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 1.11 by root, Sun Jun 23 22:57:25 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#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 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 */
… …
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 syscall (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 syscall (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 syscall (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 syscall (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 syscall (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 (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_modfy 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);
164 iocb->io.aio_reqprio = 0; 245 iocb->io.aio_reqprio = 0;
165 } 246 }
166#endif
167 247
168 if (iocb->io.aio_buf) 248 if (iocb->io.aio_buf)
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 249 /* io_cancel always returns some error on relevant kernels, but works */
250 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
170 251
171 if (nev) 252 if (nev)
172 { 253 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf = 254 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0) 255 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0); 256 | (nev & EV_WRITE ? POLLOUT : 0);
178 257
179 /* queue iocb up for io_submit */ 258 /* queue iocb up for io_submit */
180 /* 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 */
… …
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 262 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 263 }
185} 264}
186 265
187static 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
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 281linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 282{
190 while (nr) 283 while (nr)
191 { 284 {
192 int fd = ev->data; 285 int fd = ev->data;
193 int res = ev->res; 286 int res = ev->res;
194 287
195 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));
196 289
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 */ 290 /* 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 ( 291 fd_event (
207 EV_A_ 292 EV_A_
208 fd, 293 fd,
209 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 294 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
210 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 295 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
211 ); 296 );
297
298 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
299 linuxaio_fd_rearm (EV_A_ fd);
212 300
213 --nr; 301 --nr;
214 ++ev; 302 ++ev;
215 } 303 }
216} 304}
217 305
218/* get any events from ringbuffer, return true if any were handled */ 306/* get any events from ring buffer, return true if any were handled */
219static int 307static int
220linuxaio_get_events_from_ring (EV_P) 308linuxaio_get_events_from_ring (EV_P)
221{ 309{
222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 310 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
223 311
224 unsigned head = ring->head; 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;
225 unsigned tail = *(volatile unsigned *)&ring->tail; 316 unsigned tail = *(volatile unsigned *)&ring->tail;
226 317
227 if (head == tail) 318 if (head == tail)
228 return 0; 319 return 0;
229 320
230 /* bail out if the ring buffer doesn't match the expected layout */ 321 /* bail out if the ring buffer doesn't match the expected layout */
231 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 322 if (expect_false (ring->magic != AIO_RING_MAGIC)
232 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 323 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
233 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ 324 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
234 return 0; 325 return 0;
235 326
327 /* make sure the events up to tail are visible */
236 ECB_MEMORY_FENCE_ACQUIRE; 328 ECB_MEMORY_FENCE_ACQUIRE;
237 329
238 /* parse all available events, but only once, to avoid starvation */ 330 /* parse all available events, but only once, to avoid starvation */
239 if (tail > head) /* normal case around */ 331 if (tail > head) /* normal case around */
240 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 332 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
… …
242 { 334 {
243 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);
244 linuxaio_parse_events (EV_A_ ring->io_events, tail); 336 linuxaio_parse_events (EV_A_ ring->io_events, tail);
245 } 337 }
246 338
339 ECB_MEMORY_FENCE_RELEASE;
340 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
247 *(volatile unsigned *)&ring->head = tail; 341 *(volatile unsigned *)&ring->head = tail;
248
249 /* again, other implementations don't do this, but I think it's required for the kernel to see free slots */
250 ECB_MEMORY_FENCE_RELEASE;
251 342
252 return 1; 343 return 1;
253} 344}
254 345
255/* read at least one event from kernel, or timeout */ 346/* read at least one event from kernel, or timeout */
256inline_size 347inline_size
257void 348void
258linuxaio_get_events (EV_P_ ev_tstamp timeout) 349linuxaio_get_events (EV_P_ ev_tstamp timeout)
259{ 350{
260 struct timespec ts; 351 struct timespec ts;
261 struct io_event ioev; 352 struct io_event ioev[1];
262 int res; 353 int res;
263 354
264 if (linuxaio_get_events_from_ring (EV_A)) 355 if (linuxaio_get_events_from_ring (EV_A))
265 return; 356 return;
266 357
267 /* 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 */
268 /* this degrades to one event per loop iteration */ 359 /* this degrades to one event per loop iteration */
269 /* if the ring buffer changes layout, but so be it */ 360 /* if the ring buffer changes layout, but so be it */
270 361
362 EV_RELEASE_CB;
363
271 ts.tv_sec = (long)timeout; 364 ts.tv_sec = (long)timeout;
272 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 365 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
273 366
274 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;
275 370
276 if (res < 0) 371 if (res < 0)
277 if (errno == EINTR) 372 if (errno == EINTR)
278 /* ignored */; 373 /* ignored */;
279 else 374 else
280 ev_syserr ("(libev) linuxaio io_getevents"); 375 ev_syserr ("(libev) linuxaio io_getevents");
281 else if (res) 376 else if (res)
282 { 377 {
283 /* 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 */
284 linuxaio_parse_events (EV_A_ &ioev, 1); 379 linuxaio_parse_events (EV_A_ ioev, res);
285 linuxaio_get_events_from_ring (EV_A); 380 linuxaio_get_events_from_ring (EV_A);
286 } 381 }
287} 382}
288 383
289#if EPOLL_FALLBACK
290static void 384static int
291linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) 385linuxaio_io_setup (EV_P)
292{ 386{
293 struct epoll_event eev; 387 linuxaio_ctx = 0;
294 388 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
295 eev.events = EPOLLONESHOT;
296 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
297 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
298 eev.data.fd = iocb->aio_fildes;
299
300 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
301 ev_syserr ("(libeio) linuxaio epoll_ctl");
302} 389}
303#endif
304 390
305static void 391static void
306linuxaio_poll (EV_P_ ev_tstamp timeout) 392linuxaio_poll (EV_P_ ev_tstamp timeout)
307{ 393{
308 int submitted; 394 int submitted;
309 395
310 /* first phase: submit new iocbs */ 396 /* first phase: submit new iocbs */
311 397
312 /* io_submit might return less than the requested number of iocbs */ 398 /* io_submit might return less than the requested number of iocbs */
313 /* 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, */
314 /* which allows us to pinpoint the errornous iocb */ 400 /* which allows us to pinpoint the erroneous iocb */
315 for (submitted = 0; submitted < linuxaio_submitcnt; ) 401 for (submitted = 0; submitted < linuxaio_submitcnt; )
316 { 402 {
317 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);
318 404
319 if (ecb_expect_false (res < 0)) 405 if (expect_false (res < 0))
320 if (errno == EAGAIN) 406 if (errno == EINVAL)
321 { 407 {
322 /* This happens when the ring buffer is full, at least. I assume this means
323 * that the event was queued synchronously during io_submit, and thus
324 * the buffer overflowed.
325 * In this case, we just try in next loop iteration.
326 * This should not result in a few fds taking priority, as the interface
327 * is one-shot, and we submit iocb's in a round-robin fashion.
328 */
329 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
330 linuxaio_submitcnt -= submitted;
331 timeout = 0;
332 break;
333 }
334#if EPOLL_FALLBACK
335 else if (errno == EINVAL)
336 {
337 /* This hapΓΌpens for unsupported fds, officially, but in my testing, 408 /* This happens for unsupported fds, officially, but in my testing,
338 * also randomly happens for supported fds. We fall back to good old 409 * also randomly happens for supported fds. We fall back to good old
339 * poll() here, under the assumption that this is a very rare case. 410 * poll() here, under the assumption that this is a very rare case.
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.
340 */ 414 */
341 struct iocb *iocb = linuxaio_submits [submitted]; 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);
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
454 timeout = 0;
455 /* it's easiest to handle this mess in another iteration */
456 return;
457 }
458 else if (errno == EBADF)
459 {
460 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
461
342 res = 1; /* skip this iocb */ 462 res = 1; /* skip this iocb */
343
344 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
345 iocb->aio_reqprio = -1; /* mark iocb as epoll */
346 } 463 }
347#endif
348 else 464 else
349 ev_syserr ("(libev) linuxaio io_submit"); 465 ev_syserr ("(libev) linuxaio io_submit");
350 466
351 submitted += res; 467 submitted += res;
352 } 468 }
… …
356 /* second phase: fetch and parse events */ 472 /* second phase: fetch and parse events */
357 473
358 linuxaio_get_events (EV_A_ timeout); 474 linuxaio_get_events (EV_A_ timeout);
359} 475}
360 476
361#if EPOLL_FALLBACK
362
363static void
364linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
365{
366 struct epoll_event events[16];
367
368 for (;;)
369 {
370 int idx;
371 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
372
373 if (ecb_expect_false (res < 0))
374 ev_syserr ("(libev) linuxaio epoll_wait");
375 else if (!res)
376 break;
377
378 for (idx = res; idx--; )
379 {
380 int fd = events [idx].data.fd;
381 uint32_t ev = events [idx].events;
382
383 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
384
385 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
386
387 fd_event (EV_A_ fd,
388 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
389 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
390 }
391
392 if (res < sizeof (events) / sizeof (events [0]))
393 break;
394 }
395}
396
397#endif
398
399inline_size 477inline_size
400int 478int
401linuxaio_init (EV_P_ int flags) 479linuxaio_init (EV_P_ int flags)
402{ 480{
403 /* 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 */
404 /* 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 */
405 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)
406 return 0; 485 return 0;
407 486
408 linuxaio_ctx = 0; 487 if (!epoll_init (EV_A_ 0))
409 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
410 return 0; 488 return 0;
411 489
412#if EPOLL_FALLBACK 490 linuxaio_iteration = 0;
413 backend_fd = ev_epoll_create (); 491
414 if (backend_fd < 0) 492 if (linuxaio_io_setup (EV_A) < 0)
415 { 493 {
416 ev_io_destroy (linuxaio_ctx); 494 epoll_destroy (EV_A);
417 return 0; 495 return 0;
418 } 496 }
419 497
420 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 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);
421 ev_io_start (EV_A_ &linuxaio_epoll_w); 500 ev_io_start (EV_A_ &linuxaio_epoll_w);
422#endif 501 ev_unref (EV_A); /* watcher should not keep loop alive */
423 502
424 backend_modify = linuxaio_modify; 503 backend_modify = linuxaio_modify;
425 backend_poll = linuxaio_poll; 504 backend_poll = linuxaio_poll;
426 505
427 linuxaio_iocbpmax = 0; 506 linuxaio_iocbpmax = 0;
… …
436 515
437inline_size 516inline_size
438void 517void
439linuxaio_destroy (EV_P) 518linuxaio_destroy (EV_P)
440{ 519{
441#if EPOLL_FALLBACK 520 epoll_destroy (EV_A);
442 close (backend_fd);
443#endif
444 linuxaio_free_iocbp (EV_A); 521 linuxaio_free_iocbp (EV_A);
445 ev_io_destroy (linuxaio_ctx); 522 evsys_io_destroy (linuxaio_ctx);
446} 523}
447 524
448inline_size 525inline_size
449void 526void
450linuxaio_fork (EV_P) 527linuxaio_fork (EV_P)
451{ 528{
452 /* this frees all iocbs, which is very heavy-handed */ 529 /* this frees all iocbs, which is very heavy-handed */
453 linuxaio_destroy (EV_A); 530 linuxaio_destroy (EV_A);
454 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 531 linuxaio_submitcnt = 0; /* all pointers were invalidated */
455 532
456 linuxaio_ctx = 0; 533 linuxaio_iteration = 0; /* we start over in the child */
457 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 534
535 while (linuxaio_io_setup (EV_A) < 0)
458 ev_syserr ("(libev) linuxaio io_setup"); 536 ev_syserr ("(libev) linuxaio io_setup");
459 537
460#if EPOLL_FALLBACK 538 epoll_fork (EV_A);
461 while ((backend_fd = ev_epoll_create ()) < 0)
462 ev_syserr ("(libev) linuxaio epoll_create");
463 539
464 ev_io_stop (EV_A_ &linuxaio_epoll_w); 540 ev_io_stop (EV_A_ &linuxaio_epoll_w);
465 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 541 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
466 ev_io_start (EV_A_ &linuxaio_epoll_w); 542 ev_io_start (EV_A_ &linuxaio_epoll_w);
467 ev_unref (EV_A); /* watcher should not keep loop alive */
468#endif
469 543
544 /* epoll_fork already did this. hopefully */
470 fd_rearm_all (EV_A); 545 /*fd_rearm_all (EV_A);*/
471} 546}
472 547

Diff Legend

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