On Sun, 2013-01-13 at 21:31 +0100, Samuel Thibault wrote: > Svante Signell, le Sun 13 Jan 2013 21:18:07 +0100, a écrit : > > > - In the poll case, on error on one descriptor, the loop shall mark > > > POLLERR and continue, not fail altogether. > > > > > > Is that right? > > > > The current code errors out if it finds one FD not OK, which is a bug. > > Irrespective of how many FDs there are, no checks are made of the input > > data given. > > Sure, but what you refer to is only one part of the story. EBADF is one > of the "error"s I mention above. Errors returned by translators also > need the same fix.
One problem is already fixed by me and you for pflocal. Remains to fix the corresponding issues for pfinet. (and maybe other ones too). > > > > - Hopefully the code is more readable now, without causing regressions. > > > > > > But it duplicates things. As I told you, we might still keep the code > > > as-is, and just apply your poll fixes over it. The 3-split still being > > > useful for reviewing. > > > > Well, slitting into three parts made the poll case faster > > Uh? Why? I don't know, I haven't analyzed the differences, just observed them. > > the poll case _is_ different from select, see the POSIX specs. > > In a very limited way only. I would say that even splitting more could be usable, e.g. creating a hurdpoll.c in addition to hurdselect.c (as well as adding io_poll server code of course) A comment: You removed my comments about readability and maintainability of if-then-else constructs vs case oriented code. We seem to have different views on code constructs.
