On Tue, Aug 01, 2023 at 02:48:38PM +0200, David Hildenbrand wrote:
> We shouldn't be using a GUP-internal helper if it can be avoided.
>
> Similar to smaps_pte_entry() that uses vm_normal_page(), let's use
> vm_normal_page_pmd() that similarly refuses to return the huge zeropage.
>
> In contrast to follow_trans_huge_pmd(), vm_normal_page_pmd():
>
> (1) Will always return the head page, not a tail page of a THP.
>
> If we'd ever call smaps_account with a tail page while setting "compound
> = true", we could be in trouble, because smaps_account() would look at
> the memmap of unrelated pages.
>
> If we're unlucky, that memmap does not exist at all. Before we removed
> PG_doublemap, we could have triggered something similar as in
> commit 24d7275ce279 ("fs/proc: task_mmu.c: don't read mapcount for
> migration entry").
>
> This can theoretically happen ever since commit ff9f47f6f00c ("mm: proc:
> smaps_rollup: do not stall write attempts on mmap_lock"):
>
> (a) We're in show_smaps_rollup() and processed a VMA
> (b) We release the mmap lock in show_smaps_rollup() because it is
> contended
> (c) We merged that VMA with another VMA
> (d) We collapsed a THP in that merged VMA at that position
>
> If the end address of the original VMA falls into the middle of a THP
> area, we would call smap_gather_stats() with a start address that falls
> into a PMD-mapped THP. It's probably very rare to trigger when not
> really forced.
>
> (2) Will succeed on a is_pci_p2pdma_page(), like vm_normal_page()
>
> Treat such PMDs here just like smaps_pte_entry() would treat such PTEs.
> If such pages would be anonymous, we most certainly would want to
> account them.
>
> (3) Will skip over pmd_devmap(), like vm_normal_page() for pte_devmap()
>
> As noted in vm_normal_page(), that is only for handling legacy ZONE_DEVICE
> pages. So just like smaps_pte_entry(), we'll now also ignore such PMD
> entries.
>
> Especially, follow_pmd_mask() never ends up calling
> follow_trans_huge_pmd() on pmd_devmap(). Instead it calls
> follow_devmap_pmd() -- which will fail if neither FOLL_GET nor FOLL_PIN
> is set.
>
> So skipping pmd_devmap() pages seems to be the right thing to do.
>
> (4) Will properly handle VM_MIXEDMAP/VM_PFNMAP, like vm_normal_page()
>
> We won't be returning a memmap that should be ignored by core-mm, or
> worse, a memmap that does not even exist. Note that while
> walk_page_range() will skip VM_PFNMAP mappings, walk_page_vma() won't.
>
> Most probably this case doesn't currently really happen on the PMD level,
> otherwise we'd already be able to trigger kernel crashes when reading
> smaps / smaps_rollup.
>
> So most probably only (1) is relevant in practice as of now, but could only
> cause trouble in extreme corner cases.
>
> Fixes: ff9f47f6f00c ("mm: proc: smaps_rollup: do not stall write attempts on mmap_lock")
> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
Maybe move the follow_trans_huge_pmd() declaration from linux/huge_mm.h
to mm/internal.h to discourage future mistakes? Otherwise
Acked-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
--
Mel Gorman
SUSE Labs
