On Tue, 2013-02-12 at 22:55 +0100, Samuel Thibault wrote: > Svante Signell, le Tue 12 Feb 2013 09:09:25 +0100, a écrit : > > Attached is a patch wrt the latest patch (step2) on January 24 2013. The > > indentation is different, therefore the diff blocks are a little large > > (but easier to read). Hopefully this is better. Including also the > > resulting file hurdselect_step7x.c. Explanations later in a separate > > mail. > > Explanations don't need to be very long. All we need is the rationale > and the way you handle it in the patch. > > AIUI, there are three things: > > - pass 1ms as timeout. This will be fixed by Richard's patch. > - in the poll() case, do not fail completely if an FD is bogus. Instead, > poll on the remaining FDs, and later on set POLLNVAL. It seems your > patch does this by using the i_index array indirection. > - in the poll() case, on error returned by io_select, instead of setting > readiness, set POLLHUP or POLLERR according to the error. > > Is that right?
Yes, mainly. Here it goes: - introduce the dfd struct, to be instantiated later in _hurd_select() - introduce two helper functions, _io_select_request() and _wait_for replies() - in _io_select_request(): * this function calls __io_select() * store the errors EPIPE and EPOLL for poll in errvec[nfds] to fill in the revents field of the struct pollfd later. * return with value -1 for other errors. * add a delay of 1ms in the call to __io_select() for POLL to account for the round-trip delay (will be fixed by the changes by Richard to move time-outs to the server) - in _wait_for_replies(): * move all union and MDGID definitions here, since they are local to this routine. * this function calls __mach_msg() - in _hurd_select(): * split the code into two cases: POLL and SELECT * change the FD_SETSIZE upper value check to larger than or equal from larger than. (from POSIX definition of select: http://pubs.opengroup.org/onlinepubs/009604499/functions/select.html ) * call the helper functions for both cases. - for the POLL case: * mark all FDs in i_index[nfds] with -1 to start. * fill in the corresponding revents according to if the FDs are: good, broken (EPIPE or EIO) or bad i_index[] = -1. * fix a bug causing test_poll in python2.6 to fail: Clear out all revents fields before fiilng in. The caller might have set the fd and events correctly with crap in the revents field. * remove the i < nfds check, it is redundant since all nfds FDs are walked through. - for the SELECT case * exit with EBADF if nfds are larger than _hurd_dtablesize (fixes one test case in test-select from gnulib. - the POLL changes works with the Hurd built on January 28, see test code to be sent in a forthcoming mail. - some minor tweaks will still be made to be mor POSIX consistent. Svante
