Context Navigation

io.cfa @ 77fde9d5

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

Last change on this file since 77fde9d5 was ece0e80, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Added prints. Naive implementation of cancel. Server now shutdown cleanly.
Property mode set to `100644`
File size: 16.8 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/syscall.h>
35
36	#include <linux/io_uring.h>
37	}
38
39	#include "stats.hfa"
40	#include "kernel.hfa"
41	#include "kernel/fwd.hfa"
42	#include "io/types.hfa"
43
44	// returns true of acquired as leader or second leader
45	static inline bool try_lock( __leaderlock_t & this ) {
46	const uintptr_t thrd = 1z \| (uintptr_t)active_thread();
47	bool block;
48	disable_interrupts();
49	for() {
50	struct $thread * expected = this.value;
51	if( 1p != expected && 0p != expected ) {
52	/* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader
53	enable_interrupts( __cfaabi_dbg_ctx );
54	return false;
55	}
56	struct $thread * desired;
57	if( 0p == expected ) {
58	// If the lock isn't locked acquire it, no need to block
59	desired = 1p;
60	block = false;
61	}
62	else {
63	// If the lock is already locked try becomming the next leader
64	desired = (struct $thread *)thrd;
65	block = true;
66	}
67	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
68	}
69	if( block ) {
70	enable_interrupts( __cfaabi_dbg_ctx );
71	park();
72	disable_interrupts();
73	}
74	return true;
75	}
76
77	static inline bool next( __leaderlock_t & this ) {
78	/* paranoid */ verify( ! __preemption_enabled() );
79	struct $thread * nextt;
80	for() {
81	struct $thread * expected = this.value;
82	/* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked
83
84	struct $thread * desired;
85	if( 1p == expected ) {
86	// No next leader, just unlock
87	desired = 0p;
88	nextt = 0p;
89	}
90	else {
91	// There is a next leader, remove but keep locked
92	desired = 1p;
93	nextt = (struct $thread *)(~1z & (uintptr_t)expected);
94	}
95	if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
96	}
97
98	if(nextt) {
99	unpark( nextt );
100	enable_interrupts( __cfaabi_dbg_ctx );
101	return true;
102	}
103	enable_interrupts( __cfaabi_dbg_ctx );
104	return false;
105	}
106
107	//=============================================================================================
108	// I/O Syscall
109	//=============================================================================================
110	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
111	bool need_sys_to_submit = false;
112	bool need_sys_to_complete = false;
113	unsigned flags = 0;
114
115	TO_SUBMIT:
116	if( to_submit > 0 ) {
117	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
118	need_sys_to_submit = true;
119	break TO_SUBMIT;
120	}
121	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
122	need_sys_to_submit = true;
123	flags \|= IORING_ENTER_SQ_WAKEUP;
124	}
125	}
126
127	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
128	flags \|= IORING_ENTER_GETEVENTS;
129	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
130	need_sys_to_complete = true;
131	}
132	}
133
134	int ret = 0;
135	if( need_sys_to_submit \|\| need_sys_to_complete ) {
136	__cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags);
137	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8);
138	if( ret < 0 ) {
139	switch((int)errno) {
140	case EAGAIN:
141	case EINTR:
142	ret = -1;
143	break;
144	default:
145	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
146	}
147	}
148	}
149
150	// Memory barrier
151	__atomic_thread_fence( __ATOMIC_SEQ_CST );
152	return ret;
153	}
154
155	//=============================================================================================
156	// I/O Polling
157	//=============================================================================================
158	static unsigned __collect_submitions( struct __io_data & ring );
159	static __u32 __release_consumed_submission( struct __io_data & ring );
160
161	static inline void process(struct io_uring_cqe & cqe ) {
162	struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
163	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
164
165	fulfil( *future, cqe.res );
166	}
167
168	// Process a single completion message from the io_uring
169	// This is NOT thread-safe
170	static [int, bool] __drain_io( & struct __io_data ring ) {
171	/* paranoid */ verify( ! __preemption_enabled() );
172
173	unsigned to_submit = 0;
174	if( ring.poller_submits ) {
175	// If the poller thread also submits, then we need to aggregate the submissions which are ready
176	to_submit = __collect_submitions( ring );
177	}
178
179	int ret = __io_uring_enter(ring, to_submit, true);
180	if( ret < 0 ) {
181	return [0, true];
182	}
183
184	// update statistics
185	if (to_submit > 0) {
186	__STATS__( true,
187	if( to_submit > 0 ) {
188	io.submit_q.submit_avg.rdy += to_submit;
189	io.submit_q.submit_avg.csm += ret;
190	io.submit_q.submit_avg.cnt += 1;
191	}
192	)
193	}
194
195	// Release the consumed SQEs
196	__release_consumed_submission( ring );
197
198	// Drain the queue
199	unsigned head = *ring.completion_q.head;
200	unsigned tail = *ring.completion_q.tail;
201	const __u32 mask = *ring.completion_q.mask;
202
203	// Nothing was new return 0
204	if (head == tail) {
205	return [0, to_submit > 0];
206	}
207
208	__u32 count = tail - head;
209	/* paranoid */ verify( count != 0 );
210	for(i; count) {
211	unsigned idx = (head + i) & mask;
212	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
213
214	/* paranoid */ verify(&cqe);
215
216	process( cqe );
217	}
218
219	// Mark to the kernel that the cqe has been seen
220	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
221	__atomic_thread_fence( __ATOMIC_SEQ_CST );
222	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
223
224	return [count, count > 0 \|\| to_submit > 0];
225	}
226
227	void main( $io_ctx_thread & this ) {
228	epoll_event ev;
229	__ioctx_register( this, ev );
230
231	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %p for ring %p ready\n", &this, &this.ring);
232
233	const int reset_cnt = 5;
234	int reset = reset_cnt;
235	// Then loop until we need to start
236	LOOP:
237	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
238	// Drain the io
239	int count;
240	bool again;
241	disable_interrupts();
242	[count, again] = __drain_io( *this.ring );
243
244	if(!again) reset--;
245
246	// Update statistics
247	__STATS__( true,
248	io.complete_q.completed_avg.val += count;
249	io.complete_q.completed_avg.cnt += 1;
250	)
251	enable_interrupts( __cfaabi_dbg_ctx );
252
253	// If we got something, just yield and check again
254	if(reset > 1) {
255	yield();
256	continue LOOP;
257	}
258
259	// We alread failed to find events a few time.
260	if(reset == 1) {
261	// Rearm the context so it can block
262	// but don't block right away
263	// we need to retry one last time in case
264	// something completed just now
265	__ioctx_prepare_block( this, ev );
266	continue LOOP;
267	}
268
269	__STATS__( false,
270	io.complete_q.blocks += 1;
271	)
272	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %p\n", &this.self);
273
274	// block this thread
275	wait( this.sem );
276
277	// restore counter
278	reset = reset_cnt;
279	}
280
281	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
282	}
283
284	//=============================================================================================
285	// I/O Submissions
286	//=============================================================================================
287
288	// Submition steps :
289	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
290	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
291	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
292	// need to write an allocator that allows allocating concurrently.
293	//
294	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
295	//
296	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
297	// needs to arrive to two concensus at the same time:
298	// A - The order in which entries are listed in the array: no two threads must pick the
299	// same index for their entries
300	// B - When can the tail be update for the kernel. EVERY entries in the array between
301	// head and tail must be fully filled and shouldn't ever be touched again.
302	//
303
304	[* struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
305	/* paranoid */ verify( data != 0 );
306
307	// Prepare the data we need
308	__attribute((unused)) int len = 0;
309	__attribute((unused)) int block = 0;
310	__u32 cnt = *ring.submit_q.num;
311	__u32 mask = *ring.submit_q.mask;
312
313	__u32 off = thread_rand();
314
315	// Loop around looking for an available spot
316	for() {
317	// Look through the list starting at some offset
318	for(i; cnt) {
319	__u64 expected = 0;
320	__u32 idx = (i + off) & mask;
321	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
322	volatile __u64 * udata = &sqe->user_data;
323
324	if( *udata == expected &&
325	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
326	{
327	// update statistics
328	__STATS__( false,
329	io.submit_q.alloc_avg.val += len;
330	io.submit_q.alloc_avg.block += block;
331	io.submit_q.alloc_avg.cnt += 1;
332	)
333
334	__cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
335
336	// Success return the data
337	return [sqe, idx];
338	}
339	verify(expected != data);
340
341	len ++;
342	}
343
344	block++;
345	yield();
346	}
347	}
348
349	static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
350	/* paranoid */ verify( idx <= mask );
351	/* paranoid */ verify( idx != -1ul32 );
352
353	// We need to find a spot in the ready array
354	__attribute((unused)) int len = 0;
355	__attribute((unused)) int block = 0;
356	__u32 ready_mask = ring.submit_q.ready_cnt - 1;
357
358	__u32 off = thread_rand();
359
360	__u32 picked;
361	LOOKING: for() {
362	for(i; ring.submit_q.ready_cnt) {
363	picked = (i + off) & ready_mask;
364	__u32 expected = -1ul32;
365	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
366	break LOOKING;
367	}
368	verify(expected != idx);
369
370	len ++;
371	}
372
373	block++;
374
375	__u32 released = __release_consumed_submission( ring );
376	if( released == 0 ) {
377	yield();
378	}
379	}
380
381	// update statistics
382	__STATS__( false,
383	io.submit_q.look_avg.val += len;
384	io.submit_q.look_avg.block += block;
385	io.submit_q.look_avg.cnt += 1;
386	)
387
388	return picked;
389	}
390
391	void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
392	__cfadbg_print_safe( io, "Kernel I/O : submitting %u for %p\n", idx, active_thread() );
393
394	__io_data & ring = *ctx->thrd.ring;
395	// Get now the data we definetely need
396	volatile __u32 * const tail = ring.submit_q.tail;
397	const __u32 mask = *ring.submit_q.mask;
398
399	// There are 2 submission schemes, check which one we are using
400	if( ring.poller_submits ) {
401	// If the poller thread submits, then we just need to add this to the ready array
402	__submit_to_ready_array( ring, idx, mask );
403
404	post( ctx->thrd.sem );
405
406	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
407	}
408	else if( ring.eager_submits ) {
409	__u32 picked = __submit_to_ready_array( ring, idx, mask );
410
411	#if defined(LEADER_LOCK)
412	if( !try_lock(ring.submit_q.submit_lock) ) {
413	__STATS__( false,
414	io.submit_q.helped += 1;
415	)
416	return;
417	}
418	/* paranoid */ verify( ! __preemption_enabled() );
419	__STATS__( true,
420	io.submit_q.leader += 1;
421	)
422	#else
423	for() {
424	yield();
425
426	if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
427	__STATS__( false,
428	io.submit_q.leader += 1;
429	)
430	break;
431	}
432
433	// If some one else collected our index, we are done
434	#warning ABA problem
435	if( ring.submit_q.ready[picked] != idx ) {
436	__STATS__( false,
437	io.submit_q.helped += 1;
438	)
439	return;
440	}
441
442	__STATS__( false,
443	io.submit_q.busy += 1;
444	)
445	}
446	#endif
447
448	// We got the lock
449	// Collect the submissions
450	unsigned to_submit = __collect_submitions( ring );
451
452	// Actually submit
453	int ret = __io_uring_enter( ring, to_submit, false );
454
455	#if defined(LEADER_LOCK)
456	/* paranoid */ verify( ! __preemption_enabled() );
457	next(ring.submit_q.submit_lock);
458	#else
459	unlock(ring.submit_q.submit_lock);
460	#endif
461	if( ret < 0 ) {
462	return;
463	}
464
465	// Release the consumed SQEs
466	__release_consumed_submission( ring );
467
468	// update statistics
469	__STATS__( false,
470	io.submit_q.submit_avg.rdy += to_submit;
471	io.submit_q.submit_avg.csm += ret;
472	io.submit_q.submit_avg.cnt += 1;
473	)
474
475	__cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
476	}
477	else {
478	// get mutual exclusion
479	#if defined(LEADER_LOCK)
480	while(!try_lock(ring.submit_q.submit_lock));
481	#else
482	lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
483	#endif
484
485	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 0,
486	/* paranoid */ "index %u already reclaimed\n"
487	/* paranoid */ "head %u, prev %u, tail %u\n"
488	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
489	/* paranoid */ idx,
490	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
491	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
492	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
493	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
494	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
495	/* paranoid */ );
496
497	// Append to the list of ready entries
498
499	/* paranoid */ verify( idx <= mask );
500	ring.submit_q.array[ (*tail) & mask ] = idx;
501	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
502
503	// Submit however, many entries need to be submitted
504	int ret = __io_uring_enter( ring, 1, false );
505	if( ret < 0 ) {
506	switch((int)errno) {
507	default:
508	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
509	}
510	}
511
512	// update statistics
513	__STATS__( false,
514	io.submit_q.submit_avg.csm += 1;
515	io.submit_q.submit_avg.cnt += 1;
516	)
517
518	// Release the consumed SQEs
519	__release_consumed_submission( ring );
520
521	#if defined(LEADER_LOCK)
522	next(ring.submit_q.submit_lock);
523	#else
524	unlock(ring.submit_q.submit_lock);
525	#endif
526
527	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
528	}
529	}
530
531	// #define PARTIAL_SUBMIT 32
532	static unsigned __collect_submitions( struct __io_data & ring ) {
533	/* paranoid */ verify( ring.submit_q.ready != 0p );
534	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
535
536	unsigned to_submit = 0;
537	__u32 tail = *ring.submit_q.tail;
538	const __u32 mask = *ring.submit_q.mask;
539	#if defined(PARTIAL_SUBMIT)
540	#if defined(LEADER_LOCK)
541	#error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
542	#endif
543	const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
544	const __u32 offset = ring.submit_q.prev_ready;
545	ring.submit_q.prev_ready += cnt;
546	#else
547	const __u32 cnt = ring.submit_q.ready_cnt;
548	const __u32 offset = 0;
549	#endif
550
551	// Go through the list of ready submissions
552	for( c; cnt ) {
553	__u32 i = (offset + c) % ring.submit_q.ready_cnt;
554
555	// replace any submission with the sentinel, to consume it.
556	__u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
557
558	// If it was already the sentinel, then we are done
559	if( idx == -1ul32 ) continue;
560
561	// If we got a real submission, append it to the list
562	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
563	to_submit++;
564	}
565
566	// Increment the tail based on how many we are ready to submit
567	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
568
569	return to_submit;
570	}
571
572	static __u32 __release_consumed_submission( struct __io_data & ring ) {
573	const __u32 smask = *ring.submit_q.mask;
574
575	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
576	__u32 chead = *ring.submit_q.head;
577	__u32 phead = ring.submit_q.prev_head;
578	ring.submit_q.prev_head = chead;
579	unlock(ring.submit_q.release_lock);
580
581	__u32 count = chead - phead;
582	for( i; count ) {
583	__u32 idx = ring.submit_q.array[ (phead + i) & smask ];
584	ring.submit_q.sqes[ idx ].user_data = 0;
585	}
586	return count;
587	}
588	#endif

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

Download in other formats:

Original Format