source: libcfa/src/concurrency/kernel.cfa@ 375fb46

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 375fb46 was b808625, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added option to ready-queue to push ignoring locality.

  • Property mode set to 100644
File size: 29.5 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() {
[57f70ab]184 #define OLD_MAIN 1
185 #if OLD_MAIN
[dddb3dd0]186 // Check if there is pending io
187 __maybe_io_drain( this );
188
[92e7631]189 // Try to get the next thread
[8c50aed]190 readyThread = __next_thread( this->cltr );
[75f3522]191
[1eb239e4]192 if( !readyThread ) {
[dddb3dd0]193 __cfa_io_flush( this );
[1eb239e4]194 readyThread = __next_thread_slow( this->cltr );
195 }
[4e6fb8e]196
[1eb239e4]197 HALT:
198 if( !readyThread ) {
199 // Don't block if we are done
200 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[c81ebf9]201
[1eb239e4]202 #if !defined(__CFA_NO_STATISTICS__)
203 __tls_stats()->ready.sleep.halts++;
204 #endif
[398e8e9]205
[1eb239e4]206 // Push self to idle stack
[6a9b12b]207 mark_idle(this->cltr->procs, * this);
[398e8e9]208
[1eb239e4]209 // Confirm the ready-queue is empty
210 readyThread = __next_thread_slow( this->cltr );
211 if( readyThread ) {
212 // A thread was found, cancel the halt
[6a9b12b]213 mark_awake(this->cltr->procs, * this);
[1eb239e4]214
215 #if !defined(__CFA_NO_STATISTICS__)
216 __tls_stats()->ready.sleep.cancels++;
217 #endif
218
219 // continue the mai loop
220 break HALT;
221 }
222
223 #if !defined(__CFA_NO_STATISTICS__)
224 if(this->print_halts) {
[c993b15]225 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
[1eb239e4]226 }
227 #endif
228
[dddb3dd0]229 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
230
231 __disable_interrupts_hard();
232 eventfd_t val;
233 eventfd_read( this->idle, &val );
234 __enable_interrupts_hard();
[1eb239e4]235
236 #if !defined(__CFA_NO_STATISTICS__)
237 if(this->print_halts) {
[c993b15]238 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
[1eb239e4]239 }
240 #endif
241
242 // We were woken up, remove self from idle
[6a9b12b]243 mark_awake(this->cltr->procs, * this);
[1eb239e4]244
245 // DON'T just proceed, start looking again
246 continue MAIN_LOOP;
[64a7146]247 }
[1eb239e4]248
249 /* paranoid */ verify( readyThread );
250
[dddb3dd0]251 // Reset io dirty bit
252 this->io.dirty = false;
253
[1eb239e4]254 // We found a thread run it
255 __run_thread(this, readyThread);
256
257 // Are we done?
258 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[325e6ea]259
[dddb3dd0]260 if(this->io.pending && !this->io.dirty) {
261 __cfa_io_flush( this );
262 }
[89eff25]263
[34b2796]264 #else
[57f70ab]265 #warning new kernel loop
[34b2796]266 SEARCH: {
267 /* paranoid */ verify( ! __preemption_enabled() );
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
[57f70ab]280 for(5) {
[34b2796]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
[57f70ab]287 for(5) {
[34b2796]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
[fb4ccdf]325 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
[34b2796]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
[fb4ccdf]333 __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
[34b2796]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( ! __preemption_enabled() );
344 /* paranoid */ verify( readyThread );
345
346 // Reset io dirty bit
347 this->io.dirty = false;
348
349 // We found a thread run it
350 __run_thread(this, readyThread);
351
352 // Are we done?
353 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
354
355 if(this->io.pending && !this->io.dirty) {
356 __cfa_io_flush( this );
357 }
358
359 ready_schedule_lock();
360 __maybe_io_drain( this );
361 ready_schedule_unlock();
362 #endif
[c81ebf9]363 }
364
[4069faad]365 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
[c84e80a]366 }
[8118303]367
[dddb3dd0]368 __cfa_io_stop( this );
369
[454f478]370 post( this->terminated );
[bdeba0b]371
[28d73c1]372 if(this == mainProcessor) {
[6a490b2]373 // HACK : the coroutine context switch expects this_thread to be set
374 // and it make sense for it to be set in all other cases except here
375 // fake it
[8fc652e0]376 __cfaabi_tls.this_thread = mainThread;
[6a490b2]377 }
[7768b8d]378
[4069faad]379 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]380}
381
[5c1a531]382static int * __volatile_errno() __attribute__((noinline));
383static int * __volatile_errno() { asm(""); return &errno; }
384
[14a61b5]385// KERNEL ONLY
[1c273d0]386// runThread runs a thread by context switching
387// from the processor coroutine to the target thread
[ac2b598]388static void __run_thread(processor * this, $thread * thrd_dst) {
[8fc652e0]389 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]390 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
391 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
392 __builtin_prefetch( thrd_dst->context.SP );
393
[dddb3dd0]394 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
395
[ac2b598]396 $coroutine * proc_cor = get_coroutine(this->runner);
[8fcbb4c]397
[9f575ea]398 // set state of processor coroutine to inactive
399 verify(proc_cor->state == Active);
[ae7be7a]400 proc_cor->state = Blocked;
[e8e457e]401
[9f575ea]402 // Actually run the thread
[3381ed7]403 RUNNING: while(true) {
[ff79d5e]404 thrd_dst->preempted = __NO_PREEMPTION;
405 thrd_dst->state = Active;
[e8e457e]406
[58d64a4]407 // Update global state
[8fc652e0]408 kernelTLS().this_thread = thrd_dst;
[75f3522]409
[8fc652e0]410 /* paranoid */ verify( ! __preemption_enabled() );
411 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]412 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]413 /* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]414 /* paranoid */ verify( thrd_dst->state != Halted );
[210b8b3]415 /* 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
416 /* 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]417 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]418
[3381ed7]419
[58d64a4]420
[9f575ea]421 // set context switch to the thread that the processor is executing
[c7a900a]422 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
423 // when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]424
[ac12f1f]425 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[210b8b3]426 /* 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 );
427 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
[b4b63e8]428 /* paranoid */ verify( thrd_dst->context.SP );
[9d6e1b8a]429 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[8fc652e0]430 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
431 /* paranoid */ verify( ! __preemption_enabled() );
[75f3522]432
[58d64a4]433 // Reset global state
[8fc652e0]434 kernelTLS().this_thread = 0p;
[3381ed7]435
436 // We just finished running a thread, there are a few things that could have happened.
437 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
438 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
439 // 4 - Preempted
440 // In case 1, we may have won a race so we can't write to the state again.
441 // In case 2, we lost the race so we now own the thread.
442
443 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
444 // The thread was preempted, reschedule it and reset the flag
[254ad1b]445 schedule_thread$( thrd_dst );
[3381ed7]446 break RUNNING;
447 }
[75f3522]448
[3ea8ad1]449 if(unlikely(thrd_dst->state == Halting)) {
[ff79d5e]450 // The thread has halted, it should never be scheduled/run again
[5afb49a]451 // finish the thread
452 __thread_finish( thrd_dst );
[ff79d5e]453 break RUNNING;
454 }
455
456 /* paranoid */ verify( thrd_dst->state == Active );
457 thrd_dst->state = Blocked;
458
[3381ed7]459 // set state of processor coroutine to active and the thread to inactive
[ff79d5e]460 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
461 switch(old_ticket) {
[6a77224]462 case TICKET_RUNNING:
[3381ed7]463 // This is case 1, the regular case, nothing more is needed
464 break RUNNING;
[6a77224]465 case TICKET_UNBLOCK:
[ec43cf9]466 #if !defined(__CFA_NO_STATISTICS__)
467 __tls_stats()->ready.threads.threads++;
[73f4d08]468 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]469 #endif
[3381ed7]470 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
471 // In this case, just run it again.
472 continue RUNNING;
473 default:
474 // This makes no sense, something is wrong abort
[ff79d5e]475 abort();
[3381ed7]476 }
[9f575ea]477 }
[e8e457e]478
[9f575ea]479 // Just before returning to the processor, set the processor coroutine to active
[e8e457e]480 proc_cor->state = Active;
[1eb239e4]481
[dddb3dd0]482 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
483
[ec43cf9]484 #if !defined(__CFA_NO_STATISTICS__)
485 __tls_stats()->ready.threads.threads--;
[73f4d08]486 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]487 #endif
488
[8fc652e0]489 /* paranoid */ verify( ! __preemption_enabled() );
[82c948c]490}
491
[14a61b5]492// KERNEL_ONLY
[b0c7419]493void returnToKernel() {
[8fc652e0]494 /* paranoid */ verify( ! __preemption_enabled() );
495 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
496 $thread * thrd_src = kernelTLS().this_thread;
[e8e457e]497
[89eff25]498 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]499
[9f575ea]500 // Run the thread on this processor
501 {
502 int local_errno = *__volatile_errno();
503 #if defined( __i386 ) || defined( __x86_64 )
504 __x87_store;
505 #endif
[b4b63e8]506 /* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]507 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]508 __cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]509 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]510 #if defined( __i386 ) || defined( __x86_64 )
511 __x87_load;
512 #endif
513 *__volatile_errno() = local_errno;
[8fcbb4c]514 }
[deca0f5]515
[29cb302]516 #if !defined(__CFA_NO_STATISTICS__)
[89eff25]517 /* paranoid */ verify( thrd_src->last_proc != 0p );
518 if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]519 __tls_stats()->ready.threads.migration++;
520 }
521 #endif
522
[8fc652e0]523 /* paranoid */ verify( ! __preemption_enabled() );
[210b8b3]524 /* 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 );
525 /* 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]526}
527
[8def349]528//-----------------------------------------------------------------------------
529// Scheduler routines
[14a61b5]530// KERNEL ONLY
[e9c0b4c]531static void __schedule_thread( $thread * thrd ) {
[8fc652e0]532 /* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]533 /* paranoid */ verify( ready_schedule_islocked());
[6a490b2]534 /* paranoid */ verify( thrd );
535 /* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]536 /* paranoid */ verify( thrd->curr_cluster );
[3381ed7]537 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]538 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
539 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]540 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]541 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]542 /* paranoid */ #endif
[6a490b2]543 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
[ac12f1f]544 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]545
[b808625]546 const bool local = thrd->state != Start;
[ae7be7a]547 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]548
[ec43cf9]549 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
550 struct cluster * cl = thrd->curr_cluster;
[89eff25]551 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]552
[254ad1b]553 // push the thread to the cluster ready-queue
[b808625]554 push( cl, thrd, local );
[32a8b61]555
[254ad1b]556 // variable thrd is no longer safe to use
[734908c]557 thrd = 0xdeaddeaddeaddeadp;
[32a8b61]558
[254ad1b]559 // wake the cluster using the save variable.
560 __wake_one( cl );
[1c273d0]561
[ec43cf9]562 #if !defined(__CFA_NO_STATISTICS__)
563 if( kernelTLS().this_stats ) {
564 __tls_stats()->ready.threads.threads++;
[89eff25]565 if(outside) {
566 __tls_stats()->ready.threads.extunpark++;
567 }
[73f4d08]568 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
[ec43cf9]569 }
570 else {
571 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]572 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[73f4d08]573 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
[ec43cf9]574 }
575 #endif
576
[254ad1b]577 /* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]578 /* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]579}
580
[254ad1b]581void schedule_thread$( $thread * thrd ) {
582 ready_schedule_lock();
583 __schedule_thread( thrd );
584 ready_schedule_unlock();
585}
586
[14a61b5]587// KERNEL ONLY
[1eb239e4]588static inline $thread * __next_thread(cluster * this) with( *this ) {
[8fc652e0]589 /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]590
[e873838]591 ready_schedule_lock();
[431cd4f]592 $thread * thrd = pop_fast( this );
[e873838]593 ready_schedule_unlock();
[7768b8d]594
[8fc652e0]595 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]596 return thrd;
[eb2e723]597}
598
[64a7146]599// KERNEL ONLY
[1eb239e4]600static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]601 /* paranoid */ verify( ! __preemption_enabled() );
[64a7146]602
[e873838]603 ready_schedule_lock();
[fc59df78]604 $thread * thrd;
605 for(25) {
606 thrd = pop_slow( this );
607 if(thrd) goto RET;
608 }
609 thrd = pop_search( this );
610
611 RET:
[e873838]612 ready_schedule_unlock();
[64a7146]613
[8fc652e0]614 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]615 return thrd;
[64a7146]616}
617
[c6c7e6c]618static inline bool __must_unpark( $thread * thrd ) {
[ff79d5e]619 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
620 switch(old_ticket) {
[6a77224]621 case TICKET_RUNNING:
[3381ed7]622 // Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]623 return false;
[6a77224]624 case TICKET_BLOCKED:
[3381ed7]625 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]626 /* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]627 return true;
[3381ed7]628 default:
629 // This makes no sense, something is wrong abort
[7ee8153]630 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]631 }
[eb2e723]632}
633
[e9c0b4c]634void __kernel_unpark( $thread * thrd ) {
635 /* paranoid */ verify( ! __preemption_enabled() );
636 /* paranoid */ verify( ready_schedule_islocked());
637
638 if( !thrd ) return;
639
640 if(__must_unpark(thrd)) {
641 // Wake lost the race,
642 __schedule_thread( thrd );
643 }
644
645 /* paranoid */ verify( ready_schedule_islocked());
646 /* paranoid */ verify( ! __preemption_enabled() );
647}
648
[c6c7e6c]649void unpark( $thread * thrd ) {
650 if( !thrd ) return;
[0b33412]651
[c6c7e6c]652 if(__must_unpark(thrd)) {
653 disable_interrupts();
[a3821fa]654 // Wake lost the race,
[254ad1b]655 schedule_thread$( thrd );
[a3821fa]656 enable_interrupts(false);
[de6319f]657 }
[eb2e723]658}
[0b33412]659
[e235429]660void park( void ) {
[a3821fa]661 __disable_interrupts_checked();
662 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
663 returnToKernel();
664 __enable_interrupts_checked();
[0c78741]665
666}
[09800e9]667
[5afb49a]668extern "C" {
669 // Leave the thread monitor
670 // last routine called by a thread.
671 // Should never return
672 void __cfactx_thrd_leave() {
[8fc652e0]673 $thread * thrd = active_thread();
[5afb49a]674 $monitor * this = &thrd->self_mon;
675
676 // Lock the monitor now
677 lock( this->lock __cfaabi_dbg_ctx2 );
678
679 disable_interrupts();
680
[9d6e1b8a]681 /* paranoid */ verify( ! __preemption_enabled() );
682 /* paranoid */ verify( thrd->state == Active );
683 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
684 /* paranoid */ verify( kernelTLS().this_thread == thrd );
685 /* paranoid */ verify( thrd->context.SP );
686 /* 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 );
687 /* 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 );
688
[3ea8ad1]689 thrd->state = Halting;
[58688bf]690 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]691 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
692 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]693
694 // Leave the thread
695 returnToKernel();
696
697 // Control flow should never reach here!
[9d6e1b8a]698 abort();
[5afb49a]699 }
[09800e9]700}
701
[14a61b5]702// KERNEL ONLY
[3381ed7]703bool force_yield( __Preemption_Reason reason ) {
[a3821fa]704 __disable_interrupts_checked();
705 $thread * thrd = kernelTLS().this_thread;
706 /* paranoid */ verify(thrd->state == Active);
707
708 // SKULLDUGGERY: It is possible that we are preempting this thread just before
709 // it was going to park itself. If that is the case and it is already using the
710 // intrusive fields then we can't use them to preempt the thread
711 // If that is the case, abandon the preemption.
712 bool preempted = false;
713 if(thrd->link.next == 0p) {
714 preempted = true;
715 thrd->preempted = reason;
716 returnToKernel();
717 }
718 __enable_interrupts_checked( false );
[3381ed7]719 return preempted;
[f2b12406]720}
721
[14a61b5]722//=============================================================================================
[92e7631]723// Kernel Idle Sleep
[14a61b5]724//=============================================================================================
[64a7146]725// Wake a thread from the front if there are any
[e873838]726static void __wake_one(cluster * this) {
[8fc652e0]727 /* paranoid */ verify( ! __preemption_enabled() );
[e873838]728 /* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]729
[64a7146]730 // Check if there is a sleeping processor
[1eb239e4]731 processor * p;
732 unsigned idle;
733 unsigned total;
[6a9b12b]734 [idle, total, p] = query_idles(this->procs);
[14a61b5]735
[64a7146]736 // If no one is sleeping, we are done
[1eb239e4]737 if( idle == 0 ) return;
[14a61b5]738
[64a7146]739 // We found a processor, wake it up
[dddb3dd0]740 eventfd_t val;
741 val = 1;
742 eventfd_write( p->idle, val );
[14a61b5]743
[1eb239e4]744 #if !defined(__CFA_NO_STATISTICS__)
[5cb51502]745 if( kernelTLS().this_stats ) {
746 __tls_stats()->ready.sleep.wakes++;
747 }
748 else {
749 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
750 }
[1eb239e4]751 #endif
752
[e873838]753 /* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]754 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]755
756 return;
[64a7146]757}
[14a61b5]758
[64a7146]759// Unconditionnaly wake a thread
[1eb239e4]760void __wake_proc(processor * this) {
[64a7146]761 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]762
[a3821fa]763 __disable_interrupts_checked();
[8fc652e0]764 /* paranoid */ verify( ! __preemption_enabled() );
[dddb3dd0]765 eventfd_t val;
766 val = 1;
[55d6affb]767 eventfd_write( this->idle, val );
[a3821fa]768 __enable_interrupts_checked();
[92e7631]769}
770
[6a9b12b]771static void mark_idle(__cluster_proc_list & this, processor & proc) {
[8fc652e0]772 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]773 lock( this );
774 this.idle++;
775 /* paranoid */ verify( this.idle <= this.total );
[fc59b580]776 remove(proc);
[6a9b12b]777 insert_first(this.idles, proc);
[1eb239e4]778 unlock( this );
[8fc652e0]779 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]780}
[8e16177]781
[6a9b12b]782static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]783 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]784 lock( this );
785 this.idle--;
786 /* paranoid */ verify( this.idle >= 0 );
787 remove(proc);
[fc59b580]788 insert_last(this.actives, proc);
[1eb239e4]789 unlock( this );
[8fc652e0]790 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]791}
[c34ebf2]792
[6a9b12b]793static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
794 /* paranoid */ verify( ! __preemption_enabled() );
795 /* paranoid */ verify( ready_schedule_islocked() );
796
[1eb239e4]797 for() {
798 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
799 if( 1 == (l % 2) ) { Pause(); continue; }
800 unsigned idle = this.idle;
801 unsigned total = this.total;
[6a9b12b]802 processor * proc = &this.idles`first;
[7fdae38]803 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
804 asm volatile("": : :"memory");
[1eb239e4]805 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
806 return [idle, total, proc];
807 }
[6a9b12b]808
809 /* paranoid */ verify( ready_schedule_islocked() );
810 /* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]811}
812
[dbe9b08]813//=============================================================================================
814// Unexpected Terminating logic
815//=============================================================================================
[92bfda0]816void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
817 $thread * thrd = __cfaabi_tls.this_thread;
[9d944b2]818
[de94a60]819 if(thrd) {
820 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]821 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]822
[212c2187]823 if ( &thrd->self_cor != thrd->curr_cor ) {
824 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]825 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]826 }
827 else {
[1c40091]828 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]829 }
[1c273d0]830 }
[9d944b2]831 else {
[de94a60]832 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]833 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]834 }
835}
836
[92bfda0]837int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
838 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]839}
840
[de94a60]841static __spinlock_t kernel_debug_lock;
842
[9d944b2]843extern "C" {
[1c40091]844 void __cfaabi_bits_acquire() {
[36982fc]845 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]846 }
847
[1c40091]848 void __cfaabi_bits_release() {
[ea7d2b0]849 unlock( kernel_debug_lock );
[9d944b2]850 }
[8118303]851}
852
[fa21ac9]853//=============================================================================================
854// Kernel Utilities
855//=============================================================================================
[dddb3dd0]856#if defined(CFA_HAVE_LINUX_IO_URING_H)
857#include "io/types.hfa"
858#endif
859
[c1c95b1]860static inline bool __maybe_io_drain( processor * proc ) {
[e9c0b4c]861 bool ret = false;
[dddb3dd0]862 #if defined(CFA_HAVE_LINUX_IO_URING_H)
863 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
864
865 // Check if we should drain the queue
866 $io_context * ctx = proc->io.ctx;
867 unsigned head = *ctx->cq.head;
868 unsigned tail = *ctx->cq.tail;
[c1c95b1]869 if(head == tail) return false;
[57f70ab]870 #if OLD_MAIN
[b7fd2db6]871 ready_schedule_lock();
[e9c0b4c]872 ret = __cfa_io_drain( proc );
[b7fd2db6]873 ready_schedule_unlock();
[57f70ab]874 #else
875 ret = __cfa_io_drain( proc );
876 #endif
[dddb3dd0]877 #endif
[e9c0b4c]878 return ret;
[dddb3dd0]879}
880
[de94a60]881//-----------------------------------------------------------------------------
882// Debug
883__cfaabi_dbg_debug_do(
[1997b4e]884 extern "C" {
[ae66348]885 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
[1997b4e]886 this.prev_name = prev_name;
[8fc652e0]887 this.prev_thrd = kernelTLS().this_thread;
[1997b4e]888 }
[9181f1d]889 }
[f7d6bb0]890)
[2026bb6]891
892//-----------------------------------------------------------------------------
893// Debug
[8c50aed]894bool threading_enabled(void) __attribute__((const)) {
[2026bb6]895 return true;
896}
[c34ebf2]897
898//-----------------------------------------------------------------------------
899// Statistics
900#if !defined(__CFA_NO_STATISTICS__)
901 void print_halts( processor & this ) {
902 this.print_halts = true;
903 }
[58688bf]904
[69914cbc]905 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
[8464edf]906 /* paranoid */ verify( cltr->stats );
907
908 processor * it = &list`first;
909 for(unsigned i = 0; i < count; i++) {
910 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
911 /* paranoid */ verify( it->local_data->this_stats );
912 __tally_stats( cltr->stats, it->local_data->this_stats );
913 it = &(*it)`next;
914 }
915 }
916
917 void crawl_cluster_stats( cluster & this ) {
918 // Stop the world, otherwise stats could get really messed-up
919 // this doesn't solve all problems but does solve many
920 // so it's probably good enough
921 uint_fast32_t last_size = ready_mutate_lock();
922
923 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
924 crawl_list(&this, this.procs.idles , this.procs.idle );
925
926 // Unlock the RWlock
927 ready_mutate_unlock( last_size );
928 }
929
930
[58688bf]931 void print_stats_now( cluster & this, int flags ) {
[8464edf]932 crawl_cluster_stats( this );
[1b033b8]933 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
934 }
[c34ebf2]935#endif
[8118303]936// Local Variables: //
937// mode: c //
938// tab-width: 4 //
939// End: //
Note: See TracBrowser for help on using the repository browser.