Re: [PATCH bpf-next v3] bpf: make the attach target more accurate

[Jiri Olsa <olsajiri@gmail.com>]

On Fri, Jul 11, 2025 at 01:51:31PM +0800, Menglong Dong wrote:
> On Fri, Jul 11, 2025 at 11:46 AM Yonghong Song <yonghong.song@linux.dev> wrote:
> >
> >
> >
> > On 7/10/25 8:10 PM, Yonghong Song wrote:
> > >
> > >
> > > On 7/10/25 12:08 AM, Menglong Dong wrote:
> > >> For now, we lookup the address of the attach target in
> > >> bpf_check_attach_target() with find_kallsyms_symbol_value or
> > >> kallsyms_lookup_name, which is not accurate in some cases.
> > >>
> > >> For example, we want to attach to the target "t_next", but there are
> > >> multiple symbols with the name "t_next" exist in the kallsyms, which
> > >> makes
> > >> the attach target ambiguous, and the attach should fail.
> > >>
> > >> Introduce the function bpf_lookup_attach_addr() to do the address
> > >> lookup,
> > >> which will return -EADDRNOTAVAIL when the symbol is not unique.
> > >>
> > >> We can do the testing with following shell:
> > >>
> > >> for s in $(cat /proc/kallsyms | awk '{print $3}' | sort | uniq -d)
> > >> do
> > >>    if grep -q "^$s\$"
> > >> /sys/kernel/debug/tracing/available_filter_functions
> > >>    then
> > >>      bpftrace -e "fentry:$s {printf(\"1\");}" -v
> > >>    fi
> > >> done
> > >>
> > >> The script will find all the duplicated symbols in /proc/kallsyms, which
> > >> is also in /sys/kernel/debug/tracing/available_filter_functions, and
> > >> attach them with bpftrace.
> > >>
> > >> After this patch, all the attaching fail with the error:
> > >>
> > >> The address of function xxx cannot be found
> > >> or
> > >> No BTF found for xxx
> > >>
> > >> Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
> > >
> > > Maybe we should prevent vmlinux BTF generation for such symbols
> > > which are static and have more than one instances? This can
> > > be done in pahole and downstream libbpf/kernel do not
> > > need to do anything. This can avoid libbpf/kernel runtime overhead
> > > since bpf_lookup_attach_addr() could be expensive as it needs
> > > to go through ALL symbols, even for unique symbols.
> 
> Hi, yonghong. You are right, the best solution is to solve
> this problem in the pahole, just like what Jiri said in the V2:
>   https://lore.kernel.org/bpf/aG5hzvaqXi7uI4GL@krava/
> 
> I wonder will we focus the users to use the latest pahole
> that supports duplicate symbols filter after we fix this problem
> in pahole? If so, this patch is useless, and just ignore it. If
> not, the only usage of this patch is for the users that build
> the kernel with an old pahole.
> 
> >
> > There is a multi-link effort:
> >    https://lore.kernel.org/bpf/20250703121521.1874196-1-dongml2@chinatelecom.cn/
> > which tries to do similar thing for multi-kprobe. For example, for fentry,
> > multi-link may pass an array of btf_id's to the kernel. For such cases,
> > this patch may cause significant performance overhead.
> 
> For the symbol in the vmlinux, there will be no additional overhead,
> as the logic is the same as previous. If the symbol is in the
> modules, it does have additional overhead. Following is the
> testing that hooks all the symbols with fentry-multi.
> 
> Without this patch, the time to attach all the symbols:
> kernel: 0.372660s for 48857 symbols
> modules: 0.135543s for 8631 symbols
> 
> And with this patch, the time is:
> kernel: 0.380087s for 48857 symbols
> modules: 0.176904s for 8631 symbols
> 
> One more thing, is there anyone to fix the problem in pahole?
> I mean, I'm not good at pahole. But if there is nobody, I still can
> do this job, but I need to learn it first :/

I'm testing change below, I'll send the patch after some more testing

jirka


---
diff --git a/btf_encoder.c b/btf_encoder.c
index 16739066caae..29ff86bac7de 100644
--- a/btf_encoder.c
+++ b/btf_encoder.c
@@ -2143,6 +2143,31 @@ int btf_encoder__encode(struct btf_encoder *encoder, struct conf_load *conf)
 	return err;
 }
 
+static void remove_dups(struct elf_functions *functions)
+{
+	struct elf_function *n = &functions->entries[0];
+	bool matched = false;
+	int i, j;
+
+	for (i = 0, j = 1; i < functions->cnt && j < functions->cnt; i++, j++) {
+		struct elf_function *a = &functions->entries[i];
+		struct elf_function *b = &functions->entries[j];
+
+		if (!strcmp(a->name, b->name)) {
+			matched = true;
+			continue;
+		}
+
+		if (!matched)
+			*n++ = *a;
+		matched = false;
+	}
+
+	if (!matched)
+		*n++ = functions->entries[functions->cnt - 1];
+	functions->cnt = n - &functions->entries[0];
+}
+
 static int elf_functions__collect(struct elf_functions *functions)
 {
 	uint32_t nr_symbols = elf_symtab__nr_symbols(functions->symtab);
@@ -2168,6 +2193,7 @@ static int elf_functions__collect(struct elf_functions *functions)
 
 	if (functions->cnt) {
 		qsort(functions->entries, functions->cnt, sizeof(*functions->entries), functions_cmp);
+		remove_dups(functions);
 	} else {
 		err = 0;
 		goto out_free;