Christian Kruse wrote:

[ ... ]

[ ... ]

NOTE: This may or may not be your problem; it was mine, on similar hardware.

--

With 1G of RAM, you will want to increase the number of open files, sockets, etc., that are created on static allocations, in order to force some TLBs which are not flushed by a reload of %CR3 or an explicit INVTLB to be flushed, to end up being recycled instead (effectively flushing them), since there are some subtle problems in the code in locore.s, machdep.c, and pmap.c in the mp case, when interacting with the PG_G and PG_PSE bits.

An alternative workaround is to add:

options DISABLE_PSE

to your configuration file.

I've locally corrected the problem, and run with 4G of RAM+PG_PSE without incident, but I had to do some evil to make it work, which you may not want in your kernel (e.g. I changed the machdep.c swap and reserve allocations, which get ungodly huge when you have a lot of memory, and end up being largely useless, when your physical address space is so close to populate -- Matt made some changes there in -current, but I don't agree with them, so my code is "weird", starting with a 3G/1G instead of 1G/3G split on KVA vs. user space addresses).

-- Terry