Context Navigation

io.cfa @ 0a92c78

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

Last change on this file since 0a92c78 was fb98462, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Fixed io.cfa for machines supporting single map feature
Property mode set to `100644`
File size: 27.7 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
443	unsigned to_submit = 0;
444	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
445	// If the poller thread also submits, then we need to aggregate the submissions which are ready
446	to_submit = __collect_submitions( ring );
447	}
448
449	if (to_submit > 0 \|\| waitcnt > 0) {
450	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
451	if( ret < 0 ) {
452	switch((int)errno) {
453	case EAGAIN:
454	case EINTR:
455	return [0, true];
456	default:
457	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
458	}
459	}
460
461	// Release the consumed SQEs
462	__release_consumed_submission( ring );
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	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
741	__release_consumed_submission( ring );
742	unlock( ring.submit_q.lock );
743	}
744	else {
745	yield();
746	}
747	}
748
749	// update statistics
750	__STATS__( false,
751	io.submit_q.look_avg.val += len;
752	io.submit_q.look_avg.block += block;
753	io.submit_q.look_avg.cnt += 1;
754	)
755
756	return picked;
757	}
758
759	void __submit( struct __io_data & ring, uint32_t idx ) {
760	// Get now the data we definetely need
761	uint32_t * const tail = ring.submit_q.tail;
762	const uint32_t mask = *ring.submit_q.mask;
763
764	// There are 2 submission schemes, check which one we are using
765	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) {
766	// If the poller thread submits, then we just need to add this to the ready array
767	__submit_to_ready_array( ring, idx, mask );
768
769	__wake_poller( ring );
770
771	__cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
772	}
773	else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) {
774	uint32_t picked = __submit_to_ready_array( ring, idx, mask );
775
776	for() {
777	yield();
778
779	// If some one else collected our index, we are done
780	#warning ABA problem
781	if( ring.submit_q.ready[picked] != idx ) {
782	__STATS__( false,
783	io.submit_q.helped += 1;
784	)
785	return;
786	}
787
788	if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) {
789	__STATS__( false,
790	io.submit_q.leader += 1;
791	)
792	break;
793	}
794
795	__STATS__( false,
796	io.submit_q.busy += 1;
797	)
798	}
799
800	// We got the lock
801	unsigned to_submit = __collect_submitions( ring );
802	int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8);
803	if( ret < 0 ) {
804	switch((int)errno) {
805	case EAGAIN:
806	case EINTR:
807	unlock(ring.submit_q.lock);
808	return;
809	default:
810	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
811	}
812	}
813
814	/* paranoid */ verify( ret > 0 );
815
816	// Release the consumed SQEs
817	__release_consumed_submission( ring );
818
819	// update statistics
820	__STATS__( true,
821	io.submit_q.submit_avg.rdy += to_submit;
822	io.submit_q.submit_avg.csm += ret;
823	io.submit_q.submit_avg.cnt += 1;
824	)
825
826	unlock(ring.submit_q.lock);
827	}
828	else {
829	// get mutual exclusion
830	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
831
832	// Append to the list of ready entries
833
834	/* paranoid */ verify( idx <= mask );
835
836	ring.submit_q.array[ (*tail) & mask ] = idx & mask;
837	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
838
839	/* paranoid */ verify( ring.submit_q.sqes[ idx ].user_data != 0 );
840
841	// Submit however, many entries need to be submitted
842	int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
843	if( ret < 0 ) {
844	switch((int)errno) {
845	default:
846	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
847	}
848	}
849
850	// update statistics
851	__STATS__( false,
852	io.submit_q.submit_avg.csm += 1;
853	io.submit_q.submit_avg.cnt += 1;
854	)
855
856	// Release the consumed SQEs
857	__release_consumed_submission( ring );
858
859	unlock(ring.submit_q.lock);
860
861	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
862	}
863	}
864
865	static unsigned __collect_submitions( struct __io_data & ring ) {
866	/* paranoid */ verify( ring.submit_q.ready != 0p );
867	/* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
868
869	unsigned to_submit = 0;
870	uint32_t tail = *ring.submit_q.tail;
871	const uint32_t mask = *ring.submit_q.mask;
872
873	// Go through the list of ready submissions
874	for( i; ring.submit_q.ready_cnt ) {
875	// replace any submission with the sentinel, to consume it.
876	uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
877
878	// If it was already the sentinel, then we are done
879	if( idx == -1ul32 ) continue;
880
881	// If we got a real submission, append it to the list
882	ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
883	to_submit++;
884	}
885
886	// Increment the tail based on how many we are ready to submit
887	__atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
888
889	return to_submit;
890	}
891
892	static uint32_t __release_consumed_submission( struct __io_data & ring ) {
893	const uint32_t smask = *ring.submit_q.mask;
894
895	if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
896	uint32_t chead = *ring.submit_q.head;
897	uint32_t phead = ring.submit_q.prev_head;
898	ring.submit_q.prev_head = chead;
899	unlock(ring.submit_q.release_lock);
900
901	uint32_t count = chead - phead;
902	for( i; count ) {
903	uint32_t idx = ring.submit_q.array[ (phead + i) & smask ];
904	ring.submit_q.sqes[ idx ].user_data = 0;
905	}
906	return count;
907	}
908	#endif

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

Download in other formats:

Original Format