On Wed, Nov 26, 2014 at 1:28 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> For the BMIPS case:
>>
>> plat_map_dma_mem* and plat_dma_addr_to_phys are just performing
>> remapping, so dma-ranges would work.
>>
>> plat_unmap_dma_mem is used to perform an extra BMIPS-specific
>> cacheflush operation.
>
> Yes, the cacheflush again would have to be abstracted. This is normally
> done using either a platform-specific dma_map_ops struct, or using
> a further abstraction with another function pointer.
>
> I'm surprised that you need the special flush operation only
> for 'unmap' and not for 'dma_sync_*_for_cpu'. Can you check that
> you are actually doing the right thing for drivers that reuse
> a DMA buffer with the streaming API?
That's a fantastic question. I checked an older BCM7xxx kernel tree
and noticed that we used to implement plat_extra_sync_for_device()
locally to do this, but the API was "garbage collected" last year:
commit 4e7f72660c39a81cc5745d5c6f23f9500f80d8d8
Author: Felix Fietkau <nbd@xxxxxxxxxxx>
Date: Thu Aug 15 11:28:30 2013 +0200
MIPS: Remove unnecessary platform dma helper functions
The semantics stay the same - on Cavium Octeon the functions were dead
code (it overrides the MIPS DMA ops) - on other platforms they contained
no code at all.
Signed-off-by: Felix Fietkau <nbd@xxxxxxxxxxx>
Cc: linux-mips@xxxxxxxxxxxxxx
Patchwork: https://patchwork.linux-mips.org/patch/5720/
Signed-off-by: Ralf Baechle <ralf@xxxxxxxxxxxxxx>
If nobody objects, I can revert this change and then add the extra
flush to arch/mips/bmips/dma.c.
Or I can add some BMIPS-specific code to dma-default.c, similar to
cpu_needs_post_dma_flush(). Could even add it to that function
directly, although that is less intuitive:
diff --git a/arch/mips/mm/dma-default.c b/arch/mips/mm/dma-default.c
index af5f046..ee6d12c 100644
--- a/arch/mips/mm/dma-default.c
+++ b/arch/mips/mm/dma-default.c
@@ -18,6 +18,7 @@
#include <linux/highmem.h>
#include <linux/dma-contiguous.h>
+#include <asm/bmips.h>
#include <asm/cache.h>
#include <asm/cpu-type.h>
#include <asm/io.h>
@@ -69,6 +70,18 @@ static inline struct page *dma_addr_to_page(struct
device *dev,
*/
static inline int cpu_needs_post_dma_flush(struct device *dev)
{
+ if (boot_cpu_type() == CPU_BMIPS3300 ||
+ boot_cpu_type() == CPU_BMIPS4350 ||
+ boot_cpu_type() == CPU_BMIPS4380) {
+ void __iomem *cbr = BMIPS_GET_CBR();
+
+ /* Flush stale data out of the readahead cache */
+ __raw_writel(0x100, cbr + BMIPS_RAC_CONFIG);
+ __raw_readl(cbr + BMIPS_RAC_CONFIG);
+
+ return 0;
+ }
+
return !plat_device_is_coherent(dev) &&
(boot_cpu_type() == CPU_R10000 ||
boot_cpu_type() == CPU_R12000 ||
This would allow for eliminating all flushes from
arch/mips/bmips/dma.c, so the entire file could go away when we switch
to the common dma-ranges code.
Any preferences?
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
