Context Navigation

setup.cfa @ 2cd784a

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

Last change on this file since 2cd784a was dddb3dd0, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Changed io to use ring per kernel threads.
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 /* See feature_test_macros(7) */
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	$io_arbiter * create(void) { return 0p; }
34	void destroy($io_arbiter *) {}
35
36	#else
37	#include <errno.h>
38	#include <stdint.h>
39	#include <string.h>
40	#include <signal.h>
41	#include <unistd.h>
42
43	extern "C" {
44	#include <pthread.h>
45	#include <sys/epoll.h>
46	#include <sys/eventfd.h>
47	#include <sys/mman.h>
48	#include <sys/syscall.h>
49
50	#include <linux/io_uring.h>
51	}
52
53	#include "bitmanip.hfa"
54	#include "kernel_private.hfa"
55	#include "thread.hfa"
56
57	void ?{}(io_context_params & this) {
58	this.num_entries = 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	//=============================================================================================
91	// I/O Context Constrution/Destruction
92	//=============================================================================================
93
94
95
96	static void __io_uring_setup ( $io_context & this, const io_context_params & params_in, int procfd );
97	static void __io_uring_teardown( $io_context & this );
98	static void __epoll_register($io_context & ctx);
99	static void __epoll_unregister($io_context & ctx);
100	void __ioarbiter_register( $io_arbiter & mutex, $io_context & ctx );
101	void __ioarbiter_unregister( $io_arbiter & mutex, $io_context & ctx );
102
103	void ?{}($io_context & this, processor * proc, struct cluster & cl) {
104	/* paranoid */ verify( cl.io.arbiter );
105	this.proc = proc;
106	this.arbiter = cl.io.arbiter;
107	this.ext_sq.empty = true;
108	(this.ext_sq.blocked){};
109	__io_uring_setup( this, cl.io.params, proc->idle );
110	__cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
111	}
112
113	void ^?{}($io_context & this) {
114	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
115
116	__io_uring_teardown( this );
117	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
118	}
119
120	extern void __disable_interrupts_hard();
121	extern void __enable_interrupts_hard();
122
123	static void __io_uring_setup( $io_context & this, const io_context_params & params_in, int procfd ) {
124	// Step 1 : call to setup
125	struct io_uring_params params;
126	memset(&params, 0, sizeof(params));
127	// if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
128	// if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
129
130	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
131	if( !is_pow2(nentries) ) {
132	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
133	}
134
135	int fd = syscall(__NR_io_uring_setup, nentries, &params );
136	if(fd < 0) {
137	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
138	}
139
140	// Step 2 : mmap result
141	struct __sub_ring_t & sq = this.sq;
142	struct __cmp_ring_t & cq = this.cq;
143
144	// calculate the right ring size
145	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
146	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
147
148	// Requires features
149	#if defined(IORING_FEAT_SINGLE_MMAP)
150	// adjust the size according to the parameters
151	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
152	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
153	}
154	#endif
155
156	// mmap the Submit Queue into existence
157	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
158	if (sq.ring_ptr == (void*)MAP_FAILED) {
159	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
160	}
161
162	// Requires features
163	#if defined(IORING_FEAT_SINGLE_MMAP)
164	// mmap the Completion Queue into existence (may or may not be needed)
165	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
166	cq.ring_ptr = sq.ring_ptr;
167	}
168	else
169	#endif
170	{
171	// We need multiple call to MMAP
172	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
173	if (cq.ring_ptr == (void*)MAP_FAILED) {
174	munmap(sq.ring_ptr, sq.ring_sz);
175	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
176	}
177	}
178
179	// mmap the submit queue entries
180	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
181	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
182	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
183	munmap(sq.ring_ptr, sq.ring_sz);
184	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
185	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
186	}
187
188	// Step 3 : Initialize the data structure
189	// Get the pointers from the kernel to fill the structure
190	// submit queue
191	sq.kring.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
192	sq.kring.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
193	sq.kring.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
194	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
195	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
196	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
197	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
198
199	sq.kring.released = 0;
200
201	sq.free_ring.head = 0;
202	sq.free_ring.tail = *sq.num;
203	sq.free_ring.array = alloc( *sq.num, 128`align );
204	for(i; (__u32)*sq.num) {
205	sq.free_ring.array[i] = i;
206	}
207
208	sq.to_submit = 0;
209
210	// completion queue
211	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
212	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
213	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
214	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
215	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
216	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
217
218	// Step 4 : eventfd
219	// io_uring_register is so f*cking slow on some machine that it
220	// will never succeed if preemption isn't hard blocked
221	__cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);
222
223	__disable_interrupts_hard();
224
225	int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &procfd, 1);
226	if (ret < 0) {
227	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
228	}
229
230	__enable_interrupts_hard();
231
232	__cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
233
234	// some paranoid checks
235	/* 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 );
236	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
237	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
238	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
239
240	/* 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 );
241	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
242	/* paranoid / verifyf( (sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
243	/* paranoid / verifyf( (sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
244
245	// Update the global ring info
246	this.ring_flags = 0;
247	this.fd = fd;
248	}
249
250	static void __io_uring_teardown( $io_context & this ) {
251	// Shutdown the io rings
252	struct __sub_ring_t & sq = this.sq;
253	struct __cmp_ring_t & cq = this.cq;
254
255	// unmap the submit queue entries
256	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
257
258	// unmap the Submit Queue ring
259	munmap(sq.ring_ptr, sq.ring_sz);
260
261	// unmap the Completion Queue ring, if it is different
262	if (cq.ring_ptr != sq.ring_ptr) {
263	munmap(cq.ring_ptr, cq.ring_sz);
264	}
265
266	// close the file descriptor
267	close(this.fd);
268
269	free( this.sq.free_ring.array ); // Maybe null, doesn't matter
270	}
271
272	void __cfa_io_start( processor * proc ) {
273	proc->io.ctx = alloc();
274	(proc->io.ctx){proc, proc->cltr};
275	}
276	void __cfa_io_stop ( processor * proc ) {
277	^(*proc->io.ctx){};
278	free(proc->io.ctx);
279	}
280
281	//=============================================================================================
282	// I/O Context Sleep
283	//=============================================================================================
284	// static inline void __epoll_ctl($io_context & ctx, int op, const char * error) {
285	// struct epoll_event ev;
286	// ev.events = EPOLLIN \| EPOLLONESHOT;
287	// ev.data.u64 = (__u64)&ctx;
288	// int ret = epoll_ctl(iopoll.epollfd, op, ctx.efd, &ev);
289	// if (ret < 0) {
290	// abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
291	// }
292	// }
293
294	// static void __epoll_register($io_context & ctx) {
295	// __epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
296	// }
297
298	// static void __epoll_unregister($io_context & ctx) {
299	// // Read the current epoch so we know when to stop
300	// size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
301
302	// // Remove the fd from the iopoller
303	// __epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
304
305	// // Notify the io poller thread of the shutdown
306	// iopoll.run = false;
307	// sigval val = { 1 };
308	// pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
309
310	// // Make sure all this is done
311	// __atomic_thread_fence(__ATOMIC_SEQ_CST);
312
313	// // Wait for the next epoch
314	// while(curr == iopoll.epoch && !iopoll.stopped) Pause();
315	// }
316
317	// void __ioctx_prepare_block($io_context & ctx) {
318	// __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.fd, &ctx);
319	// __epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
320	// }
321
322
323	//=============================================================================================
324	// I/O Context Misc Setup
325	//=============================================================================================
326	void ?{}( $io_arbiter & this ) {
327	this.pending.flag = false;
328	}
329
330	void ^?{}( $io_arbiter & mutex this ) {
331	// /* paranoid */ verify( empty(this.assigned) );
332	// /* paranoid */ verify( empty(this.available) );
333	/* paranoid */ verify( is_empty(this.pending.blocked) );
334	}
335
336	$io_arbiter * create(void) {
337	return new();
338	}
339	void destroy($io_arbiter * arbiter) {
340	delete(arbiter);
341	}
342
343	//=============================================================================================
344	// I/O Context Misc Setup
345	//=============================================================================================
346
347	#endif

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

Download in other formats:

Original Format