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 1.18 by root, Mon Jun 24 04:54:08 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)
… …
183 linuxaio_iocbps [fd]->io.aio_buf = 0; 198 linuxaio_iocbps [fd]->io.aio_buf = 0;
184 anfds [fd].events = 0; 199 anfds [fd].events = 0;
185 fd_change (EV_A_ fd, 0); 200 fd_change (EV_A_ fd, 0);
186 201
187 /* feed events, we do not expect or handle POLLNVAL */ 202 /* feed events, we do not expect or handle POLLNVAL */
188 if (ecb_expect_false (res & POLLNVAL)) 203 if (expect_false (res & POLLNVAL))
189 fd_kill (EV_A_ fd); 204 fd_kill (EV_A_ fd);
190 else 205 else
191 fd_event ( 206 fd_event (
192 EV_A_ 207 EV_A_
193 fd, 208 fd,
… …
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 (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;
238
239 /* make sure the events up to tail are visible */
240 ECB_MEMORY_FENCE_ACQUIRE;
221 241
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 /* TODO: we only need a compiler barrier here, not a read fence */
252 ECB_MEMORY_FENCE_RELEASE;
253 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
254 *(volatile unsigned *)&ring->head = tail;
255 /* make sure kernel can see our new head value - probably not required */
256 ECB_MEMORY_FENCE_RELEASE;
232 257
233 return 1; 258 return 1;
234} 259}
235 260
236/* read at least one event from kernel, or timeout */ 261/* read at least one event from kernel, or timeout */
237inline_size 262inline_size
238void 263void
239linuxaio_get_events (EV_P_ ev_tstamp timeout) 264linuxaio_get_events (EV_P_ ev_tstamp timeout)
240{ 265{
241 struct timespec ts; 266 struct timespec ts;
242 struct io_event ioev; 267 struct io_event ioev[1];
243 int res; 268 int res;
244 269
245 if (linuxaio_get_events_from_ring (EV_A)) 270 if (linuxaio_get_events_from_ring (EV_A))
246 return; 271 return;
247 272
248 /* no events, so wait for at least one, then poll ring buffer again */ 273 /* no events, so wait for at least one, then poll ring buffer again */
249 /* this degraded to one event per loop iteration */ 274 /* this degrades to one event per loop iteration */
250 /* if the ring buffer changes layout, but so be it */ 275 /* if the ring buffer changes layout, but so be it */
251 276
252 ts.tv_sec = (long)timeout; 277 ts.tv_sec = (long)timeout;
253 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 278 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
254 279
255 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 280 res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
256 281
257 if (res < 0) 282 if (res < 0)
283 if (errno == EINTR)
284 /* ignored */;
285 else
258 ev_syserr ("(libev) io_getevents"); 286 ev_syserr ("(libev) linuxaio io_getevents");
259 else if (res) 287 else if (res)
260 { 288 {
261 /* at least one event received, handle it and any remaining ones in the ring buffer */ 289 /* at least one event received, handle it and any remaining ones in the ring buffer */
262 linuxaio_parse_events (EV_A_ &ioev, 1); 290 linuxaio_parse_events (EV_A_ ioev, res);
263 linuxaio_get_events_from_ring (EV_A); 291 linuxaio_get_events_from_ring (EV_A);
264 } 292 }
265} 293}
294
295#if EPOLL_FALLBACK
296static void
297linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
298{
299 struct epoll_event eev;
300
301 eev.events = EPOLLONESHOT;
302 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
303 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
304 eev.data.fd = iocb->aio_fildes;
305
306 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
307 ev_syserr ("(libeio) linuxaio epoll_ctl");
308}
309#endif
266 310
267static void 311static void
268linuxaio_poll (EV_P_ ev_tstamp timeout) 312linuxaio_poll (EV_P_ ev_tstamp timeout)
269{ 313{
270 int submitted; 314 int submitted;
… …
276 /* which allows us to pinpoint the errornous iocb */ 320 /* which allows us to pinpoint the errornous iocb */
277 for (submitted = 0; submitted < linuxaio_submitcnt; ) 321 for (submitted = 0; submitted < linuxaio_submitcnt; )
278 { 322 {
279 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 323 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
280 324
281 if (ecb_expect_false (res < 0)) 325 if (expect_false (res < 0))
282 if (errno == EAGAIN) 326 if (errno == EAGAIN)
283 { 327 {
284 /* This happens when the ring buffer is full, at least. I assume this means 328 /* 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 329 * that the event was queued synchronously during io_submit, and thus
286 * the buffer overflowd. 330 * the buffer overflowed.
287 * In this case, we just try next loop iteration. 331 * In this case, we just try in next loop iteration.
332 * This should not result in a few fds taking priority, as the interface
333 * is one-shot, and we submit iocb's in a round-robin fashion.
288 */ 334 */
289 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 335 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
290 linuxaio_submitcnt -= submitted; 336 linuxaio_submitcnt -= submitted;
291 timeout = 0; 337 timeout = 0;
292 break; 338 break;
293 } 339 }
340#if EPOLL_FALLBACK
341 else if (errno == EINVAL)
342 {
343 /* This happens for unsupported fds, officially, but in my testing,
344 * also randomly happens for supported fds. We fall back to good old
345 * poll() here, under the assumption that this is a very rare case.
346 * See https://lore.kernel.org/patchwork/patch/1047453/ for evidence
347 * that the problem is known, but ignored.
348 */
349 struct iocb *iocb = linuxaio_submits [submitted];
350 res = 1; /* skip this iocb */
351
352 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
353 iocb->aio_reqprio = -1; /* mark iocb as epoll */
354 }
355#endif
294 else 356 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"); 357 ev_syserr ("(libev) linuxaio io_submit");
298 358
299 submitted += res; 359 submitted += res;
300 } 360 }
301 361
302 linuxaio_submitcnt = 0; 362 linuxaio_submitcnt = 0;
303 363
304 /* second phase: fetch and parse events */ 364 /* second phase: fetch and parse events */
305 365
306 linuxaio_get_events (EV_A_ timeout); 366 linuxaio_get_events (EV_A_ timeout);
307} 367}
368
369#if EPOLL_FALLBACK
370
371static void
372linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
373{
374 struct epoll_event events[16];
375
376 for (;;)
377 {
378 int idx;
379 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
380
381 if (expect_false (res < 0))
382 ev_syserr ("(libev) linuxaio epoll_wait");
383 else if (!res)
384 break;
385
386 for (idx = res; idx--; )
387 {
388 int fd = events [idx].data.fd;
389 uint32_t ev = events [idx].events;
390
391 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
392
393 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
394
395 fd_event (EV_A_ fd,
396 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
397 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
398 }
399
400 if (res < sizeof (events) / sizeof (events [0]))
401 break;
402 }
403}
404
405#endif
308 406
309inline_size 407inline_size
310int 408int
311linuxaio_init (EV_P_ int flags) 409linuxaio_init (EV_P_ int flags)
312{ 410{
313 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 411 /* would be great to have a nice test for IOCB_CMD_POLL instead */
314 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 412 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
315 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 413#if EPOLL_FALLBACK
414 /* 4.19 made epoll work */
415 if (ev_linux_version () < 0x041300)
316 return 0; 416 return 0;
417#else
418 /* 4.18 introduced IOCB_CMD_POLL */
419 if (ev_linux_version () < 0x041200)
420 return 0;
421#endif
317 422
318 linuxaio_ctx = 0; 423 linuxaio_ctx = 0;
319 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 424 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
320 return 0; 425 return 0;
321 426
427#if EPOLL_FALLBACK
428 backend_fd = ev_epoll_create ();
429 if (backend_fd < 0)
430 {
431 ev_io_destroy (linuxaio_ctx);
432 return 0;
433 }
434
435 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
436 ev_io_start (EV_A_ &linuxaio_epoll_w);
437 ev_unref (EV_A); /* watcher should not keep loop alive */
438#endif
439
322 backend_modify = linuxaio_modify; 440 backend_modify = linuxaio_modify;
323 backend_poll = linuxaio_poll; 441 backend_poll = linuxaio_poll;
324 442
325 linuxaio_iocbpmax = 0; 443 linuxaio_iocbpmax = 0;
326 linuxaio_iocbps = 0; 444 linuxaio_iocbps = 0;
… …
334 452
335inline_size 453inline_size
336void 454void
337linuxaio_destroy (EV_P) 455linuxaio_destroy (EV_P)
338{ 456{
457#if EPOLL_FALLBACK
458 close (backend_fd);
459#endif
339 linuxaio_free_iocbp (EV_A); 460 linuxaio_free_iocbp (EV_A);
340 ev_io_destroy (linuxaio_ctx); 461 ev_io_destroy (linuxaio_ctx);
341} 462}
342 463
343inline_size 464inline_size
344void 465void
345linuxaio_fork (EV_P) 466linuxaio_fork (EV_P)
346{ 467{
347 /* TODO: verify and test */
348
349 /* this frees all iocbs, which is very heavy-handed */ 468 /* this frees all iocbs, which is very heavy-handed */
350 linuxaio_destroy (EV_A); 469 linuxaio_destroy (EV_A);
351 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 470 linuxaio_submitcnt = 0; /* all pointers were invalidated */
352 471
353 linuxaio_ctx = 0; 472 linuxaio_ctx = 0;
354 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 473 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
355 ev_syserr ("(libev) io_setup"); 474 ev_syserr ("(libev) linuxaio io_setup");
475
476#if EPOLL_FALLBACK
477 while ((backend_fd = ev_epoll_create ()) < 0)
478 ev_syserr ("(libev) linuxaio epoll_create");
479
480 ev_io_stop (EV_A_ &linuxaio_epoll_w);
481 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
482 ev_io_start (EV_A_ &linuxaio_epoll_w);
483#endif
356 484
357 fd_rearm_all (EV_A); 485 fd_rearm_all (EV_A);
358} 486}
359 487

Diff Legend

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