Context Navigation

io.cfa @ 35285fd

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

Last change on this file since 35285fd was 35285fd, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Refactor sqe reclamation to systematically thrash the sqe in debug
Property mode set to `100644`
File size: 21.5 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" {
	33	#include <sys/syscall.h>
	34
	35	#include <linux/io_uring.h>
	36	}
	37
[3e2b9c9]	38	#include "stats.hfa"
	39	#include "kernel.hfa"
	40	#include "kernel/fwd.hfa"
	41	#include "io/types.hfa"
[185efe6]	42
[426f60c]	43	static const char * opcodes[] = {
	44	"OP_NOP",
	45	"OP_READV",
	46	"OP_WRITEV",
	47	"OP_FSYNC",
	48	"OP_READ_FIXED",
	49	"OP_WRITE_FIXED",
	50	"OP_POLL_ADD",
	51	"OP_POLL_REMOVE",
	52	"OP_SYNC_FILE_RANGE",
	53	"OP_SENDMSG",
	54	"OP_RECVMSG",
	55	"OP_TIMEOUT",
	56	"OP_TIMEOUT_REMOVE",
	57	"OP_ACCEPT",
	58	"OP_ASYNC_CANCEL",
	59	"OP_LINK_TIMEOUT",
	60	"OP_CONNECT",
	61	"OP_FALLOCATE",
	62	"OP_OPENAT",
	63	"OP_CLOSE",
	64	"OP_FILES_UPDATE",
	65	"OP_STATX",
	66	"OP_READ",
	67	"OP_WRITE",
	68	"OP_FADVISE",
	69	"OP_MADVISE",
	70	"OP_SEND",
	71	"OP_RECV",
	72	"OP_OPENAT2",
	73	"OP_EPOLL_CTL",
	74	"OP_SPLICE",
	75	"OP_PROVIDE_BUFFERS",
	76	"OP_REMOVE_BUFFERS",
	77	"OP_TEE",
	78	"INVALID_OP"
	79	};
	80
[2fafe7e]	81	// returns true of acquired as leader or second leader
	82	static inline bool try_lock( __leaderlock_t & this ) {
	83	const uintptr_t thrd = 1z \| (uintptr_t)active_thread();
	84	bool block;
	85	disable_interrupts();
	86	for() {
	87	struct $thread * expected = this.value;
	88	if( 1p != expected && 0p != expected ) {
	89	/* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader
	90	enable_interrupts( __cfaabi_dbg_ctx );
	91	return false;
	92	}
	93	struct $thread * desired;
	94	if( 0p == expected ) {
	95	// If the lock isn't locked acquire it, no need to block
	96	desired = 1p;
	97	block = false;
	98	}
	99	else {
	100	// If the lock is already locked try becomming the next leader
	101	desired = (struct $thread *)thrd;
	102	block = true;
	103	}
	104	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
	105	}
	106	if( block ) {
	107	enable_interrupts( __cfaabi_dbg_ctx );
[e235429]	108	park();
[2fafe7e]	109	disable_interrupts();
	110	}
	111	return true;
	112	}
	113
	114	static inline bool next( __leaderlock_t & this ) {
[8fc652e0]	115	/* paranoid */ verify( ! __preemption_enabled() );
[2fafe7e]	116	struct $thread * nextt;
	117	for() {
	118	struct $thread * expected = this.value;
	119	/* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked
	120
	121	struct $thread * desired;
	122	if( 1p == expected ) {
	123	// No next leader, just unlock
	124	desired = 0p;
	125	nextt = 0p;
	126	}
	127	else {
	128	// There is a next leader, remove but keep locked
	129	desired = 1p;
	130	nextt = (struct $thread *)(~1z & (uintptr_t)expected);
	131	}
	132	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
	133	}
	134
	135	if(nextt) {
[b4b63e8]	136	unpark( nextt );
[2fafe7e]	137	enable_interrupts( __cfaabi_dbg_ctx );
	138	return true;
	139	}
	140	enable_interrupts( __cfaabi_dbg_ctx );
	141	return false;
	142	}
	143
[92976d9]	144	//=============================================================================================
[3e2b9c9]	145	// I/O Syscall
[92976d9]	146	//=============================================================================================
[3e2b9c9]	147	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
[20ab637]	148	bool need_sys_to_submit = false;
	149	bool need_sys_to_complete = false;
	150	unsigned flags = 0;
	151
	152	TO_SUBMIT:
	153	if( to_submit > 0 ) {
	154	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
	155	need_sys_to_submit = true;
	156	break TO_SUBMIT;
	157	}
	158	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
	159	need_sys_to_submit = true;
	160	flags \|= IORING_ENTER_SQ_WAKEUP;
	161	}
	162	}
	163
	164	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
	165	flags \|= IORING_ENTER_GETEVENTS;
	166	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
	167	need_sys_to_complete = true;
	168	}
	169	}
	170
	171	int ret = 0;
	172	if( need_sys_to_submit \|\| need_sys_to_complete ) {
[ece0e80]	173	__cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags);
[b982fb2]	174	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8);
[20ab637]	175	if( ret < 0 ) {
	176	switch((int)errno) {
	177	case EAGAIN:
	178	case EINTR:
	179	ret = -1;
	180	break;
	181	default:
	182	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	183	}
	184	}
	185	}
	186
	187	// Memory barrier
	188	__atomic_thread_fence( __ATOMIC_SEQ_CST );
	189	return ret;
	190	}
	191
[92976d9]	192	//=============================================================================================
	193	// I/O Polling
	194	//=============================================================================================
[1d5e4711]	195	static unsigned __collect_submitions( struct __io_data & ring );
[4998155]	196	static __u32 __release_consumed_submission( struct __io_data & ring );
[35285fd]	197	static inline void __clean( volatile struct io_uring_sqe * sqe );
[1d5e4711]	198
[426f60c]	199	// Process a single completion message from the io_uring
	200	// This is NOT thread-safe
	201	static inline void process( volatile struct io_uring_cqe & cqe ) {
[c402739f]	202	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
[fe9468e2]	203	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
[5751a56]	204
[c402739f]	205	fulfil( *future, cqe.res );
[5751a56]	206	}
	207
[f00b26d4]	208	static [int, bool] __drain_io( & struct __io_data ring ) {
[8fc652e0]	209	/* paranoid */ verify( ! __preemption_enabled() );
[e46c753]	210
[5dadc9b]	211	unsigned to_submit = 0;
[f00b26d4]	212	if( ring.poller_submits ) {
[5dadc9b]	213	// If the poller thread also submits, then we need to aggregate the submissions which are ready
[e46c753]	214	to_submit = __collect_submitions( ring );
[5dadc9b]	215	}
	216
[f00b26d4]	217	int ret = __io_uring_enter(ring, to_submit, true);
[20ab637]	218	if( ret < 0 ) {
	219	return [0, true];
	220	}
[1d5e4711]	221
[20ab637]	222	// update statistics
	223	if (to_submit > 0) {
[1d5e4711]	224	__STATS__( true,
	225	if( to_submit > 0 ) {
	226	io.submit_q.submit_avg.rdy += to_submit;
	227	io.submit_q.submit_avg.csm += ret;
	228	io.submit_q.submit_avg.cnt += 1;
	229	}
	230	)
[6f121b8]	231	}
	232
[426f60c]	233	__atomic_thread_fence( __ATOMIC_SEQ_CST );
	234
[20ab637]	235	// Release the consumed SQEs
	236	__release_consumed_submission( ring );
[6f121b8]	237
[d384787]	238	// Drain the queue
[92976d9]	239	unsigned head = *ring.completion_q.head;
[6f121b8]	240	unsigned tail = *ring.completion_q.tail;
[4998155]	241	const __u32 mask = *ring.completion_q.mask;
[6f121b8]	242
[d384787]	243	// Nothing was new return 0
	244	if (head == tail) {
[e46c753]	245	return [0, to_submit > 0];
[d384787]	246	}
[92976d9]	247
[4998155]	248	__u32 count = tail - head;
[1d5e4711]	249	/* paranoid */ verify( count != 0 );
[d384787]	250	for(i; count) {
[6f121b8]	251	unsigned idx = (head + i) & mask;
[426f60c]	252	volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
[92976d9]	253
[d384787]	254	/* paranoid */ verify(&cqe);
[92976d9]	255
[f00b26d4]	256	process( cqe );
[d384787]	257	}
[2d8f7b0]	258
[92976d9]	259	// Mark to the kernel that the cqe has been seen
	260	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
[426f60c]	261	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_SEQ_CST );
[92976d9]	262
[5dadc9b]	263	return [count, count > 0 \|\| to_submit > 0];
[92976d9]	264	}
	265
[f00b26d4]	266	void main( $io_ctx_thread & this ) {
[d48b174]	267	__ioctx_register( this );
[1539bbd]	268
[426f60c]	269	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %d (%p) ready\n", this.ring->fd, &this);
[1539bbd]	270
[ece0e80]	271	const int reset_cnt = 5;
	272	int reset = reset_cnt;
[61dd73d]	273	// Then loop until we need to start
[ece0e80]	274	LOOP:
[f00b26d4]	275	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
[61dd73d]	276	// Drain the io
[5dadc9b]	277	int count;
	278	bool again;
[13c5e19]	279	disable_interrupts();
[f00b26d4]	280	[count, again] = __drain_io( *this.ring );
[5dadc9b]	281
[ece0e80]	282	if(!again) reset--;
[3c039b0]	283
[13c5e19]	284	// Update statistics
[47746a2]	285	__STATS__( true,
	286	io.complete_q.completed_avg.val += count;
[dcb5f8d]	287	io.complete_q.completed_avg.cnt += 1;
[47746a2]	288	)
[13c5e19]	289	enable_interrupts( __cfaabi_dbg_ctx );
[3c039b0]	290
[5dadc9b]	291	// If we got something, just yield and check again
[ece0e80]	292	if(reset > 1) {
[61dd73d]	293	yield();
[ece0e80]	294	continue LOOP;
[61dd73d]	295	}
[ece0e80]	296
[426f60c]	297	// We alread failed to find completed entries a few time.
[ece0e80]	298	if(reset == 1) {
	299	// Rearm the context so it can block
	300	// but don't block right away
	301	// we need to retry one last time in case
	302	// something completed just now
[d48b174]	303	__ioctx_prepare_block( this );
[ece0e80]	304	continue LOOP;
	305	}
	306
[93526ef]	307	__STATS__( false,
[dcb5f8d]	308	io.complete_q.blocks += 1;
	309	)
[426f60c]	310	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %d (%p)\n", this.ring->fd, &this);
[5dadc9b]	311
[3e2b9c9]	312	// block this thread
[f00b26d4]	313	wait( this.sem );
[ece0e80]	314
	315	// restore counter
	316	reset = reset_cnt;
[f6660520]	317	}
[61dd73d]	318
[426f60c]	319	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller %d (%p) stopping\n", this.ring->fd, &this);
[61dd73d]	320	}
[f6660520]	321
[92976d9]	322	//=============================================================================================
	323	// I/O Submissions
	324	//=============================================================================================
	325
[2d8f7b0]	326	// Submition steps :
[e46c753]	327	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
[2d8f7b0]	328	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
	329	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
	330	// need to write an allocator that allows allocating concurrently.
	331	//
[e46c753]	332	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
[2d8f7b0]	333	//
[e46c753]	334	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
[2d8f7b0]	335	// needs to arrive to two concensus at the same time:
	336	// A - The order in which entries are listed in the array: no two threads must pick the
	337	// same index for their entries
	338	// B - When can the tail be update for the kernel. EVERY entries in the array between
	339	// head and tail must be fully filled and shouldn't ever be touched again.
	340	//
	341
[426f60c]	342	// Allocate an submit queue entry.
	343	// The kernel cannot see these entries until they are submitted, but other threads must be
	344	// able to see which entries can be used and which are already un used by an other thread
	345	// for convenience, return both the index and the pointer to the sqe
	346	// sqe == &sqes[idx]
	347	[* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
[e46c753]	348	/* paranoid */ verify( data != 0 );
[13c5e19]	349
[6f121b8]	350	// Prepare the data we need
	351	__attribute((unused)) int len = 0;
	352	__attribute((unused)) int block = 0;
[4998155]	353	__u32 cnt = *ring.submit_q.num;
	354	__u32 mask = *ring.submit_q.mask;
[8ae4165]	355
[fe9468e2]	356	__u32 off = thread_rand();
[6f121b8]	357
	358	// Loop around looking for an available spot
[13c5e19]	359	for() {
[6f121b8]	360	// Look through the list starting at some offset
	361	for(i; cnt) {
[426f60c]	362	__u64 expected = 3;
	363	__u32 idx = (i + off) & mask; // Get an index from a random
	364	volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
[4998155]	365	volatile __u64 * udata = &sqe->user_data;
[6f121b8]	366
[426f60c]	367	// Allocate the entry by CASing the user_data field from 0 to the future address
[6f121b8]	368	if( *udata == expected &&
	369	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
	370	{
	371	// update statistics
[47746a2]	372	__STATS__( false,
	373	io.submit_q.alloc_avg.val += len;
	374	io.submit_q.alloc_avg.block += block;
	375	io.submit_q.alloc_avg.cnt += 1;
	376	)
[6f121b8]	377
[426f60c]	378	// debug log
[ece0e80]	379	__cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
[13c5e19]	380
[6f121b8]	381	// Success return the data
[426f60c]	382	sqe->opcode = 0;
	383	sqe->flags = 0;
	384	sqe->ioprio = 0;
	385	sqe->fd = 0;
	386	sqe->off = 0;
	387	sqe->addr = 0;
	388	sqe->len = 0;
	389	sqe->accept_flags = 0;
	390	sqe->__pad2[0] = 0;
	391	sqe->__pad2[1] = 0;
	392	sqe->__pad2[2] = 0;
[6f121b8]	393	return [sqe, idx];
	394	}
	395	verify(expected != data);
[2489d31]	396
[426f60c]	397	// This one was used
[6f121b8]	398	len ++;
	399	}
[2489d31]	400
[6f121b8]	401	block++;
[426f60c]	402
	403	abort( "Kernel I/O : all submit queue entries used, yielding\n" );
	404
[6f121b8]	405	yield();
	406	}
[2489d31]	407	}
	408
[4998155]	409	static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
[df40a56]	410	/* paranoid */ verify( idx <= mask );
	411	/* paranoid */ verify( idx != -1ul32 );
	412
	413	// We need to find a spot in the ready array
	414	__attribute((unused)) int len = 0;
	415	__attribute((unused)) int block = 0;
[4998155]	416	__u32 ready_mask = ring.submit_q.ready_cnt - 1;
[df40a56]	417
[fe9468e2]	418	__u32 off = thread_rand();
[df40a56]	419
[4998155]	420	__u32 picked;
[df40a56]	421	LOOKING: for() {
	422	for(i; ring.submit_q.ready_cnt) {
	423	picked = (i + off) & ready_mask;
[4998155]	424	__u32 expected = -1ul32;
[df40a56]	425	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
	426	break LOOKING;
	427	}
	428	verify(expected != idx);
	429
	430	len ++;
	431	}
	432
	433	block++;
[2fafe7e]	434
	435	__u32 released = __release_consumed_submission( ring );
	436	if( released == 0 ) {
[34b61882]	437	yield();
	438	}
[df40a56]	439	}
	440
	441	// update statistics
[47746a2]	442	__STATS__( false,
	443	io.submit_q.look_avg.val += len;
	444	io.submit_q.look_avg.block += block;
	445	io.submit_q.look_avg.cnt += 1;
	446	)
[df40a56]	447
	448	return picked;
	449	}
	450
[4998155]	451	void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
[f00b26d4]	452	__io_data & ring = *ctx->thrd.ring;
[426f60c]	453
	454	{
	455	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
	456	__cfadbg_print_safe( io,
	457	"Kernel I/O : submitting %u (%p) for %p\n"
	458	" data: %p\n"
	459	" opcode: %s\n"
	460	" fd: %d\n"
	461	" flags: %d\n"
	462	" prio: %d\n"
	463	" off: %p\n"
	464	" addr: %p\n"
	465	" len: %d\n"
	466	" other flags: %d\n"
	467	" splice fd: %d\n"
	468	" pad[0]: %llu\n"
	469	" pad[1]: %llu\n"
	470	" pad[2]: %llu\n",
	471	idx, sqe,
	472	active_thread(),
	473	(void*)sqe->user_data,
	474	opcodes[sqe->opcode],
	475	sqe->fd,
	476	sqe->flags,
	477	sqe->ioprio,
	478	sqe->off,
	479	sqe->addr,
	480	sqe->len,
	481	sqe->accept_flags,
	482	sqe->splice_fd_in,
	483	sqe->__pad2[0],
	484	sqe->__pad2[1],
	485	sqe->__pad2[2]
	486	);
	487	}
	488
	489
[5dadc9b]	490	// Get now the data we definetely need
[4998155]	491	volatile __u32 * const tail = ring.submit_q.tail;
	492	const __u32 mask = *ring.submit_q.mask;
[2489d31]	493
[5dadc9b]	494	// There are 2 submission schemes, check which one we are using
[f00b26d4]	495	if( ring.poller_submits ) {
[5dadc9b]	496	// If the poller thread submits, then we just need to add this to the ready array
[df40a56]	497	__submit_to_ready_array( ring, idx, mask );
[5dadc9b]	498
[f00b26d4]	499	post( ctx->thrd.sem );
[5dadc9b]	500
[dd4e2d7]	501	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
[2d8f7b0]	502	}
[f00b26d4]	503	else if( ring.eager_submits ) {
[4998155]	504	__u32 picked = __submit_to_ready_array( ring, idx, mask );
[e46c753]	505
[2fafe7e]	506	#if defined(LEADER_LOCK)
	507	if( !try_lock(ring.submit_q.submit_lock) ) {
[47746a2]	508	__STATS__( false,
	509	io.submit_q.helped += 1;
	510	)
[e46c753]	511	return;
	512	}
[8fc652e0]	513	/* paranoid */ verify( ! __preemption_enabled() );
[2fafe7e]	514	__STATS__( true,
	515	io.submit_q.leader += 1;
	516	)
	517	#else
	518	for() {
	519	yield();
	520
	521	if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
	522	__STATS__( false,
	523	io.submit_q.leader += 1;
	524	)
	525	break;
	526	}
	527
	528	// If some one else collected our index, we are done
	529	#warning ABA problem
	530	if( ring.submit_q.ready[picked] != idx ) {
	531	__STATS__( false,
	532	io.submit_q.helped += 1;
	533	)
	534	return;
	535	}
[e46c753]	536
[47746a2]	537	__STATS__( false,
[2fafe7e]	538	io.submit_q.busy += 1;
[47746a2]	539	)
[e46c753]	540	}
[2fafe7e]	541	#endif
[e46c753]	542
	543	// We got the lock
[be36ec3]	544	// Collect the submissions
[e46c753]	545	unsigned to_submit = __collect_submitions( ring );
	546
[be36ec3]	547	// Actually submit
	548	int ret = __io_uring_enter( ring, to_submit, false );
[309d814]	549
[2fafe7e]	550	#if defined(LEADER_LOCK)
[8fc652e0]	551	/* paranoid */ verify( ! __preemption_enabled() );
[2fafe7e]	552	next(ring.submit_q.submit_lock);
	553	#else
	554	unlock(ring.submit_q.submit_lock);
	555	#endif
[ece0e80]	556	if( ret < 0 ) {
	557	return;
	558	}
[e46c753]	559
	560	// Release the consumed SQEs
[34b61882]	561	__release_consumed_submission( ring );
[e46c753]	562
	563	// update statistics
[be36ec3]	564	__STATS__( false,
[47746a2]	565	io.submit_q.submit_avg.rdy += to_submit;
	566	io.submit_q.submit_avg.csm += ret;
	567	io.submit_q.submit_avg.cnt += 1;
	568	)
[ece0e80]	569
	570	__cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
[e46c753]	571	}
[426f60c]	572	else
	573	{
[5dadc9b]	574	// get mutual exclusion
[2fafe7e]	575	#if defined(LEADER_LOCK)
	576	while(!try_lock(ring.submit_q.submit_lock));
	577	#else
	578	lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
	579	#endif
[2489d31]	580
[426f60c]	581	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64,
[20ab637]	582	/* paranoid */ "index %u already reclaimed\n"
	583	/* paranoid */ "head %u, prev %u, tail %u\n"
	584	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
	585	/* paranoid */ idx,
	586	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
	587	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
	588	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
	589	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
	590	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
	591	/* paranoid */ );
	592
[5dadc9b]	593	// Append to the list of ready entries
	594
	595	/* paranoid */ verify( idx <= mask );
[20ab637]	596	ring.submit_q.array[ (*tail) & mask ] = idx;
[5dadc9b]	597	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
[d384787]	598
[5dadc9b]	599	// Submit however, many entries need to be submitted
[f00b26d4]	600	int ret = __io_uring_enter( ring, 1, false );
[5dadc9b]	601	if( ret < 0 ) {
	602	switch((int)errno) {
	603	default:
	604	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
	605	}
	606	}
[d384787]	607
[426f60c]	608	/* paranoid */ verify(ret == 1);
	609
[5dadc9b]	610	// update statistics
[47746a2]	611	__STATS__( false,
	612	io.submit_q.submit_avg.csm += 1;
	613	io.submit_q.submit_avg.cnt += 1;
	614	)
[5dadc9b]	615
[426f60c]	616	{
	617	__attribute__((unused)) volatile __u32 * const head = ring.submit_q.head;
	618	__attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ];
	619	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx];
	620
	621	__cfadbg_print_safe( io,
	622	"Kernel I/O : last submitted is %u (%p)\n"
	623	" data: %p\n"
	624	" opcode: %s\n"
	625	" fd: %d\n"
	626	" flags: %d\n"
	627	" prio: %d\n"
	628	" off: %p\n"
	629	" addr: %p\n"
	630	" len: %d\n"
	631	" other flags: %d\n"
	632	" splice fd: %d\n"
	633	" pad[0]: %llu\n"
	634	" pad[1]: %llu\n"
	635	" pad[2]: %llu\n",
	636	last_idx, sqe,
	637	(void*)sqe->user_data,
	638	opcodes[sqe->opcode],
	639	sqe->fd,
	640	sqe->flags,
	641	sqe->ioprio,
	642	sqe->off,
	643	sqe->addr,
	644	sqe->len,
	645	sqe->accept_flags,
	646	sqe->splice_fd_in,
	647	sqe->__pad2[0],
	648	sqe->__pad2[1],
	649	sqe->__pad2[2]
	650	);
	651	}
	652
	653	__atomic_thread_fence( __ATOMIC_SEQ_CST );
[34b61882]	654	// Release the consumed SQEs
	655	__release_consumed_submission( ring );
[426f60c]	656	// ring.submit_q.sqes[idx].user_data = 3ul64;
[7bfc849]	657
[2fafe7e]	658	#if defined(LEADER_LOCK)
	659	next(ring.submit_q.submit_lock);
	660	#else
	661	unlock(ring.submit_q.submit_lock);
	662	#endif
[dd4e2d7]	663
[426f60c]	664	__cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() );
[5dadc9b]	665	}
[2489d31]	666	}
[e46c753]	667
[d2b5d2d]	668	// #define PARTIAL_SUBMIT 32
[426f60c]	669
	670	// go through the list of submissions in the ready array and moved them into
	671	// the ring's submit queue
[e46c753]	672	static unsigned __collect_submitions( struct __io_data & ring ) {
	673	/* paranoid */ verify( ring.submit_q.ready != 0p );
	674	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
	675
	676	unsigned to_submit = 0;
[4998155]	677	__u32 tail = *ring.submit_q.tail;
	678	const __u32 mask = *ring.submit_q.mask;
[1095ccd]	679	#if defined(PARTIAL_SUBMIT)
	680	#if defined(LEADER_LOCK)
	681	#error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
	682	#endif
	683	const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
	684	const __u32 offset = ring.submit_q.prev_ready;
	685	ring.submit_q.prev_ready += cnt;
	686	#else
	687	const __u32 cnt = ring.submit_q.ready_cnt;
	688	const __u32 offset = 0;
	689	#endif
[e46c753]	690
	691	// Go through the list of ready submissions
[1095ccd]	692	for( c; cnt ) {
	693	__u32 i = (offset + c) % ring.submit_q.ready_cnt;
	694
[e46c753]	695	// replace any submission with the sentinel, to consume it.
[4998155]	696	__u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
[e46c753]	697
	698	// If it was already the sentinel, then we are done
	699	if( idx == -1ul32 ) continue;
	700
	701	// If we got a real submission, append it to the list
	702	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
	703	to_submit++;
	704	}
	705
	706	// Increment the tail based on how many we are ready to submit
	707	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
	708
	709	return to_submit;
	710	}
[34b61882]	711
[426f60c]	712	// Go through the ring's submit queue and release everything that has already been consumed
	713	// by io_uring
[4998155]	714	static __u32 __release_consumed_submission( struct __io_data & ring ) {
	715	const __u32 smask = *ring.submit_q.mask;
[732b406]	716
[426f60c]	717	// We need to get the lock to copy the old head and new head
[732b406]	718	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
[426f60c]	719	__attribute__((unused))
	720	__u32 ctail = *ring.submit_q.tail; // get the current tail of the queue
	721	__u32 chead = *ring.submit_q.head; // get the current head of the queue
	722	__u32 phead = ring.submit_q.prev_head; // get the head the last time we were here
	723	ring.submit_q.prev_head = chead; // note up to were we processed
[732b406]	724	unlock(ring.submit_q.release_lock);
	725
[426f60c]	726	// the 3 fields are organized like this diagram
	727	// except it's are ring
	728	// ---+--------+--------+----
	729	// ---+--------+--------+----
	730	// ^ ^ ^
	731	// phead chead ctail
	732
	733	// make sure ctail doesn't wrap around and reach phead
	734	/* paranoid */ verify(
	735	(ctail >= chead && chead >= phead)
	736	\|\| (chead >= phead && phead >= ctail)
	737	\|\| (phead >= ctail && ctail >= chead)
	738	);
	739
	740	// find the range we need to clear
[4998155]	741	__u32 count = chead - phead;
[426f60c]	742
	743	// We acquired an previous-head/current-head range
	744	// go through the range and release the sqes
[34b61882]	745	for( i; count ) {
[4998155]	746	__u32 idx = ring.submit_q.array[ (phead + i) & smask ];
[426f60c]	747
	748	/* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data );
[35285fd]	749	__clean( &ring.submit_q.sqes[ idx ] );
[34b61882]	750	}
	751	return count;
	752	}
[35285fd]	753
	754	void __sqe_clean( volatile struct io_uring_sqe * sqe ) {
	755	__clean( sqe );
	756	}
	757
	758	static inline void __clean( volatile struct io_uring_sqe * sqe ) {
	759	// If we are in debug mode, thrash the fields to make sure we catch reclamation errors
	760	__cfaabi_dbg_debug_do(
	761	memset(sqe, 0xde, sizeof(*sqe));
	762	sqe->opcode = IORING_OP_LAST;
	763	);
	764
	765	// Mark the entry as unused
	766	__atomic_store_n(&sqe->user_data, 3ul64, __ATOMIC_SEQ_CST);
	767	}
[47746a2]	768	#endif

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

Download in other formats:

Original Format