source: libcfa/src/concurrency/io/setup.cfa@ 339e30a

ADT ast-experimental
Last change on this file since 339e30a was 26544f9, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

added helping and lock to allow remote processors to flush unresponsive procs
  • Property mode set to 100644
File size: 11.8 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) libcfa_public {}
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 ) { return false; }
35 bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1))) { return false; }
36 void __cfa_io_stop ( processor * proc ) {}
37
38 io_arbiter$ * create(void) { return 0p; }
39 void destroy(io_arbiter$ *) {}
40
41#else
42#pragma GCC diagnostic push
43#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
44 #include <errno.h>
45 #include <stdint.h>
46 #include <string.h>
47 #include <signal.h>
48 #include <unistd.h>
49
50 extern "C" {
51 #include <pthread.h>
52 #include <sys/epoll.h>
53 #include <sys/eventfd.h>
54 #include <sys/mman.h>
55 #include <sys/syscall.h>
56
57 #include <linux/io_uring.h>
58 }
59
60 #include "bitmanip.hfa"
61 #include "fstream.hfa"
62 #include "kernel/private.hfa"
63 #include "limits.hfa"
64 #include "thread.hfa"
65#pragma GCC diagnostic pop
66
67 void ?{}(io_context_params & this) libcfa_public {
68 this.num_entries = 256;
69 }
70
71 static void * __io_poller_slow( void * arg );
72
73 // Weirdly, some systems that do support io_uring don't actually define these
74 #ifdef __alpha__
75 /*
76 * alpha is the only exception, all other architectures
77 * have common numbers for new system calls.
78 */
79 #ifndef __NR_io_uring_setup
80 #define __NR_io_uring_setup 535
81 #endif
82 #ifndef __NR_io_uring_enter
83 #define __NR_io_uring_enter 536
84 #endif
85 #ifndef __NR_io_uring_register
86 #define __NR_io_uring_register 537
87 #endif
88 #else /* !__alpha__ */
89 #ifndef __NR_io_uring_setup
90 #define __NR_io_uring_setup 425
91 #endif
92 #ifndef __NR_io_uring_enter
93 #define __NR_io_uring_enter 426
94 #endif
95 #ifndef __NR_io_uring_register
96 #define __NR_io_uring_register 427
97 #endif
98 #endif
99
100//=============================================================================================
101// I/O Context Constrution/Destruction
102//=============================================================================================
103
104
105
106 static void __io_uring_setup ( io_context$ & this, const io_context_params & params_in, int procfd );
107 static void __io_uring_teardown( io_context$ & this );
108 static void __epoll_register(io_context$ & ctx);
109 static void __epoll_unregister(io_context$ & ctx);
110 void __ioarbiter_register( io_arbiter$ & mutex, io_context$ & ctx );
111 void __ioarbiter_unregister( io_arbiter$ & mutex, io_context$ & ctx );
112
113 void ?{}(io_context$ & this, processor * proc, struct cluster & cl) {
114 /* paranoid */ verify( cl.io.arbiter );
115 this.proc = proc;
116 this.arbiter = cl.io.arbiter;
117 this.ext_sq.empty = true;
118 (this.ext_sq.queue){};
119 __io_uring_setup( this, cl.io.params, proc->idle_wctx.evfd );
120 __cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
121 }
122
123 void ^?{}(io_context$ & this) {
124 __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
125
126 __io_uring_teardown( this );
127 __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
128 }
129
130 static void __io_uring_setup( io_context$ & this, const io_context_params & params_in, int procfd ) {
131 // Step 1 : call to setup
132 struct io_uring_params params;
133 memset(&params, 0, sizeof(params));
134 // if( params_in.poll_submit ) params.flags |= IORING_SETUP_SQPOLL;
135 // if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL;
136
137 __u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
138 if( !is_pow2(nentries) ) {
139 abort("ERROR: I/O setup 'num_entries' must be a power of 2, was %u\n", nentries);
140 }
141
142 int fd = syscall(__NR_io_uring_setup, nentries, &params );
143 if(fd < 0) {
144 abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
145 }
146
147 // Step 2 : mmap result
148 struct __sub_ring_t & sq = this.sq;
149 struct __cmp_ring_t & cq = this.cq;
150
151 // calculate the right ring size
152 sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
153 cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
154
155 // Requires features
156 #if defined(IORING_FEAT_SINGLE_MMAP)
157 // adjust the size according to the parameters
158 if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
159 cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
160 }
161 #endif
162
163 // mmap the Submit Queue into existence
164 sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
165 if (sq.ring_ptr == (void*)MAP_FAILED) {
166 abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
167 }
168
169 // Requires features
170 #if defined(IORING_FEAT_SINGLE_MMAP)
171 // mmap the Completion Queue into existence (may or may not be needed)
172 if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
173 cq.ring_ptr = sq.ring_ptr;
174 }
175 else
176 #endif
177 {
178 // We need multiple call to MMAP
179 cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
180 if (cq.ring_ptr == (void*)MAP_FAILED) {
181 munmap(sq.ring_ptr, sq.ring_sz);
182 abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
183 }
184 }
185
186 // mmap the submit queue entries
187 size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
188 sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
189 if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
190 munmap(sq.ring_ptr, sq.ring_sz);
191 if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
192 abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
193 }
194
195 // Step 3 : Initialize the data structure
196 // Get the pointers from the kernel to fill the structure
197 // submit queue
198 sq.kring.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
199 sq.kring.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
200 sq.kring.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
201 sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
202 sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
203 sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
204 sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
205
206 sq.kring.released = 0;
207
208 sq.free_ring.head = 0;
209 sq.free_ring.tail = *sq.num;
210 sq.free_ring.array = alloc( *sq.num, 128`align );
211 for(i; (__u32)*sq.num) {
212 sq.free_ring.array[i] = i;
213 }
214
215 sq.to_submit = 0;
216
217 // completion queue
218 cq.try_lock = false;
219 cq.id = MAX;
220 cq.ts = rdtscl();
221 cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
222 cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
223 cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
224 cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
225 cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
226 cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
227
228 #if !defined(CFA_WITH_IO_URING_IDLE)
229 {
230 // Step 4 : eventfd
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 }
240 #endif
241
242 // TODO: implement a proper version of this.
243 // I have not found a better maximum that works in general but users should be able to configure it
244 // the same way they configure other I/O options
245 // #if defined(CFA_HAVE_IORING_REGISTER_IOWQ_MAX_WORKERS)
246 // {
247 // // Step 5 : max worker count
248 // __cfadbg_print_safe(io_core, "Kernel I/O : lmiting max workers for ring %d\n", fd);
249
250 // unsigned int maxes[2];
251 // maxes[0] = 64; // max number of bounded workers (Regular files / block)
252 // maxes[1] = 64; // max number of unbounded workers (IOSQE_ASYNC)
253 // int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_IOWQ_MAX_WORKERS, maxes, 2);
254 // if (ret < 0) {
255 // abort("KERNEL ERROR: IO_URING MAX WORKER REGISTER - %s\n", strerror(errno));
256 // }
257
258 // __cfadbg_print_safe(io_core, "Kernel I/O : lmited max workers for ring %d\n", fd);
259 // }
260 // #endif
261
262 // some paranoid checks
263 /* 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 );
264 /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
265 /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
266 /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
267
268 /* 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 );
269 /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
270 /* paranoid */ verifyf( (*sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
271 /* paranoid */ verifyf( (*sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
272
273 // Update the global ring info
274 this.ring_flags = 0;
275 this.fd = fd;
276 }
277
278 static void __io_uring_teardown( io_context$ & this ) {
279 // Shutdown the io rings
280 struct __sub_ring_t & sq = this.sq;
281 struct __cmp_ring_t & cq = this.cq;
282 {
283 __u32 fhead = sq.free_ring.head;
284 __u32 ftail = sq.free_ring.tail;
285
286 __u32 total = *sq.num;
287 __u32 avail = ftail - fhead;
288
289 if(avail != total) abort | "Processor (" | (void*)this.proc | ") tearing down ring with" | (total - avail) | "entries allocated but not submitted, out of" | total;
290 }
291
292 // unmap the submit queue entries
293 munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));
294
295 // unmap the Submit Queue ring
296 munmap(sq.ring_ptr, sq.ring_sz);
297
298 // unmap the Completion Queue ring, if it is different
299 if (cq.ring_ptr != sq.ring_ptr) {
300 munmap(cq.ring_ptr, cq.ring_sz);
301 }
302
303 // close the file descriptor
304 close(this.fd);
305
306 free( this.sq.free_ring.array ); // Maybe null, doesn't matter
307 }
308
309 void __cfa_io_start( processor * proc ) {
310 proc->io.ctx = alloc();
311 (*proc->io.ctx){proc, *proc->cltr};
312 }
313 void __cfa_io_stop ( processor * proc ) {
314 ^(*proc->io.ctx){};
315 free(proc->io.ctx);
316 }
317
318
319//=============================================================================================
320// I/O Context Misc Setup
321//=============================================================================================
322 void ?{}( io_arbiter$ & this ) {
323 this.pending.empty = true;
324 }
325
326 void ^?{}( io_arbiter$ & mutex this ) {}
327
328 io_arbiter$ * create(void) {
329 return new();
330 }
331 void destroy(io_arbiter$ * arbiter) {
332 delete(arbiter);
333 }
334
335//=============================================================================================
336// I/O Context Misc Setup
337//=============================================================================================
338
339#endif
Note: See TracBrowser for help on using the repository browser.