source: libcfa/src/concurrency/kernel.cfa@ b28ce93

Last change on this file since b28ce93 was 6b33e89, checked in by Peter A. Buhr <pabuhr@…>, 5 months ago

change backquote call to regular call
  • Property mode set to 100644
File size: 28.7 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 : Fri Apr 25 07:02:42 2025
13// Update Count : 82
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:
297 while( true ) {
298 thrd_dst->preempted = __NO_PREEMPTION;
299
300 // Update global state
301 kernelTLS().this_thread = thrd_dst;
302
303 // Update the state after setting this_thread
304 // so that the debugger can find all active threads
305 // in tls storage
306 thrd_dst->state = Active;
307
308 /* paranoid */ verify( ! __preemption_enabled() );
309 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
310 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
311 /* paranoid */ verify( thrd_dst->context.SP );
312 /* paranoid */ verify( thrd_dst->state != Halted );
313 /* 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
314 /* 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
315 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
316 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
317
318
319
320 // set context switch to the thread that the processor is executing
321 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
322 // when __cfactx_switch returns we are back in the processor coroutine
323
324
325
326 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
327 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
328 /* 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 );
329 /* 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 );
330 /* paranoid */ verify( thrd_dst->state != Halted );
331 /* paranoid */ verify( thrd_dst->context.SP );
332 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
333 /* paranoid */ verify( ! __preemption_enabled() );
334
335 // We just finished running a thread, there are a few things that could have happened.
336 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
337 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
338 // 4 - Preempted
339 // In case 1, we may have won a race so we can't write to the state again.
340 // In case 2, we lost the race so we now own the thread.
341
342 if (unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
343 // Reset the this_thread now that we know
344 // the state isn't active anymore
345 kernelTLS().this_thread = 0p;
346
347 // The thread was preempted, reschedule it and reset the flag
348 schedule_thread$( thrd_dst, UNPARK_LOCAL );
349 break RUNNING;
350 }
351
352 if (unlikely(thrd_dst->state == Halting)) {
353 // Reset the this_thread now that we know
354 // the state isn't active anymore
355 kernelTLS().this_thread = 0p;
356
357 // The thread has halted, it should never be scheduled/run again
358 // finish the thread
359 __thread_finish( thrd_dst );
360 break RUNNING;
361 }
362
363 /* paranoid */ verify( thrd_dst->state == Active );
364 thrd_dst->state = Blocked;
365
366 // Reset the this_thread now that we know
367 // the state isn't active anymore
368 kernelTLS().this_thread = 0p;
369
370 // set state of processor coroutine to active and the thread to inactive
371 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
372 switch(old_ticket) {
373 case TICKET_RUNNING:
374 // This is case 1, the regular case, nothing more is needed
375 break RUNNING;
376 case TICKET_UNBLOCK:
377 __STATS__(true, ready.threads.threads++; )
378 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
379 // In this case, just run it again.
380 continue RUNNING;
381 default:
382 // This makes no sense, something is wrong abort
383 abort();
384 }
385 }
386
387 // Just before returning to the processor, set the processor coroutine to active
388 proc_cor->state = Active;
389
390 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
391
392 __STATS__(true, ready.threads.threads--; )
393
394 /* paranoid */ verify( ! __preemption_enabled() );
395}
396
397// KERNEL_ONLY
398static void returnToKernel() {
399 /* paranoid */ verify( ! __preemption_enabled() );
400 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
401 thread$ * thrd_src = kernelTLS().this_thread;
402
403 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
404
405 // Run the thread on this processor
406 {
407 int local_errno = *__volatile_errno();
408 #if defined( __i386 ) || defined( __x86_64 )
409 __x87_store;
410 #endif
411 /* paranoid */ verify( proc_cor->context.SP );
412 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
413 __cfactx_switch( &thrd_src->context, &proc_cor->context );
414 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
415 #if defined( __i386 ) || defined( __x86_64 )
416 __x87_load;
417 #endif
418 *__volatile_errno() = local_errno;
419 }
420
421 #if ! defined(__CFA_NO_STATISTICS__)
422 /* paranoid */ verify( thrd_src->last_proc != 0p );
423 if (thrd_src->last_proc != kernelTLS().this_processor) {
424 __tls_stats()->ready.threads.migration++;
425 }
426 #endif
427
428 /* paranoid */ verify( ! __preemption_enabled() );
429 /* 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 );
430 /* 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 );
431}
432
433//-----------------------------------------------------------------------------
434// Scheduler routines
435// KERNEL ONLY
436static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
437 /* paranoid */ verify( ! __preemption_enabled() );
438 /* paranoid */ verify( ready_schedule_islocked());
439 /* paranoid */ verify( thrd );
440 /* paranoid */ verify( thrd->state != Halted );
441 /* paranoid */ verify( thrd->curr_cluster );
442 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
443 /* paranoid */ if ( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
444 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
445 /* paranoid */ if ( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
446 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
447 /* paranoid */ #endif
448 /* paranoid */ verifyf( thrd->rdy_link.next == 0p, "Expected null got %p", thrd->rdy_link.next );
449 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
450
451 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
452
453 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
454 struct cluster * cl = thrd->curr_cluster;
455 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
456
457 // push the thread to the cluster ready-queue
458 push( cl, thrd, hint );
459
460 // variable thrd is no longer safe to use
461 thrd = 0xdeaddeaddeaddeadp;
462
463 // wake the cluster using the save variable.
464 __wake_one( cl );
465
466 #if ! defined(__CFA_NO_STATISTICS__)
467 if ( kernelTLS().this_stats ) {
468 __tls_stats()->ready.threads.threads++;
469 if (outside) {
470 __tls_stats()->ready.threads.extunpark++;
471 }
472 }
473 else {
474 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
475 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
476 }
477 #endif
478
479 /* paranoid */ verify( ready_schedule_islocked());
480 /* paranoid */ verify( ! __preemption_enabled() );
481}
482
483void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
484 ready_schedule_lock();
485 __schedule_thread( thrd, hint );
486 ready_schedule_unlock();
487}
488
489// KERNEL ONLY
490static inline thread$ * __next_thread(cluster * this) with( *this ) {
491 /* paranoid */ verify( ! __preemption_enabled() );
492
493 ready_schedule_lock();
494 thread$ * thrd = pop_fast( this );
495 ready_schedule_unlock();
496
497 /* paranoid */ verify( ! __preemption_enabled() );
498 return thrd;
499}
500
501// KERNEL ONLY
502static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
503 /* paranoid */ verify( ! __preemption_enabled() );
504
505 ready_schedule_lock();
506 thread$ * thrd = pop_slow( this );
507 ready_schedule_unlock();
508
509 /* paranoid */ verify( ! __preemption_enabled() );
510 return thrd;
511}
512
513// KERNEL ONLY
514static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
515 /* paranoid */ verify( ! __preemption_enabled() );
516
517 ready_schedule_lock();
518 thread$ * thrd = pop_search( this );
519 ready_schedule_unlock();
520
521 /* paranoid */ verify( ! __preemption_enabled() );
522 return thrd;
523}
524
525static inline bool __must_unpark( thread$ * thrd ) {
526 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
527 switch(old_ticket) {
528 case TICKET_RUNNING:
529 // Wake won the race, the thread will reschedule/rerun itself
530 return false;
531 case TICKET_BLOCKED:
532 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
533 /* paranoid */ verify( thrd->state == Blocked );
534 return true;
535 default:
536 // This makes no sense, something is wrong abort
537 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
538 }
539}
540
541void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
542 /* paranoid */ verify( ! __preemption_enabled() );
543 /* paranoid */ verify( ready_schedule_islocked());
544
545 if ( ! thrd ) return;
546
547 if (__must_unpark(thrd)) {
548 // Wake lost the race,
549 __schedule_thread( thrd, hint );
550 }
551
552 /* paranoid */ verify( ready_schedule_islocked());
553 /* paranoid */ verify( ! __preemption_enabled() );
554}
555
556void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
557 if ( ! thrd ) return;
558
559 if (__must_unpark(thrd)) {
560 disable_interrupts();
561 // Wake lost the race,
562 schedule_thread$( thrd, hint );
563 enable_interrupts(false);
564 }
565}
566
567void park( void ) libcfa_public {
568 __disable_interrupts_checked();
569 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
570 returnToKernel();
571 __enable_interrupts_checked();
572}
573
574extern "C" {
575 // Leave the thread monitor
576 // last routine called by a thread.
577 // Should never return
578 void __cfactx_thrd_leave() {
579 thread$ * thrd = active_thread();
580 monitor$ * this = &thrd->self_mon;
581
582 // Lock the monitor now
583 lock( this->lock __cfaabi_dbg_ctx2 );
584
585 disable_interrupts();
586
587 /* paranoid */ verify( ! __preemption_enabled() );
588 /* paranoid */ verify( thrd->state == Active );
589 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
590 /* paranoid */ verify( kernelTLS().this_thread == thrd );
591 /* paranoid */ verify( thrd->context.SP );
592 /* 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 );
593 /* 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 );
594
595 if ( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
596 if ( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
597 if ( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
598
599 thrd->state = Halting;
600 thrd->ticket = TICKET_DEAD;
601
602 // Leave the thread
603 returnToKernel();
604
605 // Control flow should never reach here!
606 abort();
607 }
608}
609
610// KERNEL ONLY
611bool force_yield( __Preemption_Reason reason ) libcfa_public {
612 __disable_interrupts_checked();
613 thread$ * thrd = kernelTLS().this_thread;
614 /* paranoid */ verify(thrd->state == Active);
615
616 // SKULLDUGGERY: It is possible that we are preempting this thread just before
617 // it was going to park itself. If that is the case and it is already using the
618 // intrusive fields then we can't use them to preempt the thread
619 // If that is the case, abandon the preemption.
620 bool preempted = false;
621 if (thrd->rdy_link.next == 0p) {
622 preempted = true;
623 thrd->preempted = reason;
624 returnToKernel();
625 }
626 __enable_interrupts_checked( false );
627 return preempted;
628}
629
630//=============================================================================================
631// Kernel Idle Sleep
632//=============================================================================================
633// Wake a thread from the front if there are any
634static void __wake_one(cluster * this) {
635 eventfd_t val;
636
637 /* paranoid */ verify( ! __preemption_enabled() );
638 /* paranoid */ verify( ready_schedule_islocked() );
639
640 // Check if there is a sleeping processor
641 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
642
643 // If no one is sleeping: we are done
644 if ( fdp == 0p ) return;
645
646 int fd = 1;
647 if ( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
648 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
649 }
650
651 switch(fd) {
652 __attribute__((unused)) int ret;
653 case 0:
654 // If the processor isn't ready to sleep then the exchange will already wake it up
655 #if ! defined(__CFA_NO_STATISTICS__)
656 if ( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
657 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
658 #endif
659 break;
660 case 1:
661 // If someone else already said they will wake them: we are done
662 #if ! defined(__CFA_NO_STATISTICS__)
663 if ( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
664 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
665 #endif
666 break;
667 default:
668 // If the processor was ready to sleep, we need to wake it up with an actual write
669 val = 1;
670 ret = eventfd_write( fd, val );
671 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
672
673 #if ! defined(__CFA_NO_STATISTICS__)
674 if ( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
675 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
676 #endif
677 break;
678 }
679
680 /* paranoid */ verify( ready_schedule_islocked() );
681 /* paranoid */ verify( ! __preemption_enabled() );
682
683 return;
684}
685
686// Unconditionnaly wake a thread
687void __wake_proc(processor * this) {
688 /* paranoid */ verify( ! __preemption_enabled() );
689
690 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
691
692 this->idle_wctx.sem = 1;
693
694 this->idle_wctx.wake__time = rdtscl();
695
696 eventfd_t val;
697 val = 1;
698 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
699
700 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
701 /* paranoid */ verify( ! __preemption_enabled() );
702}
703
704static void idle_sleep(processor * this) {
705 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
706 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
707
708 // Tell everyone we are ready to go do sleep
709 for() {
710 int expected = this->idle_wctx.sem;
711
712 // Someone already told us to wake-up! No time for a nap.
713 if (expected == 1) { return; }
714
715 // Try to mark that we are going to sleep
716 if (__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
717 // Every one agreed, taking a nap
718 break;
719 }
720 }
721
722
723 #if ! defined(__CFA_NO_STATISTICS__)
724 if (this->print_halts) {
725 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
726 }
727 #endif
728
729 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
730
731 {
732 eventfd_t val;
733 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
734 if (ret < 0) {
735 switch((int)errno) {
736 case EAGAIN:
737 #if EAGAIN != EWOULDBLOCK
738 case EWOULDBLOCK:
739 #endif
740 case EINTR:
741 // No need to do anything special here, just assume it's a legitimate wake-up
742 break;
743 default:
744 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
745 }
746 }
747 }
748
749 #if ! defined(__CFA_NO_STATISTICS__)
750 if (this->print_halts) {
751 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
752 }
753 #endif
754}
755
756static bool mark_idle(__cluster_proc_list & this, processor & proc) {
757 __STATS__(true, ready.sleep.halts++; )
758
759 proc.idle_wctx.sem = 0;
760
761 /* paranoid */ verify( ! __preemption_enabled() );
762 if ( ! try_lock( this )) return false;
763 this.idle++;
764 /* paranoid */ verify( this.idle <= this.total );
765 remove(proc);
766 insert_first(this.idles, proc);
767
768 // update the pointer to the head wait context, which should now point to this proc.
769 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
770 unlock( this );
771 /* paranoid */ verify( ! __preemption_enabled() );
772
773 return true;
774}
775
776static void mark_awake(__cluster_proc_list & this, processor & proc) {
777 /* paranoid */ verify( ! __preemption_enabled() );
778 lock( this );
779 this.idle--;
780 /* paranoid */ verify( this.idle >= 0 );
781 remove(proc);
782 insert_last(this.actives, proc);
783
784 {
785 // update the pointer to the head wait context
786 struct __fd_waitctx * wctx = 0;
787 if ( ! isEmpty( this.idles )) wctx = &first( this. idles ).idle_wctx;
788 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
789 }
790
791 unlock( this );
792 /* paranoid */ verify( ! __preemption_enabled() );
793}
794
795//=============================================================================================
796// Unexpected Terminating logic
797//=============================================================================================
798void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
799 thread$ * thrd = __cfaabi_tls.this_thread;
800
801 if (thrd) {
802 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
803 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
804
805 if ( &thrd->self_cor != thrd->curr_cor ) {
806 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
807 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
808 }
809 else {
810 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
811 }
812 }
813 else {
814 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
815 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
816 }
817}
818
819int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
820 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
821}
822
823static __spinlock_t kernel_debug_lock;
824
825extern "C" {
826 void __cfaabi_bits_acquire() {
827 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
828 }
829
830 void __cfaabi_bits_release() {
831 unlock( kernel_debug_lock );
832 }
833}
834
835//=============================================================================================
836// Kernel Utilities
837//=============================================================================================
838#if defined(CFA_HAVE_LINUX_IO_URING_H)
839#include "io/types.hfa"
840#endif
841
842//-----------------------------------------------------------------------------
843// Debug
844bool threading_enabled(void) __attribute__((const)) libcfa_public {
845 return true;
846}
847
848//-----------------------------------------------------------------------------
849// Statistics
850#if ! defined(__CFA_NO_STATISTICS__)
851 void print_halts( processor & this ) libcfa_public {
852 this.print_halts = true;
853 }
854
855 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
856 /* paranoid */ verify( cltr->stats );
857
858 processor * it = &first( list );
859 for(unsigned i = 0; i < count; i++) {
860 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
861 /* paranoid */ verify( it->local_data->this_stats );
862 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
863 __tally_stats( cltr->stats, it->local_data->this_stats );
864 it = &next( *it );
865 }
866 }
867
868 static void crawl_cluster_stats( cluster & this ) {
869 // Stop the world, otherwise stats could get really messed-up
870 // this doesn't solve all problems but does solve many
871 // so it's probably good enough
872 disable_interrupts();
873 uint_fast32_t last_size = ready_mutate_lock();
874
875 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
876 crawl_list(&this, this.procs.idles , this.procs.idle );
877
878 // Unlock the RWlock
879 ready_mutate_unlock( last_size );
880 enable_interrupts();
881 }
882
883
884 void print_stats_now( cluster & this, int flags ) libcfa_public {
885 crawl_cluster_stats( this );
886 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
887 }
888#endif
889// Local Variables: //
890// mode: c //
891// tab-width: 4 //
892// End: //
Note: See TracBrowser for help on using the repository browser.