Context Navigation

setup.cfa @ 74807be

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-astnew-ast-unique-exprpthread-emulationqualifiedEnum

Last change on this file since 74807be was 1eb239e4, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Removed snzi and replaced it with a fast/slow path
Property mode set to `100644`
File size: 15.9 KB

Rev	Line
[3e2b9c9]	1	//
	2	// Cforall Version 1.0.0 Copyright (C) 2020 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	// io/setup.cfa --
	8	//
	9	// Author : Thierry Delisle
	10	// Created On : Fri Jul 31 16:25:51 2020
	11	// Last Modified By :
	12	// Last Modified On :
	13	// Update Count :
	14	//
	15
	16	#define __cforall_thread__
	17	#define _GNU_SOURCE /* See feature_test_macros(7) */
	18
	19	#include "io/types.hfa"
[c44d652]	20	#include "kernel.hfa"
[3e2b9c9]	21
	22	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
	23	void __kernel_io_startup() {
	24	// Nothing to do without io_uring
	25	}
	26
	27	void __kernel_io_shutdown() {
	28	// Nothing to do without io_uring
	29	}
	30
[f277633e]	31	void ?{}(io_context_params & this) {}
	32
[3e2b9c9]	33	void ?{}(io_context & this, struct cluster & cl) {}
	34	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
	35
	36	void ^?{}(io_context & this) {}
	37	void ^?{}(io_context & this, bool cluster_context) {}
	38
	39	#else
	40	#include <errno.h>
	41	#include <stdint.h>
	42	#include <string.h>
	43	#include <signal.h>
	44	#include <unistd.h>
	45
	46	extern "C" {
	47	#include <pthread.h>
	48	#include <sys/epoll.h>
	49	#include <sys/mman.h>
	50	#include <sys/syscall.h>
	51
	52	#include <linux/io_uring.h>
	53	}
	54
	55	#include "bitmanip.hfa"
	56	#include "kernel_private.hfa"
	57	#include "thread.hfa"
	58
	59	void ?{}(io_context_params & this) {
	60	this.num_entries = 256;
	61	this.num_ready = 256;
	62	this.submit_aff = -1;
	63	this.eager_submits = false;
	64	this.poller_submits = false;
	65	this.poll_submit = false;
	66	this.poll_complete = false;
	67	}
	68
	69	static void * __io_poller_slow( void * arg );
	70
	71	// Weirdly, some systems that do support io_uring don't actually define these
	72	#ifdef __alpha__
	73	/*
	74	* alpha is the only exception, all other architectures
	75	* have common numbers for new system calls.
	76	*/
	77	#ifndef __NR_io_uring_setup
	78	#define __NR_io_uring_setup 535
	79	#endif
	80	#ifndef __NR_io_uring_enter
	81	#define __NR_io_uring_enter 536
	82	#endif
	83	#ifndef __NR_io_uring_register
	84	#define __NR_io_uring_register 537
	85	#endif
	86	#else /* !__alpha__ */
	87	#ifndef __NR_io_uring_setup
	88	#define __NR_io_uring_setup 425
	89	#endif
	90	#ifndef __NR_io_uring_enter
	91	#define __NR_io_uring_enter 426
	92	#endif
	93	#ifndef __NR_io_uring_register
	94	#define __NR_io_uring_register 427
	95	#endif
	96	#endif
	97
	98	//=============================================================================================
	99	// I/O Startup / Shutdown logic + Master Poller
	100	//=============================================================================================
	101
	102	// IO Master poller loop forward
	103	static void * iopoll_loop( __attribute__((unused)) void * args );
	104
	105	static struct {
	106	pthread_t thrd; // pthread handle to io poller thread
	107	void * stack; // pthread stack for io poller thread
	108	int epollfd; // file descriptor to the epoll instance
	109	volatile bool run; // Whether or not to continue
	110	} iopoll;
	111
	112	void __kernel_io_startup(void) {
	113	__cfaabi_dbg_print_safe( "Kernel : Creating EPOLL instance\n" );
	114
	115	iopoll.epollfd = epoll_create1(0);
	116	if (iopoll.epollfd == -1) {
	117	abort( "internal error, epoll_create1\n");
	118	}
	119
	120	__cfaabi_dbg_print_safe( "Kernel : Starting io poller thread\n" );
	121
	122	iopoll.run = true;
	123	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
	124	}
	125
	126	void __kernel_io_shutdown(void) {
	127	// Notify the io poller thread of the shutdown
	128	iopoll.run = false;
	129	sigval val = { 1 };
	130	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
	131
	132	// Wait for the io poller thread to finish
	133
	134	pthread_join( iopoll.thrd, 0p );
	135	free( iopoll.stack );
	136
	137	int ret = close(iopoll.epollfd);
	138	if (ret == -1) {
	139	abort( "internal error, close epoll\n");
	140	}
	141
	142	// Io polling is now fully stopped
	143
	144	__cfaabi_dbg_print_safe( "Kernel : IO poller stopped\n" );
	145	}
	146
	147	static void * iopoll_loop( __attribute__((unused)) void * args ) {
	148	__processor_id_t id;
	149	id.id = doregister(&id);
	150	__cfaabi_dbg_print_safe( "Kernel : IO poller thread starting\n" );
	151
	152	// Block signals to control when they arrive
	153	sigset_t mask;
	154	sigfillset(&mask);
	155	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
	156	abort( "internal error, pthread_sigmask" );
	157	}
	158
	159	sigdelset( &mask, SIGUSR1 );
	160
	161	// Create sufficient events
	162	struct epoll_event events[10];
	163	// Main loop
	164	while( iopoll.run ) {
	165	// Wait for events
	166	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
	167
	168	// Check if an error occured
	169	if (nfds == -1) {
	170	if( errno == EINTR ) continue;
	171	abort( "internal error, pthread_sigmask" );
	172	}
	173
	174	for(i; nfds) {
	175	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
	176	/* paranoid */ verify( io_ctx );
	177	__cfadbg_print_safe(io_core, "Kernel I/O : Unparking io poller %p\n", io_ctx);
	178	#if !defined( __CFA_NO_STATISTICS__ )
	179	kernelTLS.this_stats = io_ctx->self.curr_cluster->stats;
	180	#endif
	181	__post( io_ctx->sem, &id );
	182	}
	183	}
	184
	185	__cfaabi_dbg_print_safe( "Kernel : IO poller thread stopping\n" );
	186	unregister(&id);
	187	return 0p;
	188	}
	189
	190	//=============================================================================================
	191	// I/O Context Constrution/Destruction
	192	//=============================================================================================
	193
	194	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
	195	void main( $io_ctx_thread & this );
	196	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
	197	void ^?{}( $io_ctx_thread & mutex this ) {}
	198
	199	static void __io_create ( __io_data & this, const io_context_params & params_in );
	200	static void __io_destroy( __io_data & this );
	201
	202	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
	203	(this.thrd){ cl };
	204	this.thrd.ring = malloc();
	205	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
	206	__io_create( *this.thrd.ring, params );
	207
	208	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
	209	this.thrd.done = false;
	210	__thrd_start( this.thrd, main );
	211
	212	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
	213	}
	214
	215	void ?{}(io_context & this, struct cluster & cl) {
	216	io_context_params params;
	217	(this){ cl, params };
	218	}
	219
	220	void ^?{}(io_context & this, bool cluster_context) {
	221	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
	222
	223	// Notify the thread of the shutdown
	224	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
	225
	226	// If this is an io_context within a cluster, things get trickier
	227	$thread & thrd = this.thrd.self;
	228	if( cluster_context ) {
	229	cluster & cltr = *thrd.curr_cluster;
[1eb239e4]	230	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
[3e2b9c9]	231	/* paranoid */ verify( !ready_mutate_islocked() );
	232
	233	// We need to adjust the clean-up based on where the thread is
	234	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
	235
	236	ready_schedule_lock( (struct __processor_id_t *)active_processor() );
	237
	238	// This is the tricky case
	239	// The thread was preempted and now it is on the ready queue
	240	// The thread should be the last on the list
	241	/* paranoid */ verify( thrd.link.next != 0p );
	242
	243	// Remove the thread from the ready queue of this cluster
	244	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
	245	/* paranoid */ verify( removed );
	246	thrd.link.next = 0p;
	247	thrd.link.prev = 0p;
	248	__cfaabi_dbg_debug_do( thrd.unpark_stale = true );
	249
	250	// Fixup the thread state
	251	thrd.state = Blocked;
	252	thrd.ticket = 0;
	253	thrd.preempted = __NO_PREEMPTION;
	254
	255	ready_schedule_unlock( (struct __processor_id_t *)active_processor() );
	256
	257	// Pretend like the thread was blocked all along
	258	}
	259	// !!! This is not an else if !!!
	260	if( thrd.state == Blocked ) {
	261
	262	// This is the "easy case"
	263	// The thread is parked and can easily be moved to active cluster
	264	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
	265	thrd.curr_cluster = active_cluster();
	266
	267	// unpark the fast io_poller
	268	unpark( &thrd __cfaabi_dbg_ctx2 );
	269	}
	270	else {
	271
	272	// The thread is in a weird state
	273	// I don't know what to do here
	274	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
	275	}
	276	} else {
	277	unpark( &thrd __cfaabi_dbg_ctx2 );
	278	}
	279
	280	^(this.thrd){};
	281	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
	282
	283	__io_destroy( *this.thrd.ring );
	284	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
	285
	286	free(this.thrd.ring);
	287	}
	288
	289	void ^?{}(io_context & this) {
	290	^(this){ false };
	291	}
	292
	293	static void __io_create( __io_data & this, const io_context_params & params_in ) {
	294	// Step 1 : call to setup
	295	struct io_uring_params params;
	296	memset(&params, 0, sizeof(params));
	297	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
	298	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
	299
	300	uint32_t nentries = params_in.num_entries;
	301
	302	int fd = syscall(__NR_io_uring_setup, nentries, &params );
	303	if(fd < 0) {
	304	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
	305	}
	306
	307	// Step 2 : mmap result
	308	memset( &this, 0, sizeof(struct __io_data) );
	309	struct __submition_data & sq = this.submit_q;
	310	struct __completion_data & cq = this.completion_q;
	311
	312	// calculate the right ring size
	313	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
	314	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
	315
	316	// Requires features
	317	#if defined(IORING_FEAT_SINGLE_MMAP)
	318	// adjust the size according to the parameters
	319	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
	320	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
	321	}
	322	#endif
	323
	324	// mmap the Submit Queue into existence
	325	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
	326	if (sq.ring_ptr == (void*)MAP_FAILED) {
	327	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
	328	}
	329
	330	// Requires features
	331	#if defined(IORING_FEAT_SINGLE_MMAP)
	332	// mmap the Completion Queue into existence (may or may not be needed)
	333	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
	334	cq.ring_ptr = sq.ring_ptr;
	335	}
	336	else
	337	#endif
	338	{
	339	// We need multiple call to MMAP
	340	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
	341	if (cq.ring_ptr == (void*)MAP_FAILED) {
	342	munmap(sq.ring_ptr, sq.ring_sz);
	343	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
	344	}
	345	}
	346
	347	// mmap the submit queue entries
	348	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
	349	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
	350	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
	351	munmap(sq.ring_ptr, sq.ring_sz);
	352	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
	353	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
	354	}
	355
	356	// Get the pointers from the kernel to fill the structure
	357	// submit queue
	358	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
	359	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
	360	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
	361	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
	362	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
	363	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
	364	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
	365	sq.prev_head = *sq.head;
	366
	367	{
	368	const uint32_t num = *sq.num;
	369	for( i; num ) {
	370	sq.sqes[i].user_data = 0ul64;
	371	}
	372	}
	373
	374	(sq.lock){};
	375	(sq.release_lock){};
	376
	377	if( params_in.poller_submits \|\| params_in.eager_submits ) {
	378	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
	379	sq.ready_cnt = max( params_in.num_ready, 8 );
	380	sq.ready = alloc_align( 64, sq.ready_cnt );
	381	for(i; sq.ready_cnt) {
	382	sq.ready[i] = -1ul32;
	383	}
	384	}
	385	else {
	386	sq.ready_cnt = 0;
	387	sq.ready = 0p;
	388	}
	389
	390	// completion queue
	391	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
	392	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
	393	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
	394	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
	395	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
	396	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
	397
	398	// some paranoid checks
	399	/* paranoid / verifyf( (cq.mask) == ((cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (cq.num) - 1ul32, cq.num, cq.mask );
	400	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
	401	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
	402	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
	403
	404	/* paranoid / verifyf( (sq.mask) == ((sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (sq.num) - 1ul32, sq.num, sq.mask );
	405	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
	406	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
	407	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
	408
	409	// Update the global ring info
	410	this.ring_flags = params.flags;
	411	this.fd = fd;
	412	this.eager_submits = params_in.eager_submits;
	413	this.poller_submits = params_in.poller_submits;
	414	}
	415
	416	static void __io_destroy( __io_data & this ) {
	417	// Shutdown the io rings
	418	struct __submition_data & sq = this.submit_q;
	419	struct __completion_data & cq = this.completion_q;
	420
	421	// unmap the submit queue entries
	422	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
	423
	424	// unmap the Submit Queue ring
	425	munmap(sq.ring_ptr, sq.ring_sz);
	426
	427	// unmap the Completion Queue ring, if it is different
	428	if (cq.ring_ptr != sq.ring_ptr) {
	429	munmap(cq.ring_ptr, cq.ring_sz);
	430	}
	431
	432	// close the file descriptor
	433	close(this.fd);
	434
	435	free( this.submit_q.ready ); // Maybe null, doesn't matter
	436	}
	437
	438	//=============================================================================================
	439	// I/O Context Sleep
	440	//=============================================================================================
	441
	442	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
	443	ev.events = EPOLLIN \| EPOLLONESHOT;
	444	ev.data.u64 = (uint64_t)&ctx;
	445	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->fd, &ev);
	446	if (ret < 0) {
	447	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
	448	}
	449	}
	450
	451	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
	452	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->fd, &ev);
	453	if (ret < 0) {
	454	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
	455	}
	456	}
	457
	458	//=============================================================================================
	459	// I/O Context Misc Setup
	460	//=============================================================================================
	461	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
	462	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
	463	if( ret < 0 ) {
	464	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	465	}
	466
	467	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
	468	}
	469
	470	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
	471	for(i; cltr.io.cnt) {
	472	register_fixed_files( cltr.io.ctxs[i], files, count );
	473	}
	474	}
[c44d652]	475	#endif

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

Download in other formats:

Original Format