Re: selftest/bpf/test_verifier_log fails on v5.11-rc5

[Jiri Olsa <jolsa@redhat.com>]

On Mon, Feb 01, 2021 at 09:25:32AM -0300, Arnaldo Carvalho de Melo wrote:
> Em Sun, Jan 31, 2021 at 10:36:41PM +0100, Jiri Olsa escreveu:
> > On Sat, Jan 30, 2021 at 09:48:13PM +0100, Jiri Olsa wrote:
> > 
> > SNIP
> > 
> > > > > > > % uname -r
> > > > > > > 5.11.0-0.rc5.134.fc34.x86_64
> > > > > > > % pwd
> > > > > > > /.../linux/tools/testing/selftests/bpf
> > > > > > > % git log --oneline | head -n 1
> > > > > > > 6ee1d745b7c9 Linux 5.11-rc5
> > > > > > > % make test_verifier_log
> > > > > > >   ...
> > > > > > >   BINARY   test_verifier_log
> > > > > > > % ./test_verifier_log
> > > > > > > Test log_level 0...
> > > > > > > Test log_size < 128...
> > > > > > > Test log_buff = NULL...
> > > > > > > Test oversized buffer...
> > > > > > > ERROR: Program load returned: ret:-1/errno:22, expected ret:-1/errno:13
> > > > > >
> > > > > > Thanks for reporting.
> > > > > > bpf and bpf-next don't have this issue. Not sure what changed.
> > > > >
> > > > > I haven't had a chance to look into this any further, but Ondrej
> > > > > Mosnacek (CC'd) found the following today:
> > > > >
> > > > > "So I was trying to debug this further and I think I've identified what
> > > > > triggers the problem. It seems that the BTF debuginfo generation
> > > > > became broken with CONFIG_DEBUG_INFO_DWARF4=n somewhere between -rc4
> > > > > and -rc5. It also seems to depend on a recent (Fedora Rawhide) version
> > > > > of some component of the build system (GCC, probably), because the
> > > > > problem disappeared when I tried to build the "bad" kernel in F33
> > > > > buildroot instead of Rawhide."
> > > > 
> > > > I see. There were fixes for dwarf and btf, but I lost the track.
> > > > I believe it was a combination of gcc bug that was worked around in pahole.
> > > > Arnaldo, Jiri, Andrii,
> > > > what is the status? Did all fixes land in pahole?
> > > 
> > > I checked on rawhide and besides many pahole warnings,
> > > the resulted BTF data have many duplications in core structs
> > > 
> > > 	  BTFIDS  vmlinux
> > > 	WARN: multiple IDs found for 'task_struct': 132, 1247 - using 132
> > > 	WARN: multiple IDs found for 'file': 440, 1349 - using 440
> > > 	WARN: multiple IDs found for 'inode': 698, 1645 - using 698
> > > 	WARN: multiple IDs found for 'path': 729, 1672 - using 729
> > > 	WARN: multiple IDs found for 'task_struct': 132, 2984 - using 132
> > > 	WARN: multiple IDs found for 'task_struct': 132, 3043 - using 132
> > > 	WARN: multiple IDs found for 'file': 440, 3085 - using 440
> > > 	WARN: multiple IDs found for 'seq_file': 1469, 3125 - using 1469
> > > 	WARN: multiple IDs found for 'inode': 698, 3336 - using 698
> > > 	WARN: multiple IDs found for 'path': 729, 3366 - using 729
> > > 	WARN: multiple IDs found for 'task_struct': 132, 5337 - using 132
> > > 	WARN: multiple IDs found for 'inode': 698, 5360 - using 698
> > > 	WARN: multiple IDs found for 'path': 729, 5388 - using 729
> > > 	WARN: multiple IDs found for 'file': 440, 5412 - using 440
> > > 	WARN: multiple IDs found for 'seq_file': 1469, 5639 - using 1469
> > > 	WARN: multiple IDs found for 'task_struct': 132, 6243 - using 132
> > > 	...
> > > 
> > > 	# gcc --version
> > > 	gcc (GCC) 11.0.0 20210123 (Red Hat 11.0.0-0)
> > > 
> > > I'm guessing there are some DWARF changes that screwed BTF
> > > generation.. I'll check
> > > 
> > > it's not covered by the fix I posted recently, but I think
> > > Arnaldo is now fixing some related stuff.. Arnaldo, maybe
> > > you are seeing same errors?
> > 
> > with Arnaldo's fixes I see less struct duplications,
> > but still there's some
> > 
> > > 
> > > I uploaded the build log from linking part to:
> > >   http://people.redhat.com/~jolsa/build.out.gz
> > 
> > however looks like we don't handle DW_FORM_implicit_const
> > when counting the byte offset.. it was used for some struct
> > members in my vmlinux, so we got zero for byte offset and
> > that created another unique struct
> > 
> > with patch below I no longer see any struct duplication,
> > also test_verifier_log is working for me, but I could
> > not reproduce the error before
> > 
> > I'll post full dwarves patch after some more testing
> > 
> > also I wonder we could somehow use btf_check_all_metas
> > from kernel after we build BTF data, that'd help to catch
> > this earlier/easier ;-) I'll check on this
> 
> Seems like a good idea indeed :-)
> 
> I'm applying the patch below with your Signed-off-by, etc, ok?

ok, thanks

jirka