Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1503415 > unrolled thread

[patch] mm, thp: avoid unlikely branches for split_huge_pmd

Back to article view | Back to linux.kernel

Contents

  [patch] mm, thp: avoid unlikely branches for split_huge_pmd David Rientjes <rientjes@google.com> - 2016-10-19 01:10 +0200
    Re: [patch] mm, thp: avoid unlikely branches for split_huge_pmd Andrew Morton <akpm@linux-foundation.org> - 2016-10-19 01:20 +0200
    Re: [patch] mm, thp: avoid unlikely branches for split_huge_pmd Vlastimil Babka <vbabka@suse.cz> - 2016-10-19 16:20 +0200

#1503415 — [patch] mm, thp: avoid unlikely branches for split_huge_pmd

While doing MADV_DONTNEED on a large area of thp memory, I noticed we 
encountered many unlikely() branches in profiles for each backing 
hugepage.  This is because zap_pmd_range() would call split_huge_pmd(), 
which rechecked the conditions that were already validated, but as part of 
an unlikely() branch.

Avoid the unlikely() branch when in a context where pmd is known to be 
good for __split_huge_pmd() directly.

Signed-off-by: David Rientjes <rientjes@google.com>
---
 include/linux/huge_mm.h | 2 ++
 mm/memory.c             | 4 ++--
 mm/mempolicy.c          | 2 +-
 mm/mprotect.c           | 2 +-
 4 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
--- a/include/linux/huge_mm.h
+++ b/include/linux/huge_mm.h
@@ -189,6 +189,8 @@ static inline void deferred_split_huge_page(struct page *page) {}
 #define split_huge_pmd(__vma, __pmd, __address)	\
 	do { } while (0)
 
+static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
+		unsigned long address, bool freeze, struct page *page) {}
 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 		unsigned long address, bool freeze, struct page *page) {}
 
diff --git a/mm/memory.c b/mm/memory.c
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1240,7 +1240,7 @@ static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
 			if (next - addr != HPAGE_PMD_SIZE) {
 				VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
 				    !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
-				split_huge_pmd(vma, pmd, addr);
+				__split_huge_pmd(vma, pmd, addr, false, NULL);
 			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
 				goto next;
 			/* fall through */
@@ -3454,7 +3454,7 @@ static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
 
 	/* COW handled on pte level: split pmd */
 	VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
-	split_huge_pmd(fe->vma, fe->pmd, fe->address);
+	__split_huge_pmd(fe->vma, fe->pmd, fe->address, false, NULL);
 
 	return VM_FAULT_FALLBACK;
 }
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -496,7 +496,7 @@ static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 			page = pmd_page(*pmd);
 			if (is_huge_zero_page(page)) {
 				spin_unlock(ptl);
-				split_huge_pmd(vma, pmd, addr);
+				__split_huge_pmd(vma, pmd, addr, false, NULL);
 			} else {
 				get_page(page);
 				spin_unlock(ptl);
diff --git a/mm/mprotect.c b/mm/mprotect.c
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -164,7 +164,7 @@ static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 
 		if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 			if (next - addr != HPAGE_PMD_SIZE) {
-				split_huge_pmd(vma, pmd, addr);
+				__split_huge_pmd(vma, pmd, addr, false, NULL);
 				if (pmd_trans_unstable(pmd))
 					continue;
 			} else {

[toc] | [next] | [standalone]

#1503435

On Tue, 18 Oct 2016 16:04:06 -0700 (PDT) David Rientjes <rientjes@google.com> wrote:

> While doing MADV_DONTNEED on a large area of thp memory, I noticed we 
> encountered many unlikely() branches in profiles for each backing 
> hugepage.  This is because zap_pmd_range() would call split_huge_pmd(), 
> which rechecked the conditions that were already validated, but as part of 
> an unlikely() branch.
> 
> Avoid the unlikely() branch when in a context where pmd is known to be 
> good for __split_huge_pmd() directly.

Before:

   text    data     bss     dec     hex filename
  38442      75      48   38565    96a5 mm/memory.o
  21755    2369   18464   42588    a65c mm/mempolicy.o
   4557    1816       0    6373    18e5 mm/mprotect.o

After:

  38362      75      48   38485    9655 mm/memory.o
  21714    2369   18464   42547    a633 mm/mempolicy.o
   4541    1816       0    6357    18d5 mm/mprotect.o


So there's a size improvment too.  gcc-4.4.4.

[toc] | [prev] | [next] | [standalone]

#1503624

On 10/19/2016 01:04 AM, David Rientjes wrote:
> While doing MADV_DONTNEED on a large area of thp memory, I noticed we
> encountered many unlikely() branches in profiles for each backing
> hugepage.  This is because zap_pmd_range() would call split_huge_pmd(),
> which rechecked the conditions that were already validated, but as part of
> an unlikely() branch.

I'm not sure which unlikely() branch you mean here, as I don't see any in the 
split_huge_pmd() macro or the functions it calls? So is it the branches that the 
profiler flagged as mispredicted using some PMC event? In that case it's perhaps 
confusing to call it "unlikely()".

> Avoid the unlikely() branch when in a context where pmd is known to be
> good for __split_huge_pmd() directly.
>
> Signed-off-by: David Rientjes <rientjes@google.com>

That said, this makes sense. You could probably convert also:

    3    281  mm/gup.c <<follow_page_mask>>
              split_huge_pmd(vma, pmd, address);
   11    212  mm/mremap.c <<move_page_tables>>
              split_huge_pmd(vma, old_pmd, old_addr);

Acked-by: Vlastimil Babka <vbabka@suse.cz>

> ---
>  include/linux/huge_mm.h | 2 ++
>  mm/memory.c             | 4 ++--
>  mm/mempolicy.c          | 2 +-
>  mm/mprotect.c           | 2 +-
>  4 files changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> --- a/include/linux/huge_mm.h
> +++ b/include/linux/huge_mm.h
> @@ -189,6 +189,8 @@ static inline void deferred_split_huge_page(struct page *page) {}
>  #define split_huge_pmd(__vma, __pmd, __address)	\
>  	do { } while (0)
>
> +static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> +		unsigned long address, bool freeze, struct page *page) {}
>  static inline void split_huge_pmd_address(struct vm_area_struct *vma,
>  		unsigned long address, bool freeze, struct page *page) {}
>
> diff --git a/mm/memory.c b/mm/memory.c
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1240,7 +1240,7 @@ static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
>  			if (next - addr != HPAGE_PMD_SIZE) {
>  				VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
>  				    !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
> -				split_huge_pmd(vma, pmd, addr);
> +				__split_huge_pmd(vma, pmd, addr, false, NULL);
>  			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
>  				goto next;
>  			/* fall through */
> @@ -3454,7 +3454,7 @@ static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
>
>  	/* COW handled on pte level: split pmd */
>  	VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
> -	split_huge_pmd(fe->vma, fe->pmd, fe->address);
> +	__split_huge_pmd(fe->vma, fe->pmd, fe->address, false, NULL);
>
>  	return VM_FAULT_FALLBACK;
>  }
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -496,7 +496,7 @@ static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
>  			page = pmd_page(*pmd);
>  			if (is_huge_zero_page(page)) {
>  				spin_unlock(ptl);
> -				split_huge_pmd(vma, pmd, addr);
> +				__split_huge_pmd(vma, pmd, addr, false, NULL);
>  			} else {
>  				get_page(page);
>  				spin_unlock(ptl);
> diff --git a/mm/mprotect.c b/mm/mprotect.c
> --- a/mm/mprotect.c
> +++ b/mm/mprotect.c
> @@ -164,7 +164,7 @@ static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
>
>  		if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
>  			if (next - addr != HPAGE_PMD_SIZE) {
> -				split_huge_pmd(vma, pmd, addr);
> +				__split_huge_pmd(vma, pmd, addr, false, NULL);
>  				if (pmd_trans_unstable(pmd))
>  					continue;
>  			} else {
>

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web