Context Navigation

io.cfa @ 58f99b3

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

Last change on this file since 58f99b3 was 426f60c, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Web server seems to work
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
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/epoll.h>
34	#include <sys/eventfd.h>
35	#include <sys/syscall.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 "io/types.hfa"
44
45	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	// returns true of acquired as leader or second leader
84	static inline bool try_lock( __leaderlock_t & this ) {
85	const uintptr_t thrd = 1z \| (uintptr_t)active_thread();
86	bool block;
87	disable_interrupts();
88	for() {
89	struct $thread * expected = this.value;
90	if( 1p != expected && 0p != expected ) {
91	/* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader
92	enable_interrupts( __cfaabi_dbg_ctx );
93	return false;
94	}
95	struct $thread * desired;
96	if( 0p == expected ) {
97	// If the lock isn't locked acquire it, no need to block
98	desired = 1p;
99	block = false;
100	}
101	else {
102	// If the lock is already locked try becomming the next leader
103	desired = (struct $thread *)thrd;
104	block = true;
105	}
106	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
107	}
108	if( block ) {
109	enable_interrupts( __cfaabi_dbg_ctx );
110	park();
111	disable_interrupts();
112	}
113	return true;
114	}
115
116	static inline bool next( __leaderlock_t & this ) {
117	/* paranoid */ verify( ! __preemption_enabled() );
118	struct $thread * nextt;
119	for() {
120	struct $thread * expected = this.value;
121	/* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked
122
123	struct $thread * desired;
124	if( 1p == expected ) {
125	// No next leader, just unlock
126	desired = 0p;
127	nextt = 0p;
128	}
129	else {
130	// There is a next leader, remove but keep locked
131	desired = 1p;
132	nextt = (struct $thread *)(~1z & (uintptr_t)expected);
133	}
134	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
135	}
136
137	if(nextt) {
138	unpark( nextt );
139	enable_interrupts( __cfaabi_dbg_ctx );
140	return true;
141	}
142	enable_interrupts( __cfaabi_dbg_ctx );
143	return false;
144	}
145
146	//=============================================================================================
147	// I/O Syscall
148	//=============================================================================================
149	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
150	bool need_sys_to_submit = false;
151	bool need_sys_to_complete = false;
152	unsigned flags = 0;
153
154	TO_SUBMIT:
155	if( to_submit > 0 ) {
156	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
157	need_sys_to_submit = true;
158	break TO_SUBMIT;
159	}
160	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
161	need_sys_to_submit = true;
162	flags \|= IORING_ENTER_SQ_WAKEUP;
163	}
164	}
165
166	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
167	flags \|= IORING_ENTER_GETEVENTS;
168	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
169	need_sys_to_complete = true;
170	}
171	}
172
173	int ret = 0;
174	if( need_sys_to_submit \|\| need_sys_to_complete ) {
175	__cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags);
176	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8);
177	if( ret < 0 ) {
178	switch((int)errno) {
179	case EAGAIN:
180	case EINTR:
181	ret = -1;
182	break;
183	default:
184	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
185	}
186	}
187	}
188
189	// Memory barrier
190	__atomic_thread_fence( __ATOMIC_SEQ_CST );
191	return ret;
192	}
193
194	//=============================================================================================
195	// I/O Polling
196	//=============================================================================================
197	static unsigned __collect_submitions( struct __io_data & ring );
198	static __u32 __release_consumed_submission( struct __io_data & ring );
199
200	// Process a single completion message from the io_uring
201	// This is NOT thread-safe
202	static inline void process( volatile struct io_uring_cqe & cqe ) {
203	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
204	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
205
206	fulfil( *future, cqe.res );
207	}
208
209	static [int, bool] __drain_io( & struct __io_data ring ) {
210	/* paranoid */ verify( ! __preemption_enabled() );
211
212	unsigned to_submit = 0;
213	if( ring.poller_submits ) {
214	// If the poller thread also submits, then we need to aggregate the submissions which are ready
215	to_submit = __collect_submitions( ring );
216	}
217
218	int ret = __io_uring_enter(ring, to_submit, true);
219	if( ret < 0 ) {
220	return [0, true];
221	}
222
223	// update statistics
224	if (to_submit > 0) {
225	__STATS__( true,
226	if( to_submit > 0 ) {
227	io.submit_q.submit_avg.rdy += to_submit;
228	io.submit_q.submit_avg.csm += ret;
229	io.submit_q.submit_avg.cnt += 1;
230	}
231	)
232	}
233
234	__atomic_thread_fence( __ATOMIC_SEQ_CST );
235
236	// Release the consumed SQEs
237	__release_consumed_submission( ring );
238
239	// Drain the queue
240	unsigned head = *ring.completion_q.head;
241	unsigned tail = *ring.completion_q.tail;
242	const __u32 mask = *ring.completion_q.mask;
243
244	// Nothing was new return 0
245	if (head == tail) {
246	return [0, to_submit > 0];
247	}
248
249	__u32 count = tail - head;
250	/* paranoid */ verify( count != 0 );
251	for(i; count) {
252	unsigned idx = (head + i) & mask;
253	volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
254
255	/* paranoid */ verify(&cqe);
256
257	process( cqe );
258	}
259
260	// Mark to the kernel that the cqe has been seen
261	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
262	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_SEQ_CST );
263
264	return [count, count > 0 \|\| to_submit > 0];
265	}
266
267	void main( $io_ctx_thread & this ) {
268	epoll_event ev;
269	__ioctx_register( this, ev );
270
271	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %d (%p) ready\n", this.ring->fd, &this);
272
273	const int reset_cnt = 5;
274	int reset = reset_cnt;
275	// Then loop until we need to start
276	LOOP:
277	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
278	// Drain the io
279	int count;
280	bool again;
281	disable_interrupts();
282	[count, again] = __drain_io( *this.ring );
283
284	if(!again) reset--;
285
286	// Update statistics
287	__STATS__( true,
288	io.complete_q.completed_avg.val += count;
289	io.complete_q.completed_avg.cnt += 1;
290	)
291	enable_interrupts( __cfaabi_dbg_ctx );
292
293	// If we got something, just yield and check again
294	if(reset > 1) {
295	yield();
296	continue LOOP;
297	}
298
299	// We alread failed to find completed entries a few time.
300	if(reset == 1) {
301	// Rearm the context so it can block
302	// but don't block right away
303	// we need to retry one last time in case
304	// something completed just now
305	__ioctx_prepare_block( this, ev );
306	continue LOOP;
307	}
308
309	__STATS__( false,
310	io.complete_q.blocks += 1;
311	)
312	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %d (%p)\n", this.ring->fd, &this);
313
314	// block this thread
315	wait( this.sem );
316
317	eventfd_t v;
318	eventfd_read(this.ring->efd, &v);
319
320	// restore counter
321	reset = reset_cnt;
322	}
323
324	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller %d (%p) stopping\n", this.ring->fd, &this);
325	}
326
327	//=============================================================================================
328	// I/O Submissions
329	//=============================================================================================
330
331	// Submition steps :
332	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
333	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
334	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
335	// need to write an allocator that allows allocating concurrently.
336	//
337	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
338	//
339	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
340	// needs to arrive to two concensus at the same time:
341	// A - The order in which entries are listed in the array: no two threads must pick the
342	// same index for their entries
343	// B - When can the tail be update for the kernel. EVERY entries in the array between
344	// head and tail must be fully filled and shouldn't ever be touched again.
345	//
346
347	// Allocate an submit queue entry.
348	// The kernel cannot see these entries until they are submitted, but other threads must be
349	// able to see which entries can be used and which are already un used by an other thread
350	// for convenience, return both the index and the pointer to the sqe
351	// sqe == &sqes[idx]
352	[* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
353	/* paranoid */ verify( data != 0 );
354
355	// Prepare the data we need
356	__attribute((unused)) int len = 0;
357	__attribute((unused)) int block = 0;
358	__u32 cnt = *ring.submit_q.num;
359	__u32 mask = *ring.submit_q.mask;
360
361	__u32 off = thread_rand();
362
363	// Loop around looking for an available spot
364	for() {
365	// Look through the list starting at some offset
366	for(i; cnt) {
367	__u64 expected = 3;
368	__u32 idx = (i + off) & mask; // Get an index from a random
369	volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
370	volatile __u64 * udata = &sqe->user_data;
371
372	// Allocate the entry by CASing the user_data field from 0 to the future address
373	if( *udata == expected &&
374	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
375	{
376	// update statistics
377	__STATS__( false,
378	io.submit_q.alloc_avg.val += len;
379	io.submit_q.alloc_avg.block += block;
380	io.submit_q.alloc_avg.cnt += 1;
381	)
382
383	// debug log
384	__cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
385
386	// Success return the data
387	sqe->opcode = 0;
388	sqe->flags = 0;
389	sqe->ioprio = 0;
390	sqe->fd = 0;
391	sqe->off = 0;
392	sqe->addr = 0;
393	sqe->len = 0;
394	sqe->accept_flags = 0;
395	sqe->__pad2[0] = 0;
396	sqe->__pad2[1] = 0;
397	sqe->__pad2[2] = 0;
398	return [sqe, idx];
399	}
400	verify(expected != data);
401
402	// This one was used
403	len ++;
404	}
405
406	block++;
407
408	abort( "Kernel I/O : all submit queue entries used, yielding\n" );
409
410	yield();
411	}
412	}
413
414	static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
415	/* paranoid */ verify( idx <= mask );
416	/* paranoid */ verify( idx != -1ul32 );
417
418	// We need to find a spot in the ready array
419	__attribute((unused)) int len = 0;
420	__attribute((unused)) int block = 0;
421	__u32 ready_mask = ring.submit_q.ready_cnt - 1;
422
423	__u32 off = thread_rand();
424
425	__u32 picked;
426	LOOKING: for() {
427	for(i; ring.submit_q.ready_cnt) {
428	picked = (i + off) & ready_mask;
429	__u32 expected = -1ul32;
430	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
431	break LOOKING;
432	}
433	verify(expected != idx);
434
435	len ++;
436	}
437
438	block++;
439
440	__u32 released = __release_consumed_submission( ring );
441	if( released == 0 ) {
442	yield();
443	}
444	}
445
446	// update statistics
447	__STATS__( false,
448	io.submit_q.look_avg.val += len;
449	io.submit_q.look_avg.block += block;
450	io.submit_q.look_avg.cnt += 1;
451	)
452
453	return picked;
454	}
455
456	void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
457	__io_data & ring = *ctx->thrd.ring;
458
459	{
460	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
461	__cfadbg_print_safe( io,
462	"Kernel I/O : submitting %u (%p) for %p\n"
463	" data: %p\n"
464	" opcode: %s\n"
465	" fd: %d\n"
466	" flags: %d\n"
467	" prio: %d\n"
468	" off: %p\n"
469	" addr: %p\n"
470	" len: %d\n"
471	" other flags: %d\n"
472	" splice fd: %d\n"
473	" pad[0]: %llu\n"
474	" pad[1]: %llu\n"
475	" pad[2]: %llu\n",
476	idx, sqe,
477	active_thread(),
478	(void*)sqe->user_data,
479	opcodes[sqe->opcode],
480	sqe->fd,
481	sqe->flags,
482	sqe->ioprio,
483	sqe->off,
484	sqe->addr,
485	sqe->len,
486	sqe->accept_flags,
487	sqe->splice_fd_in,
488	sqe->__pad2[0],
489	sqe->__pad2[1],
490	sqe->__pad2[2]
491	);
492	}
493
494
495	// Get now the data we definetely need
496	volatile __u32 * const tail = ring.submit_q.tail;
497	const __u32 mask = *ring.submit_q.mask;
498
499	// There are 2 submission schemes, check which one we are using
500	if( ring.poller_submits ) {
501	// If the poller thread submits, then we just need to add this to the ready array
502	__submit_to_ready_array( ring, idx, mask );
503
504	post( ctx->thrd.sem );
505
506	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
507	}
508	else if( ring.eager_submits ) {
509	__u32 picked = __submit_to_ready_array( ring, idx, mask );
510
511	#if defined(LEADER_LOCK)
512	if( !try_lock(ring.submit_q.submit_lock) ) {
513	__STATS__( false,
514	io.submit_q.helped += 1;
515	)
516	return;
517	}
518	/* paranoid */ verify( ! __preemption_enabled() );
519	__STATS__( true,
520	io.submit_q.leader += 1;
521	)
522	#else
523	for() {
524	yield();
525
526	if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
527	__STATS__( false,
528	io.submit_q.leader += 1;
529	)
530	break;
531	}
532
533	// If some one else collected our index, we are done
534	#warning ABA problem
535	if( ring.submit_q.ready[picked] != idx ) {
536	__STATS__( false,
537	io.submit_q.helped += 1;
538	)
539	return;
540	}
541
542	__STATS__( false,
543	io.submit_q.busy += 1;
544	)
545	}
546	#endif
547
548	// We got the lock
549	// Collect the submissions
550	unsigned to_submit = __collect_submitions( ring );
551
552	// Actually submit
553	int ret = __io_uring_enter( ring, to_submit, false );
554
555	#if defined(LEADER_LOCK)
556	/* paranoid */ verify( ! __preemption_enabled() );
557	next(ring.submit_q.submit_lock);
558	#else
559	unlock(ring.submit_q.submit_lock);
560	#endif
561	if( ret < 0 ) {
562	return;
563	}
564
565	// Release the consumed SQEs
566	__release_consumed_submission( ring );
567
568	// update statistics
569	__STATS__( false,
570	io.submit_q.submit_avg.rdy += to_submit;
571	io.submit_q.submit_avg.csm += ret;
572	io.submit_q.submit_avg.cnt += 1;
573	)
574
575	__cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
576	}
577	else
578	{
579	// get mutual exclusion
580	#if defined(LEADER_LOCK)
581	while(!try_lock(ring.submit_q.submit_lock));
582	#else
583	lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
584	#endif
585
586	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64,
587	/* paranoid */ "index %u already reclaimed\n"
588	/* paranoid */ "head %u, prev %u, tail %u\n"
589	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
590	/* paranoid */ idx,
591	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
592	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
593	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
594	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
595	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
596	/* paranoid */ );
597
598	// Append to the list of ready entries
599
600	/* paranoid */ verify( idx <= mask );
601	ring.submit_q.array[ (*tail) & mask ] = idx;
602	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
603
604	// Submit however, many entries need to be submitted
605	int ret = __io_uring_enter( ring, 1, false );
606	if( ret < 0 ) {
607	switch((int)errno) {
608	default:
609	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
610	}
611	}
612
613	/* paranoid */ verify(ret == 1);
614
615	// update statistics
616	__STATS__( false,
617	io.submit_q.submit_avg.csm += 1;
618	io.submit_q.submit_avg.cnt += 1;
619	)
620
621	{
622	__attribute__((unused)) volatile __u32 * const head = ring.submit_q.head;
623	__attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ];
624	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx];
625
626	__cfadbg_print_safe( io,
627	"Kernel I/O : last submitted is %u (%p)\n"
628	" data: %p\n"
629	" opcode: %s\n"
630	" fd: %d\n"
631	" flags: %d\n"
632	" prio: %d\n"
633	" off: %p\n"
634	" addr: %p\n"
635	" len: %d\n"
636	" other flags: %d\n"
637	" splice fd: %d\n"
638	" pad[0]: %llu\n"
639	" pad[1]: %llu\n"
640	" pad[2]: %llu\n",
641	last_idx, sqe,
642	(void*)sqe->user_data,
643	opcodes[sqe->opcode],
644	sqe->fd,
645	sqe->flags,
646	sqe->ioprio,
647	sqe->off,
648	sqe->addr,
649	sqe->len,
650	sqe->accept_flags,
651	sqe->splice_fd_in,
652	sqe->__pad2[0],
653	sqe->__pad2[1],
654	sqe->__pad2[2]
655	);
656	}
657
658	__atomic_thread_fence( __ATOMIC_SEQ_CST );
659	// Release the consumed SQEs
660	__release_consumed_submission( ring );
661	// ring.submit_q.sqes[idx].user_data = 3ul64;
662
663	#if defined(LEADER_LOCK)
664	next(ring.submit_q.submit_lock);
665	#else
666	unlock(ring.submit_q.submit_lock);
667	#endif
668
669	__cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() );
670	}
671	}
672
673	// #define PARTIAL_SUBMIT 32
674
675	// go through the list of submissions in the ready array and moved them into
676	// the ring's submit queue
677	static unsigned __collect_submitions( struct __io_data & ring ) {
678	/* paranoid */ verify( ring.submit_q.ready != 0p );
679	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
680
681	unsigned to_submit = 0;
682	__u32 tail = *ring.submit_q.tail;
683	const __u32 mask = *ring.submit_q.mask;
684	#if defined(PARTIAL_SUBMIT)
685	#if defined(LEADER_LOCK)
686	#error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
687	#endif
688	const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
689	const __u32 offset = ring.submit_q.prev_ready;
690	ring.submit_q.prev_ready += cnt;
691	#else
692	const __u32 cnt = ring.submit_q.ready_cnt;
693	const __u32 offset = 0;
694	#endif
695
696	// Go through the list of ready submissions
697	for( c; cnt ) {
698	__u32 i = (offset + c) % ring.submit_q.ready_cnt;
699
700	// replace any submission with the sentinel, to consume it.
701	__u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
702
703	// If it was already the sentinel, then we are done
704	if( idx == -1ul32 ) continue;
705
706	// If we got a real submission, append it to the list
707	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
708	to_submit++;
709	}
710
711	// Increment the tail based on how many we are ready to submit
712	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
713
714	return to_submit;
715	}
716
717	// Go through the ring's submit queue and release everything that has already been consumed
718	// by io_uring
719	static __u32 __release_consumed_submission( struct __io_data & ring ) {
720	const __u32 smask = *ring.submit_q.mask;
721
722	// We need to get the lock to copy the old head and new head
723	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
724	__attribute__((unused))
725	__u32 ctail = *ring.submit_q.tail; // get the current tail of the queue
726	__u32 chead = *ring.submit_q.head; // get the current head of the queue
727	__u32 phead = ring.submit_q.prev_head; // get the head the last time we were here
728	ring.submit_q.prev_head = chead; // note up to were we processed
729	unlock(ring.submit_q.release_lock);
730
731	// the 3 fields are organized like this diagram
732	// except it's are ring
733	// ---+--------+--------+----
734	// ---+--------+--------+----
735	// ^ ^ ^
736	// phead chead ctail
737
738	// make sure ctail doesn't wrap around and reach phead
739	/* paranoid */ verify(
740	(ctail >= chead && chead >= phead)
741	\|\| (chead >= phead && phead >= ctail)
742	\|\| (phead >= ctail && ctail >= chead)
743	);
744
745	// find the range we need to clear
746	__u32 count = chead - phead;
747
748	// We acquired an previous-head/current-head range
749	// go through the range and release the sqes
750	for( i; count ) {
751	__u32 idx = ring.submit_q.array[ (phead + i) & smask ];
752
753	/* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data );
754	ring.submit_q.sqes[ idx ].user_data = 3ul64;
755	}
756	return count;
757	}
758	#endif

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

Download in other formats:

Original Format