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 1.40 by root, Mon Jul 1 20:35:17 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 */
302 for (;;) 252 for (;;)
303 { 253 {
304 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ 254 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
305 if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) 255 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
306 break; 256 break;
307 257
308 if (expect_true (errno == EINPROGRESS)) 258 if (ecb_expect_true (errno == EINPROGRESS))
309 break; 259 break;
310 260
311 /* the EINPROGRESS test is for nicer error message. clumsy. */ 261 /* the EINPROGRESS test is for nicer error message. clumsy. */
262 if (errno != EINTR)
263 {
312 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); 264 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
265 break;
266 }
313 } 267 }
268
269 /* increment generation counter to avoid handling old events */
270 ++anfd->egen;
314 } 271 }
272
273 iocb->io.aio_buf = (nev & EV_READ ? POLLIN : 0)
274 | (nev & EV_WRITE ? POLLOUT : 0);
315 275
316 if (nev) 276 if (nev)
317 { 277 {
318 iocb->io.aio_buf = 278 iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
319 (nev & EV_READ ? POLLIN : 0)
320 | (nev & EV_WRITE ? POLLOUT : 0);
321 279
322 /* queue iocb up for io_submit */ 280 /* queue iocb up for io_submit */
323 /* 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 */
324 ++linuxaio_submitcnt; 282 ++linuxaio_submitcnt;
325 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);
… …
345static void 303static void
346linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 304linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
347{ 305{
348 while (nr) 306 while (nr)
349 { 307 {
350 int fd = ev->data; 308 int fd = ev->data & 0xffffffff;
309 uint32_t gen = ev->data >> 32;
351 int res = ev->res; 310 int res = ev->res;
352 311
353 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));
354 313
314 /* only accept events if generation counter matches */
315 if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
316 {
355 /* feed events, we do not expect or handle POLLNVAL */ 317 /* feed events, we do not expect or handle POLLNVAL */
356 fd_event ( 318 fd_event (
357 EV_A_ 319 EV_A_
358 fd, 320 fd,
359 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 321 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
360 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 322 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
361 ); 323 );
362 324
363 /* 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 */
364 linuxaio_fd_rearm (EV_A_ fd); 326 linuxaio_fd_rearm (EV_A_ fd);
327 }
365 328
366 --nr; 329 --nr;
367 ++ev; 330 ++ev;
368 } 331 }
369} 332}
… …
371/* 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 */
372static int 335static int
373linuxaio_get_events_from_ring (EV_P) 336linuxaio_get_events_from_ring (EV_P)
374{ 337{
375 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 338 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
339 unsigned head, tail;
376 340
377 /* the kernel reads and writes both of these variables, */ 341 /* the kernel reads and writes both of these variables, */
378 /* as a C extension, we assume that volatile use here */ 342 /* as a C extension, we assume that volatile use here */
379 /* both makes reads atomic and once-only */ 343 /* both makes reads atomic and once-only */
380 unsigned head = *(volatile unsigned *)&ring->head; 344 head = *(volatile unsigned *)&ring->head;
345 ECB_MEMORY_FENCE_ACQUIRE;
381 unsigned tail = *(volatile unsigned *)&ring->tail; 346 tail = *(volatile unsigned *)&ring->tail;
382 347
383 if (head == tail) 348 if (head == tail)
384 return 0; 349 return 0;
385 350
386 /* make sure the events up to tail are visible */
387 ECB_MEMORY_FENCE_ACQUIRE;
388
389 /* parse all available events, but only once, to avoid starvation */ 351 /* parse all available events, but only once, to avoid starvation */
390 if (tail > head) /* normal case around */ 352 if (ecb_expect_true (tail > head)) /* normal case around */
391 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 353 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
392 else /* wrapped around */ 354 else /* wrapped around */
393 { 355 {
394 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);
395 linuxaio_parse_events (EV_A_ ring->io_events, tail); 357 linuxaio_parse_events (EV_A_ ring->io_events, tail);
… …
406int 368int
407linuxaio_ringbuf_valid (EV_P) 369linuxaio_ringbuf_valid (EV_P)
408{ 370{
409 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 371 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
410 372
411 return expect_true (ring->magic == AIO_RING_MAGIC) 373 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
412 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES 374 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
413 && 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]? */
414} 376}
415 377
416/* read at least one event from kernel, or timeout */ 378/* read at least one event from kernel, or timeout */
… …
421 struct timespec ts; 383 struct timespec ts;
422 struct io_event ioev[8]; /* 256 octet stack space */ 384 struct io_event ioev[8]; /* 256 octet stack space */
423 int want = 1; /* how many events to request */ 385 int want = 1; /* how many events to request */
424 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); 386 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
425 387
426 if (expect_true (ringbuf_valid)) 388 if (ecb_expect_true (ringbuf_valid))
427 { 389 {
428 /* 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 */
429 if (linuxaio_get_events_from_ring (EV_A)) 391 if (linuxaio_get_events_from_ring (EV_A))
430 return; 392 return;
431 393
… …
444 { 406 {
445 int res; 407 int res;
446 408
447 EV_RELEASE_CB; 409 EV_RELEASE_CB;
448 410
449 ts.tv_sec = (long)timeout; 411 EV_TS_SET (ts, timeout);
450 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
451
452 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); 412 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
453 413
454 EV_ACQUIRE_CB; 414 EV_ACQUIRE_CB;
455 415
456 if (res < 0) 416 if (res < 0)
… …
461 else if (res) 421 else if (res)
462 { 422 {
463 /* at least one event available, handle them */ 423 /* at least one event available, handle them */
464 linuxaio_parse_events (EV_A_ ioev, res); 424 linuxaio_parse_events (EV_A_ ioev, res);
465 425
466 if (expect_true (ringbuf_valid)) 426 if (ecb_expect_true (ringbuf_valid))
467 { 427 {
468 /* 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 */
469 linuxaio_get_events_from_ring (EV_A); 429 linuxaio_get_events_from_ring (EV_A);
470 430
471 /* at this point, we should have handled all outstanding events */ 431 /* at this point, we should have handled all outstanding events */
… …
476 break; 436 break;
477 } 437 }
478 else 438 else
479 break; /* no events from the kernel, we are done */ 439 break; /* no events from the kernel, we are done */
480 440
481 timeout = 0; /* only wait in the first iteration */ 441 timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
482 } 442 }
483} 443}
484 444
485inline_size 445inline_size
486int 446int
… …
502 /* which allows us to pinpoint the erroneous iocb */ 462 /* which allows us to pinpoint the erroneous iocb */
503 for (submitted = 0; submitted < linuxaio_submitcnt; ) 463 for (submitted = 0; submitted < linuxaio_submitcnt; )
504 { 464 {
505 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);
506 466
507 if (expect_false (res < 0)) 467 if (ecb_expect_false (res < 0))
508 if (errno == EINVAL) 468 if (errno == EINVAL)
509 { 469 {
510 /* This happens for unsupported fds, officially, but in my testing, 470 /* This happens for unsupported fds, officially, but in my testing,
511 * 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
512 * 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.
… …
542 } 502 }
543 503
544 ++linuxaio_iteration; 504 ++linuxaio_iteration;
545 if (linuxaio_io_setup (EV_A) < 0) 505 if (linuxaio_io_setup (EV_A) < 0)
546 { 506 {
507 /* TODO: rearm all and recreate epoll backend from scratch */
508 /* TODO: might be more prudent? */
509
547 /* 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 */
548 linuxaio_free_iocbp (EV_A); 511 linuxaio_free_iocbp (EV_A);
549 ev_io_stop (EV_A_ &linuxaio_epoll_w); 512 ev_io_stop (EV_A_ &linuxaio_epoll_w);
550 ev_ref (EV_A); 513 ev_ref (EV_A);
551 linuxaio_ctx = 0; 514 linuxaio_ctx = 0;
515
516 backend = EVBACKEND_EPOLL;
552 backend_modify = epoll_modify; 517 backend_modify = epoll_modify;
553 backend_poll = epoll_poll; 518 backend_poll = epoll_poll;
554 } 519 }
555 520
556 timeout = 0; 521 timeout = EV_TS_CONST (0.);
557 /* it's easiest to handle this mess in another iteration */ 522 /* it's easiest to handle this mess in another iteration */
558 return; 523 return;
559 } 524 }
560 else if (errno == EBADF) 525 else if (errno == EBADF)
561 { 526 {
… …
565 res = 1; /* skip this iocb */ 530 res = 1; /* skip this iocb */
566 } 531 }
567 else if (errno == EINTR) /* not seen in reality, not documented */ 532 else if (errno == EINTR) /* not seen in reality, not documented */
568 res = 0; /* silently ignore and retry */ 533 res = 0; /* silently ignore and retry */
569 else 534 else
535 {
570 ev_syserr ("(libev) linuxaio io_submit"); 536 ev_syserr ("(libev) linuxaio io_submit");
537 res = 0;
538 }
571 539
572 submitted += res; 540 submitted += res;
573 } 541 }
574 542
575 linuxaio_submitcnt = 0; 543 linuxaio_submitcnt = 0;
… …
598 { 566 {
599 epoll_destroy (EV_A); 567 epoll_destroy (EV_A);
600 return 0; 568 return 0;
601 } 569 }
602 570
603 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);
604 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 572 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
605 ev_io_start (EV_A_ &linuxaio_epoll_w); 573 ev_io_start (EV_A_ &linuxaio_epoll_w);
606 ev_unref (EV_A); /* watcher should not keep loop alive */ 574 ev_unref (EV_A); /* watcher should not keep loop alive */
607 575
608 backend_modify = linuxaio_modify; 576 backend_modify = linuxaio_modify;
609 backend_poll = linuxaio_poll; 577 backend_poll = linuxaio_poll;
610 578
611 linuxaio_iocbpmax = 0; 579 linuxaio_iocbpmax = 0;
612 linuxaio_iocbps = 0; 580 linuxaio_iocbps = 0;
613 581
614 linuxaio_submits = 0; 582 linuxaio_submits = 0;
… …
625 epoll_destroy (EV_A); 593 epoll_destroy (EV_A);
626 linuxaio_free_iocbp (EV_A); 594 linuxaio_free_iocbp (EV_A);
627 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 */
628} 596}
629 597
630inline_size 598ecb_cold
631void 599static void
632linuxaio_fork (EV_P) 600linuxaio_fork (EV_P)
633{ 601{
634 /* this frees all iocbs, which is very heavy-handed */
635 linuxaio_destroy (EV_A);
636 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 */
637 605
638 linuxaio_iteration = 0; /* we start over in the child */ 606 linuxaio_iteration = 0; /* we start over in the child */
639 607
640 while (linuxaio_io_setup (EV_A) < 0) 608 while (linuxaio_io_setup (EV_A) < 0)
641 ev_syserr ("(libev) linuxaio io_setup"); 609 ev_syserr ("(libev) linuxaio io_setup");
642 610
643 /* forking epoll should also effectively unregister all fds from the backend */ 611 /* forking epoll should also effectively unregister all fds from the backend */
644 epoll_fork (EV_A); 612 epoll_fork (EV_A);
613 /* epoll_fork already did this. hopefully */
614 /*fd_rearm_all (EV_A);*/
645 615
646 ev_io_stop (EV_A_ &linuxaio_epoll_w); 616 ev_io_stop (EV_A_ &linuxaio_epoll_w);
647 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);
648 ev_io_start (EV_A_ &linuxaio_epoll_w); 618 ev_io_start (EV_A_ &linuxaio_epoll_w);
649
650 /* epoll_fork already did this. hopefully */
651 /*fd_rearm_all (EV_A);*/
652} 619}
653 620

Diff Legend

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