source: libcfa/src/concurrency/kernel.cfa@ 52a5262e

Last change on this file since 52a5262e was b93bf85, checked in by caparsons <caparson@…>, 2 years ago

fixed spurious channel close waituntil error case. Was caused by a race condition causing an exception to be thrown while another was in flight
  • Property mode set to 100644
File size: 28.6 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// kernel.c --
8//
9// Author : Thierry Delisle
10// Created On : Tue Jan 17 12:27:26 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : Mon Jan 9 08:42:05 2023
13// Update Count : 77
14//
15
16#define __cforall_thread__
17
18// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
19
20#pragma GCC diagnostic push
21#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
22
23//C Includes
24#include <errno.h>
25#include <stdio.h>
26#include <string.h>
27#include <signal.h>
28#include <unistd.h>
29
30extern "C" {
31 #include <sys/eventfd.h>
32 #include <sys/uio.h>
33}
34
35//CFA Includes
36#include "kernel/private.hfa"
37#include "preemption.hfa"
38#include "strstream.hfa"
39#include "device/cpu.hfa"
40#include "io/types.hfa"
41
42//Private includes
43#define __CFA_INVOKE_PRIVATE__
44#include "invoke.h"
45#pragma GCC diagnostic pop
46
47#if !defined(__CFA_NO_STATISTICS__)
48 #define __STATS_DEF( ...) __VA_ARGS__
49#else
50 #define __STATS_DEF( ...)
51#endif
52
53//-----------------------------------------------------------------------------
54// Some assembly required
55#if defined( __i386 )
56 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
57 // fcw : X87 FPU control word (preserved across function calls)
58 #define __x87_store \
59 uint32_t __mxcr; \
60 uint16_t __fcw; \
61 __asm__ volatile ( \
62 "stmxcsr %0\n" \
63 "fnstcw %1\n" \
64 : "=m" (__mxcr),\
65 "=m" (__fcw) \
66 )
67
68 #define __x87_load \
69 __asm__ volatile ( \
70 "fldcw %1\n" \
71 "ldmxcsr %0\n" \
72 ::"m" (__mxcr),\
73 "m" (__fcw) \
74 )
75
76#elif defined( __x86_64 )
77 #define __x87_store \
78 uint32_t __mxcr; \
79 uint16_t __fcw; \
80 __asm__ volatile ( \
81 "stmxcsr %0\n" \
82 "fnstcw %1\n" \
83 : "=m" (__mxcr),\
84 "=m" (__fcw) \
85 )
86
87 #define __x87_load \
88 __asm__ volatile ( \
89 "fldcw %1\n" \
90 "ldmxcsr %0\n" \
91 :: "m" (__mxcr),\
92 "m" (__fcw) \
93 )
94
95#elif defined( __arm__ )
96 #define __x87_store
97 #define __x87_load
98
99#elif defined( __aarch64__ )
100 #define __x87_store \
101 uint32_t __fpcntl[2]; \
102 __asm__ volatile ( \
103 "mrs x9, FPCR\n" \
104 "mrs x10, FPSR\n" \
105 "stp x9, x10, %0\n" \
106 : "=m" (__fpcntl) : : "x9", "x10" \
107 )
108
109 #define __x87_load \
110 __asm__ volatile ( \
111 "ldp x9, x10, %0\n" \
112 "msr FPSR, x10\n" \
113 "msr FPCR, x9\n" \
114 : "=m" (__fpcntl) : : "x9", "x10" \
115 )
116
117#else
118 #error unsupported hardware architecture
119#endif
120
121extern thread$ * mainThread;
122extern processor * mainProcessor;
123
124//-----------------------------------------------------------------------------
125// Kernel Scheduling logic
126static thread$ * __next_thread(cluster * this);
127static thread$ * __next_thread_slow(cluster * this);
128static thread$ * __next_thread_search(cluster * this);
129static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
130static void __run_thread(processor * this, thread$ * dst);
131static void __wake_one(cluster * cltr);
132
133static void idle_sleep(processor * proc);
134static bool mark_idle (__cluster_proc_list & idles, processor & proc);
135static void mark_awake(__cluster_proc_list & idles, processor & proc);
136
137extern bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1)));
138extern bool __cfa_io_flush( processor * ) __attribute__((nonnull (1)));
139
140
141extern void __disable_interrupts_hard();
142extern void __enable_interrupts_hard();
143
144
145//=============================================================================================
146// Kernel Scheduling logic
147//=============================================================================================
148//Main of the processor contexts
149void main(processorCtx_t & runner) {
150 // Because of a bug, we couldn't initialized the seed on construction
151 // Do it here
152 PRNG_SET_SEED( __cfaabi_tls.random_state, rdtscl() );
153 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
154 __tls_rand_advance_bck();
155
156 processor * this = runner.proc;
157 verify(this);
158
159 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
160 #if !defined(__CFA_NO_STATISTICS__)
161 if( this->print_halts ) {
162 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
163 }
164 #endif
165
166 {
167 // Setup preemption data
168 preemption_scope scope = { this };
169
170 // if we need to run some special setup, now is the time to do it.
171 if(this->init.thrd) {
172 this->init.thrd->curr_cluster = this->cltr;
173 __run_thread(this, this->init.thrd);
174 }
175
176 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
177
178 thread$ * readyThread = 0p;
179 MAIN_LOOP:
180 for() {
181 // Check if there is pending io
182 __cfa_io_drain( this );
183
184 // Try to get the next thread
185 readyThread = __next_thread( this->cltr );
186
187 if( !readyThread ) {
188 // there is no point in holding submissions if we are idle
189 __IO_STATS__(true, io.flush.idle++; )
190 __cfa_io_flush( this );
191
192 // drain again in case something showed up
193 __cfa_io_drain( this );
194
195 readyThread = __next_thread( this->cltr );
196 }
197
198 if( !readyThread ) for(5) {
199 readyThread = __next_thread_slow( this->cltr );
200
201 if( readyThread ) break;
202
203 // It's unlikely we still I/O to submit, but the arbiter could
204 __IO_STATS__(true, io.flush.idle++; )
205 __cfa_io_flush( this );
206
207 // drain again in case something showed up
208 __cfa_io_drain( this );
209 }
210
211 HALT:
212 if( !readyThread ) {
213 // Don't block if we are done
214 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
215
216 // Push self to idle stack
217 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
218
219 // Confirm the ready-queue is empty
220 readyThread = __next_thread_search( this->cltr );
221 if( readyThread ) {
222 // A thread was found, cancel the halt
223 mark_awake(this->cltr->procs, * this);
224
225 __STATS__(true, ready.sleep.cancels++; )
226
227 // continue the mai loop
228 break HALT;
229 }
230
231 idle_sleep( this );
232
233 // We were woken up, remove self from idle
234 mark_awake(this->cltr->procs, * this);
235
236 // DON'T just proceed, start looking again
237 continue MAIN_LOOP;
238 }
239
240 /* paranoid */ verify( readyThread );
241
242 // Reset io dirty bit
243 this->io.dirty = false;
244
245 // We found a thread run it
246 __run_thread(this, readyThread);
247
248 // Are we done?
249 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
250
251 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
252 __IO_STATS__(true, io.flush.dirty++; )
253 __cfa_io_flush( this );
254 }
255 }
256
257 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
258 }
259
260 __cfa_io_flush( this );
261 __cfa_io_drain( this );
262
263 post( this->terminated );
264
265 if(this == mainProcessor) {
266 // HACK : the coroutine context switch expects this_thread to be set
267 // and it make sense for it to be set in all other cases except here
268 // fake it
269 __cfaabi_tls.this_thread = mainThread;
270 }
271
272 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
273}
274
275static int * __volatile_errno() __attribute__((noinline));
276static int * __volatile_errno() { asm(""); return &errno; }
277
278// KERNEL ONLY
279// runThread runs a thread by context switching
280// from the processor coroutine to the target thread
281static void __run_thread(processor * this, thread$ * thrd_dst) {
282 /* paranoid */ verify( ! __preemption_enabled() );
283 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
284 /* paranoid */ verifyf( thrd_dst->rdy_link.next == 0p, "Expected null got %p", thrd_dst->rdy_link.next );
285 __builtin_prefetch( thrd_dst->context.SP );
286
287 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
288
289 coroutine$ * proc_cor = get_coroutine(this->runner);
290
291 // set state of processor coroutine to inactive
292 verify(proc_cor->state == Active);
293 proc_cor->state = Blocked;
294
295 // Actually run the thread
296 RUNNING: while(true) {
297 thrd_dst->preempted = __NO_PREEMPTION;
298
299 // Update global state
300 kernelTLS().this_thread = thrd_dst;
301
302 // Update the state after setting this_thread
303 // so that the debugger can find all active threads
304 // in tls storage
305 thrd_dst->state = Active;
306
307 /* paranoid */ verify( ! __preemption_enabled() );
308 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
309 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
310 /* paranoid */ verify( thrd_dst->context.SP );
311 /* paranoid */ verify( thrd_dst->state != Halted );
312 /* 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
313 /* 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
314 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
315 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
316
317
318
319 // set context switch to the thread that the processor is executing
320 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
321 // when __cfactx_switch returns we are back in the processor coroutine
322
323
324
325 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
326 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
327 /* 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 );
328 /* 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 );
329 /* paranoid */ verify( thrd_dst->state != Halted );
330 /* paranoid */ verify( thrd_dst->context.SP );
331 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
332 /* paranoid */ verify( ! __preemption_enabled() );
333
334 // We just finished running a thread, there are a few things that could have happened.
335 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
336 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
337 // 4 - Preempted
338 // In case 1, we may have won a race so we can't write to the state again.
339 // In case 2, we lost the race so we now own the thread.
340
341 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
342 // Reset the this_thread now that we know
343 // the state isn't active anymore
344 kernelTLS().this_thread = 0p;
345
346 // The thread was preempted, reschedule it and reset the flag
347 schedule_thread$( thrd_dst, UNPARK_LOCAL );
348 break RUNNING;
349 }
350
351 if(unlikely(thrd_dst->state == Halting)) {
352 // Reset the this_thread now that we know
353 // the state isn't active anymore
354 kernelTLS().this_thread = 0p;
355
356 // The thread has halted, it should never be scheduled/run again
357 // finish the thread
358 __thread_finish( thrd_dst );
359 break RUNNING;
360 }
361
362 /* paranoid */ verify( thrd_dst->state == Active );
363 thrd_dst->state = Blocked;
364
365 // Reset the this_thread now that we know
366 // the state isn't active anymore
367 kernelTLS().this_thread = 0p;
368
369 // set state of processor coroutine to active and the thread to inactive
370 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
371 switch(old_ticket) {
372 case TICKET_RUNNING:
373 // This is case 1, the regular case, nothing more is needed
374 break RUNNING;
375 case TICKET_UNBLOCK:
376 __STATS__(true, ready.threads.threads++; )
377 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
378 // In this case, just run it again.
379 continue RUNNING;
380 default:
381 // This makes no sense, something is wrong abort
382 abort();
383 }
384 }
385
386 // Just before returning to the processor, set the processor coroutine to active
387 proc_cor->state = Active;
388
389 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
390
391 __STATS__(true, ready.threads.threads--; )
392
393 /* paranoid */ verify( ! __preemption_enabled() );
394}
395
396// KERNEL_ONLY
397static void returnToKernel() {
398 /* paranoid */ verify( ! __preemption_enabled() );
399 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
400 thread$ * thrd_src = kernelTLS().this_thread;
401
402 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
403
404 // Run the thread on this processor
405 {
406 int local_errno = *__volatile_errno();
407 #if defined( __i386 ) || defined( __x86_64 )
408 __x87_store;
409 #endif
410 /* paranoid */ verify( proc_cor->context.SP );
411 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
412 __cfactx_switch( &thrd_src->context, &proc_cor->context );
413 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
414 #if defined( __i386 ) || defined( __x86_64 )
415 __x87_load;
416 #endif
417 *__volatile_errno() = local_errno;
418 }
419
420 #if !defined(__CFA_NO_STATISTICS__)
421 /* paranoid */ verify( thrd_src->last_proc != 0p );
422 if(thrd_src->last_proc != kernelTLS().this_processor) {
423 __tls_stats()->ready.threads.migration++;
424 }
425 #endif
426
427 /* paranoid */ verify( ! __preemption_enabled() );
428 /* 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 );
429 /* 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 );
430}
431
432//-----------------------------------------------------------------------------
433// Scheduler routines
434// KERNEL ONLY
435static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
436 /* paranoid */ verify( ! __preemption_enabled() );
437 /* paranoid */ verify( ready_schedule_islocked());
438 /* paranoid */ verify( thrd );
439 /* paranoid */ verify( thrd->state != Halted );
440 /* paranoid */ verify( thrd->curr_cluster );
441 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
442 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
443 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
444 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
445 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
446 /* paranoid */ #endif
447 /* paranoid */ verifyf( thrd->rdy_link.next == 0p, "Expected null got %p", thrd->rdy_link.next );
448 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
449
450 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
451
452 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
453 struct cluster * cl = thrd->curr_cluster;
454 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
455
456 // push the thread to the cluster ready-queue
457 push( cl, thrd, hint );
458
459 // variable thrd is no longer safe to use
460 thrd = 0xdeaddeaddeaddeadp;
461
462 // wake the cluster using the save variable.
463 __wake_one( cl );
464
465 #if !defined(__CFA_NO_STATISTICS__)
466 if( kernelTLS().this_stats ) {
467 __tls_stats()->ready.threads.threads++;
468 if(outside) {
469 __tls_stats()->ready.threads.extunpark++;
470 }
471 }
472 else {
473 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
474 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
475 }
476 #endif
477
478 /* paranoid */ verify( ready_schedule_islocked());
479 /* paranoid */ verify( ! __preemption_enabled() );
480}
481
482void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
483 ready_schedule_lock();
484 __schedule_thread( thrd, hint );
485 ready_schedule_unlock();
486}
487
488// KERNEL ONLY
489static inline thread$ * __next_thread(cluster * this) with( *this ) {
490 /* paranoid */ verify( ! __preemption_enabled() );
491
492 ready_schedule_lock();
493 thread$ * thrd = pop_fast( this );
494 ready_schedule_unlock();
495
496 /* paranoid */ verify( ! __preemption_enabled() );
497 return thrd;
498}
499
500// KERNEL ONLY
501static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
502 /* paranoid */ verify( ! __preemption_enabled() );
503
504 ready_schedule_lock();
505 thread$ * thrd = pop_slow( this );
506 ready_schedule_unlock();
507
508 /* paranoid */ verify( ! __preemption_enabled() );
509 return thrd;
510}
511
512// KERNEL ONLY
513static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
514 /* paranoid */ verify( ! __preemption_enabled() );
515
516 ready_schedule_lock();
517 thread$ * thrd = pop_search( this );
518 ready_schedule_unlock();
519
520 /* paranoid */ verify( ! __preemption_enabled() );
521 return thrd;
522}
523
524static inline bool __must_unpark( thread$ * thrd ) {
525 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
526 switch(old_ticket) {
527 case TICKET_RUNNING:
528 // Wake won the race, the thread will reschedule/rerun itself
529 return false;
530 case TICKET_BLOCKED:
531 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
532 /* paranoid */ verify( thrd->state == Blocked );
533 return true;
534 default:
535 // This makes no sense, something is wrong abort
536 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
537 }
538}
539
540void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
541 /* paranoid */ verify( ! __preemption_enabled() );
542 /* paranoid */ verify( ready_schedule_islocked());
543
544 if( !thrd ) return;
545
546 if(__must_unpark(thrd)) {
547 // Wake lost the race,
548 __schedule_thread( thrd, hint );
549 }
550
551 /* paranoid */ verify( ready_schedule_islocked());
552 /* paranoid */ verify( ! __preemption_enabled() );
553}
554
555void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
556 if( !thrd ) return;
557
558 if(__must_unpark(thrd)) {
559 disable_interrupts();
560 // Wake lost the race,
561 schedule_thread$( thrd, hint );
562 enable_interrupts(false);
563 }
564}
565
566void park( void ) libcfa_public {
567 __disable_interrupts_checked();
568 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
569 returnToKernel();
570 __enable_interrupts_checked();
571}
572
573extern "C" {
574 // Leave the thread monitor
575 // last routine called by a thread.
576 // Should never return
577 void __cfactx_thrd_leave() {
578 thread$ * thrd = active_thread();
579 monitor$ * this = &thrd->self_mon;
580
581 // Lock the monitor now
582 lock( this->lock __cfaabi_dbg_ctx2 );
583
584 disable_interrupts();
585
586 /* paranoid */ verify( ! __preemption_enabled() );
587 /* paranoid */ verify( thrd->state == Active );
588 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
589 /* paranoid */ verify( kernelTLS().this_thread == thrd );
590 /* paranoid */ verify( thrd->context.SP );
591 /* 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 );
592 /* 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 );
593
594 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
595 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
596 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
597
598 thrd->state = Halting;
599 thrd->ticket = TICKET_DEAD;
600
601 // Leave the thread
602 returnToKernel();
603
604 // Control flow should never reach here!
605 abort();
606 }
607}
608
609// KERNEL ONLY
610bool force_yield( __Preemption_Reason reason ) libcfa_public {
611 __disable_interrupts_checked();
612 thread$ * thrd = kernelTLS().this_thread;
613 /* paranoid */ verify(thrd->state == Active);
614
615 // SKULLDUGGERY: It is possible that we are preempting this thread just before
616 // it was going to park itself. If that is the case and it is already using the
617 // intrusive fields then we can't use them to preempt the thread
618 // If that is the case, abandon the preemption.
619 bool preempted = false;
620 if(thrd->rdy_link.next == 0p) {
621 preempted = true;
622 thrd->preempted = reason;
623 returnToKernel();
624 }
625 __enable_interrupts_checked( false );
626 return preempted;
627}
628
629//=============================================================================================
630// Kernel Idle Sleep
631//=============================================================================================
632// Wake a thread from the front if there are any
633static void __wake_one(cluster * this) {
634 eventfd_t val;
635
636 /* paranoid */ verify( ! __preemption_enabled() );
637 /* paranoid */ verify( ready_schedule_islocked() );
638
639 // Check if there is a sleeping processor
640 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
641
642 // If no one is sleeping: we are done
643 if( fdp == 0p ) return;
644
645 int fd = 1;
646 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
647 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
648 }
649
650 switch(fd) {
651 __attribute__((unused)) int ret;
652 case 0:
653 // If the processor isn't ready to sleep then the exchange will already wake it up
654 #if !defined(__CFA_NO_STATISTICS__)
655 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
656 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
657 #endif
658 break;
659 case 1:
660 // If someone else already said they will wake them: we are done
661 #if !defined(__CFA_NO_STATISTICS__)
662 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
663 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
664 #endif
665 break;
666 default:
667 // If the processor was ready to sleep, we need to wake it up with an actual write
668 val = 1;
669 ret = eventfd_write( fd, val );
670 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
671
672 #if !defined(__CFA_NO_STATISTICS__)
673 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
674 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
675 #endif
676 break;
677 }
678
679 /* paranoid */ verify( ready_schedule_islocked() );
680 /* paranoid */ verify( ! __preemption_enabled() );
681
682 return;
683}
684
685// Unconditionnaly wake a thread
686void __wake_proc(processor * this) {
687 /* paranoid */ verify( ! __preemption_enabled() );
688
689 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
690
691 this->idle_wctx.sem = 1;
692
693 this->idle_wctx.wake__time = rdtscl();
694
695 eventfd_t val;
696 val = 1;
697 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
698
699 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
700 /* paranoid */ verify( ! __preemption_enabled() );
701}
702
703static void idle_sleep(processor * this) {
704 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
705 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
706
707 // Tell everyone we are ready to go do sleep
708 for() {
709 int expected = this->idle_wctx.sem;
710
711 // Someone already told us to wake-up! No time for a nap.
712 if(expected == 1) { return; }
713
714 // Try to mark that we are going to sleep
715 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
716 // Every one agreed, taking a nap
717 break;
718 }
719 }
720
721
722 #if !defined(__CFA_NO_STATISTICS__)
723 if(this->print_halts) {
724 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
725 }
726 #endif
727
728 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
729
730 {
731 eventfd_t val;
732 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
733 if(ret < 0) {
734 switch((int)errno) {
735 case EAGAIN:
736 #if EAGAIN != EWOULDBLOCK
737 case EWOULDBLOCK:
738 #endif
739 case EINTR:
740 // No need to do anything special here, just assume it's a legitimate wake-up
741 break;
742 default:
743 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
744 }
745 }
746 }
747
748 #if !defined(__CFA_NO_STATISTICS__)
749 if(this->print_halts) {
750 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
751 }
752 #endif
753}
754
755static bool mark_idle(__cluster_proc_list & this, processor & proc) {
756 __STATS__(true, ready.sleep.halts++; )
757
758 proc.idle_wctx.sem = 0;
759
760 /* paranoid */ verify( ! __preemption_enabled() );
761 if(!try_lock( this )) return false;
762 this.idle++;
763 /* paranoid */ verify( this.idle <= this.total );
764 remove(proc);
765 insert_first(this.idles, proc);
766
767 // update the pointer to the head wait context, which should now point to this proc.
768 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
769 unlock( this );
770 /* paranoid */ verify( ! __preemption_enabled() );
771
772 return true;
773}
774
775static void mark_awake(__cluster_proc_list & this, processor & proc) {
776 /* paranoid */ verify( ! __preemption_enabled() );
777 lock( this );
778 this.idle--;
779 /* paranoid */ verify( this.idle >= 0 );
780 remove(proc);
781 insert_last(this.actives, proc);
782
783 {
784 // update the pointer to the head wait context
785 struct __fd_waitctx * wctx = 0;
786 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
787 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
788 }
789
790 unlock( this );
791 /* paranoid */ verify( ! __preemption_enabled() );
792}
793
794//=============================================================================================
795// Unexpected Terminating logic
796//=============================================================================================
797void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
798 thread$ * thrd = __cfaabi_tls.this_thread;
799
800 if(thrd) {
801 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
802 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
803
804 if ( &thrd->self_cor != thrd->curr_cor ) {
805 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
806 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
807 }
808 else {
809 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
810 }
811 }
812 else {
813 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
814 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
815 }
816}
817
818int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
819 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
820}
821
822static __spinlock_t kernel_debug_lock;
823
824extern "C" {
825 void __cfaabi_bits_acquire() {
826 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
827 }
828
829 void __cfaabi_bits_release() {
830 unlock( kernel_debug_lock );
831 }
832}
833
834//=============================================================================================
835// Kernel Utilities
836//=============================================================================================
837#if defined(CFA_HAVE_LINUX_IO_URING_H)
838#include "io/types.hfa"
839#endif
840
841//-----------------------------------------------------------------------------
842// Debug
843bool threading_enabled(void) __attribute__((const)) libcfa_public {
844 return true;
845}
846
847//-----------------------------------------------------------------------------
848// Statistics
849#if !defined(__CFA_NO_STATISTICS__)
850 void print_halts( processor & this ) libcfa_public {
851 this.print_halts = true;
852 }
853
854 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
855 /* paranoid */ verify( cltr->stats );
856
857 processor * it = &list`first;
858 for(unsigned i = 0; i < count; i++) {
859 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
860 /* paranoid */ verify( it->local_data->this_stats );
861 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
862 __tally_stats( cltr->stats, it->local_data->this_stats );
863 it = &(*it)`next;
864 }
865 }
866
867 static void crawl_cluster_stats( cluster & this ) {
868 // Stop the world, otherwise stats could get really messed-up
869 // this doesn't solve all problems but does solve many
870 // so it's probably good enough
871 disable_interrupts();
872 uint_fast32_t last_size = ready_mutate_lock();
873
874 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
875 crawl_list(&this, this.procs.idles , this.procs.idle );
876
877 // Unlock the RWlock
878 ready_mutate_unlock( last_size );
879 enable_interrupts();
880 }
881
882
883 void print_stats_now( cluster & this, int flags ) libcfa_public {
884 crawl_cluster_stats( this );
885 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
886 }
887#endif
888// Local Variables: //
889// mode: c //
890// tab-width: 4 //
891// End: //
Note: See TracBrowser for help on using the repository browser.