Context Navigation

kernel.cfa @ 70b4aeb9

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since 70b4aeb9 was 70b4aeb9, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Commit last changes before moving off plg7a
Property mode set to `100644`
File size: 31.8 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	#define OLD_MAIN 1
190	#if OLD_MAIN
191	// Check if there is pending io
192	__maybe_io_drain( this );
193
194	// Try to get the next thread
195	readyThread = __next_thread( this->cltr );
196
197	if( !readyThread ) {
198	__tls_stats()->io.flush.idle++;
199	__cfa_io_flush( this, 0 );
200
201	readyThread = __next_thread_slow( this->cltr );
202	}
203
204	HALT:
205	if( !readyThread ) {
206	// Don't block if we are done
207	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
208
209	// Push self to idle stack
210	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
211
212	// Confirm the ready-queue is empty
213	readyThread = __next_thread_slow( this->cltr );
214	if( readyThread ) {
215	// A thread was found, cancel the halt
216	mark_awake(this->cltr->procs, * this);
217
218	#if !defined(__CFA_NO_STATISTICS__)
219	__tls_stats()->ready.sleep.cancels++;
220	#endif
221
222	// continue the mai loop
223	break HALT;
224	}
225
226	idle_sleep( this, future, idle_iovec );
227
228	// We were woken up, remove self from idle
229	mark_awake(this->cltr->procs, * this);
230
231	// DON'T just proceed, start looking again
232	continue MAIN_LOOP;
233	}
234
235	/* paranoid */ verify( readyThread );
236
237	// Reset io dirty bit
238	this->io.dirty = false;
239
240	// We found a thread run it
241	__run_thread(this, readyThread);
242
243	// Are we done?
244	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
245
246	if(this->io.pending && !this->io.dirty) {
247	__tls_stats()->io.flush.dirty++;
248	__cfa_io_flush( this, 0 );
249	}
250
251	#else
252	#warning new kernel loop
253	SEARCH: {
254	/* paranoid */ verify( ! __preemption_enabled() );
255
256	// First, lock the scheduler since we are searching for a thread
257	ready_schedule_lock();
258
259	// Try to get the next thread
260	readyThread = pop_fast( this->cltr );
261	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
262
263	// If we can't find a thread, might as well flush any outstanding I/O
264	if(this->io.pending) { __cfa_io_flush( this, 0 ); }
265
266	// Spin a little on I/O, just in case
267	for(5) {
268	__maybe_io_drain( this );
269	readyThread = pop_fast( this->cltr );
270	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
271	}
272
273	// no luck, try stealing a few times
274	for(5) {
275	if( __maybe_io_drain( this ) ) {
276	readyThread = pop_fast( this->cltr );
277	} else {
278	readyThread = pop_slow( this->cltr );
279	}
280	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
281	}
282
283	// still no luck, search for a thread
284	readyThread = pop_search( this->cltr );
285	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
286
287	// Don't block if we are done
288	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
289	ready_schedule_unlock();
290	break MAIN_LOOP;
291	}
292
293	__STATS( __tls_stats()->ready.sleep.halts++; )
294
295	// Push self to idle stack
296	ready_schedule_unlock();
297	if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
298	ready_schedule_lock();
299
300	// Confirm the ready-queue is empty
301	__maybe_io_drain( this );
302	readyThread = pop_search( this->cltr );
303	ready_schedule_unlock();
304
305	if( readyThread ) {
306	// A thread was found, cancel the halt
307	mark_awake(this->cltr->procs, * this);
308
309	__STATS( __tls_stats()->ready.sleep.cancels++; )
310
311	// continue the main loop
312	break SEARCH;
313	}
314
315	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
316	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
317
318	{
319	eventfd_t val;
320	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
321	if(ret < 0) {
322	switch((int)errno) {
323	case EAGAIN:
324	#if EAGAIN != EWOULDBLOCK
325	case EWOULDBLOCK:
326	#endif
327	case EINTR:
328	// No need to do anything special here, just assume it's a legitimate wake-up
329	break;
330	default:
331	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
332	}
333	}
334	}
335
336	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
337
338	// We were woken up, remove self from idle
339	mark_awake(this->cltr->procs, * this);
340
341	// DON'T just proceed, start looking again
342	continue MAIN_LOOP;
343	}
344
345	RUN_THREAD:
346	/* paranoid */ verify( ! __preemption_enabled() );
347	/* paranoid */ verify( readyThread );
348
349	// Reset io dirty bit
350	this->io.dirty = false;
351
352	// We found a thread run it
353	__run_thread(this, readyThread);
354
355	// Are we done?
356	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
357
358	if(this->io.pending && !this->io.dirty) {
359	__cfa_io_flush( this, 0 );
360	}
361
362	ready_schedule_lock();
363	__maybe_io_drain( this );
364	ready_schedule_unlock();
365	#endif
366	}
367
368	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
369	}
370
371	for(int i = 0; !available(future); i++) {
372	if(i > 1000) __cfaabi_dbg_write( "ERROR: kernel has bin spinning on a flush after exit loop.\n", 60);
373	__cfa_io_flush( this, 1 );
374	}
375
376	__cfa_io_stop( this );
377
378	post( this->terminated );
379
380	if(this == mainProcessor) {
381	// HACK : the coroutine context switch expects this_thread to be set
382	// and it make sense for it to be set in all other cases except here
383	// fake it
384	__cfaabi_tls.this_thread = mainThread;
385	}
386
387	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
388	}
389
390	static int * __volatile_errno() __attribute__((noinline));
391	static int * __volatile_errno() { asm(""); return &errno; }
392
393	// KERNEL ONLY
394	// runThread runs a thread by context switching
395	// from the processor coroutine to the target thread
396	static void __run_thread(processor * this, thread$ * thrd_dst) {
397	/* paranoid */ verify( ! __preemption_enabled() );
398	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
399	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
400	__builtin_prefetch( thrd_dst->context.SP );
401
402	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
403
404	coroutine$ * proc_cor = get_coroutine(this->runner);
405
406	// set state of processor coroutine to inactive
407	verify(proc_cor->state == Active);
408	proc_cor->state = Blocked;
409
410	// Actually run the thread
411	RUNNING: while(true) {
412	thrd_dst->preempted = __NO_PREEMPTION;
413	thrd_dst->state = Active;
414
415	// Update global state
416	kernelTLS().this_thread = thrd_dst;
417
418	/* paranoid */ verify( ! __preemption_enabled() );
419	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
420	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
421	/* paranoid */ verify( thrd_dst->context.SP );
422	/* paranoid */ verify( thrd_dst->state != Halted );
423	/* 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
424	/* 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
425	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
426
427
428
429	// set context switch to the thread that the processor is executing
430	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
431	// when __cfactx_switch returns we are back in the processor coroutine
432
433
434
435	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
436	/* 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 );
437	/* 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 );
438	/* paranoid */ verify( thrd_dst->context.SP );
439	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
440	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
441	/* paranoid */ verify( ! __preemption_enabled() );
442
443	// Reset global state
444	kernelTLS().this_thread = 0p;
445
446	// We just finished running a thread, there are a few things that could have happened.
447	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
448	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
449	// 4 - Preempted
450	// In case 1, we may have won a race so we can't write to the state again.
451	// In case 2, we lost the race so we now own the thread.
452
453	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
454	// The thread was preempted, reschedule it and reset the flag
455	schedule_thread$( thrd_dst, UNPARK_LOCAL );
456	break RUNNING;
457	}
458
459	if(unlikely(thrd_dst->state == Halting)) {
460	// The thread has halted, it should never be scheduled/run again
461	// finish the thread
462	__thread_finish( thrd_dst );
463	break RUNNING;
464	}
465
466	/* paranoid */ verify( thrd_dst->state == Active );
467	thrd_dst->state = Blocked;
468
469	// set state of processor coroutine to active and the thread to inactive
470	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
471	switch(old_ticket) {
472	case TICKET_RUNNING:
473	// This is case 1, the regular case, nothing more is needed
474	break RUNNING;
475	case TICKET_UNBLOCK:
476	#if !defined(__CFA_NO_STATISTICS__)
477	__tls_stats()->ready.threads.threads++;
478	#endif
479	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
480	// In this case, just run it again.
481	continue RUNNING;
482	default:
483	// This makes no sense, something is wrong abort
484	abort();
485	}
486	}
487
488	// Just before returning to the processor, set the processor coroutine to active
489	proc_cor->state = Active;
490
491	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
492
493	#if !defined(__CFA_NO_STATISTICS__)
494	__tls_stats()->ready.threads.threads--;
495	#endif
496
497	/* paranoid */ verify( ! __preemption_enabled() );
498	}
499
500	// KERNEL_ONLY
501	void returnToKernel() {
502	/* paranoid */ verify( ! __preemption_enabled() );
503	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
504	thread$ * thrd_src = kernelTLS().this_thread;
505
506	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
507
508	// Run the thread on this processor
509	{
510	int local_errno = *__volatile_errno();
511	#if defined( __i386 ) \|\| defined( __x86_64 )
512	__x87_store;
513	#endif
514	/* paranoid */ verify( proc_cor->context.SP );
515	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
516	__cfactx_switch( &thrd_src->context, &proc_cor->context );
517	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
518	#if defined( __i386 ) \|\| defined( __x86_64 )
519	__x87_load;
520	#endif
521	*__volatile_errno() = local_errno;
522	}
523
524	#if !defined(__CFA_NO_STATISTICS__)
525	/* paranoid */ verify( thrd_src->last_proc != 0p );
526	if(thrd_src->last_proc != kernelTLS().this_processor) {
527	__tls_stats()->ready.threads.migration++;
528	}
529	#endif
530
531	/* paranoid */ verify( ! __preemption_enabled() );
532	/* 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 );
533	/* 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 );
534	}
535
536	//-----------------------------------------------------------------------------
537	// Scheduler routines
538	// KERNEL ONLY
539	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
540	/* paranoid */ verify( ! __preemption_enabled() );
541	/* paranoid */ verify( ready_schedule_islocked());
542	/* paranoid */ verify( thrd );
543	/* paranoid */ verify( thrd->state != Halted );
544	/* paranoid */ verify( thrd->curr_cluster );
545	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
546	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
547	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
548	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
549	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
550	/* paranoid */ #endif
551	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
552	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
553
554	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
555
556	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
557	struct cluster * cl = thrd->curr_cluster;
558	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
559
560	// push the thread to the cluster ready-queue
561	push( cl, thrd, hint );
562
563	// variable thrd is no longer safe to use
564	thrd = 0xdeaddeaddeaddeadp;
565
566	// wake the cluster using the save variable.
567	__wake_one( cl );
568
569	#if !defined(__CFA_NO_STATISTICS__)
570	if( kernelTLS().this_stats ) {
571	__tls_stats()->ready.threads.threads++;
572	if(outside) {
573	__tls_stats()->ready.threads.extunpark++;
574	}
575	}
576	else {
577	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
578	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
579	}
580	#endif
581
582	/* paranoid */ verify( ready_schedule_islocked());
583	/* paranoid */ verify( ! __preemption_enabled() );
584	}
585
586	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
587	ready_schedule_lock();
588	__schedule_thread( thrd, hint );
589	ready_schedule_unlock();
590	}
591
592	// KERNEL ONLY
593	static inline thread$ * __next_thread(cluster * this) with( *this ) {
594	/* paranoid */ verify( ! __preemption_enabled() );
595
596	ready_schedule_lock();
597	thread$ * thrd = pop_fast( this );
598	ready_schedule_unlock();
599
600	/* paranoid */ verify( ! __preemption_enabled() );
601	return thrd;
602	}
603
604	// KERNEL ONLY
605	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
606	/* paranoid */ verify( ! __preemption_enabled() );
607
608	ready_schedule_lock();
609	thread$ * thrd;
610	for(25) {
611	thrd = pop_slow( this );
612	if(thrd) goto RET;
613	}
614	thrd = pop_search( this );
615
616	RET:
617	ready_schedule_unlock();
618
619	/* paranoid */ verify( ! __preemption_enabled() );
620	return thrd;
621	}
622
623	static inline bool __must_unpark( thread$ * thrd ) {
624	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
625	switch(old_ticket) {
626	case TICKET_RUNNING:
627	// Wake won the race, the thread will reschedule/rerun itself
628	return false;
629	case TICKET_BLOCKED:
630	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
631	/* paranoid */ verify( thrd->state == Blocked );
632	return true;
633	default:
634	// This makes no sense, something is wrong abort
635	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
636	}
637	}
638
639	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
640	/* paranoid */ verify( ! __preemption_enabled() );
641	/* paranoid */ verify( ready_schedule_islocked());
642
643	if( !thrd ) return;
644
645	if(__must_unpark(thrd)) {
646	// Wake lost the race,
647	__schedule_thread( thrd, hint );
648	}
649
650	/* paranoid */ verify( ready_schedule_islocked());
651	/* paranoid */ verify( ! __preemption_enabled() );
652	}
653
654	void unpark( thread$ * thrd, unpark_hint hint ) {
655	if( !thrd ) return;
656
657	if(__must_unpark(thrd)) {
658	disable_interrupts();
659	// Wake lost the race,
660	schedule_thread$( thrd, hint );
661	enable_interrupts(false);
662	}
663	}
664
665	void park( void ) {
666	__disable_interrupts_checked();
667	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
668	returnToKernel();
669	__enable_interrupts_checked();
670
671	}
672
673	extern "C" {
674	// Leave the thread monitor
675	// last routine called by a thread.
676	// Should never return
677	void __cfactx_thrd_leave() {
678	thread$ * thrd = active_thread();
679	monitor$ * this = &thrd->self_mon;
680
681	// Lock the monitor now
682	lock( this->lock __cfaabi_dbg_ctx2 );
683
684	disable_interrupts();
685
686	/* paranoid */ verify( ! __preemption_enabled() );
687	/* paranoid */ verify( thrd->state == Active );
688	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
689	/* paranoid */ verify( kernelTLS().this_thread == thrd );
690	/* paranoid */ verify( thrd->context.SP );
691	/* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) > ((uintptr_t)__get_stack(thrd->curr_cor)->limit), "ERROR : thread$ %p has been corrupted.\n StackPointer too large.\n", thrd );
692	/* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) < ((uintptr_t)__get_stack(thrd->curr_cor)->base ), "ERROR : thread$ %p has been corrupted.\n StackPointer too small.\n", thrd );
693
694	thrd->state = Halting;
695	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
696	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
697	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
698
699	// Leave the thread
700	returnToKernel();
701
702	// Control flow should never reach here!
703	abort();
704	}
705	}
706
707	// KERNEL ONLY
708	bool force_yield( __Preemption_Reason reason ) {
709	__disable_interrupts_checked();
710	thread$ * thrd = kernelTLS().this_thread;
711	/* paranoid */ verify(thrd->state == Active);
712
713	// SKULLDUGGERY: It is possible that we are preempting this thread just before
714	// it was going to park itself. If that is the case and it is already using the
715	// intrusive fields then we can't use them to preempt the thread
716	// If that is the case, abandon the preemption.
717	bool preempted = false;
718	if(thrd->link.next == 0p) {
719	preempted = true;
720	thrd->preempted = reason;
721	returnToKernel();
722	}
723	__enable_interrupts_checked( false );
724	return preempted;
725	}
726
727	//=============================================================================================
728	// Kernel Idle Sleep
729	//=============================================================================================
730	// Wake a thread from the front if there are any
731	static void __wake_one(cluster * this) {
732	eventfd_t val;
733
734	/* paranoid */ verify( ! __preemption_enabled() );
735	/* paranoid */ verify( ready_schedule_islocked() );
736
737	// Check if there is a sleeping processor
738	struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
739
740	// If no one is sleeping: we are done
741	if( fdp == 0p ) return;
742
743	int fd = 1;
744	if( __atomic_load_n(&fdp->fd, __ATOMIC_SEQ_CST) != 1 ) {
745	fd = __atomic_exchange_n(&fdp->fd, 1, __ATOMIC_RELAXED);
746	}
747
748	switch(fd) {
749	case 0:
750	// If the processor isn't ready to sleep then the exchange will already wake it up
751	#if !defined(__CFA_NO_STATISTICS__)
752	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
753	} else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
754	#endif
755	break;
756	case 1:
757	// If someone else already said they will wake them: we are done
758	#if !defined(__CFA_NO_STATISTICS__)
759	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
760	} else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
761	#endif
762	break;
763	default:
764	// If the processor was ready to sleep, we need to wake it up with an actual write
765	val = 1;
766	eventfd_write( fd, val );
767
768	#if !defined(__CFA_NO_STATISTICS__)
769	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
770	} else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
771	#endif
772	break;
773	}
774
775	/* paranoid */ verify( ready_schedule_islocked() );
776	/* paranoid */ verify( ! __preemption_enabled() );
777
778	return;
779	}
780
781	// Unconditionnaly wake a thread
782	void __wake_proc(processor * this) {
783	/* paranoid */ verify( ! __preemption_enabled() );
784
785	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
786
787	this->idle_wctx.fd = 1;
788
789	eventfd_t val;
790	val = 1;
791	eventfd_write( this->idle_fd, val );
792
793	/* paranoid */ verify( ! __preemption_enabled() );
794	}
795
796	static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
797	// Tell everyone we are ready to go do sleep
798	for() {
799	int expected = this->idle_wctx.fd;
800
801	// Someone already told us to wake-up! No time for a nap.
802	if(expected == 1) { return; }
803
804	// Try to mark that we are going to sleep
805	if(__atomic_compare_exchange_n(&this->idle_wctx.fd, &expected, this->idle_fd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
806	// Every one agreed, taking a nap
807	break;
808	}
809	}
810
811
812	#if !defined(CFA_WITH_IO_URING_IDLE)
813	#if !defined(__CFA_NO_STATISTICS__)
814	if(this->print_halts) {
815	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
816	}
817	#endif
818
819	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
820
821	{
822	eventfd_t val;
823	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
824	if(ret < 0) {
825	switch((int)errno) {
826	case EAGAIN:
827	#if EAGAIN != EWOULDBLOCK
828	case EWOULDBLOCK:
829	#endif
830	case EINTR:
831	// No need to do anything special here, just assume it's a legitimate wake-up
832	break;
833	default:
834	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
835	}
836	}
837	}
838
839	#if !defined(__CFA_NO_STATISTICS__)
840	if(this->print_halts) {
841	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
842	}
843	#endif
844	#else
845	// Do we already have a pending read
846	if(available(future)) {
847	// There is no pending read, we need to add one
848	reset(future);
849
850	__kernel_read(this, future, iov, this->idle_fd );
851	}
852
853	__cfa_io_flush( this, 1 );
854	#endif
855	}
856
857	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
858	#if !defined(__CFA_NO_STATISTICS__)
859	__tls_stats()->ready.sleep.halts++;
860	#endif
861
862	proc.idle_wctx.fd = 0;
863
864	/* paranoid */ verify( ! __preemption_enabled() );
865	if(!try_lock( this )) return false;
866	this.idle++;
867	/* paranoid */ verify( this.idle <= this.total );
868	remove(proc);
869	insert_first(this.idles, proc);
870
871	__atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
872	unlock( this );
873	/* paranoid */ verify( ! __preemption_enabled() );
874
875	return true;
876	}
877
878	static void mark_awake(__cluster_proc_list & this, processor & proc) {
879	/* paranoid */ verify( ! __preemption_enabled() );
880	lock( this );
881	this.idle--;
882	/* paranoid */ verify( this.idle >= 0 );
883	remove(proc);
884	insert_last(this.actives, proc);
885
886	{
887	struct __fd_waitctx * wctx = 0;
888	if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
889	__atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
890	}
891
892	unlock( this );
893	/* paranoid */ verify( ! __preemption_enabled() );
894	}
895
896	//=============================================================================================
897	// Unexpected Terminating logic
898	//=============================================================================================
899	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
900	thread$ * thrd = __cfaabi_tls.this_thread;
901
902	if(thrd) {
903	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
904	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
905
906	if ( &thrd->self_cor != thrd->curr_cor ) {
907	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
908	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
909	}
910	else {
911	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
912	}
913	}
914	else {
915	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
916	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
917	}
918	}
919
920	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
921	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
922	}
923
924	static __spinlock_t kernel_debug_lock;
925
926	extern "C" {
927	void __cfaabi_bits_acquire() {
928	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
929	}
930
931	void __cfaabi_bits_release() {
932	unlock( kernel_debug_lock );
933	}
934	}
935
936	//=============================================================================================
937	// Kernel Utilities
938	//=============================================================================================
939	#if defined(CFA_HAVE_LINUX_IO_URING_H)
940	#include "io/types.hfa"
941	#endif
942
943	static inline bool __maybe_io_drain( processor * proc ) {
944	bool ret = false;
945	#if defined(CFA_HAVE_LINUX_IO_URING_H)
946	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
947
948	// Check if we should drain the queue
949	$io_context * ctx = proc->io.ctx;
950	unsigned head = *ctx->cq.head;
951	unsigned tail = *ctx->cq.tail;
952	if(head == tail) return false;
953	#if OLD_MAIN
954	ready_schedule_lock();
955	ret = __cfa_io_drain( proc );
956	ready_schedule_unlock();
957	#else
958	ret = __cfa_io_drain( proc );
959	#endif
960	#endif
961	return ret;
962	}
963
964	//-----------------------------------------------------------------------------
965	// Debug
966	__cfaabi_dbg_debug_do(
967	extern "C" {
968	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
969	this.prev_name = prev_name;
970	this.prev_thrd = kernelTLS().this_thread;
971	}
972	}
973	)
974
975	//-----------------------------------------------------------------------------
976	// Debug
977	bool threading_enabled(void) __attribute__((const)) {
978	return true;
979	}
980
981	//-----------------------------------------------------------------------------
982	// Statistics
983	#if !defined(__CFA_NO_STATISTICS__)
984	void print_halts( processor & this ) {
985	this.print_halts = true;
986	}
987
988	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
989	/* paranoid */ verify( cltr->stats );
990
991	processor * it = &list`first;
992	for(unsigned i = 0; i < count; i++) {
993	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
994	/* paranoid */ verify( it->local_data->this_stats );
995	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
996	__tally_stats( cltr->stats, it->local_data->this_stats );
997	it = &(*it)`next;
998	}
999	}
1000
1001	void crawl_cluster_stats( cluster & this ) {
1002	// Stop the world, otherwise stats could get really messed-up
1003	// this doesn't solve all problems but does solve many
1004	// so it's probably good enough
1005	disable_interrupts();
1006	uint_fast32_t last_size = ready_mutate_lock();
1007
1008	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
1009	crawl_list(&this, this.procs.idles , this.procs.idle );
1010
1011	// Unlock the RWlock
1012	ready_mutate_unlock( last_size );
1013	enable_interrupts();
1014	}
1015
1016
1017	void print_stats_now( cluster & this, int flags ) {
1018	crawl_cluster_stats( this );
1019	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
1020	}
1021	#endif
1022	// Local Variables: //
1023	// mode: c //
1024	// tab-width: 4 //
1025	// End: //

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

Download in other formats:

Original Format