Context Navigation

kernel.cfa@ 07e3652

Visit:

ADT ast-experimental

Last change on this file since 07e3652 was 26544f9, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
added helping and lock to allow remote processors to flush unresponsive procs
Property mode set to `100644`
File size: 28.6 KB

Rev	Line
[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
[dd46fd3]	12	// Last Modified On : Wed Nov 30 18:14:08 2022
	13	// Update Count : 76
[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]	31	extern "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]	122	extern thread$ * mainThread;
[e660761]	123	extern processor * mainProcessor;
[2ac095d]	124
[92e7631]	125	//-----------------------------------------------------------------------------
	126	// Kernel Scheduling logic
[e84ab3d]	127	static thread$ * __next_thread(cluster * this);
	128	static thread$ * __next_thread_slow(cluster * this);
[c9c1c1cb]	129	static thread$ * __next_thread_search(cluster * this);
[e84ab3d]	130	static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
	131	static void __run_thread(processor * this, thread$ * dst);
[e873838]	132	static void __wake_one(cluster * cltr);
[1eb239e4]	133
[18f7858]	134	static void idle_sleep(processor * proc);
[5f5a729]	135	static bool mark_idle (__cluster_proc_list & idles, processor & proc);
[6a9b12b]	136	static void mark_awake(__cluster_proc_list & idles, processor & proc);
[1eb239e4]	137
[4479890]	138	extern bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1)));
[18f7858]	139	extern bool __cfa_io_flush( processor * ) __attribute__((nonnull (1)));
[dddb3dd0]	140
[7ef162b2]	141
[dddb3dd0]	142	extern void __disable_interrupts_hard();
	143	extern void __enable_interrupts_hard();
[c84e80a]	144
[7ef162b2]	145
[75f3522]	146	//=============================================================================================
	147	// Kernel Scheduling logic
	148	//=============================================================================================
[8fcbb4c]	149	//Main of the processor contexts
[83a071f9]	150	void main(processorCtx_t & runner) {
[21184e3]	151	// Because of a bug, we couldn't initialized the seed on construction
	152	// Do it here
[dd46fd3]	153	PRNG_SET_SEED( __cfaabi_tls.random_state, rdtscl() );
[8fc652e0]	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
[26544f9]	261	__cfa_io_flush( this );
	262	__cfa_io_drain( this );
	263
[454f478]	264	post( this->terminated );
[bdeba0b]	265
[28d73c1]	266	if(this == mainProcessor) {
[6a490b2]	267	// HACK : the coroutine context switch expects this_thread to be set
	268	// and it make sense for it to be set in all other cases except here
	269	// fake it
[8fc652e0]	270	__cfaabi_tls.this_thread = mainThread;
[6a490b2]	271	}
[7768b8d]	272
[4069faad]	273	__cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
[c84e80a]	274	}
	275
[5c1a531]	276	static int * __volatile_errno() __attribute__((noinline));
	277	static int * __volatile_errno() { asm(""); return &errno; }
	278
[14a61b5]	279	// KERNEL ONLY
[1c273d0]	280	// runThread runs a thread by context switching
	281	// from the processor coroutine to the target thread
[e84ab3d]	282	static void __run_thread(processor * this, thread$ * thrd_dst) {
[8fc652e0]	283	/* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]	284	/* paranoid */ verifyf( thrd_dst->state == Ready \|\| thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
[15c93d8]	285	/* paranoid */ verifyf( thrd_dst->rdy_link.next == 0p, "Expected null got %p", thrd_dst->rdy_link.next );
[1eb239e4]	286	__builtin_prefetch( thrd_dst->context.SP );
	287
[dddb3dd0]	288	__cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
	289
[e84ab3d]	290	coroutine$ * proc_cor = get_coroutine(this->runner);
[8fcbb4c]	291
[9f575ea]	292	// set state of processor coroutine to inactive
	293	verify(proc_cor->state == Active);
[ae7be7a]	294	proc_cor->state = Blocked;
[e8e457e]	295
[9f575ea]	296	// Actually run the thread
[3381ed7]	297	RUNNING: while(true) {
[ff79d5e]	298	thrd_dst->preempted = __NO_PREEMPTION;
[e8e457e]	299
[58d64a4]	300	// Update global state
[8fc652e0]	301	kernelTLS().this_thread = thrd_dst;
[75f3522]	302
[8bee858]	303	// Update the state after setting this_thread
	304	// so that the debugger can find all active threads
	305	// in tls storage
	306	thrd_dst->state = Active;
	307
[8fc652e0]	308	/* paranoid */ verify( ! __preemption_enabled() );
	309	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
[9d6e1b8a]	310	/* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
[b4b63e8]	311	/* paranoid */ verify( thrd_dst->context.SP );
[5afb49a]	312	/* paranoid */ verify( thrd_dst->state != Halted );
[e84ab3d]	313	/* 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
	314	/* 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]	315	/* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, this, __ATOMIC_SEQ_CST) == 0p );
[ac12f1f]	316	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[b4b63e8]	317
[3381ed7]	318
[58d64a4]	319
[9f575ea]	320	// set context switch to the thread that the processor is executing
[c7a900a]	321	__cfactx_switch( &proc_cor->context, &thrd_dst->context );
	322	// when __cfactx_switch returns we are back in the processor coroutine
[9f575ea]	323
[50871b4]	324
	325
[ac12f1f]	326	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
[878cfcc]	327	/* paranoid */ verify( __atomic_exchange_n( &thrd_dst->executing, 0p, __ATOMIC_SEQ_CST) == this );
[e84ab3d]	328	/* 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 );
	329	/* 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]	330	/* paranoid */ verify( thrd_dst->state != Halted );
[b4b63e8]	331	/* paranoid */ verify( thrd_dst->context.SP );
[8fc652e0]	332	/* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
	333	/* paranoid */ verify( ! __preemption_enabled() );
[75f3522]	334
[3381ed7]	335	// We just finished running a thread, there are a few things that could have happened.
	336	// 1 - Regular case : the thread has blocked and now one has scheduled it yet.
	337	// 2 - Racy case : the thread has blocked but someone has already tried to schedule it.
	338	// 4 - Preempted
	339	// In case 1, we may have won a race so we can't write to the state again.
	340	// In case 2, we lost the race so we now own the thread.
	341
	342	if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
[8bee858]	343	// Reset the this_thread now that we know
	344	// the state isn't active anymore
	345	kernelTLS().this_thread = 0p;
	346
[3381ed7]	347	// The thread was preempted, reschedule it and reset the flag
[24e321c]	348	schedule_thread$( thrd_dst, UNPARK_LOCAL );
[3381ed7]	349	break RUNNING;
	350	}
[75f3522]	351
[3ea8ad1]	352	if(unlikely(thrd_dst->state == Halting)) {
[8bee858]	353	// Reset the this_thread now that we know
	354	// the state isn't active anymore
	355	kernelTLS().this_thread = 0p;
	356
[ff79d5e]	357	// The thread has halted, it should never be scheduled/run again
[5afb49a]	358	// finish the thread
	359	__thread_finish( thrd_dst );
[ff79d5e]	360	break RUNNING;
	361	}
	362
	363	/* paranoid */ verify( thrd_dst->state == Active );
	364	thrd_dst->state = Blocked;
	365
[8bee858]	366	// Reset the this_thread now that we know
	367	// the state isn't active anymore
	368	kernelTLS().this_thread = 0p;
	369
[3381ed7]	370	// set state of processor coroutine to active and the thread to inactive
[ff79d5e]	371	int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
	372	switch(old_ticket) {
[6a77224]	373	case TICKET_RUNNING:
[3381ed7]	374	// This is case 1, the regular case, nothing more is needed
	375	break RUNNING;
[6a77224]	376	case TICKET_UNBLOCK:
[c9c1c1cb]	377	__STATS__(true, ready.threads.threads++; )
[3381ed7]	378	// This is case 2, the racy case, someone tried to run this thread before it finished blocking
	379	// In this case, just run it again.
	380	continue RUNNING;
	381	default:
	382	// This makes no sense, something is wrong abort
[ff79d5e]	383	abort();
[3381ed7]	384	}
[9f575ea]	385	}
[e8e457e]	386
[9f575ea]	387	// Just before returning to the processor, set the processor coroutine to active
[e8e457e]	388	proc_cor->state = Active;
[1eb239e4]	389
[dddb3dd0]	390	__cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
	391
[c9c1c1cb]	392	__STATS__(true, ready.threads.threads--; )
[ec43cf9]	393
[8fc652e0]	394	/* paranoid */ verify( ! __preemption_enabled() );
[82c948c]	395	}
	396
[14a61b5]	397	// KERNEL_ONLY
[c18bf9e]	398	static void returnToKernel() {
[8fc652e0]	399	/* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]	400	coroutine$ * proc_cor = get_coroutine(kernelTLS().this_processor->runner);
	401	thread$ * thrd_src = kernelTLS().this_thread;
[e8e457e]	402
[c9c1c1cb]	403	__STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
[29cb302]	404
[9f575ea]	405	// Run the thread on this processor
	406	{
	407	int local_errno = *__volatile_errno();
	408	#if defined( __i386 ) \|\| defined( __x86_64 )
	409	__x87_store;
	410	#endif
[b4b63e8]	411	/* paranoid */ verify( proc_cor->context.SP );
[ac12f1f]	412	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[c7a900a]	413	__cfactx_switch( &thrd_src->context, &proc_cor->context );
[ac12f1f]	414	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
[9f575ea]	415	#if defined( __i386 ) \|\| defined( __x86_64 )
	416	__x87_load;
	417	#endif
	418	*__volatile_errno() = local_errno;
[8fcbb4c]	419	}
[deca0f5]	420
[29cb302]	421	#if !defined(__CFA_NO_STATISTICS__)
[89eff25]	422	/* paranoid */ verify( thrd_src->last_proc != 0p );
	423	if(thrd_src->last_proc != kernelTLS().this_processor) {
[29cb302]	424	__tls_stats()->ready.threads.migration++;
	425	}
	426	#endif
	427
[8fc652e0]	428	/* paranoid */ verify( ! __preemption_enabled() );
[e84ab3d]	429	/* 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 );
	430	/* 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]	431	}
	432
[8def349]	433	//-----------------------------------------------------------------------------
	434	// Scheduler routines
[14a61b5]	435	// KERNEL ONLY
[24e321c]	436	static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
[8fc652e0]	437	/* paranoid */ verify( ! __preemption_enabled() );
[254ad1b]	438	/* paranoid */ verify( ready_schedule_islocked());
[6a490b2]	439	/* paranoid */ verify( thrd );
	440	/* paranoid */ verify( thrd->state != Halted );
[9d6e1b8a]	441	/* paranoid */ verify( thrd->curr_cluster );
[3381ed7]	442	/* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
[504a7dc]	443	/* paranoid */ if( thrd->state == Blocked \|\| thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
	444	"Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
[ff79d5e]	445	/* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
[504a7dc]	446	"Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
[3381ed7]	447	/* paranoid */ #endif
[15c93d8]	448	/* paranoid */ verifyf( thrd->rdy_link.next == 0p, "Expected null got %p", thrd->rdy_link.next );
[ac12f1f]	449	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
[b4b63e8]	450
[ae7be7a]	451	if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
[6b4cdd3]	452
[ec43cf9]	453	// Dereference the thread now because once we push it, there is not guaranteed it's still valid.
	454	struct cluster * cl = thrd->curr_cluster;
[c9c1c1cb]	455	__STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
[32a8b61]	456
[254ad1b]	457	// push the thread to the cluster ready-queue
[24e321c]	458	push( cl, thrd, hint );
[32a8b61]	459
[254ad1b]	460	// variable thrd is no longer safe to use
[734908c]	461	thrd = 0xdeaddeaddeaddeadp;
[32a8b61]	462
[254ad1b]	463	// wake the cluster using the save variable.
	464	__wake_one( cl );
[1c273d0]	465
[ec43cf9]	466	#if !defined(__CFA_NO_STATISTICS__)
	467	if( kernelTLS().this_stats ) {
	468	__tls_stats()->ready.threads.threads++;
[89eff25]	469	if(outside) {
	470	__tls_stats()->ready.threads.extunpark++;
	471	}
[ec43cf9]	472	}
	473	else {
	474	__atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
[89eff25]	475	__atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
[ec43cf9]	476	}
	477	#endif
	478
[254ad1b]	479	/* paranoid */ verify( ready_schedule_islocked());
[8fc652e0]	480	/* paranoid */ verify( ! __preemption_enabled() );
[db6f06a]	481	}
	482
[24e321c]	483	void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
[254ad1b]	484	ready_schedule_lock();
[24e321c]	485	__schedule_thread( thrd, hint );
[254ad1b]	486	ready_schedule_unlock();
	487	}
	488
[14a61b5]	489	// KERNEL ONLY
[e84ab3d]	490	static inline thread$ * __next_thread(cluster * this) with( *this ) {
[8fc652e0]	491	/* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]	492
[e873838]	493	ready_schedule_lock();
[e84ab3d]	494	thread$ * thrd = pop_fast( this );
[e873838]	495	ready_schedule_unlock();
[7768b8d]	496
[8fc652e0]	497	/* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]	498	return thrd;
[eb2e723]	499	}
	500
[64a7146]	501	// KERNEL ONLY
[e84ab3d]	502	static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
[8fc652e0]	503	/* paranoid */ verify( ! __preemption_enabled() );
[64a7146]	504
[e873838]	505	ready_schedule_lock();
[c9c1c1cb]	506	thread$ * thrd = pop_slow( this );
	507	ready_schedule_unlock();
	508
	509	/* paranoid */ verify( ! __preemption_enabled() );
	510	return thrd;
	511	}
[fc59df78]	512
[c9c1c1cb]	513	// KERNEL ONLY
	514	static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
	515	/* paranoid */ verify( ! __preemption_enabled() );
	516
	517	ready_schedule_lock();
	518	thread$ * thrd = pop_search( this );
[e873838]	519	ready_schedule_unlock();
[64a7146]	520
[8fc652e0]	521	/* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]	522	return thrd;
[64a7146]	523	}
	524
[e84ab3d]	525	static inline bool __must_unpark( thread$ * thrd ) {
[ff79d5e]	526	int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
	527	switch(old_ticket) {
[6a77224]	528	case TICKET_RUNNING:
[3381ed7]	529	// Wake won the race, the thread will reschedule/rerun itself
[c6c7e6c]	530	return false;
[6a77224]	531	case TICKET_BLOCKED:
[3381ed7]	532	/* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
[ff79d5e]	533	/* paranoid */ verify( thrd->state == Blocked );
[c6c7e6c]	534	return true;
[3381ed7]	535	default:
	536	// This makes no sense, something is wrong abort
[7ee8153]	537	abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
[de6319f]	538	}
[eb2e723]	539	}
	540
[24e321c]	541	void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
[e9c0b4c]	542	/* paranoid */ verify( ! __preemption_enabled() );
	543	/* paranoid */ verify( ready_schedule_islocked());
	544
	545	if( !thrd ) return;
	546
	547	if(__must_unpark(thrd)) {
	548	// Wake lost the race,
[24e321c]	549	__schedule_thread( thrd, hint );
[e9c0b4c]	550	}
	551
	552	/* paranoid */ verify( ready_schedule_islocked());
	553	/* paranoid */ verify( ! __preemption_enabled() );
	554	}
	555
[c18bf9e]	556	void unpark( thread$ * thrd, unpark_hint hint ) libcfa_public {
[c6c7e6c]	557	if( !thrd ) return;
[0b33412]	558
[c6c7e6c]	559	if(__must_unpark(thrd)) {
	560	disable_interrupts();
[a3821fa]	561	// Wake lost the race,
[24e321c]	562	schedule_thread$( thrd, hint );
[a3821fa]	563	enable_interrupts(false);
[de6319f]	564	}
[eb2e723]	565	}
[0b33412]	566
[c18bf9e]	567	void park( void ) libcfa_public {
[a3821fa]	568	__disable_interrupts_checked();
	569	/* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
	570	returnToKernel();
	571	__enable_interrupts_checked();
[0c78741]	572
	573	}
[09800e9]	574
[5afb49a]	575	extern "C" {
	576	// Leave the thread monitor
	577	// last routine called by a thread.
	578	// Should never return
	579	void __cfactx_thrd_leave() {
[e84ab3d]	580	thread$ * thrd = active_thread();
	581	monitor$ * this = &thrd->self_mon;
[5afb49a]	582
	583	// Lock the monitor now
	584	lock( this->lock __cfaabi_dbg_ctx2 );
	585
	586	disable_interrupts();
	587
[9d6e1b8a]	588	/* paranoid */ verify( ! __preemption_enabled() );
	589	/* paranoid */ verify( thrd->state == Active );
	590	/* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
	591	/* paranoid */ verify( kernelTLS().this_thread == thrd );
	592	/* paranoid */ verify( thrd->context.SP );
[e84ab3d]	593	/* 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 );
	594	/* 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]	595
[58688bf]	596	if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
[9d6e1b8a]	597	if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
	598	if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }
[5afb49a]	599
[28372f7]	600	thrd->state = Halting;
	601	thrd->ticket = TICKET_DEAD;
	602
[5afb49a]	603	// Leave the thread
	604	returnToKernel();
	605
	606	// Control flow should never reach here!
[9d6e1b8a]	607	abort();
[5afb49a]	608	}
[09800e9]	609	}
	610
[14a61b5]	611	// KERNEL ONLY
[c18bf9e]	612	bool force_yield( __Preemption_Reason reason ) libcfa_public {
[a3821fa]	613	__disable_interrupts_checked();
[e84ab3d]	614	thread$ * thrd = kernelTLS().this_thread;
[a3821fa]	615	/* paranoid */ verify(thrd->state == Active);
	616
	617	// SKULLDUGGERY: It is possible that we are preempting this thread just before
	618	// it was going to park itself. If that is the case and it is already using the
	619	// intrusive fields then we can't use them to preempt the thread
	620	// If that is the case, abandon the preemption.
	621	bool preempted = false;
[15c93d8]	622	if(thrd->rdy_link.next == 0p) {
[a3821fa]	623	preempted = true;
	624	thrd->preempted = reason;
	625	returnToKernel();
	626	}
	627	__enable_interrupts_checked( false );
[3381ed7]	628	return preempted;
[f2b12406]	629	}
	630
[14a61b5]	631	//=============================================================================================
[92e7631]	632	// Kernel Idle Sleep
[14a61b5]	633	//=============================================================================================
[64a7146]	634	// Wake a thread from the front if there are any
[e873838]	635	static void __wake_one(cluster * this) {
[7cf3b1d]	636	eventfd_t val;
	637
[8fc652e0]	638	/* paranoid */ verify( ! __preemption_enabled() );
[e873838]	639	/* paranoid */ verify( ready_schedule_islocked() );
[14a61b5]	640
[64a7146]	641	// Check if there is a sleeping processor
[7cf3b1d]	642	struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);
[14a61b5]	643
[7cf3b1d]	644	// If no one is sleeping: we are done
	645	if( fdp == 0p ) return;
[14a61b5]	646
[7cf3b1d]	647	int fd = 1;
[22226e4]	648	if( __atomic_load_n(&fdp->sem, __ATOMIC_SEQ_CST) != 1 ) {
	649	fd = __atomic_exchange_n(&fdp->sem, 1, __ATOMIC_RELAXED);
[7cf3b1d]	650	}
[14a61b5]	651
[7cf3b1d]	652	switch(fd) {
[4ccc150]	653	__attribute__((unused)) int ret;
[7cf3b1d]	654	case 0:
	655	// If the processor isn't ready to sleep then the exchange will already wake it up
	656	#if !defined(__CFA_NO_STATISTICS__)
	657	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
	658	} else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
	659	#endif
	660	break;
	661	case 1:
	662	// If someone else already said they will wake them: we are done
	663	#if !defined(__CFA_NO_STATISTICS__)
	664	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
	665	} else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
	666	#endif
	667	break;
	668	default:
	669	// If the processor was ready to sleep, we need to wake it up with an actual write
	670	val = 1;
[4ccc150]	671	ret = eventfd_write( fd, val );
[77adaee]	672	/* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7cf3b1d]	673
	674	#if !defined(__CFA_NO_STATISTICS__)
	675	if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
	676	} else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
	677	#endif
	678	break;
	679	}
[1eb239e4]	680
[e873838]	681	/* paranoid */ verify( ready_schedule_islocked() );
[8fc652e0]	682	/* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]	683
	684	return;
[64a7146]	685	}
[14a61b5]	686
[64a7146]	687	// Unconditionnaly wake a thread
[1eb239e4]	688	void __wake_proc(processor * this) {
[7d0ebd0]	689	/* paranoid */ verify( ! __preemption_enabled() );
	690
[64a7146]	691	__cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);
[14a61b5]	692
[22226e4]	693	this->idle_wctx.sem = 1;
[7cf3b1d]	694
[efa28d5]	695	this->idle_wctx.wake__time = rdtscl();
[262fafd9]	696
[7d0ebd0]	697	eventfd_t val;
	698	val = 1;
[d080549]	699	__attribute__((unused)) int ret = eventfd_write( this->idle_wctx.evfd, val );
[7d0ebd0]	700
[d080549]	701	/* paranoid */ verifyf( ret == 0, "Expected return to be 0, was %d\n", ret );
[7d0ebd0]	702	/* paranoid */ verify( ! __preemption_enabled() );
[92e7631]	703	}
	704
[18f7858]	705	static void idle_sleep(processor * this) {
[202c80b]	706	/* paranoid */ verify( this->idle_wctx.evfd != 1 );
	707	/* paranoid */ verify( this->idle_wctx.evfd != 2 );
	708
[7cf3b1d]	709	// Tell everyone we are ready to go do sleep
	710	for() {
[22226e4]	711	int expected = this->idle_wctx.sem;
[7cf3b1d]	712
	713	// Someone already told us to wake-up! No time for a nap.
	714	if(expected == 1) { return; }
	715
	716	// Try to mark that we are going to sleep
[22226e4]	717	if(__atomic_compare_exchange_n(&this->idle_wctx.sem, &expected, this->idle_wctx.evfd, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
[7cf3b1d]	718	// Every one agreed, taking a nap
	719	break;
	720	}
	721	}
	722
	723
[a757ba1]	724	#if !defined(__CFA_NO_STATISTICS__)
	725	if(this->print_halts) {
	726	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl());
	727	}
	728	#endif
[1757f98]	729
[a757ba1]	730	__cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
[1757f98]	731
[a757ba1]	732	{
	733	eventfd_t val;
	734	ssize_t ret = read( this->idle_wctx.evfd, &val, sizeof(val) );
	735	if(ret < 0) {
	736	switch((int)errno) {
	737	case EAGAIN:
	738	#if EAGAIN != EWOULDBLOCK
	739	case EWOULDBLOCK:
	740	#endif
	741	case EINTR:
	742	// No need to do anything special here, just assume it's a legitimate wake-up
	743	break;
	744	default:
	745	abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
[1757f98]	746	}
	747	}
[a757ba1]	748	}
[1757f98]	749
[a757ba1]	750	#if !defined(__CFA_NO_STATISTICS__)
	751	if(this->print_halts) {
	752	__cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
	753	}
[1757f98]	754	#endif
	755	}
	756
[5f5a729]	757	static bool mark_idle(__cluster_proc_list & this, processor & proc) {
[c9c1c1cb]	758	__STATS__(true, ready.sleep.halts++; )
[7cf3b1d]	759
[22226e4]	760	proc.idle_wctx.sem = 0;
[7cf3b1d]	761
[8fc652e0]	762	/* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]	763	if(!try_lock( this )) return false;
[1eb239e4]	764	this.idle++;
	765	/* paranoid */ verify( this.idle <= this.total );
[fc59b580]	766	remove(proc);
[6a9b12b]	767	insert_first(this.idles, proc);
[34b8cb7]	768
[a757ba1]	769	// update the pointer to the head wait context, which should now point to this proc.
[7cf3b1d]	770	__atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
[1eb239e4]	771	unlock( this );
[8fc652e0]	772	/* paranoid */ verify( ! __preemption_enabled() );
[5f5a729]	773
	774	return true;
[1eb239e4]	775	}
[8e16177]	776
[6a9b12b]	777	static void mark_awake(__cluster_proc_list & this, processor & proc) {
[8fc652e0]	778	/* paranoid */ verify( ! __preemption_enabled() );
[1eb239e4]	779	lock( this );
	780	this.idle--;
	781	/* paranoid */ verify( this.idle >= 0 );
	782	remove(proc);
[fc59b580]	783	insert_last(this.actives, proc);
[6a9b12b]	784
[a633f6f]	785	{
[a757ba1]	786	// update the pointer to the head wait context
[7cf3b1d]	787	struct __fd_waitctx * wctx = 0;
	788	if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
	789	__atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
[a633f6f]	790	}
	791
[34b8cb7]	792	unlock( this );
[6a9b12b]	793	/* paranoid */ verify( ! __preemption_enabled() );
[6b4cdd3]	794	}
	795
[dbe9b08]	796	//=============================================================================================
	797	// Unexpected Terminating logic
	798	//=============================================================================================
[92bfda0]	799	void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
[e84ab3d]	800	thread$ * thrd = __cfaabi_tls.this_thread;
[9d944b2]	801
[de94a60]	802	if(thrd) {
	803	int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
[1c40091]	804	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]	805
[212c2187]	806	if ( &thrd->self_cor != thrd->curr_cor ) {
	807	len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
[1c40091]	808	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[de94a60]	809	}
	810	else {
[1c40091]	811	__cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
[de94a60]	812	}
[1c273d0]	813	}
[9d944b2]	814	else {
[de94a60]	815	int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
[1c40091]	816	__cfaabi_bits_write( STDERR_FILENO, abort_text, len );
[9d944b2]	817	}
	818	}
	819
[92bfda0]	820	int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
	821	return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
[2b8bc41]	822	}
	823
[de94a60]	824	static __spinlock_t kernel_debug_lock;
	825
[9d944b2]	826	extern "C" {
[1c40091]	827	void __cfaabi_bits_acquire() {
[36982fc]	828	lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
[9d944b2]	829	}
	830
[1c40091]	831	void __cfaabi_bits_release() {
[ea7d2b0]	832	unlock( kernel_debug_lock );
[9d944b2]	833	}
[8118303]	834	}
	835
[fa21ac9]	836	//=============================================================================================
	837	// Kernel Utilities
	838	//=============================================================================================
[dddb3dd0]	839	#if defined(CFA_HAVE_LINUX_IO_URING_H)
	840	#include "io/types.hfa"
	841	#endif
	842
[2026bb6]	843	//-----------------------------------------------------------------------------
	844	// Debug
[c18bf9e]	845	bool threading_enabled(void) __attribute__((const)) libcfa_public {
[2026bb6]	846	return true;
	847	}
[c34ebf2]	848
	849	//-----------------------------------------------------------------------------
	850	// Statistics
	851	#if !defined(__CFA_NO_STATISTICS__)
[c18bf9e]	852	void print_halts( processor & this ) libcfa_public {
[c34ebf2]	853	this.print_halts = true;
	854	}
[58688bf]	855
[69914cbc]	856	static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
[8464edf]	857	/* paranoid */ verify( cltr->stats );
	858
	859	processor * it = &list`first;
	860	for(unsigned i = 0; i < count; i++) {
	861	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
	862	/* paranoid */ verify( it->local_data->this_stats );
[c33c2af]	863	// __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
[8464edf]	864	__tally_stats( cltr->stats, it->local_data->this_stats );
	865	it = &(*it)`next;
	866	}
	867	}
	868
[c18bf9e]	869	static void crawl_cluster_stats( cluster & this ) {
[8464edf]	870	// Stop the world, otherwise stats could get really messed-up
	871	// this doesn't solve all problems but does solve many
	872	// so it's probably good enough
[c33c2af]	873	disable_interrupts();
[8464edf]	874	uint_fast32_t last_size = ready_mutate_lock();
	875
	876	crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
	877	crawl_list(&this, this.procs.idles , this.procs.idle );
	878
	879	// Unlock the RWlock
	880	ready_mutate_unlock( last_size );
[c33c2af]	881	enable_interrupts();
[8464edf]	882	}
	883
	884
[c18bf9e]	885	void print_stats_now( cluster & this, int flags ) libcfa_public {
[8464edf]	886	crawl_cluster_stats( this );
[202c80b]	887	__print_stats( this.stats, flags, "Cluster", this.name, (void*)&this );
[1b033b8]	888	}
[c34ebf2]	889	#endif
[8118303]	890	// Local Variables: //
	891	// mode: c //
	892	// tab-width: 4 //
	893	// End: //

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format