On Thu, Mar 28, 2013 at 03:44:50PM +0200, Ruslan Bilovol wrote:
> Hi Felipe,
>
> On Wed, Mar 27, 2013 at 3:17 PM, Felipe Balbi <balbi@xxxxxx> wrote:
> > Hi,
> >
> > On Thu, Mar 14, 2013 at 08:12:09PM +0200, Ruslan Bilovol wrote:
> >> MUSB controller cannot work in DMA mode with misaligned buffers,
> >> switching in PIO mode.
> >>
> >> HCD core has hooks that allow to override the default DMA
> >> mapping and unmapping routines for host controllers that have
> >> special DMA requirements, such as alignment contraints.
> >>
> >> It is observed that work in PIO mode is slow and it's better
> >> to align buffers properly before passing them to MUSB
> >>
> >> This increased throughput 80->120 MBits/s over musb@omap4 with
> >> USB Gigabit ethernet adapter attached.
> >>
> >> Some ideas taken from ehci-tegra.c
> >>
> >> Signed-off-by: Ruslan Bilovol <ruslan.bilovol@xxxxxx>
> >> ---
> >> drivers/usb/musb/musb_core.c | 14 ++++++
> >> drivers/usb/musb/musb_host.c | 102 +++++++++++++++++++++++++++++++++++++++++-
> >> drivers/usb/musb/musb_host.h | 2 +-
> >> 3 files changed, 116 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/usb/musb/musb_core.c b/drivers/usb/musb/musb_core.c
> >> index 60b41cc..91ac166 100644
> >> --- a/drivers/usb/musb/musb_core.c
> >> +++ b/drivers/usb/musb/musb_core.c
> >> @@ -1431,6 +1431,20 @@ static int musb_core_init(u16 musb_type, struct musb *musb)
> >>
> >> /* log release info */
> >> musb->hwvers = musb_read_hwvers(mbase);
> >> +
> >> +#ifndef CONFIG_MUSB_PIO_ONLY
> >> + /*
> >> + * The DMA engine in RTL1.8 and above cannot handle
> >> + * DMA addresses that are not aligned to a 4 byte boundary.
> >> + * For such engine implemented (un)map_urb_for_dma hooks.
> >> + * Do not use these hooks for RTL<1.8
> >> + */
> >> + if (musb->hwvers < MUSB_HWVERS_1800) {
> >
> > if you move this check to map/unmap and always return error if this is
> > true, you can avoid removing 'const' from our struct hc_driver. Would
> > that work ?
>
> If we return an error in map/unmap callbacks, this will break urb transferring,
> however I can call core function usb_hcd_(un)map_urb_for_dma() instead of
> returning the error (and that is default behavior if we do not have
> map/unmap callbacks
> set for the hc driver) so I can avoid removing 'const' from our struct
> hc_driver and this will work.
> The side effect will be only in small overhead for this path.
>
> So, will be this OK for you? I will send v3 in this case.
should be alright. Thanks
--
balbi
Attachment:
signature.asc
Description: Digital signature
