John Fastabend <john.fastabend@xxxxxxxxx> writes:
> Toke Høiland-Jørgensen wrote:
>> Hangbin Liu <liuhangbin@xxxxxxxxx> writes:
>>
>> > On Mon, Apr 05, 2021 at 05:24:48PM -0700, John Fastabend wrote:
>> >> Hangbin Liu wrote:
>> >> > This patch add two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend
>> >> > xdp_redirect_map for broadcast support.
>> >> >
>> >> > Keep the general data path in net/core/filter.c and the native data
>> >> > path in kernel/bpf/devmap.c so we can use direct calls to get better
>> >> > performace.
>> >> >
>> >> > Here is the performance result by using xdp_redirect_{map, map_multi} in
>> >> > sample/bpf and send pkts via pktgen cmd:
>> >> > ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64
>> >> >
>> >> > There are some drop back as we need to loop the map and get each interface.
>> >> >
>> >> > Version | Test | Generic | Native
>> >> > 5.12 rc2 | redirect_map i40e->i40e | 2.0M | 9.8M
>> >> > 5.12 rc2 | redirect_map i40e->veth | 1.8M | 12.0M
>> >>
>> >> Are these are 10gbps i40e ports? Sorry if I asked this earlier, maybe
>> >> add a note in the commit if another respin is needed.
>> >
>> > Yes, I will add it if there is an update.
>> >
>> >> > diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
>> >> > index 3980fb3bfb09..c8452c5f40f8 100644
>> >> > --- a/kernel/bpf/devmap.c
>> >> > +++ b/kernel/bpf/devmap.c
>> >> > @@ -198,6 +198,7 @@ static void dev_map_free(struct bpf_map *map)
>> >> > list_del_rcu(&dtab->list);
>> >> > spin_unlock(&dev_map_lock);
>> >> >
>> >> > + bpf_clear_redirect_map(map);
>> >>
>> >> Is this a bugfix? If its needed here wouldn't we also need it in the
>> >> devmap case.
>> >
>> > No, in ee75aef23afe ("bpf, xdp: Restructure redirect actions") this function
>> > was removed. I added it back as we use ri->map again.
>> >
>> > What devmap case you mean?
>> >
>> >>
>> >> > synchronize_rcu();
>> >> >
>> >> > /* Make sure prior __dev_map_entry_free() have completed. */
>> >>
>> >> [...]
>> >>
>> >> > +
>> >> > +static struct bpf_dtab_netdev *devmap_get_next_obj(struct xdp_buff *xdp,
>> >> > + struct bpf_map *map,
>> >> > + u32 *key, u32 *next_key,
>> >> > + int ex_ifindex)
>> >> > +{
>> >> > + struct bpf_dtab_netdev *obj;
>> >> > + struct net_device *dev;
>> >> > + u32 index;
>> >> > + int err;
>> >> > +
>> >> > + err = devmap_get_next_key(map, key, next_key);
>> >> > + if (err)
>> >> > + return NULL;
>> >> > +
>> >> > + /* When using dev map hash, we could restart the hashtab traversal
>> >> > + * in case the key has been updated/removed in the mean time.
>> >> > + * So we may end up potentially looping due to traversal restarts
>> >> > + * from first elem.
>> >> > + *
>> >> > + * Let's use map's max_entries to limit the loop number.
>> >> > + */
>> >> > + for (index = 0; index < map->max_entries; index++) {
>> >> > + obj = devmap_lookup_elem(map, *next_key);
>> >> > + if (!obj || dst_dev_is_ingress(obj, ex_ifindex))
>> >> > + goto find_next;
>> >> > +
>> >> > + dev = obj->dev;
>> >> > +
>> >> > + if (!dev->netdev_ops->ndo_xdp_xmit)
>> >> > + goto find_next;
>> >> > +
>> >> > + err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
>> >> > + if (unlikely(err))
>> >> > + goto find_next;
>> >> > +
>> >> > + return obj;
>> >> > +
>> >> > +find_next:
>> >> > + key = next_key;
>> >> > + err = devmap_get_next_key(map, key, next_key);
>> >> > + if (err)
>> >> > + break;
>> >> > + }
>> >>
>> >> I'm missing something. Either an elaborated commit message or comment
>> >> is probably needed. I've been looking at this block for 30 minutes and
>> >> can't see how we avoid sending duplicate frames on a single interface?
>> >> Can you check this code flow,
>> >>
>> >> dev_map_enqueue_multi()
>> >> for (;;) {
>> >> next_obj = devmap_get_next_obj(...)
>> >> for (index = 0; index < map->max_entries; index++) {
>> >> obj = devmap_lookup_elem();
>> >> if (!obj) goto find_next
>> >> key = next_key;
>> >> err = devmap_get_next_key()
>> >> if (!key) goto find_first
>> >> for (i = 0; i < dtab->n_buckets; i++)
>> >> return *next <- now *next_key is point back
>> >> at first entry
>> >> // loop back through and find first obj and return that
>> >
>> > devmap_get_next_key() will loop to find the first one if there is no
>> > key or dev. In normal time it will stop after the latest one.
>> >> }
>> >> bq_enqueue(...) // enqueue original obj
>> >> obj = next_obj;
>> >> key = next_key;
>> >> ... // we are going to enqueue first obj, but how do we know
>> >> // this hasn't already been sent? Presumably if we have
>> >> // a delete in the hash table in the middle of a multicast
>> >> // operation this might happen?
>> >> }
>> >
>> > And yes, there is an corner case that if we removed a dev during multicast,
>> > there is an possibility that restart from the first key. But given that
>> > this is an unlikely case, and in normal internet there is also a possibility
>> > of duplicate/lost packet. This should also be acceptable?
>>
>> In my mind this falls under "acceptable corner cases". I.e., if you're
>> going to use the map for redirect and you expect to be updating it while
>> you're doing so, don't use a hashmap. But if you will not be updating
>> the map (or find the possible duplication acceptable), you can use the
>> hashmap and gain the benefit of being able to index by ifindex.
>
> In a Kubernetes setup its going to be hard, if possible at all, to restrict
> the map from moving as interfaces/IPs are going to be dynamic. Using a
> hash map has nice benefits of not having to figure out how to put ifindex's
> into the array. Although on some early implementations I wrote a small
> hashing algorithm over the top of array, so that could work.
>
> I don't know how well multicast applications might handle duplicate packets.
> I wouldn't be too surprised if it was problematic. On the other hand missing
> an entry that was just added is likely OK. There is no way to know from
> network/user side if the entry was actually added before multicast op and
> skipped or insert happened just after multicast op. And vice versa for a
> delete dev, no way to know the multicast op happened before/after the
> delete.
>
> Have we consider doing something like the batch lookup ops over hashtab?
> I don't mind "missing" values so if we just walk the list?
>
> head = dev_map_index_hash(dtab, i)
> // collect all my devs and get ready to send multicast
> hlist_nulls_for_each_entry_safe(dev, next, head, index_hlist) {
> enqueue(dev, skb)
> }
> // submit the queue of entries and do all the work to actually xmit
> submit_enqueued();
>
> We don't have to care about keys just walk the hash list?
So you'd wrap that in a loop like:
for (i = 0; i < dtab->n_buckets; i++) {
head = dev_map_index_hash(dtab, i);
hlist_nulls_for_each_entry_safe(dev, next, head, index_hlist) {
bq_enqueue(dev, xdpf, dev_rx, obj->xdp_prog);
}
}
or? Yeah, I guess that would work!
It would mean that dev_map_enqueue_multi() would need more in-depth
knowledge into the map type, so would likely need to be two different
functions for the two different map types, living in devmap.c - but
that's probably acceptable.
And while we're doing that, the array-map version can also loop over all
indexes up to max_entries, instead of stopping at the first index that
doesn't have an entry like it does now (right now, it looks like if you
populate entries 0 and 2 in an array-map only one copy of the packet
will be sent, to index 0).
It makes it a bit more awkward to do Hangbin's clever trick to avoid
doing an extra copy by aborting the loop early. But I guess the same
technique could apply...
-Toke
