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 1.4 by root, Fri Jun 21 03:14:25 2019 UTC vs.
Revision 1.35 by root, Wed Jun 26 00:11:24 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 4kn pages exactly. 91/*****************************************************************************/
45 * the ring buffer header is 32 bytes, every io event is 32 bytes. 92/* syscall wrapdadoop - this section has the raw api/abi definitions */
46 * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer 93
47 * 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.
48 */ 103 */
49#define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */
50
51/*****************************************************************************/
52/* syscall wrapdadoop */
53
54#include <sys/syscall.h> /* no glibc wrappers */
55
56/* aio_abi.h is not verioned in any way, so we cannot test for its existance */
57#define IOCB_CMD_POLL 5
58
59/* taken from linux/fs/aio.c */
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
77static 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
78ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
79{ 175{
80 return syscall (SYS_io_setup, nr_events, ctx_idp); 176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
81} 177}
82 178
83static int 179inline_size
180int
84ev_io_destroy (aio_context_t ctx_id) 181evsys_io_destroy (aio_context_t ctx_id)
85{ 182{
86 return syscall (SYS_io_destroy, ctx_id); 183 return ev_syscall1 (SYS_io_destroy, ctx_id);
87} 184}
88 185
89static int 186inline_size
187int
90ev_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[])
91{ 189{
92 return syscall (SYS_io_submit, ctx_id, nr, cbp); 190 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
93} 191}
94 192
95static int 193inline_size
194int
96ev_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)
97{ 196{
98 return syscall (SYS_io_cancel, ctx_id, cbp, result); 197 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
99} 198}
100 199
101static int 200inline_size
201int
102ev_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)
103{ 203{
104 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);
105} 205}
106
107typedef void (*ev_io_cb) (long nr, struct io_event *events);
108 206
109/*****************************************************************************/ 207/*****************************************************************************/
110/* actual backed implementation */ 208/* actual backed implementation */
111 209
112/* 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" */
113typedef struct aniocb 247typedef struct aniocb
114{ 248{
115 struct iocb io; 249 struct iocb io;
116 /*int inuse;*/ 250 /*int inuse;*/
117} *ANIOCBP; 251} *ANIOCBP;
118 252
119inline_size 253inline_size
120void 254void
121linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 255linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
122{ 256{
123 /* TODO: quite the overhead to allocate every iocb separately */
124 while (count--) 257 while (count--)
125 { 258 {
259 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
126 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 260 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
127 /* 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 */
128 memset (*base, 0, sizeof (**base)); 265 memset (iocb, 0, sizeof (*iocb));
129 /* would be nice to initialize fd/data as well */ 266
130 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 267 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
131 ++base; 268 iocb->io.aio_data = offset;
132 } 269 iocb->io.aio_fildes = offset;
133}
134 270
271 base [offset++] = iocb;
272 }
273}
274
275ecb_cold
135static void 276static void
136linuxaio_free_iocbp (EV_P) 277linuxaio_free_iocbp (EV_P)
137{ 278{
138 while (linuxaio_iocbpmax--) 279 while (linuxaio_iocbpmax--)
139 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); 280 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
140 281
141 /* next resize will completely reallocate the array */ 282 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
142 linuxaio_iocbpmax = 0;
143 linuxaio_submitcnt = 0; /* all pointers invalidated */
144} 283}
145 284
146static void 285static void
147linuxaio_modify (EV_P_ int fd, int oev, int nev) 286linuxaio_modify (EV_P_ int fd, int oev, int nev)
148{ 287{
149 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);
150 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 }
151 299
152 if (iocb->io.aio_buf) 300 if (iocb->io.aio_buf)
153 ev_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 }
154 306
155 if (nev) 307 if (nev)
156 { 308 {
157 iocb->io.aio_data = fd;
158 iocb->io.aio_fildes = fd;
159 iocb->io.aio_buf = 309 iocb->io.aio_buf =
160 (nev & EV_READ ? POLLIN : 0) 310 (nev & EV_READ ? POLLIN : 0)
161 | (nev & EV_WRITE ? POLLOUT : 0); 311 | (nev & EV_WRITE ? POLLOUT : 0);
162 312
163 /* queue iocb up for io_submit */ 313 /* queue iocb up for io_submit */
164 /* 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 */
… …
167 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 317 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
168 } 318 }
169} 319}
170 320
171static void 321static void
322linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
323{
324 epoll_poll (EV_A_ 0);
325}
326
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}
335
336static void
172linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 337linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
173{ 338{
174 while (nr) 339 while (nr)
175 { 340 {
176 int fd = ev->data; 341 int fd = ev->data;
177 int res = ev->res; 342 int res = ev->res;
178 343
179 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));
180 345
181 /* linux aio is oneshot: rearm fd */
182 linuxaio_iocbps [fd]->io.aio_buf = 0;
183 anfds [fd].events = 0;
184 fd_change (EV_A_ fd, 0);
185
186 /* feed events, we do not expect or handle POLLNVAL */ 346 /* feed events, we do not expect or handle POLLNVAL */
187 if (ecb_expect_false (res & POLLNVAL))
188 fd_kill (EV_A_ fd);
189 else
190 fd_event ( 347 fd_event (
191 EV_A_ 348 EV_A_
192 fd, 349 fd,
193 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 350 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
194 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 351 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
195 ); 352 );
353
354 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
355 linuxaio_fd_rearm (EV_A_ fd);
196 356
197 --nr; 357 --nr;
198 ++ev; 358 ++ev;
199 } 359 }
200} 360}
201 361
202/* get any events from ringbuffer, return true if any were handled */ 362/* get any events from ring buffer, return true if any were handled */
203static int 363static int
204linuxaio_get_events_from_ring (EV_P) 364linuxaio_get_events_from_ring (EV_P)
205{ 365{
206 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 366 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
207 367
368 /* the kernel reads and writes both of these variables, */
369 /* as a C extension, we assume that volatile use here */
370 /* both makes reads atomic and once-only */
371 unsigned head = *(volatile unsigned *)&ring->head;
372 unsigned tail = *(volatile unsigned *)&ring->tail;
373
374 if (head == tail)
375 return 0;
376
377 /* bail out if the ring buffer doesn't match the expected layout */
378 if (expect_false (ring->magic != AIO_RING_MAGIC)
379 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
380 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
381 return 0;
382
383 /* make sure the events up to tail are visible */
208 ECB_MEMORY_FENCE_ACQUIRE; 384 ECB_MEMORY_FENCE_ACQUIRE;
209
210 unsigned head = ring->head;
211 unsigned tail = *(volatile unsigned *)&ring->tail;
212
213 if (ring->magic != AIO_RING_MAGIC
214 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
215 || ring->header_length != sizeof (struct aio_ring) /* TODO: or use it to find io_event[0]? */
216 || head == tail)
217 return 0;
218 385
219 /* parse all available events, but only once, to avoid starvation */ 386 /* parse all available events, but only once, to avoid starvation */
220 if (tail > head) /* normal case around */ 387 if (tail > head) /* normal case around */
221 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 388 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
222 else
223 {
224 /* wrapped around */ 389 else /* wrapped around */
390 {
225 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);
226 linuxaio_parse_events (EV_A_ ring->io_events, tail); 392 linuxaio_parse_events (EV_A_ ring->io_events, tail);
227 } 393 }
228 394
229 ring->head = tail; 395 ECB_MEMORY_FENCE_RELEASE;
396 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
397 *(volatile unsigned *)&ring->head = tail;
230 398
231 return 1; 399 return 1;
232} 400}
233 401
234/* read at least one event from kernel, or timeout */ 402/* read at least one event from kernel, or timeout */
235inline_size 403inline_size
236void 404void
237linuxaio_get_events (EV_P_ ev_tstamp timeout) 405linuxaio_get_events (EV_P_ ev_tstamp timeout)
238{ 406{
239 struct timespec ts; 407 struct timespec ts;
240 struct io_event ioev; 408 struct io_event ioev[1];
241 int res; 409 int res;
242 410
243 if (linuxaio_get_events_from_ring (EV_A)) 411 if (linuxaio_get_events_from_ring (EV_A))
244 return; 412 return;
245 413
246 /* no events, so wait for at least one, then poll ring buffer again */ 414 /* no events, so wait for at least one, then poll ring buffer again */
247 /* this degraded to one event per loop iteration */ 415 /* this degrades to one event per loop iteration */
248 /* if the ring buffer changes layout, but so be it */ 416 /* if the ring buffer changes layout, but so be it */
417
418 EV_RELEASE_CB;
249 419
250 ts.tv_sec = (long)timeout; 420 ts.tv_sec = (long)timeout;
251 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 421 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
252 422
253 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 423 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
424
425 EV_ACQUIRE_CB;
254 426
255 if (res < 0) 427 if (res < 0)
428 if (errno == EINTR)
429 /* ignored */;
430 else
256 ev_syserr ("(libev) io_getevents"); 431 ev_syserr ("(libev) linuxaio io_getevents");
257 else if (res) 432 else if (res)
258 { 433 {
259 /* 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 */
260 linuxaio_parse_events (EV_A_ &ioev, 1); 435 linuxaio_parse_events (EV_A_ ioev, res);
261 linuxaio_get_events_from_ring (EV_A); 436 linuxaio_get_events_from_ring (EV_A);
262 } 437 }
438}
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);
263} 446}
264 447
265static void 448static void
266linuxaio_poll (EV_P_ ev_tstamp timeout) 449linuxaio_poll (EV_P_ ev_tstamp timeout)
267{ 450{
… …
269 452
270 /* first phase: submit new iocbs */ 453 /* first phase: submit new iocbs */
271 454
272 /* io_submit might return less than the requested number of iocbs */ 455 /* io_submit might return less than the requested number of iocbs */
273 /* 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, */
274 /* which allows us to pinpoint the errornous iocb */ 457 /* which allows us to pinpoint the erroneous iocb */
275 for (submitted = 0; submitted < linuxaio_submitcnt; ) 458 for (submitted = 0; submitted < linuxaio_submitcnt; )
276 { 459 {
277 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 460 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
278 461
279 if (res < 0) 462 if (expect_false (res < 0))
280 if (errno == EAGAIN) 463 if (errno == EINVAL)
281 { 464 {
282 /* This happens when the ring buffer is full, at least. I assume this means 465 /* This happens for unsupported fds, officially, but in my testing,
283 * that the event was queued synchronously during io_submit, and thus 466 * also randomly happens for supported fds. We fall back to good old
284 * the buffer overflowd. 467 * poll() here, under the assumption that this is a very rare case.
285 * In this case, we just try next loop iteration. 468 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
469 * discussion about such a case (ttys) where polling for POLLIN
470 * fails but POLLIN|POLLOUT works.
286 */ 471 */
287 memcpy (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 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);
288 linuxaio_submitcnt -= submitted; 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
289 timeout = 0; 511 timeout = 0;
512 /* it's easiest to handle this mess in another iteration */
290 break; 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
520 res = 1; /* skip this iocb */
291 } 521 }
292 else 522 else
293 /* TODO: we get EAGAIN when the ring buffer is full for some reason */
294 /* TODO: should we always just try next time? */
295 ev_syserr ("(libev) io_submit"); 523 ev_syserr ("(libev) linuxaio io_submit");
296 524
297 submitted += res; 525 submitted += res;
298 } 526 }
299 527
300 linuxaio_submitcnt = 0; 528 linuxaio_submitcnt = 0;
… …
308int 536int
309linuxaio_init (EV_P_ int flags) 537linuxaio_init (EV_P_ int flags)
310{ 538{
311 /* 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 */
312 /* 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 */
313 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 541 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
542 if (ev_linux_version () < 0x041300)
314 return 0; 543 return 0;
315 544
316 linuxaio_ctx = 0; 545 if (!epoll_init (EV_A_ 0))
317 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
318 return 0; 546 return 0;
547
548 linuxaio_iteration = 0;
549
550 if (linuxaio_io_setup (EV_A) < 0)
551 {
552 epoll_destroy (EV_A);
553 return 0;
554 }
555
556 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
557 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
558 ev_io_start (EV_A_ &linuxaio_epoll_w);
559 ev_unref (EV_A); /* watcher should not keep loop alive */
319 560
320 backend_modify = linuxaio_modify; 561 backend_modify = linuxaio_modify;
321 backend_poll = linuxaio_poll; 562 backend_poll = linuxaio_poll;
322 563
323 linuxaio_iocbpmax = 0; 564 linuxaio_iocbpmax = 0;
… …
332 573
333inline_size 574inline_size
334void 575void
335linuxaio_destroy (EV_P) 576linuxaio_destroy (EV_P)
336{ 577{
578 epoll_destroy (EV_A);
337 linuxaio_free_iocbp (EV_A); 579 linuxaio_free_iocbp (EV_A);
338 ev_io_destroy (linuxaio_ctx); 580 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
339} 581}
340 582
341inline_size 583inline_size
342void 584void
343linuxaio_fork (EV_P) 585linuxaio_fork (EV_P)
344{ 586{
345 /* TODO: verify and test */ 587 /* this frees all iocbs, which is very heavy-handed */
346 linuxaio_destroy (EV_A); 588 linuxaio_destroy (EV_A);
589 linuxaio_submitcnt = 0; /* all pointers were invalidated */
347 590
348 linuxaio_ctx = 0; 591 linuxaio_iteration = 0; /* we start over in the child */
349 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 592
593 while (linuxaio_io_setup (EV_A) < 0)
350 ev_syserr ("(libev) io_setup"); 594 ev_syserr ("(libev) linuxaio io_setup");
351 595
596 /* forking epoll should also effectively unregister all fds from the backend */
597 epoll_fork (EV_A);
598
599 ev_io_stop (EV_A_ &linuxaio_epoll_w);
600 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
601 ev_io_start (EV_A_ &linuxaio_epoll_w);
602
603 /* epoll_fork already did this. hopefully */
352 fd_rearm_all (EV_A); 604 /*fd_rearm_all (EV_A);*/
353} 605}
354 606

Diff Legend

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