Re: build problems on architectures where FIXADDR_* stuff is not

[Andrew Morton]

Oleg Drokin <[e-mail]> wrote:
>
>    Since there are architectures where FIXADDR_* stuff is not constant (e.g. UML),
> ...
> +			fixmap_vma.vm_start = FIXADDR_START;
> +			fixmap_vma.vm_end = FIXADDR_TOP;
> +			fixmap_vma.vm_page_prot = PAGE_READONLY;
>  			pgd = pgd_offset_k(pg);
>  			if (!pgd)
>  				return i ? : -EFAULT;

That's modifying static storage which other, unrelated processes or CPUs
may be playing with.

The new code in get_user_pages() is rather rude - it's returning a
statically allocated VMA which isn't in the VMA tree - the caller (who
holds mmap_sem()) could reasonably expect that the VMA can be located via
find_vma(), or removed from the tree or whatever.  But it cannot.

I think it needs to be redone.  Either by stuffing a VMA into every
process's mm which describes the fixmap area, or by failing
get_user_pages() if the caller has passed in a non-NULL `vmas' and is
requesting access to the fixmap area.

Probably the latter.  That'll require that access_process_vm() be changed
to not require a vma.  It's only using the vma for cache flushing, but the
flishing in there is borked anyway.