brian@uwvax.UUCP (Brian Pinkerton) (07/25/85)
Has anyone gotten the serial-line IP driver from Seismo to work? If so,
I'd like to hear your experiences...
I'm installing it into two plain-vanilla 4.2 kernels, both of which panic
in mget after an arbitrary number of minutes... anyone have any ideas?
thanks,
brian
--
Brian Pinkerton @ wisconsin
...!{allegra,heurikon,ihnp4,seismo,sfwin,ucbvax,uwm-evax}!uwvax!brian
brian@wisc-rsch.arpachris@umcp-cs.UUCP (Chris Torek) (07/25/85)
Sounds like you're running out of mbufs. Perhaps there is a leak in your version of if_sl.c. (I had this happen once or twice on eneevax before I rewrote most of if_sl.c. Hasn't done it since then....) -- In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 4251) UUCP: seismo!umcp-cs!chris CSNet: chris@umcp-cs ARPA: chris@maryland
richards@uiucdcsb.Uiuc.ARPA (07/28/85)
You also may be running out of CLIST space for TTY output -- we sure did.
A way to improve on that (though not fix it) is to reduce the max packet size
for that interface; we set ours at ... 256, I think.
(Of course doing this you drastically increase network overhead, but it's
better than crashing or running out of terminal character buffers)
Also, you might look at an open connection with TRPT to see how frequently
you re-transmit packets -- one bug fix that came over the net a while ago
left us retransmitting packets several times before they even got out of
the machine via slow links (e.g. 9.6kb serial line or IMP connections).
If you are resending a lot, we might have some fixes that would help.
Paul Richards University of Illinois at Urbana-Champaign, Dept of Comp Sci
UUCP: {pur-ee,convex,inhp4}!uiucdcs!richards
ARPA: richards@uiuc.arpa
CSNET: richards%uiuc@csnet-relaychris@umcp-cs.UUCP (Chris Torek) (07/31/85)
>You also may be running out of CLIST space for TTY output [...]
Oh yeah... that reminds me: one of the other things I changed
involved that. Rick's code moves all packets from the interface
queue to the tty output queue as soon as possible. I thought it
was a better idea to convert mbufs to clists one packet at a time.
This is easy to do by using the line switch start routine.
(Rick---why didn't you do this in the first place?)
The only bad consequence is that the packet queue on the interface
builds up and you can get some fake collisions, but if your TCP
retransmit algorithm has been fixed the times quickly adjust and
that stops happening.* (By the way, DON'T ignore the round trip
times for packets that are retransmitted; this causes the srtt
value to stay too low forever on really slow connections.)
I dredged up my 4.2 version of the driver, and anyone who wants it
is welcome to a copy. I have a (probably incomplete) list of the
other things that need to be done to use my version of the code,
which is substantially different inside (the packet format is the
same though).
-----
*Well, OK, it happens a few times per connection, depending on
load, baud rate, MTU, and probably other factors.
--
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 4251)
UUCP: seismo!umcp-cs!chris
CSNet: chris@umcp-cs ARPA: chris@marylandlee@rocksanne.UUCP (08/01/85)
I installed Seismo's serial line code one four machines and they all are very stable. There is a slight bug in the documentation about how to declare the device driver but that is all.