source: libcfa/src/concurrency/kernel.cfa@ 20be782

ADT ast-experimental pthread-emulation
Last change on this file since 20be782 was ced5e2a, checked in by caparsons <caparson@…>, 3 years ago

removed debug fields from spinlock_t
  • Property mode set to 100644
File size: 28.4 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 Aug 31 07:08:20 2020
13// Update Count : 71
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)));
140extern void __cfa_io_idle( processor * ) __attribute__((nonnull (1)));
141
142#if defined(CFA_WITH_IO_URING_IDLE)
143 extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
144#endif
145
146extern void __disable_interrupts_hard();
147extern void __enable_interrupts_hard();
148
149
150//=============================================================================================
151// Kernel Scheduling logic
152//=============================================================================================
153//Main of the processor contexts
154void main(processorCtx_t & runner) {
155 // Because of a bug, we couldn't initialized the seed on construction
156 // Do it here
157 __cfaabi_tls.rand_seed ^= rdtscl();
158 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
159 __tls_rand_advance_bck();
160
161 processor * this = runner.proc;
162 verify(this);
163
164 /* paranoid */ verify( this->idle_wctx.ftr != 0p );
165 /* paranoid */ verify( this->idle_wctx.rdbuf != 0p );
166
167 // used for idle sleep when io_uring is present
168 // mark it as already fulfilled so we know if there is a pending request or not
169 this->idle_wctx.ftr->self.ptr = 1p;
170
171 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
172 #if !defined(__CFA_NO_STATISTICS__)
173 if( this->print_halts ) {
174 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
175 }
176 #endif
177
178 {
179 // Setup preemption data
180 preemption_scope scope = { this };
181
182 // if we need to run some special setup, now is the time to do it.
183 if(this->init.thrd) {
184 this->init.thrd->curr_cluster = this->cltr;
185 __run_thread(this, this->init.thrd);
186 }
187
188 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
189
190 thread$ * readyThread = 0p;
191 MAIN_LOOP:
192 for() {
193 // Check if there is pending io
194 __cfa_io_drain( this );
195
196 // Try to get the next thread
197 readyThread = __next_thread( this->cltr );
198
199 if( !readyThread ) {
200 // there is no point in holding submissions if we are idle
201 __IO_STATS__(true, io.flush.idle++; )
202 __cfa_io_flush( this );
203
204 // drain again in case something showed up
205 __cfa_io_drain( this );
206
207 readyThread = __next_thread( this->cltr );
208 }
209
210 if( !readyThread ) for(5) {
211 readyThread = __next_thread_slow( this->cltr );
212
213 if( readyThread ) break;
214
215 // It's unlikely we still I/O to submit, but the arbiter could
216 __IO_STATS__(true, io.flush.idle++; )
217 __cfa_io_flush( this );
218
219 // drain again in case something showed up
220 __cfa_io_drain( this );
221 }
222
223 HALT:
224 if( !readyThread ) {
225 // Don't block if we are done
226 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
227
228 // Push self to idle stack
229 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
230
231 // Confirm the ready-queue is empty
232 readyThread = __next_thread_search( this->cltr );
233 if( readyThread ) {
234 // A thread was found, cancel the halt
235 mark_awake(this->cltr->procs, * this);
236
237 __STATS__(true, ready.sleep.cancels++; )
238
239 // continue the mai loop
240 break HALT;
241 }
242
243 idle_sleep( this );
244
245 // We were woken up, remove self from idle
246 mark_awake(this->cltr->procs, * this);
247
248 // DON'T just proceed, start looking again
249 continue MAIN_LOOP;
250 }
251
252 /* paranoid */ verify( readyThread );
253
254 // Reset io dirty bit
255 this->io.dirty = false;
256
257 // We found a thread run it
258 __run_thread(this, readyThread);
259
260 // Are we done?
261 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
262
263 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
264 __IO_STATS__(true, io.flush.dirty++; )
265 __cfa_io_flush( this );
266 }
267 }
268
269 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
270 }
271
272 post( this->terminated );
273
274 if(this == mainProcessor) {
275 // HACK : the coroutine context switch expects this_thread to be set
276 // and it make sense for it to be set in all other cases except here
277 // fake it
278 __cfaabi_tls.this_thread = mainThread;
279 }
280
281 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
282}
283
284static int * __volatile_errno() __attribute__((noinline));
285static int * __volatile_errno() { asm(""); return &errno; }
286
287// KERNEL ONLY
288// runThread runs a thread by context switching
289// from the processor coroutine to the target thread
290static void __run_thread(processor * this, thread$ * thrd_dst) {
291 /* paranoid */ verify( ! __preemption_enabled() );
292 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
293 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
294 __builtin_prefetch( thrd_dst->context.SP );
295
296 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
297
298 coroutine$ * proc_cor = get_coroutine(this->runner);
299
300 // set state of processor coroutine to inactive
301 verify(proc_cor->state == Active);
302 proc_cor->state = Blocked;
303
304 // Actually run the thread
305 RUNNING: while(true) {
306 thrd_dst->preempted = __NO_PREEMPTION;
307 thrd_dst->state = Active;
308
309 // Update global state
310 kernelTLS().this_thread = thrd_dst;
311
312 /* paranoid */ verify( ! __preemption_enabled() );
313 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
314 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
315 /* paranoid */ verify( thrd_dst->context.SP );
316 /* paranoid */ verify( thrd_dst->state != Halted );
317 /* 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
318 /* 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
319 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
320
321
322
323 // set context switch to the thread that the processor is executing
324 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
325 // when __cfactx_switch returns we are back in the processor coroutine
326
327
328
329 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
330 /* 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 );
331 /* 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 );
332 /* paranoid */ verify( thrd_dst->context.SP );
333 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
334 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
335 /* paranoid */ verify( ! __preemption_enabled() );
336
337 // Reset global state
338 kernelTLS().this_thread = 0p;
339
340 // We just finished running a thread, there are a few things that could have happened.
341 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
342 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
343 // 4 - Preempted
344 // In case 1, we may have won a race so we can't write to the state again.
345 // In case 2, we lost the race so we now own the thread.
346
347 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
348 // The thread was preempted, reschedule it and reset the flag
349 schedule_thread$( thrd_dst, UNPARK_LOCAL );
350 break RUNNING;
351 }
352
353 if(unlikely(thrd_dst->state == Halting)) {
354 // The thread has halted, it should never be scheduled/run again
355 // finish the thread
356 __thread_finish( thrd_dst );
357 break RUNNING;
358 }
359
360 /* paranoid */ verify( thrd_dst->state == Active );
361 thrd_dst->state = Blocked;
362
363 // set state of processor coroutine to active and the thread to inactive
364 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
365 switch(old_ticket) {
366 case TICKET_RUNNING:
367 // This is case 1, the regular case, nothing more is needed
368 break RUNNING;
369 case TICKET_UNBLOCK:
370 __STATS__(true, ready.threads.threads++; )
371 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
372 // In this case, just run it again.
373 continue RUNNING;
374 default:
375 // This makes no sense, something is wrong abort
376 abort();
377 }
378 }
379
380 // Just before returning to the processor, set the processor coroutine to active
381 proc_cor->state = Active;
382
383 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
384
385 __STATS__(true, ready.threads.threads--; )
386
387 /* paranoid */ verify( ! __preemption_enabled() );
388}
389
390// KERNEL_ONLY
391static void returnToKernel() {
392 /* paranoid */ verify( ! __preemption_enabled() );
393 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
394 thread$ * thrd_src = kernelTLS().this_thread;
395
396 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
397
398 // Run the thread on this processor
399 {
400 int local_errno = *__volatile_errno();
401 #if defined( __i386 ) || defined( __x86_64 )
402 __x87_store;
403 #endif
404 /* paranoid */ verify( proc_cor->context.SP );
405 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
406 __cfactx_switch( &thrd_src->context, &proc_cor->context );
407 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
408 #if defined( __i386 ) || defined( __x86_64 )
409 __x87_load;
410 #endif
411 *__volatile_errno() = local_errno;
412 }
413
414 #if !defined(__CFA_NO_STATISTICS__)
415 /* paranoid */ verify( thrd_src->last_proc != 0p );
416 if(thrd_src->last_proc != kernelTLS().this_processor) {
417 __tls_stats()->ready.threads.migration++;
418 }
419 #endif
420
421 /* paranoid */ verify( ! __preemption_enabled() );
422 /* 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 );
423 /* 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 );
424}
425
426//-----------------------------------------------------------------------------
427// Scheduler routines
428// KERNEL ONLY
429static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
430 /* paranoid */ verify( ! __preemption_enabled() );
431 /* paranoid */ verify( ready_schedule_islocked());
432 /* paranoid */ verify( thrd );
433 /* paranoid */ verify( thrd->state != Halted );
434 /* paranoid */ verify( thrd->curr_cluster );
435 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
436 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
437 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
438 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
439 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
440 /* paranoid */ #endif
441 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
442 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
443
444 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
445
446 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
447 struct cluster * cl = thrd->curr_cluster;
448 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
449
450 // push the thread to the cluster ready-queue
451 push( cl, thrd, hint );
452
453 // variable thrd is no longer safe to use
454 thrd = 0xdeaddeaddeaddeadp;
455
456 // wake the cluster using the save variable.
457 __wake_one( cl );
458
459 #if !defined(__CFA_NO_STATISTICS__)
460 if( kernelTLS().this_stats ) {
461 __tls_stats()->ready.threads.threads++;
462 if(outside) {
463 __tls_stats()->ready.threads.extunpark++;
464 }
465 }
466 else {
467 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
468 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
469 }
470 #endif
471
472 /* paranoid */ verify( ready_schedule_islocked());
473 /* paranoid */ verify( ! __preemption_enabled() );
474}
475
476void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
477 ready_schedule_lock();
478 __schedule_thread( thrd, hint );
479 ready_schedule_unlock();
480}
481
482// KERNEL ONLY
483static inline thread$ * __next_thread(cluster * this) with( *this ) {
484 /* paranoid */ verify( ! __preemption_enabled() );
485
486 ready_schedule_lock();
487 thread$ * thrd = pop_fast( this );
488 ready_schedule_unlock();
489
490 /* paranoid */ verify( ! __preemption_enabled() );
491 return thrd;
492}
493
494// KERNEL ONLY
495static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
496 /* paranoid */ verify( ! __preemption_enabled() );
497
498 ready_schedule_lock();
499 thread$ * thrd = pop_slow( this );
500 ready_schedule_unlock();
501
502 /* paranoid */ verify( ! __preemption_enabled() );
503 return thrd;
504}
505
506// KERNEL ONLY
507static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
508 /* paranoid */ verify( ! __preemption_enabled() );
509
510 ready_schedule_lock();
511 thread$ * thrd = pop_search( this );
512 ready_schedule_unlock();
513
514 /* paranoid */ verify( ! __preemption_enabled() );
515 return thrd;
516}
517
518static inline bool __must_unpark( thread$ * thrd ) {
519 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
520 switch(old_ticket) {
521 case TICKET_RUNNING:
522 // Wake won the race, the thread will reschedule/rerun itself
523 return false;
524 case TICKET_BLOCKED:
525 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
526 /* paranoid */ verify( thrd->state == Blocked );
527 return true;
528 default:
529 // This makes no sense, something is wrong abort
530 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
531 }
532}
533
534void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
535 /* paranoid */ verify( ! __preemption_enabled() );
536 /* paranoid */ verify( ready_schedule_islocked());
537
538 if( !thrd ) return;
539
540 if(__must_unpark(thrd)) {
541 // Wake lost the race,
542 __schedule_thread( thrd, hint );
543 }
544
545 /* paranoid */ verify( ready_schedule_islocked());
546 /* paranoid */ verify( ! __preemption_enabled() );
547}
548
549void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
550 if( !thrd ) return;
551
552 if(__must_unpark(thrd)) {
553 disable_interrupts();
554 // Wake lost the race,
555 schedule_thread$( thrd, hint );
556 enable_interrupts(false);
557 }
558}
559
560void park( void ) libcfa_public {
561 __disable_interrupts_checked();
562 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
563 returnToKernel();
564 __enable_interrupts_checked();
565
566}
567
568extern "C" {
569 // Leave the thread monitor
570 // last routine called by a thread.
571 // Should never return
572 void __cfactx_thrd_leave() {
573 thread$ * thrd = active_thread();
574 monitor$ * this = &thrd->self_mon;
575
576 // Lock the monitor now
577 lock( this->lock __cfaabi_dbg_ctx2 );
578
579 disable_interrupts();
580
581 /* paranoid */ verify( ! __preemption_enabled() );
582 /* paranoid */ verify( thrd->state == Active );
583 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
584 /* paranoid */ verify( kernelTLS().this_thread == thrd );
585 /* paranoid */ verify( thrd->context.SP );
586 /* 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 );
587 /* 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 );
588
589 thrd->state = Halting;
590 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
591 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
592 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
593
594 // Leave the thread
595 returnToKernel();
596
597 // Control flow should never reach here!
598 abort();
599 }
600}
601
602// KERNEL ONLY
603bool force_yield( __Preemption_Reason reason ) libcfa_public {
604 __disable_interrupts_checked();
605 thread$ * thrd = kernelTLS().this_thread;
606 /* paranoid */ verify(thrd->state == Active);
607
608 // SKULLDUGGERY: It is possible that we are preempting this thread just before
609 // it was going to park itself. If that is the case and it is already using the
610 // intrusive fields then we can't use them to preempt the thread
611 // If that is the case, abandon the preemption.
612 bool preempted = false;
613 if(thrd->link.next == 0p) {
614 preempted = true;
615 thrd->preempted = reason;
616 returnToKernel();
617 }
618 __enable_interrupts_checked( false );
619 return preempted;
620}
621
622//=============================================================================================
623// Kernel Idle Sleep
624//=============================================================================================
625// Wake a thread from the front if there are any
626static void __wake_one(cluster * this) {
627 eventfd_t val;
628
629 /* paranoid */ verify( ! __preemption_enabled() );
630 /* paranoid */ verify( ready_schedule_islocked() );
631
632 // Check if there is a sleeping processor
633 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
634
635 // If no one is sleeping: we are done
636 if( fdp == 0p ) return;
637
638 int fd = 1;
639 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
640 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
641 }
642
643 switch(fd) {
644 __attribute__((unused)) int ret;
645 case 0:
646 // If the processor isn't ready to sleep then the exchange will already wake it up
647 #if !defined(__CFA_NO_STATISTICS__)
648 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
649 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
650 #endif
651 break;
652 case 1:
653 // If someone else already said they will wake them: we are done
654 #if !defined(__CFA_NO_STATISTICS__)
655 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
656 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
657 #endif
658 break;
659 default:
660 // If the processor was ready to sleep, we need to wake it up with an actual write
661 val = 1;
662 ret = eventfd_write( fd, val );
663 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
664
665 #if !defined(__CFA_NO_STATISTICS__)
666 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
667 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
668 #endif
669 break;
670 }
671
672 /* paranoid */ verify( ready_schedule_islocked() );
673 /* paranoid */ verify( ! __preemption_enabled() );
674
675 return;
676}
677
678// Unconditionnaly wake a thread
679void __wake_proc(processor * this) {
680 /* paranoid */ verify( ! __preemption_enabled() );
681
682 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
683
684 this->idle_wctx.sem = 1;
685
686 this->idle_wctx.wake__time = rdtscl();
687
688 eventfd_t val;
689 val = 1;
690 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
691
692 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
693 /* paranoid */ verify( ! __preemption_enabled() );
694}
695
696static void idle_sleep(processor * this) {
697 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
698 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
699
700 // Tell everyone we are ready to go do sleep
701 for() {
702 int expected = this->idle_wctx.sem;
703
704 // Someone already told us to wake-up! No time for a nap.
705 if(expected == 1) { return; }
706
707 // Try to mark that we are going to sleep
708 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
709 // Every one agreed, taking a nap
710 break;
711 }
712 }
713
714
715 #if !defined(CFA_WITH_IO_URING_IDLE)
716 #if !defined(__CFA_NO_STATISTICS__)
717 if(this->print_halts) {
718 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
719 }
720 #endif
721
722 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
723
724 {
725 eventfd_t val;
726 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
727 if(ret < 0) {
728 switch((int)errno) {
729 case EAGAIN:
730 #if EAGAIN != EWOULDBLOCK
731 case EWOULDBLOCK:
732 #endif
733 case EINTR:
734 // No need to do anything special here, just assume it's a legitimate wake-up
735 break;
736 default:
737 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
738 }
739 }
740 }
741
742 #if !defined(__CFA_NO_STATISTICS__)
743 if(this->print_halts) {
744 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
745 }
746 #endif
747 #else
748 __cfa_io_idle( this );
749 #endif
750}
751
752static bool mark_idle(__cluster_proc_list & this, processor & proc) {
753 __STATS__(true, ready.sleep.halts++; )
754
755 proc.idle_wctx.sem = 0;
756
757 /* paranoid */ verify( ! __preemption_enabled() );
758 if(!try_lock( this )) return false;
759 this.idle++;
760 /* paranoid */ verify( this.idle <= this.total );
761 remove(proc);
762 insert_first(this.idles, proc);
763
764 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
765 unlock( this );
766 /* paranoid */ verify( ! __preemption_enabled() );
767
768 return true;
769}
770
771static void mark_awake(__cluster_proc_list & this, processor & proc) {
772 /* paranoid */ verify( ! __preemption_enabled() );
773 lock( this );
774 this.idle--;
775 /* paranoid */ verify( this.idle >= 0 );
776 remove(proc);
777 insert_last(this.actives, proc);
778
779 {
780 struct __fd_waitctx * wctx = 0;
781 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
782 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
783 }
784
785 unlock( this );
786 /* paranoid */ verify( ! __preemption_enabled() );
787}
788
789//=============================================================================================
790// Unexpected Terminating logic
791//=============================================================================================
792void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
793 thread$ * thrd = __cfaabi_tls.this_thread;
794
795 if(thrd) {
796 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
797 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
798
799 if ( &thrd->self_cor != thrd->curr_cor ) {
800 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
801 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
802 }
803 else {
804 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
805 }
806 }
807 else {
808 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
809 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
810 }
811}
812
813int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
814 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
815}
816
817static __spinlock_t kernel_debug_lock;
818
819extern "C" {
820 void __cfaabi_bits_acquire() {
821 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
822 }
823
824 void __cfaabi_bits_release() {
825 unlock( kernel_debug_lock );
826 }
827}
828
829//=============================================================================================
830// Kernel Utilities
831//=============================================================================================
832#if defined(CFA_HAVE_LINUX_IO_URING_H)
833#include "io/types.hfa"
834#endif
835
836//-----------------------------------------------------------------------------
837// Debug
838bool threading_enabled(void) __attribute__((const)) libcfa_public {
839 return true;
840}
841
842//-----------------------------------------------------------------------------
843// Statistics
844#if !defined(__CFA_NO_STATISTICS__)
845 void print_halts( processor & this ) libcfa_public {
846 this.print_halts = true;
847 }
848
849 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
850 /* paranoid */ verify( cltr->stats );
851
852 processor * it = &list`first;
853 for(unsigned i = 0; i < count; i++) {
854 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
855 /* paranoid */ verify( it->local_data->this_stats );
856 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
857 __tally_stats( cltr->stats, it->local_data->this_stats );
858 it = &(*it)`next;
859 }
860 }
861
862 static void crawl_cluster_stats( cluster & this ) {
863 // Stop the world, otherwise stats could get really messed-up
864 // this doesn't solve all problems but does solve many
865 // so it's probably good enough
866 disable_interrupts();
867 uint_fast32_t last_size = ready_mutate_lock();
868
869 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
870 crawl_list(&this, this.procs.idles , this.procs.idle );
871
872 // Unlock the RWlock
873 ready_mutate_unlock( last_size );
874 enable_interrupts();
875 }
876
877
878 void print_stats_now( cluster & this, int flags ) libcfa_public {
879 crawl_cluster_stats( this );
880 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
881 }
882#endif
883// Local Variables: //
884// mode: c //
885// tab-width: 4 //
886// End: //
Note: See TracBrowser for help on using the repository browser.