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

[Alexei Starovoitov]

On Mon, Nov 1, 2021 at 10:32 AM Yonghong Song <yhs@xxxxxx> wrote:
>
>
>
> On 11/1/21 8:01 AM, Florent Revest wrote:
> > On Mon, Nov 1, 2021 at 2:17 PM Hengqi Chen <hengqi.chen@xxxxxxxxx> wrote:
> >>
> >> Hi,
> >>
> >>
> >> On 2021/10/30 1:02 AM, Florent Revest wrote:
> >>> On Fri, Oct 29, 2021 at 12:47 AM Martin KaFai Lau <kafai@xxxxxx> wrote:
> >>>>
> >>>> On Thu, Oct 28, 2021 at 06:43:57PM +0200, Florent Revest wrote:
> >>>>> Allow the helper to be called from the perf_event_mmap hook. This is
> >>>>> convenient to lookup vma->vm_file and implement a similar logic as
> >>>>> perf_event_mmap_event in BPF.
> >>>>  From struct vm_area_struct:
> >>>>          struct file * vm_file;          /* File we map to (can be NULL). */
> >>>>
> >>>> Under perf_event_mmap, vm_file won't be NULL or bpf_d_path can handle it?
> >>>
> >>> Thanks Martin, this is a very good point. :) Yes, vm_file can be NULL
> >>> in perf_event_mmap.
> >>> I wonder what would happen (and what we could do about it? :|).
> >>> bpf_d_path is called on &vma->vm_file->f_path So without NULL checks
> >>> (of vm_file) in BPF, the helper wouldn't be called with a NULL pointer
> >>> but rather with an address that is offsetof(struct file, f_path).
> >>>
> >>
> >> I tested this patch with the following BCC script:
> >>
> >>      bpf_text = '''
> >>      #include <linux/mm_types.h>
> >>
> >>      KFUNC_PROBE(perf_event_mmap, struct vm_area_struct *vma)
> >>      {
> >>          char path[256] = {};
> >>
> >>          bpf_d_path(&vma->vm_file->f_path, path, sizeof(path));
> >>          bpf_trace_printk("perf_event_mmap %s", path);
> >>          return 0;
> >>      }
> >>      '''
> >>
> >>      b = BPF(text=bpf_text)
> >>      print("BPF program loaded")
> >>      b.trace_print()
> >>
> >> This change causes kernel panic. I think it's because of this NULL pointer.
> >
> > Thank you for the testing and repro Hengqi :)
> > Indeed, I was able to reproduce this panic. When vma->vm_file is NULL,
> > &vma->vm_file->f_path ends up being 0x18 so d_path causes a panic.
> > I suppose that this sort of issue must be relatively common in helpers
> > that take a PTR_TO_BTF_ID though ? I wonder if there is anything that
>
> Most non-tracing ARG_PTR_TO_BTF_ID argument has strict helper/prog_type
> protection and should be okay although I didn't check them 100%.
>
> For some tracing helpers with ARG_PTR_TO_BTF_ID argument, we have
> bpf_seq_printf/bpf_seq_write which has strict context as well and should
> not be NULL.
>
> For helper bpf_task_pt_regs() which can attach to ANY kernel function,
> we kind of assume "task" is not NULL which should be the case in "almost
> all* cases from kernel internal data structure.
>
> > the verifier could do about this ? For example if vma->vm_file could
> > be PTR_TO_BTF_ID_OR_NULL and therefore vma->vm_file->f_path somehow
> > considered invalid ?
>
> Verifier has no way to know whether vma->vm_file is NULL or not during
> verification time. So in your case, if we have to be conservative, that
> means verifier will reject the program.
>
> One possible way could be add a mode in verifier, we still *go through*
> the process for direct memory access but we require user explicit
> checking NULL pointers. This way, user will be forced to write code like
>
>     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.