Context Navigation

kernel.cfa@ 9bbbc8e

Visit:

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast new-ast-unique-expr pthread-emulation qualifiedEnum

Last change on this file since 9bbbc8e was f2b18d01, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
First version of tools to view halts
Property mode set to `100644`
File size: 38.3 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 : Tue May 26 22:05:19 2020
13	// Update Count : 59
14	//
15
16	#define __cforall_thread__
17	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
18
19	//C Includes
20	#include <stddef.h>
21	#include <errno.h>
22	#include <string.h>
23	#include <stdio.h>
24	#include <fenv.h>
25	#include <signal.h>
26	#include <unistd.h>
27	#include <limits.h> // PTHREAD_STACK_MIN
28	#include <sys/mman.h> // mprotect
29	extern "C" {
30	#include <sys/resource.h>
31	}
32
33	//CFA Includes
34	#include "time.hfa"
35	#include "kernel_private.hfa"
36	#include "preemption.hfa"
37	#include "startup.hfa"
38
39	//Private includes
40	#define __CFA_INVOKE_PRIVATE__
41	#include "invoke.h"
42
43
44	//-----------------------------------------------------------------------------
45	// Some assembly required
46	#if defined( __i386 )
47	#define CtxGet( ctx ) \
48	__asm__ volatile ( \
49	"movl %%esp,%0\n"\
50	"movl %%ebp,%1\n"\
51	: "=rm" (ctx.SP),\
52	"=rm" (ctx.FP) \
53	)
54
55	// mxcr : SSE Status and Control bits (control bits are preserved across function calls)
56	// fcw : X87 FPU control word (preserved across function calls)
57	#define __x87_store \
58	uint32_t __mxcr; \
59	uint16_t __fcw; \
60	__asm__ volatile ( \
61	"stmxcsr %0\n" \
62	"fnstcw %1\n" \
63	: "=m" (__mxcr),\
64	"=m" (__fcw) \
65	)
66
67	#define __x87_load \
68	__asm__ volatile ( \
69	"fldcw %1\n" \
70	"ldmxcsr %0\n" \
71	::"m" (__mxcr),\
72	"m" (__fcw) \
73	)
74
75	#elif defined( __x86_64 )
76	#define CtxGet( ctx ) \
77	__asm__ volatile ( \
78	"movq %%rsp,%0\n"\
79	"movq %%rbp,%1\n"\
80	: "=rm" (ctx.SP),\
81	"=rm" (ctx.FP) \
82	)
83
84	#define __x87_store \
85	uint32_t __mxcr; \
86	uint16_t __fcw; \
87	__asm__ volatile ( \
88	"stmxcsr %0\n" \
89	"fnstcw %1\n" \
90	: "=m" (__mxcr),\
91	"=m" (__fcw) \
92	)
93
94	#define __x87_load \
95	__asm__ volatile ( \
96	"fldcw %1\n" \
97	"ldmxcsr %0\n" \
98	:: "m" (__mxcr),\
99	"m" (__fcw) \
100	)
101
102
103	#elif defined( __ARM_ARCH )
104	#define CtxGet( ctx ) __asm__ ( \
105	"mov %0,%%sp\n" \
106	"mov %1,%%r11\n" \
107	: "=rm" (ctx.SP), "=rm" (ctx.FP) )
108	#else
109	#error unknown hardware architecture
110	#endif
111
112	//-----------------------------------------------------------------------------
113	//Start and stop routine for the kernel, declared first to make sure they run first
114	static void __kernel_startup (void) __attribute__(( constructor( STARTUP_PRIORITY_KERNEL ) ));
115	static void __kernel_shutdown(void) __attribute__(( destructor ( STARTUP_PRIORITY_KERNEL ) ));
116
117	//-----------------------------------------------------------------------------
118	// Kernel Scheduling logic
119	static $thread * __next_thread(cluster * this);
120	static bool __has_next_thread(cluster * this);
121	static void __run_thread(processor * this, $thread * dst);
122	static bool __wake_proc(processor *);
123	static bool __wake_one(struct __processor_id_t * id, cluster * cltr);
124	static void __halt(processor * this);
125
126	//-----------------------------------------------------------------------------
127	// Kernel storage
128	KERNEL_STORAGE(cluster, mainCluster);
129	KERNEL_STORAGE(processor, mainProcessor);
130	KERNEL_STORAGE($thread, mainThread);
131	KERNEL_STORAGE(__stack_t, mainThreadCtx);
132	KERNEL_STORAGE(__scheduler_RWLock_t, __scheduler_lock);
133	#if !defined(__CFA_NO_STATISTICS__)
134	KERNEL_STORAGE(__stats_t, mainProcStats);
135	#endif
136
137	cluster * mainCluster;
138	processor * mainProcessor;
139	$thread * mainThread;
140	__scheduler_RWLock_t * __scheduler_lock;
141
142	extern "C" {
143	struct { __dllist_t(cluster) list; __spinlock_t lock; } __cfa_dbg_global_clusters;
144	}
145
146	size_t __page_size = 0;
147
148	//-----------------------------------------------------------------------------
149	// Global state
150	thread_local struct KernelThreadData kernelTLS __attribute__ ((tls_model ( "initial-exec" ))) = {
151	NULL, // cannot use 0p
152	NULL,
153	NULL,
154	{ 1, false, false },
155	6u //this should be seeded better but due to a bug calling rdtsc doesn't work
156	};
157
158	//-----------------------------------------------------------------------------
159	// Struct to steal stack
160	struct current_stack_info_t {
161	__stack_t * storage; // pointer to stack object
162	void * base; // base of stack
163	void * limit; // stack grows towards stack limit
164	void * context; // address of cfa_context_t
165	};
166
167	void ?{}( current_stack_info_t & this ) {
168	__stack_context_t ctx;
169	CtxGet( ctx );
170	this.base = ctx.FP;
171
172	rlimit r;
173	getrlimit( RLIMIT_STACK, &r);
174	size_t size = r.rlim_cur;
175
176	this.limit = (void *)(((intptr_t)this.base) - size);
177	this.context = &storage_mainThreadCtx;
178	}
179
180	//-----------------------------------------------------------------------------
181	// Main thread construction
182
183	void ?{}( $coroutine & this, current_stack_info_t * info) with( this ) {
184	stack.storage = info->storage;
185	with(*stack.storage) {
186	limit = info->limit;
187	base = info->base;
188	}
189	__attribute__((may_alias)) intptr_t * istorage = (intptr_t*) &stack.storage;
190	*istorage \|= 0x1;
191	name = "Main Thread";
192	state = Start;
193	starter = 0p;
194	last = 0p;
195	cancellation = 0p;
196	}
197
198	void ?{}( $thread & this, current_stack_info_t * info) with( this ) {
199	ticket = 1;
200	state = Start;
201	self_cor{ info };
202	curr_cor = &self_cor;
203	curr_cluster = mainCluster;
204	self_mon.owner = &this;
205	self_mon.recursion = 1;
206	self_mon_p = &self_mon;
207	link.next = 0p;
208	link.prev = 0p;
209
210	node.next = 0p;
211	node.prev = 0p;
212	doregister(curr_cluster, this);
213
214	monitors{ &self_mon_p, 1, (fptr_t)0 };
215	}
216
217	//-----------------------------------------------------------------------------
218	// Processor coroutine
219	void ?{}(processorCtx_t & this) {
220
221	}
222
223	// Construct the processor context of non-main processors
224	static void ?{}(processorCtx_t & this, processor * proc, current_stack_info_t * info) {
225	(this.__cor){ info };
226	this.proc = proc;
227	}
228
229	static void * __invoke_processor(void * arg);
230
231	void ?{}(processor & this, const char name[], cluster & _cltr) with( this ) {
232	this.name = name;
233	this.cltr = &_cltr;
234	id = -1u;
235	terminated{ 0 };
236	destroyer = 0p;
237	do_terminate = false;
238	preemption_alarm = 0p;
239	pending_preemption = false;
240	runner.proc = &this;
241
242	#if !defined(__CFA_NO_STATISTICS__)
243	print_stats = 0;
244	print_halts = false;
245	#endif
246
247	idle{};
248
249	__cfadbg_print_safe(runtime_core, "Kernel : Starting core %p\n", &this);
250
251	this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
252	__atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
253
254	__cfadbg_print_safe(runtime_core, "Kernel : core %p created\n", &this);
255	}
256
257	void ^?{}(processor & this) with( this ){
258	if( ! __atomic_load_n(&do_terminate, __ATOMIC_ACQUIRE) ) {
259	__cfadbg_print_safe(runtime_core, "Kernel : core %p signaling termination\n", &this);
260
261	__atomic_store_n(&do_terminate, true, __ATOMIC_RELAXED);
262	__wake_proc( &this );
263
264	P( terminated );
265	verify( kernelTLS.this_processor != &this);
266	}
267
268	int err = pthread_join( kernel_thread, 0p );
269	if( err != 0 ) abort("KERNEL ERROR: joining processor %p caused error %s\n", &this, strerror(err));
270
271	free( this.stack );
272
273	__atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
274	}
275
276	void ?{}(cluster & this, const char name[], Duration preemption_rate, unsigned io_flags) with( this ) {
277	this.name = name;
278	this.preemption_rate = preemption_rate;
279	this.nprocessors = 0;
280	ready_queue{};
281
282	#if !defined(__CFA_NO_STATISTICS__)
283	print_stats = 0;
284	stats = alloc();
285	__init_stats( stats );
286	#endif
287
288	threads{ __get };
289
290	__kernel_io_startup( this, io_flags, &this == mainCluster );
291
292	doregister(this);
293	}
294
295	void ^?{}(cluster & this) {
296	__kernel_io_shutdown( this, &this == mainCluster );
297
298	#if !defined(__CFA_NO_STATISTICS__)
299	if( 0 != this.print_stats ) {
300	__print_stats( this.stats, this.print_stats, true, this.name, (void*)&this );
301	}
302	free( this.stats );
303	#endif
304
305	unregister(this);
306	}
307
308	//=============================================================================================
309	// Kernel Scheduling logic
310	//=============================================================================================
311	//Main of the processor contexts
312	void main(processorCtx_t & runner) {
313	// Because of a bug, we couldn't initialized the seed on construction
314	// Do it here
315	kernelTLS.rand_seed ^= rdtscl();
316
317	processor * this = runner.proc;
318	verify(this);
319
320	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
321
322	// register the processor unless it's the main thread which is handled in the boot sequence
323	if(this != mainProcessor) {
324	this->id = doregister((__processor_id_t*)this);
325	#if !defined(__CFA_NO_STATISTICS__)
326	if( this->print_halts ) {
327	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->id, this->name, (void*)this);
328	}
329	#endif
330
331	// Lock the RWlock so no-one pushes/pops while we are changing the queue
332	uint_fast32_t last_size = ready_mutate_lock();
333
334	// Adjust the ready queue size
335	ready_queue_grow( this->cltr );
336
337	// Unlock the RWlock
338	ready_mutate_unlock( last_size );
339	}
340
341	{
342	// Setup preemption data
343	preemption_scope scope = { this };
344
345	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
346
347	$thread * readyThread = 0p;
348	for( unsigned int spin_count = 0; ! __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST); spin_count++ ) {
349	// Try to get the next thread
350	readyThread = __next_thread( this->cltr );
351
352	// Check if we actually found a thread
353	if( readyThread ) {
354	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
355	/* paranoid */ verifyf( readyThread->state == Ready \|\| readyThread->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", readyThread->state, readyThread->preempted);
356	/* paranoid */ verifyf( readyThread->link.next == 0p, "Expected null got %p", readyThread->link.next );
357	__builtin_prefetch( readyThread->context.SP );
358
359	// We found a thread run it
360	__run_thread(this, readyThread);
361
362	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
363	}
364	else {
365	// Block until a thread is ready
366	__halt(this);
367	}
368	}
369
370	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
371	}
372
373	V( this->terminated );
374
375	// unregister the processor unless it's the main thread which is handled in the boot sequence
376	if(this != mainProcessor) {
377	// Lock the RWlock so no-one pushes/pops while we are changing the queue
378	uint_fast32_t last_size = ready_mutate_lock();
379
380	// Adjust the ready queue size
381	ready_queue_shrink( this->cltr );
382
383	// Make sure we aren't on the idle queue
384	#if !defined(__CFA_NO_STATISTICS__)
385	bool removed =
386	#endif
387	unsafe_remove( this->cltr->idles, this );
388
389	#if !defined(__CFA_NO_STATISTICS__)
390	if(removed) __tls_stats()->ready.sleep.exits++;
391	#endif
392
393	// Unlock the RWlock
394	ready_mutate_unlock( last_size );
395
396	// Finally we don't need the read_lock any more
397	unregister((__processor_id_t*)this);
398	}
399	else {
400	// HACK : the coroutine context switch expects this_thread to be set
401	// and it make sense for it to be set in all other cases except here
402	// fake it
403	kernelTLS.this_thread = mainThread;
404	}
405
406	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
407	}
408
409	static int * __volatile_errno() __attribute__((noinline));
410	static int * __volatile_errno() { asm(""); return &errno; }
411
412	// KERNEL ONLY
413	// runThread runs a thread by context switching
414	// from the processor coroutine to the target thread
415	static void __run_thread(processor * this, $thread * thrd_dst) {
416	$coroutine * proc_cor = get_coroutine(this->runner);
417
418	// Update global state
419	kernelTLS.this_thread = thrd_dst;
420
421	// set state of processor coroutine to inactive
422	verify(proc_cor->state == Active);
423	proc_cor->state = Blocked;
424
425	// Actually run the thread
426	RUNNING: while(true) {
427	thrd_dst->preempted = __NO_PREEMPTION;
428	thrd_dst->state = Active;
429
430	__cfaabi_dbg_debug_do(
431	thrd_dst->park_stale = true;
432	thrd_dst->unpark_stale = true;
433	)
434
435	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
436	/* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
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
440	// set context switch to the thread that the processor is executing
441	verify( thrd_dst->context.SP );
442	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
443	// when __cfactx_switch returns we are back in the processor coroutine
444
445	/* 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 );
446	/* 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 );
447	/* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
448	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
449
450
451	// We just finished running a thread, there are a few things that could have happened.
452	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
453	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
454	// 4 - Preempted
455	// In case 1, we may have won a race so we can't write to the state again.
456	// In case 2, we lost the race so we now own the thread.
457
458	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
459	// The thread was preempted, reschedule it and reset the flag
460	__schedule_thread( (__processor_id_t*)this, thrd_dst );
461	break RUNNING;
462	}
463
464	if(unlikely(thrd_dst->state == Halted)) {
465	// The thread has halted, it should never be scheduled/run again
466	// We may need to wake someone up here since
467	unpark( this->destroyer __cfaabi_dbg_ctx2 );
468	this->destroyer = 0p;
469	break RUNNING;
470	}
471
472	/* paranoid */ verify( thrd_dst->state == Active );
473	thrd_dst->state = Blocked;
474
475	// set state of processor coroutine to active and the thread to inactive
476	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
477	__cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
478	switch(old_ticket) {
479	case 1:
480	// This is case 1, the regular case, nothing more is needed
481	break RUNNING;
482	case 2:
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	kernelTLS.this_thread = 0p;
495	}
496
497	// KERNEL_ONLY
498	void returnToKernel() {
499	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
500	$coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
501	$thread * thrd_src = kernelTLS.this_thread;
502
503	#if !defined(__CFA_NO_STATISTICS__)
504	struct processor * last_proc = kernelTLS.this_processor;
505	#endif
506
507	// Run the thread on this processor
508	{
509	int local_errno = *__volatile_errno();
510	#if defined( __i386 ) \|\| defined( __x86_64 )
511	__x87_store;
512	#endif
513	verify( proc_cor->context.SP );
514	__cfactx_switch( &thrd_src->context, &proc_cor->context );
515	#if defined( __i386 ) \|\| defined( __x86_64 )
516	__x87_load;
517	#endif
518	*__volatile_errno() = local_errno;
519	}
520
521	#if !defined(__CFA_NO_STATISTICS__)
522	if(last_proc != kernelTLS.this_processor) {
523	__tls_stats()->ready.threads.migration++;
524	}
525	#endif
526
527	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
528	/* 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 );
529	/* 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 );
530	}
531
532	// KERNEL_ONLY
533	// Context invoker for processors
534	// This is the entry point for processors (kernel threads)
535	// It effectively constructs a coroutine by stealing the pthread stack
536	static void * __invoke_processor(void * arg) {
537	#if !defined( __CFA_NO_STATISTICS__ )
538	__stats_t local_stats;
539	__init_stats( &local_stats );
540	kernelTLS.this_stats = &local_stats;
541	#endif
542
543	processor * proc = (processor *) arg;
544	kernelTLS.this_processor = proc;
545	kernelTLS.this_thread = 0p;
546	kernelTLS.preemption_state.[enabled, disable_count] = [false, 1];
547	// SKULLDUGGERY: We want to create a context for the processor coroutine
548	// which is needed for the 2-step context switch. However, there is no reason
549	// to waste the perfectly valid stack create by pthread.
550	current_stack_info_t info;
551	__stack_t ctx;
552	info.storage = &ctx;
553	(proc->runner){ proc, &info };
554
555	__cfaabi_dbg_print_safe("Coroutine : created stack %p\n", get_coroutine(proc->runner)->stack.storage);
556
557	//Set global state
558	kernelTLS.this_thread = 0p;
559
560	//We now have a proper context from which to schedule threads
561	__cfadbg_print_safe(runtime_core, "Kernel : core %p created (%p, %p)\n", proc, &proc->runner, &ctx);
562
563	// SKULLDUGGERY: Since the coroutine doesn't have its own stack, we can't
564	// resume it to start it like it normally would, it will just context switch
565	// back to here. Instead directly call the main since we already are on the
566	// appropriate stack.
567	get_coroutine(proc->runner)->state = Active;
568	main( proc->runner );
569	get_coroutine(proc->runner)->state = Halted;
570
571	// Main routine of the core returned, the core is now fully terminated
572	__cfadbg_print_safe(runtime_core, "Kernel : core %p main ended (%p)\n", proc, &proc->runner);
573
574	#if !defined(__CFA_NO_STATISTICS__)
575	__tally_stats(proc->cltr->stats, &local_stats);
576	if( 0 != proc->print_stats ) {
577	__print_stats( &local_stats, proc->print_stats, true, proc->name, (void*)proc );
578	}
579	#endif
580
581	return 0p;
582	}
583
584	static void Abort( int ret, const char func[] ) {
585	if ( ret ) { // pthread routines return errno values
586	abort( "%s : internal error, error(%d) %s.", func, ret, strerror( ret ) );
587	} // if
588	} // Abort
589
590	void * __create_pthread( pthread_t * pthread, void * (start)(void ), void * arg ) {
591	pthread_attr_t attr;
592
593	Abort( pthread_attr_init( &attr ), "pthread_attr_init" ); // initialize attribute
594
595	size_t stacksize;
596	// default stack size, normally defined by shell limit
597	Abort( pthread_attr_getstacksize( &attr, &stacksize ), "pthread_attr_getstacksize" );
598	assert( stacksize >= PTHREAD_STACK_MIN );
599
600	void * stack;
601	__cfaabi_dbg_debug_do(
602	stack = memalign( __page_size, stacksize + __page_size );
603	// pthread has no mechanism to create the guard page in user supplied stack.
604	if ( mprotect( stack, __page_size, PROT_NONE ) == -1 ) {
605	abort( "mprotect : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
606	} // if
607	);
608	__cfaabi_dbg_no_debug_do(
609	stack = malloc( stacksize );
610	);
611
612	Abort( pthread_attr_setstack( &attr, stack, stacksize ), "pthread_attr_setstack" );
613
614	Abort( pthread_create( pthread, &attr, start, arg ), "pthread_create" );
615	return stack;
616	}
617
618	// KERNEL_ONLY
619	static void __kernel_first_resume( processor * this ) {
620	$thread * src = mainThread;
621	$coroutine * dst = get_coroutine(this->runner);
622
623	verify( ! kernelTLS.preemption_state.enabled );
624
625	kernelTLS.this_thread->curr_cor = dst;
626	__stack_prepare( &dst->stack, 65000 );
627	__cfactx_start(main, dst, this->runner, __cfactx_invoke_coroutine);
628
629	verify( ! kernelTLS.preemption_state.enabled );
630
631	dst->last = &src->self_cor;
632	dst->starter = dst->starter ? dst->starter : &src->self_cor;
633
634	// make sure the current state is still correct
635	/* paranoid */ verify(src->state == Ready);
636
637	// context switch to specified coroutine
638	verify( dst->context.SP );
639	__cfactx_switch( &src->context, &dst->context );
640	// when __cfactx_switch returns we are back in the src coroutine
641
642	mainThread->curr_cor = &mainThread->self_cor;
643
644	// make sure the current state has been update
645	/* paranoid */ verify(src->state == Active);
646
647	verify( ! kernelTLS.preemption_state.enabled );
648	}
649
650	// KERNEL_ONLY
651	static void __kernel_last_resume( processor * this ) {
652	$coroutine * src = &mainThread->self_cor;
653	$coroutine * dst = get_coroutine(this->runner);
654
655	verify( ! kernelTLS.preemption_state.enabled );
656	verify( dst->starter == src );
657	verify( dst->context.SP );
658
659	// SKULLDUGGERY in debug the processors check that the
660	// stack is still within the limit of the stack limits after running a thread.
661	// that check doesn't make sense if we context switch to the processor using the
662	// coroutine semantics. Since this is a special case, use the current context
663	// info to populate these fields.
664	__cfaabi_dbg_debug_do(
665	__stack_context_t ctx;
666	CtxGet( ctx );
667	mainThread->context.SP = ctx.SP;
668	mainThread->context.FP = ctx.FP;
669	)
670
671	// context switch to the processor
672	__cfactx_switch( &src->context, &dst->context );
673	}
674
675	//-----------------------------------------------------------------------------
676	// Scheduler routines
677	// KERNEL ONLY
678	void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
679	/* paranoid */ verify( thrd );
680	/* paranoid */ verify( thrd->state != Halted );
681	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
682	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
683	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
684	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
685	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
686	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
687	/* paranoid */ #endif
688	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
689
690	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
691
692	ready_schedule_lock ( id );
693	push( thrd->curr_cluster, thrd );
694
695	#if !defined(__CFA_NO_STATISTICS__)
696	bool woke =
697	#endif
698	__wake_one(id, thrd->curr_cluster);
699
700	#if !defined(__CFA_NO_STATISTICS__)
701	if(woke) __tls_stats()->ready.sleep.wakes++;
702	#endif
703	ready_schedule_unlock( id );
704
705	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
706	}
707
708	// KERNEL ONLY
709	static $thread * __next_thread(cluster * this) with( *this ) {
710	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
711
712	ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
713	$thread * head = pop( this );
714	ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
715
716	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
717	return head;
718	}
719
720	// KERNEL ONLY
721	static bool __has_next_thread(cluster * this) with( *this ) {
722	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
723
724	ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
725	bool not_empty = query( this );
726	ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
727
728	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
729	return not_empty;
730	}
731
732	// KERNEL ONLY unpark with out disabling interrupts
733	void __unpark( struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
734	// record activity
735	__cfaabi_dbg_debug_do( char * old_caller = thrd->unpark_caller; )
736	__cfaabi_dbg_record_thrd( *thrd, false, caller );
737
738	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
739	__cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
740	switch(old_ticket) {
741	case 1:
742	// Wake won the race, the thread will reschedule/rerun itself
743	break;
744	case 0:
745	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
746	/* paranoid */ verify( thrd->state == Blocked );
747
748	// Wake lost the race,
749	__schedule_thread( id, thrd );
750	break;
751	default:
752	// This makes no sense, something is wrong abort
753	abort();
754	}
755	}
756
757	void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
758	if( !thrd ) return;
759
760	disable_interrupts();
761	__unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
762	enable_interrupts( __cfaabi_dbg_ctx );
763	}
764
765	void park( __cfaabi_dbg_ctx_param ) {
766	/* paranoid */ verify( kernelTLS.preemption_state.enabled );
767	disable_interrupts();
768	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
769	/* paranoid */ verify( kernelTLS.this_thread->preempted == __NO_PREEMPTION );
770
771	// record activity
772	__cfaabi_dbg_record_thrd( *kernelTLS.this_thread, true, caller );
773
774	returnToKernel();
775
776	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
777	enable_interrupts( __cfaabi_dbg_ctx );
778	/* paranoid */ verify( kernelTLS.preemption_state.enabled );
779
780	}
781
782	// KERNEL ONLY
783	void __leave_thread() {
784	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
785	returnToKernel();
786	abort();
787	}
788
789	// KERNEL ONLY
790	bool force_yield( __Preemption_Reason reason ) {
791	/* paranoid */ verify( kernelTLS.preemption_state.enabled );
792	disable_interrupts();
793	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
794
795	$thread * thrd = kernelTLS.this_thread;
796	/* paranoid */ verify(thrd->state == Active);
797
798	// SKULLDUGGERY: It is possible that we are preempting this thread just before
799	// it was going to park itself. If that is the case and it is already using the
800	// intrusive fields then we can't use them to preempt the thread
801	// If that is the case, abandon the preemption.
802	bool preempted = false;
803	if(thrd->link.next == 0p) {
804	preempted = true;
805	thrd->preempted = reason;
806	returnToKernel();
807	}
808
809	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
810	enable_interrupts_noPoll();
811	/* paranoid */ verify( kernelTLS.preemption_state.enabled );
812
813	return preempted;
814	}
815
816	//=============================================================================================
817	// Kernel Setup logic
818	//=============================================================================================
819	//-----------------------------------------------------------------------------
820	// Kernel boot procedures
821	static void __kernel_startup(void) {
822	verify( ! kernelTLS.preemption_state.enabled );
823	__cfadbg_print_safe(runtime_core, "Kernel : Starting\n");
824
825	__page_size = sysconf( _SC_PAGESIZE );
826
827	__cfa_dbg_global_clusters.list{ __get };
828	__cfa_dbg_global_clusters.lock{};
829
830	// Initialize the global scheduler lock
831	__scheduler_lock = (__scheduler_RWLock_t*)&storage___scheduler_lock;
832	(*__scheduler_lock){};
833
834	// Initialize the main cluster
835	mainCluster = (cluster *)&storage_mainCluster;
836	(*mainCluster){"Main Cluster"};
837
838	__cfadbg_print_safe(runtime_core, "Kernel : Main cluster ready\n");
839
840	// Start by initializing the main thread
841	// SKULLDUGGERY: the mainThread steals the process main thread
842	// which will then be scheduled by the mainProcessor normally
843	mainThread = ($thread *)&storage_mainThread;
844	current_stack_info_t info;
845	info.storage = (__stack_t*)&storage_mainThreadCtx;
846	(*mainThread){ &info };
847
848	__cfadbg_print_safe(runtime_core, "Kernel : Main thread ready\n");
849
850
851
852	// Construct the processor context of the main processor
853	void ?{}(processorCtx_t & this, processor * proc) {
854	(this.__cor){ "Processor" };
855	this.__cor.starter = 0p;
856	this.proc = proc;
857	}
858
859	void ?{}(processor & this) with( this ) {
860	name = "Main Processor";
861	cltr = mainCluster;
862	terminated{ 0 };
863	do_terminate = false;
864	preemption_alarm = 0p;
865	pending_preemption = false;
866	kernel_thread = pthread_self();
867	id = -1u;
868
869	#if !defined(__CFA_NO_STATISTICS__)
870	print_stats = false;
871	print_halts = false;
872	#endif
873
874	runner{ &this };
875	__cfadbg_print_safe(runtime_core, "Kernel : constructed main processor context %p\n", &runner);
876
877	__atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
878	}
879
880	// Initialize the main processor and the main processor ctx
881	// (the coroutine that contains the processing control flow)
882	mainProcessor = (processor *)&storage_mainProcessor;
883	(*mainProcessor){};
884
885	mainProcessor->id = doregister( (__processor_id_t*)mainProcessor);
886
887	//initialize the global state variables
888	kernelTLS.this_processor = mainProcessor;
889	kernelTLS.this_thread = mainThread;
890
891	#if !defined( __CFA_NO_STATISTICS__ )
892	kernelTLS.this_stats = (__stats_t *)& storage_mainProcStats;
893	__init_stats( kernelTLS.this_stats );
894	#endif
895
896	// Enable preemption
897	kernel_start_preemption();
898
899	// Add the main thread to the ready queue
900	// once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
901	__schedule_thread((__processor_id_t *)mainProcessor, mainThread);
902
903	// SKULLDUGGERY: Force a context switch to the main processor to set the main thread's context to the current UNIX
904	// context. Hence, the main thread does not begin through __cfactx_invoke_thread, like all other threads. The trick here is that
905	// mainThread is on the ready queue when this call is made.
906	__kernel_first_resume( kernelTLS.this_processor );
907
908
909	// THE SYSTEM IS NOW COMPLETELY RUNNING
910
911
912	// Now that the system is up, finish creating systems that need threading
913	__kernel_io_finish_start( *mainCluster );
914
915
916	__cfadbg_print_safe(runtime_core, "Kernel : Started\n--------------------------------------------------\n\n");
917
918	verify( ! kernelTLS.preemption_state.enabled );
919	enable_interrupts( __cfaabi_dbg_ctx );
920	verify( TL_GET( preemption_state.enabled ) );
921	}
922
923	static void __kernel_shutdown(void) {
924	//Before we start shutting things down, wait for systems that need threading to shutdown
925	__kernel_io_prepare_stop( *mainCluster );
926
927	/* paranoid */ verify( TL_GET( preemption_state.enabled ) );
928	disable_interrupts();
929	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
930
931	__cfadbg_print_safe(runtime_core, "\n--------------------------------------------------\nKernel : Shutting down\n");
932
933	// SKULLDUGGERY: Notify the mainProcessor it needs to terminates.
934	// When its coroutine terminates, it return control to the mainThread
935	// which is currently here
936	__atomic_store_n(&mainProcessor->do_terminate, true, __ATOMIC_RELEASE);
937	__kernel_last_resume( kernelTLS.this_processor );
938	mainThread->self_cor.state = Halted;
939
940	// THE SYSTEM IS NOW COMPLETELY STOPPED
941
942	// Disable preemption
943	kernel_stop_preemption();
944
945	unregister((__processor_id_t*)mainProcessor);
946
947	// Destroy the main processor and its context in reverse order of construction
948	// These were manually constructed so we need manually destroy them
949	void ^?{}(processor & this) with( this ){
950	/* paranoid */ verify( this.do_terminate == true );
951	__atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
952	__cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
953	}
954
955	^(*mainProcessor){};
956
957	// Final step, destroy the main thread since it is no longer needed
958
959	// Since we provided a stack to this taxk it will not destroy anything
960	/* paranoid / verify(mainThread->self_cor.stack.storage == (__stack_t)(((uintptr_t)&storage_mainThreadCtx)\| 0x1));
961	^(*mainThread){};
962
963	^(*mainCluster){};
964
965	^(*__scheduler_lock){};
966
967	^(__cfa_dbg_global_clusters.list){};
968	^(__cfa_dbg_global_clusters.lock){};
969
970	__cfadbg_print_safe(runtime_core, "Kernel : Shutdown complete\n");
971	}
972
973	//=============================================================================================
974	// Kernel Idle Sleep
975	//=============================================================================================
976	// Wake a thread from the front if there are any
977	static bool __wake_one(struct __processor_id_t * id, cluster * this) {
978	/* paranoid */ verify( ready_schedule_islocked( id ) );
979
980	// Check if there is a sleeping processor
981	processor * p = pop(this->idles);
982
983	// If no one is sleeping, we are done
984	if( 0p == p ) return false;
985
986	// We found a processor, wake it up
987	post( p->idle );
988
989	return true;
990	}
991
992	// Unconditionnaly wake a thread
993	static bool __wake_proc(processor * this) {
994	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
995
996	disable_interrupts();
997	/* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
998	bool ret = post( this->idle );
999	enable_interrupts( __cfaabi_dbg_ctx );
1000
1001	return ret;
1002	}
1003
1004	static void __halt(processor * this) with( *this ) {
1005	if( do_terminate ) return;
1006
1007	#if !defined(__CFA_NO_STATISTICS__)
1008	__tls_stats()->ready.sleep.halts++;
1009	#endif
1010	// Push self to queue
1011	push(cltr->idles, *this);
1012
1013	// Makre sure we don't miss a thread
1014	if( __has_next_thread(cltr) ) {
1015	// A thread was posted, make sure a processor is woken up
1016	struct __processor_id_t id = (struct __processor_id_t ) this;
1017	ready_schedule_lock ( id );
1018	__wake_one( id, cltr );
1019	ready_schedule_unlock( id );
1020	#if !defined(__CFA_NO_STATISTICS__)
1021	__tls_stats()->ready.sleep.cancels++;
1022	#endif
1023	}
1024
1025	#if !defined(__CFA_NO_STATISTICS__)
1026	if(this->print_halts) {
1027	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
1028	}
1029	#endif
1030
1031	wait( idle );
1032
1033	#if !defined(__CFA_NO_STATISTICS__)
1034	if(this->print_halts) {
1035	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
1036	}
1037	#endif
1038	}
1039
1040	//=============================================================================================
1041	// Unexpected Terminating logic
1042	//=============================================================================================
1043	static __spinlock_t kernel_abort_lock;
1044	static bool kernel_abort_called = false;
1045
1046	void * kernel_abort(void) __attribute__ ((__nothrow__)) {
1047	// abort cannot be recursively entered by the same or different processors because all signal handlers return when
1048	// the globalAbort flag is true.
1049	lock( kernel_abort_lock __cfaabi_dbg_ctx2 );
1050
1051	// first task to abort ?
1052	if ( kernel_abort_called ) { // not first task to abort ?
1053	unlock( kernel_abort_lock );
1054
1055	sigset_t mask;
1056	sigemptyset( &mask );
1057	sigaddset( &mask, SIGALRM ); // block SIGALRM signals
1058	sigaddset( &mask, SIGUSR1 ); // block SIGALRM signals
1059	sigsuspend( &mask ); // block the processor to prevent further damage during abort
1060	_exit( EXIT_FAILURE ); // if processor unblocks before it is killed, terminate it
1061	}
1062	else {
1063	kernel_abort_called = true;
1064	unlock( kernel_abort_lock );
1065	}
1066
1067	return kernelTLS.this_thread;
1068	}
1069
1070	void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
1071	$thread * thrd = kernel_data;
1072
1073	if(thrd) {
1074	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
1075	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1076
1077	if ( &thrd->self_cor != thrd->curr_cor ) {
1078	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
1079	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1080	}
1081	else {
1082	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
1083	}
1084	}
1085	else {
1086	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
1087	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1088	}
1089	}
1090
1091	int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
1092	return get_coroutine(kernelTLS.this_thread) == get_coroutine(mainThread) ? 4 : 2;
1093	}
1094
1095	static __spinlock_t kernel_debug_lock;
1096
1097	extern "C" {
1098	void __cfaabi_bits_acquire() {
1099	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
1100	}
1101
1102	void __cfaabi_bits_release() {
1103	unlock( kernel_debug_lock );
1104	}
1105	}
1106
1107	//=============================================================================================
1108	// Kernel Utilities
1109	//=============================================================================================
1110	//-----------------------------------------------------------------------------
1111	// Locks
1112	void ?{}( semaphore & this, int count = 1 ) {
1113	(this.lock){};
1114	this.count = count;
1115	(this.waiting){};
1116	}
1117	void ^?{}(semaphore & this) {}
1118
1119	bool P(semaphore & this) with( this ){
1120	lock( lock __cfaabi_dbg_ctx2 );
1121	count -= 1;
1122	if ( count < 0 ) {
1123	// queue current task
1124	append( waiting, kernelTLS.this_thread );
1125
1126	// atomically release spin lock and block
1127	unlock( lock );
1128	park( __cfaabi_dbg_ctx );
1129	return true;
1130	}
1131	else {
1132	unlock( lock );
1133	return false;
1134	}
1135	}
1136
1137	bool V(semaphore & this) with( this ) {
1138	$thread * thrd = 0p;
1139	lock( lock __cfaabi_dbg_ctx2 );
1140	count += 1;
1141	if ( count <= 0 ) {
1142	// remove task at head of waiting list
1143	thrd = pop_head( waiting );
1144	}
1145
1146	unlock( lock );
1147
1148	// make new owner
1149	unpark( thrd __cfaabi_dbg_ctx2 );
1150
1151	return thrd != 0p;
1152	}
1153
1154	bool V(semaphore & this, unsigned diff) with( this ) {
1155	$thread * thrd = 0p;
1156	lock( lock __cfaabi_dbg_ctx2 );
1157	int release = max(-count, (int)diff);
1158	count += diff;
1159	for(release) {
1160	unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 );
1161	}
1162
1163	unlock( lock );
1164
1165	return thrd != 0p;
1166	}
1167
1168	//-----------------------------------------------------------------------------
1169	// Global Queues
1170	void doregister( cluster & cltr ) {
1171	lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1172	push_front( __cfa_dbg_global_clusters.list, cltr );
1173	unlock ( __cfa_dbg_global_clusters.lock );
1174	}
1175
1176	void unregister( cluster & cltr ) {
1177	lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1178	remove( __cfa_dbg_global_clusters.list, cltr );
1179	unlock( __cfa_dbg_global_clusters.lock );
1180	}
1181
1182	void doregister( cluster * cltr, $thread & thrd ) {
1183	lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1184	cltr->nthreads += 1;
1185	push_front(cltr->threads, thrd);
1186	unlock (cltr->thread_list_lock);
1187	}
1188
1189	void unregister( cluster * cltr, $thread & thrd ) {
1190	lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1191	remove(cltr->threads, thrd );
1192	cltr->nthreads -= 1;
1193	unlock(cltr->thread_list_lock);
1194	}
1195
1196	//-----------------------------------------------------------------------------
1197	// Debug
1198	__cfaabi_dbg_debug_do(
1199	extern "C" {
1200	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
1201	this.prev_name = prev_name;
1202	this.prev_thrd = kernelTLS.this_thread;
1203	}
1204
1205	void __cfaabi_dbg_record_thrd($thread & this, bool park, const char prev_name[]) {
1206	if(park) {
1207	this.park_caller = prev_name;
1208	this.park_stale = false;
1209	}
1210	else {
1211	this.unpark_caller = prev_name;
1212	this.unpark_stale = false;
1213	}
1214	}
1215	}
1216	)
1217
1218	//-----------------------------------------------------------------------------
1219	// Debug
1220	bool threading_enabled(void) __attribute__((const)) {
1221	return true;
1222	}
1223
1224	//-----------------------------------------------------------------------------
1225	// Statistics
1226	#if !defined(__CFA_NO_STATISTICS__)
1227	void print_halts( processor & this ) {
1228	this.print_halts = true;
1229	}
1230	#endif
1231	// Local Variables: //
1232	// mode: c //
1233	// tab-width: 4 //
1234	// End: //

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

Download in other formats:

Original Format