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 1.17 by root, Mon Jun 24 02:02:35 2019 UTC vs.
Revision 1.43 by root, Mon Jul 1 21:47:42 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
167 248
249 if (ecb_expect_false (iocb->io.aio_buf))
250 {
251 /* iocb active, so cancel it first before resubmit */
252 for (;;)
253 {
254 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
255 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
256 break;
257
258 if (ecb_expect_true (errno == EINPROGRESS))
259 break;
260
261 /* the EINPROGRESS test is for nicer error message. clumsy. */
262 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR));
263 }
264 }
265
168 if (iocb->io.aio_buf) 266 iocb->io.aio_buf =
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 267 (nev & EV_READ ? POLLIN : 0)
268 | (nev & EV_WRITE ? POLLOUT : 0);
170 269
171 if (nev) 270 if (nev)
172 { 271 {
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 */ 272 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */ 273 /* this assumes we only ever get one call per fd per loop iteration */
181 ++linuxaio_submitcnt; 274 ++linuxaio_submitcnt;
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 275 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 276 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 277 }
185} 278}
186 279
187static void 280static void
281linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
282{
283 epoll_poll (EV_A_ 0);
284}
285
286inline_speed
287void
288linuxaio_fd_rearm (EV_P_ int fd)
289{
290 anfds [fd].events = 0;
291 linuxaio_iocbps [fd]->io.aio_buf = 0;
292 fd_change (EV_A_ fd, EV_ANFD_REIFY);
293}
294
295static void
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 296linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 297{
190 while (nr) 298 while (nr)
191 { 299 {
192 int fd = ev->data; 300 int fd = ev->data;
193 int res = ev->res; 301 int res = ev->res;
194 302
195 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 303 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
196 304
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 */ 305 /* feed events, we do not expect or handle POLLNVAL */
203 if (expect_false (res & POLLNVAL))
204 fd_kill (EV_A_ fd);
205 else
206 fd_event ( 306 fd_event (
207 EV_A_ 307 EV_A_
208 fd, 308 fd,
209 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 309 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
210 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 310 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
211 ); 311 );
312
313 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
314 linuxaio_fd_rearm (EV_A_ fd);
212 315
213 --nr; 316 --nr;
214 ++ev; 317 ++ev;
215 } 318 }
216} 319}
217 320
218/* get any events from ringbuffer, return true if any were handled */ 321/* get any events from ring buffer, return true if any were handled */
219static int 322static int
220linuxaio_get_events_from_ring (EV_P) 323linuxaio_get_events_from_ring (EV_P)
221{ 324{
222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 325 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
223 326
… …
228 unsigned tail = *(volatile unsigned *)&ring->tail; 331 unsigned tail = *(volatile unsigned *)&ring->tail;
229 332
230 if (head == tail) 333 if (head == tail)
231 return 0; 334 return 0;
232 335
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 */ 336 /* make sure the events up to tail are visible */
240 ECB_MEMORY_FENCE_ACQUIRE; 337 ECB_MEMORY_FENCE_ACQUIRE;
241 338
242 /* parse all available events, but only once, to avoid starvation */ 339 /* parse all available events, but only once, to avoid starvation */
243 if (tail > head) /* normal case around */ 340 if (tail > head) /* normal case around */
… …
246 { 343 {
247 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 344 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
248 linuxaio_parse_events (EV_A_ ring->io_events, tail); 345 linuxaio_parse_events (EV_A_ ring->io_events, tail);
249 } 346 }
250 347
251 /* TODO: we only need a compiler barrier here, not a read fence */
252 ECB_MEMORY_FENCE_RELEASE; 348 ECB_MEMORY_FENCE_RELEASE;
253 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 349 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
254 *(volatile unsigned *)&ring->head = tail; 350 *(volatile unsigned *)&ring->head = tail;
255 /* make sure kernel can see our new head value - probably not required */
256 ECB_MEMORY_FENCE_RELEASE;
257 351
258 return 1; 352 return 1;
353}
354
355inline_size
356int
357linuxaio_ringbuf_valid (EV_P)
358{
359 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
360
361 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
362 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
363 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
259} 364}
260 365
261/* read at least one event from kernel, or timeout */ 366/* read at least one event from kernel, or timeout */
262inline_size 367inline_size
263void 368void
264linuxaio_get_events (EV_P_ ev_tstamp timeout) 369linuxaio_get_events (EV_P_ ev_tstamp timeout)
265{ 370{
266 struct timespec ts; 371 struct timespec ts;
267 struct io_event ioev; 372 struct io_event ioev[8]; /* 256 octet stack space */
268 int res; 373 int want = 1; /* how many events to request */
374 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
269 375
376 if (ecb_expect_true (ringbuf_valid))
377 {
378 /* if the ring buffer has any events, we don't wait or call the kernel at all */
270 if (linuxaio_get_events_from_ring (EV_A)) 379 if (linuxaio_get_events_from_ring (EV_A))
271 return; 380 return;
272 381
273 /* no events, so wait for at least one, then poll ring buffer again */ 382 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
274 /* this degrades to one event per loop iteration */ 383 if (!timeout)
275 /* if the ring buffer changes layout, but so be it */ 384 return;
385 }
386 else
387 /* no ringbuffer, request slightly larger batch */
388 want = sizeof (ioev) / sizeof (ioev [0]);
276 389
390 /* no events, so wait for some
391 * for fairness reasons, we do this in a loop, to fetch all events
392 */
393 for (;;)
394 {
395 int res;
396
397 EV_RELEASE_CB;
398
277 ts.tv_sec = (long)timeout; 399 ts.tv_sec = (long)timeout;
278 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 400 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
279 401
280 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 402 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
281 403
404 EV_ACQUIRE_CB;
405
282 if (res < 0) 406 if (res < 0)
283 if (errno == EINTR) 407 if (errno == EINTR)
284 /* ignored */; 408 /* ignored, retry */;
285 else 409 else
286 ev_syserr ("(libev) linuxaio io_getevents"); 410 ev_syserr ("(libev) linuxaio io_getevents");
287 else if (res) 411 else if (res)
288 { 412 {
289 /* at least one event received, handle it and any remaining ones in the ring buffer */ 413 /* at least one event available, handle them */
290 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 */
291 linuxaio_get_events_from_ring (EV_A); 419 linuxaio_get_events_from_ring (EV_A);
292 }
293}
294 420
295#if EPOLL_FALLBACK 421 /* at this point, we should have handled all outstanding events */
296static void 422 break;
297linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) 423 }
298{ 424 else if (res < want)
299 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 */
300 430
301 eev.events = EPOLLONESHOT; 431 timeout = 0; /* only wait in the first iteration */
302 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; 432 }
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} 433}
309#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}
310 442
311static void 443static void
312linuxaio_poll (EV_P_ ev_tstamp timeout) 444linuxaio_poll (EV_P_ ev_tstamp timeout)
313{ 445{
314 int submitted; 446 int submitted;
315 447
316 /* first phase: submit new iocbs */ 448 /* first phase: submit new iocbs */
317 449
318 /* io_submit might return less than the requested number of iocbs */ 450 /* io_submit might return less than the requested number of iocbs */
319 /* 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, */
320 /* which allows us to pinpoint the errornous iocb */ 452 /* which allows us to pinpoint the erroneous iocb */
321 for (submitted = 0; submitted < linuxaio_submitcnt; ) 453 for (submitted = 0; submitted < linuxaio_submitcnt; )
322 { 454 {
323 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);
324 456
325 if (expect_false (res < 0)) 457 if (ecb_expect_false (res < 0))
326 if (errno == EAGAIN)
327 {
328 /* This happens when the ring buffer is full, at least. I assume this means
329 * that the event was queued synchronously during io_submit, and thus
330 * the buffer overflowed.
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.
334 */
335 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
336 linuxaio_submitcnt -= submitted;
337 timeout = 0;
338 break;
339 }
340#if EPOLL_FALLBACK
341 else if (errno == EINVAL) 458 if (errno == EINVAL)
342 { 459 {
343 /* This happens for unsupported fds, officially, but in my testing, 460 /* This happens for unsupported fds, officially, but in my testing,
344 * 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
345 * 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.
346 * See https://lore.kernel.org/patchwork/patch/1047453/ for evidence 463 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
347 * that the problem is known, but ignored. 464 * discussion about such a case (ttys) where polling for POLLIN
465 * fails but POLLIN|POLLOUT works.
348 */ 466 */
349 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 /* to bad, we can't get a new aio context, go 100% epoll */
498 linuxaio_free_iocbp (EV_A);
499 ev_io_stop (EV_A_ &linuxaio_epoll_w);
500 ev_ref (EV_A);
501 linuxaio_ctx = 0;
502 backend_modify = epoll_modify;
503 backend_poll = epoll_poll;
504 }
505
506 timeout = 0;
507 /* it's easiest to handle this mess in another iteration */
508 return;
509 }
510 else if (errno == EBADF)
511 {
512 assert (("libev: event loop rejected bad fd", errno != EBADF));
513 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
514
350 res = 1; /* skip this iocb */ 515 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 } 516 }
355#endif 517 else if (errno == EINTR) /* not seen in reality, not documented */
518 res = 0; /* silently ignore and retry */
356 else 519 else
357 ev_syserr ("(libev) linuxaio io_submit"); 520 ev_syserr ("(libev) linuxaio io_submit");
358 521
359 submitted += res; 522 submitted += res;
360 } 523 }
… …
364 /* second phase: fetch and parse events */ 527 /* second phase: fetch and parse events */
365 528
366 linuxaio_get_events (EV_A_ timeout); 529 linuxaio_get_events (EV_A_ timeout);
367} 530}
368 531
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
407inline_size 532inline_size
408int 533int
409linuxaio_init (EV_P_ int flags) 534linuxaio_init (EV_P_ int flags)
410{ 535{
411 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 536 /* would be great to have a nice test for IOCB_CMD_POLL instead */
412 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 537 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
413#if EPOLL_FALLBACK 538 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
414 /* 4.19 made epoll work */
415 if (ev_linux_version () < 0x041300) 539 if (ev_linux_version () < 0x041300)
416 return 0; 540 return 0;
417#else 541
418 /* 4.18 introduced IOCB_CMD_POLL */ 542 if (!epoll_init (EV_A_ 0))
419 if (ev_linux_version () < 0x041200)
420 return 0; 543 return 0;
421#endif
422 544
423 linuxaio_ctx = 0; 545 linuxaio_iteration = 0;
424 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
425 return 0;
426 546
427#if EPOLL_FALLBACK 547 if (linuxaio_io_setup (EV_A) < 0)
428 backend_fd = ev_epoll_create ();
429 if (backend_fd < 0)
430 { 548 {
431 ev_io_destroy (linuxaio_ctx); 549 epoll_destroy (EV_A);
432 return 0; 550 return 0;
433 } 551 }
434 552
435 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 553 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
554 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
436 ev_io_start (EV_A_ &linuxaio_epoll_w); 555 ev_io_start (EV_A_ &linuxaio_epoll_w);
437 ev_unref (EV_A); /* watcher should not keep loop alive */ 556 ev_unref (EV_A); /* watcher should not keep loop alive */
438#endif
439 557
440 backend_modify = linuxaio_modify; 558 backend_modify = linuxaio_modify;
441 backend_poll = linuxaio_poll; 559 backend_poll = linuxaio_poll;
442 560
443 linuxaio_iocbpmax = 0; 561 linuxaio_iocbpmax = 0;
… …
452 570
453inline_size 571inline_size
454void 572void
455linuxaio_destroy (EV_P) 573linuxaio_destroy (EV_P)
456{ 574{
457#if EPOLL_FALLBACK 575 epoll_destroy (EV_A);
458 close (backend_fd);
459#endif
460 linuxaio_free_iocbp (EV_A); 576 linuxaio_free_iocbp (EV_A);
461 ev_io_destroy (linuxaio_ctx); 577 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
462} 578}
463 579
464inline_size 580inline_size
465void 581void
466linuxaio_fork (EV_P) 582linuxaio_fork (EV_P)
467{ 583{
468 /* this frees all iocbs, which is very heavy-handed */ 584 /* this frees all iocbs, which is very heavy-handed */
469 linuxaio_destroy (EV_A); 585 linuxaio_destroy (EV_A);
470 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 586 linuxaio_submitcnt = 0; /* all pointers were invalidated */
471 587
472 linuxaio_ctx = 0; 588 linuxaio_iteration = 0; /* we start over in the child */
473 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 589
590 while (linuxaio_io_setup (EV_A) < 0)
474 ev_syserr ("(libev) linuxaio io_setup"); 591 ev_syserr ("(libev) linuxaio io_setup");
475 592
476#if EPOLL_FALLBACK 593 /* forking epoll should also effectively unregister all fds from the backend */
477 while ((backend_fd = ev_epoll_create ()) < 0) 594 epoll_fork (EV_A);
478 ev_syserr ("(libev) linuxaio epoll_create");
479 595
480 ev_io_stop (EV_A_ &linuxaio_epoll_w); 596 ev_io_stop (EV_A_ &linuxaio_epoll_w);
481 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 597 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
482 ev_io_start (EV_A_ &linuxaio_epoll_w); 598 ev_io_start (EV_A_ &linuxaio_epoll_w);
483#endif
484 599
600 /* epoll_fork already did this. hopefully */
485 fd_rearm_all (EV_A); 601 /*fd_rearm_all (EV_A);*/
486} 602}
487 603

Diff Legend

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