source: libcfa/src/concurrency/kernel.cfa@ eaf269d

ADT ast-experimental enum forall-pointer-decay pthread-emulation qualifiedEnum
Last change on this file since eaf269d was d3605f8, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Reworked io_uring idle sleep to work with either read or readv depending on what's available.
  • Property mode set to 100644
File size: 30.9 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//C Includes
22#include <errno.h>
23#include <stdio.h>
24#include <string.h>
25#include <signal.h>
26#include <unistd.h>
27extern "C" {
28 #include <sys/eventfd.h>
29 #include <sys/uio.h>
30}
31
32//CFA Includes
33#include "kernel_private.hfa"
34#include "preemption.hfa"
35#include "strstream.hfa"
36#include "device/cpu.hfa"
37#include "io/types.hfa"
38
39//Private includes
40#define __CFA_INVOKE_PRIVATE__
41#include "invoke.h"
42
43#if !defined(__CFA_NO_STATISTICS__)
44 #define __STATS( ...) __VA_ARGS__
45#else
46 #define __STATS( ...)
47#endif
48
49//-----------------------------------------------------------------------------
50// Some assembly required
51#if defined( __i386 )
52 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
53 // fcw : X87 FPU control word (preserved across function calls)
54 #define __x87_store \
55 uint32_t __mxcr; \
56 uint16_t __fcw; \
57 __asm__ volatile ( \
58 "stmxcsr %0\n" \
59 "fnstcw %1\n" \
60 : "=m" (__mxcr),\
61 "=m" (__fcw) \
62 )
63
64 #define __x87_load \
65 __asm__ volatile ( \
66 "fldcw %1\n" \
67 "ldmxcsr %0\n" \
68 ::"m" (__mxcr),\
69 "m" (__fcw) \
70 )
71
72#elif defined( __x86_64 )
73 #define __x87_store \
74 uint32_t __mxcr; \
75 uint16_t __fcw; \
76 __asm__ volatile ( \
77 "stmxcsr %0\n" \
78 "fnstcw %1\n" \
79 : "=m" (__mxcr),\
80 "=m" (__fcw) \
81 )
82
83 #define __x87_load \
84 __asm__ volatile ( \
85 "fldcw %1\n" \
86 "ldmxcsr %0\n" \
87 :: "m" (__mxcr),\
88 "m" (__fcw) \
89 )
90
91#elif defined( __arm__ )
92 #define __x87_store
93 #define __x87_load
94
95#elif defined( __aarch64__ )
96 #define __x87_store \
97 uint32_t __fpcntl[2]; \
98 __asm__ volatile ( \
99 "mrs x9, FPCR\n" \
100 "mrs x10, FPSR\n" \
101 "stp x9, x10, %0\n" \
102 : "=m" (__fpcntl) : : "x9", "x10" \
103 )
104
105 #define __x87_load \
106 __asm__ volatile ( \
107 "ldp x9, x10, %0\n" \
108 "msr FPSR, x10\n" \
109 "msr FPCR, x9\n" \
110 : "=m" (__fpcntl) : : "x9", "x10" \
111 )
112
113#else
114 #error unsupported hardware architecture
115#endif
116
117extern thread$ * mainThread;
118extern processor * mainProcessor;
119
120//-----------------------------------------------------------------------------
121// Kernel Scheduling logic
122static thread$ * __next_thread(cluster * this);
123static thread$ * __next_thread_slow(cluster * this);
124static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
125static void __run_thread(processor * this, thread$ * dst);
126static void __wake_one(cluster * cltr);
127
128static void idle_sleep(processor * proc, io_future_t & future, iovec & iov);
129static bool mark_idle (__cluster_proc_list & idles, processor & proc);
130static void mark_awake(__cluster_proc_list & idles, processor & proc);
131
132extern void __cfa_io_start( processor * );
133extern bool __cfa_io_drain( processor * );
134extern bool __cfa_io_flush( processor *, int min_comp );
135extern void __cfa_io_stop ( processor * );
136static inline bool __maybe_io_drain( processor * );
137
138#if defined(CFA_WITH_IO_URING_IDLE)
139 extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
140#endif
141
142extern void __disable_interrupts_hard();
143extern void __enable_interrupts_hard();
144
145static inline void __disable_interrupts_checked() {
146 /* paranoid */ verify( __preemption_enabled() );
147 disable_interrupts();
148 /* paranoid */ verify( ! __preemption_enabled() );
149}
150
151static inline void __enable_interrupts_checked( bool poll = true ) {
152 /* paranoid */ verify( ! __preemption_enabled() );
153 enable_interrupts( poll );
154 /* paranoid */ verify( __preemption_enabled() );
155}
156
157
158//=============================================================================================
159// Kernel Scheduling logic
160//=============================================================================================
161//Main of the processor contexts
162void main(processorCtx_t & runner) {
163 // Because of a bug, we couldn't initialized the seed on construction
164 // Do it here
165 __cfaabi_tls.rand_seed ^= rdtscl();
166 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
167 __tls_rand_advance_bck();
168
169 processor * this = runner.proc;
170 verify(this);
171
172 io_future_t future; // used for idle sleep when io_uring is present
173 future.self.ptr = 1p; // mark it as already fulfilled so we know if there is a pending request or not
174 eventfd_t idle_val;
175 iovec idle_iovec = { &idle_val, sizeof(idle_val) };
176
177 __cfa_io_start( this );
178
179 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
180 #if !defined(__CFA_NO_STATISTICS__)
181 if( this->print_halts ) {
182 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
183 }
184 #endif
185
186 {
187 // Setup preemption data
188 preemption_scope scope = { this };
189
190 // if we need to run some special setup, now is the time to do it.
191 if(this->init.thrd) {
192 this->init.thrd->curr_cluster = this->cltr;
193 __run_thread(this, this->init.thrd);
194 }
195
196 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
197
198 thread$ * readyThread = 0p;
199 MAIN_LOOP:
200 for() {
201 #define OLD_MAIN 1
202 #if OLD_MAIN
203 // Check if there is pending io
204 __maybe_io_drain( this );
205
206 // Try to get the next thread
207 readyThread = __next_thread( this->cltr );
208
209 if( !readyThread ) {
210 __cfa_io_flush( this, 0 );
211
212 readyThread = __next_thread_slow( this->cltr );
213 }
214
215 HALT:
216 if( !readyThread ) {
217 // Don't block if we are done
218 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
219
220 #if !defined(__CFA_NO_STATISTICS__)
221 __tls_stats()->ready.sleep.halts++;
222 #endif
223
224 // Push self to idle stack
225 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
226
227 // Confirm the ready-queue is empty
228 readyThread = __next_thread_slow( this->cltr );
229 if( readyThread ) {
230 // A thread was found, cancel the halt
231 mark_awake(this->cltr->procs, * this);
232
233 #if !defined(__CFA_NO_STATISTICS__)
234 __tls_stats()->ready.sleep.cancels++;
235 #endif
236
237 // continue the mai loop
238 break HALT;
239 }
240
241 idle_sleep( this, future, idle_iovec );
242
243 // We were woken up, remove self from idle
244 mark_awake(this->cltr->procs, * this);
245
246 // DON'T just proceed, start looking again
247 continue MAIN_LOOP;
248 }
249
250 /* paranoid */ verify( readyThread );
251
252 // Reset io dirty bit
253 this->io.dirty = false;
254
255 // We found a thread run it
256 __run_thread(this, readyThread);
257
258 // Are we done?
259 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
260
261 if(this->io.pending && !this->io.dirty) {
262 __cfa_io_flush( this, 0 );
263 }
264
265 #else
266 #warning new kernel loop
267 SEARCH: {
268 /* paranoid */ verify( ! __preemption_enabled() );
269
270 // First, lock the scheduler since we are searching for a thread
271 ready_schedule_lock();
272
273 // Try to get the next thread
274 readyThread = pop_fast( this->cltr );
275 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
276
277 // If we can't find a thread, might as well flush any outstanding I/O
278 if(this->io.pending) { __cfa_io_flush( this, 0 ); }
279
280 // Spin a little on I/O, just in case
281 for(5) {
282 __maybe_io_drain( this );
283 readyThread = pop_fast( this->cltr );
284 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
285 }
286
287 // no luck, try stealing a few times
288 for(5) {
289 if( __maybe_io_drain( this ) ) {
290 readyThread = pop_fast( this->cltr );
291 } else {
292 readyThread = pop_slow( this->cltr );
293 }
294 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
295 }
296
297 // still no luck, search for a thread
298 readyThread = pop_search( this->cltr );
299 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
300
301 // Don't block if we are done
302 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
303 ready_schedule_unlock();
304 break MAIN_LOOP;
305 }
306
307 __STATS( __tls_stats()->ready.sleep.halts++; )
308
309 // Push self to idle stack
310 ready_schedule_unlock();
311 if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
312 ready_schedule_lock();
313
314 // Confirm the ready-queue is empty
315 __maybe_io_drain( this );
316 readyThread = pop_search( this->cltr );
317 ready_schedule_unlock();
318
319 if( readyThread ) {
320 // A thread was found, cancel the halt
321 mark_awake(this->cltr->procs, * this);
322
323 __STATS( __tls_stats()->ready.sleep.cancels++; )
324
325 // continue the main loop
326 break SEARCH;
327 }
328
329 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
330 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
331
332 {
333 eventfd_t val;
334 ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
335 if(ret < 0) {
336 switch((int)errno) {
337 case EAGAIN:
338 #if EAGAIN != EWOULDBLOCK
339 case EWOULDBLOCK:
340 #endif
341 case EINTR:
342 // No need to do anything special here, just assume it's a legitimate wake-up
343 break;
344 default:
345 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
346 }
347 }
348 }
349
350 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
351
352 // We were woken up, remove self from idle
353 mark_awake(this->cltr->procs, * this);
354
355 // DON'T just proceed, start looking again
356 continue MAIN_LOOP;
357 }
358
359 RUN_THREAD:
360 /* paranoid */ verify( ! __preemption_enabled() );
361 /* paranoid */ verify( readyThread );
362
363 // Reset io dirty bit
364 this->io.dirty = false;
365
366 // We found a thread run it
367 __run_thread(this, readyThread);
368
369 // Are we done?
370 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
371
372 if(this->io.pending && !this->io.dirty) {
373 __cfa_io_flush( this, 0 );
374 }
375
376 ready_schedule_lock();
377 __maybe_io_drain( this );
378 ready_schedule_unlock();
379 #endif
380 }
381
382 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
383 }
384
385 for(int i = 0; !available(future); i++) {
386 if(i > 1000) __cfaabi_dbg_write( "ERROR: kernel has bin spinning on a flush after exit loop.\n", 60);
387 __cfa_io_flush( this, 1 );
388 }
389
390 __cfa_io_stop( this );
391
392 post( this->terminated );
393
394 if(this == mainProcessor) {
395 // HACK : the coroutine context switch expects this_thread to be set
396 // and it make sense for it to be set in all other cases except here
397 // fake it
398 __cfaabi_tls.this_thread = mainThread;
399 }
400
401 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
402}
403
404static int * __volatile_errno() __attribute__((noinline));
405static int * __volatile_errno() { asm(""); return &errno; }
406
407// KERNEL ONLY
408// runThread runs a thread by context switching
409// from the processor coroutine to the target thread
410static void __run_thread(processor * this, thread$ * thrd_dst) {
411 /* paranoid */ verify( ! __preemption_enabled() );
412 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
413 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
414 __builtin_prefetch( thrd_dst->context.SP );
415
416 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
417
418 coroutine$ * proc_cor = get_coroutine(this->runner);
419
420 // set state of processor coroutine to inactive
421 verify(proc_cor->state == Active);
422 proc_cor->state = Blocked;
423
424 // Actually run the thread
425 RUNNING: while(true) {
426 thrd_dst->preempted = __NO_PREEMPTION;
427 thrd_dst->state = Active;
428
429 // Update global state
430 kernelTLS().this_thread = thrd_dst;
431
432 /* paranoid */ verify( ! __preemption_enabled() );
433 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
434 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
435 /* paranoid */ verify( thrd_dst->context.SP );
436 /* paranoid */ verify( thrd_dst->state != Halted );
437 /* 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
438 /* 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
439 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
440
441
442
443 // set context switch to the thread that the processor is executing
444 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
445 // when __cfactx_switch returns we are back in the processor coroutine
446
447
448
449 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
450 /* 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 );
451 /* 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 );
452 /* paranoid */ verify( thrd_dst->context.SP );
453 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
454 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
455 /* paranoid */ verify( ! __preemption_enabled() );
456
457 // Reset global state
458 kernelTLS().this_thread = 0p;
459
460 // We just finished running a thread, there are a few things that could have happened.
461 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
462 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
463 // 4 - Preempted
464 // In case 1, we may have won a race so we can't write to the state again.
465 // In case 2, we lost the race so we now own the thread.
466
467 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
468 // The thread was preempted, reschedule it and reset the flag
469 schedule_thread$( thrd_dst, UNPARK_LOCAL );
470 break RUNNING;
471 }
472
473 if(unlikely(thrd_dst->state == Halting)) {
474 // The thread has halted, it should never be scheduled/run again
475 // finish the thread
476 __thread_finish( thrd_dst );
477 break RUNNING;
478 }
479
480 /* paranoid */ verify( thrd_dst->state == Active );
481 thrd_dst->state = Blocked;
482
483 // set state of processor coroutine to active and the thread to inactive
484 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
485 switch(old_ticket) {
486 case TICKET_RUNNING:
487 // This is case 1, the regular case, nothing more is needed
488 break RUNNING;
489 case TICKET_UNBLOCK:
490 #if !defined(__CFA_NO_STATISTICS__)
491 __tls_stats()->ready.threads.threads++;
492 #endif
493 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
494 // In this case, just run it again.
495 continue RUNNING;
496 default:
497 // This makes no sense, something is wrong abort
498 abort();
499 }
500 }
501
502 // Just before returning to the processor, set the processor coroutine to active
503 proc_cor->state = Active;
504
505 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
506
507 #if !defined(__CFA_NO_STATISTICS__)
508 __tls_stats()->ready.threads.threads--;
509 #endif
510
511 /* paranoid */ verify( ! __preemption_enabled() );
512}
513
514// KERNEL_ONLY
515void returnToKernel() {
516 /* paranoid */ verify( ! __preemption_enabled() );
517 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
518 thread$ * thrd_src = kernelTLS().this_thread;
519
520 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
521
522 // Run the thread on this processor
523 {
524 int local_errno = *__volatile_errno();
525 #if defined( __i386 ) || defined( __x86_64 )
526 __x87_store;
527 #endif
528 /* paranoid */ verify( proc_cor->context.SP );
529 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
530 __cfactx_switch( &thrd_src->context, &proc_cor->context );
531 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
532 #if defined( __i386 ) || defined( __x86_64 )
533 __x87_load;
534 #endif
535 *__volatile_errno() = local_errno;
536 }
537
538 #if !defined(__CFA_NO_STATISTICS__)
539 /* paranoid */ verify( thrd_src->last_proc != 0p );
540 if(thrd_src->last_proc != kernelTLS().this_processor) {
541 __tls_stats()->ready.threads.migration++;
542 }
543 #endif
544
545 /* paranoid */ verify( ! __preemption_enabled() );
546 /* 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 );
547 /* 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 );
548}
549
550//-----------------------------------------------------------------------------
551// Scheduler routines
552// KERNEL ONLY
553static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
554 /* paranoid */ verify( ! __preemption_enabled() );
555 /* paranoid */ verify( ready_schedule_islocked());
556 /* paranoid */ verify( thrd );
557 /* paranoid */ verify( thrd->state != Halted );
558 /* paranoid */ verify( thrd->curr_cluster );
559 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
560 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
561 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
562 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
563 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
564 /* paranoid */ #endif
565 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
566 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
567
568 const bool local = thrd->state != Start;
569 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
570
571 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
572 struct cluster * cl = thrd->curr_cluster;
573 __STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
574
575 // push the thread to the cluster ready-queue
576 push( cl, thrd, hint );
577
578 // variable thrd is no longer safe to use
579 thrd = 0xdeaddeaddeaddeadp;
580
581 // wake the cluster using the save variable.
582 __wake_one( cl );
583
584 #if !defined(__CFA_NO_STATISTICS__)
585 if( kernelTLS().this_stats ) {
586 __tls_stats()->ready.threads.threads++;
587 if(outside) {
588 __tls_stats()->ready.threads.extunpark++;
589 }
590 }
591 else {
592 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
593 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
594 }
595 #endif
596
597 /* paranoid */ verify( ready_schedule_islocked());
598 /* paranoid */ verify( ! __preemption_enabled() );
599}
600
601void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
602 ready_schedule_lock();
603 __schedule_thread( thrd, hint );
604 ready_schedule_unlock();
605}
606
607// KERNEL ONLY
608static inline thread$ * __next_thread(cluster * this) with( *this ) {
609 /* paranoid */ verify( ! __preemption_enabled() );
610
611 ready_schedule_lock();
612 thread$ * thrd = pop_fast( this );
613 ready_schedule_unlock();
614
615 /* paranoid */ verify( ! __preemption_enabled() );
616 return thrd;
617}
618
619// KERNEL ONLY
620static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
621 /* paranoid */ verify( ! __preemption_enabled() );
622
623 ready_schedule_lock();
624 thread$ * thrd;
625 for(25) {
626 thrd = pop_slow( this );
627 if(thrd) goto RET;
628 }
629 thrd = pop_search( this );
630
631 RET:
632 ready_schedule_unlock();
633
634 /* paranoid */ verify( ! __preemption_enabled() );
635 return thrd;
636}
637
638static inline bool __must_unpark( thread$ * thrd ) {
639 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
640 switch(old_ticket) {
641 case TICKET_RUNNING:
642 // Wake won the race, the thread will reschedule/rerun itself
643 return false;
644 case TICKET_BLOCKED:
645 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
646 /* paranoid */ verify( thrd->state == Blocked );
647 return true;
648 default:
649 // This makes no sense, something is wrong abort
650 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
651 }
652}
653
654void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
655 /* paranoid */ verify( ! __preemption_enabled() );
656 /* paranoid */ verify( ready_schedule_islocked());
657
658 if( !thrd ) return;
659
660 if(__must_unpark(thrd)) {
661 // Wake lost the race,
662 __schedule_thread( thrd, hint );
663 }
664
665 /* paranoid */ verify( ready_schedule_islocked());
666 /* paranoid */ verify( ! __preemption_enabled() );
667}
668
669void unpark( thread$ * thrd, unpark_hint hint ) {
670 if( !thrd ) return;
671
672 if(__must_unpark(thrd)) {
673 disable_interrupts();
674 // Wake lost the race,
675 schedule_thread$( thrd, hint );
676 enable_interrupts(false);
677 }
678}
679
680void park( void ) {
681 __disable_interrupts_checked();
682 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
683 returnToKernel();
684 __enable_interrupts_checked();
685
686}
687
688extern "C" {
689 // Leave the thread monitor
690 // last routine called by a thread.
691 // Should never return
692 void __cfactx_thrd_leave() {
693 thread$ * thrd = active_thread();
694 monitor$ * this = &thrd->self_mon;
695
696 // Lock the monitor now
697 lock( this->lock __cfaabi_dbg_ctx2 );
698
699 disable_interrupts();
700
701 /* paranoid */ verify( ! __preemption_enabled() );
702 /* paranoid */ verify( thrd->state == Active );
703 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
704 /* paranoid */ verify( kernelTLS().this_thread == thrd );
705 /* paranoid */ verify( thrd->context.SP );
706 /* 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 );
707 /* 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 );
708
709 thrd->state = Halting;
710 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
711 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
712 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
713
714 // Leave the thread
715 returnToKernel();
716
717 // Control flow should never reach here!
718 abort();
719 }
720}
721
722// KERNEL ONLY
723bool force_yield( __Preemption_Reason reason ) {
724 __disable_interrupts_checked();
725 thread$ * thrd = kernelTLS().this_thread;
726 /* paranoid */ verify(thrd->state == Active);
727
728 // SKULLDUGGERY: It is possible that we are preempting this thread just before
729 // it was going to park itself. If that is the case and it is already using the
730 // intrusive fields then we can't use them to preempt the thread
731 // If that is the case, abandon the preemption.
732 bool preempted = false;
733 if(thrd->link.next == 0p) {
734 preempted = true;
735 thrd->preempted = reason;
736 returnToKernel();
737 }
738 __enable_interrupts_checked( false );
739 return preempted;
740}
741
742//=============================================================================================
743// Kernel Idle Sleep
744//=============================================================================================
745// Wake a thread from the front if there are any
746static void __wake_one(cluster * this) {
747 /* paranoid */ verify( ! __preemption_enabled() );
748 /* paranoid */ verify( ready_schedule_islocked() );
749
750 // Check if there is a sleeping processor
751 int fd = __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST);
752
753 // If no one is sleeping, we are done
754 if( fd == 0 ) return;
755
756 // We found a processor, wake it up
757 eventfd_t val;
758 val = 1;
759 eventfd_write( fd, val );
760
761 #if !defined(__CFA_NO_STATISTICS__)
762 if( kernelTLS().this_stats ) {
763 __tls_stats()->ready.sleep.wakes++;
764 }
765 else {
766 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
767 }
768 #endif
769
770 /* paranoid */ verify( ready_schedule_islocked() );
771 /* paranoid */ verify( ! __preemption_enabled() );
772
773 return;
774}
775
776// Unconditionnaly wake a thread
777void __wake_proc(processor * this) {
778 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
779
780 __disable_interrupts_checked();
781 /* paranoid */ verify( ! __preemption_enabled() );
782 eventfd_t val;
783 val = 1;
784 eventfd_write( this->idle_fd, val );
785 __enable_interrupts_checked();
786}
787
788static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
789 #if !defined(CFA_WITH_IO_URING_IDLE)
790 #if !defined(__CFA_NO_STATISTICS__)
791 if(this->print_halts) {
792 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
793 }
794 #endif
795
796 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
797
798 {
799 eventfd_t val;
800 ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
801 if(ret < 0) {
802 switch((int)errno) {
803 case EAGAIN:
804 #if EAGAIN != EWOULDBLOCK
805 case EWOULDBLOCK:
806 #endif
807 case EINTR:
808 // No need to do anything special here, just assume it's a legitimate wake-up
809 break;
810 default:
811 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
812 }
813 }
814 }
815
816 #if !defined(__CFA_NO_STATISTICS__)
817 if(this->print_halts) {
818 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
819 }
820 #endif
821 #else
822 // Do we already have a pending read
823 if(available(future)) {
824 // There is no pending read, we need to add one
825 reset(future);
826
827 __kernel_read(this, future, iov, this->idle_fd );
828 }
829
830 __cfa_io_flush( this, 1 );
831 #endif
832}
833
834static bool mark_idle(__cluster_proc_list & this, processor & proc) {
835 /* paranoid */ verify( ! __preemption_enabled() );
836 if(!try_lock( this )) return false;
837 this.idle++;
838 /* paranoid */ verify( this.idle <= this.total );
839 remove(proc);
840 insert_first(this.idles, proc);
841
842 __atomic_store_n(&this.fd, proc.idle_fd, __ATOMIC_SEQ_CST);
843 unlock( this );
844 /* paranoid */ verify( ! __preemption_enabled() );
845
846 return true;
847}
848
849static void mark_awake(__cluster_proc_list & this, processor & proc) {
850 /* paranoid */ verify( ! __preemption_enabled() );
851 lock( this );
852 this.idle--;
853 /* paranoid */ verify( this.idle >= 0 );
854 remove(proc);
855 insert_last(this.actives, proc);
856
857 {
858 int fd = 0;
859 if(!this.idles`isEmpty) fd = this.idles`first.idle_fd;
860 __atomic_store_n(&this.fd, fd, __ATOMIC_SEQ_CST);
861 }
862
863 unlock( this );
864 /* paranoid */ verify( ! __preemption_enabled() );
865}
866
867//=============================================================================================
868// Unexpected Terminating logic
869//=============================================================================================
870void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
871 thread$ * thrd = __cfaabi_tls.this_thread;
872
873 if(thrd) {
874 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
875 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
876
877 if ( &thrd->self_cor != thrd->curr_cor ) {
878 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
879 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
880 }
881 else {
882 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
883 }
884 }
885 else {
886 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
887 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
888 }
889}
890
891int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
892 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
893}
894
895static __spinlock_t kernel_debug_lock;
896
897extern "C" {
898 void __cfaabi_bits_acquire() {
899 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
900 }
901
902 void __cfaabi_bits_release() {
903 unlock( kernel_debug_lock );
904 }
905}
906
907//=============================================================================================
908// Kernel Utilities
909//=============================================================================================
910#if defined(CFA_HAVE_LINUX_IO_URING_H)
911#include "io/types.hfa"
912#endif
913
914static inline bool __maybe_io_drain( processor * proc ) {
915 bool ret = false;
916 #if defined(CFA_HAVE_LINUX_IO_URING_H)
917 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
918
919 // Check if we should drain the queue
920 $io_context * ctx = proc->io.ctx;
921 unsigned head = *ctx->cq.head;
922 unsigned tail = *ctx->cq.tail;
923 if(head == tail) return false;
924 #if OLD_MAIN
925 ready_schedule_lock();
926 ret = __cfa_io_drain( proc );
927 ready_schedule_unlock();
928 #else
929 ret = __cfa_io_drain( proc );
930 #endif
931 #endif
932 return ret;
933}
934
935//-----------------------------------------------------------------------------
936// Debug
937__cfaabi_dbg_debug_do(
938 extern "C" {
939 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
940 this.prev_name = prev_name;
941 this.prev_thrd = kernelTLS().this_thread;
942 }
943 }
944)
945
946//-----------------------------------------------------------------------------
947// Debug
948bool threading_enabled(void) __attribute__((const)) {
949 return true;
950}
951
952//-----------------------------------------------------------------------------
953// Statistics
954#if !defined(__CFA_NO_STATISTICS__)
955 void print_halts( processor & this ) {
956 this.print_halts = true;
957 }
958
959 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
960 /* paranoid */ verify( cltr->stats );
961
962 processor * it = &list`first;
963 for(unsigned i = 0; i < count; i++) {
964 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
965 /* paranoid */ verify( it->local_data->this_stats );
966 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
967 __tally_stats( cltr->stats, it->local_data->this_stats );
968 it = &(*it)`next;
969 }
970 }
971
972 void crawl_cluster_stats( cluster & this ) {
973 // Stop the world, otherwise stats could get really messed-up
974 // this doesn't solve all problems but does solve many
975 // so it's probably good enough
976 disable_interrupts();
977 uint_fast32_t last_size = ready_mutate_lock();
978
979 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
980 crawl_list(&this, this.procs.idles , this.procs.idle );
981
982 // Unlock the RWlock
983 ready_mutate_unlock( last_size );
984 enable_interrupts();
985 }
986
987
988 void print_stats_now( cluster & this, int flags ) {
989 crawl_cluster_stats( this );
990 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
991 }
992#endif
993// Local Variables: //
994// mode: c //
995// tab-width: 4 //
996// End: //
Note: See TracBrowser for help on using the repository browser.