> I used to be able to have two COM lines firing at 115200 baud  with almost no errors reported (all were silo errors
> and occurred when both lines were doing serious web browsing), as well as 20 or so browsers using the web proxy
> with no hassles.  Not even a flinch.  I have had 120 browers hitting the proxy as well as people coming in the ISDN
> line and have then seen peaks of about 96% utilization, which was really cranking!!!  I love this system!
> 
> BUT, since Christmas or very soon after (29/12/95 I think) I can't get one COM port running at 57600 on a stand alone
> machine to be error free for any length of time.

If this was one of the hardware projects I design for living, I would
definately start looking for a place which raises the interrupt level
and forgets to return it back to normal when the short, critical code
segment has been run.  Isn't it possible that some modification done
around Christmas leaves the IPL too high (maybe every time a certain
routine is called, maybe only when certain conditions are met)?

I don't know the NetBSD internals, but a pentium does a lot of things
within one millisecond and the interrupts are blocked for more than
that.

Where does the kernel disable interrupts or raise the interrupt level?
In device drivers, yes.. but the drivers should be pretty fast doing
their trick.  Paging?  That can't be it, the com problems are present
on systems with 10 MB of RAM free and no swap in use.

How are the interrupt priorities organized?  If, say, an Ethernet card
causes lots of interrupts, will they block com interrupts completely
or vice-versa?

Gee, I wish I knew more. :)

> I was new to ppp servers when I set the thing up, and I had a separate IP for the local end of each line as well as the 
> NIC (.10, .11, and .12) and I had static routes set up to go between these.  At christmas I changed my ppp configuration 
> at that time to go point-to-point from each line to .10, which is the NIC address.  It seemed to work fine except for the
> frame errors, so I figured I had done OK.  Is this right?

I have only one IP address that is shared by the PPP links and the Ethernet.
Works fine.  Only when I wanted to route a subnet (sub-C) over a PPP link,
I used a separate address for the PPP link at both ends (was easier to make
the other machines understand what was going on).

  -jm