Re: [PATCH RESEND v3 1/2] mm/tlb: skip redundant IPI when TLB flush already synchronized

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

On 1/7/26 07:37, Lance Yang wrote:
> Hi David,
> 
> On 2026/1/7 00:10, Lance Yang wrote:
> [..]
>>> What could work is tracking "tlb_table_flush_sent_ipi" really when we
>>> are flushing the TLB for removed/unshared tables, and maybe resetting
>>> it ... I don't know when from the top of my head.
>>
> 
> Seems like we could fix the issue that the flag lifetime was broken
> if the MMU gather gets reused by splitting the flush and reset. This
> ensures the flag stays valid between flush and sync.
> 
> Now tlb_flush_unshared_tables() does:
>     1) __tlb_flush_mmu_tlbonly() - flush only, keeps flags alive
>     2) tlb_gather_remove_table_sync_one() - can check the flag
>     3) __tlb_reset_range() - reset everything after sync
> 
> Something like this:
> 
> ---8<---
> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
> index 3975f7d11553..a95b054dfcca 100644
> --- a/include/asm-generic/tlb.h
> +++ b/include/asm-generic/tlb.h
> @@ -415,6 +415,7 @@ static inline void __tlb_reset_range(struct
> mmu_gather *tlb)
>    	tlb->cleared_puds = 0;
>    	tlb->cleared_p4ds = 0;
>    	tlb->unshared_tables = 0;
> +	tlb->tlb_flush_sent_ipi = 0;


As raised, the "tlb_flush_sent_ipi" is confusing when we sent to 
different CPUs based on whether we are removing page tables or not.

I think you would really want to track that explicitly 
"tlb_table_flush_sent_ipi" ?

>    	/*
>    	 * Do not reset mmu_gather::vma_* fields here, we do not
>    	 * call into tlb_start_vma() again to set them if there is an
> @@ -492,7 +493,7 @@ tlb_update_vma_flags(struct mmu_gather *tlb, struct
> vm_area_struct *vma)
>    	tlb->vma_pfn |= !!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP));
>    }
> 
> -static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
> +static inline void __tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
>    {
>    	/*
>    	 * Anything calling __tlb_adjust_range() also sets at least one of
> @@ -503,6 +504,11 @@ static inline void tlb_flush_mmu_tlbonly(struct
> mmu_gather *tlb)
>    		return;
> 
>    	tlb_flush(tlb);
> +}
> +
> +static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
> +{
> +	__tlb_flush_mmu_tlbonly(tlb);
>    	__tlb_reset_range(tlb);
>    }
> 
> @@ -824,7 +830,7 @@ static inline void tlb_flush_unshared_tables(struct
> mmu_gather *tlb)
>    	 * flush the TLB for the unsharer now.
>    	 */
>    	if (tlb->unshared_tables)
> -		tlb_flush_mmu_tlbonly(tlb);
> +		__tlb_flush_mmu_tlbonly(tlb);
> 
>    	/*
>    	 * Similarly, we must make sure that concurrent GUP-fast will not
> @@ -834,14 +840,16 @@ static inline void
> tlb_flush_unshared_tables(struct mmu_gather *tlb)
>    	 * We only perform this when we are the last sharer of a page table,
>    	 * as the IPI will reach all CPUs: any GUP-fast.
>    	 *
> -	 * Note that on configs where tlb_remove_table_sync_one() is a NOP,
> -	 * the expectation is that the tlb_flush_mmu_tlbonly() would have issued
> -	 * required IPIs already for us.
> +	 * Use tlb_gather_remove_table_sync_one() instead of
> +	 * tlb_remove_table_sync_one() to skip the redundant IPI if the
> +	 * TLB flush above already sent one.
>    	 */
>    	if (tlb->fully_unshared_tables) {
> -		tlb_remove_table_sync_one();
> +		tlb_gather_remove_table_sync_one(tlb);
>    		tlb->fully_unshared_tables = false;
>    	}
> +
> +	__tlb_reset_range(tlb);
>    }
>    #endif /* CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING */
> ---
> 
> For khugepaged, it should be fine - it uses a local mmu_gather that
> doesn't get reused. The lifetime is simply:
> 
>     tlb_gather_mmu() → flush → sync → tlb_finish_mmu()
> 
> Let me know if this addresses your concern :)

I'll probably have to see the full picture. But this lifetime stuff in 
core-mm ends up getting more complicated than v2 without a clear benefit 
to me (except maybe handling some x86 oddities better ;) )

-- 
Cheers

David