Context Navigation

io.cfa @ 1d5e4711

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

Last change on this file since 1d5e4711 was 1d5e4711, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Optim: No longer call io_uring_enter to poll for events if

We don't want to block
There is nothing to submit

This is only valid for "interrupt driven I/O", polling driven I/O not yet supported.

Property mode set to 100644

File size: 27.1 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 __CFA_DEBUG_PRINT_IO__
17	// #define __CFA_DEBUG_PRINT_IO_CORE__
18
19	#include "kernel.hfa"
20	#include "bitmanip.hfa"
21
22	#if !defined(HAVE_LINUX_IO_URING_H)
23	void __kernel_io_startup( cluster &, unsigned, bool ) {
24	// Nothing to do without io_uring
25	}
26
27	void __kernel_io_finish_start( cluster & ) {
28	// Nothing to do without io_uring
29	}
30
31	void __kernel_io_prepare_stop( cluster & ) {
32	// Nothing to do without io_uring
33	}
34
35	void __kernel_io_shutdown( cluster &, bool ) {
36	// Nothing to do without io_uring
37	}
38
39	#else
40	#define _GNU_SOURCE /* See feature_test_macros(7) */
41	#include <errno.h>
42	#include <stdint.h>
43	#include <string.h>
44	#include <unistd.h>
45	#include <sys/mman.h>
46
47	extern "C" {
48	#include <sys/syscall.h>
49
50	#include <linux/io_uring.h>
51	}
52
53	#include "bits/signal.hfa"
54	#include "kernel_private.hfa"
55	#include "thread.hfa"
56
57	uint32_t entries_per_cluster() {
58	return 256;
59	}
60
61	static void * __io_poller_slow( void * arg );
62
63	// Weirdly, some systems that do support io_uring don't actually define these
64	#ifdef __alpha__
65	/*
66	* alpha is the only exception, all other architectures
67	* have common numbers for new system calls.
68	*/
69	#ifndef __NR_io_uring_setup
70	#define __NR_io_uring_setup 535
71	#endif
72	#ifndef __NR_io_uring_enter
73	#define __NR_io_uring_enter 536
74	#endif
75	#ifndef __NR_io_uring_register
76	#define __NR_io_uring_register 537
77	#endif
78	#else /* !__alpha__ */
79	#ifndef __NR_io_uring_setup
80	#define __NR_io_uring_setup 425
81	#endif
82	#ifndef __NR_io_uring_enter
83	#define __NR_io_uring_enter 426
84	#endif
85	#ifndef __NR_io_uring_register
86	#define __NR_io_uring_register 427
87	#endif
88	#endif
89
90	// Fast poller user-thread
91	// Not using the "thread" keyword because we want to control
92	// more carefully when to start/stop it
93	struct __io_poller_fast {
94	struct __io_data * ring;
95	$thread thrd;
96	};
97
98	void ?{}( __io_poller_fast & this, struct cluster & cltr ) {
99	this.ring = cltr.io;
100	(this.thrd){ "Fast I/O Poller", cltr };
101	}
102	void ^?{}( __io_poller_fast & mutex this );
103	void main( __io_poller_fast & this );
104	static inline $thread * get_thread( __io_poller_fast & this ) { return &this.thrd; }
105	void ^?{}( __io_poller_fast & mutex this ) {}
106
107	struct __submition_data {
108	// Head and tail of the ring (associated with array)
109	volatile uint32_t * head;
110	volatile uint32_t * tail;
111
112	// The actual kernel ring which uses head/tail
113	// indexes into the sqes arrays
114	uint32_t * array;
115
116	// number of entries and mask to go with it
117	const uint32_t * num;
118	const uint32_t * mask;
119
120	// Submission flags (Not sure what for)
121	uint32_t * flags;
122
123	// number of sqes not submitted (whatever that means)
124	uint32_t * dropped;
125
126	// Like head/tail but not seen by the kernel
127	volatile uint32_t * ready;
128	uint32_t ready_cnt;
129
130	__spinlock_t lock;
131
132	// A buffer of sqes (not the actual ring)
133	struct io_uring_sqe * sqes;
134
135	// The location and size of the mmaped area
136	void * ring_ptr;
137	size_t ring_sz;
138	};
139
140	struct __completion_data {
141	// Head and tail of the ring
142	volatile uint32_t * head;
143	volatile uint32_t * tail;
144
145	// number of entries and mask to go with it
146	const uint32_t * mask;
147	const uint32_t * num;
148
149	// number of cqes not submitted (whatever that means)
150	uint32_t * overflow;
151
152	// the kernel ring
153	struct io_uring_cqe * cqes;
154
155	// The location and size of the mmaped area
156	void * ring_ptr;
157	size_t ring_sz;
158	};
159
160	struct __io_data {
161	struct __submition_data submit_q;
162	struct __completion_data completion_q;
163	uint32_t ring_flags;
164	int cltr_flags;
165	int fd;
166	semaphore submit;
167	volatile bool done;
168	struct {
169	struct {
170	__processor_id_t id;
171	void * stack;
172	pthread_t kthrd;
173	volatile bool blocked;
174	} slow;
175	__io_poller_fast fast;
176	__bin_sem_t sem;
177	} poller;
178	};
179
180	//=============================================================================================
181	// I/O Startup / Shutdown logic
182	//=============================================================================================
183	void __kernel_io_startup( cluster & this, unsigned io_flags, bool main_cluster ) {
184	if( (io_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS) && (io_flags & CFA_CLUSTER_IO_EAGER_SUBMITS) ) {
185	abort("CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS and CFA_CLUSTER_IO_EAGER_SUBMITS cannot be mixed\n");
186	}
187
188	this.io = malloc();
189
190	// Step 1 : call to setup
191	struct io_uring_params params;
192	memset(&params, 0, sizeof(params));
193	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_SUBMITS ) params.flags \|= IORING_SETUP_SQPOLL;
194	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_COMPLETES ) params.flags \|= IORING_SETUP_IOPOLL;
195
196	uint32_t nentries = entries_per_cluster();
197
198	int fd = syscall(__NR_io_uring_setup, nentries, &params );
199	if(fd < 0) {
200	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
201	}
202
203	// Step 2 : mmap result
204	memset( this.io, 0, sizeof(struct __io_data) );
205	struct __submition_data & sq = this.io->submit_q;
206	struct __completion_data & cq = this.io->completion_q;
207
208	// calculate the right ring size
209	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
210	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
211
212	// Requires features
213	#if defined(IORING_FEAT_SINGLE_MMAP)
214	// adjust the size according to the parameters
215	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
216	cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz);
217	}
218	#endif
219
220	// mmap the Submit Queue into existence
221	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
222	if (sq.ring_ptr == (void*)MAP_FAILED) {
223	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
224	}
225
226	// Requires features
227	#if defined(IORING_FEAT_SINGLE_MMAP)
228	// mmap the Completion Queue into existence (may or may not be needed)
229	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
230	cq->ring_ptr = sq->ring_ptr;
231	}
232	else
233	#endif
234	{
235	// We need multiple call to MMAP
236	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
237	if (cq.ring_ptr == (void*)MAP_FAILED) {
238	munmap(sq.ring_ptr, sq.ring_sz);
239	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
240	}
241	}
242
243	// mmap the submit queue entries
244	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
245	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
246	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
247	munmap(sq.ring_ptr, sq.ring_sz);
248	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
249	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
250	}
251
252	// Get the pointers from the kernel to fill the structure
253	// submit queue
254	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
255	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
256	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
257	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
258	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
259	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
260	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
261
262	{
263	const uint32_t num = *sq.num;
264	for( i; num ) {
265	sq.sqes[i].user_data = 0ul64;
266	}
267	}
268
269	(sq.lock){};
270
271	if( io_flags & ( CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS \| CFA_CLUSTER_IO_EAGER_SUBMITS ) ) {
272	/* paranoid */ verify( is_pow2( io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET ) \|\| ((io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET) < 8) );
273	sq.ready_cnt = max(io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET, 8);
274	sq.ready = alloc_align( 64, sq.ready_cnt );
275	for(i; sq.ready_cnt) {
276	sq.ready[i] = -1ul32;
277	}
278	}
279	else {
280	sq.ready_cnt = 0;
281	sq.ready = 0p;
282	}
283
284	// completion queue
285	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
286	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
287	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
288	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
289	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
290	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
291
292	// some paranoid checks
293	/* paranoid / verifyf( (cq.mask) == ((cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (cq.num) - 1ul32, cq.num, cq.mask );
294	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
295	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
296	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
297
298	/* paranoid / verifyf( (sq.mask) == ((sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (sq.num) - 1ul32, sq.num, sq.mask );
299	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
300	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
301	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
302
303	// Update the global ring info
304	this.io->ring_flags = params.flags;
305	this.io->cltr_flags = io_flags;
306	this.io->fd = fd;
307	this.io->done = false;
308	(this.io->submit){ min(sq.num, cq.num) };
309
310	if(!main_cluster) {
311	__kernel_io_finish_start( this );
312	}
313	}
314
315	void __kernel_io_finish_start( cluster & this ) {
316	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
317	__cfadbg_print_safe(io_core, "Kernel I/O : Creating fast poller for cluter %p\n", &this);
318	(this.io->poller.fast){ this };
319	__thrd_start( this.io->poller.fast, main );
320	}
321
322	// Create the poller thread
323	__cfadbg_print_safe(io_core, "Kernel I/O : Creating slow poller for cluter %p\n", &this);
324	this.io->poller.slow.blocked = false;
325	this.io->poller.slow.stack = __create_pthread( &this.io->poller.slow.kthrd, __io_poller_slow, &this );
326	}
327
328	void __kernel_io_prepare_stop( cluster & this ) {
329	__cfadbg_print_safe(io_core, "Kernel I/O : Stopping pollers for cluster\n", &this);
330	// Notify the poller thread of the shutdown
331	__atomic_store_n(&this.io->done, true, __ATOMIC_SEQ_CST);
332
333	// Stop the IO Poller
334	sigval val = { 1 };
335	pthread_sigqueue( this.io->poller.slow.kthrd, SIGUSR1, val );
336	post( this.io->poller.sem );
337
338	// Wait for the poller thread to finish
339	pthread_join( this.io->poller.slow.kthrd, 0p );
340	free( this.io->poller.slow.stack );
341
342	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller stopped for cluster\n", &this);
343
344	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
345	with( this.io->poller.fast ) {
346	/* paranoid */ verify( this.nprocessors == 0 \|\| &this == mainCluster );
347	/* paranoid */ verify( !ready_mutate_islocked() );
348
349	// We need to adjust the clean-up based on where the thread is
350	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
351
352	ready_schedule_lock( (struct __processor_id_t *)active_processor() );
353
354	// This is the tricky case
355	// The thread was preempted and now it is on the ready queue
356	// The thread should be the last on the list
357	/* paranoid */ verify( thrd.link.next != 0p );
358
359	// Remove the thread from the ready queue of this cluster
360	__attribute__((unused)) bool removed = remove_head( &this, &thrd );
361	/* paranoid */ verify( removed );
362	thrd.link.next = 0p;
363	thrd.link.prev = 0p;
364	__cfaabi_dbg_debug_do( thrd.unpark_stale = true );
365
366	// Fixup the thread state
367	thrd.state = Blocked;
368	thrd.ticket = 0;
369	thrd.preempted = __NO_PREEMPTION;
370
371	ready_schedule_unlock( (struct __processor_id_t *)active_processor() );
372
373	// Pretend like the thread was blocked all along
374	}
375	// !!! This is not an else if !!!
376	if( thrd.state == Blocked ) {
377
378	// This is the "easy case"
379	// The thread is parked and can easily be moved to active cluster
380	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
381	thrd.curr_cluster = active_cluster();
382
383	// unpark the fast io_poller
384	unpark( &thrd __cfaabi_dbg_ctx2 );
385	}
386	else {
387
388	// The thread is in a weird state
389	// I don't know what to do here
390	abort("Fast poller thread is in unexpected state, cannot clean-up correctly\n");
391	}
392
393	}
394
395	^(this.io->poller.fast){};
396
397	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller stopped for cluster\n", &this);
398	}
399	}
400
401	void __kernel_io_shutdown( cluster & this, bool main_cluster ) {
402	if(!main_cluster) {
403	__kernel_io_prepare_stop( this );
404	}
405
406	// Shutdown the io rings
407	struct __submition_data & sq = this.io->submit_q;
408	struct __completion_data & cq = this.io->completion_q;
409
410	// unmap the submit queue entries
411	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
412
413	// unmap the Submit Queue ring
414	munmap(sq.ring_ptr, sq.ring_sz);
415
416	// unmap the Completion Queue ring, if it is different
417	if (cq.ring_ptr != sq.ring_ptr) {
418	munmap(cq.ring_ptr, cq.ring_sz);
419	}
420
421	// close the file descriptor
422	close(this.io->fd);
423
424	free( this.io->submit_q.ready ); // Maybe null, doesn't matter
425	free( this.io );
426	}
427
428	//=============================================================================================
429	// I/O Polling
430	//=============================================================================================
431	static unsigned __collect_submitions( struct __io_data & ring );
432
433	// Process a single completion message from the io_uring
434	// This is NOT thread-safe
435	static [int, bool] __drain_io( & struct __io_data ring, * sigset_t mask, int waitcnt, bool in_kernel ) {
436	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
437	const uint32_t smask = *ring.submit_q.mask;
438
439	unsigned to_submit = 0;
440	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
441	// If the poller thread also submits, then we need to aggregate the submissions which are ready
442	to_submit = __collect_submitions( ring );
443	}
444
445	if (to_submit > 0 \|\| waitcnt > 0) {
446	uint32_t shead = *ring.submit_q.head;
447	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
448	if( ret < 0 ) {
449	switch((int)errno) {
450	case EAGAIN:
451	case EINTR:
452	return [0, true];
453	default:
454	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
455	}
456	}
457
458	// Release the consumed SQEs
459	for( i; ret ) {
460	uint32_t idx = ring.submit_q.array[ (i + shead) & smask ];
461	ring.submit_q.sqes[ idx ].user_data = 0;
462	}
463
464	// update statistics
465	__STATS__( true,
466	if( to_submit > 0 ) {
467	io.submit_q.submit_avg.rdy += to_submit;
468	io.submit_q.submit_avg.csm += ret;
469	io.submit_q.submit_avg.cnt += 1;
470	}
471	)
472	}
473
474	// Memory barrier
475	__atomic_thread_fence( __ATOMIC_SEQ_CST );
476
477	// Drain the queue
478	unsigned head = *ring.completion_q.head;
479	unsigned tail = *ring.completion_q.tail;
480	const uint32_t mask = *ring.completion_q.mask;
481
482	// Nothing was new return 0
483	if (head == tail) {
484	return [0, to_submit > 0];
485	}
486
487	uint32_t count = tail - head;
488	/* paranoid */ verify( count != 0 );
489	for(i; count) {
490	unsigned idx = (head + i) & mask;
491	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
492
493	/* paranoid */ verify(&cqe);
494
495	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
496	__cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
497
498	data->result = cqe.res;
499	if(!in_kernel) { unpark( data->thrd __cfaabi_dbg_ctx2 ); }
500	else { __unpark( &ring.poller.slow.id, data->thrd __cfaabi_dbg_ctx2 ); }
501	}
502
503	// Allow new submissions to happen
504	// V(ring.submit, count);
505
506	// Mark to the kernel that the cqe has been seen
507	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
508	__atomic_thread_fence( __ATOMIC_SEQ_CST );
509	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
510
511	return [count, count > 0 \|\| to_submit > 0];
512	}
513
514	static void * __io_poller_slow( void * arg ) {
515	#if !defined( __CFA_NO_STATISTICS__ )
516	__stats_t local_stats;
517	__init_stats( &local_stats );
518	kernelTLS.this_stats = &local_stats;
519	#endif
520
521	cluster * cltr = (cluster *)arg;
522	struct __io_data & ring = *cltr->io;
523
524	ring.poller.slow.id.id = doregister( &ring.poller.slow.id );
525
526	sigset_t mask;
527	sigfillset(&mask);
528	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
529	abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
530	}
531
532	sigdelset( &mask, SIGUSR1 );
533
534	verify( (ring.submit_q.head) == (ring.submit_q.tail) );
535	verify( (ring.completion_q.head) == (ring.completion_q.tail) );
536
537	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p ready\n", &ring);
538
539	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
540	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
541
542	__atomic_store_n( &ring.poller.slow.blocked, true, __ATOMIC_SEQ_CST );
543
544	// In the user-thread approach drain and if anything was drained,
545	// batton pass to the user-thread
546	int count;
547	bool again;
548	[count, again] = __drain_io( ring, &mask, 1, true );
549
550	__atomic_store_n( &ring.poller.slow.blocked, false, __ATOMIC_SEQ_CST );
551
552	// Update statistics
553	__STATS__( true,
554	io.complete_q.completed_avg.val += count;
555	io.complete_q.completed_avg.slow_cnt += 1;
556	)
557
558	if(again) {
559	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to fast poller\n", &ring);
560	__unpark( &ring.poller.slow.id, &ring.poller.fast.thrd __cfaabi_dbg_ctx2 );
561	wait( ring.poller.sem );
562	}
563	}
564	}
565	else {
566	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
567	//In the naive approach, just poll the io completion queue directly
568	int count;
569	bool again;
570	[count, again] = __drain_io( ring, &mask, 1, true );
571
572	// Update statistics
573	__STATS__( true,
574	io.complete_q.completed_avg.val += count;
575	io.complete_q.completed_avg.slow_cnt += 1;
576	)
577	}
578	}
579
580	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p stopping\n", &ring);
581
582	unregister( &ring.poller.slow.id );
583
584	#if !defined(__CFA_NO_STATISTICS__)
585	__tally_stats(cltr->stats, &local_stats);
586	#endif
587
588	return 0p;
589	}
590
591	void main( __io_poller_fast & this ) {
592	verify( this.ring->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD );
593
594	// Start parked
595	park( __cfaabi_dbg_ctx );
596
597	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p ready\n", &this.ring);
598
599	int reset = 0;
600
601	// Then loop until we need to start
602	while(!__atomic_load_n(&this.ring->done, __ATOMIC_SEQ_CST)) {
603
604	// Drain the io
605	int count;
606	bool again;
607	disable_interrupts();
608	[count, again] = __drain_io( *this.ring, 0p, 0, false );
609
610	if(!again) reset++;
611
612	// Update statistics
613	__STATS__( true,
614	io.complete_q.completed_avg.val += count;
615	io.complete_q.completed_avg.fast_cnt += 1;
616	)
617	enable_interrupts( __cfaabi_dbg_ctx );
618
619	// If we got something, just yield and check again
620	if(reset < 5) {
621	yield();
622	}
623	// We didn't get anything baton pass to the slow poller
624	else {
625	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to slow poller\n", &this.ring);
626	reset = 0;
627
628	// wake up the slow poller
629	post( this.ring->poller.sem );
630
631	// park this thread
632	park( __cfaabi_dbg_ctx );
633	}
634	}
635
636	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
637	}
638
639	static inline void __wake_poller( struct __io_data & ring ) __attribute__((artificial));
640	static inline void __wake_poller( struct __io_data & ring ) {
641	if(!__atomic_load_n( &ring.poller.slow.blocked, __ATOMIC_SEQ_CST)) return;
642
643	sigval val = { 1 };
644	pthread_sigqueue( ring.poller.slow.kthrd, SIGUSR1, val );
645	}
646
647	//=============================================================================================
648	// I/O Submissions
649	//=============================================================================================
650
651	// Submition steps :
652	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
653	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
654	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
655	// need to write an allocator that allows allocating concurrently.
656	//
657	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
658	//
659	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
660	// needs to arrive to two concensus at the same time:
661	// A - The order in which entries are listed in the array: no two threads must pick the
662	// same index for their entries
663	// B - When can the tail be update for the kernel. EVERY entries in the array between
664	// head and tail must be fully filled and shouldn't ever be touched again.
665	//
666
667	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
668	/* paranoid */ verify( data != 0 );
669
670	// Prepare the data we need
671	__attribute((unused)) int len = 0;
672	__attribute((unused)) int block = 0;
673	uint32_t cnt = *ring.submit_q.num;
674	uint32_t mask = *ring.submit_q.mask;
675
676	disable_interrupts();
677	uint32_t off = __tls_rand();
678	enable_interrupts( __cfaabi_dbg_ctx );
679
680	// Loop around looking for an available spot
681	for() {
682	// Look through the list starting at some offset
683	for(i; cnt) {
684	uint64_t expected = 0;
685	uint32_t idx = (i + off) & mask;
686	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
687	volatile uint64_t * udata = &sqe->user_data;
688
689	if( *udata == expected &&
690	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
691	{
692	// update statistics
693	__STATS__( false,
694	io.submit_q.alloc_avg.val += len;
695	io.submit_q.alloc_avg.block += block;
696	io.submit_q.alloc_avg.cnt += 1;
697	)
698
699
700	// Success return the data
701	return [sqe, idx];
702	}
703	verify(expected != data);
704
705	len ++;
706	}
707
708	block++;
709	yield();
710	}
711	}
712
713	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
714	/* paranoid */ verify( idx <= mask );
715	/* paranoid */ verify( idx != -1ul32 );
716
717	// We need to find a spot in the ready array
718	__attribute((unused)) int len = 0;
719	__attribute((unused)) int block = 0;
720	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
721
722	disable_interrupts();
723	uint32_t off = __tls_rand();
724	enable_interrupts( __cfaabi_dbg_ctx );
725
726	uint32_t picked;
727	LOOKING: for() {
728	for(i; ring.submit_q.ready_cnt) {
729	picked = (i + off) & ready_mask;
730	uint32_t expected = -1ul32;
731	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
732	break LOOKING;
733	}
734	verify(expected != idx);
735
736	len ++;
737	}
738
739	block++;
740	yield();
741	}
742
743	// update statistics
744	__STATS__( false,
745	io.submit_q.look_avg.val += len;
746	io.submit_q.look_avg.block += block;
747	io.submit_q.look_avg.cnt += 1;
748	)
749
750	return picked;
751	}
752
753	void __submit( struct __io_data & ring, uint32_t idx ) {
754	// Get now the data we definetely need
755	uint32_t * const tail = ring.submit_q.tail;
756	const uint32_t mask = *ring.submit_q.mask;
757
758	// There are 2 submission schemes, check which one we are using
759	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
760	// If the poller thread submits, then we just need to add this to the ready array
761	__submit_to_ready_array( ring, idx, mask );
762
763	__wake_poller( ring );
764
765	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
766	}
767	else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) {
768	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
769
770	for() {
771	yield();
772
773	// If some one else collected our index, we are done
774	#warning ABA problem
775	if( ring.submit_q.ready[picked] != idx ) {
776	__STATS__( false,
777	io.submit_q.helped += 1;
778	)
779	return;
780	}
781
782	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
783	__STATS__( false,
784	io.submit_q.leader += 1;
785	)
786	break;
787	}
788
789	__STATS__( false,
790	io.submit_q.busy += 1;
791	)
792	}
793
794	// We got the lock
795	unsigned to_submit = __collect_submitions( ring );
796	uint32_t shead = *ring.submit_q.head;
797	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8);
798	if( ret < 0 ) {
799	switch((int)errno) {
800	case EAGAIN:
801	case EINTR:
802	unlock(ring.submit_q.lock);
803	return;
804	default:
805	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
806	}
807	}
808
809	/* paranoid */ verify( ret > 0 );
810
811	// Release the consumed SQEs
812	const uint32_t smask = *ring.submit_q.mask;
813	for( i; ret ) {
814	uint32_t idx = ring.submit_q.array[ (i + shead) & smask ];
815	ring.submit_q.sqes[ idx ].user_data = 0;
816	}
817
818	// update statistics
819	__STATS__( true,
820	io.submit_q.submit_avg.rdy += to_submit;
821	io.submit_q.submit_avg.csm += ret;
822	io.submit_q.submit_avg.cnt += 1;
823	)
824
825	unlock(ring.submit_q.lock);
826	}
827	else {
828	// get mutual exclusion
829	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
830
831	// Append to the list of ready entries
832
833	/* paranoid */ verify( idx <= mask );
834
835	ring.submit_q.array[ (*tail) & mask ] = idx & mask;
836	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
837
838	// Submit however, many entries need to be submitted
839	int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
840	if( ret < 0 ) {
841	switch((int)errno) {
842	default:
843	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
844	}
845	}
846
847	// update statistics
848	__STATS__( false,
849	io.submit_q.submit_avg.csm += 1;
850	io.submit_q.submit_avg.cnt += 1;
851	)
852
853	ring.submit_q.sqes[ idx & mask ].user_data = 0;
854
855	unlock(ring.submit_q.lock);
856
857	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
858	}
859	}
860
861	static unsigned __collect_submitions( struct __io_data & ring ) {
862	/* paranoid */ verify( ring.submit_q.ready != 0p );
863	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
864
865	unsigned to_submit = 0;
866	uint32_t tail = *ring.submit_q.tail;
867	const uint32_t mask = *ring.submit_q.mask;
868
869	// Go through the list of ready submissions
870	for( i; ring.submit_q.ready_cnt ) {
871	// replace any submission with the sentinel, to consume it.
872	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
873
874	// If it was already the sentinel, then we are done
875	if( idx == -1ul32 ) continue;
876
877	// If we got a real submission, append it to the list
878	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
879	to_submit++;
880	}
881
882	// Increment the tail based on how many we are ready to submit
883	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
884
885	return to_submit;
886	}
887	#endif

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

Download in other formats:

Original Format