Hello,

Sorry for dealy. I was on vacation during last week.

On Thu, Oct 26, 2017 at 07:16:27AM -0700, Tony Lindgren wrote:

* Joonsoo Kim [off-list ref] [171025 21:45]:

quoted

On Wed, Oct 25, 2017 at 10:31:38AM -0700, Tony Lindgren wrote:

quoted

Great, this branch boots on n900! Early parts of the dmesg attached
below.

Sound good. Perhaps, problem is due to the cache management. With my
patchset, direct mapping for CMA area is cleared in
dma_contiguous_remap() if CONFIG_HIGHMEM. I guess that proper cache
operation is required there.

Could you test my newly updated branch again? I re-enable
dma_contiguous_remap() to clear direct mapping for CMA area and add
proper(?) cache operation although I'm not the expert on that area.

https://github.com/JoonsooKim/linux/tree/cma-debug4-next-20180901

Yeah that one boots too, dmesg below. I wonder why flushing the range
with dmac_flush_range() and outer_flush_range() no longer seems enough
though?

I don't know the reason. AFAIK, it should be enough to flush cpu cache
before clearing page table entry, however, it doesn't work for n900.

flush_cache_all() has not only d-cache flushing but also i-cache
flushing. So, I'd like to test effect of i-cache flushing.

Could you test follwing updated branch?

https://github.com/JoonsooKim/linux/tree/cma-debug4-next-20180901

It has three relevant commits on top and enables CMA memory use.

"arm/dma: remove i-cache flush"
"arm/dma: flush i-cache and d-cache separately"
"arm/dma: call flush_cache_all() and don't do outer cache operation"

I think that flush_cache_all() here looks fine because size of CMA
area is usually large enough to use flush_cache_all() but
understanding root cause would be important so I ask this test. If u
don't mind, please test each commit with reverting one by one and let
me know the result of each test.

Reverting the arch/arm/mm/dma-mapping.c changes in your patch
"arm/dma: re-enable to remap the CMA area and flush cache before
remapping" also boots, but produces the following build warnings:

arch/arm/mm/dma-mapping.c: In function 'dma_contiguous_remap':
arch/arm/mm/dma-mapping.c:503:20: warning: passing argument 1 of
'cpu_cache.dma_flush_range' makes pointer from integer without a
cast [-Wint-conversion]
dmac_flush_range(map.virtual, map.virtual + map.length);
                    ^~~
arch/arm/mm/dma-mapping.c:503:20: note: expected 'const void *' but
argument is of type 'long unsigned int'
arch/arm/mm/dma-mapping.c:503:33: warning: passing argument 2 of
'cpu_cache.dma_flush_range' makes pointer from integer without a
cast [-Wint-conversion]
dmac_flush_range(map.virtual, map.virtual + map.length);

Thanks for info.
I think that incorrect type would not be related to this problem.

Thanks.