On Wed, Feb 3, 2021 at 7:05 PM Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote: > > On Wed, Feb 3, 2021 at 5:21 PM Chris Murphy <lists@xxxxxxxxxxxxxxxxx> wrote: > > > > On Wed, Feb 3, 2021 at 6:00 PM Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote: > > > > > > Em Wed, Feb 03, 2021 at 05:46:48PM -0700, Chris Murphy escreveu: > > > > This is just the vmlinuz-5.11.0-0.rc6.141.fc34.x86_64 file > > > > > > > > https://drive.google.com/file/d/1G_2qLVRIy-ExaJI1-cTqDssrDu3sWo-m/view?usp=sharing > > > > > > Can you please provide the vmlinux for this file as well? > > > > Used this > > $ /usr/src/kernels/5.11.0-0.rc6.141.fc34.x86_64/scripts/extract-vmlinux > > /boot/vmlinuz-5.11.0-0.rc6.141.fc34.x86_64 > vmlinux > > > > https://drive.google.com/file/d/1h6cC9oZ16oLbR6NyPqKkVGaoUQ2u1UQz/view?usp=sharing > > > > I recompiled with gcc 10.2.1 and I'm not having these problems, so it > > might be that. > > > > sched_reset_on_fork is a bitfield in task_struct. You don't see issue > on gcc 10.2, but see it on gcc 11. GCC 11 started emitting DWARF5 by > default, right? It's something that Arnaldo probably already fixed in > the latest pahole. Could you please try to build pahole from sources > and see if you run into the same problem again? > > Verifier is catching a real issue with offsets going backwards. Here's > excerpt from BTF dump of task_struct: > > 'pdeath_signal' type_id=17 bits_offset=18048 > 'jobctl' type_id=1 bits_offset=18112 > 'personality' type_id=6 bits_offset=18176 > 'sched_reset_on_fork' type_id=6 bits_offset=0 > 'sched_contributes_to_load' type_id=6 bits_offset=0 > 'sched_migrated' type_id=6 bits_offset=0 > 'sched_psi_wake_requeue' type_id=6 bits_offset=0 > > ... > > eventually we get to non-bitfield field > > 'atomic_flags' type_id=1 bits_offset=18304 > > So it's a bitfield offset breakage that should be fixed in pahole 1.20. Rawhide is currently still on dwarves-1.19-2.fc34.x86_64 This might also be related: https://bugzilla.redhat.com/show_bug.cgi?id=1922707#c9 -- Chris Murphy
