Re: [PATCH v1] x86/mm/pat: fix VM_PAT handling in COW mappings

[David Hildenbrand <david@redhat.com>]

On 12.03.24 20:38, David Hildenbrand wrote:
> On 12.03.24 20:22, Matthew Wilcox wrote:
>> On Tue, Mar 12, 2024 at 07:11:18PM +0100, David Hildenbrand wrote:
>>> PAT handling won't do the right thing in COW mappings: the first PTE
>>> (or, in fact, all PTEs) can be replaced during write faults to point at
>>> anon folios. Reliably recovering the correct PFN and cachemode using
>>> follow_phys() from PTEs will not work in COW mappings.
>>
>> I guess the first question is: Why do we want to support COW mappings
>> of VM_PAT areas?  What breaks if we just disallow it?
> 
> Well, that was my first approach. Then I decided to be less radical (IOW
> make my life easier by breaking less user space) and "fix it" with
> minimal effort.
> 
> Chances of breaking some weird user space is possible, although I
> believe for most such mappings MAP_PRIVATE doesn't make too much sense
> sense.
> 
> Nasty COW support for VM_PFNMAP mappings dates back forever. So does PAT
> support.
> 
> I can try finding digging through some possible user space users tomorrow.

As discussed, MAP_PRIVATE doesn't make too much sense for most PFNMAP 
mappings.

However, /dev/mem and /proc/vmcore are still used with MAP_PRIVATE in 
some cases.

Side note: /proc/vmcore is a bit weird: mmap_vmcore() sets VM_MIXEDMAP, 
and then we might call remap_pfn_range(), which sets VM_PFNMAP. I'm not 
so sure if that's what we want to happen ...

As far as I can see, makedumpfile always mmap's memory to be dumped 
(/dev/mem, /proc/vmcore) using PROT_READ+MAP_PRIVATE, resulting in a COW 
mapping.


In my opinion, we should use this fairly simple fix to keep it working 
for now and look into disabling any MAP_PRIVATE of VM_PFNMAP separately, 
for all architectures.

But I'll leave the decision to x86 maintainers.

-- 
Cheers,

David / dhildenb