Re: [PATCH v2] usb: core: setup dma_pfn_offset for USB devices and, interfaces

[Alan Stern]

On Mon, 12 Sep 2016, Roger Quadros wrote:

> If dma_pfn_offset is not inherited correctly from the host controller,
> it might result in sub-optimal configuration as bounce
> buffer limit might be set to less than optimal level.
> 
> Consider the mass storage device case.
> USB storage driver creates a scsi host for the mass storage interface in
> drivers/usb/storage/usb.c
> The scsi host parent device is nothing but the the USB interface device.
> Now, __scsi_init_queue() calls scsi_calculate_bounce_limit() to find out
> and set the block layer bounce limit.
> scsi_calculate_bounce_limit() uses dma_max_pfn(host_dev) to get the
> bounce_limit. host_dev is nothing but the device representing the
> mass storage interface.
> If that device doesn't have the right dma_pfn_offset, then dma_max_pfn()
> is messed up and the bounce buffer limit is wrong.
> 
> e.g. On Keystone 2 systems, dma_max_pfn() is 0x87FFFF and dma_mask_pfn
> is 0xFFFFF. Consider a mass storage use case: Without this patch,
> usb scsi host device (usb-storage) will get a dma_pfn_offset of 0 resulting
> in a dma_max_pfn() of 0xFFFFF within the scsi layer
> (scsi_calculate_bounce_limit()).
> This will result in bounce buffers being unnecessarily used.
> 
> Hint: On 32-bit ARM platforms dma_max_pfn() = dma_mask_pfn + dma_pfn_offset
> 
> Signed-off-by: Roger Quadros <rogerq@xxxxxx>
> ---

How does v2 of this patch differ from v1?

>  drivers/usb/core/message.c | 6 ++++++
>  drivers/usb/core/usb.c     | 7 +++++++
>  2 files changed, 13 insertions(+)
> 
> diff --git a/drivers/usb/core/message.c b/drivers/usb/core/message.c
> index 0406a59..66364ea 100644
> --- a/drivers/usb/core/message.c
> +++ b/drivers/usb/core/message.c
> @@ -1863,6 +1863,12 @@ free_interfaces:
>  		intf->dev.type = &usb_if_device_type;
>  		intf->dev.groups = usb_interface_groups;
>  		intf->dev.dma_mask = dev->dev.dma_mask;
> +		/* Propagate dma_pfn_offset to USB interface.
> +		 * This is especially required by mass storage interface
> +		 * which relies on SCSI layer and scsi_calculate_bounce_limit()
> +		 * to set the bounce buffer limit based on dma_pfn_offset.
> +		 */
> +		intf->dev.dma_pfn_offset = dev->dev.dma_pfn_offset;

I really think this comment isn't necessary.  It seems pretty obvious 
that if you're copying DMA-related fields from one device to another 
then you'll want to copy the dma_pfn_offset along with everything else.
No?

The explanation in the Changelog is enough, IMO.

>  		INIT_WORK(&intf->reset_ws, __usb_queue_reset_device);
>  		intf->minor = -1;
>  		device_initialize(&intf->dev);
> diff --git a/drivers/usb/core/usb.c b/drivers/usb/core/usb.c
> index 5e80697..7dde93d 100644
> --- a/drivers/usb/core/usb.c
> +++ b/drivers/usb/core/usb.c
> @@ -441,6 +441,13 @@ struct usb_device *usb_alloc_dev(struct usb_device *parent,
>  	dev->dev.type = &usb_device_type;
>  	dev->dev.groups = usb_device_groups;
>  	dev->dev.dma_mask = bus->controller->dma_mask;
> +	/* Propagate bus controller's dma_pfn_offset to USB device,
> +	 * as it would be needed to be propagated to USB interfaces.
> +	 * This is especially required by mass storage interface
> +	 * which relies on SCSI layer and scsi_calculate_bounce_limit()
> +	 * to set the bounce buffer limit based on dma_pfn_offset.
> +	 */
> +	dev->dev.dma_pfn_offset = bus->controller->dma_pfn_offset;

Same here.

>  	set_dev_node(&dev->dev, dev_to_node(bus->controller));
>  	dev->state = USB_STATE_ATTACHED;
>  	dev->lpm_disable_count = 1;

Alan Stern