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 1.35 by root, Wed Jun 26 00:11:24 2019 UTC vs.
Revision 1.52 by root, Wed Aug 28 10:05:23 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{
… …
263 * this is not well documented, so we better do it. 212 * this is not well documented, so we better do it.
264 */ 213 */
265 memset (iocb, 0, sizeof (*iocb)); 214 memset (iocb, 0, sizeof (*iocb));
266 215
267 iocb->io.aio_lio_opcode = IOCB_CMD_POLL; 216 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
268 iocb->io.aio_data = offset;
269 iocb->io.aio_fildes = offset; 217 iocb->io.aio_fildes = offset;
270 218
271 base [offset++] = iocb; 219 base [offset++] = iocb;
272 } 220 }
273} 221}
… …
285static void 233static void
286linuxaio_modify (EV_P_ int fd, int oev, int nev) 234linuxaio_modify (EV_P_ int fd, int oev, int nev)
287{ 235{
288 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 236 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
289 ANIOCBP iocb = linuxaio_iocbps [fd]; 237 ANIOCBP iocb = linuxaio_iocbps [fd];
238 ANFD *anfd = &anfds [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_modify 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 anfd->emask = 0;
297 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
298 } 247 }
299 248 else if (ecb_expect_false (iocb->io.aio_buf))
300 if (iocb->io.aio_buf)
301 { 249 {
250 /* iocb active, so cancel it first before resubmit */
251 /* this assumes we only ever get one call per fd per loop iteration */
252 for (;;)
253 {
254 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
302 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); 255 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
303 /* on relevant kernels, io_cancel fails with EINPROGRES if everything is fine */ 256 break;
257
258 if (ecb_expect_true (errno == EINPROGRESS))
259 break;
260
261 /* the EINPROGRESS test is for nicer error message. clumsy. */
262 if (errno != EINTR)
263 {
304 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS)); 264 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
265 break;
266 }
305 } 267 }
268
269 /* increment generation counter to avoid handling old events */
270 ++anfd->egen;
271 }
272
273 iocb->io.aio_buf =
274 (nev & EV_READ ? POLLIN : 0)
275 | (nev & EV_WRITE ? POLLOUT : 0);
306 276
307 if (nev) 277 if (nev)
308 { 278 {
309 iocb->io.aio_buf = 279 iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
310 (nev & EV_READ ? POLLIN : 0)
311 | (nev & EV_WRITE ? POLLOUT : 0);
312 280
313 /* queue iocb up for io_submit */ 281 /* queue iocb up for io_submit */
314 /* this assumes we only ever get one call per fd per loop iteration */ 282 /* this assumes we only ever get one call per fd per loop iteration */
315 ++linuxaio_submitcnt; 283 ++linuxaio_submitcnt;
316 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 284 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
… …
336static void 304static void
337linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 305linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
338{ 306{
339 while (nr) 307 while (nr)
340 { 308 {
341 int fd = ev->data; 309 int fd = ev->data & 0xffffffff;
310 uint32_t gen = ev->data >> 32;
342 int res = ev->res; 311 int res = ev->res;
343 312
344 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 313 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
345 314
315 /* only accept events if generation counter matches */
316 if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
317 {
346 /* feed events, we do not expect or handle POLLNVAL */ 318 /* feed events, we do not expect or handle POLLNVAL */
347 fd_event ( 319 fd_event (
348 EV_A_ 320 EV_A_
349 fd, 321 fd,
350 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 322 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
351 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 323 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
352 ); 324 );
353 325
354 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ 326 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
355 linuxaio_fd_rearm (EV_A_ fd); 327 linuxaio_fd_rearm (EV_A_ fd);
328 }
356 329
357 --nr; 330 --nr;
358 ++ev; 331 ++ev;
359 } 332 }
360} 333}
… …
362/* get any events from ring buffer, return true if any were handled */ 335/* get any events from ring buffer, return true if any were handled */
363static int 336static int
364linuxaio_get_events_from_ring (EV_P) 337linuxaio_get_events_from_ring (EV_P)
365{ 338{
366 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 339 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
340 unsigned head, tail;
367 341
368 /* the kernel reads and writes both of these variables, */ 342 /* the kernel reads and writes both of these variables, */
369 /* as a C extension, we assume that volatile use here */ 343 /* as a C extension, we assume that volatile use here */
370 /* both makes reads atomic and once-only */ 344 /* both makes reads atomic and once-only */
371 unsigned head = *(volatile unsigned *)&ring->head; 345 head = *(volatile unsigned *)&ring->head;
346 ECB_MEMORY_FENCE_ACQUIRE;
372 unsigned tail = *(volatile unsigned *)&ring->tail; 347 tail = *(volatile unsigned *)&ring->tail;
373 348
374 if (head == tail) 349 if (head == tail)
375 return 0; 350 return 0;
376 351
377 /* bail out if the ring buffer doesn't match the expected layout */
378 if (expect_false (ring->magic != AIO_RING_MAGIC)
379 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
380 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
381 return 0;
382
383 /* make sure the events up to tail are visible */
384 ECB_MEMORY_FENCE_ACQUIRE;
385
386 /* parse all available events, but only once, to avoid starvation */ 352 /* parse all available events, but only once, to avoid starvation */
387 if (tail > head) /* normal case around */ 353 if (ecb_expect_true (tail > head)) /* normal case around */
388 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 354 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
389 else /* wrapped around */ 355 else /* wrapped around */
390 { 356 {
391 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 357 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
392 linuxaio_parse_events (EV_A_ ring->io_events, tail); 358 linuxaio_parse_events (EV_A_ ring->io_events, tail);
… …
397 *(volatile unsigned *)&ring->head = tail; 363 *(volatile unsigned *)&ring->head = tail;
398 364
399 return 1; 365 return 1;
400} 366}
401 367
368inline_size
369int
370linuxaio_ringbuf_valid (EV_P)
371{
372 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
373
374 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
375 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
376 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
377}
378
402/* read at least one event from kernel, or timeout */ 379/* read at least one event from kernel, or timeout */
403inline_size 380inline_size
404void 381void
405linuxaio_get_events (EV_P_ ev_tstamp timeout) 382linuxaio_get_events (EV_P_ ev_tstamp timeout)
406{ 383{
407 struct timespec ts; 384 struct timespec ts;
408 struct io_event ioev[1]; 385 struct io_event ioev[8]; /* 256 octet stack space */
409 int res; 386 int want = 1; /* how many events to request */
387 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
410 388
389 if (ecb_expect_true (ringbuf_valid))
390 {
391 /* if the ring buffer has any events, we don't wait or call the kernel at all */
411 if (linuxaio_get_events_from_ring (EV_A)) 392 if (linuxaio_get_events_from_ring (EV_A))
412 return; 393 return;
413 394
414 /* no events, so wait for at least one, then poll ring buffer again */ 395 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
415 /* this degrades to one event per loop iteration */ 396 if (!timeout)
416 /* if the ring buffer changes layout, but so be it */ 397 return;
398 }
399 else
400 /* no ringbuffer, request slightly larger batch */
401 want = sizeof (ioev) / sizeof (ioev [0]);
417 402
403 /* no events, so wait for some
404 * for fairness reasons, we do this in a loop, to fetch all events
405 */
406 for (;;)
407 {
408 int res;
409
418 EV_RELEASE_CB; 410 EV_RELEASE_CB;
419 411
420 ts.tv_sec = (long)timeout; 412 EV_TS_SET (ts, timeout);
421 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 413 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
422 414
423 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
424
425 EV_ACQUIRE_CB; 415 EV_ACQUIRE_CB;
426 416
427 if (res < 0) 417 if (res < 0)
428 if (errno == EINTR) 418 if (errno == EINTR)
429 /* ignored */; 419 /* ignored, retry */;
430 else 420 else
431 ev_syserr ("(libev) linuxaio io_getevents"); 421 ev_syserr ("(libev) linuxaio io_getevents");
432 else if (res) 422 else if (res)
433 { 423 {
434 /* at least one event available, handle it and any remaining ones in the ring buffer */ 424 /* at least one event available, handle them */
435 linuxaio_parse_events (EV_A_ ioev, res); 425 linuxaio_parse_events (EV_A_ ioev, res);
426
427 if (ecb_expect_true (ringbuf_valid))
428 {
429 /* if we have a ring buffer, handle any remaining events in it */
436 linuxaio_get_events_from_ring (EV_A); 430 linuxaio_get_events_from_ring (EV_A);
431
432 /* at this point, we should have handled all outstanding events */
433 break;
434 }
435 else if (res < want)
436 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
437 break;
438 }
439 else
440 break; /* no events from the kernel, we are done */
441
442 timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
437 } 443 }
438} 444}
439 445
440inline_size 446inline_size
441int 447int
… …
457 /* which allows us to pinpoint the erroneous iocb */ 463 /* which allows us to pinpoint the erroneous iocb */
458 for (submitted = 0; submitted < linuxaio_submitcnt; ) 464 for (submitted = 0; submitted < linuxaio_submitcnt; )
459 { 465 {
460 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 466 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
461 467
462 if (expect_false (res < 0)) 468 if (ecb_expect_false (res < 0))
463 if (errno == EINVAL) 469 if (errno == EINVAL)
464 { 470 {
465 /* This happens for unsupported fds, officially, but in my testing, 471 /* This happens for unsupported fds, officially, but in my testing,
466 * also randomly happens for supported fds. We fall back to good old 472 * also randomly happens for supported fds. We fall back to good old
467 * poll() here, under the assumption that this is a very rare case. 473 * poll() here, under the assumption that this is a very rare case.
… …
497 } 503 }
498 504
499 ++linuxaio_iteration; 505 ++linuxaio_iteration;
500 if (linuxaio_io_setup (EV_A) < 0) 506 if (linuxaio_io_setup (EV_A) < 0)
501 { 507 {
508 /* TODO: rearm all and recreate epoll backend from scratch */
509 /* TODO: might be more prudent? */
510
502 /* to bad, we can't get a new aio context, go 100% epoll */ 511 /* to bad, we can't get a new aio context, go 100% epoll */
503 linuxaio_free_iocbp (EV_A); 512 linuxaio_free_iocbp (EV_A);
504 ev_io_stop (EV_A_ &linuxaio_epoll_w); 513 ev_io_stop (EV_A_ &linuxaio_epoll_w);
505 ev_ref (EV_A); 514 ev_ref (EV_A);
506 linuxaio_ctx = 0; 515 linuxaio_ctx = 0;
516
517 backend = EVBACKEND_EPOLL;
507 backend_modify = epoll_modify; 518 backend_modify = epoll_modify;
508 backend_poll = epoll_poll; 519 backend_poll = epoll_poll;
509 } 520 }
510 521
511 timeout = 0; 522 timeout = EV_TS_CONST (0.);
512 /* it's easiest to handle this mess in another iteration */ 523 /* it's easiest to handle this mess in another iteration */
513 return; 524 return;
514 } 525 }
515 else if (errno == EBADF) 526 else if (errno == EBADF)
516 { 527 {
517 assert (("libev: event loop rejected bad fd", errno != EBADF)); 528 assert (("libev: event loop rejected bad fd", errno != EBADF));
518 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 529 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
519 530
520 res = 1; /* skip this iocb */ 531 res = 1; /* skip this iocb */
521 } 532 }
533 else if (errno == EINTR) /* not seen in reality, not documented */
534 res = 0; /* silently ignore and retry */
522 else 535 else
536 {
523 ev_syserr ("(libev) linuxaio io_submit"); 537 ev_syserr ("(libev) linuxaio io_submit");
538 res = 0;
539 }
524 540
525 submitted += res; 541 submitted += res;
526 } 542 }
527 543
528 linuxaio_submitcnt = 0; 544 linuxaio_submitcnt = 0;
… …
551 { 567 {
552 epoll_destroy (EV_A); 568 epoll_destroy (EV_A);
553 return 0; 569 return 0;
554 } 570 }
555 571
556 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 572 ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
557 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 573 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
558 ev_io_start (EV_A_ &linuxaio_epoll_w); 574 ev_io_start (EV_A_ &linuxaio_epoll_w);
559 ev_unref (EV_A); /* watcher should not keep loop alive */ 575 ev_unref (EV_A); /* watcher should not keep loop alive */
560 576
561 backend_modify = linuxaio_modify; 577 backend_modify = linuxaio_modify;
562 backend_poll = linuxaio_poll; 578 backend_poll = linuxaio_poll;
563 579
564 linuxaio_iocbpmax = 0; 580 linuxaio_iocbpmax = 0;
565 linuxaio_iocbps = 0; 581 linuxaio_iocbps = 0;
566 582
567 linuxaio_submits = 0; 583 linuxaio_submits = 0;
… …
578 epoll_destroy (EV_A); 594 epoll_destroy (EV_A);
579 linuxaio_free_iocbp (EV_A); 595 linuxaio_free_iocbp (EV_A);
580 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ 596 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
581} 597}
582 598
583inline_size 599ecb_cold
584void 600static void
585linuxaio_fork (EV_P) 601linuxaio_fork (EV_P)
586{ 602{
587 /* this frees all iocbs, which is very heavy-handed */
588 linuxaio_destroy (EV_A);
589 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 603 linuxaio_submitcnt = 0; /* all pointers were invalidated */
604 linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
605 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
590 606
591 linuxaio_iteration = 0; /* we start over in the child */ 607 linuxaio_iteration = 0; /* we start over in the child */
592 608
593 while (linuxaio_io_setup (EV_A) < 0) 609 while (linuxaio_io_setup (EV_A) < 0)
594 ev_syserr ("(libev) linuxaio io_setup"); 610 ev_syserr ("(libev) linuxaio io_setup");
595 611
596 /* forking epoll should also effectively unregister all fds from the backend */ 612 /* forking epoll should also effectively unregister all fds from the backend */
597 epoll_fork (EV_A); 613 epoll_fork (EV_A);
614 /* epoll_fork already did this. hopefully */
615 /*fd_rearm_all (EV_A);*/
598 616
599 ev_io_stop (EV_A_ &linuxaio_epoll_w); 617 ev_io_stop (EV_A_ &linuxaio_epoll_w);
600 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); 618 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
601 ev_io_start (EV_A_ &linuxaio_epoll_w); 619 ev_io_start (EV_A_ &linuxaio_epoll_w);
602
603 /* epoll_fork already did this. hopefully */
604 /*fd_rearm_all (EV_A);*/
605} 620}
606 621

Diff Legend

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