Context Navigation

io.cfa @ f302d80

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum

Last change on this file since f302d80 was e9c0b4c, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
I/O drain now keeps the schedlock for the duration of the call. (Rather than locking/unlocking for each thread).
Property mode set to `100644`
File size: 15.1 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
18	#if defined(__CFA_DEBUG__)
19	// #define __CFA_DEBUG_PRINT_IO__
20	// #define __CFA_DEBUG_PRINT_IO_CORE__
21	#endif
22
23
24	#if defined(CFA_HAVE_LINUX_IO_URING_H)
25	#define _GNU_SOURCE /* See feature_test_macros(7) */
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 );
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 ) );
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	void __cfa_io_flush( processor * proc ) {
136	/* paranoid */ verify( ! __preemption_enabled() );
137	/* paranoid */ verify( proc );
138	/* paranoid */ verify( proc->io.ctx );
139
140	$io_context & ctx = *proc->io.ctx;
141
142	__ioarbiter_flush( ctx );
143
144	__STATS__( true, io.calls.flush++; )
145	int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, 0, 0, (sigset_t *)0p, _NSIG / 8);
146	if( ret < 0 ) {
147	switch((int)errno) {
148	case EAGAIN:
149	case EINTR:
150	case EBUSY:
151	// Update statistics
152	__STATS__( false, io.calls.errors.busy ++; )
153	return;
154	default:
155	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
156	}
157	}
158
159	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
160	__STATS__( true, io.calls.submitted += ret; )
161	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
162	/* paranoid */ verify( ctx.sq.to_submit >= ret );
163
164	ctx.sq.to_submit -= ret;
165
166	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
167
168	// Release the consumed SQEs
169	__release_sqes( ctx );
170
171	/* paranoid */ verify( ! __preemption_enabled() );
172
173	ctx.proc->io.pending = false;
174	}
175
176	//=============================================================================================
177	// I/O Submissions
178	//=============================================================================================
179
180	// Submition steps :
181	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
182	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
183	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
184	// need to write an allocator that allows allocating concurrently.
185	//
186	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
187	//
188	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
189	// needs to arrive to two concensus at the same time:
190	// A - The order in which entries are listed in the array: no two threads must pick the
191	// same index for their entries
192	// B - When can the tail be update for the kernel. EVERY entries in the array between
193	// head and tail must be fully filled and shouldn't ever be touched again.
194	//
195	//=============================================================================================
196	// Allocation
197	// for user's convenience fill the sqes from the indexes
198	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
199	struct io_uring_sqe * sqes = ctx->sq.sqes;
200	for(i; want) {
201	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
202	out_sqes[i] = &sqes[idxs[i]];
203	}
204	}
205
206	// Try to directly allocate from the a given context
207	// Not thread-safe
208	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
209	__sub_ring_t & sq = ctx->sq;
210	const __u32 mask = *sq.mask;
211	__u32 fhead = sq.free_ring.head; // get the current head of the queue
212	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
213
214	// If we don't have enough sqes, fail
215	if((ftail - fhead) < want) { return false; }
216
217	// copy all the indexes we want from the available list
218	for(i; want) {
219	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
220	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
221	}
222
223	// Advance the head to mark the indexes as consumed
224	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
225
226	// return success
227	return true;
228	}
229
230	// Allocate an submit queue entry.
231	// The kernel cannot see these entries until they are submitted, but other threads must be
232	// able to see which entries can be used and which are already un used by an other thread
233	// for convenience, return both the index and the pointer to the sqe
234	// sqe == &sqes[idx]
235	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
236	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
237
238	disable_interrupts();
239	processor * proc = __cfaabi_tls.this_processor;
240	$io_context * ctx = proc->io.ctx;
241	/* paranoid */ verify( __cfaabi_tls.this_processor );
242	/* paranoid */ verify( ctx );
243
244	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
245
246	// We can proceed to the fast path
247	if( __alloc(ctx, idxs, want) ) {
248	// Allocation was successful
249	__STATS__( true, io.alloc.fast += 1; )
250	enable_interrupts();
251
252	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
253
254	__fill( sqes, want, idxs, ctx );
255	return ctx;
256	}
257	// The fast path failed, fallback
258	__STATS__( true, io.alloc.fail += 1; )
259
260	// Fast path failed, fallback on arbitration
261	__STATS__( true, io.alloc.slow += 1; )
262	enable_interrupts();
263
264	$io_arbiter * ioarb = proc->cltr->io.arbiter;
265	/* paranoid */ verify( ioarb );
266
267	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
268
269	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
270
271	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
272
273	__fill( sqes, want, idxs,ret );
274	return ret;
275	}
276
277
278	//=============================================================================================
279	// submission
280	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
281	// We can proceed to the fast path
282	// Get the right objects
283	__sub_ring_t & sq = ctx->sq;
284	const __u32 mask = *sq.mask;
285	__u32 tail = *sq.kring.tail;
286
287	// Add the sqes to the array
288	for( i; have ) {
289	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
290	sq.kring.array[ (tail + i) & mask ] = idxs[i];
291	}
292
293	// Make the sqes visible to the submitter
294	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
295	sq.to_submit++;
296
297	ctx->proc->io.pending = true;
298	ctx->proc->io.dirty = true;
299	if(sq.to_submit > 30 \|\| !lazy) {
300	__cfa_io_flush( ctx->proc );
301	}
302	}
303
304	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
305	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
306
307	disable_interrupts();
308	processor * proc = __cfaabi_tls.this_processor;
309	$io_context * ctx = proc->io.ctx;
310	/* paranoid */ verify( __cfaabi_tls.this_processor );
311	/* paranoid */ verify( ctx );
312
313	// Can we proceed to the fast path
314	if( ctx == inctx ) // We have the right instance?
315	{
316	__submit(ctx, idxs, have, lazy);
317
318	// Mark the instance as no longer in-use, re-enable interrupts and return
319	__STATS__( true, io.submit.fast += 1; )
320	enable_interrupts();
321
322	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
323	return;
324	}
325
326	// Fast path failed, fallback on arbitration
327	__STATS__( true, io.submit.slow += 1; )
328	enable_interrupts();
329
330	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
331
332	__ioarbiter_submit(inctx, idxs, have, lazy);
333	}
334
335	//=============================================================================================
336	// Flushing
337	// Go through the ring's submit queue and release everything that has already been consumed
338	// by io_uring
339	// This cannot be done by multiple threads
340	static __u32 __release_sqes( struct $io_context & ctx ) {
341	const __u32 mask = *ctx.sq.mask;
342
343	__attribute__((unused))
344	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
345	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
346	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
347
348	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
349
350	// the 3 fields are organized like this diagram
351	// except it's are ring
352	// ---+--------+--------+----
353	// ---+--------+--------+----
354	// ^ ^ ^
355	// phead chead ctail
356
357	// make sure ctail doesn't wrap around and reach phead
358	/* paranoid */ verify(
359	(ctail >= chead && chead >= phead)
360	\|\| (chead >= phead && phead >= ctail)
361	\|\| (phead >= ctail && ctail >= chead)
362	);
363
364	// find the range we need to clear
365	__u32 count = chead - phead;
366
367	if(count == 0) {
368	return 0;
369	}
370
371	// We acquired an previous-head/current-head range
372	// go through the range and release the sqes
373	for( i; count ) {
374	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
375	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
376	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
377	}
378
379	ctx.sq.kring.released = chead; // note up to were we processed
380	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
381
382	__ioarbiter_notify(ctx);
383
384	return count;
385	}
386
387	//=============================================================================================
388	// I/O Arbiter
389	//=============================================================================================
390	static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
391	// Lock the list, it's not thread safe
392	lock( queue.lock __cfaabi_dbg_ctx2 );
393	{
394	// Add our request to the list
395	add( queue.queue, item );
396
397	// Mark as pending
398	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
399	}
400	unlock( queue.lock );
401
402	wait( item.sem );
403	}
404
405	static inline bool empty(__outstanding_io_queue & queue ) {
406	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
407	}
408
409	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
410	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
411
412	__STATS__( false, io.alloc.block += 1; )
413
414	// No one has any resources left, wait for something to finish
415	// We need to add ourself to a list of pending allocs and wait for an answer
416	__pending_alloc pa;
417	pa.idxs = idxs;
418	pa.want = want;
419
420	block(this.pending, (__outstanding_io&)pa);
421
422	return pa.ctx;
423
424	}
425
426	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
427	/* paranoid */ verify( !empty(this.pending.queue) );
428
429	lock( this.pending.lock __cfaabi_dbg_ctx2 );
430	{
431	while( !empty(this.pending.queue) ) {
432	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
433	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
434	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
435
436	if( have > pa.want ) goto DONE;
437	drop( this.pending.queue );
438
439	/* paranoid */__attribute__((unused)) bool ret =
440
441	__alloc(ctx, pa.idxs, pa.want);
442
443	/* paranoid */ verify( ret );
444
445	pa.ctx = ctx;
446
447	post( pa.sem );
448	}
449
450	this.pending.empty = true;
451	DONE:;
452	}
453	unlock( this.pending.lock );
454	}
455
456	static void __ioarbiter_notify( $io_context & ctx ) {
457	if(!empty( ctx.arbiter->pending )) {
458	__ioarbiter_notify( *ctx.arbiter, &ctx );
459	}
460	}
461
462	// Simply append to the pending
463	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
464	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
465
466	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
467
468	__external_io ei;
469	ei.idxs = idxs;
470	ei.have = have;
471	ei.lazy = lazy;
472
473	block(ctx->ext_sq, (__outstanding_io&)ei);
474
475	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
476	}
477
478	static void __ioarbiter_flush( $io_context & ctx ) {
479	if(!empty( ctx.ext_sq )) {
480	__STATS__( false, io.flush.external += 1; )
481
482	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
483
484	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
485	{
486	while( !empty(ctx.ext_sq.queue) ) {
487	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
488
489	__submit(&ctx, ei.idxs, ei.have, ei.lazy);
490
491	post( ei.sem );
492	}
493
494	ctx.ext_sq.empty = true;
495	}
496	unlock(ctx.ext_sq.lock );
497	}
498	}
499	#endif

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

Download in other formats:

Original Format