[PATCH 1/2] dma-mapping: add a dma_ops_bypass flag to struct device

[Christoph Hellwig]

On Fri, Apr 03, 2020 at 07:38:11PM +1100, Alexey Kardashevskiy wrote:
> 
> 
> On 26/03/2020 12:26, Alexey Kardashevskiy wrote:
> > 
> > 
> > On 25/03/2020 19:37, Christoph Hellwig wrote:
> >> On Wed, Mar 25, 2020 at 03:51:36PM +1100, Alexey Kardashevskiy wrote:
> >>>>> This is for persistent memory which you can DMA to/from but yet it does
> >>>>> not appear in the system as a normal memory and therefore requires
> >>>>> special handling anyway (O_DIRECT or DAX, I do not know the exact
> >>>>> mechanics). All other devices in the system should just run as usual,
> >>>>> i.e. use 1:1 mapping if possible.
> >>>>
> >>>> On other systems (x86 and arm) pmem as long as it is page backed does
> >>>> not require any special handling.  This must be some weird way powerpc
> >>>> fucked up again, and I suspect you'll have to suffer from it.
> >>>
> >>>
> >>> It does not matter if it is backed by pages or not, the problem may also
> >>> appear if we wanted for example p2p PCI via IOMMU (between PHBs) and
> >>> MMIO might be mapped way too high in the system address space and make
> >>> 1:1 impossible.
> >>
> >> How can it be mapped too high for a direct mapping with a 64-bit DMA
> >> mask?
> > 
> > The window size is limited and often it is not even sparse. It requires
> > an 8 byte entry per an IOMMU page (which is most commonly is 64k max) so
> > 1TB limit (a guest RAM size) is a quite real thing. MMIO is mapped to
> > guest physical address space outside of this 1TB (on PPC).
> > 
> > 
> 
> I am trying now this approach on top of yours "dma-bypass.3" (it is
> "wip", needs an upper boundary check):
> 
> https://github.com/aik/linux/commit/49d73c7771e3f6054804f6cfa80b4e320111662d
> 
> Do you see any serious problem with this approach? Thanks!

Do you have a link to the whole branch?  The github UI is unfortunately
unusable for that (or I'm missing something).