--- libev/ev_epoll.c 2008/10/27 13:39:18 1.37 +++ libev/ev_epoll.c 2012/06/20 12:39:53 1.67 @@ -1,19 +1,19 @@ /* * libev epoll fd activity backend * - * Copyright (c) 2007,2008 Marc Alexander Lehmann + * Copyright (c) 2007,2008,2009,2010,2011 Marc Alexander Lehmann * All rights reserved. * * Redistribution and use in source and binary forms, with or without modifica- * tion, are permitted provided that the following conditions are met: - * + * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. - * + * * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. - * + * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO @@ -49,6 +49,12 @@ * c) the inability to handle fork or file descriptors (think dup) * limits the applicability over poll, so this is not a generic * poll replacement. + * d) epoll doesn't work the same as select with many file descriptors + * (such as files). while not critical, no other advanced interface + * seems to share this (rather non-unixy) limitation. + * e) epoll claims to be embeddable, but in practise you never get + * a ready event for the epoll fd (broken: <=2.6.26, working: >=2.6.32). + * f) epoll_ctl returning EPERM means the fd is always ready. * * lots of "weird code" and complication handling in this file is due * to these design problems with epoll, as we try very hard to avoid @@ -59,6 +65,8 @@ #include +#define EV_EMASK_EPERM 0x80 + static void epoll_modify (EV_P_ int fd, int oev, int nev) { @@ -70,7 +78,7 @@ * on the assumption that the fd is gone anyways * if that is wrong, we have to handle the spurious * event in epoll_poll. - * the fd is later added, we try to ADD it, and, if that + * if the fd is added again, we try to ADD it, and, if that * fails, we assume it still has the same eventmask. */ if (!nev) @@ -79,12 +87,13 @@ oldmask = anfds [fd].emask; anfds [fd].emask = nev; - /* store the generation counter in the upper 32 bits */ - ev.data.u64 = fd | ((uint64_t)++anfds [fd].egen << 32); + /* store the generation counter in the upper 32 bits, the fd in the lower 32 bits */ + ev.data.u64 = (uint64_t)(uint32_t)fd + | ((uint64_t)(uint32_t)++anfds [fd].egen << 32); ev.events = (nev & EV_READ ? EPOLLIN : 0) | (nev & EV_WRITE ? EPOLLOUT : 0); - if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) + if (expect_true (!epoll_ctl (backend_fd, oev && oldmask != nev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev))) return; if (expect_true (errno == ENOENT)) @@ -106,6 +115,21 @@ if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev)) return; } + else if (expect_true (errno == EPERM)) + { + /* EPERM means the fd is always ready, but epoll is too snobbish */ + /* to handle it, unlike select or poll. */ + anfds [fd].emask = EV_EMASK_EPERM; + + /* add fd to epoll_eperms, if not already inside */ + if (!(oldmask & EV_EMASK_EPERM)) + { + array_needsize (int, epoll_eperms, epoll_epermmax, epoll_epermcnt + 1, EMPTY2); + epoll_eperms [epoll_epermcnt++] = fd; + } + + return; + } fd_kill (EV_A_ fd); @@ -118,12 +142,21 @@ epoll_poll (EV_P_ ev_tstamp timeout) { int i; - int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.)); + int eventcnt; + + if (expect_false (epoll_epermcnt)) + timeout = 0.; + + /* epoll wait times cannot be larger than (LONG_MAX - 999UL) / HZ msecs, which is below */ + /* the default libev max wait time, however. */ + EV_RELEASE_CB; + eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, timeout * 1e3); + EV_ACQUIRE_CB; if (expect_false (eventcnt < 0)) { if (errno != EINTR) - syserr ("(libev) epoll_wait"); + ev_syserr ("(libev) epoll_wait"); return; } @@ -137,8 +170,13 @@ int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); - /* check for spurious notification */ - if (expect_false (anfds [fd].egen != (unsigned char)(ev->data.u64 >> 32))) + /* + * check for spurious notification. + * this only finds spurious notifications on egen updates + * other spurious notifications will be found by epoll_ctl, below + * we assume that fd is always in range, as we never shrink the anfds array + */ + if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) { /* recreate kernel state */ postfork = 1; @@ -149,14 +187,23 @@ { anfds [fd].emask = want; - /* we received an event but are not interested in it, try mod or del */ - /* I don't think we ever need MOD, but let's handle it anyways */ + /* + * we received an event but are not interested in it, try mod or del + * this often happens because we optimistically do not unregister fds + * when we are no longer interested in them, but also when we get spurious + * notifications for fds from another process. this is partially handled + * above with the gencounter check (== our fd is not the event fd), and + * partially here, when epoll_ctl returns an error (== a child has the fd + * but we closed it). + */ ev->events = (want & EV_READ ? EPOLLIN : 0) | (want & EV_WRITE ? EPOLLOUT : 0); + /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ + /* which is fortunately easy to do for us. */ if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) { - postfork = 1; /* an error occured, recreate kernel state */ + postfork = 1; /* an error occurred, recreate kernel state */ continue; } } @@ -171,21 +218,38 @@ epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1); epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); } + + /* now synthesize events for all fds where epoll fails, while select works... */ + for (i = epoll_epermcnt; i--; ) + { + int fd = epoll_eperms [i]; + unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE); + + if (anfds [fd].emask & EV_EMASK_EPERM && events) + fd_event (EV_A_ fd, events); + else + epoll_eperms [i] = epoll_eperms [--epoll_epermcnt]; + } } int inline_size epoll_init (EV_P_ int flags) { - backend_fd = epoll_create (256); +#ifdef EPOLL_CLOEXEC + backend_fd = epoll_create1 (EPOLL_CLOEXEC); + + if (backend_fd < 0 && (errno == EINVAL || errno == ENOSYS)) +#endif + backend_fd = epoll_create (256); if (backend_fd < 0) return 0; fcntl (backend_fd, F_SETFD, FD_CLOEXEC); - backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */ - backend_modify = epoll_modify; - backend_poll = epoll_poll; + backend_mintime = 1e-3; /* epoll does sometimes return early, this is just to avoid the worst */ + backend_modify = epoll_modify; + backend_poll = epoll_poll; epoll_eventmax = 64; /* initial number of events receivable per poll */ epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); @@ -197,6 +261,7 @@ epoll_destroy (EV_P) { ev_free (epoll_events); + array_free (epoll_eperm, EMPTY); } void inline_size @@ -205,7 +270,7 @@ close (backend_fd); while ((backend_fd = epoll_create (256)) < 0) - syserr ("(libev) epoll_create"); + ev_syserr ("(libev) epoll_create"); fcntl (backend_fd, F_SETFD, FD_CLOEXEC);