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 1.19 by root, Mon Jun 24 19:53:47 2019 UTC vs.
Revision 1.23 by root, Mon Jun 24 21:34:47 2019 UTC

… …
53 * therefore the calculation below will use "exactly" 4kB for the ring buffer 53 * therefore the calculation below will use "exactly" 4kB for the ring buffer
54 */ 54 */
55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ 55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56 56
57/*****************************************************************************/ 57/*****************************************************************************/
58/* syscall wrapdadoop */ 58/* syscall wrapdadoop - this section has the raw syscall definitions */
59 59
60#include <sys/syscall.h> /* no glibc wrappers */ 60#include <sys/syscall.h> /* no glibc wrappers */
61 61
62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */ 62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63#define IOCB_CMD_POLL 5 63#define IOCB_CMD_POLL 5
… …
80 struct io_event io_events[0]; 80 struct io_event io_events[0];
81}; 81};
82 82
83inline_size 83inline_size
84int 84int
85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 85evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86{ 86{
87 return syscall (SYS_io_setup, nr_events, ctx_idp); 87 return syscall (SYS_io_setup, nr_events, ctx_idp);
88} 88}
89 89
90inline_size 90inline_size
91int 91int
92ev_io_destroy (aio_context_t ctx_id) 92evsys_io_destroy (aio_context_t ctx_id)
93{ 93{
94 return syscall (SYS_io_destroy, ctx_id); 94 return syscall (SYS_io_destroy, ctx_id);
95} 95}
96 96
97inline_size 97inline_size
98int 98int
99ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 99evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100{ 100{
101 return syscall (SYS_io_submit, ctx_id, nr, cbp); 101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
102} 102}
103 103
104inline_size 104inline_size
105int 105int
106ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 106evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
107{ 107{
108 return syscall (SYS_io_cancel, ctx_id, cbp, result); 108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
109} 109}
110 110
111inline_size 111inline_size
112int 112int
113ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 113evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
114{ 114{
115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116} 116}
117 117
118/*****************************************************************************/ 118/*****************************************************************************/
… …
125 /*int inuse;*/ 125 /*int inuse;*/
126} *ANIOCBP; 126} *ANIOCBP;
127 127
128inline_size 128inline_size
129void 129void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 130linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 131{
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--) 132 while (count--)
134 { 133 {
134 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
135 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 135 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 /* TODO: full zero initialize required? */ 136
137 /* full zero initialise is probably not required at the moment, but
138 * this is not well documented, so we better do it.
139 */
137 memset (*base, 0, sizeof (**base)); 140 memset (iocb, 0, sizeof (*iocb));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 141
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 142 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base; 143 iocb->io.aio_data = offset;
144 iocb->io.aio_fildes = offset;
145
146 base [offset++] = iocb;
141 } 147 }
142} 148}
143 149
144ecb_cold 150ecb_cold
145static void 151static void
… …
153 159
154static void 160static void
155linuxaio_modify (EV_P_ int fd, int oev, int nev) 161linuxaio_modify (EV_P_ int fd, int oev, int nev)
156{ 162{
157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 163 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158 struct aniocb *iocb = linuxaio_iocbps [fd]; 164 ANIOCBP iocb = linuxaio_iocbps [fd];
159 165
160#if EPOLL_FALLBACK 166#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0) 167 if (iocb->io.aio_reqprio < 0)
162 { 168 {
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
164 iocb->io.aio_reqprio = 0; 170 iocb->io.aio_reqprio = 0;
165 } 171 }
166#endif 172#endif
167 173
168 if (iocb->io.aio_buf) 174 if (iocb->io.aio_buf)
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
170 176
171 if (nev) 177 if (nev)
172 { 178 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf = 179 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0) 180 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0); 181 | (nev & EV_WRITE ? POLLOUT : 0);
178 182
179 /* queue iocb up for io_submit */ 183 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */ 184 /* this assumes we only ever get one call per fd per loop iteration */
… …
250 linuxaio_iocbps [fd]->io.aio_buf = 0; 254 linuxaio_iocbps [fd]->io.aio_buf = 0;
251 anfds [fd].events = 0; 255 anfds [fd].events = 0;
252 fd_change (EV_A_ fd, 0); 256 fd_change (EV_A_ fd, 0);
253 257
254 /* feed events, we do not expect or handle POLLNVAL */ 258 /* feed events, we do not expect or handle POLLNVAL */
255 if (expect_false (res & POLLNVAL))
256 fd_kill (EV_A_ fd);
257 else
258 fd_event ( 259 fd_event (
259 EV_A_ 260 EV_A_
260 fd, 261 fd,
261 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
262 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
263 ); 264 );
264 265
265 --nr; 266 --nr;
266 ++ev; 267 ++ev;
267 } 268 }
268} 269}
… …
329 EV_RELEASE_CB; 330 EV_RELEASE_CB;
330 331
331 ts.tv_sec = (long)timeout; 332 ts.tv_sec = (long)timeout;
332 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 333 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
333 334
334 res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); 335 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
335 336
336 EV_ACQUIRE_CB; 337 EV_ACQUIRE_CB;
337 338
338 if (res < 0) 339 if (res < 0)
339 if (errno == EINTR) 340 if (errno == EINTR)
… …
361 for (submitted = 0; submitted < linuxaio_submitcnt; ) 362 for (submitted = 0; submitted < linuxaio_submitcnt; )
362 { 363 {
363#if 0 364#if 0
364 int res; 365 int res;
365 if (linuxaio_submits[submitted]->aio_fildes == backend_fd) 366 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
366 res = ev_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); 367 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
367 else 368 else
368 { res = -1; errno = EINVAL; }; 369 { res = -1; errno = EINVAL; };
369#else 370#else
370 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 371 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371#endif 372#endif
372 373
373 if (expect_false (res < 0)) 374 if (expect_false (res < 0))
374 if (errno == EAGAIN) 375 if (errno == EAGAIN)
375 { 376 {
… …
399 * See https://lore.kernel.org/patchwork/patch/1047453/ to see 400 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
400 * discussion about such a case (ttys) where polling for POLLIN 401 * discussion about such a case (ttys) where polling for POLLIN
401 * fails but POLLIN|POLLOUT works. 402 * fails but POLLIN|POLLOUT works.
402 */ 403 */
403 struct iocb *iocb = linuxaio_submits [submitted]; 404 struct iocb *iocb = linuxaio_submits [submitted];
405
406 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408
404 res = 1; /* skip this iocb */ 409 res = 1; /* skip this iocb */
405
406 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408 } 410 }
409#endif 411#endif
412 else if (errno == EBADF)
413 {
414 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
415
416 res = 1; /* skip this iocb */
417 }
410 else 418 else
411 ev_syserr ("(libev) linuxaio io_submit"); 419 ev_syserr ("(libev) linuxaio io_submit");
412 420
413 submitted += res; 421 submitted += res;
414 } 422 }
… …
435 if (ev_linux_version () < 0x041200) 443 if (ev_linux_version () < 0x041200)
436 return 0; 444 return 0;
437#endif 445#endif
438 446
439 linuxaio_ctx = 0; 447 linuxaio_ctx = 0;
440 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 448 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
441 return 0; 449 return 0;
442 450
443#if EPOLL_FALLBACK 451#if EPOLL_FALLBACK
444 backend_fd = ev_epoll_create (); 452 backend_fd = ev_epoll_create ();
445 if (backend_fd < 0) 453 if (backend_fd < 0)
446 { 454 {
447 ev_io_destroy (linuxaio_ctx); 455 evsys_io_destroy (linuxaio_ctx);
448 return 0; 456 return 0;
449 } 457 }
450 458
451 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 459 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
452 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 460 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
… …
473{ 481{
474#if EPOLL_FALLBACK 482#if EPOLL_FALLBACK
475 close (backend_fd); 483 close (backend_fd);
476#endif 484#endif
477 linuxaio_free_iocbp (EV_A); 485 linuxaio_free_iocbp (EV_A);
478 ev_io_destroy (linuxaio_ctx); 486 evsys_io_destroy (linuxaio_ctx);
479} 487}
480 488
481inline_size 489inline_size
482void 490void
483linuxaio_fork (EV_P) 491linuxaio_fork (EV_P)
… …
485 /* this frees all iocbs, which is very heavy-handed */ 493 /* this frees all iocbs, which is very heavy-handed */
486 linuxaio_destroy (EV_A); 494 linuxaio_destroy (EV_A);
487 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 495 linuxaio_submitcnt = 0; /* all pointers were invalidated */
488 496
489 linuxaio_ctx = 0; 497 linuxaio_ctx = 0;
490 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 498 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
491 ev_syserr ("(libev) linuxaio io_setup"); 499 ev_syserr ("(libev) linuxaio io_setup");
492 500
493#if EPOLL_FALLBACK 501#if EPOLL_FALLBACK
494 while ((backend_fd = ev_epoll_create ()) < 0) 502 while ((backend_fd = ev_epoll_create ()) < 0)
495 ev_syserr ("(libev) linuxaio epoll_create"); 503 ev_syserr ("(libev) linuxaio epoll_create");

Diff Legend

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