[PATCH v6 08/12] device core: Introduce DMA range map, supplanting dma_pfn_offset

[Jim Quinlan]

Hi Andy,

Sorry for the delay in response.  I will do what you suggest in your
email.  I do have one response to one of your comments below.

On Thu, Jul 2, 2020 at 4:43 AM Andy Shevchenko
<andriy.shevchenko at linux.intePHYSl.com> wrote:
>
> On Wed, Jul 01, 2020 at 05:21:38PM -0400, Jim Quinlan wrote:
> > The new field 'dma_range_map' in struct device is used to facilitate the
> > use of single or multiple offsets between mapping regions of cpu addrs and
> > dma addrs.  It subsumes the role of "dev->dma_pfn_offset" which was only
> > capable of holding a single uniform offset and had no region bounds
> > checking.
> >
> > The function of_dma_get_range() has been modified so that it takes a single
> > argument -- the device node -- and returns a map, NULL, or an error code.
> > The map is an array that holds the information regarding the DMA regions.
> > Each range entry contains the address offset, the cpu_start address, the
> > dma_start address, and the size of the region.
> >
> > of_dma_configure() is the typical manner to set range offsets but there are
> > a number of ad hoc assignments to "dev->dma_pfn_offset" in the kernel
> > driver code.  These cases now invoke the function
> > dma_attach_offset_range(dev, cpu_addr, dma_addr, size).
>
> ...
>
> > +     if (dev && dev->dma_range_map)
> > +             pfn -= (unsigned long)PFN_DOWN(dma_offset_from_phys_addr(dev, PFN_PHYS(pfn)));
>
> Instead of casting use PHYS_PFN() and it will be consistent with latter in the same line.
>
> > +     if (dev && dev->dma_range_map)
> > +             pfn += (unsigned long)PFN_DOWN(dma_offset_from_dma_addr(dev, addr));
>
> Ditto.
>
> ...
>
> > +             dev_err(dev, "set dma_offset%08llx%s\n", KEYSTONE_HIGH_PHYS_START
> > +                     - KEYSTONE_LOW_PHYS_START, ret ? " failed" : "");
>
> Please, avoid such indentation.
> Better split it to the three lines, argument per line (expect dev which will go
> on the first one).
>
> This applies to all similar places.
>
> ...
>
> >       unsigned long pfn = (dma_handle >> PAGE_SHIFT);
>
> PHYS_PFN() / PFN_DOWN() ?
>
> > +     if (!WARN_ON(!dev) && dev->dma_range_map)
> > +             pfn += (unsigned long)PFN_DOWN(dma_offset_from_dma_addr(dev, dma_handle));
>
> PHYS_PFN() ?
>
> ...
>
> > +     r = kcalloc(num_ranges + 1, sizeof(struct bus_dma_region), GFP_KERNEL);
>
> sizeof(*r) ?
>
> > +     if (!r)
> > +             return ERR_PTR(-ENOMEM);
>
> ...
>
> > +     ret = IS_ERR(map) ? PTR_ERR(map) : 0;
>
> PTR_ERR_OR_ZERO()
>
> ...
>
> > +             /* We want the offset map to be device-managed, so alloc & copy */
> > +             dev->dma_range_map = devm_kcalloc(dev, num_ranges + 1, sizeof(*r),
> > +                                               GFP_KERNEL);
>
> The question is how many times per device lifetime this can be called?
Someone else questioned this.  There are two cases that come to mind:
our overnight test which load/unloads the RC driver over and over, and
a customer that does an unbind/bind  of the RC driver when their EP is
hung and wants to restart.  Both cases are atypical and are weak
arguments to justify the second copy.  I will drop the copy.

Thanks,
Jim Quinlan
Broadcom STB
>
> ...
>
>
> > +             if (!dev->dma_range_map)
> > +                     return -ENOMEM;
> > +             memcpy((void *)dev->dma_range_map, map, sizeof(*r) * num_ranges + 1);
>
> If it's continuous, perhaps kmemdup() ?
>
> ...
>
> > +     rc = IS_ERR(map) ? PTR_ERR(map) : 0;
>
> PTR_ERR_OR_ZERO()
>
> ...
>
> > +                     = dma_offset_from_phys_addr(dev, PFN_PHYS(mem->pfn_base));
> > +
> > +             return (dma_addr_t)PFN_PHYS(mem->pfn_base) - dma_offset;
>
> Looking at this more, I think you need to introduce in the same header (pfn.h)
> something like:
>
>         #define PFN_DMA_ADDR()
>         #define DMA_ADDR_PFN()
>
> --
> With Best Regards,
> Andy Shevchenko
>
>