ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_epoll.c
(Generate patch)

Comparing libev/ev_epoll.c (file contents):
Revision 1.35 by root, Thu Oct 23 04:56:49 2008 UTC vs.
Revision 1.56 by root, Wed Jan 5 04:25:12 2011 UTC

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 *
… …
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. 61 * epoll_ctl syscalls for common usage patterns and handle the breakage
62 * ensuing from receiving events for closed and otherwise long gone
63 * file descriptors.
56 */ 64 */
57 65
58#include <sys/epoll.h> 66#include <sys/epoll.h>
59 67
60void inline_size 68#define EV_EMASK_EPERM 0x80
61unsigned_char_init (unsigned char *base, int count)
62{
63 /* memset might be overkill */
64 while (count--)
65 *base++ = 0;
66}
67 69
68static void 70static void
69epoll_modify (EV_P_ int fd, int oev, int nev) 71epoll_modify (EV_P_ int fd, int oev, int nev)
70{ 72{
71 struct epoll_event ev; 73 struct epoll_event ev;
… …
74 /* 76 /*
75 * we handle EPOLL_CTL_DEL by ignoring it here 77 * we handle EPOLL_CTL_DEL by ignoring it here
76 * on the assumption that the fd is gone anyways 78 * on the assumption that the fd is gone anyways
77 * if that is wrong, we have to handle the spurious 79 * if that is wrong, we have to handle the spurious
78 * event in epoll_poll. 80 * event in epoll_poll.
79 * 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
80 * fails, we assume it still has the same eventmask. 82 * fails, we assume it still has the same eventmask.
81 */ 83 */
82 if (!nev) 84 if (!nev)
83 return; 85 return;
84 86
85 oldmask = anfds [fd].emask; 87 oldmask = anfds [fd].emask;
86 anfds [fd].emask = nev; 88 anfds [fd].emask = nev;
87 89
88 ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */ 90 /* store the generation counter in the upper 32 bits, the fd in the lower 32 bits */
91 ev.data.u64 = (uint64_t)(uint32_t)fd
92 | ((uint64_t)(uint32_t)++anfds [fd].egen << 32);
89 ev.events = (nev & EV_READ ? EPOLLIN : 0) 93 ev.events = (nev & EV_READ ? EPOLLIN : 0)
90 | (nev & EV_WRITE ? EPOLLOUT : 0); 94 | (nev & EV_WRITE ? EPOLLOUT : 0);
91 95
92 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)))
93 return; 97 return;
94 98
95 if (expect_true (errno == ENOENT)) 99 if (expect_true (errno == ENOENT))
96 { 100 {
97 /* 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 */
98 if (!nev) 102 if (!nev)
99 return; 103 goto dec_egen;
100 104
101 if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev)) 105 if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev))
102 return; 106 return;
103 } 107 }
104 else if (expect_true (errno == EEXIST)) 108 else if (expect_true (errno == EEXIST))
105 { 109 {
106 /* EEXIST means we ignored a previous DEL, but the fd is still active */ 110 /* 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 */ 111 /* if the kernel mask is the same as the new mask, we assume it hasn't changed */
112 if (oldmask == nev)
113 goto dec_egen;
114
108 if (oldmask == nev || !epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) 115 if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev))
109 return; 116 return;
110 } 117 }
118 else if (expect_true (errno == EPERM))
119 {
120 anfds [fd].emask = EV_EMASK_EPERM;
121
122 /* add fd to epoll_eperms, if not already inside */
123 if (!(oldmask & EV_EMASK_EPERM))
124 {
125 array_needsize (int, epoll_eperms, epoll_epermmax, epoll_epermcnt + 1, EMPTY2);
126 epoll_eperms [epoll_epermcnt++] = fd;
127 }
128
129 return;
130 }
111 131
112 fd_kill (EV_A_ fd); 132 fd_kill (EV_A_ fd);
133
134dec_egen:
135 /* we didn't successfully call epoll_ctl, so decrement the generation counter again */
136 --anfds [fd].egen;
113} 137}
114 138
115static void 139static void
116epoll_poll (EV_P_ ev_tstamp timeout) 140epoll_poll (EV_P_ ev_tstamp timeout)
117{ 141{
118 int i; 142 int i;
143 int eventcnt;
144
145 /* epoll wait times cannot be larger than (LONG_MAX - 999UL) / HZ msecs, which is below */
146 /* the default libev max wait time, however. */
147 EV_RELEASE_CB;
119 int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); 148 eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.));
149 EV_ACQUIRE_CB;
120 150
121 if (expect_false (eventcnt < 0)) 151 if (expect_false (eventcnt < 0))
122 { 152 {
123 if (errno != EINTR) 153 if (errno != EINTR)
124 syserr ("(libev) epoll_wait"); 154 ev_syserr ("(libev) epoll_wait");
125 155
126 return; 156 return;
127 } 157 }
128 158
129 for (i = 0; i < eventcnt; ++i) 159 for (i = 0; i < eventcnt; ++i)
130 { 160 {
131 struct epoll_event *ev = epoll_events + i; 161 struct epoll_event *ev = epoll_events + i;
132 162
133 int fd = ev->data.u64; 163 int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */
164 int want = anfds [fd].events;
134 int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) 165 int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
135 | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); 166 | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0);
136 int want = anfds [fd].events; 167
168 /* check for spurious notification */
169 /* we assume that fd is always in range, as we never shrink the anfds array */
170 if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32)))
171 {
172 /* recreate kernel state */
173 postfork = 1;
174 continue;
175 }
137 176
138 if (expect_false (got & ~want)) 177 if (expect_false (got & ~want))
139 { 178 {
140 anfds [fd].emask = want; 179 anfds [fd].emask = want;
141 180
142 /* we received an event but are not interested in it, try mod or del */ 181 /* 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 */ 182 /* I don't think we ever need MOD, but let's handle it anyways */
144 ev->events = (want & EV_READ ? EPOLLIN : 0) 183 ev->events = (want & EV_READ ? EPOLLIN : 0)
145 | (want & EV_WRITE ? EPOLLOUT : 0); 184 | (want & EV_WRITE ? EPOLLOUT : 0);
146 185
186 /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */
187 /* which is fortunately easy to do for us. */
147 epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev); 188 if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev))
189 {
190 postfork = 1; /* an error occurred, recreate kernel state */
191 continue;
192 }
148 } 193 }
149 194
150 fd_event (EV_A_ fd, got); 195 fd_event (EV_A_ fd, got);
151 } 196 }
152 197
… …
155 { 200 {
156 ev_free (epoll_events); 201 ev_free (epoll_events);
157 epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1); 202 epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1);
158 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); 203 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax);
159 } 204 }
205
206 /* now synthesize events for all fds where epoll fails, while select works... */
207 for (i = epoll_epermcnt; i--; )
208 {
209 int fd = epoll_eperms [i];
210 unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE);
211
212 if (anfds [fd].emask & EV_EMASK_EPERM && events)
213 fd_event (EV_A_ fd, events);
214 else
215 epoll_eperms [i] = epoll_eperms [--epoll_epermcnt];
216 }
160} 217}
161 218
162int inline_size 219int inline_size
163epoll_init (EV_P_ int flags) 220epoll_init (EV_P_ int flags)
164{ 221{
222#ifdef EPOLL_CLOEXEC
223 backend_fd = epoll_create1 (EPOLL_CLOEXEC);
224
225 if (backend_fd <= 0)
226#endif
165 backend_fd = epoll_create (256); 227 backend_fd = epoll_create (256);
166 228
167 if (backend_fd < 0) 229 if (backend_fd < 0)
168 return 0; 230 return 0;
169 231
170 fcntl (backend_fd, F_SETFD, FD_CLOEXEC); 232 fcntl (backend_fd, F_SETFD, FD_CLOEXEC);
171 233
172 backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ 234 backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */
173 backend_modify = epoll_modify; 235 backend_modify = epoll_modify;
174 backend_poll = epoll_poll; 236 backend_poll = epoll_poll;
175 237
176 epoll_eventmax = 64; /* intiial number of events receivable per poll */ 238 epoll_eventmax = 64; /* initial number of events receivable per poll */
177 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); 239 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax);
178 240
179 return EVBACKEND_EPOLL; 241 return EVBACKEND_EPOLL;
180} 242}
181 243
182void inline_size 244void inline_size
183epoll_destroy (EV_P) 245epoll_destroy (EV_P)
184{ 246{
185 ev_free (epoll_events); 247 ev_free (epoll_events);
248 array_free (epoll_eperm, EMPTY);
186} 249}
187 250
188void inline_size 251void inline_size
189epoll_fork (EV_P) 252epoll_fork (EV_P)
190{ 253{
191 close (backend_fd); 254 close (backend_fd);
192 255
193 while ((backend_fd = epoll_create (256)) < 0) 256 while ((backend_fd = epoll_create (256)) < 0)
194 syserr ("(libev) epoll_create"); 257 ev_syserr ("(libev) epoll_create");
195 258
196 fcntl (backend_fd, F_SETFD, FD_CLOEXEC); 259 fcntl (backend_fd, F_SETFD, FD_CLOEXEC);
197 260
198 fd_rearm_all (EV_A); 261 fd_rearm_all (EV_A);
199} 262}

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines