Re: [PATCH v2] mtd: onenand: omap2: Disable DMA for HIGHMEM buffers

[Ladislav Michl]

Hi Boris,

(and apologies for delay)

On Fri, Apr 20, 2018 at 10:01:34PM +0200, Boris Brezillon wrote:
> Hi Ladislav,
> 
> On Mon, 16 Apr 2018 08:52:59 +0200
> Ladislav Michl <ladis@xxxxxxxxxxxxxx> wrote:
> 
> > dma_map_single doesn't get the proper DMA address for vmalloced area,
> 
> That's not true, it returns the right DMA (physical) address, it's just
> that:

To be honest I used log message from commit dcf08227e964 which is dealing
with the same issue.

> 1/ the memory location is not necessarily physically contiguous
> 2/ in case your arch is VIVT ot VIPT, there may be several entries in
>    the cache pointing to the same physical location, and the cache
>    maintenance operations done by dma_map_single() will only operate on
>    one of these entries.

Well, there are few things suspicious here: DMA is used to transfer data
to system memory from bufferram which is 3kB in size. Nikolaus reported
Internal error: Oops: 805 [#1] PREEMPT SMP ARM
PC is at v7_dma_inv_range+0x30/0x48
LR is at dma_cache_maint_page+0xd0/0xe0
but I do not see how could it happen if it would be caused by 1/ or 2/ above.
Cache issue would cause data inconsistencies which should be caught by upper
MTD layers.

And given that replacing dma_map_single with dma_map_page done in first
version of patch also fixed oops, it is not caused by 1/
CPU is OMAP3630 based on Cortex-A8, D-cache is often described as PIPT but
seems to be VIPT with built-in alias detection mechanism - anyone with
better knowledge will eventually correct me, thank you.

> > so disable DMA in this case.
> > 
> 
> The fix looks good though. Can you rephrase your commit message to make
> it clearer.

Sure, I'd like to, but I do not know what's root cause yet :)

> Thanks,
> 
> Boris
	
Best regards,
	ladis