source: libcfa/src/concurrency/kernel.cfa@ 8b58bae

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 8b58bae was 29cb302, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

cfa stats now count number of migrations
  • Property mode set to 100644
File size: 37.2 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 : Tue May 26 22:05:19 2020
13// Update Count : 59
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 6u //this should be seeded better but due to a bug calling rdtsc doesn't work
156};
157
158//-----------------------------------------------------------------------------
159// Struct to steal stack
160struct current_stack_info_t {
161 __stack_t * storage; // pointer to stack object
162 void * base; // base of stack
163 void * limit; // stack grows towards stack limit
164 void * context; // address of cfa_context_t
165};
166
167void ?{}( current_stack_info_t & this ) {
168 __stack_context_t ctx;
169 CtxGet( ctx );
170 this.base = ctx.FP;
171
172 rlimit r;
173 getrlimit( RLIMIT_STACK, &r);
174 size_t size = r.rlim_cur;
175
176 this.limit = (void *)(((intptr_t)this.base) - size);
177 this.context = &storage_mainThreadCtx;
178}
179
180//-----------------------------------------------------------------------------
181// Main thread construction
182
183void ?{}( $coroutine & this, current_stack_info_t * info) with( this ) {
184 stack.storage = info->storage;
185 with(*stack.storage) {
186 limit = info->limit;
187 base = info->base;
188 }
189 __attribute__((may_alias)) intptr_t * istorage = (intptr_t*) &stack.storage;
190 *istorage |= 0x1;
191 name = "Main Thread";
192 state = Start;
193 starter = 0p;
194 last = 0p;
195 cancellation = 0p;
196}
197
198void ?{}( $thread & this, current_stack_info_t * info) with( this ) {
199 ticket = 1;
200 state = Start;
201 self_cor{ info };
202 curr_cor = &self_cor;
203 curr_cluster = mainCluster;
204 self_mon.owner = &this;
205 self_mon.recursion = 1;
206 self_mon_p = &self_mon;
207 link.next = 0p;
208 link.prev = 0p;
209
210 node.next = 0p;
211 node.prev = 0p;
212 doregister(curr_cluster, this);
213
214 monitors{ &self_mon_p, 1, (fptr_t)0 };
215}
216
217//-----------------------------------------------------------------------------
218// Processor coroutine
219void ?{}(processorCtx_t & this) {
220
221}
222
223// Construct the processor context of non-main processors
224static void ?{}(processorCtx_t & this, processor * proc, current_stack_info_t * info) {
225 (this.__cor){ info };
226 this.proc = proc;
227}
228
229static void * __invoke_processor(void * arg);
230
231void ?{}(processor & this, const char name[], cluster & _cltr) with( this ) {
232 this.name = name;
233 this.cltr = &_cltr;
234 id = -1u;
235 terminated{ 0 };
236 destroyer = 0p;
237 do_terminate = false;
238 preemption_alarm = 0p;
239 pending_preemption = false;
240 runner.proc = &this;
241
242 idle{};
243
244 __cfadbg_print_safe(runtime_core, "Kernel : Starting core %p\n", &this);
245
246 this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
247 __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
248
249 __cfadbg_print_safe(runtime_core, "Kernel : core %p created\n", &this);
250}
251
252void ^?{}(processor & this) with( this ){
253 if( ! __atomic_load_n(&do_terminate, __ATOMIC_ACQUIRE) ) {
254 __cfadbg_print_safe(runtime_core, "Kernel : core %p signaling termination\n", &this);
255
256 __atomic_store_n(&do_terminate, true, __ATOMIC_RELAXED);
257 __wake_proc( &this );
258
259 P( terminated );
260 verify( kernelTLS.this_processor != &this);
261 }
262
263 int err = pthread_join( kernel_thread, 0p );
264 if( err != 0 ) abort("KERNEL ERROR: joining processor %p caused error %s\n", &this, strerror(err));
265
266 free( this.stack );
267
268 __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
269}
270
271void ?{}(cluster & this, const char name[], Duration preemption_rate, unsigned io_flags) with( this ) {
272 this.name = name;
273 this.preemption_rate = preemption_rate;
274 this.nprocessors = 0;
275 ready_queue{};
276
277 #if !defined(__CFA_NO_STATISTICS__)
278 print_stats = false;
279 stats = alloc();
280 __init_stats( stats );
281 #endif
282
283 threads{ __get };
284
285 __kernel_io_startup( this, io_flags, &this == mainCluster );
286
287 doregister(this);
288}
289
290void ^?{}(cluster & this) {
291 __kernel_io_shutdown( this, &this == mainCluster );
292
293 #if !defined(__CFA_NO_STATISTICS__)
294 if(this.print_stats) {
295 __print_stats( this.stats );
296 }
297 free( this.stats );
298 #endif
299
300 unregister(this);
301}
302
303//=============================================================================================
304// Kernel Scheduling logic
305//=============================================================================================
306//Main of the processor contexts
307void main(processorCtx_t & runner) {
308 // Because of a bug, we couldn't initialized the seed on construction
309 // Do it here
310 kernelTLS.rand_seed ^= rdtscl();
311
312 processor * this = runner.proc;
313 verify(this);
314
315 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
316
317 // register the processor unless it's the main thread which is handled in the boot sequence
318 if(this != mainProcessor) {
319 this->id = doregister((__processor_id_t*)this);
320 // Lock the RWlock so no-one pushes/pops while we are changing the queue
321 uint_fast32_t last_size = ready_mutate_lock();
322
323 // Adjust the ready queue size
324 ready_queue_grow( this->cltr );
325
326 // Unlock the RWlock
327 ready_mutate_unlock( last_size );
328 }
329
330 {
331 // Setup preemption data
332 preemption_scope scope = { this };
333
334 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
335
336 $thread * readyThread = 0p;
337 for( unsigned int spin_count = 0; ! __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST); spin_count++ ) {
338 // Try to get the next thread
339 readyThread = __next_thread( this->cltr );
340
341 // Check if we actually found a thread
342 if( readyThread ) {
343 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
344 /* paranoid */ verifyf( readyThread->state == Ready || readyThread->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", readyThread->state, readyThread->preempted);
345 /* paranoid */ verifyf( readyThread->link.next == 0p, "Expected null got %p", readyThread->link.next );
346 __builtin_prefetch( readyThread->context.SP );
347
348 // We found a thread run it
349 __run_thread(this, readyThread);
350
351 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
352 }
353 else {
354 // Block until a thread is ready
355 __halt(this);
356 }
357 }
358
359 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
360 }
361
362 V( this->terminated );
363
364 // unregister the processor unless it's the main thread which is handled in the boot sequence
365 if(this != mainProcessor) {
366 // Lock the RWlock so no-one pushes/pops while we are changing the queue
367 uint_fast32_t last_size = ready_mutate_lock();
368
369 // Adjust the ready queue size
370 ready_queue_shrink( this->cltr );
371
372 // Make sure we aren't on the idle queue
373 #if !defined(__CFA_NO_STATISTICS__)
374 bool removed =
375 #endif
376 unsafe_remove( this->cltr->idles, this );
377
378 #if !defined(__CFA_NO_STATISTICS__)
379 if(removed) __tls_stats()->ready.sleep.exits++;
380 #endif
381
382 // Unlock the RWlock
383 ready_mutate_unlock( last_size );
384
385 // Finally we don't need the read_lock any more
386 unregister((__processor_id_t*)this);
387 }
388 else {
389 // HACK : the coroutine context switch expects this_thread to be set
390 // and it make sense for it to be set in all other cases except here
391 // fake it
392 kernelTLS.this_thread = mainThread;
393 }
394
395 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
396}
397
398static int * __volatile_errno() __attribute__((noinline));
399static int * __volatile_errno() { asm(""); return &errno; }
400
401// KERNEL ONLY
402// runThread runs a thread by context switching
403// from the processor coroutine to the target thread
404static void __run_thread(processor * this, $thread * thrd_dst) {
405 $coroutine * proc_cor = get_coroutine(this->runner);
406
407 // Update global state
408 kernelTLS.this_thread = thrd_dst;
409
410 // set state of processor coroutine to inactive
411 verify(proc_cor->state == Active);
412 proc_cor->state = Blocked;
413
414 // Actually run the thread
415 RUNNING: while(true) {
416 thrd_dst->preempted = __NO_PREEMPTION;
417 thrd_dst->state = Active;
418
419 __cfaabi_dbg_debug_do(
420 thrd_dst->park_stale = true;
421 thrd_dst->unpark_stale = true;
422 )
423
424 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
425 /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
426 /* 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
427 /* 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
428
429 // set context switch to the thread that the processor is executing
430 verify( thrd_dst->context.SP );
431 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
432 // when __cfactx_switch returns we are back in the processor coroutine
433
434 /* 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 );
435 /* 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 );
436 /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
437 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
438
439
440 // We just finished running a thread, there are a few things that could have happened.
441 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
442 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
443 // 4 - Preempted
444 // In case 1, we may have won a race so we can't write to the state again.
445 // In case 2, we lost the race so we now own the thread.
446
447 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
448 // The thread was preempted, reschedule it and reset the flag
449 __schedule_thread( (__processor_id_t*)this, thrd_dst );
450 break RUNNING;
451 }
452
453 if(unlikely(thrd_dst->state == Halted)) {
454 // The thread has halted, it should never be scheduled/run again
455 // We may need to wake someone up here since
456 unpark( this->destroyer __cfaabi_dbg_ctx2 );
457 this->destroyer = 0p;
458 break RUNNING;
459 }
460
461 /* paranoid */ verify( thrd_dst->state == Active );
462 thrd_dst->state = Blocked;
463
464 // set state of processor coroutine to active and the thread to inactive
465 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
466 __cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
467 switch(old_ticket) {
468 case 1:
469 // This is case 1, the regular case, nothing more is needed
470 break RUNNING;
471 case 2:
472 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
473 // In this case, just run it again.
474 continue RUNNING;
475 default:
476 // This makes no sense, something is wrong abort
477 abort();
478 }
479 }
480
481 // Just before returning to the processor, set the processor coroutine to active
482 proc_cor->state = Active;
483 kernelTLS.this_thread = 0p;
484}
485
486// KERNEL_ONLY
487void returnToKernel() {
488 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
489 $coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
490 $thread * thrd_src = kernelTLS.this_thread;
491
492 #if !defined(__CFA_NO_STATISTICS__)
493 struct processor * last_proc = kernelTLS.this_processor;
494 #endif
495
496 // Run the thread on this processor
497 {
498 int local_errno = *__volatile_errno();
499 #if defined( __i386 ) || defined( __x86_64 )
500 __x87_store;
501 #endif
502 verify( proc_cor->context.SP );
503 __cfactx_switch( &thrd_src->context, &proc_cor->context );
504 #if defined( __i386 ) || defined( __x86_64 )
505 __x87_load;
506 #endif
507 *__volatile_errno() = local_errno;
508 }
509
510 #if !defined(__CFA_NO_STATISTICS__)
511 if(last_proc != kernelTLS.this_processor) {
512 __tls_stats()->ready.threads.migration++;
513 }
514 #endif
515
516 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
517 /* 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 );
518 /* 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 );
519}
520
521// KERNEL_ONLY
522// Context invoker for processors
523// This is the entry point for processors (kernel threads)
524// It effectively constructs a coroutine by stealing the pthread stack
525static void * __invoke_processor(void * arg) {
526 #if !defined( __CFA_NO_STATISTICS__ )
527 __stats_t local_stats;
528 __init_stats( &local_stats );
529 kernelTLS.this_stats = &local_stats;
530 #endif
531
532 processor * proc = (processor *) arg;
533 kernelTLS.this_processor = proc;
534 kernelTLS.this_thread = 0p;
535 kernelTLS.preemption_state.[enabled, disable_count] = [false, 1];
536 // SKULLDUGGERY: We want to create a context for the processor coroutine
537 // which is needed for the 2-step context switch. However, there is no reason
538 // to waste the perfectly valid stack create by pthread.
539 current_stack_info_t info;
540 __stack_t ctx;
541 info.storage = &ctx;
542 (proc->runner){ proc, &info };
543
544 __cfaabi_dbg_print_safe("Coroutine : created stack %p\n", get_coroutine(proc->runner)->stack.storage);
545
546 //Set global state
547 kernelTLS.this_thread = 0p;
548
549 //We now have a proper context from which to schedule threads
550 __cfadbg_print_safe(runtime_core, "Kernel : core %p created (%p, %p)\n", proc, &proc->runner, &ctx);
551
552 // SKULLDUGGERY: Since the coroutine doesn't have its own stack, we can't
553 // resume it to start it like it normally would, it will just context switch
554 // back to here. Instead directly call the main since we already are on the
555 // appropriate stack.
556 get_coroutine(proc->runner)->state = Active;
557 main( proc->runner );
558 get_coroutine(proc->runner)->state = Halted;
559
560 // Main routine of the core returned, the core is now fully terminated
561 __cfadbg_print_safe(runtime_core, "Kernel : core %p main ended (%p)\n", proc, &proc->runner);
562
563 #if !defined(__CFA_NO_STATISTICS__)
564 __tally_stats(proc->cltr->stats, &local_stats);
565 #endif
566
567 return 0p;
568}
569
570static void Abort( int ret, const char func[] ) {
571 if ( ret ) { // pthread routines return errno values
572 abort( "%s : internal error, error(%d) %s.", func, ret, strerror( ret ) );
573 } // if
574} // Abort
575
576void * __create_pthread( pthread_t * pthread, void * (*start)(void *), void * arg ) {
577 pthread_attr_t attr;
578
579 Abort( pthread_attr_init( &attr ), "pthread_attr_init" ); // initialize attribute
580
581 size_t stacksize;
582 // default stack size, normally defined by shell limit
583 Abort( pthread_attr_getstacksize( &attr, &stacksize ), "pthread_attr_getstacksize" );
584 assert( stacksize >= PTHREAD_STACK_MIN );
585
586 void * stack;
587 __cfaabi_dbg_debug_do(
588 stack = memalign( __page_size, stacksize + __page_size );
589 // pthread has no mechanism to create the guard page in user supplied stack.
590 if ( mprotect( stack, __page_size, PROT_NONE ) == -1 ) {
591 abort( "mprotect : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
592 } // if
593 );
594 __cfaabi_dbg_no_debug_do(
595 stack = malloc( stacksize );
596 );
597
598 Abort( pthread_attr_setstack( &attr, stack, stacksize ), "pthread_attr_setstack" );
599
600 Abort( pthread_create( pthread, &attr, start, arg ), "pthread_create" );
601 return stack;
602}
603
604// KERNEL_ONLY
605static void __kernel_first_resume( processor * this ) {
606 $thread * src = mainThread;
607 $coroutine * dst = get_coroutine(this->runner);
608
609 verify( ! kernelTLS.preemption_state.enabled );
610
611 kernelTLS.this_thread->curr_cor = dst;
612 __stack_prepare( &dst->stack, 65000 );
613 __cfactx_start(main, dst, this->runner, __cfactx_invoke_coroutine);
614
615 verify( ! kernelTLS.preemption_state.enabled );
616
617 dst->last = &src->self_cor;
618 dst->starter = dst->starter ? dst->starter : &src->self_cor;
619
620 // make sure the current state is still correct
621 /* paranoid */ verify(src->state == Ready);
622
623 // context switch to specified coroutine
624 verify( dst->context.SP );
625 __cfactx_switch( &src->context, &dst->context );
626 // when __cfactx_switch returns we are back in the src coroutine
627
628 mainThread->curr_cor = &mainThread->self_cor;
629
630 // make sure the current state has been update
631 /* paranoid */ verify(src->state == Active);
632
633 verify( ! kernelTLS.preemption_state.enabled );
634}
635
636// KERNEL_ONLY
637static void __kernel_last_resume( processor * this ) {
638 $coroutine * src = &mainThread->self_cor;
639 $coroutine * dst = get_coroutine(this->runner);
640
641 verify( ! kernelTLS.preemption_state.enabled );
642 verify( dst->starter == src );
643 verify( dst->context.SP );
644
645 // SKULLDUGGERY in debug the processors check that the
646 // stack is still within the limit of the stack limits after running a thread.
647 // that check doesn't make sense if we context switch to the processor using the
648 // coroutine semantics. Since this is a special case, use the current context
649 // info to populate these fields.
650 __cfaabi_dbg_debug_do(
651 __stack_context_t ctx;
652 CtxGet( ctx );
653 mainThread->context.SP = ctx.SP;
654 mainThread->context.FP = ctx.FP;
655 )
656
657 // context switch to the processor
658 __cfactx_switch( &src->context, &dst->context );
659}
660
661//-----------------------------------------------------------------------------
662// Scheduler routines
663// KERNEL ONLY
664void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
665 /* paranoid */ verify( thrd );
666 /* paranoid */ verify( thrd->state != Halted );
667 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
668 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
669 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
670 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
671 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
672 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
673 /* paranoid */ #endif
674 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
675
676 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
677
678 ready_schedule_lock ( id );
679 push( thrd->curr_cluster, thrd );
680
681 #if !defined(__CFA_NO_STATISTICS__)
682 bool woke =
683 #endif
684 __wake_one(id, thrd->curr_cluster);
685
686 #if !defined(__CFA_NO_STATISTICS__)
687 if(woke) __tls_stats()->ready.sleep.wakes++;
688 #endif
689 ready_schedule_unlock( id );
690
691 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
692}
693
694// KERNEL ONLY
695static $thread * __next_thread(cluster * this) with( *this ) {
696 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
697
698 ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
699 $thread * head = pop( this );
700 ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
701
702 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
703 return head;
704}
705
706// KERNEL ONLY
707static bool __has_next_thread(cluster * this) with( *this ) {
708 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
709
710 ready_schedule_lock ( (__processor_id_t*)kernelTLS.this_processor );
711 bool not_empty = query( this );
712 ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
713
714 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
715 return not_empty;
716}
717
718// KERNEL ONLY unpark with out disabling interrupts
719void __unpark( struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
720 // record activity
721 __cfaabi_dbg_debug_do( char * old_caller = thrd->unpark_caller; )
722 __cfaabi_dbg_record_thrd( *thrd, false, caller );
723
724 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
725 __cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
726 switch(old_ticket) {
727 case 1:
728 // Wake won the race, the thread will reschedule/rerun itself
729 break;
730 case 0:
731 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
732 /* paranoid */ verify( thrd->state == Blocked );
733
734 // Wake lost the race,
735 __schedule_thread( id, thrd );
736 break;
737 default:
738 // This makes no sense, something is wrong abort
739 abort();
740 }
741}
742
743void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
744 if( !thrd ) return;
745
746 disable_interrupts();
747 __unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
748 enable_interrupts( __cfaabi_dbg_ctx );
749}
750
751void park( __cfaabi_dbg_ctx_param ) {
752 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
753 disable_interrupts();
754 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
755 /* paranoid */ verify( kernelTLS.this_thread->preempted == __NO_PREEMPTION );
756
757 // record activity
758 __cfaabi_dbg_record_thrd( *kernelTLS.this_thread, true, caller );
759
760 returnToKernel();
761
762 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
763 enable_interrupts( __cfaabi_dbg_ctx );
764 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
765
766}
767
768// KERNEL ONLY
769void __leave_thread() {
770 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
771 returnToKernel();
772 abort();
773}
774
775// KERNEL ONLY
776bool force_yield( __Preemption_Reason reason ) {
777 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
778 disable_interrupts();
779 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
780
781 $thread * thrd = kernelTLS.this_thread;
782 /* paranoid */ verify(thrd->state == Active);
783
784 // SKULLDUGGERY: It is possible that we are preempting this thread just before
785 // it was going to park itself. If that is the case and it is already using the
786 // intrusive fields then we can't use them to preempt the thread
787 // If that is the case, abandon the preemption.
788 bool preempted = false;
789 if(thrd->link.next == 0p) {
790 preempted = true;
791 thrd->preempted = reason;
792 returnToKernel();
793 }
794
795 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
796 enable_interrupts_noPoll();
797 /* paranoid */ verify( kernelTLS.preemption_state.enabled );
798
799 return preempted;
800}
801
802//=============================================================================================
803// Kernel Setup logic
804//=============================================================================================
805//-----------------------------------------------------------------------------
806// Kernel boot procedures
807static void __kernel_startup(void) {
808 verify( ! kernelTLS.preemption_state.enabled );
809 __cfadbg_print_safe(runtime_core, "Kernel : Starting\n");
810
811 __page_size = sysconf( _SC_PAGESIZE );
812
813 __cfa_dbg_global_clusters.list{ __get };
814 __cfa_dbg_global_clusters.lock{};
815
816 // Initialize the global scheduler lock
817 __scheduler_lock = (__scheduler_RWLock_t*)&storage___scheduler_lock;
818 (*__scheduler_lock){};
819
820 // Initialize the main cluster
821 mainCluster = (cluster *)&storage_mainCluster;
822 (*mainCluster){"Main Cluster"};
823
824 __cfadbg_print_safe(runtime_core, "Kernel : Main cluster ready\n");
825
826 // Start by initializing the main thread
827 // SKULLDUGGERY: the mainThread steals the process main thread
828 // which will then be scheduled by the mainProcessor normally
829 mainThread = ($thread *)&storage_mainThread;
830 current_stack_info_t info;
831 info.storage = (__stack_t*)&storage_mainThreadCtx;
832 (*mainThread){ &info };
833
834 __cfadbg_print_safe(runtime_core, "Kernel : Main thread ready\n");
835
836
837
838 // Construct the processor context of the main processor
839 void ?{}(processorCtx_t & this, processor * proc) {
840 (this.__cor){ "Processor" };
841 this.__cor.starter = 0p;
842 this.proc = proc;
843 }
844
845 void ?{}(processor & this) with( this ) {
846 name = "Main Processor";
847 cltr = mainCluster;
848 terminated{ 0 };
849 do_terminate = false;
850 preemption_alarm = 0p;
851 pending_preemption = false;
852 kernel_thread = pthread_self();
853 id = -1u;
854
855 runner{ &this };
856 __cfadbg_print_safe(runtime_core, "Kernel : constructed main processor context %p\n", &runner);
857
858 __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
859 }
860
861 // Initialize the main processor and the main processor ctx
862 // (the coroutine that contains the processing control flow)
863 mainProcessor = (processor *)&storage_mainProcessor;
864 (*mainProcessor){};
865
866 mainProcessor->id = doregister( (__processor_id_t*)mainProcessor);
867
868 //initialize the global state variables
869 kernelTLS.this_processor = mainProcessor;
870 kernelTLS.this_thread = mainThread;
871
872 #if !defined( __CFA_NO_STATISTICS__ )
873 kernelTLS.this_stats = (__stats_t *)& storage_mainProcStats;
874 __init_stats( kernelTLS.this_stats );
875 #endif
876
877 // Enable preemption
878 kernel_start_preemption();
879
880 // Add the main thread to the ready queue
881 // once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
882 __schedule_thread((__processor_id_t *)mainProcessor, mainThread);
883
884 // SKULLDUGGERY: Force a context switch to the main processor to set the main thread's context to the current UNIX
885 // context. Hence, the main thread does not begin through __cfactx_invoke_thread, like all other threads. The trick here is that
886 // mainThread is on the ready queue when this call is made.
887 __kernel_first_resume( kernelTLS.this_processor );
888
889
890 // THE SYSTEM IS NOW COMPLETELY RUNNING
891
892
893 // Now that the system is up, finish creating systems that need threading
894 __kernel_io_finish_start( *mainCluster );
895
896
897 __cfadbg_print_safe(runtime_core, "Kernel : Started\n--------------------------------------------------\n\n");
898
899 verify( ! kernelTLS.preemption_state.enabled );
900 enable_interrupts( __cfaabi_dbg_ctx );
901 verify( TL_GET( preemption_state.enabled ) );
902}
903
904static void __kernel_shutdown(void) {
905 //Before we start shutting things down, wait for systems that need threading to shutdown
906 __kernel_io_prepare_stop( *mainCluster );
907
908 /* paranoid */ verify( TL_GET( preemption_state.enabled ) );
909 disable_interrupts();
910 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
911
912 __cfadbg_print_safe(runtime_core, "\n--------------------------------------------------\nKernel : Shutting down\n");
913
914 // SKULLDUGGERY: Notify the mainProcessor it needs to terminates.
915 // When its coroutine terminates, it return control to the mainThread
916 // which is currently here
917 __atomic_store_n(&mainProcessor->do_terminate, true, __ATOMIC_RELEASE);
918 __kernel_last_resume( kernelTLS.this_processor );
919 mainThread->self_cor.state = Halted;
920
921 // THE SYSTEM IS NOW COMPLETELY STOPPED
922
923 // Disable preemption
924 kernel_stop_preemption();
925
926 unregister((__processor_id_t*)mainProcessor);
927
928 // Destroy the main processor and its context in reverse order of construction
929 // These were manually constructed so we need manually destroy them
930 void ^?{}(processor & this) with( this ){
931 /* paranoid */ verify( this.do_terminate == true );
932 __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
933 __cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
934 }
935
936 ^(*mainProcessor){};
937
938 // Final step, destroy the main thread since it is no longer needed
939
940 // Since we provided a stack to this taxk it will not destroy anything
941 /* paranoid */ verify(mainThread->self_cor.stack.storage == (__stack_t*)(((uintptr_t)&storage_mainThreadCtx)| 0x1));
942 ^(*mainThread){};
943
944 ^(*mainCluster){};
945
946 ^(*__scheduler_lock){};
947
948 ^(__cfa_dbg_global_clusters.list){};
949 ^(__cfa_dbg_global_clusters.lock){};
950
951 __cfadbg_print_safe(runtime_core, "Kernel : Shutdown complete\n");
952}
953
954//=============================================================================================
955// Kernel Idle Sleep
956//=============================================================================================
957// Wake a thread from the front if there are any
958static bool __wake_one(struct __processor_id_t * id, cluster * this) {
959 /* paranoid */ verify( ready_schedule_islocked( id ) );
960
961 // Check if there is a sleeping processor
962 processor * p = pop(this->idles);
963
964 // If no one is sleeping, we are done
965 if( 0p == p ) return false;
966
967 // We found a processor, wake it up
968 post( p->idle );
969
970 return true;
971}
972
973// Unconditionnaly wake a thread
974static bool __wake_proc(processor * this) {
975 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
976
977 disable_interrupts();
978 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
979 bool ret = post( this->idle );
980 enable_interrupts( __cfaabi_dbg_ctx );
981
982 return ret;
983}
984
985static void __halt(processor * this) with( *this ) {
986 if( do_terminate ) return;
987
988 #if !defined(__CFA_NO_STATISTICS__)
989 __tls_stats()->ready.sleep.halts++;
990 #endif
991 // Push self to queue
992 push(cltr->idles, *this);
993
994 // Makre sure we don't miss a thread
995 if( __has_next_thread(cltr) ) {
996 // A thread was posted, make sure a processor is woken up
997 struct __processor_id_t *id = (struct __processor_id_t *) this;
998 ready_schedule_lock ( id );
999 __wake_one( id, cltr );
1000 ready_schedule_unlock( id );
1001 #if !defined(__CFA_NO_STATISTICS__)
1002 __tls_stats()->ready.sleep.cancels++;
1003 #endif
1004 }
1005
1006 wait( idle );
1007}
1008
1009//=============================================================================================
1010// Unexpected Terminating logic
1011//=============================================================================================
1012static __spinlock_t kernel_abort_lock;
1013static bool kernel_abort_called = false;
1014
1015void * kernel_abort(void) __attribute__ ((__nothrow__)) {
1016 // abort cannot be recursively entered by the same or different processors because all signal handlers return when
1017 // the globalAbort flag is true.
1018 lock( kernel_abort_lock __cfaabi_dbg_ctx2 );
1019
1020 // first task to abort ?
1021 if ( kernel_abort_called ) { // not first task to abort ?
1022 unlock( kernel_abort_lock );
1023
1024 sigset_t mask;
1025 sigemptyset( &mask );
1026 sigaddset( &mask, SIGALRM ); // block SIGALRM signals
1027 sigaddset( &mask, SIGUSR1 ); // block SIGALRM signals
1028 sigsuspend( &mask ); // block the processor to prevent further damage during abort
1029 _exit( EXIT_FAILURE ); // if processor unblocks before it is killed, terminate it
1030 }
1031 else {
1032 kernel_abort_called = true;
1033 unlock( kernel_abort_lock );
1034 }
1035
1036 return kernelTLS.this_thread;
1037}
1038
1039void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
1040 $thread * thrd = kernel_data;
1041
1042 if(thrd) {
1043 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
1044 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1045
1046 if ( &thrd->self_cor != thrd->curr_cor ) {
1047 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
1048 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1049 }
1050 else {
1051 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
1052 }
1053 }
1054 else {
1055 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
1056 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
1057 }
1058}
1059
1060int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
1061 return get_coroutine(kernelTLS.this_thread) == get_coroutine(mainThread) ? 4 : 2;
1062}
1063
1064static __spinlock_t kernel_debug_lock;
1065
1066extern "C" {
1067 void __cfaabi_bits_acquire() {
1068 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
1069 }
1070
1071 void __cfaabi_bits_release() {
1072 unlock( kernel_debug_lock );
1073 }
1074}
1075
1076//=============================================================================================
1077// Kernel Utilities
1078//=============================================================================================
1079//-----------------------------------------------------------------------------
1080// Locks
1081void ?{}( semaphore & this, int count = 1 ) {
1082 (this.lock){};
1083 this.count = count;
1084 (this.waiting){};
1085}
1086void ^?{}(semaphore & this) {}
1087
1088bool P(semaphore & this) with( this ){
1089 lock( lock __cfaabi_dbg_ctx2 );
1090 count -= 1;
1091 if ( count < 0 ) {
1092 // queue current task
1093 append( waiting, kernelTLS.this_thread );
1094
1095 // atomically release spin lock and block
1096 unlock( lock );
1097 park( __cfaabi_dbg_ctx );
1098 return true;
1099 }
1100 else {
1101 unlock( lock );
1102 return false;
1103 }
1104}
1105
1106bool V(semaphore & this) with( this ) {
1107 $thread * thrd = 0p;
1108 lock( lock __cfaabi_dbg_ctx2 );
1109 count += 1;
1110 if ( count <= 0 ) {
1111 // remove task at head of waiting list
1112 thrd = pop_head( waiting );
1113 }
1114
1115 unlock( lock );
1116
1117 // make new owner
1118 unpark( thrd __cfaabi_dbg_ctx2 );
1119
1120 return thrd != 0p;
1121}
1122
1123bool V(semaphore & this, unsigned diff) with( this ) {
1124 $thread * thrd = 0p;
1125 lock( lock __cfaabi_dbg_ctx2 );
1126 int release = max(-count, (int)diff);
1127 count += diff;
1128 for(release) {
1129 unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 );
1130 }
1131
1132 unlock( lock );
1133
1134 return thrd != 0p;
1135}
1136
1137//-----------------------------------------------------------------------------
1138// Global Queues
1139void doregister( cluster & cltr ) {
1140 lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1141 push_front( __cfa_dbg_global_clusters.list, cltr );
1142 unlock ( __cfa_dbg_global_clusters.lock );
1143}
1144
1145void unregister( cluster & cltr ) {
1146 lock ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
1147 remove( __cfa_dbg_global_clusters.list, cltr );
1148 unlock( __cfa_dbg_global_clusters.lock );
1149}
1150
1151void doregister( cluster * cltr, $thread & thrd ) {
1152 lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1153 cltr->nthreads += 1;
1154 push_front(cltr->threads, thrd);
1155 unlock (cltr->thread_list_lock);
1156}
1157
1158void unregister( cluster * cltr, $thread & thrd ) {
1159 lock (cltr->thread_list_lock __cfaabi_dbg_ctx2);
1160 remove(cltr->threads, thrd );
1161 cltr->nthreads -= 1;
1162 unlock(cltr->thread_list_lock);
1163}
1164
1165//-----------------------------------------------------------------------------
1166// Debug
1167__cfaabi_dbg_debug_do(
1168 extern "C" {
1169 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
1170 this.prev_name = prev_name;
1171 this.prev_thrd = kernelTLS.this_thread;
1172 }
1173
1174 void __cfaabi_dbg_record_thrd($thread & this, bool park, const char prev_name[]) {
1175 if(park) {
1176 this.park_caller = prev_name;
1177 this.park_stale = false;
1178 }
1179 else {
1180 this.unpark_caller = prev_name;
1181 this.unpark_stale = false;
1182 }
1183 }
1184 }
1185)
1186
1187//-----------------------------------------------------------------------------
1188// Debug
1189bool threading_enabled(void) __attribute__((const)) {
1190 return true;
1191}
1192// Local Variables: //
1193// mode: c //
1194// tab-width: 4 //
1195// End: //
Note: See TracBrowser for help on using the repository browser.