Re: NFSv4 boot support?

[Sascha Hauer <sha@pengutronix.de>]

On Wed, Feb 28, 2024 at 02:50:58PM +0300, Antony Pavlov wrote:
> On Wed, 28 Feb 2024 10:20:27 +0100
> Sascha Hauer <sha@pengutronix.de> wrote:
> 
> 
> Hi Sascha!
> 
> > On Wed, Feb 28, 2024 at 10:26:15AM +0300, Antony Pavlov wrote:
> > > On Sat, 17 Feb 2024 09:51:02 +0100
> > > Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:
> > > 
> > > Hi All!
> > > 
> > > > Hello Antony,
> > > > 
> > > > On 05.02.24 10:59, Antony Pavlov wrote:
> > > > > On Wed, 31 Jan 2024 22:37:50 +0100
> > > > > Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:
> > > > > 
> > > > > Hi All!
> > > > > 
> > > > >> Hello Dan,
> > > > >>
> > > > >> On 31.01.24 22:03, Dan Shelton wrote:
> > > > >>> Hello!
> > > > >>>
> > > > >>> Does barebox support booting from a NFSv4 filesystem, e.g. boot from
> > > > >>> NFSv4 filesystem into a Linux NFSv4 netroot (diskless machine)?
> > > > >>
> > > > >> The barebox network stack only does UDP/IP. There have been attempts to
> > > > >> bring a TCP stack into barebox, but none have so far succeeded to
> > > > >> make it mainline. This is a hard requirement before we can consider
> > > > >> supporting NFSv4. I hope that lwIP could fill this gap in the future,
> > > > >> but no one is actively continuing this work as far as I am aware[1].
> > > > > 
> > > > > I have started integration on picotcp into barebox in 2015, see
> > > > >   https://lore.barebox.org/barebox/1436991230-14251-10-git-send-email-antonynpavlov@gmail.com/T/
> > > 
> > > Actually I made the first attempt to integrate picotcp into barebox in 2014,
> > > see http://lists.infradead.org/pipermail/barebox/2014-May/019243.html
> > > 
> > > 10 years is too long for this task :)
> > > 
> > > In the message http://lists.infradead.org/pipermail/barebox/2015-July/024244.html
> > > if I understand correctly Sascha asked me to keep network stuff
> > > users (tftp, nfs, netconsole) as intact as possible.
> > > 
> > > At the moment I understand that this task is too hard.
> > > 
> > > The problem is: the network stuff users don't rely on "a network stack"
> > > in the true sense. E.g. tftp_handler() takes an ETHERNET PACKET on
> > > it's input, tftp_handler() skips Ethernet and IP stuff by itself
> > > and modifies UDP fields directly!
> > > 
> > > This week I have connected picotcp code to the existing network code
> > > in the way that makes it possible to keep dhcp_handler() and
> > > ftp_handler() intact. The result is ugly: barebox netdevice driver
> > > receives frame from network, pass it to picotcp, picotcp parses
> > > network protocol headers and extracts udp payload, next
> > > picotcp passes udp payload back to my picotcp-to-barebox adapter,
> > > the adapter RECONSTRUCTS ETHERNET PACKET and give it to tftp/dhcp_handler()!
> > > This horrible approach creates more problems than it solves!
> 
> > So if I understand correctly you tried passing *all* incoming packets to
> > picotcp and route some of them back to the barebox network stack.
> 
> for every barebox network device I create picotcp network device
> which can send packets by picotcp initiative.
> net_receive() just calls pico_stack_recv() with corresponding
> picotcp network device as an argument, so no code change in drivers/net/.
> net_poll() just calls pico_stack_tick().
> So all incoming packets go to picotcp only.
> Picotcp can send packet if necessary without barebox interraction.
> 
> UDP datagrams processed by picotcp routed to corresponding dhcp/tftp_handler().
> Are dhcp/tftp_handlers parts of barebox network stack?
> 
> > Instead of passing all packets to picotcp, can't we just dispatch the
> > incoming packets on a per-port basis in barebox and only pass the ones
> > picotcp shall handle to picotcp?
> 
> Good idea. I have to try. We can reroute IPv6 traffic to picotcp in net_receive()
> and reroute TCP/IPv4 traffic to picotcp in net_handle_ip().

You could route all TCP to picotcp. Also everything that falls through
net_handle_udp() can be routed to picotctp as well.

> 
> > So basically a struct net_connection with the handler set to the picotcp
> > receive function?
> 
> net_connection handler is set to the normal unamended dhcp/tftp_handlers.
> 
> > That way it might be possible to have the barebox network stack and
> > picotcp in parallel and port the handlers over to pictotcp one by one.
> 
> But I see a problem here. If there is a ARP respond in the incoming traffic
> I have no idea there to send it. Both barebox network code and picotcp
> can make ARP request. I suppose can just route ARP respond to the both stacks.

Yes, sending it to both stacks should be fine. When a response is coming
in and nobody wait for it then both barebox and picotctp should just
discard it.

> 
> But what I have to do if an ARP or ICMP echo __request__ is received?

Should go to only one of both stacks, otherwise we would get duplicate
ping responses.

Sascha

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |