[PATCH uprobe, thp 4/4] uprobe: collapse THP pmd after removing all uprobes

[Song Liu <songliubraving@fb.com>]

After all uprobes are removed from the huge page (with PTE pgtable), it
is possible to collapse the pmd and benefit from THP again. This patch
does the collapse.

An issue on earlier version was discovered by kbuild test robot.

Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
---
 include/linux/huge_mm.h |  9 ++++++++
 kernel/events/uprobes.c |  3 +++
 mm/huge_memory.c        | 47 +++++++++++++++++++++++++++++++++++++++++
 3 files changed, 59 insertions(+)

diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
index 4832d6580969..61f6d574d9b4 100644
--- a/include/linux/huge_mm.h
+++ b/include/linux/huge_mm.h
@@ -252,6 +252,10 @@ static inline bool thp_migration_supported(void)
 	return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 }
 
+extern inline void try_collapse_huge_pmd(struct mm_struct *mm,
+					 struct vm_area_struct *vma,
+					 unsigned long vaddr);
+
 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
@@ -377,6 +381,11 @@ static inline bool thp_migration_supported(void)
 {
 	return false;
 }
+
+static inline void try_collapse_huge_pmd(struct mm_struct *mm,
+					 struct vm_area_struct *vma,
+					 unsigned long vaddr) {}
+
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 
 #endif /* _LINUX_HUGE_MM_H */
diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
index 56eeccc2f7a2..422617bdd5ff 100644
--- a/kernel/events/uprobes.c
+++ b/kernel/events/uprobes.c
@@ -564,6 +564,9 @@ int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
 	if (ret && is_register && ref_ctr_updated)
 		update_ref_ctr(uprobe, mm, -1);
 
+	if (!ret && orig_page && PageTransCompound(orig_page))
+		try_collapse_huge_pmd(mm, vma, vaddr);
+
 	return ret;
 }
 
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 4714871353c0..e2edec3ffd43 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2923,6 +2923,53 @@ static struct shrinker deferred_split_shrinker = {
 	.flags = SHRINKER_NUMA_AWARE,
 };
 
+/**
+ * This function only checks whether all PTEs in this PMD point to
+ * continuous pages, the caller should make sure at least of these PTEs
+ * points to a huge page, e.g. PageTransCompound(one_page) != 0.
+ */
+void try_collapse_huge_pmd(struct mm_struct *mm,
+			   struct vm_area_struct *vma,
+			   unsigned long vaddr)
+{
+	struct mmu_notifier_range range;
+	unsigned long addr;
+	pmd_t *pmd, _pmd;
+	spinlock_t *ptl;
+	long long head;
+	int i;
+
+	pmd = mm_find_pmd(mm, vaddr);
+	if (!pmd)
+		return;
+
+	addr = vaddr & HPAGE_PMD_MASK;
+	head = pte_val(*pte_offset_map(pmd, addr));
+	ptl = pmd_lock(mm, pmd);
+	for (i = 0; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
+		pte_t *pte = pte_offset_map(pmd, addr);
+
+		if (pte_val(*pte) != head + i * PAGE_SIZE) {
+			spin_unlock(ptl);
+			return;
+		}
+	}
+
+	addr = vaddr & HPAGE_PMD_MASK;
+	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
+				addr, addr + HPAGE_PMD_SIZE);
+	mmu_notifier_invalidate_range_start(&range);
+
+	_pmd = pmdp_collapse_flush(vma, addr, pmd);
+	spin_unlock(ptl);
+	mmu_notifier_invalidate_range_end(&range);
+	mm_dec_nr_ptes(mm);
+	pte_free(mm, pmd_pgtable(_pmd));
+	add_mm_counter(mm,
+		       shmem_file(vma->vm_file) ? MM_SHMEMPAGES : MM_FILEPAGES,
+		       -HPAGE_PMD_NR);
+}
+
 #ifdef CONFIG_DEBUG_FS
 static int split_huge_pages_set(void *data, u64 val)
 {
-- 
2.17.1