This patch is to speed up flipping of pages in and out of the AGP aperture as
needed by the new drm memory manager.
A number of global cache flushes are removed as well as some PCI posting flushes.
The following guidelines have been used:
1) Memory that is only mapped uncached and that has been subject to a global
cache flush after the mapping was changed to uncached does not need any more
cache flushes. Neither before binding to the aperture nor after unbinding.
2) Only do one PCI posting flush after a sequence of writes modifying page
entries in the GATT.
Signed-off-by: Thomas Hellstrom <thomas@tungstengraphics.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Not all graphic page remappers support physical addresses over the 4GB
mark for remapping, so while some do (the AMD64 GART always did, and I
just fixed the i965 to do so properly), we're safest off just forcing
GFP_DMA32 allocations to make sure graphics pages get allocated in the
low 32-bit address space by default.
AGP sub-drivers that really care, and can do better, could just choose
to implement their own allocator (or we could add another "64-bit safe"
default allocator for their use), but quite frankly, you're not likely
to care in practice.
So for now, this trivial change means that we won't be allocating pages
that we can't map correctly by mistake on x86-64.
[ On traditional 32-bit x86, this could never happen, because GFP_KERNEL
would never allocate any highmem memory anyway ]
Acked-by: Andi Kleen <ak@suse.de>
Acked-by: Dave Jones <davej@redhat.com>
Cc: Eric Anholt <eric@anholt.net>
Cc: Keith Packard <keithp@keithp.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Some dumb bridges are programmed to disobey the AGP2 spec.
This is likely a BIOS misprogramming rather than poweron default, or
it would be a lot more common.
AGPv2 spec 6.1.9 states:
"The RATE field indicates the data transfer rates supported by this
device. A.G.P. devices must report all that apply."
Fix them up as best we can.
This will prevent errors like..
agpgart: Found an AGP 3.5 compliant device at 0000:00:00.0.
agpgart: req mode 1f000201 bridge_agpstat 1f000a14 vga_agpstat 2f000217.
agpgart: Device is in legacy mode, falling back to 2.x
agpgart: Putting AGP V2 device at 0000:00:00.0 into 0x mode
agpgart: Putting AGP V2 device at 0000:01:00.0 into 0x mode
agpgart: Putting AGP V2 device at 0000:01:00.1 into 0x mode
https://bugs.freedesktop.org/show_bug.cgi?id=8816
Signed-off-by: Dave Jones <davej@redhat.com>
Sometimes the logic to handle AGPx8->AGPx4 fallback failed, as can
be seen in https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=197346
The failures occured if the bridge was in AGPx8 mode, but the
user hadn't specified a mode in their X config. We weren't
setting the mode to the highest mode capable by the video card+bridge
(as we do in the AGPv2 case), which was leading to all kinds of
mayhem including us believing that after falling back from AGPx8, that
we couldn't do x4 mode (which is disastrous in AGPv3, as those are
the only two modes possible).
Signed-off-by: Dave Jones <davej@redhat.com>
AGP shouldn't use "global_flush_tlb()" to flush the AGP mappings, that i
spurely an x86'ism. The proper AGP mapping flusher that should be used
is "flush_agp_mappings()", which on x86 obviously happens to do a global
TLB flush.
This makes AGP (or at least the config _I_ happen to use) compile again
on ppc64.
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
AGP allocation/deallocation is suffering major performance issues due to
the nature of global_flush_tlb() being called on every change_page_attr()
call.
For small allocations this isn't really seen, but when you start allocating
50000 pages of AGP space, for say, texture memory, then things can take
seconds to complete.
In some cases the situation is doubled or even quadrupled in the time due
to SMP, or a deallocation, then a new reallocation. I've had a case of
upto 20 seconds wait time to deallocate and reallocate AGP space.
This patch fixes the problem by making it the caller's responsibility to
call global_flush_tlb(), and so removes it from every instance of mapping a
page into AGP space until the time that all change_page_attr() changes are
done.
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Spotted by Jeremy Fitzhardinge, this change crept in with the multiple
backend support. It's clearly incorrect to overwrite info->mode after
we just went to lengths to determine which bits to mask out.
Signed-off-by: Dave Jones <davej@redhat.com>
[AGPGART] Replace check_bridge_mode() with (bridge->mode & AGSTAT_MODE_3_0).
As mentioned earlier, the current check_bridge_mode() code assumes
that AGP bridges are PCI devices. This isn't always true. Definitely
not for HP zx1 chipset and the same seems to be the case for SGI's AGP
bridge.
The patch below fixes the problem by picking up the AGP_MODE_3_0 bit
from bridge->mode. I feel like I may be missing something, since I
can't see any reason why check_bridge_mode() wasn't doing that in the
first place. According to the AGP 3.0 specs, the AGP_MODE_3_0 bit is
determined during the hardware reset and cannot be changed, so it
seems to me it should be safe to pick it up from bridge->mode.
With the patch applied, I can definitely use AGP acceleration both
with AGP 2.0 and AGP 3.0 (one with an Nvidia card, the other with an
ATI FireGL card).
Unless someone spots a problem, please apply this patch so 3d
acceleration can work on zx1 boxes again.
This makes AGP work again on machines with an AGP bridge that isn't a
PCI device.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Dave Jones <davej@redhat.com>
When Linux is running on the Xen virtual machine monitor, physical
addresses are virtualised and cannot be directly referenced by the AGP
GART. This patch fixes the GART driver for Xen by adding a layer of
abstraction between physical addresses and 'GART addresses'.
Architecture-specific functions are also defined for allocating and freeing
the GATT. Xen requires this to ensure that table really is contiguous from
the point of view of the GART.
These extra interface functions are defined as 'no-ops' for all existing
architectures that use the GART driver.
Signed-off-by: Keir Fraser <keir@xensource.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Dave Jones <davej@redhat.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
