[PATCH fpga 9/9] fpga: Remove support for non-sg drivers

[atull]

On Thu, 10 Nov 2016, Joshua Clayton wrote:

> 
> 
> On 11/10/2016 08:33 AM, Jason Gunthorpe wrote:
> >>>  struct fpga_manager_ops {
> >>>  	enum fpga_mgr_states (*state)(struct fpga_manager *mgr);
> >>> -	int (*write_init)(struct fpga_manager *mgr, u32 flags,
> >>> -			  const char *buf, size_t count);
> >>> -	int (*write)(struct fpga_manager *mgr, const char *buf, size_t count);
> >>>  	int (*write_init_sg)(struct fpga_manager *mgr, u32 flags,
> >>>  			     struct sg_table *sgt);
> >>>  	int (*write_sg)(struct fpga_manager *mgr, struct sg_table *sgt);
> >>> @@ -118,6 +113,8 @@ struct fpga_manager {
> >>>  
> >>>  int fpga_mgr_buf_load(struct fpga_manager *mgr, u32 flags,
> >>>  		      const char *buf, size_t count);
> >>> +int fpga_mgr_buf_load_sg(struct fpga_manager *mgr, u32 flags,
> >>> +			 struct sg_table *sgt);
> >>>  
> >>>  int fpga_mgr_firmware_load(struct fpga_manager *mgr, u32 flags,
> >>>  			   const char *image_name);
> >> I don't have any feeling either way about switching to scatter-gather.
> >> (Not zynq or socfpga user)
> >> But I do object to renaming the API.
> >> write_init() and write() do not imply a particular implementation, nor even that
> >> the buffer is coherent.
> > Neither the sg or old linear interface imply any particular
> > underlying driver implementation.
> >
> > All that is being changed is how the list of physical pages gets
> > passed to the driver. The linear interface requires them to be
> > contiguously mapped (eg in a vmap) while the SG interface
> > directly passes a list of physical page addresses.
> >
> > Any alogrithm that works with the old interface can run on the new
> > interface, and the new interface can support much better options for
> > DMA drivers, while not requiring the higher layers to perform a high
> > order allocation (vmap or otherwise) to create the contiguous memory.
> >
> > The reason the old interface is being deleted here is so the fpga mgr
> > API can be expanded to accept a sg list directly. Since we cannot
> > convert a general sg list to linear memory the liner option must be
> > totally removed.
> OK. That sounds reasonable.
> >> I am working to merge an fpga manager which uses SPI to load the bitstream
> >> (see https://www.spinics.net/lists/arm-kernel/msg539328.html)
> >> Any dma in use there would come from the spi driver. write_init_sg, and write_sg
> >> don't make any sense in my case.
> > No, it still make lots of sense.
> >
> > SPI has been slowly transforming to use the same sort of SG scheme
> > universally, including facing the client. (see
> > 6ad45a27cbe343ec8d7888e5edf6335499a4b555)
> Thanks for sharing that link.
> >
> > Some day your driver can just pass the SGs directly to spi and
> > everything will be great.
> >
> > In the mean time it can do sg_miter_next to get mapped buffers.
> ..so I have a way forward if this series gets merged.
> That was my main concern.
> And as a dma n00b, learning to use dma engine structures to do
> non-dma xfer was not very high on my list.
> >> Would it not make sense to keep the top level API the same?
> > Fundamentally no.

NACK for now.  Let's slow down here a bit.

I don't see any rush in getting rid of the contiguous
buffer interface.

At the very least, my socfpga-a10 driver is using the old
interface and has been just applied.  And currently your
socfpga changes are untested whereas my original driver
has been in use and is known to work.

Let's wait on that patch until we have quite a few FPGA's
supported and can be sure that we won't miss the old
interface.

Alan

> >
> > Jason
> Joshua
>