Context Navigation

setup.cfa @ d7e9c12

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since d7e9c12 was d3605f8, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Reworked io_uring idle sleep to work with either read or readv depending on what's available.
Property mode set to `100644`
File size: 11.9 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/setup.cfa --
8	//
9	// Author : Thierry Delisle
10	// Created On : Fri Jul 31 16:25:51 2020
11	// Last Modified By :
12	// Last Modified On :
13	// Update Count :
14	//
15
16	#define __cforall_thread__
17	#define _GNU_SOURCE
18
19	#if defined(__CFA_DEBUG__)
20	// #define __CFA_DEBUG_PRINT_IO__
21	// #define __CFA_DEBUG_PRINT_IO_CORE__
22	#endif
23
24	#include "io/types.hfa"
25	#include "kernel.hfa"
26
27	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
28	void ?{}(io_context_params & this) {}
29
30	void ?{}($io_context & this, struct cluster & cl) {}
31	void ^?{}($io_context & this) {}
32
33	void __cfa_io_start( processor * proc ) {}
34	bool __cfa_io_flush( processor * proc, int ) {}
35	void __cfa_io_stop ( processor * proc ) {}
36
37	$io_arbiter * create(void) { return 0p; }
38	void destroy($io_arbiter *) {}
39
40	#else
41	#include <errno.h>
42	#include <stdint.h>
43	#include <string.h>
44	#include <signal.h>
45	#include <unistd.h>
46
47	extern "C" {
48	#include <pthread.h>
49	#include <sys/epoll.h>
50	#include <sys/eventfd.h>
51	#include <sys/mman.h>
52	#include <sys/syscall.h>
53
54	#include <linux/io_uring.h>
55	}
56
57	#include "bitmanip.hfa"
58	#include "kernel_private.hfa"
59	#include "thread.hfa"
60
61	void ?{}(io_context_params & this) {
62	this.num_entries = 256;
63	}
64
65	static void * __io_poller_slow( void * arg );
66
67	// Weirdly, some systems that do support io_uring don't actually define these
68	#ifdef __alpha__
69	/*
70	* alpha is the only exception, all other architectures
71	* have common numbers for new system calls.
72	*/
73	#ifndef __NR_io_uring_setup
74	#define __NR_io_uring_setup 535
75	#endif
76	#ifndef __NR_io_uring_enter
77	#define __NR_io_uring_enter 536
78	#endif
79	#ifndef __NR_io_uring_register
80	#define __NR_io_uring_register 537
81	#endif
82	#else /* !__alpha__ */
83	#ifndef __NR_io_uring_setup
84	#define __NR_io_uring_setup 425
85	#endif
86	#ifndef __NR_io_uring_enter
87	#define __NR_io_uring_enter 426
88	#endif
89	#ifndef __NR_io_uring_register
90	#define __NR_io_uring_register 427
91	#endif
92	#endif
93
94	//=============================================================================================
95	// I/O Context Constrution/Destruction
96	//=============================================================================================
97
98
99
100	static void __io_uring_setup ( $io_context & this, const io_context_params & params_in, int procfd );
101	static void __io_uring_teardown( $io_context & this );
102	static void __epoll_register($io_context & ctx);
103	static void __epoll_unregister($io_context & ctx);
104	void __ioarbiter_register( $io_arbiter & mutex, $io_context & ctx );
105	void __ioarbiter_unregister( $io_arbiter & mutex, $io_context & ctx );
106
107	void ?{}($io_context & this, processor * proc, struct cluster & cl) {
108	/* paranoid */ verify( cl.io.arbiter );
109	this.proc = proc;
110	this.arbiter = cl.io.arbiter;
111	this.ext_sq.empty = true;
112	(this.ext_sq.queue){};
113	__io_uring_setup( this, cl.io.params, proc->idle_fd );
114	__cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
115	}
116
117	void ^?{}($io_context & this) {
118	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
119
120	__io_uring_teardown( this );
121	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
122	}
123
124	extern void __disable_interrupts_hard();
125	extern void __enable_interrupts_hard();
126
127	static void __io_uring_setup( $io_context & this, const io_context_params & params_in, int procfd ) {
128	// Step 1 : call to setup
129	struct io_uring_params params;
130	memset(&params, 0, sizeof(params));
131	// if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
132	// if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
133
134	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
135	if( !is_pow2(nentries) ) {
136	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
137	}
138
139	int fd = syscall(__NR_io_uring_setup, nentries, &params );
140	if(fd < 0) {
141	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
142	}
143
144	// Step 2 : mmap result
145	struct __sub_ring_t & sq = this.sq;
146	struct __cmp_ring_t & cq = this.cq;
147
148	// calculate the right ring size
149	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
150	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
151
152	// Requires features
153	#if defined(IORING_FEAT_SINGLE_MMAP)
154	// adjust the size according to the parameters
155	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
156	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
157	}
158	#endif
159
160	// mmap the Submit Queue into existence
161	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
162	if (sq.ring_ptr == (void*)MAP_FAILED) {
163	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
164	}
165
166	// Requires features
167	#if defined(IORING_FEAT_SINGLE_MMAP)
168	// mmap the Completion Queue into existence (may or may not be needed)
169	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
170	cq.ring_ptr = sq.ring_ptr;
171	}
172	else
173	#endif
174	{
175	// We need multiple call to MMAP
176	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
177	if (cq.ring_ptr == (void*)MAP_FAILED) {
178	munmap(sq.ring_ptr, sq.ring_sz);
179	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
180	}
181	}
182
183	// mmap the submit queue entries
184	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
185	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
186	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
187	munmap(sq.ring_ptr, sq.ring_sz);
188	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
189	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
190	}
191
192	// Step 3 : Initialize the data structure
193	// Get the pointers from the kernel to fill the structure
194	// submit queue
195	sq.kring.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
196	sq.kring.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
197	sq.kring.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
198	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
199	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
200	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
201	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
202
203	sq.kring.released = 0;
204
205	sq.free_ring.head = 0;
206	sq.free_ring.tail = *sq.num;
207	sq.free_ring.array = alloc( *sq.num, 128`align );
208	for(i; (__u32)*sq.num) {
209	sq.free_ring.array[i] = i;
210	}
211
212	sq.to_submit = 0;
213
214	// completion queue
215	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
216	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
217	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
218	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
219	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
220	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
221
222	#if !defined(CFA_WITH_IO_URING_IDLE)
223	// Step 4 : eventfd
224	// io_uring_register is so f*cking slow on some machine that it
225	// will never succeed if preemption isn't hard blocked
226	__cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);
227
228	__disable_interrupts_hard();
229
230	int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &procfd, 1);
231	if (ret < 0) {
232	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
233	}
234
235	__enable_interrupts_hard();
236
237	__cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
238	#endif
239
240	// some paranoid checks
241	/* 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 );
242	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
243	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
244	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
245
246	/* 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 );
247	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
248	/* paranoid / verifyf( (sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
249	/* paranoid / verifyf( (sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
250
251	// Update the global ring info
252	this.ring_flags = 0;
253	this.fd = fd;
254	}
255
256	static void __io_uring_teardown( $io_context & this ) {
257	// Shutdown the io rings
258	struct __sub_ring_t & sq = this.sq;
259	struct __cmp_ring_t & cq = this.cq;
260
261	// unmap the submit queue entries
262	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
263
264	// unmap the Submit Queue ring
265	munmap(sq.ring_ptr, sq.ring_sz);
266
267	// unmap the Completion Queue ring, if it is different
268	if (cq.ring_ptr != sq.ring_ptr) {
269	munmap(cq.ring_ptr, cq.ring_sz);
270	}
271
272	// close the file descriptor
273	close(this.fd);
274
275	free( this.sq.free_ring.array ); // Maybe null, doesn't matter
276	}
277
278	void __cfa_io_start( processor * proc ) {
279	proc->io.ctx = alloc();
280	(proc->io.ctx){proc, proc->cltr};
281	}
282	void __cfa_io_stop ( processor * proc ) {
283	^(*proc->io.ctx){};
284	free(proc->io.ctx);
285	}
286
287	//=============================================================================================
288	// I/O Context Sleep
289	//=============================================================================================
290	// static inline void __epoll_ctl($io_context & ctx, int op, const char * error) {
291	// struct epoll_event ev;
292	// ev.events = EPOLLIN \| EPOLLONESHOT;
293	// ev.data.u64 = (__u64)&ctx;
294	// int ret = epoll_ctl(iopoll.epollfd, op, ctx.efd, &ev);
295	// if (ret < 0) {
296	// abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
297	// }
298	// }
299
300	// static void __epoll_register($io_context & ctx) {
301	// __epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
302	// }
303
304	// static void __epoll_unregister($io_context & ctx) {
305	// // Read the current epoch so we know when to stop
306	// size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
307
308	// // Remove the fd from the iopoller
309	// __epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
310
311	// // Notify the io poller thread of the shutdown
312	// iopoll.run = false;
313	// sigval val = { 1 };
314	// pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
315
316	// // Make sure all this is done
317	// __atomic_thread_fence(__ATOMIC_SEQ_CST);
318
319	// // Wait for the next epoch
320	// while(curr == iopoll.epoch && !iopoll.stopped) Pause();
321	// }
322
323	// void __ioctx_prepare_block($io_context & ctx) {
324	// __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.fd, &ctx);
325	// __epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
326	// }
327
328
329	//=============================================================================================
330	// I/O Context Misc Setup
331	//=============================================================================================
332	void ?{}( $io_arbiter & this ) {
333	this.pending.empty = true;
334	}
335
336	void ^?{}( $io_arbiter & this ) {}
337
338	$io_arbiter * create(void) {
339	return new();
340	}
341	void destroy($io_arbiter * arbiter) {
342	delete(arbiter);
343	}
344
345	//=============================================================================================
346	// I/O Context Misc Setup
347	//=============================================================================================
348
349	#endif

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

Download in other formats:

Original Format