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 1.49 by root, Tue Nov 24 06:37:23 2009 UTC vs.
Revision 1.57 by root, Sat Jan 8 10:25:16 2011 UTC

1/* 1/*
2 * libev epoll fd activity backend 2 * libev epoll fd activity backend
3 * 3 *
4 * Copyright (c) 2007,2008,2009 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 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>
67
68#define EV_EMASK_EPERM 0x80
61 69
62static void 70static void
63epoll_modify (EV_P_ int fd, int oev, int nev) 71epoll_modify (EV_P_ int fd, int oev, int nev)
64{ 72{
65 struct epoll_event ev; 73 struct epoll_event ev;
83 ev.data.u64 = (uint64_t)(uint32_t)fd 91 ev.data.u64 = (uint64_t)(uint32_t)fd
84 | ((uint64_t)(uint32_t)++anfds [fd].egen << 32); 92 | ((uint64_t)(uint32_t)++anfds [fd].egen << 32);
85 ev.events = (nev & EV_READ ? EPOLLIN : 0) 93 ev.events = (nev & EV_READ ? EPOLLIN : 0)
86 | (nev & EV_WRITE ? EPOLLOUT : 0); 94 | (nev & EV_WRITE ? EPOLLOUT : 0);
87 95
88 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)))
89 return; 97 return;
90 98
91 if (expect_true (errno == ENOENT)) 99 if (expect_true (errno == ENOENT))
92 { 100 {
93 /* 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 if (oldmask == nev) 112 if (oldmask == nev)
105 goto dec_egen; 113 goto dec_egen;
106 114
107 if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) 115 if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev))
108 return; 116 return;
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;
109 } 132 }
110 133
111 fd_kill (EV_A_ fd); 134 fd_kill (EV_A_ fd);
112 135
113dec_egen: 136dec_egen:
143 int want = anfds [fd].events; 166 int want = anfds [fd].events;
144 int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) 167 int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
145 | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); 168 | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0);
146 169
147 /* check for spurious notification */ 170 /* check for spurious notification */
171 /* we assume that fd is always in range, as we never shrink the anfds array */
148 if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) 172 if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32)))
149 { 173 {
150 /* recreate kernel state */ 174 /* recreate kernel state */
151 postfork = 1; 175 postfork = 1;
152 continue; 176 continue;
163 187
164 /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ 188 /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */
165 /* which is fortunately easy to do for us. */ 189 /* which is fortunately easy to do for us. */
166 if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) 190 if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev))
167 { 191 {
168 postfork = 1; /* an error occured, recreate kernel state */ 192 postfork = 1; /* an error occurred, recreate kernel state */
169 continue; 193 continue;
170 } 194 }
171 } 195 }
172 196
173 fd_event (EV_A_ fd, got); 197 fd_event (EV_A_ fd, got);
178 { 202 {
179 ev_free (epoll_events); 203 ev_free (epoll_events);
180 epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1); 204 epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1);
181 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); 205 epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax);
182 } 206 }
207
208 /* now synthesize events for all fds where epoll fails, while select works... */
209 for (i = epoll_epermcnt; i--; )
210 {
211 int fd = epoll_eperms [i];
212 unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE);
213
214 if (anfds [fd].emask & EV_EMASK_EPERM && events)
215 fd_event (EV_A_ fd, events);
216 else
217 epoll_eperms [i] = epoll_eperms [--epoll_epermcnt];
218 }
183} 219}
184 220
185int inline_size 221int inline_size
186epoll_init (EV_P_ int flags) 222epoll_init (EV_P_ int flags)
187{ 223{
209 245
210void inline_size 246void inline_size
211epoll_destroy (EV_P) 247epoll_destroy (EV_P)
212{ 248{
213 ev_free (epoll_events); 249 ev_free (epoll_events);
250 array_free (epoll_eperm, EMPTY);
214} 251}
215 252
216void inline_size 253void inline_size
217epoll_fork (EV_P) 254epoll_fork (EV_P)
218{ 255{

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines