ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
Revision: 1.23
Committed: Mon Jun 24 21:34:47 2019 UTC (4 years, 10 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.22: +1 -1 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 #define EPOLL_FALLBACK 1
41
42 #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
43 #include <poll.h>
44 #include <linux/aio_abi.h>
45
46 #if EPOLL_FALLBACK
47 # include <sys/epoll.h>
48 #endif
49
50 /* we try to fill 4kB pages exactly.
51 * the ring buffer header is 32 bytes, every io event is 32 bytes.
52 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
53 * therefore the calculation below will use "exactly" 4kB for the ring buffer
54 */
55 #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56
57 /*****************************************************************************/
58 /* syscall wrapdadoop - this section has the raw syscall definitions */
59
60 #include <sys/syscall.h> /* no glibc wrappers */
61
62 /* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63 #define IOCB_CMD_POLL 5
64
65 /* taken from linux/fs/aio.c */
66 #define AIO_RING_MAGIC 0xa10a10a1
67 #define AIO_RING_INCOMPAT_FEATURES 0
68 struct aio_ring
69 {
70 unsigned id; /* kernel internal index number */
71 unsigned nr; /* number of io_events */
72 unsigned head; /* Written to by userland or by kernel. */
73 unsigned tail;
74
75 unsigned magic;
76 unsigned compat_features;
77 unsigned incompat_features;
78 unsigned header_length; /* size of aio_ring */
79
80 struct io_event io_events[0];
81 };
82
83 inline_size
84 int
85 evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86 {
87 return syscall (SYS_io_setup, nr_events, ctx_idp);
88 }
89
90 inline_size
91 int
92 evsys_io_destroy (aio_context_t ctx_id)
93 {
94 return syscall (SYS_io_destroy, ctx_id);
95 }
96
97 inline_size
98 int
99 evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100 {
101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
102 }
103
104 inline_size
105 int
106 evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
107 {
108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
109 }
110
111 inline_size
112 int
113 evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
114 {
115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116 }
117
118 /*****************************************************************************/
119 /* actual backed implementation */
120
121 /* we use out own wrapper structure in acse we ever want to do something "clever" */
122 typedef struct aniocb
123 {
124 struct iocb io;
125 /*int inuse;*/
126 } *ANIOCBP;
127
128 inline_size
129 void
130 linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131 {
132 while (count--)
133 {
134 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
135 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136
137 /* full zero initialise is probably not required at the moment, but
138 * this is not well documented, so we better do it.
139 */
140 memset (iocb, 0, sizeof (*iocb));
141
142 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
143 iocb->io.aio_data = offset;
144 iocb->io.aio_fildes = offset;
145
146 base [offset++] = iocb;
147 }
148 }
149
150 ecb_cold
151 static void
152 linuxaio_free_iocbp (EV_P)
153 {
154 while (linuxaio_iocbpmax--)
155 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
156
157 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
158 }
159
160 static void
161 linuxaio_modify (EV_P_ int fd, int oev, int nev)
162 {
163 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
164 ANIOCBP iocb = linuxaio_iocbps [fd];
165
166 #if EPOLL_FALLBACK
167 if (iocb->io.aio_reqprio < 0)
168 {
169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
170 iocb->io.aio_reqprio = 0;
171 }
172 #endif
173
174 if (iocb->io.aio_buf)
175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
176
177 if (nev)
178 {
179 iocb->io.aio_buf =
180 (nev & EV_READ ? POLLIN : 0)
181 | (nev & EV_WRITE ? POLLOUT : 0);
182
183 /* queue iocb up for io_submit */
184 /* this assumes we only ever get one call per fd per loop iteration */
185 ++linuxaio_submitcnt;
186 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
187 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
188 }
189 }
190
191 #if EPOLL_FALLBACK
192
193 static void
194 linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
195 {
196 struct epoll_event eev;
197
198 eev.events = EPOLLONESHOT;
199 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
200 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
201 eev.data.fd = iocb->aio_fildes;
202
203 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
204 ev_syserr ("(libeio) linuxaio epoll_ctl");
205 }
206
207 static void
208 linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
209 {
210 struct epoll_event events[16];
211
212 for (;;)
213 {
214 int idx;
215 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
216
217 if (expect_false (res < 0))
218 ev_syserr ("(libev) linuxaio epoll_wait");
219 else if (!res)
220 break;
221
222 for (idx = res; idx--; )
223 {
224 int fd = events [idx].data.fd;
225 uint32_t ev = events [idx].events;
226
227 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
228
229 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
230
231 fd_event (EV_A_ fd,
232 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
233 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
234 }
235
236 if (res < sizeof (events) / sizeof (events [0]))
237 break;
238 }
239 }
240
241 #endif
242
243 static void
244 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
245 {
246 while (nr)
247 {
248 int fd = ev->data;
249 int res = ev->res;
250
251 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
252
253 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
254 linuxaio_iocbps [fd]->io.aio_buf = 0;
255 anfds [fd].events = 0;
256 fd_change (EV_A_ fd, 0);
257
258 /* feed events, we do not expect or handle POLLNVAL */
259 fd_event (
260 EV_A_
261 fd,
262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
264 );
265
266 --nr;
267 ++ev;
268 }
269 }
270
271 /* get any events from ringbuffer, return true if any were handled */
272 static int
273 linuxaio_get_events_from_ring (EV_P)
274 {
275 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
276
277 /* the kernel reads and writes both of these variables, */
278 /* as a C extension, we assume that volatile use here */
279 /* both makes reads atomic and once-only */
280 unsigned head = *(volatile unsigned *)&ring->head;
281 unsigned tail = *(volatile unsigned *)&ring->tail;
282
283 if (head == tail)
284 return 0;
285
286 /* bail out if the ring buffer doesn't match the expected layout */
287 if (expect_false (ring->magic != AIO_RING_MAGIC)
288 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
289 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
290 return 0;
291
292 /* make sure the events up to tail are visible */
293 ECB_MEMORY_FENCE_ACQUIRE;
294
295 /* parse all available events, but only once, to avoid starvation */
296 if (tail > head) /* normal case around */
297 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
298 else /* wrapped around */
299 {
300 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
301 linuxaio_parse_events (EV_A_ ring->io_events, tail);
302 }
303
304 /* TODO: we only need a compiler barrier here, not a read fence */
305 ECB_MEMORY_FENCE_RELEASE;
306 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
307 *(volatile unsigned *)&ring->head = tail;
308 /* make sure kernel can see our new head value - probably not required */
309 ECB_MEMORY_FENCE_RELEASE;
310
311 return 1;
312 }
313
314 /* read at least one event from kernel, or timeout */
315 inline_size
316 void
317 linuxaio_get_events (EV_P_ ev_tstamp timeout)
318 {
319 struct timespec ts;
320 struct io_event ioev[1];
321 int res;
322
323 if (linuxaio_get_events_from_ring (EV_A))
324 return;
325
326 /* no events, so wait for at least one, then poll ring buffer again */
327 /* this degrades to one event per loop iteration */
328 /* if the ring buffer changes layout, but so be it */
329
330 EV_RELEASE_CB;
331
332 ts.tv_sec = (long)timeout;
333 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
334
335 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
336
337 EV_ACQUIRE_CB;
338
339 if (res < 0)
340 if (errno == EINTR)
341 /* ignored */;
342 else
343 ev_syserr ("(libev) linuxaio io_getevents");
344 else if (res)
345 {
346 /* at least one event received, handle it and any remaining ones in the ring buffer */
347 linuxaio_parse_events (EV_A_ ioev, res);
348 linuxaio_get_events_from_ring (EV_A);
349 }
350 }
351
352 static void
353 linuxaio_poll (EV_P_ ev_tstamp timeout)
354 {
355 int submitted;
356
357 /* first phase: submit new iocbs */
358
359 /* io_submit might return less than the requested number of iocbs */
360 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
361 /* which allows us to pinpoint the errornous iocb */
362 for (submitted = 0; submitted < linuxaio_submitcnt; )
363 {
364 #if 0
365 int res;
366 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
367 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
368 else
369 { res = -1; errno = EINVAL; };
370 #else
371 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
372 #endif
373
374 if (expect_false (res < 0))
375 if (errno == EAGAIN)
376 {
377 /* This happens when the ring buffer is full, at least. I assume this means
378 * that the event was queued synchronously during io_submit, and thus
379 * the buffer overflowed.
380 * In this case, we just try in next loop iteration.
381 * This should not result in a few fds taking priority, as the interface
382 * is one-shot, and we submit iocb's in a round-robin fashion.
383 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
384 */
385 if (ecb_expect_false (submitted))
386 {
387 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
388 linuxaio_submitcnt -= submitted;
389 }
390
391 timeout = 0;
392 break;
393 }
394 #if EPOLL_FALLBACK
395 else if (errno == EINVAL)
396 {
397 /* This happens for unsupported fds, officially, but in my testing,
398 * also randomly happens for supported fds. We fall back to good old
399 * poll() here, under the assumption that this is a very rare case.
400 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
401 * discussion about such a case (ttys) where polling for POLLIN
402 * fails but POLLIN|POLLOUT works.
403 */
404 struct iocb *iocb = linuxaio_submits [submitted];
405
406 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408
409 res = 1; /* skip this iocb */
410 }
411 #endif
412 else if (errno == EBADF)
413 {
414 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
415
416 res = 1; /* skip this iocb */
417 }
418 else
419 ev_syserr ("(libev) linuxaio io_submit");
420
421 submitted += res;
422 }
423
424 linuxaio_submitcnt = 0;
425
426 /* second phase: fetch and parse events */
427
428 linuxaio_get_events (EV_A_ timeout);
429 }
430
431 inline_size
432 int
433 linuxaio_init (EV_P_ int flags)
434 {
435 /* would be great to have a nice test for IOCB_CMD_POLL instead */
436 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
437 #if EPOLL_FALLBACK
438 /* 4.19 made epoll work */
439 if (ev_linux_version () < 0x041300)
440 return 0;
441 #else
442 /* 4.18 introduced IOCB_CMD_POLL */
443 if (ev_linux_version () < 0x041200)
444 return 0;
445 #endif
446
447 linuxaio_ctx = 0;
448 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
449 return 0;
450
451 #if EPOLL_FALLBACK
452 backend_fd = ev_epoll_create ();
453 if (backend_fd < 0)
454 {
455 evsys_io_destroy (linuxaio_ctx);
456 return 0;
457 }
458
459 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
460 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
461 ev_io_start (EV_A_ &linuxaio_epoll_w);
462 ev_unref (EV_A); /* watcher should not keep loop alive */
463 #endif
464
465 backend_modify = linuxaio_modify;
466 backend_poll = linuxaio_poll;
467
468 linuxaio_iocbpmax = 0;
469 linuxaio_iocbps = 0;
470
471 linuxaio_submits = 0;
472 linuxaio_submitmax = 0;
473 linuxaio_submitcnt = 0;
474
475 return EVBACKEND_LINUXAIO;
476 }
477
478 inline_size
479 void
480 linuxaio_destroy (EV_P)
481 {
482 #if EPOLL_FALLBACK
483 close (backend_fd);
484 #endif
485 linuxaio_free_iocbp (EV_A);
486 evsys_io_destroy (linuxaio_ctx);
487 }
488
489 inline_size
490 void
491 linuxaio_fork (EV_P)
492 {
493 /* this frees all iocbs, which is very heavy-handed */
494 linuxaio_destroy (EV_A);
495 linuxaio_submitcnt = 0; /* all pointers were invalidated */
496
497 linuxaio_ctx = 0;
498 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
499 ev_syserr ("(libev) linuxaio io_setup");
500
501 #if EPOLL_FALLBACK
502 while ((backend_fd = ev_epoll_create ()) < 0)
503 ev_syserr ("(libev) linuxaio epoll_create");
504
505 ev_io_stop (EV_A_ &linuxaio_epoll_w);
506 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
507 ev_io_start (EV_A_ &linuxaio_epoll_w);
508 #endif
509
510 fd_rearm_all (EV_A);
511 }
512