source: libcfa/src/concurrency/kernel.cfa@ fa2e183

ADT ast-experimental
Last change on this file since fa2e183 was 878cfcc, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Added extra check to make sure threads is never double executed

  • Property mode set to 100644
File size: 29.0 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)));
140extern void __cfa_io_idle( processor * ) __attribute__((nonnull (1)));
[dddb3dd0]141
[d3605f8]142#if defined(CFA_WITH_IO_URING_IDLE)
143 extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
[6ddef36]144#endif
[7ef162b2]145
[dddb3dd0]146extern void __disable_interrupts_hard();
147extern void __enable_interrupts_hard();
[c84e80a]148
[7ef162b2]149
[75f3522]150//=============================================================================================
151// Kernel Scheduling logic
152//=============================================================================================
[8fcbb4c]153//Main of the processor contexts
[83a071f9]154void main(processorCtx_t & runner) {
[21184e3]155 // Because of a bug, we couldn't initialized the seed on construction
156 // Do it here
[8fc652e0]157 __cfaabi_tls.rand_seed ^= rdtscl();
158 __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
[f2384c9a]159 __tls_rand_advance_bck();
[21184e3]160
[83a071f9]161 processor * this = runner.proc;
[094476d]162 verify(this);
[c81ebf9]163
[22226e4]164 /* paranoid */ verify( this->idle_wctx.ftr != 0p );
165 /* paranoid */ verify( this->idle_wctx.rdbuf != 0p );
[7ef162b2]166
[22226e4]167 // used for idle sleep when io_uring is present
168 // mark it as already fulfilled so we know if there is a pending request or not
169 this->idle_wctx.ftr->self.ptr = 1p;
[dddb3dd0]170
[4069faad]171 __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
[28d73c1]172 #if !defined(__CFA_NO_STATISTICS__)
173 if( this->print_halts ) {
[c993b15]174 __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->unique_id, this->name, (void*)this);
[28d73c1]175 }
176 #endif
[b798713]177
[75f3522]178 {
[c81ebf9]179 // Setup preemption data
180 preemption_scope scope = { this };
181
[a5e7233]182 // if we need to run some special setup, now is the time to do it.
183 if(this->init.thrd) {
184 this->init.thrd->curr_cluster = this->cltr;
185 __run_thread(this, this->init.thrd);
186 }
[325e6ea]187
[4069faad]188 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
[8118303]189
[e84ab3d]190 thread$ * readyThread = 0p;
[1eb239e4]191 MAIN_LOOP:
192 for() {
[dddb3dd0]193 // Check if there is pending io
[18f7858]194 __cfa_io_drain( this );
[dddb3dd0]195
[92e7631]196 // Try to get the next thread
[8c50aed]197 readyThread = __next_thread( this->cltr );
[75f3522]198
[1eb239e4]199 if( !readyThread ) {
[18f7858]200 // there is no point in holding submissions if we are idle
[c9c1c1cb]201 __IO_STATS__(true, io.flush.idle++; )
[18f7858]202 __cfa_io_flush( this );
203
204 // drain again in case something showed up
205 __cfa_io_drain( this );
[c33c2af]206
[c9c1c1cb]207 readyThread = __next_thread( this->cltr );
208 }
209
210 if( !readyThread ) for(5) {
[1eb239e4]211 readyThread = __next_thread_slow( this->cltr );
[c9c1c1cb]212
213 if( readyThread ) break;
214
[18f7858]215 // It's unlikely we still I/O to submit, but the arbiter could
216 __IO_STATS__(true, io.flush.idle++; )
217 __cfa_io_flush( this );
218
219 // drain again in case something showed up
220 __cfa_io_drain( this );
[1eb239e4]221 }
[4e6fb8e]222
[1eb239e4]223 HALT:
224 if( !readyThread ) {
225 // Don't block if we are done
226 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[c81ebf9]227
[1eb239e4]228 // Push self to idle stack
[5f5a729]229 if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;
[398e8e9]230
[1eb239e4]231 // Confirm the ready-queue is empty
[c9c1c1cb]232 readyThread = __next_thread_search( this->cltr );
[1eb239e4]233 if( readyThread ) {
234 // A thread was found, cancel the halt
[6a9b12b]235 mark_awake(this->cltr->procs, * this);
[1eb239e4]236
[c9c1c1cb]237 __STATS__(true, ready.sleep.cancels++; )
[1eb239e4]238
239 // continue the mai loop
240 break HALT;
241 }
242
[18f7858]243 idle_sleep( this );
[1eb239e4]244
245 // We were woken up, remove self from idle
[6a9b12b]246 mark_awake(this->cltr->procs, * this);
[1eb239e4]247
248 // DON'T just proceed, start looking again
249 continue MAIN_LOOP;
[64a7146]250 }
[1eb239e4]251
252 /* paranoid */ verify( readyThread );
253
[dddb3dd0]254 // Reset io dirty bit
255 this->io.dirty = false;
256
[1eb239e4]257 // We found a thread run it
258 __run_thread(this, readyThread);
259
260 // Are we done?
261 if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
[325e6ea]262
[d529ad0]263 if(__atomic_load_n(&this->io.pending, __ATOMIC_RELAXED) && !__atomic_load_n(&this->io.dirty, __ATOMIC_RELAXED)) {
[c9c1c1cb]264 __IO_STATS__(true, io.flush.dirty++; )
[18f7858]265 __cfa_io_flush( this );
[dddb3dd0]266 }
[c81ebf9]267 }
268
[4069faad]269 __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
[c84e80a]270 }
[8118303]271
[454f478]272 post( this->terminated );
[bdeba0b]273
[28d73c1]274 if(this == mainProcessor) {
[6a490b2]275 // HACK : the coroutine context switch expects this_thread to be set
276 // and it make sense for it to be set in all other cases except here
277 // fake it
[8fc652e0]278 __cfaabi_tls.this_thread = mainThread;
[6a490b2]279 }
[7768b8d]280
[4069faad]281 __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]282}
283
[5c1a531]284static int * __volatile_errno() __attribute__((noinline));
285static int * __volatile_errno() { asm(""); return &errno; }
286
[14a61b5]287// KERNEL ONLY
[1c273d0]288// runThread runs a thread by context switching
289// from the processor coroutine to the target thread
[e84ab3d]290static void __run_thread(processor * this, thread$ * thrd_dst) {
[8fc652e0]291 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]292 /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
293 /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
294 __builtin_prefetch( thrd_dst->context.SP );
295
[dddb3dd0]296 __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
297
[e84ab3d]298 coroutine$ * proc_cor = get_coroutine(this->runner);
[8fcbb4c]299
[9f575ea]300 // set state of processor coroutine to inactive
301 verify(proc_cor->state == Active);
[ae7be7a]302 proc_cor->state = Blocked;
[e8e457e]303
[9f575ea]304 // Actually run the thread
[3381ed7]305 RUNNING: while(true) {
[ff79d5e]306 thrd_dst->preempted = __NO_PREEMPTION;
[e8e457e]307
[58d64a4]308 // Update global state
[8fc652e0]309 kernelTLS().this_thread = thrd_dst;
[75f3522]310
[8bee858]311 // Update the state after setting this_thread
312 // so that the debugger can find all active threads
313 // in tls storage
314 thrd_dst->state = Active;
315
[8fc652e0]316 /* paranoid */ verify( ! __preemption_enabled() );
317 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]318 /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]319 /* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]320 /* paranoid */ verify( thrd_dst->state != Halted );
[e84ab3d]321 /* 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
322 /* 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]323 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
[ac12f1f]324 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]325
[3381ed7]326
[58d64a4]327
[9f575ea]328 // set context switch to the thread that the processor is executing
[c7a900a]329 __cfactx_switch( &proc_cor->context, &thrd_dst->context );
330 // when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]331
[50871b4]332
333
[ac12f1f]334 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[878cfcc]335 /* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
[e84ab3d]336 /* 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 );
337 /* 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]338 /* paranoid */ verify( thrd_dst->state != Halted );
[b4b63e8]339 /* paranoid */ verify( thrd_dst->context.SP );
[8fc652e0]340 /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
341 /* paranoid */ verify( ! __preemption_enabled() );
[75f3522]342
[3381ed7]343 // We just finished running a thread, there are a few things that could have happened.
344 // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
345 // 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
346 // 4 - Preempted
347 // In case 1, we may have won a race so we can't write to the state again.
348 // In case 2, we lost the race so we now own the thread.
349
350 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
[8bee858]351 // Reset the this_thread now that we know
352 // the state isn't active anymore
353 kernelTLS().this_thread = 0p;
354
[3381ed7]355 // The thread was preempted, reschedule it and reset the flag
[24e321c]356 schedule_thread$( thrd_dst, UNPARK_LOCAL );
[3381ed7]357 break RUNNING;
358 }
[75f3522]359
[3ea8ad1]360 if(unlikely(thrd_dst->state == Halting)) {
[8bee858]361 // Reset the this_thread now that we know
362 // the state isn't active anymore
363 kernelTLS().this_thread = 0p;
364
[ff79d5e]365 // The thread has halted, it should never be scheduled/run again
[5afb49a]366 // finish the thread
367 __thread_finish( thrd_dst );
[ff79d5e]368 break RUNNING;
369 }
370
371 /* paranoid */ verify( thrd_dst->state == Active );
372 thrd_dst->state = Blocked;
373
[8bee858]374 // Reset the this_thread now that we know
375 // the state isn't active anymore
376 kernelTLS().this_thread = 0p;
377
[3381ed7]378 // set state of processor coroutine to active and the thread to inactive
[ff79d5e]379 int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
380 switch(old_ticket) {
[6a77224]381 case TICKET_RUNNING:
[3381ed7]382 // This is case 1, the regular case, nothing more is needed
383 break RUNNING;
[6a77224]384 case TICKET_UNBLOCK:
[c9c1c1cb]385 __STATS__(true, ready.threads.threads++; )
[3381ed7]386 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
387 // In this case, just run it again.
388 continue RUNNING;
389 default:
390 // This makes no sense, something is wrong abort
[ff79d5e]391 abort();
[3381ed7]392 }
[9f575ea]393 }
[e8e457e]394
[9f575ea]395 // Just before returning to the processor, set the processor coroutine to active
[e8e457e]396 proc_cor->state = Active;
[1eb239e4]397
[dddb3dd0]398 __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
399
[c9c1c1cb]400 __STATS__(true, ready.threads.threads--; )
[ec43cf9]401
[8fc652e0]402 /* paranoid */ verify( ! __preemption_enabled() );
[82c948c]403}
404
[14a61b5]405// KERNEL_ONLY
[c18bf9e]406static void returnToKernel() {
[8fc652e0]407 /* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]408 coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
409 thread$ * thrd_src = kernelTLS().this_thread;
[e8e457e]410
[c9c1c1cb]411 __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]412
[9f575ea]413 // Run the thread on this processor
414 {
415 int local_errno = *__volatile_errno();
416 #if defined( __i386 ) || defined( __x86_64 )
417 __x87_store;
418 #endif
[b4b63e8]419 /* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]420 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]421 __cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]422 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]423 #if defined( __i386 ) || defined( __x86_64 )
424 __x87_load;
425 #endif
426 *__volatile_errno() = local_errno;
[8fcbb4c]427 }
[deca0f5]428
[29cb302]429 #if !defined(__CFA_NO_STATISTICS__)
[89eff25]430 /* paranoid */ verify( thrd_src->last_proc != 0p );
431 if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]432 __tls_stats()->ready.threads.migration++;
433 }
434 #endif
435
[8fc652e0]436 /* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]437 /* 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 );
438 /* 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]439}
440
[8def349]441//-----------------------------------------------------------------------------
442// Scheduler routines
[14a61b5]443// KERNEL ONLY
[24e321c]444static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
[8fc652e0]445 /* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]446 /* paranoid */ verify( ready_schedule_islocked());
[6a490b2]447 /* paranoid */ verify( thrd );
448 /* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]449 /* paranoid */ verify( thrd->curr_cluster );
[3381ed7]450 /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]451 /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
452 "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]453 /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]454 "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]455 /* paranoid */ #endif
[6a490b2]456 /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
[ac12f1f]457 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]458
[ae7be7a]459 if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]460
[ec43cf9]461 // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
462 struct cluster * cl = thrd->curr_cluster;
[c9c1c1cb]463 __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]464
[254ad1b]465 // push the thread to the cluster ready-queue
[24e321c]466 push( cl, thrd, hint );
[32a8b61]467
[254ad1b]468 // variable thrd is no longer safe to use
[734908c]469 thrd = 0xdeaddeaddeaddeadp;
[32a8b61]470
[254ad1b]471 // wake the cluster using the save variable.
472 __wake_one( cl );
[1c273d0]473
[ec43cf9]474 #if !defined(__CFA_NO_STATISTICS__)
475 if( kernelTLS().this_stats ) {
476 __tls_stats()->ready.threads.threads++;
[89eff25]477 if(outside) {
478 __tls_stats()->ready.threads.extunpark++;
479 }
[ec43cf9]480 }
481 else {
482 __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]483 __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[ec43cf9]484 }
485 #endif
486
[254ad1b]487 /* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]488 /* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]489}
490
[24e321c]491void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
[254ad1b]492 ready_schedule_lock();
[24e321c]493 __schedule_thread( thrd, hint );
[254ad1b]494 ready_schedule_unlock();
495}
496
[14a61b5]497// KERNEL ONLY
[e84ab3d]498static inline thread$ * __next_thread(cluster * this) with( *this ) {
[8fc652e0]499 /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]500
[e873838]501 ready_schedule_lock();
[e84ab3d]502 thread$ * thrd = pop_fast( this );
[e873838]503 ready_schedule_unlock();
[7768b8d]504
[8fc652e0]505 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]506 return thrd;
[eb2e723]507}
508
[64a7146]509// KERNEL ONLY
[e84ab3d]510static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]511 /* paranoid */ verify( ! __preemption_enabled() );
[64a7146]512
[e873838]513 ready_schedule_lock();
[c9c1c1cb]514 thread$ * thrd = pop_slow( this );
515 ready_schedule_unlock();
516
517 /* paranoid */ verify( ! __preemption_enabled() );
518 return thrd;
519}
[fc59df78]520
[c9c1c1cb]521// KERNEL ONLY
522static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
523 /* paranoid */ verify( ! __preemption_enabled() );
524
525 ready_schedule_lock();
526 thread$ * thrd = pop_search( this );
[e873838]527 ready_schedule_unlock();
[64a7146]528
[8fc652e0]529 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]530 return thrd;
[64a7146]531}
532
[e84ab3d]533static inline bool __must_unpark( thread$ * thrd ) {
[ff79d5e]534 int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
535 switch(old_ticket) {
[6a77224]536 case TICKET_RUNNING:
[3381ed7]537 // Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]538 return false;
[6a77224]539 case TICKET_BLOCKED:
[3381ed7]540 /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]541 /* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]542 return true;
[3381ed7]543 default:
544 // This makes no sense, something is wrong abort
[7ee8153]545 abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]546 }
[eb2e723]547}
548
[24e321c]549void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
[e9c0b4c]550 /* paranoid */ verify( ! __preemption_enabled() );
551 /* paranoid */ verify( ready_schedule_islocked());
552
553 if( !thrd ) return;
554
555 if(__must_unpark(thrd)) {
556 // Wake lost the race,
[24e321c]557 __schedule_thread( thrd, hint );
[e9c0b4c]558 }
559
560 /* paranoid */ verify( ready_schedule_islocked());
561 /* paranoid */ verify( ! __preemption_enabled() );
562}
563
[c18bf9e]564void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
[c6c7e6c]565 if( !thrd ) return;
[0b33412]566
[c6c7e6c]567 if(__must_unpark(thrd)) {
568 disable_interrupts();
[a3821fa]569 // Wake lost the race,
[24e321c]570 schedule_thread$( thrd, hint );
[a3821fa]571 enable_interrupts(false);
[de6319f]572 }
[eb2e723]573}
[0b33412]574
[c18bf9e]575void park( void ) libcfa_public {
[a3821fa]576 __disable_interrupts_checked();
577 /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
578 returnToKernel();
579 __enable_interrupts_checked();
[0c78741]580
581}
[09800e9]582
[5afb49a]583extern "C" {
584 // Leave the thread monitor
585 // last routine called by a thread.
586 // Should never return
587 void __cfactx_thrd_leave() {
[e84ab3d]588 thread$ * thrd = active_thread();
589 monitor$ * this = &thrd->self_mon;
[5afb49a]590
591 // Lock the monitor now
592 lock( this->lock __cfaabi_dbg_ctx2 );
593
594 disable_interrupts();
595
[9d6e1b8a]596 /* paranoid */ verify( ! __preemption_enabled() );
597 /* paranoid */ verify( thrd->state == Active );
598 /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
599 /* paranoid */ verify( kernelTLS().this_thread == thrd );
600 /* paranoid */ verify( thrd->context.SP );
[e84ab3d]601 /* 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 );
602 /* 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]603
[3ea8ad1]604 thrd->state = Halting;
[58688bf]605 if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]606 if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
607 if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]608
609 // Leave the thread
610 returnToKernel();
611
612 // Control flow should never reach here!
[9d6e1b8a]613 abort();
[5afb49a]614 }
[09800e9]615}
616
[14a61b5]617// KERNEL ONLY
[c18bf9e]618bool force_yield( __Preemption_Reason reason ) libcfa_public {
[a3821fa]619 __disable_interrupts_checked();
[e84ab3d]620 thread$ * thrd = kernelTLS().this_thread;
[a3821fa]621 /* paranoid */ verify(thrd->state == Active);
622
623 // SKULLDUGGERY: It is possible that we are preempting this thread just before
624 // it was going to park itself. If that is the case and it is already using the
625 // intrusive fields then we can't use them to preempt the thread
626 // If that is the case, abandon the preemption.
627 bool preempted = false;
628 if(thrd->link.next == 0p) {
629 preempted = true;
630 thrd->preempted = reason;
631 returnToKernel();
632 }
633 __enable_interrupts_checked( false );
[3381ed7]634 return preempted;
[f2b12406]635}
636
[14a61b5]637//=============================================================================================
[92e7631]638// Kernel Idle Sleep
[14a61b5]639//=============================================================================================
[64a7146]640// Wake a thread from the front if there are any
[e873838]641static void __wake_one(cluster * this) {
[7cf3b1d]642 eventfd_t val;
643
[8fc652e0]644 /* paranoid */ verify( ! __preemption_enabled() );
[e873838]645 /* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]646
[64a7146]647 // Check if there is a sleeping processor
[7cf3b1d]648 struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
[14a61b5]649
[7cf3b1d]650 // If no one is sleeping: we are done
651 if( fdp == 0p ) return;
[14a61b5]652
[7cf3b1d]653 int fd = 1;
[22226e4]654 if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
655 fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
[7cf3b1d]656 }
[14a61b5]657
[7cf3b1d]658 switch(fd) {
[4ccc150]659 __attribute__((unused)) int ret;
[7cf3b1d]660 case 0:
661 // If the processor isn't ready to sleep then the exchange will already wake it up
662 #if !defined(__CFA_NO_STATISTICS__)
663 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
664 } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
665 #endif
666 break;
667 case 1:
668 // If someone else already said they will wake them: we are done
669 #if !defined(__CFA_NO_STATISTICS__)
670 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
671 } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
672 #endif
673 break;
674 default:
675 // If the processor was ready to sleep, we need to wake it up with an actual write
676 val = 1;
[4ccc150]677 ret = eventfd_write( fd, val );
[77adaee]678 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7cf3b1d]679
680 #if !defined(__CFA_NO_STATISTICS__)
681 if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
682 } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
683 #endif
684 break;
685 }
[1eb239e4]686
[e873838]687 /* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]688 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]689
690 return;
[64a7146]691}
[14a61b5]692
[64a7146]693// Unconditionnaly wake a thread
[1eb239e4]694void __wake_proc(processor * this) {
[7d0ebd0]695 /* paranoid */ verify( ! __preemption_enabled() );
696
[64a7146]697 __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]698
[22226e4]699 this->idle_wctx.sem = 1;
[7cf3b1d]700
[efa28d5]701 this->idle_wctx.wake__time = rdtscl();
[262fafd9]702
[7d0ebd0]703 eventfd_t val;
704 val = 1;
[d080549]705 __attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
[7d0ebd0]706
[d080549]707 /* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7d0ebd0]708 /* paranoid */ verify( ! __preemption_enabled() );
[92e7631]709}
710
[18f7858]711static void idle_sleep(processor * this) {
[202c80b]712 /* paranoid */ verify( this->idle_wctx.evfd != 1 );
713 /* paranoid */ verify( this->idle_wctx.evfd != 2 );
714
[7cf3b1d]715 // Tell everyone we are ready to go do sleep
716 for() {
[22226e4]717 int expected = this->idle_wctx.sem;
[7cf3b1d]718
719 // Someone already told us to wake-up! No time for a nap.
720 if(expected == 1) { return; }
721
722 // Try to mark that we are going to sleep
[22226e4]723 if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
[7cf3b1d]724 // Every one agreed, taking a nap
725 break;
726 }
727 }
728
729
[d3605f8]730 #if !defined(CFA_WITH_IO_URING_IDLE)
[7ef162b2]731 #if !defined(__CFA_NO_STATISTICS__)
732 if(this->print_halts) {
733 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
734 }
735 #endif
[1757f98]736
[7ef162b2]737 __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
[1757f98]738
[7ef162b2]739 {
740 eventfd_t val;
[22226e4]741 ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
[7ef162b2]742 if(ret < 0) {
743 switch((int)errno) {
744 case EAGAIN:
745 #if EAGAIN != EWOULDBLOCK
746 case EWOULDBLOCK:
747 #endif
748 case EINTR:
749 // No need to do anything special here, just assume it's a legitimate wake-up
750 break;
751 default:
752 abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
753 }
[1757f98]754 }
755 }
756
[7ef162b2]757 #if !defined(__CFA_NO_STATISTICS__)
758 if(this->print_halts) {
759 __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
760 }
761 #endif
762 #else
[18f7858]763 __cfa_io_idle( this );
[1757f98]764 #endif
765}
766
[5f5a729]767static bool mark_idle(__cluster_proc_list & this, processor & proc) {
[c9c1c1cb]768 __STATS__(true, ready.sleep.halts++; )
[7cf3b1d]769
[22226e4]770 proc.idle_wctx.sem = 0;
[7cf3b1d]771
[8fc652e0]772 /* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]773 if(!try_lock( this )) return false;
[1eb239e4]774 this.idle++;
775 /* paranoid */ verify( this.idle <= this.total );
[fc59b580]776 remove(proc);
[6a9b12b]777 insert_first(this.idles, proc);
[34b8cb7]778
[7cf3b1d]779 __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
[1eb239e4]780 unlock( this );
[8fc652e0]781 /* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]782
783 return true;
[1eb239e4]784}
[8e16177]785
[6a9b12b]786static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]787 /* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]788 lock( this );
789 this.idle--;
790 /* paranoid */ verify( this.idle >= 0 );
791 remove(proc);
[fc59b580]792 insert_last(this.actives, proc);
[6a9b12b]793
[a633f6f]794 {
[7cf3b1d]795 struct __fd_waitctx * wctx = 0;
796 if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
797 __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
[a633f6f]798 }
799
[34b8cb7]800 unlock( this );
[6a9b12b]801 /* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]802}
803
[dbe9b08]804//=============================================================================================
805// Unexpected Terminating logic
806//=============================================================================================
[92bfda0]807void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
[e84ab3d]808 thread$ * thrd = __cfaabi_tls.this_thread;
[9d944b2]809
[de94a60]810 if(thrd) {
811 int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]812 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]813
[212c2187]814 if ( &thrd->self_cor != thrd->curr_cor ) {
815 len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]816 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]817 }
818 else {
[1c40091]819 __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]820 }
[1c273d0]821 }
[9d944b2]822 else {
[de94a60]823 int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]824 __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]825 }
826}
827
[92bfda0]828int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
829 return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]830}
831
[de94a60]832static __spinlock_t kernel_debug_lock;
833
[9d944b2]834extern "C" {
[1c40091]835 void __cfaabi_bits_acquire() {
[36982fc]836 lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]837 }
838
[1c40091]839 void __cfaabi_bits_release() {
[ea7d2b0]840 unlock( kernel_debug_lock );
[9d944b2]841 }
[8118303]842}
843
[fa21ac9]844//=============================================================================================
845// Kernel Utilities
846//=============================================================================================
[dddb3dd0]847#if defined(CFA_HAVE_LINUX_IO_URING_H)
848#include "io/types.hfa"
849#endif
850
[2026bb6]851//-----------------------------------------------------------------------------
852// Debug
[c18bf9e]853bool threading_enabled(void) __attribute__((const)) libcfa_public {
[2026bb6]854 return true;
855}
[c34ebf2]856
857//-----------------------------------------------------------------------------
858// Statistics
859#if !defined(__CFA_NO_STATISTICS__)
[c18bf9e]860 void print_halts( processor & this ) libcfa_public {
[c34ebf2]861 this.print_halts = true;
862 }
[58688bf]863
[69914cbc]864 static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
[8464edf]865 /* paranoid */ verify( cltr->stats );
866
867 processor * it = &list`first;
868 for(unsigned i = 0; i < count; i++) {
869 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
870 /* paranoid */ verify( it->local_data->this_stats );
[c33c2af]871 // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
[8464edf]872 __tally_stats( cltr->stats, it->local_data->this_stats );
873 it = &(*it)`next;
874 }
875 }
876
[c18bf9e]877 static void crawl_cluster_stats( cluster & this ) {
[8464edf]878 // Stop the world, otherwise stats could get really messed-up
879 // this doesn't solve all problems but does solve many
880 // so it's probably good enough
[c33c2af]881 disable_interrupts();
[8464edf]882 uint_fast32_t last_size = ready_mutate_lock();
883
884 crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
885 crawl_list(&this, this.procs.idles , this.procs.idle );
886
887 // Unlock the RWlock
888 ready_mutate_unlock( last_size );
[c33c2af]889 enable_interrupts();
[8464edf]890 }
891
892
[c18bf9e]893 void print_stats_now( cluster & this, int flags ) libcfa_public {
[8464edf]894 crawl_cluster_stats( this );
[202c80b]895 __print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
[1b033b8]896 }
[c34ebf2]897#endif
[8118303]898// Local Variables: //
899// mode: c //
900// tab-width: 4 //
901// End: //
Note: See TracBrowser for help on using the repository browser.