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 1.28 by root, Tue Jun 25 06:36:59 2019 UTC vs.
Revision 1.41 by root, Mon Jul 1 20:41:46 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. */
… …
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
121inline_size 172inline_size
122int 173int
123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
124{ 175{
125 return syscall (SYS_io_setup, nr_events, ctx_idp); 176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
126} 177}
127 178
128inline_size 179inline_size
129int 180int
130evsys_io_destroy (aio_context_t ctx_id) 181evsys_io_destroy (aio_context_t ctx_id)
131{ 182{
132 return syscall (SYS_io_destroy, ctx_id); 183 return ev_syscall1 (SYS_io_destroy, ctx_id);
133} 184}
134 185
135inline_size 186inline_size
136int 187int
137evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 188evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
138{ 189{
139 return syscall (SYS_io_submit, ctx_id, nr, cbp); 190 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
140} 191}
141 192
142inline_size 193inline_size
143int 194int
144evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 195evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
145{ 196{
146 return syscall (SYS_io_cancel, ctx_id, cbp, result); 197 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
147} 198}
148 199
149inline_size 200inline_size
150int 201int
151evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 202evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
152{ 203{
153 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 204 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
154} 205}
155 206
156/*****************************************************************************/ 207/*****************************************************************************/
157/* actual backed implementation */ 208/* actual backed implementation */
158 209
… …
235linuxaio_modify (EV_P_ int fd, int oev, int nev) 286linuxaio_modify (EV_P_ int fd, int oev, int nev)
236{ 287{
237 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 288 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
238 ANIOCBP iocb = linuxaio_iocbps [fd]; 289 ANIOCBP iocb = linuxaio_iocbps [fd];
239 290
240 if (iocb->io.aio_reqprio < 0) 291 if (expect_false (iocb->io.aio_reqprio < 0))
241 { 292 {
242 /* we handed this fd over to epoll, so undo this first */ 293 /* we handed this fd over to epoll, so undo this first */
243 /* we do it manually because the optimisations on epoll_modfy won't do us any good */ 294 /* we do it manually because the optimisations on epoll_modify won't do us any good */
244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0;
245 iocb->io.aio_reqprio = 0; 297 iocb->io.aio_reqprio = 0;
246 } 298 }
247 299
300 if (expect_false (iocb->io.aio_buf))
301 {
302 /* iocb active, so cancel it first before resubmit */
303 for (;;)
304 {
305 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
306 if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
307 break;
308
309 if (expect_true (errno == EINPROGRESS))
310 break;
311
312 /* the EINPROGRESS test is for nicer error message. clumsy. */
313 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR));
314 }
315 }
316
248 if (iocb->io.aio_buf) 317 iocb->io.aio_buf =
249 /* io_cancel always returns some error on relevant kernels, but works */ 318 (nev & EV_READ ? POLLIN : 0)
250 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); 319 | (nev & EV_WRITE ? POLLOUT : 0);
251 320
252 if (nev) 321 if (nev)
253 { 322 {
254 iocb->io.aio_buf =
255 (nev & EV_READ ? POLLIN : 0)
256 | (nev & EV_WRITE ? POLLOUT : 0);
257
258 /* queue iocb up for io_submit */ 323 /* queue iocb up for io_submit */
259 /* this assumes we only ever get one call per fd per loop iteration */ 324 /* this assumes we only ever get one call per fd per loop iteration */
260 ++linuxaio_submitcnt; 325 ++linuxaio_submitcnt;
261 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 326 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
262 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 327 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
… …
267linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 332linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
268{ 333{
269 epoll_poll (EV_A_ 0); 334 epoll_poll (EV_A_ 0);
270} 335}
271 336
272static void 337inline_speed
338void
273linuxaio_fd_rearm (EV_P_ int fd) 339linuxaio_fd_rearm (EV_P_ int fd)
274{ 340{
275 anfds [fd].events = 0; 341 anfds [fd].events = 0;
276 linuxaio_iocbps [fd]->io.aio_buf = 0; 342 linuxaio_iocbps [fd]->io.aio_buf = 0;
277 fd_change (EV_A_ fd, EV_ANFD_REIFY); 343 fd_change (EV_A_ fd, EV_ANFD_REIFY);
… …
293 fd, 359 fd,
294 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 360 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
295 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 361 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
296 ); 362 );
297 363
298 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ 364 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
299 linuxaio_fd_rearm (EV_A_ fd); 365 linuxaio_fd_rearm (EV_A_ fd);
300 366
301 --nr; 367 --nr;
302 ++ev; 368 ++ev;
303 } 369 }
… …
316 unsigned tail = *(volatile unsigned *)&ring->tail; 382 unsigned tail = *(volatile unsigned *)&ring->tail;
317 383
318 if (head == tail) 384 if (head == tail)
319 return 0; 385 return 0;
320 386
321 /* bail out if the ring buffer doesn't match the expected layout */
322 if (expect_false (ring->magic != AIO_RING_MAGIC)
323 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
324 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
325 return 0;
326
327 /* make sure the events up to tail are visible */ 387 /* make sure the events up to tail are visible */
328 ECB_MEMORY_FENCE_ACQUIRE; 388 ECB_MEMORY_FENCE_ACQUIRE;
329 389
330 /* parse all available events, but only once, to avoid starvation */ 390 /* parse all available events, but only once, to avoid starvation */
331 if (tail > head) /* normal case around */ 391 if (tail > head) /* normal case around */
… …
337 } 397 }
338 398
339 ECB_MEMORY_FENCE_RELEASE; 399 ECB_MEMORY_FENCE_RELEASE;
340 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 400 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
341 *(volatile unsigned *)&ring->head = tail; 401 *(volatile unsigned *)&ring->head = tail;
342 /* make sure kernel can see our new head value - probably not required */
343 402
344 return 1; 403 return 1;
404}
405
406inline_size
407int
408linuxaio_ringbuf_valid (EV_P)
409{
410 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
411
412 return expect_true (ring->magic == AIO_RING_MAGIC)
413 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
414 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
345} 415}
346 416
347/* read at least one event from kernel, or timeout */ 417/* read at least one event from kernel, or timeout */
348inline_size 418inline_size
349void 419void
350linuxaio_get_events (EV_P_ ev_tstamp timeout) 420linuxaio_get_events (EV_P_ ev_tstamp timeout)
351{ 421{
352 struct timespec ts; 422 struct timespec ts;
353 struct io_event ioev[1]; 423 struct io_event ioev[8]; /* 256 octet stack space */
354 int res; 424 int want = 1; /* how many events to request */
425 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
355 426
427 if (expect_true (ringbuf_valid))
428 {
429 /* if the ring buffer has any events, we don't wait or call the kernel at all */
356 if (linuxaio_get_events_from_ring (EV_A)) 430 if (linuxaio_get_events_from_ring (EV_A))
357 return; 431 return;
358 432
359 /* no events, so wait for at least one, then poll ring buffer again */ 433 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
360 /* this degrades to one event per loop iteration */ 434 if (!timeout)
361 /* if the ring buffer changes layout, but so be it */ 435 return;
436 }
437 else
438 /* no ringbuffer, request slightly larger batch */
439 want = sizeof (ioev) / sizeof (ioev [0]);
362 440
441 /* no events, so wait for some
442 * for fairness reasons, we do this in a loop, to fetch all events
443 */
444 for (;;)
445 {
446 int res;
447
363 EV_RELEASE_CB; 448 EV_RELEASE_CB;
364 449
365 ts.tv_sec = (long)timeout; 450 ts.tv_sec = (long)timeout;
366 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 451 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
367 452
368 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); 453 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
369 454
370 EV_ACQUIRE_CB; 455 EV_ACQUIRE_CB;
371 456
372 if (res < 0) 457 if (res < 0)
373 if (errno == EINTR) 458 if (errno == EINTR)
374 /* ignored */; 459 /* ignored, retry */;
375 else 460 else
376 ev_syserr ("(libev) linuxaio io_getevents"); 461 ev_syserr ("(libev) linuxaio io_getevents");
377 else if (res) 462 else if (res)
378 { 463 {
379 /* at least one event received, handle it and any remaining ones in the ring buffer */ 464 /* at least one event available, handle them */
380 linuxaio_parse_events (EV_A_ ioev, res); 465 linuxaio_parse_events (EV_A_ ioev, res);
466
467 if (expect_true (ringbuf_valid))
468 {
469 /* if we have a ring buffer, handle any remaining events in it */
381 linuxaio_get_events_from_ring (EV_A); 470 linuxaio_get_events_from_ring (EV_A);
382 }
383}
384 471
385static int 472 /* at this point, we should have handled all outstanding events */
473 break;
474 }
475 else if (res < want)
476 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
477 break;
478 }
479 else
480 break; /* no events from the kernel, we are done */
481
482 timeout = 0; /* only wait in the first iteration */
483 }
484}
485
486inline_size
487int
386linuxaio_io_setup (EV_P) 488linuxaio_io_setup (EV_P)
387{ 489{
388 linuxaio_ctx = 0; 490 linuxaio_ctx = 0;
389 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); 491 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
390} 492}
… …
456 /* it's easiest to handle this mess in another iteration */ 558 /* it's easiest to handle this mess in another iteration */
457 return; 559 return;
458 } 560 }
459 else if (errno == EBADF) 561 else if (errno == EBADF)
460 { 562 {
563 assert (("libev: event loop rejected bad fd", errno != EBADF));
461 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 564 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
462 565
463 res = 1; /* skip this iocb */ 566 res = 1; /* skip this iocb */
464 } 567 }
568 else if (errno == EINTR) /* not seen in reality, not documented */
569 res = 0; /* silently ignore and retry */
465 else 570 else
466 ev_syserr ("(libev) linuxaio io_submit"); 571 ev_syserr ("(libev) linuxaio io_submit");
467 572
468 submitted += res; 573 submitted += res;
469 } 574 }
… …
518void 623void
519linuxaio_destroy (EV_P) 624linuxaio_destroy (EV_P)
520{ 625{
521 epoll_destroy (EV_A); 626 epoll_destroy (EV_A);
522 linuxaio_free_iocbp (EV_A); 627 linuxaio_free_iocbp (EV_A);
523 evsys_io_destroy (linuxaio_ctx); 628 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
524} 629}
525 630
526inline_size 631inline_size
527void 632void
528linuxaio_fork (EV_P) 633linuxaio_fork (EV_P)
… …
534 linuxaio_iteration = 0; /* we start over in the child */ 639 linuxaio_iteration = 0; /* we start over in the child */
535 640
536 while (linuxaio_io_setup (EV_A) < 0) 641 while (linuxaio_io_setup (EV_A) < 0)
537 ev_syserr ("(libev) linuxaio io_setup"); 642 ev_syserr ("(libev) linuxaio io_setup");
538 643
644 /* forking epoll should also effectively unregister all fds from the backend */
539 epoll_fork (EV_A); 645 epoll_fork (EV_A);
540 646
541 ev_io_stop (EV_A_ &linuxaio_epoll_w); 647 ev_io_stop (EV_A_ &linuxaio_epoll_w);
542 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); 648 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
543 ev_io_start (EV_A_ &linuxaio_epoll_w); 649 ev_io_start (EV_A_ &linuxaio_epoll_w);

Diff Legend

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