Re: [RFC PATCH bpf-next 15/16] tools/bpf: selftests: add dumper progs for bpf_map/task/task_file

[Yonghong Song <yhs@fb.com>]

On 4/9/20 8:33 PM, Alexei Starovoitov wrote:
> On Wed, Apr 08, 2020 at 04:25:38PM -0700, Yonghong Song wrote:
>> For task/file, the dumper prints out:
>>    $ cat /sys/kernel/bpfdump/task/file/my1
>>      tgid      gid       fd      file
>>         1        1        0 ffffffff95c97600
>>         1        1        1 ffffffff95c97600
>>         1        1        2 ffffffff95c97600
>>      ....
>>      1895     1895      255 ffffffff95c8fe00
>>      1932     1932        0 ffffffff95c8fe00
>>      1932     1932        1 ffffffff95c8fe00
>>      1932     1932        2 ffffffff95c8fe00
>>      1932     1932        3 ffffffff95c185c0
> ...
>> +SEC("dump//sys/kernel/bpfdump/task/file")
>> +int BPF_PROG(dump_tasks, struct task_struct *task, __u32 fd, struct file *file,
>> +	     struct seq_file *seq, u64 seq_num)
>> +{
>> +	static char const banner[] = "    tgid      gid       fd      file\n";
>> +	static char const fmt1[] = "%8d %8d";
>> +	static char const fmt2[] = " %8d %lx\n";
>> +
>> +	if (seq_num == 0)
>> +		bpf_seq_printf(seq, banner, sizeof(banner));
>> +
>> +	bpf_seq_printf(seq, fmt1, sizeof(fmt1), task->tgid, task->pid);
>> +	bpf_seq_printf(seq, fmt2, sizeof(fmt2), fd, (long)file->f_op);
>> +	return 0;
>> +}
> 
> I wonder what is the speed of walking all files in all tasks with an empty
> program? If it's fast I can imagine a million use cases for such searching bpf
> prog. Like finding which task owns particular socket. This could be a massive
> feature.
> 
> With one redundant spin_lock removed it seems it will be one spin_lock per prog
> invocation? May be eventually it can be amortized within seq_file iterating
> logic. Would be really awesome if the cost is just refcnt ++/-- per call and
> rcu_read_lock.

The main seq_read() loop is below:
         while (1) {
                 size_t offs = m->count;
                 loff_t pos = m->index;

                 p = m->op->next(m, p, &m->index);
                 if (pos == m->index)
                         /* Buggy ->next function */
                         m->index++;
                 if (!p || IS_ERR(p)) {
                         err = PTR_ERR(p);
                         break;
                 }
                 if (m->count >= size)
                         break;
                 err = m->op->show(m, p);
                 if (seq_has_overflowed(m) || err) {
                         m->count = offs;
                         if (likely(err <= 0))
                                 break;
                 }
         }

If we remove the spin_lock() as in another email comment,
we won't have spin_lock() in seq_ops->next() function, only
refcnt ++/-- and rcu_read_{lock, unlock}s. The seq_ops->show() does
not have any spin_lock() either.

I have not got time to do a perf measurement yet.
Will do in the next revision.