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 1.49 by root, Tue Jul 9 00:04:35 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 =
274 (nev & EV_READ ? POLLIN : 0)
275 | (nev & EV_WRITE ? POLLOUT : 0);
315 276
316 if (nev) 277 if (nev)
317 { 278 {
318 iocb->io.aio_buf = 279 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);
321 280
322 /* queue iocb up for io_submit */ 281 /* queue iocb up for io_submit */
323 /* 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 */
324 ++linuxaio_submitcnt; 283 ++linuxaio_submitcnt;
325 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);
… …
345static void 304static void
346linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 305linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
347{ 306{
348 while (nr) 307 while (nr)
349 { 308 {
350 int fd = ev->data; 309 int fd = ev->data & 0xffffffff;
310 uint32_t gen = ev->data >> 32;
351 int res = ev->res; 311 int res = ev->res;
352 312
353 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));
314
315 /* ignore event if generation doesn't match */
316 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
317 continue;
354 318
355 /* feed events, we do not expect or handle POLLNVAL */ 319 /* feed events, we do not expect or handle POLLNVAL */
356 fd_event ( 320 fd_event (
357 EV_A_ 321 EV_A_
358 fd, 322 fd,
… …
371/* 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 */
372static int 336static int
373linuxaio_get_events_from_ring (EV_P) 337linuxaio_get_events_from_ring (EV_P)
374{ 338{
375 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 339 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
340 unsigned head, tail;
376 341
377 /* the kernel reads and writes both of these variables, */ 342 /* the kernel reads and writes both of these variables, */
378 /* as a C extension, we assume that volatile use here */ 343 /* as a C extension, we assume that volatile use here */
379 /* both makes reads atomic and once-only */ 344 /* both makes reads atomic and once-only */
380 unsigned head = *(volatile unsigned *)&ring->head; 345 head = *(volatile unsigned *)&ring->head;
346 ECB_MEMORY_FENCE_ACQUIRE;
381 unsigned tail = *(volatile unsigned *)&ring->tail; 347 tail = *(volatile unsigned *)&ring->tail;
382 348
383 if (head == tail) 349 if (head == tail)
384 return 0; 350 return 0;
385 351
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 */ 352 /* parse all available events, but only once, to avoid starvation */
390 if (tail > head) /* normal case around */ 353 if (ecb_expect_true (tail > head)) /* normal case around */
391 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 354 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
392 else /* wrapped around */ 355 else /* wrapped around */
393 { 356 {
394 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);
395 linuxaio_parse_events (EV_A_ ring->io_events, tail); 358 linuxaio_parse_events (EV_A_ ring->io_events, tail);
… …
406int 369int
407linuxaio_ringbuf_valid (EV_P) 370linuxaio_ringbuf_valid (EV_P)
408{ 371{
409 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 372 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
410 373
411 return expect_true (ring->magic == AIO_RING_MAGIC) 374 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
412 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES 375 && 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]? */ 376 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
414} 377}
415 378
416/* read at least one event from kernel, or timeout */ 379/* read at least one event from kernel, or timeout */
… …
421 struct timespec ts; 384 struct timespec ts;
422 struct io_event ioev[8]; /* 256 octet stack space */ 385 struct io_event ioev[8]; /* 256 octet stack space */
423 int want = 1; /* how many events to request */ 386 int want = 1; /* how many events to request */
424 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); 387 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
425 388
426 if (expect_true (ringbuf_valid)) 389 if (ecb_expect_true (ringbuf_valid))
427 { 390 {
428 /* if the ring buffer has any events, we don't wait or call the kernel at all */ 391 /* 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)) 392 if (linuxaio_get_events_from_ring (EV_A))
430 return; 393 return;
431 394
… …
444 { 407 {
445 int res; 408 int res;
446 409
447 EV_RELEASE_CB; 410 EV_RELEASE_CB;
448 411
449 ts.tv_sec = (long)timeout; 412 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); 413 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
453 414
454 EV_ACQUIRE_CB; 415 EV_ACQUIRE_CB;
455 416
456 if (res < 0) 417 if (res < 0)
… …
461 else if (res) 422 else if (res)
462 { 423 {
463 /* at least one event available, handle them */ 424 /* at least one event available, handle them */
464 linuxaio_parse_events (EV_A_ ioev, res); 425 linuxaio_parse_events (EV_A_ ioev, res);
465 426
466 if (expect_true (ringbuf_valid)) 427 if (ecb_expect_true (ringbuf_valid))
467 { 428 {
468 /* if we have a ring buffer, handle any remaining events in it */ 429 /* if we have a ring buffer, handle any remaining events in it */
469 linuxaio_get_events_from_ring (EV_A); 430 linuxaio_get_events_from_ring (EV_A);
470 431
471 /* at this point, we should have handled all outstanding events */ 432 /* at this point, we should have handled all outstanding events */
… …
502 /* which allows us to pinpoint the erroneous iocb */ 463 /* which allows us to pinpoint the erroneous iocb */
503 for (submitted = 0; submitted < linuxaio_submitcnt; ) 464 for (submitted = 0; submitted < linuxaio_submitcnt; )
504 { 465 {
505 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);
506 467
507 if (expect_false (res < 0)) 468 if (ecb_expect_false (res < 0))
508 if (errno == EINVAL) 469 if (errno == EINVAL)
509 { 470 {
510 /* This happens for unsupported fds, officially, but in my testing, 471 /* This happens for unsupported fds, officially, but in my testing,
511 * 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
512 * 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.
… …
542 } 503 }
543 504
544 ++linuxaio_iteration; 505 ++linuxaio_iteration;
545 if (linuxaio_io_setup (EV_A) < 0) 506 if (linuxaio_io_setup (EV_A) < 0)
546 { 507 {
508 /* TODO: rearm all and recreate epoll backend from scratch */
509 /* TODO: might be more prudent? */
510
547 /* 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 */
548 linuxaio_free_iocbp (EV_A); 512 linuxaio_free_iocbp (EV_A);
549 ev_io_stop (EV_A_ &linuxaio_epoll_w); 513 ev_io_stop (EV_A_ &linuxaio_epoll_w);
550 ev_ref (EV_A); 514 ev_ref (EV_A);
551 linuxaio_ctx = 0; 515 linuxaio_ctx = 0;
516
517 backend = EVBACKEND_EPOLL;
552 backend_modify = epoll_modify; 518 backend_modify = epoll_modify;
553 backend_poll = epoll_poll; 519 backend_poll = epoll_poll;
554 } 520 }
555 521
556 timeout = 0; 522 timeout = 0;
… …
565 res = 1; /* skip this iocb */ 531 res = 1; /* skip this iocb */
566 } 532 }
567 else if (errno == EINTR) /* not seen in reality, not documented */ 533 else if (errno == EINTR) /* not seen in reality, not documented */
568 res = 0; /* silently ignore and retry */ 534 res = 0; /* silently ignore and retry */
569 else 535 else
536 {
570 ev_syserr ("(libev) linuxaio io_submit"); 537 ev_syserr ("(libev) linuxaio io_submit");
538 res = 0;
539 }
571 540
572 submitted += res; 541 submitted += res;
573 } 542 }
574 543
575 linuxaio_submitcnt = 0; 544 linuxaio_submitcnt = 0;
… …
603 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 572 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
604 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 573 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
605 ev_io_start (EV_A_ &linuxaio_epoll_w); 574 ev_io_start (EV_A_ &linuxaio_epoll_w);
606 ev_unref (EV_A); /* watcher should not keep loop alive */ 575 ev_unref (EV_A); /* watcher should not keep loop alive */
607 576
608 backend_modify = linuxaio_modify; 577 backend_modify = linuxaio_modify;
609 backend_poll = linuxaio_poll; 578 backend_poll = linuxaio_poll;
610 579
611 linuxaio_iocbpmax = 0; 580 linuxaio_iocbpmax = 0;
612 linuxaio_iocbps = 0; 581 linuxaio_iocbps = 0;
613 582
614 linuxaio_submits = 0; 583 linuxaio_submits = 0;
… …
625 epoll_destroy (EV_A); 594 epoll_destroy (EV_A);
626 linuxaio_free_iocbp (EV_A); 595 linuxaio_free_iocbp (EV_A);
627 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 */
628} 597}
629 598
630inline_size 599ecb_cold
631void 600static void
632linuxaio_fork (EV_P) 601linuxaio_fork (EV_P)
633{ 602{
634 /* this frees all iocbs, which is very heavy-handed */
635 linuxaio_destroy (EV_A);
636 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 */
637 606
638 linuxaio_iteration = 0; /* we start over in the child */ 607 linuxaio_iteration = 0; /* we start over in the child */
639 608
640 while (linuxaio_io_setup (EV_A) < 0) 609 while (linuxaio_io_setup (EV_A) < 0)
641 ev_syserr ("(libev) linuxaio io_setup"); 610 ev_syserr ("(libev) linuxaio io_setup");
642 611
643 /* forking epoll should also effectively unregister all fds from the backend */ 612 /* forking epoll should also effectively unregister all fds from the backend */
644 epoll_fork (EV_A); 613 epoll_fork (EV_A);
614 /* epoll_fork already did this. hopefully */
615 /*fd_rearm_all (EV_A);*/
645 616
646 ev_io_stop (EV_A_ &linuxaio_epoll_w); 617 ev_io_stop (EV_A_ &linuxaio_epoll_w);
647 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);
648 ev_io_start (EV_A_ &linuxaio_epoll_w); 619 ev_io_start (EV_A_ &linuxaio_epoll_w);
649
650 /* epoll_fork already did this. hopefully */
651 /*fd_rearm_all (EV_A);*/
652} 620}
653 621

Diff Legend

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