On Tue, 24 Jul 2007 10:38:55 +0200 Ingo Molnar <mingo@elte.hu> wrote:

quoted text
> 
> the patch below has been tested in -rt for a long time - lets get it 
> upstream please. It fixes some bad latencies on PREEMPT too.
> 
> 	Ingo
> 
> ----------------------->
> From: Hugh Dickins <hugh@veritas.com>
> Subject: mm: reduce pagetable-freeing latencies
> 
> 2.6.15-rc1 moved the unlinking of a vma from its prio_tree and anon_vma
> into free_pgtables: so the vma is hidden from rmap and vmtruncate before
> freeing its page tables, allowing safe descent without page table lock.
> But free_pgtables is still called with preemption disabled, and Lee
> Revell has now detected high latency there.
> 
> The right fix will be to rework the mmu_gathering, not to need preemption
> disabled; but for now an ugly CONFIG_PREEMPT block in free_pgtables, to
> make an initial unlinking pass with preemption enabled - made uglier by
> CONFIG_IA64 definitions (only ia64 actually uses the start and end given
> to tlb_finish_mmu, and our floor and ceiling don't quite work for those).
> These CONFIG choices being to minimize the additional TLB flushing.
> 
> Signed-off-by: Hugh Dickins <hugh@veritas.com>
> Signed-off-by: Ingo Molnar <mingo@elte.hu>
> --
> 
>  mm/memory.c |   32 ++++++++++++++++++++++++++++++++
>  1 file changed, 32 insertions(+)
> 
> Index: linux-rt-rebase.q/mm/memory.c
> ===================================================================
> --- linux-rt-rebase.q.orig/mm/memory.c
> +++ linux-rt-rebase.q/mm/memory.c
> @@ -264,18 +264,48 @@ void free_pgd_range(struct mmu_gather **
>  		flush_tlb_pgtables((*tlb)->mm, start, end);
>  }
>  
> +#ifdef CONFIG_IA64
> +#define tlb_start_addr(tlb)	(tlb)->start_addr
> +#define tlb_end_addr(tlb)	(tlb)->end_addr
> +#else
> +#define tlb_start_addr(tlb)	0UL	/* only ia64 really uses it */
> +#define tlb_end_addr(tlb)	0UL	/* only ia64 really uses it */
> +#endif
> +
>  void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
>  		unsigned long floor, unsigned long ceiling)
>  {
> +#ifdef CONFIG_PREEMPT
> +	struct vm_area_struct *unlink = vma;
> +	int fullmm = (*tlb)->fullmm;
> +
> +	if (!vma)	/* Sometimes when exiting after an oops */
> +		return;
> +	if (vma->vm_next)
> +		tlb_finish_mmu(*tlb, tlb_start_addr(*tlb), tlb_end_addr(*tlb));
> +	/*
> +	 * Hide vma from rmap and vmtruncate before freeeing pgtables,
> +	 * with preemption enabled, except when unmapping just one area.
> +	 */
> +	while (unlink) {
> +		anon_vma_unlink(unlink);
> +		unlink_file_vma(unlink);
> +		unlink = unlink->vm_next;
> +	}
> +	if (vma->vm_next)
> +		*tlb = tlb_gather_mmu(vma->vm_mm, fullmm);
> +#endif
>  	while (vma) {
>  		struct vm_area_struct *next = vma->vm_next;
>  		unsigned long addr = vma->vm_start;
>  
> +#ifndef CONFIG_PREEMPT
>  		/*
>  		 * Hide vma from rmap and vmtruncate before freeing pgtables
>  		 */
>  		anon_vma_unlink(vma);
>  		unlink_file_vma(vma);
> +#endif
>  
>  		if (is_vm_hugetlb_page(vma)) {
>  			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
> @@ -288,8 +318,10 @@ void free_pgtables(struct mmu_gather **t
>  			       && !is_vm_hugetlb_page(next)) {
>  				vma = next;
>  				next = vma->vm_next;
> +#ifndef CONFIG_PREEMPT
>  				anon_vma_unlink(vma);
>  				unlink_file_vma(vma);
> +#endif
>  			}
>  			free_pgd_range(tlb, addr, vma->vm_end,
>  				floor, next? next->vm_start: ceiling);

What a truly putrid patch.  I am suspecting that this was a quick
get-you-out-of-trouble thing, which then got forgotten about.

We have two months to do the "right fix".  Please?
-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/