Wed, Jun 28, 2006 at 10:44:25AM -0700, Jason Thorpe wrote:

Jun 28, 2006, at 3:18 AM, Andrew Doran wrote:

>I made a few more changes An uncompiled & untested diff is attached:
>
>- pass l->l_proc to bus_dmamap_load() so we don't need to double
>buffer

do this if you vslock the buffer!

When calling uvm_vslock(), do we also need to PHLD() the process ?

Wed, Jun 28, 2006 at 11:51:31AM -0700, Jason Thorpe wrote:

Jun 28, 2006, at 10:50 AM, Manuel Bouyer wrote:

>do this if you vslock the buffer!
>
>When calling uvm_vslock(), do we also need to PHLD() the process ?

You only need to PHLD if you going to do DMA to/from the stack.

The kernel stack, I guess ? So to do dma from/to a user-supplied buffer
only uvm_vslock() is needed, right ?

Jun 28, 2006, at 12:06 PM, Manuel Bouyer wrote:

Wed, Jun 28, 2006 at 11:51:31AM -0700, Jason Thorpe wrote:
>>
>Jun 28, 2006, at 10:50 AM, Manuel Bouyer wrote:
>>
do this if you vslock the buffer!

When calling uvm_vslock(), do we also need to PHLD() the process ?
>>
>You only need to PHLD if you going to do DMA to/from the stack.
>
The kernel stack, I guess ?

Right.

So to do dma from/to a user-supplied buffer
only uvm_vslock() is needed, right ?

Right. Also, make sure you never load in interrupt context (because
you have to lock the user's pmap to extract the physical addresses.)

Thu, Jun 29, 2006 at 08:01:27AM +0900, YAMAMT Takashi wrote:
I made a few more changes An uncompiled & untested diff is attached:

- pass l->l_proc to bus_dmamap_load() so we don't need to double
buffer

do this if you vslock the buffer!

and vslock is fragile

Hum. So we should probably stick with the double-buffer then ?