Context Navigation

io.cfa @ ff7b2de

ADTast-experimentalenumpthread-emulationqualifiedEnum

Last change on this file since ff7b2de was 78a580d, checked in by Thierry Delisle <tdelisle@…>, 2 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

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 "kernel/cluster.hfa"
45	#include "io/types.hfa"
46
47	__attribute__((unused)) static const char * opcodes[] = {
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
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 & );
88	static inline void __ioarbiter_notify( $io_context & ctx );
89	//=============================================================================================
90	// I/O Polling
91	//=============================================================================================
92	static inline unsigned __flush( struct $io_context & );
93	static inline __u32 __release_sqes( struct $io_context & );
94	extern void __kernel_unpark( thread$ * thrd, unpark_hint );
95
96	bool __cfa_io_drain( $io_context * ctx, cluster * cltr ) {
97	/* paranoid */ verify( ! __preemption_enabled() );
98	/* paranoid */ verify( ready_schedule_islocked() );
99	/* paranoid */ verify( ctx );
100
101	// Drain the queue
102	unsigned head = *ctx->cq.head;
103	unsigned tail = *ctx->cq.tail;
104	const __u32 mask = *ctx->cq.mask;
105
106	__u32 count = tail - head;
107	__STATS__( false, io.calls.drain++; io.calls.completed += count; )
108
109	if(count == 0) return false;
110
111	if(!__atomic_try_acquire(&ctx->cq.lock)) {
112	return false;
113	}
114
115	unsigned long long ts_prev = ctx->cq.ts;
116
117	for(i; count) {
118	unsigned idx = (head + i) & mask;
119	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
120
121	/* paranoid */ verify(&cqe);
122
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
126	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
127	}
128
129	__cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
130	unsigned long long ts_next = ctx->cq.ts = rdtscl();
131
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.
134	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
135
136	/* paranoid */ verify( ready_schedule_islocked() );
137	/* paranoid */ verify( ! __preemption_enabled() );
138
139	__atomic_unlock(&ctx->cq.lock);
140
141	touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next );
142
143	return true;
144	}
145
146	bool __cfa_io_flush( processor * proc, int min_comp ) {
147	/* paranoid */ verify( ! __preemption_enabled() );
148	/* paranoid */ verify( proc );
149	/* paranoid */ verify( proc->io.ctx );
150
151	cluster * cltr = proc->cltr;
152	$io_context & ctx = *proc->io.ctx;
153
154	__ioarbiter_flush( ctx );
155
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	}
171	}
172
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 );
177
178	ctx.sq.to_submit -= ret;
179
180	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
181
182	// Release the consumed SQEs
183	__release_sqes( ctx );
184
185	/* paranoid */ verify( ! __preemption_enabled() );
186
187	__atomic_store_n(&ctx.proc->io.pending, false, __ATOMIC_RELAXED);
188	}
189
190	ready_schedule_lock();
191	bool ret = __cfa_io_drain( &ctx, cltr );
192	ready_schedule_unlock();
193	return ret;
194	}
195
196	//=============================================================================================
197	// I/O Submissions
198	//=============================================================================================
199
200	// Submition steps :
201	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
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	//
206	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
207	//
208	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
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	//
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) {
221	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
222	out_sqes[i] = &sqes[idxs[i]];
223	}
224	}
225
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
233
234	// If we don't have enough sqes, fail
235	if((ftail - fhead) < want) { return false; }
236
237	// copy all the indexes we want from the available list
238	for(i; want) {
239	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
240	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
241	}
242
243	// Advance the head to mark the indexes as consumed
244	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
245
246	// return success
247	return true;
248	}
249
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);
257
258	disable_interrupts();
259	processor * proc = __cfaabi_tls.this_processor;
260	$io_context * ctx = proc->io.ctx;
261	/* paranoid */ verify( __cfaabi_tls.this_processor );
262	/* paranoid */ verify( ctx );
263
264	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
265
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; )
270	enable_interrupts();
271
272	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
273
274	__fill( sqes, want, idxs, ctx );
275	return ctx;
276	}
277	// The fast path failed, fallback
278	__STATS__( true, io.alloc.fail += 1; )
279
280	// Fast path failed, fallback on arbitration
281	__STATS__( true, io.alloc.slow += 1; )
282	enable_interrupts();
283
284	$io_arbiter * ioarb = proc->cltr->io.arbiter;
285	/* paranoid */ verify( ioarb );
286
287	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
288
289	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
290
291	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
292
293	__fill( sqes, want, idxs,ret );
294	return ret;
295	}
296
297	//=============================================================================================
298	// submission
299	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
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;
304	__u32 tail = *sq.kring.tail;
305
306	// Add the sqes to the array
307	for( i; have ) {
308	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
309	sq.kring.array[ (tail + i) & mask ] = idxs[i];
310	}
311
312	// Make the sqes visible to the submitter
313	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
314	sq.to_submit += have;
315
316	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED);
317	__atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED);
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
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++;
330	__cfa_io_flush( ctx->proc, 0 );
331	}
332	}
333
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");
336
337	disable_interrupts();
338	processor * proc = __cfaabi_tls.this_processor;
339	$io_context * ctx = proc->io.ctx;
340	/* paranoid */ verify( __cfaabi_tls.this_processor );
341	/* paranoid */ verify( ctx );
342
343	// Can we proceed to the fast path
344	if( ctx == inctx ) // We have the right instance?
345	{
346	__submit(ctx, idxs, have, lazy);
347
348	// Mark the instance as no longer in-use, re-enable interrupts and return
349	__STATS__( true, io.submit.fast += 1; )
350	enable_interrupts();
351
352	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
353	return;
354	}
355
356	// Fast path failed, fallback on arbitration
357	__STATS__( true, io.submit.slow += 1; )
358	enable_interrupts();
359
360	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
361
362	__ioarbiter_submit(inctx, idxs, have, lazy);
363	}
364
365	//=============================================================================================
366	// Flushing
367	// Go through the ring's submit queue and release everything that has already been consumed
368	// by io_uring
369	// This cannot be done by multiple threads
370	static __u32 __release_sqes( struct $io_context & ctx ) {
371	const __u32 mask = *ctx.sq.mask;
372
373	__attribute__((unused))
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
379
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
395	__u32 count = chead - phead;
396
397	if(count == 0) {
398	return 0;
399	}
400
401	// We acquired an previous-head/current-head range
402	// go through the range and release the sqes
403	for( i; count ) {
404	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
405	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
406	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
407	}
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
414	return count;
415	}
416
417	//=============================================================================================
418	// I/O Arbiter
419	//=============================================================================================
420	static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) {
421	bool was_empty;
422
423	// Lock the list, it's not thread safe
424	lock( queue.lock __cfaabi_dbg_ctx2 );
425	{
426	was_empty = empty(queue.queue);
427
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
436	return was_empty;
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 ) {
444	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
445
446	__STATS__( false, io.alloc.block += 1; )
447
448	// No one has any resources left, wait for something to finish
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;
453
454	enqueue(this.pending, (__outstanding_io&)pa);
455
456	wait( pa.sem );
457
458	return pa.ctx;
459
460	}
461
462	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
463	/* paranoid */ verify( !empty(this.pending.queue) );
464
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 );
471
472	if( have > pa.want ) goto DONE;
473	drop( this.pending.queue );
474
475	/* paranoid */__attribute__((unused)) bool ret =
476
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 );
490	}
491
492	static void __ioarbiter_notify( $io_context & ctx ) {
493	if(!empty( ctx.arbiter->pending )) {
494	__ioarbiter_notify( *ctx.arbiter, &ctx );
495	}
496	}
497
498	// Simply append to the pending
499	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
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
504	__external_io ei;
505	ei.idxs = idxs;
506	ei.have = have;
507	ei.lazy = lazy;
508
509	bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei);
510
511	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST);
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 );
519
520	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
521	}
522
523	static void __ioarbiter_flush( $io_context & ctx ) {
524	if(!empty( ctx.ext_sq )) {
525	__STATS__( false, io.flush.external += 1; )
526
527	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
528
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 );
533
534	__submit_only(&ctx, ei.idxs, ei.have);
535
536	post( ei.sem );
537	}
538
539	ctx.ext_sq.empty = true;
540	}
541	unlock(ctx.ext_sq.lock );
542	}
543	}
544
545	#if defined(CFA_WITH_IO_URING_IDLE)
546	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
547	$io_context * ctx = proc->io.ctx;
548	/* paranoid */ verify( ! __preemption_enabled() );
549	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
550	/* paranoid */ verify( ctx );
551
552	__u32 idx;
553	struct io_uring_sqe * sqe;
554
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;
563	sqe->fd = fd;
564	sqe->off = 0;
565	sqe->ioprio = 0;
566	sqe->fsync_flags = 0;
567	sqe->__pad2[0] = 0;
568	sqe->__pad2[1] = 0;
569	sqe->__pad2[2] = 0;
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
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() );
590
591	return true;
592	}
593	#endif
594	#endif

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

Download in other formats:

Original Format