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 1.1 by root, Thu Jun 20 22:44:59 2019 UTC vs.
Revision 1.34 by root, Wed Jun 26 00:01:46 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 */
88#include <poll.h>
41#include <linux/aio_abi.h> 89#include <linux/aio_abi.h>
42 90
43/* we try to fill 4kn pages exactly. 91/*****************************************************************************/
44 * the ring buffer header is 32 bytes, every io event is 32 bytes. 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
45 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer 93
46 * so the calculation below will use "exactly" 8kB for the ring buffer 94#include <sys/syscall.h> /* no glibc wrappers */
95
96/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
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.
47 */ 103 */
48#define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */
49
50/*****************************************************************************/
51/* syscall wrapdadoop */
52
53#include <sys/syscall.h> /* no glibc wrappers */
54
55/* aio_abi.h is not verioned in any way, so we cannot test for its existance */
56#define IOCB_CMD_POLL 5
57
58/* taken from linux/fs/aio.c */
59#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
60#define AIO_RING_INCOMPAT_FEATURES 0 105#define AIO_RING_INCOMPAT_FEATURES 0
61struct aio_ring 106struct aio_ring
62{ 107{
63 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
… …
71 unsigned header_length; /* size of aio_ring */ 116 unsigned header_length; /* size of aio_ring */
72 117
73 struct io_event io_events[0]; 118 struct io_event io_events[0];
74}; 119};
75 120
76static int 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
172inline_size
173int
77ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
78{ 175{
79 return syscall (SYS_io_setup, nr_events, ctx_idp); 176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
80} 177}
81 178
82static int 179inline_size
180int
83ev_io_destroy (aio_context_t ctx_id) 181evsys_io_destroy (aio_context_t ctx_id)
84{ 182{
85 return syscall (SYS_io_destroy, ctx_id); 183 return ev_syscall1 (SYS_io_destroy, ctx_id);
86} 184}
87 185
88static int 186inline_size
187int
89ev_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[])
90{ 189{
91 return syscall (SYS_io_submit, ctx_id, nr, cbp); 190 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
92} 191}
93 192
94static int 193inline_size
194int
95ev_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)
96{ 196{
97 return syscall (SYS_io_cancel, ctx_id, cbp, result); 197 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
98} 198}
99 199
100static int 200inline_size
201int
101ev_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)
102{ 203{
103 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);
104} 205}
105
106typedef void (*ev_io_cb) (long nr, struct io_event *events);
107 206
108/*****************************************************************************/ 207/*****************************************************************************/
109/* actual backed implementation */ 208/* actual backed implementation */
110 209
111/* two iocbs for every fd, one for read, one for write */ 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
246/* we use out own wrapper structure in case we ever want to do something "clever" */
112typedef struct aniocb 247typedef struct aniocb
113{ 248{
114 struct iocb io; 249 struct iocb io;
115 /*int inuse;*/ 250 /*int inuse;*/
116} *ANIOCBP; 251} *ANIOCBP;
117 252
118inline_size 253inline_size
119void 254void
120linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 255linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
121{ 256{
122 /* TODO: quite the overhead to allocate every iocb separately */
123 while (count--) 257 while (count--)
124 { 258 {
259 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
125 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 260 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
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 */
126 memset (*base, 0, sizeof (**base)); 265 memset (iocb, 0, sizeof (*iocb));
127 /* would be nice to initialize fd/data as well */ 266
128 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 267 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
129 ++base; 268 iocb->io.aio_data = offset;
130 } 269 iocb->io.aio_fildes = offset;
131}
132 270
271 base [offset++] = iocb;
272 }
273}
274
275ecb_cold
133static void 276static void
134linuxaio_free_iocbp (EV_P) 277linuxaio_free_iocbp (EV_P)
135{ 278{
136 while (linuxaio_iocbpmax--) 279 while (linuxaio_iocbpmax--)
137 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); 280 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
138 281
139 linuxaio_iocbpmax = 0; 282 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
140} 283}
141 284
142static void 285static void
143linuxaio_modify (EV_P_ int fd, int oev, int nev) 286linuxaio_modify (EV_P_ int fd, int oev, int nev)
144{ 287{
145 /* TODO: full zero initialize required? */
146 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);
147 struct aniocb *iocb = linuxaio_iocbps [fd]; 289 ANIOCBP iocb = linuxaio_iocbps [fd];
290
291 if (iocb->io.aio_reqprio < 0)
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 */
295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0;
297 iocb->io.aio_reqprio = 0;
298 }
148 299
149 if (iocb->io.aio_buf) 300 if (iocb->io.aio_buf)
301 {
150 ev_io_cancel (linuxaio_ctx, &iocb->io, (void *)0); 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 }
151 306
152 if (nev) 307 if (nev)
153 { 308 {
154 iocb->io.aio_data = fd;
155 iocb->io.aio_fildes = fd;
156 iocb->io.aio_buf = 309 iocb->io.aio_buf =
157 (nev & EV_READ ? POLLIN : 0) 310 (nev & EV_READ ? POLLIN : 0)
158 | (nev & EV_WRITE ? POLLOUT : 0); 311 | (nev & EV_WRITE ? POLLOUT : 0);
159 312
160 /* queue iocb up for io_submit */ 313 /* queue iocb up for io_submit */
161 /* 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 */
… …
164 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 317 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
165 } 318 }
166} 319}
167 320
168static void 321static void
322linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
323{
324 epoll_poll (EV_A_ 0);
325}
326
327static void
328linuxaio_fd_rearm (EV_P_ int fd)
329{
330 anfds [fd].events = 0;
331 linuxaio_iocbps [fd]->io.aio_buf = 0;
332 fd_change (EV_A_ fd, EV_ANFD_REIFY);
333}
334
335static void
169linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 336linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
170{ 337{
171 while (nr) 338 while (nr)
172 { 339 {
173 int fd = ev->data; 340 int fd = ev->data;
174 int res = ev->res; 341 int res = ev->res;
175 342
176 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdxmax)); 343 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
177
178 /* linux aio is oneshot: rearm fd */
179 linuxaio_iocbps [fd]->io.aio_buf = 0;
180 anfds [fd].events = 0;
181 fd_change (EV_A_ fd, 0);
182 344
183 /* feed events, we do not expect or handle POLLNVAL */ 345 /* feed events, we do not expect or handle POLLNVAL */
184 if (ecb_expect_false (res & POLLNVAL))
185 fd_kill (EV_A_ fd);
186 else
187 fd_event ( 346 fd_event (
188 EV_A_ 347 EV_A_
189 fd, 348 fd,
190 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 349 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
191 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 350 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
192 ); 351 );
352
353 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
354 linuxaio_fd_rearm (EV_A_ fd);
193 355
194 --nr; 356 --nr;
195 ++ev; 357 ++ev;
196 } 358 }
197} 359}
198 360
199/* get any events from ringbuffer, return true if any were handled */ 361/* get any events from ring buffer, return true if any were handled */
200static int 362static int
201linuxaio_get_events_from_ring (EV_P) 363linuxaio_get_events_from_ring (EV_P)
202{ 364{
203 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 365 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
204 366
367 /* the kernel reads and writes both of these variables, */
368 /* as a C extension, we assume that volatile use here */
369 /* both makes reads atomic and once-only */
370 unsigned head = *(volatile unsigned *)&ring->head;
371 unsigned tail = *(volatile unsigned *)&ring->tail;
372
373 if (head == tail)
374 return 0;
375
376 /* bail out if the ring buffer doesn't match the expected layout */
377 if (expect_false (ring->magic != AIO_RING_MAGIC)
378 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
379 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
380 return 0;
381
382 /* make sure the events up to tail are visible */
205 ECB_MEMORY_FENCE_ACQUIRE; 383 ECB_MEMORY_FENCE_ACQUIRE;
206
207 unsigned head = ring->head;
208 unsigned tail = *(volatile unsigned *)&ring->tail;
209
210 if (ring->magic != AIO_RING_MAGIC
211 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
212 || ring->header_length != sizeof (struct aio_ring) /* TODO: or use it to find io_event[0]? */
213 || head == tail)
214 return 0;
215 384
216 /* parse all available events, but only once, to avoid starvation */ 385 /* parse all available events, but only once, to avoid starvation */
217 if (tail > head) /* normal case around */ 386 if (tail > head) /* normal case around */
218 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 387 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
219 else
220 {
221 /* wrapped around */ 388 else /* wrapped around */
389 {
222 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 390 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
223 linuxaio_parse_events (EV_A_ ring->io_events, tail); 391 linuxaio_parse_events (EV_A_ ring->io_events, tail);
224 } 392 }
225 393
226 ring->head = tail; 394 ECB_MEMORY_FENCE_RELEASE;
395 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
396 *(volatile unsigned *)&ring->head = tail;
227 397
228 return 1; 398 return 1;
229} 399}
230 400
231/* read at least one event from kernel, or timeout */ 401/* read at least one event from kernel, or timeout */
232inline_size 402inline_size
233void 403void
234linuxaio_get_events (EV_P_ ev_tstamp timeout) 404linuxaio_get_events (EV_P_ ev_tstamp timeout)
235{ 405{
236 struct timespec ts; 406 struct timespec ts;
237 struct io_event ioev; 407 struct io_event ioev[1];
238 int res; 408 int res;
239 409
240 if (linuxaio_get_events_from_ring (EV_A)) 410 if (linuxaio_get_events_from_ring (EV_A))
241 return; 411 return;
242 412
243 /* no events, so wait for at least one, then poll ring buffer again */ 413 /* no events, so wait for at least one, then poll ring buffer again */
244 /* this degraded to one event per loop iteration */ 414 /* this degrades to one event per loop iteration */
245 /* if the ring buffer changes layout, but so be it */ 415 /* if the ring buffer changes layout, but so be it */
416
417 EV_RELEASE_CB;
246 418
247 ts.tv_sec = (long)timeout; 419 ts.tv_sec = (long)timeout;
248 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 420 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
249 421
250 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 422 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
423
424 EV_ACQUIRE_CB;
251 425
252 if (res < 0) 426 if (res < 0)
427 if (errno == EINTR)
428 /* ignored */;
429 else
253 ev_syserr ("(libev) io_getevents"); 430 ev_syserr ("(libev) linuxaio io_getevents");
254 else if (res) 431 else if (res)
255 { 432 {
256 /* at least one event received, handle it and any remaining ones in the ring buffer */ 433 /* at least one event available, handle it and any remaining ones in the ring buffer */
257 linuxaio_parse_events (EV_A_ &ioev, 1); 434 linuxaio_parse_events (EV_A_ ioev, res);
258 linuxaio_get_events_from_ring (EV_A); 435 linuxaio_get_events_from_ring (EV_A);
259 } 436 }
437}
438
439inline_size
440int
441linuxaio_io_setup (EV_P)
442{
443 linuxaio_ctx = 0;
444 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
260} 445}
261 446
262static void 447static void
263linuxaio_poll (EV_P_ ev_tstamp timeout) 448linuxaio_poll (EV_P_ ev_tstamp timeout)
264{ 449{
… …
266 451
267 /* first phase: submit new iocbs */ 452 /* first phase: submit new iocbs */
268 453
269 /* io_submit might return less than the requested number of iocbs */ 454 /* io_submit might return less than the requested number of iocbs */
270 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 455 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
271 /* which allows us to pinpoint the errornous iocb */ 456 /* which allows us to pinpoint the erroneous iocb */
272 for (submitted = 0; submitted < linuxaio_submitcnt; ) 457 for (submitted = 0; submitted < linuxaio_submitcnt; )
273 { 458 {
274 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 459 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
275 460
276 if (res < 0) 461 if (expect_false (res < 0))
277 if (errno == EAGAIN) 462 if (errno == EINVAL)
278 { 463 {
279 /* This happens when the ring buffer is full, at least. I assume this means 464 /* This happens for unsupported fds, officially, but in my testing,
280 * that the event was queued synchronously during io_submit, and thus 465 * also randomly happens for supported fds. We fall back to good old
281 * the buffer overflowd. 466 * poll() here, under the assumption that this is a very rare case.
282 * In this case, we just try next loop iteration. 467 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
468 * discussion about such a case (ttys) where polling for POLLIN
469 * fails but POLLIN|POLLOUT works.
283 */ 470 */
284 memcpy (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 471 struct iocb *iocb = linuxaio_submits [submitted];
472 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
473 iocb->aio_reqprio = -1; /* mark iocb as epoll */
474
475 res = 1; /* skip this iocb - another iocb, another chance */
476 }
477 else if (errno == EAGAIN)
478 {
479 /* This happens when the ring buffer is full, or some other shit we
480 * don't know and isn't documented. Most likely because we have too
481 * many requests and linux aio can't be assed to handle them.
482 * In this case, we try to allocate a larger ring buffer, freeing
483 * ours first. This might fail, in which case we have to fall back to 100%
484 * epoll.
485 * God, how I hate linux not getting its act together. Ever.
486 */
487 evsys_io_destroy (linuxaio_ctx);
285 linuxaio_submitcnt -= submitted; 488 linuxaio_submitcnt = 0;
489
490 /* rearm all fds with active iocbs */
491 {
492 int fd;
493 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
494 if (linuxaio_iocbps [fd]->io.aio_buf)
495 linuxaio_fd_rearm (EV_A_ fd);
496 }
497
498 ++linuxaio_iteration;
499 if (linuxaio_io_setup (EV_A) < 0)
500 {
501 /* to bad, we can't get a new aio context, go 100% epoll */
502 linuxaio_free_iocbp (EV_A);
503 ev_io_stop (EV_A_ &linuxaio_epoll_w);
504 ev_ref (EV_A);
505 linuxaio_ctx = 0;
506 backend_modify = epoll_modify;
507 backend_poll = epoll_poll;
508 }
509
286 timeout = 0; 510 timeout = 0;
511 /* it's easiest to handle this mess in another iteration */
287 break; 512 return;
513 }
514 else if (errno == EBADF)
515 {
516 assert (("libev: event loop rejected bad fd", errno != EBADF));
517 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
518
519 res = 1; /* skip this iocb */
288 } 520 }
289 else 521 else
290 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
291 /* TODO: should we always just try next time? */
292 ev_syserr ("(libev) io_submit"); 522 ev_syserr ("(libev) linuxaio io_submit");
293 523
294 submitted += res; 524 submitted += res;
295 } 525 }
296 526
297 linuxaio_submitcnt = 0; 527 linuxaio_submitcnt = 0;
… …
304inline_size 534inline_size
305int 535int
306linuxaio_init (EV_P_ int flags) 536linuxaio_init (EV_P_ int flags)
307{ 537{
308 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 538 /* would be great to have a nice test for IOCB_CMD_POLL instead */
309 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 539 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
540 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
541 if (ev_linux_version () < 0x041300)
310 return 0; 542 return 0;
311 543
312 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 544 if (!epoll_init (EV_A_ 0))
313 return 0; 545 return 0;
546
547 linuxaio_iteration = 0;
548
549 if (linuxaio_io_setup (EV_A) < 0)
550 {
551 epoll_destroy (EV_A);
552 return 0;
553 }
554
555 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
556 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
557 ev_io_start (EV_A_ &linuxaio_epoll_w);
558 ev_unref (EV_A); /* watcher should not keep loop alive */
314 559
315 backend_modify = linuxaio_modify; 560 backend_modify = linuxaio_modify;
316 backend_poll = linuxaio_poll; 561 backend_poll = linuxaio_poll;
317 562
318 linuxaio_iocbpmax = 0; 563 linuxaio_iocbpmax = 0;
… …
327 572
328inline_size 573inline_size
329void 574void
330linuxaio_destroy (EV_P) 575linuxaio_destroy (EV_P)
331{ 576{
577 epoll_destroy (EV_A);
332 linuxaio_free_iocbp (EV_A); 578 linuxaio_free_iocbp (EV_A);
333 ev_io_destroy (linuxaio_ctx); 579 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
334} 580}
335 581
336inline_size 582inline_size
337void 583void
338linuxaio_fork (EV_P) 584linuxaio_fork (EV_P)
339{ 585{
340 abort ();//D 586 /* this frees all iocbs, which is very heavy-handed */
341} 587 linuxaio_destroy (EV_A);
588 linuxaio_submitcnt = 0; /* all pointers were invalidated */
342 589
590 linuxaio_iteration = 0; /* we start over in the child */
591
592 while (linuxaio_io_setup (EV_A) < 0)
593 ev_syserr ("(libev) linuxaio io_setup");
594
595 /* forking epoll should also effectively unregister all fds from the backend */
596 epoll_fork (EV_A);
597
598 ev_io_stop (EV_A_ &linuxaio_epoll_w);
599 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
600 ev_io_start (EV_A_ &linuxaio_epoll_w);
601
602 /* epoll_fork already did this. hopefully */
603 /*fd_rearm_all (EV_A);*/
604}
605

Diff Legend

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