Re: Re: [PATCH] tools/resolve_btfids: fix OUTPUT leakage from kselftest causing corrupted build paths

[Jiangshan Yi]

Hi Ihor Solodrai,

Thanks for the review -- you are right on both counts, and v1 is dropped.

On the nested sub-make: the `override OUTPUT :=` only changes the
variable inside the outer resolve_btfids make; it does not touch
MAKEOVERRIDES, so the nested `$(MAKE) $(build)=resolve_btfids`
(resolve_btfids/Makefile:90) still inherits the original command-line
OUTPUT through MAKEFLAGS. The pattern rules that produce the corrupted
`$(OUTPUT)%.o` paths run in that nested sub-make, so they still see the
un-slashed value. Your parent.mk/child.mk repro captures this exactly.

On bpf/hid: selftests/bpf/Makefile:440 and selftests/hid/Makefile:159
build resolve_btfids directly with
`OUTPUT=$(HOST_BUILD_DIR)/resolve_btfids/` on the command line. An
origin check cannot tell those deliberate command-line assignments apart
from a leaked one, so v1 clobbered them -- exactly the breakage you
showed. The origin is simply the wrong place to make this decision.

v2 takes the approach you suggested: drop the generic kselftest OUTPUT
at the gen_mods_dir descent in selftests/lib.mk, so it never reaches the
kernel module build (and thus resolve_btfids) as a command-line
variable:

      gen_mods_dir: MAKEOVERRIDES := $(filter-out OUTPUT=%,$(MAKEOVERRIDES))
      gen_mods_dir:
      	$(Q)$(MAKE) -C $(TEST_GEN_MODS_DIR)

      clean_mods_dir: MAKEOVERRIDES := $(filter-out OUTPUT=%,$(MAKEOVERRIDES))
      clean_mods_dir:
      	$(Q)$(MAKE) -C $(TEST_GEN_MODS_DIR) clean

gen_mods_dir is the only path through which kselftest's per-test OUTPUT
leaks into a kernel build (the TEST_GEN_MODS_DIR module build, which can
the other selftest sub-makes -- and the test programs that rely on
OUTPUT -- are unaffected.

With the leaked command-line OUTPUT gone, resolve_btfids is reached with
only the O= that the kernel build passes on its tools/% descent, so the
`OUTPUT := $(ABSOLUTE_O)/...` assignment in scripts/Makefile.include (a
simple assignment) wins at every make level, including the nested
`$(build)=resolve_btfids` sub-make where the pattern rules run. bpf and
hid invoke resolve_btfids directly with their own OUTPUT= and do not go
through gen_mods_dir, so they are untouched.

Tested from a clean tree:
    * kselftest build of a TEST_GEN_MODS_DIR test (mm/page_frag) with
      CONFIG_DEBUG_INFO_BTF: resolve_btfids now builds into its
      O=-derived directory; no more "mmmain.o" / missing
      resolve_btfids-in.o.
    * make -C tools/bpf/resolve_btfids
      make -C tools/bpf/resolve_btfids OUTPUT=/tmp/o/resolve_btfids/
      both still build (bpf/hid-style explicit OUTPUT= preserved).
    * make OUTPUT=/tmp/build/kcmp -C tools/testing/selftests/kcmp still
      honours OUTPUT for its own programs.

v2 posted as a separate thread.

Thanks,
Jiangshan