Re: [RFC PATCH v2 8/9] USB: add HCD_NO_COHERENT_MEM host controller driver flag

[Albert Herranz <albert_herranz@yahoo.es>]

Alan Stern wrote:
> On Mon, 1 Mar 2010, Albert Herranz wrote:
> 
>>> 	If transfer_buffer_length is 0 then do nothing.
>>> 	Otherwise if num_sgs > 0 then do nothing.
>>> 	Otherwise if URB_NO_TRANSFER_DMA_MAP and transfer_dma
>>> 		are both set (this avoids your HCD_NO_COHERENT_MEM
>>> 		case) then do nothing.
>>>
>> I see. This case would include the PIO case too (for which dma_handle
>> is set to all 1s).
> 
> The test above should be transfer_dma != ~0, not transfer_dma != 0,
> since ~0 means the DMA address isn't set.  In fact I forgot to 
> include the PIO case; it should be handled by changing the remaining 
> tests as follows:
> 
> 	Otherwise if hcd->self.uses_dma is set then
> 		If this URB doesn't require PIO then call dma_map_single
> 	Otherwise if HCD_LOCAL_MEM is set then call hcd_alloc_coherent
> 	Otherwise do nothing (PIO case).
> 
> Currently "this URB doesn't require PIO" is always true, but in the 
> future it won't be.
> 

Can this be currently tested?
Should I make provisions for this check now too?

>> So this assumes that transfer_dma should be set initially to 0 when
>> allocating USB buffers for HCD_NO_COHERENT_MEM.
> 
> No, it should be set to ~0, the same as when buffers are allocated for 
> a PIO-based controller.
> 

This logic now resembles more the one in my v2 proposal although with different formal checks.
I think I'll code and post another iteration of the 8/9 patch with your proposed checks and then we can continue further discussion on it.
I'll try to add explanatory comments for each check.

> Alan Stern
> 

Thanks a lot for your input,
Albert