CVS commit: src/sys/arch/sparc/sparc

[matthew green]

Module Name: src
Committed By: chs
Date: Sun Nov 4 00:32:47 UTC 2012
src/sys/arch/sparc/sparc: locore.s pmap.c
in cpu_switchto(), remove the MP-unsafe code to mark a pmap active on a CPU,
pmap_activate() already does this. add MP locking to pmap_activate()
and pmap_deactivate(). move flushing of user windows and virtual caches
from pamp_activate() to pmap_deactivate().

hmm. this makes pmap_deactivate() do things in UP that weren't being
done before at all. switching to/from kernel/lwp for the same lwp will
now be flushing the ctx every time. that seems suboptimal?

do you just mean that I should have left the #ifdef MULTIPROCESSOR
around the sp_tlb_flush()? or something more than that?

i'm not sure exactly. i had a quick look at editing it back again
but i didn't like the result. i'm imaging a system that only flushes
the previous context in pmap_activate() if it was not the same. ie,
pmap_deactivate() shouldn't always flush. i don't think this should
be MULTIPROCESSOR-related either.

we should really update PMAP_LOCK() to not be kernel lock.

agreed.
beyond that, there's still plenty of room for improvement on most platforms.
I think the best example in our tree at this point is the x86 scheme
of deferring the MMU-switch work to points where we're actually about to
reference user mappings, eg. returning to user mode, copyin/copyout, etc.

many pmaps can have simpler locking now, since the rmind-uvmplock
branch was merged. i have a started reducing the use of
sparc64/pmap.c:pmap_lock though there is plenty more to go.

thanks.


.mrg.