source: libcfa/src/concurrency/kernel.cfa@ 78de1e5

ADT ast-experimental
Last change on this file since 78de1e5 was a757ba1, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Removed wait-with-io_uring feature which was slow, broken and ifdefed out.

  • Property mode set to 100644
File size: 28.6 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__
[43784ac]17#define _GNU_SOURCE
18
[4069faad]19// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__
[2026bb6]20
[bfb9bf5]21#pragma GCC diagnostic push
22#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
23
[8118303]24//C Includes
[214e8da]25#include <errno.h>
[9d944b2]26#include <stdio.h>
[445f984]27#include <string.h>
[58b6d1b]28#include <signal.h>
[9d944b2]29#include <unistd.h>
[bfb9bf5]30
[dddb3dd0]31extern "C" {
32 #include <sys/eventfd.h>
[d3605f8]33 #include <sys/uio.h>
[dddb3dd0]34}
[8118303]35
36//CFA Includes
[708ae38]37#include "kernel/private.hfa"
[73abe95]38#include "preemption.hfa"
[445f984]39#include "strstream.hfa"
40#include "device/cpu.hfa"
[7ef162b2]41#include "io/types.hfa"
[8118303]42
43//Private includes
44#define __CFA_INVOKE_PRIVATE__
45#include "invoke.h"
[bfb9bf5]46#pragma GCC diagnostic pop
[8118303]47
[89eff25]48#if !defined(__CFA_NO_STATISTICS__)
[c9c1c1cb]49 #define __STATS_DEF( ...) __VA_ARGS__
[89eff25]50#else
[c9c1c1cb]51 #define __STATS_DEF( ...)
[89eff25]52#endif
[4069faad]53
[deca0f5]54//-----------------------------------------------------------------------------
55// Some assembly required
[1805b1b]56#if defined( __i386 )
[deca0f5]57 // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
58 // fcw : X87 FPU control word (preserved across function calls)
59 #define __x87_store \
60 uint32_t __mxcr; \
61 uint16_t __fcw; \
62 __asm__ volatile ( \
63 "stmxcsr %0\n" \
64 "fnstcw %1\n" \
65 : "=m" (__mxcr),\
66 "=m" (__fcw) \
67 )
68
69 #define __x87_load \
70 __asm__ volatile ( \
71 "fldcw %1\n" \
72 "ldmxcsr %0\n" \
73 ::"m" (__mxcr),\
74 "m" (__fcw) \
75 )
76
77#elif defined( __x86_64 )
78 #define __x87_store \
79 uint32_t __mxcr; \
80 uint16_t __fcw; \
81 __asm__ volatile ( \
82 "stmxcsr %0\n" \
83 "fnstcw %1\n" \
84 : "=m" (__mxcr),\
85 "=m" (__fcw) \
86 )
87
88 #define __x87_load \
89 __asm__ volatile ( \
90 "fldcw %1\n" \
91 "ldmxcsr %0\n" \
92 :: "m" (__mxcr),\
93 "m" (__fcw) \
94 )
95
[0190480]96#elif defined( __arm__ )
97 #define __x87_store
98 #define __x87_load
99
100#elif defined( __aarch64__ )
[74f5c83]101 #define __x87_store \
102 uint32_t __fpcntl[2]; \
103 __asm__ volatile ( \
104 "mrs x9, FPCR\n" \
105 "mrs x10, FPSR\n" \
106 "stp x9, x10, %0\n" \
107 : "=m" (__fpcntl) : : "x9", "x10" \
108 )
109
110 #define __x87_load \
111 __asm__ volatile ( \
112 "ldp x9, x10, %0\n" \
113 "msr FPSR, x10\n" \
114 "msr FPCR, x9\n" \
115 : "=m" (__fpcntl) : : "x9", "x10" \
116 )
117
[deca0f5]118#else
[0190480]119 #error unsupported hardware architecture
[deca0f5]120#endif
121
[e84ab3d]122extern thread$ * mainThread;
[e660761]123extern processor * mainProcessor;
[2ac095d]124
[92e7631]125//-----------------------------------------------------------------------------
126// Kernel Scheduling logic
[e84ab3d]127static thread$ * __next_thread(cluster * this);
128static thread$ * __next_thread_slow(cluster * this);
[c9c1c1cb]129static thread$ * __next_thread_search(cluster * this);
[e84ab3d]130static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
131static void __run_thread(processor * this, thread$ * dst);
[e873838]132static void __wake_one(cluster * cltr);
[1eb239e4]133
[18f7858]134static void idle_sleep(processor * proc);
[5f5a729]135static bool mark_idle (__cluster_proc_list & idles, processor & proc);
[6a9b12b]136static void mark_awake(__cluster_proc_list & idles, processor & proc);
[1eb239e4]137
[4479890]138extern bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1)));
[18f7858]139extern bool __cfa_io_flush( processor * ) __attribute__((nonnull (1)));
[dddb3dd0]140
[7ef162b2]141
[dddb3dd0]142extern void __disable_interrupts_hard();
143extern void __enable_interrupts_hard();
[c84e80a]144
[7ef162b2]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
[4069faad]160 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
[28d73c1]161 #if !defined(__CFA_NO_STATISTICS__)
162 if( this->print_halts ) {
[c993b15]163 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
[28d73c1]164 }
165 #endif
[b798713]166
[75f3522]167 {
[c81ebf9]168 // Setup preemption data
169 preemption_scope scope = { this };
170
[a5e7233]171 // if we need to run some special setup, now is the time to do it.
172 if(this->init.thrd) {
173 this->init.thrd->curr_cluster = this->cltr;
174 __run_thread(this, this->init.thrd);
175 }
[325e6ea]176
[4069faad]177 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
[8118303]178
[e84ab3d]179 thread$ * readyThread = 0p;
[1eb239e4]180 MAIN_LOOP:
181 for() {
[dddb3dd0]182 // Check if there is pending io
[18f7858]183 __cfa_io_drain( this );
[dddb3dd0]184
[92e7631]185 // Try to get the next thread
[8c50aed]186 readyThread = __next_thread( this->cltr );
[75f3522]187
[1eb239e4]188 if( !readyThread ) {
[18f7858]189 // there is no point in holding submissions if we are idle
[c9c1c1cb]190 __IO_STATS__(true, io.flush.idle++; )
[18f7858]191 __cfa_io_flush( this );
192
193 // drain again in case something showed up
194 __cfa_io_drain( this );
[c33c2af]195
[c9c1c1cb]196 readyThread = __next_thread( this->cltr );
197 }
198
199 if( !readyThread ) for(5) {
[1eb239e4]200 readyThread = __next_thread_slow( this->cltr );
[c9c1c1cb]201
202 if( readyThread ) break;
203
[18f7858]204 // It's unlikely we still I/O to submit, but the arbiter could
205 __IO_STATS__(true, io.flush.idle++; )
206 __cfa_io_flush( this );
207
208 // drain again in case something showed up
209 __cfa_io_drain( this );
[1eb239e4]210 }
[4e6fb8e]211
[1eb239e4]212 HALT:
213 if( !readyThread ) {
214 // Don't block if we are done
215 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[c81ebf9]216
[1eb239e4]217 // Push self to idle stack
[5f5a729]218 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
[398e8e9]219
[1eb239e4]220 // Confirm the ready-queue is empty
[c9c1c1cb]221 readyThread = __next_thread_search( this->cltr );
[1eb239e4]222 if( readyThread ) {
223 // A thread was found, cancel the halt
[6a9b12b]224 mark_awake(this->cltr->procs, * this);
[1eb239e4]225
[c9c1c1cb]226 __STATS__(true, ready.sleep.cancels++; )
[1eb239e4]227
228 // continue the mai loop
229 break HALT;
230 }
231
[18f7858]232 idle_sleep( this );
[1eb239e4]233
234 // We were woken up, remove self from idle
[6a9b12b]235 mark_awake(this->cltr->procs, * this);
[1eb239e4]236
237 // DON'T just proceed, start looking again
238 continue MAIN_LOOP;
[64a7146]239 }
[1eb239e4]240
241 /* paranoid */ verify( readyThread );
242
[dddb3dd0]243 // Reset io dirty bit
244 this->io.dirty = false;
245
[1eb239e4]246 // We found a thread run it
247 __run_thread(this, readyThread);
248
249 // Are we done?
250 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[325e6ea]251
[d529ad0]252 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
[c9c1c1cb]253 __IO_STATS__(true, io.flush.dirty++; )
[18f7858]254 __cfa_io_flush( this );
[dddb3dd0]255 }
[c81ebf9]256 }
257
[4069faad]258 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
[c84e80a]259 }
[8118303]260
[454f478]261 post( this->terminated );
[bdeba0b]262
[28d73c1]263 if(this == mainProcessor) {
[6a490b2]264 // HACK : the coroutine context switch expects this_thread to be set
265 // and it make sense for it to be set in all other cases except here
266 // fake it
[8fc652e0]267 __cfaabi_tls.this_thread = mainThread;
[6a490b2]268 }
[7768b8d]269
[4069faad]270 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]271}
272
[5c1a531]273static int * __volatile_errno() __attribute__((noinline));
274static int * __volatile_errno() { asm(""); return &errno; }
275
[14a61b5]276// KERNEL ONLY
[1c273d0]277// runThread runs a thread by context switching
278// from the processor coroutine to the target thread
[e84ab3d]279static void __run_thread(processor * this, thread$ * thrd_dst) {
[8fc652e0]280 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]281 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
[15c93d8]282 /* paranoid */ verifyf( thrd_dst->rdy_link.next == 0p, "Expected null got %p", thrd_dst->rdy_link.next );
[1eb239e4]283 __builtin_prefetch( thrd_dst->context.SP );
284
[dddb3dd0]285 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
286
[e84ab3d]287 coroutine$ * proc_cor = get_coroutine(this->runner);
[8fcbb4c]288
[9f575ea]289 // set state of processor coroutine to inactive
290 verify(proc_cor->state == Active);
[ae7be7a]291 proc_cor->state = Blocked;
[e8e457e]292
[9f575ea]293 // Actually run the thread
[3381ed7]294 RUNNING: while(true) {
[ff79d5e]295 thrd_dst->preempted = __NO_PREEMPTION;
[e8e457e]296
[58d64a4]297 // Update global state
[8fc652e0]298 kernelTLS().this_thread = thrd_dst;
[75f3522]299
[8bee858]300 // Update the state after setting this_thread
301 // so that the debugger can find all active threads
302 // in tls storage
303 thrd_dst->state = Active;
304
[8fc652e0]305 /* paranoid */ verify( ! __preemption_enabled() );
306 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]307 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]308 /* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]309 /* paranoid */ verify( thrd_dst->state != Halted );
[e84ab3d]310 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
311 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
[878cfcc]312 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
[ac12f1f]313 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]314
[3381ed7]315
[58d64a4]316
[9f575ea]317 // set context switch to the thread that the processor is executing
[c7a900a]318 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
319 // when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]320
[50871b4]321
322
[ac12f1f]323 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[878cfcc]324 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
[e84ab3d]325 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
326 /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
[878cfcc]327 /* paranoid */ verify( thrd_dst->state != Halted );
[b4b63e8]328 /* paranoid */ verify( thrd_dst->context.SP );
[8fc652e0]329 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
330 /* paranoid */ verify( ! __preemption_enabled() );
[75f3522]331
[3381ed7]332 // We just finished running a thread, there are a few things that could have happened.
333 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
334 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
335 // 4 - Preempted
336 // In case 1, we may have won a race so we can't write to the state again.
337 // In case 2, we lost the race so we now own the thread.
338
339 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
[8bee858]340 // Reset the this_thread now that we know
341 // the state isn't active anymore
342 kernelTLS().this_thread = 0p;
343
[3381ed7]344 // The thread was preempted, reschedule it and reset the flag
[24e321c]345 schedule_thread$( thrd_dst, UNPARK_LOCAL );
[3381ed7]346 break RUNNING;
347 }
[75f3522]348
[3ea8ad1]349 if(unlikely(thrd_dst->state == Halting)) {
[8bee858]350 // Reset the this_thread now that we know
351 // the state isn't active anymore
352 kernelTLS().this_thread = 0p;
353
[ff79d5e]354 // The thread has halted, it should never be scheduled/run again
[5afb49a]355 // finish the thread
356 __thread_finish( thrd_dst );
[ff79d5e]357 break RUNNING;
358 }
359
360 /* paranoid */ verify( thrd_dst->state == Active );
361 thrd_dst->state = Blocked;
362
[8bee858]363 // Reset the this_thread now that we know
364 // the state isn't active anymore
365 kernelTLS().this_thread = 0p;
366
[3381ed7]367 // set state of processor coroutine to active and the thread to inactive
[ff79d5e]368 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
369 switch(old_ticket) {
[6a77224]370 case TICKET_RUNNING:
[3381ed7]371 // This is case 1, the regular case, nothing more is needed
372 break RUNNING;
[6a77224]373 case TICKET_UNBLOCK:
[c9c1c1cb]374 __STATS__(true, ready.threads.threads++; )
[3381ed7]375 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
376 // In this case, just run it again.
377 continue RUNNING;
378 default:
379 // This makes no sense, something is wrong abort
[ff79d5e]380 abort();
[3381ed7]381 }
[9f575ea]382 }
[e8e457e]383
[9f575ea]384 // Just before returning to the processor, set the processor coroutine to active
[e8e457e]385 proc_cor->state = Active;
[1eb239e4]386
[dddb3dd0]387 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
388
[c9c1c1cb]389 __STATS__(true, ready.threads.threads--; )
[ec43cf9]390
[8fc652e0]391 /* paranoid */ verify( ! __preemption_enabled() );
[82c948c]392}
393
[14a61b5]394// KERNEL_ONLY
[c18bf9e]395static void returnToKernel() {
[8fc652e0]396 /* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]397 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
398 thread$ * thrd_src = kernelTLS().this_thread;
[e8e457e]399
[c9c1c1cb]400 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]401
[9f575ea]402 // Run the thread on this processor
403 {
404 int local_errno = *__volatile_errno();
405 #if defined( __i386 ) || defined( __x86_64 )
406 __x87_store;
407 #endif
[b4b63e8]408 /* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]409 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]410 __cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]411 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]412 #if defined( __i386 ) || defined( __x86_64 )
413 __x87_load;
414 #endif
415 *__volatile_errno() = local_errno;
[8fcbb4c]416 }
[deca0f5]417
[29cb302]418 #if !defined(__CFA_NO_STATISTICS__)
[89eff25]419 /* paranoid */ verify( thrd_src->last_proc != 0p );
420 if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]421 __tls_stats()->ready.threads.migration++;
422 }
423 #endif
424
[8fc652e0]425 /* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]426 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ) || thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_src );
427 /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit) || thrd_src->corctx_flag, "ERROR : Returning thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_src );
[c84e80a]428}
429
[8def349]430//-----------------------------------------------------------------------------
431// Scheduler routines
[14a61b5]432// KERNEL ONLY
[24e321c]433static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
[8fc652e0]434 /* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]435 /* paranoid */ verify( ready_schedule_islocked());
[6a490b2]436 /* paranoid */ verify( thrd );
437 /* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]438 /* paranoid */ verify( thrd->curr_cluster );
[3381ed7]439 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]440 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
441 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]442 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]443 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]444 /* paranoid */ #endif
[15c93d8]445 /* paranoid */ verifyf( thrd->rdy_link.next == 0p, "Expected null got %p", thrd->rdy_link.next );
[ac12f1f]446 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]447
[ae7be7a]448 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]449
[ec43cf9]450 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
451 struct cluster * cl = thrd->curr_cluster;
[c9c1c1cb]452 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]453
[254ad1b]454 // push the thread to the cluster ready-queue
[24e321c]455 push( cl, thrd, hint );
[32a8b61]456
[254ad1b]457 // variable thrd is no longer safe to use
[734908c]458 thrd = 0xdeaddeaddeaddeadp;
[32a8b61]459
[254ad1b]460 // wake the cluster using the save variable.
461 __wake_one( cl );
[1c273d0]462
[ec43cf9]463 #if !defined(__CFA_NO_STATISTICS__)
464 if( kernelTLS().this_stats ) {
465 __tls_stats()->ready.threads.threads++;
[89eff25]466 if(outside) {
467 __tls_stats()->ready.threads.extunpark++;
468 }
[ec43cf9]469 }
470 else {
471 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]472 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[ec43cf9]473 }
474 #endif
475
[254ad1b]476 /* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]477 /* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]478}
479
[24e321c]480void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
[254ad1b]481 ready_schedule_lock();
[24e321c]482 __schedule_thread( thrd, hint );
[254ad1b]483 ready_schedule_unlock();
484}
485
[14a61b5]486// KERNEL ONLY
[e84ab3d]487static inline thread$ * __next_thread(cluster * this) with( *this ) {
[8fc652e0]488 /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]489
[e873838]490 ready_schedule_lock();
[e84ab3d]491 thread$ * thrd = pop_fast( this );
[e873838]492 ready_schedule_unlock();
[7768b8d]493
[8fc652e0]494 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]495 return thrd;
[eb2e723]496}
497
[64a7146]498// KERNEL ONLY
[e84ab3d]499static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]500 /* paranoid */ verify( ! __preemption_enabled() );
[64a7146]501
[e873838]502 ready_schedule_lock();
[c9c1c1cb]503 thread$ * thrd = pop_slow( this );
504 ready_schedule_unlock();
505
506 /* paranoid */ verify( ! __preemption_enabled() );
507 return thrd;
508}
[fc59df78]509
[c9c1c1cb]510// KERNEL ONLY
511static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
512 /* paranoid */ verify( ! __preemption_enabled() );
513
514 ready_schedule_lock();
515 thread$ * thrd = pop_search( this );
[e873838]516 ready_schedule_unlock();
[64a7146]517
[8fc652e0]518 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]519 return thrd;
[64a7146]520}
521
[e84ab3d]522static inline bool __must_unpark( thread$ * thrd ) {
[ff79d5e]523 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
524 switch(old_ticket) {
[6a77224]525 case TICKET_RUNNING:
[3381ed7]526 // Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]527 return false;
[6a77224]528 case TICKET_BLOCKED:
[3381ed7]529 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]530 /* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]531 return true;
[3381ed7]532 default:
533 // This makes no sense, something is wrong abort
[7ee8153]534 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]535 }
[eb2e723]536}
537
[24e321c]538void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
[e9c0b4c]539 /* paranoid */ verify( ! __preemption_enabled() );
540 /* paranoid */ verify( ready_schedule_islocked());
541
542 if( !thrd ) return;
543
544 if(__must_unpark(thrd)) {
545 // Wake lost the race,
[24e321c]546 __schedule_thread( thrd, hint );
[e9c0b4c]547 }
548
549 /* paranoid */ verify( ready_schedule_islocked());
550 /* paranoid */ verify( ! __preemption_enabled() );
551}
552
[c18bf9e]553void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
[c6c7e6c]554 if( !thrd ) return;
[0b33412]555
[c6c7e6c]556 if(__must_unpark(thrd)) {
557 disable_interrupts();
[a3821fa]558 // Wake lost the race,
[24e321c]559 schedule_thread$( thrd, hint );
[a3821fa]560 enable_interrupts(false);
[de6319f]561 }
[eb2e723]562}
[0b33412]563
[c18bf9e]564void park( void ) libcfa_public {
[a3821fa]565 __disable_interrupts_checked();
566 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
567 returnToKernel();
568 __enable_interrupts_checked();
[0c78741]569
570}
[09800e9]571
[5afb49a]572extern "C" {
573 // Leave the thread monitor
574 // last routine called by a thread.
575 // Should never return
576 void __cfactx_thrd_leave() {
[e84ab3d]577 thread$ * thrd = active_thread();
578 monitor$ * this = &thrd->self_mon;
[5afb49a]579
580 // Lock the monitor now
581 lock( this->lock __cfaabi_dbg_ctx2 );
582
583 disable_interrupts();
584
[9d6e1b8a]585 /* paranoid */ verify( ! __preemption_enabled() );
586 /* paranoid */ verify( thrd->state == Active );
587 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
588 /* paranoid */ verify( kernelTLS().this_thread == thrd );
589 /* paranoid */ verify( thrd->context.SP );
[e84ab3d]590 /* 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 );
591 /* 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 );
[9d6e1b8a]592
[58688bf]593 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]594 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
595 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]596
[28372f7]597 thrd->state = Halting;
598 thrd->ticket = TICKET_DEAD;
599
[5afb49a]600 // Leave the thread
601 returnToKernel();
602
603 // Control flow should never reach here!
[9d6e1b8a]604 abort();
[5afb49a]605 }
[09800e9]606}
607
[14a61b5]608// KERNEL ONLY
[c18bf9e]609bool force_yield( __Preemption_Reason reason ) libcfa_public {
[a3821fa]610 __disable_interrupts_checked();
[e84ab3d]611 thread$ * thrd = kernelTLS().this_thread;
[a3821fa]612 /* paranoid */ verify(thrd->state == Active);
613
614 // SKULLDUGGERY: It is possible that we are preempting this thread just before
615 // it was going to park itself. If that is the case and it is already using the
616 // intrusive fields then we can't use them to preempt the thread
617 // If that is the case, abandon the preemption.
618 bool preempted = false;
[15c93d8]619 if(thrd->rdy_link.next == 0p) {
[a3821fa]620 preempted = true;
621 thrd->preempted = reason;
622 returnToKernel();
623 }
624 __enable_interrupts_checked( false );
[3381ed7]625 return preempted;
[f2b12406]626}
627
[14a61b5]628//=============================================================================================
[92e7631]629// Kernel Idle Sleep
[14a61b5]630//=============================================================================================
[64a7146]631// Wake a thread from the front if there are any
[e873838]632static void __wake_one(cluster * this) {
[7cf3b1d]633 eventfd_t val;
634
[8fc652e0]635 /* paranoid */ verify( ! __preemption_enabled() );
[e873838]636 /* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]637
[64a7146]638 // Check if there is a sleeping processor
[7cf3b1d]639 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
[14a61b5]640
[7cf3b1d]641 // If no one is sleeping: we are done
642 if( fdp == 0p ) return;
[14a61b5]643
[7cf3b1d]644 int fd = 1;
[22226e4]645 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
646 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
[7cf3b1d]647 }
[14a61b5]648
[7cf3b1d]649 switch(fd) {
[4ccc150]650 __attribute__((unused)) int ret;
[7cf3b1d]651 case 0:
652 // If the processor isn't ready to sleep then the exchange will already wake it up
653 #if !defined(__CFA_NO_STATISTICS__)
654 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
655 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
656 #endif
657 break;
658 case 1:
659 // If someone else already said they will wake them: we are done
660 #if !defined(__CFA_NO_STATISTICS__)
661 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
662 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
663 #endif
664 break;
665 default:
666 // If the processor was ready to sleep, we need to wake it up with an actual write
667 val = 1;
[4ccc150]668 ret = eventfd_write( fd, val );
[77adaee]669 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7cf3b1d]670
671 #if !defined(__CFA_NO_STATISTICS__)
672 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
673 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
674 #endif
675 break;
676 }
[1eb239e4]677
[e873838]678 /* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]679 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]680
681 return;
[64a7146]682}
[14a61b5]683
[64a7146]684// Unconditionnaly wake a thread
[1eb239e4]685void __wake_proc(processor * this) {
[7d0ebd0]686 /* paranoid */ verify( ! __preemption_enabled() );
687
[64a7146]688 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]689
[22226e4]690 this->idle_wctx.sem = 1;
[7cf3b1d]691
[efa28d5]692 this->idle_wctx.wake__time = rdtscl();
[262fafd9]693
[7d0ebd0]694 eventfd_t val;
695 val = 1;
[d080549]696 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
[7d0ebd0]697
[d080549]698 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7d0ebd0]699 /* paranoid */ verify( ! __preemption_enabled() );
[92e7631]700}
701
[18f7858]702static void idle_sleep(processor * this) {
[202c80b]703 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
704 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
705
[7cf3b1d]706 // Tell everyone we are ready to go do sleep
707 for() {
[22226e4]708 int expected = this->idle_wctx.sem;
[7cf3b1d]709
710 // Someone already told us to wake-up! No time for a nap.
711 if(expected == 1) { return; }
712
713 // Try to mark that we are going to sleep
[22226e4]714 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
[7cf3b1d]715 // Every one agreed, taking a nap
716 break;
717 }
718 }
719
720
[a757ba1]721 #if !defined(__CFA_NO_STATISTICS__)
722 if(this->print_halts) {
723 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
724 }
725 #endif
[1757f98]726
[a757ba1]727 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
[1757f98]728
[a757ba1]729 {
730 eventfd_t val;
731 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
732 if(ret < 0) {
733 switch((int)errno) {
734 case EAGAIN:
735 #if EAGAIN != EWOULDBLOCK
736 case EWOULDBLOCK:
737 #endif
738 case EINTR:
739 // No need to do anything special here, just assume it's a legitimate wake-up
740 break;
741 default:
742 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
[1757f98]743 }
744 }
[a757ba1]745 }
[1757f98]746
[a757ba1]747 #if !defined(__CFA_NO_STATISTICS__)
748 if(this->print_halts) {
749 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
750 }
[1757f98]751 #endif
752}
753
[5f5a729]754static bool mark_idle(__cluster_proc_list & this, processor & proc) {
[c9c1c1cb]755 __STATS__(true, ready.sleep.halts++; )
[7cf3b1d]756
[22226e4]757 proc.idle_wctx.sem = 0;
[7cf3b1d]758
[8fc652e0]759 /* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]760 if(!try_lock( this )) return false;
[1eb239e4]761 this.idle++;
762 /* paranoid */ verify( this.idle <= this.total );
[fc59b580]763 remove(proc);
[6a9b12b]764 insert_first(this.idles, proc);
[34b8cb7]765
[a757ba1]766 // update the pointer to the head wait context, which should now point to this proc.
[7cf3b1d]767 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
[1eb239e4]768 unlock( this );
[8fc652e0]769 /* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]770
771 return true;
[1eb239e4]772}
[8e16177]773
[6a9b12b]774static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]775 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]776 lock( this );
777 this.idle--;
778 /* paranoid */ verify( this.idle >= 0 );
779 remove(proc);
[fc59b580]780 insert_last(this.actives, proc);
[6a9b12b]781
[a633f6f]782 {
[a757ba1]783 // update the pointer to the head wait context
[7cf3b1d]784 struct __fd_waitctx * wctx = 0;
785 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
786 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
[a633f6f]787 }
788
[34b8cb7]789 unlock( this );
[6a9b12b]790 /* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]791}
792
[dbe9b08]793//=============================================================================================
794// Unexpected Terminating logic
795//=============================================================================================
[92bfda0]796void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
[e84ab3d]797 thread$ * thrd = __cfaabi_tls.this_thread;
[9d944b2]798
[de94a60]799 if(thrd) {
800 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]801 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]802
[212c2187]803 if ( &thrd->self_cor != thrd->curr_cor ) {
804 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]805 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]806 }
807 else {
[1c40091]808 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]809 }
[1c273d0]810 }
[9d944b2]811 else {
[de94a60]812 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]813 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]814 }
815}
816
[92bfda0]817int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
818 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]819}
820
[de94a60]821static __spinlock_t kernel_debug_lock;
822
[9d944b2]823extern "C" {
[1c40091]824 void __cfaabi_bits_acquire() {
[36982fc]825 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]826 }
827
[1c40091]828 void __cfaabi_bits_release() {
[ea7d2b0]829 unlock( kernel_debug_lock );
[9d944b2]830 }
[8118303]831}
832
[fa21ac9]833//=============================================================================================
834// Kernel Utilities
835//=============================================================================================
[dddb3dd0]836#if defined(CFA_HAVE_LINUX_IO_URING_H)
837#include "io/types.hfa"
838#endif
839
[2026bb6]840//-----------------------------------------------------------------------------
841// Debug
[c18bf9e]842bool threading_enabled(void) __attribute__((const)) libcfa_public {
[2026bb6]843 return true;
844}
[c34ebf2]845
846//-----------------------------------------------------------------------------
847// Statistics
848#if !defined(__CFA_NO_STATISTICS__)
[c18bf9e]849 void print_halts( processor & this ) libcfa_public {
[c34ebf2]850 this.print_halts = true;
851 }
[58688bf]852
[69914cbc]853 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
[8464edf]854 /* paranoid */ verify( cltr->stats );
855
856 processor * it = &list`first;
857 for(unsigned i = 0; i < count; i++) {
858 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
859 /* paranoid */ verify( it->local_data->this_stats );
[c33c2af]860 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
[8464edf]861 __tally_stats( cltr->stats, it->local_data->this_stats );
862 it = &(*it)`next;
863 }
864 }
865
[c18bf9e]866 static void crawl_cluster_stats( cluster & this ) {
[8464edf]867 // Stop the world, otherwise stats could get really messed-up
868 // this doesn't solve all problems but does solve many
869 // so it's probably good enough
[c33c2af]870 disable_interrupts();
[8464edf]871 uint_fast32_t last_size = ready_mutate_lock();
872
873 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
874 crawl_list(&this, this.procs.idles , this.procs.idle );
875
876 // Unlock the RWlock
877 ready_mutate_unlock( last_size );
[c33c2af]878 enable_interrupts();
[8464edf]879 }
880
881
[c18bf9e]882 void print_stats_now( cluster & this, int flags ) libcfa_public {
[8464edf]883 crawl_cluster_stats( this );
[202c80b]884 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
[1b033b8]885 }
[c34ebf2]886#endif
[8118303]887// Local Variables: //
888// mode: c //
889// tab-width: 4 //
890// End: //
Note: See TracBrowser for help on using the repository browser.