Re: mlockall() and small memory systems

[Martin Husemann <martin%duskware.de@localhost>]

So, some further investigation shows that on sparc (but not sparc64)
the vmspace for userland processes is pretty huge - the t_mincore test
already has a vmspace vm->vm_map.size value of 9118 pages before it even
does the first mmap. Given the low resource limits cited upthread (~3500
locked pages) we are already beyound that when we try mlockall() (though
I guess most pages are r/o and shared, i.e. libc, but it is not clear
to me, how that is accounted against the resource limits).

The values on sparc64 (with 8k pages) for comparision:

1767 pages vm_map.size, 190 of that resident. At the end of the test,
there are exactly 128 additional pages (1895) and 745 are resident.

Is there an easy way to not count the shared r/o pages? So I could do
dynamic scaling of the test area size, or even skip the test completely
if it is hopeless.

And, more important: why are the sparc space *that* big?

Martin