[pups] Progress on 2.11BSD kernel
Steven M. Schultz
sms at 2BSD.COM
Thu Mar 20 09:52:15 AEST 2003
> From: Johnny Billquist <bqt at update.uu.se>
> Are the UMRs allocated and set up statically?
The clist and buffer cache UMRs are initialized at system boot time.
> (I haven't looked inside 2BSD for a while now, and can't remember much of
> the internals anymore.)
And I haven't been around a UNIBUS system for a long time and the
memory of having to deal with UMRs has faded.
> How many UMRs does the system use? 8 for buffer cache, you might expect
> one or two DH11s, that would require a few more, ethernet takes another
> few, but it seems there shouldn't be such a shortage.
There are only 31 total - so a few allocated here, a few allocated
there could result in not having enough.
> What did I miss?
Swapping. The kernel has to always be able to get UMRs to be
able to swap out a process. The memory is fuzzy but I think the
kernel reserved 7 as to be able to swap 56KB at a time of a process
but it could get by with less.
One UMR is 'reserved' by the hardware to cover the "I/O Page".
The clist takes up 1 UMR for DMA terminal devices.
Ethernet of course takes a UMR or two. At one time a couple of the
ethernet drivers (DEUNA) were a bit braindamaged and allocated more
UMRs than they needed - I fixed that though.
MSCP and TMSCP devices need UMRs to map their command and response
rings. The initial MSCP driver was definitely suboptimal in that
regard - again, something I tended to (the problem showed up when
more than 1 controller was present in a system). Those UMRs are
permanently allocated at boot time by the drivers when they initialize.
Tape drivers need to have UMRs (although I don't know of too many
folks with the old TE16 drives around on a UNIBUS system) available.
It started adding up and on a system with several types of devices
each of which wanted to reserve a couple UMRs the available pool for
dynamic use (by tapes and swapping) was sometimes too small.
More information about the TUHS