[systemd-devel] [PATCH 00/11] Finalize initial DHCP support

[Patrik Flykt]

On Fri, 2013-12-20 at 09:47 -0800, Marcel Holtmann wrote:
> Hi Tom,
> 
> >>> The first seven patches fix a few issues with the current code.
> >>> 
> >>> Patch 09 adds DHCP lease renewing support when timer T1 triggers.
> Using
> >>> the UDP socket sending implementation in patch 08, the DHCP lease
> >>> renewal is unicasted to the DHCP server. This means that
> systemd-network
> >>> should have applied the acquired IP address and default route to
> the
> >>> proper interface before timer T1 triggers.
> >> 
> >> this could become racy and we might end up in funny cases if the
> lease is really small. I think networkd and the DHCP need some way of
> communicating a) I set the IP you told me and/or b) we have T1
> triggering, have you set the address or should I just redo the DHCP
> process.
> > 
> > Makes sense to me for networkd to call (something like)
> > sd_dhcp_client_address_configured(client, true) whenever it has
> > successfully set the addresses/routes. I.e., I'd go with option (a).
> > Or is there a reason to go with (b) that I'm not seeing?
> 
> the case I see is that T1 is triggered, but the IP address is not set.
> Then of course we should not be setting it anymore since it might not
> stay valid.

Leases with lifetime less than 10 seconds are not accepted by the code
currently. With the default T1 of half lease time, this gives 5 seconds
to react. Now the funny part is that the server can suggest other values
for T1 and T2 that the code will use, so yes, the server can try to
suggest a 1s T1 expiry time.

If the address is not set at the time of T1 expiring, it is treated as a
temprorary error and T1 retry is rescheduled with a minimum time of 60
seconds, this from section 4.4.5 in RFC 2131. Thus both T2 and the whole
lifetime will expire before that with a short lease < 60s. If T2 is
reached, the DHCP Request is broadcasted anyway, so if the address is
not set even at this point, a new DHCP Request is still sent.

The somewhat fuzzy idea was to use T1 from the server if given and to
set the next T1 timeout to the minimum 60 seconds proposed by the RFC if
the previous reacquisition fails.

With this the code should still work nicely with short lease timeouts.
But more testing with really short lease times are anyway appropriate.

Cheers,

	Patrik