busdma problem

[Andrew Thompson]

On Fri, Jan 30, 2009 at 02:07:14PM -0800, Andrew Thompson wrote:
> On Fri, Jan 30, 2009 at 09:31:47AM -0800, Andrew Thompson wrote:
> > >> I am having an issue with busdma when bounce buffers are used. I have
> > >> patched _bus_dmamap_sync_bp() to print out the details when a bounce
> > >> happens and also print the driver buffer before and after.
> > >> 
> > >> During normal dma everything is fine,
> > >> 
> > >> Before: 0xc7c1ab40 data=c1:4b:a4:80:c0:5d:ed:78:00:00:08:0d:c1:1f:46:78:00:00:20:02:00:00:20:02:
> > >> [...do dma...]
> > >> After:  0xc7c1ab40 data=2c:03:4e:00:6f:00:76:00:61:00:74:00:65:00:6c:00:20:00:57:00:69:00:72:00:
> > >> 
> > >> The buffer 2c:03:4e:00:... is the correct response from the hardware.
> > >> When a bounce buffer is used I see the correct data come in and be
> > >> bcopy'd to my memory region but it is not visible when read later.
> > >> 
> > >> Before: 0xc7c29b40 data=c1:50:19:00:c0:5d:ed:f8:00:00:08:0d:c1:1f:46:78:00:00:20:02:00:00:20:02:
> > >> dma bounced 0x1271000 -> 0xc7c29b40 len=193 data=2c:03:4e:00:6f:00:76:00:61:00:74:00:65:00:6c:00:20:00:57:00:69:00:72:00:
> > >> After:  0xc7c29b40 data=c1:50:19:00:c0:5d:ed:f8:00:00:08:0d:c1:1f:46:78:00:00:20:02:00:00:20:02:
> > >> 
> > >> 
> > >> This is on an xscale ixp425 with 128m memory, the PCI dma tag is limited
> > >> to 64m.
> > >>   
> > > What device is involved?  Is this on HEAD?
> > 
> > This is usb/ehci. The specific function I am looking at is
> > usbd_get_string() in usbdi.c, it does a usb request to fill
> > usb_string_descriptor_t that is a stack variable.
> 
> As suggested by Sam, this works properly when the buffer is malloc'd
> instead of taken from the stack. So is this now a bug or a feature??

As a test I removed the checking of KENTER_CACHE in
arm/arm/pmap.c:pmap_kenter_internal() so all memory is uncached and now
dma bounces to a stack variable work.

busdma remaps the address nocache into vaddr_nocache pointer and uses
that for the bcopy, obviously arm_remap_nocache() is not working
correctly.

Andrew