Context Navigation

kernel.cfa @ 8f1e035

ADTast-experimentalpthread-emulation

Last change on this file since 8f1e035 was 8bee858, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Changed io types to have trailing $ instead of leading
Property mode set to `100644`
File size: 28.8 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	#pragma GCC diagnostic push
22	#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
23
24	//C Includes
25	#include <errno.h>
26	#include <stdio.h>
27	#include <string.h>
28	#include <signal.h>
29	#include <unistd.h>
30
31	extern "C" {
32	#include <sys/eventfd.h>
33	#include <sys/uio.h>
34	}
35
36	//CFA Includes
37	#include "kernel/private.hfa"
38	#include "preemption.hfa"
39	#include "strstream.hfa"
40	#include "device/cpu.hfa"
41	#include "io/types.hfa"
42
43	//Private includes
44	#define __CFA_INVOKE_PRIVATE__
45	#include "invoke.h"
46	#pragma GCC diagnostic pop
47
48	#if !defined(__CFA_NO_STATISTICS__)
49	#define __STATS_DEF( ...) __VA_ARGS__
50	#else
51	#define __STATS_DEF( ...)
52	#endif
53
54	//-----------------------------------------------------------------------------
55	// Some assembly required
56	#if defined( __i386 )
57	// mxcr : SSE Status and Control bits (control bits are preserved across function calls)
58	// fcw : X87 FPU control word (preserved across function calls)
59	#define __x87_store \
60	uint32_t __mxcr; \
61	uint16_t __fcw; \
62	__asm__ volatile ( \
63	"stmxcsr %0\n" \
64	"fnstcw %1\n" \
65	: "=m" (__mxcr),\
66	"=m" (__fcw) \
67	)
68
69	#define __x87_load \
70	__asm__ volatile ( \
71	"fldcw %1\n" \
72	"ldmxcsr %0\n" \
73	::"m" (__mxcr),\
74	"m" (__fcw) \
75	)
76
77	#elif defined( __x86_64 )
78	#define __x87_store \
79	uint32_t __mxcr; \
80	uint16_t __fcw; \
81	__asm__ volatile ( \
82	"stmxcsr %0\n" \
83	"fnstcw %1\n" \
84	: "=m" (__mxcr),\
85	"=m" (__fcw) \
86	)
87
88	#define __x87_load \
89	__asm__ volatile ( \
90	"fldcw %1\n" \
91	"ldmxcsr %0\n" \
92	:: "m" (__mxcr),\
93	"m" (__fcw) \
94	)
95
96	#elif defined( __arm__ )
97	#define __x87_store
98	#define __x87_load
99
100	#elif defined( __aarch64__ )
101	#define __x87_store \
102	uint32_t __fpcntl[2]; \
103	__asm__ volatile ( \
104	"mrs x9, FPCR\n" \
105	"mrs x10, FPSR\n" \
106	"stp x9, x10, %0\n" \
107	: "=m" (__fpcntl) : : "x9", "x10" \
108	)
109
110	#define __x87_load \
111	__asm__ volatile ( \
112	"ldp x9, x10, %0\n" \
113	"msr FPSR, x10\n" \
114	"msr FPCR, x9\n" \
115	: "=m" (__fpcntl) : : "x9", "x10" \
116	)
117
118	#else
119	#error unsupported hardware architecture
120	#endif
121
122	extern thread$ * mainThread;
123	extern processor * mainProcessor;
124
125	//-----------------------------------------------------------------------------
126	// Kernel Scheduling logic
127	static thread$ * __next_thread(cluster * this);
128	static thread$ * __next_thread_slow(cluster * this);
129	static thread$ * __next_thread_search(cluster * this);
130	static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
131	static void __run_thread(processor * this, thread$ * dst);
132	static void __wake_one(cluster * cltr);
133
134	static void idle_sleep(processor * proc);
135	static bool mark_idle (__cluster_proc_list & idles, processor & proc);
136	static void mark_awake(__cluster_proc_list & idles, processor & proc);
137
138	extern bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1)));
139	extern bool __cfa_io_flush( processor * ) __attribute__((nonnull (1)));
140	extern void __cfa_io_idle( processor * ) __attribute__((nonnull (1)));
141
142	#if defined(CFA_WITH_IO_URING_IDLE)
143	extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
144	#endif
145
146	extern void __disable_interrupts_hard();
147	extern void __enable_interrupts_hard();
148
149
150	//=============================================================================================
151	// Kernel Scheduling logic
152	//=============================================================================================
153	//Main of the processor contexts
154	void main(processorCtx_t & runner) {
155	// Because of a bug, we couldn't initialized the seed on construction
156	// Do it here
157	__cfaabi_tls.rand_seed ^= rdtscl();
158	__cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
159	__tls_rand_advance_bck();
160
161	processor * this = runner.proc;
162	verify(this);
163
164	/* paranoid */ verify( this->idle_wctx.ftr != 0p );
165	/* paranoid */ verify( this->idle_wctx.rdbuf != 0p );
166
167	// used for idle sleep when io_uring is present
168	// mark it as already fulfilled so we know if there is a pending request or not
169	this->idle_wctx.ftr->self.ptr = 1p;
170
171	__cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
172	#if !defined(__CFA_NO_STATISTICS__)
173	if( this->print_halts ) {
174	__cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
175	}
176	#endif
177
178	{
179	// Setup preemption data
180	preemption_scope scope = { this };
181
182	// if we need to run some special setup, now is the time to do it.
183	if(this->init.thrd) {
184	this->init.thrd->curr_cluster = this->cltr;
185	__run_thread(this, this->init.thrd);
186	}
187
188	__cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
189
190	thread$ * readyThread = 0p;
191	MAIN_LOOP:
192	for() {
193	// Check if there is pending io
194	__cfa_io_drain( this );
195
196	// Try to get the next thread
197	readyThread = __next_thread( this->cltr );
198
199	if( !readyThread ) {
200	// there is no point in holding submissions if we are idle
201	__IO_STATS__(true, io.flush.idle++; )
202	__cfa_io_flush( this );
203
204	// drain again in case something showed up
205	__cfa_io_drain( this );
206
207	readyThread = __next_thread( this->cltr );
208	}
209
210	if( !readyThread ) for(5) {
211	readyThread = __next_thread_slow( this->cltr );
212
213	if( readyThread ) break;
214
215	// It's unlikely we still I/O to submit, but the arbiter could
216	__IO_STATS__(true, io.flush.idle++; )
217	__cfa_io_flush( this );
218
219	// drain again in case something showed up
220	__cfa_io_drain( this );
221	}
222
223	HALT:
224	if( !readyThread ) {
225	// Don't block if we are done
226	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
227
228	// Push self to idle stack
229	if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
230
231	// Confirm the ready-queue is empty
232	readyThread = __next_thread_search( this->cltr );
233	if( readyThread ) {
234	// A thread was found, cancel the halt
235	mark_awake(this->cltr->procs, * this);
236
237	__STATS__(true, ready.sleep.cancels++; )
238
239	// continue the mai loop
240	break HALT;
241	}
242
243	idle_sleep( this );
244
245	// We were woken up, remove self from idle
246	mark_awake(this->cltr->procs, * this);
247
248	// DON'T just proceed, start looking again
249	continue MAIN_LOOP;
250	}
251
252	/* paranoid */ verify( readyThread );
253
254	// Reset io dirty bit
255	this->io.dirty = false;
256
257	// We found a thread run it
258	__run_thread(this, readyThread);
259
260	// Are we done?
261	if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
262
263	if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
264	__IO_STATS__(true, io.flush.dirty++; )
265	__cfa_io_flush( this );
266	}
267	}
268
269	__cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
270	}
271
272	post( this->terminated );
273
274	if(this == mainProcessor) {
275	// HACK : the coroutine context switch expects this_thread to be set
276	// and it make sense for it to be set in all other cases except here
277	// fake it
278	__cfaabi_tls.this_thread = mainThread;
279	}
280
281	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
282	}
283
284	static int * __volatile_errno() __attribute__((noinline));
285	static int * __volatile_errno() { asm(""); return &errno; }
286
287	// KERNEL ONLY
288	// runThread runs a thread by context switching
289	// from the processor coroutine to the target thread
290	static void __run_thread(processor * this, thread$ * thrd_dst) {
291	/* paranoid */ verify( ! __preemption_enabled() );
292	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
293	/* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
294	__builtin_prefetch( thrd_dst->context.SP );
295
296	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
297
298	coroutine$ * proc_cor = get_coroutine(this->runner);
299
300	// set state of processor coroutine to inactive
301	verify(proc_cor->state == Active);
302	proc_cor->state = Blocked;
303
304	// Actually run the thread
305	RUNNING: while(true) {
306	thrd_dst->preempted = __NO_PREEMPTION;
307
308	// Update global state
309	kernelTLS().this_thread = thrd_dst;
310
311	// Update the state after setting this_thread
312	// so that the debugger can find all active threads
313	// in tls storage
314	thrd_dst->state = Active;
315
316	/* paranoid */ verify( ! __preemption_enabled() );
317	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
318	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
319	/* paranoid */ verify( thrd_dst->context.SP );
320	/* paranoid */ verify( thrd_dst->state != Halted );
321	/* 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
322	/* 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
323	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
324
325
326
327	// set context switch to the thread that the processor is executing
328	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
329	// when __cfactx_switch returns we are back in the processor coroutine
330
331
332
333	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
334	/* 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 );
335	/* 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 );
336	/* paranoid */ verify( thrd_dst->context.SP );
337	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
338	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
339	/* paranoid */ verify( ! __preemption_enabled() );
340
341	// We just finished running a thread, there are a few things that could have happened.
342	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
343	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
344	// 4 - Preempted
345	// In case 1, we may have won a race so we can't write to the state again.
346	// In case 2, we lost the race so we now own the thread.
347
348	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
349	// Reset the this_thread now that we know
350	// the state isn't active anymore
351	kernelTLS().this_thread = 0p;
352
353	// The thread was preempted, reschedule it and reset the flag
354	schedule_thread$( thrd_dst, UNPARK_LOCAL );
355	break RUNNING;
356	}
357
358	if(unlikely(thrd_dst->state == Halting)) {
359	// Reset the this_thread now that we know
360	// the state isn't active anymore
361	kernelTLS().this_thread = 0p;
362
363	// The thread has halted, it should never be scheduled/run again
364	// finish the thread
365	__thread_finish( thrd_dst );
366	break RUNNING;
367	}
368
369	/* paranoid */ verify( thrd_dst->state == Active );
370	thrd_dst->state = Blocked;
371
372	// Reset the this_thread now that we know
373	// the state isn't active anymore
374	kernelTLS().this_thread = 0p;
375
376	// set state of processor coroutine to active and the thread to inactive
377	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
378	switch(old_ticket) {
379	case TICKET_RUNNING:
380	// This is case 1, the regular case, nothing more is needed
381	break RUNNING;
382	case TICKET_UNBLOCK:
383	__STATS__(true, ready.threads.threads++; )
384	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
385	// In this case, just run it again.
386	continue RUNNING;
387	default:
388	// This makes no sense, something is wrong abort
389	abort();
390	}
391	}
392
393	// Just before returning to the processor, set the processor coroutine to active
394	proc_cor->state = Active;
395
396	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
397
398	__STATS__(true, ready.threads.threads--; )
399
400	/* paranoid */ verify( ! __preemption_enabled() );
401	}
402
403	// KERNEL_ONLY
404	static void returnToKernel() {
405	/* paranoid */ verify( ! __preemption_enabled() );
406	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
407	thread$ * thrd_src = kernelTLS().this_thread;
408
409	__STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
410
411	// Run the thread on this processor
412	{
413	int local_errno = *__volatile_errno();
414	#if defined( __i386 ) \|\| defined( __x86_64 )
415	__x87_store;
416	#endif
417	/* paranoid */ verify( proc_cor->context.SP );
418	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
419	__cfactx_switch( &thrd_src->context, &proc_cor->context );
420	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
421	#if defined( __i386 ) \|\| defined( __x86_64 )
422	__x87_load;
423	#endif
424	*__volatile_errno() = local_errno;
425	}
426
427	#if !defined(__CFA_NO_STATISTICS__)
428	/* paranoid */ verify( thrd_src->last_proc != 0p );
429	if(thrd_src->last_proc != kernelTLS().this_processor) {
430	__tls_stats()->ready.threads.migration++;
431	}
432	#endif
433
434	/* paranoid */ verify( ! __preemption_enabled() );
435	/* 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 );
436	/* 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 );
437	}
438
439	//-----------------------------------------------------------------------------
440	// Scheduler routines
441	// KERNEL ONLY
442	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
443	/* paranoid */ verify( ! __preemption_enabled() );
444	/* paranoid */ verify( ready_schedule_islocked());
445	/* paranoid */ verify( thrd );
446	/* paranoid */ verify( thrd->state != Halted );
447	/* paranoid */ verify( thrd->curr_cluster );
448	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
449	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
450	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
451	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
452	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
453	/* paranoid */ #endif
454	/* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
455	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
456
457	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
458
459	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
460	struct cluster * cl = thrd->curr_cluster;
461	__STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
462
463	// push the thread to the cluster ready-queue
464	push( cl, thrd, hint );
465
466	// variable thrd is no longer safe to use
467	thrd = 0xdeaddeaddeaddeadp;
468
469	// wake the cluster using the save variable.
470	__wake_one( cl );
471
472	#if !defined(__CFA_NO_STATISTICS__)
473	if( kernelTLS().this_stats ) {
474	__tls_stats()->ready.threads.threads++;
475	if(outside) {
476	__tls_stats()->ready.threads.extunpark++;
477	}
478	}
479	else {
480	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
481	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
482	}
483	#endif
484
485	/* paranoid */ verify( ready_schedule_islocked());
486	/* paranoid */ verify( ! __preemption_enabled() );
487	}
488
489	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
490	ready_schedule_lock();
491	__schedule_thread( thrd, hint );
492	ready_schedule_unlock();
493	}
494
495	// KERNEL ONLY
496	static inline thread$ * __next_thread(cluster * this) with( *this ) {
497	/* paranoid */ verify( ! __preemption_enabled() );
498
499	ready_schedule_lock();
500	thread$ * thrd = pop_fast( this );
501	ready_schedule_unlock();
502
503	/* paranoid */ verify( ! __preemption_enabled() );
504	return thrd;
505	}
506
507	// KERNEL ONLY
508	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
509	/* paranoid */ verify( ! __preemption_enabled() );
510
511	ready_schedule_lock();
512	thread$ * thrd = pop_slow( this );
513	ready_schedule_unlock();
514
515	/* paranoid */ verify( ! __preemption_enabled() );
516	return thrd;
517	}
518
519	// KERNEL ONLY
520	static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
521	/* paranoid */ verify( ! __preemption_enabled() );
522
523	ready_schedule_lock();
524	thread$ * thrd = pop_search( this );
525	ready_schedule_unlock();
526
527	/* paranoid */ verify( ! __preemption_enabled() );
528	return thrd;
529	}
530
531	static inline bool __must_unpark( thread$ * thrd ) {
532	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
533	switch(old_ticket) {
534	case TICKET_RUNNING:
535	// Wake won the race, the thread will reschedule/rerun itself
536	return false;
537	case TICKET_BLOCKED:
538	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
539	/* paranoid */ verify( thrd->state == Blocked );
540	return true;
541	default:
542	// This makes no sense, something is wrong abort
543	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
544	}
545	}
546
547	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
548	/* paranoid */ verify( ! __preemption_enabled() );
549	/* paranoid */ verify( ready_schedule_islocked());
550
551	if( !thrd ) return;
552
553	if(__must_unpark(thrd)) {
554	// Wake lost the race,
555	__schedule_thread( thrd, hint );
556	}
557
558	/* paranoid */ verify( ready_schedule_islocked());
559	/* paranoid */ verify( ! __preemption_enabled() );
560	}
561
562	void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
563	if( !thrd ) return;
564
565	if(__must_unpark(thrd)) {
566	disable_interrupts();
567	// Wake lost the race,
568	schedule_thread$( thrd, hint );
569	enable_interrupts(false);
570	}
571	}
572
573	void park( void ) libcfa_public {
574	__disable_interrupts_checked();
575	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
576	returnToKernel();
577	__enable_interrupts_checked();
578
579	}
580
581	extern "C" {
582	// Leave the thread monitor
583	// last routine called by a thread.
584	// Should never return
585	void __cfactx_thrd_leave() {
586	thread$ * thrd = active_thread();
587	monitor$ * this = &thrd->self_mon;
588
589	// Lock the monitor now
590	lock( this->lock __cfaabi_dbg_ctx2 );
591
592	disable_interrupts();
593
594	/* paranoid */ verify( ! __preemption_enabled() );
595	/* paranoid */ verify( thrd->state == Active );
596	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
597	/* paranoid */ verify( kernelTLS().this_thread == thrd );
598	/* paranoid */ verify( thrd->context.SP );
599	/* 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 );
600	/* 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 );
601
602	thrd->state = Halting;
603	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
604	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
605	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
606
607	// Leave the thread
608	returnToKernel();
609
610	// Control flow should never reach here!
611	abort();
612	}
613	}
614
615	// KERNEL ONLY
616	bool force_yield( __Preemption_Reason reason ) libcfa_public {
617	__disable_interrupts_checked();
618	thread$ * thrd = kernelTLS().this_thread;
619	/* paranoid */ verify(thrd->state == Active);
620
621	// SKULLDUGGERY: It is possible that we are preempting this thread just before
622	// it was going to park itself. If that is the case and it is already using the
623	// intrusive fields then we can't use them to preempt the thread
624	// If that is the case, abandon the preemption.
625	bool preempted = false;
626	if(thrd->link.next == 0p) {
627	preempted = true;
628	thrd->preempted = reason;
629	returnToKernel();
630	}
631	__enable_interrupts_checked( false );
632	return preempted;
633	}
634
635	//=============================================================================================
636	// Kernel Idle Sleep
637	//=============================================================================================
638	// Wake a thread from the front if there are any
639	static void __wake_one(cluster * this) {
640	eventfd_t val;
641
642	/* paranoid */ verify( ! __preemption_enabled() );
643	/* paranoid */ verify( ready_schedule_islocked() );
644
645	// Check if there is a sleeping processor
646	struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
647
648	// If no one is sleeping: we are done
649	if( fdp == 0p ) return;
650
651	int fd = 1;
652	if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
653	fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
654	}
655
656	switch(fd) {
657	__attribute__((unused)) int ret;
658	case 0:
659	// If the processor isn't ready to sleep then the exchange will already wake it up
660	#if !defined(__CFA_NO_STATISTICS__)
661	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
662	} else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
663	#endif
664	break;
665	case 1:
666	// If someone else already said they will wake them: we are done
667	#if !defined(__CFA_NO_STATISTICS__)
668	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
669	} else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
670	#endif
671	break;
672	default:
673	// If the processor was ready to sleep, we need to wake it up with an actual write
674	val = 1;
675	ret = eventfd_write( fd, val );
676	/* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
677
678	#if !defined(__CFA_NO_STATISTICS__)
679	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
680	} else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
681	#endif
682	break;
683	}
684
685	/* paranoid */ verify( ready_schedule_islocked() );
686	/* paranoid */ verify( ! __preemption_enabled() );
687
688	return;
689	}
690
691	// Unconditionnaly wake a thread
692	void __wake_proc(processor * this) {
693	/* paranoid */ verify( ! __preemption_enabled() );
694
695	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
696
697	this->idle_wctx.sem = 1;
698
699	this->idle_wctx.wake__time = rdtscl();
700
701	eventfd_t val;
702	val = 1;
703	__attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
704
705	/* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
706	/* paranoid */ verify( ! __preemption_enabled() );
707	}
708
709	static void idle_sleep(processor * this) {
710	/* paranoid */ verify( this->idle_wctx.evfd != 1 );
711	/* paranoid */ verify( this->idle_wctx.evfd != 2 );
712
713	// Tell everyone we are ready to go do sleep
714	for() {
715	int expected = this->idle_wctx.sem;
716
717	// Someone already told us to wake-up! No time for a nap.
718	if(expected == 1) { return; }
719
720	// Try to mark that we are going to sleep
721	if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
722	// Every one agreed, taking a nap
723	break;
724	}
725	}
726
727
728	#if !defined(CFA_WITH_IO_URING_IDLE)
729	#if !defined(__CFA_NO_STATISTICS__)
730	if(this->print_halts) {
731	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
732	}
733	#endif
734
735	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
736
737	{
738	eventfd_t val;
739	ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
740	if(ret < 0) {
741	switch((int)errno) {
742	case EAGAIN:
743	#if EAGAIN != EWOULDBLOCK
744	case EWOULDBLOCK:
745	#endif
746	case EINTR:
747	// No need to do anything special here, just assume it's a legitimate wake-up
748	break;
749	default:
750	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
751	}
752	}
753	}
754
755	#if !defined(__CFA_NO_STATISTICS__)
756	if(this->print_halts) {
757	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
758	}
759	#endif
760	#else
761	__cfa_io_idle( this );
762	#endif
763	}
764
765	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
766	__STATS__(true, ready.sleep.halts++; )
767
768	proc.idle_wctx.sem = 0;
769
770	/* paranoid */ verify( ! __preemption_enabled() );
771	if(!try_lock( this )) return false;
772	this.idle++;
773	/* paranoid */ verify( this.idle <= this.total );
774	remove(proc);
775	insert_first(this.idles, proc);
776
777	__atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
778	unlock( this );
779	/* paranoid */ verify( ! __preemption_enabled() );
780
781	return true;
782	}
783
784	static void mark_awake(__cluster_proc_list & this, processor & proc) {
785	/* paranoid */ verify( ! __preemption_enabled() );
786	lock( this );
787	this.idle--;
788	/* paranoid */ verify( this.idle >= 0 );
789	remove(proc);
790	insert_last(this.actives, proc);
791
792	{
793	struct __fd_waitctx * wctx = 0;
794	if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
795	__atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
796	}
797
798	unlock( this );
799	/* paranoid */ verify( ! __preemption_enabled() );
800	}
801
802	//=============================================================================================
803	// Unexpected Terminating logic
804	//=============================================================================================
805	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
806	thread$ * thrd = __cfaabi_tls.this_thread;
807
808	if(thrd) {
809	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
810	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
811
812	if ( &thrd->self_cor != thrd->curr_cor ) {
813	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
814	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
815	}
816	else {
817	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
818	}
819	}
820	else {
821	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
822	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
823	}
824	}
825
826	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
827	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
828	}
829
830	static __spinlock_t kernel_debug_lock;
831
832	extern "C" {
833	void __cfaabi_bits_acquire() {
834	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
835	}
836
837	void __cfaabi_bits_release() {
838	unlock( kernel_debug_lock );
839	}
840	}
841
842	//=============================================================================================
843	// Kernel Utilities
844	//=============================================================================================
845	#if defined(CFA_HAVE_LINUX_IO_URING_H)
846	#include "io/types.hfa"
847	#endif
848
849	//-----------------------------------------------------------------------------
850	// Debug
851	bool threading_enabled(void) __attribute__((const)) libcfa_public {
852	return true;
853	}
854
855	//-----------------------------------------------------------------------------
856	// Statistics
857	#if !defined(__CFA_NO_STATISTICS__)
858	void print_halts( processor & this ) libcfa_public {
859	this.print_halts = true;
860	}
861
862	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
863	/* paranoid */ verify( cltr->stats );
864
865	processor * it = &list`first;
866	for(unsigned i = 0; i < count; i++) {
867	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
868	/* paranoid */ verify( it->local_data->this_stats );
869	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
870	__tally_stats( cltr->stats, it->local_data->this_stats );
871	it = &(*it)`next;
872	}
873	}
874
875	static void crawl_cluster_stats( cluster & this ) {
876	// Stop the world, otherwise stats could get really messed-up
877	// this doesn't solve all problems but does solve many
878	// so it's probably good enough
879	disable_interrupts();
880	uint_fast32_t last_size = ready_mutate_lock();
881
882	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
883	crawl_list(&this, this.procs.idles , this.procs.idle );
884
885	// Unlock the RWlock
886	ready_mutate_unlock( last_size );
887	enable_interrupts();
888	}
889
890
891	void print_stats_now( cluster & this, int flags ) libcfa_public {
892	crawl_cluster_stats( this );
893	__print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
894	}
895	#endif
896	// Local Variables: //
897	// mode: c //
898	// tab-width: 4 //
899	// End: //

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

Download in other formats:

Original Format