ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
Revision: 1.9
Committed: Sat Jun 22 22:29:38 2019 UTC (4 years, 10 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.8: +4 -4 lines
Log Message:
*** empty log message ***

File Contents

# Content
1 /*
2 * libev linux aio fd activity backend
3 *
4 * Copyright (c) 2019 Marc Alexander Lehmann <libev@schmorp.de>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without modifica-
8 * tion, are permitted provided that the following conditions are met:
9 *
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
12 *
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
19 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
20 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
21 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
23 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
24 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
25 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
26 * OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Alternatively, the contents of this file may be used under the terms of
29 * the GNU General Public License ("GPL") version 2 or any later version,
30 * in which case the provisions of the GPL are applicable instead of
31 * the above. If you wish to allow the use of your version of this file
32 * only under the terms of the GPL and not to allow others to use your
33 * version of this file under the BSD license, indicate your decision
34 * by deleting the provisions above and replace them with the notice
35 * and other provisions required by the GPL. If you do not delete the
36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL.
38 */
39
40 #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
41 #include <poll.h>
42 #include <linux/aio_abi.h>
43
44 /* we try to fill 4kB pages exactly.
45 * the ring buffer header is 32 bytes, every io event is 32 bytes.
46 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
47 * therefore the calculation below will use "exactly" 4kB for the ring buffer
48 */
49 #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
50
51 /*****************************************************************************/
52 /* syscall wrapdadoop */
53
54 #include <sys/syscall.h> /* no glibc wrappers */
55
56 /* aio_abi.h is not versioned in any way, so we cannot test for its existance */
57 #define IOCB_CMD_POLL 5
58
59 /* taken from linux/fs/aio.c */
60 #define AIO_RING_MAGIC 0xa10a10a1
61 #define AIO_RING_INCOMPAT_FEATURES 0
62 struct aio_ring
63 {
64 unsigned id; /* kernel internal index number */
65 unsigned nr; /* number of io_events */
66 unsigned head; /* Written to by userland or by kernel. */
67 unsigned tail;
68
69 unsigned magic;
70 unsigned compat_features;
71 unsigned incompat_features;
72 unsigned header_length; /* size of aio_ring */
73
74 struct io_event io_events[0];
75 };
76
77 inline_size
78 int
79 ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
80 {
81 return syscall (SYS_io_setup, nr_events, ctx_idp);
82 }
83
84 inline_size
85 int
86 ev_io_destroy (aio_context_t ctx_id)
87 {
88 return syscall (SYS_io_destroy, ctx_id);
89 }
90
91 inline_size
92 int
93 ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
94 {
95 return syscall (SYS_io_submit, ctx_id, nr, cbp);
96 }
97
98 inline_size
99 int
100 ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
101 {
102 return syscall (SYS_io_cancel, ctx_id, cbp, result);
103 }
104
105 inline_size
106 int
107 ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
108 {
109 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
110 }
111
112 /*****************************************************************************/
113 /* actual backed implementation */
114
115 /* we use out own wrapper structure in acse we ever want to do something "clever" */
116 typedef struct aniocb
117 {
118 struct iocb io;
119 /*int inuse;*/
120 } *ANIOCBP;
121
122 inline_size
123 void
124 linuxaio_array_needsize_iocbp (ANIOCBP *base, int count)
125 {
126 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
127 while (count--)
128 {
129 *base = (ANIOCBP)ev_malloc (sizeof (**base));
130 /* TODO: full zero initialize required? */
131 memset (*base, 0, sizeof (**base));
132 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */
133 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
134 ++base;
135 }
136 }
137
138 ecb_cold
139 static void
140 linuxaio_free_iocbp (EV_P)
141 {
142 while (linuxaio_iocbpmax--)
143 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
144
145 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
146 }
147
148 static void
149 linuxaio_modify (EV_P_ int fd, int oev, int nev)
150 {
151 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
152 struct aniocb *iocb = linuxaio_iocbps [fd];
153
154 if (iocb->io.aio_buf)
155 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
156
157 if (nev)
158 {
159 iocb->io.aio_data = fd;
160 iocb->io.aio_fildes = fd;
161 iocb->io.aio_buf =
162 (nev & EV_READ ? POLLIN : 0)
163 | (nev & EV_WRITE ? POLLOUT : 0);
164
165 /* queue iocb up for io_submit */
166 /* this assumes we only ever get one call per fd per loop iteration */
167 ++linuxaio_submitcnt;
168 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
169 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
170 }
171 }
172
173 static void
174 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
175 {
176 while (nr)
177 {
178 int fd = ev->data;
179 int res = ev->res;
180
181 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
182
183 /* linux aio is oneshot: rearm fd */
184 linuxaio_iocbps [fd]->io.aio_buf = 0;
185 anfds [fd].events = 0;
186 fd_change (EV_A_ fd, 0);
187
188 /* feed events, we do not expect or handle POLLNVAL */
189 if (ecb_expect_false (res & POLLNVAL))
190 fd_kill (EV_A_ fd);
191 else
192 fd_event (
193 EV_A_
194 fd,
195 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
196 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
197 );
198
199 --nr;
200 ++ev;
201 }
202 }
203
204 /* get any events from ringbuffer, return true if any were handled */
205 static int
206 linuxaio_get_events_from_ring (EV_P)
207 {
208 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
209
210 unsigned head = ring->head;
211 unsigned tail = *(volatile unsigned *)&ring->tail;
212
213 if (head == tail)
214 return 0;
215
216 /* bail out if the ring buffer doesn't match the expected layout */
217 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC)
218 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
219 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
220 return 0;
221
222 ECB_MEMORY_FENCE_ACQUIRE;
223
224 /* parse all available events, but only once, to avoid starvation */
225 if (tail > head) /* normal case around */
226 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
227 else /* wrapped around */
228 {
229 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
230 linuxaio_parse_events (EV_A_ ring->io_events, tail);
231 }
232
233 ring->head = tail;
234
235 return 1;
236 }
237
238 /* read at least one event from kernel, or timeout */
239 inline_size
240 void
241 linuxaio_get_events (EV_P_ ev_tstamp timeout)
242 {
243 struct timespec ts;
244 struct io_event ioev;
245 int res;
246
247 if (linuxaio_get_events_from_ring (EV_A))
248 return;
249
250 /* no events, so wait for at least one, then poll ring buffer again */
251 /* this degrades to one event per loop iteration */
252 /* if the ring buffer changes layout, but so be it */
253
254 ts.tv_sec = (long)timeout;
255 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
256
257 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);
258
259 if (res < 0)
260 ev_syserr ("(libev) linuxaio io_getevents");
261 else if (res)
262 {
263 /* at least one event received, handle it and any remaining ones in the ring buffer */
264 linuxaio_parse_events (EV_A_ &ioev, 1);
265 linuxaio_get_events_from_ring (EV_A);
266 }
267 }
268
269 static void
270 linuxaio_poll (EV_P_ ev_tstamp timeout)
271 {
272 int submitted;
273
274 /* first phase: submit new iocbs */
275
276 /* io_submit might return less than the requested number of iocbs */
277 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
278 /* which allows us to pinpoint the errornous iocb */
279 for (submitted = 0; submitted < linuxaio_submitcnt; )
280 {
281 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
282
283 if (ecb_expect_false (res < 0))
284 if (errno == EAGAIN)
285 {
286 /* This happens when the ring buffer is full, at least. I assume this means
287 * that the event was queued synchronously during io_submit, and thus
288 * the buffer overflowd.
289 * In this case, we just try next loop iteration.
290 * This should not result in a few fds taking priority, as the interface
291 * is one-shot, and we submit iocb's in a round-robin fashion.
292 */
293 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
294 linuxaio_submitcnt -= submitted;
295 timeout = 0;
296 break;
297 }
298 else
299 ev_syserr ("(libev) linuxaio io_submit");
300
301 submitted += res;
302 }
303
304 linuxaio_submitcnt = 0;
305
306 /* second phase: fetch and parse events */
307
308 linuxaio_get_events (EV_A_ timeout);
309 }
310
311 inline_size
312 int
313 linuxaio_init (EV_P_ int flags)
314 {
315 /* would be great to have a nice test for IOCB_CMD_POLL instead */
316 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
317 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */
318 return 0;
319
320 linuxaio_ctx = 0;
321 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
322 return 0;
323
324 backend_modify = linuxaio_modify;
325 backend_poll = linuxaio_poll;
326
327 linuxaio_iocbpmax = 0;
328 linuxaio_iocbps = 0;
329
330 linuxaio_submits = 0;
331 linuxaio_submitmax = 0;
332 linuxaio_submitcnt = 0;
333
334 return EVBACKEND_LINUXAIO;
335 }
336
337 inline_size
338 void
339 linuxaio_destroy (EV_P)
340 {
341 linuxaio_free_iocbp (EV_A);
342 ev_io_destroy (linuxaio_ctx);
343 }
344
345 inline_size
346 void
347 linuxaio_fork (EV_P)
348 {
349 /* this frees all iocbs, which is very heavy-handed */
350 linuxaio_destroy (EV_A);
351 linuxaio_submitcnt = 0; /* all pointers were invalidated */
352
353 linuxaio_ctx = 0;
354 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
355 ev_syserr ("(libev) linuxaio io_setup");
356
357 fd_rearm_all (EV_A);
358 }
359