… | |
… | |
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 4kB pages exactly. |
|
|
45 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
|
|
46 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. |
|
|
47 | * therefore the calculation below will use "exactly" 8kB for the ring buffer |
|
|
48 | */ |
|
|
49 | #define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */ |
|
|
50 | |
|
|
51 | /*****************************************************************************/ |
91 | /*****************************************************************************/ |
52 | /* syscall wrapdadoop */ |
92 | /* syscall wrapdadoop - this section has the raw api/abi definitions */ |
53 | |
93 | |
54 | #include <sys/syscall.h> /* no glibc wrappers */ |
94 | #include <sys/syscall.h> /* no glibc wrappers */ |
55 | |
95 | |
56 | /* 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 */ |
57 | #define IOCB_CMD_POLL 5 |
97 | #define IOCB_CMD_POLL 5 |
58 | |
98 | |
59 | /* 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 | */ |
60 | #define AIO_RING_MAGIC 0xa10a10a1 |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
61 | #define AIO_RING_INCOMPAT_FEATURES 0 |
105 | #define EV_AIO_RING_INCOMPAT_FEATURES 0 |
62 | struct aio_ring |
106 | struct aio_ring |
63 | { |
107 | { |
64 | unsigned id; /* kernel internal index number */ |
108 | unsigned id; /* kernel internal index number */ |
65 | unsigned nr; /* number of io_events */ |
109 | unsigned nr; /* number of io_events */ |
66 | unsigned head; /* Written to by userland or by kernel. */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
… | |
… | |
74 | struct io_event io_events[0]; |
118 | struct io_event io_events[0]; |
75 | }; |
119 | }; |
76 | |
120 | |
77 | inline_size |
121 | inline_size |
78 | int |
122 | int |
79 | ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
80 | { |
124 | { |
81 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
125 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
82 | } |
126 | } |
83 | |
127 | |
84 | inline_size |
128 | inline_size |
85 | int |
129 | int |
86 | ev_io_destroy (aio_context_t ctx_id) |
130 | evsys_io_destroy (aio_context_t ctx_id) |
87 | { |
131 | { |
88 | return syscall (SYS_io_destroy, ctx_id); |
132 | return ev_syscall1 (SYS_io_destroy, ctx_id); |
89 | } |
133 | } |
90 | |
134 | |
91 | inline_size |
135 | inline_size |
92 | int |
136 | int |
93 | ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
94 | { |
138 | { |
95 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
139 | return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
96 | } |
140 | } |
97 | |
141 | |
98 | inline_size |
142 | inline_size |
99 | int |
143 | int |
100 | ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
101 | { |
145 | { |
102 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
146 | return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
103 | } |
147 | } |
104 | |
148 | |
105 | inline_size |
149 | inline_size |
106 | int |
150 | int |
107 | ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
108 | { |
152 | { |
109 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
153 | return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
110 | } |
154 | } |
111 | |
155 | |
112 | /*****************************************************************************/ |
156 | /*****************************************************************************/ |
113 | /* actual backed implementation */ |
157 | /* actual backed implementation */ |
114 | |
158 | |
|
|
159 | ecb_cold |
|
|
160 | static int |
|
|
161 | linuxaio_nr_events (EV_P) |
|
|
162 | { |
|
|
163 | /* we start with 16 iocbs and incraese from there |
|
|
164 | * that's tiny, but the kernel has a rather low system-wide |
|
|
165 | * limit that can be reached quickly, so let's be parsimonious |
|
|
166 | * with this resource. |
|
|
167 | * Rest assured, the kernel generously rounds up small and big numbers |
|
|
168 | * in different ways (but doesn't seem to charge you for it). |
|
|
169 | * The 15 here is because the kernel usually has a power of two as aio-max-nr, |
|
|
170 | * and this helps to take advantage of that limit. |
|
|
171 | */ |
|
|
172 | |
|
|
173 | /* we try to fill 4kB pages exactly. |
|
|
174 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
|
|
175 | * the kernel takes the io requests number, doubles it, adds 2 |
|
|
176 | * and adds the ring buffer. |
|
|
177 | * the way we use this is by starting low, and then roughly doubling the |
|
|
178 | * size each time we hit a limit. |
|
|
179 | */ |
|
|
180 | |
|
|
181 | int requests = 15 << linuxaio_iteration; |
|
|
182 | int one_page = (4096 |
|
|
183 | / sizeof (struct io_event) ) / 2; /* how many fit into one page */ |
|
|
184 | int first_page = ((4096 - sizeof (struct aio_ring)) |
|
|
185 | / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */ |
|
|
186 | |
|
|
187 | /* if everything fits into one page, use count exactly */ |
|
|
188 | if (requests > first_page) |
|
|
189 | /* otherwise, round down to full pages and add the first page */ |
|
|
190 | requests = requests / one_page * one_page + first_page; |
|
|
191 | |
|
|
192 | return requests; |
|
|
193 | } |
|
|
194 | |
115 | /* we use out own wrapper structure in acse we ever want to do something "clever" */ |
195 | /* we use out own wrapper structure in case we ever want to do something "clever" */ |
116 | typedef struct aniocb |
196 | typedef struct aniocb |
117 | { |
197 | { |
118 | struct iocb io; |
198 | struct iocb io; |
119 | /*int inuse;*/ |
199 | /*int inuse;*/ |
120 | } *ANIOCBP; |
200 | } *ANIOCBP; |
121 | |
201 | |
122 | inline_size |
202 | inline_size |
123 | void |
203 | void |
124 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) |
204 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
125 | { |
205 | { |
126 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
|
|
127 | while (count--) |
206 | while (count--) |
128 | { |
207 | { |
|
|
208 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */ |
129 | *base = (ANIOCBP)ev_malloc (sizeof (**base)); |
209 | ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
130 | /* TODO: full zero initialize required? */ |
210 | |
|
|
211 | /* full zero initialise is probably not required at the moment, but |
|
|
212 | * this is not well documented, so we better do it. |
|
|
213 | */ |
131 | memset (*base, 0, sizeof (**base)); |
214 | memset (iocb, 0, sizeof (*iocb)); |
132 | /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ |
215 | |
133 | (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; |
216 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
134 | ++base; |
217 | iocb->io.aio_fildes = offset; |
|
|
218 | |
|
|
219 | base [offset++] = iocb; |
135 | } |
220 | } |
136 | } |
221 | } |
137 | |
222 | |
138 | ecb_cold |
223 | ecb_cold |
139 | static void |
224 | static void |
… | |
… | |
147 | |
232 | |
148 | static void |
233 | static void |
149 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
234 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
150 | { |
235 | { |
151 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
236 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
152 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
237 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
|
|
238 | ANFD *anfd = &anfds [fd]; |
153 | |
239 | |
154 | if (iocb->io.aio_buf) |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
155 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
241 | { |
|
|
242 | /* we handed this fd over to epoll, so undo this first */ |
|
|
243 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
|
|
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
245 | anfd->emask = 0; |
|
|
246 | iocb->io.aio_reqprio = 0; |
|
|
247 | } |
|
|
248 | else if (ecb_expect_false (iocb->io.aio_buf)) |
|
|
249 | { |
|
|
250 | /* iocb active, so cancel it first before resubmit */ |
|
|
251 | /* this assumes we only ever get one call per fd per loop iteration */ |
|
|
252 | for (;;) |
|
|
253 | { |
|
|
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
|
|
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
|
|
256 | break; |
|
|
257 | |
|
|
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
|
|
259 | break; |
|
|
260 | |
|
|
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
262 | if (errno != EINTR) |
|
|
263 | { |
|
|
264 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS)); |
|
|
265 | break; |
|
|
266 | } |
|
|
267 | } |
|
|
268 | |
|
|
269 | /* increment generation counter to avoid handling old events */ |
|
|
270 | ++anfd->egen; |
|
|
271 | } |
|
|
272 | |
|
|
273 | iocb->io.aio_buf = (nev & EV_READ ? POLLIN : 0) |
|
|
274 | | (nev & EV_WRITE ? POLLOUT : 0); |
156 | |
275 | |
157 | if (nev) |
276 | if (nev) |
158 | { |
277 | { |
159 | iocb->io.aio_data = fd; |
278 | iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32); |
160 | iocb->io.aio_fildes = fd; |
|
|
161 | iocb->io.aio_buf = |
|
|
162 | (nev & EV_READ ? POLLIN : 0) |
|
|
163 | | (nev & EV_WRITE ? POLLOUT : 0); |
|
|
164 | |
279 | |
165 | /* queue iocb up for io_submit */ |
280 | /* queue iocb up for io_submit */ |
166 | /* this assumes we only ever get one call per fd per loop iteration */ |
281 | /* this assumes we only ever get one call per fd per loop iteration */ |
167 | ++linuxaio_submitcnt; |
282 | ++linuxaio_submitcnt; |
168 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
283 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
169 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
284 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
170 | } |
285 | } |
171 | } |
286 | } |
172 | |
287 | |
173 | static void |
288 | static void |
|
|
289 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
290 | { |
|
|
291 | epoll_poll (EV_A_ 0); |
|
|
292 | } |
|
|
293 | |
|
|
294 | inline_speed |
|
|
295 | void |
|
|
296 | linuxaio_fd_rearm (EV_P_ int fd) |
|
|
297 | { |
|
|
298 | anfds [fd].events = 0; |
|
|
299 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
|
|
300 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
|
|
301 | } |
|
|
302 | |
|
|
303 | static void |
174 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
304 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
175 | { |
305 | { |
176 | while (nr) |
306 | while (nr) |
177 | { |
307 | { |
178 | int fd = ev->data; |
308 | int fd = ev->data & 0xffffffff; |
|
|
309 | uint32_t gen = ev->data >> 32; |
179 | int res = ev->res; |
310 | int res = ev->res; |
180 | |
311 | |
181 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
312 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
182 | |
313 | |
183 | /* linux aio is oneshot: rearm fd */ |
314 | /* only accept events if generation counter matches */ |
184 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
315 | if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen)) |
185 | anfds [fd].events = 0; |
316 | { |
186 | fd_change (EV_A_ fd, 0); |
|
|
187 | |
|
|
188 | /* feed events, we do not expect or handle POLLNVAL */ |
317 | /* feed events, we do not expect or handle POLLNVAL */ |
189 | if (ecb_expect_false (res & POLLNVAL)) |
|
|
190 | fd_kill (EV_A_ fd); |
|
|
191 | else |
|
|
192 | fd_event ( |
318 | fd_event ( |
193 | EV_A_ |
319 | EV_A_ |
194 | fd, |
320 | fd, |
195 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
321 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
196 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
322 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
197 | ); |
323 | ); |
|
|
324 | |
|
|
325 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
|
|
326 | linuxaio_fd_rearm (EV_A_ fd); |
|
|
327 | } |
198 | |
328 | |
199 | --nr; |
329 | --nr; |
200 | ++ev; |
330 | ++ev; |
201 | } |
331 | } |
202 | } |
332 | } |
203 | |
333 | |
204 | /* get any events from ringbuffer, return true if any were handled */ |
334 | /* get any events from ring buffer, return true if any were handled */ |
205 | static int |
335 | static int |
206 | linuxaio_get_events_from_ring (EV_P) |
336 | linuxaio_get_events_from_ring (EV_P) |
207 | { |
337 | { |
208 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
338 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
339 | unsigned head, tail; |
209 | |
340 | |
|
|
341 | /* the kernel reads and writes both of these variables, */ |
|
|
342 | /* as a C extension, we assume that volatile use here */ |
|
|
343 | /* both makes reads atomic and once-only */ |
|
|
344 | head = *(volatile unsigned *)&ring->head; |
210 | ECB_MEMORY_FENCE_ACQUIRE; |
345 | ECB_MEMORY_FENCE_ACQUIRE; |
211 | |
|
|
212 | unsigned head = ring->head; |
|
|
213 | unsigned tail = *(volatile unsigned *)&ring->tail; |
346 | tail = *(volatile unsigned *)&ring->tail; |
214 | |
347 | |
215 | if (head == tail) |
348 | if (head == tail) |
216 | return 0; |
349 | return 0; |
217 | |
350 | |
218 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
219 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
220 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
221 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
222 | return 0; |
|
|
223 | |
|
|
224 | /* parse all available events, but only once, to avoid starvation */ |
351 | /* parse all available events, but only once, to avoid starvation */ |
225 | if (tail > head) /* normal case around */ |
352 | if (ecb_expect_true (tail > head)) /* normal case around */ |
226 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
353 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
227 | else /* wrapped around */ |
354 | else /* wrapped around */ |
228 | { |
355 | { |
229 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
356 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
230 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
357 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
231 | } |
358 | } |
232 | |
359 | |
233 | ring->head = tail; |
360 | ECB_MEMORY_FENCE_RELEASE; |
|
|
361 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
|
|
362 | *(volatile unsigned *)&ring->head = tail; |
234 | |
363 | |
235 | return 1; |
364 | return 1; |
|
|
365 | } |
|
|
366 | |
|
|
367 | inline_size |
|
|
368 | int |
|
|
369 | linuxaio_ringbuf_valid (EV_P) |
|
|
370 | { |
|
|
371 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
372 | |
|
|
373 | return ecb_expect_true (ring->magic == AIO_RING_MAGIC) |
|
|
374 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
|
|
375 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
236 | } |
376 | } |
237 | |
377 | |
238 | /* read at least one event from kernel, or timeout */ |
378 | /* read at least one event from kernel, or timeout */ |
239 | inline_size |
379 | inline_size |
240 | void |
380 | void |
241 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
381 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
242 | { |
382 | { |
243 | struct timespec ts; |
383 | struct timespec ts; |
244 | struct io_event ioev; |
384 | struct io_event ioev[8]; /* 256 octet stack space */ |
245 | int res; |
385 | int want = 1; /* how many events to request */ |
|
|
386 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
246 | |
387 | |
|
|
388 | if (ecb_expect_true (ringbuf_valid)) |
|
|
389 | { |
|
|
390 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
247 | if (linuxaio_get_events_from_ring (EV_A)) |
391 | if (linuxaio_get_events_from_ring (EV_A)) |
248 | return; |
392 | return; |
249 | |
393 | |
250 | /* no events, so wait for at least one, then poll ring buffer again */ |
394 | /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
251 | /* this degrades to one event per loop iteration */ |
395 | if (!timeout) |
252 | /* if the ring buffer changes layout, but so be it */ |
396 | return; |
|
|
397 | } |
|
|
398 | else |
|
|
399 | /* no ringbuffer, request slightly larger batch */ |
|
|
400 | want = sizeof (ioev) / sizeof (ioev [0]); |
253 | |
401 | |
254 | ts.tv_sec = (long)timeout; |
402 | /* no events, so wait for some |
255 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
403 | * for fairness reasons, we do this in a loop, to fetch all events |
|
|
404 | */ |
|
|
405 | for (;;) |
|
|
406 | { |
|
|
407 | int res; |
256 | |
408 | |
|
|
409 | EV_RELEASE_CB; |
|
|
410 | |
|
|
411 | EV_TS_SET (ts, timeout); |
257 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
412 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
258 | |
413 | |
|
|
414 | EV_ACQUIRE_CB; |
|
|
415 | |
259 | if (res < 0) |
416 | if (res < 0) |
|
|
417 | if (errno == EINTR) |
|
|
418 | /* ignored, retry */; |
|
|
419 | else |
260 | ev_syserr ("(libev) io_getevents"); |
420 | ev_syserr ("(libev) linuxaio io_getevents"); |
261 | else if (res) |
421 | else if (res) |
262 | { |
422 | { |
263 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
423 | /* at least one event available, handle them */ |
264 | linuxaio_parse_events (EV_A_ &ioev, 1); |
424 | linuxaio_parse_events (EV_A_ ioev, res); |
|
|
425 | |
|
|
426 | if (ecb_expect_true (ringbuf_valid)) |
|
|
427 | { |
|
|
428 | /* if we have a ring buffer, handle any remaining events in it */ |
265 | linuxaio_get_events_from_ring (EV_A); |
429 | linuxaio_get_events_from_ring (EV_A); |
|
|
430 | |
|
|
431 | /* at this point, we should have handled all outstanding events */ |
|
|
432 | break; |
|
|
433 | } |
|
|
434 | else if (res < want) |
|
|
435 | /* otherwise, if there were fewere events than we wanted, we assume there are no more */ |
|
|
436 | break; |
|
|
437 | } |
|
|
438 | else |
|
|
439 | break; /* no events from the kernel, we are done */ |
|
|
440 | |
|
|
441 | timeout = EV_TS_CONST (0.); /* only wait in the first iteration */ |
266 | } |
442 | } |
|
|
443 | } |
|
|
444 | |
|
|
445 | inline_size |
|
|
446 | int |
|
|
447 | linuxaio_io_setup (EV_P) |
|
|
448 | { |
|
|
449 | linuxaio_ctx = 0; |
|
|
450 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
267 | } |
451 | } |
268 | |
452 | |
269 | static void |
453 | static void |
270 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
454 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
271 | { |
455 | { |
… | |
… | |
273 | |
457 | |
274 | /* first phase: submit new iocbs */ |
458 | /* first phase: submit new iocbs */ |
275 | |
459 | |
276 | /* io_submit might return less than the requested number of iocbs */ |
460 | /* io_submit might return less than the requested number of iocbs */ |
277 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
461 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
278 | /* which allows us to pinpoint the errornous iocb */ |
462 | /* which allows us to pinpoint the erroneous iocb */ |
279 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
463 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
280 | { |
464 | { |
281 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
465 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
282 | |
466 | |
283 | if (ecb_expect_false (res < 0)) |
467 | if (ecb_expect_false (res < 0)) |
284 | if (errno == EAGAIN) |
468 | if (errno == EINVAL) |
285 | { |
469 | { |
286 | /* This happens when the ring buffer is full, at least. I assume this means |
470 | /* This happens for unsupported fds, officially, but in my testing, |
287 | * that the event was queued synchronously during io_submit, and thus |
471 | * also randomly happens for supported fds. We fall back to good old |
288 | * the buffer overflowd. |
472 | * poll() here, under the assumption that this is a very rare case. |
289 | * In this case, we just try next loop iteration. |
473 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
290 | * This should not result in a few fds taking priority, as the interface |
474 | * discussion about such a case (ttys) where polling for POLLIN |
291 | * is one-shot, and we submit iocb's in a round-robin fashion. |
475 | * fails but POLLIN|POLLOUT works. |
292 | */ |
476 | */ |
293 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
477 | struct iocb *iocb = linuxaio_submits [submitted]; |
294 | linuxaio_submitcnt -= submitted; |
478 | epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events); |
295 | timeout = 0; |
479 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
296 | break; |
480 | |
|
|
481 | res = 1; /* skip this iocb - another iocb, another chance */ |
297 | } |
482 | } |
|
|
483 | else if (errno == EAGAIN) |
|
|
484 | { |
|
|
485 | /* This happens when the ring buffer is full, or some other shit we |
|
|
486 | * don't know and isn't documented. Most likely because we have too |
|
|
487 | * many requests and linux aio can't be assed to handle them. |
|
|
488 | * In this case, we try to allocate a larger ring buffer, freeing |
|
|
489 | * ours first. This might fail, in which case we have to fall back to 100% |
|
|
490 | * epoll. |
|
|
491 | * God, how I hate linux not getting its act together. Ever. |
|
|
492 | */ |
|
|
493 | evsys_io_destroy (linuxaio_ctx); |
|
|
494 | linuxaio_submitcnt = 0; |
|
|
495 | |
|
|
496 | /* rearm all fds with active iocbs */ |
|
|
497 | { |
|
|
498 | int fd; |
|
|
499 | for (fd = 0; fd < linuxaio_iocbpmax; ++fd) |
|
|
500 | if (linuxaio_iocbps [fd]->io.aio_buf) |
|
|
501 | linuxaio_fd_rearm (EV_A_ fd); |
|
|
502 | } |
|
|
503 | |
|
|
504 | ++linuxaio_iteration; |
|
|
505 | if (linuxaio_io_setup (EV_A) < 0) |
|
|
506 | { |
|
|
507 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
508 | /* TODO: might be more prudent? */ |
|
|
509 | |
|
|
510 | /* to bad, we can't get a new aio context, go 100% epoll */ |
|
|
511 | linuxaio_free_iocbp (EV_A); |
|
|
512 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
513 | ev_ref (EV_A); |
|
|
514 | linuxaio_ctx = 0; |
|
|
515 | |
|
|
516 | backend = EVBACKEND_EPOLL; |
|
|
517 | backend_modify = epoll_modify; |
|
|
518 | backend_poll = epoll_poll; |
|
|
519 | } |
|
|
520 | |
|
|
521 | timeout = EV_TS_CONST (0.); |
|
|
522 | /* it's easiest to handle this mess in another iteration */ |
|
|
523 | return; |
|
|
524 | } |
|
|
525 | else if (errno == EBADF) |
|
|
526 | { |
|
|
527 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
|
|
528 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
|
|
529 | |
|
|
530 | res = 1; /* skip this iocb */ |
|
|
531 | } |
|
|
532 | else if (errno == EINTR) /* not seen in reality, not documented */ |
|
|
533 | res = 0; /* silently ignore and retry */ |
298 | else |
534 | else |
|
|
535 | { |
299 | ev_syserr ("(libev) io_submit"); |
536 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
537 | res = 0; |
|
|
538 | } |
300 | |
539 | |
301 | submitted += res; |
540 | submitted += res; |
302 | } |
541 | } |
303 | |
542 | |
304 | linuxaio_submitcnt = 0; |
543 | linuxaio_submitcnt = 0; |
… | |
… | |
312 | int |
551 | int |
313 | linuxaio_init (EV_P_ int flags) |
552 | linuxaio_init (EV_P_ int flags) |
314 | { |
553 | { |
315 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
554 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
316 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
555 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
317 | if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ |
556 | /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */ |
|
|
557 | if (ev_linux_version () < 0x041300) |
318 | return 0; |
558 | return 0; |
319 | |
559 | |
320 | linuxaio_ctx = 0; |
560 | if (!epoll_init (EV_A_ 0)) |
321 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
|
|
322 | return 0; |
561 | return 0; |
323 | |
562 | |
|
|
563 | linuxaio_iteration = 0; |
|
|
564 | |
|
|
565 | if (linuxaio_io_setup (EV_A) < 0) |
|
|
566 | { |
|
|
567 | epoll_destroy (EV_A); |
|
|
568 | return 0; |
|
|
569 | } |
|
|
570 | |
|
|
571 | ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
572 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
|
|
573 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
574 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
575 | |
324 | backend_modify = linuxaio_modify; |
576 | backend_modify = linuxaio_modify; |
325 | backend_poll = linuxaio_poll; |
577 | backend_poll = linuxaio_poll; |
326 | |
578 | |
327 | linuxaio_iocbpmax = 0; |
579 | linuxaio_iocbpmax = 0; |
328 | linuxaio_iocbps = 0; |
580 | linuxaio_iocbps = 0; |
329 | |
581 | |
330 | linuxaio_submits = 0; |
582 | linuxaio_submits = 0; |
… | |
… | |
336 | |
588 | |
337 | inline_size |
589 | inline_size |
338 | void |
590 | void |
339 | linuxaio_destroy (EV_P) |
591 | linuxaio_destroy (EV_P) |
340 | { |
592 | { |
|
|
593 | epoll_destroy (EV_A); |
341 | linuxaio_free_iocbp (EV_A); |
594 | linuxaio_free_iocbp (EV_A); |
342 | ev_io_destroy (linuxaio_ctx); |
595 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
343 | } |
596 | } |
344 | |
597 | |
345 | inline_size |
598 | ecb_cold |
346 | void |
599 | static void |
347 | linuxaio_fork (EV_P) |
600 | linuxaio_fork (EV_P) |
348 | { |
601 | { |
349 | /* this frees all iocbs, which is very heavy-handed */ |
|
|
350 | linuxaio_destroy (EV_A); |
|
|
351 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
602 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
|
|
603 | linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */ |
|
|
604 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
352 | |
605 | |
353 | linuxaio_ctx = 0; |
606 | linuxaio_iteration = 0; /* we start over in the child */ |
354 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
607 | |
|
|
608 | while (linuxaio_io_setup (EV_A) < 0) |
355 | ev_syserr ("(libev) io_setup"); |
609 | ev_syserr ("(libev) linuxaio io_setup"); |
356 | |
610 | |
|
|
611 | /* forking epoll should also effectively unregister all fds from the backend */ |
|
|
612 | epoll_fork (EV_A); |
|
|
613 | /* epoll_fork already did this. hopefully */ |
357 | fd_rearm_all (EV_A); |
614 | /*fd_rearm_all (EV_A);*/ |
358 | } |
|
|
359 | |
615 | |
|
|
616 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
617 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
|
|
618 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
619 | } |
|
|
620 | |