Re: [PATCH v7 21/27] udp: Handle "spliced" datagrams with per-flow sockets

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 3168 bytes --]

On Wed, Jul 10, 2024 at 07:13:26PM +0200, Stefano Brivio wrote:
> On Wed, 10 Jul 2024 10:23:14 +1000
> David Gibson <david@gibson.dropbear.id.au> wrote:
> 
> > On Wed, Jul 10, 2024 at 12:32:33AM +0200, Stefano Brivio wrote:
> > > On Fri,  5 Jul 2024 12:07:18 +1000
> > > David Gibson <david@gibson.dropbear.id.au> wrote:
[snip]
> > > > +		uflow->s[INISIDE] = fcntl(s_ini, F_DUPFD_CLOEXEC, 0);  
> > > 
> > > There's one aspect of this I don't understand: if s_ini is closed while
> > > checking for bound ports (is it? I didn't really reach the end of this
> > > series), aren't duplicates also closed?
> > > 
> > > That is, the documentation of dup2(2), which should be the same for
> > > this purpose, states that the duplicate inherits "file status flags",
> > > which I would assume also includes the fact that a socket is closed. I
> > > didn't test that though.  
> > 
> > I don't believe so.  My understanding is that dup() (and the rest)
> > make a new fd referencing the same underlying file object, yes.  But
> > AIUI, close() just closes one fd - the underlying object is only
> > closed only when all fds are gone.
> 
> Ah, probably, yes.
> 
> > > If duplicates are closed, I guess an alternative solution could be to
> > > introduce some kind of reference counting for sockets... somewhere.  
> > 
> > .. in other words, I believe the kernel does the reference counting.
> > 
> > I should verify this though, I'll try to come up with something new
> > for doc/platform-requirements.
> 
> I didn't really find the time to sketch this but I guess the easiest
> way to check this behaviour is to have a TCP connection between a
> socket pair, with one socket having two descriptors, then closing one
> descriptor and check if the peer socket sees a closed connection
> (recv() returning 0 or similar).

So.. yes, this would check whether close() on a non-last fd for a
socket triggers socket closing actions, but that's much stricter than
what we actually need here.  I would, for example, expect shutdown()
on a TCP socket to affect all dups - and I don't actually know if a
close() on one dup might trigger that.

But we're dealing with UDP here, so there's no "on wire" effect of a
close.  So all we actually need to check is:
   1. Open a "listening" udp socket
   2. Dup it
   3. Close a dup
   4. Can the remaining dup still receive datagrams?

I've written a test program for this, which I'll include in the next
spin.

> I was wondering whether it's worth to use the vforked namespaced peer
> trick I drafted here:
>   https://archives.passt.top/passt-dev/20231206160808.3d312733@elisabeth/
> 
> just in case we want to use some of those test cases for actual tests,
> where we don't want to bind an actual TCP port on the machine we're
> running on. But if it adds complexity I'd say it's not worth it.

Yeah, I tend to think we can add that sort of sandboxing when and if
we need it.

-- 
David Gibson (he or they)	| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you, not the other way
				| around.
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]