Context Navigation

io.cfa @ 2432e8e

ADTast-experimentalenumpthread-emulationqualifiedEnum

Last change on this file since 2432e8e was 2432e8e, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Fixed deadlock on io slowpath due to spin lock recursion.
Property mode set to `100644`
File size: 17.0 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	if(ctx.sq.to_submit != 0 \|\| min_comp > 0) {
147
148	__STATS__( true, io.calls.flush++; )
149	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);
150	if( ret < 0 ) {
151	switch((int)errno) {
152	case EAGAIN:
153	case EINTR:
154	case EBUSY:
155	// Update statistics
156	__STATS__( false, io.calls.errors.busy ++; )
157	return false;
158	default:
159	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
160	}
161	}
162
163	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
164	__STATS__( true, io.calls.submitted += ret; )
165	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
166	/* paranoid */ verify( ctx.sq.to_submit >= ret );
167
168	ctx.sq.to_submit -= ret;
169
170	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
171
172	// Release the consumed SQEs
173	__release_sqes( ctx );
174
175	/* paranoid */ verify( ! __preemption_enabled() );
176
177	ctx.proc->io.pending = false;
178	}
179
180	ready_schedule_lock();
181	bool ret = __cfa_io_drain( proc );
182	ready_schedule_unlock();
183	return ret;
184	}
185
186	//=============================================================================================
187	// I/O Submissions
188	//=============================================================================================
189
190	// Submition steps :
191	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
192	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
193	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
194	// need to write an allocator that allows allocating concurrently.
195	//
196	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
197	//
198	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
199	// needs to arrive to two concensus at the same time:
200	// A - The order in which entries are listed in the array: no two threads must pick the
201	// same index for their entries
202	// B - When can the tail be update for the kernel. EVERY entries in the array between
203	// head and tail must be fully filled and shouldn't ever be touched again.
204	//
205	//=============================================================================================
206	// Allocation
207	// for user's convenience fill the sqes from the indexes
208	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
209	struct io_uring_sqe * sqes = ctx->sq.sqes;
210	for(i; want) {
211	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
212	out_sqes[i] = &sqes[idxs[i]];
213	}
214	}
215
216	// Try to directly allocate from the a given context
217	// Not thread-safe
218	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
219	__sub_ring_t & sq = ctx->sq;
220	const __u32 mask = *sq.mask;
221	__u32 fhead = sq.free_ring.head; // get the current head of the queue
222	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
223
224	// If we don't have enough sqes, fail
225	if((ftail - fhead) < want) { return false; }
226
227	// copy all the indexes we want from the available list
228	for(i; want) {
229	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
230	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
231	}
232
233	// Advance the head to mark the indexes as consumed
234	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
235
236	// return success
237	return true;
238	}
239
240	// Allocate an submit queue entry.
241	// The kernel cannot see these entries until they are submitted, but other threads must be
242	// able to see which entries can be used and which are already un used by an other thread
243	// for convenience, return both the index and the pointer to the sqe
244	// sqe == &sqes[idx]
245	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
246	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
247
248	disable_interrupts();
249	processor * proc = __cfaabi_tls.this_processor;
250	$io_context * ctx = proc->io.ctx;
251	/* paranoid */ verify( __cfaabi_tls.this_processor );
252	/* paranoid */ verify( ctx );
253
254	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
255
256	// We can proceed to the fast path
257	if( __alloc(ctx, idxs, want) ) {
258	// Allocation was successful
259	__STATS__( true, io.alloc.fast += 1; )
260	enable_interrupts();
261
262	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
263
264	__fill( sqes, want, idxs, ctx );
265	return ctx;
266	}
267	// The fast path failed, fallback
268	__STATS__( true, io.alloc.fail += 1; )
269
270	// Fast path failed, fallback on arbitration
271	__STATS__( true, io.alloc.slow += 1; )
272	enable_interrupts();
273
274	$io_arbiter * ioarb = proc->cltr->io.arbiter;
275	/* paranoid */ verify( ioarb );
276
277	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
278
279	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
280
281	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
282
283	__fill( sqes, want, idxs,ret );
284	return ret;
285	}
286
287	//=============================================================================================
288	// submission
289	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
290	// We can proceed to the fast path
291	// Get the right objects
292	__sub_ring_t & sq = ctx->sq;
293	const __u32 mask = *sq.mask;
294	__u32 tail = *sq.kring.tail;
295
296	// Add the sqes to the array
297	for( i; have ) {
298	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
299	sq.kring.array[ (tail + i) & mask ] = idxs[i];
300	}
301
302	// Make the sqes visible to the submitter
303	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
304	sq.to_submit += have;
305
306	ctx->proc->io.pending = true;
307	ctx->proc->io.dirty = true;
308	}
309
310	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
311	__sub_ring_t & sq = ctx->sq;
312	__submit_only(ctx, idxs, have);
313
314	if(sq.to_submit > 30) {
315	__tls_stats()->io.flush.full++;
316	__cfa_io_flush( ctx->proc, 0 );
317	}
318	if(!lazy) {
319	__tls_stats()->io.flush.eager++;
320	__cfa_io_flush( ctx->proc, 0 );
321	}
322	}
323
324	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
325	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
326
327	disable_interrupts();
328	processor * proc = __cfaabi_tls.this_processor;
329	$io_context * ctx = proc->io.ctx;
330	/* paranoid */ verify( __cfaabi_tls.this_processor );
331	/* paranoid */ verify( ctx );
332
333	// Can we proceed to the fast path
334	if( ctx == inctx ) // We have the right instance?
335	{
336	__submit(ctx, idxs, have, lazy);
337
338	// Mark the instance as no longer in-use, re-enable interrupts and return
339	__STATS__( true, io.submit.fast += 1; )
340	enable_interrupts();
341
342	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
343	return;
344	}
345
346	// Fast path failed, fallback on arbitration
347	__STATS__( true, io.submit.slow += 1; )
348	enable_interrupts();
349
350	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
351
352	__ioarbiter_submit(inctx, idxs, have, lazy);
353	}
354
355	//=============================================================================================
356	// Flushing
357	// Go through the ring's submit queue and release everything that has already been consumed
358	// by io_uring
359	// This cannot be done by multiple threads
360	static __u32 __release_sqes( struct $io_context & ctx ) {
361	const __u32 mask = *ctx.sq.mask;
362
363	__attribute__((unused))
364	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
365	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
366	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
367
368	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
369
370	// the 3 fields are organized like this diagram
371	// except it's are ring
372	// ---+--------+--------+----
373	// ---+--------+--------+----
374	// ^ ^ ^
375	// phead chead ctail
376
377	// make sure ctail doesn't wrap around and reach phead
378	/* paranoid */ verify(
379	(ctail >= chead && chead >= phead)
380	\|\| (chead >= phead && phead >= ctail)
381	\|\| (phead >= ctail && ctail >= chead)
382	);
383
384	// find the range we need to clear
385	__u32 count = chead - phead;
386
387	if(count == 0) {
388	return 0;
389	}
390
391	// We acquired an previous-head/current-head range
392	// go through the range and release the sqes
393	for( i; count ) {
394	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
395	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
396	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
397	}
398
399	ctx.sq.kring.released = chead; // note up to were we processed
400	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
401
402	__ioarbiter_notify(ctx);
403
404	return count;
405	}
406
407	//=============================================================================================
408	// I/O Arbiter
409	//=============================================================================================
410	static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
411	// Lock the list, it's not thread safe
412	lock( queue.lock __cfaabi_dbg_ctx2 );
413	{
414	// Add our request to the list
415	add( queue.queue, item );
416
417	// Mark as pending
418	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
419	}
420	unlock( queue.lock );
421
422	wait( item.sem );
423	}
424
425	static inline bool empty(__outstanding_io_queue & queue ) {
426	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
427	}
428
429	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
430	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
431
432	__STATS__( false, io.alloc.block += 1; )
433
434	// No one has any resources left, wait for something to finish
435	// We need to add ourself to a list of pending allocs and wait for an answer
436	__pending_alloc pa;
437	pa.idxs = idxs;
438	pa.want = want;
439
440	block(this.pending, (__outstanding_io&)pa);
441
442	return pa.ctx;
443
444	}
445
446	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
447	/* paranoid */ verify( !empty(this.pending.queue) );
448
449	lock( this.pending.lock __cfaabi_dbg_ctx2 );
450	{
451	while( !empty(this.pending.queue) ) {
452	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
453	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
454	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
455
456	if( have > pa.want ) goto DONE;
457	drop( this.pending.queue );
458
459	/* paranoid */__attribute__((unused)) bool ret =
460
461	__alloc(ctx, pa.idxs, pa.want);
462
463	/* paranoid */ verify( ret );
464
465	pa.ctx = ctx;
466
467	post( pa.sem );
468	}
469
470	this.pending.empty = true;
471	DONE:;
472	}
473	unlock( this.pending.lock );
474	}
475
476	static void __ioarbiter_notify( $io_context & ctx ) {
477	if(!empty( ctx.arbiter->pending )) {
478	__ioarbiter_notify( *ctx.arbiter, &ctx );
479	}
480	}
481
482	// Simply append to the pending
483	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
484	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
485
486	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
487
488	__external_io ei;
489	ei.idxs = idxs;
490	ei.have = have;
491	ei.lazy = lazy;
492
493	block(ctx->ext_sq, (__outstanding_io&)ei);
494
495	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
496	}
497
498	static void __ioarbiter_flush( $io_context & ctx ) {
499	if(!empty( ctx.ext_sq )) {
500	__STATS__( false, io.flush.external += 1; )
501
502	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
503
504	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
505	{
506	while( !empty(ctx.ext_sq.queue) ) {
507	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
508
509	__submit_only(&ctx, ei.idxs, ei.have);
510
511	post( ei.sem );
512	}
513
514	ctx.ext_sq.empty = true;
515	}
516	unlock(ctx.ext_sq.lock );
517	}
518	}
519
520	#if defined(CFA_WITH_IO_URING_IDLE)
521	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
522	$io_context * ctx = proc->io.ctx;
523	/* paranoid */ verify( ! __preemption_enabled() );
524	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
525	/* paranoid */ verify( ctx );
526
527	__u32 idx;
528	struct io_uring_sqe * sqe;
529
530	// We can proceed to the fast path
531	if( !__alloc(ctx, &idx, 1) ) return false;
532
533	// Allocation was successful
534	__fill( &sqe, 1, &idx, ctx );
535
536	sqe->user_data = (uintptr_t)&future;
537	sqe->flags = 0;
538	sqe->fd = fd;
539	sqe->off = 0;
540	sqe->ioprio = 0;
541	sqe->fsync_flags = 0;
542	sqe->__pad2[0] = 0;
543	sqe->__pad2[1] = 0;
544	sqe->__pad2[2] = 0;
545
546	#if defined(CFA_HAVE_IORING_OP_READ)
547	sqe->opcode = IORING_OP_READ;
548	sqe->addr = (uint64_t)iov.iov_base;
549	sqe->len = iov.iov_len;
550	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
551	sqe->opcode = IORING_OP_READV;
552	sqe->addr = (uintptr_t)&iov;
553	sqe->len = 1;
554	#else
555	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
556	#endif
557
558	asm volatile("": : :"memory");
559
560	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
561	__submit( ctx, &idx, 1, true );
562
563	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
564	/* paranoid */ verify( ! __preemption_enabled() );
565
566	return true;
567	}
568	#endif
569	#endif

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

Download in other formats:

Original Format