Context Navigation

io.cfa@ 0521a1a

Visit:

ADT ast-experimental enum pthread-emulation qualifiedEnum

Last change on this file since 0521a1a was 78a580d, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
I/O now updates the timestamps when draining. Timestamps are not used yet.
Property mode set to `100644`
File size: 17.6 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__
[43784ac]	17	#define _GNU_SOURCE
[3e2b9c9]	18
[20ab637]	19	#if defined(__CFA_DEBUG__)
[d60d30e]	20	// #define __CFA_DEBUG_PRINT_IO__
	21	// #define __CFA_DEBUG_PRINT_IO_CORE__
[20ab637]	22	#endif
[4069faad]	23
[f6660520]	24
[3e2b9c9]	25	#if defined(CFA_HAVE_LINUX_IO_URING_H)
[31bb2e1]	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>
[dddb3dd0]	34	#include <sys/eventfd.h>
[d3605f8]	35	#include <sys/uio.h>
[92976d9]	36
	37	#include <linux/io_uring.h>
	38	}
	39
[3e2b9c9]	40	#include "stats.hfa"
	41	#include "kernel.hfa"
	42	#include "kernel/fwd.hfa"
[708ae38]	43	#include "kernel/private.hfa"
[78a580d]	44	#include "kernel/cluster.hfa"
[3e2b9c9]	45	#include "io/types.hfa"
[185efe6]	46
[2fab24e3]	47	__attribute__((unused)) static const char * opcodes[] = {
[426f60c]	48	"OP_NOP",
	49	"OP_READV",
	50	"OP_WRITEV",
	51	"OP_FSYNC",
	52	"OP_READ_FIXED",
	53	"OP_WRITE_FIXED",
	54	"OP_POLL_ADD",
	55	"OP_POLL_REMOVE",
	56	"OP_SYNC_FILE_RANGE",
	57	"OP_SENDMSG",
	58	"OP_RECVMSG",
	59	"OP_TIMEOUT",
	60	"OP_TIMEOUT_REMOVE",
	61	"OP_ACCEPT",
	62	"OP_ASYNC_CANCEL",
	63	"OP_LINK_TIMEOUT",
	64	"OP_CONNECT",
	65	"OP_FALLOCATE",
	66	"OP_OPENAT",
	67	"OP_CLOSE",
	68	"OP_FILES_UPDATE",
	69	"OP_STATX",
	70	"OP_READ",
	71	"OP_WRITE",
	72	"OP_FADVISE",
	73	"OP_MADVISE",
	74	"OP_SEND",
	75	"OP_RECV",
	76	"OP_OPENAT2",
	77	"OP_EPOLL_CTL",
	78	"OP_SPLICE",
	79	"OP_PROVIDE_BUFFERS",
	80	"OP_REMOVE_BUFFERS",
	81	"OP_TEE",
	82	"INVALID_OP"
	83	};
	84
[11054eb]	85	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want );
	86	static void __ioarbiter_submit( $io_context * , __u32 idxs[], __u32 have, bool lazy );
	87	static void __ioarbiter_flush ( $io_context & );
[dddb3dd0]	88	static inline void __ioarbiter_notify( $io_context & ctx );
[92976d9]	89	//=============================================================================================
	90	// I/O Polling
	91	//=============================================================================================
[78da4ab]	92	static inline unsigned __flush( struct $io_context & );
	93	static inline __u32 __release_sqes( struct $io_context & );
[24e321c]	94	extern void __kernel_unpark( thread$ * thrd, unpark_hint );
[1d5e4711]	95
[78a580d]	96	bool __cfa_io_drain( $io_context * ctx, cluster * cltr ) {
[dddb3dd0]	97	/* paranoid */ verify( ! __preemption_enabled() );
[e9c0b4c]	98	/* paranoid */ verify( ready_schedule_islocked() );
[4ecc35a]	99	/* paranoid */ verify( ctx );
[6f121b8]	100
[d384787]	101	// Drain the queue
[dddb3dd0]	102	unsigned head = *ctx->cq.head;
	103	unsigned tail = *ctx->cq.tail;
	104	const __u32 mask = *ctx->cq.mask;
[92976d9]	105
[4998155]	106	__u32 count = tail - head;
[dddb3dd0]	107	__STATS__( false, io.calls.drain++; io.calls.completed += count; )
[d60d30e]	108
[c1c95b1]	109	if(count == 0) return false;
	110
[4ecc35a]	111	if(!__atomic_try_acquire(&ctx->cq.lock)) {
	112	return false;
	113	}
	114
[78a580d]	115	unsigned long long ts_prev = ctx->cq.ts;
	116
[d384787]	117	for(i; count) {
[6f121b8]	118	unsigned idx = (head + i) & mask;
[dddb3dd0]	119	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
[92976d9]	120
[d384787]	121	/* paranoid */ verify(&cqe);
[92976d9]	122
[78da4ab]	123	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
	124	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
	125
[24e321c]	126	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
[78da4ab]	127	}
	128
[dddb3dd0]	129	__cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
[78a580d]	130	unsigned long long ts_next = ctx->cq.ts = rdtscl();
[2d8f7b0]	131
[92976d9]	132	// Mark to the kernel that the cqe has been seen
	133	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
[dddb3dd0]	134	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
[92976d9]	135
[e9c0b4c]	136	/* paranoid */ verify( ready_schedule_islocked() );
[dddb3dd0]	137	/* paranoid */ verify( ! __preemption_enabled() );
	138
[4ecc35a]	139	__atomic_unlock(&ctx->cq.lock);
	140
[78a580d]	141	touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next );
	142
[c1c95b1]	143	return true;
[92976d9]	144	}
	145
[c7b2215]	146	bool __cfa_io_flush( processor * proc, int min_comp ) {
[dddb3dd0]	147	/* paranoid */ verify( ! __preemption_enabled() );
	148	/* paranoid */ verify( proc );
	149	/* paranoid */ verify( proc->io.ctx );
[1539bbd]	150
[78a580d]	151	cluster * cltr = proc->cltr;
[dddb3dd0]	152	$io_context & ctx = *proc->io.ctx;
[78da4ab]	153
[11054eb]	154	__ioarbiter_flush( ctx );
[3c039b0]	155
[21a5bfb7]	156	if(ctx.sq.to_submit != 0 \|\| min_comp > 0) {
	157
	158	__STATS__( true, io.calls.flush++; )
	159	int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, min_comp, min_comp > 0 ? IORING_ENTER_GETEVENTS : 0, (sigset_t *)0p, _NSIG / 8);
	160	if( ret < 0 ) {
	161	switch((int)errno) {
	162	case EAGAIN:
	163	case EINTR:
	164	case EBUSY:
	165	// Update statistics
	166	__STATS__( false, io.calls.errors.busy ++; )
	167	return false;
	168	default:
	169	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	170	}
[61dd73d]	171	}
[ece0e80]	172
[21a5bfb7]	173	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
	174	__STATS__( true, io.calls.submitted += ret; )
	175	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
	176	/* paranoid */ verify( ctx.sq.to_submit >= ret );
[dddb3dd0]	177
[21a5bfb7]	178	ctx.sq.to_submit -= ret;
[ece0e80]	179
[21a5bfb7]	180	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
[5dadc9b7]	181
[21a5bfb7]	182	// Release the consumed SQEs
	183	__release_sqes( ctx );
[ece0e80]	184
[21a5bfb7]	185	/* paranoid */ verify( ! __preemption_enabled() );
	186
[d529ad0]	187	__atomic_store_n(&ctx.proc->io.pending, false, __ATOMIC_RELAXED);
[21a5bfb7]	188	}
[61dd73d]	189
[7ef162b2]	190	ready_schedule_lock();
[78a580d]	191	bool ret = __cfa_io_drain( &ctx, cltr );
[7ef162b2]	192	ready_schedule_unlock();
	193	return ret;
[61dd73d]	194	}
[f6660520]	195
[92976d9]	196	//=============================================================================================
	197	// I/O Submissions
	198	//=============================================================================================
	199
[2d8f7b0]	200	// Submition steps :
[e46c753]	201	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
[2d8f7b0]	202	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
	203	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
	204	// need to write an allocator that allows allocating concurrently.
	205	//
[e46c753]	206	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
[2d8f7b0]	207	//
[e46c753]	208	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
[2d8f7b0]	209	// needs to arrive to two concensus at the same time:
	210	// A - The order in which entries are listed in the array: no two threads must pick the
	211	// same index for their entries
	212	// B - When can the tail be update for the kernel. EVERY entries in the array between
	213	// head and tail must be fully filled and shouldn't ever be touched again.
	214	//
[78da4ab]	215	//=============================================================================================
	216	// Allocation
	217	// for user's convenience fill the sqes from the indexes
	218	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
	219	struct io_uring_sqe * sqes = ctx->sq.sqes;
	220	for(i; want) {
[dddb3dd0]	221	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
[78da4ab]	222	out_sqes[i] = &sqes[idxs[i]];
	223	}
	224	}
[2489d31]	225
[78da4ab]	226	// Try to directly allocate from the a given context
	227	// Not thread-safe
	228	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
	229	__sub_ring_t & sq = ctx->sq;
	230	const __u32 mask = *sq.mask;
	231	__u32 fhead = sq.free_ring.head; // get the current head of the queue
	232	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
[2489d31]	233
[78da4ab]	234	// If we don't have enough sqes, fail
	235	if((ftail - fhead) < want) { return false; }
[426f60c]	236
[78da4ab]	237	// copy all the indexes we want from the available list
	238	for(i; want) {
[dddb3dd0]	239	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
[78da4ab]	240	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
[6f121b8]	241	}
[2489d31]	242
[78da4ab]	243	// Advance the head to mark the indexes as consumed
	244	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
[df40a56]	245
[78da4ab]	246	// return success
	247	return true;
	248	}
[df40a56]	249
[78da4ab]	250	// Allocate an submit queue entry.
	251	// The kernel cannot see these entries until they are submitted, but other threads must be
	252	// able to see which entries can be used and which are already un used by an other thread
	253	// for convenience, return both the index and the pointer to the sqe
	254	// sqe == &sqes[idx]
	255	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
	256	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
[df40a56]	257
[78da4ab]	258	disable_interrupts();
	259	processor * proc = __cfaabi_tls.this_processor;
[dddb3dd0]	260	$io_context * ctx = proc->io.ctx;
[78da4ab]	261	/* paranoid */ verify( __cfaabi_tls.this_processor );
[dddb3dd0]	262	/* paranoid */ verify( ctx );
[78da4ab]	263
[dddb3dd0]	264	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
[78da4ab]	265
[dddb3dd0]	266	// We can proceed to the fast path
	267	if( __alloc(ctx, idxs, want) ) {
	268	// Allocation was successful
	269	__STATS__( true, io.alloc.fast += 1; )
[a3821fa]	270	enable_interrupts();
[df40a56]	271
[dddb3dd0]	272	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
[2fafe7e]	273
[dddb3dd0]	274	__fill( sqes, want, idxs, ctx );
	275	return ctx;
[df40a56]	276	}
[dddb3dd0]	277	// The fast path failed, fallback
	278	__STATS__( true, io.alloc.fail += 1; )
[df40a56]	279
[78da4ab]	280	// Fast path failed, fallback on arbitration
[d60d30e]	281	__STATS__( true, io.alloc.slow += 1; )
[a3821fa]	282	enable_interrupts();
[78da4ab]	283
[dddb3dd0]	284	$io_arbiter * ioarb = proc->cltr->io.arbiter;
	285	/* paranoid */ verify( ioarb );
	286
[78da4ab]	287	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
	288
[11054eb]	289	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
[78da4ab]	290
[dddb3dd0]	291	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
[df40a56]	292
[78da4ab]	293	__fill( sqes, want, idxs,ret );
	294	return ret;
[df40a56]	295	}
	296
[78da4ab]	297	//=============================================================================================
	298	// submission
[2432e8e]	299	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
[78da4ab]	300	// We can proceed to the fast path
	301	// Get the right objects
	302	__sub_ring_t & sq = ctx->sq;
	303	const __u32 mask = *sq.mask;
[dddb3dd0]	304	__u32 tail = *sq.kring.tail;
[78da4ab]	305
	306	// Add the sqes to the array
	307	for( i; have ) {
[dddb3dd0]	308	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
[78da4ab]	309	sq.kring.array[ (tail + i) & mask ] = idxs[i];
[426f60c]	310	}
	311
[78da4ab]	312	// Make the sqes visible to the submitter
[dddb3dd0]	313	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
[e8ac228]	314	sq.to_submit += have;
[426f60c]	315
[d529ad0]	316	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED);
	317	__atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED);
[2432e8e]	318	}
	319
	320	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
	321	__sub_ring_t & sq = ctx->sq;
	322	__submit_only(ctx, idxs, have);
	323
[70b4aeb9]	324	if(sq.to_submit > 30) {
	325	__tls_stats()->io.flush.full++;
	326	__cfa_io_flush( ctx->proc, 0 );
	327	}
	328	if(!lazy) {
	329	__tls_stats()->io.flush.eager++;
[c7b2215]	330	__cfa_io_flush( ctx->proc, 0 );
[dddb3dd0]	331	}
[78da4ab]	332	}
[2489d31]	333
[dddb3dd0]	334	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
	335	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
[5dadc9b7]	336
[78da4ab]	337	disable_interrupts();
	338	processor * proc = __cfaabi_tls.this_processor;
	339	$io_context * ctx = proc->io.ctx;
[dddb3dd0]	340	/* paranoid */ verify( __cfaabi_tls.this_processor );
	341	/* paranoid */ verify( ctx );
[e46c753]	342
[78da4ab]	343	// Can we proceed to the fast path
[dddb3dd0]	344	if( ctx == inctx ) // We have the right instance?
[78da4ab]	345	{
[dddb3dd0]	346	__submit(ctx, idxs, have, lazy);
[e46c753]	347
[78da4ab]	348	// Mark the instance as no longer in-use, re-enable interrupts and return
[d60d30e]	349	__STATS__( true, io.submit.fast += 1; )
[a3821fa]	350	enable_interrupts();
[ece0e80]	351
[78da4ab]	352	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
	353	return;
[e46c753]	354	}
[d384787]	355
[78da4ab]	356	// Fast path failed, fallback on arbitration
[d60d30e]	357	__STATS__( true, io.submit.slow += 1; )
[a3821fa]	358	enable_interrupts();
[5dadc9b7]	359
[78da4ab]	360	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
[426f60c]	361
[11054eb]	362	__ioarbiter_submit(inctx, idxs, have, lazy);
[78da4ab]	363	}
[2fab24e3]	364
[78da4ab]	365	//=============================================================================================
	366	// Flushing
[426f60c]	367	// Go through the ring's submit queue and release everything that has already been consumed
	368	// by io_uring
[78da4ab]	369	// This cannot be done by multiple threads
	370	static __u32 __release_sqes( struct $io_context & ctx ) {
	371	const __u32 mask = *ctx.sq.mask;
[732b406]	372
[426f60c]	373	__attribute__((unused))
[78da4ab]	374	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
	375	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
	376	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
	377
	378	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
[732b406]	379
[426f60c]	380	// the 3 fields are organized like this diagram
	381	// except it's are ring
	382	// ---+--------+--------+----
	383	// ---+--------+--------+----
	384	// ^ ^ ^
	385	// phead chead ctail
	386
	387	// make sure ctail doesn't wrap around and reach phead
	388	/* paranoid */ verify(
	389	(ctail >= chead && chead >= phead)
	390	\|\| (chead >= phead && phead >= ctail)
	391	\|\| (phead >= ctail && ctail >= chead)
	392	);
	393
	394	// find the range we need to clear
[4998155]	395	__u32 count = chead - phead;
[426f60c]	396
[78da4ab]	397	if(count == 0) {
	398	return 0;
	399	}
	400
[426f60c]	401	// We acquired an previous-head/current-head range
	402	// go through the range and release the sqes
[34b61882]	403	for( i; count ) {
[dddb3dd0]	404	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
[78da4ab]	405	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
	406	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
[34b61882]	407	}
[78da4ab]	408
	409	ctx.sq.kring.released = chead; // note up to were we processed
	410	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
	411
	412	__ioarbiter_notify(ctx);
	413
[34b61882]	414	return count;
	415	}
[35285fd]	416
[78da4ab]	417	//=============================================================================================
	418	// I/O Arbiter
	419	//=============================================================================================
[9f5a71eb]	420	static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) {
	421	bool was_empty;
	422
[11054eb]	423	// Lock the list, it's not thread safe
	424	lock( queue.lock __cfaabi_dbg_ctx2 );
	425	{
[9f5a71eb]	426	was_empty = empty(queue.queue);
	427
[11054eb]	428	// Add our request to the list
	429	add( queue.queue, item );
	430
	431	// Mark as pending
	432	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
	433	}
	434	unlock( queue.lock );
	435
[9f5a71eb]	436	return was_empty;
[11054eb]	437	}
	438
	439	static inline bool empty(__outstanding_io_queue & queue ) {
	440	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
	441	}
	442
	443	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
[78da4ab]	444	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
	445
[d60d30e]	446	__STATS__( false, io.alloc.block += 1; )
	447
[78da4ab]	448	// No one has any resources left, wait for something to finish
[11054eb]	449	// We need to add ourself to a list of pending allocs and wait for an answer
	450	__pending_alloc pa;
	451	pa.idxs = idxs;
	452	pa.want = want;
[78da4ab]	453
[9f5a71eb]	454	enqueue(this.pending, (__outstanding_io&)pa);
	455
	456	wait( pa.sem );
[78da4ab]	457
[11054eb]	458	return pa.ctx;
[dddb3dd0]	459
[78da4ab]	460	}
	461
[11054eb]	462	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
	463	/* paranoid */ verify( !empty(this.pending.queue) );
[78da4ab]	464
[11054eb]	465	lock( this.pending.lock __cfaabi_dbg_ctx2 );
	466	{
	467	while( !empty(this.pending.queue) ) {
	468	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
	469	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
	470	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
[78da4ab]	471
[11054eb]	472	if( have > pa.want ) goto DONE;
	473	drop( this.pending.queue );
[78da4ab]	474
[11054eb]	475	/* paranoid */__attribute__((unused)) bool ret =
[78da4ab]	476
[11054eb]	477	__alloc(ctx, pa.idxs, pa.want);
	478
	479	/* paranoid */ verify( ret );
	480
	481	pa.ctx = ctx;
	482
	483	post( pa.sem );
	484	}
	485
	486	this.pending.empty = true;
	487	DONE:;
	488	}
	489	unlock( this.pending.lock );
[78da4ab]	490	}
	491
	492	static void __ioarbiter_notify( $io_context & ctx ) {
[11054eb]	493	if(!empty( ctx.arbiter->pending )) {
[78da4ab]	494	__ioarbiter_notify( *ctx.arbiter, &ctx );
	495	}
	496	}
	497
	498	// Simply append to the pending
[11054eb]	499	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
[78da4ab]	500	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
	501
	502	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
	503
[11054eb]	504	__external_io ei;
	505	ei.idxs = idxs;
	506	ei.have = have;
	507	ei.lazy = lazy;
[78da4ab]	508
[9f5a71eb]	509	bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei);
	510
[d529ad0]	511	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST);
[9f5a71eb]	512
	513	if( we ) {
	514	sigval_t value = { PREEMPT_IO };
	515	pthread_sigqueue(ctx->proc->kernel_thread, SIGUSR1, value);
	516	}
	517
	518	wait( ei.sem );
[78da4ab]	519
	520	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
	521	}
	522
[11054eb]	523	static void __ioarbiter_flush( $io_context & ctx ) {
	524	if(!empty( ctx.ext_sq )) {
	525	__STATS__( false, io.flush.external += 1; )
[78da4ab]	526
[11054eb]	527	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
[d60d30e]	528
[11054eb]	529	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
	530	{
	531	while( !empty(ctx.ext_sq.queue) ) {
	532	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
[78da4ab]	533
[2432e8e]	534	__submit_only(&ctx, ei.idxs, ei.have);
[78da4ab]	535
[11054eb]	536	post( ei.sem );
	537	}
	538
	539	ctx.ext_sq.empty = true;
	540	}
	541	unlock(ctx.ext_sq.lock );
	542	}
[78da4ab]	543	}
[7ef162b2]	544
[d3605f8]	545	#if defined(CFA_WITH_IO_URING_IDLE)
	546	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
[6ddef36]	547	$io_context * ctx = proc->io.ctx;
	548	/* paranoid */ verify( ! __preemption_enabled() );
	549	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
	550	/* paranoid */ verify( ctx );
[7ef162b2]	551
[6ddef36]	552	__u32 idx;
	553	struct io_uring_sqe * sqe;
[7ef162b2]	554
[6ddef36]	555	// We can proceed to the fast path
	556	if( !__alloc(ctx, &idx, 1) ) return false;
	557
	558	// Allocation was successful
	559	__fill( &sqe, 1, &idx, ctx );
	560
	561	sqe->user_data = (uintptr_t)&future;
	562	sqe->flags = 0;
[a1f3d93]	563	sqe->fd = fd;
[6ddef36]	564	sqe->off = 0;
[d3605f8]	565	sqe->ioprio = 0;
[6ddef36]	566	sqe->fsync_flags = 0;
	567	sqe->__pad2[0] = 0;
	568	sqe->__pad2[1] = 0;
	569	sqe->__pad2[2] = 0;
[d3605f8]	570
	571	#if defined(CFA_HAVE_IORING_OP_READ)
	572	sqe->opcode = IORING_OP_READ;
	573	sqe->addr = (uint64_t)iov.iov_base;
	574	sqe->len = iov.iov_len;
	575	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
	576	sqe->opcode = IORING_OP_READV;
	577	sqe->addr = (uintptr_t)&iov;
	578	sqe->len = 1;
	579	#else
	580	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
	581	#endif
[6ddef36]	582
	583	asm volatile("": : :"memory");
	584
	585	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
	586	__submit( ctx, &idx, 1, true );
	587
	588	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
	589	/* paranoid */ verify( ! __preemption_enabled() );
[078fb05]	590
	591	return true;
[6ddef36]	592	}
	593	#endif
[47746a2]	594	#endif

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

Download in other formats:

Original Format