Hello Andrea, On Thu, Feb 02, 2017 at 07:02:47PM +0100, Andrea Arcangeli wrote: > On Fri, Jan 27, 2017 at 08:44:32PM +0200, Mike Rapoport wrote: > > - err = -EINVAL; > > + err = -ENOENT; > > dst_vma = find_vma(dst_mm, dst_start); > > if (!dst_vma || !is_vm_hugetlb_page(dst_vma)) > > goto out_unlock; > > + /* > > + * Only allow __mcopy_atomic_hugetlb on userfaultfd > > + * registered ranges. > > + */ > > + if (!dst_vma->vm_userfaultfd_ctx.ctx) > > + goto out_unlock; > > > > + err = -EINVAL; > > if (vma_hpagesize != vma_kernel_pagesize(dst_vma)) > > goto out_unlock; > > That's correct, if a new vma emerges with a different page size it > cannot have a not null dst_vma->vm_userfaultfd_ctx.ctx in the non > cooperative case. > > > @@ -219,12 +226,6 @@ static __always_inline ssize_t > > __mcopy_atomic_hugetlb(struct mm_struct *dst_mm, > > goto out_unlock; > > > > /* > > - * Only allow __mcopy_atomic_hugetlb on userfaultfd registered ranges. > > - */ > > - if (!dst_vma->vm_userfaultfd_ctx.ctx) > > - goto out_unlock; > > - > > - /* > > but this is buggy and it shouldn't be removed, we need this check also > if dst_vma was found not NULL.
The check for not-NULL uffd context is done in __mcopy_atomic, between find_vma and call to __mcopy_atomic_hugetlb. Sp, at this point we verified that dst_vma->vm_userfaultfd_ctx.ctx is not NULL either in the caller, or for the 'retry' case in the hunk above. > > * Ensure the dst_vma has a anon_vma. > > */ > > err = -ENOMEM; > > @@ -368,10 +369,23 @@ static __always_inline ssize_t __mcopy_atomic(struct > > mm_struct *dst_mm, > > * Make sure the vma is not shared, that the dst range is > > * both valid and fully within a single existing vma. > > */ > > - err = -EINVAL; > > + err = -ENOENT; > > dst_vma = find_vma(dst_mm, dst_start); > > if (!dst_vma) > > goto out_unlock; > > + /* > > + * Be strict and only allow __mcopy_atomic on userfaultfd > > + * registered ranges to prevent userland errors going > > + * unnoticed. As far as the VM consistency is concerned, it > > + * would be perfectly safe to remove this check, but there's > > + * no useful usage for __mcopy_atomic ouside of userfaultfd > > + * registered ranges. This is after all why these are ioctls > > + * belonging to the userfaultfd and not syscalls. > > + */ > > + if (!dst_vma->vm_userfaultfd_ctx.ctx) > > + goto out_unlock; > > + > > + err = -EINVAL; > > if (!vma_is_shmem(dst_vma) && dst_vma->vm_flags & VM_SHARED) > > goto out_unlock; > > if (dst_start < dst_vma->vm_start || > > This isn't enough, the -ENOENT should be returned also if the address > doesn't isn't in the range of the found vma, instead of -EINVAL. "vma" > may be a completely different vma just it happen to be way above the > fault address, and the vma previously covering the "addr" (which was > below the found "vma") was already munmapped, so you'd be returning > -EINVAL after munmap still unless the -EINVAL is moved down below. Will fix, thanks. > The check on !vma_is_shmem(dst_vma) && dst_vma->vm_flags & VM_SHARED > instead can be shifted down below after setting err to -EINVAL as then > we know the vma is really the one we were looking for but it's of a > type we can't handle. -- Sincerely yours, Mike.
