[PATCH V2] arm64/mm: Intercept pfn changes in set_pte_at()

[Will Deacon]

On Fri, Feb 03, 2023 at 10:40:18AM +0800, Muchun Song wrote:
> 
> 
> > On Feb 2, 2023, at 18:45, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > 
> > On Thu, Feb 02, 2023 at 05:51:39PM +0800, Muchun Song wrote:
> >>> On Feb 1, 2023, at 20:20, Catalin Marinas <catalin.marinas at arm.com> wrote:
> >>>> Bah, sorry! Catalin reckons it may have been him talking about the vmemmap.
> >>> 
> >>> Indeed. The discussion with Anshuman started from this thread:
> >>> 
> >>> https://lore.kernel.org/all/20221025014215.3466904-1-mawupeng1@huawei.com/
> >>> 
> >>> We already trip over the existing checks even without Anshuman's patch,
> >>> though only by chance. We are not setting the software PTE_DIRTY on the
> >>> new pte (we don't bother with this bit for kernel mappings).
> >>> 
> >>> Given that the vmemmap ptes are still live when such change happens and
> >>> no-one came with a solution to the break-before-make problem, I propose
> >>> we revert the arm64 part of commit 47010c040dec ("mm: hugetlb_vmemmap:
> >>> cleanup CONFIG_HUGETLB_PAGE_FREE_VMEMMAP*"). We just need this hunk:
> >>> 
> >>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> >>> index 27b2592698b0..5263454a5794 100644
> >>> --- a/arch/arm64/Kconfig
> >>> +++ b/arch/arm64/Kconfig
> >>> @@ -100,7 +100,6 @@ config ARM64
> >>> select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> >>> select ARCH_WANT_FRAME_POINTERS
> >>> select ARCH_WANT_HUGE_PMD_SHARE if ARM64_4K_PAGES || (ARM64_16K_PAGES && !ARM64_VA_BITS_36)
> >>> - select ARCH_WANT_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
> >> 
> >> Maybe it is a little overkill for HVO as it can significantly minimize the
> >> overhead of vmemmap on ARM64 servers for some workloads (like qemu, DPDK).
> >> So I don't think disabling it is a good approach. Indeed, HVO broke BBM,
> >> but the waring does not affect anything since the tail vmemmap pages are
> >> supposed to be read-only. So, I suggest skipping warnings if it is the
> >> vmemmap address in set_pte_at(). What do you think of?
> > 
> > IIUC, vmemmap_remap_pte() not only makes the pte read-only but also
> > changes the output address. Architecturally, this needs a BBM sequence.
> > We can avoid going through an invalid pte if we first make the pte
> > read-only, TLBI but keeping the same pfn, followed by a change of the
> > pfn while keeping the pte readonly. This also assumes that the content
> > of the page pointed at by the pte is the same at both old and new pfn.
> 
> Right. I think using BBM is to avoid possibly creating multiple TLB entries
> for the same address for a extremely short period. But accessing either the
> old page or the new page is fine in this case. Is it acceptable for this
> special case without using BBM?

Sadly, the architecture allows the CPU to conjure up a mapping based on a
combination of the old and the new descriptor (a process known as
"amalgamation") so we _really_ need the BBM sequence.

I'm in favour of disabling the optimisation now and bringing it back once
we've got this fixed.

Will