powerpc/mm/thp: Make page table walk safe against thp split/collapse

We can disable a THP split or a hugepage collapse by disabling irq.
We do send IPI to all the cpus in the early part of split/collapse,
and disabling local irq ensure we don't make progress with
split/collapse. If the THP is getting split we return NULL from
find_linux_pte_or_hugepte(). For all the current callers it should be ok.
We need to be careful if we want to use returned pte_t pointer outside
the irq disabled region. W.r.t to THP split, the pfn remains the same,
but then a hugepage collapse will result in a pfn change. There are
few steps we can take to avoid a hugepage collapse.One way is to take page
reference inside the irq disable region. Other option is to take
mmap_sem so that a parallel collapse will not happen. We can also
disable collapse by taking pmd_lock. Another method used by kvm
subsystem is to check whether we had a mmu_notifer update in between
using mmu_notifier_retry().

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

1 files changed, 16 insertions, 6 deletions

diff --git a/arch/powerpc/mm/hugetlbpage.c b/arch/powerpc/mm/hugetlbpage.c
index 2cb278a2f658..a9dbb27ca887 100644
--- a/arch/powerpc/mm/hugetlbpage.c
+++ b/arch/powerpc/mm/hugetlbpage.c
@@ -109,7 +109,7 @@ int pgd_huge(pgd_t pgd)
 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 {
 	/* Only called for hugetlbfs pages, hence can ignore THP */
-	return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
+	return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
 }
 
 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
@@ -682,28 +682,35 @@ void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 	} while (addr = next, addr != end);
 }
 
+/*
+ * We are holding mmap_sem, so a parallel huge page collapse cannot run.
+ * To prevent hugepage split, disable irq.
+ */
 struct page *
 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 {
 	pte_t *ptep;
 	struct page *page;
 	unsigned shift;
-	unsigned long mask;
+	unsigned long mask, flags;
 	/*
 	 * Transparent hugepages are handled by generic code. We can skip them
 	 * here.
 	 */
+	local_irq_save(flags);
 	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
 
 	/* Verify it is a huge page else bail. */
-	if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
+	if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep)) {
+		local_irq_restore(flags);
 		return ERR_PTR(-EINVAL);
-
+	}
 	mask = (1UL << shift) - 1;
 	page = pte_page(*ptep);
 	if (page)
 		page += (address & mask) / PAGE_SIZE;
 
+	local_irq_restore(flags);
 	return page;
 }
 
@@ -950,9 +957,12 @@ void flush_dcache_icache_hugepage(struct page *page)
  *
  * So long as we atomically load page table pointers we are safe against teardown,
  * we can follow the address down to the the page and take a ref on it.
+ * This function need to be called with interrupts disabled. We use this variant
+ * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
  */
 
-pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
+pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
+				   unsigned *shift)
 {
 	pgd_t pgd, *pgdp;
 	pud_t pud, *pudp;
@@ -1031,7 +1041,7 @@ out:
 		*shift = pdshift;
 	return ret_pte;
 }
-EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
+EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
 
 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 		unsigned long end, int write, struct page **pages, int *nr)