Context Navigation

setup.cfa @ ece0e80

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

Last change on this file since ece0e80 was ece0e80, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Added prints. Naive implementation of cancel. Server now shutdown cleanly.
Property mode set to `100644`
File size: 17.0 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
[80444bb]	19	#if defined(__CFA_DEBUG__)
	20	// #define __CFA_DEBUG_PRINT_IO__
	21	// #define __CFA_DEBUG_PRINT_IO_CORE__
	22	#endif
	23
[3e2b9c9]	24	#include "io/types.hfa"
[c44d652]	25	#include "kernel.hfa"
[3e2b9c9]	26
	27	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
	28	void __kernel_io_startup() {
	29	// Nothing to do without io_uring
	30	}
	31
	32	void __kernel_io_shutdown() {
	33	// Nothing to do without io_uring
	34	}
	35
[f277633e]	36	void ?{}(io_context_params & this) {}
	37
[3e2b9c9]	38	void ?{}(io_context & this, struct cluster & cl) {}
	39	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
	40
	41	void ^?{}(io_context & this) {}
	42	void ^?{}(io_context & this, bool cluster_context) {}
	43
	44	#else
	45	#include <errno.h>
	46	#include <stdint.h>
	47	#include <string.h>
	48	#include <signal.h>
	49	#include <unistd.h>
	50
	51	extern "C" {
	52	#include <pthread.h>
	53	#include <sys/epoll.h>
	54	#include <sys/mman.h>
	55	#include <sys/syscall.h>
	56
	57	#include <linux/io_uring.h>
	58	}
	59
	60	#include "bitmanip.hfa"
	61	#include "kernel_private.hfa"
	62	#include "thread.hfa"
	63
	64	void ?{}(io_context_params & this) {
	65	this.num_entries = 256;
	66	this.num_ready = 256;
	67	this.submit_aff = -1;
	68	this.eager_submits = false;
	69	this.poller_submits = false;
	70	this.poll_submit = false;
	71	this.poll_complete = false;
	72	}
	73
	74	static void * __io_poller_slow( void * arg );
	75
	76	// Weirdly, some systems that do support io_uring don't actually define these
	77	#ifdef __alpha__
	78	/*
	79	* alpha is the only exception, all other architectures
	80	* have common numbers for new system calls.
	81	*/
	82	#ifndef __NR_io_uring_setup
	83	#define __NR_io_uring_setup 535
	84	#endif
	85	#ifndef __NR_io_uring_enter
	86	#define __NR_io_uring_enter 536
	87	#endif
	88	#ifndef __NR_io_uring_register
	89	#define __NR_io_uring_register 537
	90	#endif
	91	#else /* !__alpha__ */
	92	#ifndef __NR_io_uring_setup
	93	#define __NR_io_uring_setup 425
	94	#endif
	95	#ifndef __NR_io_uring_enter
	96	#define __NR_io_uring_enter 426
	97	#endif
	98	#ifndef __NR_io_uring_register
	99	#define __NR_io_uring_register 427
	100	#endif
	101	#endif
	102
	103	//=============================================================================================
	104	// I/O Startup / Shutdown logic + Master Poller
	105	//=============================================================================================
	106
	107	// IO Master poller loop forward
	108	static void * iopoll_loop( __attribute__((unused)) void * args );
	109
	110	static struct {
	111	pthread_t thrd; // pthread handle to io poller thread
	112	void * stack; // pthread stack for io poller thread
	113	int epollfd; // file descriptor to the epoll instance
	114	volatile bool run; // Whether or not to continue
	115	} iopoll;
	116
	117	void __kernel_io_startup(void) {
[80444bb]	118	__cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
[3e2b9c9]	119
	120	iopoll.epollfd = epoll_create1(0);
	121	if (iopoll.epollfd == -1) {
	122	abort( "internal error, epoll_create1\n");
	123	}
	124
[80444bb]	125	__cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
[3e2b9c9]	126
	127	iopoll.run = true;
	128	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
	129	}
	130
	131	void __kernel_io_shutdown(void) {
	132	// Notify the io poller thread of the shutdown
	133	iopoll.run = false;
	134	sigval val = { 1 };
	135	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
	136
	137	// Wait for the io poller thread to finish
	138
[bfcf6b9]	139	__destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
[3e2b9c9]	140
	141	int ret = close(iopoll.epollfd);
	142	if (ret == -1) {
	143	abort( "internal error, close epoll\n");
	144	}
	145
	146	// Io polling is now fully stopped
	147
[80444bb]	148	__cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
[3e2b9c9]	149	}
	150
	151	static void * iopoll_loop( __attribute__((unused)) void * args ) {
	152	__processor_id_t id;
[58d64a4]	153	id.full_proc = false;
[3e2b9c9]	154	id.id = doregister(&id);
[8fc652e0]	155	__cfaabi_tls.this_proc_id = &id;
[80444bb]	156	__cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
[3e2b9c9]	157
	158	// Block signals to control when they arrive
	159	sigset_t mask;
	160	sigfillset(&mask);
	161	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
	162	abort( "internal error, pthread_sigmask" );
	163	}
	164
	165	sigdelset( &mask, SIGUSR1 );
	166
	167	// Create sufficient events
	168	struct epoll_event events[10];
	169	// Main loop
	170	while( iopoll.run ) {
[ece0e80]	171	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
	172
[3e2b9c9]	173	// Wait for events
	174	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
	175
[ece0e80]	176	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
	177
[3e2b9c9]	178	// Check if an error occured
	179	if (nfds == -1) {
	180	if( errno == EINTR ) continue;
	181	abort( "internal error, pthread_sigmask" );
	182	}
	183
	184	for(i; nfds) {
	185	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
	186	/* paranoid */ verify( io_ctx );
[ece0e80]	187	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %p\n", io_ctx);
[3e2b9c9]	188	#if !defined( __CFA_NO_STATISTICS__ )
[8fc652e0]	189	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
[3e2b9c9]	190	#endif
[e873838]	191	post( io_ctx->sem );
[3e2b9c9]	192	}
	193	}
	194
[80444bb]	195	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
[3e2b9c9]	196	unregister(&id);
	197	return 0p;
	198	}
	199
	200	//=============================================================================================
	201	// I/O Context Constrution/Destruction
	202	//=============================================================================================
	203
	204	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
	205	void main( $io_ctx_thread & this );
	206	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
	207	void ^?{}( $io_ctx_thread & mutex this ) {}
	208
	209	static void __io_create ( __io_data & this, const io_context_params & params_in );
	210	static void __io_destroy( __io_data & this );
	211
	212	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
	213	(this.thrd){ cl };
	214	this.thrd.ring = malloc();
	215	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
	216	__io_create( *this.thrd.ring, params );
	217
	218	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
	219	this.thrd.done = false;
	220	__thrd_start( this.thrd, main );
	221
	222	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
	223	}
	224
	225	void ?{}(io_context & this, struct cluster & cl) {
	226	io_context_params params;
	227	(this){ cl, params };
	228	}
	229
	230	void ^?{}(io_context & this, bool cluster_context) {
	231	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
	232
	233	// Notify the thread of the shutdown
	234	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
	235
	236	// If this is an io_context within a cluster, things get trickier
	237	$thread & thrd = this.thrd.self;
	238	if( cluster_context ) {
[ece0e80]	239	// We are about to do weird things with the threads
	240	// we don't need interrupts to complicate everything
	241	disable_interrupts();
	242
	243	// Get cluster info
[3e2b9c9]	244	cluster & cltr = *thrd.curr_cluster;
[1eb239e4]	245	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
[3e2b9c9]	246	/* paranoid */ verify( !ready_mutate_islocked() );
	247
	248	// We need to adjust the clean-up based on where the thread is
	249	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
[ece0e80]	250	// This is the tricky case
	251	// The thread was preempted or ready to run and now it is on the ready queue
	252	// but the cluster is shutting down, so there aren't any processors to run the ready queue
	253	// the solution is to steal the thread from the ready-queue and pretend it was blocked all along
[3e2b9c9]	254
[e873838]	255	ready_schedule_lock();
[ece0e80]	256	// The thread should on the list
[3e2b9c9]	257	/* paranoid */ verify( thrd.link.next != 0p );
	258
	259	// Remove the thread from the ready queue of this cluster
[ece0e80]	260	// The thread should be the last on the list
[3e2b9c9]	261	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
	262	/* paranoid */ verify( removed );
	263	thrd.link.next = 0p;
	264	thrd.link.prev = 0p;
	265
	266	// Fixup the thread state
	267	thrd.state = Blocked;
[6a77224]	268	thrd.ticket = TICKET_BLOCKED;
[3e2b9c9]	269	thrd.preempted = __NO_PREEMPTION;
	270
[e873838]	271	ready_schedule_unlock();
[3e2b9c9]	272
	273	// Pretend like the thread was blocked all along
	274	}
	275	// !!! This is not an else if !!!
[ece0e80]	276	// Ok, now the thread is blocked (whether we cheated to get here or not)
[3e2b9c9]	277	if( thrd.state == Blocked ) {
	278	// This is the "easy case"
	279	// The thread is parked and can easily be moved to active cluster
	280	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
	281	thrd.curr_cluster = active_cluster();
	282
	283	// unpark the fast io_poller
[0d72d45]	284	unpark( &thrd );
[3e2b9c9]	285	}
	286	else {
	287	// The thread is in a weird state
	288	// I don't know what to do here
	289	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
	290	}
[ece0e80]	291
	292	// The weird thread kidnapping stuff is over, restore interrupts.
	293	enable_interrupts( __cfaabi_dbg_ctx );
[3e2b9c9]	294	} else {
[3c80ccc]	295	post( this.thrd.sem );
[3e2b9c9]	296	}
	297
	298	^(this.thrd){};
	299	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
	300
	301	__io_destroy( *this.thrd.ring );
	302	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
	303
	304	free(this.thrd.ring);
	305	}
	306
	307	void ^?{}(io_context & this) {
	308	^(this){ false };
	309	}
	310
	311	static void __io_create( __io_data & this, const io_context_params & params_in ) {
	312	// Step 1 : call to setup
	313	struct io_uring_params params;
	314	memset(&params, 0, sizeof(params));
	315	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
	316	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
	317
[4998155]	318	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
[63fe427c]	319	if( !is_pow2(nentries) ) {
	320	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
	321	}
	322	if( params_in.poller_submits && params_in.eager_submits ) {
	323	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
	324	}
[3e2b9c9]	325
	326	int fd = syscall(__NR_io_uring_setup, nentries, &params );
	327	if(fd < 0) {
	328	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
	329	}
	330
	331	// Step 2 : mmap result
	332	memset( &this, 0, sizeof(struct __io_data) );
	333	struct __submition_data & sq = this.submit_q;
	334	struct __completion_data & cq = this.completion_q;
	335
	336	// calculate the right ring size
	337	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
	338	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
	339
	340	// Requires features
	341	#if defined(IORING_FEAT_SINGLE_MMAP)
	342	// adjust the size according to the parameters
	343	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
	344	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
	345	}
	346	#endif
	347
	348	// mmap the Submit Queue into existence
	349	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
	350	if (sq.ring_ptr == (void*)MAP_FAILED) {
	351	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
	352	}
	353
	354	// Requires features
	355	#if defined(IORING_FEAT_SINGLE_MMAP)
	356	// mmap the Completion Queue into existence (may or may not be needed)
	357	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
	358	cq.ring_ptr = sq.ring_ptr;
	359	}
	360	else
	361	#endif
	362	{
	363	// We need multiple call to MMAP
	364	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
	365	if (cq.ring_ptr == (void*)MAP_FAILED) {
	366	munmap(sq.ring_ptr, sq.ring_sz);
	367	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
	368	}
	369	}
	370
	371	// mmap the submit queue entries
	372	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
	373	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
	374	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
	375	munmap(sq.ring_ptr, sq.ring_sz);
	376	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
	377	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
	378	}
	379
	380	// Get the pointers from the kernel to fill the structure
	381	// submit queue
[4998155]	382	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
	383	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
	384	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
	385	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
	386	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
	387	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
	388	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
[3e2b9c9]	389	sq.prev_head = *sq.head;
	390
	391	{
[4998155]	392	const __u32 num = *sq.num;
[3e2b9c9]	393	for( i; num ) {
	394	sq.sqes[i].user_data = 0ul64;
	395	}
	396	}
	397
[2fafe7e]	398	(sq.submit_lock){};
[3e2b9c9]	399	(sq.release_lock){};
	400
	401	if( params_in.poller_submits \|\| params_in.eager_submits ) {
	402	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
	403	sq.ready_cnt = max( params_in.num_ready, 8 );
[ceb7db8]	404	sq.ready = alloc( sq.ready_cnt, 64`align );
[3e2b9c9]	405	for(i; sq.ready_cnt) {
	406	sq.ready[i] = -1ul32;
	407	}
[1095ccd]	408	sq.prev_ready = 0;
[3e2b9c9]	409	}
	410	else {
	411	sq.ready_cnt = 0;
	412	sq.ready = 0p;
[1095ccd]	413	sq.prev_ready = 0;
[3e2b9c9]	414	}
	415
	416	// completion queue
[4998155]	417	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
	418	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
	419	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
	420	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
	421	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
	422	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
[3e2b9c9]	423
	424	// some paranoid checks
	425	/* 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 );
	426	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
	427	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
	428	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
	429
	430	/* 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 );
	431	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
	432	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
	433	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
	434
	435	// Update the global ring info
	436	this.ring_flags = params.flags;
	437	this.fd = fd;
	438	this.eager_submits = params_in.eager_submits;
	439	this.poller_submits = params_in.poller_submits;
	440	}
	441
	442	static void __io_destroy( __io_data & this ) {
	443	// Shutdown the io rings
	444	struct __submition_data & sq = this.submit_q;
	445	struct __completion_data & cq = this.completion_q;
	446
	447	// unmap the submit queue entries
	448	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
	449
	450	// unmap the Submit Queue ring
	451	munmap(sq.ring_ptr, sq.ring_sz);
	452
	453	// unmap the Completion Queue ring, if it is different
	454	if (cq.ring_ptr != sq.ring_ptr) {
	455	munmap(cq.ring_ptr, cq.ring_sz);
	456	}
	457
	458	// close the file descriptor
	459	close(this.fd);
	460
	461	free( this.submit_q.ready ); // Maybe null, doesn't matter
	462	}
	463
	464	//=============================================================================================
	465	// I/O Context Sleep
	466	//=============================================================================================
	467
	468	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
	469	ev.events = EPOLLIN \| EPOLLONESHOT;
[4998155]	470	ev.data.u64 = (__u64)&ctx;
[3e2b9c9]	471	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->fd, &ev);
	472	if (ret < 0) {
	473	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
	474	}
	475	}
	476
	477	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
[ece0e80]	478	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %p\n", &ctx);
[3e2b9c9]	479	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->fd, &ev);
	480	if (ret < 0) {
	481	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
	482	}
	483	}
	484
	485	//=============================================================================================
	486	// I/O Context Misc Setup
	487	//=============================================================================================
	488	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
	489	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
	490	if( ret < 0 ) {
	491	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	492	}
	493
	494	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
	495	}
	496
	497	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
	498	for(i; cltr.io.cnt) {
	499	register_fixed_files( cltr.io.ctxs[i], files, count );
	500	}
	501	}
[c44d652]	502	#endif

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

Download in other formats:

Original Format