ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
(Generate patch)

Comparing libev/ev_linuxaio.c (file contents):
Revision 1.24 by root, Mon Jun 24 22:27:29 2019 UTC vs.
Revision 1.25 by root, Tue Jun 25 04:52:30 2019 UTC

… …
35 * and other provisions required by the GPL. If you do not delete the 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 36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL. 37 * either the BSD or the GPL.
38 */ 38 */
39 39
40#define EPOLL_FALLBACK 1 40/*
41 * general notes about linux aio:
42 *
43 * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
44 * 4.18 looks too good to be true: both watchers and events can be
45 * batched, and events can even be handled in userspace using
46 * a ring buffer shared with the kernel. watchers can be canceled
47 * regardless of whether the fd has been closed. no problems with fork.
48 * ok, the ring buffer is 200% undocumented (there isn't even a
49 * header file), but otherwise, it's pure bliss!
50 * b) ok, watchers are one-shot, so you have to re-arm active ones
51 * on every iteration. so much for syscall-less event handling,
52 * but at least these re-arms can be batched, no big deal, right?
53 * c) well, linux as usual: the documentation lies to you: io_submit
54 * sometimes returns EINVAL because the kernel doesn't feel like
55 * handling your poll mask - ttys can be polled for POLLOUT,
56 * POLLOUT|POLLIN, but polling for POLLIN fails. just great,
57 * so we have to fall back to something else (hello, epoll),
58 * but at least the fallback can be slow, because these are
59 * exceptional cases, right?
60 * d) hmm, you have to tell the kernel the maximum number of watchers
61 * you want to queue when initialiasing the aio context. but of
62 * course the real limit is magically calculated in the kernel, and
63 * is often higher then we asked for. so we just have to destroy
64 * the aio context and re-create it a bit larger if we hit the limit.
65 * (starts to remind you of epoll? well, it's a bit more deterministic
66 * and less gambling, but still ugly as hell).
67 * e) that's when you find out you can also hit an arbitrary system-wide
68 * limit. or the kernel simply doesn't want to handle your watchers.
69 * what the fuck do we do then? you guessed it, in the middle
70 * of event handling we have to switch to 100% epoll polling. and
71 * that better is as fast as normal epoll polling, so you practically
72 * have to use the normal epoll backend with all its quirks.
73 * f) end result of this trainwreck: it inherits all the disadvantages
74 * from epoll, while adding a number on its own. why even bother to use
75 * it? because if conditions are right and your fds are supported and you
76 * don't hit a limit, this backend is actually faster, doesn't gamble with
77 * your fds, batches watchers and events and doesn't require costly state
78 * recreates. well, until it does.
79 * g) all of this makes this backend use almost twice as much code as epoll.
80 * which in turn uses twice as much code as poll. and thats not counting
81 * the fact that this backend also depends on the epoll backend, making
82 * it three times as much code as poll, or kqueue.
83 * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
84 * it's clear that whwetaver linux does is far, far far worse.
85 */
41 86
42#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ 87#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
43#include <poll.h> 88#include <poll.h>
44#include <linux/aio_abi.h> 89#include <linux/aio_abi.h>
45 90
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/*****************************************************************************/ 91/*****************************************************************************/
58/* syscall wrapdadoop - this section has the raw syscall definitions */ 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
59 93
60#include <sys/syscall.h> /* no glibc wrappers */ 94#include <sys/syscall.h> /* no glibc wrappers */
61 95
62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */ 96/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63#define IOCB_CMD_POLL 5 97#define IOCB_CMD_POLL 5
64 98
65/* taken from linux/fs/aio.c */ 99/* taken from linux/fs/aio.c. yup, that's a .c file.
100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for.
103 */
66#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
67#define AIO_RING_INCOMPAT_FEATURES 0 105#define AIO_RING_INCOMPAT_FEATURES 0
68struct aio_ring 106struct aio_ring
69{ 107{
70 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
… …
116} 154}
117 155
118/*****************************************************************************/ 156/*****************************************************************************/
119/* actual backed implementation */ 157/* actual backed implementation */
120 158
159ecb_cold
160static int
161linuxaio_nr_events (EV_P)
162{
163 /* we start with 16 iocbs and incraese from there
164 * that's tiny, but the kernel has a rather low system-wide
165 * limit that can be reached quickly, so let's be parsimonious
166 * with this resource.
167 * Rest assured, the kernel generously rounds up small and big numbers
168 * in different ways (but doesn't seem to charge you for it).
169 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
170 * and this helps to take advantage of that limit.
171 */
172
173 /* we try to fill 4kB pages exactly.
174 * the ring buffer header is 32 bytes, every io event is 32 bytes.
175 * the kernel takes the io requests number, doubles it, adds 2
176 * and adds the ring buffer.
177 * the way we use this is by starting low, and then roughly doubling the
178 * size each time we hit a limit.
179 */
180
181 int requests = 15 << linuxaio_iteration;
182 int one_page = (4096
183 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
184 int first_page = ((4096 - sizeof (struct aio_ring))
185 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
186
187 /* if everything fits into one page, use count exactly */
188 if (requests > first_page)
189 /* otherwise, round down to full pages and add the first page */
190 requests = requests / one_page * one_page + first_page;
191
192 return requests;
193}
194
121/* we use out own wrapper structure in acse we ever want to do something "clever" */ 195/* we use out own wrapper structure in acse we ever want to do something "clever" */
122typedef struct aniocb 196typedef struct aniocb
123{ 197{
124 struct iocb io; 198 struct iocb io;
125 /*int inuse;*/ 199 /*int inuse;*/
… …
161linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
162{ 236{
163 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 237 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
164 ANIOCBP iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
165 239
166#if EPOLL_FALLBACK
167 if (iocb->io.aio_reqprio < 0) 240 if (iocb->io.aio_reqprio < 0)
168 { 241 {
242 /* we handed this fd over to epoll, so undo this first */
243 /* we do it manually becvause the optimisations on epoll_modfy won't do us any good */
169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
170 iocb->io.aio_reqprio = 0; 245 iocb->io.aio_reqprio = 0;
171 } 246 }
172#endif
173 247
174 if (iocb->io.aio_buf) 248 if (iocb->io.aio_buf)
175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 249 /* io_cancel always returns some error on relevant kernels, but works */
250 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
176 251
177 if (nev) 252 if (nev)
178 { 253 {
179 iocb->io.aio_buf = 254 iocb->io.aio_buf =
180 (nev & EV_READ ? POLLIN : 0) 255 (nev & EV_READ ? POLLIN : 0)
… …
186 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 261 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
187 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 262 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
188 } 263 }
189} 264}
190 265
191#if EPOLL_FALLBACK
192
193static void
194linuxaio_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
207static void 266static void
208linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 267linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
209{ 268{
210 struct epoll_event events[16]; 269 epoll_poll (EV_A_ 0);
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} 270}
240 271
241#endif 272static void
273linuxaio_fd_rearm (EV_P_ int fd)
274{
275 anfds [fd].events = 0;
276 linuxaio_iocbps [fd]->io.aio_buf = 0;
277 fd_change (EV_A_ fd, EV_ANFD_REIFY);
278}
242 279
243static void 280static void
244linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 281linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
245{ 282{
246 while (nr) 283 while (nr)
247 { 284 {
248 int fd = ev->data; 285 int fd = ev->data;
249 int res = ev->res; 286 int res = ev->res;
250 287
251 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 288 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);
257 289
258 /* feed events, we do not expect or handle POLLNVAL */ 290 /* feed events, we do not expect or handle POLLNVAL */
259 fd_event ( 291 fd_event (
260 EV_A_ 292 EV_A_
261 fd, 293 fd,
262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 294 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 295 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
264 ); 296 );
265 297
298 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
299 linuxaio_fd_rearm (EV_A_ fd);
300
266 --nr; 301 --nr;
267 ++ev; 302 ++ev;
268 } 303 }
269} 304}
270 305
… …
346 linuxaio_parse_events (EV_A_ ioev, res); 381 linuxaio_parse_events (EV_A_ ioev, res);
347 linuxaio_get_events_from_ring (EV_A); 382 linuxaio_get_events_from_ring (EV_A);
348 } 383 }
349} 384}
350 385
386static int
387linuxaio_io_setup (EV_P)
388{
389 linuxaio_ctx = 0;
390 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
391}
392
351static void 393static void
352linuxaio_poll (EV_P_ ev_tstamp timeout) 394linuxaio_poll (EV_P_ ev_tstamp timeout)
353{ 395{
354 int submitted; 396 int submitted;
355 397
… …
358 /* io_submit might return less than the requested number of iocbs */ 400 /* 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, */ 401 /* 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 */ 402 /* which allows us to pinpoint the errornous iocb */
361 for (submitted = 0; submitted < linuxaio_submitcnt; ) 403 for (submitted = 0; submitted < linuxaio_submitcnt; )
362 { 404 {
363#if 0
364 int res;
365 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
366 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
367 else
368 { res = -1; errno = EINVAL; };
369#else
370 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 405 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371#endif
372 406
373 if (expect_false (res < 0)) 407 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) 408 if (errno == EINVAL)
395 { 409 {
396 /* This happens for unsupported fds, officially, but in my testing, 410 /* This happens for unsupported fds, officially, but in my testing,
397 * also randomly happens for supported fds. We fall back to good old 411 * 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. 412 * poll() here, under the assumption that this is a very rare case.
399 * See https://lore.kernel.org/patchwork/patch/1047453/ to see 413 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
400 * discussion about such a case (ttys) where polling for POLLIN 414 * discussion about such a case (ttys) where polling for POLLIN
401 * fails but POLLIN|POLLOUT works. 415 * fails but POLLIN|POLLOUT works.
402 */ 416 */
403 struct iocb *iocb = linuxaio_submits [submitted]; 417 struct iocb *iocb = linuxaio_submits [submitted];
404 418 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
405 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
406 iocb->aio_reqprio = -1; /* mark iocb as epoll */ 419 iocb->aio_reqprio = -1; /* mark iocb as epoll */
407 420
408 res = 1; /* skip this iocb */ 421 res = 1; /* skip this iocb - another iocb, another chance */
409 } 422 }
410#endif 423 else if (errno == EAGAIN)
424 {
425 /* This happens when the ring buffer is full, or some other shit we
426 * dont' know and isn't documented. Most likely because we have too
427 * many requests and linux aio can't be assed to handle them.
428 * In this case, we try to allocate a larger ring buffer, freeing
429 * ours first. This might fail, in which case we have to fall back to 100%
430 * epoll.
431 * God, how I hate linux not getting its act together. Ever.
432 */
433 evsys_io_destroy (linuxaio_ctx);
434 linuxaio_submitcnt = 0;
435
436 /* rearm all fds with active iocbs */
437 {
438 int fd;
439 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
440 if (linuxaio_iocbps [fd]->io.aio_buf)
441 linuxaio_fd_rearm (EV_A_ fd);
442 }
443
444 ++linuxaio_iteration;
445 if (linuxaio_io_setup (EV_A) < 0)
446 {
447 /* to bad, we can't get a new aio context, go 100% epoll */
448 linuxaio_free_iocbp (EV_A);
449 ev_io_stop (EV_A_ &linuxaio_epoll_w);
450 ev_ref (EV_A);
451 linuxaio_ctx = 0;
452 backend_modify = epoll_modify;
453 backend_poll = epoll_poll;
454 }
455
456 timeout = 0;
457 /* it's easiest to handle this mess in another iteration */
458 return;
459 }
411 else if (errno == EBADF) 460 else if (errno == EBADF)
412 { 461 {
413 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 462 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
414 463
415 res = 1; /* skip this iocb */ 464 res = 1; /* skip this iocb */
… …
431int 480int
432linuxaio_init (EV_P_ int flags) 481linuxaio_init (EV_P_ int flags)
433{ 482{
434 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 483 /* would be great to have a nice test for IOCB_CMD_POLL instead */
435 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 484 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
436#if EPOLL_FALLBACK 485 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work */
437 /* 4.19 made epoll work */
438 if (ev_linux_version () < 0x041300) 486 if (ev_linux_version () < 0x041300)
439 return 0; 487 return 0;
440#else 488
441 /* 4.18 introduced IOCB_CMD_POLL */ 489 if (!epoll_init (EV_A_ 0))
442 if (ev_linux_version () < 0x041200)
443 return 0; 490 return 0;
444#endif
445 491
446 linuxaio_ctx = 0; 492 linuxaio_iteration = 0;
447 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
448 return 0;
449 493
450#if EPOLL_FALLBACK 494 if (linuxaio_io_setup (EV_A) < 0)
451 backend_fd = ev_epoll_create ();
452 if (backend_fd < 0)
453 { 495 {
454 evsys_io_destroy (linuxaio_ctx); 496 epoll_destroy (EV_A);
455 return 0; 497 return 0;
456 } 498 }
457 499
458 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 500 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
459 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 501 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
460 ev_io_start (EV_A_ &linuxaio_epoll_w); 502 ev_io_start (EV_A_ &linuxaio_epoll_w);
461 ev_unref (EV_A); /* watcher should not keep loop alive */ 503 ev_unref (EV_A); /* watcher should not keep loop alive */
462#endif
463 504
464 backend_modify = linuxaio_modify; 505 backend_modify = linuxaio_modify;
465 backend_poll = linuxaio_poll; 506 backend_poll = linuxaio_poll;
466 507
467 linuxaio_iocbpmax = 0; 508 linuxaio_iocbpmax = 0;
… …
476 517
477inline_size 518inline_size
478void 519void
479linuxaio_destroy (EV_P) 520linuxaio_destroy (EV_P)
480{ 521{
481#if EPOLL_FALLBACK 522 epoll_destroy (EV_A);
482 close (backend_fd);
483#endif
484 linuxaio_free_iocbp (EV_A); 523 linuxaio_free_iocbp (EV_A);
485 evsys_io_destroy (linuxaio_ctx); 524 evsys_io_destroy (linuxaio_ctx);
486} 525}
487 526
488inline_size 527inline_size
… …
491{ 530{
492 /* this frees all iocbs, which is very heavy-handed */ 531 /* this frees all iocbs, which is very heavy-handed */
493 linuxaio_destroy (EV_A); 532 linuxaio_destroy (EV_A);
494 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 533 linuxaio_submitcnt = 0; /* all pointers were invalidated */
495 534
496 linuxaio_ctx = 0; 535 linuxaio_iteration = 0; /* we start over in the child */
497 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 536
537 while (linuxaio_io_setup (EV_A) < 0)
498 ev_syserr ("(libev) linuxaio io_setup"); 538 ev_syserr ("(libev) linuxaio io_setup");
499 539
500#if EPOLL_FALLBACK 540 epoll_fork (EV_A);
501 while ((backend_fd = ev_epoll_create ()) < 0)
502 ev_syserr ("(libev) linuxaio epoll_create");
503 541
504 ev_io_stop (EV_A_ &linuxaio_epoll_w); 542 ev_io_stop (EV_A_ &linuxaio_epoll_w);
505 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 543 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
506 ev_io_start (EV_A_ &linuxaio_epoll_w); 544 ev_io_start (EV_A_ &linuxaio_epoll_w);
507#endif
508 545
546 /* epoll_fork already did this. hopefully */
509 fd_rearm_all (EV_A); 547 /*fd_rearm_all (EV_A);*/
510} 548}
511 549

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines