> On May 17, 2019, at 2:06 PM, Alexei Starovoitov <ast@xxxxxx> wrote:
>
> On 5/17/19 11:40 AM, Song Liu wrote:
>> +Alexei, Daniel, and bpf
>>
>>> On May 17, 2019, at 2:10 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>>
>>> On Fri, May 17, 2019 at 04:15:39PM +0800, Kairui Song wrote:
>>>> Hi, I think the actual problem is that bpf_get_stackid_tp (and maybe
>>>> some other bfp functions) is now broken, or, strating an unwind
>>>> directly inside a bpf program will end up strangely. It have following
>>>> kernel message:
>>>
>>> Urgh, what is that bpf_get_stackid_tp() doing to get the regs? I can't
>>> follow.
>>
>> I guess we need something like the following? (we should be able to
>> optimize the PER_CPU stuff).
>>
>> Thanks,
>> Song
>>
>>
>> diff --git i/kernel/trace/bpf_trace.c w/kernel/trace/bpf_trace.c
>> index f92d6ad5e080..c525149028a7 100644
>> --- i/kernel/trace/bpf_trace.c
>> +++ w/kernel/trace/bpf_trace.c
>> @@ -696,11 +696,13 @@ static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
>> .arg5_type = ARG_CONST_SIZE_OR_ZERO,
>> };
>>
>> +static DEFINE_PER_CPU(struct pt_regs, bpf_stackid_tp_regs);
>> BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
>> u64, flags)
>> {
>> - struct pt_regs *regs = *(struct pt_regs **)tp_buff;
>> + struct pt_regs *regs = this_cpu_ptr(&bpf_stackid_tp_regs);
>>
>> + perf_fetch_caller_regs(regs);
>
> No. pt_regs is already passed in. It's the first argument.
> If we call perf_fetch_caller_regs() again the stack trace will be wrong.
> bpf prog should not see itself, interpreter or all the frames in between.
Thanks Alexei! I get it now.
In bpf_get_stackid_tp(), the pt_regs is get by dereferencing the first field
of tp_buff:
struct pt_regs *regs = *(struct pt_regs **)tp_buff;
tp_buff points to something like
struct sched_switch_args {
unsigned long long pad;
char prev_comm[16];
int prev_pid;
int prev_prio;
long long prev_state;
char next_comm[16];
int next_pid;
int next_prio;
};
where the first field "pad" is a pointer to pt_regs.
@Kairui, I think you confirmed that current code will give empty call trace
with ORC unwinder? If that's the case, can we add regs->ip back? (as in the
first email of this thread.
Thanks,
Song
