> At 57600 baud, it takes 354 ms to overflow a 2k ring buffer. Something
> is massively keeping interrupts blocked for way to long (data come out
> of the ring at softinterrupt level, which is at about spl 0.5.
> 
> I think upping the ring buffer's not the solution, We need to figure out
> what's really wrong.

While the following probably isn't the real solution either, I think
it does sound useful... I think the i386 com driver does this, but I'm
not sure; I read about it somewhere though :) When the ring buffer is
about to overflow, and flow control is turned on (crtscts or mdmbuf or
whatever there is), could you twiddle the flow control line so the
other end stops sending? This would at least prevent dropping
characters, which messes up tcp/ip throughput pretty badly :)

Name: Dave Huang     |   Mammal, mammal / their names are called /
INet: khym@bga.com   |   they raise a paw / the bat, the cat /
FurryMUCK: Dahan     |   dolphin and dog / koala bear and hog -- TMBG
Dahan: Hani G Y+C 21 Y++ L+++ W- C++ T++ A+ E+ S++ V++ F- Q+++ P+ B+ PA+ PL++