2.6.10-rc2-mm4
(sorry for breaking the email thread, I tried to grab the mbox archive for responding, but ftp.uwsg.iu.edu seems to be having problems with the mail archive).
in response to Alan Cox' voiced concerns from Tue Nov 30:
Quoted 1 line
That would also be a proprietary hook wouldn't it 8)
nvidia really isn't interested in proprietary hooks; such a hook really wouldn't be good for nvidia or the linux kernel. but we don't feel that a 32-bit zone is such a hook.
Quoted 2 lines
beyond the supported open source hardware PCI-Express is the non-AGPGART user and that is 64bit.
I assume you mean traditional pci in this case, but I remain confused. the pci spec calls for 32-bits of addressing, although there is an optional extension for 64-bit bus extension pins. I can't speak for other pci devices, but all of our pci devices are 32-bit.
additionally, the pci-express spec defines legacy and non-legacy devices. legacy devices are only required to address 32-bits, whereas non-legacy devices are required to handle 64-bit addresses.
Quoted 1 line
In the video space it's even stranger because DRI doesn't need it
I'm unclear how traditional pci video cards function without being able to allocate < 32-bit addresses for dma purposes. are the video cards you've tested 64-bit capable or is dma disabled? what about 32-bit cards? I'm afraid I don't understand how this isn't an issue.
Quoted 2 lines
it seems a risky path because of the past problems trying to get zone balancing working.
I certainly understand the concerns with this, although I was led to believe that recent 2.6 work made the zone balancing much less expensive. is that not the case?
Quoted 1 line
I can find users for a 512Mb or 1Gb DMA region
there was some brief discussion of this when we originally discussed 32-bit addressing issues, but I don't know if a satisfactory solution was reached. If a 1Gb region was prefered for this reason, that should satisfy nvidia's needs for 32-bit addressing, but I couldn't speak for any other device drivers.
for reference, the previous discussion was here:
http://www.uwsg.indiana.edu/hypermail/linux/kernel/0406.2/2093.html
Thanks,
Terence
Quoted 4 lines · Terence Ripperda
I assume you mean traditional pci in this case, but I remain confused. the pci spec calls for 32-bits of addressing, although there is an optional extension for 64-bit bus extension pins. I can't speak for other pci devices, but all of our pci devices are 32-bit.
The current DRI drivers don't really deal much with PCI devices. A pure PCI video card on 64bit boxes might be problematic although I'd question the sanity of anyone doing this 8)
Quoted 4 lines
additionally, the pci-express spec defines legacy and non-legacy devices. legacy devices are only required to address 32-bits, whereas non-legacy devices are required to handle 64-bit addresses.
I'd assumed video card vendors were non-legacy but ok
Quoted 3 lines
I certainly understand the concerns with this, although I was led to believe that recent 2.6 work made the zone balancing much less expensive. is that not the case?
Andrew certainly believes this is. Certainly in 2.4 it was not.
Quoted 7 lines
Quoted 1 line
I can find users for a 512Mb or 1Gb DMA region
there was some brief discussion of this when we originally discussed 32-bit addressing issues, but I don't know if a satisfactory solution was reached. If a 1Gb region was prefered for this reason, that should satisfy nvidia's needs for 32-bit addressing, but I couldn't speak for any other device drivers.
If the VM can take it and get it right I am all for a 512Mb or 1Gb DMA region to fix the various devices that have 29-31bit DMA issues. If it fixes Nvidia's needs to then fine.
Alan