Context Navigation

io.cfa @ e8ac228

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since e8ac228 was e8ac228, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Fix major bug where to few items were submitted.
Property mode set to `100644`
File size: 15.5 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	void __cfa_io_flush( processor * proc ) {
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	// for(i; 2) {
144	// unsigned idx = proc->rdq.id + i;
145	// cltr->ready_queue.lanes.tscs[idx].tv = -1ull;
146	// }
147
148	__ioarbiter_flush( ctx );
149
150	__STATS__( true, io.calls.flush++; )
151	int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, 0, 0, (sigset_t *)0p, _NSIG / 8);
152	if( ret < 0 ) {
153	switch((int)errno) {
154	case EAGAIN:
155	case EINTR:
156	case EBUSY:
157	// Update statistics
158	__STATS__( false, io.calls.errors.busy ++; )
159	// for(i; 2) {
160	// unsigned idx = proc->rdq.id + i;
161	// cltr->ready_queue.lanes.tscs[idx].tv = rdtscl();
162	// }
163	return;
164	default:
165	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
166	}
167	}
168
169	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
170	__STATS__( true, io.calls.submitted += ret; )
171	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
172	/* paranoid */ verify( ctx.sq.to_submit >= ret );
173
174	ctx.sq.to_submit -= ret;
175
176	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
177
178	// Release the consumed SQEs
179	__release_sqes( ctx );
180
181	/* paranoid */ verify( ! __preemption_enabled() );
182
183	ctx.proc->io.pending = false;
184
185	__cfa_io_drain( proc );
186	// for(i; 2) {
187	// unsigned idx = proc->rdq.id + i;
188	// cltr->ready_queue.lanes.tscs[idx].tv = rdtscl();
189	// }
190	}
191
192	//=============================================================================================
193	// I/O Submissions
194	//=============================================================================================
195
196	// Submition steps :
197	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
198	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
199	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
200	// need to write an allocator that allows allocating concurrently.
201	//
202	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
203	//
204	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
205	// needs to arrive to two concensus at the same time:
206	// A - The order in which entries are listed in the array: no two threads must pick the
207	// same index for their entries
208	// B - When can the tail be update for the kernel. EVERY entries in the array between
209	// head and tail must be fully filled and shouldn't ever be touched again.
210	//
211	//=============================================================================================
212	// Allocation
213	// for user's convenience fill the sqes from the indexes
214	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
215	struct io_uring_sqe * sqes = ctx->sq.sqes;
216	for(i; want) {
217	__cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
218	out_sqes[i] = &sqes[idxs[i]];
219	}
220	}
221
222	// Try to directly allocate from the a given context
223	// Not thread-safe
224	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
225	__sub_ring_t & sq = ctx->sq;
226	const __u32 mask = *sq.mask;
227	__u32 fhead = sq.free_ring.head; // get the current head of the queue
228	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
229
230	// If we don't have enough sqes, fail
231	if((ftail - fhead) < want) { return false; }
232
233	// copy all the indexes we want from the available list
234	for(i; want) {
235	__cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
236	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
237	}
238
239	// Advance the head to mark the indexes as consumed
240	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
241
242	// return success
243	return true;
244	}
245
246	// Allocate an submit queue entry.
247	// The kernel cannot see these entries until they are submitted, but other threads must be
248	// able to see which entries can be used and which are already un used by an other thread
249	// for convenience, return both the index and the pointer to the sqe
250	// sqe == &sqes[idx]
251	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
252	__cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
253
254	disable_interrupts();
255	processor * proc = __cfaabi_tls.this_processor;
256	$io_context * ctx = proc->io.ctx;
257	/* paranoid */ verify( __cfaabi_tls.this_processor );
258	/* paranoid */ verify( ctx );
259
260	__cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
261
262	// We can proceed to the fast path
263	if( __alloc(ctx, idxs, want) ) {
264	// Allocation was successful
265	__STATS__( true, io.alloc.fast += 1; )
266	enable_interrupts();
267
268	__cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
269
270	__fill( sqes, want, idxs, ctx );
271	return ctx;
272	}
273	// The fast path failed, fallback
274	__STATS__( true, io.alloc.fail += 1; )
275
276	// Fast path failed, fallback on arbitration
277	__STATS__( true, io.alloc.slow += 1; )
278	enable_interrupts();
279
280	$io_arbiter * ioarb = proc->cltr->io.arbiter;
281	/* paranoid */ verify( ioarb );
282
283	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
284
285	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
286
287	__cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
288
289	__fill( sqes, want, idxs,ret );
290	return ret;
291	}
292
293
294	//=============================================================================================
295	// submission
296	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
297	// We can proceed to the fast path
298	// Get the right objects
299	__sub_ring_t & sq = ctx->sq;
300	const __u32 mask = *sq.mask;
301	__u32 tail = *sq.kring.tail;
302
303	// Add the sqes to the array
304	for( i; have ) {
305	__cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
306	sq.kring.array[ (tail + i) & mask ] = idxs[i];
307	}
308
309	// Make the sqes visible to the submitter
310	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
311	sq.to_submit += have;
312
313	ctx->proc->io.pending = true;
314	ctx->proc->io.dirty = true;
315	if(sq.to_submit > 30 \|\| !lazy) {
316	ready_schedule_lock();
317	__cfa_io_flush( ctx->proc );
318	ready_schedule_unlock();
319	}
320	}
321
322	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
323	__cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
324
325	disable_interrupts();
326	processor * proc = __cfaabi_tls.this_processor;
327	$io_context * ctx = proc->io.ctx;
328	/* paranoid */ verify( __cfaabi_tls.this_processor );
329	/* paranoid */ verify( ctx );
330
331	// Can we proceed to the fast path
332	if( ctx == inctx ) // We have the right instance?
333	{
334	__submit(ctx, idxs, have, lazy);
335
336	// Mark the instance as no longer in-use, re-enable interrupts and return
337	__STATS__( true, io.submit.fast += 1; )
338	enable_interrupts();
339
340	__cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
341	return;
342	}
343
344	// Fast path failed, fallback on arbitration
345	__STATS__( true, io.submit.slow += 1; )
346	enable_interrupts();
347
348	__cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
349
350	__ioarbiter_submit(inctx, idxs, have, lazy);
351	}
352
353	//=============================================================================================
354	// Flushing
355	// Go through the ring's submit queue and release everything that has already been consumed
356	// by io_uring
357	// This cannot be done by multiple threads
358	static __u32 __release_sqes( struct $io_context & ctx ) {
359	const __u32 mask = *ctx.sq.mask;
360
361	__attribute__((unused))
362	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
363	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
364	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
365
366	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
367
368	// the 3 fields are organized like this diagram
369	// except it's are ring
370	// ---+--------+--------+----
371	// ---+--------+--------+----
372	// ^ ^ ^
373	// phead chead ctail
374
375	// make sure ctail doesn't wrap around and reach phead
376	/* paranoid */ verify(
377	(ctail >= chead && chead >= phead)
378	\|\| (chead >= phead && phead >= ctail)
379	\|\| (phead >= ctail && ctail >= chead)
380	);
381
382	// find the range we need to clear
383	__u32 count = chead - phead;
384
385	if(count == 0) {
386	return 0;
387	}
388
389	// We acquired an previous-head/current-head range
390	// go through the range and release the sqes
391	for( i; count ) {
392	__cfadbg_print_safe(io, "Kernel I/O : release loop\n");
393	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
394	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
395	}
396
397	ctx.sq.kring.released = chead; // note up to were we processed
398	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
399
400	__ioarbiter_notify(ctx);
401
402	return count;
403	}
404
405	//=============================================================================================
406	// I/O Arbiter
407	//=============================================================================================
408	static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
409	// Lock the list, it's not thread safe
410	lock( queue.lock __cfaabi_dbg_ctx2 );
411	{
412	// Add our request to the list
413	add( queue.queue, item );
414
415	// Mark as pending
416	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
417	}
418	unlock( queue.lock );
419
420	wait( item.sem );
421	}
422
423	static inline bool empty(__outstanding_io_queue & queue ) {
424	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
425	}
426
427	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
428	__cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
429
430	__STATS__( false, io.alloc.block += 1; )
431
432	// No one has any resources left, wait for something to finish
433	// We need to add ourself to a list of pending allocs and wait for an answer
434	__pending_alloc pa;
435	pa.idxs = idxs;
436	pa.want = want;
437
438	block(this.pending, (__outstanding_io&)pa);
439
440	return pa.ctx;
441
442	}
443
444	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
445	/* paranoid */ verify( !empty(this.pending.queue) );
446
447	lock( this.pending.lock __cfaabi_dbg_ctx2 );
448	{
449	while( !empty(this.pending.queue) ) {
450	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
451	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
452	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
453
454	if( have > pa.want ) goto DONE;
455	drop( this.pending.queue );
456
457	/* paranoid */__attribute__((unused)) bool ret =
458
459	__alloc(ctx, pa.idxs, pa.want);
460
461	/* paranoid */ verify( ret );
462
463	pa.ctx = ctx;
464
465	post( pa.sem );
466	}
467
468	this.pending.empty = true;
469	DONE:;
470	}
471	unlock( this.pending.lock );
472	}
473
474	static void __ioarbiter_notify( $io_context & ctx ) {
475	if(!empty( ctx.arbiter->pending )) {
476	__ioarbiter_notify( *ctx.arbiter, &ctx );
477	}
478	}
479
480	// Simply append to the pending
481	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
482	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
483
484	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
485
486	__external_io ei;
487	ei.idxs = idxs;
488	ei.have = have;
489	ei.lazy = lazy;
490
491	block(ctx->ext_sq, (__outstanding_io&)ei);
492
493	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
494	}
495
496	static void __ioarbiter_flush( $io_context & ctx ) {
497	if(!empty( ctx.ext_sq )) {
498	__STATS__( false, io.flush.external += 1; )
499
500	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
501
502	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
503	{
504	while( !empty(ctx.ext_sq.queue) ) {
505	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
506
507	__submit(&ctx, ei.idxs, ei.have, ei.lazy);
508
509	post( ei.sem );
510	}
511
512	ctx.ext_sq.empty = true;
513	}
514	unlock(ctx.ext_sq.lock );
515	}
516	}
517	#endif

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

Download in other formats:

Original Format