Re: [PATCH v2 4/4] mm/hugetlb: fix excessive IPI broadcasts when unsharing PMD tables using mmu_gather

["David Hildenbrand (Red Hat)" <david@kernel.org>]

On 12/19/25 13:37, Harry Yoo wrote:
> On Fri, Dec 12, 2025 at 08:10:19AM +0100, David Hildenbrand (Red Hat) wrote:
>> As reported, ever since commit 1013af4f585f ("mm/hugetlb: fix
>> huge_pmd_unshare() vs GUP-fast race") we can end up in some situations
>> where we perform so many IPI broadcasts when unsharing hugetlb PMD page
>> tables that it severely regresses some workloads.
>>
>> In particular, when we fork()+exit(), or when we munmap() a large
>> area backed by many shared PMD tables, we perform one IPI broadcast per
>> unshared PMD table.
>>
> 
> [...snip...]
> 
>> Fixes: 1013af4f585f ("mm/hugetlb: fix huge_pmd_unshare() vs GUP-fast race")
>> Reported-by: Uschakow, Stanislav" <suschako@amazon.de>
>> Closes: https://lore.kernel.org/all/4d3878531c76479d9f8ca9789dc6485d@amazon.de/
>> Tested-by: Laurence Oberman <loberman@redhat.com>
>> Cc: <stable@vger.kernel.org>
>> Signed-off-by: David Hildenbrand (Red Hat) <david@kernel.org>
>> ---
>>   include/asm-generic/tlb.h |  74 ++++++++++++++++++++++-
>>   include/linux/hugetlb.h   |  19 +++---
>>   mm/hugetlb.c              | 121 ++++++++++++++++++++++----------------
>>   mm/mmu_gather.c           |   7 +++
>>   mm/mprotect.c             |   2 +-
>>   mm/rmap.c                 |  25 +++++---
>>   6 files changed, 179 insertions(+), 69 deletions(-)
>>
>> @@ -6522,22 +6511,16 @@ long hugetlb_change_protection(struct vm_area_struct *vma,
>>   				pte = huge_pte_clear_uffd_wp(pte);
>>   			huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
>>   			pages++;
>> +			tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
>>   		}
>>   
>>   next:
>>   		spin_unlock(ptl);
>>   		cond_resched();
>>   	}
>> -	/*
>> -	 * There is nothing protecting a previously-shared page table that we
>> -	 * unshared through huge_pmd_unshare() from getting freed after we
>> -	 * release i_mmap_rwsem, so flush the TLB now. If huge_pmd_unshare()
>> -	 * succeeded, flush the range corresponding to the pud.
>> -	 */
>> -	if (shared_pmd)
>> -		flush_hugetlb_tlb_range(vma, range.start, range.end);
>> -	else
>> -		flush_hugetlb_tlb_range(vma, start, end);
>> +
>> +	tlb_flush_mmu_tlbonly(tlb);
>> +	huge_pmd_unshare_flush(tlb, vma);
> 
> Shouldn't we teach mmu_gather that it has to call

I hope not :) In the worst case we could keep the 
flush_hugetlb_tlb_range() in the !shared case in. Suboptimal but I am 
sick and tired of dealing with this hugetlb mess.


Let me CC Ryan and Catalin for the arm64 pieces and Christophe on the 
ppc pieces: See [1] where we convert away from some 
flush_hugetlb_tlb_range() users to operate on mmu_gather using
* tlb_remove_huge_tlb_entry() for mremap() and mprotect(). Before we
   would only use it in __unmap_hugepage_range().
* tlb_flush_pmd_range() for unsharing of shared PMD tables. We already
   used that in one call path.

[1] https://lore.kernel.org/all/20251212071019.471146-5-david@kernel.org/


> flush_hugetlb_tlb_range() instead of ordinary TLB flush routine,
> otherwise it will break ARCHes that has "special requirements"
> for evicting hugetlb backing TLB entries?

Yeah, I was briefly wondering about that myself (and the inconsistency 
we had in the code). I would hope that we're good, but maybe there are 
some nasty corner cases we're missing. So thanks for raising that.


Given tlb_remove_huge_tlb_entry() exist (and is already getting used) I 
would assume that it does the right thing.

In tlb_unshare_pmd_ptdesc(), I am now using tlb_flush_pmd_range(), 
because we know that we are dealing with PMD-sized hugetlb folios.

And in fact, we were already doing that in case of 
__unmap_hugepage_range(), where we did exactly what I do now:

	tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);

So, again, something would already be broken there unless I am missing 
something important.


Looking at it, I wonder whether we must do the 
tlb_remove_huge_tlb_entry() in move_hugetlb_page_tables() after the
move_huge_pte(). Looks like tlb_remove_huge_tlb_entry() might do some 
flushing on ppc (and not just updating the mmu_gather) through 
__tlb_remove_tlb_entry(). But it's a bit confusing.

-- 
Cheers

David