On Tue, Jun 25, 2019 at 02:26:41PM +0200, Christoph Hellwig wrote:

Hi Paul, Dale and Mark (I hope this reaches the right Mark),

Hi Christoph.  Yes, you did reach the right Mark.  :)

I've started auditing all users of DMA_ATTR_NON_CONSISTENT ot prepare
for major API improvements in that area.

One of the odd users is the mpsc ѕerial driver, which allocates DMA
memory with the above flag, and then actually properly calls
dma_cache_sync.  So far, so good.  But it turns out it also has
"#if defined(CONFIG_PPC32) && !defined(CONFIG_NOT_COHERENT_CACHE)"
ifdef blocks next to the dma_cache_sync calls that perform cache
maintainance for platforms that according to the ifdef claim to
be cache coherent.  According to the Kconfig the driver can
only build if the MV64X60 symbol is set, which is a ppc embedded 6xx
SOC, which appears to be configurable as either cache coherent, or
not.  But according to the code in the driver at least this device
always is not cache coherent.

It seems like we need to always mark that platform as potentially
not coherent, and then use the per-device flag to mark all device
except for this one as coherent.  Or did I miss anything?  Maybe
all this is actually dead code and can go away?

Yeah, the mpsc driver had lots of ugly cache related hacks because of
cache coherency bugs in the early version of the MV64x60 bridge chips
that it was embedded in.  That chip is pretty much dead now and I've
removed core support for it from the powerpc tree.  Removing the mpsc
driver is on my todo list but I've been busy and lazy.  So, to sum it
up, don't spend any more time worrying about it as it should be removed.

I'll post a patch to do that tonight and I'm sorry for any time you've
spent looking at it so far.

Mark
--