Context Navigation

io.cfa@ 8e819a9

Visit:

ADT ast-experimental enum pthread-emulation qualifiedEnum

Last change on this file since 8e819a9 was 54c1196, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Added io helping stats
Property mode set to `100644`
File size: 19.8 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
[4479890]	96	static bool __cfa_do_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
[dddb3dd0]	101	const __u32 mask = *ctx->cq.mask;
[92976d9]	102
[d60d30e]	103
[3caf5e3]	104	{
	105	const __u32 head = *ctx->cq.head;
	106	const __u32 tail = *ctx->cq.tail;
	107
	108	if(head == tail) return false;
	109	}
[c1c95b1]	110
[3caf5e3]	111	// Drain the queue
[4ecc35a]	112	if(!__atomic_try_acquire(&ctx->cq.lock)) {
[54c1196]	113	__STATS__( false, io.calls.locked++; )
[4ecc35a]	114	return false;
	115	}
	116
[78a580d]	117	unsigned long long ts_prev = ctx->cq.ts;
	118
[3caf5e3]	119	// re-read the head and tail in case it already changed.
	120	const __u32 head = *ctx->cq.head;
	121	const __u32 tail = *ctx->cq.tail;
	122	const __u32 count = tail - head;
	123	__STATS__( false, io.calls.drain++; io.calls.completed += count; )
	124
[d384787]	125	for(i; count) {
[6f121b8]	126	unsigned idx = (head + i) & mask;
[dddb3dd0]	127	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
[92976d9]	128
[d384787]	129	/* paranoid */ verify(&cqe);
[92976d9]	130
[78da4ab]	131	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
	132	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
	133
[24e321c]	134	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
[78da4ab]	135	}
	136
[dddb3dd0]	137	__cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
[78a580d]	138	unsigned long long ts_next = ctx->cq.ts = rdtscl();
[2d8f7b0]	139
[92976d9]	140	// Mark to the kernel that the cqe has been seen
	141	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
[dddb3dd0]	142	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
[92976d9]	143
[e9c0b4c]	144	/* paranoid */ verify( ready_schedule_islocked() );
[dddb3dd0]	145	/* paranoid */ verify( ! __preemption_enabled() );
	146
[4ecc35a]	147	__atomic_unlock(&ctx->cq.lock);
	148
[78a580d]	149	touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next );
	150
[c1c95b1]	151	return true;
[92976d9]	152	}
	153
[4479890]	154	bool __cfa_io_drain( processor * proc ) {
	155	bool local = false;
	156	bool remote = false;
	157
	158	cluster * const cltr = proc->cltr;
	159	$io_context * const ctx = proc->io.ctx;
	160	/* paranoid */ verify( cltr );
	161	/* paranoid */ verify( ctx );
	162
	163	with(cltr->sched) {
	164	const size_t ctxs_count = io.count;
	165
	166	/* paranoid */ verify( ready_schedule_islocked() );
	167	/* paranoid */ verify( ! __preemption_enabled() );
	168	/* paranoid */ verify( active_processor() == proc );
	169	/* paranoid */ verify( __shard_factor.io > 0 );
	170	/* paranoid */ verify( ctxs_count > 0 );
	171	/* paranoid */ verify( ctx->cq.id < ctxs_count );
	172
	173	const unsigned this_cache = cache_id(cltr, ctx->cq.id / __shard_factor.io);
	174	const unsigned long long ctsc = rdtscl();
	175
	176	if(proc->io.target == MAX) {
	177	uint64_t chaos = __tls_rand();
	178	unsigned ext = chaos & 0xff;
	179	unsigned other = (chaos >> 8) % (ctxs_count);
	180
	181	if(ext < 3 \|\| __atomic_load_n(&caches[other / __shard_factor.io].id, __ATOMIC_RELAXED) == this_cache) {
	182	proc->io.target = other;
	183	}
	184	}
	185	else {
	186	const unsigned target = proc->io.target;
	187	/* paranoid */ verify( io.tscs[target].tv != MAX );
	188	if(target < ctxs_count) {
	189	const unsigned long long cutoff = calc_cutoff(ctsc, ctx->cq.id, ctxs_count, io.data, io.tscs, __shard_factor.io);
	190	const unsigned long long age = moving_average(ctsc, io.tscs[target].tv, io.tscs[target].ma);
	191	// __cfadbg_print_safe(ready_queue, "Kernel : Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no");
	192	if(age > cutoff) {
	193	remote = __cfa_do_drain( io.data[target], cltr );
[54c1196]	194	if(remote) __STATS__( false, io.calls.helped++; )
[4479890]	195	}
	196	}
	197	proc->io.target = MAX;
	198	}
	199	}
	200
	201
	202	// Drain the local queue
	203	local = __cfa_do_drain( proc->io.ctx, cltr );
	204
	205	/* paranoid */ verify( ready_schedule_islocked() );
	206	/* paranoid */ verify( ! __preemption_enabled() );
	207	/* paranoid */ verify( active_processor() == proc );
	208	return local \|\| remote;
	209	}
	210
[c7b2215]	211	bool __cfa_io_flush( processor * proc, int min_comp ) {
[dddb3dd0]	212	/* paranoid */ verify( ! __preemption_enabled() );
	213	/* paranoid */ verify( proc );
	214	/* paranoid */ verify( proc->io.ctx );
[1539bbd]	215
[78a580d]	216	cluster * cltr = proc->cltr;
[dddb3dd0]	217	$io_context & ctx = *proc->io.ctx;
[78da4ab]	218
[11054eb]	219	__ioarbiter_flush( ctx );
[3c039b0]	220
[21a5bfb7]	221	if(ctx.sq.to_submit != 0 \|\| min_comp > 0) {
	222
	223	__STATS__( true, io.calls.flush++; )
	224	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);
	225	if( ret < 0 ) {
	226	switch((int)errno) {
	227	case EAGAIN:
	228	case EINTR:
	229	case EBUSY:
	230	// Update statistics
	231	__STATS__( false, io.calls.errors.busy ++; )
	232	return false;
	233	default:
	234	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
	235	}
[61dd73d]	236	}
[ece0e80]	237
[21a5bfb7]	238	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
	239	__STATS__( true, io.calls.submitted += ret; )
	240	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
	241	/* paranoid */ verify( ctx.sq.to_submit >= ret );
[dddb3dd0]	242
[21a5bfb7]	243	ctx.sq.to_submit -= ret;
[ece0e80]	244
[21a5bfb7]	245	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
[5dadc9b7]	246
[21a5bfb7]	247	// Release the consumed SQEs
	248	__release_sqes( ctx );
[ece0e80]	249
[21a5bfb7]	250	/* paranoid */ verify( ! __preemption_enabled() );
	251
[d529ad0]	252	__atomic_store_n(&ctx.proc->io.pending, false, __ATOMIC_RELAXED);
[21a5bfb7]	253	}
[61dd73d]	254
[7ef162b2]	255	ready_schedule_lock();
[4479890]	256	bool ret = __cfa_io_drain( proc );
[7ef162b2]	257	ready_schedule_unlock();
	258	return ret;
[61dd73d]	259	}
[f6660520]	260
[92976d9]	261	//=============================================================================================
	262	// I/O Submissions
	263	//=============================================================================================
	264
[2d8f7b0]	265	// Submition steps :
[e46c753]	266	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
[2d8f7b0]	267	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
	268	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
	269	// need to write an allocator that allows allocating concurrently.
	270	//
[e46c753]	271	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
[2d8f7b0]	272	//
[e46c753]	273	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
[2d8f7b0]	274	// needs to arrive to two concensus at the same time:
	275	// A - The order in which entries are listed in the array: no two threads must pick the
	276	// same index for their entries
	277	// B - When can the tail be update for the kernel. EVERY entries in the array between
	278	// head and tail must be fully filled and shouldn't ever be touched again.
	279	//
[78da4ab]	280	//=============================================================================================
	281	// Allocation
	282	// for user's convenience fill the sqes from the indexes
	283	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
	284	struct io_uring_sqe * sqes = ctx->sq.sqes;
	285	for(i; want) {
[dddb3dd0]	286	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
[78da4ab]	287	out_sqes[i] = &sqes[idxs[i]];
	288	}
	289	}
[2489d31]	290
[78da4ab]	291	// Try to directly allocate from the a given context
	292	// Not thread-safe
	293	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
	294	__sub_ring_t & sq = ctx->sq;
	295	const __u32 mask = *sq.mask;
	296	__u32 fhead = sq.free_ring.head; // get the current head of the queue
	297	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
[2489d31]	298
[78da4ab]	299	// If we don't have enough sqes, fail
	300	if((ftail - fhead) < want) { return false; }
[426f60c]	301
[78da4ab]	302	// copy all the indexes we want from the available list
	303	for(i; want) {
[dddb3dd0]	304	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
[78da4ab]	305	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
[6f121b8]	306	}
[2489d31]	307
[78da4ab]	308	// Advance the head to mark the indexes as consumed
	309	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
[df40a56]	310
[78da4ab]	311	// return success
	312	return true;
	313	}
[df40a56]	314
[78da4ab]	315	// Allocate an submit queue entry.
	316	// The kernel cannot see these entries until they are submitted, but other threads must be
	317	// able to see which entries can be used and which are already un used by an other thread
	318	// for convenience, return both the index and the pointer to the sqe
	319	// sqe == &sqes[idx]
	320	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
	321	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
[df40a56]	322
[78da4ab]	323	disable_interrupts();
	324	processor * proc = __cfaabi_tls.this_processor;
[dddb3dd0]	325	$io_context * ctx = proc->io.ctx;
[78da4ab]	326	/* paranoid */ verify( __cfaabi_tls.this_processor );
[dddb3dd0]	327	/* paranoid */ verify( ctx );
[78da4ab]	328
[dddb3dd0]	329	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
[78da4ab]	330
[dddb3dd0]	331	// We can proceed to the fast path
	332	if( __alloc(ctx, idxs, want) ) {
	333	// Allocation was successful
	334	__STATS__( true, io.alloc.fast += 1; )
[a3821fa]	335	enable_interrupts();
[df40a56]	336
[dddb3dd0]	337	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
[2fafe7e]	338
[dddb3dd0]	339	__fill( sqes, want, idxs, ctx );
	340	return ctx;
[df40a56]	341	}
[dddb3dd0]	342	// The fast path failed, fallback
	343	__STATS__( true, io.alloc.fail += 1; )
[df40a56]	344
[78da4ab]	345	// Fast path failed, fallback on arbitration
[d60d30e]	346	__STATS__( true, io.alloc.slow += 1; )
[a3821fa]	347	enable_interrupts();
[78da4ab]	348
[dddb3dd0]	349	$io_arbiter * ioarb = proc->cltr->io.arbiter;
	350	/* paranoid */ verify( ioarb );
	351
[78da4ab]	352	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
	353
[11054eb]	354	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
[78da4ab]	355
[dddb3dd0]	356	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
[df40a56]	357
[78da4ab]	358	__fill( sqes, want, idxs,ret );
	359	return ret;
[df40a56]	360	}
	361
[78da4ab]	362	//=============================================================================================
	363	// submission
[2432e8e]	364	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
[78da4ab]	365	// We can proceed to the fast path
	366	// Get the right objects
	367	__sub_ring_t & sq = ctx->sq;
	368	const __u32 mask = *sq.mask;
[dddb3dd0]	369	__u32 tail = *sq.kring.tail;
[78da4ab]	370
	371	// Add the sqes to the array
	372	for( i; have ) {
[dddb3dd0]	373	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
[78da4ab]	374	sq.kring.array[ (tail + i) & mask ] = idxs[i];
[426f60c]	375	}
	376
[78da4ab]	377	// Make the sqes visible to the submitter
[dddb3dd0]	378	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
[e8ac228]	379	sq.to_submit += have;
[426f60c]	380
[d529ad0]	381	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED);
	382	__atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED);
[2432e8e]	383	}
	384
	385	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
	386	__sub_ring_t & sq = ctx->sq;
	387	__submit_only(ctx, idxs, have);
	388
[70b4aeb9]	389	if(sq.to_submit > 30) {
	390	__tls_stats()->io.flush.full++;
	391	__cfa_io_flush( ctx->proc, 0 );
	392	}
	393	if(!lazy) {
	394	__tls_stats()->io.flush.eager++;
[c7b2215]	395	__cfa_io_flush( ctx->proc, 0 );
[dddb3dd0]	396	}
[78da4ab]	397	}
[2489d31]	398
[dddb3dd0]	399	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
	400	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
[5dadc9b7]	401
[78da4ab]	402	disable_interrupts();
	403	processor * proc = __cfaabi_tls.this_processor;
	404	$io_context * ctx = proc->io.ctx;
[dddb3dd0]	405	/* paranoid */ verify( __cfaabi_tls.this_processor );
	406	/* paranoid */ verify( ctx );
[e46c753]	407
[78da4ab]	408	// Can we proceed to the fast path
[dddb3dd0]	409	if( ctx == inctx ) // We have the right instance?
[78da4ab]	410	{
[dddb3dd0]	411	__submit(ctx, idxs, have, lazy);
[e46c753]	412
[78da4ab]	413	// Mark the instance as no longer in-use, re-enable interrupts and return
[d60d30e]	414	__STATS__( true, io.submit.fast += 1; )
[a3821fa]	415	enable_interrupts();
[ece0e80]	416
[78da4ab]	417	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
	418	return;
[e46c753]	419	}
[d384787]	420
[78da4ab]	421	// Fast path failed, fallback on arbitration
[d60d30e]	422	__STATS__( true, io.submit.slow += 1; )
[a3821fa]	423	enable_interrupts();
[5dadc9b7]	424
[78da4ab]	425	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
[426f60c]	426
[11054eb]	427	__ioarbiter_submit(inctx, idxs, have, lazy);
[78da4ab]	428	}
[2fab24e3]	429
[78da4ab]	430	//=============================================================================================
	431	// Flushing
[426f60c]	432	// Go through the ring's submit queue and release everything that has already been consumed
	433	// by io_uring
[78da4ab]	434	// This cannot be done by multiple threads
	435	static __u32 __release_sqes( struct $io_context & ctx ) {
	436	const __u32 mask = *ctx.sq.mask;
[732b406]	437
[426f60c]	438	__attribute__((unused))
[78da4ab]	439	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
	440	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
	441	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
	442
	443	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
[732b406]	444
[426f60c]	445	// the 3 fields are organized like this diagram
	446	// except it's are ring
	447	// ---+--------+--------+----
	448	// ---+--------+--------+----
	449	// ^ ^ ^
	450	// phead chead ctail
	451
	452	// make sure ctail doesn't wrap around and reach phead
	453	/* paranoid */ verify(
	454	(ctail >= chead && chead >= phead)
	455	\|\| (chead >= phead && phead >= ctail)
	456	\|\| (phead >= ctail && ctail >= chead)
	457	);
	458
	459	// find the range we need to clear
[4998155]	460	__u32 count = chead - phead;
[426f60c]	461
[78da4ab]	462	if(count == 0) {
	463	return 0;
	464	}
	465
[426f60c]	466	// We acquired an previous-head/current-head range
	467	// go through the range and release the sqes
[34b61882]	468	for( i; count ) {
[dddb3dd0]	469	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
[78da4ab]	470	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
	471	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
[34b61882]	472	}
[78da4ab]	473
	474	ctx.sq.kring.released = chead; // note up to were we processed
	475	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
	476
	477	__ioarbiter_notify(ctx);
	478
[34b61882]	479	return count;
	480	}
[35285fd]	481
[78da4ab]	482	//=============================================================================================
	483	// I/O Arbiter
	484	//=============================================================================================
[9f5a71eb]	485	static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) {
	486	bool was_empty;
	487
[11054eb]	488	// Lock the list, it's not thread safe
	489	lock( queue.lock __cfaabi_dbg_ctx2 );
	490	{
[9f5a71eb]	491	was_empty = empty(queue.queue);
	492
[11054eb]	493	// Add our request to the list
	494	add( queue.queue, item );
	495
	496	// Mark as pending
	497	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
	498	}
	499	unlock( queue.lock );
	500
[9f5a71eb]	501	return was_empty;
[11054eb]	502	}
	503
	504	static inline bool empty(__outstanding_io_queue & queue ) {
	505	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
	506	}
	507
	508	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
[78da4ab]	509	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
	510
[d60d30e]	511	__STATS__( false, io.alloc.block += 1; )
	512
[78da4ab]	513	// No one has any resources left, wait for something to finish
[11054eb]	514	// We need to add ourself to a list of pending allocs and wait for an answer
	515	__pending_alloc pa;
	516	pa.idxs = idxs;
	517	pa.want = want;
[78da4ab]	518
[9f5a71eb]	519	enqueue(this.pending, (__outstanding_io&)pa);
	520
	521	wait( pa.sem );
[78da4ab]	522
[11054eb]	523	return pa.ctx;
[dddb3dd0]	524
[78da4ab]	525	}
	526
[11054eb]	527	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
	528	/* paranoid */ verify( !empty(this.pending.queue) );
[78da4ab]	529
[11054eb]	530	lock( this.pending.lock __cfaabi_dbg_ctx2 );
	531	{
	532	while( !empty(this.pending.queue) ) {
	533	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
	534	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
	535	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
[78da4ab]	536
[11054eb]	537	if( have > pa.want ) goto DONE;
	538	drop( this.pending.queue );
[78da4ab]	539
[11054eb]	540	/* paranoid */__attribute__((unused)) bool ret =
[78da4ab]	541
[11054eb]	542	__alloc(ctx, pa.idxs, pa.want);
	543
	544	/* paranoid */ verify( ret );
	545
	546	pa.ctx = ctx;
	547
	548	post( pa.sem );
	549	}
	550
	551	this.pending.empty = true;
	552	DONE:;
	553	}
	554	unlock( this.pending.lock );
[78da4ab]	555	}
	556
	557	static void __ioarbiter_notify( $io_context & ctx ) {
[11054eb]	558	if(!empty( ctx.arbiter->pending )) {
[78da4ab]	559	__ioarbiter_notify( *ctx.arbiter, &ctx );
	560	}
	561	}
	562
	563	// Simply append to the pending
[11054eb]	564	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
[78da4ab]	565	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
	566
	567	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
	568
[11054eb]	569	__external_io ei;
	570	ei.idxs = idxs;
	571	ei.have = have;
	572	ei.lazy = lazy;
[78da4ab]	573
[9f5a71eb]	574	bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei);
	575
[d529ad0]	576	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST);
[9f5a71eb]	577
	578	if( we ) {
	579	sigval_t value = { PREEMPT_IO };
	580	pthread_sigqueue(ctx->proc->kernel_thread, SIGUSR1, value);
	581	}
	582
	583	wait( ei.sem );
[78da4ab]	584
	585	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
	586	}
	587
[11054eb]	588	static void __ioarbiter_flush( $io_context & ctx ) {
	589	if(!empty( ctx.ext_sq )) {
	590	__STATS__( false, io.flush.external += 1; )
[78da4ab]	591
[11054eb]	592	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
[d60d30e]	593
[11054eb]	594	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
	595	{
	596	while( !empty(ctx.ext_sq.queue) ) {
	597	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
[78da4ab]	598
[2432e8e]	599	__submit_only(&ctx, ei.idxs, ei.have);
[78da4ab]	600
[11054eb]	601	post( ei.sem );
	602	}
	603
	604	ctx.ext_sq.empty = true;
	605	}
	606	unlock(ctx.ext_sq.lock );
	607	}
[78da4ab]	608	}
[7ef162b2]	609
[d3605f8]	610	#if defined(CFA_WITH_IO_URING_IDLE)
	611	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
[6ddef36]	612	$io_context * ctx = proc->io.ctx;
	613	/* paranoid */ verify( ! __preemption_enabled() );
	614	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
	615	/* paranoid */ verify( ctx );
[7ef162b2]	616
[6ddef36]	617	__u32 idx;
	618	struct io_uring_sqe * sqe;
[7ef162b2]	619
[6ddef36]	620	// We can proceed to the fast path
	621	if( !__alloc(ctx, &idx, 1) ) return false;
	622
	623	// Allocation was successful
	624	__fill( &sqe, 1, &idx, ctx );
	625
	626	sqe->user_data = (uintptr_t)&future;
	627	sqe->flags = 0;
[a1f3d93]	628	sqe->fd = fd;
[6ddef36]	629	sqe->off = 0;
[d3605f8]	630	sqe->ioprio = 0;
[6ddef36]	631	sqe->fsync_flags = 0;
	632	sqe->__pad2[0] = 0;
	633	sqe->__pad2[1] = 0;
	634	sqe->__pad2[2] = 0;
[d3605f8]	635
	636	#if defined(CFA_HAVE_IORING_OP_READ)
	637	sqe->opcode = IORING_OP_READ;
	638	sqe->addr = (uint64_t)iov.iov_base;
	639	sqe->len = iov.iov_len;
	640	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
	641	sqe->opcode = IORING_OP_READV;
	642	sqe->addr = (uintptr_t)&iov;
	643	sqe->len = 1;
	644	#else
	645	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
	646	#endif
[6ddef36]	647
	648	asm volatile("": : :"memory");
	649
	650	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
	651	__submit( ctx, &idx, 1, true );
	652
	653	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
	654	/* paranoid */ verify( ! __preemption_enabled() );
[078fb05]	655
	656	return true;
[6ddef36]	657	}
	658	#endif
[47746a2]	659	#endif

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

Download in other formats:

Original Format