Context Navigation

io.cfa@ 4fc3343

Visit:

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum

Last change on this file since 4fc3343 was d611995, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
Fix crash on cluster exit. (Added epoch based blocking to iopoll.)
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
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	__ioctx_unregister( this );
322	}
323
324	//=============================================================================================
325	// I/O Submissions
326	//=============================================================================================
327
328	// Submition steps :
329	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
330	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
331	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
332	// need to write an allocator that allows allocating concurrently.
333	//
334	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
335	//
336	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
337	// needs to arrive to two concensus at the same time:
338	// A - The order in which entries are listed in the array: no two threads must pick the
339	// same index for their entries
340	// B - When can the tail be update for the kernel. EVERY entries in the array between
341	// head and tail must be fully filled and shouldn't ever be touched again.
342	//
343
344	// Allocate an submit queue entry.
345	// The kernel cannot see these entries until they are submitted, but other threads must be
346	// able to see which entries can be used and which are already un used by an other thread
347	// for convenience, return both the index and the pointer to the sqe
348	// sqe == &sqes[idx]
349	[* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
350	/* paranoid */ verify( data != 0 );
351
352	// Prepare the data we need
353	__attribute((unused)) int len = 0;
354	__attribute((unused)) int block = 0;
355	__u32 cnt = *ring.submit_q.num;
356	__u32 mask = *ring.submit_q.mask;
357
358	__u32 off = thread_rand();
359
360	// Loop around looking for an available spot
361	for() {
362	// Look through the list starting at some offset
363	for(i; cnt) {
364	__u64 expected = 3;
365	__u32 idx = (i + off) & mask; // Get an index from a random
366	volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
367	volatile __u64 * udata = &sqe->user_data;
368
369	// Allocate the entry by CASing the user_data field from 0 to the future address
370	if( *udata == expected &&
371	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
372	{
373	// update statistics
374	__STATS__( false,
375	io.submit_q.alloc_avg.val += len;
376	io.submit_q.alloc_avg.block += block;
377	io.submit_q.alloc_avg.cnt += 1;
378	)
379
380	// debug log
381	__cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
382
383	// Success return the data
384	return [sqe, idx];
385	}
386	verify(expected != data);
387
388	// This one was used
389	len ++;
390	}
391
392	block++;
393
394	yield();
395	}
396	}
397
398	static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
399	/* paranoid */ verify( idx <= mask );
400	/* paranoid */ verify( idx != -1ul32 );
401
402	// We need to find a spot in the ready array
403	__attribute((unused)) int len = 0;
404	__attribute((unused)) int block = 0;
405	__u32 ready_mask = ring.submit_q.ready_cnt - 1;
406
407	__u32 off = thread_rand();
408
409	__u32 picked;
410	LOOKING: for() {
411	for(i; ring.submit_q.ready_cnt) {
412	picked = (i + off) & ready_mask;
413	__u32 expected = -1ul32;
414	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
415	break LOOKING;
416	}
417	verify(expected != idx);
418
419	len ++;
420	}
421
422	block++;
423
424	__u32 released = __release_consumed_submission( ring );
425	if( released == 0 ) {
426	yield();
427	}
428	}
429
430	// update statistics
431	__STATS__( false,
432	io.submit_q.look_avg.val += len;
433	io.submit_q.look_avg.block += block;
434	io.submit_q.look_avg.cnt += 1;
435	)
436
437	return picked;
438	}
439
440	void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
441	__io_data & ring = *ctx->thrd.ring;
442
443	{
444	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
445	__cfadbg_print_safe( io,
446	"Kernel I/O : submitting %u (%p) for %p\n"
447	" data: %p\n"
448	" opcode: %s\n"
449	" fd: %d\n"
450	" flags: %d\n"
451	" prio: %d\n"
452	" off: %p\n"
453	" addr: %p\n"
454	" len: %d\n"
455	" other flags: %d\n"
456	" splice fd: %d\n"
457	" pad[0]: %llu\n"
458	" pad[1]: %llu\n"
459	" pad[2]: %llu\n",
460	idx, sqe,
461	active_thread(),
462	(void*)sqe->user_data,
463	opcodes[sqe->opcode],
464	sqe->fd,
465	sqe->flags,
466	sqe->ioprio,
467	sqe->off,
468	sqe->addr,
469	sqe->len,
470	sqe->accept_flags,
471	sqe->splice_fd_in,
472	sqe->__pad2[0],
473	sqe->__pad2[1],
474	sqe->__pad2[2]
475	);
476	}
477
478
479	// Get now the data we definetely need
480	volatile __u32 * const tail = ring.submit_q.tail;
481	const __u32 mask = *ring.submit_q.mask;
482
483	// There are 2 submission schemes, check which one we are using
484	if( ring.poller_submits ) {
485	// If the poller thread submits, then we just need to add this to the ready array
486	__submit_to_ready_array( ring, idx, mask );
487
488	post( ctx->thrd.sem );
489
490	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
491	}
492	else if( ring.eager_submits ) {
493	__u32 picked = __submit_to_ready_array( ring, idx, mask );
494
495	#if defined(LEADER_LOCK)
496	if( !try_lock(ring.submit_q.submit_lock) ) {
497	__STATS__( false,
498	io.submit_q.helped += 1;
499	)
500	return;
501	}
502	/* paranoid */ verify( ! __preemption_enabled() );
503	__STATS__( true,
504	io.submit_q.leader += 1;
505	)
506	#else
507	for() {
508	yield();
509
510	if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
511	__STATS__( false,
512	io.submit_q.leader += 1;
513	)
514	break;
515	}
516
517	// If some one else collected our index, we are done
518	#warning ABA problem
519	if( ring.submit_q.ready[picked] != idx ) {
520	__STATS__( false,
521	io.submit_q.helped += 1;
522	)
523	return;
524	}
525
526	__STATS__( false,
527	io.submit_q.busy += 1;
528	)
529	}
530	#endif
531
532	// We got the lock
533	// Collect the submissions
534	unsigned to_submit = __collect_submitions( ring );
535
536	// Actually submit
537	int ret = __io_uring_enter( ring, to_submit, false );
538
539	#if defined(LEADER_LOCK)
540	/* paranoid */ verify( ! __preemption_enabled() );
541	next(ring.submit_q.submit_lock);
542	#else
543	unlock(ring.submit_q.submit_lock);
544	#endif
545	if( ret < 0 ) {
546	return;
547	}
548
549	// Release the consumed SQEs
550	__release_consumed_submission( ring );
551
552	// update statistics
553	__STATS__( false,
554	io.submit_q.submit_avg.rdy += to_submit;
555	io.submit_q.submit_avg.csm += ret;
556	io.submit_q.submit_avg.cnt += 1;
557	)
558
559	__cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
560	}
561	else
562	{
563	// get mutual exclusion
564	#if defined(LEADER_LOCK)
565	while(!try_lock(ring.submit_q.submit_lock));
566	#else
567	lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
568	#endif
569
570	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64,
571	/* paranoid */ "index %u already reclaimed\n"
572	/* paranoid */ "head %u, prev %u, tail %u\n"
573	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
574	/* paranoid */ idx,
575	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
576	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
577	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
578	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
579	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
580	/* paranoid */ );
581
582	// Append to the list of ready entries
583
584	/* paranoid */ verify( idx <= mask );
585	ring.submit_q.array[ (*tail) & mask ] = idx;
586	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
587
588	// Submit however, many entries need to be submitted
589	int ret = __io_uring_enter( ring, 1, false );
590	if( ret < 0 ) {
591	switch((int)errno) {
592	default:
593	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
594	}
595	}
596
597	/* paranoid */ verify(ret == 1);
598
599	// update statistics
600	__STATS__( false,
601	io.submit_q.submit_avg.csm += 1;
602	io.submit_q.submit_avg.cnt += 1;
603	)
604
605	{
606	__attribute__((unused)) volatile __u32 * const head = ring.submit_q.head;
607	__attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ];
608	__attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx];
609
610	__cfadbg_print_safe( io,
611	"Kernel I/O : last submitted is %u (%p)\n"
612	" data: %p\n"
613	" opcode: %s\n"
614	" fd: %d\n"
615	" flags: %d\n"
616	" prio: %d\n"
617	" off: %p\n"
618	" addr: %p\n"
619	" len: %d\n"
620	" other flags: %d\n"
621	" splice fd: %d\n"
622	" pad[0]: %llu\n"
623	" pad[1]: %llu\n"
624	" pad[2]: %llu\n",
625	last_idx, sqe,
626	(void*)sqe->user_data,
627	opcodes[sqe->opcode],
628	sqe->fd,
629	sqe->flags,
630	sqe->ioprio,
631	sqe->off,
632	sqe->addr,
633	sqe->len,
634	sqe->accept_flags,
635	sqe->splice_fd_in,
636	sqe->__pad2[0],
637	sqe->__pad2[1],
638	sqe->__pad2[2]
639	);
640	}
641
642	__atomic_thread_fence( __ATOMIC_SEQ_CST );
643	// Release the consumed SQEs
644	__release_consumed_submission( ring );
645	// ring.submit_q.sqes[idx].user_data = 3ul64;
646
647	#if defined(LEADER_LOCK)
648	next(ring.submit_q.submit_lock);
649	#else
650	unlock(ring.submit_q.submit_lock);
651	#endif
652
653	__cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() );
654	}
655	}
656
657	// #define PARTIAL_SUBMIT 32
658
659	// go through the list of submissions in the ready array and moved them into
660	// the ring's submit queue
661	static unsigned __collect_submitions( struct __io_data & ring ) {
662	/* paranoid */ verify( ring.submit_q.ready != 0p );
663	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
664
665	unsigned to_submit = 0;
666	__u32 tail = *ring.submit_q.tail;
667	const __u32 mask = *ring.submit_q.mask;
668	#if defined(PARTIAL_SUBMIT)
669	#if defined(LEADER_LOCK)
670	#error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
671	#endif
672	const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
673	const __u32 offset = ring.submit_q.prev_ready;
674	ring.submit_q.prev_ready += cnt;
675	#else
676	const __u32 cnt = ring.submit_q.ready_cnt;
677	const __u32 offset = 0;
678	#endif
679
680	// Go through the list of ready submissions
681	for( c; cnt ) {
682	__u32 i = (offset + c) % ring.submit_q.ready_cnt;
683
684	// replace any submission with the sentinel, to consume it.
685	__u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
686
687	// If it was already the sentinel, then we are done
688	if( idx == -1ul32 ) continue;
689
690	// If we got a real submission, append it to the list
691	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
692	to_submit++;
693	}
694
695	// Increment the tail based on how many we are ready to submit
696	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
697
698	return to_submit;
699	}
700
701	// Go through the ring's submit queue and release everything that has already been consumed
702	// by io_uring
703	static __u32 __release_consumed_submission( struct __io_data & ring ) {
704	const __u32 smask = *ring.submit_q.mask;
705
706	// We need to get the lock to copy the old head and new head
707	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
708	__attribute__((unused))
709	__u32 ctail = *ring.submit_q.tail; // get the current tail of the queue
710	__u32 chead = *ring.submit_q.head; // get the current head of the queue
711	__u32 phead = ring.submit_q.prev_head; // get the head the last time we were here
712	ring.submit_q.prev_head = chead; // note up to were we processed
713	unlock(ring.submit_q.release_lock);
714
715	// the 3 fields are organized like this diagram
716	// except it's are ring
717	// ---+--------+--------+----
718	// ---+--------+--------+----
719	// ^ ^ ^
720	// phead chead ctail
721
722	// make sure ctail doesn't wrap around and reach phead
723	/* paranoid */ verify(
724	(ctail >= chead && chead >= phead)
725	\|\| (chead >= phead && phead >= ctail)
726	\|\| (phead >= ctail && ctail >= chead)
727	);
728
729	// find the range we need to clear
730	__u32 count = chead - phead;
731
732	// We acquired an previous-head/current-head range
733	// go through the range and release the sqes
734	for( i; count ) {
735	__u32 idx = ring.submit_q.array[ (phead + i) & smask ];
736
737	/* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data );
738	__clean( &ring.submit_q.sqes[ idx ] );
739	}
740	return count;
741	}
742
743	void __sqe_clean( volatile struct io_uring_sqe * sqe ) {
744	__clean( sqe );
745	}
746
747	static inline void __clean( volatile struct io_uring_sqe * sqe ) {
748	// If we are in debug mode, thrash the fields to make sure we catch reclamation errors
749	__cfaabi_dbg_debug_do(
750	memset(sqe, 0xde, sizeof(*sqe));
751	sqe->opcode = (sizeof(opcodes) / sizeof(const char *)) - 1;
752	);
753
754	// Mark the entry as unused
755	__atomic_store_n(&sqe->user_data, 3ul64, __ATOMIC_SEQ_CST);
756	}
757	#endif

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

Download in other formats:

Original Format