Re: accept after select for RTNet/TCP

[Per Oberg <pero@wolfram.com>]

----- Den 27 sep 2023, på kl 22:21, Florian Bezdeka florian.bezdeka@siemens.com skrev:

> Hi Per,

> On Wed, 2023-09-27 at 10:13 -0500, Per Oberg wrote:
> > Hi

>> I'm currently porting parts of an application that uses some TCP during startup
>> to Xenomai. I have a working example of a TCP server but i cannot make the
> > "select" part work.

> > If I do

> > fd = __COBALT(accept(list_fd, (struct sockaddr *) &client_addr, &client_len)))

> The usage of __COBALT() should not be necessary, the Xenomai function
> wrapping will take care of redirecting the call to the cobalt world.
> (Assuming you are using the LDFLAGS and CFLAGS delivered by xeno-
> config)

> That allows you to compile the same code - but with different flags -
> for both worlds without modifications.


>> without the select part it works. If I do a select before it returns "Invalid
> > argument".

> I can remember that I saw something similar in the past. If I recall
> correctly there was an ordering problem. It somehow crashed when the
> first connection request came in before the accept() call happened, or
> something in this direction. The socket state was somehow "corrupted".

> I should have it on one of my TODO lists... Unable to find it right
> now.

> > The "select" part:
> > ------------------------------------
> > fd_set in_fds;
> > FD_ZERO(&in_fds);
> > FD_SET(list_fd, &in_fds);
> > int selRet = __COBALT(select(list_fd + 1, &in_fds, 0, 0, 0));
> > (FD_ISSET(list_fd, &in_fds))
> > {
> > printf("list_fd (%i) is in FDSET\n", list_fd);
> > }
> > ------------------------------------

> > The program works when compiled for regular linux.
> > Is select for TCP RTNet a no no ?


> select() should work as expected.

> Can you share a complete reproducer? That would simplify things and
> increases the chance that someone is able to look into it.

> Best regards,
> Florian


> > Best Regards
> > Per Öberg

Hi Florian and thanks for answering

The explicit use __COBALT / __STD wrappers is for some of my code that sometimes go on a real time socket and sometimes on a regular socket (depending on the destination). I think I found an issue when leaving it out for some special cases. (But it might have been the missing pselect that was the real issue)

That said, I also use it to ensure that there are no missing syscalls when debugging. E.g. I think I recall that pselect was missing in the implementation and that became apparent when putting __COBALT in front of it. 

My issue is 100% reproducible (afaik) so no timing issues show up for me in the regard. 

I can certainly share the code. My example is an adapted version of a TCP Client/Server demo. I will put it on github if that would work?

A few other notes that came up during the porting:

1. It seems like accept does not like beeing called with null pointers to client address 

2. Accept returns the same socket number as the listen socket, and thus closes the socket when the connection closes. I guessed that this was a simplification deemed ok for real time. This effectively makes the TCP server single client because SOCK_REUSE is not supported, right? 

Best Regards
Per Öberg