Re: [PATCHv4 bpf-next 2/4] xdp: extend xdp_redirect_map with broadcast support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



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





[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux