Context Navigation

io.cfa@ 7f6e9eb

Visit:

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

Last change on this file since 7f6e9eb was be36ec3, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
No longer hold the submit lock when doing the io_uring_enter system call
Property mode set to `100644`
File size: 13.6 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	//=============================================================================================
45	// I/O Syscall
46	//=============================================================================================
47	static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
48	bool need_sys_to_submit = false;
49	bool need_sys_to_complete = false;
50	unsigned flags = 0;
51
52	TO_SUBMIT:
53	if( to_submit > 0 ) {
54	if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
55	need_sys_to_submit = true;
56	break TO_SUBMIT;
57	}
58	if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
59	need_sys_to_submit = true;
60	flags \|= IORING_ENTER_SQ_WAKEUP;
61	}
62	}
63
64	if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
65	flags \|= IORING_ENTER_GETEVENTS;
66	if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
67	need_sys_to_complete = true;
68	}
69	}
70
71	int ret = 0;
72	if( need_sys_to_submit \|\| need_sys_to_complete ) {
73	ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, 0p, _NSIG / 8);
74	if( ret < 0 ) {
75	switch((int)errno) {
76	case EAGAIN:
77	case EINTR:
78	ret = -1;
79	break;
80	default:
81	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
82	}
83	}
84	}
85
86	// Memory barrier
87	__atomic_thread_fence( __ATOMIC_SEQ_CST );
88	return ret;
89	}
90
91	//=============================================================================================
92	// I/O Polling
93	//=============================================================================================
94	static unsigned __collect_submitions( struct __io_data & ring );
95	static uint32_t __release_consumed_submission( struct __io_data & ring );
96
97	static inline void process(struct io_uring_cqe & cqe ) {
98	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
99	__cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
100
101	data->result = cqe.res;
102	unpark( data->thrd __cfaabi_dbg_ctx2 );
103	}
104
105	// Process a single completion message from the io_uring
106	// This is NOT thread-safe
107	static [int, bool] __drain_io( & struct __io_data ring ) {
108	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
109
110	unsigned to_submit = 0;
111	if( ring.poller_submits ) {
112	// If the poller thread also submits, then we need to aggregate the submissions which are ready
113	to_submit = __collect_submitions( ring );
114	}
115
116	int ret = __io_uring_enter(ring, to_submit, true);
117	if( ret < 0 ) {
118	return [0, true];
119	}
120
121	// update statistics
122	if (to_submit > 0) {
123	__STATS__( true,
124	if( to_submit > 0 ) {
125	io.submit_q.submit_avg.rdy += to_submit;
126	io.submit_q.submit_avg.csm += ret;
127	io.submit_q.submit_avg.cnt += 1;
128	}
129	)
130	}
131
132	// Release the consumed SQEs
133	__release_consumed_submission( ring );
134
135	// Drain the queue
136	unsigned head = *ring.completion_q.head;
137	unsigned tail = *ring.completion_q.tail;
138	const uint32_t mask = *ring.completion_q.mask;
139
140	// Nothing was new return 0
141	if (head == tail) {
142	return [0, to_submit > 0];
143	}
144
145	uint32_t count = tail - head;
146	/* paranoid */ verify( count != 0 );
147	for(i; count) {
148	unsigned idx = (head + i) & mask;
149	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
150
151	/* paranoid */ verify(&cqe);
152
153	process( cqe );
154	}
155
156	// Mark to the kernel that the cqe has been seen
157	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
158	__atomic_thread_fence( __ATOMIC_SEQ_CST );
159	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
160
161	return [count, count > 0 \|\| to_submit > 0];
162	}
163
164	void main( $io_ctx_thread & this ) {
165	epoll_event ev;
166	__ioctx_register( this, ev );
167
168	__cfadbg_print_safe(io_core, "Kernel I/O : IO poller %p for ring %p ready\n", &this, &this.ring);
169
170	int reset = 0;
171	// Then loop until we need to start
172	while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
173	// Drain the io
174	int count;
175	bool again;
176	disable_interrupts();
177	[count, again] = __drain_io( *this.ring );
178
179	if(!again) reset++;
180
181	// Update statistics
182	__STATS__( true,
183	io.complete_q.completed_avg.val += count;
184	io.complete_q.completed_avg.fast_cnt += 1;
185	)
186	enable_interrupts( __cfaabi_dbg_ctx );
187
188	// If we got something, just yield and check again
189	if(reset < 5) {
190	yield();
191	}
192	// We didn't get anything baton pass to the slow poller
193	else {
194	__cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %p\n", &this.self);
195	reset = 0;
196
197	// block this thread
198	__ioctx_prepare_block( this, ev );
199	wait( this.sem );
200	}
201	}
202
203	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
204	}
205
206	//=============================================================================================
207	// I/O Submissions
208	//=============================================================================================
209
210	// Submition steps :
211	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
212	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
213	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
214	// need to write an allocator that allows allocating concurrently.
215	//
216	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
217	//
218	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
219	// needs to arrive to two concensus at the same time:
220	// A - The order in which entries are listed in the array: no two threads must pick the
221	// same index for their entries
222	// B - When can the tail be update for the kernel. EVERY entries in the array between
223	// head and tail must be fully filled and shouldn't ever be touched again.
224	//
225
226	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
227	/* paranoid */ verify( data != 0 );
228
229	// Prepare the data we need
230	__attribute((unused)) int len = 0;
231	__attribute((unused)) int block = 0;
232	uint32_t cnt = *ring.submit_q.num;
233	uint32_t mask = *ring.submit_q.mask;
234
235	disable_interrupts();
236	uint32_t off = __tls_rand();
237	enable_interrupts( __cfaabi_dbg_ctx );
238
239	// Loop around looking for an available spot
240	for() {
241	// Look through the list starting at some offset
242	for(i; cnt) {
243	uint64_t expected = 0;
244	uint32_t idx = (i + off) & mask;
245	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
246	volatile uint64_t * udata = (volatile uint64_t *)&sqe->user_data;
247
248	if( *udata == expected &&
249	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
250	{
251	// update statistics
252	__STATS__( false,
253	io.submit_q.alloc_avg.val += len;
254	io.submit_q.alloc_avg.block += block;
255	io.submit_q.alloc_avg.cnt += 1;
256	)
257
258
259	// Success return the data
260	return [sqe, idx];
261	}
262	verify(expected != data);
263
264	len ++;
265	}
266
267	block++;
268	yield();
269	}
270	}
271
272	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
273	/* paranoid */ verify( idx <= mask );
274	/* paranoid */ verify( idx != -1ul32 );
275
276	// We need to find a spot in the ready array
277	__attribute((unused)) int len = 0;
278	__attribute((unused)) int block = 0;
279	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
280
281	disable_interrupts();
282	uint32_t off = __tls_rand();
283	enable_interrupts( __cfaabi_dbg_ctx );
284
285	uint32_t picked;
286	LOOKING: for() {
287	for(i; ring.submit_q.ready_cnt) {
288	picked = (i + off) & ready_mask;
289	uint32_t expected = -1ul32;
290	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
291	break LOOKING;
292	}
293	verify(expected != idx);
294
295	len ++;
296	}
297
298	block++;
299	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
300	__release_consumed_submission( ring );
301	unlock( ring.submit_q.lock );
302	}
303	else {
304	yield();
305	}
306	}
307
308	// update statistics
309	__STATS__( false,
310	io.submit_q.look_avg.val += len;
311	io.submit_q.look_avg.block += block;
312	io.submit_q.look_avg.cnt += 1;
313	)
314
315	return picked;
316	}
317
318	void __submit( struct io_context * ctx, uint32_t idx ) __attribute__((nonnull (1))) {
319	__io_data & ring = *ctx->thrd.ring;
320	// Get now the data we definetely need
321	volatile uint32_t * const tail = ring.submit_q.tail;
322	const uint32_t mask = *ring.submit_q.mask;
323
324	// There are 2 submission schemes, check which one we are using
325	if( ring.poller_submits ) {
326	// If the poller thread submits, then we just need to add this to the ready array
327	__submit_to_ready_array( ring, idx, mask );
328
329	post( ctx->thrd.sem );
330
331	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
332	}
333	else if( ring.eager_submits ) {
334	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
335
336	for() {
337	yield();
338
339	// If some one else collected our index, we are done
340	#warning ABA problem
341	if( ring.submit_q.ready[picked] != idx ) {
342	__STATS__( false,
343	io.submit_q.helped += 1;
344	)
345	return;
346	}
347
348	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
349	__STATS__( false,
350	io.submit_q.leader += 1;
351	)
352	break;
353	}
354
355	__STATS__( false,
356	io.submit_q.busy += 1;
357	)
358	}
359
360	// We got the lock
361	// Collect the submissions
362	unsigned to_submit = __collect_submitions( ring );
363
364	// Release the lock now so syscalls can overlap
365	unlock(ring.submit_q.lock);
366
367	// Actually submit
368	int ret = __io_uring_enter( ring, to_submit, false );
369	if( ret < 0 ) return;
370
371	// Release the consumed SQEs
372	__release_consumed_submission( ring );
373
374	// update statistics
375	__STATS__( false,
376	io.submit_q.submit_avg.rdy += to_submit;
377	io.submit_q.submit_avg.csm += ret;
378	io.submit_q.submit_avg.cnt += 1;
379	)
380	}
381	else {
382	// get mutual exclusion
383	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
384
385	/* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 0,
386	/* paranoid */ "index %u already reclaimed\n"
387	/* paranoid */ "head %u, prev %u, tail %u\n"
388	/* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
389	/* paranoid */ idx,
390	/* paranoid / ring.submit_q.head, ring.submit_q.prev_head, *tail
391	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
392	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
393	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
394	/* paranoid / ,ring.submit_q.array[ ((ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
395	/* paranoid */ );
396
397	// Append to the list of ready entries
398
399	/* paranoid */ verify( idx <= mask );
400	ring.submit_q.array[ (*tail) & mask ] = idx;
401	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
402
403	// Submit however, many entries need to be submitted
404	int ret = __io_uring_enter( ring, 1, false );
405	if( ret < 0 ) {
406	switch((int)errno) {
407	default:
408	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
409	}
410	}
411
412	// update statistics
413	__STATS__( false,
414	io.submit_q.submit_avg.csm += 1;
415	io.submit_q.submit_avg.cnt += 1;
416	)
417
418	// Release the consumed SQEs
419	__release_consumed_submission( ring );
420
421	unlock(ring.submit_q.lock);
422
423	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
424	}
425	}
426
427	static unsigned __collect_submitions( struct __io_data & ring ) {
428	/* paranoid */ verify( ring.submit_q.ready != 0p );
429	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
430
431	unsigned to_submit = 0;
432	uint32_t tail = *ring.submit_q.tail;
433	const uint32_t mask = *ring.submit_q.mask;
434
435	// Go through the list of ready submissions
436	for( i; ring.submit_q.ready_cnt ) {
437	// replace any submission with the sentinel, to consume it.
438	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
439
440	// If it was already the sentinel, then we are done
441	if( idx == -1ul32 ) continue;
442
443	// If we got a real submission, append it to the list
444	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
445	to_submit++;
446	}
447
448	// Increment the tail based on how many we are ready to submit
449	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
450
451	return to_submit;
452	}
453
454	static uint32_t __release_consumed_submission( struct __io_data & ring ) {
455	const uint32_t smask = *ring.submit_q.mask;
456
457	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
458	uint32_t chead = *ring.submit_q.head;
459	uint32_t phead = ring.submit_q.prev_head;
460	ring.submit_q.prev_head = chead;
461	unlock(ring.submit_q.release_lock);
462
463	uint32_t count = chead - phead;
464	for( i; count ) {
465	uint32_t idx = ring.submit_q.array[ (phead + i) & smask ];
466	ring.submit_q.sqes[ idx ].user_data = 0;
467	}
468	return count;
469	}
470	#endif

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

Download in other formats:

Original Format