* Re: TLB entries > 4kb
@ 1998-02-12 21:52 David Woodhouse
1998-02-12 21:57 ` Alan Cox
1998-02-12 22:00 ` MOLNAR Ingo
0 siblings, 2 replies; 10+ messages in thread
From: David Woodhouse @ 1998-02-12 21:52 UTC (permalink / raw)
To: MOLNAR Ingo; +Cc: Alan Cox, ralf, linux-kernel, linux
mingo@chiara.csoma.elte.hu said:
> On Thu, 12 Feb 1998, Alan Cox wrote:
> > On the intel its a huge win to map PCI frame buffers using 4Mbyte pages, but
> > the kernel mmap really can't hack the idea right now. A pity cos if you
> > hack it in and make sure you never unmap it you get 2-3% better X
> > performance
>
> i tried to hack this half a year ago, but had to find out that 4MB pages
> are only valid for kernel mappings, which means we'd have to make X a ring
> 0 process too ...
---------------
alan@lxorguk.ukuu.org.uk said:
> The way its done is ok normally. I guess you need to make X over
> Linux/RT call the rtlinux kernel to handle it. Note X _has_ to do
> those global cli/sti's to meet hardware needs
---------------
GGI (well, KGI) should be able to fix both of these, shouldn't it?
---- ---- ----
David Woodhouse, Robinson College, CB3 9AN, England. (+44) 0976 658355
Dave@imladris.demon.co.uk http://dwmw2.robinson.cam.ac.uk
finger pgp@dwmw2.robinson.cam.ac.uk for PGP key.
^ permalink raw reply [flat|nested] 10+ messages in thread* Re: TLB entries > 4kb
1998-02-12 21:52 TLB entries > 4kb David Woodhouse
@ 1998-02-12 21:57 ` Alan Cox
1998-02-12 21:57 ` Alan Cox
1998-02-12 22:00 ` MOLNAR Ingo
1 sibling, 1 reply; 10+ messages in thread
From: Alan Cox @ 1998-02-12 21:57 UTC (permalink / raw)
To: David Woodhouse; +Cc: mingo, alan, ralf, linux-kernel, linux
> > The way its done is ok normally. I guess you need to make X over
> > Linux/RT call the rtlinux kernel to handle it. Note X _has_ to do
> > those global cli/sti's to meet hardware needs
>
> GGI (well, KGI) should be able to fix both of these, shouldn't it?
Not for Linux/RT - the kernel cant do real cli/sti pairs. The stuff
would have to become an RTLinux graphics setup call set
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 21:57 ` Alan Cox
@ 1998-02-12 21:57 ` Alan Cox
0 siblings, 0 replies; 10+ messages in thread
From: Alan Cox @ 1998-02-12 21:57 UTC (permalink / raw)
To: David Woodhouse; +Cc: mingo, alan, ralf, linux-kernel, linux
> > The way its done is ok normally. I guess you need to make X over
> > Linux/RT call the rtlinux kernel to handle it. Note X _has_ to do
> > those global cli/sti's to meet hardware needs
>
> GGI (well, KGI) should be able to fix both of these, shouldn't it?
Not for Linux/RT - the kernel cant do real cli/sti pairs. The stuff
would have to become an RTLinux graphics setup call set
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 21:52 TLB entries > 4kb David Woodhouse
1998-02-12 21:57 ` Alan Cox
@ 1998-02-12 22:00 ` MOLNAR Ingo
1 sibling, 0 replies; 10+ messages in thread
From: MOLNAR Ingo @ 1998-02-12 22:00 UTC (permalink / raw)
To: David Woodhouse; +Cc: Alan Cox, ralf, linux-kernel, linux
On Thu, 12 Feb 1998, David Woodhouse wrote:
> GGI (well, KGI) should be able to fix both of these, shouldn't it?
the hw virtualization (cli/sti) thing yes, but the 4MB thing not, as
GGILib lives in user-space, GGI does not spend too much time in the
kernel, rendering is done in user-space.
but i suspect things doing direct rendering like Quake could benefit from
4MB pages big time. For this we have to implement ring-0 processes though
... XFree86 is nicely accelerated (and the accelerator chip has no
business with the CPU).
-- mingo
^ permalink raw reply [flat|nested] 10+ messages in thread
* TLB entries > 4kb
@ 1998-02-12 4:56 ralf
1998-02-12 5:50 ` David S. Miller
` (2 more replies)
0 siblings, 3 replies; 10+ messages in thread
From: ralf @ 1998-02-12 4:56 UTC (permalink / raw)
To: linux-kernel, linux
Some architectures can use multiple page sizes in the TLB at the same time.
This would for example allow to map memory allocations > PAGE_SIZE using just
a single TLB entry if the circumstances are just right, thereby
reducing / eleminating TLB trashing. This should improve the performance
for huge apps quite a bit. Some architectures could partially get rid of
the sick effects of their virtual indexed primary caches as well. All that
is needed for this to work is to have sufficiently large physical pages with
sufficient alignment at hand.
Has anybody ever looked into implementing that? What architectures besides
MIPS could take advantage of such a feature?
Ralf
^ permalink raw reply [flat|nested] 10+ messages in thread* Re: TLB entries > 4kb
1998-02-12 4:56 ralf
@ 1998-02-12 5:50 ` David S. Miller
1998-02-12 8:51 ` Alan Cox
1998-02-12 12:10 ` Michael Lausch
2 siblings, 0 replies; 10+ messages in thread
From: David S. Miller @ 1998-02-12 5:50 UTC (permalink / raw)
To: ralf; +Cc: linux-kernel, linux
From: ralf@uni-koblenz.de
Date: Thu, 12 Feb 1998 05:56:48 +0100
Has anybody ever looked into implementing that? What architectures
besides MIPS could take advantage of such a feature?
We already use it on Sparc64 for mmap()'ing the frame buffers, we
could easily extend it to work for mmap()'s of real memory. Check out
arch/sparc64/mm/generic.c in the vger CVS tree for details.
Later,
David S. Miller
davem@dm.cobaltmicro.com
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 4:56 ralf
1998-02-12 5:50 ` David S. Miller
@ 1998-02-12 8:51 ` Alan Cox
1998-02-12 18:00 ` MOLNAR Ingo
1998-02-12 12:10 ` Michael Lausch
2 siblings, 1 reply; 10+ messages in thread
From: Alan Cox @ 1998-02-12 8:51 UTC (permalink / raw)
To: ralf; +Cc: linux-kernel, linux
> Has anybody ever looked into implementing that? What architectures besides
> MIPS could take advantage of such a feature?
On the intel its a huge win to map PCI frame buffers using 4Mbyte pages, but
the kernel mmap really can't hack the idea right now. A pity cos if you
hack it in and make sure you never unmap it you get 2-3% better X performance
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 8:51 ` Alan Cox
@ 1998-02-12 18:00 ` MOLNAR Ingo
0 siblings, 0 replies; 10+ messages in thread
From: MOLNAR Ingo @ 1998-02-12 18:00 UTC (permalink / raw)
To: Alan Cox; +Cc: ralf, linux-kernel, linux
On Thu, 12 Feb 1998, Alan Cox wrote:
> > Has anybody ever looked into implementing that? What architectures besides
> > MIPS could take advantage of such a feature?
>
> On the intel its a huge win to map PCI frame buffers using 4Mbyte pages, but
> the kernel mmap really can't hack the idea right now. A pity cos if you
> hack it in and make sure you never unmap it you get 2-3% better X performance
i tried to hack this half a year ago, but had to find out that 4MB pages
are only valid for kernel mappings, which means we'd have to make X a ring
0 process too ...
-- mingo
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 4:56 ralf
1998-02-12 5:50 ` David S. Miller
1998-02-12 8:51 ` Alan Cox
@ 1998-02-12 12:10 ` Michael Lausch
1998-02-12 22:05 ` J.D. Bakker
2 siblings, 1 reply; 10+ messages in thread
From: Michael Lausch @ 1998-02-12 12:10 UTC (permalink / raw)
To: ralf; +Cc: linux-kernel, linux
>>>>> "r" == ralf <ralf@uni-koblenz.de>
>>>>> wrote the following on Thu, 12 Feb 1998 05:56:48 +0100
r> Some architectures can use multiple page sizes in the TLB at the same time.
r> This would for example allow to map memory allocations > PAGE_SIZE using just
r> a single TLB entry if the circumstances are just right, thereby
r> reducing / eleminating TLB trashing. This should improve the performance
r> for huge apps quite a bit. Some architectures could partially get rid of
r> the sick effects of their virtual indexed primary caches as well. All that
r> is needed for this to work is to have sufficiently large physical pages with
r> sufficient alignment at hand.
r> Has anybody ever looked into implementing that? What architectures besides
r> MIPS could take advantage of such a feature?
The MPC860. We now have 8 MByte Page Table entries for the Kernel Address
Space and the Dual Ported RAM used to communicate with the CPM.
r> Ralf
r> -
r> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
r> the body of a message to majordomo@vger.rutgers.edu
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: TLB entries > 4kb
1998-02-12 12:10 ` Michael Lausch
@ 1998-02-12 22:05 ` J.D. Bakker
0 siblings, 0 replies; 10+ messages in thread
From: J.D. Bakker @ 1998-02-12 22:05 UTC (permalink / raw)
To: Michael Lausch; +Cc: ralf, linux-kernel, linux
>r> Has anybody ever looked into implementing that? What architectures besides
>r> MIPS could take advantage of such a feature?
>
>The MPC860. We now have 8 MByte Page Table entries for the Kernel Address
>Space and the Dual Ported RAM used to communicate with the CPM.
...and the IBM PPC403GC/GCX. Variable page sizes between 1k and 16M (in
eight steps, variable per TLB IIRC). This MMU, "designed for embedded apps"
according to IBM, also handles the process ID per TLB that Ralf mentioned
on the MIPS.
JDB (hunting for round tuits for that PPC board design)
Jan-Derk Bakker
Official Usenet Net.scum; see http://www.netscum.net
^ permalink raw reply [flat|nested] 10+ messages in thread
end of thread, other threads:[~1998-02-13 7:54 UTC | newest]
Thread overview: 10+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
1998-02-12 21:52 TLB entries > 4kb David Woodhouse
1998-02-12 21:57 ` Alan Cox
1998-02-12 21:57 ` Alan Cox
1998-02-12 22:00 ` MOLNAR Ingo
-- strict thread matches above, loose matches on Subject: below --
1998-02-12 4:56 ralf
1998-02-12 5:50 ` David S. Miller
1998-02-12 8:51 ` Alan Cox
1998-02-12 18:00 ` MOLNAR Ingo
1998-02-12 12:10 ` Michael Lausch
1998-02-12 22:05 ` J.D. Bakker
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox