[RESEND,v5,3/3] mm: avoid unnecessary flush on change_huge_pmd()

Commit Message

Nadav Amit March 20, 2022, 12:07 a.m. UTC

From: Nadav Amit <[e-mail]>

Calls to change_protection_range() on THP can trigger, at least on x86,
two TLB flushes for one page: one immediately, when pmdp_invalidate() is
called by change_huge_pmd(), and then another one later (that can be
batched) when change_protection_range() finishes.

The first TLB flush is only necessary to prevent the dirty bit (and with
a lesser importance the access bit) from changing while the PTE is
modified. However, this is not necessary as the x86 CPUs set the
dirty-bit atomically with an additional check that the PTE is (still)
present. One caveat is Intel's Knights Landing that has a bug and does
not do so.

Leverage this behavior to eliminate the unnecessary TLB flush in
change_huge_pmd(). Introduce a new arch specific pmdp_invalidate_ad()
that only invalidates the access and dirty bit from further changes.

Cc: Andrea Arcangeli <[e-mail]>
Cc: Andrew Cooper <[e-mail]>
Cc: Andrew Morton <[e-mail]>
Cc: Andy Lutomirski <[e-mail]>
Cc: Dave Hansen <[e-mail]>
Cc: Peter Xu <[e-mail]>
Cc: Peter Zijlstra <[e-mail]>
Cc: Thomas Gleixner <[e-mail]>
Cc: Will Deacon <[e-mail]>
Cc: Yu Zhao <[e-mail]>
Cc: Nick Piggin <[e-mail]>
Cc: [e-mail]
Signed-off-by: Nadav Amit <[e-mail]>
---
 arch/x86/include/asm/pgtable.h |  5 +++++
 arch/x86/mm/pgtable.c          | 10 ++++++++++
 include/linux/pgtable.h        | 20 ++++++++++++++++++++
 mm/huge_memory.c               |  4 ++--
 mm/pgtable-generic.c           |  8 ++++++++
 5 files changed, 45 insertions(+), 2 deletions(-)

Patch

diff --git a/arch/x86/include/asm/pgtable.h b/arch/x86/include/asm/pgtable.h
index 62ab07e24aef..23ad34edcc4b 100644
--- a/arch/x86/include/asm/pgtable.h
+++ b/arch/x86/include/asm/pgtable.h
@@ -1173,6 +1173,11 @@  static inline pmd_t pmdp_establish(struct vm_area_struct *vma,
 	}
 }
 #endif
+
+#define __HAVE_ARCH_PMDP_INVALIDATE_AD
+extern pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma,
+				unsigned long address, pmd_t *pmdp);
+
 /*
  * Page table pages are page-aligned.  The lower half of the top
  * level is used for userspace and the top half for the kernel.
diff --git a/arch/x86/mm/pgtable.c b/arch/x86/mm/pgtable.c
index 3481b35cb4ec..f16059e9a85e 100644
--- a/arch/x86/mm/pgtable.c
+++ b/arch/x86/mm/pgtable.c
@@ -608,6 +608,16 @@  int pmdp_clear_flush_young(struct vm_area_struct *vma,
 
 	return young;
 }
+
+pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma, unsigned long address,
+			 pmd_t *pmdp)
+{
+	/*
+	 * No flush is necessary. Once an invalid PTE is established, the PTE's
+	 * access and dirty bits cannot be updated.
+	 */
+	return pmdp_establish(vma, address, pmdp, pmd_mkinvalid(*pmdp));
+}
 #endif
 
 /**
diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
index f4f4077b97aa..5826e8e52619 100644
--- a/include/linux/pgtable.h
+++ b/include/linux/pgtable.h
@@ -570,6 +570,26 @@  extern pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
 			    pmd_t *pmdp);
 #endif
 
+#ifndef __HAVE_ARCH_PMDP_INVALIDATE_AD
+
+/*
+ * pmdp_invalidate_ad() invalidates the PMD while changing a transparent
+ * hugepage mapping in the page tables. This function is similar to
+ * pmdp_invalidate(), but should only be used if the access and dirty bits would
+ * not be cleared by the software in the new PMD value. The function ensures
+ * that hardware changes of the access and dirty bits updates would not be lost.
+ *
+ * Doing so can allow in certain architectures to avoid a TLB flush in most
+ * cases. Yet, another TLB flush might be necessary later if the PMD update
+ * itself requires such flush (e.g., if protection was set to be stricter). Yet,
+ * even when a TLB flush is needed because of the update, the caller may be able
+ * to batch these TLB flushing operations, so fewer TLB flush operations are
+ * needed.
+ */
+extern pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma,
+				unsigned long address, pmd_t *pmdp);
+#endif
+
 #ifndef __HAVE_ARCH_PTE_SAME
 static inline int pte_same(pte_t pte_a, pte_t pte_b)
 {
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 51b0f3cb1ba0..691d80edcfd7 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1781,10 +1781,10 @@  int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
 	 * which may break userspace.
 	 *
-	 * pmdp_invalidate() is required to make sure we don't miss
+	 * pmdp_invalidate_ad() is required to make sure we don't miss
 	 * dirty/young flags set by hardware.
 	 */
-	oldpmd = pmdp_invalidate(vma, addr, pmd);
+	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
 
 	entry = pmd_modify(oldpmd, newprot);
 	if (preserve_write)
diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c
index 6523fda274e5..90ab721a12a8 100644
--- a/mm/pgtable-generic.c
+++ b/mm/pgtable-generic.c
@@ -201,6 +201,14 @@  pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
 }
 #endif
 
+#ifndef __HAVE_ARCH_PMDP_INVALIDATE_AD
+pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma, unsigned long address,
+			 pmd_t *pmdp)
+{
+	return pmdp_invalidate(vma, address, pmdp);
+}
+#endif
+
 #ifndef pmdp_collapse_flush
 pmd_t pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
 			  pmd_t *pmdp)