Re: [PATCH] basic drm driver for Trident CyberBlade

[<uni...@sh...>]

Jesse Barnes wrote:

>I'm trying to get the CyberBlade EXA driver into a bit better shape, and 
>that means some sort of decent host<->fb DMA support.  The device 
>itself is apparently capable of that, but EXA has trouble since it only 
>has the user virtual address of the transfers it wants to make, so I 
>thought a DRM driver might be in order.  It's also necessary for decent 
>VBLANK support, which will hopefully be pretty straightforward.  Note 
>that at this point, the DRM driver won't do you any good since I 
>haven't checked the EXA stuff that uses it into X.Org CVS yet (and 
>afaik there's no DRI driver for this device either).
>
>Any thoughts on how the host<->fb DMA stuff should work?  Basically EXA 
>UploadToScreen wants to hand the card a virtual address range and have 
>it loaded up to the card's VRAM; DownloadFromScreen wants to do the 
>same in reverse.  Should I just add a device specific DRM_IOCTL to 
>accomplish this (using get_user_pages() etc.) or is there some existing 
>code I could leverage?
>
>Thanks,
>Jesse
>  
>
The via driver has code that does this (via_dmablit.c) as a 
device-specific IOCTL.

It maintains a queue of blit operations and fire them off when the 
previous one is completed. The user calls a sync IOCTL to verify that 
the operation is finished.

The via-specific code is quite limited, so it should be easy to port it 
to other hardware but the card needs to have a descriptor-based PCI DMA 
engine. Note that the throughput is quite low. With EXA you get about 
65-70 MB/s either direction. Limited by the PCI bus speed.

If your card can do AGP->VRAM transfers in hardware the best option for 
upload is probably to use that functionality. Software copies to AGP can 
be pipelined with AGP->VRAM blits using two bounce-buffers in AGP space. 
You split up the transfer and copy to one buffer while you're blitting 
to VRAM from the other.

For downloads from VRAM we will soon have an option to blit from VRAM to 
AGP, and then take the pages out of AGP space and read from them with 
full speed, but that's still under development.

/Thomas