On Mon, Dec 09, 2024 at 05:42:26PM +0800, Zhenhua Huang wrote:
> Commit c1cc1552616d ("arm64: MMU initialisation")
> optimizes the vmemmap to populate at the PMD section level. However, if
> start or end is not aligned to a section boundary, such as when a
> subsection is hot added, populating the entire section is wasteful. For
> instance, if only one subsection hot-added, the entire section's struct
> page metadata will still be populated.In such cases, it is more effective
> to populate at page granularity.
OK, so from the vmemmap perspective, we waste up to 2MB memory that has
been allocated even if a 2MB hot-plugged subsection required only 32KB
of struct page. I don't mind this much really. I hope all those
subsections are not scattered around to amplify this waste.
> This change also addresses mismatch issues during vmemmap_free(): When
> pmd_sect() is true, the entire PMD section is cleared, even if there is
> other effective subsection. For example, pagemap1 and pagemap2 are part
> of a single PMD entry and they are hot-added sequentially. Then pagemap1
> is removed, vmemmap_free() will clear the entire PMD entry, freeing the
> struct page metadata for the whole section, even though pagemap2 is still
> active.
I think that's the bigger issue. We can't unplug a subsection only.
Looking at unmap_hotplug_pmd_range(), it frees a 2MB vmemmap section but
that may hold struct page for the equivalent of 128MB of memory. So any
struct page accesses for the other subsections will fault.
> Fixes: c1cc1552616d ("arm64: MMU initialisation")
I wouldn't add a fix for the first commit adding arm64 support, we did
not even have memory hotplug at the time (added later in 5.7 by commit
bbd6ec605c0f ("arm64/mm: Enable memory hot remove")). IIUC, this hasn't
been a problem until commit ba72b4c8cf60 ("mm/sparsemem: support
sub-section hotplug"). That commit broke some arm64 assumptions.
> Signed-off-by: Zhenhua Huang <quic_zhenhuah@xxxxxxxxxxx>
> ---
> arch/arm64/mm/mmu.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index e2739b69e11b..fd59ee44960e 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -1177,7 +1177,9 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
> {
> WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
>
> - if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES))
> + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) ||
> + !IS_ALIGNED(page_to_pfn((struct page *)start), PAGES_PER_SECTION) ||
> + !IS_ALIGNED(page_to_pfn((struct page *)end), PAGES_PER_SECTION))
> return vmemmap_populate_basepages(start, end, node, altmap);
> else
> return vmemmap_populate_hugepages(start, end, node, altmap);
An alternative would be to fix unmap_hotplug_pmd_range() etc. to avoid
nuking the whole vmemmap pmd section if it's not empty. Not sure how
easy that is, whether we have the necessary information (I haven't
looked in detail).
A potential issue - can we hotplug 128MB of RAM and only unplug 2MB? If
that's possible, the problem isn't solved by this patch.
--
Catalin
