranoid */ verify( ! __preemption_enabled() );
657	}
658
659	void unpark( thread$ * thrd ) {
660	if( !thrd ) return;
661
662	if(__must_unpark(thrd)) {
663	disable_interrupts();
664	// Wake lost the race,
665	schedule_thread$( thrd );
666	enable_interrupts(false);
667	}
668	}
669
670	void park( void ) {
671	__disable_interrupts_checked();
672	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
673	returnToKernel();
674	__enable_interrupts_checked();
675
676	}
677
678	extern "C" {
679	// Leave the thread monitor
680	// last routine called by a thread.
681	// Should never return
682	void __cfactx_thrd_leave() {
683	thread$ * thrd = active_thread();
684	monitor$ * this = &thrd->self_mon;
685
686	// Lock the monitor now
687	lock( this->lock __cfaabi_dbg_ctx2 );
688
689	disable_interrupts();
690
691	/* paranoid */ verify( ! __preemption_enabled() );
692	/* paranoid */ verify( thrd->state == Active );
693	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
694	/* paranoid */ verify( kernelTLS().this_thread == thrd );
695	/* paranoid */ verify( thrd->context.SP );
696	/* 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 );
697	/* 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 );
698
699	thrd->state = Halting;
700	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
701	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
702	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
703
704	// Leave the thread
705	returnToKernel();
706
707	// Control flow should never reach here!
708	abort();
709	}
710	}
711
712	// KERNEL ONLY
713	bool force_yield( __Preemption_Reason reason ) {
714	__disable_interrupts_checked();
715	thread$ * thrd = kernelTLS().this_thread;
716	/* paranoid */ verify(thrd->state == Active);
717
718	// SKULLDUGGERY: It is possible that we are preempting this thread just before
719	// it was going to park itself. If that is the case and it is already using the
720	// intrusive fields then we can't use them to preempt the thread
721	// If that is the case, abandon the preemption.
722	bool preempted = false;
723	if(thrd->link.next == 0p) {
724	preempted = true;
725	thrd->preempted = reason;
726	returnToKernel();
727	}
728	__enable_interrupts_checked( false );
729	return preempted;
730	}
731
732	//=============================================================================================
733	// Kernel Idle Sleep
734	//=============================================================================================
735	// Wake a thread from the front if there are any
736	static void __wake_one(cluster * this) {
737	/* paranoid */ verify( ! __preemption_enabled() );
738	/* paranoid */ verify( ready_schedule_islocked() );
739
740	// Check if there is a sleeping processor
741	processor * p;
742	unsigned idle;
743	unsigned total;
744	[idle, total, p] = query_idles(this->procs);
745
746	// If no one is sleeping, we are done
747	if( idle == 0 ) return;
748
749	// We found a processor, wake it up
750	eventfd_t val;
751	val = 1;
752	eventfd_write( p->idle, val );
753
754	#if !defined(__CFA_NO_STATISTICS__)
755	if( kernelTLS().this_stats ) {
756	__tls_stats()->ready.sleep.wakes++;
757	}
758	else {
759	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
760	}
761	#endif
762
763	/* paranoid */ verify( ready_schedule_islocked() );
764	/* paranoid */ verify( ! __preemption_enabled() );
765
766	return;
767	}
768
769	// Unconditionnaly wake a thread
770	void __wake_proc(processor * this) {
771	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
772
773	__disable_interrupts_checked();
774	/* paranoid */ verify( ! __preemption_enabled() );
775	eventfd_t val;
776	val = 1;
777	eventfd_write( this->idle, val );
778	__enable_interrupts_checked();
779	}
780
781	static void mark_idle(__cluster_proc_list & this, processor & proc) {
782	/* paranoid */ verify( ! __preemption_enabled() );
783	lock( this );
784	this.idle++;
785	/* paranoid */ verify( this.idle <= this.total );
786	remove(proc);
787	insert_first(this.idles, proc);
788	unlock( this );
789	/* paranoid */ verify( ! __preemption_enabled() );
790	}
791
792	static void mark_awake(__cluster_proc_list & this, processor & proc) {
793	/* paranoid */ verify( ! __preemption_enabled() );
794	lock( this );
795	this.idle--;
796	/* paranoid */ verify( this.idle >= 0 );
797	remove(proc);
798	insert_last(this.actives, proc);
799	unlock( this );
800	/* paranoid */ verify( ! __preemption_enabled() );
801	}
802
803	static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
804	/* paranoid */ verify( ! __preemption_enabled() );
805	/* paranoid */ verify( ready_schedule_islocked() );
806
807	for() {
808	uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
809	if( 1 == (l % 2) ) { Pause(); continue; }
810	unsigned idle = this.idle;
811	unsigned total = this.total;
812	processor * proc = &this.idles`first;
813	// Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
814	asm volatile("": : :"memory");
815	if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
816	return [idle, total, proc];
817	}
818
819	/* paranoid */ verify( ready_schedule_islocked() );
820	/* paranoid */ verify( ! __preemption_enabled() );
821	}
822
823	//=============================================================================================
824	// Unexpected Terminating logic
825	//=============================================================================================
826	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
827	thread$ * thrd = __cfaabi_tls.this_thread;
828
829	if(thrd) {
830	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
831	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
832
833	if ( &thrd->self_cor != thrd->curr_cor ) {
834	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
835	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
836	}
837	else {
838	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
839	}
840	}
841	else {
842	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
843	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
844	}
845	}
846
847	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
848	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
849	}
850
851	static __spinlock_t kernel_debug_lock;
852
853	extern "C" {
854	void __cfaabi_bits_acquire() {
855	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
856	}
857
858	void __cfaabi_bits_release() {
859	unlock( kernel_debug_lock );
860	}
861	}
862
863	//=============================================================================================
864	// Kernel Utilities
865	//=============================================================================================
866	#if defined(CFA_HAVE_LINUX_IO_URING_H)
867	#include "io/types.hfa"
868	#endif
869
870	static inline bool __maybe_io_drain( processor * proc ) {
871	bool ret = false;
872	#if defined(CFA_HAVE_LINUX_IO_URING_H)
873	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
874
875	// Check if we should drain the queue
876	$io_context * ctx = proc->io.ctx;
877	unsigned head = *ctx->cq.head;
878	unsigned tail = *ctx->cq.tail;
879	if(head == tail) return false;
880	#if OLD_MAIN
881	ready_schedule_lock();
882	ret = __cfa_io_drain( proc );
883	ready_schedule_unlock();
884	#else
885	ret = __cfa_io_drain( proc );
886	#endif
887	#endif
888	return ret;
889	}
890
891	//-----------------------------------------------------------------------------
892	// Debug
893	__cfaabi_dbg_debug_do(
894	extern "C" {
895	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
896	this.prev_name = prev_name;
897	this.prev_thrd = kernelTLS().this_thread;
898	}
899	}
900	)
901
902	//-----------------------------------------------------------------------------
903	// Debug
904	bool threading_enabled(void) __attribute__((const)) {
905	return true;
906	}
907
908	//-----------------------------------------------------------------------------
909	// Statistics
910	#if !defined(__CFA_NO_STATISTICS__)
911	void print_halts( processor & this ) {
912	this.print_halts = true;
913	}
914
915	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
916	/* paranoid */ verify( cltr->stats );
917
918	processor * it = &list`first;
919	for(unsigned i = 0; i < count; i++) {
920	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
921	/* paranoid */ verify( it->local_data->this_stats );
922	__tally_stats( cltr->stats, it->local_data->this_stats );
923	it = &(*it)`next;
924	}
925	}
926
927	void crawl_cluster_stats( cluster & this ) {
928	// Stop the world, otherwise stats could get really messed-up
929	// this doesn't solve all problems but does solve many
930	// so it's probably good enough
931	uint_fast32_t last_size = ready_mutate_lock();
932
933	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
934	crawl_list(&this, this.procs.idles , this.procs.idle );
935
936	// Unlock the RWlock
937	ready_mutate_unlock( last_size );
938	}
939
940
941	void print_stats_now( cluster & this, int flags ) {
942	crawl_cluster_stats( this );
943	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
944	}
945	#endif
946	// Local Variables: //
947	// mode: c //
948	// tab-width: 4 //
949	// End: //

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

Download in other formats:

Original Format