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 1.37 by root, Wed Jun 26 07:20:09 2019 UTC vs.
Revision 1.53 by root, Fri Dec 27 16:12:55 2019 UTC

… …
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};
120 120
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
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{
176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); 125 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
… …
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 = (nev & EV_READ ? POLLIN : 0)
274 | (nev & EV_WRITE ? POLLOUT : 0);
306 275
307 if (nev) 276 if (nev)
308 { 277 {
309 iocb->io.aio_buf = 278 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);
312 279
313 /* queue iocb up for io_submit */ 280 /* queue iocb up for io_submit */
314 /* this assumes we only ever get one call per fd per loop iteration */ 281 /* this assumes we only ever get one call per fd per loop iteration */
315 ++linuxaio_submitcnt; 282 ++linuxaio_submitcnt;
316 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 283 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
… …
336static void 303static void
337linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 304linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
338{ 305{
339 while (nr) 306 while (nr)
340 { 307 {
341 int fd = ev->data; 308 int fd = ev->data & 0xffffffff;
309 uint32_t gen = ev->data >> 32;
342 int res = ev->res; 310 int res = ev->res;
343 311
344 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 312 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
345 313
314 /* only accept events if generation counter matches */
315 if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
316 {
346 /* feed events, we do not expect or handle POLLNVAL */ 317 /* feed events, we do not expect or handle POLLNVAL */
347 fd_event ( 318 fd_event (
348 EV_A_ 319 EV_A_
349 fd, 320 fd,
350 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 321 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
351 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 322 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
352 ); 323 );
353 324
354 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ 325 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
355 linuxaio_fd_rearm (EV_A_ fd); 326 linuxaio_fd_rearm (EV_A_ fd);
327 }
356 328
357 --nr; 329 --nr;
358 ++ev; 330 ++ev;
359 } 331 }
360} 332}
… …
362/* get any events from ring buffer, return true if any were handled */ 334/* get any events from ring buffer, return true if any were handled */
363static int 335static int
364linuxaio_get_events_from_ring (EV_P) 336linuxaio_get_events_from_ring (EV_P)
365{ 337{
366 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 338 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
339 unsigned head, tail;
367 340
368 /* the kernel reads and writes both of these variables, */ 341 /* the kernel reads and writes both of these variables, */
369 /* as a C extension, we assume that volatile use here */ 342 /* as a C extension, we assume that volatile use here */
370 /* both makes reads atomic and once-only */ 343 /* both makes reads atomic and once-only */
371 unsigned head = *(volatile unsigned *)&ring->head; 344 head = *(volatile unsigned *)&ring->head;
345 ECB_MEMORY_FENCE_ACQUIRE;
372 unsigned tail = *(volatile unsigned *)&ring->tail; 346 tail = *(volatile unsigned *)&ring->tail;
373 347
374 if (head == tail) 348 if (head == tail)
375 return 0; 349 return 0;
376 350
377 /* make sure the events up to tail are visible */
378 ECB_MEMORY_FENCE_ACQUIRE;
379
380 /* parse all available events, but only once, to avoid starvation */ 351 /* parse all available events, but only once, to avoid starvation */
381 if (tail > head) /* normal case around */ 352 if (ecb_expect_true (tail > head)) /* normal case around */
382 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 353 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
383 else /* wrapped around */ 354 else /* wrapped around */
384 { 355 {
385 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 356 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
386 linuxaio_parse_events (EV_A_ ring->io_events, tail); 357 linuxaio_parse_events (EV_A_ ring->io_events, tail);
… …
397int 368int
398linuxaio_ringbuf_valid (EV_P) 369linuxaio_ringbuf_valid (EV_P)
399{ 370{
400 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 371 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
401 372
402 return expect_true (ring->magic == AIO_RING_MAGIC) 373 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
403 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES 374 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
404 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ 375 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
405} 376}
406 377
407/* read at least one event from kernel, or timeout */ 378/* read at least one event from kernel, or timeout */
… …
412 struct timespec ts; 383 struct timespec ts;
413 struct io_event ioev[8]; /* 256 octet stack space */ 384 struct io_event ioev[8]; /* 256 octet stack space */
414 int want = 1; /* how many events to request */ 385 int want = 1; /* how many events to request */
415 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); 386 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
416 387
417 if (expect_true (ringbuf_valid)) 388 if (ecb_expect_true (ringbuf_valid))
418 { 389 {
419 /* if the ring buffer has any events, we don't wait or call the kernel at all */ 390 /* if the ring buffer has any events, we don't wait or call the kernel at all */
420 if (linuxaio_get_events_from_ring (EV_A)) 391 if (linuxaio_get_events_from_ring (EV_A))
421 return; 392 return;
422 393
… …
435 { 406 {
436 int res; 407 int res;
437 408
438 EV_RELEASE_CB; 409 EV_RELEASE_CB;
439 410
440 ts.tv_sec = (long)timeout; 411 EV_TS_SET (ts, timeout);
441 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
442
443 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); 412 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
444 413
445 EV_ACQUIRE_CB; 414 EV_ACQUIRE_CB;
446 415
447 if (res < 0) 416 if (res < 0)
… …
452 else if (res) 421 else if (res)
453 { 422 {
454 /* at least one event available, handle them */ 423 /* at least one event available, handle them */
455 linuxaio_parse_events (EV_A_ ioev, res); 424 linuxaio_parse_events (EV_A_ ioev, res);
456 425
457 if (expect_true (ringbuf_valid)) 426 if (ecb_expect_true (ringbuf_valid))
458 { 427 {
459 /* if we have a ring buffer, handle any remaining events in it */ 428 /* if we have a ring buffer, handle any remaining events in it */
460 linuxaio_get_events_from_ring (EV_A); 429 linuxaio_get_events_from_ring (EV_A);
461 430
462 /* at this point, we should have handled all outstanding events */ 431 /* at this point, we should have handled all outstanding events */
… …
467 break; 436 break;
468 } 437 }
469 else 438 else
470 break; /* no events from the kernel, we are done */ 439 break; /* no events from the kernel, we are done */
471 440
472 timeout = 0; /* only wait in the first iteration */ 441 timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
473 } 442 }
474} 443}
475 444
476inline_size 445inline_size
477int 446int
… …
493 /* which allows us to pinpoint the erroneous iocb */ 462 /* which allows us to pinpoint the erroneous iocb */
494 for (submitted = 0; submitted < linuxaio_submitcnt; ) 463 for (submitted = 0; submitted < linuxaio_submitcnt; )
495 { 464 {
496 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 465 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
497 466
498 if (expect_false (res < 0)) 467 if (ecb_expect_false (res < 0))
499 if (errno == EINVAL) 468 if (errno == EINVAL)
500 { 469 {
501 /* This happens for unsupported fds, officially, but in my testing, 470 /* This happens for unsupported fds, officially, but in my testing,
502 * also randomly happens for supported fds. We fall back to good old 471 * also randomly happens for supported fds. We fall back to good old
503 * poll() here, under the assumption that this is a very rare case. 472 * poll() here, under the assumption that this is a very rare case.
… …
533 } 502 }
534 503
535 ++linuxaio_iteration; 504 ++linuxaio_iteration;
536 if (linuxaio_io_setup (EV_A) < 0) 505 if (linuxaio_io_setup (EV_A) < 0)
537 { 506 {
507 /* TODO: rearm all and recreate epoll backend from scratch */
508 /* TODO: might be more prudent? */
509
538 /* to bad, we can't get a new aio context, go 100% epoll */ 510 /* to bad, we can't get a new aio context, go 100% epoll */
539 linuxaio_free_iocbp (EV_A); 511 linuxaio_free_iocbp (EV_A);
540 ev_io_stop (EV_A_ &linuxaio_epoll_w); 512 ev_io_stop (EV_A_ &linuxaio_epoll_w);
541 ev_ref (EV_A); 513 ev_ref (EV_A);
542 linuxaio_ctx = 0; 514 linuxaio_ctx = 0;
515
516 backend = EVBACKEND_EPOLL;
543 backend_modify = epoll_modify; 517 backend_modify = epoll_modify;
544 backend_poll = epoll_poll; 518 backend_poll = epoll_poll;
545 } 519 }
546 520
547 timeout = 0; 521 timeout = EV_TS_CONST (0.);
548 /* it's easiest to handle this mess in another iteration */ 522 /* it's easiest to handle this mess in another iteration */
549 return; 523 return;
550 } 524 }
551 else if (errno == EBADF) 525 else if (errno == EBADF)
552 { 526 {
553 assert (("libev: event loop rejected bad fd", errno != EBADF)); 527 assert (("libev: event loop rejected bad fd", errno != EBADF));
554 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 528 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
555 529
556 res = 1; /* skip this iocb */ 530 res = 1; /* skip this iocb */
557 } 531 }
532 else if (errno == EINTR) /* not seen in reality, not documented */
533 res = 0; /* silently ignore and retry */
558 else 534 else
535 {
559 ev_syserr ("(libev) linuxaio io_submit"); 536 ev_syserr ("(libev) linuxaio io_submit");
537 res = 0;
538 }
560 539
561 submitted += res; 540 submitted += res;
562 } 541 }
563 542
564 linuxaio_submitcnt = 0; 543 linuxaio_submitcnt = 0;
… …
587 { 566 {
588 epoll_destroy (EV_A); 567 epoll_destroy (EV_A);
589 return 0; 568 return 0;
590 } 569 }
591 570
592 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 571 ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
593 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 572 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
594 ev_io_start (EV_A_ &linuxaio_epoll_w); 573 ev_io_start (EV_A_ &linuxaio_epoll_w);
595 ev_unref (EV_A); /* watcher should not keep loop alive */ 574 ev_unref (EV_A); /* watcher should not keep loop alive */
596 575
597 backend_modify = linuxaio_modify; 576 backend_modify = linuxaio_modify;
598 backend_poll = linuxaio_poll; 577 backend_poll = linuxaio_poll;
599 578
600 linuxaio_iocbpmax = 0; 579 linuxaio_iocbpmax = 0;
601 linuxaio_iocbps = 0; 580 linuxaio_iocbps = 0;
602 581
603 linuxaio_submits = 0; 582 linuxaio_submits = 0;
… …
614 epoll_destroy (EV_A); 593 epoll_destroy (EV_A);
615 linuxaio_free_iocbp (EV_A); 594 linuxaio_free_iocbp (EV_A);
616 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ 595 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
617} 596}
618 597
619inline_size 598ecb_cold
620void 599static void
621linuxaio_fork (EV_P) 600linuxaio_fork (EV_P)
622{ 601{
623 /* this frees all iocbs, which is very heavy-handed */
624 linuxaio_destroy (EV_A);
625 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 602 linuxaio_submitcnt = 0; /* all pointers were invalidated */
603 linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
604 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
626 605
627 linuxaio_iteration = 0; /* we start over in the child */ 606 linuxaio_iteration = 0; /* we start over in the child */
628 607
629 while (linuxaio_io_setup (EV_A) < 0) 608 while (linuxaio_io_setup (EV_A) < 0)
630 ev_syserr ("(libev) linuxaio io_setup"); 609 ev_syserr ("(libev) linuxaio io_setup");
631 610
632 /* forking epoll should also effectively unregister all fds from the backend */ 611 /* forking epoll should also effectively unregister all fds from the backend */
633 epoll_fork (EV_A); 612 epoll_fork (EV_A);
613 /* epoll_fork already did this. hopefully */
614 /*fd_rearm_all (EV_A);*/
634 615
635 ev_io_stop (EV_A_ &linuxaio_epoll_w); 616 ev_io_stop (EV_A_ &linuxaio_epoll_w);
636 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); 617 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
637 ev_io_start (EV_A_ &linuxaio_epoll_w); 618 ev_io_start (EV_A_ &linuxaio_epoll_w);
638
639 /* epoll_fork already did this. hopefully */
640 /*fd_rearm_all (EV_A);*/
641} 619}
642 620

Diff Legend

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