1 | /* |
1 | /* |
2 | * libev epoll fd activity backend |
2 | * libev epoll fd activity backend |
3 | * |
3 | * |
4 | * Copyright (c) 2007 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 | |
66 | |
60 | static void |
67 | static void |
61 | epoll_modify (EV_P_ int fd, int oev, int nev) |
68 | epoll_modify (EV_P_ int fd, int oev, int nev) |
62 | { |
69 | { |
63 | struct epoll_event ev; |
70 | struct epoll_event ev; |
|
|
71 | unsigned char oldmask; |
64 | |
72 | |
65 | /* |
73 | /* |
66 | * we handle EPOLL_CTL_DEL by ignoring it here |
74 | * we handle EPOLL_CTL_DEL by ignoring it here |
67 | * on the assumption that the fd is gone anyways |
75 | * on the assumption that the fd is gone anyways |
68 | * if that is wrong, we have to handle the spurious |
76 | * if that is wrong, we have to handle the spurious |
69 | * event in epoll_poll. |
77 | * event in epoll_poll. |
|
|
78 | * if the fd is added again, we try to ADD it, and, if that |
|
|
79 | * fails, we assume it still has the same eventmask. |
70 | */ |
80 | */ |
71 | if (!nev) |
81 | if (!nev) |
72 | return; |
82 | return; |
73 | |
83 | |
74 | ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */ |
84 | oldmask = anfds [fd].emask; |
|
|
85 | anfds [fd].emask = nev; |
|
|
86 | |
|
|
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); |
75 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
90 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
76 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
91 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
77 | |
92 | |
78 | 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))) |
79 | return; |
94 | return; |
80 | |
95 | |
81 | if (expect_true (errno == ENOENT)) |
96 | if (expect_true (errno == ENOENT)) |
82 | { |
97 | { |
83 | /* on ENOENT 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 */ |
84 | if (!nev) |
99 | if (!nev) |
85 | return; |
100 | goto dec_egen; |
86 | |
101 | |
87 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
102 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
88 | return; |
103 | return; |
89 | } |
104 | } |
90 | else if (expect_true (errno == EEXIST)) |
105 | else if (expect_true (errno == EEXIST)) |
91 | { |
106 | { |
92 | /* on EEXIST we ignored a previous DEL */ |
107 | /* EEXIST means we ignored a previous DEL, but the fd is still active */ |
|
|
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 | |
93 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
112 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
94 | return; |
113 | return; |
95 | } |
114 | } |
96 | |
115 | |
97 | 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; |
98 | } |
121 | } |
99 | |
122 | |
100 | static void |
123 | static void |
101 | epoll_poll (EV_P_ ev_tstamp timeout) |
124 | epoll_poll (EV_P_ ev_tstamp timeout) |
102 | { |
125 | { |
103 | 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; |
104 | 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; |
105 | |
134 | |
106 | if (expect_false (eventcnt < 0)) |
135 | if (expect_false (eventcnt < 0)) |
107 | { |
136 | { |
108 | if (errno != EINTR) |
137 | if (errno != EINTR) |
109 | syserr ("(libev) epoll_wait"); |
138 | ev_syserr ("(libev) epoll_wait"); |
110 | |
139 | |
111 | return; |
140 | return; |
112 | } |
141 | } |
113 | |
142 | |
114 | for (i = 0; i < eventcnt; ++i) |
143 | for (i = 0; i < eventcnt; ++i) |
115 | { |
144 | { |
116 | struct epoll_event *ev = epoll_events + i; |
145 | struct epoll_event *ev = epoll_events + i; |
117 | |
146 | |
118 | 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; |
119 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
149 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
120 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
150 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
121 | int want = anfds [fd].events; |
151 | |
|
|
152 | /* check for spurious notification */ |
|
|
153 | /* we assume that fd is always in range, as we never shrink the anfds array */ |
|
|
154 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
|
|
155 | { |
|
|
156 | /* recreate kernel state */ |
|
|
157 | postfork = 1; |
|
|
158 | continue; |
|
|
159 | } |
122 | |
160 | |
123 | if (expect_false (got & ~want)) |
161 | if (expect_false (got & ~want)) |
124 | { |
162 | { |
|
|
163 | anfds [fd].emask = want; |
|
|
164 | |
125 | /* we received an event but are not interested in it, try mod or del */ |
165 | /* we received an event but are not interested in it, try mod or del */ |
|
|
166 | /* I don't think we ever need MOD, but let's handle it anyways */ |
126 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
167 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
127 | | (want & EV_WRITE ? EPOLLOUT : 0); |
168 | | (want & EV_WRITE ? EPOLLOUT : 0); |
128 | |
169 | |
|
|
170 | /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ |
|
|
171 | /* which is fortunately easy to do for us. */ |
129 | epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev); |
172 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
|
|
173 | { |
|
|
174 | postfork = 1; /* an error occurred, recreate kernel state */ |
|
|
175 | continue; |
|
|
176 | } |
130 | } |
177 | } |
131 | |
178 | |
132 | fd_event (EV_A_ fd, got); |
179 | fd_event (EV_A_ fd, got); |
133 | } |
180 | } |
134 | |
181 | |
… | |
… | |
142 | } |
189 | } |
143 | |
190 | |
144 | int inline_size |
191 | int inline_size |
145 | epoll_init (EV_P_ int flags) |
192 | epoll_init (EV_P_ int flags) |
146 | { |
193 | { |
|
|
194 | #ifdef EPOLL_CLOEXEC |
|
|
195 | backend_fd = epoll_create1 (EPOLL_CLOEXEC); |
|
|
196 | |
|
|
197 | if (backend_fd <= 0) |
|
|
198 | #endif |
147 | backend_fd = epoll_create (256); |
199 | backend_fd = epoll_create (256); |
148 | |
200 | |
149 | if (backend_fd < 0) |
201 | if (backend_fd < 0) |
150 | return 0; |
202 | return 0; |
151 | |
203 | |
152 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
204 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
153 | |
205 | |
154 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
206 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
155 | backend_modify = epoll_modify; |
207 | backend_modify = epoll_modify; |
156 | backend_poll = epoll_poll; |
208 | backend_poll = epoll_poll; |
157 | |
209 | |
158 | epoll_eventmax = 64; /* intiial number of events receivable per poll */ |
210 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
159 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
211 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
160 | |
212 | |
161 | return EVBACKEND_EPOLL; |
213 | return EVBACKEND_EPOLL; |
162 | } |
214 | } |
163 | |
215 | |
… | |
… | |
171 | epoll_fork (EV_P) |
223 | epoll_fork (EV_P) |
172 | { |
224 | { |
173 | close (backend_fd); |
225 | close (backend_fd); |
174 | |
226 | |
175 | while ((backend_fd = epoll_create (256)) < 0) |
227 | while ((backend_fd = epoll_create (256)) < 0) |
176 | syserr ("(libev) epoll_create"); |
228 | ev_syserr ("(libev) epoll_create"); |
177 | |
229 | |
178 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
230 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
179 | |
231 | |
180 | fd_rearm_all (EV_A); |
232 | fd_rearm_all (EV_A); |
181 | } |
233 | } |