Hello,
On Sun, 11 May 2014 19:07:01 +0100
Nick Hudson <nick.hud...@gmx.co.uk> wrote:
> On 11/05/2014 16:47, Michael wrote:
> > Hello,
> >
> > On Sun, 11 May 2014 07:53:28 +0000
> > "Nick Hudson" <sk...@netbsd.org> wrote:
> >
> >> Module Name: src
> >> Committed By: skrll
> >> Date: Sun May 11 07:53:28 UTC 2014
> >>
> >> Modified Files:
> >> src/sys/arch/mips/include: pmap.h
> >> src/sys/arch/mips/mips: pmap.c pmap_segtab.c
> >>
> >> Log Message:
> >> Deal with incompatible cache aliases. Specifically,
> >>
> >> - always flush an ephemeral page on unmap
> >> - track unmanaged mappings (mappings entered via pmap_kenter_pa) for
> >> aliases where required and handle appropriately (via pmap_enter_pv)
> >>
> >> Hopefully this (finally) addresses the instability reported in the
> >> following PRs:
> >>
> >> PR/44900 - R5000/Rm5200 mips ports are broken
> >> PR/46890 - upcoming NetBSD 6.0 release is very unstable / unusable on
> >> cobalt qube 2
> >> PR/48628 - cobalt and hpcmips ports are dead
> >>
> >>
> >> To generate a diff of this commit:
> >> cvs rdiff -u -r1.62 -r1.63 src/sys/arch/mips/include/pmap.h
> >> cvs rdiff -u -r1.213 -r1.214 src/sys/arch/mips/mips/pmap.c
> >> cvs rdiff -u -r1.7 -r1.8 src/sys/arch/mips/mips/pmap_segtab.c
> > My O2's been building stuff from pkgsrc for the last hour without any
> > of the glitches we've seen before ( fingers crossed )
> Good news.
It's doing much better but something is still not quite right. Saw a
SIGILL from /usr/bin/ar, disassemply showed a bunch of seemingly random
0xwhatever between what looked like sane MIPS instructions, problem
disappeared after reboot.
> > Building perl now, let's see how that goes.
> > I'm using an n32 userland and kernel, with the bus_dma.c changes
> > tsutsui@ mentioned adapted to sgimips/bus.c,
>
> How about switching to generic bus_{dma,space}?
Not entirely trivial, sgimips' bus_{dma,space} is weird. Should
probably be part of the larger project of making IP2x and IP3x kernels
run LP64.
> > gcc 4.5, gmp patched as
> > posted earlier.
>
> Is this ready to commit?
Good question, I'd like to hear from one of the Matts first. All the
patch does is to disable using assembly routines when building n32
binaries ( as in, mips64* but not _LP64 ). Seems to work, at least on
sgimips, needs more torturing on Loongson.
have fun
Michael
