Re: [PATCH v3] USB: Improve usb_fill_* documentation

From: Ricardo Ribalda
Date: Wed Dec 21 2022 - 14:35:03 EST


Hi Alan

On Wed, 21 Dec 2022 at 18:28, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Wed, Dec 21, 2022 at 11:15:14AM +0100, Ricardo Ribalda wrote:
> > Document the transfer buffer requirement. That is, the buffer must be
> > DMAble - otherwise data corruption might occur.
> >
> > Acked-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> > Signed-off-by: Ricardo Ribalda <ribalda@xxxxxxxxxxxx>
> > ---
> > USB: Improve usb_fill_* documentation
> >
> > After trying to "cleanup" the uvc code, I was patiently explained about
> > the requirements of the urb transfer buffers.
> >
> > Lets make this explicit, so other developers do not make the same mistake.
> >
> > To: Bagas Sanjaya <bagasdotme@xxxxxxxxx>
> > To: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> > To: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
> > To: Christoph Hellwig <hch@xxxxxx>
> > To: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
> > To: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> > Cc: linux-usb@xxxxxxxxxxxxxxx
> > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > ---
> > Changes in v3:
> > - Improve commit message. Thanks Bagas!
> > - Improve field description. Thanks Alan!
> > - Link to v2: https://lore.kernel.org/r/20221220-usb-dmadoc-v2-0-4dd4f198113e@xxxxxxxxxxxx
> >
> > Changes in v2:
> > - s/allocatiing/allocating/ Thanks Randy
> > - Link to v1: https://lore.kernel.org/r/20221220-usb-dmadoc-v1-0-28386d2eb6cd@xxxxxxxxxxxx
> > ---
> > include/linux/usb.h | 27 +++++++++++++++++++++++----
> > 1 file changed, 23 insertions(+), 4 deletions(-)
> >
> > diff --git a/include/linux/usb.h b/include/linux/usb.h
> > index 7d5325d47c45..06cde9ddca97 100644
> > --- a/include/linux/usb.h
> > +++ b/include/linux/usb.h
> > @@ -1626,14 +1626,25 @@ struct urb {
> > * @urb: pointer to the urb to initialize.
> > * @dev: pointer to the struct usb_device for this urb.
> > * @pipe: the endpoint pipe
> > - * @setup_packet: pointer to the setup_packet buffer
> > - * @transfer_buffer: pointer to the transfer buffer
> > + * @setup_packet: pointer to the setup_packet buffer. The buffer must be
> > + * suitable for DMA.
> > + * @transfer_buffer: pointer to the transfer buffer. The buffer must be
> > + * suitable for DMA.
> > * @buffer_length: length of the transfer buffer
> > * @complete_fn: pointer to the usb_complete_t function
> > * @context: what to set the urb context to.
> > *
> > * Initializes a control urb with the proper information needed to submit
> > * it to a device.
> > + *
> > + * The transfer buffer and the setup_packet buffer will most likely be filled
> > + * via DMA.
>
> No, no! The setup_packet buffer will never be filled via DMA -- it is
> _read_ via DMA.
>
> The transfer buffer may be filled via DMA; it depends on the direction
> of the transfer. For OUT transfers, the buffer is read via DMA; for IN
> transfers it is filled via DMA.

I tried to have a short sentence :), but you are right, let's make it
explicit, at the end of the day that is the whole purpose of this
patch :)

I have written now: will most likely be filled or read via DMA.

>
> > The simplest way to get a buffer that can be DMAed to is
> > + * allocating it via kmalloc() or equivalent, even for very small buffers.
> > + * If the buffers are embedded in a bigger structure, there is a risk that
> > + * the buffer itself, the previous field and/or the next field are corrupted
> > + * due to cache incoherencies; or slowed down if they are evicted from the
> > + * cache.
> > + *
> > */
> > static inline void usb_fill_control_urb(struct urb *urb,
> > struct usb_device *dev,
> > @@ -1658,13 +1669,17 @@ static inline void usb_fill_control_urb(struct urb *urb,
> > * @urb: pointer to the urb to initialize.
> > * @dev: pointer to the struct usb_device for this urb.
> > * @pipe: the endpoint pipe
> > - * @transfer_buffer: pointer to the transfer buffer
> > + * @transfer_buffer: pointer to the transfer buffer. The buffer must be
> > + * suitable for DMA.
> > * @buffer_length: length of the transfer buffer
> > * @complete_fn: pointer to the usb_complete_t function
> > * @context: what to set the urb context to.
> > *
> > * Initializes a bulk urb with the proper information needed to submit it
> > * to a device.
> > + *
> > + * Please refer to usb_fill_control_urb() for a description of the
> > + * requirements for transfer_buffer.
>
> _This_ right here is the kerneldoc for usb_fill_control_urb()! You
> should have written "Please refer to the kerneldoc for struct urb".
>

Please try applying the patch. I believe it is on the right location.
diff does not know about kerneldoc and this is why it copies the
previous header (struct urb), instead of usb_fill_control_urb.

This is why it seems that all the messages are shifted one location up.

I would also prefer to keep the: refer to usb_fill_control_urb, as the
functions are more indexed than the structures: Eg:
https://manpages.debian.org/testing/linux-manual-4.8/usb_fill_control_urb.9.en.html

What about adding a reference to struct urb at usb_fill_control_urb() ?



> > */
> > static inline void usb_fill_bulk_urb(struct urb *urb,
> > struct usb_device *dev,
> > @@ -1687,7 +1702,8 @@ static inline void usb_fill_bulk_urb(struct urb *urb,
> > * @urb: pointer to the urb to initialize.
> > * @dev: pointer to the struct usb_device for this urb.
> > * @pipe: the endpoint pipe
> > - * @transfer_buffer: pointer to the transfer buffer
> > + * @transfer_buffer: pointer to the transfer buffer. The buffer must be
> > + * suitable for DMA.
> > * @buffer_length: length of the transfer buffer
> > * @complete_fn: pointer to the usb_complete_t function
> > * @context: what to set the urb context to.
> > @@ -1697,6 +1713,9 @@ static inline void usb_fill_bulk_urb(struct urb *urb,
> > * Initializes a interrupt urb with the proper information needed to submit
> > * it to a device.
> > *
> > + * Please refer to usb_fill_control_urb() for a description of the
> > + * requirements for transfer_buffer.
>
> Similar to the above.
>
> > + *
> > * Note that High Speed and SuperSpeed(+) interrupt endpoints use a logarithmic
> > * encoding of the endpoint interval, and express polling intervals in
> > * microframes (eight per millisecond) rather than in frames (one per
>
> Why didn't you also make the same change to usb_fill_int_urb()?
>
> Alan Stern
>
> >
> > ---
> > base-commit: b6bb9676f2165d518b35ba3bea5f1fcfc0d969bf
> > change-id: 20221220-usb-dmadoc-29384acebd48
> >
> > Best regards,
> > --
> > Ricardo Ribalda <ribalda@xxxxxxxxxxxx>



--
Ricardo Ribalda