Re: [PATCH bpf-next v7 4/7] bpftool: Implement "gen min_core_btf" logic

[Quentin Monnet <quentin@isovalent.com>]

2022-02-15 17:58 UTC-0500 ~ Mauricio Vásquez <mauricio@kinvolk.io>
> This commit implements the logic for the gen min_core_btf command.
> Specifically, it implements the following functions:
> 
> - minimize_btf(): receives the path of a source and destination BTF
> files and a list of BPF objects. This function records the relocations
> for all objects and then generates the BTF file by calling
> btfgen_get_btf() (implemented in the following commit).
> 
> - btfgen_record_obj(): loads the BTF and BTF.ext sections of the BPF
> objects and loops through all CO-RE relocations. It uses
> bpf_core_calc_relo_insn() from libbpf and passes the target spec to
> btfgen_record_reloc(), that calls one of the following functions
> depending on the relocation kind.
> 
> - btfgen_record_field_relo(): uses the target specification to mark all
> the types that are involved in a field-based CO-RE relocation. In this
> case types resolved and marked recursively using btfgen_mark_type().
> Only the struct and union members (and their types) involved in the
> relocation are marked to optimize the size of the generated BTF file.
> 
> - btfgen_record_type_relo(): marks the types involved in a type-based
> CO-RE relocation. In this case no members for the struct and union types
> are marked as libbpf doesn't use them while performing this kind of
> relocation. Pointed types are marked as they are used by libbpf in this
> case.
> 
> - btfgen_record_enumval_relo(): marks the whole enum type for enum-based
> relocations.
> 
> Signed-off-by: Mauricio Vásquez <mauricio@kinvolk.io>
> Signed-off-by: Rafael David Tinoco <rafael.tinoco@aquasec.com>
> Signed-off-by: Lorenzo Fontana <lorenzo.fontana@elastic.co>
> Signed-off-by: Leonardo Di Donato <leonardo.didonato@elastic.co>
> ---
>  tools/bpf/bpftool/Makefile |   8 +-
>  tools/bpf/bpftool/gen.c    | 455 ++++++++++++++++++++++++++++++++++++-
>  2 files changed, 457 insertions(+), 6 deletions(-)
> 
> diff --git a/tools/bpf/bpftool/Makefile b/tools/bpf/bpftool/Makefile
> index 94b2c2f4ad43..a137db96bd56 100644
> --- a/tools/bpf/bpftool/Makefile
> +++ b/tools/bpf/bpftool/Makefile
> @@ -34,10 +34,10 @@ LIBBPF_BOOTSTRAP_INCLUDE := $(LIBBPF_BOOTSTRAP_DESTDIR)/include
>  LIBBPF_BOOTSTRAP_HDRS_DIR := $(LIBBPF_BOOTSTRAP_INCLUDE)/bpf
>  LIBBPF_BOOTSTRAP := $(LIBBPF_BOOTSTRAP_OUTPUT)libbpf.a
>  
> -# We need to copy hashmap.h and nlattr.h which is not otherwise exported by
> -# libbpf, but still required by bpftool.
> -LIBBPF_INTERNAL_HDRS := $(addprefix $(LIBBPF_HDRS_DIR)/,hashmap.h nlattr.h)
> -LIBBPF_BOOTSTRAP_INTERNAL_HDRS := $(addprefix $(LIBBPF_BOOTSTRAP_HDRS_DIR)/,hashmap.h)
> +# We need to copy hashmap.h, nlattr.h, relo_core.h and libbpf_internal.h
> +# which are not otherwise exported by libbpf, but still required by bpftool.
> +LIBBPF_INTERNAL_HDRS := $(addprefix $(LIBBPF_HDRS_DIR)/,hashmap.h nlattr.h relo_core.h libbpf_internal.h)
> +LIBBPF_BOOTSTRAP_INTERNAL_HDRS := $(addprefix $(LIBBPF_BOOTSTRAP_HDRS_DIR)/,hashmap.h relo_core.h libbpf_internal.h)
>  
>  $(LIBBPF_OUTPUT) $(BOOTSTRAP_OUTPUT) $(LIBBPF_BOOTSTRAP_OUTPUT) $(LIBBPF_HDRS_DIR) $(LIBBPF_BOOTSTRAP_HDRS_DIR):
>  	$(QUIET_MKDIR)mkdir -p $@
> diff --git a/tools/bpf/bpftool/gen.c b/tools/bpf/bpftool/gen.c
> index 8e066c747691..806001020841 100644
> --- a/tools/bpf/bpftool/gen.c
> +++ b/tools/bpf/bpftool/gen.c
> @@ -14,6 +14,7 @@
>  #include <unistd.h>
>  #include <bpf/bpf.h>
>  #include <bpf/libbpf.h>
> +#include <bpf/libbpf_internal.h>
>  #include <sys/types.h>
>  #include <sys/stat.h>
>  #include <sys/mman.h>

Mauricio, did you try this patch on a system with an old Glibc (< 2.26)
by any chance? Haven't tried yet but I expect this might break bpftool's
build when COMPAT_NEED_REALLOCARRAY is set, because in that case gen.c
pulls <bpf/libbpf_internal.h>, and then <tools/libc_compat.h> (through
main.h). And libc_compat.h defines reallocarray(), which
libbpf_internal.h poisons with a GCC pragma.

At least this is what I observe when trying to add your patches to the
kernel mirror, where reallocarray() is redefined unconditionally. I'm
trying to figure out if we should fix this mirror-side, or kernel-side.
(I suppose we still need this compatibility layer, Ubuntu 16.04 seems to
use Glibc 2.23).

Quentin