Re: [PATCH] drm/prime: Only call dma_map_sgtable() for devices with DMA support

[Alan Stern <stern@rowland.harvard.edu>]

On Fri, Feb 19, 2021 at 04:56:16PM +0100, Christian König wrote:
> 
> 
> Am 19.02.21 um 16:53 schrieb Alan Stern:
> > On Fri, Feb 19, 2021 at 02:45:54PM +0100, Christian König wrote:
> > > Well as far as I can see this is a relative clear NAK.
> > > 
> > > When a device can't do DMA and has no DMA mask then why it is requesting an
> > > sg-table in the first place?
> > This may not be important for your discussion, but I'd like to give an
> > answer to the question -- at least, for the case of USB.
> > 
> > A USB device cannot do DMA and has no DMA mask.  Nevertheless, if you
> > want to send large amounts of bulk data to/from a USB device then using
> > an SG table is often a good way to do it.  The reason is simple: All
> > communication with a USB device has to go through a USB host controller,
> > and many (though not all) host controllers _can_ do DMA and _do_ have a
> > DMA mask.
> > 
> > The USB mass-storage and uas drivers in particular make heavy use of
> > this mechanism.
> 
> Yeah, I was assuming something like that would work.
> 
> But in this case the USB device should give the host controllers device
> structure to the dma_buf_attach function so that the sg_table can be filled
> in with DMA addresses properly.

Indeed.  Although in the contexts I'm familiar with, the host controller 
device is actually passed to routines like dma_pool_create, 
dma_alloc_coherent, dma_map_sg, or dma_map_single.

Alan Stern