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 1.3 by root, Thu Jun 20 23:59:30 2019 UTC vs.
Revision 1.29 by root, Tue Jun 25 06:43:04 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/*
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 */
86
40#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 */
41#include <poll.h> 88#include <poll.h>
42#include <linux/aio_abi.h> 89#include <linux/aio_abi.h>
43 90
44/* we try to fill 4kn pages exactly. 91/*****************************************************************************/
45 * the ring buffer header is 32 bytes, every io event is 32 bytes. 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
46 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer 93
47 * so the calculation below will use "exactly" 8kB for the ring buffer 94#include <sys/syscall.h> /* no glibc wrappers */
95
96/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
97#define IOCB_CMD_POLL 5
98
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.
48 */ 103 */
49#define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */
50
51/*****************************************************************************/
52/* syscall wrapdadoop */
53
54#include <sys/syscall.h> /* no glibc wrappers */
55
56/* aio_abi.h is not verioned in any way, so we cannot test for its existance */
57#define IOCB_CMD_POLL 5
58
59/* taken from linux/fs/aio.c */
60#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
61#define AIO_RING_INCOMPAT_FEATURES 0 105#define AIO_RING_INCOMPAT_FEATURES 0
62struct aio_ring 106struct aio_ring
63{ 107{
64 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
… …
72 unsigned header_length; /* size of aio_ring */ 116 unsigned header_length; /* size of aio_ring */
73 117
74 struct io_event io_events[0]; 118 struct io_event io_events[0];
75}; 119};
76 120
77static int 121inline_size
122int
78ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
79{ 124{
80 return syscall (SYS_io_setup, nr_events, ctx_idp); 125 return syscall (SYS_io_setup, nr_events, ctx_idp);
81} 126}
82 127
83static int 128inline_size
129int
84ev_io_destroy (aio_context_t ctx_id) 130evsys_io_destroy (aio_context_t ctx_id)
85{ 131{
86 return syscall (SYS_io_destroy, ctx_id); 132 return syscall (SYS_io_destroy, ctx_id);
87} 133}
88 134
89static int 135inline_size
136int
90ev_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[])
91{ 138{
92 return syscall (SYS_io_submit, ctx_id, nr, cbp); 139 return syscall (SYS_io_submit, ctx_id, nr, cbp);
93} 140}
94 141
95static int 142inline_size
143int
96ev_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)
97{ 145{
98 return syscall (SYS_io_cancel, ctx_id, cbp, result); 146 return syscall (SYS_io_cancel, ctx_id, cbp, result);
99} 147}
100 148
101static int 149inline_size
150int
102ev_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)
103{ 152{
104 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 153 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
105} 154}
106
107typedef void (*ev_io_cb) (long nr, struct io_event *events);
108 155
109/*****************************************************************************/ 156/*****************************************************************************/
110/* actual backed implementation */ 157/* actual backed implementation */
111 158
112/* two iocbs for every fd, one for read, one for write */ 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
195/* we use out own wrapper structure in case we ever want to do something "clever" */
113typedef struct aniocb 196typedef struct aniocb
114{ 197{
115 struct iocb io; 198 struct iocb io;
116 /*int inuse;*/ 199 /*int inuse;*/
117} *ANIOCBP; 200} *ANIOCBP;
118 201
119inline_size 202inline_size
120void 203void
121linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 204linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
122{ 205{
123 /* TODO: quite the overhead to allocate every iocb separately */
124 while (count--) 206 while (count--)
125 { 207 {
208 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
126 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 209 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
127 /* 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 */
128 memset (*base, 0, sizeof (**base)); 214 memset (iocb, 0, sizeof (*iocb));
129 /* would be nice to initialize fd/data as well */ 215
130 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 216 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
131 ++base; 217 iocb->io.aio_data = offset;
132 } 218 iocb->io.aio_fildes = offset;
133}
134 219
220 base [offset++] = iocb;
221 }
222}
223
224ecb_cold
135static void 225static void
136linuxaio_free_iocbp (EV_P) 226linuxaio_free_iocbp (EV_P)
137{ 227{
138 while (linuxaio_iocbpmax--) 228 while (linuxaio_iocbpmax--)
139 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); 229 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
140 230
141 /* next resize will completely reallocate the array */ 231 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
142 linuxaio_iocbpmax = 0;
143 linuxaio_submitcnt = 0; /* all pointers invalidated */
144} 232}
145 233
146static void 234static void
147linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
148{ 236{
149 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);
150 struct aniocb *iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
239
240 if (iocb->io.aio_reqprio < 0)
241 {
242 /* we handed this fd over to epoll, so undo this first */
243 /* we do it manually because the optimisations on epoll_modfy won't do us any good */
244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
245 iocb->io.aio_reqprio = 0;
246 }
151 247
152 if (iocb->io.aio_buf) 248 if (iocb->io.aio_buf)
249 /* io_cancel always returns some error on relevant kernels, but works */
153 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always fails in relevant kernels */ 250 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
154 251
155 if (nev) 252 if (nev)
156 { 253 {
157 iocb->io.aio_data = fd;
158 iocb->io.aio_fildes = fd;
159 iocb->io.aio_buf = 254 iocb->io.aio_buf =
160 (nev & EV_READ ? POLLIN : 0) 255 (nev & EV_READ ? POLLIN : 0)
161 | (nev & EV_WRITE ? POLLOUT : 0); 256 | (nev & EV_WRITE ? POLLOUT : 0);
162 257
163 /* queue iocb up for io_submit */ 258 /* queue iocb up for io_submit */
164 /* this assumes we only ever get one call per fd per loop iteration */ 259 /* this assumes we only ever get one call per fd per loop iteration */
… …
167 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 262 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
168 } 263 }
169} 264}
170 265
171static void 266static void
267linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
268{
269 epoll_poll (EV_A_ 0);
270}
271
272static void
273linuxaio_fd_rearm (EV_P_ int fd)
274{
275 anfds [fd].events = 0;
276 linuxaio_iocbps [fd]->io.aio_buf = 0;
277 fd_change (EV_A_ fd, EV_ANFD_REIFY);
278}
279
280static void
172linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 281linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
173{ 282{
174 while (nr) 283 while (nr)
175 { 284 {
176 int fd = ev->data; 285 int fd = ev->data;
177 int res = ev->res; 286 int res = ev->res;
178 287
179 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 288 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
180 289
181 /* linux aio is oneshot: rearm fd */
182 linuxaio_iocbps [fd]->io.aio_buf = 0;
183 anfds [fd].events = 0;
184 fd_change (EV_A_ fd, 0);
185
186 /* feed events, we do not expect or handle POLLNVAL */ 290 /* feed events, we do not expect or handle POLLNVAL */
187 if (ecb_expect_false (res & POLLNVAL))
188 fd_kill (EV_A_ fd);
189 else
190 fd_event ( 291 fd_event (
191 EV_A_ 292 EV_A_
192 fd, 293 fd,
193 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 294 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
194 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 295 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
195 ); 296 );
297
298 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
299 linuxaio_fd_rearm (EV_A_ fd);
196 300
197 --nr; 301 --nr;
198 ++ev; 302 ++ev;
199 } 303 }
200} 304}
201 305
202/* get any events from ringbuffer, return true if any were handled */ 306/* get any events from ring buffer, return true if any were handled */
203static int 307static int
204linuxaio_get_events_from_ring (EV_P) 308linuxaio_get_events_from_ring (EV_P)
205{ 309{
206 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 310 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
207 311
312 /* the kernel reads and writes both of these variables, */
313 /* as a C extension, we assume that volatile use here */
314 /* both makes reads atomic and once-only */
315 unsigned head = *(volatile unsigned *)&ring->head;
316 unsigned tail = *(volatile unsigned *)&ring->tail;
317
318 if (head == tail)
319 return 0;
320
321 /* bail out if the ring buffer doesn't match the expected layout */
322 if (expect_false (ring->magic != AIO_RING_MAGIC)
323 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
324 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
325 return 0;
326
327 /* make sure the events up to tail are visible */
208 ECB_MEMORY_FENCE_ACQUIRE; 328 ECB_MEMORY_FENCE_ACQUIRE;
209
210 unsigned head = ring->head;
211 unsigned tail = *(volatile unsigned *)&ring->tail;
212
213 if (ring->magic != AIO_RING_MAGIC
214 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
215 || ring->header_length != sizeof (struct aio_ring) /* TODO: or use it to find io_event[0]? */
216 || head == tail)
217 return 0;
218 329
219 /* parse all available events, but only once, to avoid starvation */ 330 /* parse all available events, but only once, to avoid starvation */
220 if (tail > head) /* normal case around */ 331 if (tail > head) /* normal case around */
221 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 332 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
222 else
223 {
224 /* wrapped around */ 333 else /* wrapped around */
334 {
225 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 335 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
226 linuxaio_parse_events (EV_A_ ring->io_events, tail); 336 linuxaio_parse_events (EV_A_ ring->io_events, tail);
227 } 337 }
228 338
229 ring->head = tail; 339 ECB_MEMORY_FENCE_RELEASE;
340 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
341 *(volatile unsigned *)&ring->head = tail;
230 342
231 return 1; 343 return 1;
232} 344}
233 345
234/* read at least one event from kernel, or timeout */ 346/* read at least one event from kernel, or timeout */
235inline_size 347inline_size
236void 348void
237linuxaio_get_events (EV_P_ ev_tstamp timeout) 349linuxaio_get_events (EV_P_ ev_tstamp timeout)
238{ 350{
239 struct timespec ts; 351 struct timespec ts;
240 struct io_event ioev; 352 struct io_event ioev[1];
241 int res; 353 int res;
242 354
243 if (linuxaio_get_events_from_ring (EV_A)) 355 if (linuxaio_get_events_from_ring (EV_A))
244 return; 356 return;
245 357
246 /* no events, so wait for at least one, then poll ring buffer again */ 358 /* no events, so wait for at least one, then poll ring buffer again */
247 /* this degraded to one event per loop iteration */ 359 /* this degrades to one event per loop iteration */
248 /* if the ring buffer changes layout, but so be it */ 360 /* if the ring buffer changes layout, but so be it */
361
362 EV_RELEASE_CB;
249 363
250 ts.tv_sec = (long)timeout; 364 ts.tv_sec = (long)timeout;
251 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 365 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
252 366
253 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 367 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
368
369 EV_ACQUIRE_CB;
254 370
255 if (res < 0) 371 if (res < 0)
372 if (errno == EINTR)
373 /* ignored */;
374 else
256 ev_syserr ("(libev) io_getevents"); 375 ev_syserr ("(libev) linuxaio io_getevents");
257 else if (res) 376 else if (res)
258 { 377 {
259 /* at least one event received, handle it and any remaining ones in the ring buffer */ 378 /* at least one event received, handle it and any remaining ones in the ring buffer */
260 linuxaio_parse_events (EV_A_ &ioev, 1); 379 linuxaio_parse_events (EV_A_ ioev, res);
261 linuxaio_get_events_from_ring (EV_A); 380 linuxaio_get_events_from_ring (EV_A);
262 } 381 }
382}
383
384static int
385linuxaio_io_setup (EV_P)
386{
387 linuxaio_ctx = 0;
388 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
263} 389}
264 390
265static void 391static void
266linuxaio_poll (EV_P_ ev_tstamp timeout) 392linuxaio_poll (EV_P_ ev_tstamp timeout)
267{ 393{
… …
269 395
270 /* first phase: submit new iocbs */ 396 /* first phase: submit new iocbs */
271 397
272 /* io_submit might return less than the requested number of iocbs */ 398 /* io_submit might return less than the requested number of iocbs */
273 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 399 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
274 /* which allows us to pinpoint the errornous iocb */ 400 /* which allows us to pinpoint the erroneous iocb */
275 for (submitted = 0; submitted < linuxaio_submitcnt; ) 401 for (submitted = 0; submitted < linuxaio_submitcnt; )
276 { 402 {
277 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 403 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
278 404
279 if (res < 0) 405 if (expect_false (res < 0))
280 if (errno == EAGAIN) 406 if (errno == EINVAL)
281 { 407 {
282 /* This happens when the ring buffer is full, at least. I assume this means 408 /* This happens for unsupported fds, officially, but in my testing,
283 * that the event was queued synchronously during io_submit, and thus 409 * also randomly happens for supported fds. We fall back to good old
284 * the buffer overflowd. 410 * poll() here, under the assumption that this is a very rare case.
285 * In this case, we just try next loop iteration. 411 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
412 * discussion about such a case (ttys) where polling for POLLIN
413 * fails but POLLIN|POLLOUT works.
286 */ 414 */
287 memcpy (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 415 struct iocb *iocb = linuxaio_submits [submitted];
416 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
417 iocb->aio_reqprio = -1; /* mark iocb as epoll */
418
419 res = 1; /* skip this iocb - another iocb, another chance */
420 }
421 else if (errno == EAGAIN)
422 {
423 /* This happens when the ring buffer is full, or some other shit we
424 * don't know and isn't documented. Most likely because we have too
425 * many requests and linux aio can't be assed to handle them.
426 * In this case, we try to allocate a larger ring buffer, freeing
427 * ours first. This might fail, in which case we have to fall back to 100%
428 * epoll.
429 * God, how I hate linux not getting its act together. Ever.
430 */
431 evsys_io_destroy (linuxaio_ctx);
288 linuxaio_submitcnt -= submitted; 432 linuxaio_submitcnt = 0;
433
434 /* rearm all fds with active iocbs */
435 {
436 int fd;
437 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
438 if (linuxaio_iocbps [fd]->io.aio_buf)
439 linuxaio_fd_rearm (EV_A_ fd);
440 }
441
442 ++linuxaio_iteration;
443 if (linuxaio_io_setup (EV_A) < 0)
444 {
445 /* to bad, we can't get a new aio context, go 100% epoll */
446 linuxaio_free_iocbp (EV_A);
447 ev_io_stop (EV_A_ &linuxaio_epoll_w);
448 ev_ref (EV_A);
449 linuxaio_ctx = 0;
450 backend_modify = epoll_modify;
451 backend_poll = epoll_poll;
452 }
453
289 timeout = 0; 454 timeout = 0;
455 /* it's easiest to handle this mess in another iteration */
290 break; 456 return;
457 }
458 else if (errno == EBADF)
459 {
460 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
461
462 res = 1; /* skip this iocb */
291 } 463 }
292 else 464 else
293 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
294 /* TODO: should we always just try next time? */
295 ev_syserr ("(libev) io_submit"); 465 ev_syserr ("(libev) linuxaio io_submit");
296 466
297 submitted += res; 467 submitted += res;
298 } 468 }
299 469
300 linuxaio_submitcnt = 0; 470 linuxaio_submitcnt = 0;
… …
308int 478int
309linuxaio_init (EV_P_ int flags) 479linuxaio_init (EV_P_ int flags)
310{ 480{
311 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 481 /* would be great to have a nice test for IOCB_CMD_POLL instead */
312 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 482 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
313 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 483 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
484 if (ev_linux_version () < 0x041300)
314 return 0; 485 return 0;
315 486
316 linuxaio_ctx = 0; 487 if (!epoll_init (EV_A_ 0))
317 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
318 return 0; 488 return 0;
489
490 linuxaio_iteration = 0;
491
492 if (linuxaio_io_setup (EV_A) < 0)
493 {
494 epoll_destroy (EV_A);
495 return 0;
496 }
497
498 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
499 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
500 ev_io_start (EV_A_ &linuxaio_epoll_w);
501 ev_unref (EV_A); /* watcher should not keep loop alive */
319 502
320 backend_modify = linuxaio_modify; 503 backend_modify = linuxaio_modify;
321 backend_poll = linuxaio_poll; 504 backend_poll = linuxaio_poll;
322 505
323 linuxaio_iocbpmax = 0; 506 linuxaio_iocbpmax = 0;
… …
332 515
333inline_size 516inline_size
334void 517void
335linuxaio_destroy (EV_P) 518linuxaio_destroy (EV_P)
336{ 519{
520 epoll_destroy (EV_A);
337 linuxaio_free_iocbp (EV_A); 521 linuxaio_free_iocbp (EV_A);
338 ev_io_destroy (linuxaio_ctx); 522 evsys_io_destroy (linuxaio_ctx);
339} 523}
340 524
341inline_size 525inline_size
342void 526void
343linuxaio_fork (EV_P) 527linuxaio_fork (EV_P)
344{ 528{
345 /* TODO: verify and test */ 529 /* this frees all iocbs, which is very heavy-handed */
346 linuxaio_destroy (EV_A); 530 linuxaio_destroy (EV_A);
531 linuxaio_submitcnt = 0; /* all pointers were invalidated */
347 532
348 linuxaio_ctx = 0; 533 linuxaio_iteration = 0; /* we start over in the child */
349 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 534
535 while (linuxaio_io_setup (EV_A) < 0)
350 ev_syserr ("(libev) io_setup"); 536 ev_syserr ("(libev) linuxaio io_setup");
351 537
538 epoll_fork (EV_A);
539
540 ev_io_stop (EV_A_ &linuxaio_epoll_w);
541 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
542 ev_io_start (EV_A_ &linuxaio_epoll_w);
543
544 /* epoll_fork already did this. hopefully */
352 fd_rearm_all (EV_A); 545 /*fd_rearm_all (EV_A);*/
353} 546}
354 547

Diff Legend

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