thp: implement split_huge_pmd()
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Sat, 16 Jan 2016 00:53:53 +0000 (16:53 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 16 Jan 2016 01:56:32 +0000 (17:56 -0800)
Original split_huge_page() combined two operations: splitting PMDs into
tables of PTEs and splitting underlying compound page.  This patch
implements split_huge_pmd() which split given PMD without splitting
other PMDs this page mapped with or underlying compound page.

Without tail page refcounting, implementation of split_huge_pmd() is
pretty straight-forward.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Tested-by: Sasha Levin <sasha.levin@oracle.com>
Tested-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Jerome Marchand <jmarchan@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Steve Capper <steve.capper@linaro.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/huge_mm.h
mm/huge_memory.c

index 333b058..f1fa1c2 100644 (file)
@@ -92,7 +92,16 @@ extern unsigned long transparent_hugepage_flags;
 
 #define split_huge_page_to_list(page, list) BUILD_BUG()
 #define split_huge_page(page) BUILD_BUG()
-#define split_huge_pmd(__vma, __pmd, __address) BUILD_BUG()
+
+void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
+               unsigned long address);
+
+#define split_huge_pmd(__vma, __pmd, __address)                                \
+       do {                                                            \
+               pmd_t *____pmd = (__pmd);                               \
+               if (pmd_trans_huge(*____pmd))                           \
+                       __split_huge_pmd(__vma, __pmd, __address);      \
+       }  while (0)
 
 #if HPAGE_PMD_ORDER >= MAX_ORDER
 #error "hugepages can't be allocated by the buddy allocator"
index 1588f68..22ab365 100644 (file)
@@ -2666,6 +2666,130 @@ static int khugepaged(void *none)
        return 0;
 }
 
+static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
+               unsigned long haddr, pmd_t *pmd)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       pgtable_t pgtable;
+       pmd_t _pmd;
+       int i;
+
+       /* leave pmd empty until pte is filled */
+       pmdp_huge_clear_flush_notify(vma, haddr, pmd);
+
+       pgtable = pgtable_trans_huge_withdraw(mm, pmd);
+       pmd_populate(mm, &_pmd, pgtable);
+
+       for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
+               pte_t *pte, entry;
+               entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
+               entry = pte_mkspecial(entry);
+               pte = pte_offset_map(&_pmd, haddr);
+               VM_BUG_ON(!pte_none(*pte));
+               set_pte_at(mm, haddr, pte, entry);
+               pte_unmap(pte);
+       }
+       smp_wmb(); /* make pte visible before pmd */
+       pmd_populate(mm, pmd, pgtable);
+       put_huge_zero_page();
+}
+
+static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
+               unsigned long haddr)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       struct page *page;
+       pgtable_t pgtable;
+       pmd_t _pmd;
+       bool young, write;
+       int i;
+
+       VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
+       VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
+       VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
+       VM_BUG_ON(!pmd_trans_huge(*pmd));
+
+       count_vm_event(THP_SPLIT_PMD);
+
+       if (vma_is_dax(vma)) {
+               pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
+               if (is_huge_zero_pmd(_pmd))
+                       put_huge_zero_page();
+               return;
+       } else if (is_huge_zero_pmd(*pmd)) {
+               return __split_huge_zero_page_pmd(vma, haddr, pmd);
+       }
+
+       page = pmd_page(*pmd);
+       VM_BUG_ON_PAGE(!page_count(page), page);
+       atomic_add(HPAGE_PMD_NR - 1, &page->_count);
+       write = pmd_write(*pmd);
+       young = pmd_young(*pmd);
+
+       /* leave pmd empty until pte is filled */
+       pmdp_huge_clear_flush_notify(vma, haddr, pmd);
+
+       pgtable = pgtable_trans_huge_withdraw(mm, pmd);
+       pmd_populate(mm, &_pmd, pgtable);
+
+       for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
+               pte_t entry, *pte;
+               /*
+                * Note that NUMA hinting access restrictions are not
+                * transferred to avoid any possibility of altering
+                * permissions across VMAs.
+                */
+               entry = mk_pte(page + i, vma->vm_page_prot);
+               entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+               if (!write)
+                       entry = pte_wrprotect(entry);
+               if (!young)
+                       entry = pte_mkold(entry);
+               pte = pte_offset_map(&_pmd, haddr);
+               BUG_ON(!pte_none(*pte));
+               set_pte_at(mm, haddr, pte, entry);
+               atomic_inc(&page[i]._mapcount);
+               pte_unmap(pte);
+       }
+
+       /*
+        * Set PG_double_map before dropping compound_mapcount to avoid
+        * false-negative page_mapped().
+        */
+       if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
+               for (i = 0; i < HPAGE_PMD_NR; i++)
+                       atomic_inc(&page[i]._mapcount);
+       }
+
+       if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
+               /* Last compound_mapcount is gone. */
+               __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
+               if (TestClearPageDoubleMap(page)) {
+                       /* No need in mapcount reference anymore */
+                       for (i = 0; i < HPAGE_PMD_NR; i++)
+                               atomic_dec(&page[i]._mapcount);
+               }
+       }
+
+       smp_wmb(); /* make pte visible before pmd */
+       pmd_populate(mm, pmd, pgtable);
+}
+
+void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
+               unsigned long address)
+{
+       spinlock_t *ptl;
+       struct mm_struct *mm = vma->vm_mm;
+       unsigned long haddr = address & HPAGE_PMD_MASK;
+
+       mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
+       ptl = pmd_lock(mm, pmd);
+       if (likely(pmd_trans_huge(*pmd)))
+               __split_huge_pmd_locked(vma, pmd, haddr);
+       spin_unlock(ptl);
+       mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
+}
+
 static void split_huge_pmd_address(struct vm_area_struct *vma,
                                    unsigned long address)
 {