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 1.10 by root, Sun Jun 23 02:02:24 2019 UTC vs.
Revision 1.47 by root, Mon Jul 8 02:45:43 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 40/*
41 * general notes about linux aio:
42 *
43 * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
44 * 4.18 looks too good to be true: both watchers and events can be
45 * batched, and events can even be handled in userspace using
46 * a ring buffer shared with the kernel. watchers can be canceled
47 * regardless of whether the fd has been closed. no problems with fork.
48 * ok, the ring buffer is 200% undocumented (there isn't even a
49 * header file), but otherwise, it's pure bliss!
50 * b) ok, watchers are one-shot, so you have to re-arm active ones
51 * on every iteration. so much for syscall-less event handling,
52 * but at least these re-arms can be batched, no big deal, right?
53 * c) well, linux as usual: the documentation lies to you: io_submit
54 * sometimes returns EINVAL because the kernel doesn't feel like
55 * handling your poll mask - ttys can be polled for POLLOUT,
56 * POLLOUT|POLLIN, but polling for POLLIN fails. just great,
57 * so we have to fall back to something else (hello, epoll),
58 * but at least the fallback can be slow, because these are
59 * exceptional cases, right?
60 * d) hmm, you have to tell the kernel the maximum number of watchers
61 * you want to queue when initialising the aio context. but of
62 * course the real limit is magically calculated in the kernel, and
63 * is often higher then we asked for. so we just have to destroy
64 * the aio context and re-create it a bit larger if we hit the limit.
65 * (starts to remind you of epoll? well, it's a bit more deterministic
66 * and less gambling, but still ugly as hell).
67 * e) that's when you find out you can also hit an arbitrary system-wide
68 * limit. or the kernel simply doesn't want to handle your watchers.
69 * what the fuck do we do then? you guessed it, in the middle
70 * of event handling we have to switch to 100% epoll polling. and
71 * that better is as fast as normal epoll polling, so you practically
72 * have to use the normal epoll backend with all its quirks.
73 * f) end result of this train wreck: it inherits all the disadvantages
74 * from epoll, while adding a number on its own. why even bother to use
75 * it? because if conditions are right and your fds are supported and you
76 * don't hit a limit, this backend is actually faster, doesn't gamble with
77 * your fds, batches watchers and events and doesn't require costly state
78 * recreates. well, until it does.
79 * g) all of this makes this backend use almost twice as much code as epoll.
80 * which in turn uses twice as much code as poll. and that#s not counting
81 * the fact that this backend also depends on the epoll backend, making
82 * it three times as much code as poll, or kqueue.
83 * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
84 * it's clear that whatever linux comes up with is far, far, far worse.
85 */
41 86
42#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ 87#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
43#include <poll.h> 88#include <poll.h>
44#include <linux/aio_abi.h> 89#include <linux/aio_abi.h>
45 90
46#if EPOLL_FALLBACK
47# include <sys/epoll.h>
48#endif
49
50/* we try to fill 4kB pages exactly.
51 * the ring buffer header is 32 bytes, every io event is 32 bytes.
52 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
53 * therefore the calculation below will use "exactly" 4kB for the ring buffer
54 */
55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56
57/*****************************************************************************/ 91/*****************************************************************************/
58/* syscall wrapdadoop */ 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
59 93
60#include <sys/syscall.h> /* no glibc wrappers */ 94#include <sys/syscall.h> /* no glibc wrappers */
61 95
62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */ 96/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63#define IOCB_CMD_POLL 5 97#define IOCB_CMD_POLL 5
64 98
65/* taken from linux/fs/aio.c */ 99/* taken from linux/fs/aio.c. yup, that's a .c file.
100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for.
103 */
66#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
67#define AIO_RING_INCOMPAT_FEATURES 0 105#define EV_AIO_RING_INCOMPAT_FEATURES 0
68struct aio_ring 106struct aio_ring
69{ 107{
70 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
71 unsigned nr; /* number of io_events */ 109 unsigned nr; /* number of io_events */
72 unsigned head; /* Written to by userland or by kernel. */ 110 unsigned head; /* Written to by userland or by kernel. */
… …
80 struct io_event io_events[0]; 118 struct io_event io_events[0];
81}; 119};
82 120
83inline_size 121inline_size
84int 122int
85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86{ 124{
87 return syscall (SYS_io_setup, nr_events, ctx_idp); 125 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
88} 126}
89 127
90inline_size 128inline_size
91int 129int
92ev_io_destroy (aio_context_t ctx_id) 130evsys_io_destroy (aio_context_t ctx_id)
93{ 131{
94 return syscall (SYS_io_destroy, ctx_id); 132 return ev_syscall1 (SYS_io_destroy, ctx_id);
95} 133}
96 134
97inline_size 135inline_size
98int 136int
99ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 137evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100{ 138{
101 return syscall (SYS_io_submit, ctx_id, nr, cbp); 139 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
102} 140}
103 141
104inline_size 142inline_size
105int 143int
106ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 144evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
107{ 145{
108 return syscall (SYS_io_cancel, ctx_id, cbp, result); 146 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
109} 147}
110 148
111inline_size 149inline_size
112int 150int
113ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 151evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
114{ 152{
115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 153 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116} 154}
117 155
118/*****************************************************************************/ 156/*****************************************************************************/
119/* actual backed implementation */ 157/* actual backed implementation */
120 158
159ecb_cold
160static int
161linuxaio_nr_events (EV_P)
162{
163 /* we start with 16 iocbs and incraese from there
164 * that's tiny, but the kernel has a rather low system-wide
165 * limit that can be reached quickly, so let's be parsimonious
166 * with this resource.
167 * Rest assured, the kernel generously rounds up small and big numbers
168 * in different ways (but doesn't seem to charge you for it).
169 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
170 * and this helps to take advantage of that limit.
171 */
172
173 /* we try to fill 4kB pages exactly.
174 * the ring buffer header is 32 bytes, every io event is 32 bytes.
175 * the kernel takes the io requests number, doubles it, adds 2
176 * and adds the ring buffer.
177 * the way we use this is by starting low, and then roughly doubling the
178 * size each time we hit a limit.
179 */
180
181 int requests = 15 << linuxaio_iteration;
182 int one_page = (4096
183 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
184 int first_page = ((4096 - sizeof (struct aio_ring))
185 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
186
187 /* if everything fits into one page, use count exactly */
188 if (requests > first_page)
189 /* otherwise, round down to full pages and add the first page */
190 requests = requests / one_page * one_page + first_page;
191
192 return requests;
193}
194
121/* we use out own wrapper structure in acse we ever want to do something "clever" */ 195/* we use out own wrapper structure in case we ever want to do something "clever" */
122typedef struct aniocb 196typedef struct aniocb
123{ 197{
124 struct iocb io; 198 struct iocb io;
125 /*int inuse;*/ 199 /*int inuse;*/
126} *ANIOCBP; 200} *ANIOCBP;
127 201
128inline_size 202inline_size
129void 203void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 204linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 205{
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--) 206 while (count--)
134 { 207 {
208 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
135 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 209 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 /* TODO: full zero initialize required? */ 210
211 /* full zero initialise is probably not required at the moment, but
212 * this is not well documented, so we better do it.
213 */
137 memset (*base, 0, sizeof (**base)); 214 memset (iocb, 0, sizeof (*iocb));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 215
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 216 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base; 217 iocb->io.aio_data = offset;
218 iocb->io.aio_fildes = offset;
219
220 base [offset++] = iocb;
141 } 221 }
142} 222}
143 223
144ecb_cold 224ecb_cold
145static void 225static void
… …
153 233
154static void 234static void
155linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
156{ 236{
157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 237 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158 struct aniocb *iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
159 239
160#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0) 240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
162 { 241 {
242 /* we handed this fd over to epoll, so undo this first */
243 /* we do it manually because the optimisations on epoll_modify won't do us any good */
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
245 anfds [fd].emask = 0;
164 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
165 } 247 }
166#endif 248 else if (ecb_expect_false (iocb->io.aio_buf))
249 {
250 /* iocb active, so cancel it first before resubmit */
251 for (;;)
252 {
253 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
254 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
255 break;
167 256
257 if (ecb_expect_true (errno == EINPROGRESS))
258 break;
259
260 /* the EINPROGRESS test is for nicer error message. clumsy. */
261 if (errno != EINTR)
262 {
263 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
264 break;
265 }
266 }
267 }
268
168 if (iocb->io.aio_buf) 269 iocb->io.aio_buf =
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 270 (nev & EV_READ ? POLLIN : 0)
271 | (nev & EV_WRITE ? POLLOUT : 0);
170 272
171 if (nev) 273 if (nev)
172 { 274 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0);
178
179 /* queue iocb up for io_submit */ 275 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */ 276 /* this assumes we only ever get one call per fd per loop iteration */
181 ++linuxaio_submitcnt; 277 ++linuxaio_submitcnt;
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 278 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 279 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 280 }
185} 281}
186 282
187static void 283static void
284linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
285{
286 epoll_poll (EV_A_ 0);
287}
288
289inline_speed
290void
291linuxaio_fd_rearm (EV_P_ int fd)
292{
293 anfds [fd].events = 0;
294 linuxaio_iocbps [fd]->io.aio_buf = 0;
295 fd_change (EV_A_ fd, EV_ANFD_REIFY);
296}
297
298static void
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 299linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 300{
190 while (nr) 301 while (nr)
191 { 302 {
192 int fd = ev->data; 303 int fd = ev->data;
193 int res = ev->res; 304 int res = ev->res;
194 305
195 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 306 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
196 307
197 /* linux aio is oneshot: rearm fd */
198 linuxaio_iocbps [fd]->io.aio_buf = 0;
199 anfds [fd].events = 0;
200 fd_change (EV_A_ fd, 0);
201
202 /* feed events, we do not expect or handle POLLNVAL */ 308 /* feed events, we do not expect or handle POLLNVAL */
203 if (ecb_expect_false (res & POLLNVAL))
204 fd_kill (EV_A_ fd);
205 else
206 fd_event ( 309 fd_event (
207 EV_A_ 310 EV_A_
208 fd, 311 fd,
209 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 312 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
210 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 313 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
211 ); 314 );
315
316 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
317 linuxaio_fd_rearm (EV_A_ fd);
212 318
213 --nr; 319 --nr;
214 ++ev; 320 ++ev;
215 } 321 }
216} 322}
217 323
218/* get any events from ringbuffer, return true if any were handled */ 324/* get any events from ring buffer, return true if any were handled */
219static int 325static int
220linuxaio_get_events_from_ring (EV_P) 326linuxaio_get_events_from_ring (EV_P)
221{ 327{
222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 328 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
329 unsigned head, tail;
223 330
224 unsigned head = ring->head; 331 /* the kernel reads and writes both of these variables, */
332 /* as a C extension, we assume that volatile use here */
333 /* both makes reads atomic and once-only */
334 head = *(volatile unsigned *)&ring->head;
335 ECB_MEMORY_FENCE_ACQUIRE;
225 unsigned tail = *(volatile unsigned *)&ring->tail; 336 tail = *(volatile unsigned *)&ring->tail;
226 337
227 if (head == tail) 338 if (head == tail)
228 return 0; 339 return 0;
229 340
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 ECB_MEMORY_FENCE_ACQUIRE;
237
238 /* parse all available events, but only once, to avoid starvation */ 341 /* parse all available events, but only once, to avoid starvation */
239 if (tail > head) /* normal case around */ 342 if (ecb_expect_true (tail > head)) /* normal case around */
240 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 343 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
241 else /* wrapped around */ 344 else /* wrapped around */
242 { 345 {
243 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 346 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
244 linuxaio_parse_events (EV_A_ ring->io_events, tail); 347 linuxaio_parse_events (EV_A_ ring->io_events, tail);
245 } 348 }
246 349
247 ring->head = tail; 350 ECB_MEMORY_FENCE_RELEASE;
351 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
352 *(volatile unsigned *)&ring->head = tail;
248 353
249 return 1; 354 return 1;
355}
356
357inline_size
358int
359linuxaio_ringbuf_valid (EV_P)
360{
361 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
362
363 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
364 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
365 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
250} 366}
251 367
252/* read at least one event from kernel, or timeout */ 368/* read at least one event from kernel, or timeout */
253inline_size 369inline_size
254void 370void
255linuxaio_get_events (EV_P_ ev_tstamp timeout) 371linuxaio_get_events (EV_P_ ev_tstamp timeout)
256{ 372{
257 struct timespec ts; 373 struct timespec ts;
258 struct io_event ioev; 374 struct io_event ioev[8]; /* 256 octet stack space */
259 int res; 375 int want = 1; /* how many events to request */
376 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
260 377
378 if (ecb_expect_true (ringbuf_valid))
379 {
380 /* if the ring buffer has any events, we don't wait or call the kernel at all */
261 if (linuxaio_get_events_from_ring (EV_A)) 381 if (linuxaio_get_events_from_ring (EV_A))
262 return; 382 return;
263 383
264 /* no events, so wait for at least one, then poll ring buffer again */ 384 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
265 /* this degrades to one event per loop iteration */ 385 if (!timeout)
266 /* if the ring buffer changes layout, but so be it */ 386 return;
387 }
388 else
389 /* no ringbuffer, request slightly larger batch */
390 want = sizeof (ioev) / sizeof (ioev [0]);
267 391
268 ts.tv_sec = (long)timeout; 392 /* no events, so wait for some
269 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 393 * for fairness reasons, we do this in a loop, to fetch all events
394 */
395 for (;;)
396 {
397 int res;
270 398
399 EV_RELEASE_CB;
400
401 EV_TS_SET (ts, timeout);
271 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 402 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
272 403
404 EV_ACQUIRE_CB;
405
273 if (res < 0) 406 if (res < 0)
274 if (errno == EINTR) 407 if (errno == EINTR)
275 /* ignored */; 408 /* ignored, retry */;
276 else 409 else
277 ev_syserr ("(libev) linuxaio io_getevents"); 410 ev_syserr ("(libev) linuxaio io_getevents");
278 else if (res) 411 else if (res)
279 { 412 {
280 /* at least one event received, handle it and any remaining ones in the ring buffer */ 413 /* at least one event available, handle them */
281 linuxaio_parse_events (EV_A_ &ioev, 1); 414 linuxaio_parse_events (EV_A_ ioev, res);
415
416 if (ecb_expect_true (ringbuf_valid))
417 {
418 /* if we have a ring buffer, handle any remaining events in it */
282 linuxaio_get_events_from_ring (EV_A); 419 linuxaio_get_events_from_ring (EV_A);
283 }
284}
285 420
286#if EPOLL_FALLBACK 421 /* at this point, we should have handled all outstanding events */
287static void 422 break;
288linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) 423 }
289{ 424 else if (res < want)
290 struct epoll_event eev; 425 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
426 break;
427 }
428 else
429 break; /* no events from the kernel, we are done */
291 430
292 eev.events = EPOLLONESHOT; 431 timeout = 0; /* only wait in the first iteration */
293 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; 432 }
294 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
295 eev.data.fd = iocb->aio_fildes;
296
297 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
298 ev_syserr ("(libeio) linuxaio epoll_ctl");
299} 433}
300#endif 434
435inline_size
436int
437linuxaio_io_setup (EV_P)
438{
439 linuxaio_ctx = 0;
440 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
441}
301 442
302static void 443static void
303linuxaio_poll (EV_P_ ev_tstamp timeout) 444linuxaio_poll (EV_P_ ev_tstamp timeout)
304{ 445{
305 int submitted; 446 int submitted;
306 447
307 /* first phase: submit new iocbs */ 448 /* first phase: submit new iocbs */
308 449
309 /* io_submit might return less than the requested number of iocbs */ 450 /* io_submit might return less than the requested number of iocbs */
310 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 451 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
311 /* which allows us to pinpoint the errornous iocb */ 452 /* which allows us to pinpoint the erroneous iocb */
312 for (submitted = 0; submitted < linuxaio_submitcnt; ) 453 for (submitted = 0; submitted < linuxaio_submitcnt; )
313 { 454 {
314 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 455 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
315 456
316 if (ecb_expect_false (res < 0)) 457 if (ecb_expect_false (res < 0))
317 if (errno == EAGAIN) 458 if (errno == EINVAL)
318 { 459 {
319 /* This happens when the ring buffer is full, at least. I assume this means
320 * that the event was queued synchronously during io_submit, and thus
321 * the buffer overflowed.
322 * In this case, we just try in next loop iteration.
323 * This should not result in a few fds taking priority, as the interface
324 * is one-shot, and we submit iocb's in a round-robin fashion.
325 */
326 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
327 linuxaio_submitcnt -= submitted;
328 timeout = 0;
329 break;
330 }
331#if EPOLL_FALLBACK
332 else if (errno == EINVAL)
333 {
334 /* This hapΓΌpens for unsupported fds, officially, but in my testing, 460 /* This happens for unsupported fds, officially, but in my testing,
335 * also randomly happens for supported fds. We fall back to good old 461 * also randomly happens for supported fds. We fall back to good old
336 * poll() here, under the assumption that this is a very rare case. 462 * poll() here, under the assumption that this is a very rare case.
463 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
464 * discussion about such a case (ttys) where polling for POLLIN
465 * fails but POLLIN|POLLOUT works.
337 */ 466 */
338 struct iocb *iocb = linuxaio_submits [submitted]; 467 struct iocb *iocb = linuxaio_submits [submitted];
468 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
469 iocb->aio_reqprio = -1; /* mark iocb as epoll */
470
471 res = 1; /* skip this iocb - another iocb, another chance */
472 }
473 else if (errno == EAGAIN)
474 {
475 /* This happens when the ring buffer is full, or some other shit we
476 * don't know and isn't documented. Most likely because we have too
477 * many requests and linux aio can't be assed to handle them.
478 * In this case, we try to allocate a larger ring buffer, freeing
479 * ours first. This might fail, in which case we have to fall back to 100%
480 * epoll.
481 * God, how I hate linux not getting its act together. Ever.
482 */
483 evsys_io_destroy (linuxaio_ctx);
484 linuxaio_submitcnt = 0;
485
486 /* rearm all fds with active iocbs */
487 {
488 int fd;
489 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
490 if (linuxaio_iocbps [fd]->io.aio_buf)
491 linuxaio_fd_rearm (EV_A_ fd);
492 }
493
494 ++linuxaio_iteration;
495 if (linuxaio_io_setup (EV_A) < 0)
496 {
497 /* TODO: rearm all and recreate epoll backend from scratch */
498 /* TODO: might be more prudent? */
499
500 /* to bad, we can't get a new aio context, go 100% epoll */
501 linuxaio_free_iocbp (EV_A);
502 ev_io_stop (EV_A_ &linuxaio_epoll_w);
503 ev_ref (EV_A);
504 linuxaio_ctx = 0;
505
506 backend = EVBACKEND_EPOLL;
507 backend_modify = epoll_modify;
508 backend_poll = epoll_poll;
509 }
510
511 timeout = 0;
512 /* it's easiest to handle this mess in another iteration */
513 return;
514 }
515 else if (errno == EBADF)
516 {
517 assert (("libev: event loop rejected bad fd", errno != EBADF));
518 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
519
339 res = 1; /* skip this iocb */ 520 res = 1; /* skip this iocb */
340
341 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
342 iocb->aio_reqprio = -1; /* mark iocb as epoll */
343 } 521 }
344#endif 522 else if (errno == EINTR) /* not seen in reality, not documented */
523 res = 0; /* silently ignore and retry */
345 else 524 else
525 {
346 ev_syserr ("(libev) linuxaio io_submit"); 526 ev_syserr ("(libev) linuxaio io_submit");
527 res = 0;
528 }
347 529
348 submitted += res; 530 submitted += res;
349 } 531 }
350 532
351 linuxaio_submitcnt = 0; 533 linuxaio_submitcnt = 0;
352 534
353 /* second phase: fetch and parse events */ 535 /* second phase: fetch and parse events */
354 536
355 linuxaio_get_events (EV_A_ timeout); 537 linuxaio_get_events (EV_A_ timeout);
356} 538}
357
358#if EPOLL_FALLBACK
359
360static void
361linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
362{
363 struct epoll_event events[16];
364
365 for (;;)
366 {
367 int idx;
368 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
369
370 if (ecb_expect_false (res < 0))
371 ev_syserr ("(libev) linuxaio epoll_wait");
372 else if (!res)
373 break;
374
375 for (idx = res; idx--; )
376 {
377 int fd = events [idx].data.fd;
378 uint32_t ev = events [idx].events;
379
380 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
381
382 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
383
384 fd_event (EV_A_ fd,
385 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
386 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
387 }
388
389 if (res < sizeof (events) / sizeof (events [0]))
390 break;
391 }
392}
393
394#endif
395 539
396inline_size 540inline_size
397int 541int
398linuxaio_init (EV_P_ int flags) 542linuxaio_init (EV_P_ int flags)
399{ 543{
400 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 544 /* would be great to have a nice test for IOCB_CMD_POLL instead */
401 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 545 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
402 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 546 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
547 if (ev_linux_version () < 0x041300)
403 return 0; 548 return 0;
404 549
405 linuxaio_ctx = 0; 550 if (!epoll_init (EV_A_ 0))
406 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
407 return 0; 551 return 0;
408 552
409#if EPOLL_FALLBACK 553 linuxaio_iteration = 0;
410 backend_fd = ev_epoll_create (); 554
411 if (backend_fd < 0) 555 if (linuxaio_io_setup (EV_A) < 0)
412 { 556 {
413 ev_io_destroy (linuxaio_ctx); 557 epoll_destroy (EV_A);
414 return 0; 558 return 0;
415 } 559 }
416 560
417 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 561 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
562 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
418 ev_io_start (EV_A_ &linuxaio_epoll_w); 563 ev_io_start (EV_A_ &linuxaio_epoll_w);
419#endif 564 ev_unref (EV_A); /* watcher should not keep loop alive */
420 565
421 backend_modify = linuxaio_modify; 566 backend_modify = linuxaio_modify;
422 backend_poll = linuxaio_poll; 567 backend_poll = linuxaio_poll;
423 568
424 linuxaio_iocbpmax = 0; 569 linuxaio_iocbpmax = 0;
425 linuxaio_iocbps = 0; 570 linuxaio_iocbps = 0;
426 571
427 linuxaio_submits = 0; 572 linuxaio_submits = 0;
… …
433 578
434inline_size 579inline_size
435void 580void
436linuxaio_destroy (EV_P) 581linuxaio_destroy (EV_P)
437{ 582{
438#if EPOLL_FALLBACK 583 epoll_destroy (EV_A);
439 close (backend_fd);
440#endif
441 linuxaio_free_iocbp (EV_A); 584 linuxaio_free_iocbp (EV_A);
442 ev_io_destroy (linuxaio_ctx); 585 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
443} 586}
444 587
445inline_size 588ecb_cold
446void 589static void
447linuxaio_fork (EV_P) 590linuxaio_fork (EV_P)
448{ 591{
449 /* this frees all iocbs, which is very heavy-handed */
450 linuxaio_destroy (EV_A);
451 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 592 linuxaio_submitcnt = 0; /* all pointers were invalidated */
593 linuxaio_free_iocp (EV_A); /* this frees all iocbs, which is very heavy-handed */
594 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
452 595
453 linuxaio_ctx = 0; 596 linuxaio_iteration = 0; /* we start over in the child */
454 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 597
598 while (linuxaio_io_setup (EV_A) < 0)
455 ev_syserr ("(libev) linuxaio io_setup"); 599 ev_syserr ("(libev) linuxaio io_setup");
456 600
457#if EPOLL_FALLBACK 601 /* forking epoll should also effectively unregister all fds from the backend */
458 while ((backend_fd = ev_epoll_create ()) < 0) 602 epoll_fork (EV_A);
459 ev_syserr ("(libev) linuxaio epoll_create"); 603 /* epoll_fork already did this. hopefully */
604 /*fd_rearm_all (EV_A);*/
460 605
461 ev_io_stop (EV_A_ &linuxaio_epoll_w); 606 ev_io_stop (EV_A_ &linuxaio_epoll_w);
462 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 607 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
463 ev_io_start (EV_A_ &linuxaio_epoll_w); 608 ev_io_start (EV_A_ &linuxaio_epoll_w);
464 ev_unref (EV_A); /* watcher should not keep loop alive */
465#endif
466
467 fd_rearm_all (EV_A);
468} 609}
469 610

Diff Legend

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