On Tue, 2 Mar 2010, Albert Herranz wrote:
> The HCD_NO_COHERENT_MEM USB host controller driver flag can be enabled
> to instruct the USB stack to avoid allocating coherent memory for USB
> buffers.
This patch appears to be based on your earlier work, not on the current
development kernel.
> --- a/drivers/usb/core/buffer.c
> +++ b/drivers/usb/core/buffer.c
> @@ -53,8 +53,9 @@ int hcd_buffer_create(struct usb_hcd *hcd)
> char name[16];
> int i, size;
>
> - if (!hcd->self.controller->dma_mask &&
> - !(hcd->driver->flags & HCD_LOCAL_MEM))
> + if ((!hcd->self.controller->dma_mask &&
> + !(hcd->driver->flags & HCD_LOCAL_MEM)) ||
> + (hcd->driver->flags & HCD_NO_COHERENT_MEM))
> return 0;
>
> for (i = 0; i < HCD_BUFFER_POOLS; i++) {
> @@ -109,8 +110,9 @@ void *hcd_buffer_alloc(
> int i;
>
> /* some USB hosts just use PIO */
> - if (!bus->controller->dma_mask &&
> - !(hcd->driver->flags & HCD_LOCAL_MEM)) {
> + if ((!bus->controller->dma_mask &&
> + !(hcd->driver->flags & HCD_LOCAL_MEM)) ||
> + (hcd->driver->flags & HCD_NO_COHERENT_MEM)) {
> *dma = ~(dma_addr_t) 0;
> return kmalloc(size, mem_flags);
> }
> @@ -135,8 +137,9 @@ void hcd_buffer_free(
> if (!addr)
> return;
>
> - if (!bus->controller->dma_mask &&
> - !(hcd->driver->flags & HCD_LOCAL_MEM)) {
> + if ((!bus->controller->dma_mask &&
> + !(hcd->driver->flags & HCD_LOCAL_MEM)) ||
> + (hcd->driver->flags & HCD_NO_COHERENT_MEM)) {
> kfree(addr);
> return;
> }
These three tests should be encapsulated in a single subroutine.
> --- a/drivers/usb/core/hcd.c
> +++ b/drivers/usb/core/hcd.c
> @@ -1267,6 +1267,16 @@ static void unmap_urb_setup_packet(struct usb_hcd *hcd, struct urb *urb)
> dma_unmap_single(hcd->self.controller, urb->setup_dma,
> sizeof(struct usb_ctrlrequest),
> DMA_TO_DEVICE);
> + /*
> + * For the HCD_NO_COHERENT_MEM case we need to reset
> + * setup_dma to its original value if the buffer was
> + * allocated via usb_buffer_alloc().
> + * This is necessary to force a new streaming DMA mapping
> + * when the same buffer is used again as the setup packet.
This is a little fragile. I think it would be better to move the test
for HCD_NO_COHERENT_MEM into the submission path, by combining it with
the URB_NO_SETUP_DMA_MAP test. Evidently this case has to be covered
either during submission or during giveback, and doing it during
submission would be cleaner.
The same goes for URB_NO_TRANSFER_DMA_SETUP, of course.
Alan Stern
--
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
