thp: fix split_huge_page() after mremap() of THP
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Sat, 16 Jan 2016 00:55:46 +0000 (16:55 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 16 Jan 2016 01:56:32 +0000 (17:56 -0800)
Sasha Levin has reported KASAN out-of-bounds bug[1].  It points to "if
(!is_swap_pte(pte[i]))" in unfreeze_page_vma() as a problematic access.

The cause is that split_huge_page() doesn't handle THP correctly if it's
not allingned to PMD boundary.  It can happen after mremap().

Test-case (not always triggers the bug):

#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <sys/mman.h>

#define MB (1024UL*1024)
#define SIZE (2*MB)
#define BASE ((void *)0x400000000000)

int main()
{
char *p;

p = mmap(BASE, SIZE, PROT_READ | PROT_WRITE,
MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE,
-1, 0);
if (p == MAP_FAILED)
perror("mmap"), exit(1);
p = mremap(BASE, SIZE, SIZE, MREMAP_FIXED | MREMAP_MAYMOVE,
BASE + SIZE + 8192);
if (p == MAP_FAILED)
perror("mremap"), exit(1);
system("echo 1 > /sys/kernel/debug/split_huge_pages");
return 0;
}

The patch fixes freeze and unfreeze paths to handle page table boundary
crossing.

It also makes mapcount vs count check in split_huge_page_to_list()
stricter:
 - after freeze we don't expect any subpage mapped as we remove them
   from rmap when setting up migration entries;
 - count must be 1, meaning only caller has reference to the page;

[1] https://gist.github.com/sashalevin/c67fbea55e7c0576972a

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c

index 1a4989fef08fd83b06565e8c4b4106d8661eaf53..9d12d63a0ddddc9c9d52fb11643f31a5252d18ea 100644 (file)
@@ -2990,6 +2990,7 @@ void vma_adjust_trans_huge(struct vm_area_struct *vma,
 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
                unsigned long address)
 {
+       unsigned long haddr = address & HPAGE_PMD_MASK;
        spinlock_t *ptl;
        pgd_t *pgd;
        pud_t *pud;
@@ -3019,34 +3020,47 @@ static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
        }
        if (pmd_trans_huge(*pmd)) {
                if (page == pmd_page(*pmd))
-                       __split_huge_pmd_locked(vma, pmd, address, true);
+                       __split_huge_pmd_locked(vma, pmd, haddr, true);
                spin_unlock(ptl);
                return;
        }
        spin_unlock(ptl);
 
        pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
-       for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
+       for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
                pte_t entry, swp_pte;
                swp_entry_t swp_entry;
 
-               if (!pte_present(pte[i]))
+               /*
+                * We've just crossed page table boundary: need to map next one.
+                * It can happen if THP was mremaped to non PMD-aligned address.
+                */
+               if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
+                       pte_unmap_unlock(pte - 1, ptl);
+                       pmd = mm_find_pmd(vma->vm_mm, address);
+                       if (!pmd)
+                               return;
+                       pte = pte_offset_map_lock(vma->vm_mm, pmd,
+                                       address, &ptl);
+               }
+
+               if (!pte_present(*pte))
                        continue;
-               if (page_to_pfn(page) != pte_pfn(pte[i]))
+               if (page_to_pfn(page) != pte_pfn(*pte))
                        continue;
                flush_cache_page(vma, address, page_to_pfn(page));
-               entry = ptep_clear_flush(vma, address, pte + i);
+               entry = ptep_clear_flush(vma, address, pte);
                if (pte_dirty(entry))
                        SetPageDirty(page);
                swp_entry = make_migration_entry(page, pte_write(entry));
                swp_pte = swp_entry_to_pte(swp_entry);
                if (pte_soft_dirty(entry))
                        swp_pte = pte_swp_mksoft_dirty(swp_pte);
-               set_pte_at(vma->vm_mm, address, pte + i, swp_pte);
+               set_pte_at(vma->vm_mm, address, pte, swp_pte);
                page_remove_rmap(page, false);
                put_page(page);
        }
-       pte_unmap_unlock(pte, ptl);
+       pte_unmap_unlock(pte - 1, ptl);
 }
 
 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
@@ -3058,14 +3072,13 @@ static void freeze_page(struct anon_vma *anon_vma, struct page *page)
 
        anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
                        pgoff + HPAGE_PMD_NR - 1) {
-               unsigned long haddr;
+               unsigned long address = __vma_address(page, avc->vma);
 
-               haddr = __vma_address(page, avc->vma) & HPAGE_PMD_MASK;
                mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
-                               haddr, haddr + HPAGE_PMD_SIZE);
-               freeze_page_vma(avc->vma, page, haddr);
+                               address, address + HPAGE_PMD_SIZE);
+               freeze_page_vma(avc->vma, page, address);
                mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
-                               haddr, haddr + HPAGE_PMD_SIZE);
+                               address, address + HPAGE_PMD_SIZE);
        }
 }
 
@@ -3076,6 +3089,7 @@ static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
        pmd_t *pmd;
        pte_t *pte, entry;
        swp_entry_t swp_entry;
+       unsigned long haddr = address & HPAGE_PMD_MASK;
        int i, nr = HPAGE_PMD_NR;
 
        /* Skip pages which doesn't belong to the VMA */
@@ -3089,12 +3103,26 @@ static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
        pmd = mm_find_pmd(vma->vm_mm, address);
        if (!pmd)
                return;
+
        pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
-       for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) {
-               if (!is_swap_pte(pte[i]))
+       for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
+               /*
+                * We've just crossed page table boundary: need to map next one.
+                * It can happen if THP was mremaped to non-PMD aligned address.
+                */
+               if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
+                       pte_unmap_unlock(pte - 1, ptl);
+                       pmd = mm_find_pmd(vma->vm_mm, address);
+                       if (!pmd)
+                               return;
+                       pte = pte_offset_map_lock(vma->vm_mm, pmd,
+                                       address, &ptl);
+               }
+
+               if (!is_swap_pte(*pte))
                        continue;
 
-               swp_entry = pte_to_swp_entry(pte[i]);
+               swp_entry = pte_to_swp_entry(*pte);
                if (!is_migration_entry(swp_entry))
                        continue;
                if (migration_entry_to_page(swp_entry) != page)
@@ -3110,12 +3138,12 @@ static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
                        entry = maybe_mkwrite(entry, vma);
 
                flush_dcache_page(page);
-               set_pte_at(vma->vm_mm, address, pte + i, entry);
+               set_pte_at(vma->vm_mm, address, pte, entry);
 
                /* No need to invalidate - it was non-present before */
-               update_mmu_cache(vma, address, pte + i);
+               update_mmu_cache(vma, address, pte);
        }
-       pte_unmap_unlock(pte, ptl);
+       pte_unmap_unlock(pte - 1, ptl);
 }
 
 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
@@ -3321,7 +3349,7 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
        spin_lock(&split_queue_lock);
        count = page_count(head);
        mapcount = total_mapcount(head);
-       if (mapcount == count - 1) {
+       if (!mapcount && count == 1) {
                if (!list_empty(page_deferred_list(head))) {
                        split_queue_len--;
                        list_del(page_deferred_list(head));
@@ -3329,13 +3357,13 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
                spin_unlock(&split_queue_lock);
                __split_huge_page(page, list);
                ret = 0;
-       } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount > count - 1) {
+       } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
                spin_unlock(&split_queue_lock);
                pr_alert("total_mapcount: %u, page_count(): %u\n",
                                mapcount, count);
                if (PageTail(page))
                        dump_page(head, NULL);
-               dump_page(page, "total_mapcount(head) > page_count(head) - 1");
+               dump_page(page, "total_mapcount(head) > 0");
                BUG();
        } else {
                spin_unlock(&split_queue_lock);