ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
Revision: 1.19
Committed: Mon Jun 24 19:53:47 2019 UTC (4 years, 10 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.18: +76 -59 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 */
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 ev_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 ev_io_destroy (aio_context_t ctx_id)
93 {
94 return syscall (SYS_io_destroy, ctx_id);
95 }
96
97 inline_size
98 int
99 ev_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 ev_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 ev_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 count)
131 {
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--)
134 {
135 *base = (ANIOCBP)ev_malloc (sizeof (**base));
136 /* TODO: full zero initialize required? */
137 memset (*base, 0, sizeof (**base));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base;
141 }
142 }
143
144 ecb_cold
145 static void
146 linuxaio_free_iocbp (EV_P)
147 {
148 while (linuxaio_iocbpmax--)
149 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
150
151 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
152 }
153
154 static void
155 linuxaio_modify (EV_P_ int fd, int oev, int nev)
156 {
157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158 struct aniocb *iocb = linuxaio_iocbps [fd];
159
160 #if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0)
162 {
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
164 iocb->io.aio_reqprio = 0;
165 }
166 #endif
167
168 if (iocb->io.aio_buf)
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
170
171 if (nev)
172 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0);
178
179 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */
181 ++linuxaio_submitcnt;
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 }
185 }
186
187 #if EPOLL_FALLBACK
188
189 static void
190 linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
191 {
192 struct epoll_event eev;
193
194 eev.events = EPOLLONESHOT;
195 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
196 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
197 eev.data.fd = iocb->aio_fildes;
198
199 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
200 ev_syserr ("(libeio) linuxaio epoll_ctl");
201 }
202
203 static void
204 linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
205 {
206 struct epoll_event events[16];
207
208 for (;;)
209 {
210 int idx;
211 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
212
213 if (expect_false (res < 0))
214 ev_syserr ("(libev) linuxaio epoll_wait");
215 else if (!res)
216 break;
217
218 for (idx = res; idx--; )
219 {
220 int fd = events [idx].data.fd;
221 uint32_t ev = events [idx].events;
222
223 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
224
225 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
226
227 fd_event (EV_A_ fd,
228 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
229 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
230 }
231
232 if (res < sizeof (events) / sizeof (events [0]))
233 break;
234 }
235 }
236
237 #endif
238
239 static void
240 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
241 {
242 while (nr)
243 {
244 int fd = ev->data;
245 int res = ev->res;
246
247 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
248
249 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
250 linuxaio_iocbps [fd]->io.aio_buf = 0;
251 anfds [fd].events = 0;
252 fd_change (EV_A_ fd, 0);
253
254 /* feed events, we do not expect or handle POLLNVAL */
255 if (expect_false (res & POLLNVAL))
256 fd_kill (EV_A_ fd);
257 else
258 fd_event (
259 EV_A_
260 fd,
261 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
262 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
263 );
264
265 --nr;
266 ++ev;
267 }
268 }
269
270 /* get any events from ringbuffer, return true if any were handled */
271 static int
272 linuxaio_get_events_from_ring (EV_P)
273 {
274 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
275
276 /* the kernel reads and writes both of these variables, */
277 /* as a C extension, we assume that volatile use here */
278 /* both makes reads atomic and once-only */
279 unsigned head = *(volatile unsigned *)&ring->head;
280 unsigned tail = *(volatile unsigned *)&ring->tail;
281
282 if (head == tail)
283 return 0;
284
285 /* bail out if the ring buffer doesn't match the expected layout */
286 if (expect_false (ring->magic != AIO_RING_MAGIC)
287 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
288 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
289 return 0;
290
291 /* make sure the events up to tail are visible */
292 ECB_MEMORY_FENCE_ACQUIRE;
293
294 /* parse all available events, but only once, to avoid starvation */
295 if (tail > head) /* normal case around */
296 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
297 else /* wrapped around */
298 {
299 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
300 linuxaio_parse_events (EV_A_ ring->io_events, tail);
301 }
302
303 /* TODO: we only need a compiler barrier here, not a read fence */
304 ECB_MEMORY_FENCE_RELEASE;
305 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
306 *(volatile unsigned *)&ring->head = tail;
307 /* make sure kernel can see our new head value - probably not required */
308 ECB_MEMORY_FENCE_RELEASE;
309
310 return 1;
311 }
312
313 /* read at least one event from kernel, or timeout */
314 inline_size
315 void
316 linuxaio_get_events (EV_P_ ev_tstamp timeout)
317 {
318 struct timespec ts;
319 struct io_event ioev[1];
320 int res;
321
322 if (linuxaio_get_events_from_ring (EV_A))
323 return;
324
325 /* no events, so wait for at least one, then poll ring buffer again */
326 /* this degrades to one event per loop iteration */
327 /* if the ring buffer changes layout, but so be it */
328
329 EV_RELEASE_CB;
330
331 ts.tv_sec = (long)timeout;
332 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
333
334 res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
335
336 EV_ACQUIRE_CB;
337
338 if (res < 0)
339 if (errno == EINTR)
340 /* ignored */;
341 else
342 ev_syserr ("(libev) linuxaio io_getevents");
343 else if (res)
344 {
345 /* at least one event received, handle it and any remaining ones in the ring buffer */
346 linuxaio_parse_events (EV_A_ ioev, res);
347 linuxaio_get_events_from_ring (EV_A);
348 }
349 }
350
351 static void
352 linuxaio_poll (EV_P_ ev_tstamp timeout)
353 {
354 int submitted;
355
356 /* first phase: submit new iocbs */
357
358 /* io_submit might return less than the requested number of iocbs */
359 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
360 /* which allows us to pinpoint the errornous iocb */
361 for (submitted = 0; submitted < linuxaio_submitcnt; )
362 {
363 #if 0
364 int res;
365 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
366 res = ev_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
367 else
368 { res = -1; errno = EINVAL; };
369 #else
370 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371 #endif
372
373 if (expect_false (res < 0))
374 if (errno == EAGAIN)
375 {
376 /* This happens when the ring buffer is full, at least. I assume this means
377 * that the event was queued synchronously during io_submit, and thus
378 * the buffer overflowed.
379 * In this case, we just try in next loop iteration.
380 * This should not result in a few fds taking priority, as the interface
381 * is one-shot, and we submit iocb's in a round-robin fashion.
382 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
383 */
384 if (ecb_expect_false (submitted))
385 {
386 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
387 linuxaio_submitcnt -= submitted;
388 }
389
390 timeout = 0;
391 break;
392 }
393 #if EPOLL_FALLBACK
394 else if (errno == EINVAL)
395 {
396 /* This happens for unsupported fds, officially, but in my testing,
397 * also randomly happens for supported fds. We fall back to good old
398 * poll() here, under the assumption that this is a very rare case.
399 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
400 * discussion about such a case (ttys) where polling for POLLIN
401 * fails but POLLIN|POLLOUT works.
402 */
403 struct iocb *iocb = linuxaio_submits [submitted];
404 res = 1; /* skip this iocb */
405
406 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408 }
409 #endif
410 else
411 ev_syserr ("(libev) linuxaio io_submit");
412
413 submitted += res;
414 }
415
416 linuxaio_submitcnt = 0;
417
418 /* second phase: fetch and parse events */
419
420 linuxaio_get_events (EV_A_ timeout);
421 }
422
423 inline_size
424 int
425 linuxaio_init (EV_P_ int flags)
426 {
427 /* would be great to have a nice test for IOCB_CMD_POLL instead */
428 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
429 #if EPOLL_FALLBACK
430 /* 4.19 made epoll work */
431 if (ev_linux_version () < 0x041300)
432 return 0;
433 #else
434 /* 4.18 introduced IOCB_CMD_POLL */
435 if (ev_linux_version () < 0x041200)
436 return 0;
437 #endif
438
439 linuxaio_ctx = 0;
440 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
441 return 0;
442
443 #if EPOLL_FALLBACK
444 backend_fd = ev_epoll_create ();
445 if (backend_fd < 0)
446 {
447 ev_io_destroy (linuxaio_ctx);
448 return 0;
449 }
450
451 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
452 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
453 ev_io_start (EV_A_ &linuxaio_epoll_w);
454 ev_unref (EV_A); /* watcher should not keep loop alive */
455 #endif
456
457 backend_modify = linuxaio_modify;
458 backend_poll = linuxaio_poll;
459
460 linuxaio_iocbpmax = 0;
461 linuxaio_iocbps = 0;
462
463 linuxaio_submits = 0;
464 linuxaio_submitmax = 0;
465 linuxaio_submitcnt = 0;
466
467 return EVBACKEND_LINUXAIO;
468 }
469
470 inline_size
471 void
472 linuxaio_destroy (EV_P)
473 {
474 #if EPOLL_FALLBACK
475 close (backend_fd);
476 #endif
477 linuxaio_free_iocbp (EV_A);
478 ev_io_destroy (linuxaio_ctx);
479 }
480
481 inline_size
482 void
483 linuxaio_fork (EV_P)
484 {
485 /* this frees all iocbs, which is very heavy-handed */
486 linuxaio_destroy (EV_A);
487 linuxaio_submitcnt = 0; /* all pointers were invalidated */
488
489 linuxaio_ctx = 0;
490 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
491 ev_syserr ("(libev) linuxaio io_setup");
492
493 #if EPOLL_FALLBACK
494 while ((backend_fd = ev_epoll_create ()) < 0)
495 ev_syserr ("(libev) linuxaio epoll_create");
496
497 ev_io_stop (EV_A_ &linuxaio_epoll_w);
498 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
499 ev_io_start (EV_A_ &linuxaio_epoll_w);
500 #endif
501
502 fd_rearm_all (EV_A);
503 }
504