sendto: don't clear transient errors for atomic protocols

[glebius]

Summary

The changeset 65572cade35 uncovered the fact that top layer of sendto(2)
would clear a transient error code if some data was copied out of uio.
The clearing of the error makes sense for non-atomic protocols, since
they have sent some data. The atomic protocols send all or nothing.

The current implementation of unix/dgram uses sosend_generic(), which
would always copyout and only then it may fail to deliver a message.
The sosend_dgram(), currently used by UDP only, also has same behavior.

Reported by: pho

[pho]

The patch fixed the issue I had observed.

[markj]

I wonder why we don't use sosend_dgram() for unix dgram sockets? I tried it and don't see any immediate problems. Though, the problem applies to SEQPACKET sockets too.

[glebius]

I wonder why we don't use sosend_dgram() for unix dgram sockets? I tried it and don't see any immediate problems. Though, the problem applies to SEQPACKET sockets too.

This question is on my table, too. UDP definitely can use sosend_dgram. I will get back to this question after more work on unix/dgram.

Thanks for noticing SEQPACKET. I believe the check for PR_ATOMIC will address it, too.