[PATCH v4 1/2] Revert "powerpc: Set max_mapnr correctly"

[Michael Ellerman]

Christophe Leroy <christophe.leroy at csgroup.eu> writes:
> Le 28/03/2022 à 12:37, Michael Ellerman a écrit :
>> Kefeng Wang <wangkefeng.wang at huawei.com> writes:
>>> Hi maintainers，
>>>
>>> I saw the patches has been reviewed[1], could they be merged?
>>
>> Maybe I'm just misreading the change log, but it seems wrong that we
>> need to add extra checks. pfn_valid() shouldn't return true for vmalloc
>> addresses in the first place, shouldn't we fix that instead? Who knows
>> what else that might be broken because of that.
>
> pfn_valid() doesn't take an address but a PFN

Yeah sorry that was unclear wording on my part.

What I mean is that pfn_valid(virt_to_pfn(some_vmalloc_addr)) should be
false, because virt_to_pfn(vmalloc_addr) should fail.

The right way to convert a vmalloc address to a pfn is with
vmalloc_to_pfn(), which walks the page tables to find the actual pfn
backing that vmalloc addr.

> If you have 1Gbyte of memory you have 256k PFNs.
>
> In a generic config the kernel will map 768 Mbytes of mémory (From
> 0xc0000000 to 0xe0000000) and will use 0xf0000000-0xffffffff for
> everything else including vmalloc.
>
> If you take a page above that 768 Mbytes limit, and tries to linarly
> convert it's PFN to a va, you'll hip vmalloc space. Anyway that PFN is
> valid.

That's true, but it's just some random page in vmalloc space, there's no
guarantee that it's the same page as the PFN you started with.

Note it's not true on 64-bit Book3S. There if you take a valid PFN (ie.
backed by RAM) and convert it to a virtual address (with __va()), you
will never get a vmalloc address.

> So the check really needs to be done in virt_addr_valid().

I don't think it has to, but with the way our virt_to_pfn()/__pa() works
I guess for now it's the easiest solution.

> There is another thing however that would be worth fixing (in another
> patch): that's the virt_to_pfn() in PPC64:
>
> #define virt_to_pfn(kaddr)	(__pa(kaddr) >> PAGE_SHIFT)
>
> #define __pa(x)								\
> ({									\
> 	VIRTUAL_BUG_ON((unsigned long)(x) < PAGE_OFFSET);		\
> 	(unsigned long)(x) & 0x0fffffffffffffffUL;			\
> })
>
>
> So 0xc000000000000000 and 0xd000000000000000 have the same PFN. That's
> wrong.

Yes it was wrong. But we don't use 0xd000000000000000 anymore, so it
shouldn't be an issue in practice.

See 0034d395f89d ("powerpc/mm/hash64: Map all the kernel regions in the same 0xc range").

I guess it is still a problem for 64-bit Book3E, because they use 0xc
and 0x8.

I looked at fixing __pa()/__va() to use +/- a few years back, but from
memory it still didn't work and/or generated bad code. There's a comment
about it working around a GCC miscompile.

The other thing that makes me reluctant to change it is that I worry we
have places that inadvertantly use __pa() on addresses that are already
physical addresses. If we changed __pa() to use subtraction that would
break, ie. we'd end up with a negative address.

See eg. a6e2c226c3d5 ("powerpc: Fix kernel crash in show_instructions() w/DEBUG_VIRTUAL")

So to fix it we'd have to 1) verify that the compiler bug is no longer
an issue and 2) audit uses of __pa()/__va().

cheers