Re: [PATCH bpf-next] selftests/bpf: Check trampoline execution in d_path test

[Jiri Olsa <jolsa@redhat.com>]

On Mon, Sep 14, 2020 at 07:30:33PM -0700, Alexei Starovoitov wrote:
> On Fri, Sep 11, 2020 at 6:16 AM Jiri Olsa <jolsa@redhat.com> wrote:
> >
> > On Thu, Sep 10, 2020 at 05:46:21PM -0700, Alexei Starovoitov wrote:
> > > On Thu, Sep 10, 2020 at 5:22 AM Jiri Olsa <jolsa@kernel.org> wrote:
> > > >
> > > > Some kernels builds might inline vfs_getattr call within
> > > > fstat syscall code path, so fentry/vfs_getattr trampoline
> > > > is not called.
> > > >
> > > > I'm not sure how to handle this in some generic way other
> > > > than use some other function, but that might get inlined at
> > > > some point as well.
> > >
> > > It's great that we had the test and it failed.
> > > Doing the test skipping will only hide the problem.
> > > Please don't do it here and in the future.
> > > Instead let's figure out the real solution.
> > > Assuming that vfs_getattr was added to btf_allowlist_d_path
> > > for a reason we have to make this introspection place
> > > reliable regardless of compiler inlining decisions.
> > > We can mark it as 'noinline', but that's undesirable.
> > > I suggest we remove it from the allowlist and replace it with
> > > security_inode_getattr.
> > > I think that is a better long term fix.
> >
> > in my case vfs_getattr got inlined in vfs_statx_fd and both
> > of them are defined in fs/stat.c
> >
> > so the idea is that inlining will not happen if the function
> > is defined in another object? or less likely..?
> 
> when it's in a different .o file. yes.
> Very few folks build LTO kernels, so I propose to cross that bridge when
> we get there.
> Eventually we can replace security_inode_getattr
> with bpf_lsm_inode_getattr or simply add noinline to security_inode_getattr.
> 
> > we should be safe when it's called from module
> 
> what do you mean?

it's external call, so it will not get inlined

> 
> > > While at it I would apply the same critical thinking to other
> > > functions in the allowlist. They might suffer the same issue.
> > > So s/vfs_truncate/security_path_truncate/ and so on?
> > > Things won't work when CONFIG_SECURITY is off, but that is a rare kernel config?
> > > Or add both security_* and vfs_* variants and switch tests to use security_* ?
> > > but it feels fragile to allow inline-able funcs in allowlist.
> >
> > hm, what's the difference between vfs_getattr and security_inode_getattr
> > in this regard? I'd expect compiler could inline it same way as for vfs_getattr
> 
> not really because they're in different files and LTO is not on.
> Even with LTO the chances of inlining are small. The compiler will
> consider profitability of it. Since there is a loop inside, it's unlikely.

ok, thanks for info

I'll use that security_inode_getattr instead of vfs_getattr

jirka