Context Navigation

io.cfa @ 34b61882

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

Last change on this file since 34b61882 was 34b61882, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Submission release is now based on previous seen head rather than the return value of io_uring_enter This will make it easier to support SQ POLLING It is not currently thread safe. It is unclear if it can be made thread safe
Property mode set to `100644`
File size: 27.6 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 __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	volatile uint32_t prev_head;
112
113	// The actual kernel ring which uses head/tail
114	// indexes into the sqes arrays
115	uint32_t * array;
116
117	// number of entries and mask to go with it
118	const uint32_t * num;
119	const uint32_t * mask;
120
121	// Submission flags (Not sure what for)
122	uint32_t * flags;
123
124	// number of sqes not submitted (whatever that means)
125	uint32_t * dropped;
126
127	// Like head/tail but not seen by the kernel
128	volatile uint32_t * ready;
129	uint32_t ready_cnt;
130
131	__spinlock_t lock;
132
133	// A buffer of sqes (not the actual ring)
134	struct io_uring_sqe * sqes;
135
136	// The location and size of the mmaped area
137	void * ring_ptr;
138	size_t ring_sz;
139	};
140
141	struct __completion_data {
142	// Head and tail of the ring
143	volatile uint32_t * head;
144	volatile uint32_t * tail;
145
146	// number of entries and mask to go with it
147	const uint32_t * mask;
148	const uint32_t * num;
149
150	// number of cqes not submitted (whatever that means)
151	uint32_t * overflow;
152
153	// the kernel ring
154	struct io_uring_cqe * cqes;
155
156	// The location and size of the mmaped area
157	void * ring_ptr;
158	size_t ring_sz;
159	};
160
161	struct __io_data {
162	struct __submition_data submit_q;
163	struct __completion_data completion_q;
164	uint32_t ring_flags;
165	int cltr_flags;
166	int fd;
167	semaphore submit;
168	volatile bool done;
169	struct {
170	struct {
171	__processor_id_t id;
172	void * stack;
173	pthread_t kthrd;
174	volatile bool blocked;
175	} slow;
176	__io_poller_fast fast;
177	__bin_sem_t sem;
178	} poller;
179	};
180
181	//=============================================================================================
182	// I/O Startup / Shutdown logic
183	//=============================================================================================
184	void __kernel_io_startup( cluster & this, unsigned io_flags, bool main_cluster ) {
185	if( (io_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS) && (io_flags & CFA_CLUSTER_IO_EAGER_SUBMITS) ) {
186	abort("CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS and CFA_CLUSTER_IO_EAGER_SUBMITS cannot be mixed\n");
187	}
188
189	this.io = malloc();
190
191	// Step 1 : call to setup
192	struct io_uring_params params;
193	memset(&params, 0, sizeof(params));
194	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_SUBMITS ) params.flags \|= IORING_SETUP_SQPOLL;
195	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_COMPLETES ) params.flags \|= IORING_SETUP_IOPOLL;
196
197	uint32_t nentries = entries_per_cluster();
198
199	int fd = syscall(__NR_io_uring_setup, nentries, &params );
200	if(fd < 0) {
201	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
202	}
203
204	// Step 2 : mmap result
205	memset( this.io, 0, sizeof(struct __io_data) );
206	struct __submition_data & sq = this.io->submit_q;
207	struct __completion_data & cq = this.io->completion_q;
208
209	// calculate the right ring size
210	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
211	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
212
213	// Requires features
214	#if defined(IORING_FEAT_SINGLE_MMAP)
215	// adjust the size according to the parameters
216	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
217	cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz);
218	}
219	#endif
220
221	// mmap the Submit Queue into existence
222	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
223	if (sq.ring_ptr == (void*)MAP_FAILED) {
224	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
225	}
226
227	// Requires features
228	#if defined(IORING_FEAT_SINGLE_MMAP)
229	// mmap the Completion Queue into existence (may or may not be needed)
230	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
231	cq->ring_ptr = sq->ring_ptr;
232	}
233	else
234	#endif
235	{
236	// We need multiple call to MMAP
237	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
238	if (cq.ring_ptr == (void*)MAP_FAILED) {
239	munmap(sq.ring_ptr, sq.ring_sz);
240	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
241	}
242	}
243
244	// mmap the submit queue entries
245	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
246	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
247	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
248	munmap(sq.ring_ptr, sq.ring_sz);
249	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
250	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
251	}
252
253	// Get the pointers from the kernel to fill the structure
254	// submit queue
255	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
256	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
257	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
258	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
259	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
260	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
261	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
262	sq.prev_head = *sq.head;
263
264	{
265	const uint32_t num = *sq.num;
266	for( i; num ) {
267	sq.sqes[i].user_data = 0ul64;
268	}
269	}
270
271	(sq.lock){};
272
273	if( io_flags & ( CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS \| CFA_CLUSTER_IO_EAGER_SUBMITS ) ) {
274	/* paranoid */ verify( is_pow2( io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET ) \|\| ((io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET) < 8) );
275	sq.ready_cnt = max(io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET, 8);
276	sq.ready = alloc_align( 64, sq.ready_cnt );
277	for(i; sq.ready_cnt) {
278	sq.ready[i] = -1ul32;
279	}
280	}
281	else {
282	sq.ready_cnt = 0;
283	sq.ready = 0p;
284	}
285
286	// completion queue
287	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
288	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
289	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
290	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
291	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
292	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
293
294	// some paranoid checks
295	/* 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 );
296	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
297	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
298	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
299
300	/* 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 );
301	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
302	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
303	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
304
305	// Update the global ring info
306	this.io->ring_flags = params.flags;
307	this.io->cltr_flags = io_flags;
308	this.io->fd = fd;
309	this.io->done = false;
310	(this.io->submit){ min(sq.num, cq.num) };
311
312	if(!main_cluster) {
313	__kernel_io_finish_start( this );
314	}
315	}
316
317	void __kernel_io_finish_start( cluster & this ) {
318	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
319	__cfadbg_print_safe(io_core, "Kernel I/O : Creating fast poller for cluter %p\n", &this);
320	(this.io->poller.fast){ this };
321	__thrd_start( this.io->poller.fast, main );
322	}
323
324	// Create the poller thread
325	__cfadbg_print_safe(io_core, "Kernel I/O : Creating slow poller for cluter %p\n", &this);
326	this.io->poller.slow.blocked = false;
327	this.io->poller.slow.stack = __create_pthread( &this.io->poller.slow.kthrd, __io_poller_slow, &this );
328	}
329
330	void __kernel_io_prepare_stop( cluster & this ) {
331	__cfadbg_print_safe(io_core, "Kernel I/O : Stopping pollers for cluster\n", &this);
332	// Notify the poller thread of the shutdown
333	__atomic_store_n(&this.io->done, true, __ATOMIC_SEQ_CST);
334
335	// Stop the IO Poller
336	sigval val = { 1 };
337	pthread_sigqueue( this.io->poller.slow.kthrd, SIGUSR1, val );
338	post( this.io->poller.sem );
339
340	// Wait for the poller thread to finish
341	pthread_join( this.io->poller.slow.kthrd, 0p );
342	free( this.io->poller.slow.stack );
343
344	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller stopped for cluster\n", &this);
345
346	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
347	with( this.io->poller.fast ) {
348	/* paranoid */ verify( this.nprocessors == 0 \|\| &this == mainCluster );
349	/* paranoid */ verify( !ready_mutate_islocked() );
350
351	// We need to adjust the clean-up based on where the thread is
352	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
353
354	ready_schedule_lock( (struct __processor_id_t *)active_processor() );
355
356	// This is the tricky case
357	// The thread was preempted and now it is on the ready queue
358	// The thread should be the last on the list
359	/* paranoid */ verify( thrd.link.next != 0p );
360
361	// Remove the thread from the ready queue of this cluster
362	__attribute__((unused)) bool removed = remove_head( &this, &thrd );
363	/* paranoid */ verify( removed );
364	thrd.link.next = 0p;
365	thrd.link.prev = 0p;
366	__cfaabi_dbg_debug_do( thrd.unpark_stale = true );
367
368	// Fixup the thread state
369	thrd.state = Blocked;
370	thrd.ticket = 0;
371	thrd.preempted = __NO_PREEMPTION;
372
373	ready_schedule_unlock( (struct __processor_id_t *)active_processor() );
374
375	// Pretend like the thread was blocked all along
376	}
377	// !!! This is not an else if !!!
378	if( thrd.state == Blocked ) {
379
380	// This is the "easy case"
381	// The thread is parked and can easily be moved to active cluster
382	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
383	thrd.curr_cluster = active_cluster();
384
385	// unpark the fast io_poller
386	unpark( &thrd __cfaabi_dbg_ctx2 );
387	}
388	else {
389
390	// The thread is in a weird state
391	// I don't know what to do here
392	abort("Fast poller thread is in unexpected state, cannot clean-up correctly\n");
393	}
394
395	}
396
397	^(this.io->poller.fast){};
398
399	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller stopped for cluster\n", &this);
400	}
401	}
402
403	void __kernel_io_shutdown( cluster & this, bool main_cluster ) {
404	if(!main_cluster) {
405	__kernel_io_prepare_stop( this );
406	}
407
408	// Shutdown the io rings
409	struct __submition_data & sq = this.io->submit_q;
410	struct __completion_data & cq = this.io->completion_q;
411
412	// unmap the submit queue entries
413	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
414
415	// unmap the Submit Queue ring
416	munmap(sq.ring_ptr, sq.ring_sz);
417
418	// unmap the Completion Queue ring, if it is different
419	if (cq.ring_ptr != sq.ring_ptr) {
420	munmap(cq.ring_ptr, cq.ring_sz);
421	}
422
423	// close the file descriptor
424	close(this.io->fd);
425
426	free( this.io->submit_q.ready ); // Maybe null, doesn't matter
427	free( this.io );
428	}
429
430	//=============================================================================================
431	// I/O Polling
432	//=============================================================================================
433	static unsigned __collect_submitions( struct __io_data & ring );
434	static uint32_t __release_consumed_submission( struct __io_data & ring );
435
436	// Process a single completion message from the io_uring
437	// This is NOT thread-safe
438	static [int, bool] __drain_io( & struct __io_data ring, * sigset_t mask, int waitcnt, bool in_kernel ) {
439	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
440	const uint32_t smask = *ring.submit_q.mask;
441
442	unsigned to_submit = 0;
443	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
444	// If the poller thread also submits, then we need to aggregate the submissions which are ready
445	to_submit = __collect_submitions( ring );
446	}
447
448	if (to_submit > 0 \|\| waitcnt > 0) {
449	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
450	if( ret < 0 ) {
451	switch((int)errno) {
452	case EAGAIN:
453	case EINTR:
454	return [0, true];
455	default:
456	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
457	}
458	}
459
460	// Release the consumed SQEs
461	__release_consumed_submission( ring );
462
463	// update statistics
464	__STATS__( true,
465	if( to_submit > 0 ) {
466	io.submit_q.submit_avg.rdy += to_submit;
467	io.submit_q.submit_avg.csm += ret;
468	io.submit_q.submit_avg.cnt += 1;
469	}
470	)
471	}
472
473	// Memory barrier
474	__atomic_thread_fence( __ATOMIC_SEQ_CST );
475
476	// Drain the queue
477	unsigned head = *ring.completion_q.head;
478	unsigned tail = *ring.completion_q.tail;
479	const uint32_t mask = *ring.completion_q.mask;
480
481	// Nothing was new return 0
482	if (head == tail) {
483	return [0, to_submit > 0];
484	}
485
486	uint32_t count = tail - head;
487	/* paranoid */ verify( count != 0 );
488	for(i; count) {
489	unsigned idx = (head + i) & mask;
490	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
491
492	/* paranoid */ verify(&cqe);
493
494	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
495	__cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
496
497	data->result = cqe.res;
498	if(!in_kernel) { unpark( data->thrd __cfaabi_dbg_ctx2 ); }
499	else { __unpark( &ring.poller.slow.id, data->thrd __cfaabi_dbg_ctx2 ); }
500	}
501
502	// Allow new submissions to happen
503	// V(ring.submit, count);
504
505	// Mark to the kernel that the cqe has been seen
506	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
507	__atomic_thread_fence( __ATOMIC_SEQ_CST );
508	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
509
510	return [count, count > 0 \|\| to_submit > 0];
511	}
512
513	static void * __io_poller_slow( void * arg ) {
514	#if !defined( __CFA_NO_STATISTICS__ )
515	__stats_t local_stats;
516	__init_stats( &local_stats );
517	kernelTLS.this_stats = &local_stats;
518	#endif
519
520	cluster * cltr = (cluster *)arg;
521	struct __io_data & ring = *cltr->io;
522
523	ring.poller.slow.id.id = doregister( &ring.poller.slow.id );
524
525	sigset_t mask;
526	sigfillset(&mask);
527	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
528	abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
529	}
530
531	sigdelset( &mask, SIGUSR1 );
532
533	verify( (ring.submit_q.head) == (ring.submit_q.tail) );
534	verify( (ring.completion_q.head) == (ring.completion_q.tail) );
535
536	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p ready\n", &ring);
537
538	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
539	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
540
541	__atomic_store_n( &ring.poller.slow.blocked, true, __ATOMIC_SEQ_CST );
542
543	// In the user-thread approach drain and if anything was drained,
544	// batton pass to the user-thread
545	int count;
546	bool again;
547	[count, again] = __drain_io( ring, &mask, 1, true );
548
549	__atomic_store_n( &ring.poller.slow.blocked, false, __ATOMIC_SEQ_CST );
550
551	// Update statistics
552	__STATS__( true,
553	io.complete_q.completed_avg.val += count;
554	io.complete_q.completed_avg.slow_cnt += 1;
555	)
556
557	if(again) {
558	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to fast poller\n", &ring);
559	__unpark( &ring.poller.slow.id, &ring.poller.fast.thrd __cfaabi_dbg_ctx2 );
560	wait( ring.poller.sem );
561	}
562	}
563	}
564	else {
565	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
566	//In the naive approach, just poll the io completion queue directly
567	int count;
568	bool again;
569	[count, again] = __drain_io( ring, &mask, 1, true );
570
571	// Update statistics
572	__STATS__( true,
573	io.complete_q.completed_avg.val += count;
574	io.complete_q.completed_avg.slow_cnt += 1;
575	)
576	}
577	}
578
579	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p stopping\n", &ring);
580
581	unregister( &ring.poller.slow.id );
582
583	#if !defined(__CFA_NO_STATISTICS__)
584	__tally_stats(cltr->stats, &local_stats);
585	#endif
586
587	return 0p;
588	}
589
590	void main( __io_poller_fast & this ) {
591	verify( this.ring->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD );
592
593	// Start parked
594	park( __cfaabi_dbg_ctx );
595
596	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p ready\n", &this.ring);
597
598	int reset = 0;
599
600	// Then loop until we need to start
601	while(!__atomic_load_n(&this.ring->done, __ATOMIC_SEQ_CST)) {
602
603	// Drain the io
604	int count;
605	bool again;
606	disable_interrupts();
607	[count, again] = __drain_io( *this.ring, 0p, 0, false );
608
609	if(!again) reset++;
610
611	// Update statistics
612	__STATS__( true,
613	io.complete_q.completed_avg.val += count;
614	io.complete_q.completed_avg.fast_cnt += 1;
615	)
616	enable_interrupts( __cfaabi_dbg_ctx );
617
618	// If we got something, just yield and check again
619	if(reset < 5) {
620	yield();
621	}
622	// We didn't get anything baton pass to the slow poller
623	else {
624	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to slow poller\n", &this.ring);
625	reset = 0;
626
627	// wake up the slow poller
628	post( this.ring->poller.sem );
629
630	// park this thread
631	park( __cfaabi_dbg_ctx );
632	}
633	}
634
635	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
636	}
637
638	static inline void __wake_poller( struct __io_data & ring ) __attribute__((artificial));
639	static inline void __wake_poller( struct __io_data & ring ) {
640	if(!__atomic_load_n( &ring.poller.slow.blocked, __ATOMIC_SEQ_CST)) return;
641
642	sigval val = { 1 };
643	pthread_sigqueue( ring.poller.slow.kthrd, SIGUSR1, val );
644	}
645
646	//=============================================================================================
647	// I/O Submissions
648	//=============================================================================================
649
650	// Submition steps :
651	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
652	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
653	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
654	// need to write an allocator that allows allocating concurrently.
655	//
656	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
657	//
658	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
659	// needs to arrive to two concensus at the same time:
660	// A - The order in which entries are listed in the array: no two threads must pick the
661	// same index for their entries
662	// B - When can the tail be update for the kernel. EVERY entries in the array between
663	// head and tail must be fully filled and shouldn't ever be touched again.
664	//
665
666	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
667	/* paranoid */ verify( data != 0 );
668
669	// Prepare the data we need
670	__attribute((unused)) int len = 0;
671	__attribute((unused)) int block = 0;
672	uint32_t cnt = *ring.submit_q.num;
673	uint32_t mask = *ring.submit_q.mask;
674
675	disable_interrupts();
676	uint32_t off = __tls_rand();
677	enable_interrupts( __cfaabi_dbg_ctx );
678
679	// Loop around looking for an available spot
680	for() {
681	// Look through the list starting at some offset
682	for(i; cnt) {
683	uint64_t expected = 0;
684	uint32_t idx = (i + off) & mask;
685	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
686	volatile uint64_t * udata = &sqe->user_data;
687
688	if( *udata == expected &&
689	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
690	{
691	// update statistics
692	__STATS__( false,
693	io.submit_q.alloc_avg.val += len;
694	io.submit_q.alloc_avg.block += block;
695	io.submit_q.alloc_avg.cnt += 1;
696	)
697
698
699	// Success return the data
700	return [sqe, idx];
701	}
702	verify(expected != data);
703
704	len ++;
705	}
706
707	block++;
708	yield();
709	}
710	}
711
712	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
713	/* paranoid */ verify( idx <= mask );
714	/* paranoid */ verify( idx != -1ul32 );
715
716	// We need to find a spot in the ready array
717	__attribute((unused)) int len = 0;
718	__attribute((unused)) int block = 0;
719	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
720
721	disable_interrupts();
722	uint32_t off = __tls_rand();
723	enable_interrupts( __cfaabi_dbg_ctx );
724
725	uint32_t picked;
726	LOOKING: for() {
727	for(i; ring.submit_q.ready_cnt) {
728	picked = (i + off) & ready_mask;
729	uint32_t expected = -1ul32;
730	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
731	break LOOKING;
732	}
733	verify(expected != idx);
734
735	len ++;
736	}
737
738	block++;
739	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
740	__release_consumed_submission( ring );
741	unlock( ring.submit_q.lock );
742	}
743	else {
744	yield();
745	}
746	}
747
748	// update statistics
749	__STATS__( false,
750	io.submit_q.look_avg.val += len;
751	io.submit_q.look_avg.block += block;
752	io.submit_q.look_avg.cnt += 1;
753	)
754
755	return picked;
756	}
757
758	void __submit( struct __io_data & ring, uint32_t idx ) {
759	// Get now the data we definetely need
760	uint32_t * const tail = ring.submit_q.tail;
761	const uint32_t mask = *ring.submit_q.mask;
762
763	// There are 2 submission schemes, check which one we are using
764	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
765	// If the poller thread submits, then we just need to add this to the ready array
766	__submit_to_ready_array( ring, idx, mask );
767
768	__wake_poller( ring );
769
770	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
771	}
772	else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) {
773	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
774
775	for() {
776	yield();
777
778	// If some one else collected our index, we are done
779	#warning ABA problem
780	if( ring.submit_q.ready[picked] != idx ) {
781	__STATS__( false,
782	io.submit_q.helped += 1;
783	)
784	return;
785	}
786
787	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
788	__STATS__( false,
789	io.submit_q.leader += 1;
790	)
791	break;
792	}
793
794	__STATS__( false,
795	io.submit_q.busy += 1;
796	)
797	}
798
799	// We got the lock
800	unsigned to_submit = __collect_submitions( ring );
801	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8);
802	if( ret < 0 ) {
803	switch((int)errno) {
804	case EAGAIN:
805	case EINTR:
806	unlock(ring.submit_q.lock);
807	return;
808	default:
809	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
810	}
811	}
812
813	/* paranoid */ verify( ret > 0 );
814
815	// Release the consumed SQEs
816	__release_consumed_submission( ring );
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	// Release the consumed SQEs
854	__release_consumed_submission( ring );
855
856	unlock(ring.submit_q.lock);
857
858	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
859	}
860	}
861
862	static unsigned __collect_submitions( struct __io_data & ring ) {
863	/* paranoid */ verify( ring.submit_q.ready != 0p );
864	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
865
866	unsigned to_submit = 0;
867	uint32_t tail = *ring.submit_q.tail;
868	const uint32_t mask = *ring.submit_q.mask;
869
870	// Go through the list of ready submissions
871	for( i; ring.submit_q.ready_cnt ) {
872	// replace any submission with the sentinel, to consume it.
873	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
874
875	// If it was already the sentinel, then we are done
876	if( idx == -1ul32 ) continue;
877
878	// If we got a real submission, append it to the list
879	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
880	to_submit++;
881	}
882
883	// Increment the tail based on how many we are ready to submit
884	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
885
886	return to_submit;
887	}
888
889	static uint32_t __release_consumed_submission( struct __io_data & ring ) {
890	const uint32_t smask = *ring.submit_q.mask;
891	uint32_t chead = *ring.submit_q.head;
892	uint32_t phead = ring.submit_q.prev_head;
893	ring.submit_q.prev_head = chead;
894	uint32_t count = chead - phead;
895	for( i; count ) {
896	uint32_t idx = ring.submit_q.array[ (phead + i) & smask ];
897	ring.submit_q.sqes[ idx ].user_data = 0;
898	}
899	return count;
900	}
901	#endif

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

Download in other formats:

Original Format