source: libcfa/src/concurrency/kernel.cfa@ d672350

ADT ast-experimental enum pthread-emulation qualifiedEnum
Last change on this file since d672350 was 4ecc35a, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added spin lock to io drain.
last step before completion fairness
  • Property mode set to 100644
File size: 28.7 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// kernel.c --
8//
9// Author : Thierry Delisle
10// Created On : Tue Jan 17 12:27:26 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : 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, io_future_t & future, iovec & iov);
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( $io_context * );
139extern bool __cfa_io_flush( processor *, int min_comp );
140static inline bool __maybe_io_drain( processor * );
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 iovec idle_iovec = { this->idle_wctx.rdbuf, sizeof(eventfd_t) };
171
172 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
173 #if !defined(__CFA_NO_STATISTICS__)
174 if( this->print_halts ) {
175 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
176 }
177 #endif
178
179 {
180 // Setup preemption data
181 preemption_scope scope = { this };
182
183 // if we need to run some special setup, now is the time to do it.
184 if(this->init.thrd) {
185 this->init.thrd->curr_cluster = this->cltr;
186 __run_thread(this, this->init.thrd);
187 }
188
189 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
190
191 thread$ * readyThread = 0p;
192 MAIN_LOOP:
193 for() {
194 // Check if there is pending io
195 __maybe_io_drain( this );
196
197 // Try to get the next thread
198 readyThread = __next_thread( this->cltr );
199
200 if( !readyThread ) {
201 __IO_STATS__(true, io.flush.idle++; )
202 __cfa_io_flush( this, 0 );
203
204 readyThread = __next_thread( this->cltr );
205 }
206
207 if( !readyThread ) for(5) {
208 __IO_STATS__(true, io.flush.idle++; )
209
210 readyThread = __next_thread_slow( this->cltr );
211
212 if( readyThread ) break;
213
214 __cfa_io_flush( this, 0 );
215 }
216
217 HALT:
218 if( !readyThread ) {
219 // Don't block if we are done
220 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
221
222 // Push self to idle stack
223 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
224
225 // Confirm the ready-queue is empty
226 readyThread = __next_thread_search( this->cltr );
227 if( readyThread ) {
228 // A thread was found, cancel the halt
229 mark_awake(this->cltr->procs, * this);
230
231 __STATS__(true, ready.sleep.cancels++; )
232
233 // continue the mai loop
234 break HALT;
235 }
236
237 idle_sleep( this, *this->idle_wctx.ftr, idle_iovec );
238
239 // We were woken up, remove self from idle
240 mark_awake(this->cltr->procs, * this);
241
242 // DON'T just proceed, start looking again
243 continue MAIN_LOOP;
244 }
245
246 /* paranoid */ verify( readyThread );
247
248 // Reset io dirty bit
249 this->io.dirty = false;
250
251 // We found a thread run it
252 __run_thread(this, readyThread);
253
254 // Are we done?
255 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
256
257 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
258 __IO_STATS__(true, io.flush.dirty++; )
259 __cfa_io_flush( this, 0 );
260 }
261 }
262
263 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
264 }
265
266 post( this->terminated );
267
268 if(this == mainProcessor) {
269 // HACK : the coroutine context switch expects this_thread to be set
270 // and it make sense for it to be set in all other cases except here
271 // fake it
272 __cfaabi_tls.this_thread = mainThread;
273 }
274
275 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
276}
277
278static int * __volatile_errno() __attribute__((noinline));
279static int * __volatile_errno() { asm(""); return &errno; }
280
281// KERNEL ONLY
282// runThread runs a thread by context switching
283// from the processor coroutine to the target thread
284static void __run_thread(processor * this, thread$ * thrd_dst) {
285 /* paranoid */ verify( ! __preemption_enabled() );
286 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
287 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
288 __builtin_prefetch( thrd_dst->context.SP );
289
290 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
291
292 coroutine$ * proc_cor = get_coroutine(this->runner);
293
294 // set state of processor coroutine to inactive
295 verify(proc_cor->state == Active);
296 proc_cor->state = Blocked;
297
298 // Actually run the thread
299 RUNNING: while(true) {
300 thrd_dst->preempted = __NO_PREEMPTION;
301 thrd_dst->state = Active;
302
303 // Update global state
304 kernelTLS().this_thread = thrd_dst;
305
306 /* paranoid */ verify( ! __preemption_enabled() );
307 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
308 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
309 /* paranoid */ verify( thrd_dst->context.SP );
310 /* paranoid */ verify( thrd_dst->state != Halted );
311 /* 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
312 /* 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
313 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
314
315
316
317 // set context switch to the thread that the processor is executing
318 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
319 // when __cfactx_switch returns we are back in the processor coroutine
320
321
322
323 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
324 /* 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 );
325 /* 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 );
326 /* paranoid */ verify( thrd_dst->context.SP );
327 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
328 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
329 /* paranoid */ verify( ! __preemption_enabled() );
330
331 // Reset global state
332 kernelTLS().this_thread = 0p;
333
334 // We just finished running a thread, there are a few things that could have happened.
335 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
336 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
337 // 4 - Preempted
338 // In case 1, we may have won a race so we can't write to the state again.
339 // In case 2, we lost the race so we now own the thread.
340
341 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
342 // The thread was preempted, reschedule it and reset the flag
343 schedule_thread$( thrd_dst, UNPARK_LOCAL );
344 break RUNNING;
345 }
346
347 if(unlikely(thrd_dst->state == Halting)) {
348 // The thread has halted, it should never be scheduled/run again
349 // finish the thread
350 __thread_finish( thrd_dst );
351 break RUNNING;
352 }
353
354 /* paranoid */ verify( thrd_dst->state == Active );
355 thrd_dst->state = Blocked;
356
357 // set state of processor coroutine to active and the thread to inactive
358 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
359 switch(old_ticket) {
360 case TICKET_RUNNING:
361 // This is case 1, the regular case, nothing more is needed
362 break RUNNING;
363 case TICKET_UNBLOCK:
364 __STATS__(true, ready.threads.threads++; )
365 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
366 // In this case, just run it again.
367 continue RUNNING;
368 default:
369 // This makes no sense, something is wrong abort
370 abort();
371 }
372 }
373
374 // Just before returning to the processor, set the processor coroutine to active
375 proc_cor->state = Active;
376
377 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
378
379 __STATS__(true, ready.threads.threads--; )
380
381 /* paranoid */ verify( ! __preemption_enabled() );
382}
383
384// KERNEL_ONLY
385void returnToKernel() {
386 /* paranoid */ verify( ! __preemption_enabled() );
387 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
388 thread$ * thrd_src = kernelTLS().this_thread;
389
390 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
391
392 // Run the thread on this processor
393 {
394 int local_errno = *__volatile_errno();
395 #if defined( __i386 ) || defined( __x86_64 )
396 __x87_store;
397 #endif
398 /* paranoid */ verify( proc_cor->context.SP );
399 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
400 __cfactx_switch( &thrd_src->context, &proc_cor->context );
401 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
402 #if defined( __i386 ) || defined( __x86_64 )
403 __x87_load;
404 #endif
405 *__volatile_errno() = local_errno;
406 }
407
408 #if !defined(__CFA_NO_STATISTICS__)
409 /* paranoid */ verify( thrd_src->last_proc != 0p );
410 if(thrd_src->last_proc != kernelTLS().this_processor) {
411 __tls_stats()->ready.threads.migration++;
412 }
413 #endif
414
415 /* paranoid */ verify( ! __preemption_enabled() );
416 /* 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 );
417 /* 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 );
418}
419
420//-----------------------------------------------------------------------------
421// Scheduler routines
422// KERNEL ONLY
423static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
424 /* paranoid */ verify( ! __preemption_enabled() );
425 /* paranoid */ verify( ready_schedule_islocked());
426 /* paranoid */ verify( thrd );
427 /* paranoid */ verify( thrd->state != Halted );
428 /* paranoid */ verify( thrd->curr_cluster );
429 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
430 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
431 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
432 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
433 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
434 /* paranoid */ #endif
435 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
436 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
437
438 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
439
440 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
441 struct cluster * cl = thrd->curr_cluster;
442 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
443
444 // push the thread to the cluster ready-queue
445 push( cl, thrd, hint );
446
447 // variable thrd is no longer safe to use
448 thrd = 0xdeaddeaddeaddeadp;
449
450 // wake the cluster using the save variable.
451 __wake_one( cl );
452
453 #if !defined(__CFA_NO_STATISTICS__)
454 if( kernelTLS().this_stats ) {
455 __tls_stats()->ready.threads.threads++;
456 if(outside) {
457 __tls_stats()->ready.threads.extunpark++;
458 }
459 }
460 else {
461 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
462 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
463 }
464 #endif
465
466 /* paranoid */ verify( ready_schedule_islocked());
467 /* paranoid */ verify( ! __preemption_enabled() );
468}
469
470void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
471 ready_schedule_lock();
472 __schedule_thread( thrd, hint );
473 ready_schedule_unlock();
474}
475
476// KERNEL ONLY
477static inline thread$ * __next_thread(cluster * this) with( *this ) {
478 /* paranoid */ verify( ! __preemption_enabled() );
479
480 ready_schedule_lock();
481 thread$ * thrd = pop_fast( this );
482 ready_schedule_unlock();
483
484 /* paranoid */ verify( ! __preemption_enabled() );
485 return thrd;
486}
487
488// KERNEL ONLY
489static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
490 /* paranoid */ verify( ! __preemption_enabled() );
491
492 ready_schedule_lock();
493 thread$ * thrd = pop_slow( this );
494 ready_schedule_unlock();
495
496 /* paranoid */ verify( ! __preemption_enabled() );
497 return thrd;
498}
499
500// KERNEL ONLY
501static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
502 /* paranoid */ verify( ! __preemption_enabled() );
503
504 ready_schedule_lock();
505 thread$ * thrd = pop_search( this );
506 ready_schedule_unlock();
507
508 /* paranoid */ verify( ! __preemption_enabled() );
509 return thrd;
510}
511
512static inline bool __must_unpark( thread$ * thrd ) {
513 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
514 switch(old_ticket) {
515 case TICKET_RUNNING:
516 // Wake won the race, the thread will reschedule/rerun itself
517 return false;
518 case TICKET_BLOCKED:
519 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
520 /* paranoid */ verify( thrd->state == Blocked );
521 return true;
522 default:
523 // This makes no sense, something is wrong abort
524 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
525 }
526}
527
528void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
529 /* paranoid */ verify( ! __preemption_enabled() );
530 /* paranoid */ verify( ready_schedule_islocked());
531
532 if( !thrd ) return;
533
534 if(__must_unpark(thrd)) {
535 // Wake lost the race,
536 __schedule_thread( thrd, hint );
537 }
538
539 /* paranoid */ verify( ready_schedule_islocked());
540 /* paranoid */ verify( ! __preemption_enabled() );
541}
542
543void unpark( thread$ * thrd, unpark_hint hint ) {
544 if( !thrd ) return;
545
546 if(__must_unpark(thrd)) {
547 disable_interrupts();
548 // Wake lost the race,
549 schedule_thread$( thrd, hint );
550 enable_interrupts(false);
551 }
552}
553
554void park( void ) {
555 __disable_interrupts_checked();
556 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
557 returnToKernel();
558 __enable_interrupts_checked();
559
560}
561
562extern "C" {
563 // Leave the thread monitor
564 // last routine called by a thread.
565 // Should never return
566 void __cfactx_thrd_leave() {
567 thread$ * thrd = active_thread();
568 monitor$ * this = &thrd->self_mon;
569
570 // Lock the monitor now
571 lock( this->lock __cfaabi_dbg_ctx2 );
572
573 disable_interrupts();
574
575 /* paranoid */ verify( ! __preemption_enabled() );
576 /* paranoid */ verify( thrd->state == Active );
577 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
578 /* paranoid */ verify( kernelTLS().this_thread == thrd );
579 /* paranoid */ verify( thrd->context.SP );
580 /* 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 );
581 /* 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 );
582
583 thrd->state = Halting;
584 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
585 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
586 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
587
588 // Leave the thread
589 returnToKernel();
590
591 // Control flow should never reach here!
592 abort();
593 }
594}
595
596// KERNEL ONLY
597bool force_yield( __Preemption_Reason reason ) {
598 __disable_interrupts_checked();
599 thread$ * thrd = kernelTLS().this_thread;
600 /* paranoid */ verify(thrd->state == Active);
601
602 // SKULLDUGGERY: It is possible that we are preempting this thread just before
603 // it was going to park itself. If that is the case and it is already using the
604 // intrusive fields then we can't use them to preempt the thread
605 // If that is the case, abandon the preemption.
606 bool preempted = false;
607 if(thrd->link.next == 0p) {
608 preempted = true;
609 thrd->preempted = reason;
610 returnToKernel();
611 }
612 __enable_interrupts_checked( false );
613 return preempted;
614}
615
616//=============================================================================================
617// Kernel Idle Sleep
618//=============================================================================================
619// Wake a thread from the front if there are any
620static void __wake_one(cluster * this) {
621 eventfd_t val;
622
623 /* paranoid */ verify( ! __preemption_enabled() );
624 /* paranoid */ verify( ready_schedule_islocked() );
625
626 // Check if there is a sleeping processor
627 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
628
629 // If no one is sleeping: we are done
630 if( fdp == 0p ) return;
631
632 int fd = 1;
633 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
634 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
635 }
636
637 switch(fd) {
638 case 0:
639 // If the processor isn't ready to sleep then the exchange will already wake it up
640 #if !defined(__CFA_NO_STATISTICS__)
641 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
642 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
643 #endif
644 break;
645 case 1:
646 // If someone else already said they will wake them: we are done
647 #if !defined(__CFA_NO_STATISTICS__)
648 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
649 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
650 #endif
651 break;
652 default:
653 // If the processor was ready to sleep, we need to wake it up with an actual write
654 val = 1;
655 eventfd_write( fd, val );
656
657 #if !defined(__CFA_NO_STATISTICS__)
658 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
659 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
660 #endif
661 break;
662 }
663
664 /* paranoid */ verify( ready_schedule_islocked() );
665 /* paranoid */ verify( ! __preemption_enabled() );
666
667 return;
668}
669
670// Unconditionnaly wake a thread
671void __wake_proc(processor * this) {
672 /* paranoid */ verify( ! __preemption_enabled() );
673
674 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
675
676 this->idle_wctx.sem = 1;
677
678 eventfd_t val;
679 val = 1;
680 eventfd_write( this->idle_wctx.evfd, val );
681
682 /* paranoid */ verify( ! __preemption_enabled() );
683}
684
685static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
686 // Tell everyone we are ready to go do sleep
687 for() {
688 int expected = this->idle_wctx.sem;
689
690 // Someone already told us to wake-up! No time for a nap.
691 if(expected == 1) { return; }
692
693 // Try to mark that we are going to sleep
694 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
695 // Every one agreed, taking a nap
696 break;
697 }
698 }
699
700
701 #if !defined(CFA_WITH_IO_URING_IDLE)
702 #if !defined(__CFA_NO_STATISTICS__)
703 if(this->print_halts) {
704 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
705 }
706 #endif
707
708 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
709
710 {
711 eventfd_t val;
712 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
713 if(ret < 0) {
714 switch((int)errno) {
715 case EAGAIN:
716 #if EAGAIN != EWOULDBLOCK
717 case EWOULDBLOCK:
718 #endif
719 case EINTR:
720 // No need to do anything special here, just assume it's a legitimate wake-up
721 break;
722 default:
723 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
724 }
725 }
726 }
727
728 #if !defined(__CFA_NO_STATISTICS__)
729 if(this->print_halts) {
730 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
731 }
732 #endif
733 #else
734 // Do we already have a pending read
735 if(available(future)) {
736 // There is no pending read, we need to add one
737 reset(future);
738
739 __kernel_read(this, future, iov, this->idle_wctx.evfd );
740 }
741
742 __cfa_io_flush( this, 1 );
743 #endif
744}
745
746static bool mark_idle(__cluster_proc_list & this, processor & proc) {
747 __STATS__(true, ready.sleep.halts++; )
748
749 proc.idle_wctx.sem = 0;
750
751 /* paranoid */ verify( ! __preemption_enabled() );
752 if(!try_lock( this )) return false;
753 this.idle++;
754 /* paranoid */ verify( this.idle <= this.total );
755 remove(proc);
756 insert_first(this.idles, proc);
757
758 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
759 unlock( this );
760 /* paranoid */ verify( ! __preemption_enabled() );
761
762 return true;
763}
764
765static void mark_awake(__cluster_proc_list & this, processor & proc) {
766 /* paranoid */ verify( ! __preemption_enabled() );
767 lock( this );
768 this.idle--;
769 /* paranoid */ verify( this.idle >= 0 );
770 remove(proc);
771 insert_last(this.actives, proc);
772
773 {
774 struct __fd_waitctx * wctx = 0;
775 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
776 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
777 }
778
779 unlock( this );
780 /* paranoid */ verify( ! __preemption_enabled() );
781}
782
783//=============================================================================================
784// Unexpected Terminating logic
785//=============================================================================================
786void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
787 thread$ * thrd = __cfaabi_tls.this_thread;
788
789 if(thrd) {
790 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
791 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
792
793 if ( &thrd->self_cor != thrd->curr_cor ) {
794 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
795 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
796 }
797 else {
798 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
799 }
800 }
801 else {
802 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
803 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
804 }
805}
806
807int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
808 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
809}
810
811static __spinlock_t kernel_debug_lock;
812
813extern "C" {
814 void __cfaabi_bits_acquire() {
815 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
816 }
817
818 void __cfaabi_bits_release() {
819 unlock( kernel_debug_lock );
820 }
821}
822
823//=============================================================================================
824// Kernel Utilities
825//=============================================================================================
826#if defined(CFA_HAVE_LINUX_IO_URING_H)
827#include "io/types.hfa"
828#endif
829
830static inline bool __maybe_io_drain( processor * proc ) {
831 bool ret = false;
832 #if defined(CFA_HAVE_LINUX_IO_URING_H)
833 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
834
835 // Check if we should drain the queue
836 $io_context * ctx = proc->io.ctx;
837 unsigned head = *ctx->cq.head;
838 unsigned tail = *ctx->cq.tail;
839 if(head == tail) return false;
840 ready_schedule_lock();
841 ret = __cfa_io_drain( ctx );
842 ready_schedule_unlock();
843 #endif
844 return ret;
845}
846
847//-----------------------------------------------------------------------------
848// Debug
849__cfaabi_dbg_debug_do(
850 extern "C" {
851 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
852 this.prev_name = prev_name;
853 this.prev_thrd = kernelTLS().this_thread;
854 }
855 }
856)
857
858//-----------------------------------------------------------------------------
859// Debug
860bool threading_enabled(void) __attribute__((const)) {
861 return true;
862}
863
864//-----------------------------------------------------------------------------
865// Statistics
866#if !defined(__CFA_NO_STATISTICS__)
867 void print_halts( processor & this ) {
868 this.print_halts = true;
869 }
870
871 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
872 /* paranoid */ verify( cltr->stats );
873
874 processor * it = &list`first;
875 for(unsigned i = 0; i < count; i++) {
876 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
877 /* paranoid */ verify( it->local_data->this_stats );
878 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
879 __tally_stats( cltr->stats, it->local_data->this_stats );
880 it = &(*it)`next;
881 }
882 }
883
884 void crawl_cluster_stats( cluster & this ) {
885 // Stop the world, otherwise stats could get really messed-up
886 // this doesn't solve all problems but does solve many
887 // so it's probably good enough
888 disable_interrupts();
889 uint_fast32_t last_size = ready_mutate_lock();
890
891 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
892 crawl_list(&this, this.procs.idles , this.procs.idle );
893
894 // Unlock the RWlock
895 ready_mutate_unlock( last_size );
896 enable_interrupts();
897 }
898
899
900 void print_stats_now( cluster & this, int flags ) {
901 crawl_cluster_stats( this );
902 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
903 }
904#endif
905// Local Variables: //
906// mode: c //
907// tab-width: 4 //
908// End: //
Note: See TracBrowser for help on using the repository browser.