1 | /* |
1 | /* |
2 | * libev epoll fd activity backend |
2 | * libev epoll fd activity backend |
3 | * |
3 | * |
4 | * Copyright (c) 2007,2008 Marc Alexander Lehmann <libev@schmorp.de> |
4 | * Copyright (c) 2007,2008,2009,2010,2011 Marc Alexander Lehmann <libev@schmorp.de> |
5 | * All rights reserved. |
5 | * All rights reserved. |
6 | * |
6 | * |
7 | * Redistribution and use in source and binary forms, with or without modifica- |
7 | * Redistribution and use in source and binary forms, with or without modifica- |
8 | * tion, are permitted provided that the following conditions are met: |
8 | * tion, are permitted provided that the following conditions are met: |
9 | * |
9 | * |
10 | * 1. Redistributions of source code must retain the above copyright notice, |
10 | * 1. Redistributions of source code must retain the above copyright notice, |
11 | * this list of conditions and the following disclaimer. |
11 | * this list of conditions and the following disclaimer. |
12 | * |
12 | * |
13 | * 2. Redistributions in binary form must reproduce the above copyright |
13 | * 2. Redistributions in binary form must reproduce the above copyright |
14 | * notice, this list of conditions and the following disclaimer in the |
14 | * notice, this list of conditions and the following disclaimer in the |
15 | * documentation and/or other materials provided with the distribution. |
15 | * documentation and/or other materials provided with the distribution. |
16 | * |
16 | * |
17 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED |
17 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED |
18 | * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- |
18 | * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- |
19 | * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO |
19 | * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO |
20 | * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE- |
20 | * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE- |
21 | * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
21 | * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
… | |
… | |
47 | * b) the fact that ADD != MOD creates a lot of extra syscalls due to a) |
47 | * b) the fact that ADD != MOD creates a lot of extra syscalls due to a) |
48 | * and seems not to have any advantage. |
48 | * and seems not to have any advantage. |
49 | * c) the inability to handle fork or file descriptors (think dup) |
49 | * c) the inability to handle fork or file descriptors (think dup) |
50 | * limits the applicability over poll, so this is not a generic |
50 | * limits the applicability over poll, so this is not a generic |
51 | * poll replacement. |
51 | * poll replacement. |
|
|
52 | * d) epoll doesn't work the same as select with many file descriptors |
|
|
53 | * (such as files). while not critical, no other advanced interface |
|
|
54 | * seems to share this (rather non-unixy) limitation. |
|
|
55 | * e) epoll claims to be embeddable, but in practise you never get |
|
|
56 | * a ready event for the epoll fd (broken: <=2.6.26, working: >=2.6.32). |
|
|
57 | * f) epoll_ctl returning EPERM means the fd is always ready. |
52 | * |
58 | * |
53 | * lots of "weird code" and complication handling in this file is due |
59 | * lots of "weird code" and complication handling in this file is due |
54 | * to these design problems with epoll, as we try very hard to avoid |
60 | * to these design problems with epoll, as we try very hard to avoid |
55 | * epoll_ctl syscalls for common usage patterns and handle the breakage |
61 | * epoll_ctl syscalls for common usage patterns and handle the breakage |
56 | * ensuing from receiving events for closed and otherwise long gone |
62 | * ensuing from receiving events for closed and otherwise long gone |
57 | * file descriptors. |
63 | * file descriptors. |
58 | */ |
64 | */ |
59 | |
65 | |
60 | #include <sys/epoll.h> |
66 | #include <sys/epoll.h> |
61 | |
67 | |
|
|
68 | #define EV_EMASK_EPERM 0x80 |
|
|
69 | |
62 | static void |
70 | static void |
63 | epoll_modify (EV_P_ int fd, int oev, int nev) |
71 | epoll_modify (EV_P_ int fd, int oev, int nev) |
64 | { |
72 | { |
65 | struct epoll_event ev; |
73 | struct epoll_event ev; |
66 | unsigned char oldmask; |
74 | unsigned char oldmask; |
… | |
… | |
68 | /* |
76 | /* |
69 | * we handle EPOLL_CTL_DEL by ignoring it here |
77 | * we handle EPOLL_CTL_DEL by ignoring it here |
70 | * on the assumption that the fd is gone anyways |
78 | * on the assumption that the fd is gone anyways |
71 | * if that is wrong, we have to handle the spurious |
79 | * if that is wrong, we have to handle the spurious |
72 | * event in epoll_poll. |
80 | * event in epoll_poll. |
73 | * the fd is later added, we try to ADD it, and, if that |
81 | * if the fd is added again, we try to ADD it, and, if that |
74 | * fails, we assume it still has the same eventmask. |
82 | * fails, we assume it still has the same eventmask. |
75 | */ |
83 | */ |
76 | if (!nev) |
84 | if (!nev) |
77 | return; |
85 | return; |
78 | |
86 | |
79 | oldmask = anfds [fd].emask; |
87 | oldmask = anfds [fd].emask; |
80 | anfds [fd].emask = nev; |
88 | anfds [fd].emask = nev; |
81 | |
89 | |
82 | /* store the generation counter in the upper 32 bits */ |
90 | /* store the generation counter in the upper 32 bits, the fd in the lower 32 bits */ |
83 | ev.data.u64 = fd | ((uint64_t)++anfds [fd].egen << 32); |
91 | ev.data.u64 = (uint64_t)(uint32_t)fd |
|
|
92 | | ((uint64_t)(uint32_t)++anfds [fd].egen << 32); |
84 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
93 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
85 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
94 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
86 | |
95 | |
87 | if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
96 | if (expect_true (!epoll_ctl (backend_fd, oev && oldmask != nev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
88 | return; |
97 | return; |
89 | |
98 | |
90 | if (expect_true (errno == ENOENT)) |
99 | if (expect_true (errno == ENOENT)) |
91 | { |
100 | { |
92 | /* if ENOENT then the fd went away, so try to do the right thing */ |
101 | /* if ENOENT then the fd went away, so try to do the right thing */ |
… | |
… | |
104 | goto dec_egen; |
113 | goto dec_egen; |
105 | |
114 | |
106 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
115 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
107 | return; |
116 | return; |
108 | } |
117 | } |
|
|
118 | else if (expect_true (errno == EPERM)) |
|
|
119 | { |
|
|
120 | /* EPERM means the fd is always ready, but epoll is too snobbish */ |
|
|
121 | /* to handle it, unlike select or poll. */ |
|
|
122 | anfds [fd].emask = EV_EMASK_EPERM; |
|
|
123 | |
|
|
124 | /* add fd to epoll_eperms, if not already inside */ |
|
|
125 | if (!(oldmask & EV_EMASK_EPERM)) |
|
|
126 | { |
|
|
127 | array_needsize (int, epoll_eperms, epoll_epermmax, epoll_epermcnt + 1, EMPTY2); |
|
|
128 | epoll_eperms [epoll_epermcnt++] = fd; |
|
|
129 | } |
|
|
130 | |
|
|
131 | return; |
|
|
132 | } |
109 | |
133 | |
110 | fd_kill (EV_A_ fd); |
134 | fd_kill (EV_A_ fd); |
111 | |
135 | |
112 | dec_egen: |
136 | dec_egen: |
113 | /* we didn't successfully call epoll_ctl, so decrement the generation counter again */ |
137 | /* we didn't successfully call epoll_ctl, so decrement the generation counter again */ |
… | |
… | |
116 | |
140 | |
117 | static void |
141 | static void |
118 | epoll_poll (EV_P_ ev_tstamp timeout) |
142 | epoll_poll (EV_P_ ev_tstamp timeout) |
119 | { |
143 | { |
120 | int i; |
144 | int i; |
|
|
145 | int eventcnt; |
|
|
146 | |
|
|
147 | if (expect_false (epoll_epermcnt)) |
|
|
148 | timeout = 0.; |
|
|
149 | |
|
|
150 | /* epoll wait times cannot be larger than (LONG_MAX - 999UL) / HZ msecs, which is below */ |
|
|
151 | /* the default libev max wait time, however. */ |
|
|
152 | EV_RELEASE_CB; |
121 | int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); |
153 | eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, timeout * 1e3); |
|
|
154 | EV_ACQUIRE_CB; |
122 | |
155 | |
123 | if (expect_false (eventcnt < 0)) |
156 | if (expect_false (eventcnt < 0)) |
124 | { |
157 | { |
125 | if (errno != EINTR) |
158 | if (errno != EINTR) |
126 | syserr ("(libev) epoll_wait"); |
159 | ev_syserr ("(libev) epoll_wait"); |
127 | |
160 | |
128 | return; |
161 | return; |
129 | } |
162 | } |
130 | |
163 | |
131 | for (i = 0; i < eventcnt; ++i) |
164 | for (i = 0; i < eventcnt; ++i) |
… | |
… | |
135 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
168 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
136 | int want = anfds [fd].events; |
169 | int want = anfds [fd].events; |
137 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
170 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
138 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
171 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
139 | |
172 | |
|
|
173 | /* |
140 | /* check for spurious notification */ |
174 | * check for spurious notification. |
|
|
175 | * this only finds spurious notifications on egen updates |
|
|
176 | * other spurious notifications will be found by epoll_ctl, below |
|
|
177 | * we assume that fd is always in range, as we never shrink the anfds array |
|
|
178 | */ |
141 | if (expect_false (anfds [fd].egen != (unsigned char)(ev->data.u64 >> 32))) |
179 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
142 | { |
180 | { |
143 | /* recreate kernel state */ |
181 | /* recreate kernel state */ |
144 | postfork = 1; |
182 | postfork = 1; |
145 | continue; |
183 | continue; |
146 | } |
184 | } |
147 | |
185 | |
148 | if (expect_false (got & ~want)) |
186 | if (expect_false (got & ~want)) |
149 | { |
187 | { |
150 | anfds [fd].emask = want; |
188 | anfds [fd].emask = want; |
151 | |
189 | |
|
|
190 | /* |
152 | /* we received an event but are not interested in it, try mod or del */ |
191 | * we received an event but are not interested in it, try mod or del |
153 | /* I don't think we ever need MOD, but let's handle it anyways */ |
192 | * this often happens because we optimistically do not unregister fds |
|
|
193 | * when we are no longer interested in them, but also when we get spurious |
|
|
194 | * notifications for fds from another process. this is partially handled |
|
|
195 | * above with the gencounter check (== our fd is not the event fd), and |
|
|
196 | * partially here, when epoll_ctl returns an error (== a child has the fd |
|
|
197 | * but we closed it). |
|
|
198 | */ |
154 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
199 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
155 | | (want & EV_WRITE ? EPOLLOUT : 0); |
200 | | (want & EV_WRITE ? EPOLLOUT : 0); |
156 | |
201 | |
|
|
202 | /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ |
|
|
203 | /* which is fortunately easy to do for us. */ |
157 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
204 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
158 | { |
205 | { |
159 | postfork = 1; /* an error occured, recreate kernel state */ |
206 | postfork = 1; /* an error occurred, recreate kernel state */ |
160 | continue; |
207 | continue; |
161 | } |
208 | } |
162 | } |
209 | } |
163 | |
210 | |
164 | fd_event (EV_A_ fd, got); |
211 | fd_event (EV_A_ fd, got); |
… | |
… | |
169 | { |
216 | { |
170 | ev_free (epoll_events); |
217 | ev_free (epoll_events); |
171 | epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1); |
218 | epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1); |
172 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
219 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
173 | } |
220 | } |
|
|
221 | |
|
|
222 | /* now synthesize events for all fds where epoll fails, while select works... */ |
|
|
223 | for (i = epoll_epermcnt; i--; ) |
|
|
224 | { |
|
|
225 | int fd = epoll_eperms [i]; |
|
|
226 | unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE); |
|
|
227 | |
|
|
228 | if (anfds [fd].emask & EV_EMASK_EPERM && events) |
|
|
229 | fd_event (EV_A_ fd, events); |
|
|
230 | else |
|
|
231 | epoll_eperms [i] = epoll_eperms [--epoll_epermcnt]; |
|
|
232 | } |
174 | } |
233 | } |
175 | |
234 | |
176 | int inline_size |
235 | int inline_size |
177 | epoll_init (EV_P_ int flags) |
236 | epoll_init (EV_P_ int flags) |
178 | { |
237 | { |
|
|
238 | #ifdef EPOLL_CLOEXEC |
|
|
239 | backend_fd = epoll_create1 (EPOLL_CLOEXEC); |
|
|
240 | |
|
|
241 | if (backend_fd <= 0) |
|
|
242 | #endif |
179 | backend_fd = epoll_create (256); |
243 | backend_fd = epoll_create (256); |
180 | |
244 | |
181 | if (backend_fd < 0) |
245 | if (backend_fd < 0) |
182 | return 0; |
246 | return 0; |
183 | |
247 | |
184 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
248 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
185 | |
249 | |
186 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
250 | backend_mintime = 1./1024.; /* epoll does sometimes return early, this is just to avoid the worst */ |
187 | backend_modify = epoll_modify; |
251 | backend_modify = epoll_modify; |
188 | backend_poll = epoll_poll; |
252 | backend_poll = epoll_poll; |
189 | |
253 | |
190 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
254 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
191 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
255 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
192 | |
256 | |
193 | return EVBACKEND_EPOLL; |
257 | return EVBACKEND_EPOLL; |
… | |
… | |
195 | |
259 | |
196 | void inline_size |
260 | void inline_size |
197 | epoll_destroy (EV_P) |
261 | epoll_destroy (EV_P) |
198 | { |
262 | { |
199 | ev_free (epoll_events); |
263 | ev_free (epoll_events); |
|
|
264 | array_free (epoll_eperm, EMPTY); |
200 | } |
265 | } |
201 | |
266 | |
202 | void inline_size |
267 | void inline_size |
203 | epoll_fork (EV_P) |
268 | epoll_fork (EV_P) |
204 | { |
269 | { |
205 | close (backend_fd); |
270 | close (backend_fd); |
206 | |
271 | |
207 | while ((backend_fd = epoll_create (256)) < 0) |
272 | while ((backend_fd = epoll_create (256)) < 0) |
208 | syserr ("(libev) epoll_create"); |
273 | ev_syserr ("(libev) epoll_create"); |
209 | |
274 | |
210 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
275 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
211 | |
276 | |
212 | fd_rearm_all (EV_A); |
277 | fd_rearm_all (EV_A); |
213 | } |
278 | } |