Hi Alexei,
Thank you for the review.
On 17/08/2024 09:22, Alexei Starovoitov wrote:
> On Fri, Aug 16, 2024 at 7:56 PM Matthieu Baerts (NGI0)
> <matttbe@xxxxxxxxxx> wrote:
>>
>> Instead of duplicating UAPI header files in 'tools/include/uapi', the
>> BPF selftests can also look at the header files inside the kernel
>> source.
>>
>> To do that, the kernel selftests infrastructure provides the
>> 'KHDR_INCLUDES' variable. This is what is being used in most selftests,
>> because it is what is recommended in the documentation [1]. If the
>> selftests are not executed from the kernel sources, it is possible to
>> override the variable, e.g.
>>
>> make KHDR_INCLUDES="-I${HDR_DIR}/include" -C "${KSFT_DIR}"
>>
>> ... where ${HDR_DIR} has been generated by this command:
>>
>> make headers_install INSTALL_HDR_PATH="${HDR_DIR}"
>>
>> Thanks to 'KHDR_INCLUDES', it is no longer needed to duplicate header
>> files for userspace test programs, and these programs can include UAPI
>> header files without the 'uapi' prefix.
>>
>> Note that it is still required to use 'tools/include/uapi' -- APIDIR,
>> which corresponds to TOOLS_INCLUDES from lib.mk -- for the BPF programs,
>> not to conflict with what is already defined in vmlinux.h.
>>
>> Link: https://docs.kernel.org/dev-tools/kselftest.html#contributing-new-tests-details [1]
>> Signed-off-by: Matthieu Baerts (NGI0) <matttbe@xxxxxxxxxx>
>> ---
>> tools/testing/selftests/bpf/Makefile | 2 +-
>> tools/testing/selftests/bpf/prog_tests/assign_reuse.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/tc_links.c | 4 ++--
>> tools/testing/selftests/bpf/prog_tests/tc_netkit.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/tc_opts.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/user_ringbuf.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/xdp_bonding.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/xdp_cpumap_attach.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/xdp_devmap_attach.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/xdp_do_redirect.c | 2 +-
>> tools/testing/selftests/bpf/prog_tests/xdp_link.c | 2 +-
>> tools/testing/selftests/bpf/xdp_features.c | 4 ++--
>> 12 files changed, 14 insertions(+), 14 deletions(-)
>>
>> diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
>> index 4eceb491a8ae..6a7aeae7e206 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
>>
>> diff --git a/tools/testing/selftests/bpf/prog_tests/assign_reuse.c b/tools/testing/selftests/bpf/prog_tests/assign_reuse.c
>> index 989ee4d9785b..3d06bf5a1ba4 100644
>> --- a/tools/testing/selftests/bpf/prog_tests/assign_reuse.c
>> +++ b/tools/testing/selftests/bpf/prog_tests/assign_reuse.c
>> @@ -1,6 +1,6 @@
>> // SPDX-License-Identifier: GPL-2.0
>> /* Copyright (c) 2023 Isovalent */
>> -#include <uapi/linux/if_link.h>
>> +#include <linux/if_link.h>
>
> No. This is not an option.
> User space shouldn't include kernel headers like this.
> Long ago tools/include directory was specifically
> created to break such dependency.
> Back then it was done for perf.
I'm sorry, but I think we are not talking about the same thing here:
here, I'm only modifying the "normal" userspace programs, not the ones
used to generate the BPF objects. Perf is a special case I suppose, it
needs to know the kernel internals. It is the same with BPF programs
requiring vmlinux.h. But I think "normal" userspace programs in the
sefltests can use the UAPI headers, no?
Here, the BPF userspace programs will take the kernel headers generated
by 'make headers', not directly from 'include/uapi' in the source.
That's what is done in other kernel selftests, and what is documented in
the kselftests doc [1]. With this modification, I'm able to compile the
BPF selftests.
I see that the BPF CI didn't manage to build them, but I think that's
because it sets 'KBUILD_OUTPUT', and build the BPF selftests directly,
not via the 'tools/testing/selftests/Makefile', which is the recommended
way, which supports 'KBUILD_OUTPUT'. Please see [2] for a fix for the
BPF CI.
A bit of context: my goal here is to be able to add new BPF selftests to
cover cases where BPF is used to modify MPTCP sockets and subflows. To
setup these tests, we need to communicate with the kernel, either by
using IPRoute2 -- but the BPF CI is still using IPRoute2 5.5 from
pre-COVID time -- or with a dedicated userspace program. This program
[2] has nothing special, it needs to include <linux/mptcp.h> to get the
value of some defines and enums to be able to communicate with the
kernel via Netlink. If I want to compile this program on my system,
outside the kselftest, the compiler will use the file from
'/usr/include/linux/mptcp.h', coming from the kernel UAPI. When
validating new features in the selftests, it will use the version from
the kernel source, the result of a 'make headers'. That works fine, we
don't have to duplicate the MPTCP UAPI header file when building the
MPTCP selftests. Same with the other selftests. Without this patch, I
would need to duplicate the 'mptcp.h' header if I want to use the same
program in the BPF selftests, because KHDR_INCLUDES is not used in the
CFLAGS when compiling the userspace programs like everywhere else in the
selftests.
I understand that I could indeed fix my initial problem by duplicating
mptcp.h in tools/include/uapi/linux/, but this doesn't look to be
allowed any more by the Netdev maintainers, e.g. recently, 'ethtool.h'
has been duplicated there in commit 7effe3fdc049 ("tools: Add ethtool.h
header to tooling infra"), but removed quickly after in commit
bbe91a9f6889 ("tools: remove redundant ethtool.h from tooling infra").
In this case, it was fine to simply drop it, because the linked test
doesn't require a recent version. Jakub mentioned [4] that these
duplicated headers should be avoided, and the ones generated by 'make
headers' should be used instead: what is being suggested here.
Do you think it would not work? If not, any idea what I can do?
Duplicating the defines and enums I need in the .c file doesn't seem
like a nice solution either.
[1] https://docs.kernel.org/dev-tools/kselftest.html
[2] https://github.com/libbpf/ci/pull/131
[3]
https://elixir.bootlin.com/linux/v6.10.5/source/tools/testing/selftests/net/mptcp/pm_nl_ctl.c
[4] https://lore.kernel.org/netdev/20240429132207.58ecf430@xxxxxxxxxx/T/#u
Cheers,
Matt
--
Sponsored by the NGI0 Core fund.
