Context Navigation

setup.cfa @ 8f01ad71

ADTast-experimentalenumpthread-emulationqualifiedEnum

Last change on this file since 8f01ad71 was 4ecc35a, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Added spin lock to io drain. last step before completion fairness
Property mode set to `100644`
File size: 12.3 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 ) { return false; }
35	void __cfa_io_stop ( processor * proc ) {}
36
37	$io_arbiter * create(void) { return 0p; }
38	void destroy($io_arbiter *) {}
39
40	#else
41	#pragma GCC diagnostic push
42	#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
43	#include <errno.h>
44	#include <stdint.h>
45	#include <string.h>
46	#include <signal.h>
47	#include <unistd.h>
48
49	extern "C" {
50	#include <pthread.h>
51	#include <sys/epoll.h>
52	#include <sys/eventfd.h>
53	#include <sys/mman.h>
54	#include <sys/syscall.h>
55
56	#include <linux/io_uring.h>
57	}
58
59	#include "bitmanip.hfa"
60	#include "fstream.hfa"
61	#include "kernel/private.hfa"
62	#include "thread.hfa"
63	#pragma GCC diagnostic pop
64
65	void ?{}(io_context_params & this) {
66	this.num_entries = 256;
67	}
68
69	static void * __io_poller_slow( void * arg );
70
71	// Weirdly, some systems that do support io_uring don't actually define these
72	#ifdef __alpha__
73	/*
74	* alpha is the only exception, all other architectures
75	* have common numbers for new system calls.
76	*/
77	#ifndef __NR_io_uring_setup
78	#define __NR_io_uring_setup 535
79	#endif
80	#ifndef __NR_io_uring_enter
81	#define __NR_io_uring_enter 536
82	#endif
83	#ifndef __NR_io_uring_register
84	#define __NR_io_uring_register 537
85	#endif
86	#else /* !__alpha__ */
87	#ifndef __NR_io_uring_setup
88	#define __NR_io_uring_setup 425
89	#endif
90	#ifndef __NR_io_uring_enter
91	#define __NR_io_uring_enter 426
92	#endif
93	#ifndef __NR_io_uring_register
94	#define __NR_io_uring_register 427
95	#endif
96	#endif
97
98	//=============================================================================================
99	// I/O Context Constrution/Destruction
100	//=============================================================================================
101
102
103
104	static void __io_uring_setup ( $io_context & this, const io_context_params & params_in, int procfd );
105	static void __io_uring_teardown( $io_context & this );
106	static void __epoll_register($io_context & ctx);
107	static void __epoll_unregister($io_context & ctx);
108	void __ioarbiter_register( $io_arbiter & mutex, $io_context & ctx );
109	void __ioarbiter_unregister( $io_arbiter & mutex, $io_context & ctx );
110
111	void ?{}($io_context & this, processor * proc, struct cluster & cl) {
112	/* paranoid */ verify( cl.io.arbiter );
113	this.proc = proc;
114	this.arbiter = cl.io.arbiter;
115	this.ext_sq.empty = true;
116	(this.ext_sq.queue){};
117	__io_uring_setup( this, cl.io.params, proc->idle_wctx.evfd );
118	__cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
119	}
120
121	void ^?{}($io_context & this) {
122	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
123
124	__io_uring_teardown( this );
125	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
126	}
127
128	static void __io_uring_setup( $io_context & this, const io_context_params & params_in, int procfd ) {
129	// Step 1 : call to setup
130	struct io_uring_params params;
131	memset(&params, 0, sizeof(params));
132	// if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
133	// if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
134
135	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
136	if( !is_pow2(nentries) ) {
137	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
138	}
139
140	int fd = syscall(__NR_io_uring_setup, nentries, &params );
141	if(fd < 0) {
142	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
143	}
144
145	// Step 2 : mmap result
146	struct __sub_ring_t & sq = this.sq;
147	struct __cmp_ring_t & cq = this.cq;
148
149	// calculate the right ring size
150	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
151	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
152
153	// Requires features
154	#if defined(IORING_FEAT_SINGLE_MMAP)
155	// adjust the size according to the parameters
156	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
157	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
158	}
159	#endif
160
161	// mmap the Submit Queue into existence
162	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
163	if (sq.ring_ptr == (void*)MAP_FAILED) {
164	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
165	}
166
167	// Requires features
168	#if defined(IORING_FEAT_SINGLE_MMAP)
169	// mmap the Completion Queue into existence (may or may not be needed)
170	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
171	cq.ring_ptr = sq.ring_ptr;
172	}
173	else
174	#endif
175	{
176	// We need multiple call to MMAP
177	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
178	if (cq.ring_ptr == (void*)MAP_FAILED) {
179	munmap(sq.ring_ptr, sq.ring_sz);
180	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
181	}
182	}
183
184	// mmap the submit queue entries
185	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
186	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
187	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
188	munmap(sq.ring_ptr, sq.ring_sz);
189	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
190	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
191	}
192
193	// Step 3 : Initialize the data structure
194	// Get the pointers from the kernel to fill the structure
195	// submit queue
196	sq.kring.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
197	sq.kring.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
198	sq.kring.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
199	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
200	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
201	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
202	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
203
204	sq.kring.released = 0;
205
206	sq.free_ring.head = 0;
207	sq.free_ring.tail = *sq.num;
208	sq.free_ring.array = alloc( *sq.num, 128`align );
209	for(i; (__u32)*sq.num) {
210	sq.free_ring.array[i] = i;
211	}
212
213	sq.to_submit = 0;
214
215	// completion queue
216	cq.lock = 0;
217	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
218	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
219	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
220	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
221	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
222	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
223
224	#if !defined(CFA_WITH_IO_URING_IDLE)
225	// Step 4 : eventfd
226	// io_uring_register is so f*cking slow on some machine that it
227	// will never succeed if preemption isn't hard blocked
228	__cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);
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	__cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
236	#endif
237
238	// some paranoid checks
239	/* 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 );
240	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
241	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
242	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
243
244	/* 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 );
245	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
246	/* paranoid / verifyf( (sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
247	/* paranoid / verifyf( (sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
248
249	// Update the global ring info
250	this.ring_flags = 0;
251	this.fd = fd;
252	}
253
254	static void __io_uring_teardown( $io_context & this ) {
255	// Shutdown the io rings
256	struct __sub_ring_t & sq = this.sq;
257	struct __cmp_ring_t & cq = this.cq;
258	{
259	__u32 fhead = sq.free_ring.head;
260	__u32 ftail = sq.free_ring.tail;
261
262	__u32 total = *sq.num;
263	__u32 avail = ftail - fhead;
264
265	if(avail != total) abort \| "Processor (" \| (void*)this.proc \| ") tearing down ring with" \| (total - avail) \| "entries allocated but not submitted, out of" \| total;
266	}
267
268	// unmap the submit queue entries
269	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
270
271	// unmap the Submit Queue ring
272	munmap(sq.ring_ptr, sq.ring_sz);
273
274	// unmap the Completion Queue ring, if it is different
275	if (cq.ring_ptr != sq.ring_ptr) {
276	munmap(cq.ring_ptr, cq.ring_sz);
277	}
278
279	// close the file descriptor
280	close(this.fd);
281
282	free( this.sq.free_ring.array ); // Maybe null, doesn't matter
283	}
284
285	void __cfa_io_start( processor * proc ) {
286	proc->io.ctx = alloc();
287	(proc->io.ctx){proc, proc->cltr};
288	}
289	void __cfa_io_stop ( processor * proc ) {
290	^(*proc->io.ctx){};
291	free(proc->io.ctx);
292	}
293
294	//=============================================================================================
295	// I/O Context Sleep
296	//=============================================================================================
297	// static inline void __epoll_ctl($io_context & ctx, int op, const char * error) {
298	// struct epoll_event ev;
299	// ev.events = EPOLLIN \| EPOLLONESHOT;
300	// ev.data.u64 = (__u64)&ctx;
301	// int ret = epoll_ctl(iopoll.epollfd, op, ctx.efd, &ev);
302	// if (ret < 0) {
303	// abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
304	// }
305	// }
306
307	// static void __epoll_register($io_context & ctx) {
308	// __epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
309	// }
310
311	// static void __epoll_unregister($io_context & ctx) {
312	// // Read the current epoch so we know when to stop
313	// size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
314
315	// // Remove the fd from the iopoller
316	// __epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
317
318	// // Notify the io poller thread of the shutdown
319	// iopoll.run = false;
320	// sigval val = { 1 };
321	// pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
322
323	// // Make sure all this is done
324	// __atomic_thread_fence(__ATOMIC_SEQ_CST);
325
326	// // Wait for the next epoch
327	// while(curr == iopoll.epoch && !iopoll.stopped) Pause();
328	// }
329
330	// void __ioctx_prepare_block($io_context & ctx) {
331	// __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.fd, &ctx);
332	// __epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
333	// }
334
335
336	//=============================================================================================
337	// I/O Context Misc Setup
338	//=============================================================================================
339	void ?{}( $io_arbiter & this ) {
340	this.pending.empty = true;
341	}
342
343	void ^?{}( $io_arbiter & this ) {}
344
345	$io_arbiter * create(void) {
346	return new();
347	}
348	void destroy($io_arbiter * arbiter) {
349	delete(arbiter);
350	}
351
352	//=============================================================================================
353	// I/O Context Misc Setup
354	//=============================================================================================
355
356	#endif

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

Download in other formats:

Original Format