source: libcfa/src/concurrency/kernel.cfa@ 5e180c2

ADT ast-experimental
Last change on this file since 5e180c2 was 26544f9, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

added helping and lock to allow remote processors to flush unresponsive procs
  • 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 : Wed Nov 30 18:14:08 2022
13// Update Count : 76
14//
15
16#define __cforall_thread__
17#define _GNU_SOURCE
18
19// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
20
21#pragma GCC diagnostic push
22#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
23
24//C Includes
25#include <errno.h>
26#include <stdio.h>
27#include <string.h>
28#include <signal.h>
29#include <unistd.h>
30
31extern "C" {
32 #include <sys/eventfd.h>
33 #include <sys/uio.h>
34}
35
36//CFA Includes
37#include "kernel/private.hfa"
38#include "preemption.hfa"
39#include "strstream.hfa"
40#include "device/cpu.hfa"
41#include "io/types.hfa"
42
43//Private includes
44#define __CFA_INVOKE_PRIVATE__
45#include "invoke.h"
46#pragma GCC diagnostic pop
47
48#if !defined(__CFA_NO_STATISTICS__)
49 #define __STATS_DEF( ...) __VA_ARGS__
50#else
51 #define __STATS_DEF( ...)
52#endif
53
54//-----------------------------------------------------------------------------
55// Some assembly required
56#if defined( __i386 )
57 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
58 // fcw : X87 FPU control word (preserved across function calls)
59 #define __x87_store \
60 uint32_t __mxcr; \
61 uint16_t __fcw; \
62 __asm__ volatile ( \
63 "stmxcsr %0\n" \
64 "fnstcw %1\n" \
65 : "=m" (__mxcr),\
66 "=m" (__fcw) \
67 )
68
69 #define __x87_load \
70 __asm__ volatile ( \
71 "fldcw %1\n" \
72 "ldmxcsr %0\n" \
73 ::"m" (__mxcr),\
74 "m" (__fcw) \
75 )
76
77#elif defined( __x86_64 )
78 #define __x87_store \
79 uint32_t __mxcr; \
80 uint16_t __fcw; \
81 __asm__ volatile ( \
82 "stmxcsr %0\n" \
83 "fnstcw %1\n" \
84 : "=m" (__mxcr),\
85 "=m" (__fcw) \
86 )
87
88 #define __x87_load \
89 __asm__ volatile ( \
90 "fldcw %1\n" \
91 "ldmxcsr %0\n" \
92 :: "m" (__mxcr),\
93 "m" (__fcw) \
94 )
95
96#elif defined( __arm__ )
97 #define __x87_store
98 #define __x87_load
99
100#elif defined( __aarch64__ )
101 #define __x87_store \
102 uint32_t __fpcntl[2]; \
103 __asm__ volatile ( \
104 "mrs x9, FPCR\n" \
105 "mrs x10, FPSR\n" \
106 "stp x9, x10, %0\n" \
107 : "=m" (__fpcntl) : : "x9", "x10" \
108 )
109
110 #define __x87_load \
111 __asm__ volatile ( \
112 "ldp x9, x10, %0\n" \
113 "msr FPSR, x10\n" \
114 "msr FPCR, x9\n" \
115 : "=m" (__fpcntl) : : "x9", "x10" \
116 )
117
118#else
119 #error unsupported hardware architecture
120#endif
121
122extern thread$ * mainThread;
123extern processor * mainProcessor;
124
125//-----------------------------------------------------------------------------
126// Kernel Scheduling logic
127static thread$ * __next_thread(cluster * this);
128static thread$ * __next_thread_slow(cluster * this);
129static thread$ * __next_thread_search(cluster * this);
130static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
131static void __run_thread(processor * this, thread$ * dst);
132static void __wake_one(cluster * cltr);
133
134static void idle_sleep(processor * proc);
135static bool mark_idle (__cluster_proc_list & idles, processor & proc);
136static void mark_awake(__cluster_proc_list & idles, processor & proc);
137
138extern bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1)));
139extern bool __cfa_io_flush( processor * ) __attribute__((nonnull (1)));
140
141
142extern void __disable_interrupts_hard();
143extern void __enable_interrupts_hard();
144
145
146//=============================================================================================
147// Kernel Scheduling logic
148//=============================================================================================
149//Main of the processor contexts
150void main(processorCtx_t & runner) {
151 // Because of a bug, we couldn't initialized the seed on construction
152 // Do it here
153 PRNG_SET_SEED( __cfaabi_tls.random_state, rdtscl() );
154 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
155 __tls_rand_advance_bck();
156
157 processor * this = runner.proc;
158 verify(this);
159
160 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
161 #if !defined(__CFA_NO_STATISTICS__)
162 if( this->print_halts ) {
163 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
164 }
165 #endif
166
167 {
168 // Setup preemption data
169 preemption_scope scope = { this };
170
171 // if we need to run some special setup, now is the time to do it.
172 if(this->init.thrd) {
173 this->init.thrd->curr_cluster = this->cltr;
174 __run_thread(this, this->init.thrd);
175 }
176
177 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
178
179 thread$ * readyThread = 0p;
180 MAIN_LOOP:
181 for() {
182 // Check if there is pending io
183 __cfa_io_drain( this );
184
185 // Try to get the next thread
186 readyThread = __next_thread( this->cltr );
187
188 if( !readyThread ) {
189 // there is no point in holding submissions if we are idle
190 __IO_STATS__(true, io.flush.idle++; )
191 __cfa_io_flush( this );
192
193 // drain again in case something showed up
194 __cfa_io_drain( this );
195
196 readyThread = __next_thread( this->cltr );
197 }
198
199 if( !readyThread ) for(5) {
200 readyThread = __next_thread_slow( this->cltr );
201
202 if( readyThread ) break;
203
204 // It's unlikely we still I/O to submit, but the arbiter could
205 __IO_STATS__(true, io.flush.idle++; )
206 __cfa_io_flush( this );
207
208 // drain again in case something showed up
209 __cfa_io_drain( this );
210 }
211
212 HALT:
213 if( !readyThread ) {
214 // Don't block if we are done
215 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
216
217 // Push self to idle stack
218 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
219
220 // Confirm the ready-queue is empty
221 readyThread = __next_thread_search( this->cltr );
222 if( readyThread ) {
223 // A thread was found, cancel the halt
224 mark_awake(this->cltr->procs, * this);
225
226 __STATS__(true, ready.sleep.cancels++; )
227
228 // continue the mai loop
229 break HALT;
230 }
231
232 idle_sleep( this );
233
234 // We were woken up, remove self from idle
235 mark_awake(this->cltr->procs, * this);
236
237 // DON'T just proceed, start looking again
238 continue MAIN_LOOP;
239 }
240
241 /* paranoid */ verify( readyThread );
242
243 // Reset io dirty bit
244 this->io.dirty = false;
245
246 // We found a thread run it
247 __run_thread(this, readyThread);
248
249 // Are we done?
250 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
251
252 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
253 __IO_STATS__(true, io.flush.dirty++; )
254 __cfa_io_flush( this );
255 }
256 }
257
258 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
259 }
260
261 __cfa_io_flush( this );
262 __cfa_io_drain( this );
263
264 post( this->terminated );
265
266 if(this == mainProcessor) {
267 // HACK : the coroutine context switch expects this_thread to be set
268 // and it make sense for it to be set in all other cases except here
269 // fake it
270 __cfaabi_tls.this_thread = mainThread;
271 }
272
273 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
274}
275
276static int * __volatile_errno() __attribute__((noinline));
277static int * __volatile_errno() { asm(""); return &errno; }
278
279// KERNEL ONLY
280// runThread runs a thread by context switching
281// from the processor coroutine to the target thread
282static void __run_thread(processor * this, thread$ * thrd_dst) {
283 /* paranoid */ verify( ! __preemption_enabled() );
284 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
285 /* paranoid */ verifyf( thrd_dst->rdy_link.next == 0p, "Expected null got %p", thrd_dst->rdy_link.next );
286 __builtin_prefetch( thrd_dst->context.SP );
287
288 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
289
290 coroutine$ * proc_cor = get_coroutine(this->runner);
291
292 // set state of processor coroutine to inactive
293 verify(proc_cor->state == Active);
294 proc_cor->state = Blocked;
295
296 // Actually run the thread
297 RUNNING: 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$( thrd, hint );
563 enable_interrupts(false);
564 }
565}
566
567void park( void ) libcfa_public {
568 __disable_interrupts_checked();
569 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
570 returnToKernel();
571 __enable_interrupts_checked();
572
573}
574
575extern "C" {
576 // Leave the thread monitor
577 // last routine called by a thread.
578 // Should never return
579 void __cfactx_thrd_leave() {
580 thread$ * thrd = active_thread();
581 monitor$ * this = &thrd->self_mon;
582
583 // Lock the monitor now
584 lock( this->lock __cfaabi_dbg_ctx2 );
585
586 disable_interrupts();
587
588 /* paranoid */ verify( ! __preemption_enabled() );
589 /* paranoid */ verify( thrd->state == Active );
590 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
591 /* paranoid */ verify( kernelTLS().this_thread == thrd );
592 /* paranoid */ verify( thrd->context.SP );
593 /* 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 );
594 /* 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 );
595
596 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
597 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
598 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
599
600 thrd->state = Halting;
601 thrd->ticket = TICKET_DEAD;
602
603 // Leave the thread
604 returnToKernel();
605
606 // Control flow should never reach here!
607 abort();
608 }
609}
610
611// KERNEL ONLY
612bool force_yield( __Preemption_Reason reason ) libcfa_public {
613 __disable_interrupts_checked();
614 thread$ * thrd = kernelTLS().this_thread;
615 /* paranoid */ verify(thrd->state == Active);
616
617 // SKULLDUGGERY: It is possible that we are preempting this thread just before
618 // it was going to park itself. If that is the case and it is already using the
619 // intrusive fields then we can't use them to preempt the thread
620 // If that is the case, abandon the preemption.
621 bool preempted = false;
622 if(thrd->rdy_link.next == 0p) {
623 preempted = true;
624 thrd->preempted = reason;
625 returnToKernel();
626 }
627 __enable_interrupts_checked( false );
628 return preempted;
629}
630
631//=============================================================================================
632// Kernel Idle Sleep
633//=============================================================================================
634// Wake a thread from the front if there are any
635static void __wake_one(cluster * this) {
636 eventfd_t val;
637
638 /* paranoid */ verify( ! __preemption_enabled() );
639 /* paranoid */ verify( ready_schedule_islocked() );
640
641 // Check if there is a sleeping processor
642 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
643
644 // If no one is sleeping: we are done
645 if( fdp == 0p ) return;
646
647 int fd = 1;
648 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
649 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
650 }
651
652 switch(fd) {
653 __attribute__((unused)) int ret;
654 case 0:
655 // If the processor isn't ready to sleep then the exchange will already wake it up
656 #if !defined(__CFA_NO_STATISTICS__)
657 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
658 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
659 #endif
660 break;
661 case 1:
662 // If someone else already said they will wake them: we are done
663 #if !defined(__CFA_NO_STATISTICS__)
664 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
665 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
666 #endif
667 break;
668 default:
669 // If the processor was ready to sleep, we need to wake it up with an actual write
670 val = 1;
671 ret = eventfd_write( fd, val );
672 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
673
674 #if !defined(__CFA_NO_STATISTICS__)
675 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
676 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
677 #endif
678 break;
679 }
680
681 /* paranoid */ verify( ready_schedule_islocked() );
682 /* paranoid */ verify( ! __preemption_enabled() );
683
684 return;
685}
686
687// Unconditionnaly wake a thread
688void __wake_proc(processor * this) {
689 /* paranoid */ verify( ! __preemption_enabled() );
690
691 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
692
693 this->idle_wctx.sem = 1;
694
695 this->idle_wctx.wake__time = rdtscl();
696
697 eventfd_t val;
698 val = 1;
699 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
700
701 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
702 /* paranoid */ verify( ! __preemption_enabled() );
703}
704
705static void idle_sleep(processor * this) {
706 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
707 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
708
709 // Tell everyone we are ready to go do sleep
710 for() {
711 int expected = this->idle_wctx.sem;
712
713 // Someone already told us to wake-up! No time for a nap.
714 if(expected == 1) { return; }
715
716 // Try to mark that we are going to sleep
717 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
718 // Every one agreed, taking a nap
719 break;
720 }
721 }
722
723
724 #if !defined(__CFA_NO_STATISTICS__)
725 if(this->print_halts) {
726 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
727 }
728 #endif
729
730 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
731
732 {
733 eventfd_t val;
734 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
735 if(ret < 0) {
736 switch((int)errno) {
737 case EAGAIN:
738 #if EAGAIN != EWOULDBLOCK
739 case EWOULDBLOCK:
740 #endif
741 case EINTR:
742 // No need to do anything special here, just assume it's a legitimate wake-up
743 break;
744 default:
745 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
746 }
747 }
748 }
749
750 #if !defined(__CFA_NO_STATISTICS__)
751 if(this->print_halts) {
752 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
753 }
754 #endif
755}
756
757static bool mark_idle(__cluster_proc_list & this, processor & proc) {
758 __STATS__(true, ready.sleep.halts++; )
759
760 proc.idle_wctx.sem = 0;
761
762 /* paranoid */ verify( ! __preemption_enabled() );
763 if(!try_lock( this )) return false;
764 this.idle++;
765 /* paranoid */ verify( this.idle <= this.total );
766 remove(proc);
767 insert_first(this.idles, proc);
768
769 // update the pointer to the head wait context, which should now point to this proc.
770 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
771 unlock( this );
772 /* paranoid */ verify( ! __preemption_enabled() );
773
774 return true;
775}
776
777static void mark_awake(__cluster_proc_list & this, processor & proc) {
778 /* paranoid */ verify( ! __preemption_enabled() );
779 lock( this );
780 this.idle--;
781 /* paranoid */ verify( this.idle >= 0 );
782 remove(proc);
783 insert_last(this.actives, proc);
784
785 {
786 // update the pointer to the head wait context
787 struct __fd_waitctx * wctx = 0;
788 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
789 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
790 }
791
792 unlock( this );
793 /* paranoid */ verify( ! __preemption_enabled() );
794}
795
796//=============================================================================================
797// Unexpected Terminating logic
798//=============================================================================================
799void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
800 thread$ * thrd = __cfaabi_tls.this_thread;
801
802 if(thrd) {
803 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
804 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
805
806 if ( &thrd->self_cor != thrd->curr_cor ) {
807 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
808 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
809 }
810 else {
811 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
812 }
813 }
814 else {
815 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
816 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
817 }
818}
819
820int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
821 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
822}
823
824static __spinlock_t kernel_debug_lock;
825
826extern "C" {
827 void __cfaabi_bits_acquire() {
828 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
829 }
830
831 void __cfaabi_bits_release() {
832 unlock( kernel_debug_lock );
833 }
834}
835
836//=============================================================================================
837// Kernel Utilities
838//=============================================================================================
839#if defined(CFA_HAVE_LINUX_IO_URING_H)
840#include "io/types.hfa"
841#endif
842
843//-----------------------------------------------------------------------------
844// Debug
845bool threading_enabled(void) __attribute__((const)) libcfa_public {
846 return true;
847}
848
849//-----------------------------------------------------------------------------
850// Statistics
851#if !defined(__CFA_NO_STATISTICS__)
852 void print_halts( processor & this ) libcfa_public {
853 this.print_halts = true;
854 }
855
856 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
857 /* paranoid */ verify( cltr->stats );
858
859 processor * it = &list`first;
860 for(unsigned i = 0; i < count; i++) {
861 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
862 /* paranoid */ verify( it->local_data->this_stats );
863 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
864 __tally_stats( cltr->stats, it->local_data->this_stats );
865 it = &(*it)`next;
866 }
867 }
868
869 static void crawl_cluster_stats( cluster & this ) {
870 // Stop the world, otherwise stats could get really messed-up
871 // this doesn't solve all problems but does solve many
872 // so it's probably good enough
873 disable_interrupts();
874 uint_fast32_t last_size = ready_mutate_lock();
875
876 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
877 crawl_list(&this, this.procs.idles , this.procs.idle );
878
879 // Unlock the RWlock
880 ready_mutate_unlock( last_size );
881 enable_interrupts();
882 }
883
884
885 void print_stats_now( cluster & this, int flags ) libcfa_public {
886 crawl_cluster_stats( this );
887 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
888 }
889#endif
890// Local Variables: //
891// mode: c //
892// tab-width: 4 //
893// End: //
Note: See TracBrowser for help on using the repository browser.