Michael Kelly, le dim. 15 nov. 2026 15:20:46 +0000, a ecrit:
> Hi Samuel,
> 
> On 14/11/2026 21:40, Samuel Thibault wrote:
> > Michael Kelly, le sam. 14 nov. 2026 21:36:01 +0000, a ecrit:
> > > I'm not concerned about the performance of these prints post-boot; it's 
> > > the
> > > affect on the boot time that concerns me. The boot output is much longer 
> > > on
> > > a real machine than a virtual one with lots more ACPI  debug which should
> > > perhaps be turned off ?
> > We could probably disable some ACPI prints yes, but better just fix the
> > underlying issue :)
> Agreed.
> > > > As mentioned in a previous mail, it looks completely ok to me to
> > > > implement scrolling by just drawing the known text of the gnumach
> > > > console at the new position, and leave the rest alone. That'll be way
> > > > faster than memmoving everything.
> 
> I've prototyped (hacked!) a write combining mapping for the user land frame
> buffer. The speed up is tremendous with 10000 full writes to the frame
> buffer taking 3.5s compared with 64s without WC. I think that most of what
> I've implemented already for the gnumach console will be acceptable once
> I've added an interface to support write combining properly.
> 
> A user interface would be useful too so the Hurd console can benefit.
> 
> Have you any strong opinions on the detail of supporting write combining in
> gnumach or shall I try something and see what you think ?

pmap_enter() actually already handles setting
INTEL_PTE_NCACHE|INTEL_PTE_WTHRU for non-RAM pages, so it should already
be working for userland. It's the set up of the directmap used by
phystokv that would need to fix the flags being used to map the VGA RAM.

Michael Kelly, le dim. 15 nov. 2026 18:11:26 +0000, a ecrit:
> Integrating the write combined mapping into the gnumach console code
> achieves a scrolled output rate of approximately 800 lines per second. These
> lines are 50 characters wide which is my estimate at an average length. This
> is running on my fairly vintage desktop box which probably has lower than
> average performance. I think that's good enough for a starting point.

Yes, that looks good enough.

Samuel

Reply via email to