Hi David C,
This has been through a lot of revisions so I think for the sake of expediency
the best solution here is for me to grab this and fiddle with it a bit then post
it out myself.
TO BE VERY CLEAR - I will keep your authorship and attribution :), and simply
stick my name on a Co-developed-by tag.
So this will remain your patch.
Cheers, Lorenzo
On Mon, Jul 06, 2026 at 09:31:27PM +0100, David Carlier wrote:
> ptdump_walk_pgd() walks the kernel page tables under get_online_mems().
> That does not stop vmalloc from freeing a kernel PTE page underneath the
> walk.
>
> When vmap_try_huge_pmd() promotes a range to a huge PMD it collapses the
> existing PTE table and frees it via pmd_free_pte_page(). On x86, riscv and
> powerpc this runs without the init_mm mmap lock; only arm64 takes it, and
> not on the block-split path. So ptdump can dereference a just-freed PTE
> page, which is the use after free syzbot hit in ptdump_pte_entry().
>
> The race is not new. ptdump walks the whole kernel address space, including
> ranges other code is actively mapping, so it reads page tables it does not
> own. Commit 5ba2f0a15564 ("mm: introduce deferred freeing for kernel page
> tables") only widened the window; the Fixes tag points there for that
> reason.
>
> Every other walker works on a range it owns and is the only one mutating
> it: set_memory() on arm64/riscv/loongarch, the arm64 block-split path, the
> openrisc DMA path and the hugetlb_vmemmap remap. Nothing frees those ranges
> concurrently, so they cannot race and do not need RCU. ptdump is the only
> walker that traverses ranges it does not own.
>
> Defer the free by an RCU grace period. pagetable_free_kernel() now frees
> via call_rcu() in both the async and non-async configs. The async path
> still flushes the TLB first, then queues the per-page RCU free. The page
> stays valid until any walk that may have observed it drops its RCU read
> lock.
>
> x86, arm64 and riscv reach pagetable_free_kernel() for the collapsed PTE
> page (the ptdesc carries PT_kernel), so the deferral covers them. powerpc
> uses its own fragment allocator: pte_free_kernel() there frees the page
> synchronously via pte_fragment_free() and the ptdesc is not flagged kernel,
> so defer the kernel case in pte_fragment_free() as well. arm64 also takes
> init_mm.mmap_lock around the free under a static key; that is now redundant
> with the RCU deferral but left in place.
>
> On the read side walk_page_range_debug() walks the init_mm range in bounded
> chunks, taking rcu_read_lock() around each chunk and calling cond_resched()
> between them. It uses the lockless walker as the mmap lock is no longer
> held, and pgd_addr_end() to bound each chunk without overflowing at the top
> of the address space. A walker either sees the cleared PMD and skips, or
> keeps the page alive until it drops the lock. The owned-range walkers are
> unchanged.
>
> Stop taking mmap_write_lock() in ptdump_walk_pgd() for init_mm. It never
> guarded against this free -- most architectures free the collapsed PTE
> table without it -- and RCU now provides the synchronisation. efi_mm and
> current->mm page tables are not RCU-freed, so they keep the mmap write
> lock.
>
> ptdump callbacks run under RCU within a chunk, so they must not sleep. The
> arch note_page() and effective_prot() callbacks only format into the
> preallocated seq_file buffer; the only GFP_KERNEL marker setup runs before
> the walk, and cond_resched() happens between chunks, outside the read lock.
>
> Fixes: 5ba2f0a15564 ("mm: introduce deferred freeing for kernel page tables")
> Reported-by: [email protected]
> Closes:
> https://lore.kernel.org/all/[email protected]/T/
> Assisted-by: Claude:claude-opus-4-8
> Signed-off-by: David Carlier <[email protected]>
> ---
> v8: fix four issues raised by the Gemini review of v7 (relayed by Andrew).
> - the init_mm walk called walk_kernel_page_table_range(), which
> asserts init_mm.mmap_lock -- but that lock is now dropped; use the
> lockless walker instead.
> - only stop taking mmap_write_lock() for init_mm; efi_mm and
> current->mm are not RCU-freed and keep it.
> - bound each chunk with pgd_addr_end() so the last chunk does not
> overflow at the top of the address space.
> - powerpc frees kernel PTE pages synchronously via pte_fragment_free()
> and never sets PT_kernel, so it bypassed the RCU deferral; defer the
> kernel case there too.
> v7: no code change; add version tag and per-revision changelog (Dev).
> v6: chunk the init_mm walk in walk_page_range_debug() and take
> rcu_read_lock() per chunk (reverting v5's caller-side lock + assert)
> so the read section stays bounded on large kernel address spaces and
> can cond_resched() between chunks; drop the now-redundant
> mmap_write_lock() in ptdump_walk_pgd().
> v5: reframe changelog around the pre-existing race and range ownership;
> correct the mmap-lock description (arm64 is the exception, not x86);
> move rcu_read_lock() into ptdump_walk_pgd() and assert it in
> walk_page_range_debug(); drop walk_kernel_page_table_range_rcu(); fix
> the pgtable-generic.c comment; document the no-sleep audit of the
> callbacks.
> v4: defer the free in both the async and non-async configs, not just the
> async one; add a walk_kernel_page_table_range_rcu() helper.
> v3: take rcu_read_lock() in the init_mm branch of walk_page_range_debug().
> v2: use call_rcu() instead of synchronize_rcu().
> arch/powerpc/mm/pgtable-frag.c | 7 ++++++-
> include/linux/mm.h | 7 -------
> mm/pagewalk.c | 36 +++++++++++++++++++++++++++++-----
> mm/pgtable-generic.c | 22 ++++++++++++++++++++-
> mm/ptdump.c | 11 +++++++++--
> 5 files changed, 67 insertions(+), 16 deletions(-)
>
> diff --git a/arch/powerpc/mm/pgtable-frag.c b/arch/powerpc/mm/pgtable-frag.c
> index ae742564a3d5..1e1e88f831f7 100644
> --- a/arch/powerpc/mm/pgtable-frag.c
> +++ b/arch/powerpc/mm/pgtable-frag.c
> @@ -123,7 +123,12 @@ void pte_fragment_free(unsigned long *table, int kernel)
>
> BUG_ON(atomic_read(&ptdesc->pt_frag_refcount) <= 0);
> if (atomic_dec_and_test(&ptdesc->pt_frag_refcount)) {
> - if (kernel || !folio_test_clear_active(ptdesc_folio(ptdesc)))
> + /*
> + * Kernel page tables may be walked locklessly under RCU by
> + * ptdump, so defer their free by a grace period too, like the
> + * lockless-GUP case below for user tables.
> + */
> + if (!kernel && !folio_test_clear_active(ptdesc_folio(ptdesc)))
> pte_free_now(&ptdesc->pt_rcu_head);
> else
> call_rcu(&ptdesc->pt_rcu_head, pte_free_now);
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 485df9c2dbdd..79408a17a1b0 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -3695,14 +3695,7 @@ static inline void __pagetable_free(struct ptdesc *pt)
> __free_pages(page, compound_order(page));
> }
>
> -#ifdef CONFIG_ASYNC_KERNEL_PGTABLE_FREE
> void pagetable_free_kernel(struct ptdesc *pt);
> -#else
> -static inline void pagetable_free_kernel(struct ptdesc *pt)
> -{
> - __pagetable_free(pt);
> -}
> -#endif
> /**
> * pagetable_free - Free pagetables
> * @pt: The page table descriptor
> diff --git a/mm/pagewalk.c b/mm/pagewalk.c
> index 3ae2586ff45b..fc2fe014ac8c 100644
> --- a/mm/pagewalk.c
> +++ b/mm/pagewalk.c
> @@ -620,7 +620,7 @@ int walk_page_range(struct mm_struct *mm, unsigned long
> start,
> * Note: Be careful to walk the kernel pages tables, the caller may be need
> to
> * take other effective approaches (mmap lock may be insufficient) to prevent
> * the intermediate kernel page tables belonging to the specified address
> range
> - * from being freed (e.g. memory hot-remove).
> + * from being freed (e.g. memory hot-remove, vmap huge page promotion).
> */
> int walk_kernel_page_table_range(unsigned long start, unsigned long end,
> const struct mm_walk_ops *ops, pgd_t *pgd, void *private)
> @@ -643,7 +643,7 @@ int walk_kernel_page_table_range(unsigned long start,
> unsigned long end,
> * Use this function to walk the kernel page tables locklessly. It should be
> * guaranteed that the caller has exclusive access over the range they are
> * operating on - that there should be no concurrent access, for example,
> - * changing permissions for vmalloc objects.
> + * changing permissions for vmalloc objects, or vmap huge page promotion.
> */
> int walk_kernel_page_table_range_lockless(unsigned long start, unsigned long
> end,
> const struct mm_walk_ops *ops, pgd_t *pgd, void *private)
> @@ -692,9 +692,35 @@ int walk_page_range_debug(struct mm_struct *mm, unsigned
> long start,
> };
>
> /* For convenience, we allow traversal of kernel mappings. */
> - if (mm == &init_mm)
> - return walk_kernel_page_table_range(start, end, ops,
> - pgd, private);
> + if (mm == &init_mm) {
> + unsigned long addr = start;
> +
> + /*
> + * Walk in bounded chunks so the RCU read lock is never held
> + * across the whole kernel address space. A kernel page table
> + * freed via pagetable_free_kernel() stays valid until the walk
> + * that may have observed it drops the lock; releasing the lock
> + * between chunks is safe as no page table pointer is held
> + * across the gap. The mmap lock is not held, so use the
> + * lockless walker; RCU, not the lock, keeps the table alive.
> + */
> + while (addr < end) {
> + unsigned long next = pgd_addr_end(addr, end);
> + int err;
> +
> + rcu_read_lock();
> + err = walk_kernel_page_table_range_lockless(addr, next,
> ops,
> + pgd,
> private);
> + rcu_read_unlock();
> + if (err)
> + return err;
> +
> + addr = next;
> + cond_resched();
> + }
> + return 0;
> + }
> +
> if (start >= end || !walk.mm)
> return -EINVAL;
> if (!check_ops_safe(ops))
> diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c
> index b91b1a98029c..7a32e4821957 100644
> --- a/mm/pgtable-generic.c
> +++ b/mm/pgtable-generic.c
> @@ -410,6 +410,13 @@ pte_t *pte_offset_map_lock(struct mm_struct *mm, pmd_t
> *pmd,
> goto again;
> }
>
> +static void kernel_pgtable_free_rcu(struct rcu_head *head)
> +{
> + struct ptdesc *pt = container_of(head, struct ptdesc, pt_rcu_head);
> +
> + __pagetable_free(pt);
> +}
> +
> #ifdef CONFIG_ASYNC_KERNEL_PGTABLE_FREE
> static void kernel_pgtable_work_func(struct work_struct *work);
>
> @@ -434,8 +441,15 @@ static void kernel_pgtable_work_func(struct work_struct
> *work)
> spin_unlock(&kernel_pgtable_work.lock);
>
> iommu_sva_invalidate_kva_range(PAGE_OFFSET, TLB_FLUSH_ALL);
> +
> + /*
> + * Debug walkers (ptdump) may walk ranges they do not own and race this
> + * free, so they walk under rcu_read_lock(). Free after a grace period:
> + * a walker either already saw the cleared PMD, or keeps the page alive
> + * until it drops the RCU lock.
> + */
> list_for_each_entry_safe(pt, next, &page_list, pt_list)
> - __pagetable_free(pt);
> + call_rcu(&pt->pt_rcu_head, kernel_pgtable_free_rcu);
> }
>
> void pagetable_free_kernel(struct ptdesc *pt)
> @@ -446,4 +460,10 @@ void pagetable_free_kernel(struct ptdesc *pt)
>
> schedule_work(&kernel_pgtable_work.work);
> }
> +#else
> +void pagetable_free_kernel(struct ptdesc *pt)
> +{
> + /* Defer the free by a grace period; see kernel_pgtable_work_func(). */
> + call_rcu(&pt->pt_rcu_head, kernel_pgtable_free_rcu);
> +}
> #endif
> diff --git a/mm/ptdump.c b/mm/ptdump.c
> index 973020000096..537be9995e1a 100644
> --- a/mm/ptdump.c
> +++ b/mm/ptdump.c
> @@ -177,13 +177,20 @@ void ptdump_walk_pgd(struct ptdump_state *st, struct
> mm_struct *mm, pgd_t *pgd)
> const struct ptdump_range *range = st->range;
>
> get_online_mems();
> - mmap_write_lock(mm);
> + /*
> + * init_mm is walked locklessly under RCU by walk_page_range_debug().
> + * efi_mm / current->mm page tables are not RCU-freed, so hold the
> + * mmap write lock to keep them stable against concurrent teardown.
> + */
> + if (mm != &init_mm)
> + mmap_write_lock(mm);
> while (range->start != range->end) {
> walk_page_range_debug(mm, range->start, range->end,
> &ptdump_ops, pgd, st);
> range++;
> }
> - mmap_write_unlock(mm);
> + if (mm != &init_mm)
> + mmap_write_unlock(mm);
> put_online_mems();
>
> /* Flush out the last page */
> --
> 2.53.0
>
