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 1.12 by root, Sun Jun 23 23:02:09 2019 UTC vs.
Revision 1.31 by root, Tue Jun 25 18:00:42 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 */ 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
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
83inline_size 171inline_size
84int 172int
85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 173evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86{ 174{
87 return syscall (SYS_io_setup, nr_events, ctx_idp); 175 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
88} 176}
89 177
90inline_size 178inline_size
91int 179int
92ev_io_destroy (aio_context_t ctx_id) 180evsys_io_destroy (aio_context_t ctx_id)
93{ 181{
94 return syscall (SYS_io_destroy, ctx_id); 182 return ev_syscall1 (SYS_io_destroy, ctx_id);
95} 183}
96 184
97inline_size 185inline_size
98int 186int
99ev_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[])
100{ 188{
101 return syscall (SYS_io_submit, ctx_id, nr, cbp); 189 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
102} 190}
103 191
104inline_size 192inline_size
105int 193int
106ev_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)
107{ 195{
108 return syscall (SYS_io_cancel, ctx_id, cbp, result); 196 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
109} 197}
110 198
111inline_size 199inline_size
112int 200int
113ev_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)
114{ 202{
115 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);
116} 204}
117 205
118/*****************************************************************************/ 206/*****************************************************************************/
119/* actual backed implementation */ 207/* actual backed implementation */
120 208
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
121/* 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" */
122typedef struct aniocb 246typedef struct aniocb
123{ 247{
124 struct iocb io; 248 struct iocb io;
125 /*int inuse;*/ 249 /*int inuse;*/
126} *ANIOCBP; 250} *ANIOCBP;
127 251
128inline_size 252inline_size
129void 253void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 254linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 255{
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--) 256 while (count--)
134 { 257 {
258 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
135 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 259 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 /* 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 */
137 memset (*base, 0, sizeof (**base)); 264 memset (iocb, 0, sizeof (*iocb));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 265
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 266 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base; 267 iocb->io.aio_data = offset;
268 iocb->io.aio_fildes = offset;
269
270 base [offset++] = iocb;
141 } 271 }
142} 272}
143 273
144ecb_cold 274ecb_cold
145static void 275static void
… …
153 283
154static void 284static void
155linuxaio_modify (EV_P_ int fd, int oev, int nev) 285linuxaio_modify (EV_P_ int fd, int oev, int nev)
156{ 286{
157 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);
158 struct aniocb *iocb = linuxaio_iocbps [fd]; 288 ANIOCBP iocb = linuxaio_iocbps [fd];
159 289
160#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0) 290 if (iocb->io.aio_reqprio < 0)
162 { 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 */
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 294 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
295 anfds [fd].emask = 0;
164 iocb->io.aio_reqprio = 0; 296 iocb->io.aio_reqprio = 0;
165 } 297 }
166#endif
167 298
168 if (iocb->io.aio_buf) 299 if (iocb->io.aio_buf)
169 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);
170 302
171 if (nev) 303 if (nev)
172 { 304 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf = 305 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0) 306 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0); 307 | (nev & EV_WRITE ? POLLOUT : 0);
178 308
179 /* queue iocb up for io_submit */ 309 /* queue iocb up for io_submit */
180 /* 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 */
… …
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 313 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 314 }
185} 315}
186 316
187static 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
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 332linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 333{
190 while (nr) 334 while (nr)
191 { 335 {
192 int fd = ev->data; 336 int fd = ev->data;
193 int res = ev->res; 337 int res = ev->res;
194 338
195 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));
196 340
197 /* linux aio is oneshot: rearm fd */
198 linuxaio_iocbps [fd]->io.aio_buf = 0;
199 anfds [fd].events = 0;
200 fd_change (EV_A_ fd, 0);
201
202 /* feed events, we do not expect or handle POLLNVAL */ 341 /* feed events, we do not expect or handle POLLNVAL */
203 if (ecb_expect_false (res & POLLNVAL))
204 fd_kill (EV_A_ fd);
205 else
206 fd_event ( 342 fd_event (
207 EV_A_ 343 EV_A_
208 fd, 344 fd,
209 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 345 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
210 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 346 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
211 ); 347 );
348
349 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
350 linuxaio_fd_rearm (EV_A_ fd);
212 351
213 --nr; 352 --nr;
214 ++ev; 353 ++ev;
215 } 354 }
216} 355}
217 356
218/* get any events from ringbuffer, return true if any were handled */ 357/* get any events from ring buffer, return true if any were handled */
219static int 358static int
220linuxaio_get_events_from_ring (EV_P) 359linuxaio_get_events_from_ring (EV_P)
221{ 360{
222 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 361 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
223 362
224 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;
225 unsigned tail = *(volatile unsigned *)&ring->tail; 367 unsigned tail = *(volatile unsigned *)&ring->tail;
226 368
227 if (head == tail) 369 if (head == tail)
228 return 0; 370 return 0;
229 371
230 /* 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 */
231 if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) 373 if (expect_false (ring->magic != AIO_RING_MAGIC)
232 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES 374 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
233 || 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]? */
234 return 0; 376 return 0;
235 377
236 /* make sure the events up to tail are visible */ 378 /* make sure the events up to tail are visible */
237 ECB_MEMORY_FENCE_ACQUIRE; 379 ECB_MEMORY_FENCE_ACQUIRE;
238 380
… …
243 { 385 {
244 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);
245 linuxaio_parse_events (EV_A_ ring->io_events, tail); 387 linuxaio_parse_events (EV_A_ ring->io_events, tail);
246 } 388 }
247 389
390 ECB_MEMORY_FENCE_RELEASE;
391 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
248 *(volatile unsigned *)&ring->head = tail; 392 *(volatile unsigned *)&ring->head = tail;
249 /* make sure kernel can see our new head value - probably not required */
250 ECB_MEMORY_FENCE_RELEASE;
251 393
252 return 1; 394 return 1;
253} 395}
254 396
255/* read at least one event from kernel, or timeout */ 397/* read at least one event from kernel, or timeout */
256inline_size 398inline_size
257void 399void
258linuxaio_get_events (EV_P_ ev_tstamp timeout) 400linuxaio_get_events (EV_P_ ev_tstamp timeout)
259{ 401{
260 struct timespec ts; 402 struct timespec ts;
261 struct io_event ioev; 403 struct io_event ioev[1];
262 int res; 404 int res;
263 405
264 if (linuxaio_get_events_from_ring (EV_A)) 406 if (linuxaio_get_events_from_ring (EV_A))
265 return; 407 return;
266 408
267 /* 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 */
268 /* this degrades to one event per loop iteration */ 410 /* this degrades to one event per loop iteration */
269 /* if the ring buffer changes layout, but so be it */ 411 /* if the ring buffer changes layout, but so be it */
270 412
413 EV_RELEASE_CB;
414
271 ts.tv_sec = (long)timeout; 415 ts.tv_sec = (long)timeout;
272 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 416 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
273 417
274 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;
275 421
276 if (res < 0) 422 if (res < 0)
277 if (errno == EINTR) 423 if (errno == EINTR)
278 /* ignored */; 424 /* ignored */;
279 else 425 else
280 ev_syserr ("(libev) linuxaio io_getevents"); 426 ev_syserr ("(libev) linuxaio io_getevents");
281 else if (res) 427 else if (res)
282 { 428 {
283 /* at least one event received, handle it and any remaining ones in the ring buffer */ 429 /* at least one event available, handle it and any remaining ones in the ring buffer */
284 linuxaio_parse_events (EV_A_ &ioev, 1); 430 linuxaio_parse_events (EV_A_ ioev, res);
285 linuxaio_get_events_from_ring (EV_A); 431 linuxaio_get_events_from_ring (EV_A);
286 } 432 }
287} 433}
288 434
289#if EPOLL_FALLBACK 435inline_size
290static void 436int
291linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) 437linuxaio_io_setup (EV_P)
292{ 438{
293 struct epoll_event eev; 439 linuxaio_ctx = 0;
294 440 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
295 eev.events = EPOLLONESHOT;
296 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
297 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
298 eev.data.fd = iocb->aio_fildes;
299
300 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
301 ev_syserr ("(libeio) linuxaio epoll_ctl");
302} 441}
303#endif
304 442
305static void 443static void
306linuxaio_poll (EV_P_ ev_tstamp timeout) 444linuxaio_poll (EV_P_ ev_tstamp timeout)
307{ 445{
308 int submitted; 446 int submitted;
309 447
310 /* first phase: submit new iocbs */ 448 /* first phase: submit new iocbs */
311 449
312 /* io_submit might return less than the requested number of iocbs */ 450 /* io_submit might return less than the requested number of iocbs */
313 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 451 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
314 /* which allows us to pinpoint the errornous iocb */ 452 /* which allows us to pinpoint the erroneous iocb */
315 for (submitted = 0; submitted < linuxaio_submitcnt; ) 453 for (submitted = 0; submitted < linuxaio_submitcnt; )
316 { 454 {
317 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 455 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
318 456
319 if (ecb_expect_false (res < 0)) 457 if (expect_false (res < 0))
320 if (errno == EAGAIN) 458 if (errno == EINVAL)
321 { 459 {
322 /* This happens when the ring buffer is full, at least. I assume this means
323 * that the event was queued synchronously during io_submit, and thus
324 * the buffer overflowed.
325 * In this case, we just try in next loop iteration.
326 * This should not result in a few fds taking priority, as the interface
327 * is one-shot, and we submit iocb's in a round-robin fashion.
328 */
329 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
330 linuxaio_submitcnt -= submitted;
331 timeout = 0;
332 break;
333 }
334#if EPOLL_FALLBACK
335 else if (errno == EINVAL)
336 {
337 /* This hapΓΌpens for unsupported fds, officially, but in my testing, 460 /* This happens for unsupported fds, officially, but in my testing,
338 * also randomly happens for supported fds. We fall back to good old 461 * also randomly happens for supported fds. We fall back to good old
339 * poll() here, under the assumption that this is a very rare case. 462 * poll() here, under the assumption that this is a very rare case.
463 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
464 * discussion about such a case (ttys) where polling for POLLIN
465 * fails but POLLIN|POLLOUT works.
340 */ 466 */
341 struct iocb *iocb = linuxaio_submits [submitted]; 467 struct iocb *iocb = linuxaio_submits [submitted];
468 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
469 iocb->aio_reqprio = -1; /* mark iocb as epoll */
470
471 res = 1; /* skip this iocb - another iocb, another chance */
472 }
473 else if (errno == EAGAIN)
474 {
475 /* This happens when the ring buffer is full, or some other shit we
476 * don't know and isn't documented. Most likely because we have too
477 * many requests and linux aio can't be assed to handle them.
478 * In this case, we try to allocate a larger ring buffer, freeing
479 * ours first. This might fail, in which case we have to fall back to 100%
480 * epoll.
481 * God, how I hate linux not getting its act together. Ever.
482 */
483 evsys_io_destroy (linuxaio_ctx);
484 linuxaio_submitcnt = 0;
485
486 /* rearm all fds with active iocbs */
487 {
488 int fd;
489 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
490 if (linuxaio_iocbps [fd]->io.aio_buf)
491 linuxaio_fd_rearm (EV_A_ fd);
492 }
493
494 ++linuxaio_iteration;
495 if (linuxaio_io_setup (EV_A) < 0)
496 {
497 /* to bad, we can't get a new aio context, go 100% epoll */
498 linuxaio_free_iocbp (EV_A);
499 ev_io_stop (EV_A_ &linuxaio_epoll_w);
500 ev_ref (EV_A);
501 linuxaio_ctx = 0;
502 backend_modify = epoll_modify;
503 backend_poll = epoll_poll;
504 }
505
506 timeout = 0;
507 /* it's easiest to handle this mess in another iteration */
508 return;
509 }
510 else if (errno == EBADF)
511 {
512 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
513
342 res = 1; /* skip this iocb */ 514 res = 1; /* skip this iocb */
343
344 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
345 iocb->aio_reqprio = -1; /* mark iocb as epoll */
346 } 515 }
347#endif
348 else 516 else
349 ev_syserr ("(libev) linuxaio io_submit"); 517 ev_syserr ("(libev) linuxaio io_submit");
350 518
351 submitted += res; 519 submitted += res;
352 } 520 }
… …
356 /* second phase: fetch and parse events */ 524 /* second phase: fetch and parse events */
357 525
358 linuxaio_get_events (EV_A_ timeout); 526 linuxaio_get_events (EV_A_ timeout);
359} 527}
360 528
361#if EPOLL_FALLBACK
362
363static void
364linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
365{
366 struct epoll_event events[16];
367
368 for (;;)
369 {
370 int idx;
371 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
372
373 if (ecb_expect_false (res < 0))
374 ev_syserr ("(libev) linuxaio epoll_wait");
375 else if (!res)
376 break;
377
378 for (idx = res; idx--; )
379 {
380 int fd = events [idx].data.fd;
381 uint32_t ev = events [idx].events;
382
383 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
384
385 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
386
387 fd_event (EV_A_ fd,
388 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
389 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
390 }
391
392 if (res < sizeof (events) / sizeof (events [0]))
393 break;
394 }
395}
396
397#endif
398
399inline_size 529inline_size
400int 530int
401linuxaio_init (EV_P_ int flags) 531linuxaio_init (EV_P_ int flags)
402{ 532{
403 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 533 /* would be great to have a nice test for IOCB_CMD_POLL instead */
404 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 534 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
405 if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ 535 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
536 if (ev_linux_version () < 0x041300)
406 return 0; 537 return 0;
407 538
408 linuxaio_ctx = 0; 539 if (!epoll_init (EV_A_ 0))
409 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
410 return 0; 540 return 0;
411 541
412#if EPOLL_FALLBACK 542 linuxaio_iteration = 0;
413 backend_fd = ev_epoll_create (); 543
414 if (backend_fd < 0) 544 if (linuxaio_io_setup (EV_A) < 0)
415 { 545 {
416 ev_io_destroy (linuxaio_ctx); 546 epoll_destroy (EV_A);
417 return 0; 547 return 0;
418 } 548 }
419 549
420 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 550 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
551 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
421 ev_io_start (EV_A_ &linuxaio_epoll_w); 552 ev_io_start (EV_A_ &linuxaio_epoll_w);
422#endif 553 ev_unref (EV_A); /* watcher should not keep loop alive */
423 554
424 backend_modify = linuxaio_modify; 555 backend_modify = linuxaio_modify;
425 backend_poll = linuxaio_poll; 556 backend_poll = linuxaio_poll;
426 557
427 linuxaio_iocbpmax = 0; 558 linuxaio_iocbpmax = 0;
… …
436 567
437inline_size 568inline_size
438void 569void
439linuxaio_destroy (EV_P) 570linuxaio_destroy (EV_P)
440{ 571{
441#if EPOLL_FALLBACK 572 epoll_destroy (EV_A);
442 close (backend_fd);
443#endif
444 linuxaio_free_iocbp (EV_A); 573 linuxaio_free_iocbp (EV_A);
445 ev_io_destroy (linuxaio_ctx); 574 evsys_io_destroy (linuxaio_ctx);
446} 575}
447 576
448inline_size 577inline_size
449void 578void
450linuxaio_fork (EV_P) 579linuxaio_fork (EV_P)
451{ 580{
452 /* this frees all iocbs, which is very heavy-handed */ 581 /* this frees all iocbs, which is very heavy-handed */
453 linuxaio_destroy (EV_A); 582 linuxaio_destroy (EV_A);
454 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 583 linuxaio_submitcnt = 0; /* all pointers were invalidated */
455 584
456 linuxaio_ctx = 0; 585 linuxaio_iteration = 0; /* we start over in the child */
457 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 586
587 while (linuxaio_io_setup (EV_A) < 0)
458 ev_syserr ("(libev) linuxaio io_setup"); 588 ev_syserr ("(libev) linuxaio io_setup");
459 589
460#if EPOLL_FALLBACK 590 epoll_fork (EV_A);
461 while ((backend_fd = ev_epoll_create ()) < 0)
462 ev_syserr ("(libev) linuxaio epoll_create");
463 591
464 ev_io_stop (EV_A_ &linuxaio_epoll_w); 592 ev_io_stop (EV_A_ &linuxaio_epoll_w);
465 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 593 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
466 ev_io_start (EV_A_ &linuxaio_epoll_w); 594 ev_io_start (EV_A_ &linuxaio_epoll_w);
467 ev_unref (EV_A); /* watcher should not keep loop alive */
468#endif
469 595
596 /* epoll_fork already did this. hopefully */
470 fd_rearm_all (EV_A); 597 /*fd_rearm_all (EV_A);*/
471} 598}
472 599

Diff Legend

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