On Thu, Apr 02, 2020 at 03:21:06PM +0100, Al Viro wrote:
> On Wed, Apr 01, 2020 at 01:09:04PM +0200, Jiri Olsa wrote:
> > hi,
> > adding d_path helper to return full path for 'path' object.
> >
> > I originally added and used 'file_path' helper, which did the same,
> > but used 'struct file' object. Then realized that file_path is just
> > a wrapper for d_path, so we'd cover more calling sites if we add
> > d_path helper and allowed resolving BTF object within another object,
> > so we could call d_path also with file pointer, like:
> >
> > bpf_d_path(&file->f_path, buf, size);
> >
> > This feature is mainly to be able to add dpath (filepath originally)
> > function to bpftrace, which seems to work nicely now, like:
> >
> > # bpftrace -e 'kretfunc:fget { printf("%s\n", dpath(args->ret->f_path)); }'
> >
> > I'm not completely sure this is all safe and bullet proof and there's
> > no other way to do this, hence RFC post.
> >
> > I'd be happy also with file_path function, but I thought it'd be
> > a shame not to try to add d_path with the verifier change.
> > I'm open to any suggestions ;-)
>
> What are the locking conditions guaranteed to that sucker? Note that d_path()
> is *NOT* lockless - call it from an interrupt/NMI/etc. and you are fucked.
> It can grab rename_lock and mount_lock; usually it avoids that, so you won't
> see them grabbed on every call, but after the first seqlock mismatch it will
> fall back to grabbing the spinlock in question. And then there's ->d_dname(),
> with whatever things _that_ chooses to do....
if we limit it just to task context I think it would still be
helpful for us:
if (in_task())
d_path..
perhaps even create a d_path version without d_dname callback
if that'd be still a problem, because it seems to be there mainly
for special filesystems..?
thanks,
jirka
