Revert "[ARM] pxa: remove now unnecessary dma_needs_bounce()"

[Eric Miao]

Hrm... checked again. The IXP4xx build regression seems to have been fixed
by the below commits:

88a5810 ARM: fix IXP4xx build failure
710224f arm: fix "arm: fix pci_set_consistent_dma_mask for dmabounce devices"

Yet the simple introduction of dma_set_coherent_mask() doesn't look
like to be a correct fix to this particular issue.

Barry,

You may want to follow the lead more and help come up with a proper fix.

- eric
On Mon, Mar 21, 2011 at 2:12 PM, Barry Song <21cnbao at gmail.com> wrote:
> Hi Eric/Fujita/Russell,
> As i checked the "remove now unnecessary dma_needs_bounce()" patches
> of Eric, I found you were all involved. Due to compiling error of
> IXP4xx platforms, the patch was reverted by Russell in
> http://ftp.arm.linux.org.uk/git/gitweb.cgi?p=linux-2.6-arm.git;a=commit;h=0485e18bc4112a3b548baa314c24bfbece4d156b
>
> But I think Eric's patch made much sense. dma_needs_bounce() is very
> ambiguous in the whole arm arch, if devices use dma_mask right,
> map_single() will decide whether it needs DMA bounce by the following
> logic:
>        if (dev->dma_mask) {
>                unsigned long mask = *dev->dma_mask;
>                unsigned long limit;
>
>                limit = (mask + 1) & ~mask;
>                if (limit && size > limit) {
>                        dev_err(dev, "DMA mapping too big (requested %#x "
>                                "mask %#Lx)\n", size, *dev->dma_mask);
>                        return ~0;
>                }
>
>                /*
>                 * Figure out if we need to bounce from the DMA mask.
>                 */
>                needs_bounce = (dma_addr | (dma_addr + size - 1)) & ~mask;
>        }
>
> Basically, we don't need dma_needs_bounce() for all chips at all. For
> saa1111, it is needed only because a bug in the SoC.
> For  assabet and pfs168 machines, it gave a workaound by self-defined
> a dma_needs_bounce function.
> return ((machine_is_assabet() || machine_is_pfs168()) &&
>                (addr >= 0xc8000000 || (addr + size) >= 0xc8000000));
>
> I think the right workaround for saa1111 should be setting dma_mask of
> related devices to 0xc8000000 for the assabet and pfs168 machines but
> not add a unnecessary APIs,  which will cause other SoCs need to
> finish the function if they use dma bounce too.
>
> Eric's path make much sense, we should delete the unnecessary
> dma_needs_bounce() from the whole arm arch at all. If all related
> people agree, we need another patch to fix it.
>
> Thanks
> Barry
>