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 1.6 by root, Fri Jun 21 03:52:29 2019 UTC vs.
Revision 1.13 by root, Sun Jun 23 23:38:38 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 4kB 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 * therefore 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 */
… …
110} 116}
111 117
112/*****************************************************************************/ 118/*****************************************************************************/
113/* actual backed implementation */ 119/* actual backed implementation */
114 120
121/* we use out own wrapper structure in acse we ever want to do something "clever" */
115typedef struct aniocb 122typedef struct aniocb
116{ 123{
117 struct iocb io; 124 struct iocb io;
118 /*int inuse;*/ 125 /*int inuse;*/
119} *ANIOCBP; 126} *ANIOCBP;
120 127
121inline_size 128inline_size
122void 129void
123linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count)
124{ 131{
125 /* TODO: quite the overhead to allocate every iocb separately */ 132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
126 while (count--) 133 while (count--)
127 { 134 {
128 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 135 *base = (ANIOCBP)ev_malloc (sizeof (**base));
129 /* TODO: full zero initialize required? */ 136 /* TODO: full zero initialize required? */
130 memset (*base, 0, sizeof (**base)); 137 memset (*base, 0, sizeof (**base));
131 /* 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 */
132 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
133 ++base; 140 ++base;
134 } 141 }
135} 142}
136 143
… …
147static void 154static void
148linuxaio_modify (EV_P_ int fd, int oev, int nev) 155linuxaio_modify (EV_P_ int fd, int oev, int nev)
149{ 156{
150 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);
151 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
152 167
153 if (iocb->io.aio_buf) 168 if (iocb->io.aio_buf)
154 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 */
155 170
156 if (nev) 171 if (nev)
… …
204static int 219static int
205linuxaio_get_events_from_ring (EV_P) 220linuxaio_get_events_from_ring (EV_P)
206{ 221{
207 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
208 223
209 ECB_MEMORY_FENCE_ACQUIRE; 224 /* the kernel reads and writes both of these variables, */
210 225 /* as a C extension, we assume that volatile use here */
211 unsigned head = ring->head; 226 /* both makes reads atomic and once-only */
227 unsigned head = *(volatile unsigned *)&ring->head;
212 unsigned tail = *(volatile unsigned *)&ring->tail; 228 unsigned tail = *(volatile unsigned *)&ring->tail;
213 229
214 if (head == tail) 230 if (head == tail)
215 return 0; 231 return 0;
216 232
233 /* bail out if the ring buffer doesn't match the expected layout */
217 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 234 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC)
218 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 235 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
219 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ 236 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
220 return 0; 237 return 0;
221 238
239 /* make sure the events up to tail are visible */
240 ECB_MEMORY_FENCE_ACQUIRE;
241
222 /* parse all available events, but only once, to avoid starvation */ 242 /* parse all available events, but only once, to avoid starvation */
223 if (tail > head) /* normal case around */ 243 if (tail > head) /* normal case around */
224 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 244 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
225 else /* wrapped around */ 245 else /* wrapped around */
226 { 246 {
227 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 247 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
228 linuxaio_parse_events (EV_A_ ring->io_events, tail); 248 linuxaio_parse_events (EV_A_ ring->io_events, tail);
229 } 249 }
230 250
231 ring->head = tail; 251 /* as an extension to C, we hope that the volatile will makethis atomic and once-only */
252 *(volatile unsigned *)&ring->head = tail;
253 /* make sure kernel can see our new head value - probably not required */
254 ECB_MEMORY_FENCE_RELEASE;
232 255
233 return 1; 256 return 1;
234} 257}
235 258
236/* read at least one event from kernel, or timeout */ 259/* read at least one event from kernel, or timeout */
… …
244 267
245 if (linuxaio_get_events_from_ring (EV_A)) 268 if (linuxaio_get_events_from_ring (EV_A))
246 return; 269 return;
247 270
248 /* no events, so wait for at least one, then poll ring buffer again */ 271 /* no events, so wait for at least one, then poll ring buffer again */
249 /* this degraded to one event per loop iteration */ 272 /* this degrades to one event per loop iteration */
250 /* if the ring buffer changes layout, but so be it */ 273 /* if the ring buffer changes layout, but so be it */
251 274
252 ts.tv_sec = (long)timeout; 275 ts.tv_sec = (long)timeout;
253 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 276 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
254 277
255 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 278 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);
256 279
257 if (res < 0) 280 if (res < 0)
281 if (errno == EINTR)
282 /* ignored */;
283 else
258 ev_syserr ("(libev) io_getevents"); 284 ev_syserr ("(libev) linuxaio io_getevents");
259 else if (res) 285 else if (res)
260 { 286 {
261 /* at least one event received, handle it and any remaining ones in the ring buffer */ 287 /* at least one event received, handle it and any remaining ones in the ring buffer */
262 linuxaio_parse_events (EV_A_ &ioev, 1); 288 linuxaio_parse_events (EV_A_ &ioev, 1);
263 linuxaio_get_events_from_ring (EV_A); 289 linuxaio_get_events_from_ring (EV_A);
264 } 290 }
265} 291}
292
293#if EPOLL_FALLBACK
294static void
295linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
296{
297 struct epoll_event eev;
298
299 eev.events = EPOLLONESHOT;
300 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
301 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
302 eev.data.fd = iocb->aio_fildes;
303
304 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
305 ev_syserr ("(libeio) linuxaio epoll_ctl");
306}
307#endif
266 308
267static void 309static void
268linuxaio_poll (EV_P_ ev_tstamp timeout) 310linuxaio_poll (EV_P_ ev_tstamp timeout)
269{ 311{
270 int submitted; 312 int submitted;
… …
281 if (ecb_expect_false (res < 0)) 323 if (ecb_expect_false (res < 0))
282 if (errno == EAGAIN) 324 if (errno == EAGAIN)
283 { 325 {
284 /* This happens when the ring buffer is full, at least. I assume this means 326 /* This happens when the ring buffer is full, at least. I assume this means
285 * that the event was queued synchronously during io_submit, and thus 327 * that the event was queued synchronously during io_submit, and thus
286 * the buffer overflowd. 328 * the buffer overflowed.
287 * In this case, we just try next loop iteration. 329 * In this case, we just try in next loop iteration.
330 * This should not result in a few fds taking priority, as the interface
331 * is one-shot, and we submit iocb's in a round-robin fashion.
288 */ 332 */
289 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 333 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
290 linuxaio_submitcnt -= submitted; 334 linuxaio_submitcnt -= submitted;
291 timeout = 0; 335 timeout = 0;
292 break; 336 break;
293 } 337 }
338#if EPOLL_FALLBACK
339 else if (errno == EINVAL)
340 {
341 /* This hapüpens for unsupported fds, officially, but in my testing,
342 * also randomly happens for supported fds. We fall back to good old
343 * poll() here, under the assumption that this is a very rare case.
344 */
345 struct iocb *iocb = linuxaio_submits [submitted];
346 res = 1; /* skip this iocb */
347
348 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
349 iocb->aio_reqprio = -1; /* mark iocb as epoll */
350 }
351#endif
294 else 352 else
295 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
296 /* TODO: should we always just try next time? */
297 ev_syserr ("(libev) io_submit"); 353 ev_syserr ("(libev) linuxaio io_submit");
298 354
299 submitted += res; 355 submitted += res;
300 } 356 }
301 357
302 linuxaio_submitcnt = 0; 358 linuxaio_submitcnt = 0;
303 359
304 /* second phase: fetch and parse events */ 360 /* second phase: fetch and parse events */
305 361
306 linuxaio_get_events (EV_A_ timeout); 362 linuxaio_get_events (EV_A_ timeout);
307} 363}
364
365#if EPOLL_FALLBACK
366
367static void
368linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
369{
370 struct epoll_event events[16];
371
372 for (;;)
373 {
374 int idx;
375 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
376
377 if (ecb_expect_false (res < 0))
378 ev_syserr ("(libev) linuxaio epoll_wait");
379 else if (!res)
380 break;
381
382 for (idx = res; idx--; )
383 {
384 int fd = events [idx].data.fd;
385 uint32_t ev = events [idx].events;
386
387 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
388
389 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
390
391 fd_event (EV_A_ fd,
392 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
393 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
394 }
395
396 if (res < sizeof (events) / sizeof (events [0]))
397 break;
398 }
399}
400
401#endif
308 402
309inline_size 403inline_size
310int 404int
311linuxaio_init (EV_P_ int flags) 405linuxaio_init (EV_P_ int flags)
312{ 406{
… …
317 411
318 linuxaio_ctx = 0; 412 linuxaio_ctx = 0;
319 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 413 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
320 return 0; 414 return 0;
321 415
416#if EPOLL_FALLBACK
417 backend_fd = ev_epoll_create ();
418 if (backend_fd < 0)
419 {
420 ev_io_destroy (linuxaio_ctx);
421 return 0;
422 }
423
424 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
425 ev_io_start (EV_A_ &linuxaio_epoll_w);
426#endif
427
322 backend_modify = linuxaio_modify; 428 backend_modify = linuxaio_modify;
323 backend_poll = linuxaio_poll; 429 backend_poll = linuxaio_poll;
324 430
325 linuxaio_iocbpmax = 0; 431 linuxaio_iocbpmax = 0;
326 linuxaio_iocbps = 0; 432 linuxaio_iocbps = 0;
… …
334 440
335inline_size 441inline_size
336void 442void
337linuxaio_destroy (EV_P) 443linuxaio_destroy (EV_P)
338{ 444{
445#if EPOLL_FALLBACK
446 close (backend_fd);
447#endif
339 linuxaio_free_iocbp (EV_A); 448 linuxaio_free_iocbp (EV_A);
340 ev_io_destroy (linuxaio_ctx); 449 ev_io_destroy (linuxaio_ctx);
341} 450}
342 451
343inline_size 452inline_size
344void 453void
345linuxaio_fork (EV_P) 454linuxaio_fork (EV_P)
346{ 455{
347 /* TODO: verify and test */
348
349 /* this frees all iocbs, which is very heavy-handed */ 456 /* this frees all iocbs, which is very heavy-handed */
350 linuxaio_destroy (EV_A); 457 linuxaio_destroy (EV_A);
351 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 458 linuxaio_submitcnt = 0; /* all pointers were invalidated */
352 459
353 linuxaio_ctx = 0; 460 linuxaio_ctx = 0;
354 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 461 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
355 ev_syserr ("(libev) io_setup"); 462 ev_syserr ("(libev) linuxaio io_setup");
463
464#if EPOLL_FALLBACK
465 while ((backend_fd = ev_epoll_create ()) < 0)
466 ev_syserr ("(libev) linuxaio epoll_create");
467
468 ev_io_stop (EV_A_ &linuxaio_epoll_w);
469 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
470 ev_io_start (EV_A_ &linuxaio_epoll_w);
471 ev_unref (EV_A); /* watcher should not keep loop alive */
472#endif
356 473
357 fd_rearm_all (EV_A); 474 fd_rearm_all (EV_A);
358} 475}
359 476

Diff Legend

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