On Thu, Dec 10, 2020 at 10:54:05PM +0000, Al Viro wrote: > On Thu, Dec 10, 2020 at 11:30:24PM +0100, Christian Brauner wrote: > > (requiring btf), i.e. security_file_open, then follow > > file->f_inode->i_sb->s_type->s_magic. If we change the say struct > > super_block I'd expect these bpf programs to break. > > To break they would need to have compiled in the first place; > ->s_type is struct file_system_type and it contains no ->s_magic > (nor would it be possible, really - ->s_magic can vary between > filesystems that *do* share ->s_type). Incidentally, a lot of things in e.g. struct dentry need care when accessing; the fields are there, but e.g. blind access to name or parent really can oops. Moreover, blindly following a chain of ->d_parent pointers without taking appropriate precautions might end up reading from arbitrary kernel address, including iomem ones. I don't see anything that would prevent that... TAINT_BPF would probably be too impractical, since there's a lot of boxen using it more reasonably on the networking side. But it really looks like we *do* need annotations with their violation triggering a taint, so that BS bug reports could be discarded.
