Context Navigation

io.cfa@ 3a32b3a

Visit:

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

Last change on this file since 3a32b3a was 732b406, checked in by Thierry Delisle <tdelisle@…>, 6 years ago
Added missing lock around release_consumed_submission
Property mode set to `100644`
File size: 27.8 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// io.cfa --
8	//
9	// Author : Thierry Delisle
10	// Created On : Thu Apr 23 17:31:00 2020
11	// Last Modified By :
12	// Last Modified On :
13	// Update Count :
14	//
15
16	// #define __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	__spinlock_t release_lock;
133
134	// A buffer of sqes (not the actual ring)
135	struct io_uring_sqe * sqes;
136
137	// The location and size of the mmaped area
138	void * ring_ptr;
139	size_t ring_sz;
140	};
141
142	struct __completion_data {
143	// Head and tail of the ring
144	volatile uint32_t * head;
145	volatile uint32_t * tail;
146
147	// number of entries and mask to go with it
148	const uint32_t * mask;
149	const uint32_t * num;
150
151	// number of cqes not submitted (whatever that means)
152	uint32_t * overflow;
153
154	// the kernel ring
155	struct io_uring_cqe * cqes;
156
157	// The location and size of the mmaped area
158	void * ring_ptr;
159	size_t ring_sz;
160	};
161
162	struct __io_data {
163	struct __submition_data submit_q;
164	struct __completion_data completion_q;
165	uint32_t ring_flags;
166	int cltr_flags;
167	int fd;
168	semaphore submit;
169	volatile bool done;
170	struct {
171	struct {
172	__processor_id_t id;
173	void * stack;
174	pthread_t kthrd;
175	volatile bool blocked;
176	} slow;
177	__io_poller_fast fast;
178	__bin_sem_t sem;
179	} poller;
180	};
181
182	//=============================================================================================
183	// I/O Startup / Shutdown logic
184	//=============================================================================================
185	void __kernel_io_startup( cluster & this, unsigned io_flags, bool main_cluster ) {
186	if( (io_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS) && (io_flags & CFA_CLUSTER_IO_EAGER_SUBMITS) ) {
187	abort("CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS and CFA_CLUSTER_IO_EAGER_SUBMITS cannot be mixed\n");
188	}
189
190	this.io = malloc();
191
192	// Step 1 : call to setup
193	struct io_uring_params params;
194	memset(&params, 0, sizeof(params));
195	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_SUBMITS ) params.flags \|= IORING_SETUP_SQPOLL;
196	if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_COMPLETES ) params.flags \|= IORING_SETUP_IOPOLL;
197
198	uint32_t nentries = entries_per_cluster();
199
200	int fd = syscall(__NR_io_uring_setup, nentries, &params );
201	if(fd < 0) {
202	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
203	}
204
205	// Step 2 : mmap result
206	memset( this.io, 0, sizeof(struct __io_data) );
207	struct __submition_data & sq = this.io->submit_q;
208	struct __completion_data & cq = this.io->completion_q;
209
210	// calculate the right ring size
211	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
212	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
213
214	// Requires features
215	#if defined(IORING_FEAT_SINGLE_MMAP)
216	// adjust the size according to the parameters
217	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
218	cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz);
219	}
220	#endif
221
222	// mmap the Submit Queue into existence
223	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
224	if (sq.ring_ptr == (void*)MAP_FAILED) {
225	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
226	}
227
228	// Requires features
229	#if defined(IORING_FEAT_SINGLE_MMAP)
230	// mmap the Completion Queue into existence (may or may not be needed)
231	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
232	cq->ring_ptr = sq->ring_ptr;
233	}
234	else
235	#endif
236	{
237	// We need multiple call to MMAP
238	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
239	if (cq.ring_ptr == (void*)MAP_FAILED) {
240	munmap(sq.ring_ptr, sq.ring_sz);
241	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
242	}
243	}
244
245	// mmap the submit queue entries
246	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
247	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
248	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
249	munmap(sq.ring_ptr, sq.ring_sz);
250	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
251	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
252	}
253
254	// Get the pointers from the kernel to fill the structure
255	// submit queue
256	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
257	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
258	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
259	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
260	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
261	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
262	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
263	sq.prev_head = *sq.head;
264
265	{
266	const uint32_t num = *sq.num;
267	for( i; num ) {
268	sq.sqes[i].user_data = 0ul64;
269	}
270	}
271
272	(sq.lock){};
273	(sq.release_lock){};
274
275	if( io_flags & ( CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS \| CFA_CLUSTER_IO_EAGER_SUBMITS ) ) {
276	/* paranoid */ verify( is_pow2( io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET ) \|\| ((io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET) < 8) );
277	sq.ready_cnt = max(io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET, 8);
278	sq.ready = alloc_align( 64, sq.ready_cnt );
279	for(i; sq.ready_cnt) {
280	sq.ready[i] = -1ul32;
281	}
282	}
283	else {
284	sq.ready_cnt = 0;
285	sq.ready = 0p;
286	}
287
288	// completion queue
289	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
290	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
291	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
292	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
293	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
294	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
295
296	// some paranoid checks
297	/* 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 );
298	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
299	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
300	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
301
302	/* 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 );
303	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
304	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
305	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
306
307	// Update the global ring info
308	this.io->ring_flags = params.flags;
309	this.io->cltr_flags = io_flags;
310	this.io->fd = fd;
311	this.io->done = false;
312	(this.io->submit){ min(sq.num, cq.num) };
313
314	if(!main_cluster) {
315	__kernel_io_finish_start( this );
316	}
317	}
318
319	void __kernel_io_finish_start( cluster & this ) {
320	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
321	__cfadbg_print_safe(io_core, "Kernel I/O : Creating fast poller for cluter %p\n", &this);
322	(this.io->poller.fast){ this };
323	__thrd_start( this.io->poller.fast, main );
324	}
325
326	// Create the poller thread
327	__cfadbg_print_safe(io_core, "Kernel I/O : Creating slow poller for cluter %p\n", &this);
328	this.io->poller.slow.blocked = false;
329	this.io->poller.slow.stack = __create_pthread( &this.io->poller.slow.kthrd, __io_poller_slow, &this );
330	}
331
332	void __kernel_io_prepare_stop( cluster & this ) {
333	__cfadbg_print_safe(io_core, "Kernel I/O : Stopping pollers for cluster\n", &this);
334	// Notify the poller thread of the shutdown
335	__atomic_store_n(&this.io->done, true, __ATOMIC_SEQ_CST);
336
337	// Stop the IO Poller
338	sigval val = { 1 };
339	pthread_sigqueue( this.io->poller.slow.kthrd, SIGUSR1, val );
340	post( this.io->poller.sem );
341
342	// Wait for the poller thread to finish
343	pthread_join( this.io->poller.slow.kthrd, 0p );
344	free( this.io->poller.slow.stack );
345
346	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller stopped for cluster\n", &this);
347
348	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
349	with( this.io->poller.fast ) {
350	/* paranoid */ verify( this.nprocessors == 0 \|\| &this == mainCluster );
351	/* paranoid */ verify( !ready_mutate_islocked() );
352
353	// We need to adjust the clean-up based on where the thread is
354	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
355
356	ready_schedule_lock( (struct __processor_id_t *)active_processor() );
357
358	// This is the tricky case
359	// The thread was preempted and now it is on the ready queue
360	// The thread should be the last on the list
361	/* paranoid */ verify( thrd.link.next != 0p );
362
363	// Remove the thread from the ready queue of this cluster
364	__attribute__((unused)) bool removed = remove_head( &this, &thrd );
365	/* paranoid */ verify( removed );
366	thrd.link.next = 0p;
367	thrd.link.prev = 0p;
368	__cfaabi_dbg_debug_do( thrd.unpark_stale = true );
369
370	// Fixup the thread state
371	thrd.state = Blocked;
372	thrd.ticket = 0;
373	thrd.preempted = __NO_PREEMPTION;
374
375	ready_schedule_unlock( (struct __processor_id_t *)active_processor() );
376
377	// Pretend like the thread was blocked all along
378	}
379	// !!! This is not an else if !!!
380	if( thrd.state == Blocked ) {
381
382	// This is the "easy case"
383	// The thread is parked and can easily be moved to active cluster
384	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
385	thrd.curr_cluster = active_cluster();
386
387	// unpark the fast io_poller
388	unpark( &thrd __cfaabi_dbg_ctx2 );
389	}
390	else {
391
392	// The thread is in a weird state
393	// I don't know what to do here
394	abort("Fast poller thread is in unexpected state, cannot clean-up correctly\n");
395	}
396
397	}
398
399	^(this.io->poller.fast){};
400
401	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller stopped for cluster\n", &this);
402	}
403	}
404
405	void __kernel_io_shutdown( cluster & this, bool main_cluster ) {
406	if(!main_cluster) {
407	__kernel_io_prepare_stop( this );
408	}
409
410	// Shutdown the io rings
411	struct __submition_data & sq = this.io->submit_q;
412	struct __completion_data & cq = this.io->completion_q;
413
414	// unmap the submit queue entries
415	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
416
417	// unmap the Submit Queue ring
418	munmap(sq.ring_ptr, sq.ring_sz);
419
420	// unmap the Completion Queue ring, if it is different
421	if (cq.ring_ptr != sq.ring_ptr) {
422	munmap(cq.ring_ptr, cq.ring_sz);
423	}
424
425	// close the file descriptor
426	close(this.io->fd);
427
428	free( this.io->submit_q.ready ); // Maybe null, doesn't matter
429	free( this.io );
430	}
431
432	//=============================================================================================
433	// I/O Polling
434	//=============================================================================================
435	static unsigned __collect_submitions( struct __io_data & ring );
436	static uint32_t __release_consumed_submission( struct __io_data & ring );
437
438	// Process a single completion message from the io_uring
439	// This is NOT thread-safe
440	static [int, bool] __drain_io( & struct __io_data ring, * sigset_t mask, int waitcnt, bool in_kernel ) {
441	/* paranoid */ verify( !kernelTLS.preemption_state.enabled );
442	const uint32_t smask = *ring.submit_q.mask;
443
444	unsigned to_submit = 0;
445	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
446	// If the poller thread also submits, then we need to aggregate the submissions which are ready
447	to_submit = __collect_submitions( ring );
448	}
449
450	if (to_submit > 0 \|\| waitcnt > 0) {
451	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
452	if( ret < 0 ) {
453	switch((int)errno) {
454	case EAGAIN:
455	case EINTR:
456	return [0, true];
457	default:
458	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
459	}
460	}
461
462	// Release the consumed SQEs
463	__release_consumed_submission( ring );
464
465	// update statistics
466	__STATS__( true,
467	if( to_submit > 0 ) {
468	io.submit_q.submit_avg.rdy += to_submit;
469	io.submit_q.submit_avg.csm += ret;
470	io.submit_q.submit_avg.cnt += 1;
471	}
472	)
473	}
474
475	// Memory barrier
476	__atomic_thread_fence( __ATOMIC_SEQ_CST );
477
478	// Drain the queue
479	unsigned head = *ring.completion_q.head;
480	unsigned tail = *ring.completion_q.tail;
481	const uint32_t mask = *ring.completion_q.mask;
482
483	// Nothing was new return 0
484	if (head == tail) {
485	return [0, to_submit > 0];
486	}
487
488	uint32_t count = tail - head;
489	/* paranoid */ verify( count != 0 );
490	for(i; count) {
491	unsigned idx = (head + i) & mask;
492	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
493
494	/* paranoid */ verify(&cqe);
495
496	struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data;
497	__cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
498
499	data->result = cqe.res;
500	if(!in_kernel) { unpark( data->thrd __cfaabi_dbg_ctx2 ); }
501	else { __unpark( &ring.poller.slow.id, data->thrd __cfaabi_dbg_ctx2 ); }
502	}
503
504	// Allow new submissions to happen
505	// V(ring.submit, count);
506
507	// Mark to the kernel that the cqe has been seen
508	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
509	__atomic_thread_fence( __ATOMIC_SEQ_CST );
510	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
511
512	return [count, count > 0 \|\| to_submit > 0];
513	}
514
515	static void * __io_poller_slow( void * arg ) {
516	#if !defined( __CFA_NO_STATISTICS__ )
517	__stats_t local_stats;
518	__init_stats( &local_stats );
519	kernelTLS.this_stats = &local_stats;
520	#endif
521
522	cluster * cltr = (cluster *)arg;
523	struct __io_data & ring = *cltr->io;
524
525	ring.poller.slow.id.id = doregister( &ring.poller.slow.id );
526
527	sigset_t mask;
528	sigfillset(&mask);
529	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
530	abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
531	}
532
533	sigdelset( &mask, SIGUSR1 );
534
535	verify( (ring.submit_q.head) == (ring.submit_q.tail) );
536	verify( (ring.completion_q.head) == (ring.completion_q.tail) );
537
538	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p ready\n", &ring);
539
540	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
541	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
542
543	__atomic_store_n( &ring.poller.slow.blocked, true, __ATOMIC_SEQ_CST );
544
545	// In the user-thread approach drain and if anything was drained,
546	// batton pass to the user-thread
547	int count;
548	bool again;
549	[count, again] = __drain_io( ring, &mask, 1, true );
550
551	__atomic_store_n( &ring.poller.slow.blocked, false, __ATOMIC_SEQ_CST );
552
553	// Update statistics
554	__STATS__( true,
555	io.complete_q.completed_avg.val += count;
556	io.complete_q.completed_avg.slow_cnt += 1;
557	)
558
559	if(again) {
560	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to fast poller\n", &ring);
561	__unpark( &ring.poller.slow.id, &ring.poller.fast.thrd __cfaabi_dbg_ctx2 );
562	wait( ring.poller.sem );
563	}
564	}
565	}
566	else {
567	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
568	//In the naive approach, just poll the io completion queue directly
569	int count;
570	bool again;
571	[count, again] = __drain_io( ring, &mask, 1, true );
572
573	// Update statistics
574	__STATS__( true,
575	io.complete_q.completed_avg.val += count;
576	io.complete_q.completed_avg.slow_cnt += 1;
577	)
578	}
579	}
580
581	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p stopping\n", &ring);
582
583	unregister( &ring.poller.slow.id );
584
585	#if !defined(__CFA_NO_STATISTICS__)
586	__tally_stats(cltr->stats, &local_stats);
587	#endif
588
589	return 0p;
590	}
591
592	void main( __io_poller_fast & this ) {
593	verify( this.ring->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD );
594
595	// Start parked
596	park( __cfaabi_dbg_ctx );
597
598	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p ready\n", &this.ring);
599
600	int reset = 0;
601
602	// Then loop until we need to start
603	while(!__atomic_load_n(&this.ring->done, __ATOMIC_SEQ_CST)) {
604
605	// Drain the io
606	int count;
607	bool again;
608	disable_interrupts();
609	[count, again] = __drain_io( *this.ring, 0p, 0, false );
610
611	if(!again) reset++;
612
613	// Update statistics
614	__STATS__( true,
615	io.complete_q.completed_avg.val += count;
616	io.complete_q.completed_avg.fast_cnt += 1;
617	)
618	enable_interrupts( __cfaabi_dbg_ctx );
619
620	// If we got something, just yield and check again
621	if(reset < 5) {
622	yield();
623	}
624	// We didn't get anything baton pass to the slow poller
625	else {
626	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to slow poller\n", &this.ring);
627	reset = 0;
628
629	// wake up the slow poller
630	post( this.ring->poller.sem );
631
632	// park this thread
633	park( __cfaabi_dbg_ctx );
634	}
635	}
636
637	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
638	}
639
640	static inline void __wake_poller( struct __io_data & ring ) __attribute__((artificial));
641	static inline void __wake_poller( struct __io_data & ring ) {
642	if(!__atomic_load_n( &ring.poller.slow.blocked, __ATOMIC_SEQ_CST)) return;
643
644	sigval val = { 1 };
645	pthread_sigqueue( ring.poller.slow.kthrd, SIGUSR1, val );
646	}
647
648	//=============================================================================================
649	// I/O Submissions
650	//=============================================================================================
651
652	// Submition steps :
653	// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
654	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
655	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
656	// need to write an allocator that allows allocating concurrently.
657	//
658	// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
659	//
660	// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
661	// needs to arrive to two concensus at the same time:
662	// A - The order in which entries are listed in the array: no two threads must pick the
663	// same index for their entries
664	// B - When can the tail be update for the kernel. EVERY entries in the array between
665	// head and tail must be fully filled and shouldn't ever be touched again.
666	//
667
668	[* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) {
669	/* paranoid */ verify( data != 0 );
670
671	// Prepare the data we need
672	__attribute((unused)) int len = 0;
673	__attribute((unused)) int block = 0;
674	uint32_t cnt = *ring.submit_q.num;
675	uint32_t mask = *ring.submit_q.mask;
676
677	disable_interrupts();
678	uint32_t off = __tls_rand();
679	enable_interrupts( __cfaabi_dbg_ctx );
680
681	// Loop around looking for an available spot
682	for() {
683	// Look through the list starting at some offset
684	for(i; cnt) {
685	uint64_t expected = 0;
686	uint32_t idx = (i + off) & mask;
687	struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
688	volatile uint64_t * udata = &sqe->user_data;
689
690	if( *udata == expected &&
691	__atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
692	{
693	// update statistics
694	__STATS__( false,
695	io.submit_q.alloc_avg.val += len;
696	io.submit_q.alloc_avg.block += block;
697	io.submit_q.alloc_avg.cnt += 1;
698	)
699
700
701	// Success return the data
702	return [sqe, idx];
703	}
704	verify(expected != data);
705
706	len ++;
707	}
708
709	block++;
710	yield();
711	}
712	}
713
714	static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) {
715	/* paranoid */ verify( idx <= mask );
716	/* paranoid */ verify( idx != -1ul32 );
717
718	// We need to find a spot in the ready array
719	__attribute((unused)) int len = 0;
720	__attribute((unused)) int block = 0;
721	uint32_t ready_mask = ring.submit_q.ready_cnt - 1;
722
723	disable_interrupts();
724	uint32_t off = __tls_rand();
725	enable_interrupts( __cfaabi_dbg_ctx );
726
727	uint32_t picked;
728	LOOKING: for() {
729	for(i; ring.submit_q.ready_cnt) {
730	picked = (i + off) & ready_mask;
731	uint32_t expected = -1ul32;
732	if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
733	break LOOKING;
734	}
735	verify(expected != idx);
736
737	len ++;
738	}
739
740	block++;
741	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
742	__release_consumed_submission( ring );
743	unlock( ring.submit_q.lock );
744	}
745	else {
746	yield();
747	}
748	}
749
750	// update statistics
751	__STATS__( false,
752	io.submit_q.look_avg.val += len;
753	io.submit_q.look_avg.block += block;
754	io.submit_q.look_avg.cnt += 1;
755	)
756
757	return picked;
758	}
759
760	void __submit( struct __io_data & ring, uint32_t idx ) {
761	// Get now the data we definetely need
762	uint32_t * const tail = ring.submit_q.tail;
763	const uint32_t mask = *ring.submit_q.mask;
764
765	// There are 2 submission schemes, check which one we are using
766	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
767	// If the poller thread submits, then we just need to add this to the ready array
768	__submit_to_ready_array( ring, idx, mask );
769
770	__wake_poller( ring );
771
772	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
773	}
774	else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) {
775	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
776
777	for() {
778	yield();
779
780	// If some one else collected our index, we are done
781	#warning ABA problem
782	if( ring.submit_q.ready[picked] != idx ) {
783	__STATS__( false,
784	io.submit_q.helped += 1;
785	)
786	return;
787	}
788
789	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
790	__STATS__( false,
791	io.submit_q.leader += 1;
792	)
793	break;
794	}
795
796	__STATS__( false,
797	io.submit_q.busy += 1;
798	)
799	}
800
801	// We got the lock
802	unsigned to_submit = __collect_submitions( ring );
803	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8);
804	if( ret < 0 ) {
805	switch((int)errno) {
806	case EAGAIN:
807	case EINTR:
808	unlock(ring.submit_q.lock);
809	return;
810	default:
811	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
812	}
813	}
814
815	/* paranoid */ verify( ret > 0 );
816
817	// Release the consumed SQEs
818	__release_consumed_submission( ring );
819
820	// update statistics
821	__STATS__( true,
822	io.submit_q.submit_avg.rdy += to_submit;
823	io.submit_q.submit_avg.csm += ret;
824	io.submit_q.submit_avg.cnt += 1;
825	)
826
827	unlock(ring.submit_q.lock);
828	}
829	else {
830	// get mutual exclusion
831	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
832
833	// Append to the list of ready entries
834
835	/* paranoid */ verify( idx <= mask );
836
837	ring.submit_q.array[ (*tail) & mask ] = idx & mask;
838	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
839
840	/* paranoid */ verify( ring.submit_q.sqes[ idx ].user_data != 0 );
841
842	// Submit however, many entries need to be submitted
843	int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
844	if( ret < 0 ) {
845	switch((int)errno) {
846	default:
847	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
848	}
849	}
850
851	// update statistics
852	__STATS__( false,
853	io.submit_q.submit_avg.csm += 1;
854	io.submit_q.submit_avg.cnt += 1;
855	)
856
857	// Release the consumed SQEs
858	__release_consumed_submission( ring );
859
860	unlock(ring.submit_q.lock);
861
862	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
863	}
864	}
865
866	static unsigned __collect_submitions( struct __io_data & ring ) {
867	/* paranoid */ verify( ring.submit_q.ready != 0p );
868	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
869
870	unsigned to_submit = 0;
871	uint32_t tail = *ring.submit_q.tail;
872	const uint32_t mask = *ring.submit_q.mask;
873
874	// Go through the list of ready submissions
875	for( i; ring.submit_q.ready_cnt ) {
876	// replace any submission with the sentinel, to consume it.
877	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
878
879	// If it was already the sentinel, then we are done
880	if( idx == -1ul32 ) continue;
881
882	// If we got a real submission, append it to the list
883	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
884	to_submit++;
885	}
886
887	// Increment the tail based on how many we are ready to submit
888	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
889
890	return to_submit;
891	}
892
893	static uint32_t __release_consumed_submission( struct __io_data & ring ) {
894	const uint32_t smask = *ring.submit_q.mask;
895
896	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
897	uint32_t chead = *ring.submit_q.head;
898	uint32_t phead = ring.submit_q.prev_head;
899	ring.submit_q.prev_head = chead;
900	unlock(ring.submit_q.release_lock);
901
902	uint32_t count = chead - phead;
903	for( i; count ) {
904	uint32_t idx = ring.submit_q.array[ (phead + i) & smask ];
905	ring.submit_q.sqes[ idx ].user_data = 0;
906	}
907	return count;
908	}
909	#endif

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

Download in other formats:

Original Format