Context Navigation

kernel.cfa @ 040334e

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since 040334e was 040334e, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Removed so-called 'new proc main' which wasn't useful.
Property mode set to `100644`
File size: 28.4 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 <string.h>
25	#include <signal.h>
26	#include <unistd.h>
27	extern "C" {
28	#include <sys/eventfd.h>
29	#include <sys/uio.h>
30	}
31
32	//CFA Includes
33	#include "kernel_private.hfa"
34	#include "preemption.hfa"
35	#include "strstream.hfa"
36	#include "device/cpu.hfa"
37	#include "io/types.hfa"
38
39	//Private includes
40	#define __CFA_INVOKE_PRIVATE__
41	#include "invoke.h"
42
43	#if !defined(__CFA_NO_STATISTICS__)
44	#define __STATS( ...) __VA_ARGS__
45	#else
46	#define __STATS( ...)
47	#endif
48
49	//-----------------------------------------------------------------------------
50	// Some assembly required
51	#if defined( __i386 )
52	// mxcr : SSE Status and Control bits (control bits are preserved across function calls)
53	// fcw : X87 FPU control word (preserved across function calls)
54	#define __x87_store \
55	uint32_t __mxcr; \
56	uint16_t __fcw; \
57	__asm__ volatile ( \
58	"stmxcsr %0\n" \
59	"fnstcw %1\n" \
60	: "=m" (__mxcr),\
61	"=m" (__fcw) \
62	)
63
64	#define __x87_load \
65	__asm__ volatile ( \
66	"fldcw %1\n" \
67	"ldmxcsr %0\n" \
68	::"m" (__mxcr),\
69	"m" (__fcw) \
70	)
71
72	#elif defined( __x86_64 )
73	#define __x87_store \
74	uint32_t __mxcr; \
75	uint16_t __fcw; \
76	__asm__ volatile ( \
77	"stmxcsr %0\n" \
78	"fnstcw %1\n" \
79	: "=m" (__mxcr),\
80	"=m" (__fcw) \
81	)
82
83	#define __x87_load \
84	__asm__ volatile ( \
85	"fldcw %1\n" \
86	"ldmxcsr %0\n" \
87	:: "m" (__mxcr),\
88	"m" (__fcw) \
89	)
90
91	#elif defined( __arm__ )
92	#define __x87_store
93	#define __x87_load
94
95	#elif defined( __aarch64__ )
96	#define __x87_store \
97	uint32_t __fpcntl[2]; \
98	__asm__ volatile ( \
99	"mrs x9, FPCR\n" \
100	"mrs x10, FPSR\n" \
101	"stp x9, x10, %0\n" \
102	: "=m" (__fpcntl) : : "x9", "x10" \
103	)
104
105	#define __x87_load \
106	__asm__ volatile ( \
107	"ldp x9, x10, %0\n" \
108	"msr FPSR, x10\n" \
109	"msr FPCR, x9\n" \
110	: "=m" (__fpcntl) : : "x9", "x10" \
111	)
112
113	#else
114	#error unsupported hardware architecture
115	#endif
116
117	extern thread$ * mainThread;
118	extern processor * mainProcessor;
119
120	//-----------------------------------------------------------------------------
121	// Kernel Scheduling logic
122	static thread$ * __next_thread(cluster * this);
123	static thread$ * __next_thread_slow(cluster * this);
124	static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
125	static void __run_thread(processor * this, thread$ * dst);
126	static void __wake_one(cluster * cltr);
127
128	static void idle_sleep(processor * proc, io_future_t & future, iovec & iov);
129	static bool mark_idle (__cluster_proc_list & idles, processor & proc);
130	static void mark_awake(__cluster_proc_list & idles, processor & proc);
131
132	extern void __cfa_io_start( processor * );
133	extern bool __cfa_io_drain( processor * );
134	extern bool __cfa_io_flush( processor *, int min_comp );
135	extern void __cfa_io_stop ( processor * );
136	static inline bool __maybe_io_drain( processor * );
137
138	#if defined(CFA_WITH_IO_URING_IDLE)
139	extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
140	#endif
141
142	extern void __disable_interrupts_hard();
143	extern void __enable_interrupts_hard();
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	io_future_t future; // used for idle sleep when io_uring is present
161	future.self.ptr = 1p; // mark it as already fulfilled so we know if there is a pending request or not
162	eventfd_t idle_val;
163	iovec idle_iovec = { &idle_val, sizeof(idle_val) };
164
165	__cfa_io_start( this );
166
167	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
168	#if !defined(__CFA_NO_STATISTICS__)
169	if( this->print_halts ) {
170	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
171	}
172	#endif
173
174	{
175	// Setup preemption data
176	preemption_scope scope = { this };
177
178	// if we need to run some special setup, now is the time to do it.
179	if(this->init.thrd) {
180	this->init.thrd->curr_cluster = this->cltr;
181	__run_thread(this, this->init.thrd);
182	}
183
184	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
185
186	thread$ * readyThread = 0p;
187	MAIN_LOOP:
188	for() {
189	// Check if there is pending io
190	__maybe_io_drain( this );
191
192	// Try to get the next thread
193	readyThread = __next_thread( this->cltr );
194
195	if( !readyThread ) {
196	__tls_stats()->io.flush.idle++;
197	__cfa_io_flush( this, 0 );
198
199	readyThread = __next_thread_slow( this->cltr );
200	}
201
202	HALT:
203	if( !readyThread ) {
204	// Don't block if we are done
205	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
206
207	// Push self to idle stack
208	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
209
210	// Confirm the ready-queue is empty
211	readyThread = __next_thread_slow( this->cltr );
212	if( readyThread ) {
213	// A thread was found, cancel the halt
214	mark_awake(this->cltr->procs, * this);
215
216	#if !defined(__CFA_NO_STATISTICS__)
217	__tls_stats()->ready.sleep.cancels++;
218	#endif
219
220	// continue the mai loop
221	break HALT;
222	}
223
224	idle_sleep( this, future, idle_iovec );
225
226	// We were woken up, remove self from idle
227	mark_awake(this->cltr->procs, * this);
228
229	// DON'T just proceed, start looking again
230	continue MAIN_LOOP;
231	}
232
233	/* paranoid */ verify( readyThread );
234
235	// Reset io dirty bit
236	this->io.dirty = false;
237
238	// We found a thread run it
239	__run_thread(this, readyThread);
240
241	// Are we done?
242	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
243
244	if(this->io.pending && !this->io.dirty) {
245	__tls_stats()->io.flush.dirty++;
246	__cfa_io_flush( this, 0 );
247	}
248	}
249
250	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
251	}
252
253	for(int i = 0; !available(future); i++) {
254	if(i > 1000) __cfaabi_dbg_write( "ERROR: kernel has bin spinning on a flush after exit loop.\n", 60);
255	__cfa_io_flush( this, 1 );
256	}
257
258	__cfa_io_stop( this );
259
260	post( this->terminated );
261
262	if(this == mainProcessor) {
263	// HACK : the coroutine context switch expects this_thread to be set
264	// and it make sense for it to be set in all other cases except here
265	// fake it
266	__cfaabi_tls.this_thread = mainThread;
267	}
268
269	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
270	}
271
272	static int * __volatile_errno() __attribute__((noinline));
273	static int * __volatile_errno() { asm(""); return &errno; }
274
275	// KERNEL ONLY
276	// runThread runs a thread by context switching
277	// from the processor coroutine to the target thread
278	static void __run_thread(processor * this, thread$ * thrd_dst) {
279	/* paranoid */ verify( ! __preemption_enabled() );
280	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
281	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
282	__builtin_prefetch( thrd_dst->context.SP );
283
284	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
285
286	coroutine$ * proc_cor = get_coroutine(this->runner);
287
288	// set state of processor coroutine to inactive
289	verify(proc_cor->state == Active);
290	proc_cor->state = Blocked;
291
292	// Actually run the thread
293	RUNNING: while(true) {
294	thrd_dst->preempted = __NO_PREEMPTION;
295	thrd_dst->state = Active;
296
297	// Update global state
298	kernelTLS().this_thread = thrd_dst;
299
300	/* paranoid */ verify( ! __preemption_enabled() );
301	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
302	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
303	/* paranoid */ verify( thrd_dst->context.SP );
304	/* paranoid */ verify( thrd_dst->state != Halted );
305	/* 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
306	/* 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
307	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
308
309
310
311	// set context switch to the thread that the processor is executing
312	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
313	// when __cfactx_switch returns we are back in the processor coroutine
314
315
316
317	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
318	/* 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 );
319	/* 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 );
320	/* paranoid */ verify( thrd_dst->context.SP );
321	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
322	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
323	/* paranoid */ verify( ! __preemption_enabled() );
324
325	// Reset global state
326	kernelTLS().this_thread = 0p;
327
328	// We just finished running a thread, there are a few things that could have happened.
329	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
330	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
331	// 4 - Preempted
332	// In case 1, we may have won a race so we can't write to the state again.
333	// In case 2, we lost the race so we now own the thread.
334
335	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
336	// The thread was preempted, reschedule it and reset the flag
337	schedule_thread$( thrd_dst, UNPARK_LOCAL );
338	break RUNNING;
339	}
340
341	if(unlikely(thrd_dst->state == Halting)) {
342	// The thread has halted, it should never be scheduled/run again
343	// finish the thread
344	__thread_finish( thrd_dst );
345	break RUNNING;
346	}
347
348	/* paranoid */ verify( thrd_dst->state == Active );
349	thrd_dst->state = Blocked;
350
351	// set state of processor coroutine to active and the thread to inactive
352	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
353	switch(old_ticket) {
354	case TICKET_RUNNING:
355	// This is case 1, the regular case, nothing more is needed
356	break RUNNING;
357	case TICKET_UNBLOCK:
358	#if !defined(__CFA_NO_STATISTICS__)
359	__tls_stats()->ready.threads.threads++;
360	#endif
361	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
362	// In this case, just run it again.
363	continue RUNNING;
364	default:
365	// This makes no sense, something is wrong abort
366	abort();
367	}
368	}
369
370	// Just before returning to the processor, set the processor coroutine to active
371	proc_cor->state = Active;
372
373	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
374
375	#if !defined(__CFA_NO_STATISTICS__)
376	__tls_stats()->ready.threads.threads--;
377	#endif
378
379	/* paranoid */ verify( ! __preemption_enabled() );
380	}
381
382	// KERNEL_ONLY
383	void returnToKernel() {
384	/* paranoid */ verify( ! __preemption_enabled() );
385	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
386	thread$ * thrd_src = kernelTLS().this_thread;
387
388	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
389
390	// Run the thread on this processor
391	{
392	int local_errno = *__volatile_errno();
393	#if defined( __i386 ) \|\| defined( __x86_64 )
394	__x87_store;
395	#endif
396	/* paranoid */ verify( proc_cor->context.SP );
397	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
398	__cfactx_switch( &thrd_src->context, &proc_cor->context );
399	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
400	#if defined( __i386 ) \|\| defined( __x86_64 )
401	__x87_load;
402	#endif
403	*__volatile_errno() = local_errno;
404	}
405
406	#if !defined(__CFA_NO_STATISTICS__)
407	/* paranoid */ verify( thrd_src->last_proc != 0p );
408	if(thrd_src->last_proc != kernelTLS().this_processor) {
409	__tls_stats()->ready.threads.migration++;
410	}
411	#endif
412
413	/* paranoid */ verify( ! __preemption_enabled() );
414	/* 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 );
415	/* 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 );
416	}
417
418	//-----------------------------------------------------------------------------
419	// Scheduler routines
420	// KERNEL ONLY
421	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
422	/* paranoid */ verify( ! __preemption_enabled() );
423	/* paranoid */ verify( ready_schedule_islocked());
424	/* paranoid */ verify( thrd );
425	/* paranoid */ verify( thrd->state != Halted );
426	/* paranoid */ verify( thrd->curr_cluster );
427	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
428	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
429	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
430	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
431	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
432	/* paranoid */ #endif
433	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
434	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
435
436	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
437
438	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
439	struct cluster * cl = thrd->curr_cluster;
440	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
441
442	// push the thread to the cluster ready-queue
443	push( cl, thrd, hint );
444
445	// variable thrd is no longer safe to use
446	thrd = 0xdeaddeaddeaddeadp;
447
448	// wake the cluster using the save variable.
449	__wake_one( cl );
450
451	#if !defined(__CFA_NO_STATISTICS__)
452	if( kernelTLS().this_stats ) {
453	__tls_stats()->ready.threads.threads++;
454	if(outside) {
455	__tls_stats()->ready.threads.extunpark++;
456	}
457	}
458	else {
459	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
460	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
461	}
462	#endif
463
464	/* paranoid */ verify( ready_schedule_islocked());
465	/* paranoid */ verify( ! __preemption_enabled() );
466	}
467
468	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
469	ready_schedule_lock();
470	__schedule_thread( thrd, hint );
471	ready_schedule_unlock();
472	}
473
474	// KERNEL ONLY
475	static inline thread$ * __next_thread(cluster * this) with( *this ) {
476	/* paranoid */ verify( ! __preemption_enabled() );
477
478	ready_schedule_lock();
479	thread$ * thrd = pop_fast( this );
480	ready_schedule_unlock();
481
482	/* paranoid */ verify( ! __preemption_enabled() );
483	return thrd;
484	}
485
486	// KERNEL ONLY
487	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
488	/* paranoid */ verify( ! __preemption_enabled() );
489
490	ready_schedule_lock();
491	thread$ * thrd;
492	for(25) {
493	thrd = pop_slow( this );
494	if(thrd) goto RET;
495	}
496	thrd = pop_search( this );
497
498	RET:
499	ready_schedule_unlock();
500
501	/* paranoid */ verify( ! __preemption_enabled() );
502	return thrd;
503	}
504
505	static inline bool __must_unpark( thread$ * thrd ) {
506	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
507	switch(old_ticket) {
508	case TICKET_RUNNING:
509	// Wake won the race, the thread will reschedule/rerun itself
510	return false;
511	case TICKET_BLOCKED:
512	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
513	/* paranoid */ verify( thrd->state == Blocked );
514	return true;
515	default:
516	// This makes no sense, something is wrong abort
517	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
518	}
519	}
520
521	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
522	/* paranoid */ verify( ! __preemption_enabled() );
523	/* paranoid */ verify( ready_schedule_islocked());
524
525	if( !thrd ) return;
526
527	if(__must_unpark(thrd)) {
528	// Wake lost the race,
529	__schedule_thread( thrd, hint );
530	}
531
532	/* paranoid */ verify( ready_schedule_islocked());
533	/* paranoid */ verify( ! __preemption_enabled() );
534	}
535
536	void unpark( thread$ * thrd, unpark_hint hint ) {
537	if( !thrd ) return;
538
539	if(__must_unpark(thrd)) {
540	disable_interrupts();
541	// Wake lost the race,
542	schedule_thread$( thrd, hint );
543	enable_interrupts(false);
544	}
545	}
546
547	void park( void ) {
548	__disable_interrupts_checked();
549	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
550	returnToKernel();
551	__enable_interrupts_checked();
552
553	}
554
555	extern "C" {
556	// Leave the thread monitor
557	// last routine called by a thread.
558	// Should never return
559	void __cfactx_thrd_leave() {
560	thread$ * thrd = active_thread();
561	monitor$ * this = &thrd->self_mon;
562
563	// Lock the monitor now
564	lock( this->lock __cfaabi_dbg_ctx2 );
565
566	disable_interrupts();
567
568	/* paranoid */ verify( ! __preemption_enabled() );
569	/* paranoid */ verify( thrd->state == Active );
570	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
571	/* paranoid */ verify( kernelTLS().this_thread == thrd );
572	/* paranoid */ verify( thrd->context.SP );
573	/* 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 );
574	/* 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 );
575
576	thrd->state = Halting;
577	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
578	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
579	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
580
581	// Leave the thread
582	returnToKernel();
583
584	// Control flow should never reach here!
585	abort();
586	}
587	}
588
589	// KERNEL ONLY
590	bool force_yield( __Preemption_Reason reason ) {
591	__disable_interrupts_checked();
592	thread$ * thrd = kernelTLS().this_thread;
593	/* paranoid */ verify(thrd->state == Active);
594
595	// SKULLDUGGERY: It is possible that we are preempting this thread just before
596	// it was going to park itself. If that is the case and it is already using the
597	// intrusive fields then we can't use them to preempt the thread
598	// If that is the case, abandon the preemption.
599	bool preempted = false;
600	if(thrd->link.next == 0p) {
601	preempted = true;
602	thrd->preempted = reason;
603	returnToKernel();
604	}
605	__enable_interrupts_checked( false );
606	return preempted;
607	}
608
609	//=============================================================================================
610	// Kernel Idle Sleep
611	//=============================================================================================
612	// Wake a thread from the front if there are any
613	static void __wake_one(cluster * this) {
614	eventfd_t val;
615
616	/* paranoid */ verify( ! __preemption_enabled() );
617	/* paranoid */ verify( ready_schedule_islocked() );
618
619	// Check if there is a sleeping processor
620	struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
621
622	// If no one is sleeping: we are done
623	if( fdp == 0p ) return;
624
625	int fd = 1;
626	if( __atomic_load_n(&fdp->fd, __ATOMIC_SEQ_CST) != 1 ) {
627	fd = __atomic_exchange_n(&fdp->fd, 1, __ATOMIC_RELAXED);
628	}
629
630	switch(fd) {
631	case 0:
632	// If the processor isn't ready to sleep then the exchange will already wake it up
633	#if !defined(__CFA_NO_STATISTICS__)
634	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
635	} else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
636	#endif
637	break;
638	case 1:
639	// If someone else already said they will wake them: we are done
640	#if !defined(__CFA_NO_STATISTICS__)
641	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
642	} else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
643	#endif
644	break;
645	default:
646	// If the processor was ready to sleep, we need to wake it up with an actual write
647	val = 1;
648	eventfd_write( fd, val );
649
650	#if !defined(__CFA_NO_STATISTICS__)
651	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
652	} else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
653	#endif
654	break;
655	}
656
657	/* paranoid */ verify( ready_schedule_islocked() );
658	/* paranoid */ verify( ! __preemption_enabled() );
659
660	return;
661	}
662
663	// Unconditionnaly wake a thread
664	void __wake_proc(processor * this) {
665	/* paranoid */ verify( ! __preemption_enabled() );
666
667	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
668
669	this->idle_wctx.fd = 1;
670
671	eventfd_t val;
672	val = 1;
673	eventfd_write( this->idle_fd, val );
674
675	/* paranoid */ verify( ! __preemption_enabled() );
676	}
677
678	static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
679	// Tell everyone we are ready to go do sleep
680	for() {
681	int expected = this->idle_wctx.fd;
682
683	// Someone already told us to wake-up! No time for a nap.
684	if(expected == 1) { return; }
685
686	// Try to mark that we are going to sleep
687	if(__atomic_compare_exchange_n(&this->idle_wctx.fd, &expected, this->idle_fd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
688	// Every one agreed, taking a nap
689	break;
690	}
691	}
692
693
694	#if !defined(CFA_WITH_IO_URING_IDLE)
695	#if !defined(__CFA_NO_STATISTICS__)
696	if(this->print_halts) {
697	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
698	}
699	#endif
700
701	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
702
703	{
704	eventfd_t val;
705	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
706	if(ret < 0) {
707	switch((int)errno) {
708	case EAGAIN:
709	#if EAGAIN != EWOULDBLOCK
710	case EWOULDBLOCK:
711	#endif
712	case EINTR:
713	// No need to do anything special here, just assume it's a legitimate wake-up
714	break;
715	default:
716	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
717	}
718	}
719	}
720
721	#if !defined(__CFA_NO_STATISTICS__)
722	if(this->print_halts) {
723	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
724	}
725	#endif
726	#else
727	// Do we already have a pending read
728	if(available(future)) {
729	// There is no pending read, we need to add one
730	reset(future);
731
732	__kernel_read(this, future, iov, this->idle_fd );
733	}
734
735	__cfa_io_flush( this, 1 );
736	#endif
737	}
738
739	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
740	#if !defined(__CFA_NO_STATISTICS__)
741	__tls_stats()->ready.sleep.halts++;
742	#endif
743
744	proc.idle_wctx.fd = 0;
745
746	/* paranoid */ verify( ! __preemption_enabled() );
747	if(!try_lock( this )) return false;
748	this.idle++;
749	/* paranoid */ verify( this.idle <= this.total );
750	remove(proc);
751	insert_first(this.idles, proc);
752
753	__atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
754	unlock( this );
755	/* paranoid */ verify( ! __preemption_enabled() );
756
757	return true;
758	}
759
760	static void mark_awake(__cluster_proc_list & this, processor & proc) {
761	/* paranoid */ verify( ! __preemption_enabled() );
762	lock( this );
763	this.idle--;
764	/* paranoid */ verify( this.idle >= 0 );
765	remove(proc);
766	insert_last(this.actives, proc);
767
768	{
769	struct __fd_waitctx * wctx = 0;
770	if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
771	__atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
772	}
773
774	unlock( this );
775	/* paranoid */ verify( ! __preemption_enabled() );
776	}
777
778	//=============================================================================================
779	// Unexpected Terminating logic
780	//=============================================================================================
781	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
782	thread$ * thrd = __cfaabi_tls.this_thread;
783
784	if(thrd) {
785	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
786	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
787
788	if ( &thrd->self_cor != thrd->curr_cor ) {
789	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
790	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
791	}
792	else {
793	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
794	}
795	}
796	else {
797	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
798	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
799	}
800	}
801
802	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
803	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
804	}
805
806	static __spinlock_t kernel_debug_lock;
807
808	extern "C" {
809	void __cfaabi_bits_acquire() {
810	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
811	}
812
813	void __cfaabi_bits_release() {
814	unlock( kernel_debug_lock );
815	}
816	}
817
818	//=============================================================================================
819	// Kernel Utilities
820	//=============================================================================================
821	#if defined(CFA_HAVE_LINUX_IO_URING_H)
822	#include "io/types.hfa"
823	#endif
824
825	static inline bool __maybe_io_drain( processor * proc ) {
826	bool ret = false;
827	#if defined(CFA_HAVE_LINUX_IO_URING_H)
828	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
829
830	// Check if we should drain the queue
831	$io_context * ctx = proc->io.ctx;
832	unsigned head = *ctx->cq.head;
833	unsigned tail = *ctx->cq.tail;
834	if(head == tail) return false;
835	ready_schedule_lock();
836	ret = __cfa_io_drain( proc );
837	ready_schedule_unlock();
838	#endif
839	return ret;
840	}
841
842	//-----------------------------------------------------------------------------
843	// Debug
844	__cfaabi_dbg_debug_do(
845	extern "C" {
846	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
847	this.prev_name = prev_name;
848	this.prev_thrd = kernelTLS().this_thread;
849	}
850	}
851	)
852
853	//-----------------------------------------------------------------------------
854	// Debug
855	bool threading_enabled(void) __attribute__((const)) {
856	return true;
857	}
858
859	//-----------------------------------------------------------------------------
860	// Statistics
861	#if !defined(__CFA_NO_STATISTICS__)
862	void print_halts( processor & this ) {
863	this.print_halts = true;
864	}
865
866	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
867	/* paranoid */ verify( cltr->stats );
868
869	processor * it = &list`first;
870	for(unsigned i = 0; i < count; i++) {
871	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
872	/* paranoid */ verify( it->local_data->this_stats );
873	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
874	__tally_stats( cltr->stats, it->local_data->this_stats );
875	it = &(*it)`next;
876	}
877	}
878
879	void crawl_cluster_stats( cluster & this ) {
880	// Stop the world, otherwise stats could get really messed-up
881	// this doesn't solve all problems but does solve many
882	// so it's probably good enough
883	disable_interrupts();
884	uint_fast32_t last_size = ready_mutate_lock();
885
886	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
887	crawl_list(&this, this.procs.idles , this.procs.idle );
888
889	// Unlock the RWlock
890	ready_mutate_unlock( last_size );
891	enable_interrupts();
892	}
893
894
895	void print_stats_now( cluster & this, int flags ) {
896	crawl_cluster_stats( this );
897	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
898	}
899	#endif
900	// Local Variables: //
901	// mode: c //
902	// tab-width: 4 //
903	// End: //

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

Download in other formats:

Original Format