Context Navigation

io.cfa@ 919a6b2

Visit:

ADT ast-experimental enum forall-pointer-decay pthread-emulation qualifiedEnum

Last change on this file since 919a6b2 was 078fb05, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Fixed a few warnings
Property mode set to `100644`
File size: 16.7 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( processor * proc ) {
96	/* paranoid */ verify( ! __preemption_enabled() );
97	/* paranoid */ verify( ready_schedule_islocked() );
98	/* paranoid */ verify( proc );
99	/* paranoid */ verify( proc->io.ctx );
100
101	// Drain the queue
102	$io_context * ctx = proc->io.ctx;
103	unsigned head = *ctx->cq.head;
104	unsigned tail = *ctx->cq.tail;
105	const __u32 mask = *ctx->cq.mask;
106
107	__u32 count = tail - head;
108	__STATS__( false, io.calls.drain++; io.calls.completed += count; )
109
110	if(count == 0) return false;
111
112	for(i; count) {
113	unsigned idx = (head + i) & mask;
114	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
115
116	/* paranoid */ verify(&cqe);
117
118	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
119	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
120
121	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
122	}
123
124	__cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
125
126	// Mark to the kernel that the cqe has been seen
127	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
128	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
129
130	/* paranoid */ verify( ready_schedule_islocked() );
131	/* paranoid */ verify( ! __preemption_enabled() );
132
133	return true;
134	}
135
136	bool __cfa_io_flush( processor * proc, int min_comp ) {
137	/* paranoid */ verify( ! __preemption_enabled() );
138	/* paranoid */ verify( proc );
139	/* paranoid */ verify( proc->io.ctx );
140
141	__attribute__((unused)) cluster * cltr = proc->cltr;
142	$io_context & ctx = *proc->io.ctx;
143
144	__ioarbiter_flush( ctx );
145
146	__STATS__( true, io.calls.flush++; )
147	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);
148	if( ret < 0 ) {
149	switch((int)errno) {
150	case EAGAIN:
151	case EINTR:
152	case EBUSY:
153	// Update statistics
154	__STATS__( false, io.calls.errors.busy ++; )
155	return false;
156	default:
157	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
158	}
159	}
160
161	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
162	__STATS__( true, io.calls.submitted += ret; )
163	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
164	/* paranoid */ verify( ctx.sq.to_submit >= ret );
165
166	ctx.sq.to_submit -= ret;
167
168	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
169
170	// Release the consumed SQEs
171	__release_sqes( ctx );
172
173	/* paranoid */ verify( ! __preemption_enabled() );
174
175	ctx.proc->io.pending = false;
176	ready_schedule_lock();
177	bool ret = __cfa_io_drain( proc );
178	ready_schedule_unlock();
179	return ret;
180	}
181
182	//=============================================================================================
183	// I/O Submissions
184	//=============================================================================================
185
186	// Submition steps :
187	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
188	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
189	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
190	// need to write an allocator that allows allocating concurrently.
191	//
192	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
193	//
194	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
195	// needs to arrive to two concensus at the same time:
196	// A - The order in which entries are listed in the array: no two threads must pick the
197	// same index for their entries
198	// B - When can the tail be update for the kernel. EVERY entries in the array between
199	// head and tail must be fully filled and shouldn't ever be touched again.
200	//
201	//=============================================================================================
202	// Allocation
203	// for user's convenience fill the sqes from the indexes
204	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
205	struct io_uring_sqe * sqes = ctx->sq.sqes;
206	for(i; want) {
207	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
208	out_sqes[i] = &sqes[idxs[i]];
209	}
210	}
211
212	// Try to directly allocate from the a given context
213	// Not thread-safe
214	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
215	__sub_ring_t & sq = ctx->sq;
216	const __u32 mask = *sq.mask;
217	__u32 fhead = sq.free_ring.head; // get the current head of the queue
218	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
219
220	// If we don't have enough sqes, fail
221	if((ftail - fhead) < want) { return false; }
222
223	// copy all the indexes we want from the available list
224	for(i; want) {
225	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
226	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
227	}
228
229	// Advance the head to mark the indexes as consumed
230	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
231
232	// return success
233	return true;
234	}
235
236	// Allocate an submit queue entry.
237	// The kernel cannot see these entries until they are submitted, but other threads must be
238	// able to see which entries can be used and which are already un used by an other thread
239	// for convenience, return both the index and the pointer to the sqe
240	// sqe == &sqes[idx]
241	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
242	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
243
244	disable_interrupts();
245	processor * proc = __cfaabi_tls.this_processor;
246	$io_context * ctx = proc->io.ctx;
247	/* paranoid */ verify( __cfaabi_tls.this_processor );
248	/* paranoid */ verify( ctx );
249
250	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
251
252	// We can proceed to the fast path
253	if( __alloc(ctx, idxs, want) ) {
254	// Allocation was successful
255	__STATS__( true, io.alloc.fast += 1; )
256	enable_interrupts();
257
258	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
259
260	__fill( sqes, want, idxs, ctx );
261	return ctx;
262	}
263	// The fast path failed, fallback
264	__STATS__( true, io.alloc.fail += 1; )
265
266	// Fast path failed, fallback on arbitration
267	__STATS__( true, io.alloc.slow += 1; )
268	enable_interrupts();
269
270	$io_arbiter * ioarb = proc->cltr->io.arbiter;
271	/* paranoid */ verify( ioarb );
272
273	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
274
275	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
276
277	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
278
279	__fill( sqes, want, idxs,ret );
280	return ret;
281	}
282
283	//=============================================================================================
284	// submission
285	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
286	// We can proceed to the fast path
287	// Get the right objects
288	__sub_ring_t & sq = ctx->sq;
289	const __u32 mask = *sq.mask;
290	__u32 tail = *sq.kring.tail;
291
292	// Add the sqes to the array
293	for( i; have ) {
294	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
295	sq.kring.array[ (tail + i) & mask ] = idxs[i];
296	}
297
298	// Make the sqes visible to the submitter
299	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
300	sq.to_submit += have;
301
302	ctx->proc->io.pending = true;
303	ctx->proc->io.dirty = true;
304	if(sq.to_submit > 30 \|\| !lazy) {
305	__cfa_io_flush( ctx->proc, 0 );
306	}
307	}
308
309	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
310	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
311
312	disable_interrupts();
313	processor * proc = __cfaabi_tls.this_processor;
314	$io_context * ctx = proc->io.ctx;
315	/* paranoid */ verify( __cfaabi_tls.this_processor );
316	/* paranoid */ verify( ctx );
317
318	// Can we proceed to the fast path
319	if( ctx == inctx ) // We have the right instance?
320	{
321	__submit(ctx, idxs, have, lazy);
322
323	// Mark the instance as no longer in-use, re-enable interrupts and return
324	__STATS__( true, io.submit.fast += 1; )
325	enable_interrupts();
326
327	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
328	return;
329	}
330
331	// Fast path failed, fallback on arbitration
332	__STATS__( true, io.submit.slow += 1; )
333	enable_interrupts();
334
335	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
336
337	__ioarbiter_submit(inctx, idxs, have, lazy);
338	}
339
340	//=============================================================================================
341	// Flushing
342	// Go through the ring's submit queue and release everything that has already been consumed
343	// by io_uring
344	// This cannot be done by multiple threads
345	static __u32 __release_sqes( struct $io_context & ctx ) {
346	const __u32 mask = *ctx.sq.mask;
347
348	__attribute__((unused))
349	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
350	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
351	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
352
353	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
354
355	// the 3 fields are organized like this diagram
356	// except it's are ring
357	// ---+--------+--------+----
358	// ---+--------+--------+----
359	// ^ ^ ^
360	// phead chead ctail
361
362	// make sure ctail doesn't wrap around and reach phead
363	/* paranoid */ verify(
364	(ctail >= chead && chead >= phead)
365	\|\| (chead >= phead && phead >= ctail)
366	\|\| (phead >= ctail && ctail >= chead)
367	);
368
369	// find the range we need to clear
370	__u32 count = chead - phead;
371
372	if(count == 0) {
373	return 0;
374	}
375
376	// We acquired an previous-head/current-head range
377	// go through the range and release the sqes
378	for( i; count ) {
379	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
380	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
381	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
382	}
383
384	ctx.sq.kring.released = chead; // note up to were we processed
385	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
386
387	__ioarbiter_notify(ctx);
388
389	return count;
390	}
391
392	//=============================================================================================
393	// I/O Arbiter
394	//=============================================================================================
395	static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
396	// Lock the list, it's not thread safe
397	lock( queue.lock __cfaabi_dbg_ctx2 );
398	{
399	// Add our request to the list
400	add( queue.queue, item );
401
402	// Mark as pending
403	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
404	}
405	unlock( queue.lock );
406
407	wait( item.sem );
408	}
409
410	static inline bool empty(__outstanding_io_queue & queue ) {
411	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
412	}
413
414	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
415	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
416
417	__STATS__( false, io.alloc.block += 1; )
418
419	// No one has any resources left, wait for something to finish
420	// We need to add ourself to a list of pending allocs and wait for an answer
421	__pending_alloc pa;
422	pa.idxs = idxs;
423	pa.want = want;
424
425	block(this.pending, (__outstanding_io&)pa);
426
427	return pa.ctx;
428
429	}
430
431	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
432	/* paranoid */ verify( !empty(this.pending.queue) );
433
434	lock( this.pending.lock __cfaabi_dbg_ctx2 );
435	{
436	while( !empty(this.pending.queue) ) {
437	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
438	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
439	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
440
441	if( have > pa.want ) goto DONE;
442	drop( this.pending.queue );
443
444	/* paranoid */__attribute__((unused)) bool ret =
445
446	__alloc(ctx, pa.idxs, pa.want);
447
448	/* paranoid */ verify( ret );
449
450	pa.ctx = ctx;
451
452	post( pa.sem );
453	}
454
455	this.pending.empty = true;
456	DONE:;
457	}
458	unlock( this.pending.lock );
459	}
460
461	static void __ioarbiter_notify( $io_context & ctx ) {
462	if(!empty( ctx.arbiter->pending )) {
463	__ioarbiter_notify( *ctx.arbiter, &ctx );
464	}
465	}
466
467	// Simply append to the pending
468	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
469	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
470
471	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
472
473	__external_io ei;
474	ei.idxs = idxs;
475	ei.have = have;
476	ei.lazy = lazy;
477
478	block(ctx->ext_sq, (__outstanding_io&)ei);
479
480	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
481	}
482
483	static void __ioarbiter_flush( $io_context & ctx ) {
484	if(!empty( ctx.ext_sq )) {
485	__STATS__( false, io.flush.external += 1; )
486
487	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
488
489	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
490	{
491	while( !empty(ctx.ext_sq.queue) ) {
492	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
493
494	__submit(&ctx, ei.idxs, ei.have, ei.lazy);
495
496	post( ei.sem );
497	}
498
499	ctx.ext_sq.empty = true;
500	}
501	unlock(ctx.ext_sq.lock );
502	}
503	}
504
505	#if defined(CFA_WITH_IO_URING_IDLE)
506	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
507	$io_context * ctx = proc->io.ctx;
508	/* paranoid */ verify( ! __preemption_enabled() );
509	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
510	/* paranoid */ verify( ctx );
511
512	__u32 idx;
513	struct io_uring_sqe * sqe;
514
515	// We can proceed to the fast path
516	if( !__alloc(ctx, &idx, 1) ) return false;
517
518	// Allocation was successful
519	__fill( &sqe, 1, &idx, ctx );
520
521	sqe->user_data = (uintptr_t)&future;
522	sqe->flags = 0;
523	sqe->fd = fd;
524	sqe->off = 0;
525	sqe->ioprio = 0;
526	sqe->fsync_flags = 0;
527	sqe->__pad2[0] = 0;
528	sqe->__pad2[1] = 0;
529	sqe->__pad2[2] = 0;
530
531	#if defined(CFA_HAVE_IORING_OP_READ)
532	sqe->opcode = IORING_OP_READ;
533	sqe->addr = (uint64_t)iov.iov_base;
534	sqe->len = iov.iov_len;
535	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
536	sqe->opcode = IORING_OP_READV;
537	sqe->addr = (uintptr_t)&iov;
538	sqe->len = 1;
539	#else
540	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
541	#endif
542
543	asm volatile("": : :"memory");
544
545	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
546	__submit( ctx, &idx, 1, true );
547
548	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
549	/* paranoid */ verify( ! __preemption_enabled() );
550
551	return true;
552	}
553	#endif
554	#endif

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

Download in other formats:

Original Format