[BUG 5.14] arm64/mm: dma memory mapping fails (in some cases)

[Catalin Marinas]

On Sat, Sep 18, 2021 at 02:39:49PM +0300, Mike Rapoport wrote:
> On Sat, Sep 18, 2021 at 11:37:22AM +0300, Mike Rapoport wrote:
> > On Sat, Sep 18, 2021 at 07:18:43AM +0200, Christoph Hellwig wrote:
> > > On Sat, Sep 18, 2021 at 12:22:47AM +0300, Mike Rapoport wrote:
> > > > I did some digging and it seems that the most "generic" way to check if a
> > > > page is in RAM is page_is_ram(). It's not 100% bullet proof as it'll give
> > > > false negatives for architectures that do not register "System RAM", but
> > > > those are not using dma_map_resource() anyway and, apparently, never would.
> > > 
> > > The downside of page_is_ram is that it looks really expensiv for
> > > something done at dma mapping time.
> > 
> > Indeed :(
> > But pfn_valid is plain wrong...
> > I'll keep digging.
> 
> I did some more archaeology and it that check for pfn_valid() was requested
> by arm folks because their MMU may have troubles with alias mappings with
> different attributes and so they made the check to use a false assumption
> that pfn_valid() == "RAM".
> 
> As this WARN_ON(pfn_valid()) is only present in dma_map_resource() it's
> probably safe to drop it entirely. 

I agree, we should drop it. IIUC dma_map_resource() does not create any
kernel mapping to cause problems with attribute aliasing. You'd need a
prior devm_ioremap_resource() if you want access to that range from the
CPU side. For arm64 at least, the latter ends up with a
pfn_is_map_memory() check.

-- 
Catalin