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 1.1 by root, Thu Jun 20 22:44:59 2019 UTC vs.
Revision 1.17 by root, Mon Jun 24 02:02:35 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 */
43#include <poll.h>
41#include <linux/aio_abi.h> 44#include <linux/aio_abi.h>
42 45
46#if EPOLL_FALLBACK
47# include <sys/epoll.h>
48#endif
49
43/* we try to fill 4kn pages exactly. 50/* we try to fill 4kB pages exactly.
44 * 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.
45 * 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.
46 * 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
47 */ 54 */
48#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 */
49 56
50/*****************************************************************************/ 57/*****************************************************************************/
51/* syscall wrapdadoop */ 58/* syscall wrapdadoop */
52 59
53#include <sys/syscall.h> /* no glibc wrappers */ 60#include <sys/syscall.h> /* no glibc wrappers */
54 61
55/* aio_abi.h is not verioned 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 */
56#define IOCB_CMD_POLL 5 63#define IOCB_CMD_POLL 5
57 64
58/* taken from linux/fs/aio.c */ 65/* taken from linux/fs/aio.c */
59#define AIO_RING_MAGIC 0xa10a10a1 66#define AIO_RING_MAGIC 0xa10a10a1
60#define AIO_RING_INCOMPAT_FEATURES 0 67#define AIO_RING_INCOMPAT_FEATURES 0
… …
71 unsigned header_length; /* size of aio_ring */ 78 unsigned header_length; /* size of aio_ring */
72 79
73 struct io_event io_events[0]; 80 struct io_event io_events[0];
74}; 81};
75 82
76static int 83inline_size
84int
77ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
78{ 86{
79 return syscall (SYS_io_setup, nr_events, ctx_idp); 87 return syscall (SYS_io_setup, nr_events, ctx_idp);
80} 88}
81 89
82static int 90inline_size
91int
83ev_io_destroy (aio_context_t ctx_id) 92ev_io_destroy (aio_context_t ctx_id)
84{ 93{
85 return syscall (SYS_io_destroy, ctx_id); 94 return syscall (SYS_io_destroy, ctx_id);
86} 95}
87 96
88static int 97inline_size
98int
89ev_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[])
90{ 100{
91 return syscall (SYS_io_submit, ctx_id, nr, cbp); 101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
92} 102}
93 103
94static int 104inline_size
105int
95ev_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)
96{ 107{
97 return syscall (SYS_io_cancel, ctx_id, cbp, result); 108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
98} 109}
99 110
100static int 111inline_size
112int
101ev_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)
102{ 114{
103 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);
104} 116}
105
106typedef void (*ev_io_cb) (long nr, struct io_event *events);
107 117
108/*****************************************************************************/ 118/*****************************************************************************/
109/* actual backed implementation */ 119/* actual backed implementation */
110 120
111/* 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" */
112typedef struct aniocb 122typedef struct aniocb
113{ 123{
114 struct iocb io; 124 struct iocb io;
115 /*int inuse;*/ 125 /*int inuse;*/
116} *ANIOCBP; 126} *ANIOCBP;
117 127
118inline_size 128inline_size
119void 129void
120linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count)
121{ 131{
122 /* TODO: quite the overhead to allocate every iocb separately */ 132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
123 while (count--) 133 while (count--)
124 { 134 {
125 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 135 *base = (ANIOCBP)ev_malloc (sizeof (**base));
136 /* TODO: full zero initialize required? */
126 memset (*base, 0, sizeof (**base)); 137 memset (*base, 0, sizeof (**base));
127 /* 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 */
128 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
129 ++base; 140 ++base;
130 } 141 }
131} 142}
132 143
144ecb_cold
133static void 145static void
134linuxaio_free_iocbp (EV_P) 146linuxaio_free_iocbp (EV_P)
135{ 147{
136 while (linuxaio_iocbpmax--) 148 while (linuxaio_iocbpmax--)
137 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); 149 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
138 150
139 linuxaio_iocbpmax = 0; 151 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
140} 152}
141 153
142static void 154static void
143linuxaio_modify (EV_P_ int fd, int oev, int nev) 155linuxaio_modify (EV_P_ int fd, int oev, int nev)
144{ 156{
145 /* TODO: full zero initialize required? */
146 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);
147 struct aniocb *iocb = linuxaio_iocbps [fd]; 158 struct aniocb *iocb = linuxaio_iocbps [fd];
148 159
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
149 if (iocb->io.aio_buf) 168 if (iocb->io.aio_buf)
150 ev_io_cancel (linuxaio_ctx, &iocb->io, (void *)0); 169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
151 170
152 if (nev) 171 if (nev)
153 { 172 {
154 iocb->io.aio_data = fd; 173 iocb->io.aio_data = fd;
155 iocb->io.aio_fildes = fd; 174 iocb->io.aio_fildes = fd;
… …
171 while (nr) 190 while (nr)
172 { 191 {
173 int fd = ev->data; 192 int fd = ev->data;
174 int res = ev->res; 193 int res = ev->res;
175 194
176 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdxmax)); 195 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
177 196
178 /* linux aio is oneshot: rearm fd */ 197 /* linux aio is oneshot: rearm fd */
179 linuxaio_iocbps [fd]->io.aio_buf = 0; 198 linuxaio_iocbps [fd]->io.aio_buf = 0;
180 anfds [fd].events = 0; 199 anfds [fd].events = 0;
181 fd_change (EV_A_ fd, 0); 200 fd_change (EV_A_ fd, 0);
182 201
183 /* feed events, we do not expect or handle POLLNVAL */ 202 /* feed events, we do not expect or handle POLLNVAL */
184 if (ecb_expect_false (res & POLLNVAL)) 203 if (expect_false (res & POLLNVAL))
185 fd_kill (EV_A_ fd); 204 fd_kill (EV_A_ fd);
186 else 205 else
187 fd_event ( 206 fd_event (
188 EV_A_ 207 EV_A_
189 fd, 208 fd,
… …
200static int 219static int
201linuxaio_get_events_from_ring (EV_P) 220linuxaio_get_events_from_ring (EV_P)
202{ 221{
203 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
204 223
224 /* the kernel reads and writes both of these variables, */
225 /* as a C extension, we assume that volatile use here */
226 /* both makes reads atomic and once-only */
227 unsigned head = *(volatile unsigned *)&ring->head;
228 unsigned tail = *(volatile unsigned *)&ring->tail;
229
230 if (head == tail)
231 return 0;
232
233 /* bail out if the ring buffer doesn't match the expected layout */
234 if (expect_false (ring->magic != AIO_RING_MAGIC)
235 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
236 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
237 return 0;
238
239 /* make sure the events up to tail are visible */
205 ECB_MEMORY_FENCE_ACQUIRE; 240 ECB_MEMORY_FENCE_ACQUIRE;
206
207 unsigned head = ring->head;
208 unsigned tail = *(volatile unsigned *)&ring->tail;
209
210 if (ring->magic != AIO_RING_MAGIC
211 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
212 || ring->header_length != sizeof (struct aio_ring) /* TODO: or use it to find io_event[0]? */
213 || head == tail)
214 return 0;
215 241
216 /* parse all available events, but only once, to avoid starvation */ 242 /* parse all available events, but only once, to avoid starvation */
217 if (tail > head) /* normal case around */ 243 if (tail > head) /* normal case around */
218 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 244 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
219 else
220 {
221 /* wrapped around */ 245 else /* wrapped around */
246 {
222 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);
223 linuxaio_parse_events (EV_A_ ring->io_events, tail); 248 linuxaio_parse_events (EV_A_ ring->io_events, tail);
224 } 249 }
225 250
226 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;
227 257
228 return 1; 258 return 1;
229} 259}
230 260
231/* read at least one event from kernel, or timeout */ 261/* read at least one event from kernel, or timeout */
… …
239 269
240 if (linuxaio_get_events_from_ring (EV_A)) 270 if (linuxaio_get_events_from_ring (EV_A))
241 return; 271 return;
242 272
243 /* 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 */
244 /* this degraded to one event per loop iteration */ 274 /* this degrades to one event per loop iteration */
245 /* if the ring buffer changes layout, but so be it */ 275 /* if the ring buffer changes layout, but so be it */
246 276
247 ts.tv_sec = (long)timeout; 277 ts.tv_sec = (long)timeout;
248 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 278 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
249 279
250 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 280 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);
251 281
252 if (res < 0) 282 if (res < 0)
283 if (errno == EINTR)
284 /* ignored */;
285 else
253 ev_syserr ("(libev) io_getevents"); 286 ev_syserr ("(libev) linuxaio io_getevents");
254 else if (res) 287 else if (res)
255 { 288 {
256 /* 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 */
257 linuxaio_parse_events (EV_A_ &ioev, 1); 290 linuxaio_parse_events (EV_A_ &ioev, 1);
258 linuxaio_get_events_from_ring (EV_A); 291 linuxaio_get_events_from_ring (EV_A);
259 } 292 }
260} 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
261 310
262static void 311static void
263linuxaio_poll (EV_P_ ev_tstamp timeout) 312linuxaio_poll (EV_P_ ev_tstamp timeout)
264{ 313{
265 int submitted; 314 int submitted;
… …
271 /* which allows us to pinpoint the errornous iocb */ 320 /* which allows us to pinpoint the errornous iocb */
272 for (submitted = 0; submitted < linuxaio_submitcnt; ) 321 for (submitted = 0; submitted < linuxaio_submitcnt; )
273 { 322 {
274 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);
275 324
276 if (res < 0) 325 if (expect_false (res < 0))
277 if (errno == EAGAIN) 326 if (errno == EAGAIN)
278 { 327 {
279 /* 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
280 * that the event was queued synchronously during io_submit, and thus 329 * that the event was queued synchronously during io_submit, and thus
281 * the buffer overflowd. 330 * the buffer overflowed.
282 * 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.
283 */ 334 */
284 memcpy (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 335 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
285 linuxaio_submitcnt -= submitted; 336 linuxaio_submitcnt -= submitted;
286 timeout = 0; 337 timeout = 0;
287 break; 338 break;
288 } 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
289 else 356 else
290 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
291 /* TODO: should we always just try next time? */
292 ev_syserr ("(libev) io_submit"); 357 ev_syserr ("(libev) linuxaio io_submit");
293 358
294 submitted += res; 359 submitted += res;
295 } 360 }
296 361
297 linuxaio_submitcnt = 0; 362 linuxaio_submitcnt = 0;
… …
299 /* second phase: fetch and parse events */ 364 /* second phase: fetch and parse events */
300 365
301 linuxaio_get_events (EV_A_ timeout); 366 linuxaio_get_events (EV_A_ timeout);
302} 367}
303 368
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
406
304inline_size 407inline_size
305int 408int
306linuxaio_init (EV_P_ int flags) 409linuxaio_init (EV_P_ int flags)
307{ 410{
308 /* 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 */
309 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 412 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
413#if EPOLL_FALLBACK
414 /* 4.19 made epoll work */
415 if (ev_linux_version () < 0x041300)
310 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
311 422
423 linuxaio_ctx = 0;
312 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 424 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
313 return 0; 425 return 0;
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
314 439
315 backend_modify = linuxaio_modify; 440 backend_modify = linuxaio_modify;
316 backend_poll = linuxaio_poll; 441 backend_poll = linuxaio_poll;
317 442
318 linuxaio_iocbpmax = 0; 443 linuxaio_iocbpmax = 0;
… …
327 452
328inline_size 453inline_size
329void 454void
330linuxaio_destroy (EV_P) 455linuxaio_destroy (EV_P)
331{ 456{
457#if EPOLL_FALLBACK
458 close (backend_fd);
459#endif
332 linuxaio_free_iocbp (EV_A); 460 linuxaio_free_iocbp (EV_A);
333 ev_io_destroy (linuxaio_ctx); 461 ev_io_destroy (linuxaio_ctx);
334} 462}
335 463
336inline_size 464inline_size
337void 465void
338linuxaio_fork (EV_P) 466linuxaio_fork (EV_P)
339{ 467{
340 abort ();//D 468 /* this frees all iocbs, which is very heavy-handed */
341} 469 linuxaio_destroy (EV_A);
470 linuxaio_submitcnt = 0; /* all pointers were invalidated */
342 471
472 linuxaio_ctx = 0;
473 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
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
484
485 fd_rearm_all (EV_A);
486}
487

Diff Legend

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