fcntl()

[Silvan Scherrer]

setting the O_NonBlock flag seems to work, but clearing the flag seems not to work. we found this problem in samba. as samba after clearing the flag still produced EAGAIN. now we changed the fcntl() with os2_ioctl() and the flags are set properly.

[Yuri Dario]

it seems that os/2 ioctl FIONBIO reversed data parameter, e.g. data==0 set non-blocking status.

[bird]

(In [3708]) nonblock-1.c: extended the testcase. References #210.

[bird]

Not able to reproduce the behavior with the testcase using either method: source:trunk/libc/tests/libc/smoketests/nonblock-1.c​

[bird]

I've verified the correctness of the testcase on mac os x. So, it's either some more complicated thing or a bug in the samba port.

[Yuri Dario]

I had to add similar code to exim, courier-imap and tcl to get sockets working correctly.

[Silvan Scherrer]

i guess it's more complicated, as also in samba it did not happen every time. and usually it happened after some hours/days of samba running.

[Silvan Scherrer]

i just found another evidence, that fcntl() is not clearing the O_NONBLOCK right. in samba the socket flags where retrieved with fcntl(), set to nonblocking and later is restored the saved flags with fcntl(). now when the client wanted to send some stuff larger than around 33kb we got a WOULDBLOCK back and not all bytes where sent with writev(). after setting the socket with ioctl() to nonblocking all worked as it should. so i guess with small chunks of date it always works, but with larger the problem can be seen. at least in samba we could reproduce the same behaviour on different installations.

[Silvan Scherrer]

​http://svn.netlabs.org/samba/changeset/668 fixed the problem found in comment:8