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 1.23 by root, Mon Jun 24 21:34:47 2019 UTC vs.
Revision 1.41 by root, Mon Jul 1 20:41:46 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#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 EV_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 */
71 unsigned nr; /* number of io_events */ 109 unsigned nr; /* number of io_events */
72 unsigned head; /* Written to by userland or by kernel. */ 110 unsigned head; /* Written to by userland or by kernel. */
… …
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;
… …
129void 254void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) 255linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 256{
132 while (count--) 257 while (count--)
133 { 258 {
134 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ 259 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
135 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); 260 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 261
137 /* full zero initialise is probably not required at the moment, but 262 /* full zero initialise is probably not required at the moment, but
138 * this is not well documented, so we better do it. 263 * this is not well documented, so we better do it.
139 */ 264 */
… …
161linuxaio_modify (EV_P_ int fd, int oev, int nev) 286linuxaio_modify (EV_P_ int fd, int oev, int nev)
162{ 287{
163 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);
164 ANIOCBP iocb = linuxaio_iocbps [fd]; 289 ANIOCBP iocb = linuxaio_iocbps [fd];
165 290
166#if EPOLL_FALLBACK
167 if (iocb->io.aio_reqprio < 0) 291 if (expect_false (iocb->io.aio_reqprio < 0))
168 { 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 */
169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0;
170 iocb->io.aio_reqprio = 0; 297 iocb->io.aio_reqprio = 0;
171 } 298 }
172#endif
173 299
300 if (expect_false (iocb->io.aio_buf))
301 {
302 /* iocb active, so cancel it first before resubmit */
303 for (;;)
304 {
305 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
306 if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
307 break;
308
309 if (expect_true (errno == EINPROGRESS))
310 break;
311
312 /* the EINPROGRESS test is for nicer error message. clumsy. */
313 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR));
314 }
315 }
316
174 if (iocb->io.aio_buf) 317 iocb->io.aio_buf =
175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 318 (nev & EV_READ ? POLLIN : 0)
319 | (nev & EV_WRITE ? POLLOUT : 0);
176 320
177 if (nev) 321 if (nev)
178 { 322 {
179 iocb->io.aio_buf =
180 (nev & EV_READ ? POLLIN : 0)
181 | (nev & EV_WRITE ? POLLOUT : 0);
182
183 /* queue iocb up for io_submit */ 323 /* queue iocb up for io_submit */
184 /* this assumes we only ever get one call per fd per loop iteration */ 324 /* this assumes we only ever get one call per fd per loop iteration */
185 ++linuxaio_submitcnt; 325 ++linuxaio_submitcnt;
186 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 326 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
187 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 327 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
188 } 328 }
189} 329}
190 330
191#if EPOLL_FALLBACK
192
193static void
194linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
195{
196 struct epoll_event eev;
197
198 eev.events = EPOLLONESHOT;
199 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
200 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
201 eev.data.fd = iocb->aio_fildes;
202
203 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
204 ev_syserr ("(libeio) linuxaio epoll_ctl");
205}
206
207static void 331static void
208linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 332linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
209{ 333{
210 struct epoll_event events[16]; 334 epoll_poll (EV_A_ 0);
211
212 for (;;)
213 {
214 int idx;
215 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
216
217 if (expect_false (res < 0))
218 ev_syserr ("(libev) linuxaio epoll_wait");
219 else if (!res)
220 break;
221
222 for (idx = res; idx--; )
223 {
224 int fd = events [idx].data.fd;
225 uint32_t ev = events [idx].events;
226
227 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
228
229 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
230
231 fd_event (EV_A_ fd,
232 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
233 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
234 }
235
236 if (res < sizeof (events) / sizeof (events [0]))
237 break;
238 }
239} 335}
240 336
241#endif 337inline_speed
338void
339linuxaio_fd_rearm (EV_P_ int fd)
340{
341 anfds [fd].events = 0;
342 linuxaio_iocbps [fd]->io.aio_buf = 0;
343 fd_change (EV_A_ fd, EV_ANFD_REIFY);
344}
242 345
243static void 346static void
244linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 347linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
245{ 348{
246 while (nr) 349 while (nr)
247 { 350 {
248 int fd = ev->data; 351 int fd = ev->data;
249 int res = ev->res; 352 int res = ev->res;
250 353
251 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 354 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
252
253 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
254 linuxaio_iocbps [fd]->io.aio_buf = 0;
255 anfds [fd].events = 0;
256 fd_change (EV_A_ fd, 0);
257 355
258 /* feed events, we do not expect or handle POLLNVAL */ 356 /* feed events, we do not expect or handle POLLNVAL */
259 fd_event ( 357 fd_event (
260 EV_A_ 358 EV_A_
261 fd, 359 fd,
262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 360 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 361 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
264 ); 362 );
265 363
364 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
365 linuxaio_fd_rearm (EV_A_ fd);
366
266 --nr; 367 --nr;
267 ++ev; 368 ++ev;
268 } 369 }
269} 370}
270 371
271/* get any events from ringbuffer, return true if any were handled */ 372/* get any events from ring buffer, return true if any were handled */
272static int 373static int
273linuxaio_get_events_from_ring (EV_P) 374linuxaio_get_events_from_ring (EV_P)
274{ 375{
275 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 376 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
276 377
… …
281 unsigned tail = *(volatile unsigned *)&ring->tail; 382 unsigned tail = *(volatile unsigned *)&ring->tail;
282 383
283 if (head == tail) 384 if (head == tail)
284 return 0; 385 return 0;
285 386
286 /* bail out if the ring buffer doesn't match the expected layout */
287 if (expect_false (ring->magic != AIO_RING_MAGIC)
288 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
289 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
290 return 0;
291
292 /* make sure the events up to tail are visible */ 387 /* make sure the events up to tail are visible */
293 ECB_MEMORY_FENCE_ACQUIRE; 388 ECB_MEMORY_FENCE_ACQUIRE;
294 389
295 /* parse all available events, but only once, to avoid starvation */ 390 /* parse all available events, but only once, to avoid starvation */
296 if (tail > head) /* normal case around */ 391 if (tail > head) /* normal case around */
… …
299 { 394 {
300 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); 395 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
301 linuxaio_parse_events (EV_A_ ring->io_events, tail); 396 linuxaio_parse_events (EV_A_ ring->io_events, tail);
302 } 397 }
303 398
304 /* TODO: we only need a compiler barrier here, not a read fence */
305 ECB_MEMORY_FENCE_RELEASE; 399 ECB_MEMORY_FENCE_RELEASE;
306 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 400 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
307 *(volatile unsigned *)&ring->head = tail; 401 *(volatile unsigned *)&ring->head = tail;
308 /* make sure kernel can see our new head value - probably not required */
309 ECB_MEMORY_FENCE_RELEASE;
310 402
311 return 1; 403 return 1;
404}
405
406inline_size
407int
408linuxaio_ringbuf_valid (EV_P)
409{
410 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
411
412 return expect_true (ring->magic == AIO_RING_MAGIC)
413 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
414 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
312} 415}
313 416
314/* read at least one event from kernel, or timeout */ 417/* read at least one event from kernel, or timeout */
315inline_size 418inline_size
316void 419void
317linuxaio_get_events (EV_P_ ev_tstamp timeout) 420linuxaio_get_events (EV_P_ ev_tstamp timeout)
318{ 421{
319 struct timespec ts; 422 struct timespec ts;
320 struct io_event ioev[1]; 423 struct io_event ioev[8]; /* 256 octet stack space */
321 int res; 424 int want = 1; /* how many events to request */
425 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
322 426
427 if (expect_true (ringbuf_valid))
428 {
429 /* if the ring buffer has any events, we don't wait or call the kernel at all */
323 if (linuxaio_get_events_from_ring (EV_A)) 430 if (linuxaio_get_events_from_ring (EV_A))
324 return; 431 return;
325 432
326 /* no events, so wait for at least one, then poll ring buffer again */ 433 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
327 /* this degrades to one event per loop iteration */ 434 if (!timeout)
328 /* if the ring buffer changes layout, but so be it */ 435 return;
436 }
437 else
438 /* no ringbuffer, request slightly larger batch */
439 want = sizeof (ioev) / sizeof (ioev [0]);
329 440
441 /* no events, so wait for some
442 * for fairness reasons, we do this in a loop, to fetch all events
443 */
444 for (;;)
445 {
446 int res;
447
330 EV_RELEASE_CB; 448 EV_RELEASE_CB;
331 449
332 ts.tv_sec = (long)timeout; 450 ts.tv_sec = (long)timeout;
333 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 451 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
334 452
335 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); 453 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
336 454
337 EV_ACQUIRE_CB; 455 EV_ACQUIRE_CB;
338 456
339 if (res < 0) 457 if (res < 0)
340 if (errno == EINTR) 458 if (errno == EINTR)
341 /* ignored */; 459 /* ignored, retry */;
342 else 460 else
343 ev_syserr ("(libev) linuxaio io_getevents"); 461 ev_syserr ("(libev) linuxaio io_getevents");
344 else if (res) 462 else if (res)
345 { 463 {
346 /* at least one event received, handle it and any remaining ones in the ring buffer */ 464 /* at least one event available, handle them */
347 linuxaio_parse_events (EV_A_ ioev, res); 465 linuxaio_parse_events (EV_A_ ioev, res);
466
467 if (expect_true (ringbuf_valid))
468 {
469 /* if we have a ring buffer, handle any remaining events in it */
348 linuxaio_get_events_from_ring (EV_A); 470 linuxaio_get_events_from_ring (EV_A);
471
472 /* at this point, we should have handled all outstanding events */
473 break;
474 }
475 else if (res < want)
476 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
477 break;
478 }
479 else
480 break; /* no events from the kernel, we are done */
481
482 timeout = 0; /* only wait in the first iteration */
349 } 483 }
484}
485
486inline_size
487int
488linuxaio_io_setup (EV_P)
489{
490 linuxaio_ctx = 0;
491 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
350} 492}
351 493
352static void 494static void
353linuxaio_poll (EV_P_ ev_tstamp timeout) 495linuxaio_poll (EV_P_ ev_tstamp timeout)
354{ 496{
… …
356 498
357 /* first phase: submit new iocbs */ 499 /* first phase: submit new iocbs */
358 500
359 /* io_submit might return less than the requested number of iocbs */ 501 /* io_submit might return less than the requested number of iocbs */
360 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 502 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
361 /* which allows us to pinpoint the errornous iocb */ 503 /* which allows us to pinpoint the erroneous iocb */
362 for (submitted = 0; submitted < linuxaio_submitcnt; ) 504 for (submitted = 0; submitted < linuxaio_submitcnt; )
363 { 505 {
364#if 0
365 int res;
366 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
367 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
368 else
369 { res = -1; errno = EINVAL; };
370#else
371 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 506 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
372#endif
373 507
374 if (expect_false (res < 0)) 508 if (expect_false (res < 0))
375 if (errno == EAGAIN)
376 {
377 /* This happens when the ring buffer is full, at least. I assume this means
378 * that the event was queued synchronously during io_submit, and thus
379 * the buffer overflowed.
380 * In this case, we just try in next loop iteration.
381 * This should not result in a few fds taking priority, as the interface
382 * is one-shot, and we submit iocb's in a round-robin fashion.
383 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
384 */
385 if (ecb_expect_false (submitted))
386 {
387 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
388 linuxaio_submitcnt -= submitted;
389 }
390
391 timeout = 0;
392 break;
393 }
394#if EPOLL_FALLBACK
395 else if (errno == EINVAL) 509 if (errno == EINVAL)
396 { 510 {
397 /* This happens for unsupported fds, officially, but in my testing, 511 /* This happens for unsupported fds, officially, but in my testing,
398 * also randomly happens for supported fds. We fall back to good old 512 * also randomly happens for supported fds. We fall back to good old
399 * poll() here, under the assumption that this is a very rare case. 513 * poll() here, under the assumption that this is a very rare case.
400 * See https://lore.kernel.org/patchwork/patch/1047453/ to see 514 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
401 * discussion about such a case (ttys) where polling for POLLIN 515 * discussion about such a case (ttys) where polling for POLLIN
402 * fails but POLLIN|POLLOUT works. 516 * fails but POLLIN|POLLOUT works.
403 */ 517 */
404 struct iocb *iocb = linuxaio_submits [submitted]; 518 struct iocb *iocb = linuxaio_submits [submitted];
405 519 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
406 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */ 520 iocb->aio_reqprio = -1; /* mark iocb as epoll */
521
522 res = 1; /* skip this iocb - another iocb, another chance */
523 }
524 else if (errno == EAGAIN)
525 {
526 /* This happens when the ring buffer is full, or some other shit we
527 * don't know and isn't documented. Most likely because we have too
528 * many requests and linux aio can't be assed to handle them.
529 * In this case, we try to allocate a larger ring buffer, freeing
530 * ours first. This might fail, in which case we have to fall back to 100%
531 * epoll.
532 * God, how I hate linux not getting its act together. Ever.
533 */
534 evsys_io_destroy (linuxaio_ctx);
535 linuxaio_submitcnt = 0;
536
537 /* rearm all fds with active iocbs */
538 {
539 int fd;
540 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
541 if (linuxaio_iocbps [fd]->io.aio_buf)
542 linuxaio_fd_rearm (EV_A_ fd);
543 }
544
545 ++linuxaio_iteration;
546 if (linuxaio_io_setup (EV_A) < 0)
547 {
548 /* to bad, we can't get a new aio context, go 100% epoll */
549 linuxaio_free_iocbp (EV_A);
550 ev_io_stop (EV_A_ &linuxaio_epoll_w);
551 ev_ref (EV_A);
552 linuxaio_ctx = 0;
553 backend_modify = epoll_modify;
554 backend_poll = epoll_poll;
555 }
556
557 timeout = 0;
558 /* it's easiest to handle this mess in another iteration */
559 return;
560 }
561 else if (errno == EBADF)
562 {
563 assert (("libev: event loop rejected bad fd", errno != EBADF));
564 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
408 565
409 res = 1; /* skip this iocb */ 566 res = 1; /* skip this iocb */
410 } 567 }
411#endif 568 else if (errno == EINTR) /* not seen in reality, not documented */
412 else if (errno == EBADF) 569 res = 0; /* silently ignore and retry */
413 {
414 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
415
416 res = 1; /* skip this iocb */
417 }
418 else 570 else
419 ev_syserr ("(libev) linuxaio io_submit"); 571 ev_syserr ("(libev) linuxaio io_submit");
420 572
421 submitted += res; 573 submitted += res;
422 } 574 }
… …
432int 584int
433linuxaio_init (EV_P_ int flags) 585linuxaio_init (EV_P_ int flags)
434{ 586{
435 /* would be great to have a nice test for IOCB_CMD_POLL instead */ 587 /* would be great to have a nice test for IOCB_CMD_POLL instead */
436 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ 588 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
437#if EPOLL_FALLBACK 589 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
438 /* 4.19 made epoll work */
439 if (ev_linux_version () < 0x041300) 590 if (ev_linux_version () < 0x041300)
440 return 0; 591 return 0;
441#else 592
442 /* 4.18 introduced IOCB_CMD_POLL */ 593 if (!epoll_init (EV_A_ 0))
443 if (ev_linux_version () < 0x041200)
444 return 0; 594 return 0;
445#endif
446 595
447 linuxaio_ctx = 0; 596 linuxaio_iteration = 0;
448 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
449 return 0;
450 597
451#if EPOLL_FALLBACK 598 if (linuxaio_io_setup (EV_A) < 0)
452 backend_fd = ev_epoll_create ();
453 if (backend_fd < 0)
454 { 599 {
455 evsys_io_destroy (linuxaio_ctx); 600 epoll_destroy (EV_A);
456 return 0; 601 return 0;
457 } 602 }
458 603
459 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 604 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
460 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 605 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
461 ev_io_start (EV_A_ &linuxaio_epoll_w); 606 ev_io_start (EV_A_ &linuxaio_epoll_w);
462 ev_unref (EV_A); /* watcher should not keep loop alive */ 607 ev_unref (EV_A); /* watcher should not keep loop alive */
463#endif
464 608
465 backend_modify = linuxaio_modify; 609 backend_modify = linuxaio_modify;
466 backend_poll = linuxaio_poll; 610 backend_poll = linuxaio_poll;
467 611
468 linuxaio_iocbpmax = 0; 612 linuxaio_iocbpmax = 0;
… …
477 621
478inline_size 622inline_size
479void 623void
480linuxaio_destroy (EV_P) 624linuxaio_destroy (EV_P)
481{ 625{
482#if EPOLL_FALLBACK 626 epoll_destroy (EV_A);
483 close (backend_fd);
484#endif
485 linuxaio_free_iocbp (EV_A); 627 linuxaio_free_iocbp (EV_A);
486 evsys_io_destroy (linuxaio_ctx); 628 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
487} 629}
488 630
489inline_size 631inline_size
490void 632void
491linuxaio_fork (EV_P) 633linuxaio_fork (EV_P)
492{ 634{
493 /* this frees all iocbs, which is very heavy-handed */ 635 /* this frees all iocbs, which is very heavy-handed */
494 linuxaio_destroy (EV_A); 636 linuxaio_destroy (EV_A);
495 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 637 linuxaio_submitcnt = 0; /* all pointers were invalidated */
496 638
497 linuxaio_ctx = 0; 639 linuxaio_iteration = 0; /* we start over in the child */
498 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 640
641 while (linuxaio_io_setup (EV_A) < 0)
499 ev_syserr ("(libev) linuxaio io_setup"); 642 ev_syserr ("(libev) linuxaio io_setup");
500 643
501#if EPOLL_FALLBACK 644 /* forking epoll should also effectively unregister all fds from the backend */
502 while ((backend_fd = ev_epoll_create ()) < 0) 645 epoll_fork (EV_A);
503 ev_syserr ("(libev) linuxaio epoll_create");
504 646
505 ev_io_stop (EV_A_ &linuxaio_epoll_w); 647 ev_io_stop (EV_A_ &linuxaio_epoll_w);
506 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 648 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
507 ev_io_start (EV_A_ &linuxaio_epoll_w); 649 ev_io_start (EV_A_ &linuxaio_epoll_w);
508#endif
509 650
651 /* epoll_fork already did this. hopefully */
510 fd_rearm_all (EV_A); 652 /*fd_rearm_all (EV_A);*/
511} 653}
512 654

Diff Legend

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