> On Apr 2, 2021, at 1:57 PM, Cong Wang <xiyou.wangc...@gmail.com> wrote:
> 
> On Fri, Apr 2, 2021 at 12:45 PM Song Liu <songliubrav...@fb.com> wrote:
>> 
>> 
>> 
>>> On Apr 2, 2021, at 12:08 PM, Cong Wang <xiyou.wangc...@gmail.com> wrote:
>>> 
>>> On Fri, Apr 2, 2021 at 10:57 AM Song Liu <songliubrav...@fb.com> wrote:
>>>> 
>>>> 
>>>> 
>>>>> On Apr 2, 2021, at 10:34 AM, Cong Wang <xiyou.wangc...@gmail.com> wrote:
>>>>> 
>>>>> On Thu, Apr 1, 2021 at 1:17 PM Song Liu <songliubrav...@fb.com> wrote:
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>>> On Apr 1, 2021, at 10:28 AM, Cong Wang <xiyou.wangc...@gmail.com> wrote:
>>>>>>> 
>>>>>>> On Wed, Mar 31, 2021 at 11:38 PM Song Liu <songliubrav...@fb.com> wrote:
>>>>>>>> 
>>>>>>>> 
>>>>>>>> 
>>>>>>>>> On Mar 31, 2021, at 9:26 PM, Cong Wang <xiyou.wangc...@gmail.com> 
>>>>>>>>> wrote:
>>>>>>>>> 
>>>>>>>>> From: Cong Wang <cong.w...@bytedance.com>
>>>>>>>>> 
>>>>>>>>> (This patch is still in early stage and obviously incomplete. I am 
>>>>>>>>> sending
>>>>>>>>> it out to get some high-level feedbacks. Please kindly ignore any 
>>>>>>>>> coding
>>>>>>>>> details for now and focus on the design.)
>>>>>>>> 
>>>>>>>> Could you please explain the use case of the timer? Is it the same as
>>>>>>>> earlier proposal of BPF_MAP_TYPE_TIMEOUT_HASH?
>>>>>>>> 
>>>>>>>> Assuming that is the case, I guess the use case is to assign an expire
>>>>>>>> time for each element in a hash map; and periodically remove expired
>>>>>>>> element from the map.
>>>>>>>> 
>>>>>>>> If this is still correct, my next question is: how does this compare
>>>>>>>> against a user space timer? Will the user space timer be too slow?
>>>>>>> 
>>>>>>> Yes, as I explained in timeout hashmap patchset, doing it in user-space
>>>>>>> would require a lot of syscalls (without batching) or copying (with 
>>>>>>> batching).
>>>>>>> I will add the explanation here, in case people miss why we need a 
>>>>>>> timer.
>>>>>> 
>>>>>> How about we use a user space timer to trigger a BPF program (e.g. use
>>>>>> BPF_PROG_TEST_RUN on a raw_tp program); then, in the BPF program, we can
>>>>>> use bpf_for_each_map_elem and bpf_map_delete_elem to scan and update the
>>>>>> map? With this approach, we only need one syscall per period.
>>>>> 
>>>>> Interesting, I didn't know we can explicitly trigger a BPF program running
>>>>> from user-space. Is it for testing purposes only?
>>>> 
>>>> This is not only for testing. We will use this in perf (starting in 5.13).
>>>> 
>>>> /* currently in Arnaldo's tree, tools/perf/util/bpf_counter.c: */
>>>> 
>>>> /* trigger the leader program on a cpu */
>>>> static int bperf_trigger_reading(int prog_fd, int cpu)
>>>> {
>>>>       DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts,
>>>>                           .ctx_in = NULL,
>>>>                           .ctx_size_in = 0,
>>>>                           .flags = BPF_F_TEST_RUN_ON_CPU,
>>>>                           .cpu = cpu,
>>>>                           .retval = 0,
>>>>               );
>>>> 
>>>>       return bpf_prog_test_run_opts(prog_fd, &opts);
>>>> }
>>>> 
>>>> test_run also passes return value (retval) back to user space, so we and
>>>> adjust the timer interval based on retval.
>>> 
>>> This is really odd, every name here contains a "test" but it is not for 
>>> testing
>>> purposes. You probably need to rename/alias it. ;)
>>> 
>>> So, with this we have to get a user-space daemon running just to keep
>>> this "timer" alive. If I want to run it every 1ms, it means I have to issue
>>> a syscall BPF_PROG_TEST_RUN every 1ms. Even with a timer fd, we
>>> still need poll() and timerfd_settime(). This is a considerable overhead
>>> for just a single timer.
>> 
>> sys_bpf() takes about 0.5us. I would expect poll() and timerfd_settime() to
>> be slightly faster. So the overhead is less than 0.2% of a single core
>> (0.5us x 3 / 1ms). Do we need many counters for conntrack?
> 
> This is just for one timer. The whole system may end up with many timers
> when we have more and more eBPF programs. So managing the timers
> in the use-space would be a problem too someday, clearly one daemon
> per-timer does not scale.

If we do need many timers, I agree that it doesn't make sense to create 
a thread for each of them. 

A less-flexible alternative is to create a perf_event of "cpu-clock" and 
attach BPF program to it. It is not easy to adjust the interval, I guess.

> 
>> 
>>> 
>>> With current design, user-space can just exit after installing the timer,
>>> either it can adjust itself or other eBPF code can adjust it, so the per
>>> timer overhead is the same as a kernel timer.
>> 
>> I guess we still need to hold a fd to the prog/map? Alternatively, we can
>> pin the prog/map, but then the user need to clean it up.
> 
> Yes, but I don't see how holding a fd could bring any overhead after
> initial setup.
>> 
>>> 
>>> The visibility to other BPF code is important for the conntrack case,
>>> because each time we get an expired item during a lookup, we may
>>> want to schedule the GC timer to run sooner. At least this would give
>>> users more freedom to decide when to reschedule the timer.
>> 
>> Do we plan to share the timer program among multiple processes (which can
>> start and terminate in arbitrary orders)? If that is the case, I can imagine
>> a timer program is better than a user space timer.
> 
> I mean I want other eBPF program to manage the timers in kernel-space,
> as conntrack is mostly in kernel-space. If the timer is only manageable
> in user-space, it would seriously limit its use case.
> 
> Ideally I even prefer to create timers in kernel-space too, but as I already
> explained, this seems impossible to me.

Would hrtimer (include/linux/hrtimer.h) work? 

Thanks,
Song





Reply via email to