mm/mmap: Fix error path in do_vmi_align_munmap()
authorLiam R. Howlett <Liam.Howlett@oracle.com>
Sun, 18 Jun 2023 00:47:08 +0000 (20:47 -0400)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 18 Jun 2023 16:32:17 +0000 (09:32 -0700)
The error unrolling was leaving the VMAs detached in many cases and
leaving the locked_vm statistic altered, and skipping the unrolling
entirely in the case of the vma tree write failing.

Fix the error path by re-attaching the detached VMAs and adding the
necessary goto for the failed vma tree write, and fix the locked_vm
statistic by only updating after the vma tree write succeeds.

Fixes: 763ecb035029 ("mm: remove the vma linked list")
Reported-by: Vegard Nossum <vegard.nossum@oracle.com>
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mmap.c

index 13678ed..d600404 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2318,21 +2318,6 @@ int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
        return __split_vma(vmi, vma, addr, new_below);
 }
 
-static inline int munmap_sidetree(struct vm_area_struct *vma,
-                                  struct ma_state *mas_detach)
-{
-       vma_start_write(vma);
-       mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
-       if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
-               return -ENOMEM;
-
-       vma_mark_detached(vma, true);
-       if (vma->vm_flags & VM_LOCKED)
-               vma->vm_mm->locked_vm -= vma_pages(vma);
-
-       return 0;
-}
-
 /*
  * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
  * @vmi: The vma iterator
@@ -2354,6 +2339,7 @@ do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
        struct maple_tree mt_detach;
        int count = 0;
        int error = -ENOMEM;
+       unsigned long locked_vm = 0;
        MA_STATE(mas_detach, &mt_detach, 0, 0);
        mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
        mt_set_external_lock(&mt_detach, &mm->mmap_lock);
@@ -2399,9 +2385,13 @@ do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
                        if (error)
                                goto end_split_failed;
                }
-               error = munmap_sidetree(next, &mas_detach);
-               if (error)
-                       goto munmap_sidetree_failed;
+               vma_start_write(next);
+               mas_set_range(&mas_detach, next->vm_start, next->vm_end - 1);
+               if (mas_store_gfp(&mas_detach, next, GFP_KERNEL))
+                       goto munmap_gather_failed;
+               vma_mark_detached(next, true);
+               if (next->vm_flags & VM_LOCKED)
+                       locked_vm += vma_pages(next);
 
                count++;
 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
@@ -2447,10 +2437,12 @@ do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
        }
 #endif
        /* Point of no return */
+       error = -ENOMEM;
        vma_iter_set(vmi, start);
        if (vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL))
-               return -ENOMEM;
+               goto clear_tree_failed;
 
+       mm->locked_vm -= locked_vm;
        mm->map_count -= count;
        /*
         * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
@@ -2480,9 +2472,14 @@ do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
        validate_mm(mm);
        return downgrade ? 1 : 0;
 
+clear_tree_failed:
 userfaultfd_error:
-munmap_sidetree_failed:
+munmap_gather_failed:
 end_split_failed:
+       mas_set(&mas_detach, 0);
+       mas_for_each(&mas_detach, next, end)
+               vma_mark_detached(next, false);
+
        __mt_destroy(&mt_detach);
 start_split_failed:
 map_count_exceeded: