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 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 | * |
… | |
… | |
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. |
52 | * |
57 | * |
53 | * lots of "weird code" and complication handling in this file is due |
58 | * 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 |
59 | * to these design problems with epoll, as we try very hard to avoid |
55 | * epoll_ctl syscalls for common usage patterns. |
60 | * epoll_ctl syscalls for common usage patterns and handle the breakage |
|
|
61 | * ensuing from receiving events for closed and otherwise long gone |
|
|
62 | * file descriptors. |
56 | */ |
63 | */ |
57 | |
64 | |
58 | #include <sys/epoll.h> |
65 | #include <sys/epoll.h> |
59 | |
|
|
60 | void inline_size |
|
|
61 | unsigned_char_init (unsigned char *base, int count) |
|
|
62 | { |
|
|
63 | /* memset might be overkill */ |
|
|
64 | while (count--) |
|
|
65 | *base++ = 0; |
|
|
66 | } |
|
|
67 | |
66 | |
68 | static void |
67 | static void |
69 | epoll_modify (EV_P_ int fd, int oev, int nev) |
68 | epoll_modify (EV_P_ int fd, int oev, int nev) |
70 | { |
69 | { |
71 | struct epoll_event ev; |
70 | struct epoll_event ev; |
… | |
… | |
74 | /* |
73 | /* |
75 | * we handle EPOLL_CTL_DEL by ignoring it here |
74 | * we handle EPOLL_CTL_DEL by ignoring it here |
76 | * on the assumption that the fd is gone anyways |
75 | * on the assumption that the fd is gone anyways |
77 | * if that is wrong, we have to handle the spurious |
76 | * if that is wrong, we have to handle the spurious |
78 | * event in epoll_poll. |
77 | * event in epoll_poll. |
79 | * the fd is later added, we try to ADD it, and, if that |
78 | * if the fd is added again, we try to ADD it, and, if that |
80 | * fails, we assume it still has the same eventmask. |
79 | * fails, we assume it still has the same eventmask. |
81 | */ |
80 | */ |
82 | if (!nev) |
81 | if (!nev) |
83 | return; |
82 | return; |
84 | |
83 | |
85 | oldmask = anfds [fd].emask; |
84 | oldmask = anfds [fd].emask; |
86 | anfds [fd].emask = nev; |
85 | anfds [fd].emask = nev; |
87 | |
86 | |
88 | ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */ |
87 | /* store the generation counter in the upper 32 bits, the fd in the lower 32 bits */ |
|
|
88 | ev.data.u64 = (uint64_t)(uint32_t)fd |
|
|
89 | | ((uint64_t)(uint32_t)++anfds [fd].egen << 32); |
89 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
90 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
90 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
91 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
91 | |
92 | |
92 | if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
93 | if (expect_true (!epoll_ctl (backend_fd, oev && oldmask != nev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
93 | return; |
94 | return; |
94 | |
95 | |
95 | if (expect_true (errno == ENOENT)) |
96 | if (expect_true (errno == ENOENT)) |
96 | { |
97 | { |
97 | /* if ENOENT then the fd went away, so try to do the right thing */ |
98 | /* if ENOENT then the fd went away, so try to do the right thing */ |
98 | if (!nev) |
99 | if (!nev) |
99 | return; |
100 | goto dec_egen; |
100 | |
101 | |
101 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
102 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
102 | return; |
103 | return; |
103 | } |
104 | } |
104 | else if (expect_true (errno == EEXIST)) |
105 | else if (expect_true (errno == EEXIST)) |
105 | { |
106 | { |
106 | /* EEXIST means we ignored a previous DEL, but the fd is still active */ |
107 | /* EEXIST means we ignored a previous DEL, but the fd is still active */ |
107 | /* if the kernel mask is the same as the new mask, we assume it hasn't changed */ |
108 | /* if the kernel mask is the same as the new mask, we assume it hasn't changed */ |
|
|
109 | if (oldmask == nev) |
|
|
110 | goto dec_egen; |
|
|
111 | |
108 | if (oldmask == nev || !epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
112 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
109 | return; |
113 | return; |
110 | } |
114 | } |
111 | |
115 | |
112 | fd_kill (EV_A_ fd); |
116 | fd_kill (EV_A_ fd); |
|
|
117 | |
|
|
118 | dec_egen: |
|
|
119 | /* we didn't successfully call epoll_ctl, so decrement the generation counter again */ |
|
|
120 | --anfds [fd].egen; |
113 | } |
121 | } |
114 | |
122 | |
115 | static void |
123 | static void |
116 | epoll_poll (EV_P_ ev_tstamp timeout) |
124 | epoll_poll (EV_P_ ev_tstamp timeout) |
117 | { |
125 | { |
118 | int i; |
126 | int i; |
|
|
127 | int eventcnt; |
|
|
128 | |
|
|
129 | /* epoll wait times cannot be larger than (LONG_MAX - 999UL) / HZ msecs, which is below */ |
|
|
130 | /* the default libev max wait time, however. */ |
|
|
131 | EV_RELEASE_CB; |
119 | int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); |
132 | eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); |
|
|
133 | EV_ACQUIRE_CB; |
120 | |
134 | |
121 | if (expect_false (eventcnt < 0)) |
135 | if (expect_false (eventcnt < 0)) |
122 | { |
136 | { |
123 | if (errno != EINTR) |
137 | if (errno != EINTR) |
124 | syserr ("(libev) epoll_wait"); |
138 | ev_syserr ("(libev) epoll_wait"); |
125 | |
139 | |
126 | return; |
140 | return; |
127 | } |
141 | } |
128 | |
142 | |
129 | for (i = 0; i < eventcnt; ++i) |
143 | for (i = 0; i < eventcnt; ++i) |
130 | { |
144 | { |
131 | struct epoll_event *ev = epoll_events + i; |
145 | struct epoll_event *ev = epoll_events + i; |
132 | |
146 | |
133 | int fd = ev->data.u64; |
147 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
|
|
148 | int want = anfds [fd].events; |
134 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
149 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
135 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
150 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
136 | int want = anfds [fd].events; |
151 | |
|
|
152 | /* check for spurious notification */ |
|
|
153 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
|
|
154 | { |
|
|
155 | /* recreate kernel state */ |
|
|
156 | postfork = 1; |
|
|
157 | continue; |
|
|
158 | } |
137 | |
159 | |
138 | if (expect_false (got & ~want)) |
160 | if (expect_false (got & ~want)) |
139 | { |
161 | { |
140 | anfds [fd].emask = want; |
162 | anfds [fd].emask = want; |
141 | |
163 | |
142 | /* we received an event but are not interested in it, try mod or del */ |
164 | /* we received an event but are not interested in it, try mod or del */ |
143 | /* I don't think we ever need MOD, but let's handle it anyways */ |
165 | /* I don't think we ever need MOD, but let's handle it anyways */ |
144 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
166 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
145 | | (want & EV_WRITE ? EPOLLOUT : 0); |
167 | | (want & EV_WRITE ? EPOLLOUT : 0); |
146 | |
168 | |
|
|
169 | /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ |
|
|
170 | /* which is fortunately easy to do for us. */ |
147 | epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev); |
171 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
|
|
172 | { |
|
|
173 | postfork = 1; /* an error occurred, recreate kernel state */ |
|
|
174 | continue; |
|
|
175 | } |
148 | } |
176 | } |
149 | |
177 | |
150 | fd_event (EV_A_ fd, got); |
178 | fd_event (EV_A_ fd, got); |
151 | } |
179 | } |
152 | |
180 | |
… | |
… | |
160 | } |
188 | } |
161 | |
189 | |
162 | int inline_size |
190 | int inline_size |
163 | epoll_init (EV_P_ int flags) |
191 | epoll_init (EV_P_ int flags) |
164 | { |
192 | { |
|
|
193 | #ifdef EPOLL_CLOEXEC |
|
|
194 | backend_fd = epoll_create1 (EPOLL_CLOEXEC); |
|
|
195 | |
|
|
196 | if (backend_fd <= 0) |
|
|
197 | #endif |
165 | backend_fd = epoll_create (256); |
198 | backend_fd = epoll_create (256); |
166 | |
199 | |
167 | if (backend_fd < 0) |
200 | if (backend_fd < 0) |
168 | return 0; |
201 | return 0; |
169 | |
202 | |
170 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
203 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
171 | |
204 | |
172 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
205 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
173 | backend_modify = epoll_modify; |
206 | backend_modify = epoll_modify; |
174 | backend_poll = epoll_poll; |
207 | backend_poll = epoll_poll; |
175 | |
208 | |
176 | epoll_eventmax = 64; /* intiial number of events receivable per poll */ |
209 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
177 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
210 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
178 | |
211 | |
179 | return EVBACKEND_EPOLL; |
212 | return EVBACKEND_EPOLL; |
180 | } |
213 | } |
181 | |
214 | |
… | |
… | |
189 | epoll_fork (EV_P) |
222 | epoll_fork (EV_P) |
190 | { |
223 | { |
191 | close (backend_fd); |
224 | close (backend_fd); |
192 | |
225 | |
193 | while ((backend_fd = epoll_create (256)) < 0) |
226 | while ((backend_fd = epoll_create (256)) < 0) |
194 | syserr ("(libev) epoll_create"); |
227 | ev_syserr ("(libev) epoll_create"); |
195 | |
228 | |
196 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
229 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
197 | |
230 | |
198 | fd_rearm_all (EV_A); |
231 | fd_rearm_all (EV_A); |
199 | } |
232 | } |