Re: [PATCH bpf-next v3 4/4] bpf_trace: pass array of u64 values in kprobe_multi.addrs

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Wed, May 18, 2022 at 5:30 AM Eugene Syromiatnikov <esyr@redhat.com> wrote:
>
> On Wed, May 18, 2022 at 01:24:56PM +0200, Jiri Olsa wrote:
> > On Tue, May 17, 2022 at 02:34:55PM -0700, Yonghong Song wrote:
> > > On 5/17/22 1:03 PM, Jiri Olsa wrote:
> > > > On Tue, May 17, 2022 at 02:30:50PM +0200, Eugene Syromiatnikov wrote:
> > > > > On Tue, May 17, 2022 at 11:12:34AM +0200, Jiri Olsa wrote:
> > > > > > On Tue, May 17, 2022 at 09:36:47AM +0200, Eugene Syromiatnikov wrote:
> > > > > > > With the interface as defined, it is impossible to pass 64-bit kernel
> > > > > > > addresses from a 32-bit userspace process in BPF_LINK_TYPE_KPROBE_MULTI,
> > > > > > > which severly limits the useability of the interface, change the ABI
> > > > > > > to accept an array of u64 values instead of (kernel? user?) longs.
> > > > > > > Interestingly, the rest of the libbpf infrastructure uses 64-bit values
> > > > > > > for kallsyms addresses already, so this patch also eliminates
> > > > > > > the sym_addr cast in tools/lib/bpf/libbpf.c:resolve_kprobe_multi_cb().
> > > > > >
> > > > > > so the problem is when we have 32bit user sace on 64bit kernel right?
> > > > > >
> > > > > > I think we should keep addrs as longs in uapi and have kernel to figure out
> > > > > > if it needs to read u32 or u64, like you did for symbols in previous patch
> > > > >
> > > > > No, it's not possible here, as addrs are kernel addrs and not user space
> > > > > addrs, so user space has to explicitly pass 64-bit addresses on 64-bit
> > > > > kernels (or have a notion whether it is running on a 64-bit
> > > > > or 32-bit kernel, and form the passed array accordingly, which is against
> > > > > the idea of compat layer that tries to abstract it out).
> > > >
> > > > hum :-\ I'll need to check on compat layer.. there must
> > > > be some other code doing this already somewhere, right?
> >
> > so the 32bit application running on 64bit kernel using libbpf won't
> > work at the moment, right? because it sees:
> >
> >   bpf_kprobe_multi_opts::addrs as its 'unsigned long'
> >
> > which is 4 bytes and it needs to put there 64bits kernel addresses
> >
> > if we force the libbpf interface to use u64, then we should be fine
>
> Yes, that's correct.
>
> > > I am not familiar with all these compatibility thing. But if we
> > > have 64-bit pointer for **syms, maybe we could also have
> > > 64-bit pointer for *syms for consistency?
> >
> > right, perhaps we could have one function to read both syms and addrs arrays
>
> The distinction here it that syms are user space pointers (so they are
> naturally 32-bit for 32-bit applications) and addrs are kernel-space
> pointers (so they may be 64-bit even when the application is 32-bit).
> Nothing prevents from changing the interface so that syms is an array
> of 64-bit values treated as user space pointers, of course.
>
> > > > > > we'll need to fix also bpf_kprobe_multi_cookie_swap because it assumes
> > > > > > 64bit user space pointers
> >
> > if we have both addresses and cookies 64 then this should be ok
> >
> > > > > >
> > > > > > would be gret if we could have selftest for this
> >
> > let's add selftest for this
>
> Sure, I'll try to write one.
>

Not sure how you can do that without having extra test_progs variant
that's running in compat mode?