Context Navigation

io.cfa@ 6e47b49

Visit:

ADT ast-experimental enum pthread-emulation qualifiedEnum

Last change on this file since 6e47b49 was 262fafd9, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Added debugging information to help find deadlock.
Property mode set to `100644`
File size: 21.2 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	#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
182	__cfadbg_print_safe(io, "Kernel I/O : %u completed age %llu\n", count, ts_next);
183	/* paranoid */ verify( ready_schedule_islocked() );
184	/* paranoid */ verify( ! __preemption_enabled() );
185
186	__atomic_unlock(&ctx->cq.lock);
187
188	touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next );
189
190	return true;
191	}
192
193	bool __cfa_io_drain( processor * proc ) {
194	bool local = false;
195	bool remote = false;
196
197	ready_schedule_lock();
198
199	cluster * const cltr = proc->cltr;
200	$io_context * const ctx = proc->io.ctx;
201	/* paranoid */ verify( cltr );
202	/* paranoid */ verify( ctx );
203
204	with(cltr->sched) {
205	const size_t ctxs_count = io.count;
206
207	/* paranoid */ verify( ready_schedule_islocked() );
208	/* paranoid */ verify( ! __preemption_enabled() );
209	/* paranoid */ verify( active_processor() == proc );
210	/* paranoid */ verify( __shard_factor.io > 0 );
211	/* paranoid */ verify( ctxs_count > 0 );
212	/* paranoid */ verify( ctx->cq.id < ctxs_count );
213
214	const unsigned this_cache = cache_id(cltr, ctx->cq.id / __shard_factor.io);
215	const unsigned long long ctsc = rdtscl();
216
217	if(proc->io.target == MAX) {
218	uint64_t chaos = __tls_rand();
219	unsigned ext = chaos & 0xff;
220	unsigned other = (chaos >> 8) % (ctxs_count);
221
222	if(ext < 3 \|\| __atomic_load_n(&caches[other / __shard_factor.io].id, __ATOMIC_RELAXED) == this_cache) {
223	proc->io.target = other;
224	}
225	}
226	else {
227	const unsigned target = proc->io.target;
228	/* paranoid */ verify( io.tscs[target].tv != MAX );
229	HELP: if(target < ctxs_count) {
230	const unsigned long long cutoff = calc_cutoff(ctsc, ctx->cq.id, ctxs_count, io.data, io.tscs, __shard_factor.io);
231	const unsigned long long age = moving_average(ctsc, io.tscs[target].tv, io.tscs[target].ma);
232	__cfadbg_print_safe(io, "Kernel I/O: Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no");
233	if(age <= cutoff) break HELP;
234
235	if(!try_acquire(io.data[target])) break HELP;
236
237	if(!__cfa_do_drain( io.data[target], cltr )) break HELP;
238
239	remote = true;
240	__STATS__( false, io.calls.helped++; )
241	}
242	proc->io.target = MAX;
243	}
244	}
245
246
247	// Drain the local queue
248	if(try_acquire( proc->io.ctx )) {
249	local = __cfa_do_drain( proc->io.ctx, cltr );
250	}
251
252	/* paranoid */ verify( ready_schedule_islocked() );
253	/* paranoid */ verify( ! __preemption_enabled() );
254	/* paranoid */ verify( active_processor() == proc );
255
256	ready_schedule_unlock();
257	return local \|\| remote;
258	}
259
260	bool __cfa_io_flush( processor * proc ) {
261	/* paranoid */ verify( ! __preemption_enabled() );
262	/* paranoid */ verify( proc );
263	/* paranoid */ verify( proc->io.ctx );
264
265	$io_context & ctx = *proc->io.ctx;
266
267	__ioarbiter_flush( ctx );
268
269	if(ctx.sq.to_submit != 0) {
270	ioring_syscsll(ctx, 0, 0);
271
272	}
273
274	return __cfa_io_drain( proc );
275	}
276
277	//=============================================================================================
278	// I/O Submissions
279	//=============================================================================================
280
281	// Submition steps :
282	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
283	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
284	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
285	// need to write an allocator that allows allocating concurrently.
286	//
287	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
288	//
289	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
290	// needs to arrive to two concensus at the same time:
291	// A - The order in which entries are listed in the array: no two threads must pick the
292	// same index for their entries
293	// B - When can the tail be update for the kernel. EVERY entries in the array between
294	// head and tail must be fully filled and shouldn't ever be touched again.
295	//
296	//=============================================================================================
297	// Allocation
298	// for user's convenience fill the sqes from the indexes
299	static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
300	struct io_uring_sqe * sqes = ctx->sq.sqes;
301	for(i; want) {
302	// __cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
303	out_sqes[i] = &sqes[idxs[i]];
304	}
305	}
306
307	// Try to directly allocate from the a given context
308	// Not thread-safe
309	static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
310	__sub_ring_t & sq = ctx->sq;
311	const __u32 mask = *sq.mask;
312	__u32 fhead = sq.free_ring.head; // get the current head of the queue
313	__u32 ftail = sq.free_ring.tail; // get the current tail of the queue
314
315	// If we don't have enough sqes, fail
316	if((ftail - fhead) < want) { return false; }
317
318	// copy all the indexes we want from the available list
319	for(i; want) {
320	// __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
321	idxs[i] = sq.free_ring.array[(fhead + i) & mask];
322	}
323
324	// Advance the head to mark the indexes as consumed
325	__atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
326
327	// return success
328	return true;
329	}
330
331	// Allocate an submit queue entry.
332	// The kernel cannot see these entries until they are submitted, but other threads must be
333	// able to see which entries can be used and which are already un used by an other thread
334	// for convenience, return both the index and the pointer to the sqe
335	// sqe == &sqes[idx]
336	struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
337	// __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
338
339	disable_interrupts();
340	processor * proc = __cfaabi_tls.this_processor;
341	$io_context * ctx = proc->io.ctx;
342	/* paranoid */ verify( __cfaabi_tls.this_processor );
343	/* paranoid */ verify( ctx );
344
345	// __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
346
347	// We can proceed to the fast path
348	if( __alloc(ctx, idxs, want) ) {
349	// Allocation was successful
350	__STATS__( true, io.alloc.fast += 1; )
351	enable_interrupts();
352
353	// __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
354
355	__fill( sqes, want, idxs, ctx );
356	return ctx;
357	}
358	// The fast path failed, fallback
359	__STATS__( true, io.alloc.fail += 1; )
360
361	// Fast path failed, fallback on arbitration
362	__STATS__( true, io.alloc.slow += 1; )
363	enable_interrupts();
364
365	$io_arbiter * ioarb = proc->cltr->io.arbiter;
366	/* paranoid */ verify( ioarb );
367
368	// __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
369
370	struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
371
372	// __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
373
374	__fill( sqes, want, idxs,ret );
375	return ret;
376	}
377
378	//=============================================================================================
379	// submission
380	static inline void __submit_only( struct $io_context * ctx, __u32 idxs[], __u32 have) {
381	// We can proceed to the fast path
382	// Get the right objects
383	__sub_ring_t & sq = ctx->sq;
384	const __u32 mask = *sq.mask;
385	__u32 tail = *sq.kring.tail;
386
387	// Add the sqes to the array
388	for( i; have ) {
389	// __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
390	sq.kring.array[ (tail + i) & mask ] = idxs[i];
391	}
392
393	// Make the sqes visible to the submitter
394	__atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
395	sq.to_submit += have;
396
397	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED);
398	__atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED);
399	}
400
401	static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
402	__sub_ring_t & sq = ctx->sq;
403	__submit_only(ctx, idxs, have);
404
405	if(sq.to_submit > 30) {
406	__tls_stats()->io.flush.full++;
407	__cfa_io_flush( ctx->proc );
408	}
409	if(!lazy) {
410	__tls_stats()->io.flush.eager++;
411	__cfa_io_flush( ctx->proc );
412	}
413	}
414
415	void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
416	// __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
417
418	disable_interrupts();
419	processor * proc = __cfaabi_tls.this_processor;
420	$io_context * ctx = proc->io.ctx;
421	/* paranoid */ verify( __cfaabi_tls.this_processor );
422	/* paranoid */ verify( ctx );
423
424	// Can we proceed to the fast path
425	if( ctx == inctx ) // We have the right instance?
426	{
427	__submit(ctx, idxs, have, lazy);
428
429	// Mark the instance as no longer in-use, re-enable interrupts and return
430	__STATS__( true, io.submit.fast += 1; )
431	enable_interrupts();
432
433	// __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
434	return;
435	}
436
437	// Fast path failed, fallback on arbitration
438	__STATS__( true, io.submit.slow += 1; )
439	enable_interrupts();
440
441	// __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
442
443	__ioarbiter_submit(inctx, idxs, have, lazy);
444	}
445
446	//=============================================================================================
447	// Flushing
448	// Go through the ring's submit queue and release everything that has already been consumed
449	// by io_uring
450	// This cannot be done by multiple threads
451	static __u32 __release_sqes( struct $io_context & ctx ) {
452	const __u32 mask = *ctx.sq.mask;
453
454	__attribute__((unused))
455	__u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
456	__u32 chead = *ctx.sq.kring.head; // get the current head of the queue
457	__u32 phead = ctx.sq.kring.released; // get the head the last time we were here
458
459	__u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
460
461	// the 3 fields are organized like this diagram
462	// except it's are ring
463	// ---+--------+--------+----
464	// ---+--------+--------+----
465	// ^ ^ ^
466	// phead chead ctail
467
468	// make sure ctail doesn't wrap around and reach phead
469	/* paranoid */ verify(
470	(ctail >= chead && chead >= phead)
471	\|\| (chead >= phead && phead >= ctail)
472	\|\| (phead >= ctail && ctail >= chead)
473	);
474
475	// find the range we need to clear
476	__u32 count = chead - phead;
477
478	if(count == 0) {
479	return 0;
480	}
481
482	// We acquired an previous-head/current-head range
483	// go through the range and release the sqes
484	for( i; count ) {
485	// __cfadbg_print_safe(io, "Kernel I/O : release loop\n");
486	__u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
487	ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
488	}
489
490	ctx.sq.kring.released = chead; // note up to were we processed
491	__atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
492
493	__ioarbiter_notify(ctx);
494
495	return count;
496	}
497
498	//=============================================================================================
499	// I/O Arbiter
500	//=============================================================================================
501	static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) {
502	bool was_empty;
503
504	// Lock the list, it's not thread safe
505	lock( queue.lock __cfaabi_dbg_ctx2 );
506	{
507	was_empty = empty(queue.queue);
508
509	// Add our request to the list
510	add( queue.queue, item );
511
512	// Mark as pending
513	__atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
514	}
515	unlock( queue.lock );
516
517	return was_empty;
518	}
519
520	static inline bool empty(__outstanding_io_queue & queue ) {
521	return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
522	}
523
524	static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
525	// __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
526
527	__STATS__( false, io.alloc.block += 1; )
528
529	// No one has any resources left, wait for something to finish
530	// We need to add ourself to a list of pending allocs and wait for an answer
531	__pending_alloc pa;
532	pa.idxs = idxs;
533	pa.want = want;
534
535	enqueue(this.pending, (__outstanding_io&)pa);
536
537	wait( pa.sem );
538
539	return pa.ctx;
540
541	}
542
543	static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
544	/* paranoid */ verify( !empty(this.pending.queue) );
545
546	lock( this.pending.lock __cfaabi_dbg_ctx2 );
547	{
548	while( !empty(this.pending.queue) ) {
549	__cfadbg_print_safe(io, "Kernel I/O : notifying\n");
550	__u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
551	__pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
552
553	if( have > pa.want ) goto DONE;
554	drop( this.pending.queue );
555
556	/* paranoid */__attribute__((unused)) bool ret =
557
558	__alloc(ctx, pa.idxs, pa.want);
559
560	/* paranoid */ verify( ret );
561
562	pa.ctx = ctx;
563
564	post( pa.sem );
565	}
566
567	this.pending.empty = true;
568	DONE:;
569	}
570	unlock( this.pending.lock );
571	}
572
573	static void __ioarbiter_notify( $io_context & ctx ) {
574	if(!empty( ctx.arbiter->pending )) {
575	__ioarbiter_notify( *ctx.arbiter, &ctx );
576	}
577	}
578
579	// Simply append to the pending
580	static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
581	__cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
582
583	__cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
584
585	__external_io ei;
586	ei.idxs = idxs;
587	ei.have = have;
588	ei.lazy = lazy;
589
590	bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei);
591
592	__atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST);
593
594	if( we ) {
595	sigval_t value = { PREEMPT_IO };
596	pthread_sigqueue(ctx->proc->kernel_thread, SIGUSR1, value);
597	}
598
599	wait( ei.sem );
600
601	__cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
602	}
603
604	static void __ioarbiter_flush( $io_context & ctx ) {
605	if(!empty( ctx.ext_sq )) {
606	__STATS__( false, io.flush.external += 1; )
607
608	__cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
609
610	lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
611	{
612	while( !empty(ctx.ext_sq.queue) ) {
613	__external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
614
615	__submit_only(&ctx, ei.idxs, ei.have);
616
617	post( ei.sem );
618	}
619
620	ctx.ext_sq.empty = true;
621	}
622	unlock(ctx.ext_sq.lock );
623	}
624	}
625
626	#if defined(CFA_WITH_IO_URING_IDLE)
627	bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) {
628	$io_context * ctx = proc->io.ctx;
629	/* paranoid */ verify( ! __preemption_enabled() );
630	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
631	/* paranoid */ verify( ctx );
632
633	__u32 idx;
634	struct io_uring_sqe * sqe;
635
636	// We can proceed to the fast path
637	if( !__alloc(ctx, &idx, 1) ) {
638	/* paranoid */ verify( false ); // for now check if this happens, next time just abort the sleep.
639	return false;
640	}
641
642	// Allocation was successful
643	__fill( &sqe, 1, &idx, ctx );
644
645	sqe->user_data = (uintptr_t)&future;
646	sqe->flags = 0;
647	sqe->fd = fd;
648	sqe->off = 0;
649	sqe->ioprio = 0;
650	sqe->fsync_flags = 0;
651	sqe->__pad2[0] = 0;
652	sqe->__pad2[1] = 0;
653	sqe->__pad2[2] = 0;
654
655	#if defined(CFA_HAVE_IORING_OP_READ)
656	sqe->opcode = IORING_OP_READ;
657	sqe->addr = (uint64_t)iov.iov_base;
658	sqe->len = iov.iov_len;
659	#elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)
660	sqe->opcode = IORING_OP_READV;
661	sqe->addr = (uintptr_t)&iov;
662	sqe->len = 1;
663	#else
664	#error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined
665	#endif
666
667	asm volatile("": : :"memory");
668
669	/* paranoid */ verify( sqe->user_data == (uintptr_t)&future );
670	__submit_only( ctx, &idx, 1 );
671
672	/* paranoid */ verify( proc == __cfaabi_tls.this_processor );
673	/* paranoid */ verify( ! __preemption_enabled() );
674
675	return true;
676	}
677
678	void __cfa_io_idle( processor * proc ) {
679	iovec iov;
680	__atomic_acquire( &proc->io.ctx->cq.lock );
681
682	__attribute__((used)) volatile bool was_reset = false;
683
684	with( proc->idle_wctx) {
685
686	// Do we already have a pending read
687	if(available(*ftr)) {
688	// There is no pending read, we need to add one
689	reset(*ftr);
690
691	iov.iov_base = rdbuf;
692	iov.iov_len = sizeof(eventfd_t);
693	__kernel_read(proc, *ftr, iov, evfd );
694	ftr->result = 0xDEADDEAD;
695	((eventfd_t )rdbuf) = 0xDEADDEADDEADDEAD;
696	was_reset = true;
697	}
698	}
699
700	if( !__atomic_load_n( &proc->do_terminate, __ATOMIC_SEQ_CST ) ) {
701	__ioarbiter_flush( *proc->io.ctx );
702	proc->idle_wctx.sleep_time = rdtscl();
703	ioring_syscsll( *proc->io.ctx, 1, IORING_ENTER_GETEVENTS);
704	}
705
706	ready_schedule_lock();
707	__cfa_do_drain( proc->io.ctx, proc->cltr );
708	ready_schedule_unlock();
709
710	asm volatile ("" :: "m" (was_reset));
711	}
712	#endif
713	#endif

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

Download in other formats:

Original Format