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

[Hengqi Chen]

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.