… | |
… | |
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 |
|
|
41 | |
40 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
42 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
|
|
43 | #include <poll.h> |
41 | #include <linux/aio_abi.h> |
44 | #include <linux/aio_abi.h> |
42 | |
45 | |
|
|
46 | #if EPOLL_FALLBACK |
|
|
47 | # include <sys/epoll.h> |
|
|
48 | #endif |
|
|
49 | |
43 | /* we try to fill 4kn pages exactly. |
50 | /* we try to fill 4kB pages exactly. |
44 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
51 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
45 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer |
52 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. |
46 | * so the calculation below will use "exactly" 8kB for the ring buffer |
53 | * therefore the calculation below will use "exactly" 4kB for the ring buffer |
47 | */ |
54 | */ |
48 | #define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */ |
55 | #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ |
49 | |
56 | |
50 | /*****************************************************************************/ |
57 | /*****************************************************************************/ |
51 | /* syscall wrapdadoop */ |
58 | /* syscall wrapdadoop - this section has the raw syscall definitions */ |
52 | |
59 | |
53 | #include <sys/syscall.h> /* no glibc wrappers */ |
60 | #include <sys/syscall.h> /* no glibc wrappers */ |
54 | |
61 | |
55 | /* aio_abi.h is not verioned in any way, so we cannot test for its existance */ |
62 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
56 | #define IOCB_CMD_POLL 5 |
63 | #define IOCB_CMD_POLL 5 |
57 | |
64 | |
58 | /* taken from linux/fs/aio.c */ |
65 | /* taken from linux/fs/aio.c */ |
59 | #define AIO_RING_MAGIC 0xa10a10a1 |
66 | #define AIO_RING_MAGIC 0xa10a10a1 |
60 | #define AIO_RING_INCOMPAT_FEATURES 0 |
67 | #define AIO_RING_INCOMPAT_FEATURES 0 |
… | |
… | |
71 | unsigned header_length; /* size of aio_ring */ |
78 | unsigned header_length; /* size of aio_ring */ |
72 | |
79 | |
73 | struct io_event io_events[0]; |
80 | struct io_event io_events[0]; |
74 | }; |
81 | }; |
75 | |
82 | |
76 | static int |
83 | inline_size |
|
|
84 | int |
77 | ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
85 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
78 | { |
86 | { |
79 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
87 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
80 | } |
88 | } |
81 | |
89 | |
82 | static int |
90 | inline_size |
|
|
91 | int |
83 | ev_io_destroy (aio_context_t ctx_id) |
92 | evsys_io_destroy (aio_context_t ctx_id) |
84 | { |
93 | { |
85 | return syscall (SYS_io_destroy, ctx_id); |
94 | return syscall (SYS_io_destroy, ctx_id); |
86 | } |
95 | } |
87 | |
96 | |
88 | static int |
97 | inline_size |
|
|
98 | int |
89 | ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
99 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
90 | { |
100 | { |
91 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
101 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
92 | } |
102 | } |
93 | |
103 | |
94 | static int |
104 | inline_size |
|
|
105 | int |
95 | ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
106 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
96 | { |
107 | { |
97 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
108 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
98 | } |
109 | } |
99 | |
110 | |
100 | static int |
111 | inline_size |
|
|
112 | int |
101 | ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
113 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
102 | { |
114 | { |
103 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
115 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
104 | } |
116 | } |
105 | |
|
|
106 | typedef void (*ev_io_cb) (long nr, struct io_event *events); |
|
|
107 | |
117 | |
108 | /*****************************************************************************/ |
118 | /*****************************************************************************/ |
109 | /* actual backed implementation */ |
119 | /* actual backed implementation */ |
110 | |
120 | |
111 | /* two iocbs for every fd, one for read, one for write */ |
121 | /* we use out own wrapper structure in acse we ever want to do something "clever" */ |
112 | typedef struct aniocb |
122 | typedef struct aniocb |
113 | { |
123 | { |
114 | struct iocb io; |
124 | struct iocb io; |
115 | /*int inuse;*/ |
125 | /*int inuse;*/ |
116 | } *ANIOCBP; |
126 | } *ANIOCBP; |
117 | |
127 | |
118 | inline_size |
128 | inline_size |
119 | void |
129 | void |
120 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) |
130 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) |
121 | { |
131 | { |
122 | /* TODO: quite the overhead to allocate every iocb separately */ |
132 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
123 | while (count--) |
133 | while (count--) |
124 | { |
134 | { |
125 | *base = (ANIOCBP)ev_malloc (sizeof (**base)); |
135 | *base = (ANIOCBP)ev_malloc (sizeof (**base)); |
|
|
136 | /* TODO: full zero initialize required? */ |
126 | memset (*base, 0, sizeof (**base)); |
137 | memset (*base, 0, sizeof (**base)); |
127 | /* would be nice to initialize fd/data as well */ |
138 | /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ |
128 | (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; |
139 | (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; |
129 | ++base; |
140 | ++base; |
130 | } |
141 | } |
131 | } |
142 | } |
132 | |
143 | |
|
|
144 | ecb_cold |
133 | static void |
145 | static void |
134 | linuxaio_free_iocbp (EV_P) |
146 | linuxaio_free_iocbp (EV_P) |
135 | { |
147 | { |
136 | while (linuxaio_iocbpmax--) |
148 | while (linuxaio_iocbpmax--) |
137 | ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); |
149 | ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); |
138 | |
150 | |
139 | linuxaio_iocbpmax = 0; |
151 | linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */ |
140 | } |
152 | } |
141 | |
153 | |
142 | static void |
154 | static void |
143 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
155 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
144 | { |
156 | { |
145 | /* TODO: full zero initialize required? */ |
|
|
146 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
157 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
147 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
158 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
148 | |
159 | |
|
|
160 | #if EPOLL_FALLBACK |
|
|
161 | if (iocb->io.aio_reqprio < 0) |
|
|
162 | { |
|
|
163 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
164 | iocb->io.aio_reqprio = 0; |
|
|
165 | } |
|
|
166 | #endif |
|
|
167 | |
149 | if (iocb->io.aio_buf) |
168 | if (iocb->io.aio_buf) |
150 | ev_io_cancel (linuxaio_ctx, &iocb->io, (void *)0); |
169 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
151 | |
170 | |
152 | if (nev) |
171 | if (nev) |
153 | { |
172 | { |
154 | iocb->io.aio_data = fd; |
173 | iocb->io.aio_data = fd; |
155 | iocb->io.aio_fildes = fd; |
174 | iocb->io.aio_fildes = fd; |
… | |
… | |
163 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
182 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
164 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
165 | } |
184 | } |
166 | } |
185 | } |
167 | |
186 | |
|
|
187 | #if EPOLL_FALLBACK |
|
|
188 | |
|
|
189 | static void |
|
|
190 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
191 | { |
|
|
192 | struct epoll_event eev; |
|
|
193 | |
|
|
194 | eev.events = EPOLLONESHOT; |
|
|
195 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
196 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
197 | eev.data.fd = iocb->aio_fildes; |
|
|
198 | |
|
|
199 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
200 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
201 | } |
|
|
202 | |
|
|
203 | static void |
|
|
204 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
205 | { |
|
|
206 | struct epoll_event events[16]; |
|
|
207 | |
|
|
208 | for (;;) |
|
|
209 | { |
|
|
210 | int idx; |
|
|
211 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
212 | |
|
|
213 | if (expect_false (res < 0)) |
|
|
214 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
215 | else if (!res) |
|
|
216 | break; |
|
|
217 | |
|
|
218 | for (idx = res; idx--; ) |
|
|
219 | { |
|
|
220 | int fd = events [idx].data.fd; |
|
|
221 | uint32_t ev = events [idx].events; |
|
|
222 | |
|
|
223 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
224 | |
|
|
225 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
226 | |
|
|
227 | fd_event (EV_A_ fd, |
|
|
228 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
229 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
230 | } |
|
|
231 | |
|
|
232 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
233 | break; |
|
|
234 | } |
|
|
235 | } |
|
|
236 | |
|
|
237 | #endif |
|
|
238 | |
168 | static void |
239 | static void |
169 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
240 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
170 | { |
241 | { |
171 | while (nr) |
242 | while (nr) |
172 | { |
243 | { |
173 | int fd = ev->data; |
244 | int fd = ev->data; |
174 | int res = ev->res; |
245 | int res = ev->res; |
175 | |
246 | |
176 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdxmax)); |
247 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
177 | |
248 | |
178 | /* linux aio is oneshot: rearm fd */ |
249 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
179 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
250 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
180 | anfds [fd].events = 0; |
251 | anfds [fd].events = 0; |
181 | fd_change (EV_A_ fd, 0); |
252 | fd_change (EV_A_ fd, 0); |
182 | |
253 | |
183 | /* feed events, we do not expect or handle POLLNVAL */ |
254 | /* feed events, we do not expect or handle POLLNVAL */ |
184 | if (ecb_expect_false (res & POLLNVAL)) |
255 | if (expect_false (res & POLLNVAL)) |
185 | fd_kill (EV_A_ fd); |
256 | fd_kill (EV_A_ fd); |
186 | else |
257 | else |
187 | fd_event ( |
258 | fd_event ( |
188 | EV_A_ |
259 | EV_A_ |
189 | fd, |
260 | fd, |
… | |
… | |
200 | static int |
271 | static int |
201 | linuxaio_get_events_from_ring (EV_P) |
272 | linuxaio_get_events_from_ring (EV_P) |
202 | { |
273 | { |
203 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
274 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
204 | |
275 | |
|
|
276 | /* the kernel reads and writes both of these variables, */ |
|
|
277 | /* as a C extension, we assume that volatile use here */ |
|
|
278 | /* both makes reads atomic and once-only */ |
|
|
279 | unsigned head = *(volatile unsigned *)&ring->head; |
|
|
280 | unsigned tail = *(volatile unsigned *)&ring->tail; |
|
|
281 | |
|
|
282 | if (head == tail) |
|
|
283 | return 0; |
|
|
284 | |
|
|
285 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
286 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
287 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
288 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
289 | return 0; |
|
|
290 | |
|
|
291 | /* make sure the events up to tail are visible */ |
205 | ECB_MEMORY_FENCE_ACQUIRE; |
292 | 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 | |
293 | |
216 | /* parse all available events, but only once, to avoid starvation */ |
294 | /* parse all available events, but only once, to avoid starvation */ |
217 | if (tail > head) /* normal case around */ |
295 | if (tail > head) /* normal case around */ |
218 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
296 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
219 | else |
|
|
220 | { |
|
|
221 | /* wrapped around */ |
297 | else /* wrapped around */ |
|
|
298 | { |
222 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
299 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
223 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
300 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
224 | } |
301 | } |
225 | |
302 | |
226 | ring->head = tail; |
303 | /* TODO: we only need a compiler barrier here, not a read fence */ |
|
|
304 | ECB_MEMORY_FENCE_RELEASE; |
|
|
305 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
|
|
306 | *(volatile unsigned *)&ring->head = tail; |
|
|
307 | /* make sure kernel can see our new head value - probably not required */ |
|
|
308 | ECB_MEMORY_FENCE_RELEASE; |
227 | |
309 | |
228 | return 1; |
310 | return 1; |
229 | } |
311 | } |
230 | |
312 | |
231 | /* read at least one event from kernel, or timeout */ |
313 | /* read at least one event from kernel, or timeout */ |
232 | inline_size |
314 | inline_size |
233 | void |
315 | void |
234 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
316 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
235 | { |
317 | { |
236 | struct timespec ts; |
318 | struct timespec ts; |
237 | struct io_event ioev; |
319 | struct io_event ioev[1]; |
238 | int res; |
320 | int res; |
239 | |
321 | |
240 | if (linuxaio_get_events_from_ring (EV_A)) |
322 | if (linuxaio_get_events_from_ring (EV_A)) |
241 | return; |
323 | return; |
242 | |
324 | |
243 | /* no events, so wait for at least one, then poll ring buffer again */ |
325 | /* no events, so wait for at least one, then poll ring buffer again */ |
244 | /* this degraded to one event per loop iteration */ |
326 | /* this degrades to one event per loop iteration */ |
245 | /* if the ring buffer changes layout, but so be it */ |
327 | /* if the ring buffer changes layout, but so be it */ |
|
|
328 | |
|
|
329 | EV_RELEASE_CB; |
246 | |
330 | |
247 | ts.tv_sec = (long)timeout; |
331 | ts.tv_sec = (long)timeout; |
248 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
332 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
249 | |
333 | |
250 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
334 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
335 | |
|
|
336 | EV_ACQUIRE_CB; |
251 | |
337 | |
252 | if (res < 0) |
338 | if (res < 0) |
|
|
339 | if (errno == EINTR) |
|
|
340 | /* ignored */; |
|
|
341 | else |
253 | ev_syserr ("(libev) io_getevents"); |
342 | ev_syserr ("(libev) linuxaio io_getevents"); |
254 | else if (res) |
343 | else if (res) |
255 | { |
344 | { |
256 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
345 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
257 | linuxaio_parse_events (EV_A_ &ioev, 1); |
346 | linuxaio_parse_events (EV_A_ ioev, res); |
258 | linuxaio_get_events_from_ring (EV_A); |
347 | linuxaio_get_events_from_ring (EV_A); |
259 | } |
348 | } |
260 | } |
349 | } |
261 | |
350 | |
262 | static void |
351 | static void |
… | |
… | |
269 | /* io_submit might return less than the requested number of iocbs */ |
358 | /* 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, */ |
359 | /* 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 */ |
360 | /* which allows us to pinpoint the errornous iocb */ |
272 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
361 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
273 | { |
362 | { |
|
|
363 | #if 0 |
|
|
364 | int res; |
|
|
365 | if (linuxaio_submits[submitted]->aio_fildes == backend_fd) |
|
|
366 | res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); |
|
|
367 | else |
|
|
368 | { res = -1; errno = EINVAL; }; |
|
|
369 | #else |
274 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
370 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
|
|
371 | #endif |
275 | |
372 | |
276 | if (res < 0) |
373 | if (expect_false (res < 0)) |
277 | if (errno == EAGAIN) |
374 | if (errno == EAGAIN) |
278 | { |
375 | { |
279 | /* This happens when the ring buffer is full, at least. I assume this means |
376 | /* This happens when the ring buffer is full, at least. I assume this means |
280 | * that the event was queued synchronously during io_submit, and thus |
377 | * that the event was queued synchronously during io_submit, and thus |
281 | * the buffer overflowd. |
378 | * the buffer overflowed. |
282 | * In this case, we just try next loop iteration. |
379 | * In this case, we just try in next loop iteration. |
|
|
380 | * This should not result in a few fds taking priority, as the interface |
|
|
381 | * is one-shot, and we submit iocb's in a round-robin fashion. |
|
|
382 | * TODO: maybe make "submitted" persistent, so we don't have to memmove? |
283 | */ |
383 | */ |
|
|
384 | if (ecb_expect_false (submitted)) |
|
|
385 | { |
284 | memcpy (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
386 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
285 | linuxaio_submitcnt -= submitted; |
387 | linuxaio_submitcnt -= submitted; |
|
|
388 | } |
|
|
389 | |
286 | timeout = 0; |
390 | timeout = 0; |
287 | break; |
391 | break; |
288 | } |
392 | } |
|
|
393 | #if EPOLL_FALLBACK |
|
|
394 | else if (errno == EINVAL) |
|
|
395 | { |
|
|
396 | /* This happens for unsupported fds, officially, but in my testing, |
|
|
397 | * also randomly happens for supported fds. We fall back to good old |
|
|
398 | * poll() here, under the assumption that this is a very rare case. |
|
|
399 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
|
|
400 | * discussion about such a case (ttys) where polling for POLLIN |
|
|
401 | * fails but POLLIN|POLLOUT works. |
|
|
402 | */ |
|
|
403 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
404 | res = 1; /* skip this iocb */ |
|
|
405 | |
|
|
406 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
407 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
408 | } |
|
|
409 | #endif |
289 | else |
410 | 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"); |
411 | ev_syserr ("(libev) linuxaio io_submit"); |
293 | |
412 | |
294 | submitted += res; |
413 | submitted += res; |
295 | } |
414 | } |
296 | |
415 | |
297 | linuxaio_submitcnt = 0; |
416 | linuxaio_submitcnt = 0; |
… | |
… | |
304 | inline_size |
423 | inline_size |
305 | int |
424 | int |
306 | linuxaio_init (EV_P_ int flags) |
425 | linuxaio_init (EV_P_ int flags) |
307 | { |
426 | { |
308 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
427 | /* 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 */ |
428 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
|
|
429 | #if EPOLL_FALLBACK |
|
|
430 | /* 4.19 made epoll work */ |
|
|
431 | if (ev_linux_version () < 0x041300) |
310 | return 0; |
432 | return 0; |
311 | |
433 | #else |
312 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
434 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
435 | if (ev_linux_version () < 0x041200) |
313 | return 0; |
436 | return 0; |
|
|
437 | #endif |
|
|
438 | |
|
|
439 | linuxaio_ctx = 0; |
|
|
440 | if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
|
|
441 | return 0; |
|
|
442 | |
|
|
443 | #if EPOLL_FALLBACK |
|
|
444 | backend_fd = ev_epoll_create (); |
|
|
445 | if (backend_fd < 0) |
|
|
446 | { |
|
|
447 | evsys_io_destroy (linuxaio_ctx); |
|
|
448 | return 0; |
|
|
449 | } |
|
|
450 | |
|
|
451 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
452 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
|
|
453 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
454 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
455 | #endif |
314 | |
456 | |
315 | backend_modify = linuxaio_modify; |
457 | backend_modify = linuxaio_modify; |
316 | backend_poll = linuxaio_poll; |
458 | backend_poll = linuxaio_poll; |
317 | |
459 | |
318 | linuxaio_iocbpmax = 0; |
460 | linuxaio_iocbpmax = 0; |
… | |
… | |
327 | |
469 | |
328 | inline_size |
470 | inline_size |
329 | void |
471 | void |
330 | linuxaio_destroy (EV_P) |
472 | linuxaio_destroy (EV_P) |
331 | { |
473 | { |
|
|
474 | #if EPOLL_FALLBACK |
|
|
475 | close (backend_fd); |
|
|
476 | #endif |
332 | linuxaio_free_iocbp (EV_A); |
477 | linuxaio_free_iocbp (EV_A); |
333 | ev_io_destroy (linuxaio_ctx); |
478 | ev_io_destroy (linuxaio_ctx); |
334 | } |
479 | } |
335 | |
480 | |
336 | inline_size |
481 | inline_size |
337 | void |
482 | void |
338 | linuxaio_fork (EV_P) |
483 | linuxaio_fork (EV_P) |
339 | { |
484 | { |
340 | abort ();//D |
485 | /* this frees all iocbs, which is very heavy-handed */ |
341 | } |
486 | linuxaio_destroy (EV_A); |
|
|
487 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
342 | |
488 | |
|
|
489 | linuxaio_ctx = 0; |
|
|
490 | while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
|
|
491 | ev_syserr ("(libev) linuxaio io_setup"); |
|
|
492 | |
|
|
493 | #if EPOLL_FALLBACK |
|
|
494 | while ((backend_fd = ev_epoll_create ()) < 0) |
|
|
495 | ev_syserr ("(libev) linuxaio epoll_create"); |
|
|
496 | |
|
|
497 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
498 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
499 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
500 | #endif |
|
|
501 | |
|
|
502 | fd_rearm_all (EV_A); |
|
|
503 | } |
|
|
504 | |