Re: [PATCH v2 2/2] perf-stat: enable counting events for BPF programs

[Song Liu <songliubraving@fb.com>]

> On Dec 7, 2020, at 2:07 PM, Jiri Olsa <jolsa@redhat.com> wrote:
> 
> On Thu, Dec 03, 2020 at 10:13:10PM -0800, Song Liu wrote:
> 
> SNIP
> 
>> +#include "bpf_skel/bpf_prog_profiler.skel.h"
>> +
>> +static inline void *u64_to_ptr(__u64 ptr)
>> +{
>> +	return (void *)(unsigned long)ptr;
>> +}
>> +
>> +static void set_max_rlimit(void)
>> +{
>> +	struct rlimit rinf = { RLIM_INFINITY, RLIM_INFINITY };
>> +
>> +	setrlimit(RLIMIT_MEMLOCK, &rinf);
>> +}
>> +
>> +static inline struct bpf_counter *bpf_counter_alloc(void)
> 
> why is this inlined?

We don't need the inline here. I will remove it in the next version. 

> 
> SNIP
> 
>> +static int bpf_program_profiler_load_one(struct evsel *evsel, u32 prog_id)
>> +{
>> +	struct bpf_prog_profiler_bpf *skel;
>> +	struct bpf_counter *counter;
>> +	struct bpf_program *prog;
>> +	char *prog_name;
>> +	int prog_fd;
>> +	int err;
>> +
>> +	prog_fd = bpf_prog_get_fd_by_id(prog_id);
>> +	if (prog_fd < 0) {
>> +		pr_debug("Failed to open fd for bpf prog %u\n", prog_id);
>> +		return -1;
>> +	}
>> +	counter = bpf_counter_alloc();
>> +	if (!counter)
>> +		return -1;
>> +
>> +	skel = bpf_prog_profiler_bpf__open();
>> +	if (!skel) {
>> +		pr_debug("Failed to load bpf skeleton\n");
> 
> I'm still getting
> 
> [root@dell-r440-01 perf]# ./perf stat -b 38
> libbpf: elf: skipping unrecognized data section(9) .eh_frame
> libbpf: elf: skipping relo section(15) .rel.eh_frame for section(9) .eh_frame
> libbpf: XXX is not found in vmlinux BTF
> libbpf: failed to load object 'bpf_prog_profiler_bpf'
> libbpf: failed to load BPF skeleton 'bpf_prog_profiler_bpf': -2
> ...
> 
> with id 38 being:
> 
> 38: tracepoint  name sys_enter  tag 03418b72a610af75  gpl
>        loaded_at 2020-12-07T22:54:05+0100  uid 0
>        xlated 272B  jited 153B  memlock 4096B  map_ids 1
> 
> how is this supposed to work when there's XXX in the
> program's section? libbpf is trying to find XXX in
> kernel BTF and fails of course

I think this is because this program doesn't have BTF. The actual failed
function was bpf_program__set_attach_target(). So the error message above
should be "Failed to _open_ bpf skeleton". I will fix the error messages. 

> 
> 
>> +		free(counter);
>> +		return -1;
>> +	}
>> +	skel->rodata->num_cpu = evsel__nr_cpus(evsel);
>> +
>> +	bpf_map__resize(skel->maps.events, evsel__nr_cpus(evsel));
>> +	bpf_map__resize(skel->maps.fentry_readings, 1);
>> +	bpf_map__resize(skel->maps.accum_readings, 1);
>> +
> 
> SNIP
> 
>> +static int bpf_program_profiler__read(struct evsel *evsel)
>> +{
>> +	int num_cpu = evsel__nr_cpus(evsel);
>> +	struct bpf_perf_event_value values[num_cpu];
>> +	struct bpf_counter *counter;
>> +	int reading_map_fd;
>> +	__u32 key = 0;
>> +	int err, cpu;
>> +
>> +	if (list_empty(&evsel->bpf_counter_list))
>> +		return -EAGAIN;
>> +
>> +	for (cpu = 0; cpu < num_cpu; cpu++) {
>> +		perf_counts(evsel->counts, cpu, 0)->val = 0;
>> +		perf_counts(evsel->counts, cpu, 0)->ena = 0;
>> +		perf_counts(evsel->counts, cpu, 0)->run = 0;
>> +	}
>> +	list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
>> +		struct bpf_prog_profiler_bpf *skel = counter->skel;
>> +
>> +		reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
>> +
>> +		err = bpf_map_lookup_elem(reading_map_fd, &key, values);
>> +		if (err) {
>> +			fprintf(stderr, "failed to read value\n");
>> +			return err;
>> +		}
>> +
>> +		for (cpu = 0; cpu < num_cpu; cpu++) {
>> +			perf_counts(evsel->counts, cpu, 0)->val += values[cpu].counter;
>> +			perf_counts(evsel->counts, cpu, 0)->ena += values[cpu].enabled;
>> +			perf_counts(evsel->counts, cpu, 0)->run += values[cpu].running;
>> +		}
> 
> so we sum everything up for all provided bpf IDs,
> should we count/display them separately?

I think that's the default behavior with --pid x,y,z or --cpu a,b,c. 
Do we need to separate them?

> 
> SNIP
> 
>> +SEC("fentry/XXX")
>> +int BPF_PROG(fentry_XXX)
>> +{
>> +	__u32 key = bpf_get_smp_processor_id();
>> +	struct bpf_perf_event_value reading;
>> +	struct bpf_perf_event_value *ptr;
>> +	__u32 zero = 0;
>> +	long err;
>> +
>> +	/* look up before reading, to reduce error */
>> +	ptr = bpf_map_lookup_elem(&fentry_readings, &zero);
>> +	if (!ptr)
>> +		return 0;
>> +
>> +	err = bpf_perf_event_read_value(&events, key, &reading,
>> +					sizeof(reading));
> 
> can't we read directly to ptr in here?

Yes, we can! Thanks for catching this. 

> 
> SNIP
> 
>> 	/* THREAD and SYSTEM/CPU are mutually exclusive */
>> 	if (target->per_thread && (target->system_wide || target->cpu_list)) {
>> 		target->per_thread = false;
>> @@ -109,6 +137,10 @@ static const char *target__error_str[] = {
>> 	"PID/TID switch overriding SYSTEM",
>> 	"UID switch overriding SYSTEM",
>> 	"SYSTEM/CPU switch overriding PER-THREAD",
>> +	"BPF switch overriding CPU",
>> +	"BPF switch overriding PID/TID",
>> +	"BPF switch overriding UID",
>> +	"BPF switch overriding THREAD",
>> 	"Invalid User: %s",
>> 	"Problems obtaining information for user %s",
>> };
>> @@ -134,7 +166,7 @@ int target__strerror(struct target *target, int errnum,
>> 
>> 	switch (errnum) {
>> 	case TARGET_ERRNO__PID_OVERRIDE_CPU ...
>> -	     TARGET_ERRNO__SYSTEM_OVERRIDE_THREAD:
> 
> hum, this should stay, no?

We need this to show the warning like:

~/perf stat -e cycles,instructions -b 245561 -C 0
BPF switch overriding CPU
...

Thanks,
Song