Alan Stern wrote: > On Tue, 3 Aug 2010, Praveena NADAHALLY wrote: > > > Hello, > > Can you please tell your email client to wrap lines after 72 columns or > so? > > > Some USB HOST chip drivers mix DMA and PIO mode. The USB > > Host core in Linux assumes that data is always DMA'ed, if the > > dev->dma_mask is non-zero, and will perform automatic cache > > operations on the data before and after all USB transfers. > > This works fine with chips like EHCI/OHCI, but breaks > > completely on chips like MUSB, which mix PIO and DMA. Our > > simple workaround is to set dev->dma_mask = 0, in the MUSB > > init code. Else we see regular data loss during loopback > > tests. We have been requested to make a better fix for this > > problem. Could the Linux community advice the places we need > > to patch, which gives us the possibility for the USB Host > > chip driver to fully manage cache sync/flush calls through a > > generalized API? The MUSB driver is an OTG driver, setting > > dev->dma_mask to zero will kill DMA operation in gadget mode, > > and is not a long term solution for us. > > The easiest solution is to allow usbcore to map all transfer buffers > for DMA, and then have the HCD call unmap_urb_for_dma in cases where > it wants to use PIO. (You'll have to make unmap_urb_for_dma non-static > and EXPORT it first. And you'll run into trouble if you want to use > PIO with the Setup packet of a control transfer and DMA with the Data > packet.) > > A more complicated solution would be to add intelligence to > map_urb_for_dma so that it would know which URBs will be handled by PIO > even when the DMA mask is set. This is a more fragile approach, > because it relies on keeping the core code in sync with the HCD. > > If necessary, you could add a new callback to struct hc_driver to be > invoked by map_urb_for_dma. The callback routine would return a value > indicating whether the Setup and Data packets should be mapped. > We'd brought up this issue with MUSB a while ago (past discussions are on this thread [1]), but forgot all about it. We'd been carrying a temporary hack in our code all this while (and it came back to haunt us recently). I was about to restart this thread where it left off. I'll post out a patch for RFC in a few minutes. That fixes it for MUSB, and allows co-existence with other controllers that use DMA for all transfers. I'll be glad to rework the patch till it's in an acceptable state for merge. - Anand -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
