Re: [RFC][PATCH] arm64: update iomem_resource.end

[Robin Murphy]

On 10/05/18 23:29, Nicolin Chen wrote:
> Thanks for the comments, Robin.
>
> On Thu, May 10, 2018 at 06:45:59PM +0100, Robin Murphy wrote:
> > On 09/05/18 23:58, Nicolin Chen wrote:
> > > The iomem_resource.end is -1 by default and should be updated in
> > > arch-level code.
> > >
> > > ARM64 so far hasn't updated it while core kernel code (mm/hmm.c)
> > > started to use iomem_resource.end for boundary check. So it'd be
> > > better to assign iomem_resource.end using a valid value, the end
> > > of physical address space for example because iomem_resource.end
> > > in theory should reflect that.
> > >
> > > However, VA_BITS might be smaller than PA_BITS in ARM64. So using
> > > the end of physical address space doesn't make a lot of sense in
> > > this case, or could be even harmful since virtual address cannot
> > > reach that memory region.
> >
> > Why? There's plenty of stuff in the physical address space that will
> > only ever be accessed via ioremap/memremap. There's no reason you
> > shouldn't be able to run a VA_BITS < 48 kernel on a Cavium ThunderX
>
> I'm running VA_BITS_39 and PA_BITS_48 on Tegra 210. There had
> not been any problem of it, however with hmm.....
>
> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/mm/hmm.c#n1144
>
> This hmm_devmem_add() requests a region with PFNs being outside
> of the linear region in ARM64 case which takes MAX_PHYSMEM_BITS
> (48 bits) over iomem_resource.end without this patch. Then when
> dealing with page structures in vmemmap region from a given PFN
> directly (CONFIG_SPARSEMEM_VMEMMAP=y), and the given PFN is the
> last one based on physical region (48 bits), the address of its
> page structure will go beyond vmemmap region. Does this sound a
> problem?

Yes, but as far as we're concerned here it's not a problem with arm64:

	config ARCH_HAS_HMM
		...
		depends on (X86_64 || PPC64)
		depends on ZONE_DEVICE
		...
		depends on MEMORY_HOTPLUG
		depends on MEMORY_HOTREMOVE
		...

Whatever out-of-tree changes you have to address all of those are 
clearly implemented incorrectly; *that's* your problem.

> > where *all* the I/O is in the top half of the PA space. We already
> > constrain RAM in this very function to those regions which fit into
> > the linear map, and if you're accessing anything other than RAM
> > through the linear map you're probably doing something wrong.
>
> If I understand this part correctly, since ARM64 has applied the
> memory limit already, does it mean that probably we should fix
> something in the region_intersects() or add an extra check in the
> hmm_devmem_add(), instead of limiting the iomem_resource?

It means we should implement memory hotplug correctly. Which, 
unfortunately, I happen to know is really hard (it's something I've been 
looking at from the device-DAX angle).

> > Furthermore, the physical region covered by the linear map doesn't
> > necessarily start at physical address 0 anyway - see PHYS_OFFSET.
>
> Hmm...okay...but there still should be a protection somewhere if
> it happens to access a page structure via pfn_to_page() while the
> PFN is not covered by the vmemmap linear mapping, right?

There already is: pfn_valid() will return false for anything outside the 
intersection of memblock regions and the linear map region as calculated 
by arm64_memblock_init(); anything calling pfn_to_page() without 
checking pfn_valid() first is fundamentally broken. Or if you have 
out-of-tree changes to the pfn_valid() implementation then all bets are 
off, and it's not something for mainline to work around.

Robin.