Network device driver and blocking send.

Hello,

I've developed a Linux Network device driver.
For now, this driver only provides PF_PACKET protocol family socket.

The problem is that, even when in blocking mode, if there are no more space
available the "send" function returns -1 with error code ENOBUFS.

I don't think the problem comes from the test application as this very
simple application works fine on "eth0" network device.

I should have missed something.

Do you have any idea?

Regards.

Vincent

"Vincent" <(E-Mail Removed)> wrote in message
news:LyEsd.1830$(E-Mail Removed)...

> The problem is that, even when in blocking mode, if there are no more
> space
> available the "send" function returns -1 with error code ENOBUFS.

What would you like it to do? It can't block, because doing that would
require some place to store the information about which process is waiting
on what resource, and it's out of places to store things.

> I don't think the problem comes from the test application as this very
> simple application works fine on "eth0" network device.
>
> I should have missed something.
>
> Do you have any idea?

This is behavior you have to deal with. You may wish to 'usleep' and try
again.

DS

David Schwartz wrote:
>
> What would you like it to do? It can't block, because doing that would
> require some place to store the information about which process is waiting
> on what resource, and it's out of places to store things.

And why couldn't that be on the stack?

--
Kasper Dupont

Hi David,

"David Schwartz" <(E-Mail Removed)> a écrit dans le message de
news:cp0gjv$kmi$(E-Mail Removed)...
>
> "Vincent" <(E-Mail Removed)> wrote in message
> news:LyEsd.1830$(E-Mail Removed)...
>
> > The problem is that, even when in blocking mode, if there are no more
> > space
> > available the "send" function returns -1 with error code ENOBUFS.
>
> What would you like it to do? It can't block, because doing that would
> require some place to store the information about which process is waiting
> on what resource, and it's out of places to store things.

This could be done in the network stack. And actually I think this is done
there as it works when I use the "eth0" device.
I suspect I have made a mistake in my driver that prevent the stack properly
doing its job.

> > I don't think the problem comes from the test application as this very
> > simple application works fine on "eth0" network device.
> >
> > I should have missed something.
> >
> > Do you have any idea?
>
> This is behavior you have to deal with. You may wish to 'usleep' and
try
> again.

Not possible as I will use up to 512 sockets at the same time with something
like 40000 messages per second.

In fact I do need blocking mode to use this sockets with "kqueue" API.

Regards.

Vincent