Context Navigation

io.cfa@ 77adaee

Visit:

ADT ast-experimental enum pthread-emulation qualifiedEnum

Last change on this file since 77adaee was 729df21, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Merge branch 'master' of plg.uwaterloo.ca:software/cfa/cfa-cc
Property mode set to `100644`
File size: 21.3 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 "kernel/cluster.hfa"
45	#include "io/types.hfa"
46
47	__attribute__((unused)) static const char * opcodes[] = {
48	"OP_NOP",
49	"OP_READV",
50	"OP_WRITEV",
51	"OP_FSYNC",
52	"OP_READ_FIXED",
53	"OP_WRITE_FIXED",
54	"OP_POLL_ADD",
55	"OP_POLL_REMOVE",
56	"OP_SYNC_FILE_RANGE",
57	"OP_SENDMSG",
58	"OP_RECVMSG",
59	"OP_TIMEOUT",
60	"OP_TIMEOUT_REMOVE",
61	"OP_ACCEPT",
62	"OP_ASYNC_CANCEL",
63	"OP_LINK_TIMEOUT",
64	"OP_CONNECT",
65	"OP_FALLOCATE",
66	"OP_OPENAT",
67	"OP_CLOSE",
68	"OP_FILES_UPDATE",
69	"OP_STATX",
70	"OP_READ",
71	"OP_WRITE",
72	"OP_FADVISE",
73	"OP_MADVISE",
74	"OP_SEND",
75	"OP_RECV",
76	"OP_OPENAT2",
77	"OP_EPOLL_CTL",
78	"OP_SPLICE",
79	"OP_PROVIDE_BUFFERS",
80	"OP_REMOVE_BUFFERS",
81	"OP_TEE",
82	"INVALID_OP"
83	};
84
85	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want );
86	static void __ioarbiter_submit( $io_context * , __u32 idxs[], __u32 have, bool lazy );
87	static void __ioarbiter_flush ( $io_context & );
88	static inline void __ioarbiter_notify( $io_context & ctx );
89	//=============================================================================================
90	// I/O Polling
91	//=============================================================================================
92	static inline unsigned __flush( struct $io_context & );
93	static inline __u32 __release_sqes( struct $io_context & );
94	extern void __kernel_unpark( thread$ * thrd, unpark_hint );
95
96	static void ioring_syscsll( struct $io_context & ctx, unsigned int min_comp, unsigned int flags ) {
97	__STATS__( true, io.calls.flush++; )
98	int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, min_comp, flags, (sigset_t *)0p, _NSIG / 8);
99	if( ret < 0 ) {
100	switch((int)errno) {
101	case EAGAIN:
102	case EINTR:
103	case EBUSY:
104	// Update statistics
105	__STATS__( false, io.calls.errors.busy ++; )
106	return false;
107	default:
108	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
109	}
110	}
111
112	__cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
113	__STATS__( true, io.calls.submitted += ret; )
114	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
115	/* paranoid */ verify( ctx.sq.to_submit >= ret );
116
117	ctx.sq.to_submit -= ret;
118
119	/* paranoid / verify( ctx.sq.to_submit <= ctx.sq.num );
120
121	// Release the consumed SQEs
122	__release_sqes( ctx );
123
124	/* paranoid */ verify( ! __preemption_enabled() );
125
126	__atomic_store_n(&ctx.proc->io.pending, false, __ATOMIC_RELAXED);
127	}
128
129	static bool try_acquire( $io_context * ctx ) __attribute__((nonnull(1))) {
130	/* paranoid */ verify( ! __preemption_enabled() );
131	/* paranoid */ verify( ready_schedule_islocked() );
132
133
134	{
135	const __u32 head = *ctx->cq.head;
136	const __u32 tail = *ctx->cq.tail;
137
138	if(head == tail) return false;
139	}
140
141	// Drain the queue
142	if(!__atomic_try_acquire(&ctx->cq.lock)) {
143	__STATS__( false, io.calls.locked++; )
144	return false;
145	}
146
147	return true;
148	}
149
150	static bool __cfa_do_drain( $io_context * ctx, cluster * cltr ) __attribute__((nonnull(1, 2))) {
151	/* paranoid */ verify( ! __preemption_enabled() );
152	/* paranoid */ verify( ready_schedule_islocked() );
153	/* paranoid */ verify( ctx->cq.lock == true );
154
155	const __u32 mask = *ctx->cq.mask;
156	unsigned long long ts_prev = ctx->cq.ts;
157
158	// re-read the head and tail in case it already changed.
159	const __u32 head = *ctx->cq.head;
160	const __u32 tail = *ctx->cq.tail;
161	const __u32 count = tail - head;
162	__STATS__( false, io.calls.drain++; io.calls.completed += count; )
163
164	for(i; count) {
165	unsigned idx = (head + i) & mask;
166	volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
167
168	/* paranoid */ verify(&cqe);
169
170	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
171	// __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
172
173	__kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL );
174	}
175
176	unsigned long long ts_next = ctx->cq.ts = rdtscl();
177
178	// Mark to the kernel that the cqe has been seen
179	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
180	__atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
181	ctx->proc->idle_wctx.drain_time = ts_next;
182
183	__cfadbg_print_safe(io, "Kernel I/O : %u completed age %llu\n", count, ts_next);
184	/* paranoid */ verify( ready_schedule_islocked() );
185	/* paranoid */ verify( ! __preemption_enabled() );
186
187	__atomic_unlock(&ctx->cq.lock);
188
189	touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next );
190
191	return true;
192	}
193
194	bool __cfa_io_drain( processor * proc ) {
195	bool local = false;
196	bool remote = false;
197
198	ready_schedule_lock();
199
200	cluster * const cltr = proc->cltr;
201	$io_context * const ctx = proc->io.ctx;
202	/* paranoid */ verify( cltr );
203	/* paranoid */ verify( ctx );
204
205	with(cltr->sched) {
206	const size_t ctxs_count = io.count;
207
208	/* paranoid */ verify( ready_schedule_islocked() );
209	/* paranoid */ verify( ! __preemption_enabled() );
210	/* paranoid */ verify( active_processor() == proc );
211	/* paranoid */ verify( __shard_factor.io > 0 );
212	/* paranoid */ verify( ctxs_count > 0 );
213	/* paranoid */ verify( ctx->cq.id < ctxs_count );
214
215	const unsigned this_cache = cache_id(cltr, ctx->cq.id / __shard_factor.io);
216	const unsigned long long ctsc = rdtscl();
217
218	if(proc->io.target == MAX) {
219	uint64_t chaos = __tls_rand();
220	unsigned ext = chaos & 0xff;
221	unsigned other = (chaos >> 8) % (ctxs_count);
222
223	if(ext < 3 \|\| __atomic_load_n(&caches[other / __shard_factor.io].id, __ATOMIC_RELAXED) == this_cache) {
224	proc->io.target = other;
225	}
226	}
227	else {
228	const unsigned target = proc->io.target;
229	/* paranoid */ verify( io.tscs[target].tv != MAX );
230	HELP: if(target < ctxs_count) {
231	const unsigned long long cutoff = calc_cutoff(ctsc, ctx->cq.id, ctxs_count, io.data, io.tscs, __shard_factor.io);
232	const unsigned long long age = moving_average(ctsc, io.tscs[target].tv, io.tscs[target].ma);
233	__cfadbg_print_safe(io, "Kernel I/O: Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, ctx->cq.id, age, cutoff, age > cutoff ? "yes" : "no");
234	if(age <= cutoff) break HELP;
235
236	if(!try_acquire(io.data[target])) break HELP;
237
238	if(!__cfa_do_drain( io.data[target], cltr )) break HELP;
239
240	remote = true;
241	__STATS__( false, io.calls.helped++; )
242	}
243	proc->io.target = MAX;
244	}
245	}
246
247
248	// Drain the local queue
249	if(try_acquire( proc->io.ctx )) {
250	local = __cfa_do_drain( proc->io.ctx, cltr );
251	}
252
253	/* paranoid */ verify( ready_schedule_islocked() );
254	/* paranoid */ verify( ! __preemption_enabled() );
255	/* paranoid */ verify( active_processor() == proc );
256
257	ready_schedule_unlock();
258	return local \|\| remote;
259	}
260
261	bool __cfa_io_flush( processor * proc ) {
262	/* paranoid */ verify( ! __preemption_enabled() );
263	/* paranoid */ verify( proc );
264	/* paranoid */ verify( proc->io.ctx );
265
266	$io_context & ctx = *proc->io.ctx;
267
268	__ioarbiter_flush( ctx );
269
270	if(ctx.sq.to_submit != 0) {
271	ioring_syscsll(ctx, 0, 0);
272
273	}
274
275	return __cfa_io_drain( proc );
276	}
277
278	//=============================================================================================
279	// I/O Submissions
280	//=============================================================================================
281
282	// Submition steps :
283	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
284	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
285	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
286	// need to write an allocator that allows allocating concurrently.
287	//
288	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
289	//
290	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
291	// needs to arrive to two concensus at the same time:
292	// A - The order in which entries are listed in the array: no two threads must pick the
293	// same index for their entries
294	// B - When can the tail be update for the kernel. EVERY entries in the array between
295	// head and tail must be fully filled and shouldn't ever be touched again.
296	//
297	//=============================================================================================
298	// Allocation
299	// for user's convenience fill the sqes from the indexes
300	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
301	struct io_uring_sqe * sqes = ctx->sq.sqes;
302	for(i; want) {
303	// __cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
304	out_sqes[i] = &sqes[idxs[i]];
305	}
306	}
307
308	// Try to directly allocate from the a given context
309	// Not thread-safe
310	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
311	__sub_ring_t & sq = ctx->sq;
312	const __u32 mask = *sq.mask;
313	__u32 fhead = sq.free_ring.head; // get the current head of the queue
314	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
315
316	// If we don't have enough sqes, fail
317	if((ftail - fhead) < want) { return false; }
318
319	// copy all the indexes we want from the available list
320	for(i; want) {
321	// __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
322	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
323	}
324
325	// Advance the head to mark the indexes as consumed
326	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
327
328	// return success
329	return true;
330	}
331
332	// Allocate an submit queue entry.
333	// The kernel cannot see these entries until they are submitted, but other threads must be
334	// able to see which entries can be used and which are already un used by an other thread
335	// for convenience, return both the index and the pointer to the sqe
336	// sqe == &sqes[idx]
337	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
338	// __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
339
340	disable_interrupts();
341	processor * proc = __cfaabi_tls.this_processor;
342	$io_context * ctx = proc->io.ctx;
343	/* paranoid */ verify( __cfaabi_tls.this_processor );
344	/* paranoid */ verify( ctx );
345
346	// __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
347
348	// We can proceed to the fast path
349	if( __alloc(ctx, idxs, want) ) {
350	// Allocation was successful
351	__STATS__( true, io.alloc.fast += 1; )
352	enable_interrupts();
353
354	// __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
355
356	__fill( sqes, want, idxs, ctx );
357	return ctx;
358	}
359	// The fast path failed, fallback
360	__STATS__( true, io.alloc.fail += 1; )
361
362	// Fast path failed, fallback on arbitration
363	__STATS__( true, io.alloc.slow += 1; )
364	enable_interrupts();
365
366	$io_arbiter * ioarb = proc->cltr->io.arbiter;
367	/* paranoid */ verify( ioarb );
368
369	// __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
370
371	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
372
373	// __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
374
375	__fill( sqes, want, idxs,ret );
376	return ret;
377	}
378
379	//=============================================================================================
380	// submission
381	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
382	// We can proceed to the fast path
383	// Get the right objects
384	__sub_ring_t & sq = ctx->sq;
385	const __u32 mask = *sq.mask;
386	__u32 tail = *sq.kring.tail;
387
388	// Add the sqes to the array
389	for( i; have ) {
390	// __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
391	sq.kring.array[ (tail + i) & mask ] = idxs[i];
392	}
393
394	// Make the sqes visible to the submitter
395	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
396	sq.to_submit += have;
397
398	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED);
399	__atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED);
400	}
401
402	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
403	__sub_ring_t & sq = ctx->sq;
404	__submit_only(ctx, idxs, have);
405
406	if(sq.to_submit > 30) {
407	__tls_stats()->io.flush.full++;
408	__cfa_io_flush( ctx->proc );
409	}
410	if(!lazy) {
411	__tls_stats()->io.flush.eager++;
412	__cfa_io_flush( ctx->proc );
413	}
414	}
415
416	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
417	// __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
418
419	disable_interrupts();
420	processor * proc = __cfaabi_tls.this_processor;
421	$io_context * ctx = proc->io.ctx;
422	/* paranoid */ verify( __cfaabi_tls.this_processor );
423	/* paranoid */ verify( ctx );
424
425	// Can we proceed to the fast path
426	if( ctx == inctx ) // We have the right instance?
427	{
428	__submit(ctx, idxs, have, lazy);
429
430	// Mark the instance as no longer in-use, re-enable interrupts and return
431	__STATS__( true, io.submit.fast += 1; )
432	enable_interrupts();
433
434	// __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
435	return;
436	}
437
438	// Fast path failed, fallback on arbitration
439	__STATS__( true, io.submit.slow += 1; )
440	enable_interrupts();
441
442	// __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
443
444	__ioarbiter_submit(inctx, idxs, have, lazy);
445	}
446
447	//=============================================================================================
448	// Flushing
449	// Go through the ring's submit queue and release everything that has already been consumed
450	// by io_uring
451	// This cannot be done by multiple threads
452	static __u32 __release_sqes( struct $io_context & ctx ) {
453	const __u32 mask = *ctx.sq.mask;
454
455	__attribute__((unused))
456	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
457	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
458	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
459
460	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
461
462	// the 3 fields are organized like this diagram
463	// except it's are ring
464	// ---+--------+--------+----
465	// ---+--------+--------+----
466	// ^ ^ ^
467	// phead chead ctail
468
469	// make sure ctail doesn't wrap around and reach phead
470	/* paranoid */ verify(
471	(ctail >= chead && chead >= phead)
472	\|\| (chead >= phead && phead >= ctail)
473	\|\| (phead >= ctail && ctail >= chead)
474	);
475
476	// find the range we need to clear
477	__u32 count = chead - phead;
478
479	if(count == 0) {
480	return 0;
481	}
482
483	// We acquired an previous-head/current-head range
484	// go through the range and release the sqes
485	for( i; count ) {
486	// __cfadbg_print_safe(io, "Kernel I/O : release loop\n");
487	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
488	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
489	}
490
491	ctx.sq.kring.released = chead; // note up to were we processed
492	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
493
494	__ioarbiter_notify(ctx);
495
496	return count;
497	}
498
499	//=============================================================================================
500	// I/O Arbiter
501	//=============================================================================================
502	static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) {
503	bool was_empty;
504
505	// Lock the list, it's not thread safe
506	lock( queue.lock __cfaabi_dbg_ctx2 );
507	{
508	was_empty = empty(queue.queue);
509
510	// Add our request to the list
511	add( queue.queue, item );
512
513	// Mark as pending
514	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
515	}
516	unlock( queue.lock );
517
518	return was_empty;
519	}
520
521	static inline bool empty(__outstanding_io_queue & queue ) {
522	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
523	}
524
525	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
526	// __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
527
528	__STATS__( false, io.alloc.block += 1; )
529
530	// No one has any resources left, wait for something to finish
531	// We need to add ourself to a list of pending allocs and wait for an answer
532	__pending_alloc pa;
533	pa.idxs = idxs;
534	pa.want = want;
535
536	enqueue(this.pending, (__outstanding_io&)pa);
537
538	wait( pa.sem );
539
540	return pa.ctx;
541
542	}
543
544	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
545	/* paranoid */ verify( !empty(this.pending.queue) );
546
547	lock( this.pending.lock __cfaabi_dbg_ctx2 );
548	{
549	while( !empty(this.pending.queue) ) {
550	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
551	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
552	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
553
554	if( have > pa.want ) goto DONE;
555	drop( this.pending.queue );
556
557	/* paranoid */__attribute__((unused)) bool ret =
558
559	__alloc(ctx, pa.idxs, pa.want);
560
561	/* paranoid */ verify( ret );
562
563	pa.ctx = ctx;
564
565	post( pa.sem );
566	}
567
568	this.pending.empty = true;
569	DONE:;
570	}
571	unlock( this.pending.lock );
572	}
573
574	static void __ioarbiter_notify( $io_context & ctx ) {
575	if(!empty( ctx.arbiter->pending )) {
576	__ioarbiter_notify( *ctx.arbiter, &ctx );
577	}
578	}
579
580	// Simply append to the pending
581	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
582	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
583
584	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
585
586	__external_io ei;
587	ei.idxs = idxs;
588	ei.have = have;
589	ei.lazy = lazy;
590
591	bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei);
592
593	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST);
594
595	if( we ) {
596	sigval_t value = { PREEMPT_IO };
597	pthread_sigqueue(ctx->proc->kernel_thread, SIGUSR1, value);
598	}
599
600	wait( ei.sem );
601
602	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
603	}
604
605	static void __ioarbiter_flush( $io_context & ctx ) {
606	if(!empty( ctx.ext_sq )) {
607	__STATS__( false, io.flush.external += 1; )
608
609	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
610
611	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
612	{
613	while( !empty(ctx.ext_sq.queue) ) {
614	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
615
616	__submit_only(&ctx, ei.idxs, ei.have);
617
618	post( ei.sem );
619	}
620
621	ctx.ext_sq.empty = true;
622	}
623	unlock(ctx.ext_sq.lock );
624	}
625	}
626
627	#if defined(CFA_WITH_IO_URING_IDLE)
628	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
629	$io_context * ctx = proc->io.ctx;
630	/* paranoid */ verify( ! __preemption_enabled() );
631	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
632	/* paranoid */ verify( ctx );
633
634	__u32 idx;
635	struct io_uring_sqe * sqe;
636
637	// We can proceed to the fast path
638	if( !__alloc(ctx, &idx, 1) ) {
639	/* paranoid */ verify( false ); // for now check if this happens, next time just abort the sleep.
640	return false;
641	}
642
643	// Allocation was successful
644	__fill( &sqe, 1, &idx, ctx );
645
646	sqe->user_data = (uintptr_t)&future;
647	sqe->flags = 0;
648	sqe->fd = fd;
649	sqe->off = 0;
650	sqe->ioprio = 0;
651	sqe->fsync_flags = 0;
652	sqe->__pad2[0] = 0;
653	sqe->__pad2[1] = 0;
654	sqe->__pad2[2] = 0;
655
656	#if defined(CFA_HAVE_IORING_OP_READ)
657	sqe->opcode = IORING_OP_READ;
658	sqe->addr = (uint64_t)iov.iov_base;
659	sqe->len = iov.iov_len;
660	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
661	sqe->opcode = IORING_OP_READV;
662	sqe->addr = (uintptr_t)&iov;
663	sqe->len = 1;
664	#else
665	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
666	#endif
667
668	asm volatile("": : :"memory");
669
670	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
671	__submit_only( ctx, &idx, 1 );
672
673	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
674	/* paranoid */ verify( ! __preemption_enabled() );
675
676	return true;
677	}
678
679	void __cfa_io_idle( processor * proc ) {
680	iovec iov;
681	__atomic_acquire( &proc->io.ctx->cq.lock );
682
683	__attribute__((used)) volatile bool was_reset = false;
684
685	with( proc->idle_wctx) {
686
687	// Do we already have a pending read
688	if(available(*ftr)) {
689	// There is no pending read, we need to add one
690	reset(*ftr);
691
692	iov.iov_base = rdbuf;
693	iov.iov_len = sizeof(eventfd_t);
694	__kernel_read(proc, *ftr, iov, evfd );
695	ftr->result = 0xDEADDEAD;
696	((eventfd_t )rdbuf) = 0xDEADDEADDEADDEAD;
697	was_reset = true;
698	}
699	}
700
701	if( !__atomic_load_n( &proc->do_terminate, __ATOMIC_SEQ_CST ) ) {
702	__ioarbiter_flush( *proc->io.ctx );
703	proc->idle_wctx.sleep_time = rdtscl();
704	ioring_syscsll( *proc->io.ctx, 1, IORING_ENTER_GETEVENTS);
705	}
706
707	ready_schedule_lock();
708	__cfa_do_drain( proc->io.ctx, proc->cltr );
709	ready_schedule_unlock();
710
711	asm volatile ("" :: "m" (was_reset));
712	}
713	#endif
714	#endif

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

Download in other formats:

Original Format