Context Navigation

io.cfa @ 8c43d05

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

Last change on this file since 8c43d05 was b8b0de0c, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Removed temporary debug abort that caused plg7a's untimely death
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/syscall.h>
34
35	#include <linux/io_uring.h>
36	}
37
38	#include "stats.hfa"
39	#include "kernel.hfa"
40	#include "kernel/fwd.hfa"
41	#include "io/types.hfa"
42
43	static const char * opcodes[] = {
44	"OP_NOP",
45	"OP_READV",
46	"OP_WRITEV",
47	"OP_FSYNC",
48	"OP_READ_FIXED",
49	"OP_WRITE_FIXED",
50	"OP_POLL_ADD",
51	"OP_POLL_REMOVE",
52	"OP_SYNC_FILE_RANGE",
53	"OP_SENDMSG",
54	"OP_RECVMSG",
55	"OP_TIMEOUT",
56	"OP_TIMEOUT_REMOVE",
57	"OP_ACCEPT",
58	"OP_ASYNC_CANCEL",
59	"OP_LINK_TIMEOUT",
60	"OP_CONNECT",
61	"OP_FALLOCATE",
62	"OP_OPENAT",
63	"OP_CLOSE",
64	"OP_FILES_UPDATE",
65	"OP_STATX",
66	"OP_READ",
67	"OP_WRITE",
68	"OP_FADVISE",
69	"OP_MADVISE",
70	"OP_SEND",
71	"OP_RECV",
72	"OP_OPENAT2",
73	"OP_EPOLL_CTL",
74	"OP_SPLICE",
75	"OP_PROVIDE_BUFFERS",
76	"OP_REMOVE_BUFFERS",
77	"OP_TEE",
78	"INVALID_OP"
79	};
80
81	// returns true of acquired as leader or second leader
82	static inline bool try_lock( __leaderlock_t & this ) {
83	const uintptr_t thrd = 1z \| (uintptr_t)active_thread();
84	bool block;
85	disable_interrupts();
86	for() {
87	struct $thread * expected = this.value;
88	if( 1p != expected && 0p != expected ) {
89	/* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader
90	enable_interrupts( __cfaabi_dbg_ctx );
91	return false;
92	}
93	struct $thread * desired;
94	if( 0p == expected ) {
95	// If the lock isn't locked acquire it, no need to block
96	desired = 1p;
97	block = false;
98	}
99	else {
100	// If the lock is already locked try becomming the next leader
101	desired = (struct $thread *)thrd;
102	block = true;
103	}
104	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
105	}
106	if( block ) {
107	enable_interrupts( __cfaabi_dbg_ctx );
108	park();
109	disable_interrupts();
110	}
111	return true;
112	}
113
114	static inline bool next( __leaderlock_t & this ) {
115	/* paranoid */ verify( ! __preemption_enabled() );
116	struct $thread * nextt;
117	for() {
118	struct $thread * expected = this.value;
119	/* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked
120
121	struct $thread * desired;
122	if( 1p == expected ) {
123	// No next leader, just unlock
124	desired = 0p;
125	nextt = 0p;
126	}
127	else {
128	// There is a next leader, remove but keep locked
129	desired = 1p;
130	nextt = (struct $thread *)(~1z & (uintptr_t)expected);
131	}
132	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
133	}
134
135	if(nextt) {
136	unpark( nextt );
137	enable_interrupts( __cfaabi_dbg_ctx );
138	return true;
139	}
140	enable_interrupts( __cfaabi_dbg_ctx );
141	return false;
142	}
143
144	//=============================================================================================
145	// I/O Syscall
146	//=============================================================================================
147	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
148	bool need_sys_to_submit = false;
149	bool need_sys_to_complete = false;
150	unsigned flags = 0;
151
152	TO_SUBMIT:
153	if( to_submit > 0 ) {
154	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
155	need_sys_to_submit = true;
156	break TO_SUBMIT;
157	}
158	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
159	need_sys_to_submit = true;
160	flags \|= IORING_ENTER_SQ_WAKEUP;
161	}
162	}
163
164	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
165	flags \|= IORING_ENTER_GETEVENTS;
166	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
167	need_sys_to_complete = true;
168	}
169	}
170
171	int ret = 0;
172	if( need_sys_to_submit \|\| need_sys_to_complete ) {
173	__cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags);
174	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8);
175	if( ret < 0 ) {
176	switch((int)errno) {
177	case EAGAIN:
178	case EINTR:
179	ret = -1;
180	break;
181	default:
182	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
183	}
184	}
185	}
186
187	// Memory barrier
188	__atomic_thread_fence( __ATOMIC_SEQ_CST );
189	return ret;
190	}
191
192	//=============================================================================================
193	// I/O Polling
194	//=============================================================================================
195	static unsigned __collect_submitions( struct __io_data & ring );
196	static __u32 __release_consumed_submission( struct __io_data & ring );
197	static inline void __clean( volatile struct io_uring_sqe * sqe );
198
199	// Process a single completion message from the io_uring
200	// This is NOT thread-safe
201	static inline void process( volatile struct io_uring_cqe & cqe ) {
202	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
203	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
204
205	fulfil( *future, cqe.res );
206	}
207
208	static [int, bool] __drain_io( & struct __io_data ring ) {
209	/* paranoid */ verify( ! __preemption_enabled() );
210
211	unsigned to_submit = 0;
212	if( ring.poller_submits ) {
213	// If the poller thread also submits, then we need to aggregate the submissions which are ready
214	to_submit = __collect_submitions( ring );
215	}
216
217	int ret = __io_uring_enter(ring, to_submit, true);
218	if( ret < 0 ) {
219	return [0, true];
220	}
221
222	// update statistics
223	if (to_submit > 0) {
224	__STATS__( true,
225	if( to_submit > 0 ) {
226	io.submit_q.submit_avg.rdy += to_submit;
227	io.submit_q.submit_avg.csm += ret;
228	io.submit_q.submit_avg.cnt += 1;
229	}
230	)
231	}
232
233	__atomic_thread_fence( __ATOMIC_SEQ_CST );
234
235	// Release the consumed SQEs
236	__release_consumed_submission( ring );
237
238	// Drain the queue
239	unsigned head = *ring.completion_q.head;
240	unsigned tail = *ring.completion_q.tail;
241	const __u32 mask = *ring.completion_q.mask;
242
243	// Nothing was new return 0
244	if (head == tail) {
245	return [0, to_submit > 0];
246	}
247
248	__u32 count = tail - head;
249	/* paranoid */ verify( count != 0 );
250	for(i; count) {
251	unsigned idx = (head + i) & mask;
252	volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
253
254	/* paranoid */ verify(&cqe);
255
256	process( cqe );
257	}
258
259	// Mark to the kernel that the cqe has been seen
260	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
261	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_SEQ_CST );
262
263	return [count, count > 0 \|\| to_submit > 0];
264	}
265
266	void main( $io_ctx_thread & this ) {
267	__ioctx_register( this );
268
269	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %d (%p) ready\n", this.ring->fd, &this);
270
271	const int reset_cnt = 5;
272	int reset = reset_cnt;
273	// Then loop until we need to start
274	LOOP:
275	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
276	// Drain the io
277	int count;
278	bool again;
279	disable_interrupts();
280	[count, again] = __drain_io( *this.ring );
281
282	if(!again) reset--;
283
284	// Update statistics
285	__STATS__( true,
286	io.complete_q.completed_avg.val += count;
287	io.complete_q.completed_avg.cnt += 1;
288	)
289	enable_interrupts( __cfaabi_dbg_ctx );
290
291	// If we got something, just yield and check again
292	if(reset > 1) {
293	yield();
294	continue LOOP;
295	}
296
297	// We alread failed to find completed entries a few time.
298	if(reset == 1) {
299	// Rearm the context so it can block
300	// but don't block right away
301	// we need to retry one last time in case
302	// something completed just now
303	__ioctx_prepare_block( this );
304	continue LOOP;
305	}
306
307	__STATS__( false,
308	io.complete_q.blocks += 1;
309	)
310	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %d (%p)\n", this.ring->fd, &this);
311
312	// block this thread
313	wait( this.sem );
314
315	// restore counter
316	reset = reset_cnt;
317	}
318
319	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller %d (%p) stopping\n", this.ring->fd, &this);
320	}
321
322	//=============================================================================================
323	// I/O Submissions
324	//=============================================================================================
325
326	// Submition steps :
327	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
328	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
329	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
330	// need to write an allocator that allows allocating concurrently.
331	//
332	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
333	//
334	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
335	// needs to arrive to two concensus at the same time:
336	// A - The order in which entries are listed in the array: no two threads must pick the
337	// same index for their entries
338	// B - When can the tail be update for the kernel. EVERY entries in the array between
339	// head and tail must be fully filled and shouldn't ever be touched again.
340	//
341
342	// Allocate an submit queue entry.
343	// The kernel cannot see these entries until they are submitted, but other threads must be
344	// able to see which entries can be used and which are already un used by an other thread
345	// for convenience, return both the index and the pointer to the sqe
346	// sqe == &sqes[idx]
347	[* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
348	/* paranoid */ verify( data != 0 );
349
350	// Prepare the data we need
351	__attribute((unused)) int len = 0;
352	__attribute((unused)) int block = 0;
353	__u32 cnt = *ring.submit_q.num;
354	__u32 mask = *ring.submit_q.mask;
355
356	__u32 off = thread_rand();
357
358	// Loop around looking for an available spot
359	for() {
360	// Look through the list starting at some offset
361	for(i; cnt) {
362	__u64 expected = 3;
363	__u32 idx = (i + off) & mask; // Get an index from a random
364	volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
365	volatile __u64 * udata = &sqe->user_data;
366
367	// Allocate the entry by CASing the user_data field from 0 to the future address
368	if( *udata == expected &&
369	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
370	{
371	// update statistics
372	__STATS__( false,
373	io.submit_q.alloc_avg.val += len;
374	io.submit_q.alloc_avg.block += block;
375	io.submit_q.alloc_avg.cnt += 1;
376	)
377
378	// debug log
379	__cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
380
381	// Success return the data
382	return [sqe, idx];
383	}
384	verify(expected != data);
385
386	// This one was used
387	len ++;
388	}
389
390	block++;
391
392	yield();
393	}
394	}
395
396	static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
397	/* paranoid */ verify( idx <= mask );
398	/* paranoid */ verify( idx != -1ul32 );
399
400	// We need to find a spot in the ready array
401	__attribute((unused)) int len = 0;
402	__attribute((unused)) int block = 0;
403	__u32 ready_mask = ring.submit_q.ready_cnt - 1;
404
405	__u32 off = thread_rand();
406
407	__u32 picked;
408	LOOKING: for() {
409	for(i; ring.submit_q.ready_cnt) {
410	picked = (i + off) & ready_mask;
411	__u32 expected = -1ul32;
412	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
413	break LOOKING;
414	}
415	verify(expected != idx);
416
417	len ++;
418	}
419
420	block++;
421
422	__u32 released = __release_consumed_submission( ring );
423	if( released == 0 ) {
424	yield();
425	}
426	}
427
428	// update statistics
429	__STATS__( false,
430	io.submit_q.look_avg.val += len;
431	io.submit_q.look_avg.block += block;
432	io.submit_q.look_avg.cnt += 1;
433	)
434
435	return picked;
436	}
437
438	void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
439	__io_data & ring = *ctx->thrd.ring;
440
441	{
442	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
443	__cfadbg_print_safe( io,
444	"Kernel I/O : submitting %u (%p) for %p\n"
445	" data: %p\n"
446	" opcode: %s\n"
447	" fd: %d\n"
448	" flags: %d\n"
449	" prio: %d\n"
450	" off: %p\n"
451	" addr: %p\n"
452	" len: %d\n"
453	" other flags: %d\n"
454	" splice fd: %d\n"
455	" pad[0]: %llu\n"
456	" pad[1]: %llu\n"
457	" pad[2]: %llu\n",
458	idx, sqe,
459	active_thread(),
460	(void*)sqe->user_data,
461	opcodes[sqe->opcode],
462	sqe->fd,
463	sqe->flags,
464	sqe->ioprio,
465	sqe->off,
466	sqe->addr,
467	sqe->len,
468	sqe->accept_flags,
469	sqe->splice_fd_in,
470	sqe->__pad2[0],
471	sqe->__pad2[1],
472	sqe->__pad2[2]
473	);
474	}
475
476
477	// Get now the data we definetely need
478	volatile __u32 * const tail = ring.submit_q.tail;
479	const __u32 mask = *ring.submit_q.mask;
480
481	// There are 2 submission schemes, check which one we are using
482	if( ring.poller_submits ) {
483	// If the poller thread submits, then we just need to add this to the ready array
484	__submit_to_ready_array( ring, idx, mask );
485
486	post( ctx->thrd.sem );
487
488	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
489	}
490	else if( ring.eager_submits ) {
491	__u32 picked = __submit_to_ready_array( ring, idx, mask );
492
493	#if defined(LEADER_LOCK)
494	if( !try_lock(ring.submit_q.submit_lock) ) {
495	__STATS__( false,
496	io.submit_q.helped += 1;
497	)
498	return;
499	}
500	/* paranoid */ verify( ! __preemption_enabled() );
501	__STATS__( true,
502	io.submit_q.leader += 1;
503	)
504	#else
505	for() {
506	yield();
507
508	if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
509	__STATS__( false,
510	io.submit_q.leader += 1;
511	)
512	break;
513	}
514
515	// If some one else collected our index, we are done
516	#warning ABA problem
517	if( ring.submit_q.ready[picked] != idx ) {
518	__STATS__( false,
519	io.submit_q.helped += 1;
520	)
521	return;
522	}
523
524	__STATS__( false,
525	io.submit_q.busy += 1;
526	)
527	}
528	#endif
529
530	// We got the lock
531	// Collect the submissions
532	unsigned to_submit = __collect_submitions( ring );
533
534	// Actually submit
535	int ret = __io_uring_enter( ring, to_submit, false );
536
537	#if defined(LEADER_LOCK)
538	/* paranoid */ verify( ! __preemption_enabled() );
539	next(ring.submit_q.submit_lock);
540	#else
541	unlock(ring.submit_q.submit_lock);
542	#endif
543	if( ret < 0 ) {
544	return;
545	}
546
547	// Release the consumed SQEs
548	__release_consumed_submission( ring );
549
550	// update statistics
551	__STATS__( false,
552	io.submit_q.submit_avg.rdy += to_submit;
553	io.submit_q.submit_avg.csm += ret;
554	io.submit_q.submit_avg.cnt += 1;
555	)
556
557	__cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
558	}
559	else
560	{
561	// get mutual exclusion
562	#if defined(LEADER_LOCK)
563	while(!try_lock(ring.submit_q.submit_lock));
564	#else
565	lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
566	#endif
567
568	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64,
569	/* paranoid */ "index %u already reclaimed\n"
570	/* paranoid */ "head %u, prev %u, tail %u\n"
571	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
572	/* paranoid */ idx,
573	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
574	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
575	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
576	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
577	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
578	/* paranoid */ );
579
580	// Append to the list of ready entries
581
582	/* paranoid */ verify( idx <= mask );
583	ring.submit_q.array[ (*tail) & mask ] = idx;
584	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
585
586	// Submit however, many entries need to be submitted
587	int ret = __io_uring_enter( ring, 1, false );
588	if( ret < 0 ) {
589	switch((int)errno) {
590	default:
591	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
592	}
593	}
594
595	/* paranoid */ verify(ret == 1);
596
597	// update statistics
598	__STATS__( false,
599	io.submit_q.submit_avg.csm += 1;
600	io.submit_q.submit_avg.cnt += 1;
601	)
602
603	{
604	__attribute__((unused)) volatile __u32 * const head = ring.submit_q.head;
605	__attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ];
606	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx];
607
608	__cfadbg_print_safe( io,
609	"Kernel I/O : last submitted is %u (%p)\n"
610	" data: %p\n"
611	" opcode: %s\n"
612	" fd: %d\n"
613	" flags: %d\n"
614	" prio: %d\n"
615	" off: %p\n"
616	" addr: %p\n"
617	" len: %d\n"
618	" other flags: %d\n"
619	" splice fd: %d\n"
620	" pad[0]: %llu\n"
621	" pad[1]: %llu\n"
622	" pad[2]: %llu\n",
623	last_idx, sqe,
624	(void*)sqe->user_data,
625	opcodes[sqe->opcode],
626	sqe->fd,
627	sqe->flags,
628	sqe->ioprio,
629	sqe->off,
630	sqe->addr,
631	sqe->len,
632	sqe->accept_flags,
633	sqe->splice_fd_in,
634	sqe->__pad2[0],
635	sqe->__pad2[1],
636	sqe->__pad2[2]
637	);
638	}
639
640	__atomic_thread_fence( __ATOMIC_SEQ_CST );
641	// Release the consumed SQEs
642	__release_consumed_submission( ring );
643	// ring.submit_q.sqes[idx].user_data = 3ul64;
644
645	#if defined(LEADER_LOCK)
646	next(ring.submit_q.submit_lock);
647	#else
648	unlock(ring.submit_q.submit_lock);
649	#endif
650
651	__cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() );
652	}
653	}
654
655	// #define PARTIAL_SUBMIT 32
656
657	// go through the list of submissions in the ready array and moved them into
658	// the ring's submit queue
659	static unsigned __collect_submitions( struct __io_data & ring ) {
660	/* paranoid */ verify( ring.submit_q.ready != 0p );
661	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
662
663	unsigned to_submit = 0;
664	__u32 tail = *ring.submit_q.tail;
665	const __u32 mask = *ring.submit_q.mask;
666	#if defined(PARTIAL_SUBMIT)
667	#if defined(LEADER_LOCK)
668	#error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
669	#endif
670	const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
671	const __u32 offset = ring.submit_q.prev_ready;
672	ring.submit_q.prev_ready += cnt;
673	#else
674	const __u32 cnt = ring.submit_q.ready_cnt;
675	const __u32 offset = 0;
676	#endif
677
678	// Go through the list of ready submissions
679	for( c; cnt ) {
680	__u32 i = (offset + c) % ring.submit_q.ready_cnt;
681
682	// replace any submission with the sentinel, to consume it.
683	__u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
684
685	// If it was already the sentinel, then we are done
686	if( idx == -1ul32 ) continue;
687
688	// If we got a real submission, append it to the list
689	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
690	to_submit++;
691	}
692
693	// Increment the tail based on how many we are ready to submit
694	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
695
696	return to_submit;
697	}
698
699	// Go through the ring's submit queue and release everything that has already been consumed
700	// by io_uring
701	static __u32 __release_consumed_submission( struct __io_data & ring ) {
702	const __u32 smask = *ring.submit_q.mask;
703
704	// We need to get the lock to copy the old head and new head
705	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
706	__attribute__((unused))
707	__u32 ctail = *ring.submit_q.tail; // get the current tail of the queue
708	__u32 chead = *ring.submit_q.head; // get the current head of the queue
709	__u32 phead = ring.submit_q.prev_head; // get the head the last time we were here
710	ring.submit_q.prev_head = chead; // note up to were we processed
711	unlock(ring.submit_q.release_lock);
712
713	// the 3 fields are organized like this diagram
714	// except it's are ring
715	// ---+--------+--------+----
716	// ---+--------+--------+----
717	// ^ ^ ^
718	// phead chead ctail
719
720	// make sure ctail doesn't wrap around and reach phead
721	/* paranoid */ verify(
722	(ctail >= chead && chead >= phead)
723	\|\| (chead >= phead && phead >= ctail)
724	\|\| (phead >= ctail && ctail >= chead)
725	);
726
727	// find the range we need to clear
728	__u32 count = chead - phead;
729
730	// We acquired an previous-head/current-head range
731	// go through the range and release the sqes
732	for( i; count ) {
733	__u32 idx = ring.submit_q.array[ (phead + i) & smask ];
734
735	/* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data );
736	__clean( &ring.submit_q.sqes[ idx ] );
737	}
738	return count;
739	}
740
741	void __sqe_clean( volatile struct io_uring_sqe * sqe ) {
742	__clean( sqe );
743	}
744
745	static inline void __clean( volatile struct io_uring_sqe * sqe ) {
746	// If we are in debug mode, thrash the fields to make sure we catch reclamation errors
747	__cfaabi_dbg_debug_do(
748	memset(sqe, 0xde, sizeof(*sqe));
749	sqe->opcode = (sizeof(opcodes) / sizeof(const char *)) - 1;
750	);
751
752	// Mark the entry as unused
753	__atomic_store_n(&sqe->user_data, 3ul64, __ATOMIC_SEQ_CST);
754	}
755	#endif

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

Download in other formats:

Original Format