… | |
… | |
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 | * |
52 | * |
53 | * lots of "weird code" and complication handling in this file is due |
53 | * 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 |
54 | * to these design problems with epoll, as we try very hard to avoid |
55 | * epoll_ctl syscalls for common usage patterns. |
55 | * epoll_ctl syscalls for common usage patterns and handle the breakage |
|
|
56 | * ensuing from receiving events for closed and otherwise long gone |
|
|
57 | * file descriptors. |
56 | */ |
58 | */ |
57 | |
59 | |
58 | #include <sys/epoll.h> |
60 | #include <sys/epoll.h> |
59 | |
61 | |
60 | static void |
62 | static void |
61 | epoll_modify (EV_P_ int fd, int oev, int nev) |
63 | epoll_modify (EV_P_ int fd, int oev, int nev) |
62 | { |
64 | { |
63 | struct epoll_event ev; |
65 | struct epoll_event ev; |
|
|
66 | unsigned char oldmask; |
64 | |
67 | |
65 | /* |
68 | /* |
66 | * we handle EPOLL_CTL_DEL by ignoring it here |
69 | * we handle EPOLL_CTL_DEL by ignoring it here |
67 | * on the assumption that the fd is gone anyways |
70 | * on the assumption that the fd is gone anyways |
68 | * if that is wrong, we have to handle the spurious |
71 | * if that is wrong, we have to handle the spurious |
69 | * event in epoll_poll. |
72 | * event in epoll_poll. |
|
|
73 | * if the fd is added again, we try to ADD it, and, if that |
|
|
74 | * fails, we assume it still has the same eventmask. |
70 | */ |
75 | */ |
71 | if (!nev) |
76 | if (!nev) |
72 | return; |
77 | return; |
73 | |
78 | |
74 | ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */ |
79 | oldmask = anfds [fd].emask; |
|
|
80 | anfds [fd].emask = nev; |
|
|
81 | |
|
|
82 | /* store the generation counter in the upper 32 bits, the fd in the lower 32 bits */ |
|
|
83 | ev.data.u64 = (uint64_t)(uint32_t)fd |
|
|
84 | | ((uint64_t)(uint32_t)++anfds [fd].egen << 32); |
75 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
85 | ev.events = (nev & EV_READ ? EPOLLIN : 0) |
76 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
86 | | (nev & EV_WRITE ? EPOLLOUT : 0); |
77 | |
87 | |
78 | if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
88 | if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) |
79 | return; |
89 | return; |
80 | |
90 | |
81 | if (expect_true (errno == ENOENT)) |
91 | if (expect_true (errno == ENOENT)) |
82 | { |
92 | { |
83 | /* on ENOENT the fd went away, so try to do the right thing */ |
93 | /* if ENOENT then the fd went away, so try to do the right thing */ |
84 | if (!nev) |
94 | if (!nev) |
85 | return; |
95 | goto dec_egen; |
86 | |
96 | |
87 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
97 | if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) |
88 | return; |
98 | return; |
89 | } |
99 | } |
90 | else if (expect_true (errno == EEXIST)) |
100 | else if (expect_true (errno == EEXIST)) |
91 | { |
101 | { |
92 | /* on EEXIST we ignored a previous DEL */ |
102 | /* EEXIST means we ignored a previous DEL, but the fd is still active */ |
|
|
103 | /* if the kernel mask is the same as the new mask, we assume it hasn't changed */ |
|
|
104 | if (oldmask == nev) |
|
|
105 | goto dec_egen; |
|
|
106 | |
93 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
107 | if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) |
94 | return; |
108 | return; |
95 | } |
109 | } |
96 | |
110 | |
97 | fd_kill (EV_A_ fd); |
111 | fd_kill (EV_A_ fd); |
|
|
112 | |
|
|
113 | dec_egen: |
|
|
114 | /* we didn't successfully call epoll_ctl, so decrement the generation counter again */ |
|
|
115 | --anfds [fd].egen; |
98 | } |
116 | } |
99 | |
117 | |
100 | static void |
118 | static void |
101 | epoll_poll (EV_P_ ev_tstamp timeout) |
119 | epoll_poll (EV_P_ ev_tstamp timeout) |
102 | { |
120 | { |
103 | int i; |
121 | int i; |
|
|
122 | int eventcnt; |
|
|
123 | |
|
|
124 | if (expect_false (suspend_cb)) suspend_cb (EV_A); |
104 | int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); |
125 | eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); |
|
|
126 | if (expect_false (resume_cb)) resume_cb (EV_A); |
105 | |
127 | |
106 | if (expect_false (eventcnt < 0)) |
128 | if (expect_false (eventcnt < 0)) |
107 | { |
129 | { |
108 | if (errno != EINTR) |
130 | if (errno != EINTR) |
109 | syserr ("(libev) epoll_wait"); |
131 | ev_syserr ("(libev) epoll_wait"); |
110 | |
132 | |
111 | return; |
133 | return; |
112 | } |
134 | } |
113 | |
135 | |
114 | for (i = 0; i < eventcnt; ++i) |
136 | for (i = 0; i < eventcnt; ++i) |
115 | { |
137 | { |
116 | struct epoll_event *ev = epoll_events + i; |
138 | struct epoll_event *ev = epoll_events + i; |
117 | |
139 | |
118 | int fd = ev->data.u64; |
140 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
|
|
141 | int want = anfds [fd].events; |
119 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
142 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
120 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
143 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
121 | int want = anfds [fd].events; |
144 | |
|
|
145 | /* check for spurious notification */ |
|
|
146 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
|
|
147 | { |
|
|
148 | /* recreate kernel state */ |
|
|
149 | postfork = 1; |
|
|
150 | continue; |
|
|
151 | } |
122 | |
152 | |
123 | if (expect_false (got & ~want)) |
153 | if (expect_false (got & ~want)) |
124 | { |
154 | { |
|
|
155 | anfds [fd].emask = want; |
|
|
156 | |
125 | /* we received an event but are not interested in it, try mod or del */ |
157 | /* we received an event but are not interested in it, try mod or del */ |
|
|
158 | /* I don't think we ever need MOD, but let's handle it anyways */ |
126 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
159 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
127 | | (want & EV_WRITE ? EPOLLOUT : 0); |
160 | | (want & EV_WRITE ? EPOLLOUT : 0); |
128 | |
161 | |
129 | epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev); |
162 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
|
|
163 | { |
|
|
164 | postfork = 1; /* an error occured, recreate kernel state */ |
|
|
165 | continue; |
|
|
166 | } |
130 | } |
167 | } |
131 | |
168 | |
132 | fd_event (EV_A_ fd, got); |
169 | fd_event (EV_A_ fd, got); |
133 | } |
170 | } |
134 | |
171 | |
… | |
… | |
153 | |
190 | |
154 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
191 | backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ |
155 | backend_modify = epoll_modify; |
192 | backend_modify = epoll_modify; |
156 | backend_poll = epoll_poll; |
193 | backend_poll = epoll_poll; |
157 | |
194 | |
158 | epoll_eventmax = 64; /* intiial number of events receivable per poll */ |
195 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
159 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
196 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
160 | |
197 | |
161 | return EVBACKEND_EPOLL; |
198 | return EVBACKEND_EPOLL; |
162 | } |
199 | } |
163 | |
200 | |
… | |
… | |
171 | epoll_fork (EV_P) |
208 | epoll_fork (EV_P) |
172 | { |
209 | { |
173 | close (backend_fd); |
210 | close (backend_fd); |
174 | |
211 | |
175 | while ((backend_fd = epoll_create (256)) < 0) |
212 | while ((backend_fd = epoll_create (256)) < 0) |
176 | syserr ("(libev) epoll_create"); |
213 | ev_syserr ("(libev) epoll_create"); |
177 | |
214 | |
178 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
215 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
179 | |
216 | |
180 | fd_rearm_all (EV_A); |
217 | fd_rearm_all (EV_A); |
181 | } |
218 | } |