Re: BTFIDS: FAILED unresolved symbol udp6_sock

From: Andrii Nakryiko
Date: Mon Jan 04 2021 - 18:58:53 EST


On Sat, Jan 2, 2021 at 3:07 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
>
> On Sat, Jan 02, 2021 at 02:25:34PM -0800, Andrii Nakryiko wrote:
>
> SNIP
>
> > >
> > > so your .config has
> > > CONFIG_CRYPTO_DEV_BCM_SPU=y
> > >
> > > and that defines 'struct device_private' which
> > > clashes with the same struct defined in drivers/base/base.h
> > >
> > > so several networking structs will be doubled, like net_device:
> > >
> > > $ bpftool btf dump file ../vmlinux.config | grep net_device\' | grep STRUCT
> > > [2731] STRUCT 'net_device' size=2240 vlen=133
> > > [113981] STRUCT 'net_device' size=2240 vlen=133
> > >
> > > each is using different 'struct device_private' when it's unwinded
> > >
> > > and that will confuse BTFIDS logic, becase we have multiple structs
> > > with the same name, and we can't be sure which one to pick
> > >
> > > perhaps we should check on this in pahole and warn earlier with
> > > better error message.. I'll check, but I'm not sure if pahole can
> > > survive another hastab ;-)
> > >
> > > Andrii, any ideas on this? ;-)
> >
> > It's both unavoidable and correct from the C type system's
> > perspective, so there is nothing for pahole to warn about. We used to
> > have (for a long time) a similar clash with two completely different
> > ring_buffer structs. Eventually they just got renamed to avoid
> > duplication of related structs (task_struct and tons of other). But
> > both BTF dedup and CO-RE relocation algorithms are designed to handle
> > this correctly, ...
>
> AFAIU it's all correctly dedulicated, but still all structs that
> contain (at some point) 'struct device_private' will appear twice
> in BTF data.. each with different 'struct device_private'

it's correct from the type system perspective, right. Those two
duplicates of struct device_private are parts of two different
hierarchies of types. However inconvenient it is, C allows it,
unfortunately :(

>
> > ... so perhaps BTFIDS should be able to handle this as
> > well?
>
> hm, BTFIDS sees BTF data with two same struct names and has no
> way to tell which one to use
>
> unless we have some annotation data for BTF types I don't
> see a way to handle this correctly.. but I think we can
> detect this directly in BTFIDS and print more accurate error
> message
>
> as long as we dont see this on daily basis, I think that better
> error message + following struct rename is good solution

Perhaps warning and handling this gracefully is a bit better way to
handle this. Renaming is definitely good, but shouldn't block the
kernel build process. I don't remember the exact details for why
duplicate struct would cause troubles for resolve_btfids, but maybe
just picking the struct with minimal ID (out of 2+ duplicates) would
be ok in practice most of the time. In any case, that's what most
users (including libbpf) will do, when searching for the type by name.

>
> >
> > >
> > > easy fix is the patch below that renames the bcm's structs,
> > > it makes the kernel to compile.. but of course the new name
> > > is probably wrong and we should push this through that code
> > > authors
> >
> > In this case, I think renaming generic device_private name is a good
> > thing regardless.
>
> ok, I'll send the change
>

great, thanks

> jirka
>