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 1.10 by root, Sun Jun 23 02:02:24 2019 UTC vs.
Revision 1.19 by root, Mon Jun 24 19:53:47 2019 UTC

… …
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 184 }
185} 185}
186 186
187#if EPOLL_FALLBACK
188
189static void
190linuxaio_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
203static void
204linuxaio_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
187static void 239static void
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 240linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 241{
190 while (nr) 242 while (nr)
191 { 243 {
192 int fd = ev->data; 244 int fd = ev->data;
193 int res = ev->res; 245 int res = ev->res;
194 246
195 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 247 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
196 248
197 /* linux aio is oneshot: rearm fd */ 249 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
198 linuxaio_iocbps [fd]->io.aio_buf = 0; 250 linuxaio_iocbps [fd]->io.aio_buf = 0;
199 anfds [fd].events = 0; 251 anfds [fd].events = 0;
200 fd_change (EV_A_ fd, 0); 252 fd_change (EV_A_ fd, 0);
201 253
202 /* feed events, we do not expect or handle POLLNVAL */ 254 /* feed events, we do not expect or handle POLLNVAL */
203 if (ecb_expect_false (res & POLLNVAL)) 255 if (expect_false (res & POLLNVAL))
204 fd_kill (EV_A_ fd); 256 fd_kill (EV_A_ fd);
205 else 257 else
206 fd_event ( 258 fd_event (
207 EV_A_ 259 EV_A_
208 fd, 260 fd,
… …
219static int 271static int
220linuxaio_get_events_from_ring (EV_P) 272linuxaio_get_events_from_ring (EV_P)
221{ 273{
222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 274 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
223 275
224 unsigned head = ring->head; 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;
225 unsigned tail = *(volatile unsigned *)&ring->tail; 280 unsigned tail = *(volatile unsigned *)&ring->tail;
226 281
227 if (head == tail) 282 if (head == tail)
228 return 0; 283 return 0;
229 284
230 /* bail out if the ring buffer doesn't match the expected layout */ 285 /* bail out if the ring buffer doesn't match the expected layout */
231 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 286 if (expect_false (ring->magic != AIO_RING_MAGIC)
232 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 287 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
233 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ 288 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
234 return 0; 289 return 0;
235 290
291 /* make sure the events up to tail are visible */
236 ECB_MEMORY_FENCE_ACQUIRE; 292 ECB_MEMORY_FENCE_ACQUIRE;
237 293
238 /* parse all available events, but only once, to avoid starvation */ 294 /* parse all available events, but only once, to avoid starvation */
239 if (tail > head) /* normal case around */ 295 if (tail > head) /* normal case around */
240 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 296 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
… …
242 { 298 {
243 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 299 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
244 linuxaio_parse_events (EV_A_ ring->io_events, tail); 300 linuxaio_parse_events (EV_A_ ring->io_events, tail);
245 } 301 }
246 302
247 ring->head = tail; 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;
248 309
249 return 1; 310 return 1;
250} 311}
251 312
252/* read at least one event from kernel, or timeout */ 313/* read at least one event from kernel, or timeout */
253inline_size 314inline_size
254void 315void
255linuxaio_get_events (EV_P_ ev_tstamp timeout) 316linuxaio_get_events (EV_P_ ev_tstamp timeout)
256{ 317{
257 struct timespec ts; 318 struct timespec ts;
258 struct io_event ioev; 319 struct io_event ioev[1];
259 int res; 320 int res;
260 321
261 if (linuxaio_get_events_from_ring (EV_A)) 322 if (linuxaio_get_events_from_ring (EV_A))
262 return; 323 return;
263 324
264 /* no events, so wait for at least one, then poll ring buffer again */ 325 /* no events, so wait for at least one, then poll ring buffer again */
265 /* this degrades to one event per loop iteration */ 326 /* this degrades to one event per loop iteration */
266 /* if the ring buffer changes layout, but so be it */ 327 /* if the ring buffer changes layout, but so be it */
267 328
329 EV_RELEASE_CB;
330
268 ts.tv_sec = (long)timeout; 331 ts.tv_sec = (long)timeout;
269 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 332 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
270 333
271 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 334 res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
335
336 EV_ACQUIRE_CB;
272 337
273 if (res < 0) 338 if (res < 0)
274 if (errno == EINTR) 339 if (errno == EINTR)
275 /* ignored */; 340 /* ignored */;
276 else 341 else
277 ev_syserr ("(libev) linuxaio io_getevents"); 342 ev_syserr ("(libev) linuxaio io_getevents");
278 else if (res) 343 else if (res)
279 { 344 {
280 /* at least one event received, handle it and any remaining ones in the ring buffer */ 345 /* at least one event received, handle it and any remaining ones in the ring buffer */
281 linuxaio_parse_events (EV_A_ &ioev, 1); 346 linuxaio_parse_events (EV_A_ ioev, res);
282 linuxaio_get_events_from_ring (EV_A); 347 linuxaio_get_events_from_ring (EV_A);
283 } 348 }
284} 349}
285
286#if EPOLL_FALLBACK
287static void
288linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
289{
290 struct epoll_event eev;
291
292 eev.events = EPOLLONESHOT;
293 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
294 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
295 eev.data.fd = iocb->aio_fildes;
296
297 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
298 ev_syserr ("(libeio) linuxaio epoll_ctl");
299}
300#endif
301 350
302static void 351static void
303linuxaio_poll (EV_P_ ev_tstamp timeout) 352linuxaio_poll (EV_P_ ev_tstamp timeout)
304{ 353{
305 int submitted; 354 int submitted;
… …
309 /* io_submit might return less than the requested number of iocbs */ 358 /* io_submit might return less than the requested number of iocbs */
310 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 359 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
311 /* which allows us to pinpoint the errornous iocb */ 360 /* which allows us to pinpoint the errornous iocb */
312 for (submitted = 0; submitted < linuxaio_submitcnt; ) 361 for (submitted = 0; submitted < linuxaio_submitcnt; )
313 { 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
314 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 370 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371#endif
315 372
316 if (ecb_expect_false (res < 0)) 373 if (expect_false (res < 0))
317 if (errno == EAGAIN) 374 if (errno == EAGAIN)
318 { 375 {
319 /* This happens when the ring buffer is full, at least. I assume this means 376 /* This happens when the ring buffer is full, at least. I assume this means
320 * that the event was queued synchronously during io_submit, and thus 377 * that the event was queued synchronously during io_submit, and thus
321 * the buffer overflowed. 378 * the buffer overflowed.
322 * In this case, we just try in next loop iteration. 379 * In this case, we just try in next loop iteration.
323 * This should not result in a few fds taking priority, as the interface 380 * This should not result in a few fds taking priority, as the interface
324 * is one-shot, and we submit iocb's in a round-robin fashion. 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?
325 */ 383 */
384 if (ecb_expect_false (submitted))
385 {
326 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 386 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
327 linuxaio_submitcnt -= submitted; 387 linuxaio_submitcnt -= submitted;
388 }
389
328 timeout = 0; 390 timeout = 0;
329 break; 391 break;
330 } 392 }
331#if EPOLL_FALLBACK 393#if EPOLL_FALLBACK
332 else if (errno == EINVAL) 394 else if (errno == EINVAL)
333 { 395 {
334 /* This hapΓΌpens for unsupported fds, officially, but in my testing, 396 /* This happens for unsupported fds, officially, but in my testing,
335 * also randomly happens for supported fds. We fall back to good old 397 * also randomly happens for supported fds. We fall back to good old
336 * poll() here, under the assumption that this is a very rare case. 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.
337 */ 402 */
338 struct iocb *iocb = linuxaio_submits [submitted]; 403 struct iocb *iocb = linuxaio_submits [submitted];
339 res = 1; /* skip this iocb */ 404 res = 1; /* skip this iocb */
340 405
341 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); 406 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
… …
353 /* second phase: fetch and parse events */ 418 /* second phase: fetch and parse events */
354 419
355 linuxaio_get_events (EV_A_ timeout); 420 linuxaio_get_events (EV_A_ timeout);
356} 421}
357 422
358#if EPOLL_FALLBACK
359
360static void
361linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
362{
363 struct epoll_event events[16];
364
365 for (;;)
366 {
367 int idx;
368 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
369
370 if (ecb_expect_false (res < 0))
371 ev_syserr ("(libev) linuxaio epoll_wait");
372 else if (!res)
373 break;
374
375 for (idx = res; idx--; )
376 {
377 int fd = events [idx].data.fd;
378 uint32_t ev = events [idx].events;
379
380 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
381
382 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
383
384 fd_event (EV_A_ fd,
385 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
386 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
387 }
388
389 if (res < sizeof (events) / sizeof (events [0]))
390 break;
391 }
392}
393
394#endif
395
396inline_size 423inline_size
397int 424int
398linuxaio_init (EV_P_ int flags) 425linuxaio_init (EV_P_ int flags)
399{ 426{
400 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 427 /* would be great to have a nice test for IOCB_CMD_POLL instead */
401 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 428 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
402 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 429#if EPOLL_FALLBACK
430 /* 4.19 made epoll work */
431 if (ev_linux_version () < 0x041300)
403 return 0; 432 return 0;
433#else
434 /* 4.18 introduced IOCB_CMD_POLL */
435 if (ev_linux_version () < 0x041200)
436 return 0;
437#endif
404 438
405 linuxaio_ctx = 0; 439 linuxaio_ctx = 0;
406 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 440 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
407 return 0; 441 return 0;
408 442
… …
413 ev_io_destroy (linuxaio_ctx); 447 ev_io_destroy (linuxaio_ctx);
414 return 0; 448 return 0;
415 } 449 }
416 450
417 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 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);
418 ev_io_start (EV_A_ &linuxaio_epoll_w); 453 ev_io_start (EV_A_ &linuxaio_epoll_w);
454 ev_unref (EV_A); /* watcher should not keep loop alive */
419#endif 455#endif
420 456
421 backend_modify = linuxaio_modify; 457 backend_modify = linuxaio_modify;
422 backend_poll = linuxaio_poll; 458 backend_poll = linuxaio_poll;
423 459
… …
459 ev_syserr ("(libev) linuxaio epoll_create"); 495 ev_syserr ("(libev) linuxaio epoll_create");
460 496
461 ev_io_stop (EV_A_ &linuxaio_epoll_w); 497 ev_io_stop (EV_A_ &linuxaio_epoll_w);
462 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 498 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
463 ev_io_start (EV_A_ &linuxaio_epoll_w); 499 ev_io_start (EV_A_ &linuxaio_epoll_w);
464 ev_unref (EV_A); /* watcher should not keep loop alive */
465#endif 500#endif
466 501
467 fd_rearm_all (EV_A); 502 fd_rearm_all (EV_A);
468} 503}
469 504

Diff Legend

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