Re: [PATCH bpf-next] bpf: Allow bpf_d_path in perf_event_mmap

From: Florent Revest
Date: Tue Nov 02 2021 - 19:03:59 EST


On Tue, Nov 2, 2021 at 5:06 AM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Mon, Nov 1, 2021 at 8:20 PM Alexei Starovoitov
> <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Mon, Nov 1, 2021 at 8:16 PM Andrii Nakryiko
> > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > > >
> > > > > FILE *vm_file = vma->vm_file; /* no checking is needed, vma from
> > > > > parameter which is not NULL */
> > > > > if (vm_file)
> > > > > bpf_d_path(&vm_file->f_path, path, sizeof(path));
> > > >
> > > > That should work.
> > > > The verifier can achieve that by marking certain fields as PTR_TO_BTF_ID_OR_NULL
> > > > instead of PTR_TO_BTF_ID while walking such pointers.
> > > > And then disallow pointer arithmetic on PTR_TO_BTF_ID_OR_NULL until it
> > > > goes through 'if (Rx == NULL)' check inside the program and gets converted to
> > > > PTR_TO_BTF_ID.
> > > > Initially we can hard code such fields via BTF_ID(struct, file) macro.'
> > > > So any pointer that results into a 'struct file' pointer will be
> > > > PTR_TO_BTF_ID_OR_NULL.

Right, this is what I had in mind originally. But I was afraid this
could maybe prevent some existing programs from loading on newer
kernels ? Not sure if that's an issue.

> > The helper can check that it's [0, few_pages] and declare it's bad.
>
> That's basically what happens with direct memory reads, so I guess it
> would be fine.
>
> > I guess we can do that and only do what I proposed for "more than a page"
> > math on the pointer. Or even disallow "add more than a page offset to
> > PTR_TO_BTF_ID"
> > for now, since it will cover 99% of the cases.

Otherwise this sounds like a straightforward solution, yes :)
Especially if this is how direct memory accesses already work.

I'd be happy to look into this when I get some slack time. ;)