source: libcfa/src/concurrency/kernel.cfa@ 7cf8006

ADT ast-experimental
Last change on this file since 7cf8006 was 15c93d8, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Renamed ready-queue link fields to rdy_link
  • Property mode set to 100644
File size: 29.0 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->rdy_link.next == 0p, "Expected null got %p", thrd_dst->rdy_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
308 // Update global state
309 kernelTLS().this_thread = thrd_dst;
310
311 // Update the state after setting this_thread
312 // so that the debugger can find all active threads
313 // in tls storage
314 thrd_dst->state = Active;
315
316 /* paranoid */ verify( ! __preemption_enabled() );
317 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
318 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
319 /* paranoid */ verify( thrd_dst->context.SP );
320 /* paranoid */ verify( thrd_dst->state != Halted );
321 /* 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
322 /* 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
323 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
324 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
325
326
327
328 // set context switch to the thread that the processor is executing
329 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
330 // when __cfactx_switch returns we are back in the processor coroutine
331
332
333
334 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
335 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
336 /* 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 );
337 /* 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 );
338 /* paranoid */ verify( thrd_dst->state != Halted );
339 /* paranoid */ verify( thrd_dst->context.SP );
340 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
341 /* paranoid */ verify( ! __preemption_enabled() );
342
343 // We just finished running a thread, there are a few things that could have happened.
344 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
345 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
346 // 4 - Preempted
347 // In case 1, we may have won a race so we can't write to the state again.
348 // In case 2, we lost the race so we now own the thread.
349
350 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
351 // Reset the this_thread now that we know
352 // the state isn't active anymore
353 kernelTLS().this_thread = 0p;
354
355 // The thread was preempted, reschedule it and reset the flag
356 schedule_thread$( thrd_dst, UNPARK_LOCAL );
357 break RUNNING;
358 }
359
360 if(unlikely(thrd_dst->state == Halting)) {
361 // Reset the this_thread now that we know
362 // the state isn't active anymore
363 kernelTLS().this_thread = 0p;
364
365 // The thread has halted, it should never be scheduled/run again
366 // finish the thread
367 __thread_finish( thrd_dst );
368 break RUNNING;
369 }
370
371 /* paranoid */ verify( thrd_dst->state == Active );
372 thrd_dst->state = Blocked;
373
374 // Reset the this_thread now that we know
375 // the state isn't active anymore
376 kernelTLS().this_thread = 0p;
377
378 // set state of processor coroutine to active and the thread to inactive
379 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
380 switch(old_ticket) {
381 case TICKET_RUNNING:
382 // This is case 1, the regular case, nothing more is needed
383 break RUNNING;
384 case TICKET_UNBLOCK:
385 __STATS__(true, ready.threads.threads++; )
386 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
387 // In this case, just run it again.
388 continue RUNNING;
389 default:
390 // This makes no sense, something is wrong abort
391 abort();
392 }
393 }
394
395 // Just before returning to the processor, set the processor coroutine to active
396 proc_cor->state = Active;
397
398 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
399
400 __STATS__(true, ready.threads.threads--; )
401
402 /* paranoid */ verify( ! __preemption_enabled() );
403}
404
405// KERNEL_ONLY
406static void returnToKernel() {
407 /* paranoid */ verify( ! __preemption_enabled() );
408 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
409 thread$ * thrd_src = kernelTLS().this_thread;
410
411 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
412
413 // Run the thread on this processor
414 {
415 int local_errno = *__volatile_errno();
416 #if defined( __i386 ) || defined( __x86_64 )
417 __x87_store;
418 #endif
419 /* paranoid */ verify( proc_cor->context.SP );
420 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
421 __cfactx_switch( &thrd_src->context, &proc_cor->context );
422 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
423 #if defined( __i386 ) || defined( __x86_64 )
424 __x87_load;
425 #endif
426 *__volatile_errno() = local_errno;
427 }
428
429 #if !defined(__CFA_NO_STATISTICS__)
430 /* paranoid */ verify( thrd_src->last_proc != 0p );
431 if(thrd_src->last_proc != kernelTLS().this_processor) {
432 __tls_stats()->ready.threads.migration++;
433 }
434 #endif
435
436 /* paranoid */ verify( ! __preemption_enabled() );
437 /* 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 );
438 /* 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 );
439}
440
441//-----------------------------------------------------------------------------
442// Scheduler routines
443// KERNEL ONLY
444static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
445 /* paranoid */ verify( ! __preemption_enabled() );
446 /* paranoid */ verify( ready_schedule_islocked());
447 /* paranoid */ verify( thrd );
448 /* paranoid */ verify( thrd->state != Halted );
449 /* paranoid */ verify( thrd->curr_cluster );
450 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
451 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
452 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
453 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
454 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
455 /* paranoid */ #endif
456 /* paranoid */ verifyf( thrd->rdy_link.next == 0p, "Expected null got %p", thrd->rdy_link.next );
457 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
458
459 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
460
461 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
462 struct cluster * cl = thrd->curr_cluster;
463 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
464
465 // push the thread to the cluster ready-queue
466 push( cl, thrd, hint );
467
468 // variable thrd is no longer safe to use
469 thrd = 0xdeaddeaddeaddeadp;
470
471 // wake the cluster using the save variable.
472 __wake_one( cl );
473
474 #if !defined(__CFA_NO_STATISTICS__)
475 if( kernelTLS().this_stats ) {
476 __tls_stats()->ready.threads.threads++;
477 if(outside) {
478 __tls_stats()->ready.threads.extunpark++;
479 }
480 }
481 else {
482 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
483 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
484 }
485 #endif
486
487 /* paranoid */ verify( ready_schedule_islocked());
488 /* paranoid */ verify( ! __preemption_enabled() );
489}
490
491void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
492 ready_schedule_lock();
493 __schedule_thread( thrd, hint );
494 ready_schedule_unlock();
495}
496
497// KERNEL ONLY
498static inline thread$ * __next_thread(cluster * this) with( *this ) {
499 /* paranoid */ verify( ! __preemption_enabled() );
500
501 ready_schedule_lock();
502 thread$ * thrd = pop_fast( this );
503 ready_schedule_unlock();
504
505 /* paranoid */ verify( ! __preemption_enabled() );
506 return thrd;
507}
508
509// KERNEL ONLY
510static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
511 /* paranoid */ verify( ! __preemption_enabled() );
512
513 ready_schedule_lock();
514 thread$ * thrd = pop_slow( this );
515 ready_schedule_unlock();
516
517 /* paranoid */ verify( ! __preemption_enabled() );
518 return thrd;
519}
520
521// KERNEL ONLY
522static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
523 /* paranoid */ verify( ! __preemption_enabled() );
524
525 ready_schedule_lock();
526 thread$ * thrd = pop_search( this );
527 ready_schedule_unlock();
528
529 /* paranoid */ verify( ! __preemption_enabled() );
530 return thrd;
531}
532
533static inline bool __must_unpark( thread$ * thrd ) {
534 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
535 switch(old_ticket) {
536 case TICKET_RUNNING:
537 // Wake won the race, the thread will reschedule/rerun itself
538 return false;
539 case TICKET_BLOCKED:
540 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
541 /* paranoid */ verify( thrd->state == Blocked );
542 return true;
543 default:
544 // This makes no sense, something is wrong abort
545 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
546 }
547}
548
549void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
550 /* paranoid */ verify( ! __preemption_enabled() );
551 /* paranoid */ verify( ready_schedule_islocked());
552
553 if( !thrd ) return;
554
555 if(__must_unpark(thrd)) {
556 // Wake lost the race,
557 __schedule_thread( thrd, hint );
558 }
559
560 /* paranoid */ verify( ready_schedule_islocked());
561 /* paranoid */ verify( ! __preemption_enabled() );
562}
563
564void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
565 if( !thrd ) return;
566
567 if(__must_unpark(thrd)) {
568 disable_interrupts();
569 // Wake lost the race,
570 schedule_thread$( thrd, hint );
571 enable_interrupts(false);
572 }
573}
574
575void park( void ) libcfa_public {
576 __disable_interrupts_checked();
577 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
578 returnToKernel();
579 __enable_interrupts_checked();
580
581}
582
583extern "C" {
584 // Leave the thread monitor
585 // last routine called by a thread.
586 // Should never return
587 void __cfactx_thrd_leave() {
588 thread$ * thrd = active_thread();
589 monitor$ * this = &thrd->self_mon;
590
591 // Lock the monitor now
592 lock( this->lock __cfaabi_dbg_ctx2 );
593
594 disable_interrupts();
595
596 /* paranoid */ verify( ! __preemption_enabled() );
597 /* paranoid */ verify( thrd->state == Active );
598 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
599 /* paranoid */ verify( kernelTLS().this_thread == thrd );
600 /* paranoid */ verify( thrd->context.SP );
601 /* 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 );
602 /* 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 );
603
604 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
605 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
606 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
607
608 thrd->state = Halting;
609 thrd->ticket = TICKET_DEAD;
610
611 // Leave the thread
612 returnToKernel();
613
614 // Control flow should never reach here!
615 abort();
616 }
617}
618
619// KERNEL ONLY
620bool force_yield( __Preemption_Reason reason ) libcfa_public {
621 __disable_interrupts_checked();
622 thread$ * thrd = kernelTLS().this_thread;
623 /* paranoid */ verify(thrd->state == Active);
624
625 // SKULLDUGGERY: It is possible that we are preempting this thread just before
626 // it was going to park itself. If that is the case and it is already using the
627 // intrusive fields then we can't use them to preempt the thread
628 // If that is the case, abandon the preemption.
629 bool preempted = false;
630 if(thrd->rdy_link.next == 0p) {
631 preempted = true;
632 thrd->preempted = reason;
633 returnToKernel();
634 }
635 __enable_interrupts_checked( false );
636 return preempted;
637}
638
639//=============================================================================================
640// Kernel Idle Sleep
641//=============================================================================================
642// Wake a thread from the front if there are any
643static void __wake_one(cluster * this) {
644 eventfd_t val;
645
646 /* paranoid */ verify( ! __preemption_enabled() );
647 /* paranoid */ verify( ready_schedule_islocked() );
648
649 // Check if there is a sleeping processor
650 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
651
652 // If no one is sleeping: we are done
653 if( fdp == 0p ) return;
654
655 int fd = 1;
656 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
657 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
658 }
659
660 switch(fd) {
661 __attribute__((unused)) int ret;
662 case 0:
663 // If the processor isn't ready to sleep then the exchange will already wake it up
664 #if !defined(__CFA_NO_STATISTICS__)
665 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
666 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
667 #endif
668 break;
669 case 1:
670 // If someone else already said they will wake them: we are done
671 #if !defined(__CFA_NO_STATISTICS__)
672 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
673 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
674 #endif
675 break;
676 default:
677 // If the processor was ready to sleep, we need to wake it up with an actual write
678 val = 1;
679 ret = eventfd_write( fd, val );
680 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
681
682 #if !defined(__CFA_NO_STATISTICS__)
683 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
684 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
685 #endif
686 break;
687 }
688
689 /* paranoid */ verify( ready_schedule_islocked() );
690 /* paranoid */ verify( ! __preemption_enabled() );
691
692 return;
693}
694
695// Unconditionnaly wake a thread
696void __wake_proc(processor * this) {
697 /* paranoid */ verify( ! __preemption_enabled() );
698
699 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
700
701 this->idle_wctx.sem = 1;
702
703 this->idle_wctx.wake__time = rdtscl();
704
705 eventfd_t val;
706 val = 1;
707 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
708
709 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
710 /* paranoid */ verify( ! __preemption_enabled() );
711}
712
713static void idle_sleep(processor * this) {
714 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
715 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
716
717 // Tell everyone we are ready to go do sleep
718 for() {
719 int expected = this->idle_wctx.sem;
720
721 // Someone already told us to wake-up! No time for a nap.
722 if(expected == 1) { return; }
723
724 // Try to mark that we are going to sleep
725 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
726 // Every one agreed, taking a nap
727 break;
728 }
729 }
730
731
732 #if !defined(CFA_WITH_IO_URING_IDLE)
733 #if !defined(__CFA_NO_STATISTICS__)
734 if(this->print_halts) {
735 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
736 }
737 #endif
738
739 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
740
741 {
742 eventfd_t val;
743 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
744 if(ret < 0) {
745 switch((int)errno) {
746 case EAGAIN:
747 #if EAGAIN != EWOULDBLOCK
748 case EWOULDBLOCK:
749 #endif
750 case EINTR:
751 // No need to do anything special here, just assume it's a legitimate wake-up
752 break;
753 default:
754 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
755 }
756 }
757 }
758
759 #if !defined(__CFA_NO_STATISTICS__)
760 if(this->print_halts) {
761 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
762 }
763 #endif
764 #else
765 __cfa_io_idle( this );
766 #endif
767}
768
769static bool mark_idle(__cluster_proc_list & this, processor & proc) {
770 __STATS__(true, ready.sleep.halts++; )
771
772 proc.idle_wctx.sem = 0;
773
774 /* paranoid */ verify( ! __preemption_enabled() );
775 if(!try_lock( this )) return false;
776 this.idle++;
777 /* paranoid */ verify( this.idle <= this.total );
778 remove(proc);
779 insert_first(this.idles, proc);
780
781 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
782 unlock( this );
783 /* paranoid */ verify( ! __preemption_enabled() );
784
785 return true;
786}
787
788static void mark_awake(__cluster_proc_list & this, processor & proc) {
789 /* paranoid */ verify( ! __preemption_enabled() );
790 lock( this );
791 this.idle--;
792 /* paranoid */ verify( this.idle >= 0 );
793 remove(proc);
794 insert_last(this.actives, proc);
795
796 {
797 struct __fd_waitctx * wctx = 0;
798 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
799 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
800 }
801
802 unlock( this );
803 /* paranoid */ verify( ! __preemption_enabled() );
804}
805
806//=============================================================================================
807// Unexpected Terminating logic
808//=============================================================================================
809void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
810 thread$ * thrd = __cfaabi_tls.this_thread;
811
812 if(thrd) {
813 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
814 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
815
816 if ( &thrd->self_cor != thrd->curr_cor ) {
817 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
818 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
819 }
820 else {
821 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
822 }
823 }
824 else {
825 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
826 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
827 }
828}
829
830int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
831 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
832}
833
834static __spinlock_t kernel_debug_lock;
835
836extern "C" {
837 void __cfaabi_bits_acquire() {
838 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
839 }
840
841 void __cfaabi_bits_release() {
842 unlock( kernel_debug_lock );
843 }
844}
845
846//=============================================================================================
847// Kernel Utilities
848//=============================================================================================
849#if defined(CFA_HAVE_LINUX_IO_URING_H)
850#include "io/types.hfa"
851#endif
852
853//-----------------------------------------------------------------------------
854// Debug
855bool threading_enabled(void) __attribute__((const)) libcfa_public {
856 return true;
857}
858
859//-----------------------------------------------------------------------------
860// Statistics
861#if !defined(__CFA_NO_STATISTICS__)
862 void print_halts( processor & this ) libcfa_public {
863 this.print_halts = true;
864 }
865
866 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
867 /* paranoid */ verify( cltr->stats );
868
869 processor * it = &list`first;
870 for(unsigned i = 0; i < count; i++) {
871 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
872 /* paranoid */ verify( it->local_data->this_stats );
873 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
874 __tally_stats( cltr->stats, it->local_data->this_stats );
875 it = &(*it)`next;
876 }
877 }
878
879 static void crawl_cluster_stats( cluster & this ) {
880 // Stop the world, otherwise stats could get really messed-up
881 // this doesn't solve all problems but does solve many
882 // so it's probably good enough
883 disable_interrupts();
884 uint_fast32_t last_size = ready_mutate_lock();
885
886 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
887 crawl_list(&this, this.procs.idles , this.procs.idle );
888
889 // Unlock the RWlock
890 ready_mutate_unlock( last_size );
891 enable_interrupts();
892 }
893
894
895 void print_stats_now( cluster & this, int flags ) libcfa_public {
896 crawl_cluster_stats( this );
897 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
898 }
899#endif
900// Local Variables: //
901// mode: c //
902// tab-width: 4 //
903// End: //
Note: See TracBrowser for help on using the repository browser.