Context Navigation

io.cfa@ ddd2ec9

Visit:

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

Last change on this file since ddd2ec9 was 7ef162b2, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
First attempt at using io_uring_enter for idle sleep.
Property mode set to `100644`
File size: 16.2 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
36	#include <linux/io_uring.h>
37	}
38
39	#include "stats.hfa"
40	#include "kernel.hfa"
41	#include "kernel/fwd.hfa"
42	#include "kernel_private.hfa"
43	#include "io/types.hfa"
44
45	__attribute__((unused)) static const char * opcodes[] = {
46	"OP_NOP",
47	"OP_READV",
48	"OP_WRITEV",
49	"OP_FSYNC",
50	"OP_READ_FIXED",
51	"OP_WRITE_FIXED",
52	"OP_POLL_ADD",
53	"OP_POLL_REMOVE",
54	"OP_SYNC_FILE_RANGE",
55	"OP_SENDMSG",
56	"OP_RECVMSG",
57	"OP_TIMEOUT",
58	"OP_TIMEOUT_REMOVE",
59	"OP_ACCEPT",
60	"OP_ASYNC_CANCEL",
61	"OP_LINK_TIMEOUT",
62	"OP_CONNECT",
63	"OP_FALLOCATE",
64	"OP_OPENAT",
65	"OP_CLOSE",
66	"OP_FILES_UPDATE",
67	"OP_STATX",
68	"OP_READ",
69	"OP_WRITE",
70	"OP_FADVISE",
71	"OP_MADVISE",
72	"OP_SEND",
73	"OP_RECV",
74	"OP_OPENAT2",
75	"OP_EPOLL_CTL",
76	"OP_SPLICE",
77	"OP_PROVIDE_BUFFERS",
78	"OP_REMOVE_BUFFERS",
79	"OP_TEE",
80	"INVALID_OP"
81	};
82
83	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want );
84	static void __ioarbiter_submit( $io_context * , __u32 idxs[], __u32 have, bool lazy );
85	static void __ioarbiter_flush ( $io_context & );
86	static inline void __ioarbiter_notify( $io_context & ctx );
87	//=============================================================================================
88	// I/O Polling
89	//=============================================================================================
90	static inline unsigned __flush( struct $io_context & );
91	static inline __u32 __release_sqes( struct $io_context & );
92	extern void __kernel_unpark( thread$ * thrd, unpark_hint );
93
94	bool __cfa_io_drain( processor * proc ) {
95	/* paranoid */ verify( ! __preemption_enabled() );
96	/* paranoid */ verify( ready_schedule_islocked() );
97	/* paranoid */ verify( proc );
98	/* paranoid */ verify( proc->io.ctx );
99
100	// Drain the queue
101	$io_context * ctx = proc->io.ctx;
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	for(i; count) {
112	unsigned idx = (head + i) & mask;
113	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
114
115	/* paranoid */ verify(&cqe);
116
117	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
118	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
119
120	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
121	}
122
123	__cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
124
125	// Mark to the kernel that the cqe has been seen
126	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
127	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
128
129	/* paranoid */ verify( ready_schedule_islocked() );
130	/* paranoid */ verify( ! __preemption_enabled() );
131
132	return true;
133	}
134
135	bool __cfa_io_flush( processor * proc, bool wait ) {
136	/* paranoid */ verify( ! __preemption_enabled() );
137	/* paranoid */ verify( proc );
138	/* paranoid */ verify( proc->io.ctx );
139
140	__attribute__((unused)) cluster * cltr = proc->cltr;
141	$io_context & ctx = *proc->io.ctx;
142
143	__ioarbiter_flush( ctx );
144
145	__STATS__( true, io.calls.flush++; )
146	int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, wait ? 1 : 0, 0, (sigset_t *)0p, _NSIG / 8);
147	if( ret < 0 ) {
148	switch((int)errno) {
149	case EAGAIN:
150	case EINTR:
151	case EBUSY:
152	// Update statistics
153	__STATS__( false, io.calls.errors.busy ++; )
154	return false;
155	default:
156	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
157	}
158	}
159
160	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
161	__STATS__( true, io.calls.submitted += ret; )
162	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
163	/* paranoid */ verify( ctx.sq.to_submit >= ret );
164
165	ctx.sq.to_submit -= ret;
166
167	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
168
169	// Release the consumed SQEs
170	__release_sqes( ctx );
171
172	/* paranoid */ verify( ! __preemption_enabled() );
173
174	ctx.proc->io.pending = false;
175	ready_schedule_lock();
176	bool ret = __cfa_io_drain( proc );
177	ready_schedule_unlock();
178	return ret;
179	}
180
181	//=============================================================================================
182	// I/O Submissions
183	//=============================================================================================
184
185	// Submition steps :
186	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
187	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
188	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
189	// need to write an allocator that allows allocating concurrently.
190	//
191	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
192	//
193	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
194	// needs to arrive to two concensus at the same time:
195	// A - The order in which entries are listed in the array: no two threads must pick the
196	// same index for their entries
197	// B - When can the tail be update for the kernel. EVERY entries in the array between
198	// head and tail must be fully filled and shouldn't ever be touched again.
199	//
200	//=============================================================================================
201	// Allocation
202	// for user's convenience fill the sqes from the indexes
203	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
204	struct io_uring_sqe * sqes = ctx->sq.sqes;
205	for(i; want) {
206	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
207	out_sqes[i] = &sqes[idxs[i]];
208	}
209	}
210
211	// Try to directly allocate from the a given context
212	// Not thread-safe
213	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
214	__sub_ring_t & sq = ctx->sq;
215	const __u32 mask = *sq.mask;
216	__u32 fhead = sq.free_ring.head; // get the current head of the queue
217	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
218
219	// If we don't have enough sqes, fail
220	if((ftail - fhead) < want) { return false; }
221
222	// copy all the indexes we want from the available list
223	for(i; want) {
224	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
225	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
226	}
227
228	// Advance the head to mark the indexes as consumed
229	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
230
231	// return success
232	return true;
233	}
234
235	// Allocate an submit queue entry.
236	// The kernel cannot see these entries until they are submitted, but other threads must be
237	// able to see which entries can be used and which are already un used by an other thread
238	// for convenience, return both the index and the pointer to the sqe
239	// sqe == &sqes[idx]
240	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
241	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
242
243	disable_interrupts();
244	processor * proc = __cfaabi_tls.this_processor;
245	$io_context * ctx = proc->io.ctx;
246	/* paranoid */ verify( __cfaabi_tls.this_processor );
247	/* paranoid */ verify( ctx );
248
249	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
250
251	// We can proceed to the fast path
252	if( __alloc(ctx, idxs, want) ) {
253	// Allocation was successful
254	__STATS__( true, io.alloc.fast += 1; )
255	enable_interrupts();
256
257	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
258
259	__fill( sqes, want, idxs, ctx );
260	return ctx;
261	}
262	// The fast path failed, fallback
263	__STATS__( true, io.alloc.fail += 1; )
264
265	// Fast path failed, fallback on arbitration
266	__STATS__( true, io.alloc.slow += 1; )
267	enable_interrupts();
268
269	$io_arbiter * ioarb = proc->cltr->io.arbiter;
270	/* paranoid */ verify( ioarb );
271
272	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
273
274	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
275
276	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
277
278	__fill( sqes, want, idxs,ret );
279	return ret;
280	}
281
282	//=============================================================================================
283	// submission
284	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
285	// We can proceed to the fast path
286	// Get the right objects
287	__sub_ring_t & sq = ctx->sq;
288	const __u32 mask = *sq.mask;
289	__u32 tail = *sq.kring.tail;
290
291	// Add the sqes to the array
292	for( i; have ) {
293	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
294	sq.kring.array[ (tail + i) & mask ] = idxs[i];
295	}
296
297	// Make the sqes visible to the submitter
298	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
299	sq.to_submit += have;
300
301	ctx->proc->io.pending = true;
302	ctx->proc->io.dirty = true;
303	if(sq.to_submit > 30 \|\| !lazy) {
304	__cfa_io_flush( ctx->proc, false );
305	}
306	}
307
308	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
309	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
310
311	disable_interrupts();
312	processor * proc = __cfaabi_tls.this_processor;
313	$io_context * ctx = proc->io.ctx;
314	/* paranoid */ verify( __cfaabi_tls.this_processor );
315	/* paranoid */ verify( ctx );
316
317	// Can we proceed to the fast path
318	if( ctx == inctx ) // We have the right instance?
319	{
320	__submit(ctx, idxs, have, lazy);
321
322	// Mark the instance as no longer in-use, re-enable interrupts and return
323	__STATS__( true, io.submit.fast += 1; )
324	enable_interrupts();
325
326	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
327	return;
328	}
329
330	// Fast path failed, fallback on arbitration
331	__STATS__( true, io.submit.slow += 1; )
332	enable_interrupts();
333
334	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
335
336	__ioarbiter_submit(inctx, idxs, have, lazy);
337	}
338
339	//=============================================================================================
340	// Flushing
341	// Go through the ring's submit queue and release everything that has already been consumed
342	// by io_uring
343	// This cannot be done by multiple threads
344	static __u32 __release_sqes( struct $io_context & ctx ) {
345	const __u32 mask = *ctx.sq.mask;
346
347	__attribute__((unused))
348	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
349	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
350	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
351
352	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
353
354	// the 3 fields are organized like this diagram
355	// except it's are ring
356	// ---+--------+--------+----
357	// ---+--------+--------+----
358	// ^ ^ ^
359	// phead chead ctail
360
361	// make sure ctail doesn't wrap around and reach phead
362	/* paranoid */ verify(
363	(ctail >= chead && chead >= phead)
364	\|\| (chead >= phead && phead >= ctail)
365	\|\| (phead >= ctail && ctail >= chead)
366	);
367
368	// find the range we need to clear
369	__u32 count = chead - phead;
370
371	if(count == 0) {
372	return 0;
373	}
374
375	// We acquired an previous-head/current-head range
376	// go through the range and release the sqes
377	for( i; count ) {
378	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
379	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
380	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
381	}
382
383	ctx.sq.kring.released = chead; // note up to were we processed
384	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
385
386	__ioarbiter_notify(ctx);
387
388	return count;
389	}
390
391	//=============================================================================================
392	// I/O Arbiter
393	//=============================================================================================
394	static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
395	// Lock the list, it's not thread safe
396	lock( queue.lock __cfaabi_dbg_ctx2 );
397	{
398	// Add our request to the list
399	add( queue.queue, item );
400
401	// Mark as pending
402	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
403	}
404	unlock( queue.lock );
405
406	wait( item.sem );
407	}
408
409	static inline bool empty(__outstanding_io_queue & queue ) {
410	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
411	}
412
413	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
414	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
415
416	__STATS__( false, io.alloc.block += 1; )
417
418	// No one has any resources left, wait for something to finish
419	// We need to add ourself to a list of pending allocs and wait for an answer
420	__pending_alloc pa;
421	pa.idxs = idxs;
422	pa.want = want;
423
424	block(this.pending, (__outstanding_io&)pa);
425
426	return pa.ctx;
427
428	}
429
430	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
431	/* paranoid */ verify( !empty(this.pending.queue) );
432
433	lock( this.pending.lock __cfaabi_dbg_ctx2 );
434	{
435	while( !empty(this.pending.queue) ) {
436	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
437	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
438	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
439
440	if( have > pa.want ) goto DONE;
441	drop( this.pending.queue );
442
443	/* paranoid */__attribute__((unused)) bool ret =
444
445	__alloc(ctx, pa.idxs, pa.want);
446
447	/* paranoid */ verify( ret );
448
449	pa.ctx = ctx;
450
451	post( pa.sem );
452	}
453
454	this.pending.empty = true;
455	DONE:;
456	}
457	unlock( this.pending.lock );
458	}
459
460	static void __ioarbiter_notify( $io_context & ctx ) {
461	if(!empty( ctx.arbiter->pending )) {
462	__ioarbiter_notify( *ctx.arbiter, &ctx );
463	}
464	}
465
466	// Simply append to the pending
467	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
468	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
469
470	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
471
472	__external_io ei;
473	ei.idxs = idxs;
474	ei.have = have;
475	ei.lazy = lazy;
476
477	block(ctx->ext_sq, (__outstanding_io&)ei);
478
479	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
480	}
481
482	static void __ioarbiter_flush( $io_context & ctx ) {
483	if(!empty( ctx.ext_sq )) {
484	__STATS__( false, io.flush.external += 1; )
485
486	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
487
488	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
489	{
490	while( !empty(ctx.ext_sq.queue) ) {
491	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
492
493	__submit(&ctx, ei.idxs, ei.have, ei.lazy);
494
495	post( ei.sem );
496	}
497
498	ctx.ext_sq.empty = true;
499	}
500	unlock(ctx.ext_sq.lock );
501	}
502	}
503
504	bool __kernel_read(processor * proc, io_future_t & future, char buf[], int fd) {
505	$io_context * ctx = proc->io.ctx;
506	/* paranoid */ verify( ! __preemption_enabled() );
507	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
508	/* paranoid */ verify( ctx );
509
510	__u32 idx;
511	struct io_uring_sqe * sqe;
512
513	// We can proceed to the fast path
514	if( !__alloc(ctx, &idx, 1) ) return false;
515
516	// Allocation was successful
517	__fill( &sqe, 1, &idx, ctx );
518
519	sqe->opcode = IORING_OP_READ;
520	sqe->user_data = (uintptr_t)&future;
521	sqe->flags = 0;
522	sqe->ioprio = 0;
523	sqe->fd = 0;
524	sqe->off = 0;
525	sqe->fsync_flags = 0;
526	sqe->__pad2[0] = 0;
527	sqe->__pad2[1] = 0;
528	sqe->__pad2[2] = 0;
529	sqe->addr = (uintptr_t)buf;
530	sqe->len = sizeof(uint64_t);
531
532	asm volatile("": : :"memory");
533
534	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
535	__submit( ctx, &idx, 1, true );
536
537	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
538	/* paranoid */ verify( ! __preemption_enabled() );
539	}
540	#endif

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

Download in other formats:

Original Format