source: libcfa/src/concurrency/io/setup.cfa@ 3caf5e3

ADT ast-experimental enum pthread-emulation qualifiedEnum
Last change on this file since 3caf5e3 was 78a580d, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

I/O now updates the timestamps when draining.
Timestamps are not used yet.
  • 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 "limits.hfa"
63 #include "thread.hfa"
64#pragma GCC diagnostic pop
65
66 void ?{}(io_context_params & this) {
67 this.num_entries = 256;
68 }
69
70 static void * __io_poller_slow( void * arg );
71
72 // Weirdly, some systems that do support io_uring don't actually define these
73 #ifdef __alpha__
74 /*
75 * alpha is the only exception, all other architectures
76 * have common numbers for new system calls.
77 */
78 #ifndef __NR_io_uring_setup
79 #define __NR_io_uring_setup 535
80 #endif
81 #ifndef __NR_io_uring_enter
82 #define __NR_io_uring_enter 536
83 #endif
84 #ifndef __NR_io_uring_register
85 #define __NR_io_uring_register 537
86 #endif
87 #else /* !__alpha__ */
88 #ifndef __NR_io_uring_setup
89 #define __NR_io_uring_setup 425
90 #endif
91 #ifndef __NR_io_uring_enter
92 #define __NR_io_uring_enter 426
93 #endif
94 #ifndef __NR_io_uring_register
95 #define __NR_io_uring_register 427
96 #endif
97 #endif
98
99//=============================================================================================
100// I/O Context Constrution/Destruction
101//=============================================================================================
102
103
104
105 static void __io_uring_setup ( $io_context & this, const io_context_params & params_in, int procfd );
106 static void __io_uring_teardown( $io_context & this );
107 static void __epoll_register($io_context & ctx);
108 static void __epoll_unregister($io_context & ctx);
109 void __ioarbiter_register( $io_arbiter & mutex, $io_context & ctx );
110 void __ioarbiter_unregister( $io_arbiter & mutex, $io_context & ctx );
111
112 void ?{}($io_context & this, processor * proc, struct cluster & cl) {
113 /* paranoid */ verify( cl.io.arbiter );
114 this.proc = proc;
115 this.arbiter = cl.io.arbiter;
116 this.ext_sq.empty = true;
117 (this.ext_sq.queue){};
118 __io_uring_setup( this, cl.io.params, proc->idle_wctx.evfd );
119 __cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
120 }
121
122 void ^?{}($io_context & this) {
123 __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
124
125 __io_uring_teardown( this );
126 __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
127 }
128
129 static void __io_uring_setup( $io_context & this, const io_context_params & params_in, int procfd ) {
130 // Step 1 : call to setup
131 struct io_uring_params params;
132 memset(&params, 0, sizeof(params));
133 // if( params_in.poll_submit ) params.flags |= IORING_SETUP_SQPOLL;
134 // if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL;
135
136 __u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
137 if( !is_pow2(nentries) ) {
138 abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
139 }
140
141 int fd = syscall(__NR_io_uring_setup, nentries, &params );
142 if(fd < 0) {
143 abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
144 }
145
146 // Step 2 : mmap result
147 struct __sub_ring_t & sq = this.sq;
148 struct __cmp_ring_t & cq = this.cq;
149
150 // calculate the right ring size
151 sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
152 cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
153
154 // Requires features
155 #if defined(IORING_FEAT_SINGLE_MMAP)
156 // adjust the size according to the parameters
157 if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
158 cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
159 }
160 #endif
161
162 // mmap the Submit Queue into existence
163 sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
164 if (sq.ring_ptr == (void*)MAP_FAILED) {
165 abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
166 }
167
168 // Requires features
169 #if defined(IORING_FEAT_SINGLE_MMAP)
170 // mmap the Completion Queue into existence (may or may not be needed)
171 if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
172 cq.ring_ptr = sq.ring_ptr;
173 }
174 else
175 #endif
176 {
177 // We need multiple call to MMAP
178 cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
179 if (cq.ring_ptr == (void*)MAP_FAILED) {
180 munmap(sq.ring_ptr, sq.ring_sz);
181 abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
182 }
183 }
184
185 // mmap the submit queue entries
186 size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
187 sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
188 if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
189 munmap(sq.ring_ptr, sq.ring_sz);
190 if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
191 abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
192 }
193
194 // Step 3 : Initialize the data structure
195 // Get the pointers from the kernel to fill the structure
196 // submit queue
197 sq.kring.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
198 sq.kring.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
199 sq.kring.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
200 sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
201 sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
202 sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
203 sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
204
205 sq.kring.released = 0;
206
207 sq.free_ring.head = 0;
208 sq.free_ring.tail = *sq.num;
209 sq.free_ring.array = alloc( *sq.num, 128`align );
210 for(i; (__u32)*sq.num) {
211 sq.free_ring.array[i] = i;
212 }
213
214 sq.to_submit = 0;
215
216 // completion queue
217 cq.lock = 0;
218 cq.id = MAX;
219 cq.ts = rdtscl();
220 cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
221 cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
222 cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
223 cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
224 cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
225 cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
226
227 #if !defined(CFA_WITH_IO_URING_IDLE)
228 // Step 4 : eventfd
229 // io_uring_register is so f*cking slow on some machine that it
230 // will never succeed if preemption isn't hard blocked
231 __cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);
232
233 int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &procfd, 1);
234 if (ret < 0) {
235 abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
236 }
237
238 __cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
239 #endif
240
241 // some paranoid checks
242 /* 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 );
243 /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
244 /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
245 /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
246
247 /* 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 );
248 /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
249 /* paranoid */ verifyf( (*sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
250 /* paranoid */ verifyf( (*sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
251
252 // Update the global ring info
253 this.ring_flags = 0;
254 this.fd = fd;
255 }
256
257 static void __io_uring_teardown( $io_context & this ) {
258 // Shutdown the io rings
259 struct __sub_ring_t & sq = this.sq;
260 struct __cmp_ring_t & cq = this.cq;
261 {
262 __u32 fhead = sq.free_ring.head;
263 __u32 ftail = sq.free_ring.tail;
264
265 __u32 total = *sq.num;
266 __u32 avail = ftail - fhead;
267
268 if(avail != total) abort | "Processor (" | (void*)this.proc | ") tearing down ring with" | (total - avail) | "entries allocated but not submitted, out of" | total;
269 }
270
271 // unmap the submit queue entries
272 munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));
273
274 // unmap the Submit Queue ring
275 munmap(sq.ring_ptr, sq.ring_sz);
276
277 // unmap the Completion Queue ring, if it is different
278 if (cq.ring_ptr != sq.ring_ptr) {
279 munmap(cq.ring_ptr, cq.ring_sz);
280 }
281
282 // close the file descriptor
283 close(this.fd);
284
285 free( this.sq.free_ring.array ); // Maybe null, doesn't matter
286 }
287
288 void __cfa_io_start( processor * proc ) {
289 proc->io.ctx = alloc();
290 (*proc->io.ctx){proc, *proc->cltr};
291 }
292 void __cfa_io_stop ( processor * proc ) {
293 ^(*proc->io.ctx){};
294 free(proc->io.ctx);
295 }
296
297//=============================================================================================
298// I/O Context Sleep
299//=============================================================================================
300 // static inline void __epoll_ctl($io_context & ctx, int op, const char * error) {
301 // struct epoll_event ev;
302 // ev.events = EPOLLIN | EPOLLONESHOT;
303 // ev.data.u64 = (__u64)&ctx;
304 // int ret = epoll_ctl(iopoll.epollfd, op, ctx.efd, &ev);
305 // if (ret < 0) {
306 // abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
307 // }
308 // }
309
310 // static void __epoll_register($io_context & ctx) {
311 // __epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
312 // }
313
314 // static void __epoll_unregister($io_context & ctx) {
315 // // Read the current epoch so we know when to stop
316 // size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
317
318 // // Remove the fd from the iopoller
319 // __epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
320
321 // // Notify the io poller thread of the shutdown
322 // iopoll.run = false;
323 // sigval val = { 1 };
324 // pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
325
326 // // Make sure all this is done
327 // __atomic_thread_fence(__ATOMIC_SEQ_CST);
328
329 // // Wait for the next epoch
330 // while(curr == iopoll.epoch && !iopoll.stopped) Pause();
331 // }
332
333 // void __ioctx_prepare_block($io_context & ctx) {
334 // __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.fd, &ctx);
335 // __epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
336 // }
337
338
339//=============================================================================================
340// I/O Context Misc Setup
341//=============================================================================================
342 void ?{}( $io_arbiter & this ) {
343 this.pending.empty = true;
344 }
345
346 void ^?{}( $io_arbiter & this ) {}
347
348 $io_arbiter * create(void) {
349 return new();
350 }
351 void destroy($io_arbiter * arbiter) {
352 delete(arbiter);
353 }
354
355//=============================================================================================
356// I/O Context Misc Setup
357//=============================================================================================
358
359#endif
Note: See TracBrowser for help on using the repository browser.