[U-Boot] [PATCH] net: BOOTP retry timeout improvements

[Thierry Reding]

On Fri, Aug 15, 2014 at 10:02:40AM -0600, Stephen Warren wrote:
> On 08/15/2014 06:49 AM, Thierry Reding wrote:
> >On Fri, Aug 15, 2014 at 02:39:45PM +0200, Thierry Reding wrote:
> >>On Fri, Jul 25, 2014 at 05:30:48PM -0600, Stephen Warren wrote:
> >>>From: Stephen Warren <swarren at nvidia.com>
> >>>
> >>>Currently, the BOOTP code sends out its initial request as soon as the
> >>>Ethernet driver indicates "link up". If this packet is lost or not
> >>>replied to for some reason, the code waits for a 1s timeout before
> >>>retrying. For some reason, such early packets are often lost on my
> >>>system, so this causes an annoying delay.
> >>>
> >>>To optimize this, modify the BOOTP code to have very short timeouts for
> >>>the first packet transmitted, but gradually increase the timeout each
> >>>time a timeout occurs. This way, if the first packet is lost, the second
> >>>packet is transmitted quite quickly and hence the overall delay is low.
> >>>However, if there's still no response, we don't keep spewing out packets
> >>>at an insane speed.
> >>>
> >>>It's arguably more correct to try and find out why the first packet is
> >>>lost. However, it seems to disappear inside my Ethenet chip; the TX chip
> >>>indicates no error during TX (not that it has much in the way of
> >>>reporting...), yet wireshark on the RX side doesn't see any packet.
> >>>FWIW, I'm using an ASIX USB Ethernet adapter. Perhaps "link up" is
> >>>reported too early or based on the wrong condition in HW, and we should
> >>>add some fixed extra delay into the driver. However, this would slow down
> >>>every link up event even if it ends up not being needed in some cases.
> >>>Having BOOTP retry quickly applies the fix/WAR to every possible
> >>>Ethernet device, and is quite simple to implement, so seems a better
> >>>solution.
> >>>
> >>>Signed-off-by: Stephen Warren <swarren at nvidia.com>
> >>>---
> >>>  net/bootp.c | 38 +++++++++++++++++++++++++++++++-------
> >>>  net/bootp.h |  3 +--
> >>>  net/net.c   |  4 ++--
> >>>  3 files changed, 34 insertions(+), 11 deletions(-)
> >>
> >>Sorry for jumping in a little late, but I only encountered this after
> >>rebasing on latest origin/master today.
> >>
> >>With this patch applied, it takes about 8 seconds until U-Boot manages
> >>to get a DHCP address (it needs to broadcast 14 times) whereas with the
> >>patch reverted I get a reply almost instantly (with only a single
> >>broadcast).
> >>
> >>I'm testing this on a Jetson TK1 (with local patches for PCIe and
> >>ethernet support).
> >
> >Further tests show that if I increase the initial bootp_timeout to
> >anywhere beyond 750 ms things don't regress for me. If I go down to 500
> >ms it will send out a second broadcast (presumably because the DHCP
> >server in my network hasn't replied yet) but still get a reasonably
> >quick reply (~ 1 second).
> >
> >So it would seem that this parameter depends to a large degree on the
> >network infrastructure and needs to be carefully tweaked to work well in
> >most circumstances.
> >
> >Could you try if 500, 750 and 1000 ms as initial bootp_timeout value
> >give you acceptable results?
> 
> That will still introduce extremely large delays into the boot process,
> which is exactly what this patch was trying to avoid. 500..1000ms are the
> same order of magnitude as the initial delay without this patch applied.

It will introduce a 1000 ms delay /if/ you have broken hardware/drivers
where the first packet is lost. If the first packet isn't lost there's
no additional delay except for the DHCP server's latency in replying. So
in effect the patch trades long latency on one setup for longer latency
on another set up.

Also the timeout was previously 5000 ms, so by adjusting the timeout to
1000 ms we'd still be gaining quite a few seconds.

> Re-transmitting a DHCP request shouldn't prevent a response to the previous
> request from being processed - AFAIK each request is idempotent. Can you
> debug what is causing the 8s delay? Are earlier responses received and
> rejected for some reason, or is your DHCP server getting confused by the
> multiple requests and not responding, or ...?

I haven't really tested this, but by looking at the code in net/bootp.c
(function BootpCheckPkt()) the U-Boot implementation will reject all
packets that don't match the BootpID (which is composed of the lower 4
bytes of the ethernet address plus the time in milliseconds when the
discover packet was sent, see BootpRequest()).

So indeed earlier responses will be rejected, which would also explain
the 8 second delay since that's about the time it takes for the backoff
to reach the timeout that the server requires to reply before the next
retransmission.

> (and as an aside, how on earth is your DHCP server taking >500ms to respond,
> yet still actually responding?)

It's a black box mostly but it seems to be doing the same for any of the
other devices (tcpdump on the machine that I'm typing this from gives me
roughly 430 ms in one session and 580 ms in another between the DISCOVER
message and the server's reply) on the network.

This seems absurdly long, but it's consumer network equipment (I'd even
say high-end by the price-tag) and there's not a lot I can configure to
make it faster. I also suspect that other people may have hardware with
similar latency.

I've also looked briefly at RFC 2131 (DHCP) which recommends an initial
retransmission timeout of 4 (+- 1) seconds for 10 Mbps ethernet. The RFC
goes on to say that this should be chosen based on the characteristics
of the network, so unless we really do choose the initial timeout based
on the link speed (even then you could still have many layers of bridges
and WiFi between client and server) the default initial timeout should
probably be larger than 10 ms. On 100 Mbps 10 ms is about the round-trip
time for a ping, so it's cutting it rather close as timeout for
retransmission.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.denx.de/pipermail/u-boot/attachments/20140815/abf2bf80/attachment.pgp>