Thorvald reported a WARNING [1]. And the root cause is below race: CPU 1 CPU 2 fork hugetlbfs_fallocate dup_mmap hugetlbfs_punch_hole i_mmap_lock_write(mapping); vma_interval_tree_insert_after -- Child vma is visible through i_mmap tree. i_mmap_unlock_write(mapping); hugetlb_dup_vma_private -- Clear vma_lock outside i_mmap_rwsem! i_mmap_lock_write(mapping); hugetlb_vmdelete_list vma_interval_tree_foreach hugetlb_vma_trylock_write -- Vma_lock is cleared. tmp->vm_ops->open -- Alloc new vma_lock outside i_mmap_rwsem! hugetlb_vma_unlock_write -- Vma_lock is assigned!!! i_mmap_unlock_write(mapping); hugetlb_dup_vma_private() and hugetlb_vm_op_open() are called outside i_mmap_rwsem lock while vma lock can be used in the same time. Fix this by deferring linking file vma until vma is fully initialized. Those vmas should be initialized first before they can be used.
Unified · Split
| @@ -714,6 +714,23 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm, · 713 unmodified lines | |||
| 714 | } else if (anon_vma_fork(tmp, mpnt)) | 714 | } else if (anon_vma_fork(tmp, mpnt)) |
| 715 | goto fail_nomem_anon_vma_fork; | 715 | goto fail_nomem_anon_vma_fork; |
| 716 | vm_flags_clear(tmp, VM_LOCKED_MASK); | 716 | vm_flags_clear(tmp, VM_LOCKED_MASK); |
| 717 | /* | ||
| 718 | * Copy/update hugetlb private vma information. | ||
| 719 | */ | ||
| 720 | if (is_vm_hugetlb_page(tmp)) | ||
| 721 | hugetlb_dup_vma_private(tmp); | ||
| 722 | |||
| 723 | /* | ||
| 724 | * Link the vma into the MT. After using __mt_dup(), memory | ||
| 725 | * allocation is not necessary here, so it cannot fail. | ||
| 726 | */ | ||
| 727 | vma_iter_bulk_store(&vmi, tmp); | ||
| 728 | |||
| 729 | mm->map_count++; | ||
| 730 | |||
| 731 | if (tmp->vm_ops && tmp->vm_ops->open) | ||
| 732 | tmp->vm_ops->open(tmp); | ||
| 733 | |||
| 717 | file = tmp->vm_file; | 734 | file = tmp->vm_file; |
| 718 | if (file) { | 735 | if (file) { |
| 719 | struct address_space *mapping = file->f_mapping; | 736 | struct address_space *mapping = file->f_mapping; |
| @@ -730,25 +747,9 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm, · 10 unmodified lines | |||
| 730 | i_mmap_unlock_write(mapping); | 747 | i_mmap_unlock_write(mapping); |
| 731 | } | 748 | } |
| 732 | 749 | ||
| 733 | /* | ||
| 734 | * Copy/update hugetlb private vma information. | ||
| 735 | */ | ||
| 736 | if (is_vm_hugetlb_page(tmp)) | ||
| 737 | hugetlb_dup_vma_private(tmp); | ||
| 738 | |||
| 739 | /* | ||
| 740 | * Link the vma into the MT. After using __mt_dup(), memory | ||
| 741 | * allocation is not necessary here, so it cannot fail. | ||
| 742 | */ | ||
| 743 | vma_iter_bulk_store(&vmi, tmp); | ||
| 744 | |||
| 745 | mm->map_count++; | ||
| 746 | if (!(tmp->vm_flags & VM_WIPEONFORK)) | 750 | if (!(tmp->vm_flags & VM_WIPEONFORK)) |
| 747 | retval = copy_page_range(tmp, mpnt); | 751 | retval = copy_page_range(tmp, mpnt); |
| 748 | 752 | ||
| 749 | if (tmp->vm_ops && tmp->vm_ops->open) | ||
| 750 | tmp->vm_ops->open(tmp); | ||
| 751 | |||
| 752 | if (retval) { | 753 | if (retval) { |
| 753 | mpnt = vma_next(&vmi); | 754 | mpnt = vma_next(&vmi); |
| 754 | goto loop_out; | 755 | goto loop_out; |