Anders Lindgren wrote:
>   Hi everyone, and especially Jesse Off if you're reading. :-)
> Playing around with my TS7250 board and its TS7KV framebuffer, and 
> NetBSD 4.99.20.
> 
>   I am trying to access the fb on the PC/104 bus from user space by 
> simply mmap:ing /dev/mem, following the board's [Linux] example code. 
> This way, I've been able to identify the board on the bus, load the FPGA 
> bitmask into the FPGA, ACK the loading, and activate the FPGA. It lights 
> its green LED. :-) However, I get no visible video output apart from a 
> brief flicker before the screen goes blank again.

Likely, all you need to do is enable vsync/hsync in the control 
register.  On FPGA bitstream load, those bits are cleared so the monitor 
will be in power-save.  Otherwise, you'd be seeing uninitialized SDRAM 
until you got around to zeroing out the frame buffer.

> 
>   Before I dig any deeper into this -- is this method going to work at 
> all? I've seen comments to the effect that some X drivers do pretty much 
> this, and I'd like to use the same method to bang on some EP9302 GPIO 
> pins as well. Or do I have to modify the tsarm startup or write a stub 
> open + mmap driver somehow to tell NetBSD not to cache these address 
> ranges?

It should map them uncacheable/unbufferable.  If you request an address 
that falls in SDRAM, I think only then you will get cacheable/bufferable 
mappings.

> 
>   As long as the kernel realizes that the physical ISA addresses I mmap 
> are non-cachable side effect memory locations, it should Just Work(tm)? 
> I figured the tsarm "isaio" node somehow takes care of this somewhere 
> during startup, but reading the code didn't tell me much, as I am not 
> yet familiar with how the bus space framework works.

Not quite.  I believe the ARM pmap module (pmap_enter()) will assume any 
physical address not explicitly registered with uvm_physload() is 
uncacheable/unbufferable.

//Jesse Off