Re: [PATCH net-next 0/4] bpf: bpf_htab: Add BPF_MAP_TYPE_PERCPU_HASH

[Martin KaFai Lau]

Hi Ming,

On Sat, Jan 09, 2016 at 05:39:16PM +0800, Ming Lei wrote:
> On Sat, Jan 9, 2016 at 8:44 AM, Martin KaFai Lau <kafai@xxxxxx> wrote:
> > On Fri, Jan 08, 2016 at 02:55:32PM +0800, Ming Lei wrote:
> >> On Fri, Jan 8, 2016 at 6:35 AM, Martin KaFai Lau <kafai@xxxxxx> wrote:
> >> > This patchset adds BPF_MAP_TYPE_PERCPU_HASH map type which allows
> >> > percpu value.
> >>
> >> I am also thinking about using percpu variable to ebpf map, but IMO it
> >> should be better for ARRAY map instead of HASH map, then we can
> >> avoid the atomic op in eBPF program, see example of tracex3, sockex1
> >> and sockex3 in sample/bpf/ of kernel tree.  Also looks the ARRAY map
> >> usage in bcc is wrong, strictly speaking.
> > array and hash are two different use cases. May be we should have percpu
> > value for array map too.
>
> Atomic operations are definitely expensive, which should have been avoided
> in eBPF prog by percpu way. I think ARRAY map does need percpu way more, :-)
>
> But for hash map, I am still not 100% sure if percpu is needed:
>
> - key plus cpu_id is ok
> - key often can be chosed so that the mapped value is accessed without
> race(such as, req is used as key in biolatency, ...)
> - percpu hash can cause huge memory consumption compared with the way
> of key plus cpu_id without obvious performance advantage
If the bpf prog uses the percpu map type, the expectation
is the bpf is tapping some kprobes that most cpus (if not all) will access
the same key.  I would argue that having a bigger key (due to cpu_id),
and 40 more keys (and linking to 40 buckets) is less efficient.

If the expectation is to have a small numbers of cpu accessing a key (like
biolatency), the bpf prog can stay with the regular bpf map.  The new
percpu map has almost the same interface (except the syscall added in
Patch 3).  Switching between the two map types is easy.

> >> For HASH map, it is easy to make cpu id as part of key, then the map
> >> can be thought as percpu too, and atomic op isn't needed in eBPF program.
> > Putting the cpu id as part of the key was indeed the first hack I did
> > to get a sense of potential benefit.
> >
> > However, by extending the real-key with cpu-id, it is not intuitive to
> > use and it is prone to error.  For example, how to delete a real-key for
> > all cpus?  Iterating a particular real-key for all cpu is also tricky.  What
> > does it mean if a real-key exists for cpu#0 but not cpu#1? The real-key
>
> lookup returns NULL if the key(real key plus cpu id) doesn't exist.
>
> > got deleted from all cpu while iterating? or something else?  I believe
> > there are ways to get around but it is better to provide a clean
> > implementation instead.
>
> It might be true, but I hope there is one real example, in which we can
> see the obvious disadvantage of real key plus cpu_id. In the other way,
> we still can improve current hash map to make this case easier to use.
By extend(real_key, cpu_id), the uniqueness property of a real_key (and
it should be as a key for a hashmap) is getting blurry and it is not
ideal.

The typical use case is for doing a lot of aggregation (which bpf+kprobe
is strong at) on a real_key which will exist for a reasonable long period
of time.  For example, I would like to do some packet statistics on each
IPv6 /64 prefix.  The userspace will iterate the /64 prefix (real_key)
and sum up all values (of each cpu) before reporting for a particular
prefix.

When the userspace prog can find extend(real_key, cpu0) but not
extend(real_key, cpu1).  What does it mean?
It means the real_key has been deleted from all cpu and probably
not making sense to continue reporting?
or cpu1 has not inserted the real_key yet and I should
continue with cpu2?

How to delete a real_key from all cpu? A loop to iterate all
extend(real_key, cpu_id)?

> That sounds not a good result, 40X memory consumption only save %3 CPU,
> and the approach(real_key, cpu_id) still can save %3 without extra memory
> consumption(or much less than percpu MAP)
Another way to look at it, the bpf prog costs 6% CPU and saving 3%
with percpu implemenation.

Thanks,
-- Martin