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 1.5 by root, Fri Jun 21 03:41:40 2019 UTC vs.
Revision 1.12 by root, Sun Jun 23 23:02:09 2019 UTC

… …
35 * and other provisions required by the GPL. If you do not delete the 35 * and other provisions required by the GPL. If you do not delete the
36 * provisions above, a recipient may use your version of this file under 36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL. 37 * either the BSD or the GPL.
38 */ 38 */
39 39
40#define EPOLL_FALLBACK 1
41
40#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ 42#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
41#include <poll.h> 43#include <poll.h>
42#include <linux/aio_abi.h> 44#include <linux/aio_abi.h>
43 45
46#if EPOLL_FALLBACK
47# include <sys/epoll.h>
48#endif
49
44/* we try to fill 4kn pages exactly. 50/* we try to fill 4kB pages exactly.
45 * the ring buffer header is 32 bytes, every io event is 32 bytes. 51 * the ring buffer header is 32 bytes, every io event is 32 bytes.
46 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer 52 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
47 * so the calculation below will use "exactly" 8kB for the ring buffer 53 * therefore the calculation below will use "exactly" 4kB for the ring buffer
48 */ 54 */
49#define EV_LINUXAIO_DEPTH (256 / 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 */
50 56
51/*****************************************************************************/ 57/*****************************************************************************/
52/* syscall wrapdadoop */ 58/* syscall wrapdadoop */
53 59
54#include <sys/syscall.h> /* no glibc wrappers */ 60#include <sys/syscall.h> /* no glibc wrappers */
… …
72 unsigned header_length; /* size of aio_ring */ 78 unsigned header_length; /* size of aio_ring */
73 79
74 struct io_event io_events[0]; 80 struct io_event io_events[0];
75}; 81};
76 82
77static int 83inline_size
84int
78ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
79{ 86{
80 return syscall (SYS_io_setup, nr_events, ctx_idp); 87 return syscall (SYS_io_setup, nr_events, ctx_idp);
81} 88}
82 89
83static int 90inline_size
91int
84ev_io_destroy (aio_context_t ctx_id) 92ev_io_destroy (aio_context_t ctx_id)
85{ 93{
86 return syscall (SYS_io_destroy, ctx_id); 94 return syscall (SYS_io_destroy, ctx_id);
87} 95}
88 96
89static int 97inline_size
98int
90ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 99ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
91{ 100{
92 return syscall (SYS_io_submit, ctx_id, nr, cbp); 101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
93} 102}
94 103
95static int 104inline_size
105int
96ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 106ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
97{ 107{
98 return syscall (SYS_io_cancel, ctx_id, cbp, result); 108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
99} 109}
100 110
101static int 111inline_size
112int
102ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 113ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
103{ 114{
104 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);
105} 116}
106
107typedef void (*ev_io_cb) (long nr, struct io_event *events);
108 117
109/*****************************************************************************/ 118/*****************************************************************************/
110/* actual backed implementation */ 119/* actual backed implementation */
111 120
112/* two iocbs for every fd, one for read, one for write */ 121/* we use out own wrapper structure in acse we ever want to do something "clever" */
113typedef struct aniocb 122typedef struct aniocb
114{ 123{
115 struct iocb io; 124 struct iocb io;
116 /*int inuse;*/ 125 /*int inuse;*/
117} *ANIOCBP; 126} *ANIOCBP;
118 127
119inline_size 128inline_size
120void 129void
121linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count)
122{ 131{
123 /* TODO: quite the overhead to allocate every iocb separately */ 132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
124 while (count--) 133 while (count--)
125 { 134 {
126 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 135 *base = (ANIOCBP)ev_malloc (sizeof (**base));
127 /* TODO: full zero initialize required? */ 136 /* TODO: full zero initialize required? */
128 memset (*base, 0, sizeof (**base)); 137 memset (*base, 0, sizeof (**base));
129 /* would be nice to initialize fd/data as well */ 138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */
130 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
131 ++base; 140 ++base;
132 } 141 }
133} 142}
134 143
144ecb_cold
135static void 145static void
136linuxaio_free_iocbp (EV_P) 146linuxaio_free_iocbp (EV_P)
137{ 147{
138 while (linuxaio_iocbpmax--) 148 while (linuxaio_iocbpmax--)
139 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); 149 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
140 150
141 /* next resize will completely reallocate the array */ 151 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
142 linuxaio_iocbpmax = 0;
143 linuxaio_submitcnt = 0; /* all pointers invalidated */
144} 152}
145 153
146static void 154static void
147linuxaio_modify (EV_P_ int fd, int oev, int nev) 155linuxaio_modify (EV_P_ int fd, int oev, int nev)
148{ 156{
149 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
150 struct aniocb *iocb = linuxaio_iocbps [fd]; 158 struct aniocb *iocb = linuxaio_iocbps [fd];
159
160#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0)
162 {
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
164 iocb->io.aio_reqprio = 0;
165 }
166#endif
151 167
152 if (iocb->io.aio_buf) 168 if (iocb->io.aio_buf)
153 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
154 170
155 if (nev) 171 if (nev)
… …
203static int 219static int
204linuxaio_get_events_from_ring (EV_P) 220linuxaio_get_events_from_ring (EV_P)
205{ 221{
206 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
207 223
208 ECB_MEMORY_FENCE_ACQUIRE;
209
210 unsigned head = ring->head; 224 unsigned head = ring->head;
211 unsigned tail = *(volatile unsigned *)&ring->tail; 225 unsigned tail = *(volatile unsigned *)&ring->tail;
212 226
213 if (ring->magic != AIO_RING_MAGIC 227 if (head == tail)
214 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
215 || ring->header_length != sizeof (struct aio_ring) /* TODO: or use it to find io_event[0]? */
216 || head == tail)
217 return 0; 228 return 0;
229
230 /* bail out if the ring buffer doesn't match the expected layout */
231 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC)
232 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
233 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
234 return 0;
235
236 /* make sure the events up to tail are visible */
237 ECB_MEMORY_FENCE_ACQUIRE;
218 238
219 /* parse all available events, but only once, to avoid starvation */ 239 /* parse all available events, but only once, to avoid starvation */
220 if (tail > head) /* normal case around */ 240 if (tail > head) /* normal case around */
221 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 241 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
222 else
223 {
224 /* wrapped around */ 242 else /* wrapped around */
243 {
225 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 244 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
226 linuxaio_parse_events (EV_A_ ring->io_events, tail); 245 linuxaio_parse_events (EV_A_ ring->io_events, tail);
227 } 246 }
228 247
229 ring->head = tail; 248 *(volatile unsigned *)&ring->head = tail;
249 /* make sure kernel can see our new head value - probably not required */
250 ECB_MEMORY_FENCE_RELEASE;
230 251
231 return 1; 252 return 1;
232} 253}
233 254
234/* read at least one event from kernel, or timeout */ 255/* read at least one event from kernel, or timeout */
… …
242 263
243 if (linuxaio_get_events_from_ring (EV_A)) 264 if (linuxaio_get_events_from_ring (EV_A))
244 return; 265 return;
245 266
246 /* no events, so wait for at least one, then poll ring buffer again */ 267 /* no events, so wait for at least one, then poll ring buffer again */
247 /* this degraded to one event per loop iteration */ 268 /* this degrades to one event per loop iteration */
248 /* if the ring buffer changes layout, but so be it */ 269 /* if the ring buffer changes layout, but so be it */
249 270
250 ts.tv_sec = (long)timeout; 271 ts.tv_sec = (long)timeout;
251 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 272 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
252 273
253 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 274 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);
254 275
255 if (res < 0) 276 if (res < 0)
277 if (errno == EINTR)
278 /* ignored */;
279 else
256 ev_syserr ("(libev) io_getevents"); 280 ev_syserr ("(libev) linuxaio io_getevents");
257 else if (res) 281 else if (res)
258 { 282 {
259 /* at least one event received, handle it and any remaining ones in the ring buffer */ 283 /* at least one event received, handle it and any remaining ones in the ring buffer */
260 linuxaio_parse_events (EV_A_ &ioev, 1); 284 linuxaio_parse_events (EV_A_ &ioev, 1);
261 linuxaio_get_events_from_ring (EV_A); 285 linuxaio_get_events_from_ring (EV_A);
262 } 286 }
263} 287}
288
289#if EPOLL_FALLBACK
290static void
291linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
292{
293 struct epoll_event eev;
294
295 eev.events = EPOLLONESHOT;
296 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
297 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
298 eev.data.fd = iocb->aio_fildes;
299
300 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
301 ev_syserr ("(libeio) linuxaio epoll_ctl");
302}
303#endif
264 304
265static void 305static void
266linuxaio_poll (EV_P_ ev_tstamp timeout) 306linuxaio_poll (EV_P_ ev_tstamp timeout)
267{ 307{
268 int submitted; 308 int submitted;
… …
274 /* which allows us to pinpoint the errornous iocb */ 314 /* which allows us to pinpoint the errornous iocb */
275 for (submitted = 0; submitted < linuxaio_submitcnt; ) 315 for (submitted = 0; submitted < linuxaio_submitcnt; )
276 { 316 {
277 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 317 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
278 318
279 if (res < 0) 319 if (ecb_expect_false (res < 0))
280 if (errno == EAGAIN) 320 if (errno == EAGAIN)
281 { 321 {
282 /* This happens when the ring buffer is full, at least. I assume this means 322 /* This happens when the ring buffer is full, at least. I assume this means
283 * that the event was queued synchronously during io_submit, and thus 323 * that the event was queued synchronously during io_submit, and thus
284 * the buffer overflowd. 324 * the buffer overflowed.
285 * In this case, we just try next loop iteration. 325 * In this case, we just try in next loop iteration.
326 * This should not result in a few fds taking priority, as the interface
327 * is one-shot, and we submit iocb's in a round-robin fashion.
286 */ 328 */
287 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 329 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
288 linuxaio_submitcnt -= submitted; 330 linuxaio_submitcnt -= submitted;
289 timeout = 0; 331 timeout = 0;
290 break; 332 break;
291 } 333 }
334#if EPOLL_FALLBACK
335 else if (errno == EINVAL)
336 {
337 /* This hapüpens for unsupported fds, officially, but in my testing,
338 * also randomly happens for supported fds. We fall back to good old
339 * poll() here, under the assumption that this is a very rare case.
340 */
341 struct iocb *iocb = linuxaio_submits [submitted];
342 res = 1; /* skip this iocb */
343
344 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
345 iocb->aio_reqprio = -1; /* mark iocb as epoll */
346 }
347#endif
292 else 348 else
293 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
294 /* TODO: should we always just try next time? */
295 ev_syserr ("(libev) io_submit"); 349 ev_syserr ("(libev) linuxaio io_submit");
296 350
297 submitted += res; 351 submitted += res;
298 } 352 }
299 353
300 linuxaio_submitcnt = 0; 354 linuxaio_submitcnt = 0;
301 355
302 /* second phase: fetch and parse events */ 356 /* second phase: fetch and parse events */
303 357
304 linuxaio_get_events (EV_A_ timeout); 358 linuxaio_get_events (EV_A_ timeout);
305} 359}
360
361#if EPOLL_FALLBACK
362
363static void
364linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
365{
366 struct epoll_event events[16];
367
368 for (;;)
369 {
370 int idx;
371 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
372
373 if (ecb_expect_false (res < 0))
374 ev_syserr ("(libev) linuxaio epoll_wait");
375 else if (!res)
376 break;
377
378 for (idx = res; idx--; )
379 {
380 int fd = events [idx].data.fd;
381 uint32_t ev = events [idx].events;
382
383 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
384
385 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
386
387 fd_event (EV_A_ fd,
388 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
389 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
390 }
391
392 if (res < sizeof (events) / sizeof (events [0]))
393 break;
394 }
395}
396
397#endif
306 398
307inline_size 399inline_size
308int 400int
309linuxaio_init (EV_P_ int flags) 401linuxaio_init (EV_P_ int flags)
310{ 402{
… …
315 407
316 linuxaio_ctx = 0; 408 linuxaio_ctx = 0;
317 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 409 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
318 return 0; 410 return 0;
319 411
412#if EPOLL_FALLBACK
413 backend_fd = ev_epoll_create ();
414 if (backend_fd < 0)
415 {
416 ev_io_destroy (linuxaio_ctx);
417 return 0;
418 }
419
420 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
421 ev_io_start (EV_A_ &linuxaio_epoll_w);
422#endif
423
320 backend_modify = linuxaio_modify; 424 backend_modify = linuxaio_modify;
321 backend_poll = linuxaio_poll; 425 backend_poll = linuxaio_poll;
322 426
323 linuxaio_iocbpmax = 0; 427 linuxaio_iocbpmax = 0;
324 linuxaio_iocbps = 0; 428 linuxaio_iocbps = 0;
… …
332 436
333inline_size 437inline_size
334void 438void
335linuxaio_destroy (EV_P) 439linuxaio_destroy (EV_P)
336{ 440{
441#if EPOLL_FALLBACK
442 close (backend_fd);
443#endif
337 linuxaio_free_iocbp (EV_A); 444 linuxaio_free_iocbp (EV_A);
338 ev_io_destroy (linuxaio_ctx); 445 ev_io_destroy (linuxaio_ctx);
339} 446}
340 447
341inline_size 448inline_size
342void 449void
343linuxaio_fork (EV_P) 450linuxaio_fork (EV_P)
344{ 451{
345 /* TODO: verify and test */ 452 /* this frees all iocbs, which is very heavy-handed */
346 linuxaio_destroy (EV_A); 453 linuxaio_destroy (EV_A);
454 linuxaio_submitcnt = 0; /* all pointers were invalidated */
347 455
348 linuxaio_ctx = 0; 456 linuxaio_ctx = 0;
349 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 457 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
350 ev_syserr ("(libev) io_setup"); 458 ev_syserr ("(libev) linuxaio io_setup");
459
460#if EPOLL_FALLBACK
461 while ((backend_fd = ev_epoll_create ()) < 0)
462 ev_syserr ("(libev) linuxaio epoll_create");
463
464 ev_io_stop (EV_A_ &linuxaio_epoll_w);
465 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
466 ev_io_start (EV_A_ &linuxaio_epoll_w);
467 ev_unref (EV_A); /* watcher should not keep loop alive */
468#endif
351 469
352 fd_rearm_all (EV_A); 470 fd_rearm_all (EV_A);
353} 471}
354 472

Diff Legend

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