[PATCH] arm64: tlb: Flush walk cache when unsharing PMD tables

[Catalin Marinas]

On Fri, May 22, 2026 at 01:32:07PM +0800, Zeng Heng wrote:
> On 2026/5/21 23:15, Catalin Marinas wrote:
> > On Thu, May 21, 2026 at 04:05:07PM +0100, Catalin Marinas wrote:
> > > On Thu, May 21, 2026 at 03:30:11PM +0800, Zeng Heng wrote:
> > > > From: Zeng Heng <zengheng4 at huawei.com>
> > > > 
> > > > When huge_pmd_unshare() is called to unshare a PMD table, the
> > > > tlb_unshare_pmd_ptdesc() function sets tlb->unshared_tables=true
> > > > but the aarch64 tlb_flush() only checked tlb->freed_tables to
> > > > determine whether to use TLBF_NONE (vae1is, invalidates walk
> > > > cache) or TLBF_NOWALKCACHE (vale1is, leaf-only).
> > > > 
> > > > This caused the stale PMD page table entry to remain in the walk cache
> > > > after unshare, potentially leading to incorrect page table walks.
> > > > 
> > > > Fix by including unshared_tables in the check, so that when
> > > > unsharing tables, TLBF_NONE is used and the walk cache is properly
> > > > invalidated.
> > > > 
> > > > Here is the detailed distinction between vae1is and vale1is:
> > > > 
> > > > | Instruction Combination  | Actual Invalidation Scope                         |
> > > > | ------------------------ | --------------------------------------------------|
> > > > | `VAE1IS`  + TTL=`0`      | All entries at all levels (full invalidation)     |
> > > > | `VAE1IS`  + TTL=`2` (L2) | Non-leaf at Level 0/1 + leaf at Level 2           |
> > > > | `VALE1IS` + TTL=`0`      | Leaf entries at all levels (non-leaf not cleared) |
> > > > | `VALE1IS` + TTL=`2` (L2) | Leaf entry at Level 2 only                        |
> > > > 
> > > > Signed-off-by: Zeng Heng <zengheng4 at huawei.com>
> > > The fix looks fine but does it need:
> > > 
> > > Fixes: 8ce720d5bd91 ("mm/hugetlb: fix excessive IPI broadcasts when unsharing PMD tables using mmu_gather")
> > > Cc: <stable at vger.kernel.org>
> > > 
> > > > ---
> > > >   arch/arm64/include/asm/tlb.h | 3 ++-
> > > >   1 file changed, 2 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/arch/arm64/include/asm/tlb.h b/arch/arm64/include/asm/tlb.h
> > > > index 10869d7731b8..751bd57bc3ba 100644
> > > > --- a/arch/arm64/include/asm/tlb.h
> > > > +++ b/arch/arm64/include/asm/tlb.h
> > > > @@ -53,7 +53,8 @@ static inline int tlb_get_level(struct mmu_gather *tlb)
> > > >   static inline void tlb_flush(struct mmu_gather *tlb)
> > > >   {
> > > >   	struct vm_area_struct vma = TLB_FLUSH_VMA(tlb->mm, 0);
> > > > -	tlbf_t flags = tlb->freed_tables ? TLBF_NONE : TLBF_NOWALKCACHE;
> > > > +	tlbf_t flags = (tlb->freed_tables || tlb->unshared_tables) ?
> > > > +			TLBF_NONE : TLBF_NOWALKCACHE;
> > > >   	unsigned long stride = tlb_get_unmap_size(tlb);
> > > >   	int tlb_level = tlb_get_level(tlb);
> > Do we need this as well?
> 
> The proposed fix has been validated against the issue scenarios and
> works as expected.
> 
> Per the ARM Architecture Reference Manual, whether only the last-level
> page table entry is invalidated is determined by the instruction used
> (vale1is for leaf entry only, vae1is for walk cache including leaf entry and
> non-leaf entry), rather than the TTL field. The TTL field merely specifies
> which level the leaf entry belongs to.

Ah, yes, you are right. The TTL is still 2 in this case for a huge pmd,
we just want the walk cache leading to it to be invalidated. So no need
for the additional tlb_get_level().

Thanks.

-- 
Catalin