From: Douglas Raillard <douglas.raillard@xxxxxxx>
Infer alignment of struct members when loading BTF, so that using the
header dumped by pahole results in the same offsets. This is useful for
debugging tools to be able to access pointers to structs of possibly
private types. One such example are kernel modules bolting traceevents
on existing upstream tracepoints, which typically expose "top-level"
struct pointers (e.g. struct cfs_rq) to allow maximum flexibility.
Changes since v1:
* Split alignment patch into refactor and new code.
* Test with fullcircle on vmlinux and .o of kernel tree:
export PATH=pahole/build:$PATH
for f in {kernel,drivers,arch,block,crypto,certs,fs,init,ipc,lib,mm,net,sound,virt}/**.o; do
echo $f
pahole/fullcircle $f
done
This did not produce any error but neither did it on pahole master
branch, since I assume it reads DWARF by default.
Trying with "pfunct -F btf" on object files seemed to yield empty
sources:
pfunct: file 'kernel/sched/fair.o' has no btf type information.
Running pfunct on vmlinux worked, but fullcircle seems to have bailed
out before getting to pfunct. When commenting out the bailouts, I
managed to get pfunct to work, but the header generated with --compile
contains one duplicate type, and lots of broken inlined functions
(e.g. "return 0;" in a func returning a struct).
Douglas Raillard (3):
fprintf: Fix nested struct printing
btf_loader.c: Refactor class__fixup_btf_bitfields
btf_loader.c: Infer alignment info
btf_loader.c | 65 ++++++++++++++++++++++++++++++++++++-----------
dwarves.c | 2 +-
dwarves.h | 2 ++
dwarves_fprintf.c | 22 +++++++++++++---
4 files changed, 71 insertions(+), 20 deletions(-)
--
2.25.1
