On Fri, Apr 10, 2020 at 1:51 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
>
> On Fri, Apr 10, 2020 at 2:04 AM Andrii Nakryiko <andriin@xxxxxx> wrote:
> > VM_MAYWRITE flag during initial memory mapping determines if already mmap()'ed
> > pages can be later remapped as writable ones through mprotect() call. To
> > prevent user application to rewrite contents of memory-mapped as read-only and
> > subsequently frozen BPF map, remove VM_MAYWRITE flag completely on initially
> > read-only mapping.
> >
> > Alternatively, we could treat any memory-mapping on unfrozen map as writable
> > and bump writecnt instead. But there is little legitimate reason to map
> > BPF map as read-only and then re-mmap() it as writable through mprotect(),
> > instead of just mmap()'ing it as read/write from the very beginning.
> >
> > Fixes: fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
> > Reported-by: Jann Horn <jannh@xxxxxxxxxx>
> > Signed-off-by: Andrii Nakryiko <andriin@xxxxxx>
> > ---
> > kernel/bpf/syscall.c | 4 ++++
> > 1 file changed, 4 insertions(+)
> >
> > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > index 64783da34202..f7f6db50a085 100644
> > --- a/kernel/bpf/syscall.c
> > +++ b/kernel/bpf/syscall.c
> > @@ -635,6 +635,10 @@ static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
> > /* set default open/close callbacks */
> > vma->vm_ops = &bpf_map_default_vmops;
> > vma->vm_private_data = map;
> > + vma->vm_flags &= ~VM_MAYEXEC;
> > + if (!(vma->vm_flags & VM_WRITE))
> > + /* disallow re-mapping with PROT_WRITE */
> > + vma->vm_flags &= ~VM_MAYWRITE;
>
> The .open and .close handlers for the VMA are also wrong:
Yes, it has to check VM_MAYWRITE now, my bad, thanks for catching
this! Extended selftest to validate that scenario as well.
>
> /* called for any extra memory-mapped regions (except initial) */
> static void bpf_map_mmap_open(struct vm_area_struct *vma)
> {
> struct bpf_map *map = vma->vm_file->private_data;
>
> bpf_map_inc_with_uref(map);
>
> if (vma->vm_flags & VM_WRITE) {
> mutex_lock(&map->freeze_mutex);
> map->writecnt++;
> mutex_unlock(&map->freeze_mutex);
> }
> }
>
> /* called for all unmapped memory region (including initial) */
> static void bpf_map_mmap_close(struct vm_area_struct *vma)
> {
> struct bpf_map *map = vma->vm_file->private_data;
>
> if (vma->vm_flags & VM_WRITE) {
> mutex_lock(&map->freeze_mutex);
> map->writecnt--;
> mutex_unlock(&map->freeze_mutex);
> }
>
> bpf_map_put_with_uref(map);
> }
>
> static const struct vm_operations_struct bpf_map_default_vmops = {
> .open = bpf_map_mmap_open,
> .close = bpf_map_mmap_close,
> };
>
> You can use mprotect() to flip VM_WRITE off while a VMA exists, and
> then the writecnt won't go down when you close it. Or you could even
> get the writecnt to go negative if you map as writable, then
> mprotect() to readonly, then split the VMA a few times, mprotect the
> split VMAs to writable, and then unmap them.
>
> I think you'll want to also check for MAYWRITE here.
>
> Also, the bpf_map_inc_with_uref/bpf_map_put_with_uref here look
> superfluous - the VMA holds a reference to the file, and the file
> holds a reference to the map.
Hm.. So the file from which memory-mapping was created originally will
stay referenced by VMA subsystem until the last VMA segment is
unmmaped (and bpf_map_mmap_close is called for it), even if file
itself is closed from user-space? It makes sense, though I didn't
realize it at the time I was implementing this. I'll drop refcounting
then.
