Hi Martin, Thank you for your reply! On 06/07/2024 01:10, Martin KaFai Lau wrote: > On 7/4/24 3:48 AM, Matthieu Baerts wrote: >>> diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/ >>> selftests/bpf/Makefile >>> index e0b3887b3d2d..204269d0b5b8 100644 >>> --- a/tools/testing/selftests/bpf/Makefile >>> +++ b/tools/testing/selftests/bpf/Makefile >>> @@ -144,7 +144,7 @@ TEST_GEN_PROGS_EXTENDED = test_skb_cgroup_id_user \ >>> flow_dissector_load test_flow_dissector >>> test_tcp_check_syncookie_user \ >>> test_lirc_mode2_user xdping test_cpp runqslower bench >>> bpf_testmod.ko \ >>> xskxceiver xdp_redirect_multi xdp_synproxy veristat >>> xdp_hw_metadata \ >>> - xdp_features bpf_test_no_cfi.ko >>> + xdp_features bpf_test_no_cfi.ko mptcp_pm_nl_ctl >> On the BPF CI, we have such errors: >> >> mptcp_pm_nl_ctl.c:20:10: fatal error: 'linux/mptcp.h' file not found >> 20 | #include "linux/mptcp.h" >> | ^~~~~~~~~~~~~~~ >> >> On my side, I don't have any issue, because the compiler uses the >> mptcp.h file from the system: /usr/include/linux/mptcp.h >> >> I suppose that's not OK on the BPF CI, as it looks like it doesn't have >> this file there, probably because it still uses Ubuntu 20.04 as base, >> which doesn't include this file in the linux-libc-dev package. >> >> When I look at how this 'mptcp_pm_nl_ctl' tool -- and all the other >> programs from that list -- is compiled (V=1), I see that the following >> "-I" options are given: >> >> -I${PWD}/tools/testing/selftests/bpf >> -I${BUILD}//tools/include >> -I${BUILD}/include/generated >> -I${PWD}/tools/lib >> -I${PWD}/tools/include >> -I${PWD}/tools/include/uapi >> -I${BUILD}/ >> >> It will then not look at -I${PWD}/usr/include or the directory generated >> with: >> >> make headers_install INSTALL_HDR_PATH=(...) > > It sounds like the tools/testing/selftests/net/mptcp/Makefile is looking > at this include path, so it works? Yes it does work. > iiu the bpf/Makefile correctly, it has the bpftool "make" compiled and > installed at tools/testing/selftests/bpf/tools/sbin/. May be directly > compile the pm_nl_ctl by "make tools/testing/selftests/net/mptcp/"? That could be an alternative, I didn't know it would be OK to add such dependence, good idea. >> I guess that's why people have duplicated files in 'tools/include/uapi', >> but I also understood from Jakub that it is not a good idea to continue >> to do so. >> >> What would be the best solution to avoid a copy? A symlink still looks >> like a workaround. >> >> In the other selftests, KHDR_INCLUDES is used to be able to include the >> path containing the UAPI headers. So if someone built the headers in a > > Meaning KHDR_INCLUDES should be used and -I${PWD}/tools/include/uapi can > be retired? That's the idea, yes, for "userspace programs". I mean: for BPF programs requiring vmlinux.h (BPF_CFLAGS), I guess you will still need the bpf.h file from tools/include/uapi, no? > I haven't looked into the details. I quickly tried but it > fails in my environment. Do you not have issues because some files have something like: #include <uapi/linux/(...).h> On my side, I had a working version using this patch: > diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile > index 7c5827d20c2e..112f14d40852 100644 > --- a/tools/testing/selftests/bpf/Makefile > +++ b/tools/testing/selftests/bpf/Makefile > @@ -37,7 +37,7 @@ CFLAGS += -g $(OPT_FLAGS) -rdynamic \ > -Wall -Werror -fno-omit-frame-pointer \ > $(GENFLAGS) $(SAN_CFLAGS) $(LIBELF_CFLAGS) \ > -I$(CURDIR) -I$(INCLUDE_DIR) -I$(GENDIR) -I$(LIBDIR) \ > - -I$(TOOLSINCDIR) -I$(APIDIR) -I$(OUTPUT) > + -I$(TOOLSINCDIR) $(KHDR_INCLUDES) -I$(OUTPUT) > LDFLAGS += $(SAN_LDFLAGS) > LDLIBS += $(LIBELF_LIBS) -lz -lrt -lpthread > But only after having removed these extra 'uapi/': $ git grep -l '<uapi/' -- tools/testing/selftests/bpf | \ xargs sed -i 's|#include <uapi/|#include <|g' Is it not OK for you like that? Note that I built the selftests using KHDR_INCLUDES=-I$INSTALL_HDR_PATH. >> seperated directory -- INSTALL_HDR_PATH=(...) -- KHDR_INCLUDES can be >> overridden to look there, instead of ${KERNEL_SRC}/usr/include. Would it >> be OK to do that? Would it work for the CI without extra changes? Or do >> you still prefer a copy/symlink to 'tools/include/uapi' instead? Cheers, Matt -- Sponsored by the NGI0 Core fund.