Re: DMA_ATTR_WRITE_COMBINE on mips

From: Maciej W. Rozycki
Date: Wed Sep 18 2019 - 15:48:43 EST


Hi Christoph,

> [I hope the imgtec address still works, but maybe the mips folks know
> if it moved to mips]

Alex left Imagination long before the transition to the interim MIPS
company.

> you added DMA_ATTR_WRITE_COMBINE support in dma_mmap_attrs to mips
> in commit 8c172467be36f7c9591e59b647e4cd342ce2ef41
> ("MIPS: Add implementation of dma_map_ops.mmap()"), but that commit
> only added the support in mmap, not in dma_alloc_attrs. This means
> the memory is now used in kernel space through KSEG1, and thus uncached,
> while for userspace mappings through dma_mmap_* pgprot_writebombine
> is used, which creates a write combine mapping, which on some MIPS CPUs
> sets the _CACHE_UNCACHED_ACCELERATED pte bit instead of the
> _CACHE_UNCACHED one. I know at least on arm, powerpc and x86 such
> mixed page cachability attributes can cause pretty severe problems.
> Are they ok on mips?

The uncached accelerated mode is implementation-specific, so you won't
find its definition in the architecture, however the original R10000
implementation explicitly documents[1] interactions between bus accesses
using the two modes (essentially a _CACHE_UNCACHED store acts as a barrier
for any outstanding _CACHE_UNCACHED_ACCELERATED stores; for loads the
modes are equivalent), so that's clearly supported.

I've glanced over the interAptiv manual[2] too and it seems to define the
caching modes similarly.

> Or was the DMA_ATTR_WRITE_COMBINE supported
> unintended and not correct and we should remove it?

I don't know, so regrettably I can't comment on this.

References:

[1] "MIPS R10000 Microprocessor", Version 2.0, MIPS Technologies, Inc.,
January 29, 1997
<http://techpubs.sgi.com/library/manuals/2000/007-2490-001/pdf/007-2490-001.pdf>

[2] "MIPS32 interAptiv Multiprocessing System Software User's Manual",
Imagination Technologies Ltd., Document Number: MD00904, Revision
02.01, June 15, 2016

Maciej