On 24 Apr 2002, Jose Fonseca wrote: > On Wed, 2002-04-24 at 14:13, Jens Owen wrote: > > ... > > > > Jose, > > > > I don't have a complete answer for you, but here is *some* background > > information that might help. > > > > The drmAddBufs style was created with the origianal DRI infrastructure > > that supported a partial implementation of the Gamma driver. Buffers > > were short, and based out of PCI memory. > > > > The first complete driver was the 3Dfx. It used MMIO and didn't need > > any DMA buffers. > > > > The second and third complete drivers were the i810 and mga. The were > > the first to require AGP based DMA, and that's when the drmAGPAlloc > > functionality was added. > > > > I believe the AGP implementation is flexable about where the physical > > pages reside, and can use the GART to make the pages appear contiguous > > to both the 3D driver *and* the hardware. > > > > For PCI DMA, I believe you need to allocate actual contiguous pages in > > order for the hardware to see the memory as contiguous. This may be why > > the mechanism is more cumbersome--or it could be that the PCI DMA case > > just hasn't been fully flushed out. > > Allocating contiguous pages in physical memory is complicated, and is > addressed by the drmAddBufs, assuming that one doesn't choose the size > of each individual buffer too big. The problem is that drmAddBufs makes > it difficult to separate the general purpose DMA buffers from buffer > with a special purpose, such as primary buffers or page tables.
I was looking at the code in glint_dri.c to add the page table thinking that we could set the DRM_RESTRICTED flag for the descriptor table as is done there. However, in looking at drmAddBufs and the ioctl, the last arg is the agp_offset for DRM_AGP_BUFFER and is unused for PCI, so passing DRM_RESTRICTED in the last arg to drmAddBufs is just wrong. AFAICT, you only use DRM_RESTRICTED with drmAddMap. Am I missing something or does the glint/gamma driver have a problem here? > I think I'll extract the actual buffer allocation code from drmAddBufs > into a separate IOCTL available for those cards that may need it. > > > I believe the r128 driver's PCI support still utilizes GART > > functionality to allow non-contiguous memory to be utilized. Your > > driver may be the first to need true PCI support, and *possibly* the > > last to need it unless other legacy chipsets of the same era are > > supported. > > The gamma driver could use it (but just to have a little more code > readability). I don't know if MGA has a PCI version but it would > simplify substantially extending its support in case there is, as there > would be possible to make a parallelism between the AGP and PCI code. I > believe that the Savage chips come in PCI versions too, and this might > be handy again. > > > > > I hope this helps. > > Yes. Thanks very much for the background info. > > > > > -- /\ > > Jens Owen / \/\ _ > > [EMAIL PROTECTED] / \ \ \ Steamboat Springs, Colorado > > > > Regards, > > Jose Fonseca > > > _______________________________________________ > Dri-devel mailing list > [EMAIL PROTECTED] > https://lists.sourceforge.net/lists/listinfo/dri-devel > -- Leif Delgass http://www.retinalburn.net _______________________________________________ Dri-devel mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/dri-devel
