Re: [PATCH bpf-next v4] selftests/bpf: use auto-dependencies for test objects

[Tony Ambardar <tony.ambardar@gmail.com>]

On Fri, Jul 19, 2024 at 11:18:16AM -0700, Andrii Nakryiko wrote:
> On Thu, Jul 18, 2024 at 3:57 PM Ihor Solodrai <ihor.solodrai@pm.me> wrote:
> >
> > Make use of -M compiler options when building .test.o objects to
> > generate .d files and avoid re-building all tests every time.
> >
> > Previously, if a single test bpf program under selftests/bpf/progs/*.c
> > has changed, make would rebuild all the *.bpf.o, *.skel.h and *.test.o
> > objects, which is a lot of unnecessary work.
> >
> > A typical dependency chain is:
> > progs/x.c -> x.bpf.o -> x.skel.h -> x.test.o -> trunner_binary
> >
> > However for many tests it's not a 1:1 mapping by name, and so far
> > %.test.o have been simply dependent on all %.skel.h files, and
> > %.skel.h files on all %.bpf.o objects.
> >
> > Avoid full rebuilds by instructing the compiler (via -MMD) to
> > produce *.d files with real dependencies, and appropriately including
> > them. Exploit make feature that rebuilds included makefiles if they
> > were changed by setting %.test.d as prerequisite for %.test.o files.
> >
> > A couple of examples of compilation time speedup (after the first
> > clean build):
> >
> > $ touch progs/verifier_and.c && time make -j8
> > Before: real    0m16.651s
> > After:  real    0m2.245s
> > $ touch progs/read_vsyscall.c && time make -j8
> > Before: real    0m15.743s
> > After:  real    0m1.575s
> >
> > A drawback of this change is that now there is an overhead due to make
> > processing lots of .d files, which potentially may slow down unrelated
> > targets. However a time to make all from scratch hasn't changed
> > significantly:
> >
> > $ make clean && time make -j8
> > Before: real    1m31.148s
> > After:  real    1m30.309s
> >
> > Suggested-by: Eduard Zingerman <eddyz87@gmail.com>
> > Signed-off-by: Ihor Solodrai <ihor.solodrai@pm.me>
> >
> > ---
> > v3 -> v4: Make $(TRUNNER_BPF_OBJS) order only prereq of trunner binary
> > v2 -> v3: Restore dependency on $(TRUNNER_BPF_OBJS)
> > v1 -> v2: Make %.test.d prerequisite order only
> > ---
> >  tools/testing/selftests/bpf/.gitignore |  1 +
> >  tools/testing/selftests/bpf/Makefile   | 44 +++++++++++++++++++-------
> >  2 files changed, 34 insertions(+), 11 deletions(-)
> >
> 
> It seems to behave correctly, but it reports wrong flavor when
> building .bpf.o, e.g.,:
> 
Hi Andrii,

This is actually an old, confusing bug unrelated to the current (very
nice) improvements. I have a fix as part of a larger series
targeting libc portability and MIPS support which I'll post shortly.  Or
I can send separately if you like?

Thanks,
Tony

> $ touch progs/test_vmlinux.c
> $ make -j90
>   CLNG-BPF [test_maps] test_vmlinux.bpf.o
>   CLNG-BPF [test_maps] test_vmlinux.bpf.o
>   CLNG-BPF [test_maps] test_vmlinux.bpf.o
>   GEN-SKEL [test_progs] test_vmlinux.skel.h
>   GEN-SKEL [test_progs-cpuv4] test_vmlinux.skel.h
>   GEN-SKEL [test_progs-no_alu32] test_vmlinux.skel.h
>   TEST-OBJ [test_progs] vmlinux.test.o
>   TEST-OBJ [test_progs-no_alu32] vmlinux.test.o
>   EXT-COPY [test_progs-no_alu32] urandom_read bpf_testmod.ko
> bpf_test_no_cfi.ko liburandom_read.so xdp_synproxy sign-file
> uprobe_multi ima_setup.sh verify_sig_setup.sh
> btf_dump_test_case_bitfields.c btf_dump_test_case_multidim.c
> btf_dump_test_case_namespacing.c btf_dump_test_case_ordering.c
> btf_dump_test_case_packing.c btf_dump_test_case_padding.c
> btf_dump_test_case_syntax.c
>   TEST-OBJ [test_progs-cpuv4] vmlinux.test.o
>   EXT-COPY [test_progs-cpuv4] urandom_read bpf_testmod.ko
> bpf_test_no_cfi.ko liburandom_read.so xdp_synproxy sign-file
> uprobe_multi ima_setup.sh verify_sig_setup.sh
> btf_dump_test_case_bitfields.c btf_dump_test_case_multidim.c
> btf_dump_test_case_namespacing.c btf_dump_test_case_ordering.c
> btf_dump_test_case_packing.c btf_dump_test_case_padding.c
> btf_dump_test_case_syntax.c
> make[1]: Nothing to be done for 'docs'.
>   BINARY   test_progs
>   BINARY   test_progs-no_alu32
>   BINARY   test_progs-cpuv4
> $ ls -la test_vmlinux.bpf.o no_alu32/test_vmlinux.bpf.o cpuv4/test_vmlinux.bpf.o
> -rw-r--r-- 1 andriin users 21344 Jul 19 11:08 cpuv4/test_vmlinux.bpf.o
> -rw-r--r-- 1 andriin users 21408 Jul 19 11:08 no_alu32/test_vmlinux.bpf.o
> -rw-r--r-- 1 andriin users 21408 Jul 19 11:08 test_vmlinux.bpf.o
> 
> 
> Note [test_maps] for all three variants (I expected
> test_maps/test_progs + no_alu32 + cpuv4, just like we see for skel.h).
> Can you please double check what's going on? Looking at timestamps it
> seems like they are actually regenerated, though.
> 
> 
> BTW, if you get a chance, see if you can avoid unnecessary EXT-COPY as
> well (probably a bit smarter rule dependency should be set up, e.g.,
> phony target that then depends on actual files or something like
> that).
> 
> Regardless, this is a massive improvement and seems to work correctly,
> so I've applied this and will wait for follow ups. Thanks a lot!
> 
> BTW, are you planning to look into vmlinux.h optimization as well?
> 
> > diff --git a/tools/testing/selftests/bpf/.gitignore b/tools/testing/selftests/bpf/.gitignore
> > index 5025401323af..4e4aae8aa7ec 100644
> > --- a/tools/testing/selftests/bpf/.gitignore
> > +++ b/tools/testing/selftests/bpf/.gitignore
> > @@ -31,6 +31,7 @@ test_tcp_check_syncookie_user
> >  test_sysctl
> >  xdping
> >  test_cpp
> > +*.d
> >  *.subskel.h
> >  *.skel.h
> >  *.lskel.h
> 
> [...]