Context Navigation

kernel.cfa @ b200492

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since b200492 was 5b7a3662, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Tentative optimization for wake-one
Property mode set to `100644`
File size: 30.6 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// kernel.c --
8	//
9	// Author : Thierry Delisle
10	// Created On : Tue Jan 17 12:27:26 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 31 07:08:20 2020
13	// Update Count : 71
14	//
15
16	#define __cforall_thread__
17	#define _GNU_SOURCE
18
19	// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
20
21	//C Includes
22	#include <errno.h>
23	#include <stdio.h>
24	#include <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	__cfa_io_flush( this, 0 );
199
200	readyThread = __next_thread_slow( this->cltr );
201	}
202
203	HALT:
204	if( !readyThread ) {
205	// Don't block if we are done
206	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
207
208	#if !defined(__CFA_NO_STATISTICS__)
209	__tls_stats()->ready.sleep.halts++;
210	#endif
211
212	// Push self to idle stack
213	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
214
215	// Confirm the ready-queue is empty
216	readyThread = __next_thread_slow( this->cltr );
217	if( readyThread ) {
218	// A thread was found, cancel the halt
219	mark_awake(this->cltr->procs, * this);
220
221	#if !defined(__CFA_NO_STATISTICS__)
222	__tls_stats()->ready.sleep.cancels++;
223	#endif
224
225	// continue the mai loop
226	break HALT;
227	}
228
229	idle_sleep( this, future, idle_iovec );
230
231	// We were woken up, remove self from idle
232	mark_awake(this->cltr->procs, * this);
233
234	// DON'T just proceed, start looking again
235	continue MAIN_LOOP;
236	}
237
238	/* paranoid */ verify( readyThread );
239
240	// Reset io dirty bit
241	this->io.dirty = false;
242
243	// We found a thread run it
244	__run_thread(this, readyThread);
245
246	// Are we done?
247	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
248
249	if(this->io.pending && !this->io.dirty) {
250	__cfa_io_flush( this, 0 );
251	}
252
253	#else
254	#warning new kernel loop
255	SEARCH: {
256	/* paranoid */ verify( ! __preemption_enabled() );
257
258	// First, lock the scheduler since we are searching for a thread
259	ready_schedule_lock();
260
261	// Try to get the next thread
262	readyThread = pop_fast( this->cltr );
263	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
264
265	// If we can't find a thread, might as well flush any outstanding I/O
266	if(this->io.pending) { __cfa_io_flush( this, 0 ); }
267
268	// Spin a little on I/O, just in case
269	for(5) {
270	__maybe_io_drain( this );
271	readyThread = pop_fast( this->cltr );
272	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
273	}
274
275	// no luck, try stealing a few times
276	for(5) {
277	if( __maybe_io_drain( this ) ) {
278	readyThread = pop_fast( this->cltr );
279	} else {
280	readyThread = pop_slow( this->cltr );
281	}
282	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
283	}
284
285	// still no luck, search for a thread
286	readyThread = pop_search( this->cltr );
287	if(readyThread) { ready_schedule_unlock(); break SEARCH; }
288
289	// Don't block if we are done
290	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
291	ready_schedule_unlock();
292	break MAIN_LOOP;
293	}
294
295	__STATS( __tls_stats()->ready.sleep.halts++; )
296
297	// Push self to idle stack
298	ready_schedule_unlock();
299	if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
300	ready_schedule_lock();
301
302	// Confirm the ready-queue is empty
303	__maybe_io_drain( this );
304	readyThread = pop_search( this->cltr );
305	ready_schedule_unlock();
306
307	if( readyThread ) {
308	// A thread was found, cancel the halt
309	mark_awake(this->cltr->procs, * this);
310
311	__STATS( __tls_stats()->ready.sleep.cancels++; )
312
313	// continue the main loop
314	break SEARCH;
315	}
316
317	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
318	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
319
320	{
321	eventfd_t val;
322	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
323	if(ret < 0) {
324	switch((int)errno) {
325	case EAGAIN:
326	#if EAGAIN != EWOULDBLOCK
327	case EWOULDBLOCK:
328	#endif
329	case EINTR:
330	// No need to do anything special here, just assume it's a legitimate wake-up
331	break;
332	default:
333	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
334	}
335	}
336	}
337
338	__STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
339
340	// We were woken up, remove self from idle
341	mark_awake(this->cltr->procs, * this);
342
343	// DON'T just proceed, start looking again
344	continue MAIN_LOOP;
345	}
346
347	RUN_THREAD:
348	/* paranoid */ verify( ! __preemption_enabled() );
349	/* paranoid */ verify( readyThread );
350
351	// Reset io dirty bit
352	this->io.dirty = false;
353
354	// We found a thread run it
355	__run_thread(this, readyThread);
356
357	// Are we done?
358	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
359
360	if(this->io.pending && !this->io.dirty) {
361	__cfa_io_flush( this, 0 );
362	}
363
364	ready_schedule_lock();
365	__maybe_io_drain( this );
366	ready_schedule_unlock();
367	#endif
368	}
369
370	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
371	}
372
373	for(int i = 0; !available(future); i++) {
374	if(i > 1000) __cfaabi_dbg_write( "ERROR: kernel has bin spinning on a flush after exit loop.\n", 60);
375	__cfa_io_flush( this, 1 );
376	}
377
378	__cfa_io_stop( this );
379
380	post( this->terminated );
381
382	if(this == mainProcessor) {
383	// HACK : the coroutine context switch expects this_thread to be set
384	// and it make sense for it to be set in all other cases except here
385	// fake it
386	__cfaabi_tls.this_thread = mainThread;
387	}
388
389	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
390	}
391
392	static int * __volatile_errno() __attribute__((noinline));
393	static int * __volatile_errno() { asm(""); return &errno; }
394
395	// KERNEL ONLY
396	// runThread runs a thread by context switching
397	// from the processor coroutine to the target thread
398	static void __run_thread(processor * this, thread$ * thrd_dst) {
399	/* paranoid */ verify( ! __preemption_enabled() );
400	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
401	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
402	__builtin_prefetch( thrd_dst->context.SP );
403
404	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
405
406	coroutine$ * proc_cor = get_coroutine(this->runner);
407
408	// set state of processor coroutine to inactive
409	verify(proc_cor->state == Active);
410	proc_cor->state = Blocked;
411
412	// Actually run the thread
413	RUNNING: while(true) {
414	thrd_dst->preempted = __NO_PREEMPTION;
415	thrd_dst->state = Active;
416
417	// Update global state
418	kernelTLS().this_thread = thrd_dst;
419
420	/* paranoid */ verify( ! __preemption_enabled() );
421	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
422	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
423	/* paranoid */ verify( thrd_dst->context.SP );
424	/* paranoid */ verify( thrd_dst->state != Halted );
425	/* 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
426	/* 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
427	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
428
429
430
431	// set context switch to the thread that the processor is executing
432	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
433	// when __cfactx_switch returns we are back in the processor coroutine
434
435
436
437	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
438	/* 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 );
439	/* 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 );
440	/* paranoid */ verify( thrd_dst->context.SP );
441	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
442	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
443	/* paranoid */ verify( ! __preemption_enabled() );
444
445	// Reset global state
446	kernelTLS().this_thread = 0p;
447
448	// We just finished running a thread, there are a few things that could have happened.
449	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
450	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
451	// 4 - Preempted
452	// In case 1, we may have won a race so we can't write to the state again.
453	// In case 2, we lost the race so we now own the thread.
454
455	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
456	// The thread was preempted, reschedule it and reset the flag
457	schedule_thread$( thrd_dst, UNPARK_LOCAL );
458	break RUNNING;
459	}
460
461	if(unlikely(thrd_dst->state == Halting)) {
462	// The thread has halted, it should never be scheduled/run again
463	// finish the thread
464	__thread_finish( thrd_dst );
465	break RUNNING;
466	}
467
468	/* paranoid */ verify( thrd_dst->state == Active );
469	thrd_dst->state = Blocked;
470
471	// set state of processor coroutine to active and the thread to inactive
472	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
473	switch(old_ticket) {
474	case TICKET_RUNNING:
475	// This is case 1, the regular case, nothing more is needed
476	break RUNNING;
477	case TICKET_UNBLOCK:
478	#if !defined(__CFA_NO_STATISTICS__)
479	__tls_stats()->ready.threads.threads++;
480	#endif
481	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
482	// In this case, just run it again.
483	continue RUNNING;
484	default:
485	// This makes no sense, something is wrong abort
486	abort();
487	}
488	}
489
490	// Just before returning to the processor, set the processor coroutine to active
491	proc_cor->state = Active;
492
493	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
494
495	#if !defined(__CFA_NO_STATISTICS__)
496	__tls_stats()->ready.threads.threads--;
497	#endif
498
499	/* paranoid */ verify( ! __preemption_enabled() );
500	}
501
502	// KERNEL_ONLY
503	void returnToKernel() {
504	/* paranoid */ verify( ! __preemption_enabled() );
505	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
506	thread$ * thrd_src = kernelTLS().this_thread;
507
508	__STATS( thrd_src->last_proc = kernelTLS().this_processor; )
509
510	// Run the thread on this processor
511	{
512	int local_errno = *__volatile_errno();
513	#if defined( __i386 ) \|\| defined( __x86_64 )
514	__x87_store;
515	#endif
516	/* paranoid */ verify( proc_cor->context.SP );
517	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
518	__cfactx_switch( &thrd_src->context, &proc_cor->context );
519	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
520	#if defined( __i386 ) \|\| defined( __x86_64 )
521	__x87_load;
522	#endif
523	*__volatile_errno() = local_errno;
524	}
525
526	#if !defined(__CFA_NO_STATISTICS__)
527	/* paranoid */ verify( thrd_src->last_proc != 0p );
528	if(thrd_src->last_proc != kernelTLS().this_processor) {
529	__tls_stats()->ready.threads.migration++;
530	}
531	#endif
532
533	/* paranoid */ verify( ! __preemption_enabled() );
534	/* 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 );
535	/* 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 );
536	}
537
538	//-----------------------------------------------------------------------------
539	// Scheduler routines
540	// KERNEL ONLY
541	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
542	/* paranoid */ verify( ! __preemption_enabled() );
543	/* paranoid */ verify( ready_schedule_islocked());
544	/* paranoid */ verify( thrd );
545	/* paranoid */ verify( thrd->state != Halted );
546	/* paranoid */ verify( thrd->curr_cluster );
547	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
548	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
549	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
550	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
551	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
552	/* paranoid */ #endif
553	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
554	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
555
556	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
557
558	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
559	struct cluster * cl = thrd->curr_cluster;
560	__STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
561
562	// push the thread to the cluster ready-queue
563	push( cl, thrd, hint );
564
565	// variable thrd is no longer safe to use
566	thrd = 0xdeaddeaddeaddeadp;
567
568	// wake the cluster using the save variable.
569	__wake_one( cl );
570
571	#if !defined(__CFA_NO_STATISTICS__)
572	if( kernelTLS().this_stats ) {
573	__tls_stats()->ready.threads.threads++;
574	if(outside) {
575	__tls_stats()->ready.threads.extunpark++;
576	}
577	}
578	else {
579	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
580	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
581	}
582	#endif
583
584	/* paranoid */ verify( ready_schedule_islocked());
585	/* paranoid */ verify( ! __preemption_enabled() );
586	}
587
588	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
589	ready_schedule_lock();
590	__schedule_thread( thrd, hint );
591	ready_schedule_unlock();
592	}
593
594	// KERNEL ONLY
595	static inline thread$ * __next_thread(cluster * this) with( *this ) {
596	/* paranoid */ verify( ! __preemption_enabled() );
597
598	ready_schedule_lock();
599	thread$ * thrd = pop_fast( this );
600	ready_schedule_unlock();
601
602	/* paranoid */ verify( ! __preemption_enabled() );
603	return thrd;
604	}
605
606	// KERNEL ONLY
607	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
608	/* paranoid */ verify( ! __preemption_enabled() );
609
610	ready_schedule_lock();
611	thread$ * thrd;
612	for(25) {
613	thrd = pop_slow( this );
614	if(thrd) goto RET;
615	}
616	thrd = pop_search( this );
617
618	RET:
619	ready_schedule_unlock();
620
621	/* paranoid */ verify( ! __preemption_enabled() );
622	return thrd;
623	}
624
625	static inline bool __must_unpark( thread$ * thrd ) {
626	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
627	switch(old_ticket) {
628	case TICKET_RUNNING:
629	// Wake won the race, the thread will reschedule/rerun itself
630	return false;
631	case TICKET_BLOCKED:
632	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
633	/* paranoid */ verify( thrd->state == Blocked );
634	return true;
635	default:
636	// This makes no sense, something is wrong abort
637	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
638	}
639	}
640
641	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
642	/* paranoid */ verify( ! __preemption_enabled() );
643	/* paranoid */ verify( ready_schedule_islocked());
644
645	if( !thrd ) return;
646
647	if(__must_unpark(thrd)) {
648	// Wake lost the race,
649	__schedule_thread( thrd, hint );
650	}
651
652	/* paranoid */ verify( ready_schedule_islocked());
653	/* paranoid */ verify( ! __preemption_enabled() );
654	}
655
656	void unpark( thread$ * thrd, unpark_hint hint ) {
657	if( !thrd ) return;
658
659	if(__must_unpark(thrd)) {
660	disable_interrupts();
661	// Wake lost the race,
662	schedule_thread$( thrd, hint );
663	enable_interrupts(false);
664	}
665	}
666
667	void park( void ) {
668	__disable_interrupts_checked();
669	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
670	returnToKernel();
671	__enable_interrupts_checked();
672
673	}
674
675	extern "C" {
676	// Leave the thread monitor
677	// last routine called by a thread.
678	// Should never return
679	void __cfactx_thrd_leave() {
680	thread$ * thrd = active_thread();
681	monitor$ * this = &thrd->self_mon;
682
683	// Lock the monitor now
684	lock( this->lock __cfaabi_dbg_ctx2 );
685
686	disable_interrupts();
687
688	/* paranoid */ verify( ! __preemption_enabled() );
689	/* paranoid */ verify( thrd->state == Active );
690	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
691	/* paranoid */ verify( kernelTLS().this_thread == thrd );
692	/* paranoid */ verify( thrd->context.SP );
693	/* 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 );
694	/* 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 );
695
696	thrd->state = Halting;
697	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
698	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
699	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
700
701	// Leave the thread
702	returnToKernel();
703
704	// Control flow should never reach here!
705	abort();
706	}
707	}
708
709	// KERNEL ONLY
710	bool force_yield( __Preemption_Reason reason ) {
711	__disable_interrupts_checked();
712	thread$ * thrd = kernelTLS().this_thread;
713	/* paranoid */ verify(thrd->state == Active);
714
715	// SKULLDUGGERY: It is possible that we are preempting this thread just before
716	// it was going to park itself. If that is the case and it is already using the
717	// intrusive fields then we can't use them to preempt the thread
718	// If that is the case, abandon the preemption.
719	bool preempted = false;
720	if(thrd->link.next == 0p) {
721	preempted = true;
722	thrd->preempted = reason;
723	returnToKernel();
724	}
725	__enable_interrupts_checked( false );
726	return preempted;
727	}
728
729	//=============================================================================================
730	// Kernel Idle Sleep
731	//=============================================================================================
732	// Wake a thread from the front if there are any
733	static void __wake_one(cluster * this) {
734	/* paranoid */ verify( ! __preemption_enabled() );
735	/* paranoid */ verify( ready_schedule_islocked() );
736
737	// Check if there is a sleeping processor
738	// int fd = __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST);
739	int fd = 0;
740	if( __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST) != 0 ) {
741	fd = __atomic_exchange_n(&this->procs.fd, 0, __ATOMIC_RELAXED);
742	}
743
744	// If no one is sleeping, we are done
745	if( fd == 0 ) return;
746
747	// We found a processor, wake it up
748	eventfd_t val;
749	val = 1;
750	eventfd_write( fd, val );
751
752	#if !defined(__CFA_NO_STATISTICS__)
753	if( kernelTLS().this_stats ) {
754	__tls_stats()->ready.sleep.wakes++;
755	}
756	else {
757	__atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
758	}
759	#endif
760
761	/* paranoid */ verify( ready_schedule_islocked() );
762	/* paranoid */ verify( ! __preemption_enabled() );
763
764	return;
765	}
766
767	// Unconditionnaly wake a thread
768	void __wake_proc(processor * this) {
769	/* paranoid */ verify( ! __preemption_enabled() );
770
771	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
772
773	eventfd_t val;
774	val = 1;
775	eventfd_write( this->idle_fd, val );
776
777	/* paranoid */ verify( ! __preemption_enabled() );
778	}
779
780	static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
781	#if !defined(CFA_WITH_IO_URING_IDLE)
782	#if !defined(__CFA_NO_STATISTICS__)
783	if(this->print_halts) {
784	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
785	}
786	#endif
787
788	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
789
790	{
791	eventfd_t val;
792	ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
793	if(ret < 0) {
794	switch((int)errno) {
795	case EAGAIN:
796	#if EAGAIN != EWOULDBLOCK
797	case EWOULDBLOCK:
798	#endif
799	case EINTR:
800	// No need to do anything special here, just assume it's a legitimate wake-up
801	break;
802	default:
803	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
804	}
805	}
806	}
807
808	#if !defined(__CFA_NO_STATISTICS__)
809	if(this->print_halts) {
810	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
811	}
812	#endif
813	#else
814	// Do we already have a pending read
815	if(available(future)) {
816	// There is no pending read, we need to add one
817	reset(future);
818
819	__kernel_read(this, future, iov, this->idle_fd );
820	}
821
822	__cfa_io_flush( this, 1 );
823	#endif
824	}
825
826	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
827	/* paranoid */ verify( ! __preemption_enabled() );
828	if(!try_lock( this )) return false;
829	this.idle++;
830	/* paranoid */ verify( this.idle <= this.total );
831	remove(proc);
832	insert_first(this.idles, proc);
833
834	__atomic_store_n(&this.fd, proc.idle_fd, __ATOMIC_SEQ_CST);
835	unlock( this );
836	/* paranoid */ verify( ! __preemption_enabled() );
837
838	return true;
839	}
840
841	static void mark_awake(__cluster_proc_list & this, processor & proc) {
842	/* paranoid */ verify( ! __preemption_enabled() );
843	lock( this );
844	this.idle--;
845	/* paranoid */ verify( this.idle >= 0 );
846	remove(proc);
847	insert_last(this.actives, proc);
848
849	{
850	int fd = 0;
851	if(!this.idles`isEmpty) fd = this.idles`first.idle_fd;
852	__atomic_store_n(&this.fd, fd, __ATOMIC_SEQ_CST);
853	}
854
855	unlock( this );
856	/* paranoid */ verify( ! __preemption_enabled() );
857	}
858
859	//=============================================================================================
860	// Unexpected Terminating logic
861	//=============================================================================================
862	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
863	thread$ * thrd = __cfaabi_tls.this_thread;
864
865	if(thrd) {
866	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
867	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
868
869	if ( &thrd->self_cor != thrd->curr_cor ) {
870	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
871	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
872	}
873	else {
874	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
875	}
876	}
877	else {
878	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
879	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
880	}
881	}
882
883	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
884	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
885	}
886
887	static __spinlock_t kernel_debug_lock;
888
889	extern "C" {
890	void __cfaabi_bits_acquire() {
891	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
892	}
893
894	void __cfaabi_bits_release() {
895	unlock( kernel_debug_lock );
896	}
897	}
898
899	//=============================================================================================
900	// Kernel Utilities
901	//=============================================================================================
902	#if defined(CFA_HAVE_LINUX_IO_URING_H)
903	#include "io/types.hfa"
904	#endif
905
906	static inline bool __maybe_io_drain( processor * proc ) {
907	bool ret = false;
908	#if defined(CFA_HAVE_LINUX_IO_URING_H)
909	__cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
910
911	// Check if we should drain the queue
912	$io_context * ctx = proc->io.ctx;
913	unsigned head = *ctx->cq.head;
914	unsigned tail = *ctx->cq.tail;
915	if(head == tail) return false;
916	#if OLD_MAIN
917	ready_schedule_lock();
918	ret = __cfa_io_drain( proc );
919	ready_schedule_unlock();
920	#else
921	ret = __cfa_io_drain( proc );
922	#endif
923	#endif
924	return ret;
925	}
926
927	//-----------------------------------------------------------------------------
928	// Debug
929	__cfaabi_dbg_debug_do(
930	extern "C" {
931	void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
932	this.prev_name = prev_name;
933	this.prev_thrd = kernelTLS().this_thread;
934	}
935	}
936	)
937
938	//-----------------------------------------------------------------------------
939	// Debug
940	bool threading_enabled(void) __attribute__((const)) {
941	return true;
942	}
943
944	//-----------------------------------------------------------------------------
945	// Statistics
946	#if !defined(__CFA_NO_STATISTICS__)
947	void print_halts( processor & this ) {
948	this.print_halts = true;
949	}
950
951	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
952	/* paranoid */ verify( cltr->stats );
953
954	processor * it = &list`first;
955	for(unsigned i = 0; i < count; i++) {
956	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
957	/* paranoid */ verify( it->local_data->this_stats );
958	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
959	__tally_stats( cltr->stats, it->local_data->this_stats );
960	it = &(*it)`next;
961	}
962	}
963
964	void crawl_cluster_stats( cluster & this ) {
965	// Stop the world, otherwise stats could get really messed-up
966	// this doesn't solve all problems but does solve many
967	// so it's probably good enough
968	disable_interrupts();
969	uint_fast32_t last_size = ready_mutate_lock();
970
971	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
972	crawl_list(&this, this.procs.idles , this.procs.idle );
973
974	// Unlock the RWlock
975	ready_mutate_unlock( last_size );
976	enable_interrupts();
977	}
978
979
980	void print_stats_now( cluster & this, int flags ) {
981	crawl_cluster_stats( this );
982	__print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
983	}
984	#endif
985	// Local Variables: //
986	// mode: c //
987	// tab-width: 4 //
988	// End: //

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

Download in other formats:

Original Format