Re: [BUG] Failed to obtain stack trace via bpf_get_stackid on ARM64 architecture

On Fri, 19 Sep 2025 19:56:20 -0700
Alexei Starovoitov [off-list ref] wrote:

On Fri, Sep 19, 2025 at 12:19 AM Feng Yang [off-list ref] wrote:

quoted

When I use bpf_program__attach_kprobe_multi_opts to hook a BPF program that contains the bpf_get_stackid function on the arm64 architecture,
I find that the stack trace cannot be obtained. The trace->nr in __bpf_get_stackid is 0, and the function returns -EFAULT.

For example:

diff --git a/tools/testing/selftests/bpf/progs/kprobe_multi.c b/tools/testing/selftests/bpf/progs/kprobe_multi.c
index 9e1ca8e34913..844fa88cdc4c 100644
--- a/tools/testing/selftests/bpf/progs/kprobe_multi.c
+++ b/tools/testing/selftests/bpf/progs/kprobe_multi.c

@@ -36,6 +36,15 @@ __u64 kretprobe_test6_result = 0;
 __u64 kretprobe_test7_result = 0;
 __u64 kretprobe_test8_result = 0;

+typedef __u64 stack_trace_t[2];
+
+struct {
+       __uint(type, BPF_MAP_TYPE_STACK_TRACE);
+       __uint(max_entries, 1024);
+       __type(key, __u32);
+       __type(value, stack_trace_t);
+} stacks SEC(".maps");
+
 static void kprobe_multi_check(void *ctx, bool is_return)
 {
        if (bpf_get_current_pid_tgid() >> 32 != pid)

@@ -100,7 +109,9 @@ int test_kretprobe(struct pt_regs *ctx)
 SEC("kprobe.multi")
 int test_kprobe_manual(struct pt_regs *ctx)
 {
+       int id = bpf_get_stackid(ctx, &stacks, 0);

ftrace_partial_regs() supposed to work on x86 and arm64,
but since multi-kprobe is the only user...

It should be able to unwind stack. It saves sp, pc, lr, fp.

	regs->sp = afregs->sp;
	regs->pc = afregs->pc;
	regs->regs[29] = afregs->fp;
	regs->regs[30] = afregs->lr;

I suspect the arm64 implementation wasn't really tested.
Or maybe there is some other issue.

It depends on how bpf_get_stackid() works. Some registers for that
function may not be saved.

If it returns -EFAULT, the get_perf_callchain() returns NULL.

struct perf_callchain_entry *
get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
		   u32 max_stack, bool crosstask, bool add_mark)
{
...
	entry = get_callchain_entry(&rctx);
	if (!entry)
		return NULL;


Thus the `get_callchain_entry(&rctx)` returns NULL. But if so,
this does not related to the ftrace_partial_regs(), because
get_callchain_entry() returns the per-cpu callchain woarking
buffer for the context, not decoding stack.

struct perf_callchain_entry *get_callchain_entry(int *rctx)
{
	int cpu;
	struct callchain_cpus_entries *entries;

	*rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
	if (*rctx == -1)
		return NULL;

	entries = rcu_dereference(callchain_cpus_entries);
	if (!entries) {
		put_recursion_context(this_cpu_ptr(callchain_recursion), *rctx);
		return NULL;
	}

	cpu = smp_processor_id();

	return (((void *)entries->cpu_entries[cpu]) +
		(*rctx * perf_callchain_entry__sizeof()));
}

What context does BPF expect, and how does it detect?

Thank you,

-- 
Masami Hiramatsu (Google) [off-list ref]