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 1.20 by root, Mon Jun 24 20:33:24 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 - this section has the raw syscall definitions */
53 59
54#include <sys/syscall.h> /* no glibc wrappers */ 60#include <sys/syscall.h> /* no glibc wrappers */
55 61
56/* 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 */
57#define IOCB_CMD_POLL 5 63#define IOCB_CMD_POLL 5
… …
74 struct io_event io_events[0]; 80 struct io_event io_events[0];
75}; 81};
76 82
77inline_size 83inline_size
78int 84int
79ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 85evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
80{ 86{
81 return syscall (SYS_io_setup, nr_events, ctx_idp); 87 return syscall (SYS_io_setup, nr_events, ctx_idp);
82} 88}
83 89
84inline_size 90inline_size
85int 91int
86ev_io_destroy (aio_context_t ctx_id) 92evsys_io_destroy (aio_context_t ctx_id)
87{ 93{
88 return syscall (SYS_io_destroy, ctx_id); 94 return syscall (SYS_io_destroy, ctx_id);
89} 95}
90 96
91inline_size 97inline_size
92int 98int
93ev_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[])
94{ 100{
95 return syscall (SYS_io_submit, ctx_id, nr, cbp); 101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
96} 102}
97 103
98inline_size 104inline_size
99int 105int
100ev_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)
101{ 107{
102 return syscall (SYS_io_cancel, ctx_id, cbp, result); 108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
103} 109}
104 110
105inline_size 111inline_size
106int 112int
107ev_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)
108{ 114{
109 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);
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
… …
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];
152 159
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
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 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
155 170
156 if (nev) 171 if (nev)
157 { 172 {
158 iocb->io.aio_data = fd; 173 iocb->io.aio_data = fd;
159 iocb->io.aio_fildes = fd; 174 iocb->io.aio_fildes = fd;
… …
167 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
168 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
169 } 184 }
170} 185}
171 186
187#if EPOLL_FALLBACK
188
189static void
190linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
191{
192 struct epoll_event eev;
193
194 eev.events = EPOLLONESHOT;
195 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
196 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
197 eev.data.fd = iocb->aio_fildes;
198
199 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
200 ev_syserr ("(libeio) linuxaio epoll_ctl");
201}
202
203static void
204linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
205{
206 struct epoll_event events[16];
207
208 for (;;)
209 {
210 int idx;
211 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
212
213 if (expect_false (res < 0))
214 ev_syserr ("(libev) linuxaio epoll_wait");
215 else if (!res)
216 break;
217
218 for (idx = res; idx--; )
219 {
220 int fd = events [idx].data.fd;
221 uint32_t ev = events [idx].events;
222
223 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
224
225 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
226
227 fd_event (EV_A_ fd,
228 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
229 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
230 }
231
232 if (res < sizeof (events) / sizeof (events [0]))
233 break;
234 }
235}
236
237#endif
238
172static void 239static void
173linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 240linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
174{ 241{
175 while (nr) 242 while (nr)
176 { 243 {
177 int fd = ev->data; 244 int fd = ev->data;
178 int res = ev->res; 245 int res = ev->res;
179 246
180 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 247 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
181 248
182 /* linux aio is oneshot: rearm fd */ 249 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
183 linuxaio_iocbps [fd]->io.aio_buf = 0; 250 linuxaio_iocbps [fd]->io.aio_buf = 0;
184 anfds [fd].events = 0; 251 anfds [fd].events = 0;
185 fd_change (EV_A_ fd, 0); 252 fd_change (EV_A_ fd, 0);
186 253
187 /* feed events, we do not expect or handle POLLNVAL */ 254 /* feed events, we do not expect or handle POLLNVAL */
188 if (ecb_expect_false (res & POLLNVAL)) 255 if (expect_false (res & POLLNVAL))
189 fd_kill (EV_A_ fd); 256 fd_kill (EV_A_ fd);
190 else 257 else
191 fd_event ( 258 fd_event (
192 EV_A_ 259 EV_A_
193 fd, 260 fd,
… …
204static int 271static int
205linuxaio_get_events_from_ring (EV_P) 272linuxaio_get_events_from_ring (EV_P)
206{ 273{
207 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 274 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
208 275
209 ECB_MEMORY_FENCE_ACQUIRE; 276 /* the kernel reads and writes both of these variables, */
210 277 /* as a C extension, we assume that volatile use here */
211 unsigned head = ring->head; 278 /* both makes reads atomic and once-only */
279 unsigned head = *(volatile unsigned *)&ring->head;
212 unsigned tail = *(volatile unsigned *)&ring->tail; 280 unsigned tail = *(volatile unsigned *)&ring->tail;
213 281
214 if (head == tail) 282 if (head == tail)
215 return 0; 283 return 0;
216 284
285 /* bail out if the ring buffer doesn't match the expected layout */
217 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 286 if (expect_false (ring->magic != AIO_RING_MAGIC)
218 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 287 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
219 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ 288 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
220 return 0; 289 return 0;
290
291 /* make sure the events up to tail are visible */
292 ECB_MEMORY_FENCE_ACQUIRE;
221 293
222 /* parse all available events, but only once, to avoid starvation */ 294 /* parse all available events, but only once, to avoid starvation */
223 if (tail > head) /* normal case around */ 295 if (tail > head) /* normal case around */
224 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 296 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
225 else /* wrapped around */ 297 else /* wrapped around */
226 { 298 {
227 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 299 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
228 linuxaio_parse_events (EV_A_ ring->io_events, tail); 300 linuxaio_parse_events (EV_A_ ring->io_events, tail);
229 } 301 }
230 302
231 ring->head = tail; 303 /* TODO: we only need a compiler barrier here, not a read fence */
304 ECB_MEMORY_FENCE_RELEASE;
305 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
306 *(volatile unsigned *)&ring->head = tail;
307 /* make sure kernel can see our new head value - probably not required */
308 ECB_MEMORY_FENCE_RELEASE;
232 309
233 return 1; 310 return 1;
234} 311}
235 312
236/* read at least one event from kernel, or timeout */ 313/* read at least one event from kernel, or timeout */
237inline_size 314inline_size
238void 315void
239linuxaio_get_events (EV_P_ ev_tstamp timeout) 316linuxaio_get_events (EV_P_ ev_tstamp timeout)
240{ 317{
241 struct timespec ts; 318 struct timespec ts;
242 struct io_event ioev; 319 struct io_event ioev[1];
243 int res; 320 int res;
244 321
245 if (linuxaio_get_events_from_ring (EV_A)) 322 if (linuxaio_get_events_from_ring (EV_A))
246 return; 323 return;
247 324
248 /* no events, so wait for at least one, then poll ring buffer again */ 325 /* no events, so wait for at least one, then poll ring buffer again */
249 /* this degraded to one event per loop iteration */ 326 /* this degrades to one event per loop iteration */
250 /* if the ring buffer changes layout, but so be it */ 327 /* if the ring buffer changes layout, but so be it */
328
329 EV_RELEASE_CB;
251 330
252 ts.tv_sec = (long)timeout; 331 ts.tv_sec = (long)timeout;
253 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 332 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
254 333
255 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 334 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
335
336 EV_ACQUIRE_CB;
256 337
257 if (res < 0) 338 if (res < 0)
339 if (errno == EINTR)
340 /* ignored */;
341 else
258 ev_syserr ("(libev) io_getevents"); 342 ev_syserr ("(libev) linuxaio io_getevents");
259 else if (res) 343 else if (res)
260 { 344 {
261 /* at least one event received, handle it and any remaining ones in the ring buffer */ 345 /* at least one event received, handle it and any remaining ones in the ring buffer */
262 linuxaio_parse_events (EV_A_ &ioev, 1); 346 linuxaio_parse_events (EV_A_ ioev, res);
263 linuxaio_get_events_from_ring (EV_A); 347 linuxaio_get_events_from_ring (EV_A);
264 } 348 }
265} 349}
266 350
267static void 351static void
… …
274 /* io_submit might return less than the requested number of iocbs */ 358 /* io_submit might return less than the requested number of iocbs */
275 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 359 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
276 /* which allows us to pinpoint the errornous iocb */ 360 /* which allows us to pinpoint the errornous iocb */
277 for (submitted = 0; submitted < linuxaio_submitcnt; ) 361 for (submitted = 0; submitted < linuxaio_submitcnt; )
278 { 362 {
363#if 0
364 int res;
365 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
366 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
367 else
368 { res = -1; errno = EINVAL; };
369#else
279 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 370 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371#endif
280 372
281 if (ecb_expect_false (res < 0)) 373 if (expect_false (res < 0))
282 if (errno == EAGAIN) 374 if (errno == EAGAIN)
283 { 375 {
284 /* This happens when the ring buffer is full, at least. I assume this means 376 /* 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 377 * that the event was queued synchronously during io_submit, and thus
286 * the buffer overflowd. 378 * the buffer overflowed.
287 * In this case, we just try next loop iteration. 379 * In this case, we just try in next loop iteration.
380 * This should not result in a few fds taking priority, as the interface
381 * is one-shot, and we submit iocb's in a round-robin fashion.
382 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
288 */ 383 */
384 if (ecb_expect_false (submitted))
385 {
289 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 386 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
290 linuxaio_submitcnt -= submitted; 387 linuxaio_submitcnt -= submitted;
388 }
389
291 timeout = 0; 390 timeout = 0;
292 break; 391 break;
293 } 392 }
393#if EPOLL_FALLBACK
394 else if (errno == EINVAL)
395 {
396 /* This happens for unsupported fds, officially, but in my testing,
397 * also randomly happens for supported fds. We fall back to good old
398 * poll() here, under the assumption that this is a very rare case.
399 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
400 * discussion about such a case (ttys) where polling for POLLIN
401 * fails but POLLIN|POLLOUT works.
402 */
403 struct iocb *iocb = linuxaio_submits [submitted];
404 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 }
409#endif
294 else 410 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"); 411 ev_syserr ("(libev) linuxaio io_submit");
298 412
299 submitted += res; 413 submitted += res;
300 } 414 }
301 415
302 linuxaio_submitcnt = 0; 416 linuxaio_submitcnt = 0;
… …
310int 424int
311linuxaio_init (EV_P_ int flags) 425linuxaio_init (EV_P_ int flags)
312{ 426{
313 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 427 /* 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 */ 428 /* 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 */ 429#if EPOLL_FALLBACK
430 /* 4.19 made epoll work */
431 if (ev_linux_version () < 0x041300)
316 return 0; 432 return 0;
433#else
434 /* 4.18 introduced IOCB_CMD_POLL */
435 if (ev_linux_version () < 0x041200)
436 return 0;
437#endif
317 438
318 linuxaio_ctx = 0; 439 linuxaio_ctx = 0;
319 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 440 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
320 return 0; 441 return 0;
442
443#if EPOLL_FALLBACK
444 backend_fd = ev_epoll_create ();
445 if (backend_fd < 0)
446 {
447 evsys_io_destroy (linuxaio_ctx);
448 return 0;
449 }
450
451 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
452 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
453 ev_io_start (EV_A_ &linuxaio_epoll_w);
454 ev_unref (EV_A); /* watcher should not keep loop alive */
455#endif
321 456
322 backend_modify = linuxaio_modify; 457 backend_modify = linuxaio_modify;
323 backend_poll = linuxaio_poll; 458 backend_poll = linuxaio_poll;
324 459
325 linuxaio_iocbpmax = 0; 460 linuxaio_iocbpmax = 0;
… …
334 469
335inline_size 470inline_size
336void 471void
337linuxaio_destroy (EV_P) 472linuxaio_destroy (EV_P)
338{ 473{
474#if EPOLL_FALLBACK
475 close (backend_fd);
476#endif
339 linuxaio_free_iocbp (EV_A); 477 linuxaio_free_iocbp (EV_A);
340 ev_io_destroy (linuxaio_ctx); 478 ev_io_destroy (linuxaio_ctx);
341} 479}
342 480
343inline_size 481inline_size
344void 482void
345linuxaio_fork (EV_P) 483linuxaio_fork (EV_P)
346{ 484{
347 /* TODO: verify and test */
348
349 /* this frees all iocbs, which is very heavy-handed */ 485 /* this frees all iocbs, which is very heavy-handed */
350 linuxaio_destroy (EV_A); 486 linuxaio_destroy (EV_A);
351 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 487 linuxaio_submitcnt = 0; /* all pointers were invalidated */
352 488
353 linuxaio_ctx = 0; 489 linuxaio_ctx = 0;
354 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 490 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
355 ev_syserr ("(libev) io_setup"); 491 ev_syserr ("(libev) linuxaio io_setup");
492
493#if EPOLL_FALLBACK
494 while ((backend_fd = ev_epoll_create ()) < 0)
495 ev_syserr ("(libev) linuxaio epoll_create");
496
497 ev_io_stop (EV_A_ &linuxaio_epoll_w);
498 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
499 ev_io_start (EV_A_ &linuxaio_epoll_w);
500#endif
356 501
357 fd_rearm_all (EV_A); 502 fd_rearm_all (EV_A);
358} 503}
359 504

Diff Legend

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