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 1.20 by root, Mon Jun 24 20:33:24 2019 UTC vs.
Revision 1.36 by root, Wed Jun 26 07:01:14 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 - this section has the raw syscall definitions */ 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 */
… …
78 unsigned header_length; /* size of aio_ring */ 116 unsigned header_length; /* size of aio_ring */
79 117
80 struct io_event io_events[0]; 118 struct io_event io_events[0];
81}; 119};
82 120
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 && !defined __OPTIMIZE_SIZE__
130 /* the costly errno access probably kills this for size optimisation */
131
132 #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \
133 ({ \
134 long res; \
135 register unsigned long r5 __asm__ ("r8" ); \
136 register unsigned long r4 __asm__ ("r10"); \
137 register unsigned long r3 __asm__ ("rdx"); \
138 register unsigned long r2 __asm__ ("rsi"); \
139 register unsigned long r1 __asm__ ("rdi"); \
140 if (narg >= 5) r5 = (unsigned long)(arg5); \
141 if (narg >= 4) r4 = (unsigned long)(arg4); \
142 if (narg >= 3) r3 = (unsigned long)(arg3); \
143 if (narg >= 2) r2 = (unsigned long)(arg2); \
144 if (narg >= 1) r1 = (unsigned long)(arg1); \
145 __asm__ __volatile__ ( \
146 "syscall\n\t" \
147 : "=a" (res) \
148 : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \
149 : "cc", "r11", "cx", "memory"); \
150 errno = -res; \
151 res; \
152 })
153
154#endif
155
156#ifdef ev_syscall
157 #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0
158 #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0)
159 #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0)
160 #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0)
161 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0)
162 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5)
163#else
164 #define ev_syscall0(nr) syscall (nr)
165 #define ev_syscall1(nr,arg1) syscall (nr, arg1)
166 #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2)
167 #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3)
168 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4)
169 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5)
170#endif
171
83inline_size 172inline_size
84int 173int
85evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86{ 175{
87 return syscall (SYS_io_setup, nr_events, ctx_idp); 176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
88} 177}
89 178
90inline_size 179inline_size
91int 180int
92evsys_io_destroy (aio_context_t ctx_id) 181evsys_io_destroy (aio_context_t ctx_id)
93{ 182{
94 return syscall (SYS_io_destroy, ctx_id); 183 return ev_syscall1 (SYS_io_destroy, ctx_id);
95} 184}
96 185
97inline_size 186inline_size
98int 187int
99evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 188evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100{ 189{
101 return syscall (SYS_io_submit, ctx_id, nr, cbp); 190 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
102} 191}
103 192
104inline_size 193inline_size
105int 194int
106evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 195evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
107{ 196{
108 return syscall (SYS_io_cancel, ctx_id, cbp, result); 197 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
109} 198}
110 199
111inline_size 200inline_size
112int 201int
113evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 202evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
114{ 203{
115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 204 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116} 205}
117 206
118/*****************************************************************************/ 207/*****************************************************************************/
119/* actual backed implementation */ 208/* actual backed implementation */
120 209
210ecb_cold
211static int
212linuxaio_nr_events (EV_P)
213{
214 /* we start with 16 iocbs and incraese from there
215 * that's tiny, but the kernel has a rather low system-wide
216 * limit that can be reached quickly, so let's be parsimonious
217 * with this resource.
218 * Rest assured, the kernel generously rounds up small and big numbers
219 * in different ways (but doesn't seem to charge you for it).
220 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
221 * and this helps to take advantage of that limit.
222 */
223
224 /* we try to fill 4kB pages exactly.
225 * the ring buffer header is 32 bytes, every io event is 32 bytes.
226 * the kernel takes the io requests number, doubles it, adds 2
227 * and adds the ring buffer.
228 * the way we use this is by starting low, and then roughly doubling the
229 * size each time we hit a limit.
230 */
231
232 int requests = 15 << linuxaio_iteration;
233 int one_page = (4096
234 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
235 int first_page = ((4096 - sizeof (struct aio_ring))
236 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
237
238 /* if everything fits into one page, use count exactly */
239 if (requests > first_page)
240 /* otherwise, round down to full pages and add the first page */
241 requests = requests / one_page * one_page + first_page;
242
243 return requests;
244}
245
121/* we use out own wrapper structure in acse we ever want to do something "clever" */ 246/* we use out own wrapper structure in case we ever want to do something "clever" */
122typedef struct aniocb 247typedef struct aniocb
123{ 248{
124 struct iocb io; 249 struct iocb io;
125 /*int inuse;*/ 250 /*int inuse;*/
126} *ANIOCBP; 251} *ANIOCBP;
127 252
128inline_size 253inline_size
129void 254void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 255linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 256{
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--) 257 while (count--)
134 { 258 {
259 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
135 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 260 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 /* TODO: full zero initialize required? */ 261
262 /* full zero initialise is probably not required at the moment, but
263 * this is not well documented, so we better do it.
264 */
137 memset (*base, 0, sizeof (**base)); 265 memset (iocb, 0, sizeof (*iocb));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 266
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 267 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base; 268 iocb->io.aio_data = offset;
269 iocb->io.aio_fildes = offset;
270
271 base [offset++] = iocb;
141 } 272 }
142} 273}
143 274
144ecb_cold 275ecb_cold
145static void 276static void
… …
153 284
154static void 285static void
155linuxaio_modify (EV_P_ int fd, int oev, int nev) 286linuxaio_modify (EV_P_ int fd, int oev, int nev)
156{ 287{
157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 288 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158 struct aniocb *iocb = linuxaio_iocbps [fd]; 289 ANIOCBP iocb = linuxaio_iocbps [fd];
159 290
160#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0) 291 if (iocb->io.aio_reqprio < 0)
162 { 292 {
293 /* we handed this fd over to epoll, so undo this first */
294 /* 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); 295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0;
164 iocb->io.aio_reqprio = 0; 297 iocb->io.aio_reqprio = 0;
165 } 298 }
166#endif
167 299
168 if (iocb->io.aio_buf) 300 if (iocb->io.aio_buf)
169 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 301 {
302 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
303 /* on relevant kernels, io_cancel fails with EINPROGRES if everything is fine */
304 assert (("libev: linuxaio unexpected io_cancel failed", errno == EINPROGRESS));
305 }
170 306
171 if (nev) 307 if (nev)
172 { 308 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf = 309 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0) 310 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0); 311 | (nev & EV_WRITE ? POLLOUT : 0);
178 312
179 /* queue iocb up for io_submit */ 313 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */ 314 /* this assumes we only ever get one call per fd per loop iteration */
… …
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 316 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 317 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 318 }
185} 319}
186 320
187#if EPOLL_FALLBACK
188
189static void
190linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
191{
192 struct epoll_event eev;
193
194 eev.events = EPOLLONESHOT;
195 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
196 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
197 eev.data.fd = iocb->aio_fildes;
198
199 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
200 ev_syserr ("(libeio) linuxaio epoll_ctl");
201}
202
203static void 321static void
204linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 322linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
205{ 323{
206 struct epoll_event events[16]; 324 epoll_poll (EV_A_ 0);
207
208 for (;;)
209 {
210 int idx;
211 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
212
213 if (expect_false (res < 0))
214 ev_syserr ("(libev) linuxaio epoll_wait");
215 else if (!res)
216 break;
217
218 for (idx = res; idx--; )
219 {
220 int fd = events [idx].data.fd;
221 uint32_t ev = events [idx].events;
222
223 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
224
225 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
226
227 fd_event (EV_A_ fd,
228 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
229 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
230 }
231
232 if (res < sizeof (events) / sizeof (events [0]))
233 break;
234 }
235} 325}
236 326
237#endif 327inline_speed
328void
329linuxaio_fd_rearm (EV_P_ int fd)
330{
331 anfds [fd].events = 0;
332 linuxaio_iocbps [fd]->io.aio_buf = 0;
333 fd_change (EV_A_ fd, EV_ANFD_REIFY);
334}
238 335
239static void 336static void
240linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 337linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
241{ 338{
242 while (nr) 339 while (nr)
… …
244 int fd = ev->data; 341 int fd = ev->data;
245 int res = ev->res; 342 int res = ev->res;
246 343
247 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 344 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
248 345
249 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
250 linuxaio_iocbps [fd]->io.aio_buf = 0;
251 anfds [fd].events = 0;
252 fd_change (EV_A_ fd, 0);
253
254 /* feed events, we do not expect or handle POLLNVAL */ 346 /* feed events, we do not expect or handle POLLNVAL */
255 if (expect_false (res & POLLNVAL))
256 fd_kill (EV_A_ fd);
257 else
258 fd_event ( 347 fd_event (
259 EV_A_ 348 EV_A_
260 fd, 349 fd,
261 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 350 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
262 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 351 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
263 ); 352 );
353
354 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
355 linuxaio_fd_rearm (EV_A_ fd);
264 356
265 --nr; 357 --nr;
266 ++ev; 358 ++ev;
267 } 359 }
268} 360}
269 361
270/* get any events from ringbuffer, return true if any were handled */ 362/* get any events from ring buffer, return true if any were handled */
271static int 363static int
272linuxaio_get_events_from_ring (EV_P) 364linuxaio_get_events_from_ring (EV_P)
273{ 365{
274 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 366 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
275 367
… …
298 { 390 {
299 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 391 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
300 linuxaio_parse_events (EV_A_ ring->io_events, tail); 392 linuxaio_parse_events (EV_A_ ring->io_events, tail);
301 } 393 }
302 394
303 /* TODO: we only need a compiler barrier here, not a read fence */
304 ECB_MEMORY_FENCE_RELEASE; 395 ECB_MEMORY_FENCE_RELEASE;
305 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 396 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
306 *(volatile unsigned *)&ring->head = tail; 397 *(volatile unsigned *)&ring->head = tail;
307 /* make sure kernel can see our new head value - probably not required */
308 ECB_MEMORY_FENCE_RELEASE;
309 398
310 return 1; 399 return 1;
311} 400}
312 401
313/* read at least one event from kernel, or timeout */ 402/* read at least one event from kernel, or timeout */
… …
340 /* ignored */; 429 /* ignored */;
341 else 430 else
342 ev_syserr ("(libev) linuxaio io_getevents"); 431 ev_syserr ("(libev) linuxaio io_getevents");
343 else if (res) 432 else if (res)
344 { 433 {
345 /* at least one event received, handle it and any remaining ones in the ring buffer */ 434 /* at least one event available, handle it and any remaining ones in the ring buffer */
346 linuxaio_parse_events (EV_A_ ioev, res); 435 linuxaio_parse_events (EV_A_ ioev, res);
347 linuxaio_get_events_from_ring (EV_A); 436 linuxaio_get_events_from_ring (EV_A);
348 } 437 }
349} 438}
350 439
440inline_size
441int
442linuxaio_io_setup (EV_P)
443{
444 linuxaio_ctx = 0;
445 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
446}
447
351static void 448static void
352linuxaio_poll (EV_P_ ev_tstamp timeout) 449linuxaio_poll (EV_P_ ev_tstamp timeout)
353{ 450{
354 int submitted; 451 int submitted;
355 452
356 /* first phase: submit new iocbs */ 453 /* first phase: submit new iocbs */
357 454
358 /* io_submit might return less than the requested number of iocbs */ 455 /* io_submit might return less than the requested number of iocbs */
359 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 456 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
360 /* which allows us to pinpoint the errornous iocb */ 457 /* which allows us to pinpoint the erroneous iocb */
361 for (submitted = 0; submitted < linuxaio_submitcnt; ) 458 for (submitted = 0; submitted < linuxaio_submitcnt; )
362 { 459 {
363#if 0
364 int res;
365 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
366 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
367 else
368 { res = -1; errno = EINVAL; };
369#else
370 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 460 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
371#endif
372 461
373 if (expect_false (res < 0)) 462 if (expect_false (res < 0))
374 if (errno == EAGAIN)
375 {
376 /* This happens when the ring buffer is full, at least. I assume this means
377 * that the event was queued synchronously during io_submit, and thus
378 * the buffer overflowed.
379 * In this case, we just try in next loop iteration.
380 * This should not result in a few fds taking priority, as the interface
381 * is one-shot, and we submit iocb's in a round-robin fashion.
382 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
383 */
384 if (ecb_expect_false (submitted))
385 {
386 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
387 linuxaio_submitcnt -= submitted;
388 }
389
390 timeout = 0;
391 break;
392 }
393#if EPOLL_FALLBACK
394 else if (errno == EINVAL) 463 if (errno == EINVAL)
395 { 464 {
396 /* This happens for unsupported fds, officially, but in my testing, 465 /* This happens for unsupported fds, officially, but in my testing,
397 * also randomly happens for supported fds. We fall back to good old 466 * also randomly happens for supported fds. We fall back to good old
398 * poll() here, under the assumption that this is a very rare case. 467 * poll() here, under the assumption that this is a very rare case.
399 * See https://lore.kernel.org/patchwork/patch/1047453/ to see 468 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
400 * discussion about such a case (ttys) where polling for POLLIN 469 * discussion about such a case (ttys) where polling for POLLIN
401 * fails but POLLIN|POLLOUT works. 470 * fails but POLLIN|POLLOUT works.
402 */ 471 */
403 struct iocb *iocb = linuxaio_submits [submitted]; 472 struct iocb *iocb = linuxaio_submits [submitted];
473 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
474 iocb->aio_reqprio = -1; /* mark iocb as epoll */
475
476 res = 1; /* skip this iocb - another iocb, another chance */
477 }
478 else if (errno == EAGAIN)
479 {
480 /* This happens when the ring buffer is full, or some other shit we
481 * don't know and isn't documented. Most likely because we have too
482 * many requests and linux aio can't be assed to handle them.
483 * In this case, we try to allocate a larger ring buffer, freeing
484 * ours first. This might fail, in which case we have to fall back to 100%
485 * epoll.
486 * God, how I hate linux not getting its act together. Ever.
487 */
488 evsys_io_destroy (linuxaio_ctx);
489 linuxaio_submitcnt = 0;
490
491 /* rearm all fds with active iocbs */
492 {
493 int fd;
494 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
495 if (linuxaio_iocbps [fd]->io.aio_buf)
496 linuxaio_fd_rearm (EV_A_ fd);
497 }
498
499 ++linuxaio_iteration;
500 if (linuxaio_io_setup (EV_A) < 0)
501 {
502 /* to bad, we can't get a new aio context, go 100% epoll */
503 linuxaio_free_iocbp (EV_A);
504 ev_io_stop (EV_A_ &linuxaio_epoll_w);
505 ev_ref (EV_A);
506 linuxaio_ctx = 0;
507 backend_modify = epoll_modify;
508 backend_poll = epoll_poll;
509 }
510
511 timeout = 0;
512 /* it's easiest to handle this mess in another iteration */
513 return;
514 }
515 else if (errno == EBADF)
516 {
517 assert (("libev: event loop rejected bad fd", errno != EBADF));
518 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
519
404 res = 1; /* skip this iocb */ 520 res = 1; /* skip this iocb */
405
406 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408 } 521 }
409#endif
410 else 522 else
411 ev_syserr ("(libev) linuxaio io_submit"); 523 ev_syserr ("(libev) linuxaio io_submit");
412 524
413 submitted += res; 525 submitted += res;
414 } 526 }
… …
424int 536int
425linuxaio_init (EV_P_ int flags) 537linuxaio_init (EV_P_ int flags)
426{ 538{
427 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 539 /* would be great to have a nice test for IOCB_CMD_POLL instead */
428 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 540 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
429#if EPOLL_FALLBACK 541 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
430 /* 4.19 made epoll work */
431 if (ev_linux_version () < 0x041300) 542 if (ev_linux_version () < 0x041300)
432 return 0; 543 return 0;
433#else 544
434 /* 4.18 introduced IOCB_CMD_POLL */ 545 if (!epoll_init (EV_A_ 0))
435 if (ev_linux_version () < 0x041200)
436 return 0; 546 return 0;
437#endif
438 547
439 linuxaio_ctx = 0; 548 linuxaio_iteration = 0;
440 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
441 return 0;
442 549
443#if EPOLL_FALLBACK 550 if (linuxaio_io_setup (EV_A) < 0)
444 backend_fd = ev_epoll_create ();
445 if (backend_fd < 0)
446 { 551 {
447 evsys_io_destroy (linuxaio_ctx); 552 epoll_destroy (EV_A);
448 return 0; 553 return 0;
449 } 554 }
450 555
451 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 556 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
452 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 557 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
453 ev_io_start (EV_A_ &linuxaio_epoll_w); 558 ev_io_start (EV_A_ &linuxaio_epoll_w);
454 ev_unref (EV_A); /* watcher should not keep loop alive */ 559 ev_unref (EV_A); /* watcher should not keep loop alive */
455#endif
456 560
457 backend_modify = linuxaio_modify; 561 backend_modify = linuxaio_modify;
458 backend_poll = linuxaio_poll; 562 backend_poll = linuxaio_poll;
459 563
460 linuxaio_iocbpmax = 0; 564 linuxaio_iocbpmax = 0;
… …
469 573
470inline_size 574inline_size
471void 575void
472linuxaio_destroy (EV_P) 576linuxaio_destroy (EV_P)
473{ 577{
474#if EPOLL_FALLBACK 578 epoll_destroy (EV_A);
475 close (backend_fd);
476#endif
477 linuxaio_free_iocbp (EV_A); 579 linuxaio_free_iocbp (EV_A);
478 ev_io_destroy (linuxaio_ctx); 580 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
479} 581}
480 582
481inline_size 583inline_size
482void 584void
483linuxaio_fork (EV_P) 585linuxaio_fork (EV_P)
484{ 586{
485 /* this frees all iocbs, which is very heavy-handed */ 587 /* this frees all iocbs, which is very heavy-handed */
486 linuxaio_destroy (EV_A); 588 linuxaio_destroy (EV_A);
487 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 589 linuxaio_submitcnt = 0; /* all pointers were invalidated */
488 590
489 linuxaio_ctx = 0; 591 linuxaio_iteration = 0; /* we start over in the child */
490 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 592
593 while (linuxaio_io_setup (EV_A) < 0)
491 ev_syserr ("(libev) linuxaio io_setup"); 594 ev_syserr ("(libev) linuxaio io_setup");
492 595
493#if EPOLL_FALLBACK 596 /* forking epoll should also effectively unregister all fds from the backend */
494 while ((backend_fd = ev_epoll_create ()) < 0) 597 epoll_fork (EV_A);
495 ev_syserr ("(libev) linuxaio epoll_create");
496 598
497 ev_io_stop (EV_A_ &linuxaio_epoll_w); 599 ev_io_stop (EV_A_ &linuxaio_epoll_w);
498 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 600 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
499 ev_io_start (EV_A_ &linuxaio_epoll_w); 601 ev_io_start (EV_A_ &linuxaio_epoll_w);
500#endif
501 602
603 /* epoll_fork already did this. hopefully */
502 fd_rearm_all (EV_A); 604 /*fd_rearm_all (EV_A);*/
503} 605}
504 606

Diff Legend

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