Context Navigation

io.cfa @ 1c49dc5

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

Last change on this file since 1c49dc5 was 47746a2, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Added stats macro to reduce typing when using stats Added new macros for Complete polling and submit polling (not implemented yet)
Property mode set to `100644`
File size: 27.0 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	// Process a single completion message from the io_uring
432	// This is NOT thread-safe
433	static unsigned __collect_submitions( struct __io_data & ring );
434	static [int, bool] __drain_io( & struct __io_data ring, * sigset_t mask, int waitcnt, bool in_kernel ) {
435	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
436
437	unsigned to_submit = 0;
438	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
439	// If the poller thread also submits, then we need to aggregate the submissions which are ready
440	to_submit = __collect_submitions( ring );
441	}
442
443	const uint32_t smask = *ring.submit_q.mask;
444	uint32_t shead = *ring.submit_q.head;
445	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
446	if( ret < 0 ) {
447	switch((int)errno) {
448	case EAGAIN:
449	case EINTR:
450	return [0, true];
451	default:
452	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
453	}
454	}
455
456	// Release the consumed SQEs
457	for( i; ret ) {
458	uint32_t idx = ring.submit_q.array[ (i + shead) & smask ];
459	ring.submit_q.sqes[ idx ].user_data = 0;
460	}
461
462	// update statistics
463	__STATS__( true,
464	if( to_submit > 0 ) {
465	io.submit_q.submit_avg.rdy += to_submit;
466	io.submit_q.submit_avg.csm += ret;
467	io.submit_q.submit_avg.cnt += 1;
468	}
469	)
470
471	// Drain the queue
472	unsigned head = *ring.completion_q.head;
473	unsigned tail = *ring.completion_q.tail;
474	const uint32_t mask = *ring.completion_q.mask;
475
476	// Memory barrier
477	__atomic_thread_fence( __ATOMIC_SEQ_CST );
478
479	// Nothing was new return 0
480	if (head == tail) {
481	return [0, to_submit > 0];
482	}
483
484	uint32_t count = tail - head;
485	for(i; count) {
486	unsigned idx = (head + i) & mask;
487	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
488
489	/* paranoid */ verify(&cqe);
490
491	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
492	__cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
493
494	data->result = cqe.res;
495	if(!in_kernel) { unpark( data->thrd __cfaabi_dbg_ctx2 ); }
496	else { __unpark( &ring.poller.slow.id, data->thrd __cfaabi_dbg_ctx2 ); }
497	}
498
499	// Allow new submissions to happen
500	// V(ring.submit, count);
501
502	// Mark to the kernel that the cqe has been seen
503	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
504	__atomic_thread_fence( __ATOMIC_SEQ_CST );
505	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
506
507	return [count, count > 0 \|\| to_submit > 0];
508	}
509
510	static void * __io_poller_slow( void * arg ) {
511	#if !defined( __CFA_NO_STATISTICS__ )
512	__stats_t local_stats;
513	__init_stats( &local_stats );
514	kernelTLS.this_stats = &local_stats;
515	#endif
516
517	cluster * cltr = (cluster *)arg;
518	struct __io_data & ring = *cltr->io;
519
520	ring.poller.slow.id.id = doregister( &ring.poller.slow.id );
521
522	sigset_t mask;
523	sigfillset(&mask);
524	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
525	abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
526	}
527
528	sigdelset( &mask, SIGUSR1 );
529
530	verify( (ring.submit_q.head) == (ring.submit_q.tail) );
531	verify( (ring.completion_q.head) == (ring.completion_q.tail) );
532
533	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p ready\n", &ring);
534
535	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
536	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
537
538	__atomic_store_n( &ring.poller.slow.blocked, true, __ATOMIC_SEQ_CST );
539
540	// In the user-thread approach drain and if anything was drained,
541	// batton pass to the user-thread
542	int count;
543	bool again;
544	[count, again] = __drain_io( ring, &mask, 1, true );
545
546	__atomic_store_n( &ring.poller.slow.blocked, false, __ATOMIC_SEQ_CST );
547
548	// Update statistics
549	__STATS__( true,
550	io.complete_q.completed_avg.val += count;
551	io.complete_q.completed_avg.slow_cnt += 1;
552	)
553
554	if(again) {
555	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to fast poller\n", &ring);
556	__unpark( &ring.poller.slow.id, &ring.poller.fast.thrd __cfaabi_dbg_ctx2 );
557	wait( ring.poller.sem );
558	}
559	}
560	}
561	else {
562	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
563	//In the naive approach, just poll the io completion queue directly
564	int count;
565	bool again;
566	[count, again] = __drain_io( ring, &mask, 1, true );
567
568	// Update statistics
569	__STATS__( true,
570	io.complete_q.completed_avg.val += count;
571	io.complete_q.completed_avg.slow_cnt += 1;
572	)
573	}
574	}
575
576	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p stopping\n", &ring);
577
578	unregister( &ring.poller.slow.id );
579
580	#if !defined(__CFA_NO_STATISTICS__)
581	__tally_stats(cltr->stats, &local_stats);
582	#endif
583
584	return 0p;
585	}
586
587	void main( __io_poller_fast & this ) {
588	verify( this.ring->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD );
589
590	// Start parked
591	park( __cfaabi_dbg_ctx );
592
593	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p ready\n", &this.ring);
594
595	int reset = 0;
596
597	// Then loop until we need to start
598	while(!__atomic_load_n(&this.ring->done, __ATOMIC_SEQ_CST)) {
599
600	// Drain the io
601	int count;
602	bool again;
603	disable_interrupts();
604	[count, again] = __drain_io( *this.ring, 0p, 0, false );
605
606	if(!again) reset++;
607
608	// Update statistics
609	__STATS__( true,
610	io.complete_q.completed_avg.val += count;
611	io.complete_q.completed_avg.fast_cnt += 1;
612	)
613	enable_interrupts( __cfaabi_dbg_ctx );
614
615	// If we got something, just yield and check again
616	if(reset < 5) {
617	yield();
618	}
619	// We didn't get anything baton pass to the slow poller
620	else {
621	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to slow poller\n", &this.ring);
622	reset = 0;
623
624	// wake up the slow poller
625	post( this.ring->poller.sem );
626
627	// park this thread
628	park( __cfaabi_dbg_ctx );
629	}
630	}
631
632	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
633	}
634
635	static inline void __wake_poller( struct __io_data & ring ) __attribute__((artificial));
636	static inline void __wake_poller( struct __io_data & ring ) {
637	if(!__atomic_load_n( &ring.poller.slow.blocked, __ATOMIC_SEQ_CST)) return;
638
639	sigval val = { 1 };
640	pthread_sigqueue( ring.poller.slow.kthrd, SIGUSR1, val );
641	}
642
643	//=============================================================================================
644	// I/O Submissions
645	//=============================================================================================
646
647	// Submition steps :
648	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
649	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
650	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
651	// need to write an allocator that allows allocating concurrently.
652	//
653	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
654	//
655	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
656	// needs to arrive to two concensus at the same time:
657	// A - The order in which entries are listed in the array: no two threads must pick the
658	// same index for their entries
659	// B - When can the tail be update for the kernel. EVERY entries in the array between
660	// head and tail must be fully filled and shouldn't ever be touched again.
661	//
662
663	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
664	/* paranoid */ verify( data != 0 );
665
666	// Prepare the data we need
667	__attribute((unused)) int len = 0;
668	__attribute((unused)) int block = 0;
669	uint32_t cnt = *ring.submit_q.num;
670	uint32_t mask = *ring.submit_q.mask;
671
672	disable_interrupts();
673	uint32_t off = __tls_rand();
674	enable_interrupts( __cfaabi_dbg_ctx );
675
676	// Loop around looking for an available spot
677	for() {
678	// Look through the list starting at some offset
679	for(i; cnt) {
680	uint64_t expected = 0;
681	uint32_t idx = (i + off) & mask;
682	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
683	volatile uint64_t * udata = &sqe->user_data;
684
685	if( *udata == expected &&
686	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
687	{
688	// update statistics
689	__STATS__( false,
690	io.submit_q.alloc_avg.val += len;
691	io.submit_q.alloc_avg.block += block;
692	io.submit_q.alloc_avg.cnt += 1;
693	)
694
695
696	// Success return the data
697	return [sqe, idx];
698	}
699	verify(expected != data);
700
701	len ++;
702	}
703
704	block++;
705	yield();
706	}
707	}
708
709	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
710	/* paranoid */ verify( idx <= mask );
711	/* paranoid */ verify( idx != -1ul32 );
712
713	// We need to find a spot in the ready array
714	__attribute((unused)) int len = 0;
715	__attribute((unused)) int block = 0;
716	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
717
718	disable_interrupts();
719	uint32_t off = __tls_rand();
720	enable_interrupts( __cfaabi_dbg_ctx );
721
722	uint32_t picked;
723	LOOKING: for() {
724	for(i; ring.submit_q.ready_cnt) {
725	picked = (i + off) & ready_mask;
726	uint32_t expected = -1ul32;
727	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
728	break LOOKING;
729	}
730	verify(expected != idx);
731
732	len ++;
733	}
734
735	block++;
736	yield();
737	}
738
739	// update statistics
740	__STATS__( false,
741	io.submit_q.look_avg.val += len;
742	io.submit_q.look_avg.block += block;
743	io.submit_q.look_avg.cnt += 1;
744	)
745
746	return picked;
747	}
748
749	void __submit( struct __io_data & ring, uint32_t idx ) {
750	// Get now the data we definetely need
751	uint32_t * const tail = ring.submit_q.tail;
752	const uint32_t mask = *ring.submit_q.mask;
753
754	// There are 2 submission schemes, check which one we are using
755	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
756	// If the poller thread submits, then we just need to add this to the ready array
757	__submit_to_ready_array( ring, idx, mask );
758
759	__wake_poller( ring );
760
761	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
762	}
763	else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) {
764	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
765
766	for() {
767	yield();
768
769	// If some one else collected our index, we are done
770	#warning ABA problem
771	if( ring.submit_q.ready[picked] != idx ) {
772	__STATS__( false,
773	io.submit_q.helped += 1;
774	)
775	return;
776	}
777
778	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
779	__STATS__( false,
780	io.submit_q.leader += 1;
781	)
782	break;
783	}
784
785	__STATS__( false,
786	io.submit_q.busy += 1;
787	)
788	}
789
790	// We got the lock
791	unsigned to_submit = __collect_submitions( ring );
792	const uint32_t smask = *ring.submit_q.mask;
793	uint32_t shead = *ring.submit_q.head;
794	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8);
795	if( ret < 0 ) {
796	switch((int)errno) {
797	case EAGAIN:
798	case EINTR:
799	unlock(ring.submit_q.lock);
800	return;
801	default:
802	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
803	}
804	}
805
806	/* paranoid */ verify( ret > 0 );
807
808	// Release the consumed SQEs
809	for( i; ret ) {
810	uint32_t idx = ring.submit_q.array[ (i + shead) & smask ];
811	ring.submit_q.sqes[ idx ].user_data = 0;
812	}
813
814	// update statistics
815	__STATS__( true,
816	io.submit_q.submit_avg.rdy += to_submit;
817	io.submit_q.submit_avg.csm += ret;
818	io.submit_q.submit_avg.cnt += 1;
819	)
820
821	unlock(ring.submit_q.lock);
822	}
823	else {
824	// get mutual exclusion
825	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
826
827	// Append to the list of ready entries
828
829	/* paranoid */ verify( idx <= mask );
830
831	ring.submit_q.array[ (*tail) & mask ] = idx & mask;
832	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
833
834	// Submit however, many entries need to be submitted
835	int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
836	if( ret < 0 ) {
837	switch((int)errno) {
838	default:
839	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
840	}
841	}
842
843	// update statistics
844	__STATS__( false,
845	io.submit_q.submit_avg.csm += 1;
846	io.submit_q.submit_avg.cnt += 1;
847	)
848
849	ring.submit_q.sqes[ idx & mask ].user_data = 0;
850
851	unlock(ring.submit_q.lock);
852
853	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
854	}
855	}
856
857	static unsigned __collect_submitions( struct __io_data & ring ) {
858	/* paranoid */ verify( ring.submit_q.ready != 0p );
859	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
860
861	unsigned to_submit = 0;
862	uint32_t tail = *ring.submit_q.tail;
863	const uint32_t mask = *ring.submit_q.mask;
864
865	// Go through the list of ready submissions
866	for( i; ring.submit_q.ready_cnt ) {
867	// replace any submission with the sentinel, to consume it.
868	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
869
870	// If it was already the sentinel, then we are done
871	if( idx == -1ul32 ) continue;
872
873	// If we got a real submission, append it to the list
874	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
875	to_submit++;
876	}
877
878	// Increment the tail based on how many we are ready to submit
879	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
880
881	return to_submit;
882	}
883	#endif

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

Download in other formats:

Original Format