source: libcfa/src/concurrency/kernel.cfa@ 3e1a705

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since 3e1a705 was 8464edf, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

'print_stats_now' now pulls stats from processors.

  • Property mode set to 100644
File size: 29.7 KB
RevLine 
[8118303]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
[75f3522]10// Created On : Tue Jan 17 12:27:26 2017
[6b0b624]11// Last Modified By : Peter A. Buhr
[0190480]12// Last Modified On : Mon Aug 31 07:08:20 2020
13// Update Count : 71
[8118303]14//
15
[2026bb6]16#define __cforall_thread__
[4069faad]17// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
[2026bb6]18
[8118303]19//C Includes
[214e8da]20#include <errno.h>
[9d944b2]21#include <stdio.h>
[58b6d1b]22#include <signal.h>
[9d944b2]23#include <unistd.h>
[dddb3dd0]24extern "C" {
25 #include <sys/eventfd.h>
26}
[8118303]27
28//CFA Includes
[73abe95]29#include "kernel_private.hfa"
30#include "preemption.hfa"
[8118303]31
32//Private includes
33#define __CFA_INVOKE_PRIVATE__
34#include "invoke.h"
35
[89eff25]36#if !defined(__CFA_NO_STATISTICS__)
37 #define __STATS( ...) __VA_ARGS__
38#else
39 #define __STATS( ...)
40#endif
[4069faad]41
[deca0f5]42//-----------------------------------------------------------------------------
43// Some assembly required
[1805b1b]44#if defined( __i386 )
[deca0f5]45 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
46 // fcw : X87 FPU control word (preserved across function calls)
47 #define __x87_store \
48 uint32_t __mxcr; \
49 uint16_t __fcw; \
50 __asm__ volatile ( \
51 "stmxcsr %0\n" \
52 "fnstcw %1\n" \
53 : "=m" (__mxcr),\
54 "=m" (__fcw) \
55 )
56
57 #define __x87_load \
58 __asm__ volatile ( \
59 "fldcw %1\n" \
60 "ldmxcsr %0\n" \
61 ::"m" (__mxcr),\
62 "m" (__fcw) \
63 )
64
65#elif defined( __x86_64 )
66 #define __x87_store \
67 uint32_t __mxcr; \
68 uint16_t __fcw; \
69 __asm__ volatile ( \
70 "stmxcsr %0\n" \
71 "fnstcw %1\n" \
72 : "=m" (__mxcr),\
73 "=m" (__fcw) \
74 )
75
76 #define __x87_load \
77 __asm__ volatile ( \
78 "fldcw %1\n" \
79 "ldmxcsr %0\n" \
80 :: "m" (__mxcr),\
81 "m" (__fcw) \
82 )
83
[0190480]84#elif defined( __arm__ )
85 #define __x87_store
86 #define __x87_load
87
88#elif defined( __aarch64__ )
[74f5c83]89 #define __x87_store \
90 uint32_t __fpcntl[2]; \
91 __asm__ volatile ( \
92 "mrs x9, FPCR\n" \
93 "mrs x10, FPSR\n" \
94 "stp x9, x10, %0\n" \
95 : "=m" (__fpcntl) : : "x9", "x10" \
96 )
97
98 #define __x87_load \
99 __asm__ volatile ( \
100 "ldp x9, x10, %0\n" \
101 "msr FPSR, x10\n" \
102 "msr FPCR, x9\n" \
103 : "=m" (__fpcntl) : : "x9", "x10" \
104 )
105
[deca0f5]106#else
[0190480]107 #error unsupported hardware architecture
[deca0f5]108#endif
109
[e660761]110extern $thread * mainThread;
111extern processor * mainProcessor;
[2ac095d]112
[92e7631]113//-----------------------------------------------------------------------------
114// Kernel Scheduling logic
115static $thread * __next_thread(cluster * this);
[1eb239e4]116static $thread * __next_thread_slow(cluster * this);
[c6c7e6c]117static inline bool __must_unpark( $thread * thrd ) __attribute((nonnull(1)));
[92e7631]118static void __run_thread(processor * this, $thread * dst);
[e873838]119static void __wake_one(cluster * cltr);
[1eb239e4]120
[6a9b12b]121static void mark_idle (__cluster_proc_list & idles, processor & proc);
122static void mark_awake(__cluster_proc_list & idles, processor & proc);
123static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
[1eb239e4]124
[dddb3dd0]125extern void __cfa_io_start( processor * );
[c1c95b1]126extern bool __cfa_io_drain( processor * );
[dddb3dd0]127extern void __cfa_io_flush( processor * );
128extern void __cfa_io_stop ( processor * );
[c1c95b1]129static inline bool __maybe_io_drain( processor * );
[dddb3dd0]130
131extern void __disable_interrupts_hard();
132extern void __enable_interrupts_hard();
[c84e80a]133
[a3821fa]134static inline void __disable_interrupts_checked() {
135 /* paranoid */ verify( __preemption_enabled() );
136 disable_interrupts();
137 /* paranoid */ verify( ! __preemption_enabled() );
138}
139
140static inline void __enable_interrupts_checked( bool poll = true ) {
141 /* paranoid */ verify( ! __preemption_enabled() );
142 enable_interrupts( poll );
143 /* paranoid */ verify( __preemption_enabled() );
144}
145
[75f3522]146//=============================================================================================
147// Kernel Scheduling logic
148//=============================================================================================
[8fcbb4c]149//Main of the processor contexts
[83a071f9]150void main(processorCtx_t & runner) {
[21184e3]151 // Because of a bug, we couldn't initialized the seed on construction
152 // Do it here
[8fc652e0]153 __cfaabi_tls.rand_seed ^= rdtscl();
154 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
[f2384c9a]155 __tls_rand_advance_bck();
[21184e3]156
[83a071f9]157 processor * this = runner.proc;
[094476d]158 verify(this);
[c81ebf9]159
[dddb3dd0]160 __cfa_io_start( this );
161
[4069faad]162 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
[28d73c1]163 #if !defined(__CFA_NO_STATISTICS__)
164 if( this->print_halts ) {
[c993b15]165 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
[28d73c1]166 }
167 #endif
[b798713]168
[75f3522]169 {
[c81ebf9]170 // Setup preemption data
171 preemption_scope scope = { this };
172
[a5e7233]173 // if we need to run some special setup, now is the time to do it.
174 if(this->init.thrd) {
175 this->init.thrd->curr_cluster = this->cltr;
176 __run_thread(this, this->init.thrd);
177 }
[325e6ea]178
[4069faad]179 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
[8118303]180
[ac2b598]181 $thread * readyThread = 0p;
[1eb239e4]182 MAIN_LOOP:
183 for() {
[34b2796]184 #if 1
[dddb3dd0]185 // Check if there is pending io
186 __maybe_io_drain( this );
187
[92e7631]188 // Try to get the next thread
[8c50aed]189 readyThread = __next_thread( this->cltr );
[75f3522]190
[1eb239e4]191 if( !readyThread ) {
[dddb3dd0]192 __cfa_io_flush( this );
[1eb239e4]193 readyThread = __next_thread_slow( this->cltr );
194 }
[4e6fb8e]195
[1eb239e4]196 HALT:
197 if( !readyThread ) {
198 // Don't block if we are done
199 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[c81ebf9]200
[1eb239e4]201 #if !defined(__CFA_NO_STATISTICS__)
202 __tls_stats()->ready.sleep.halts++;
203 #endif
[398e8e9]204
[1eb239e4]205 // Push self to idle stack
[6a9b12b]206 mark_idle(this->cltr->procs, * this);
[398e8e9]207
[1eb239e4]208 // Confirm the ready-queue is empty
209 readyThread = __next_thread_slow( this->cltr );
210 if( readyThread ) {
211 // A thread was found, cancel the halt
[6a9b12b]212 mark_awake(this->cltr->procs, * this);
[1eb239e4]213
214 #if !defined(__CFA_NO_STATISTICS__)
215 __tls_stats()->ready.sleep.cancels++;
216 #endif
217
218 // continue the mai loop
219 break HALT;
220 }
221
222 #if !defined(__CFA_NO_STATISTICS__)
223 if(this->print_halts) {
[c993b15]224 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
[1eb239e4]225 }
226 #endif
227
[dddb3dd0]228 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
229
230 __disable_interrupts_hard();
231 eventfd_t val;
232 eventfd_read( this->idle, &val );
233 __enable_interrupts_hard();
[1eb239e4]234
235 #if !defined(__CFA_NO_STATISTICS__)
236 if(this->print_halts) {
[c993b15]237 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
[1eb239e4]238 }
239 #endif
240
241 // We were woken up, remove self from idle
[6a9b12b]242 mark_awake(this->cltr->procs, * this);
[1eb239e4]243
244 // DON'T just proceed, start looking again
245 continue MAIN_LOOP;
[64a7146]246 }
[1eb239e4]247
248 /* paranoid */ verify( readyThread );
249
[dddb3dd0]250 // Reset io dirty bit
251 this->io.dirty = false;
252
[1eb239e4]253 // We found a thread run it
254 __run_thread(this, readyThread);
255
256 // Are we done?
257 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[325e6ea]258
[dddb3dd0]259 if(this->io.pending && !this->io.dirty) {
260 __cfa_io_flush( this );
261 }
[89eff25]262
[34b2796]263 #else
264
265 SEARCH: {
266 /* paranoid */ verify( ! __preemption_enabled() );
267 /* paranoid */ verify( kernelTLS().this_proc_id );
268
269 // First, lock the scheduler since we are searching for a thread
270 ready_schedule_lock();
271
272 // Try to get the next thread
273 readyThread = pop_fast( this->cltr );
274 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
275
276 // If we can't find a thread, might as well flush any outstanding I/O
277 if(this->io.pending) { __cfa_io_flush( this ); }
278
279 // Spin a little on I/O, just in case
280 for(25) {
281 __maybe_io_drain( this );
282 readyThread = pop_fast( this->cltr );
283 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
284 }
285
286 // no luck, try stealing a few times
287 for(25) {
288 if( __maybe_io_drain( this ) ) {
289 readyThread = pop_fast( this->cltr );
290 } else {
291 readyThread = pop_slow( this->cltr );
292 }
293 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
294 }
295
296 // still no luck, search for a thread
297 readyThread = pop_search( this->cltr );
298 if(readyThread) { ready_schedule_unlock(); break SEARCH; }
299
300 // Don't block if we are done
301 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
302
303 __STATS( __tls_stats()->ready.sleep.halts++; )
304
305 // Push self to idle stack
306 ready_schedule_unlock();
307 mark_idle(this->cltr->procs, * this);
308 ready_schedule_lock();
309
310 // Confirm the ready-queue is empty
311 __maybe_io_drain( this );
312 readyThread = pop_search( this->cltr );
313 ready_schedule_unlock();
314
315 if( readyThread ) {
316 // A thread was found, cancel the halt
317 mark_awake(this->cltr->procs, * this);
318
319 __STATS( __tls_stats()->ready.sleep.cancels++; )
320
321 // continue the main loop
322 break SEARCH;
323 }
324
325 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl()); )
326 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
327
328 // __disable_interrupts_hard();
329 eventfd_t val;
330 eventfd_read( this->idle, &val );
331 // __enable_interrupts_hard();
332
333 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl()); )
334
335 // We were woken up, remove self from idle
336 mark_awake(this->cltr->procs, * this);
337
338 // DON'T just proceed, start looking again
339 continue MAIN_LOOP;
340 }
341
342 RUN_THREAD:
343 /* paranoid */ verify( kernelTLS().this_proc_id );
344 /* paranoid */ verify( ! __preemption_enabled() );
345 /* paranoid */ verify( readyThread );
346
347 // Reset io dirty bit
348 this->io.dirty = false;
349
350 // We found a thread run it
351 __run_thread(this, readyThread);
352
353 // Are we done?
354 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
355
356 #if !defined(__CFA_NO_STATISTICS__)
357 unsigned long long curr = rdtscl();
358 if(curr > (last_tally + 500000000)) {
359 __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
360 last_tally = curr;
361 }
362 #endif
363
364 if(this->io.pending && !this->io.dirty) {
365 __cfa_io_flush( this );
366 }
367
368 ready_schedule_lock();
369 __maybe_io_drain( this );
370 ready_schedule_unlock();
371 #endif
[c81ebf9]372 }
373
[4069faad]374 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
[c84e80a]375 }
[8118303]376
[dddb3dd0]377 __cfa_io_stop( this );
378
[454f478]379 post( this->terminated );
[bdeba0b]380
[28d73c1]381 if(this == mainProcessor) {
[6a490b2]382 // HACK : the coroutine context switch expects this_thread to be set
383 // and it make sense for it to be set in all other cases except here
384 // fake it
[8fc652e0]385 __cfaabi_tls.this_thread = mainThread;
[6a490b2]386 }
[7768b8d]387
[4069faad]388 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]389}
390
[5c1a531]391static int * __volatile_errno() __attribute__((noinline));
392static int * __volatile_errno() { asm(""); return &errno; }
393
[14a61b5]394// KERNEL ONLY
[1c273d0]395// runThread runs a thread by context switching
396// from the processor coroutine to the target thread
[ac2b598]397static void __run_thread(processor * this, $thread * thrd_dst) {
[8fc652e0]398 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]399 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
400 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
401 __builtin_prefetch( thrd_dst->context.SP );
402
[dddb3dd0]403 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
404
[ac2b598]405 $coroutine * proc_cor = get_coroutine(this->runner);
[8fcbb4c]406
[9f575ea]407 // set state of processor coroutine to inactive
408 verify(proc_cor->state == Active);
[ae7be7a]409 proc_cor->state = Blocked;
[e8e457e]410
[9f575ea]411 // Actually run the thread
[3381ed7]412 RUNNING: while(true) {
[ff79d5e]413 thrd_dst->preempted = __NO_PREEMPTION;
414 thrd_dst->state = Active;
[e8e457e]415
[58d64a4]416 // Update global state
[8fc652e0]417 kernelTLS().this_thread = thrd_dst;
[75f3522]418
[8fc652e0]419 /* paranoid */ verify( ! __preemption_enabled() );
420 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]421 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]422 /* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]423 /* paranoid */ verify( thrd_dst->state != Halted );
[210b8b3]424 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
425 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
[ac12f1f]426 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]427
[3381ed7]428
[58d64a4]429
[9f575ea]430 // set context switch to the thread that the processor is executing
[c7a900a]431 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
432 // when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]433
[ac12f1f]434 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[210b8b3]435 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
436 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
[b4b63e8]437 /* paranoid */ verify( thrd_dst->context.SP );
[9d6e1b8a]438 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[8fc652e0]439 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
440 /* paranoid */ verify( ! __preemption_enabled() );
[75f3522]441
[58d64a4]442 // Reset global state
[8fc652e0]443 kernelTLS().this_thread = 0p;
[3381ed7]444
445 // We just finished running a thread, there are a few things that could have happened.
446 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
447 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
448 // 4 - Preempted
449 // In case 1, we may have won a race so we can't write to the state again.
450 // In case 2, we lost the race so we now own the thread.
451
452 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
453 // The thread was preempted, reschedule it and reset the flag
[254ad1b]454 schedule_thread$( thrd_dst );
[3381ed7]455 break RUNNING;
456 }
[75f3522]457
[3ea8ad1]458 if(unlikely(thrd_dst->state == Halting)) {
[ff79d5e]459 // The thread has halted, it should never be scheduled/run again
[5afb49a]460 // finish the thread
461 __thread_finish( thrd_dst );
[ff79d5e]462 break RUNNING;
463 }
464
465 /* paranoid */ verify( thrd_dst->state == Active );
466 thrd_dst->state = Blocked;
467
[3381ed7]468 // set state of processor coroutine to active and the thread to inactive
[ff79d5e]469 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
470 switch(old_ticket) {
[6a77224]471 case TICKET_RUNNING:
[3381ed7]472 // This is case 1, the regular case, nothing more is needed
473 break RUNNING;
[6a77224]474 case TICKET_UNBLOCK:
[ec43cf9]475 #if !defined(__CFA_NO_STATISTICS__)
476 __tls_stats()->ready.threads.threads++;
[73f4d08]477 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]478 #endif
[3381ed7]479 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
480 // In this case, just run it again.
481 continue RUNNING;
482 default:
483 // This makes no sense, something is wrong abort
[ff79d5e]484 abort();
[3381ed7]485 }
[9f575ea]486 }
[e8e457e]487
[9f575ea]488 // Just before returning to the processor, set the processor coroutine to active
[e8e457e]489 proc_cor->state = Active;
[1eb239e4]490
[dddb3dd0]491 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
492
[ec43cf9]493 #if !defined(__CFA_NO_STATISTICS__)
494 __tls_stats()->ready.threads.threads--;
[73f4d08]495 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]496 #endif
497
[8fc652e0]498 /* paranoid */ verify( ! __preemption_enabled() );
[82c948c]499}
500
[14a61b5]501// KERNEL_ONLY
[b0c7419]502void returnToKernel() {
[8fc652e0]503 /* paranoid */ verify( ! __preemption_enabled() );
504 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
505 $thread * thrd_src = kernelTLS().this_thread;
[e8e457e]506
[89eff25]507 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]508
[9f575ea]509 // Run the thread on this processor
510 {
511 int local_errno = *__volatile_errno();
512 #if defined( __i386 ) || defined( __x86_64 )
513 __x87_store;
514 #endif
[b4b63e8]515 /* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]516 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]517 __cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]518 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]519 #if defined( __i386 ) || defined( __x86_64 )
520 __x87_load;
521 #endif
522 *__volatile_errno() = local_errno;
[8fcbb4c]523 }
[deca0f5]524
[29cb302]525 #if !defined(__CFA_NO_STATISTICS__)
[89eff25]526 /* paranoid */ verify( thrd_src->last_proc != 0p );
527 if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]528 __tls_stats()->ready.threads.migration++;
529 }
530 #endif
531
[8fc652e0]532 /* paranoid */ verify( ! __preemption_enabled() );
[210b8b3]533 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
534 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
[c84e80a]535}
536
[8def349]537//-----------------------------------------------------------------------------
538// Scheduler routines
[14a61b5]539// KERNEL ONLY
[e9c0b4c]540static void __schedule_thread( $thread * thrd ) {
[8fc652e0]541 /* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]542 /* paranoid */ verify( ready_schedule_islocked());
[6a490b2]543 /* paranoid */ verify( thrd );
544 /* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]545 /* paranoid */ verify( thrd->curr_cluster );
[3381ed7]546 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]547 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
548 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]549 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]550 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]551 /* paranoid */ #endif
[6a490b2]552 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
[ac12f1f]553 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]554
[9f575ea]555
[ae7be7a]556 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]557
[ec43cf9]558 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
559 struct cluster * cl = thrd->curr_cluster;
[89eff25]560 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]561
[254ad1b]562 // push the thread to the cluster ready-queue
563 push( cl, thrd );
[32a8b61]564
[254ad1b]565 // variable thrd is no longer safe to use
[734908c]566 thrd = 0xdeaddeaddeaddeadp;
[32a8b61]567
[254ad1b]568 // wake the cluster using the save variable.
569 __wake_one( cl );
[1c273d0]570
[ec43cf9]571 #if !defined(__CFA_NO_STATISTICS__)
572 if( kernelTLS().this_stats ) {
573 __tls_stats()->ready.threads.threads++;
[89eff25]574 if(outside) {
575 __tls_stats()->ready.threads.extunpark++;
576 }
[73f4d08]577 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
[ec43cf9]578 }
579 else {
580 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]581 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[73f4d08]582 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
[ec43cf9]583 }
584 #endif
585
[254ad1b]586 /* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]587 /* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]588}
589
[254ad1b]590void schedule_thread$( $thread * thrd ) {
591 ready_schedule_lock();
592 __schedule_thread( thrd );
593 ready_schedule_unlock();
594}
595
[14a61b5]596// KERNEL ONLY
[1eb239e4]597static inline $thread * __next_thread(cluster * this) with( *this ) {
[8fc652e0]598 /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]599
[e873838]600 ready_schedule_lock();
[431cd4f]601 $thread * thrd = pop_fast( this );
[e873838]602 ready_schedule_unlock();
[7768b8d]603
[8fc652e0]604 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]605 return thrd;
[eb2e723]606}
607
[64a7146]608// KERNEL ONLY
[1eb239e4]609static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]610 /* paranoid */ verify( ! __preemption_enabled() );
[64a7146]611
[e873838]612 ready_schedule_lock();
[fc59df78]613 $thread * thrd;
614 for(25) {
615 thrd = pop_slow( this );
616 if(thrd) goto RET;
617 }
618 thrd = pop_search( this );
619
620 RET:
[e873838]621 ready_schedule_unlock();
[64a7146]622
[8fc652e0]623 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]624 return thrd;
[64a7146]625}
626
[c6c7e6c]627static inline bool __must_unpark( $thread * thrd ) {
[ff79d5e]628 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
629 switch(old_ticket) {
[6a77224]630 case TICKET_RUNNING:
[3381ed7]631 // Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]632 return false;
[6a77224]633 case TICKET_BLOCKED:
[3381ed7]634 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]635 /* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]636 return true;
[3381ed7]637 default:
638 // This makes no sense, something is wrong abort
[7ee8153]639 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]640 }
[eb2e723]641}
642
[e9c0b4c]643void __kernel_unpark( $thread * thrd ) {
644 /* paranoid */ verify( ! __preemption_enabled() );
645 /* paranoid */ verify( ready_schedule_islocked());
646
647 if( !thrd ) return;
648
649 if(__must_unpark(thrd)) {
650 // Wake lost the race,
651 __schedule_thread( thrd );
652 }
653
654 /* paranoid */ verify( ready_schedule_islocked());
655 /* paranoid */ verify( ! __preemption_enabled() );
656}
657
[c6c7e6c]658void unpark( $thread * thrd ) {
659 if( !thrd ) return;
[0b33412]660
[c6c7e6c]661 if(__must_unpark(thrd)) {
662 disable_interrupts();
[a3821fa]663 // Wake lost the race,
[254ad1b]664 schedule_thread$( thrd );
[a3821fa]665 enable_interrupts(false);
[de6319f]666 }
[eb2e723]667}
[0b33412]668
[e235429]669void park( void ) {
[a3821fa]670 __disable_interrupts_checked();
671 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
672 returnToKernel();
673 __enable_interrupts_checked();
[0c78741]674
675}
[09800e9]676
[5afb49a]677extern "C" {
678 // Leave the thread monitor
679 // last routine called by a thread.
680 // Should never return
681 void __cfactx_thrd_leave() {
[8fc652e0]682 $thread * thrd = active_thread();
[5afb49a]683 $monitor * this = &thrd->self_mon;
684
685 // Lock the monitor now
686 lock( this->lock __cfaabi_dbg_ctx2 );
687
688 disable_interrupts();
689
[9d6e1b8a]690 /* paranoid */ verify( ! __preemption_enabled() );
691 /* paranoid */ verify( thrd->state == Active );
692 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
693 /* paranoid */ verify( kernelTLS().this_thread == thrd );
694 /* paranoid */ verify( thrd->context.SP );
695 /* 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 );
696 /* 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 );
697
[3ea8ad1]698 thrd->state = Halting;
[58688bf]699 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]700 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
701 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]702
703 // Leave the thread
704 returnToKernel();
705
706 // Control flow should never reach here!
[9d6e1b8a]707 abort();
[5afb49a]708 }
[09800e9]709}
710
[14a61b5]711// KERNEL ONLY
[3381ed7]712bool force_yield( __Preemption_Reason reason ) {
[a3821fa]713 __disable_interrupts_checked();
714 $thread * thrd = kernelTLS().this_thread;
715 /* paranoid */ verify(thrd->state == Active);
716
717 // SKULLDUGGERY: It is possible that we are preempting this thread just before
718 // it was going to park itself. If that is the case and it is already using the
719 // intrusive fields then we can't use them to preempt the thread
720 // If that is the case, abandon the preemption.
721 bool preempted = false;
722 if(thrd->link.next == 0p) {
723 preempted = true;
724 thrd->preempted = reason;
725 returnToKernel();
726 }
727 __enable_interrupts_checked( false );
[3381ed7]728 return preempted;
[f2b12406]729}
730
[14a61b5]731//=============================================================================================
[92e7631]732// Kernel Idle Sleep
[14a61b5]733//=============================================================================================
[64a7146]734// Wake a thread from the front if there are any
[e873838]735static void __wake_one(cluster * this) {
[8fc652e0]736 /* paranoid */ verify( ! __preemption_enabled() );
[e873838]737 /* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]738
[64a7146]739 // Check if there is a sleeping processor
[1eb239e4]740 processor * p;
741 unsigned idle;
742 unsigned total;
[6a9b12b]743 [idle, total, p] = query_idles(this->procs);
[14a61b5]744
[64a7146]745 // If no one is sleeping, we are done
[1eb239e4]746 if( idle == 0 ) return;
[14a61b5]747
[64a7146]748 // We found a processor, wake it up
[dddb3dd0]749 eventfd_t val;
750 val = 1;
751 eventfd_write( p->idle, val );
[14a61b5]752
[1eb239e4]753 #if !defined(__CFA_NO_STATISTICS__)
[5cb51502]754 if( kernelTLS().this_stats ) {
755 __tls_stats()->ready.sleep.wakes++;
756 }
757 else {
758 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
759 }
[1eb239e4]760 #endif
761
[e873838]762 /* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]763 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]764
765 return;
[64a7146]766}
[14a61b5]767
[64a7146]768// Unconditionnaly wake a thread
[1eb239e4]769void __wake_proc(processor * this) {
[64a7146]770 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]771
[a3821fa]772 __disable_interrupts_checked();
[8fc652e0]773 /* paranoid */ verify( ! __preemption_enabled() );
[dddb3dd0]774 eventfd_t val;
775 val = 1;
[55d6affb]776 eventfd_write( this->idle, val );
[a3821fa]777 __enable_interrupts_checked();
[92e7631]778}
779
[6a9b12b]780static void mark_idle(__cluster_proc_list & this, processor & proc) {
[8fc652e0]781 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]782 lock( this );
783 this.idle++;
784 /* paranoid */ verify( this.idle <= this.total );
[fc59b580]785 remove(proc);
[6a9b12b]786 insert_first(this.idles, proc);
[1eb239e4]787 unlock( this );
[8fc652e0]788 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]789}
[8e16177]790
[6a9b12b]791static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]792 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]793 lock( this );
794 this.idle--;
795 /* paranoid */ verify( this.idle >= 0 );
796 remove(proc);
[fc59b580]797 insert_last(this.actives, proc);
[1eb239e4]798 unlock( this );
[8fc652e0]799 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]800}
[c34ebf2]801
[6a9b12b]802static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
803 /* paranoid */ verify( ! __preemption_enabled() );
804 /* paranoid */ verify( ready_schedule_islocked() );
805
[1eb239e4]806 for() {
807 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
808 if( 1 == (l % 2) ) { Pause(); continue; }
809 unsigned idle = this.idle;
810 unsigned total = this.total;
[6a9b12b]811 processor * proc = &this.idles`first;
[7fdae38]812 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
813 asm volatile("": : :"memory");
[1eb239e4]814 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
815 return [idle, total, proc];
816 }
[6a9b12b]817
818 /* paranoid */ verify( ready_schedule_islocked() );
819 /* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]820}
821
[dbe9b08]822//=============================================================================================
823// Unexpected Terminating logic
824//=============================================================================================
[92bfda0]825void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
826 $thread * thrd = __cfaabi_tls.this_thread;
[9d944b2]827
[de94a60]828 if(thrd) {
829 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]830 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]831
[212c2187]832 if ( &thrd->self_cor != thrd->curr_cor ) {
833 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]834 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]835 }
836 else {
[1c40091]837 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]838 }
[1c273d0]839 }
[9d944b2]840 else {
[de94a60]841 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]842 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]843 }
844}
845
[92bfda0]846int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
847 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]848}
849
[de94a60]850static __spinlock_t kernel_debug_lock;
851
[9d944b2]852extern "C" {
[1c40091]853 void __cfaabi_bits_acquire() {
[36982fc]854 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]855 }
856
[1c40091]857 void __cfaabi_bits_release() {
[ea7d2b0]858 unlock( kernel_debug_lock );
[9d944b2]859 }
[8118303]860}
861
[fa21ac9]862//=============================================================================================
863// Kernel Utilities
864//=============================================================================================
[dddb3dd0]865#if defined(CFA_HAVE_LINUX_IO_URING_H)
866#include "io/types.hfa"
867#endif
868
[c1c95b1]869static inline bool __maybe_io_drain( processor * proc ) {
[e9c0b4c]870 bool ret = false;
[dddb3dd0]871 #if defined(CFA_HAVE_LINUX_IO_URING_H)
872 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
873
874 // Check if we should drain the queue
875 $io_context * ctx = proc->io.ctx;
876 unsigned head = *ctx->cq.head;
877 unsigned tail = *ctx->cq.tail;
[c1c95b1]878 if(head == tail) return false;
[b7fd2db6]879 ready_schedule_lock();
[e9c0b4c]880 ret = __cfa_io_drain( proc );
[b7fd2db6]881 ready_schedule_unlock();
[dddb3dd0]882 #endif
[e9c0b4c]883 return ret;
[dddb3dd0]884}
885
[de94a60]886//-----------------------------------------------------------------------------
887// Debug
888__cfaabi_dbg_debug_do(
[1997b4e]889 extern "C" {
[ae66348]890 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
[1997b4e]891 this.prev_name = prev_name;
[8fc652e0]892 this.prev_thrd = kernelTLS().this_thread;
[1997b4e]893 }
[9181f1d]894 }
[f7d6bb0]895)
[2026bb6]896
897//-----------------------------------------------------------------------------
898// Debug
[8c50aed]899bool threading_enabled(void) __attribute__((const)) {
[2026bb6]900 return true;
901}
[c34ebf2]902
903//-----------------------------------------------------------------------------
904// Statistics
905#if !defined(__CFA_NO_STATISTICS__)
906 void print_halts( processor & this ) {
907 this.print_halts = true;
908 }
[58688bf]909
[8464edf]910 static void crawl_list( cluster * cltr, dlist(processor, processor) & list, unsigned count ) {
911 /* paranoid */ verify( cltr->stats );
912
913 processor * it = &list`first;
914 for(unsigned i = 0; i < count; i++) {
915 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
916 /* paranoid */ verify( it->local_data->this_stats );
917 __tally_stats( cltr->stats, it->local_data->this_stats );
918 it = &(*it)`next;
919 }
920 }
921
922 void crawl_cluster_stats( cluster & this ) {
923 // Stop the world, otherwise stats could get really messed-up
924 // this doesn't solve all problems but does solve many
925 // so it's probably good enough
926 uint_fast32_t last_size = ready_mutate_lock();
927
928 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
929 crawl_list(&this, this.procs.idles , this.procs.idle );
930
931 // Unlock the RWlock
932 ready_mutate_unlock( last_size );
933 }
934
935
[58688bf]936 void print_stats_now( cluster & this, int flags ) {
[8464edf]937 crawl_cluster_stats( this );
[1b033b8]938 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
939 }
[c34ebf2]940#endif
[8118303]941// Local Variables: //
942// mode: c //
943// tab-width: 4 //
944// End: //
Note: See TracBrowser for help on using the repository browser.