Context Navigation

io.cfa@ 6e33a2d

Visit:

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

Last change on this file since 6e33a2d was 4e74466, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
Fast poller thread now polls a few times before baton-passing to the slow thread. Currently doing 5 tries, the number is arbitrary and requires more work.
Property mode set to `100644`
File size: 31.5 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
21	#if !defined(HAVE_LINUX_IO_URING_H)
22	void __kernel_io_startup( cluster &, int, bool ) {
23	// Nothing to do without io_uring
24	}
25
26	void __kernel_io_finish_start( cluster & ) {
27	// Nothing to do without io_uring
28	}
29
30	void __kernel_io_prepare_stop( cluster & ) {
31	// Nothing to do without io_uring
32	}
33
34	void __kernel_io_shutdown( cluster &, bool ) {
35	// Nothing to do without io_uring
36	}
37
38	#else
39	extern "C" {
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	#include <sys/syscall.h>
47
48	#include <linux/io_uring.h>
49	}
50
51	#include "bits/signal.hfa"
52	#include "kernel_private.hfa"
53	#include "thread.hfa"
54
55	uint32_t entries_per_cluster() {
56	return 256;
57	}
58
59	static void * __io_poller_slow( void * arg );
60
61	// Weirdly, some systems that do support io_uring don't actually define these
62	#ifdef __alpha__
63	/*
64	* alpha is the only exception, all other architectures
65	* have common numbers for new system calls.
66	*/
67	#ifndef __NR_io_uring_setup
68	#define __NR_io_uring_setup 535
69	#endif
70	#ifndef __NR_io_uring_enter
71	#define __NR_io_uring_enter 536
72	#endif
73	#ifndef __NR_io_uring_register
74	#define __NR_io_uring_register 537
75	#endif
76	#else /* !__alpha__ */
77	#ifndef __NR_io_uring_setup
78	#define __NR_io_uring_setup 425
79	#endif
80	#ifndef __NR_io_uring_enter
81	#define __NR_io_uring_enter 426
82	#endif
83	#ifndef __NR_io_uring_register
84	#define __NR_io_uring_register 427
85	#endif
86	#endif
87
88	// Fast poller user-thread
89	// Not using the "thread" keyword because we want to control
90	// more carefully when to start/stop it
91	struct __io_poller_fast {
92	struct __io_data * ring;
93	bool waiting;
94	$thread thrd;
95	};
96
97	void ?{}( __io_poller_fast & this, struct cluster & cltr ) {
98	this.ring = cltr.io;
99	this.waiting = true;
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 alloc;
128	volatile uint32_t ready;
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	// Statistics
140	#if !defined(__CFA_NO_STATISTICS__)
141	struct {
142	struct {
143	unsigned long long int val;
144	unsigned long long int cnt;
145	} submit_avg;
146	} stats;
147	#endif
148	};
149
150	struct __completion_data {
151	// Head and tail of the ring
152	volatile uint32_t * head;
153	volatile uint32_t * tail;
154
155	// number of entries and mask to go with it
156	const uint32_t * mask;
157	const uint32_t * num;
158
159	// number of cqes not submitted (whatever that means)
160	uint32_t * overflow;
161
162	// the kernel ring
163	struct io_uring_cqe * cqes;
164
165	// The location and size of the mmaped area
166	void * ring_ptr;
167	size_t ring_sz;
168
169	// Statistics
170	#if !defined(__CFA_NO_STATISTICS__)
171	struct {
172	struct {
173	unsigned long long int val;
174	unsigned long long int slow_cnt;
175	unsigned long long int fast_cnt;
176	} completed_avg;
177	} stats;
178	#endif
179	};
180
181	struct __io_data {
182	struct __submition_data submit_q;
183	struct __completion_data completion_q;
184	uint32_t ring_flags;
185	int cltr_flags;
186	int fd;
187	semaphore submit;
188	volatile bool done;
189	struct {
190	struct {
191	void * stack;
192	pthread_t kthrd;
193	} slow;
194	__io_poller_fast fast;
195	__bin_sem_t sem;
196	} poller;
197	};
198
199	//=============================================================================================
200	// I/O Startup / Shutdown logic
201	//=============================================================================================
202	void __kernel_io_startup( cluster & this, int io_flags, bool main_cluster ) {
203	this.io = malloc();
204
205	// Step 1 : call to setup
206	struct io_uring_params params;
207	memset(&params, 0, sizeof(params));
208
209	uint32_t nentries = entries_per_cluster();
210
211	int fd = syscall(__NR_io_uring_setup, nentries, &params );
212	if(fd < 0) {
213	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
214	}
215
216	// Step 2 : mmap result
217	memset( this.io, 0, sizeof(struct __io_data) );
218	struct __submition_data & sq = this.io->submit_q;
219	struct __completion_data & cq = this.io->completion_q;
220
221	// calculate the right ring size
222	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
223	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
224
225	// Requires features
226	#if defined(IORING_FEAT_SINGLE_MMAP)
227	// adjust the size according to the parameters
228	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
229	cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz);
230	}
231	#endif
232
233	// mmap the Submit Queue into existence
234	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
235	if (sq.ring_ptr == (void*)MAP_FAILED) {
236	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
237	}
238
239	// Requires features
240	#if defined(IORING_FEAT_SINGLE_MMAP)
241	// mmap the Completion Queue into existence (may or may not be needed)
242	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
243	cq->ring_ptr = sq->ring_ptr;
244	}
245	else
246	#endif
247	{
248	// We need multiple call to MMAP
249	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
250	if (cq.ring_ptr == (void*)MAP_FAILED) {
251	munmap(sq.ring_ptr, sq.ring_sz);
252	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
253	}
254	}
255
256	// mmap the submit queue entries
257	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
258	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
259	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
260	munmap(sq.ring_ptr, sq.ring_sz);
261	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
262	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
263	}
264
265	// Get the pointers from the kernel to fill the structure
266	// submit queue
267	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
268	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
269	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
270	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
271	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
272	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
273	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
274	sq.alloc = *sq.tail;
275	sq.ready = *sq.tail;
276
277	// completion queue
278	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
279	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
280	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
281	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
282	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
283	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
284
285	// some paranoid checks
286	/* 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 );
287	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
288	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
289	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
290
291	/* 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 );
292	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
293	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
294	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
295
296	// Update the global ring info
297	this.io->ring_flags = params.flags;
298	this.io->cltr_flags = io_flags;
299	this.io->fd = fd;
300	this.io->done = false;
301	(this.io->submit){ min(sq.num, cq.num) };
302
303	// Initialize statistics
304	#if !defined(__CFA_NO_STATISTICS__)
305	this.io->submit_q.stats.submit_avg.val = 0;
306	this.io->submit_q.stats.submit_avg.cnt = 0;
307	this.io->completion_q.stats.completed_avg.val = 0;
308	this.io->completion_q.stats.completed_avg.slow_cnt = 0;
309	this.io->completion_q.stats.completed_avg.fast_cnt = 0;
310	#endif
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.stack = __create_pthread( &this.io->poller.slow.kthrd, __io_poller_slow, &this );
327	}
328
329	void __kernel_io_prepare_stop( cluster & this ) {
330	__cfadbg_print_safe(io_core, "Kernel I/O : Stopping pollers for cluster\n", &this);
331	// Notify the poller thread of the shutdown
332	__atomic_store_n(&this.io->done, true, __ATOMIC_SEQ_CST);
333
334	// Stop the IO Poller
335	sigval val = { 1 };
336	pthread_sigqueue( this.io->poller.slow.kthrd, SIGUSR1, val );
337	post( this.io->poller.sem );
338
339	// Wait for the poller thread to finish
340	pthread_join( this.io->poller.slow.kthrd, 0p );
341	free( this.io->poller.slow.stack );
342
343	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller stopped for cluster\n", &this);
344
345	if( this.io->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
346	verify( this.io->poller.fast.waiting );
347	verify( this.io->poller.fast.thrd.state == Blocked );
348
349	this.io->poller.fast.thrd.curr_cluster = mainCluster;
350
351	// unpark the fast io_poller
352	unpark( &this.io->poller.fast.thrd __cfaabi_dbg_ctx2 );
353
354	^(this.io->poller.fast){};
355
356	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller stopped for cluster\n", &this);
357	}
358	}
359
360	void __kernel_io_shutdown( cluster & this, bool main_cluster ) {
361	if(!main_cluster) {
362	__kernel_io_prepare_stop( this );
363	}
364
365	// print statistics
366	#if !defined(__CFA_NO_STATISTICS__)
367	if(this.print_stats) {
368	with(this.io->submit_q.stats, this.io->completion_q.stats) {
369	__cfaabi_bits_print_safe( STDERR_FILENO,
370	"----- I/O uRing Stats -----\n"
371	"- total submit calls : %llu\n"
372	"- avg submit : %lf\n"
373	"- total wait calls : %llu (%llu slow, %llu fast)\n"
374	"- avg completion/wait : %lf\n",
375	submit_avg.cnt,
376	((double)submit_avg.val) / submit_avg.cnt,
377	completed_avg.slow_cnt + completed_avg.fast_cnt,
378	completed_avg.slow_cnt, completed_avg.fast_cnt,
379	((double)completed_avg.val) / (completed_avg.slow_cnt + completed_avg.fast_cnt)
380	);
381	}
382	}
383	#endif
384
385	// Shutdown the io rings
386	struct __submition_data & sq = this.io->submit_q;
387	struct __completion_data & cq = this.io->completion_q;
388
389	// unmap the submit queue entries
390	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
391
392	// unmap the Submit Queue ring
393	munmap(sq.ring_ptr, sq.ring_sz);
394
395	// unmap the Completion Queue ring, if it is different
396	if (cq.ring_ptr != sq.ring_ptr) {
397	munmap(cq.ring_ptr, cq.ring_sz);
398	}
399
400	// close the file descriptor
401	close(this.io->fd);
402
403	free( this.io );
404	}
405
406	//=============================================================================================
407	// I/O Polling
408	//=============================================================================================
409	struct io_user_data {
410	int32_t result;
411	$thread * thrd;
412	};
413
414	// Process a single completion message from the io_uring
415	// This is NOT thread-safe
416	static int __drain_io( struct __io_data & ring, sigset_t * mask, int waitcnt, bool in_kernel ) {
417	int ret = syscall( __NR_io_uring_enter, ring.fd, 0, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8);
418	if( ret < 0 ) {
419	switch((int)errno) {
420	case EAGAIN:
421	case EINTR:
422	return -EAGAIN;
423	default:
424	abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
425	}
426	}
427
428	// Drain the queue
429	unsigned head = *ring.completion_q.head;
430	unsigned tail = __atomic_load_n(ring.completion_q.tail, __ATOMIC_ACQUIRE);
431
432	// Nothing was new return 0
433	if (head == tail) {
434	return 0;
435	}
436
437	uint32_t count = tail - head;
438	for(i; count) {
439	unsigned idx = (head + i) & (*ring.completion_q.mask);
440	struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
441
442	/* paranoid */ verify(&cqe);
443
444	struct io_user_data * data = (struct io_user_data *)cqe.user_data;
445	__cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
446
447	data->result = cqe.res;
448	if(!in_kernel) { unpark( data->thrd __cfaabi_dbg_ctx2 ); }
449	else { __unpark( data->thrd __cfaabi_dbg_ctx2 ); }
450	}
451
452	// Allow new submissions to happen
453	V(ring.submit, count);
454
455	// Mark to the kernel that the cqe has been seen
456	// Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
457	__atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED );
458
459	return count;
460	}
461
462	static void * __io_poller_slow( void * arg ) {
463	cluster * cltr = (cluster *)arg;
464	struct __io_data & ring = *cltr->io;
465
466	sigset_t mask;
467	sigfillset(&mask);
468	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
469	abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
470	}
471
472	sigdelset( &mask, SIGUSR1 );
473
474	verify( (ring.submit_q.head) == (ring.submit_q.tail) );
475	verify( (ring.completion_q.head) == (ring.completion_q.tail) );
476
477	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p ready\n", &ring);
478
479	if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD ) {
480	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
481	// In the user-thread approach drain and if anything was drained,
482	// batton pass to the user-thread
483	int count = __drain_io( ring, &mask, 1, true );
484
485	// Update statistics
486	#if !defined(__CFA_NO_STATISTICS__)
487	ring.completion_q.stats.completed_avg.val += count;
488	ring.completion_q.stats.completed_avg.slow_cnt += 1;
489	#endif
490
491	if(count > 0) {
492	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to fast poller\n", &ring);
493	__unpark( &ring.poller.fast.thrd __cfaabi_dbg_ctx2 );
494	wait( ring.poller.sem );
495	}
496	}
497	}
498	else {
499	while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
500	//In the naive approach, just poll the io completion queue directly
501	int count = __drain_io( ring, &mask, 1, true );
502
503	// Update statistics
504	#if !defined(__CFA_NO_STATISTICS__)
505	ring.completion_q.stats.completed_avg.val += count;
506	ring.completion_q.stats.completed_avg.slow_cnt += 1;
507	#endif
508	}
509	}
510
511	__cfadbg_print_safe(io_core, "Kernel I/O : Slow poller for ring %p stopping\n", &ring);
512
513	return 0p;
514	}
515
516	void main( __io_poller_fast & this ) {
517	verify( this.ring->cltr_flags & CFA_CLUSTER_IO_POLLER_USER_THREAD );
518
519	// Start parked
520	park( __cfaabi_dbg_ctx );
521
522	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p ready\n", &this.ring);
523
524	int reset = 0;
525
526	// Then loop until we need to start
527	while(!__atomic_load_n(&this.ring->done, __ATOMIC_SEQ_CST)) {
528	// Drain the io
529	this.waiting = false;
530	int count = __drain_io( *this.ring, 0p, 0, false );
531	reset += count > 0 ? 1 : 0;
532
533	// Update statistics
534	#if !defined(__CFA_NO_STATISTICS__)
535	this.ring->completion_q.stats.completed_avg.val += count;
536	this.ring->completion_q.stats.completed_avg.fast_cnt += 1;
537	#endif
538
539	this.waiting = true;
540	if(reset < 5) {
541	// If we got something, just yield and check again
542	yield();
543	}
544	else {
545	// We didn't get anything baton pass to the slow poller
546	__cfadbg_print_safe(io_core, "Kernel I/O : Moving to ring %p to slow poller\n", &this.ring);
547	post( this.ring->poller.sem );
548	park( __cfaabi_dbg_ctx );
549	reset = 0;
550	}
551	}
552
553	__cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring);
554	}
555
556	//=============================================================================================
557	// I/O Submissions
558	//=============================================================================================
559
560	// Submition steps :
561	// 1 - We need to make sure we don't overflow any of the buffer, P(ring.submit) to make sure
562	// entries are available. The semaphore make sure that there is no more operations in
563	// progress then the number of entries in the buffer. This probably limits concurrency
564	// more than necessary since submitted but not completed operations don't need any
565	// entries in user space. However, I don't know what happens if we overflow the buffers
566	// because too many requests completed at once. This is a safe approach in all cases.
567	// Furthermore, with hundreds of entries, this may be okay.
568	//
569	// 2 - Allocate a queue entry. The ring already has memory for all entries but only the ones
570	// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
571	// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
572	// need to write an allocator that allows allocating concurrently.
573	//
574	// 3 - Actually fill the submit entry, this is the only simple and straightforward step.
575	//
576	// 4 - Append the entry index to the array and adjust the tail accordingly. This operation
577	// needs to arrive to two concensus at the same time:
578	// A - The order in which entries are listed in the array: no two threads must pick the
579	// same index for their entries
580	// B - When can the tail be update for the kernel. EVERY entries in the array between
581	// head and tail must be fully filled and shouldn't ever be touched again.
582	//
583
584	static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring ) {
585	// Wait for a spot to be available
586	P(ring.submit);
587
588	// Allocate the sqe
589	uint32_t idx = __atomic_fetch_add(&ring.submit_q.alloc, 1ul32, __ATOMIC_SEQ_CST);
590
591	// Validate that we didn't overflow anything
592	// Check that nothing overflowed
593	/* paranoid */ verify( true );
594
595	// Check that it goes head -> tail -> alloc and never head -> alloc -> tail
596	/* paranoid */ verify( true );
597
598	// Return the sqe
599	return [&ring.submit_q.sqes[ idx & (*ring.submit_q.mask)], idx];
600	}
601
602	static inline void __submit( struct __io_data & ring, uint32_t idx ) {
603	// get mutual exclusion
604	lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
605
606	// Append to the list of ready entries
607	uint32_t * tail = ring.submit_q.tail;
608	const uint32_t mask = *ring.submit_q.mask;
609
610	ring.submit_q.array[ (*tail) & mask ] = idx & mask;
611	__atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
612
613	// Submit however, many entries need to be submitted
614	int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
615	if( ret < 0 ) {
616	switch((int)errno) {
617	default:
618	abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
619	}
620	}
621
622	// update statistics
623	#if !defined(__CFA_NO_STATISTICS__)
624	ring.submit_q.stats.submit_avg.val += 1;
625	ring.submit_q.stats.submit_avg.cnt += 1;
626	#endif
627
628	unlock(ring.submit_q.lock);
629	// Make sure that idx was submitted
630	// Be careful to not get false positive if we cycled the entire list or that someone else submitted for us
631	__cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret );
632	}
633
634	static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) {
635	this.opcode = opcode;
636	#if !defined(IOSQE_ASYNC)
637	this.flags = 0;
638	#else
639	this.flags = IOSQE_ASYNC;
640	#endif
641	this.ioprio = 0;
642	this.fd = fd;
643	this.off = 0;
644	this.addr = 0;
645	this.len = 0;
646	this.rw_flags = 0;
647	this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0;
648	}
649
650	static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) {
651	(this){ opcode, fd };
652	this.off = off;
653	this.addr = (uint64_t)addr;
654	this.len = len;
655	}
656
657
658	//=============================================================================================
659	// I/O Interface
660	//=============================================================================================
661
662	#define __submit_prelude \
663	struct __io_data & ring = *active_cluster()->io; \
664	struct io_uring_sqe * sqe; \
665	uint32_t idx; \
666	[sqe, idx] = __submit_alloc( ring );
667
668	#define __submit_wait \
669	io_user_data data = { 0, active_thread() }; \
670	/__cfaabi_bits_print_safe( STDERR_FILENO, "Preparing user data %p for %p\n", &data, data.thrd );/ \
671	sqe->user_data = (uint64_t)&data; \
672	__submit( ring, idx ); \
673	park( __cfaabi_dbg_ctx ); \
674	return data.result;
675	#endif
676
677	// Some forward declarations
678	extern "C" {
679	#include <unistd.h>
680	#include <sys/types.h>
681	#include <sys/socket.h>
682	#include <sys/syscall.h>
683	struct iovec;
684	extern ssize_t preadv2 (int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags);
685	extern ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags);
686
687	extern int fsync(int fd);
688	extern int sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags);
689
690	struct msghdr;
691	struct sockaddr;
692	extern ssize_t sendmsg(int sockfd, const struct msghdr *msg, int flags);
693	extern ssize_t recvmsg(int sockfd, struct msghdr *msg, int flags);
694	extern ssize_t send(int sockfd, const void *buf, size_t len, int flags);
695	extern ssize_t recv(int sockfd, void *buf, size_t len, int flags);
696	extern int accept4(int sockfd, struct sockaddr addr, socklen_t addrlen, int flags);
697	extern int connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen);
698
699	extern int fallocate(int fd, int mode, uint64_t offset, uint64_t len);
700	extern int posix_fadvise(int fd, uint64_t offset, uint64_t len, int advice);
701	extern int madvise(void *addr, size_t length, int advice);
702
703	extern int openat(int dirfd, const char *pathname, int flags, mode_t mode);
704	extern int close(int fd);
705
706	extern ssize_t read (int fd, void *buf, size_t count);
707	}
708
709	//-----------------------------------------------------------------------------
710	// Asynchronous operations
711	ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {
712	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_READV)
713	return preadv2(fd, iov, iovcnt, offset, flags);
714	#else
715	__submit_prelude
716
717	(*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset };
718
719	__submit_wait
720	#endif
721	}
722
723	ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {
724	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_WRITEV)
725	return pwritev2(fd, iov, iovcnt, offset, flags);
726	#else
727	__submit_prelude
728
729	(*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset };
730
731	__submit_wait
732	#endif
733	}
734
735	int cfa_fsync(int fd) {
736	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_FSYNC)
737	return fsync(fd);
738	#else
739	__submit_prelude
740
741	(*sqe){ IORING_OP_FSYNC, fd };
742
743	__submit_wait
744	#endif
745	}
746
747	int cfa_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) {
748	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_SYNC_FILE_RANGE)
749	return sync_file_range(fd, offset, nbytes, flags);
750	#else
751	__submit_prelude
752
753	(*sqe){ IORING_OP_SYNC_FILE_RANGE, fd };
754	sqe->off = offset;
755	sqe->len = nbytes;
756	sqe->sync_range_flags = flags;
757
758	__submit_wait
759	#endif
760	}
761
762
763	ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags) {
764	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_SENDMSG)
765	return sendmsg(sockfd, msg, flags);
766	#else
767	__submit_prelude
768
769	(*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 };
770	sqe->msg_flags = flags;
771
772	__submit_wait
773	#endif
774	}
775
776	ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags) {
777	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_RECVMSG)
778	return recvmsg(sockfd, msg, flags);
779	#else
780	__submit_prelude
781
782	(*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 };
783	sqe->msg_flags = flags;
784
785	__submit_wait
786	#endif
787	}
788
789	ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags) {
790	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_SEND)
791	return send( sockfd, buf, len, flags );
792	#else
793	__submit_prelude
794
795	(*sqe){ IORING_OP_SEND, sockfd };
796	sqe->addr = (uint64_t)buf;
797	sqe->len = len;
798	sqe->msg_flags = flags;
799
800	__submit_wait
801	#endif
802	}
803
804	ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags) {
805	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_RECV)
806	return recv( sockfd, buf, len, flags );
807	#else
808	__submit_prelude
809
810	(*sqe){ IORING_OP_RECV, sockfd };
811	sqe->addr = (uint64_t)buf;
812	sqe->len = len;
813	sqe->msg_flags = flags;
814
815	__submit_wait
816	#endif
817	}
818
819	int cfa_accept4(int sockfd, struct sockaddr addr, socklen_t addrlen, int flags) {
820	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_ACCEPT)
821	return accept4( sockfd, addr, addrlen, flags );
822	#else
823	__submit_prelude
824
825	(*sqe){ IORING_OP_ACCEPT, sockfd };
826	sqe->addr = addr;
827	sqe->addr2 = addrlen;
828	sqe->accept_flags = flags;
829
830	__submit_wait
831	#endif
832	}
833
834	int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) {
835	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_CONNECT)
836	return connect( sockfd, addr, addrlen );
837	#else
838	__submit_prelude
839
840	(*sqe){ IORING_OP_CONNECT, sockfd };
841	sqe->addr = (uint64_t)addr;
842	sqe->off = addrlen;
843
844	__submit_wait
845	#endif
846	}
847
848	int cfa_fallocate(int fd, int mode, uint64_t offset, uint64_t len) {
849	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_FALLOCATE)
850	return fallocate( fd, mode, offset, len );
851	#else
852	__submit_prelude
853
854	(*sqe){ IORING_OP_FALLOCATE, fd };
855	sqe->off = offset;
856	sqe->len = length;
857	sqe->mode = mode;
858
859	__submit_wait
860	#endif
861	}
862
863	int cfa_fadvise(int fd, uint64_t offset, uint64_t len, int advice) {
864	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_FADVISE)
865	return posix_fadvise( fd, offset, len, advice );
866	#else
867	__submit_prelude
868
869	(*sqe){ IORING_OP_FADVISE, fd };
870	sqe->off = (uint64_t)offset;
871	sqe->len = length;
872	sqe->fadvise_advice = advice;
873
874	__submit_wait
875	#endif
876	}
877
878	int cfa_madvise(void *addr, size_t length, int advice) {
879	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_MADVISE)
880	return madvise( addr, length, advice );
881	#else
882	__submit_prelude
883
884	(*sqe){ IORING_OP_MADVISE, 0 };
885	sqe->addr = (uint64_t)addr;
886	sqe->len = length;
887	sqe->fadvise_advice = advice;
888
889	__submit_wait
890	#endif
891	}
892
893	int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode) {
894	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_OPENAT)
895	return openat( dirfd, pathname, flags, mode );
896	#else
897	__submit_prelude
898
899	(*sqe){ IORING_OP_OPENAT, dirfd };
900	sqe->addr = (uint64_t)pathname;
901	sqe->open_flags = flags;
902	sqe->mode = mode;
903
904	__submit_wait
905	#endif
906	}
907
908	int cfa_close(int fd) {
909	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_CLOSE)
910	return close( fd );
911	#else
912	__submit_prelude
913
914	(*sqe){ IORING_OP_CLOSE, fd };
915
916	__submit_wait
917	#endif
918	}
919
920
921	ssize_t cfa_read(int fd, void *buf, size_t count) {
922	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_READ)
923	return read( fd, buf, count );
924	#else
925	__submit_prelude
926
927	(*sqe){ IORING_OP_READ, fd, buf, count, 0 };
928
929	__submit_wait
930	#endif
931	}
932
933	ssize_t cfa_write(int fd, void *buf, size_t count) {
934	#if !defined(HAVE_LINUX_IO_URING_H) \|\| !defined(IORING_OP_WRITE)
935	return read( fd, buf, count );
936	#else
937	__submit_prelude
938
939	(*sqe){ IORING_OP_WRITE, fd, buf, count, 0 };
940
941	__submit_wait
942	#endif
943	}
944
945	//-----------------------------------------------------------------------------
946	// Check if a function is asynchronous
947
948	// Macro magic to reduce the size of the following switch case
949	#define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__)
950	#define IS_DEFINED_SECOND(first, second, ...) second
951	#define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion
952	#define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true)
953
954	bool has_user_level_blocking( fptr_t func ) {
955	#if defined(HAVE_LINUX_IO_URING_H)
956	if( /func == (fptr_t)preadv2 \|\| /
957	func == (fptr_t)cfa_preadv2 )
958	#define _CFA_IO_FEATURE_IORING_OP_READV ,
959	return IS_DEFINED(IORING_OP_READV);
960
961	if( /func == (fptr_t)pwritev2 \|\| /
962	func == (fptr_t)cfa_pwritev2 )
963	#define _CFA_IO_FEATURE_IORING_OP_WRITEV ,
964	return IS_DEFINED(IORING_OP_WRITEV);
965
966	if( /func == (fptr_t)fsync \|\| /
967	func == (fptr_t)cfa_fsync )
968	#define _CFA_IO_FEATURE_IORING_OP_FSYNC ,
969	return IS_DEFINED(IORING_OP_FSYNC);
970
971	if( /func == (fptr_t)ync_file_range \|\| /
972	func == (fptr_t)cfa_sync_file_range )
973	#define _CFA_IO_FEATURE_IORING_OP_SYNC_FILE_RANGE ,
974	return IS_DEFINED(IORING_OP_SYNC_FILE_RANGE);
975
976	if( /func == (fptr_t)sendmsg \|\| /
977	func == (fptr_t)cfa_sendmsg )
978	#define _CFA_IO_FEATURE_IORING_OP_SENDMSG ,
979	return IS_DEFINED(IORING_OP_SENDMSG);
980
981	if( /func == (fptr_t)recvmsg \|\| /
982	func == (fptr_t)cfa_recvmsg )
983	#define _CFA_IO_FEATURE_IORING_OP_RECVMSG ,
984	return IS_DEFINED(IORING_OP_RECVMSG);
985
986	if( /func == (fptr_t)send \|\| /
987	func == (fptr_t)cfa_send )
988	#define _CFA_IO_FEATURE_IORING_OP_SEND ,
989	return IS_DEFINED(IORING_OP_SEND);
990
991	if( /func == (fptr_t)recv \|\| /
992	func == (fptr_t)cfa_recv )
993	#define _CFA_IO_FEATURE_IORING_OP_RECV ,
994	return IS_DEFINED(IORING_OP_RECV);
995
996	if( /func == (fptr_t)accept4 \|\| /
997	func == (fptr_t)cfa_accept4 )
998	#define _CFA_IO_FEATURE_IORING_OP_ACCEPT ,
999	return IS_DEFINED(IORING_OP_ACCEPT);
1000
1001	if( /func == (fptr_t)connect \|\| /
1002	func == (fptr_t)cfa_connect )
1003	#define _CFA_IO_FEATURE_IORING_OP_CONNECT ,
1004	return IS_DEFINED(IORING_OP_CONNECT);
1005
1006	if( /func == (fptr_t)fallocate \|\| /
1007	func == (fptr_t)cfa_fallocate )
1008	#define _CFA_IO_FEATURE_IORING_OP_FALLOCATE ,
1009	return IS_DEFINED(IORING_OP_FALLOCATE);
1010
1011	if( /func == (fptr_t)posix_fadvise \|\| /
1012	func == (fptr_t)cfa_fadvise )
1013	#define _CFA_IO_FEATURE_IORING_OP_FADVISE ,
1014	return IS_DEFINED(IORING_OP_FADVISE);
1015
1016	if( /func == (fptr_t)madvise \|\| /
1017	func == (fptr_t)cfa_madvise )
1018	#define _CFA_IO_FEATURE_IORING_OP_MADVISE ,
1019	return IS_DEFINED(IORING_OP_MADVISE);
1020
1021	if( /func == (fptr_t)openat \|\| /
1022	func == (fptr_t)cfa_openat )
1023	#define _CFA_IO_FEATURE_IORING_OP_OPENAT ,
1024	return IS_DEFINED(IORING_OP_OPENAT);
1025
1026	if( /func == (fptr_t)close \|\| /
1027	func == (fptr_t)cfa_close )
1028	#define _CFA_IO_FEATURE_IORING_OP_CLOSE ,
1029	return IS_DEFINED(IORING_OP_CLOSE);
1030
1031	if( /func == (fptr_t)read \|\| /
1032	func == (fptr_t)cfa_read )
1033	#define _CFA_IO_FEATURE_IORING_OP_READ ,
1034	return IS_DEFINED(IORING_OP_READ);
1035
1036	if( /func == (fptr_t)write \|\| /
1037	func == (fptr_t)cfa_write )
1038	#define _CFA_IO_FEATURE_IORING_OP_WRITE ,
1039	return IS_DEFINED(IORING_OP_WRITE);
1040	#endif
1041
1042	return false;
1043	}

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

Download in other formats:

Original Format