btf_encoder: Represent "."-suffixed functions (".isra.0") in BTF
At gcc optimization level O2 or higher, many function optimizations
occur such as:
- constant propagation (.constprop.0);
- interprocedural scalar replacement of aggregates, removal of
unused parameters and replacement of parameters passed by
reference by parameters passed by value (.isra.0)
See [1] for details.
Currently BTF encoding does not handle such optimized functions that get
renamed with a "." suffix such as ".isra.0", ".constprop.0". This is
safer because such suffixes can often indicate parameters have been
optimized out. Since we can now spot this, support matching to a "."
suffix and represent the function in BTF if it does not have
optimized-out parameters. First an attempt to match by exact name is
made; if that fails we fall back to checking for a "."-suffixed name.
The BTF representation will use the original function name "foo" not
"foo.isra.0" for consistency with DWARF representation.
There is a complication however, and this arises because we process each
CU separately and merge BTF when complete. Different CUs may optimize
differently, so in one CU, a function may have optimized-out parameters
- and thus be ineligible for BTF - while in another it does not have
optimized-out parameters - making it eligible for BTF. The NF_HOOK
function is an example of this.
To avoid disrupting BTF generation parallelism, the approach taken is to
save pointers to the function representations in the ELF function table;
it is per-encoder, but the same representation is used across encoders,
so we can use the same array index to find the same function when
merging for efficiency. Using the ELF function is ideal also as we
already have to carry out a name lookup for matching from DWARF to ELF.
At thread collection time, observations about optimizations are merged
across encoders and we know whether it is safe to add a "."-suffixed
function or not.
The result of this is we add 1180 "."-suffixed functions to the BTF
representation.
Encoding of "."-suffixed functions is done only if the
"--btf_gen_optimized" function is specified.
Because we need to ensure consistency across encoders, there is
a performance cost to the save/merge/apply approach. Comparing
baseline to test times:
$ time LLVM_OBJCOPY=objcopy pahole -J -j4 vmlinux
real 0m7.788s
user 0m17.629s
sys 0m0.652s
$ time LLVM_OBJCOPY=objcopy pahole -J -j4 --btf_gen_optimized vmlinux
real 0m10.839s
user 0m19.473s
sys 0m5.932s
[1] https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html
Signed-off-by: Alan Maguire <alan.maguire@oracle.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Hao Luo <haoluo@google.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: KP Singh <kpsingh@chromium.org>
Cc: Kui-Feng Lee <sinquersw@gmail.com>
Cc: Martin KaFai Lau <martin.lau@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Cc: Stanislav Fomichev <sdf@google.com>
Cc: Timo Beckers <timo@incline.eu>
Cc: Yonghong Song <yhs@fb.com>
Cc: bpf@vger.kernel.org
Link: https://lore.kernel.org/r/1675790102-23037-6-git-send-email-alan.maguire@oracle.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
4 files changed
