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 1.23 by root, Mon Jun 24 21:34:47 2019 UTC vs.
Revision 1.54 by sf-exg, Sun May 14 19:02:31 2023 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 91/*****************************************************************************/
47# include <sys/epoll.h> 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
48#endif
49 93
50/* we try to fill 4kB pages exactly. 94#include <sys/syscall.h> /* no glibc wrappers */
51 * the ring buffer header is 32 bytes, every io event is 32 bytes. 95
52 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. 96/* aio_abi.h is not versioned in any way, so we cannot test for its existence */
53 * therefore the calculation below will use "exactly" 4kB for the ring buffer 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.
54 */ 103 */
55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56
57/*****************************************************************************/
58/* syscall wrapdadoop - this section has the raw syscall definitions */
59
60#include <sys/syscall.h> /* no glibc wrappers */
61
62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63#define IOCB_CMD_POLL 5
64
65/* taken from linux/fs/aio.c */
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. */
… …
82 120
83inline_size 121inline_size
84int 122int
85evsys_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
92evsys_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
99evsys_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
106evsys_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
113evsys_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;
… …
129void 203void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) 204linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 205{
132 while (count--) 206 while (count--)
133 { 207 {
134 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ 208 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
135 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); 209 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 210
137 /* full zero initialise is probably not required at the moment, but 211 /* full zero initialise is probably not required at the moment, but
138 * this is not well documented, so we better do it. 212 * this is not well documented, so we better do it.
139 */ 213 */
140 memset (iocb, 0, sizeof (*iocb)); 214 memset (iocb, 0, sizeof (*iocb));
141 215
142 iocb->io.aio_lio_opcode = IOCB_CMD_POLL; 216 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
143 iocb->io.aio_data = offset;
144 iocb->io.aio_fildes = offset; 217 iocb->io.aio_fildes = offset;
145 218
146 base [offset++] = iocb; 219 base [offset++] = iocb;
147 } 220 }
148} 221}
… …
160static void 233static void
161linuxaio_modify (EV_P_ int fd, int oev, int nev) 234linuxaio_modify (EV_P_ int fd, int oev, int nev)
162{ 235{
163 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 236 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
164 ANIOCBP iocb = linuxaio_iocbps [fd]; 237 ANIOCBP iocb = linuxaio_iocbps [fd];
238 ANFD *anfd = &anfds [fd];
165 239
166#if EPOLL_FALLBACK
167 if (iocb->io.aio_reqprio < 0) 240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
168 { 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 */
169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
245 anfd->emask = 0;
170 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
171 } 247 }
172#endif 248 else if (ecb_expect_false (iocb->io.aio_buf))
249 {
250 /* iocb active, so cancel it first before resubmit */
251 /* this assumes we only ever get one call per fd per loop iteration */
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;
173 257
174 if (iocb->io.aio_buf) 258 if (ecb_expect_true (errno == EINPROGRESS))
175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 259 break;
260
261 /* the EINPROGRESS test is for nicer error message. clumsy. */
262 if (errno != EINTR)
263 {
264 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
265 break;
266 }
267 }
268
269 /* increment generation counter to avoid handling old events */
270 ++anfd->egen;
271 }
272
273 iocb->io.aio_buf = (nev & EV_READ ? POLLIN : 0)
274 | (nev & EV_WRITE ? POLLOUT : 0);
176 275
177 if (nev) 276 if (nev)
178 { 277 {
179 iocb->io.aio_buf = 278 iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
180 (nev & EV_READ ? POLLIN : 0)
181 | (nev & EV_WRITE ? POLLOUT : 0);
182 279
183 /* queue iocb up for io_submit */ 280 /* queue iocb up for io_submit */
184 /* this assumes we only ever get one call per fd per loop iteration */ 281 /* this assumes we only ever get one call per fd per loop iteration */
185 ++linuxaio_submitcnt; 282 ++linuxaio_submitcnt;
186 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 283 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
187 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 284 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
188 } 285 }
189} 286}
190 287
191#if EPOLL_FALLBACK
192
193static void
194linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
195{
196 struct epoll_event eev;
197
198 eev.events = EPOLLONESHOT;
199 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
200 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
201 eev.data.fd = iocb->aio_fildes;
202
203 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
204 ev_syserr ("(libeio) linuxaio epoll_ctl");
205}
206
207static void 288static void
208linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 289linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
209{ 290{
210 struct epoll_event events[16]; 291 epoll_poll (EV_A_ 0);
211
212 for (;;)
213 {
214 int idx;
215 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
216
217 if (expect_false (res < 0))
218 ev_syserr ("(libev) linuxaio epoll_wait");
219 else if (!res)
220 break;
221
222 for (idx = res; idx--; )
223 {
224 int fd = events [idx].data.fd;
225 uint32_t ev = events [idx].events;
226
227 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
228
229 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
230
231 fd_event (EV_A_ fd,
232 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
233 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
234 }
235
236 if (res < sizeof (events) / sizeof (events [0]))
237 break;
238 }
239} 292}
240 293
241#endif 294inline_speed
295void
296linuxaio_fd_rearm (EV_P_ int fd)
297{
298 anfds [fd].events = 0;
299 linuxaio_iocbps [fd]->io.aio_buf = 0;
300 fd_change (EV_A_ fd, EV_ANFD_REIFY);
301}
242 302
243static void 303static void
244linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 304linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
245{ 305{
246 while (nr) 306 while (nr)
247 { 307 {
248 int fd = ev->data; 308 int fd = ev->data & 0xffffffff;
309 uint32_t gen = ev->data >> 32;
249 int res = ev->res; 310 int res = ev->res;
250 311
251 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 312 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
252 313
253 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ 314 /* only accept events if generation counter matches */
254 linuxaio_iocbps [fd]->io.aio_buf = 0; 315 if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
255 anfds [fd].events = 0; 316 {
256 fd_change (EV_A_ fd, 0);
257
258 /* feed events, we do not expect or handle POLLNVAL */ 317 /* feed events, we do not expect or handle POLLNVAL */
259 fd_event ( 318 fd_event (
260 EV_A_ 319 EV_A_
261 fd, 320 fd,
262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 321 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 322 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
264 ); 323 );
324
325 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
326 linuxaio_fd_rearm (EV_A_ fd);
327 }
265 328
266 --nr; 329 --nr;
267 ++ev; 330 ++ev;
268 } 331 }
269} 332}
270 333
271/* get any events from ringbuffer, return true if any were handled */ 334/* get any events from ring buffer, return true if any were handled */
272static int 335static int
273linuxaio_get_events_from_ring (EV_P) 336linuxaio_get_events_from_ring (EV_P)
274{ 337{
275 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 338 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
339 unsigned head, tail;
276 340
277 /* the kernel reads and writes both of these variables, */ 341 /* the kernel reads and writes both of these variables, */
278 /* as a C extension, we assume that volatile use here */ 342 /* as a C extension, we assume that volatile use here */
279 /* both makes reads atomic and once-only */ 343 /* both makes reads atomic and once-only */
280 unsigned head = *(volatile unsigned *)&ring->head; 344 head = *(volatile unsigned *)&ring->head;
345 ECB_MEMORY_FENCE_ACQUIRE;
281 unsigned tail = *(volatile unsigned *)&ring->tail; 346 tail = *(volatile unsigned *)&ring->tail;
282 347
283 if (head == tail) 348 if (head == tail)
284 return 0; 349 return 0;
285 350
286 /* bail out if the ring buffer doesn't match the expected layout */
287 if (expect_false (ring->magic != AIO_RING_MAGIC)
288 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
289 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
290 return 0;
291
292 /* make sure the events up to tail are visible */
293 ECB_MEMORY_FENCE_ACQUIRE;
294
295 /* parse all available events, but only once, to avoid starvation */ 351 /* parse all available events, but only once, to avoid starvation */
296 if (tail > head) /* normal case around */ 352 if (ecb_expect_true (tail > head)) /* normal case around */
297 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 353 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
298 else /* wrapped around */ 354 else /* wrapped around */
299 { 355 {
300 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 356 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
301 linuxaio_parse_events (EV_A_ ring->io_events, tail); 357 linuxaio_parse_events (EV_A_ ring->io_events, tail);
302 } 358 }
303 359
304 /* TODO: we only need a compiler barrier here, not a read fence */
305 ECB_MEMORY_FENCE_RELEASE; 360 ECB_MEMORY_FENCE_RELEASE;
306 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 361 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
307 *(volatile unsigned *)&ring->head = tail; 362 *(volatile unsigned *)&ring->head = tail;
308 /* make sure kernel can see our new head value - probably not required */
309 ECB_MEMORY_FENCE_RELEASE;
310 363
311 return 1; 364 return 1;
365}
366
367inline_size
368int
369linuxaio_ringbuf_valid (EV_P)
370{
371 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
372
373 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
374 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
375 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
312} 376}
313 377
314/* read at least one event from kernel, or timeout */ 378/* read at least one event from kernel, or timeout */
315inline_size 379inline_size
316void 380void
317linuxaio_get_events (EV_P_ ev_tstamp timeout) 381linuxaio_get_events (EV_P_ ev_tstamp timeout)
318{ 382{
319 struct timespec ts; 383 struct timespec ts;
320 struct io_event ioev[1]; 384 struct io_event ioev[8]; /* 256 octet stack space */
321 int res; 385 int want = 1; /* how many events to request */
386 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
322 387
388 if (ecb_expect_true (ringbuf_valid))
389 {
390 /* if the ring buffer has any events, we don't wait or call the kernel at all */
323 if (linuxaio_get_events_from_ring (EV_A)) 391 if (linuxaio_get_events_from_ring (EV_A))
324 return; 392 return;
325 393
326 /* no events, so wait for at least one, then poll ring buffer again */ 394 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
327 /* this degrades to one event per loop iteration */ 395 if (!timeout)
328 /* if the ring buffer changes layout, but so be it */ 396 return;
397 }
398 else
399 /* no ringbuffer, request slightly larger batch */
400 want = sizeof (ioev) / sizeof (ioev [0]);
329 401
402 /* no events, so wait for some
403 * for fairness reasons, we do this in a loop, to fetch all events
404 */
405 for (;;)
406 {
407 int res;
408
330 EV_RELEASE_CB; 409 EV_RELEASE_CB;
331 410
332 ts.tv_sec = (long)timeout; 411 EV_TS_SET (ts, timeout);
333 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 412 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
334 413
335 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
336
337 EV_ACQUIRE_CB; 414 EV_ACQUIRE_CB;
338 415
339 if (res < 0) 416 if (res < 0)
340 if (errno == EINTR) 417 if (errno == EINTR)
341 /* ignored */; 418 /* ignored, retry */;
342 else 419 else
343 ev_syserr ("(libev) linuxaio io_getevents"); 420 ev_syserr ("(libev) linuxaio io_getevents");
344 else if (res) 421 else if (res)
345 { 422 {
346 /* at least one event received, handle it and any remaining ones in the ring buffer */ 423 /* at least one event available, handle them */
347 linuxaio_parse_events (EV_A_ ioev, res); 424 linuxaio_parse_events (EV_A_ ioev, res);
425
426 if (ecb_expect_true (ringbuf_valid))
427 {
428 /* if we have a ring buffer, handle any remaining events in it */
348 linuxaio_get_events_from_ring (EV_A); 429 linuxaio_get_events_from_ring (EV_A);
430
431 /* at this point, we should have handled all outstanding events */
432 break;
433 }
434 else if (res < want)
435 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
436 break;
437 }
438 else
439 break; /* no events from the kernel, we are done */
440
441 timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
349 } 442 }
443}
444
445inline_size
446int
447linuxaio_io_setup (EV_P)
448{
449 linuxaio_ctx = 0;
450 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
350} 451}
351 452
352static void 453static void
353linuxaio_poll (EV_P_ ev_tstamp timeout) 454linuxaio_poll (EV_P_ ev_tstamp timeout)
354{ 455{
… …
356 457
357 /* first phase: submit new iocbs */ 458 /* first phase: submit new iocbs */
358 459
359 /* io_submit might return less than the requested number of iocbs */ 460 /* io_submit might return less than the requested number of iocbs */
360 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 461 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
361 /* which allows us to pinpoint the errornous iocb */ 462 /* which allows us to pinpoint the erroneous iocb */
362 for (submitted = 0; submitted < linuxaio_submitcnt; ) 463 for (submitted = 0; submitted < linuxaio_submitcnt; )
363 { 464 {
364#if 0
365 int res;
366 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
367 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
368 else
369 { res = -1; errno = EINVAL; };
370#else
371 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 465 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
372#endif
373 466
374 if (expect_false (res < 0)) 467 if (ecb_expect_false (res < 0))
375 if (errno == EAGAIN)
376 {
377 /* This happens when the ring buffer is full, at least. I assume this means
378 * that the event was queued synchronously during io_submit, and thus
379 * the buffer overflowed.
380 * In this case, we just try in next loop iteration.
381 * This should not result in a few fds taking priority, as the interface
382 * is one-shot, and we submit iocb's in a round-robin fashion.
383 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
384 */
385 if (ecb_expect_false (submitted))
386 {
387 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
388 linuxaio_submitcnt -= submitted;
389 }
390
391 timeout = 0;
392 break;
393 }
394#if EPOLL_FALLBACK
395 else if (errno == EINVAL) 468 if (errno == EINVAL)
396 { 469 {
397 /* This happens for unsupported fds, officially, but in my testing, 470 /* This happens for unsupported fds, officially, but in my testing,
398 * also randomly happens for supported fds. We fall back to good old 471 * also randomly happens for supported fds. We fall back to good old
399 * poll() here, under the assumption that this is a very rare case. 472 * poll() here, under the assumption that this is a very rare case.
400 * See https://lore.kernel.org/patchwork/patch/1047453/ to see 473 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
401 * discussion about such a case (ttys) where polling for POLLIN 474 * discussion about such a case (ttys) where polling for POLLIN
402 * fails but POLLIN|POLLOUT works. 475 * fails but POLLIN|POLLOUT works.
403 */ 476 */
404 struct iocb *iocb = linuxaio_submits [submitted]; 477 struct iocb *iocb = linuxaio_submits [submitted];
405 478 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
406 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */ 479 iocb->aio_reqprio = -1; /* mark iocb as epoll */
480
481 res = 1; /* skip this iocb - another iocb, another chance */
482 }
483 else if (errno == EAGAIN)
484 {
485 /* This happens when the ring buffer is full, or some other shit we
486 * don't know and isn't documented. Most likely because we have too
487 * many requests and linux aio can't be assed to handle them.
488 * In this case, we try to allocate a larger ring buffer, freeing
489 * ours first. This might fail, in which case we have to fall back to 100%
490 * epoll.
491 * God, how I hate linux not getting its act together. Ever.
492 */
493 evsys_io_destroy (linuxaio_ctx);
494 linuxaio_submitcnt = 0;
495
496 /* rearm all fds with active iocbs */
497 {
498 int fd;
499 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
500 if (linuxaio_iocbps [fd]->io.aio_buf)
501 linuxaio_fd_rearm (EV_A_ fd);
502 }
503
504 ++linuxaio_iteration;
505 if (linuxaio_io_setup (EV_A) < 0)
506 {
507 /* TODO: rearm all and recreate epoll backend from scratch */
508 /* TODO: might be more prudent? */
509
510 /* to bad, we can't get a new aio context, go 100% epoll */
511 linuxaio_free_iocbp (EV_A);
512 ev_io_stop (EV_A_ &linuxaio_epoll_w);
513 ev_ref (EV_A);
514 linuxaio_ctx = 0;
515
516 backend = EVBACKEND_EPOLL;
517 backend_modify = epoll_modify;
518 backend_poll = epoll_poll;
519 }
520
521 timeout = EV_TS_CONST (0.);
522 /* it's easiest to handle this mess in another iteration */
523 return;
524 }
525 else if (errno == EBADF)
526 {
527 assert (("libev: event loop rejected bad fd", errno != EBADF));
528 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
408 529
409 res = 1; /* skip this iocb */ 530 res = 1; /* skip this iocb */
410 } 531 }
411#endif 532 else if (errno == EINTR) /* not seen in reality, not documented */
412 else if (errno == EBADF) 533 res = 0; /* silently ignore and retry */
534 else
413 { 535 {
414 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 536 ev_syserr ("(libev) linuxaio io_submit");
415 537 res = 0;
416 res = 1; /* skip this iocb */
417 } 538 }
418 else
419 ev_syserr ("(libev) linuxaio io_submit");
420 539
421 submitted += res; 540 submitted += res;
422 } 541 }
423 542
424 linuxaio_submitcnt = 0; 543 linuxaio_submitcnt = 0;
… …
432int 551int
433linuxaio_init (EV_P_ int flags) 552linuxaio_init (EV_P_ int flags)
434{ 553{
435 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 554 /* would be great to have a nice test for IOCB_CMD_POLL instead */
436 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 555 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
437#if EPOLL_FALLBACK 556 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
438 /* 4.19 made epoll work */
439 if (ev_linux_version () < 0x041300) 557 if (ev_linux_version () < 0x041300)
440 return 0; 558 return 0;
441#else 559
442 /* 4.18 introduced IOCB_CMD_POLL */ 560 if (!epoll_init (EV_A_ 0))
443 if (ev_linux_version () < 0x041200)
444 return 0; 561 return 0;
445#endif
446 562
447 linuxaio_ctx = 0; 563 linuxaio_iteration = 0;
448 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
449 return 0;
450 564
451#if EPOLL_FALLBACK 565 if (linuxaio_io_setup (EV_A) < 0)
452 backend_fd = ev_epoll_create ();
453 if (backend_fd < 0)
454 { 566 {
455 evsys_io_destroy (linuxaio_ctx); 567 epoll_destroy (EV_A);
456 return 0; 568 return 0;
457 } 569 }
458 570
459 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 571 ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
460 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 572 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
461 ev_io_start (EV_A_ &linuxaio_epoll_w); 573 ev_io_start (EV_A_ &linuxaio_epoll_w);
462 ev_unref (EV_A); /* watcher should not keep loop alive */ 574 ev_unref (EV_A); /* watcher should not keep loop alive */
463#endif
464 575
465 backend_modify = linuxaio_modify; 576 backend_modify = linuxaio_modify;
466 backend_poll = linuxaio_poll; 577 backend_poll = linuxaio_poll;
467 578
468 linuxaio_iocbpmax = 0; 579 linuxaio_iocbpmax = 0;
469 linuxaio_iocbps = 0; 580 linuxaio_iocbps = 0;
470 581
471 linuxaio_submits = 0; 582 linuxaio_submits = 0;
… …
477 588
478inline_size 589inline_size
479void 590void
480linuxaio_destroy (EV_P) 591linuxaio_destroy (EV_P)
481{ 592{
482#if EPOLL_FALLBACK 593 epoll_destroy (EV_A);
483 close (backend_fd);
484#endif
485 linuxaio_free_iocbp (EV_A); 594 linuxaio_free_iocbp (EV_A);
486 evsys_io_destroy (linuxaio_ctx); 595 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
487} 596}
488 597
489inline_size 598ecb_cold
490void 599static void
491linuxaio_fork (EV_P) 600linuxaio_fork (EV_P)
492{ 601{
493 /* this frees all iocbs, which is very heavy-handed */
494 linuxaio_destroy (EV_A);
495 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 602 linuxaio_submitcnt = 0; /* all pointers were invalidated */
603 linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
604 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
496 605
497 linuxaio_ctx = 0; 606 linuxaio_iteration = 0; /* we start over in the child */
498 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 607
608 while (linuxaio_io_setup (EV_A) < 0)
499 ev_syserr ("(libev) linuxaio io_setup"); 609 ev_syserr ("(libev) linuxaio io_setup");
500 610
501#if EPOLL_FALLBACK 611 /* forking epoll should also effectively unregister all fds from the backend */
502 while ((backend_fd = ev_epoll_create ()) < 0) 612 epoll_fork (EV_A);
503 ev_syserr ("(libev) linuxaio epoll_create"); 613 /* epoll_fork already did this. hopefully */
614 /*fd_rearm_all (EV_A);*/
504 615
505 ev_io_stop (EV_A_ &linuxaio_epoll_w); 616 ev_io_stop (EV_A_ &linuxaio_epoll_w);
506 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 617 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
507 ev_io_start (EV_A_ &linuxaio_epoll_w); 618 ev_io_start (EV_A_ &linuxaio_epoll_w);
508#endif
509
510 fd_rearm_all (EV_A);
511} 619}
512 620

Diff Legend

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