unix/dgram: use minimal possible socket buffer for PF_UNIX/SOCK_DGRAM

[glebius]

Summary

This change fully splits away PF_UNIX/SOCK_DGRAM from other socket
buffer implementations, without any behavior changes.

Generic socket implementation is reduced down to one STAILQ and very
little code.

[glebius]

Note: this patch is a step towards D35303.

[markj]

sys/kern/uipc_usrreq.c

1320

I'd prefer an explicit KASSERT which hints at why we expect m != NULL. Or restructure the code so that m == NULL will result in a page fault.

1409

I would make each of these cases set error and then break, then after the loop:

if (m != NULL) {
    SOCK_RECVBUF_UNLOCK(so);
    SOCK_IO_RECV_UNLOCK(so);
    return (error);
}

[glebius]

Improve KASSERT message.

[glebius]

sys/kern/uipc_usrreq.c

1409

Can you please look at the final code in D35303? With it the handling would be more comlex:

if (m == NULL && sb == NULL) {
        SOCK_RECVBUF_UNLOCK(so);
        SOCK_IO_RECV_UNLOCK(so);
        return (error)
}

If you agree with that, I can go forward and change.

[markj]

sys/kern/uipc_usrreq.c

1409

Hmm, maybe just ignore my suggestion then. We can do this kind of cleanup after the functional changes are done.

1425

I wish this code (and sbfree()) would check for wraparound with INVARIANTS on...

[glebius]

Rebase

[glebius]

Add pru_sopoll method. Fixes select(2).

[markj]

In D35300#802382, @glebius wrote:

Add pru_sopoll method. Fixes select(2).

Is there an existing test case which catches this?

sys/kern/uipc_usrreq.c

1516

Does so->so_rerror need to be checked also?

1517

You also checked for SBS_CNTRCVMORE below, when POLLINIGNEOF is not set. Comparing to sopoll_generic(), the check here looks wrong.

[glebius]

In D35300#802525, @markj wrote:

In D35300#802382, @glebius wrote:

Add pru_sopoll method. Fixes select(2).

Is there an existing test case which catches this?

I'm working on the event tests now. It seems I need to do substantial changes to this branch, cause there is also kevent and aio, who also just use bare sbavail().

[glebius]

There are two directions to go:

1. make protocol specific sbavail() & sbused() or soreadable() & sowriteable() and that would probably allow to leave socket poll/kevent/aio unmodified.
2. Make protocol specific so_poll, kevent filterops and aio ops.

I'm researching these two options. Any input much appreciated.

[glebius]

Abandon protocol specific select/poll for now, and instead just maintain
sb_acc + sb_ccc that generic socket code requires.

Now passes tests/kern/unix_dgram:event

[markj]

sys/kern/uipc_usrreq.c

1147

BTW, you might consider asserting on valid values for flags here. For example, I forgot that PRUS_NOTREADY cannot appear here since it's only used with stream sockets.

1266

[glebius]

sys/kern/uipc_usrreq.c

1147

I can do that in the changeset that adds uipc_sosend_dgram().

[glebius]

sys/kern/uipc_usrreq.c

1147

Nope. Those flags are not in the family of PRUS_* flags. Those are MSG_* flags. This is pru_sosend method, not pru_send. We already return error for MSG_OOB and ignore the rest. This is what original implementation does.

[glebius]

Fix unp_gc_scan().