On Fri, Oct 18, 2024 at 1:00 PM Lorenzo Stoakes
<lorenzo.stoakes@xxxxxxxxxx> wrote:
> On Fri, Oct 18, 2024 at 01:49:06PM +0300, Kirill A. Shutemov wrote:
> > On Fri, Oct 18, 2024 at 11:24:06AM +0200, Roberto Sassu wrote:
> > > Probably it is hard, @Kirill would there be any way to safely move
> > > security_mmap_file() out of the mmap_lock lock?
> >
> > What about something like this (untested):
> >
> > diff --git a/mm/mmap.c b/mm/mmap.c
> > index dd4b35a25aeb..03473e77d356 100644
> > --- a/mm/mmap.c
> > +++ b/mm/mmap.c
> > @@ -1646,6 +1646,26 @@ SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
> > if (pgoff + (size >> PAGE_SHIFT) < pgoff)
> > return ret;
> >
> > + if (mmap_read_lock_killable(mm))
> > + return -EINTR;
> > +
> > + vma = vma_lookup(mm, start);
> > +
> > + if (!vma || !(vma->vm_flags & VM_SHARED)) {
> > + mmap_read_unlock(mm);
> > + return -EINVAL;
> > + }
> > +
> > + file = get_file(vma->vm_file);
> > +
> > + mmap_read_unlock(mm);
> > +
> > + ret = security_mmap_file(vma->vm_file, prot, flags);
>
> Accessing VMA fields without any kind of lock is... very much not advised.
>
> I'm guessing you meant to say:
>
> ret = security_mmap_file(file, prot, flags);
>
> Here? :)
>
> I see the original code did this, but obviously was under an mmap lock.
>
> I guess given you check that the file is the same below this.... should be
> fine? Assuming nothing can come in and invalidate the security_mmap_file()
> check in the mean time somehow?
>
> Jann any thoughts?
The overall approach seems reasonable to me - it aligns this path with
the other security_mmap_file() checks, which also don't happen under
the lock.
