Context Navigation

io.cfa@ dab09ad

Visit:

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast new-ast-unique-expr pthread-emulation qualifiedEnum

Last change on this file since dab09ad was eafec07, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
More pointer fixes
Property mode set to `100644`
File size: 13.6 KB

Rev	Line
[ecf6b46]	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.cfa --
	8	//
	9	// Author : Thierry Delisle
	10	// Created On : Thu Apr 23 17:31:00 2020
	11	// Last Modified By :
	12	// Last Modified On :
	13	// Update Count :
	14	//
	15
[3e2b9c9]	16	#define __cforall_thread__
	17
[20ab637]	18	#if defined(__CFA_DEBUG__)
	19	// #define __CFA_DEBUG_PRINT_IO__
[e660761]	20	// #define __CFA_DEBUG_PRINT_IO_CORE__
[20ab637]	21	#endif
[4069faad]	22
[f6660520]	23
[3e2b9c9]	24	#if defined(CFA_HAVE_LINUX_IO_URING_H)
[31bb2e1]	25	#define _GNU_SOURCE /* See feature_test_macros(7) */
	26	#include <errno.h>
[3e2b9c9]	27	#include <signal.h>
[31bb2e1]	28	#include <stdint.h>
	29	#include <string.h>
	30	#include <unistd.h>
	31
[92976d9]	32	extern "C" {
[f00b26d4]	33	#include <sys/epoll.h>
[92976d9]	34	#include <sys/syscall.h>
	35
	36	#include <linux/io_uring.h>
	37	}
	38
[3e2b9c9]	39	#include "stats.hfa"
	40	#include "kernel.hfa"
	41	#include "kernel/fwd.hfa"
	42	#include "io/types.hfa"
[185efe6]	43
[92976d9]	44	//=============================================================================================
[3e2b9c9]	45	// I/O Syscall
[92976d9]	46	//=============================================================================================
[3e2b9c9]	47	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
[20ab637]	48	bool need_sys_to_submit = false;
	49	bool need_sys_to_complete = false;
	50	unsigned flags = 0;
	51
	52	TO_SUBMIT:
	53	if( to_submit > 0 ) {
	54	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
	55	need_sys_to_submit = true;
	56	break TO_SUBMIT;
	57	}
	58	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
	59	need_sys_to_submit = true;
	60	flags \|= IORING_ENTER_SQ_WAKEUP;
	61	}
	62	}
	63
	64	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
	65	flags \|= IORING_ENTER_GETEVENTS;
	66	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
	67	need_sys_to_complete = true;
	68	}
	69	}
	70
	71	int ret = 0;
	72	if( need_sys_to_submit \|\| need_sys_to_complete ) {
[f00b26d4]	73	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, 0p, _NSIG / 8);
[20ab637]	74	if( ret < 0 ) {
	75	switch((int)errno) {
	76	case EAGAIN:
	77	case EINTR:
	78	ret = -1;
	79	break;
	80	default:
	81	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	82	}
	83	}
	84	}
	85
	86	// Memory barrier
	87	__atomic_thread_fence( __ATOMIC_SEQ_CST );
	88	return ret;
	89	}
	90
[92976d9]	91	//=============================================================================================
	92	// I/O Polling
	93	//=============================================================================================
[1d5e4711]	94	static unsigned __collect_submitions( struct __io_data & ring );
[34b61882]	95	static uint32_t __release_consumed_submission( struct __io_data & ring );
[1d5e4711]	96
[f00b26d4]	97	static inline void process(struct io_uring_cqe & cqe ) {
[5751a56]	98	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
	99	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
	100
	101	data->result = cqe.res;
[f00b26d4]	102	unpark( data->thrd __cfaabi_dbg_ctx2 );
[5751a56]	103	}
	104
[92976d9]	105	// Process a single completion message from the io_uring
	106	// This is NOT thread-safe
[f00b26d4]	107	static [int, bool] __drain_io( & struct __io_data ring ) {
[e46c753]	108	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
	109
[5dadc9b7]	110	unsigned to_submit = 0;
[f00b26d4]	111	if( ring.poller_submits ) {
[5dadc9b7]	112	// If the poller thread also submits, then we need to aggregate the submissions which are ready
[e46c753]	113	to_submit = __collect_submitions( ring );
[5dadc9b7]	114	}
	115
[f00b26d4]	116	int ret = __io_uring_enter(ring, to_submit, true);
[20ab637]	117	if( ret < 0 ) {
	118	return [0, true];
	119	}
[1d5e4711]	120
[20ab637]	121	// update statistics
	122	if (to_submit > 0) {
[1d5e4711]	123	__STATS__( true,
	124	if( to_submit > 0 ) {
	125	io.submit_q.submit_avg.rdy += to_submit;
	126	io.submit_q.submit_avg.csm += ret;
	127	io.submit_q.submit_avg.cnt += 1;
	128	}
	129	)
[6f121b8]	130	}
	131
[20ab637]	132	// Release the consumed SQEs
	133	__release_consumed_submission( ring );
[6f121b8]	134
[d384787]	135	// Drain the queue
[92976d9]	136	unsigned head = *ring.completion_q.head;
[6f121b8]	137	unsigned tail = *ring.completion_q.tail;
	138	const uint32_t mask = *ring.completion_q.mask;
	139
[d384787]	140	// Nothing was new return 0
	141	if (head == tail) {
[e46c753]	142	return [0, to_submit > 0];
[d384787]	143	}
[92976d9]	144
[d384787]	145	uint32_t count = tail - head;
[1d5e4711]	146	/* paranoid */ verify( count != 0 );
[d384787]	147	for(i; count) {
[6f121b8]	148	unsigned idx = (head + i) & mask;
[d384787]	149	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
[92976d9]	150
[d384787]	151	/* paranoid */ verify(&cqe);
[92976d9]	152
[f00b26d4]	153	process( cqe );
[d384787]	154	}
[2d8f7b0]	155
[92976d9]	156	// Mark to the kernel that the cqe has been seen
	157	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
[6f121b8]	158	__atomic_thread_fence( __ATOMIC_SEQ_CST );
[d384787]	159	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
[92976d9]	160
[5dadc9b7]	161	return [count, count > 0 \|\| to_submit > 0];
[92976d9]	162	}
	163
[f00b26d4]	164	void main( $io_ctx_thread & this ) {
	165	epoll_event ev;
[3e2b9c9]	166	__ioctx_register( this, ev );
[1539bbd]	167
[f00b26d4]	168	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %p for ring %p ready\n", &this, &this.ring);
[1539bbd]	169
[4e74466]	170	int reset = 0;
[61dd73d]	171	// Then loop until we need to start
[f00b26d4]	172	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
[61dd73d]	173	// Drain the io
[5dadc9b7]	174	int count;
	175	bool again;
[13c5e19]	176	disable_interrupts();
[f00b26d4]	177	[count, again] = __drain_io( *this.ring );
[5dadc9b7]	178
[13c5e19]	179	if(!again) reset++;
[3c039b0]	180
[13c5e19]	181	// Update statistics
[47746a2]	182	__STATS__( true,
	183	io.complete_q.completed_avg.val += count;
	184	io.complete_q.completed_avg.fast_cnt += 1;
	185	)
[13c5e19]	186	enable_interrupts( __cfaabi_dbg_ctx );
[3c039b0]	187
[5dadc9b7]	188	// If we got something, just yield and check again
[4e74466]	189	if(reset < 5) {
[61dd73d]	190	yield();
	191	}
[5dadc9b7]	192	// We didn't get anything baton pass to the slow poller
[61dd73d]	193	else {
[f00b26d4]	194	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %p\n", &this.self);
[5dadc9b7]	195	reset = 0;
	196
[3e2b9c9]	197	// block this thread
	198	__ioctx_prepare_block( this, ev );
[f00b26d4]	199	wait( this.sem );
[f6660520]	200	}
	201	}
[61dd73d]	202
	203	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
	204	}
[f6660520]	205
[92976d9]	206	//=============================================================================================
	207	// I/O Submissions
	208	//=============================================================================================
	209
[2d8f7b0]	210	// Submition steps :
[e46c753]	211	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
[2d8f7b0]	212	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
	213	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
	214	// need to write an allocator that allows allocating concurrently.
	215	//
[e46c753]	216	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
[2d8f7b0]	217	//
[e46c753]	218	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
[2d8f7b0]	219	// needs to arrive to two concensus at the same time:
	220	// A - The order in which entries are listed in the array: no two threads must pick the
	221	// same index for their entries
	222	// B - When can the tail be update for the kernel. EVERY entries in the array between
	223	// head and tail must be fully filled and shouldn't ever be touched again.
	224	//
	225
[31bb2e1]	226	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
[e46c753]	227	/* paranoid */ verify( data != 0 );
[13c5e19]	228
[6f121b8]	229	// Prepare the data we need
	230	__attribute((unused)) int len = 0;
	231	__attribute((unused)) int block = 0;
	232	uint32_t cnt = *ring.submit_q.num;
	233	uint32_t mask = *ring.submit_q.mask;
[8ae4165]	234
	235	disable_interrupts();
	236	uint32_t off = __tls_rand();
	237	enable_interrupts( __cfaabi_dbg_ctx );
[6f121b8]	238
	239	// Loop around looking for an available spot
[13c5e19]	240	for() {
[6f121b8]	241	// Look through the list starting at some offset
	242	for(i; cnt) {
	243	uint64_t expected = 0;
	244	uint32_t idx = (i + off) & mask;
	245	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
[eafec07]	246	volatile uint64_t * udata = (volatile uint64_t *)&sqe->user_data;
[6f121b8]	247
	248	if( *udata == expected &&
	249	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
	250	{
	251	// update statistics
[47746a2]	252	__STATS__( false,
	253	io.submit_q.alloc_avg.val += len;
	254	io.submit_q.alloc_avg.block += block;
	255	io.submit_q.alloc_avg.cnt += 1;
	256	)
[6f121b8]	257
[13c5e19]	258
[6f121b8]	259	// Success return the data
	260	return [sqe, idx];
	261	}
	262	verify(expected != data);
[2489d31]	263
[6f121b8]	264	len ++;
	265	}
[2489d31]	266
[6f121b8]	267	block++;
	268	yield();
	269	}
[2489d31]	270	}
	271
[df40a56]	272	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
	273	/* paranoid */ verify( idx <= mask );
	274	/* paranoid */ verify( idx != -1ul32 );
	275
	276	// We need to find a spot in the ready array
	277	__attribute((unused)) int len = 0;
	278	__attribute((unused)) int block = 0;
	279	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
	280
	281	disable_interrupts();
	282	uint32_t off = __tls_rand();
	283	enable_interrupts( __cfaabi_dbg_ctx );
	284
	285	uint32_t picked;
	286	LOOKING: for() {
	287	for(i; ring.submit_q.ready_cnt) {
	288	picked = (i + off) & ready_mask;
	289	uint32_t expected = -1ul32;
	290	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
	291	break LOOKING;
	292	}
	293	verify(expected != idx);
	294
	295	len ++;
	296	}
	297
	298	block++;
[34b61882]	299	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
	300	__release_consumed_submission( ring );
	301	unlock( ring.submit_q.lock );
	302	}
	303	else {
	304	yield();
	305	}
[df40a56]	306	}
	307
	308	// update statistics
[47746a2]	309	__STATS__( false,
	310	io.submit_q.look_avg.val += len;
	311	io.submit_q.look_avg.block += block;
	312	io.submit_q.look_avg.cnt += 1;
	313	)
[df40a56]	314
	315	return picked;
	316	}
	317
[f00b26d4]	318	void __submit( struct io_context * ctx, uint32_t idx ) __attribute__((nonnull (1))) {
	319	__io_data & ring = *ctx->thrd.ring;
[5dadc9b7]	320	// Get now the data we definetely need
[eafec07]	321	volatile uint32_t * const tail = ring.submit_q.tail;
[f00b26d4]	322	const uint32_t mask = *ring.submit_q.mask;
[2489d31]	323
[5dadc9b7]	324	// There are 2 submission schemes, check which one we are using
[f00b26d4]	325	if( ring.poller_submits ) {
[5dadc9b7]	326	// If the poller thread submits, then we just need to add this to the ready array
[df40a56]	327	__submit_to_ready_array( ring, idx, mask );
[5dadc9b7]	328
[f00b26d4]	329	post( ctx->thrd.sem );
[5dadc9b7]	330
[dd4e2d7]	331	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
[2d8f7b0]	332	}
[f00b26d4]	333	else if( ring.eager_submits ) {
[e46c753]	334	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
	335
	336	for() {
	337	yield();
	338
	339	// If some one else collected our index, we are done
[8bb239d]	340	#warning ABA problem
[e46c753]	341	if( ring.submit_q.ready[picked] != idx ) {
[47746a2]	342	__STATS__( false,
	343	io.submit_q.helped += 1;
	344	)
[e46c753]	345	return;
	346	}
	347
	348	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
[47746a2]	349	__STATS__( false,
	350	io.submit_q.leader += 1;
	351	)
[e46c753]	352	break;
	353	}
[8bb239d]	354
[47746a2]	355	__STATS__( false,
	356	io.submit_q.busy += 1;
	357	)
[e46c753]	358	}
	359
	360	// We got the lock
	361	unsigned to_submit = __collect_submitions( ring );
[f00b26d4]	362	int ret = __io_uring_enter( ring, to_submit, false );
[e46c753]	363	if( ret < 0 ) {
[20ab637]	364	unlock(ring.submit_q.lock);
	365	return;
[e46c753]	366	}
	367
[eafec07]	368	/* paranoid */ verify( ret > 0 \|\| to_submit == 0 \|\| (ring.ring_flags & IORING_SETUP_SQPOLL) );
[e46c753]	369
	370	// Release the consumed SQEs
[34b61882]	371	__release_consumed_submission( ring );
[e46c753]	372
	373	// update statistics
[47746a2]	374	__STATS__( true,
	375	io.submit_q.submit_avg.rdy += to_submit;
	376	io.submit_q.submit_avg.csm += ret;
	377	io.submit_q.submit_avg.cnt += 1;
	378	)
[e46c753]	379
	380	unlock(ring.submit_q.lock);
	381	}
[5dadc9b7]	382	else {
	383	// get mutual exclusion
	384	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
[2489d31]	385
[20ab637]	386	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 0,
	387	/* paranoid */ "index %u already reclaimed\n"
	388	/* paranoid */ "head %u, prev %u, tail %u\n"
	389	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
	390	/* paranoid */ idx,
	391	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
	392	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
	393	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
	394	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
	395	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
	396	/* paranoid */ );
	397
[5dadc9b7]	398	// Append to the list of ready entries
	399
	400	/* paranoid */ verify( idx <= mask );
[20ab637]	401	ring.submit_q.array[ (*tail) & mask ] = idx;
[5dadc9b7]	402	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
[d384787]	403
[5dadc9b7]	404	// Submit however, many entries need to be submitted
[f00b26d4]	405	int ret = __io_uring_enter( ring, 1, false );
[5dadc9b7]	406	if( ret < 0 ) {
	407	switch((int)errno) {
	408	default:
	409	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
	410	}
	411	}
[d384787]	412
[5dadc9b7]	413	// update statistics
[47746a2]	414	__STATS__( false,
	415	io.submit_q.submit_avg.csm += 1;
	416	io.submit_q.submit_avg.cnt += 1;
	417	)
[5dadc9b7]	418
[34b61882]	419	// Release the consumed SQEs
	420	__release_consumed_submission( ring );
[7bfc849]	421
[5dadc9b7]	422	unlock(ring.submit_q.lock);
[dd4e2d7]	423
	424	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
[5dadc9b7]	425	}
[2489d31]	426	}
[e46c753]	427
	428	static unsigned __collect_submitions( struct __io_data & ring ) {
	429	/* paranoid */ verify( ring.submit_q.ready != 0p );
	430	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
	431
	432	unsigned to_submit = 0;
	433	uint32_t tail = *ring.submit_q.tail;
	434	const uint32_t mask = *ring.submit_q.mask;
	435
	436	// Go through the list of ready submissions
	437	for( i; ring.submit_q.ready_cnt ) {
	438	// replace any submission with the sentinel, to consume it.
	439	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
	440
	441	// If it was already the sentinel, then we are done
	442	if( idx == -1ul32 ) continue;
	443
	444	// If we got a real submission, append it to the list
	445	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
	446	to_submit++;
	447	}
	448
	449	// Increment the tail based on how many we are ready to submit
	450	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
	451
	452	return to_submit;
	453	}
[34b61882]	454
	455	static uint32_t __release_consumed_submission( struct __io_data & ring ) {
	456	const uint32_t smask = *ring.submit_q.mask;
[732b406]	457
	458	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
[34b61882]	459	uint32_t chead = *ring.submit_q.head;
	460	uint32_t phead = ring.submit_q.prev_head;
	461	ring.submit_q.prev_head = chead;
[732b406]	462	unlock(ring.submit_q.release_lock);
	463
[34b61882]	464	uint32_t count = chead - phead;
	465	for( i; count ) {
	466	uint32_t idx = ring.submit_q.array[ (phead + i) & smask ];
	467	ring.submit_q.sqes[ idx ].user_data = 0;
	468	}
	469	return count;
	470	}
[47746a2]	471	#endif

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

Download in other formats:

Original Format