On Sun, 1 Mar 2020 at 18:24, Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
>
> Hi Antti,
>
> On 3/1/20 7:51 AM, Antti Seppälä wrote:
> > On Sat, 29 Feb 2020 at 18:33, Antti Seppälä <a.seppala@xxxxxxxxx> wrote:
> >>
> >> On Sat, 29 Feb 2020 at 17:25, Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> >>>
> >>> Sigh. It would have been too simple. Too bad I can't test myself.
> >>> I'd like to know if this is because URB_NO_TRANSFER_DMA_MAP is set on a
> >>> transfer, or because the beginning of the buffer indeed needs to be aligned
> >>> to the DMA cache line size on that system. In the latter case, the question
> >>> is why the alignment to DWC2_USB_DMA_ALIGN (=4) works. In the former case,
> >>> question would be why the realignment does any good in the first place.
> >>>
> >>> Any chance you can add some test code to help figuring out what exactly
> >>> goes wrong ?
> >>>
> >>
> >> Sure, I can try to help. Just let me know what code you would like to
> >> insert and where and I'll see what I can do.
> >>
> >
> > So I did some further research on this and it turns out that:
> > - URB_NO_TRANSFER_DMA_MAP is not set on the offending transfers so
> > the issue really is buffer alignment
> > - DWC2_USB_DMA_ALIGN=4 "works" because in my limited testcase (usb
> > 4g-dongle utilized via qmi-wwan) all transfers are unaligned. That is,
> > every urb->transfer_buffer is misaligned by 2 bytes == unaligned
> > - I can fix both issues and thus make the patch work on MIPS by
> > modifying it like this:
> >
> > -#define DWC2_USB_DMA_ALIGN 4
> > +#define DWC2_USB_DMA_ALIGN dma_get_cache_alignment()
> >
> > struct dma_aligned_buffer {
> > void *old_xfer_buffer;
> > - u8 data[0];
> > + u8 data[0] ____cacheline_aligned;
> > };
> >
>
> Thanks for the additional testing. That means that the existing code
> is already broken, or am I missing something ?
>
Yes, I believe the existing code is broken if 4 byte aligned transfers
occur. I seem to have no easy way to prove that though as none of my
devices generate those.
> Updating DWC2_USB_DMA_ALIGN to dma_get_cache_alignment() was part
> of my initial fix, but then I abandoned it because I thought, well,
> the existing alignment works, so that can't be the problem.
>
> Anyway, using ____cacheline_aligned is an excellent idea. I'll make
> the changes suggested above for the next version of my series.
>
Great. In the meanwhile I'll continue testing and will report back if
I find anything new.
--
Antti
