source: libcfa/src/concurrency/kernel.cfa@ 2b96031

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 2b96031 was c993b15, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Changed RW lock to avoid hitting the global array on schedule.

  • Property mode set to 100644
File size: 29.4 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
[89eff25]173 __STATS( unsigned long long last_tally = rdtscl(); )
[325e6ea]174
[a5e7233]175 // if we need to run some special setup, now is the time to do it.
176 if(this->init.thrd) {
177 this->init.thrd->curr_cluster = this->cltr;
178 __run_thread(this, this->init.thrd);
179 }
[325e6ea]180
[4069faad]181 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
[8118303]182
[ac2b598]183 $thread * readyThread = 0p;
[1eb239e4]184 MAIN_LOOP:
185 for() {
[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
260 #if !defined(__CFA_NO_STATISTICS__)
261 unsigned long long curr = rdtscl();
262 if(curr > (last_tally + 500000000)) {
263 __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
264 last_tally = curr;
265 }
266 #endif
[dddb3dd0]267
268 if(this->io.pending && !this->io.dirty) {
269 __cfa_io_flush( this );
270 }
[89eff25]271
272 // SEARCH: {
273 // /* paranoid */ verify( ! __preemption_enabled() );
274 // /* paranoid */ verify( kernelTLS().this_proc_id );
275
276 // // First, lock the scheduler since we are searching for a thread
277
278 // // Try to get the next thread
279 // ready_schedule_lock();
280 // readyThread = pop_fast( this->cltr );
281 // ready_schedule_unlock();
282 // if(readyThread) { break SEARCH; }
283
284 // // If we can't find a thread, might as well flush any outstanding I/O
285 // if(this->io.pending) { __cfa_io_flush( this ); }
286
287 // // Spin a little on I/O, just in case
288 // for(25) {
289 // __maybe_io_drain( this );
290 // ready_schedule_lock();
291 // readyThread = pop_fast( this->cltr );
292 // ready_schedule_unlock();
293 // if(readyThread) { break SEARCH; }
294 // }
295
296 // // no luck, try stealing a few times
297 // for(25) {
298 // if( __maybe_io_drain( this ) ) {
299 // ready_schedule_lock();
300 // readyThread = pop_fast( this->cltr );
301 // } else {
302 // ready_schedule_lock();
303 // readyThread = pop_slow( this->cltr );
304 // }
305 // ready_schedule_unlock();
306 // if(readyThread) { break SEARCH; }
307 // }
308
309 // // still no luck, search for a thread
310 // ready_schedule_lock();
311 // readyThread = pop_search( this->cltr );
312 // ready_schedule_unlock();
313 // if(readyThread) { break SEARCH; }
314
315 // // Don't block if we are done
316 // if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
317
318 // __STATS( __tls_stats()->ready.sleep.halts++; )
319
320 // // Push self to idle stack
321 // mark_idle(this->cltr->procs, * this);
322
323 // // Confirm the ready-queue is empty
324 // __maybe_io_drain( this );
325 // ready_schedule_lock();
326 // readyThread = pop_search( this->cltr );
327 // ready_schedule_unlock();
328
329 // if( readyThread ) {
330 // // A thread was found, cancel the halt
331 // mark_awake(this->cltr->procs, * this);
332
333 // __STATS( __tls_stats()->ready.sleep.cancels++; )
334
335 // // continue the main loop
336 // break SEARCH;
337 // }
338
339 // __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl()); )
340 // __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
341
342 // // __disable_interrupts_hard();
343 // eventfd_t val;
344 // eventfd_read( this->idle, &val );
345 // // __enable_interrupts_hard();
346
347 // __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl()); )
348
349 // // We were woken up, remove self from idle
350 // mark_awake(this->cltr->procs, * this);
351
352 // // DON'T just proceed, start looking again
353 // continue MAIN_LOOP;
354 // }
355
356 // RUN_THREAD:
357 // /* paranoid */ verify( kernelTLS().this_proc_id );
358 // /* paranoid */ verify( ! __preemption_enabled() );
359 // /* paranoid */ verify( readyThread );
360
361 // // Reset io dirty bit
362 // this->io.dirty = false;
363
364 // // We found a thread run it
365 // __run_thread(this, readyThread);
366
367 // // Are we done?
368 // if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
369
370 // #if !defined(__CFA_NO_STATISTICS__)
371 // unsigned long long curr = rdtscl();
372 // if(curr > (last_tally + 500000000)) {
373 // __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
374 // last_tally = curr;
375 // }
376 // #endif
377
378 // if(this->io.pending && !this->io.dirty) {
379 // __cfa_io_flush( this );
380 // }
381
382 // // Check if there is pending io
383 // __maybe_io_drain( this );
[c81ebf9]384 }
385
[4069faad]386 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
[c84e80a]387 }
[8118303]388
[dddb3dd0]389 __cfa_io_stop( this );
390
[454f478]391 post( this->terminated );
[bdeba0b]392
[28d73c1]393 if(this == mainProcessor) {
[6a490b2]394 // HACK : the coroutine context switch expects this_thread to be set
395 // and it make sense for it to be set in all other cases except here
396 // fake it
[8fc652e0]397 __cfaabi_tls.this_thread = mainThread;
[6a490b2]398 }
[7768b8d]399
[4069faad]400 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]401}
402
[5c1a531]403static int * __volatile_errno() __attribute__((noinline));
404static int * __volatile_errno() { asm(""); return &errno; }
405
[14a61b5]406// KERNEL ONLY
[1c273d0]407// runThread runs a thread by context switching
408// from the processor coroutine to the target thread
[ac2b598]409static void __run_thread(processor * this, $thread * thrd_dst) {
[8fc652e0]410 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]411 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
412 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
413 __builtin_prefetch( thrd_dst->context.SP );
414
[dddb3dd0]415 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
416
[ac2b598]417 $coroutine * proc_cor = get_coroutine(this->runner);
[8fcbb4c]418
[9f575ea]419 // set state of processor coroutine to inactive
420 verify(proc_cor->state == Active);
[ae7be7a]421 proc_cor->state = Blocked;
[e8e457e]422
[9f575ea]423 // Actually run the thread
[3381ed7]424 RUNNING: while(true) {
[ff79d5e]425 thrd_dst->preempted = __NO_PREEMPTION;
426 thrd_dst->state = Active;
[e8e457e]427
[58d64a4]428 // Update global state
[8fc652e0]429 kernelTLS().this_thread = thrd_dst;
[75f3522]430
[8fc652e0]431 /* paranoid */ verify( ! __preemption_enabled() );
432 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]433 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]434 /* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]435 /* paranoid */ verify( thrd_dst->state != Halted );
[210b8b3]436 /* 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
437 /* 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]438 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]439
[3381ed7]440
[58d64a4]441
[9f575ea]442 // set context switch to the thread that the processor is executing
[c7a900a]443 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
444 // when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]445
[ac12f1f]446 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[210b8b3]447 /* 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 );
448 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
[b4b63e8]449 /* paranoid */ verify( thrd_dst->context.SP );
[9d6e1b8a]450 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[8fc652e0]451 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
452 /* paranoid */ verify( ! __preemption_enabled() );
[75f3522]453
[58d64a4]454 // Reset global state
[8fc652e0]455 kernelTLS().this_thread = 0p;
[3381ed7]456
457 // We just finished running a thread, there are a few things that could have happened.
458 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
459 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
460 // 4 - Preempted
461 // In case 1, we may have won a race so we can't write to the state again.
462 // In case 2, we lost the race so we now own the thread.
463
464 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
465 // The thread was preempted, reschedule it and reset the flag
[254ad1b]466 schedule_thread$( thrd_dst );
[3381ed7]467 break RUNNING;
468 }
[75f3522]469
[3ea8ad1]470 if(unlikely(thrd_dst->state == Halting)) {
[ff79d5e]471 // The thread has halted, it should never be scheduled/run again
[5afb49a]472 // finish the thread
473 __thread_finish( thrd_dst );
[ff79d5e]474 break RUNNING;
475 }
476
477 /* paranoid */ verify( thrd_dst->state == Active );
478 thrd_dst->state = Blocked;
479
[3381ed7]480 // set state of processor coroutine to active and the thread to inactive
[ff79d5e]481 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
482 switch(old_ticket) {
[6a77224]483 case TICKET_RUNNING:
[3381ed7]484 // This is case 1, the regular case, nothing more is needed
485 break RUNNING;
[6a77224]486 case TICKET_UNBLOCK:
[ec43cf9]487 #if !defined(__CFA_NO_STATISTICS__)
488 __tls_stats()->ready.threads.threads++;
[73f4d08]489 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]490 #endif
[3381ed7]491 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
492 // In this case, just run it again.
493 continue RUNNING;
494 default:
495 // This makes no sense, something is wrong abort
[ff79d5e]496 abort();
[3381ed7]497 }
[9f575ea]498 }
[e8e457e]499
[9f575ea]500 // Just before returning to the processor, set the processor coroutine to active
[e8e457e]501 proc_cor->state = Active;
[1eb239e4]502
[dddb3dd0]503 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
504
[ec43cf9]505 #if !defined(__CFA_NO_STATISTICS__)
506 __tls_stats()->ready.threads.threads--;
[73f4d08]507 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
[ec43cf9]508 #endif
509
[8fc652e0]510 /* paranoid */ verify( ! __preemption_enabled() );
[82c948c]511}
512
[14a61b5]513// KERNEL_ONLY
[b0c7419]514void returnToKernel() {
[8fc652e0]515 /* paranoid */ verify( ! __preemption_enabled() );
516 $coroutine * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
517 $thread * thrd_src = kernelTLS().this_thread;
[e8e457e]518
[89eff25]519 __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]520
[9f575ea]521 // Run the thread on this processor
522 {
523 int local_errno = *__volatile_errno();
524 #if defined( __i386 ) || defined( __x86_64 )
525 __x87_store;
526 #endif
[b4b63e8]527 /* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]528 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]529 __cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]530 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]531 #if defined( __i386 ) || defined( __x86_64 )
532 __x87_load;
533 #endif
534 *__volatile_errno() = local_errno;
[8fcbb4c]535 }
[deca0f5]536
[29cb302]537 #if !defined(__CFA_NO_STATISTICS__)
[89eff25]538 /* paranoid */ verify( thrd_src->last_proc != 0p );
539 if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]540 __tls_stats()->ready.threads.migration++;
541 }
542 #endif
543
[8fc652e0]544 /* paranoid */ verify( ! __preemption_enabled() );
[210b8b3]545 /* 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 );
546 /* 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]547}
548
[8def349]549//-----------------------------------------------------------------------------
550// Scheduler routines
[14a61b5]551// KERNEL ONLY
[e9c0b4c]552static void __schedule_thread( $thread * thrd ) {
[8fc652e0]553 /* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]554 /* paranoid */ verify( ready_schedule_islocked());
[6a490b2]555 /* paranoid */ verify( thrd );
556 /* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]557 /* paranoid */ verify( thrd->curr_cluster );
[3381ed7]558 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]559 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
560 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]561 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]562 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]563 /* paranoid */ #endif
[6a490b2]564 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
[ac12f1f]565 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]566
[9f575ea]567
[ae7be7a]568 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]569
[ec43cf9]570 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
571 struct cluster * cl = thrd->curr_cluster;
[89eff25]572 __STATS(bool outside = thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]573
[254ad1b]574 // push the thread to the cluster ready-queue
575 push( cl, thrd );
[32a8b61]576
[254ad1b]577 // variable thrd is no longer safe to use
[734908c]578 thrd = 0xdeaddeaddeaddeadp;
[32a8b61]579
[254ad1b]580 // wake the cluster using the save variable.
581 __wake_one( cl );
[1c273d0]582
[ec43cf9]583 #if !defined(__CFA_NO_STATISTICS__)
584 if( kernelTLS().this_stats ) {
585 __tls_stats()->ready.threads.threads++;
[89eff25]586 if(outside) {
587 __tls_stats()->ready.threads.extunpark++;
588 }
[73f4d08]589 __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
[ec43cf9]590 }
591 else {
592 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]593 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[73f4d08]594 __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
[ec43cf9]595 }
596 #endif
597
[254ad1b]598 /* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]599 /* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]600}
601
[254ad1b]602void schedule_thread$( $thread * thrd ) {
603 ready_schedule_lock();
604 __schedule_thread( thrd );
605 ready_schedule_unlock();
606}
607
[14a61b5]608// KERNEL ONLY
[1eb239e4]609static inline $thread * __next_thread(cluster * this) with( *this ) {
[8fc652e0]610 /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]611
[e873838]612 ready_schedule_lock();
[431cd4f]613 $thread * thrd = pop_fast( this );
[e873838]614 ready_schedule_unlock();
[7768b8d]615
[8fc652e0]616 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]617 return thrd;
[eb2e723]618}
619
[64a7146]620// KERNEL ONLY
[1eb239e4]621static inline $thread * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]622 /* paranoid */ verify( ! __preemption_enabled() );
[64a7146]623
[e873838]624 ready_schedule_lock();
[fc59df78]625 $thread * thrd;
626 for(25) {
627 thrd = pop_slow( this );
628 if(thrd) goto RET;
629 }
630 thrd = pop_search( this );
631
632 RET:
[e873838]633 ready_schedule_unlock();
[64a7146]634
[8fc652e0]635 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]636 return thrd;
[64a7146]637}
638
[c6c7e6c]639static inline bool __must_unpark( $thread * thrd ) {
[ff79d5e]640 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
641 switch(old_ticket) {
[6a77224]642 case TICKET_RUNNING:
[3381ed7]643 // Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]644 return false;
[6a77224]645 case TICKET_BLOCKED:
[3381ed7]646 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]647 /* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]648 return true;
[3381ed7]649 default:
650 // This makes no sense, something is wrong abort
[7ee8153]651 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]652 }
[eb2e723]653}
654
[e9c0b4c]655void __kernel_unpark( $thread * thrd ) {
656 /* paranoid */ verify( ! __preemption_enabled() );
657 /* paranoid */ verify( ready_schedule_islocked());
658
659 if( !thrd ) return;
660
661 if(__must_unpark(thrd)) {
662 // Wake lost the race,
663 __schedule_thread( thrd );
664 }
665
666 /* paranoid */ verify( ready_schedule_islocked());
667 /* paranoid */ verify( ! __preemption_enabled() );
668}
669
[c6c7e6c]670void unpark( $thread * thrd ) {
671 if( !thrd ) return;
[0b33412]672
[c6c7e6c]673 if(__must_unpark(thrd)) {
674 disable_interrupts();
[a3821fa]675 // Wake lost the race,
[254ad1b]676 schedule_thread$( thrd );
[a3821fa]677 enable_interrupts(false);
[de6319f]678 }
[eb2e723]679}
[0b33412]680
[e235429]681void park( void ) {
[a3821fa]682 __disable_interrupts_checked();
683 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
684 returnToKernel();
685 __enable_interrupts_checked();
[0c78741]686
687}
[09800e9]688
[5afb49a]689extern "C" {
690 // Leave the thread monitor
691 // last routine called by a thread.
692 // Should never return
693 void __cfactx_thrd_leave() {
[8fc652e0]694 $thread * thrd = active_thread();
[5afb49a]695 $monitor * this = &thrd->self_mon;
696
697 // Lock the monitor now
698 lock( this->lock __cfaabi_dbg_ctx2 );
699
700 disable_interrupts();
701
[9d6e1b8a]702 /* paranoid */ verify( ! __preemption_enabled() );
703 /* paranoid */ verify( thrd->state == Active );
704 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
705 /* paranoid */ verify( kernelTLS().this_thread == thrd );
706 /* paranoid */ verify( thrd->context.SP );
707 /* 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 );
708 /* 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 );
709
[3ea8ad1]710 thrd->state = Halting;
[58688bf]711 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]712 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
713 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]714
715 // Leave the thread
716 returnToKernel();
717
718 // Control flow should never reach here!
[9d6e1b8a]719 abort();
[5afb49a]720 }
[09800e9]721}
722
[14a61b5]723// KERNEL ONLY
[3381ed7]724bool force_yield( __Preemption_Reason reason ) {
[a3821fa]725 __disable_interrupts_checked();
726 $thread * thrd = kernelTLS().this_thread;
727 /* paranoid */ verify(thrd->state == Active);
728
729 // SKULLDUGGERY: It is possible that we are preempting this thread just before
730 // it was going to park itself. If that is the case and it is already using the
731 // intrusive fields then we can't use them to preempt the thread
732 // If that is the case, abandon the preemption.
733 bool preempted = false;
734 if(thrd->link.next == 0p) {
735 preempted = true;
736 thrd->preempted = reason;
737 returnToKernel();
738 }
739 __enable_interrupts_checked( false );
[3381ed7]740 return preempted;
[f2b12406]741}
742
[14a61b5]743//=============================================================================================
[92e7631]744// Kernel Idle Sleep
[14a61b5]745//=============================================================================================
[64a7146]746// Wake a thread from the front if there are any
[e873838]747static void __wake_one(cluster * this) {
[8fc652e0]748 /* paranoid */ verify( ! __preemption_enabled() );
[e873838]749 /* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]750
[64a7146]751 // Check if there is a sleeping processor
[1eb239e4]752 processor * p;
753 unsigned idle;
754 unsigned total;
[6a9b12b]755 [idle, total, p] = query_idles(this->procs);
[14a61b5]756
[64a7146]757 // If no one is sleeping, we are done
[1eb239e4]758 if( idle == 0 ) return;
[14a61b5]759
[64a7146]760 // We found a processor, wake it up
[dddb3dd0]761 eventfd_t val;
762 val = 1;
763 eventfd_write( p->idle, val );
[14a61b5]764
[1eb239e4]765 #if !defined(__CFA_NO_STATISTICS__)
[5cb51502]766 if( kernelTLS().this_stats ) {
767 __tls_stats()->ready.sleep.wakes++;
768 }
769 else {
770 __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED);
771 }
[1eb239e4]772 #endif
773
[e873838]774 /* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]775 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]776
777 return;
[64a7146]778}
[14a61b5]779
[64a7146]780// Unconditionnaly wake a thread
[1eb239e4]781void __wake_proc(processor * this) {
[64a7146]782 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]783
[a3821fa]784 __disable_interrupts_checked();
[8fc652e0]785 /* paranoid */ verify( ! __preemption_enabled() );
[dddb3dd0]786 eventfd_t val;
787 val = 1;
[55d6affb]788 eventfd_write( this->idle, val );
[a3821fa]789 __enable_interrupts_checked();
[92e7631]790}
791
[6a9b12b]792static void mark_idle(__cluster_proc_list & this, processor & proc) {
[8fc652e0]793 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]794 lock( this );
795 this.idle++;
796 /* paranoid */ verify( this.idle <= this.total );
[fc59b580]797 remove(proc);
[6a9b12b]798 insert_first(this.idles, proc);
[1eb239e4]799 unlock( this );
[8fc652e0]800 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]801}
[8e16177]802
[6a9b12b]803static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]804 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]805 lock( this );
806 this.idle--;
807 /* paranoid */ verify( this.idle >= 0 );
808 remove(proc);
[fc59b580]809 insert_last(this.actives, proc);
[1eb239e4]810 unlock( this );
[8fc652e0]811 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]812}
[c34ebf2]813
[6a9b12b]814static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
815 /* paranoid */ verify( ! __preemption_enabled() );
816 /* paranoid */ verify( ready_schedule_islocked() );
817
[1eb239e4]818 for() {
819 uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
820 if( 1 == (l % 2) ) { Pause(); continue; }
821 unsigned idle = this.idle;
822 unsigned total = this.total;
[6a9b12b]823 processor * proc = &this.idles`first;
[7fdae38]824 // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
825 asm volatile("": : :"memory");
[1eb239e4]826 if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
827 return [idle, total, proc];
828 }
[6a9b12b]829
830 /* paranoid */ verify( ready_schedule_islocked() );
831 /* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]832}
833
[dbe9b08]834//=============================================================================================
835// Unexpected Terminating logic
836//=============================================================================================
[92bfda0]837void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
838 $thread * thrd = __cfaabi_tls.this_thread;
[9d944b2]839
[de94a60]840 if(thrd) {
841 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]842 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]843
[212c2187]844 if ( &thrd->self_cor != thrd->curr_cor ) {
845 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]846 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]847 }
848 else {
[1c40091]849 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]850 }
[1c273d0]851 }
[9d944b2]852 else {
[de94a60]853 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]854 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]855 }
856}
857
[92bfda0]858int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
859 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]860}
861
[de94a60]862static __spinlock_t kernel_debug_lock;
863
[9d944b2]864extern "C" {
[1c40091]865 void __cfaabi_bits_acquire() {
[36982fc]866 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]867 }
868
[1c40091]869 void __cfaabi_bits_release() {
[ea7d2b0]870 unlock( kernel_debug_lock );
[9d944b2]871 }
[8118303]872}
873
[fa21ac9]874//=============================================================================================
875// Kernel Utilities
876//=============================================================================================
[dddb3dd0]877#if defined(CFA_HAVE_LINUX_IO_URING_H)
878#include "io/types.hfa"
879#endif
880
[c1c95b1]881static inline bool __maybe_io_drain( processor * proc ) {
[e9c0b4c]882 bool ret = false;
[dddb3dd0]883 #if defined(CFA_HAVE_LINUX_IO_URING_H)
884 __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
885
886 // Check if we should drain the queue
887 $io_context * ctx = proc->io.ctx;
888 unsigned head = *ctx->cq.head;
889 unsigned tail = *ctx->cq.tail;
[c1c95b1]890 if(head == tail) return false;
[b7fd2db6]891 ready_schedule_lock();
[e9c0b4c]892 ret = __cfa_io_drain( proc );
[b7fd2db6]893 ready_schedule_unlock();
[dddb3dd0]894 #endif
[e9c0b4c]895 return ret;
[dddb3dd0]896}
897
[de94a60]898//-----------------------------------------------------------------------------
899// Debug
900__cfaabi_dbg_debug_do(
[1997b4e]901 extern "C" {
[ae66348]902 void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
[1997b4e]903 this.prev_name = prev_name;
[8fc652e0]904 this.prev_thrd = kernelTLS().this_thread;
[1997b4e]905 }
[9181f1d]906 }
[f7d6bb0]907)
[2026bb6]908
909//-----------------------------------------------------------------------------
910// Debug
[8c50aed]911bool threading_enabled(void) __attribute__((const)) {
[2026bb6]912 return true;
913}
[c34ebf2]914
915//-----------------------------------------------------------------------------
916// Statistics
917#if !defined(__CFA_NO_STATISTICS__)
918 void print_halts( processor & this ) {
919 this.print_halts = true;
920 }
[58688bf]921
922 void print_stats_now( cluster & this, int flags ) {
[1b033b8]923 __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
924 }
[c34ebf2]925#endif
[8118303]926// Local Variables: //
927// mode: c //
928// tab-width: 4 //
929// End: //
Note: See TracBrowser for help on using the repository browser.