source: libcfa/src/concurrency/kernel.cfa@ 5b15c4f

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since 5b15c4f was 39fc03e, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

Fixed ready_queue working with 0/1 processors on cluster
  • Property mode set to 100644
File size: 38.2 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 : Thu Jul 9 06:22:54 2020
13// Update Count : 66
14//
15
16#define __cforall_thread__
17// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
18
19//C Includes
20#include <stddef.h>
21#include <errno.h>
22#include <string.h>
23#include <stdio.h>
24#include <fenv.h>
25#include <signal.h>
26#include <unistd.h>
27#include <limits.h> // PTHREAD_STACK_MIN
28#include <sys/mman.h> // mprotect
29extern "C" {
30#include <sys/resource.h>
31}
32
33//CFA Includes
34#include "time.hfa"
35#include "kernel_private.hfa"
36#include "preemption.hfa"
37#include "startup.hfa"
38
39//Private includes
40#define __CFA_INVOKE_PRIVATE__
41#include "invoke.h"
42
43
44//-----------------------------------------------------------------------------
45// Some assembly required
46#if defined( __i386 )
47 #define CtxGet( ctx ) \
48 __asm__ volatile ( \
49 "movl %%esp,%0\n"\
50 "movl %%ebp,%1\n"\
51 : "=rm" (ctx.SP),\
52 "=rm" (ctx.FP) \
53 )
54
55 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
56 // fcw : X87 FPU control word (preserved across function calls)
57 #define __x87_store \
58 uint32_t __mxcr; \
59 uint16_t __fcw; \
60 __asm__ volatile ( \
61 "stmxcsr %0\n" \
62 "fnstcw %1\n" \
63 : "=m" (__mxcr),\
64 "=m" (__fcw) \
65 )
66
67 #define __x87_load \
68 __asm__ volatile ( \
69 "fldcw %1\n" \
70 "ldmxcsr %0\n" \
71 ::"m" (__mxcr),\
72 "m" (__fcw) \
73 )
74
75#elif defined( __x86_64 )
76 #define CtxGet( ctx ) \
77 __asm__ volatile ( \
78 "movq %%rsp,%0\n"\
79 "movq %%rbp,%1\n"\
80 : "=rm" (ctx.SP),\
81 "=rm" (ctx.FP) \
82 )
83
84 #define __x87_store \
85 uint32_t __mxcr; \
86 uint16_t __fcw; \
87 __asm__ volatile ( \
88 "stmxcsr %0\n" \
89 "fnstcw %1\n" \
90 : "=m" (__mxcr),\
91 "=m" (__fcw) \
92 )
93
94 #define __x87_load \
95 __asm__ volatile ( \
96 "fldcw %1\n" \
97 "ldmxcsr %0\n" \
98 :: "m" (__mxcr),\
99 "m" (__fcw) \
100 )
101
102
103#elif defined( __ARM_ARCH )
104#define CtxGet( ctx ) __asm__ ( \
105 "mov %0,%%sp\n" \
106 "mov %1,%%r11\n" \
107 : "=rm" (ctx.SP), "=rm" (ctx.FP) )
108#else
109 #error unknown hardware architecture
110#endif
111
112//-----------------------------------------------------------------------------
113//Start and stop routine for the kernel, declared first to make sure they run first
114static void __kernel_startup (void) __attribute__(( constructor( STARTUP_PRIORITY_KERNEL ) ));
115static void __kernel_shutdown(void) __attribute__(( destructor ( STARTUP_PRIORITY_KERNEL ) ));
116
117//-----------------------------------------------------------------------------
118// Kernel Scheduling logic
119static $thread * __next_thread(cluster * this);
120static bool __has_next_thread(cluster * this);
121static void __run_thread(processor * this, $thread * dst);
122static bool __wake_proc(processor *);
123static bool __wake_one(struct __processor_id_t * id, cluster * cltr);
124static void __halt(processor * this);
125
126//-----------------------------------------------------------------------------
127// Kernel storage
128KERNEL_STORAGE(cluster, mainCluster);
129KERNEL_STORAGE(processor, mainProcessor);
130KERNEL_STORAGE($thread, mainThread);
131KERNEL_STORAGE(__stack_t, mainThreadCtx);
132KERNEL_STORAGE(__scheduler_RWLock_t, __scheduler_lock);
133#if !defined(__CFA_NO_STATISTICS__)
134KERNEL_STORAGE(__stats_t, mainProcStats);
135#endif
136
137cluster * mainCluster;
138processor * mainProcessor;
139$thread * mainThread;
140__scheduler_RWLock_t * __scheduler_lock;
141
142extern "C" {
143 struct { __dllist_t(cluster) list; __spinlock_t lock; } __cfa_dbg_global_clusters;
144}
145
146size_t __page_size = 0;
147
148//-----------------------------------------------------------------------------
149// Global state
150thread_local struct KernelThreadData kernelTLS __attribute__ ((tls_model ( "initial-exec" ))) @= {
151 NULL, // cannot use 0p
152 NULL,
153 NULL,
154 { 1, false, false },
155};
156
157//-----------------------------------------------------------------------------
158// Struct to steal stack
159struct current_stack_info_t {
160 __stack_t * storage; // pointer to stack object
161 void * base; // base of stack
162 void * limit; // stack grows towards stack limit
163 void * context; // address of cfa_context_t
164};
165
166void ?{}( current_stack_info_t & this ) {
167 __stack_context_t ctx;
168 CtxGet( ctx );
169 this.base = ctx.FP;
170
171 rlimit r;
172 getrlimit( RLIMIT_STACK, &r);
173 size_t size = r.rlim_cur;
174
175 this.limit = (void *)(((intptr_t)this.base) - size);
176 this.context = &storage_mainThreadCtx;
177}
178
179//-----------------------------------------------------------------------------
180// Main thread construction
181
182void ?{}( $coroutine & this, current_stack_info_t * info) with( this ) {
183 stack.storage = info->storage;
184 with(*stack.storage) {
185 limit = info->limit;
186 base = info->base;
187 }
188 __attribute__((may_alias)) intptr_t * istorage = (intptr_t*) &stack.storage;
189 *istorage |= 0x1;
190 name = "Main Thread";
191 state = Start;
192 starter = 0p;
193 last = 0p;
194 cancellation = 0p;
195}
196
197void ?{}( $thread & this, current_stack_info_t * info) with( this ) {
198 ticket = 1;
199 state = Start;
200 self_cor{ info };
201 curr_cor = &self_cor;
202 curr_cluster = mainCluster;
203 self_mon.owner = &this;
204 self_mon.recursion = 1;
205 self_mon_p = &self_mon;
206 link.next = 0p;
207 link.prev = 0p;
208
209 node.next = 0p;
210 node.prev = 0p;
211 doregister(curr_cluster, this);
212
213 monitors{ &self_mon_p, 1, (fptr_t)0 };
214}
215
216//-----------------------------------------------------------------------------
217// Processor coroutine
218void ?{}(processorCtx_t & this) {
219
220}
221
222// Construct the processor context of non-main processors
223static void ?{}(processorCtx_t & this, processor * proc, current_stack_info_t * info) {
224 (this.__cor){ info };
225 this.proc = proc;
226}
227
228static void * __invoke_processor(void * arg);
229
230static init(processor & this, const char name[], cluster & _cltr) with( this ) {
231 this.name = name;
232 this.cltr = &_cltr;
233 id = -1u;
234 destroyer = 0p;
235 do_terminate = false;
236 preemption_alarm = 0p;
237 pending_preemption = false;
238
239 #if !defined(__CFA_NO_STATISTICS__)
240 print_stats = 0;
241 print_halts = false;
242 #endif
243
244 int target = __atomic_add_fetch( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
245
246 id = doregister((__processor_id_t*)&this);
247
248 // Lock the RWlock so no-one pushes/pops while we are changing the queue
249 uint_fast32_t last_size = ready_mutate_lock();
250
251 // Adjust the ready queue size
252 ready_queue_grow( cltr, target );
253
254 // Unlock the RWlock
255 ready_mutate_unlock( last_size );
256
257 __cfadbg_print_safe(runtime_core, "Kernel : core %p created\n", &this);
258}
259
260// Not a ctor, it just preps the destruction but should not destroy members
261void deinit(processor & this) {
262
263 int target = __atomic_sub_fetch( &this.cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
264
265 // Lock the RWlock so no-one pushes/pops while we are changing the queue
266 uint_fast32_t last_size = ready_mutate_lock();
267
268 // Adjust the ready queue size
269 ready_queue_shrink( this.cltr, target );
270
271 // Make sure we aren't on the idle queue
272 unsafe_remove( this.cltr->idles, &this );
273
274 // Unlock the RWlock
275 ready_mutate_unlock( last_size );
276
277 // Finally we don't need the read_lock any more
278 unregister((__processor_id_t*)&this);
279}
280
281void ?{}(processor & this, const char name[], cluster & _cltr) {
282 ( this.idle ){};
283 ( this.terminated ){ 0 };
284 ( this.runner ){};
285 init( this, name, _cltr );
286
287 __cfadbg_print_safe(runtime_core, "Kernel : Starting core %p\n", &this);
288
289 this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
290
291}
292
293void ^?{}(processor & this) with( this ){
294 if( ! __atomic_load_n(&do_terminate, __ATOMIC_ACQUIRE) ) {
295 __cfadbg_print_safe(runtime_core, "Kernel : core %p signaling termination\n", &this);
296
297 __atomic_store_n(&do_terminate, true, __ATOMIC_RELAXED);
298 __wake_proc( &this );
299
300 P( terminated );
301 verify( kernelTLS.this_processor != &this);
302 }
303
304 int err = pthread_join( kernel_thread, 0p );
305 if( err != 0 ) abort("KERNEL ERROR: joining processor %p caused error %s\n", &this, strerror(err));
306
307 free( this.stack );
308
309 deinit( this );
310}
311
312void ?{}(cluster & this, const char name[], Duration preemption_rate, unsigned io_flags) with( this ) {
313 this.name = name;
314 this.preemption_rate = preemption_rate;
315 this.nprocessors = 0;
316 ready_queue{};
317
318 #if !defined(__CFA_NO_STATISTICS__)
319 print_stats = 0;
320 stats = alloc();
321 __init_stats( stats );
322 #endif
323
324 threads{ __get };
325
326 doregister(this);
327
328 // Lock the RWlock so no-one pushes/pops while we are changing the queue
329 uint_fast32_t last_size = ready_mutate_lock();
330
331 // Adjust the ready queue size
332 ready_queue_grow( &this, 0 );
333
334 // Unlock the RWlock
335 ready_mutate_unlock( last_size );
336
337
338 __kernel_io_startup( this, io_flags, &this == mainCluster );
339}
340
341void ^?{}(cluster & this) {
342 __kernel_io_shutdown( this, &this == mainCluster );
343
344 // Lock the RWlock so no-one pushes/pops while we are changing the queue
345 uint_fast32_t last_size = ready_mutate_lock();
346
347 // Adjust the ready queue size
348 ready_queue_shrink( &this, 0 );
349
350 // Unlock the RWlock
351 ready_mutate_unlock( last_size );
352
353 #if !defined(__CFA_NO_STATISTICS__)
354 if( 0 != this.print_stats ) {
355 __print_stats( this.stats, this.print_stats, true, this.name, (void*)&this );
356 }
357 free( this.stats );
358 #endif
359
360 unregister(this);
361}
362
363//=============================================================================================
364// Kernel Scheduling logic
365//=============================================================================================
366//Main of the processor contexts
367void main(processorCtx_t & runner) {
368 // Because of a bug, we couldn't initialized the seed on construction
369 // Do it here
370 kernelTLS.rand_seed ^= rdtscl();
371
372 processor * this = runner.proc;
373 verify(this);
374
375 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
376 #if !defined(__CFA_NO_STATISTICS__)
377 if( this->print_halts ) {
378 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->id, this->name, (void*)this);
379 }
380 #endif
381
382 {
383 // Setup preemption data
384 preemption_scope scope = { this };
385
386 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
387
388 $thread * readyThread = 0p;
389 for( unsigned int spin_count = 0;; spin_count++ ) {
390 // Try to get the next thread
391 readyThread = __next_thread( this->cltr );
392
393 // Check if we actually found a thread
394 if( readyThread ) {
395 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
396 /* paranoid */ verifyf( readyThread->state == Ready || readyThread->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", readyThread->state, readyThread->preempted);
397 /* paranoid */ verifyf( readyThread->link.next == 0p, "Expected null got %p", readyThread->link.next );
398 __builtin_prefetch( readyThread->context.SP );
399
400 // We found a thread run it
401 __run_thread(this, readyThread);
402
403 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
404 }
405
406 if(__atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST)) break;
407
408 if( !readyThread ) {
409 // Block until a thread is ready
410 __halt(this);
411 }
412 }
413
414 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
415 }
416
417 V( this->terminated );
418
419 if(this == mainProcessor) {
420 // HACK : the coroutine context switch expects this_thread to be set
421 // and it make sense for it to be set in all other cases except here
422 // fake it
423 kernelTLS.this_thread = mainThread;
424 }
425
426 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
427}
428
429static int * __volatile_errno() __attribute__((noinline));
430static int * __volatile_errno() { asm(""); return &errno; }
431
432// KERNEL ONLY
433// runThread runs a thread by context switching
434// from the processor coroutine to the target thread
435static void __run_thread(processor * this, $thread * thrd_dst) {
436 $coroutine * proc_cor = get_coroutine(this->runner);
437
438 // Update global state
439 kernelTLS.this_thread = thrd_dst;
440
441 // set state of processor coroutine to inactive
442 verify(proc_cor->state == Active);
443 proc_cor->state = Blocked;
444
445 // Actually run the thread
446 RUNNING: while(true) {
447 thrd_dst->preempted = __NO_PREEMPTION;
448 thrd_dst->state = Active;
449
450 __cfaabi_dbg_debug_do(
451 thrd_dst->park_stale = true;
452 thrd_dst->unpark_stale = true;
453 )
454
455 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
456 /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
457 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
458 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
459
460 // set context switch to the thread that the processor is executing
461 verify( thrd_dst->context.SP );
462 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
463 // when __cfactx_switch returns we are back in the processor coroutine
464
465 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
466 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
467 /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
468 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
469
470
471 // We just finished running a thread, there are a few things that could have happened.
472 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
473 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
474 // 4 - Preempted
475 // In case 1, we may have won a race so we can't write to the state again.
476 // In case 2, we lost the race so we now own the thread.
477
478 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
479 // The thread was preempted, reschedule it and reset the flag
480 __schedule_thread( (__processor_id_t*)this, thrd_dst );
481 break RUNNING;
482 }
483
484 if(unlikely(thrd_dst->state == Halted)) {
485 // The thread has halted, it should never be scheduled/run again
486 // We may need to wake someone up here since
487 unpark( this->destroyer __cfaabi_dbg_ctx2 );
488 this->destroyer = 0p;
489 break RUNNING;
490 }
491
492 /* paranoid */ verify( thrd_dst->state == Active );
493 thrd_dst->state = Blocked;
494
495 // set state of processor coroutine to active and the thread to inactive
496 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
497 __cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
498 switch(old_ticket) {
499 case 1:
500 // This is case 1, the regular case, nothing more is needed
501 break RUNNING;
502 case 2:
503 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
504 // In this case, just run it again.
505 continue RUNNING;
506 default:
507 // This makes no sense, something is wrong abort
508 abort();
509 }
510 }
511
512 // Just before returning to the processor, set the processor coroutine to active
513 proc_cor->state = Active;
514 kernelTLS.this_thread = 0p;
515}
516
517// KERNEL_ONLY
518void returnToKernel() {
519 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
520 $coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
521 $thread * thrd_src = kernelTLS.this_thread;
522
523 #if !defined(__CFA_NO_STATISTICS__)
524 struct processor * last_proc = kernelTLS.this_processor;
525 #endif
526
527 // Run the thread on this processor
528 {
529 int local_errno = *__volatile_errno();
530 #if defined( __i386 ) || defined( __x86_64 )
531 __x87_store;
532 #endif
533 verify( proc_cor->context.SP );
534 __cfactx_switch( &thrd_src->context, &proc_cor->context );
535 #if defined( __i386 ) || defined( __x86_64 )
536 __x87_load;
537 #endif
538 *__volatile_errno() = local_errno;
539 }
540
541 #if !defined(__CFA_NO_STATISTICS__)
542 if(last_proc != kernelTLS.this_processor) {
543 __tls_stats()->ready.threads.migration++;
544 }
545 #endif
546
547 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
548 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
549 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
550}
551
552// KERNEL_ONLY
553// Context invoker for processors
554// This is the entry point for processors (kernel threads)
555// It effectively constructs a coroutine by stealing the pthread stack
556static void * __invoke_processor(void * arg) {
557 #if !defined( __CFA_NO_STATISTICS__ )
558 __stats_t local_stats;
559 __init_stats( &local_stats );
560 kernelTLS.this_stats = &local_stats;
561 #endif
562
563 processor * proc = (processor *) arg;
564 kernelTLS.this_processor = proc;
565 kernelTLS.this_thread = 0p;
566 kernelTLS.preemption_state.[enabled, disable_count] = [false, 1];
567 // SKULLDUGGERY: We want to create a context for the processor coroutine
568 // which is needed for the 2-step context switch. However, there is no reason
569 // to waste the perfectly valid stack create by pthread.
570 current_stack_info_t info;
571 __stack_t ctx;
572 info.storage = &ctx;
573 (proc->runner){ proc, &info };
574
575 __cfaabi_dbg_print_safe("Coroutine : created stack %p\n", get_coroutine(proc->runner)->stack.storage);
576
577 //Set global state
578 kernelTLS.this_thread = 0p;
579
580 //We now have a proper context from which to schedule threads
581 __cfadbg_print_safe(runtime_core, "Kernel : core %p created (%p, %p)\n", proc, &proc->runner, &ctx);
582
583 // SKULLDUGGERY: Since the coroutine doesn't have its own stack, we can't
584 // resume it to start it like it normally would, it will just context switch
585 // back to here. Instead directly call the main since we already are on the
586 // appropriate stack.
587 get_coroutine(proc->runner)->state = Active;
588 main( proc->runner );
589 get_coroutine(proc->runner)->state = Halted;
590
591 // Main routine of the core returned, the core is now fully terminated
592 __cfadbg_print_safe(runtime_core, "Kernel : core %p main ended (%p)\n", proc, &proc->runner);
593
594 #if !defined(__CFA_NO_STATISTICS__)
595 __tally_stats(proc->cltr->stats, &local_stats);
596 if( 0 != proc->print_stats ) {
597 __print_stats( &local_stats, proc->print_stats, true, proc->name, (void*)proc );
598 }
599 #endif
600
601 return 0p;
602}
603
604static void Abort( int ret, const char func[] ) {
605 if ( ret ) { // pthread routines return errno values
606 abort( "%s : internal error, error(%d) %s.", func, ret, strerror( ret ) );
607 } // if
608} // Abort
609
610void * __create_pthread( pthread_t * pthread, void * (*start)(void *), void * arg ) {
611 pthread_attr_t attr;
612
613 Abort( pthread_attr_init( &attr ), "pthread_attr_init" ); // initialize attribute
614
615 size_t stacksize;
616 // default stack size, normally defined by shell limit
617 Abort( pthread_attr_getstacksize( &attr, &stacksize ), "pthread_attr_getstacksize" );
618 assert( stacksize >= PTHREAD_STACK_MIN );
619
620 void * stack;
621 __cfaabi_dbg_debug_do(
622 stack = memalign( __page_size, stacksize + __page_size );
623 // pthread has no mechanism to create the guard page in user supplied stack.
624 if ( mprotect( stack, __page_size, PROT_NONE ) == -1 ) {
625 abort( "mprotect : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
626 } // if
627 );
628 __cfaabi_dbg_no_debug_do(
629 stack = malloc( stacksize );
630 );
631
632 Abort( pthread_attr_setstack( &attr, stack, stacksize ), "pthread_attr_setstack" );
633
634 Abort( pthread_create( pthread, &attr, start, arg ), "pthread_create" );
635 return stack;
636}
637
638// KERNEL_ONLY
639static void __kernel_first_resume( processor * this ) {
640 $thread * src = mainThread;
641 $coroutine * dst = get_coroutine(this->runner);
642
643 verify( ! kernelTLS.preemption_state.enabled );
644
645 kernelTLS.this_thread->curr_cor = dst;
646 __stack_prepare( &dst->stack, 65000 );
647 __cfactx_start(main, dst, this->runner, __cfactx_invoke_coroutine);
648
649 verify( ! kernelTLS.preemption_state.enabled );
650
651 dst->last = &src->self_cor;
652 dst->starter = dst->starter ? dst->starter : &src->self_cor;
653
654 // make sure the current state is still correct
655 /* paranoid */ verify(src->state == Ready);
656
657 // context switch to specified coroutine
658 verify( dst->context.SP );
659 __cfactx_switch( &src->context, &dst->context );
660 // when __cfactx_switch returns we are back in the src coroutine
661
662 mainThread->curr_cor = &mainThread->self_cor;
663
664 // make sure the current state has been update
665 /* paranoid */ verify(src->state == Active);
666
667 verify( ! kernelTLS.preemption_state.enabled );
668}
669
670// KERNEL_ONLY
671static void __kernel_last_resume( processor * this ) {
672 $coroutine * src = &mainThread->self_cor;
673 $coroutine * dst = get_coroutine(this->runner);
674
675 verify( ! kernelTLS.preemption_state.enabled );
676 verify( dst->starter == src );
677 verify( dst->context.SP );
678
679 // SKULLDUGGERY in debug the processors check that the
680 // stack is still within the limit of the stack limits after running a thread.
681 // that check doesn't make sense if we context switch to the processor using the
682 // coroutine semantics. Since this is a special case, use the current context
683 // info to populate these fields.
684 __cfaabi_dbg_debug_do(
685 __stack_context_t ctx;
686 CtxGet( ctx );
687 mainThread->context.SP = ctx.SP;
688 mainThread->context.FP = ctx.FP;
689 )
690
691 // context switch to the processor
692 __cfactx_switch( &src->context, &dst->context );
693}
694
695//-----------------------------------------------------------------------------
696// Scheduler routines
697// KERNEL ONLY
698void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
699 /* paranoid */ verify( thrd );
700 /* paranoid */ verify( thrd->state != Halted );
701 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
702 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
703 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
704 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
705 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
706 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
707 /* paranoid */ #endif
708 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
709
710 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
711
712 ready_schedule_lock ( id );
713 push( thrd->curr_cluster, thrd );
714
715 #if !defined(__CFA_NO_STATISTICS__)
716 bool woke =
717 #endif
718 __wake_one(id, thrd->curr_cluster);
719
720 #if !defined(__CFA_NO_STATISTICS__)
721 if(woke) __tls_stats()->ready.sleep.wakes++;
722 #endif
723 ready_schedule_unlock( id );
724
725 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
726}
727
728// KERNEL ONLY
729static $thread * __next_thread(cluster * this) with( *this ) {
730 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
731
732 ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
733 $thread * head = pop( this );
734 ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
735
736 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
737 return head;
738}
739
740// KERNEL ONLY
741static bool __has_next_thread(cluster * this) with( *this ) {
742 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
743
744 ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
745 bool not_empty = query( this );
746 ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
747
748 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
749 return not_empty;
750}
751
752// KERNEL ONLY unpark with out disabling interrupts
753void __unpark( struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
754 // record activity
755 __cfaabi_dbg_record_thrd( *thrd, false, caller );
756
757 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
758 __cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
759 switch(old_ticket) {
760 case 1:
761 // Wake won the race, the thread will reschedule/rerun itself
762 break;
763 case 0:
764 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
765 /* paranoid */ verify( thrd->state == Blocked );
766
767 // Wake lost the race,
768 __schedule_thread( id, thrd );
769 break;
770 default:
771 // This makes no sense, something is wrong abort
772 abort();
773 }
774}
775
776void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
777 if( !thrd ) return;
778
779 disable_interrupts();
780 __unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
781 enable_interrupts( __cfaabi_dbg_ctx );
782}
783
784void park( __cfaabi_dbg_ctx_param ) {
785 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
786 disable_interrupts();
787 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
788 /* paranoid */ verify( kernelTLS.this_thread->preempted == __NO_PREEMPTION );
789
790 // record activity
791 __cfaabi_dbg_record_thrd( *kernelTLS.this_thread, true, caller );
792
793 returnToKernel();
794
795 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
796 enable_interrupts( __cfaabi_dbg_ctx );
797 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
798
799}
800
801// KERNEL ONLY
802void __leave_thread() {
803 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
804 returnToKernel();
805 abort();
806}
807
808// KERNEL ONLY
809bool force_yield( __Preemption_Reason reason ) {
810 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
811 disable_interrupts();
812 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
813
814 $thread * thrd = kernelTLS.this_thread;
815 /* paranoid */ verify(thrd->state == Active);
816
817 // SKULLDUGGERY: It is possible that we are preempting this thread just before
818 // it was going to park itself. If that is the case and it is already using the
819 // intrusive fields then we can't use them to preempt the thread
820 // If that is the case, abandon the preemption.
821 bool preempted = false;
822 if(thrd->link.next == 0p) {
823 preempted = true;
824 thrd->preempted = reason;
825 returnToKernel();
826 }
827
828 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
829 enable_interrupts_noPoll();
830 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
831
832 return preempted;
833}
834
835//=============================================================================================
836// Kernel Setup logic
837//=============================================================================================
838//-----------------------------------------------------------------------------
839// Kernel boot procedures
840static void __kernel_startup(void) {
841 verify( ! kernelTLS.preemption_state.enabled );
842 __cfadbg_print_safe(runtime_core, "Kernel : Starting\n");
843
844 __page_size = sysconf( _SC_PAGESIZE );
845
846 __cfa_dbg_global_clusters.list{ __get };
847 __cfa_dbg_global_clusters.lock{};
848
849 // Initialize the global scheduler lock
850 __scheduler_lock = (__scheduler_RWLock_t*)&storage___scheduler_lock;
851 (*__scheduler_lock){};
852
853 // Initialize the main cluster
854 mainCluster = (cluster *)&storage_mainCluster;
855 (*mainCluster){"Main Cluster"};
856
857 __cfadbg_print_safe(runtime_core, "Kernel : Main cluster ready\n");
858
859 // Start by initializing the main thread
860 // SKULLDUGGERY: the mainThread steals the process main thread
861 // which will then be scheduled by the mainProcessor normally
862 mainThread = ($thread *)&storage_mainThread;
863 current_stack_info_t info;
864 info.storage = (__stack_t*)&storage_mainThreadCtx;
865 (*mainThread){ &info };
866
867 __cfadbg_print_safe(runtime_core, "Kernel : Main thread ready\n");
868
869
870
871 // Construct the processor context of the main processor
872 void ?{}(processorCtx_t & this, processor * proc) {
873 (this.__cor){ "Processor" };
874 this.__cor.starter = 0p;
875 this.proc = proc;
876 }
877
878 void ?{}(processor & this) with( this ) {
879 ( this.idle ){};
880 ( this.terminated ){ 0 };
881 ( this.runner ){};
882 init( this, "Main Processor", *mainCluster );
883 kernel_thread = pthread_self();
884
885 runner{ &this };
886 __cfadbg_print_safe(runtime_core, "Kernel : constructed main processor context %p\n", &runner);
887 }
888
889 // Initialize the main processor and the main processor ctx
890 // (the coroutine that contains the processing control flow)
891 mainProcessor = (processor *)&storage_mainProcessor;
892 (*mainProcessor){};
893
894 //initialize the global state variables
895 kernelTLS.this_processor = mainProcessor;
896 kernelTLS.this_thread = mainThread;
897
898 #if !defined( __CFA_NO_STATISTICS__ )
899 kernelTLS.this_stats = (__stats_t *)& storage_mainProcStats;
900 __init_stats( kernelTLS.this_stats );
901 #endif
902
903 // Enable preemption
904 kernel_start_preemption();
905
906 // Add the main thread to the ready queue
907 // once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
908 __schedule_thread((__processor_id_t *)mainProcessor, mainThread);
909
910 // SKULLDUGGERY: Force a context switch to the main processor to set the main thread's context to the current UNIX
911 // context. Hence, the main thread does not begin through __cfactx_invoke_thread, like all other threads. The trick here is that
912 // mainThread is on the ready queue when this call is made.
913 __kernel_first_resume( kernelTLS.this_processor );
914
915
916 // THE SYSTEM IS NOW COMPLETELY RUNNING
917
918
919 // Now that the system is up, finish creating systems that need threading
920 __kernel_io_finish_start( *mainCluster );
921
922
923 __cfadbg_print_safe(runtime_core, "Kernel : Started\n--------------------------------------------------\n\n");
924
925 verify( ! kernelTLS.preemption_state.enabled );
926 enable_interrupts( __cfaabi_dbg_ctx );
927 verify( TL_GET( preemption_state.enabled ) );
928}
929
930static void __kernel_shutdown(void) {
931 //Before we start shutting things down, wait for systems that need threading to shutdown
932 __kernel_io_prepare_stop( *mainCluster );
933
934 /* paranoid */ verify( TL_GET( preemption_state.enabled ) );
935 disable_interrupts();
936 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
937
938 __cfadbg_print_safe(runtime_core, "\n--------------------------------------------------\nKernel : Shutting down\n");
939
940 // SKULLDUGGERY: Notify the mainProcessor it needs to terminates.
941 // When its coroutine terminates, it return control to the mainThread
942 // which is currently here
943 __atomic_store_n(&mainProcessor->do_terminate, true, __ATOMIC_RELEASE);
944 __kernel_last_resume( kernelTLS.this_processor );
945 mainThread->self_cor.state = Halted;
946
947 // THE SYSTEM IS NOW COMPLETELY STOPPED
948
949 // Disable preemption
950 kernel_stop_preemption();
951
952 // Destroy the main processor and its context in reverse order of construction
953 // These were manually constructed so we need manually destroy them
954 void ^?{}(processor & this) with( this ){
955 deinit( this );
956
957 /* paranoid */ verify( this.do_terminate == true );
958 __cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
959 }
960
961 ^(*mainProcessor){};
962
963 // Final step, destroy the main thread since it is no longer needed
964
965 // Since we provided a stack to this taxk it will not destroy anything
966 /* paranoid */ verify(mainThread->self_cor.stack.storage == (__stack_t*)(((uintptr_t)&storage_mainThreadCtx)| 0x1));
967 ^(*mainThread){};
968
969 ^(*mainCluster){};
970
971 ^(*__scheduler_lock){};
972
973 ^(__cfa_dbg_global_clusters.list){};
974 ^(__cfa_dbg_global_clusters.lock){};
975
976 __cfadbg_print_safe(runtime_core, "Kernel : Shutdown complete\n");
977}
978
979//=============================================================================================
980// Kernel Idle Sleep
981//=============================================================================================
982// Wake a thread from the front if there are any
983static bool __wake_one(struct __processor_id_t * id, cluster * this) {
984 /* paranoid */ verify( ready_schedule_islocked( id ) );
985
986 // Check if there is a sleeping processor
987 processor * p = pop(this->idles);
988
989 // If no one is sleeping, we are done
990 if( 0p == p ) return false;
991
992 // We found a processor, wake it up
993 post( p->idle );
994
995 return true;
996}
997
998// Unconditionnaly wake a thread
999static bool __wake_proc(processor * this) {
1000 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
1001
1002 disable_interrupts();
1003 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
1004 bool ret = post( this->idle );
1005 enable_interrupts( __cfaabi_dbg_ctx );
1006
1007 return ret;
1008}
1009
1010static void __halt(processor * this) with( *this ) {
1011 if( do_terminate ) return;
1012
1013 #if !defined(__CFA_NO_STATISTICS__)
1014 __tls_stats()->ready.sleep.halts++;
1015 #endif
1016 // Push self to queue
1017 push(cltr->idles, *this);
1018
1019 // Makre sure we don't miss a thread
1020 if( __has_next_thread(cltr) ) {
1021 // A thread was posted, make sure a processor is woken up
1022 struct __processor_id_t *id = (struct __processor_id_t *) this;
1023 ready_schedule_lock ( id );
1024 __wake_one( id, cltr );
1025 ready_schedule_unlock( id );
1026 #if !defined(__CFA_NO_STATISTICS__)
1027 __tls_stats()->ready.sleep.cancels++;
1028 #endif
1029 }
1030
1031 #if !defined(__CFA_NO_STATISTICS__)
1032 if(this->print_halts) {
1033 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
1034 }
1035 #endif
1036
1037 wait( idle );
1038
1039 #if !defined(__CFA_NO_STATISTICS__)
1040 if(this->print_halts) {
1041 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
1042 }
1043 #endif
1044}
1045
1046//=============================================================================================
1047// Unexpected Terminating logic
1048//=============================================================================================
1049static __spinlock_t kernel_abort_lock;
1050static bool kernel_abort_called = false;
1051
1052void * kernel_abort(void) __attribute__ ((__nothrow__)) {
1053 // abort cannot be recursively entered by the same or different processors because all signal handlers return when
1054 // the globalAbort flag is true.
1055 lock( kernel_abort_lock __cfaabi_dbg_ctx2 );
1056
1057 // first task to abort ?
1058 if ( kernel_abort_called ) { // not first task to abort ?
1059 unlock( kernel_abort_lock );
1060
1061 sigset_t mask;
1062 sigemptyset( &mask );
1063 sigaddset( &mask, SIGALRM ); // block SIGALRM signals
1064 sigaddset( &mask, SIGUSR1 ); // block SIGALRM signals
1065 sigsuspend( &mask ); // block the processor to prevent further damage during abort
1066 _exit( EXIT_FAILURE ); // if processor unblocks before it is killed, terminate it
1067 }
1068 else {
1069 kernel_abort_called = true;
1070 unlock( kernel_abort_lock );
1071 }
1072
1073 return kernelTLS.this_thread;
1074}
1075
1076void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
1077 $thread * thrd = kernel_data;
1078
1079 if(thrd) {
1080 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
1081 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1082
1083 if ( &thrd->self_cor != thrd->curr_cor ) {
1084 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
1085 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1086 }
1087 else {
1088 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
1089 }
1090 }
1091 else {
1092 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
1093 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1094 }
1095}
1096
1097int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
1098 return get_coroutine(kernelTLS.this_thread) == get_coroutine(mainThread) ? 4 : 2;
1099}
1100
1101static __spinlock_t kernel_debug_lock;
1102
1103extern "C" {
1104 void __cfaabi_bits_acquire() {
1105 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
1106 }
1107
1108 void __cfaabi_bits_release() {
1109 unlock( kernel_debug_lock );
1110 }
1111}
1112
1113//=============================================================================================
1114// Kernel Utilities
1115//=============================================================================================
1116//-----------------------------------------------------------------------------
1117// Locks
1118void ?{}( semaphore & this, int count = 1 ) {
1119 (this.lock){};
1120 this.count = count;
1121 (this.waiting){};
1122}
1123void ^?{}(semaphore & this) {}
1124
1125bool P(semaphore & this) with( this ){
1126 lock( lock __cfaabi_dbg_ctx2 );
1127 count -= 1;
1128 if ( count < 0 ) {
1129 // queue current task
1130 append( waiting, kernelTLS.this_thread );
1131
1132 // atomically release spin lock and block
1133 unlock( lock );
1134 park( __cfaabi_dbg_ctx );
1135 return true;
1136 }
1137 else {
1138 unlock( lock );
1139 return false;
1140 }
1141}
1142
1143bool V(semaphore & this) with( this ) {
1144 $thread * thrd = 0p;
1145 lock( lock __cfaabi_dbg_ctx2 );
1146 count += 1;
1147 if ( count <= 0 ) {
1148 // remove task at head of waiting list
1149 thrd = pop_head( waiting );
1150 }
1151
1152 unlock( lock );
1153
1154 // make new owner
1155 unpark( thrd __cfaabi_dbg_ctx2 );
1156
1157 return thrd != 0p;
1158}
1159
1160bool V(semaphore & this, unsigned diff) with( this ) {
1161 $thread * thrd = 0p;
1162 lock( lock __cfaabi_dbg_ctx2 );
1163 int release = max(-count, (int)diff);
1164 count += diff;
1165 for(release) {
1166 unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 );
1167 }
1168
1169 unlock( lock );
1170
1171 return thrd != 0p;
1172}
1173
1174//-----------------------------------------------------------------------------
1175// Global Queues
1176void doregister( cluster & cltr ) {
1177 lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1178 push_front( __cfa_dbg_global_clusters.list, cltr );
1179 unlock ( __cfa_dbg_global_clusters.lock );
1180}
1181
1182void unregister( cluster & cltr ) {
1183 lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1184 remove( __cfa_dbg_global_clusters.list, cltr );
1185 unlock( __cfa_dbg_global_clusters.lock );
1186}
1187
1188void doregister( cluster * cltr, $thread & thrd ) {
1189 lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1190 cltr->nthreads += 1;
1191 push_front(cltr->threads, thrd);
1192 unlock (cltr->thread_list_lock);
1193}
1194
1195void unregister( cluster * cltr, $thread & thrd ) {
1196 lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1197 remove(cltr->threads, thrd );
1198 cltr->nthreads -= 1;
1199 unlock(cltr->thread_list_lock);
1200}
1201
1202//-----------------------------------------------------------------------------
1203// Debug
1204__cfaabi_dbg_debug_do(
1205 extern "C" {
1206 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
1207 this.prev_name = prev_name;
1208 this.prev_thrd = kernelTLS.this_thread;
1209 }
1210
1211 void __cfaabi_dbg_record_thrd($thread & this, bool park, const char prev_name[]) {
1212 if(park) {
1213 this.park_caller = prev_name;
1214 this.park_stale = false;
1215 }
1216 else {
1217 this.unpark_caller = prev_name;
1218 this.unpark_stale = false;
1219 }
1220 }
1221 }
1222)
1223
1224//-----------------------------------------------------------------------------
1225// Debug
1226bool threading_enabled(void) __attribute__((const)) {
1227 return true;
1228}
1229
1230//-----------------------------------------------------------------------------
1231// Statistics
1232#if !defined(__CFA_NO_STATISTICS__)
1233 void print_halts( processor & this ) {
1234 this.print_halts = true;
1235 }
1236#endif
1237// Local Variables: //
1238// mode: c //
1239// tab-width: 4 //
1240// End: //
Note: See TracBrowser for help on using the repository browser.