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 1.32 by root, Tue Jun 25 18:07:44 2019 UTC vs.
Revision 1.48 by root, Mon Jul 8 03:00:09 2019 UTC

… …
100 * not only is this totally undocumented, not even the source code 100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and 101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for. 102 * incompat_features are, or what header_length actually is for.
103 */ 103 */
104#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
105#define AIO_RING_INCOMPAT_FEATURES 0 105#define EV_AIO_RING_INCOMPAT_FEATURES 0
106struct aio_ring 106struct aio_ring
107{ 107{
108 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
109 unsigned nr; /* number of io_events */ 109 unsigned nr; /* number of io_events */
110 unsigned head; /* Written to by userland or by kernel. */ 110 unsigned head; /* Written to by userland or by kernel. */
… …
115 unsigned incompat_features; 115 unsigned incompat_features;
116 unsigned header_length; /* size of aio_ring */ 116 unsigned header_length; /* size of aio_ring */
117 117
118 struct io_event io_events[0]; 118 struct io_event io_events[0];
119}; 119};
120
121/*
122 * define some syscall wrappers for common architectures
123 * this is mostly for nice looks during debugging, not performance.
124 * our syscalls return < 0, not == -1, on error. which is good
125 * enough for linux aio.
126 * TODO: arm is also common nowadays, maybe even mips and x86
127 * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove...
128 */
129#if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__
130 /* the costly errno access probably kills this for size optimisation */
131
132 #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \
133 ({ \
134 long res; \
135 register unsigned long r5 __asm__ ("r8" ); \
136 register unsigned long r4 __asm__ ("r10"); \
137 register unsigned long r3 __asm__ ("rdx"); \
138 register unsigned long r2 __asm__ ("rsi"); \
139 register unsigned long r1 __asm__ ("rdi"); \
140 if (narg >= 5) r5 = (unsigned long)(arg5); \
141 if (narg >= 4) r4 = (unsigned long)(arg4); \
142 if (narg >= 3) r3 = (unsigned long)(arg3); \
143 if (narg >= 2) r2 = (unsigned long)(arg2); \
144 if (narg >= 1) r1 = (unsigned long)(arg1); \
145 __asm__ __volatile__ ( \
146 "syscall\n\t" \
147 : "=a" (res) \
148 : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \
149 : "cc", "r11", "cx", "memory"); \
150 errno = -res; \
151 res; \
152 })
153
154#endif
155
156#ifdef ev_syscall
157 #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0
158 #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0)
159 #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0)
160 #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0)
161 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0)
162 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5)
163#else
164 #define ev_syscall0(nr) syscall (nr)
165 #define ev_syscall1(nr,arg1) syscall (nr, arg1)
166 #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2)
167 #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3)
168 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4)
169 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5)
170#endif
171 120
172inline_size 121inline_size
173int 122int
174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
175{ 124{
… …
286linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
287{ 236{
288 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);
289 ANIOCBP iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
290 239
291 if (iocb->io.aio_reqprio < 0) 240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
292 { 241 {
293 /* we handed this fd over to epoll, so undo this first */ 242 /* we handed this fd over to epoll, so undo this first */
294 /* we do it manually because the optimisations on epoll_modfy won't do us any good */ 243 /* we do it manually because the optimisations on epoll_modify won't do us any good */
295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0; 245 anfds [fd].emask = 0;
297 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
298 } 247 }
248 else if (ecb_expect_false (iocb->io.aio_buf))
249 {
250 /* iocb active, so cancel it first before resubmit */
251 for (;;)
252 {
253 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
254 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
255 break;
299 256
257 if (ecb_expect_true (errno == EINPROGRESS))
258 break;
259
260 /* the EINPROGRESS test is for nicer error message. clumsy. */
261 if (errno != EINTR)
262 {
263 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
264 break;
265 }
266 }
267 }
268
300 if (iocb->io.aio_buf) 269 iocb->io.aio_buf =
301 /* io_cancel always returns some error on relevant kernels, but works */ 270 (nev & EV_READ ? POLLIN : 0)
302 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); 271 | (nev & EV_WRITE ? POLLOUT : 0);
303 272
304 if (nev) 273 if (nev)
305 { 274 {
306 iocb->io.aio_buf =
307 (nev & EV_READ ? POLLIN : 0)
308 | (nev & EV_WRITE ? POLLOUT : 0);
309
310 /* queue iocb up for io_submit */ 275 /* queue iocb up for io_submit */
311 /* this assumes we only ever get one call per fd per loop iteration */ 276 /* this assumes we only ever get one call per fd per loop iteration */
312 ++linuxaio_submitcnt; 277 ++linuxaio_submitcnt;
313 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 278 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
314 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 279 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
… …
319linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 284linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
320{ 285{
321 epoll_poll (EV_A_ 0); 286 epoll_poll (EV_A_ 0);
322} 287}
323 288
324static void 289inline_speed
290void
325linuxaio_fd_rearm (EV_P_ int fd) 291linuxaio_fd_rearm (EV_P_ int fd)
326{ 292{
327 anfds [fd].events = 0; 293 anfds [fd].events = 0;
328 linuxaio_iocbps [fd]->io.aio_buf = 0; 294 linuxaio_iocbps [fd]->io.aio_buf = 0;
329 fd_change (EV_A_ fd, EV_ANFD_REIFY); 295 fd_change (EV_A_ fd, EV_ANFD_REIFY);
… …
345 fd, 311 fd,
346 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 312 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
347 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 313 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
348 ); 314 );
349 315
350 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ 316 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
351 linuxaio_fd_rearm (EV_A_ fd); 317 linuxaio_fd_rearm (EV_A_ fd);
352 318
353 --nr; 319 --nr;
354 ++ev; 320 ++ev;
355 } 321 }
… …
358/* get any events from ring buffer, return true if any were handled */ 324/* get any events from ring buffer, return true if any were handled */
359static int 325static int
360linuxaio_get_events_from_ring (EV_P) 326linuxaio_get_events_from_ring (EV_P)
361{ 327{
362 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 328 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
329 unsigned head, tail;
363 330
364 /* the kernel reads and writes both of these variables, */ 331 /* the kernel reads and writes both of these variables, */
365 /* as a C extension, we assume that volatile use here */ 332 /* as a C extension, we assume that volatile use here */
366 /* both makes reads atomic and once-only */ 333 /* both makes reads atomic and once-only */
367 unsigned head = *(volatile unsigned *)&ring->head; 334 head = *(volatile unsigned *)&ring->head;
335 ECB_MEMORY_FENCE_ACQUIRE;
368 unsigned tail = *(volatile unsigned *)&ring->tail; 336 tail = *(volatile unsigned *)&ring->tail;
369 337
370 if (head == tail) 338 if (head == tail)
371 return 0; 339 return 0;
372 340
373 /* bail out if the ring buffer doesn't match the expected layout */
374 if (expect_false (ring->magic != AIO_RING_MAGIC)
375 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
376 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
377 return 0;
378
379 /* make sure the events up to tail are visible */
380 ECB_MEMORY_FENCE_ACQUIRE;
381
382 /* parse all available events, but only once, to avoid starvation */ 341 /* parse all available events, but only once, to avoid starvation */
383 if (tail > head) /* normal case around */ 342 if (ecb_expect_true (tail > head)) /* normal case around */
384 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 343 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
385 else /* wrapped around */ 344 else /* wrapped around */
386 { 345 {
387 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 346 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
388 linuxaio_parse_events (EV_A_ ring->io_events, tail); 347 linuxaio_parse_events (EV_A_ ring->io_events, tail);
… …
393 *(volatile unsigned *)&ring->head = tail; 352 *(volatile unsigned *)&ring->head = tail;
394 353
395 return 1; 354 return 1;
396} 355}
397 356
357inline_size
358int
359linuxaio_ringbuf_valid (EV_P)
360{
361 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
362
363 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
364 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
365 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
366}
367
398/* read at least one event from kernel, or timeout */ 368/* read at least one event from kernel, or timeout */
399inline_size 369inline_size
400void 370void
401linuxaio_get_events (EV_P_ ev_tstamp timeout) 371linuxaio_get_events (EV_P_ ev_tstamp timeout)
402{ 372{
403 struct timespec ts; 373 struct timespec ts;
404 struct io_event ioev[1]; 374 struct io_event ioev[8]; /* 256 octet stack space */
405 int res; 375 int want = 1; /* how many events to request */
376 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
406 377
378 if (ecb_expect_true (ringbuf_valid))
379 {
380 /* if the ring buffer has any events, we don't wait or call the kernel at all */
407 if (linuxaio_get_events_from_ring (EV_A)) 381 if (linuxaio_get_events_from_ring (EV_A))
408 return; 382 return;
409 383
410 /* no events, so wait for at least one, then poll ring buffer again */ 384 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
411 /* this degrades to one event per loop iteration */ 385 if (!timeout)
412 /* if the ring buffer changes layout, but so be it */ 386 return;
387 }
388 else
389 /* no ringbuffer, request slightly larger batch */
390 want = sizeof (ioev) / sizeof (ioev [0]);
413 391
392 /* no events, so wait for some
393 * for fairness reasons, we do this in a loop, to fetch all events
394 */
395 for (;;)
396 {
397 int res;
398
414 EV_RELEASE_CB; 399 EV_RELEASE_CB;
415 400
416 ts.tv_sec = (long)timeout; 401 EV_TS_SET (ts, timeout);
417 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 402 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
418 403
419 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
420
421 EV_ACQUIRE_CB; 404 EV_ACQUIRE_CB;
422 405
423 if (res < 0) 406 if (res < 0)
424 if (errno == EINTR) 407 if (errno == EINTR)
425 /* ignored */; 408 /* ignored, retry */;
426 else 409 else
427 ev_syserr ("(libev) linuxaio io_getevents"); 410 ev_syserr ("(libev) linuxaio io_getevents");
428 else if (res) 411 else if (res)
429 { 412 {
430 /* at least one event available, handle it and any remaining ones in the ring buffer */ 413 /* at least one event available, handle them */
431 linuxaio_parse_events (EV_A_ ioev, res); 414 linuxaio_parse_events (EV_A_ ioev, res);
415
416 if (ecb_expect_true (ringbuf_valid))
417 {
418 /* if we have a ring buffer, handle any remaining events in it */
432 linuxaio_get_events_from_ring (EV_A); 419 linuxaio_get_events_from_ring (EV_A);
420
421 /* at this point, we should have handled all outstanding events */
422 break;
423 }
424 else if (res < want)
425 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
426 break;
427 }
428 else
429 break; /* no events from the kernel, we are done */
430
431 timeout = 0; /* only wait in the first iteration */
433 } 432 }
434} 433}
435 434
436inline_size 435inline_size
437int 436int
… …
453 /* which allows us to pinpoint the erroneous iocb */ 452 /* which allows us to pinpoint the erroneous iocb */
454 for (submitted = 0; submitted < linuxaio_submitcnt; ) 453 for (submitted = 0; submitted < linuxaio_submitcnt; )
455 { 454 {
456 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 455 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
457 456
458 if (expect_false (res < 0)) 457 if (ecb_expect_false (res < 0))
459 if (errno == EINVAL) 458 if (errno == EINVAL)
460 { 459 {
461 /* This happens for unsupported fds, officially, but in my testing, 460 /* This happens for unsupported fds, officially, but in my testing,
462 * also randomly happens for supported fds. We fall back to good old 461 * also randomly happens for supported fds. We fall back to good old
463 * poll() here, under the assumption that this is a very rare case. 462 * poll() here, under the assumption that this is a very rare case.
… …
493 } 492 }
494 493
495 ++linuxaio_iteration; 494 ++linuxaio_iteration;
496 if (linuxaio_io_setup (EV_A) < 0) 495 if (linuxaio_io_setup (EV_A) < 0)
497 { 496 {
497 /* TODO: rearm all and recreate epoll backend from scratch */
498 /* TODO: might be more prudent? */
499
498 /* to bad, we can't get a new aio context, go 100% epoll */ 500 /* to bad, we can't get a new aio context, go 100% epoll */
499 linuxaio_free_iocbp (EV_A); 501 linuxaio_free_iocbp (EV_A);
500 ev_io_stop (EV_A_ &linuxaio_epoll_w); 502 ev_io_stop (EV_A_ &linuxaio_epoll_w);
501 ev_ref (EV_A); 503 ev_ref (EV_A);
502 linuxaio_ctx = 0; 504 linuxaio_ctx = 0;
505
506 backend = EVBACKEND_EPOLL;
503 backend_modify = epoll_modify; 507 backend_modify = epoll_modify;
504 backend_poll = epoll_poll; 508 backend_poll = epoll_poll;
505 } 509 }
506 510
507 timeout = 0; 511 timeout = 0;
508 /* it's easiest to handle this mess in another iteration */ 512 /* it's easiest to handle this mess in another iteration */
509 return; 513 return;
510 } 514 }
511 else if (errno == EBADF) 515 else if (errno == EBADF)
512 { 516 {
517 assert (("libev: event loop rejected bad fd", errno != EBADF));
513 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 518 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
514 519
515 res = 1; /* skip this iocb */ 520 res = 1; /* skip this iocb */
516 } 521 }
522 else if (errno == EINTR) /* not seen in reality, not documented */
523 res = 0; /* silently ignore and retry */
517 else 524 else
525 {
518 ev_syserr ("(libev) linuxaio io_submit"); 526 ev_syserr ("(libev) linuxaio io_submit");
527 res = 0;
528 }
519 529
520 submitted += res; 530 submitted += res;
521 } 531 }
522 532
523 linuxaio_submitcnt = 0; 533 linuxaio_submitcnt = 0;
… …
551 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 561 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
552 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 562 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
553 ev_io_start (EV_A_ &linuxaio_epoll_w); 563 ev_io_start (EV_A_ &linuxaio_epoll_w);
554 ev_unref (EV_A); /* watcher should not keep loop alive */ 564 ev_unref (EV_A); /* watcher should not keep loop alive */
555 565
556 backend_modify = linuxaio_modify; 566 backend_modify = linuxaio_modify;
557 backend_poll = linuxaio_poll; 567 backend_poll = linuxaio_poll;
558 568
559 linuxaio_iocbpmax = 0; 569 linuxaio_iocbpmax = 0;
560 linuxaio_iocbps = 0; 570 linuxaio_iocbps = 0;
561 571
562 linuxaio_submits = 0; 572 linuxaio_submits = 0;
… …
570void 580void
571linuxaio_destroy (EV_P) 581linuxaio_destroy (EV_P)
572{ 582{
573 epoll_destroy (EV_A); 583 epoll_destroy (EV_A);
574 linuxaio_free_iocbp (EV_A); 584 linuxaio_free_iocbp (EV_A);
575 evsys_io_destroy (linuxaio_ctx); 585 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
576} 586}
577 587
578inline_size 588ecb_cold
579void 589static void
580linuxaio_fork (EV_P) 590linuxaio_fork (EV_P)
581{ 591{
582 /* this frees all iocbs, which is very heavy-handed */
583 linuxaio_destroy (EV_A);
584 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 592 linuxaio_submitcnt = 0; /* all pointers were invalidated */
593 linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
594 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
585 595
586 linuxaio_iteration = 0; /* we start over in the child */ 596 linuxaio_iteration = 0; /* we start over in the child */
587 597
588 while (linuxaio_io_setup (EV_A) < 0) 598 while (linuxaio_io_setup (EV_A) < 0)
589 ev_syserr ("(libev) linuxaio io_setup"); 599 ev_syserr ("(libev) linuxaio io_setup");
590 600
601 /* forking epoll should also effectively unregister all fds from the backend */
591 epoll_fork (EV_A); 602 epoll_fork (EV_A);
603 /* epoll_fork already did this. hopefully */
604 /*fd_rearm_all (EV_A);*/
592 605
593 ev_io_stop (EV_A_ &linuxaio_epoll_w); 606 ev_io_stop (EV_A_ &linuxaio_epoll_w);
594 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); 607 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
595 ev_io_start (EV_A_ &linuxaio_epoll_w); 608 ev_io_start (EV_A_ &linuxaio_epoll_w);
596
597 /* epoll_fork already did this. hopefully */
598 /*fd_rearm_all (EV_A);*/
599} 609}
600 610

Diff Legend

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