Context Navigation

io.cfa@ a76202d

Visit:

ADT ast-experimental enum pthread-emulation qualifiedEnum

Last change on this file since a76202d was 4ecc35a, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Added spin lock to io drain. last step before completion fairness
Property mode set to `100644`
File size: 17.5 KB

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

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

Download in other formats:

Original Format