| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAOrHB_DKEHOmqT1XvQ=UiF-rgsGENymkD88B=zzjLxUz0RiVyw@mail.gmail.com>
Date: Sat, 11 Apr 2020 01:14:19 -0700
From: Pravin Shelar <pravin.ovn@...il.com>
To: Tonghao Zhang <xiangxia.m.yue@...il.com>
Cc: William Tu <u9012063@...il.com>, ovs dev <dev@...nvswitch.org>,
Linux Kernel Network Developers <netdev@...r.kernel.org>
Subject: Re: [ovs-dev] [PATCH net-next v1 1/3] net: openvswitch: expand the
meters number supported
Given that we already use id-pool, we can significantly reduce
probability of the negative case of meter lookup. Therefore I do not
see need to use hash table in the datapath.
On Thu, Apr 9, 2020 at 4:29 PM Tonghao Zhang <xiangxia.m.yue@...il.com> wrote:
>
> On Fri, Apr 10, 2020 at 5:41 AM William Tu <u9012063@...il.com> wrote:
> >
> > On Wed, Apr 08, 2020 at 11:59:25PM +0800, Tonghao Zhang wrote:
> > > On Wed, Apr 8, 2020 at 11:09 PM William Tu <u9012063@...il.com> wrote:
> > > >
> > > > On Wed, Apr 01, 2020 at 06:50:09PM +0800, Tonghao Zhang wrote:
> > > > > On Tue, Mar 31, 2020 at 11:57 AM Pravin Shelar <pshelar@....org> wrote:
> > > > > >
> > > > > > On Sun, Mar 29, 2020 at 5:35 PM Tonghao Zhang <xiangxia.m.yue@...il.com> wrote:
> > > > > > >
> > > > > > > On Mon, Mar 30, 2020 at 12:46 AM Pravin Shelar <pshelar@....org> wrote:
> > > > > > > >
> > > > > > > > On Sat, Mar 28, 2020 at 8:46 AM <xiangxia.m.yue@...il.com> wrote:
> > > > > > > > >
> > > > > > > > > From: Tonghao Zhang <xiangxia.m.yue@...il.com>
> > > > > > > > >
> > > > > > > > > In kernel datapath of Open vSwitch, there are only 1024
> > > > > > > > > buckets of meter in one dp. If installing more than 1024
> > > > > > > > > (e.g. 8192) meters, it may lead to the performance drop.
> > > > > > > > > But in some case, for example, Open vSwitch used as edge
> > > > > > > > > gateway, there should be 200,000+ at least, meters used for
> > > > > > > > > IP address bandwidth limitation.
> > > > > > > > >
> > > > > > > > > [Open vSwitch userspace datapath has this issue too.]
> > > > > > > > >
> > > > > > > > > For more scalable meter, this patch expands the buckets
> > > > > > > > > when necessary, so we can install more meters in the datapath.
> > > > > > > > >
> > > > > > > > > * Introducing the struct *dp_meter_instance*, it's easy to
> > > > > > > > > expand meter though change the *ti* point in the struct
> > > > > > > > > *dp_meter_table*.
> > > > > > > > > * Using kvmalloc_array instead of kmalloc_array.
> > > > > > > > >
> > > > > > > > Thanks for working on this, I have couple of comments.
> > > > > > > >
> > > > > > > > > Cc: Pravin B Shelar <pshelar@....org>
> > > > > > > > > Cc: Andy Zhou <azhou@....org>
> > > > > > > > > Signed-off-by: Tonghao Zhang <xiangxia.m.yue@...il.com>
> > > > > > > > > ---
> > > > > > > > > net/openvswitch/datapath.h | 2 +-
> > > > > > > > > net/openvswitch/meter.c | 168 ++++++++++++++++++++++++++++++-------
> > > > > > > > > net/openvswitch/meter.h | 17 +++-
> > > > > > > > > 3 files changed, 153 insertions(+), 34 deletions(-)
> > > > > > > > >
> > > > > > > > > diff --git a/net/openvswitch/datapath.h b/net/openvswitch/datapath.h
> > > > > > > > > index e239a46c2f94..785105578448 100644
> > > > > > > > > --- a/net/openvswitch/datapath.h
> > > > > > > > > +++ b/net/openvswitch/datapath.h
> > > > > > > > > @@ -82,7 +82,7 @@ struct datapath {
> > > > > > > > > u32 max_headroom;
> > > > > > > > >
> > > > > > > > > /* Switch meters. */
> > > > > > > > > - struct hlist_head *meters;
> > > > > > > > > + struct dp_meter_table *meters;
> > > > > > > > > };
> > > > > > > > >
> > > > > > > > > /**
> > > > > > > > > diff --git a/net/openvswitch/meter.c b/net/openvswitch/meter.c
> > > > > > > > > index 5010d1ddd4bd..98003b201b45 100644
> > > > > > > > > --- a/net/openvswitch/meter.c
> > > > > > > > > +++ b/net/openvswitch/meter.c
> > > > > > > > > @@ -47,40 +47,136 @@ static void ovs_meter_free(struct dp_meter *meter)
> > > > > > > > > kfree_rcu(meter, rcu);
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > -static struct hlist_head *meter_hash_bucket(const struct datapath *dp,
> > > > > > > > > +static struct hlist_head *meter_hash_bucket(struct dp_meter_instance *ti,
> > > > > > > > > u32 meter_id)
> > > > > > > > > {
> > > > > > > > > - return &dp->meters[meter_id & (METER_HASH_BUCKETS - 1)];
> > > > > > > > > + u32 hash = jhash_1word(meter_id, ti->hash_seed);
> > > > > > > > > +
> > > > > > > > I do not see any need to hash meter-id, can you explain it.
> > > > > > > >
> > > > > > > > > + return &ti->buckets[hash & (ti->n_buckets - 1)];
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > /* Call with ovs_mutex or RCU read lock. */
> > > > > > > > > -static struct dp_meter *lookup_meter(const struct datapath *dp,
> > > > > > > > > +static struct dp_meter *lookup_meter(const struct dp_meter_table *tbl,
> > > > > > > > > u32 meter_id)
> > > > > > > > > {
> > > > > > > > > + struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
> > > > > > > > > struct dp_meter *meter;
> > > > > > > > > struct hlist_head *head;
> > > > > > > > >
> > > > > > > > > - head = meter_hash_bucket(dp, meter_id);
> > > > > > > > > - hlist_for_each_entry_rcu(meter, head, dp_hash_node,
> > > > > > > > > - lockdep_ovsl_is_held()) {
> > > > > > > > > + head = meter_hash_bucket(ti, meter_id);
> > > > > > > > > + hlist_for_each_entry_rcu(meter, head, hash_node[ti->node_ver],
> > > > > > > > > + lockdep_ovsl_is_held()) {
> > > > > > > > > if (meter->id == meter_id)
> > > > > > > > > return meter;
> > > > > > > > > }
> > > > > > > > > +
> > > > > > > > This patch is expanding meter table linearly with number meters added
> > > > > > > > to datapath. so I do not see need to have hash table. it can be a
> > > > > > > > simple array. This would also improve lookup efficiency.
> > > > > > > > For hash collision we could find next free slot in array. let me know
> > > > > > > > what do you think about this approach.
> > > > > > > Hi Pravin
> > > > > > > If we use the simple array, when inserting the meter, for hash collision, we can
> > > > > > > find next free slot, but one case, when there are many meters in the array.
> > > > > > > we may find many slot for the free slot.
> > > > > > > And when we lookup the meter, for hash collision, we may find many
> > > > > > > array slots, and
> > > > > > > then find it, or that meter does not exist in the array, In that case,
> > > > > > > there may be a lookup performance
> > > > > > > drop.
> > > > > > >
> > > > > > I was thinking that users can insure that there are no hash collision,
> > > > > > but time complexity of negative case is expensive. so I am fine with
> > > > > > the hash table.
> > > >
> > > > IIUC, there will be hash collision. meter id is an 32-bit value.
> > > > Currenly in lib/dpif-netdev.c, MAX_METERS = 65536.
> > > Hi, William
> > > but id-pool makes sure the meter id is from 0, 1, 2, 3 ... n, but not n, m, y.
> > > so if we alloc 1024 meters, the last meter id should be 1023, and then
> > > use the simple array to expand the meter is better ?
> > >
> >
> > I see, so you want to set the # of hash bucket = max # of meter id,
> > so there is no hash collision, (with the cost of using more memory)
> Not really, there are 1024 buckets as default, and will expand to
> 1024*2, and then 1024*2*2 if necessary
> if the most meter is deleted, we will shrink it.
>
> > I don't have strong opinion on which design is better. Let's wait for
> > Pravin's feedback.
> >
> > William
> >
> > > > I think what Pravin suggest is to use another hash function to make
> > > > the hash table more condense. Ex: hash1 and hash2.
> > > > For lookup, if hash1(key) misses, then try hash2(key).
> > > >
> > > > William
> > > >
> > > > > Hi Pravi
> > > > > I check again the meter implementation of ovs, ovs-vswitchd use the id-pool to
> > > > > get a valid meter-id which passed to kernel, so there is no hash collision. You
> > > > > are right. we use the single array is the better solution.
> > > > > > > For hash meter-id in meter_hash_bucket, I am not 100% sure it is
> > > > > > > useful. it just update
> > > > > > > hash_seed when expand meters. For performance, we can remove it. Thanks.
> > > > > > ok.
> >
>
>
> --
> Best regards, Tonghao
Powered by blists - more mailing lists