[PATCH] mm/memory.c: do_fault: avoid usage of stale vm_area_struct

[Jan Stancek]

LTP testcase mtest06 [1] can trigger a crash on s390x running 5.0.0-rc8.
This is a stress test, where one thread mmaps/writes/munmaps memory area
and other thread is trying to read from it:

  CPU: 0 PID: 2611 Comm: mmap1 Not tainted 5.0.0-rc8+ #51
  Hardware name: IBM 2964 N63 400 (z/VM 6.4.0)
  Krnl PSW : 0404e00180000000 00000000001ac8d8 (__lock_acquire+0x7/0x7a8)
  Call Trace:
  ([<0000000000000000>]           (null))
   [<00000000001adae4>] lock_acquire+0xec/0x258
   [<000000000080d1ac>] _raw_spin_lock_bh+0x5c/0x98
   [<000000000012a780>] page_table_free+0x48/0x1a8
   [<00000000002f6e54>] do_fault+0xdc/0x670
   [<00000000002fadae>] __handle_mm_fault+0x416/0x5f0
   [<00000000002fb138>] handle_mm_fault+0x1b0/0x320
   [<00000000001248cc>] do_dat_exception+0x19c/0x2c8
   [<000000000080e5ee>] pgm_check_handler+0x19e/0x200

page_table_free() is called with NULL mm parameter, but because
"0" is a valid address on s390 (see S390_lowcore), it keeps
going until it eventually crashes in lockdep's lock_acquire.
This crash is reproducible at least since 4.14.

Problem is that "vmf->vma" used in do_fault() can become stale.
Because mmap_sem may be released, other threads can come in,
call munmap() and cause "vma" be returned to kmem cache, and
get zeroed/re-initialized and re-used:

handle_mm_fault                           |
  __handle_mm_fault                       |
    do_fault                              |
      vma = vmf->vma                      |
      do_read_fault                       |
        __do_fault                        |
          vma->vm_ops->fault(vmf);        |
            mmap_sem is released          |
                                          |
                                          | do_munmap()
                                          |   remove_vma_list()
                                          |     remove_vma()
                                          |       vm_area_free()
                                          |         # vma is released
                                          | ...
                                          | # same vma is allocated
                                          | # from kmem cache
                                          | do_mmap()
                                          |   vm_area_alloc()
                                          |     memset(vma, 0, ...)
                                          |
      pte_free(vma->vm_mm, ...);          |
        page_table_free                   |
          spin_lock_bh(&mm->context.lock);|
            <crash>                       |

This patch pins mm_struct and stores its value, to avoid using
potentially stale "vma" when calling pte_free().

[1] https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/mem/mtest06/mmap1.c

Signed-off-by: Jan Stancek <jstancek@xxxxxxxxxx>
---
 mm/memory.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/mm/memory.c b/mm/memory.c
index e11ca9dd823f..1287ee9acbdc 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3517,12 +3517,17 @@ static vm_fault_t do_shared_fault(struct vm_fault *vmf)
  * but allow concurrent faults).
  * The mmap_sem may have been released depending on flags and our
  * return value.  See filemap_fault() and __lock_page_or_retry().
+ * If mmap_sem is released, vma may become invalid (for example
+ * by other thread calling munmap()).
  */
 static vm_fault_t do_fault(struct vm_fault *vmf)
 {
 	struct vm_area_struct *vma = vmf->vma;
+	struct mm_struct *vm_mm = READ_ONCE(vma->vm_mm);
 	vm_fault_t ret;
 
+	mmgrab(vm_mm);
+
 	/*
 	 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
 	 */
@@ -3561,9 +3566,12 @@ static vm_fault_t do_fault(struct vm_fault *vmf)
 
 	/* preallocated pagetable is unused: free it */
 	if (vmf->prealloc_pte) {
-		pte_free(vma->vm_mm, vmf->prealloc_pte);
+		pte_free(vm_mm, vmf->prealloc_pte);
 		vmf->prealloc_pte = NULL;
 	}
+
+	mmdrop(vm_mm);
+
 	return ret;
 }
 
-- 
1.8.3.1