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 1.9 by root, Sat Jun 22 22:29:38 2019 UTC vs.
Revision 1.30 by root, Tue Jun 25 17:54:02 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 4kB pages exactly.
45 * the ring buffer header is 32 bytes, every io event is 32 bytes.
46 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
47 * therefore the calculation below will use "exactly" 4kB for the ring buffer
48 */
49#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
50
51/*****************************************************************************/ 91/*****************************************************************************/
52/* syscall wrapdadoop */ 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
53 93
54#include <sys/syscall.h> /* no glibc wrappers */ 94#include <sys/syscall.h> /* no glibc wrappers */
55 95
56/* 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 */
57#define IOCB_CMD_POLL 5 97#define IOCB_CMD_POLL 5
58 98
59/* 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 */
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
121/*
122 * define some syscall wrappers for common architectures
123 * this is mostly for nice looks during debugging, not performance.
124 * our syscalls return < 0, not == -1, on error. which is good
125 * enough for linux aio.
126 * TODO: arm is also common nowadays, maybe even mips and x86
127 * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove...
128 */
129#if __GNUC__ && __linux && ECB_AMD64
130
131 #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \
132 ({ \
133 long res; \
134 register unsigned long r5 __asm__ ("r8" ); \
135 register unsigned long r4 __asm__ ("r10"); \
136 register unsigned long r3 __asm__ ("rdx"); \
137 register unsigned long r2 __asm__ ("rsi"); \
138 register unsigned long r1 __asm__ ("rdi"); \
139 if (narg >= 5) r5 = (unsigned long)(arg5); \
140 if (narg >= 4) r4 = (unsigned long)(arg4); \
141 if (narg >= 3) r3 = (unsigned long)(arg3); \
142 if (narg >= 2) r2 = (unsigned long)(arg2); \
143 if (narg >= 1) r1 = (unsigned long)(arg1); \
144 __asm__ __volatile__ ( \
145 "syscall\n\t" \
146 : "=a" (res) \
147 : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \
148 : "cc", "r11", "cx", "memory"); \
149 errno = -res; \
150 res; \
151 })
152
153#endif
154
155#ifdef ev_syscall
156 #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0
157 #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0)
158 #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0)
159 #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0)
160 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0)
161 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5)
162#else
163 #define ev_syscall0(nr) syscall (nr)
164 #define ev_syscall1(nr,arg1) syscall (nr, arg1)
165 #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2)
166 #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3)
167 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4)
168 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5)
169#endif
170
77inline_size 171inline_size
78int 172int
79ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 173evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
80{ 174{
81 return syscall (SYS_io_setup, nr_events, ctx_idp); 175 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
82} 176}
83 177
84inline_size 178inline_size
85int 179int
86ev_io_destroy (aio_context_t ctx_id) 180evsys_io_destroy (aio_context_t ctx_id)
87{ 181{
88 return syscall (SYS_io_destroy, ctx_id); 182 return ev_syscall1 (SYS_io_destroy, ctx_id);
89} 183}
90 184
91inline_size 185inline_size
92int 186int
93ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 187evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
94{ 188{
95 return syscall (SYS_io_submit, ctx_id, nr, cbp); 189 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
96} 190}
97 191
98inline_size 192inline_size
99int 193int
100ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 194evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
101{ 195{
102 return syscall (SYS_io_cancel, ctx_id, cbp, result); 196 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
103} 197}
104 198
105inline_size 199inline_size
106int 200int
107ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 201evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
108{ 202{
109 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 203 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
110} 204}
111 205
112/*****************************************************************************/ 206/*****************************************************************************/
113/* actual backed implementation */ 207/* actual backed implementation */
114 208
209ecb_cold
210static int
211linuxaio_nr_events (EV_P)
212{
213 /* we start with 16 iocbs and incraese from there
214 * that's tiny, but the kernel has a rather low system-wide
215 * limit that can be reached quickly, so let's be parsimonious
216 * with this resource.
217 * Rest assured, the kernel generously rounds up small and big numbers
218 * in different ways (but doesn't seem to charge you for it).
219 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
220 * and this helps to take advantage of that limit.
221 */
222
223 /* we try to fill 4kB pages exactly.
224 * the ring buffer header is 32 bytes, every io event is 32 bytes.
225 * the kernel takes the io requests number, doubles it, adds 2
226 * and adds the ring buffer.
227 * the way we use this is by starting low, and then roughly doubling the
228 * size each time we hit a limit.
229 */
230
231 int requests = 15 << linuxaio_iteration;
232 int one_page = (4096
233 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
234 int first_page = ((4096 - sizeof (struct aio_ring))
235 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
236
237 /* if everything fits into one page, use count exactly */
238 if (requests > first_page)
239 /* otherwise, round down to full pages and add the first page */
240 requests = requests / one_page * one_page + first_page;
241
242 return requests;
243}
244
115/* we use out own wrapper structure in acse we ever want to do something "clever" */ 245/* we use out own wrapper structure in case we ever want to do something "clever" */
116typedef struct aniocb 246typedef struct aniocb
117{ 247{
118 struct iocb io; 248 struct iocb io;
119 /*int inuse;*/ 249 /*int inuse;*/
120} *ANIOCBP; 250} *ANIOCBP;
121 251
122inline_size 252inline_size
123void 253void
124linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 254linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
125{ 255{
126 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
127 while (count--) 256 while (count--)
128 { 257 {
258 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
129 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 259 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
130 /* TODO: full zero initialize required? */ 260
261 /* full zero initialise is probably not required at the moment, but
262 * this is not well documented, so we better do it.
263 */
131 memset (*base, 0, sizeof (**base)); 264 memset (iocb, 0, sizeof (*iocb));
132 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 265
133 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 266 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
134 ++base; 267 iocb->io.aio_data = offset;
268 iocb->io.aio_fildes = offset;
269
270 base [offset++] = iocb;
135 } 271 }
136} 272}
137 273
138ecb_cold 274ecb_cold
139static void 275static void
… …
147 283
148static void 284static void
149linuxaio_modify (EV_P_ int fd, int oev, int nev) 285linuxaio_modify (EV_P_ int fd, int oev, int nev)
150{ 286{
151 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 287 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
152 struct aniocb *iocb = linuxaio_iocbps [fd]; 288 ANIOCBP iocb = linuxaio_iocbps [fd];
289
290 if (iocb->io.aio_reqprio < 0)
291 {
292 /* we handed this fd over to epoll, so undo this first */
293 /* we do it manually because the optimisations on epoll_modfy won't do us any good */
294 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
295 anfds [fd].emask = 0;
296 iocb->io.aio_reqprio = 0;
297 }
153 298
154 if (iocb->io.aio_buf) 299 if (iocb->io.aio_buf)
155 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 300 /* io_cancel always returns some error on relevant kernels, but works */
301 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
156 302
157 if (nev) 303 if (nev)
158 { 304 {
159 iocb->io.aio_data = fd;
160 iocb->io.aio_fildes = fd;
161 iocb->io.aio_buf = 305 iocb->io.aio_buf =
162 (nev & EV_READ ? POLLIN : 0) 306 (nev & EV_READ ? POLLIN : 0)
163 | (nev & EV_WRITE ? POLLOUT : 0); 307 | (nev & EV_WRITE ? POLLOUT : 0);
164 308
165 /* queue iocb up for io_submit */ 309 /* queue iocb up for io_submit */
166 /* this assumes we only ever get one call per fd per loop iteration */ 310 /* this assumes we only ever get one call per fd per loop iteration */
… …
169 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 313 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
170 } 314 }
171} 315}
172 316
173static void 317static void
318linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
319{
320 epoll_poll (EV_A_ 0);
321}
322
323static void
324linuxaio_fd_rearm (EV_P_ int fd)
325{
326 anfds [fd].events = 0;
327 linuxaio_iocbps [fd]->io.aio_buf = 0;
328 fd_change (EV_A_ fd, EV_ANFD_REIFY);
329}
330
331static void
174linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 332linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
175{ 333{
176 while (nr) 334 while (nr)
177 { 335 {
178 int fd = ev->data; 336 int fd = ev->data;
179 int res = ev->res; 337 int res = ev->res;
180 338
181 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 339 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
182 340
183 /* linux aio is oneshot: rearm fd */
184 linuxaio_iocbps [fd]->io.aio_buf = 0;
185 anfds [fd].events = 0;
186 fd_change (EV_A_ fd, 0);
187
188 /* feed events, we do not expect or handle POLLNVAL */ 341 /* feed events, we do not expect or handle POLLNVAL */
189 if (ecb_expect_false (res & POLLNVAL))
190 fd_kill (EV_A_ fd);
191 else
192 fd_event ( 342 fd_event (
193 EV_A_ 343 EV_A_
194 fd, 344 fd,
195 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 345 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
196 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 346 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
197 ); 347 );
348
349 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
350 linuxaio_fd_rearm (EV_A_ fd);
198 351
199 --nr; 352 --nr;
200 ++ev; 353 ++ev;
201 } 354 }
202} 355}
203 356
204/* get any events from ringbuffer, return true if any were handled */ 357/* get any events from ring buffer, return true if any were handled */
205static int 358static int
206linuxaio_get_events_from_ring (EV_P) 359linuxaio_get_events_from_ring (EV_P)
207{ 360{
208 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 361 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
209 362
210 unsigned head = ring->head; 363 /* the kernel reads and writes both of these variables, */
364 /* as a C extension, we assume that volatile use here */
365 /* both makes reads atomic and once-only */
366 unsigned head = *(volatile unsigned *)&ring->head;
211 unsigned tail = *(volatile unsigned *)&ring->tail; 367 unsigned tail = *(volatile unsigned *)&ring->tail;
212 368
213 if (head == tail) 369 if (head == tail)
214 return 0; 370 return 0;
215 371
216 /* bail out if the ring buffer doesn't match the expected layout */ 372 /* bail out if the ring buffer doesn't match the expected layout */
217 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 373 if (expect_false (ring->magic != AIO_RING_MAGIC)
218 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 374 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
219 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ 375 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
220 return 0; 376 return 0;
221 377
378 /* make sure the events up to tail are visible */
222 ECB_MEMORY_FENCE_ACQUIRE; 379 ECB_MEMORY_FENCE_ACQUIRE;
223 380
224 /* parse all available events, but only once, to avoid starvation */ 381 /* parse all available events, but only once, to avoid starvation */
225 if (tail > head) /* normal case around */ 382 if (tail > head) /* normal case around */
226 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 383 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
… …
228 { 385 {
229 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 386 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
230 linuxaio_parse_events (EV_A_ ring->io_events, tail); 387 linuxaio_parse_events (EV_A_ ring->io_events, tail);
231 } 388 }
232 389
233 ring->head = tail; 390 ECB_MEMORY_FENCE_RELEASE;
391 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
392 *(volatile unsigned *)&ring->head = tail;
234 393
235 return 1; 394 return 1;
236} 395}
237 396
238/* read at least one event from kernel, or timeout */ 397/* read at least one event from kernel, or timeout */
239inline_size 398inline_size
240void 399void
241linuxaio_get_events (EV_P_ ev_tstamp timeout) 400linuxaio_get_events (EV_P_ ev_tstamp timeout)
242{ 401{
243 struct timespec ts; 402 struct timespec ts;
244 struct io_event ioev; 403 struct io_event ioev[1];
245 int res; 404 int res;
246 405
247 if (linuxaio_get_events_from_ring (EV_A)) 406 if (linuxaio_get_events_from_ring (EV_A))
248 return; 407 return;
249 408
250 /* no events, so wait for at least one, then poll ring buffer again */ 409 /* no events, so wait for at least one, then poll ring buffer again */
251 /* this degrades to one event per loop iteration */ 410 /* this degrades to one event per loop iteration */
252 /* if the ring buffer changes layout, but so be it */ 411 /* if the ring buffer changes layout, but so be it */
253 412
413 EV_RELEASE_CB;
414
254 ts.tv_sec = (long)timeout; 415 ts.tv_sec = (long)timeout;
255 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 416 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
256 417
257 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 418 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
419
420 EV_ACQUIRE_CB;
258 421
259 if (res < 0) 422 if (res < 0)
423 if (errno == EINTR)
424 /* ignored */;
425 else
260 ev_syserr ("(libev) linuxaio io_getevents"); 426 ev_syserr ("(libev) linuxaio io_getevents");
261 else if (res) 427 else if (res)
262 { 428 {
263 /* at least one event received, handle it and any remaining ones in the ring buffer */ 429 /* at least one event received, handle it and any remaining ones in the ring buffer */
264 linuxaio_parse_events (EV_A_ &ioev, 1); 430 linuxaio_parse_events (EV_A_ ioev, res);
265 linuxaio_get_events_from_ring (EV_A); 431 linuxaio_get_events_from_ring (EV_A);
266 } 432 }
433}
434
435static int
436linuxaio_io_setup (EV_P)
437{
438 linuxaio_ctx = 0;
439 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
267} 440}
268 441
269static void 442static void
270linuxaio_poll (EV_P_ ev_tstamp timeout) 443linuxaio_poll (EV_P_ ev_tstamp timeout)
271{ 444{
… …
273 446
274 /* first phase: submit new iocbs */ 447 /* first phase: submit new iocbs */
275 448
276 /* io_submit might return less than the requested number of iocbs */ 449 /* io_submit might return less than the requested number of iocbs */
277 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 450 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
278 /* which allows us to pinpoint the errornous iocb */ 451 /* which allows us to pinpoint the erroneous iocb */
279 for (submitted = 0; submitted < linuxaio_submitcnt; ) 452 for (submitted = 0; submitted < linuxaio_submitcnt; )
280 { 453 {
281 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 454 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
282 455
283 if (ecb_expect_false (res < 0)) 456 if (expect_false (res < 0))
284 if (errno == EAGAIN) 457 if (errno == EINVAL)
285 { 458 {
286 /* This happens when the ring buffer is full, at least. I assume this means 459 /* This happens for unsupported fds, officially, but in my testing,
287 * that the event was queued synchronously during io_submit, and thus 460 * also randomly happens for supported fds. We fall back to good old
288 * the buffer overflowd. 461 * poll() here, under the assumption that this is a very rare case.
289 * In this case, we just try next loop iteration. 462 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
290 * This should not result in a few fds taking priority, as the interface 463 * discussion about such a case (ttys) where polling for POLLIN
291 * is one-shot, and we submit iocb's in a round-robin fashion. 464 * fails but POLLIN|POLLOUT works.
292 */ 465 */
293 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 466 struct iocb *iocb = linuxaio_submits [submitted];
467 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
468 iocb->aio_reqprio = -1; /* mark iocb as epoll */
469
470 res = 1; /* skip this iocb - another iocb, another chance */
471 }
472 else if (errno == EAGAIN)
473 {
474 /* This happens when the ring buffer is full, or some other shit we
475 * don't know and isn't documented. Most likely because we have too
476 * many requests and linux aio can't be assed to handle them.
477 * In this case, we try to allocate a larger ring buffer, freeing
478 * ours first. This might fail, in which case we have to fall back to 100%
479 * epoll.
480 * God, how I hate linux not getting its act together. Ever.
481 */
482 evsys_io_destroy (linuxaio_ctx);
294 linuxaio_submitcnt -= submitted; 483 linuxaio_submitcnt = 0;
484
485 /* rearm all fds with active iocbs */
486 {
487 int fd;
488 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
489 if (linuxaio_iocbps [fd]->io.aio_buf)
490 linuxaio_fd_rearm (EV_A_ fd);
491 }
492
493 ++linuxaio_iteration;
494 if (linuxaio_io_setup (EV_A) < 0)
495 {
496 /* to bad, we can't get a new aio context, go 100% epoll */
497 linuxaio_free_iocbp (EV_A);
498 ev_io_stop (EV_A_ &linuxaio_epoll_w);
499 ev_ref (EV_A);
500 linuxaio_ctx = 0;
501 backend_modify = epoll_modify;
502 backend_poll = epoll_poll;
503 }
504
295 timeout = 0; 505 timeout = 0;
506 /* it's easiest to handle this mess in another iteration */
296 break; 507 return;
508 }
509 else if (errno == EBADF)
510 {
511 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
512
513 res = 1; /* skip this iocb */
297 } 514 }
298 else 515 else
299 ev_syserr ("(libev) linuxaio io_submit"); 516 ev_syserr ("(libev) linuxaio io_submit");
300 517
301 submitted += res; 518 submitted += res;
… …
312int 529int
313linuxaio_init (EV_P_ int flags) 530linuxaio_init (EV_P_ int flags)
314{ 531{
315 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 532 /* would be great to have a nice test for IOCB_CMD_POLL instead */
316 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 533 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
317 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 534 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
535 if (ev_linux_version () < 0x041300)
318 return 0; 536 return 0;
319 537
320 linuxaio_ctx = 0; 538 if (!epoll_init (EV_A_ 0))
321 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
322 return 0; 539 return 0;
540
541 linuxaio_iteration = 0;
542
543 if (linuxaio_io_setup (EV_A) < 0)
544 {
545 epoll_destroy (EV_A);
546 return 0;
547 }
548
549 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
550 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
551 ev_io_start (EV_A_ &linuxaio_epoll_w);
552 ev_unref (EV_A); /* watcher should not keep loop alive */
323 553
324 backend_modify = linuxaio_modify; 554 backend_modify = linuxaio_modify;
325 backend_poll = linuxaio_poll; 555 backend_poll = linuxaio_poll;
326 556
327 linuxaio_iocbpmax = 0; 557 linuxaio_iocbpmax = 0;
… …
336 566
337inline_size 567inline_size
338void 568void
339linuxaio_destroy (EV_P) 569linuxaio_destroy (EV_P)
340{ 570{
571 epoll_destroy (EV_A);
341 linuxaio_free_iocbp (EV_A); 572 linuxaio_free_iocbp (EV_A);
342 ev_io_destroy (linuxaio_ctx); 573 evsys_io_destroy (linuxaio_ctx);
343} 574}
344 575
345inline_size 576inline_size
346void 577void
347linuxaio_fork (EV_P) 578linuxaio_fork (EV_P)
348{ 579{
349 /* this frees all iocbs, which is very heavy-handed */ 580 /* this frees all iocbs, which is very heavy-handed */
350 linuxaio_destroy (EV_A); 581 linuxaio_destroy (EV_A);
351 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 582 linuxaio_submitcnt = 0; /* all pointers were invalidated */
352 583
353 linuxaio_ctx = 0; 584 linuxaio_iteration = 0; /* we start over in the child */
354 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 585
586 while (linuxaio_io_setup (EV_A) < 0)
355 ev_syserr ("(libev) linuxaio io_setup"); 587 ev_syserr ("(libev) linuxaio io_setup");
356 588
589 epoll_fork (EV_A);
590
591 ev_io_stop (EV_A_ &linuxaio_epoll_w);
592 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
593 ev_io_start (EV_A_ &linuxaio_epoll_w);
594
595 /* epoll_fork already did this. hopefully */
357 fd_rearm_all (EV_A); 596 /*fd_rearm_all (EV_A);*/
358} 597}
359 598

Diff Legend

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