Module: sip-router Branch: sr_3.0 Commit: ea37f13d610fdab887a256960e4a056dfca45f4a URL: http://git.sip-router.org/cgi-bin/gitweb.cgi/sip-router/?a=commit;h=ea37f13… Author: Andrei Pelinescu-Onciul <andrei(a)iptel.org> Committer: Andrei Pelinescu-Onciul <andrei(a)iptel.org> Date: Thu Jun 17 16:15:10 2010 +0200 io_wait: fix: check for EV_ERROR for kqueue() Re-enabled and enhanced the check for EV_ERROR when using kqueue (*bsd). This is needed to workaround errors reported by kqueue when trying to delete (EV_DELETE) an already closed FD (this can happen in the tcp code, where we try to avoid applying immediately changes in the set of watched FDs and instead collect them and apply them after all the current kqueue events are processed => in some corner case situations it's possible to try to delete the FD from kqueue after the fd was close()'ed). This fix will ignore EV_ERROR with data == EBADF. All the other errors will result in a POLLERR flag for the callback. It fixes crashes with *bsd under tcp stress tests (lots of very short lived connections). (cherry picked from commit 05c080a53f88babad6729f79015f555f53fdf957) --- io_wait.h | 55 ++++++++++++++++++++++++++++++++++++++++--------------- 1 files changed, 40 insertions(+), 15 deletions(-) diff --git a/io_wait.h b/io_wait.h index 0557c00..44ef60c 100644 --- a/io_wait.h +++ b/io_wait.h @@ -45,6 +45,7 @@ * 2007-11-29 support for write (POLLOUT); added io_watch_chg() (andrei) * 2008-02-04 POLLRDHUP & EPOLLRDHUP support (automatically enabled if POLLIN * is set) (andrei) + * 2010-06-17 re-enabled & enhanced the EV_ERROR for kqueue (andrei) */ @@ -1097,31 +1098,55 @@ again: r, n, h->kq_array[r].ident, (long)h->kq_array[r].udata, h->kq_array[r].flags); #endif -#if 0 - if (unlikely(h->kq_array[r].flags & EV_ERROR)){ - /* error in changes: we ignore it, it can be caused by - trying to remove an already closed fd: race between - adding something to the changes array, close() and - applying the changes */ - LOG(L_INFO, "INFO: io_wait_loop_kqueue: kevent error on " - "fd %ld: %s [%ld]\n", h->kq_array[r].ident, + if (unlikely((h->kq_array[r].flags & EV_ERROR) && + (h->kq_array[r].data == EBADF || + h->kq_array[r].udata == 0))){ + /* error in changes: we ignore it if it has to do with a + bad fd or update==0. It can be caused by trying to remove an + already closed fd: race between adding something to the + changes array, close() and applying the changes. + E.g. for ser tcp: tcp_main sends a fd to child fore reading + => deletes it from the watched fds => the changes array + will contain an EV_DELETE for it. Before the changes + are applied (they are at the end of the main io_wait loop, + after all the fd events were processed), a CON_ERR sent + to tcp_main by a sender (send fail) is processed and causes + the fd to be closed. When the changes are applied => + error for the EV_DELETE attempt of a closed fd. + */ + /* + example EV_ERROR for trying to delete a read watched fd, + that was already closed: + { + ident = 63, [fd] + filter = -1, [EVFILT_READ] + flags = 16384, [EV_ERROR] + fflags = 0, + data = 9, [errno = EBADF] + udata = 0x0 + } + */ + if (h->kq_array[r].data != EBADF) + LOG(L_INFO, "INFO: io_wait_loop_kqueue: kevent error on " + "fd %ld: %s [%ld]\n", (long)h->kq_array[r].ident, strerror(h->kq_array[r].data), (long)h->kq_array[r].data); - }else{ -#endif + }else{ fm=(struct fd_map*)h->kq_array[r].udata; if (likely(h->kq_array[r].filter==EVFILT_READ)){ - revents=POLLIN | - (((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP); + revents=POLLIN | + (((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP) | + (((int)!(h->kq_array[r].flags & EV_ERROR)-1)&POLLERR); while(fm->type && (fm->events & revents) && (handle_io(fm, revents, -1)>0) && repeat); }else if (h->kq_array[r].filter==EVFILT_WRITE){ - revents=POLLOUT | - (((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP); + revents=POLLOUT | + (((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP) | + (((int)!(h->kq_array[r].flags & EV_ERROR)-1)&POLLERR); while(fm->type && (fm->events & revents) && (handle_io(fm, revents, -1)>0) && repeat); } - /*} */ + } } error: return n;