On Fri, Oct 13, 2017 at 5:12 PM, Pravin Shelar <pshe...@ovn.org> wrote:
> On Thu, Oct 12, 2017 at 3:38 PM, Andy Zhou <az...@ovn.org> wrote:
>> OVS kernel datapath so far does not support Openflow meter action.
>> This is the first stab at adding kernel datapath meter support.
>> This implementation supports only drop band type.
>>
>> Signed-off-by: Andy Zhou <az...@ovn.org>
>> ---
>> net/openvswitch/Makefile | 1 +
>> net/openvswitch/datapath.c | 14 +-
>> net/openvswitch/datapath.h | 3 +
>> net/openvswitch/meter.c | 611
>> +++++++++++++++++++++++++++++++++++++++++++++
>> net/openvswitch/meter.h | 54 ++++
>> 5 files changed, 681 insertions(+), 2 deletions(-)
>> create mode 100644 net/openvswitch/meter.c
>> create mode 100644 net/openvswitch/meter.h
>>
> ...
>
>> diff --git a/net/openvswitch/meter.c b/net/openvswitch/meter.c
>> new file mode 100644
>> index 000000000000..f24ebb5f7af4
>> --- /dev/null
>> +++ b/net/openvswitch/meter.c
>
> ....
> ....
>> +static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info
>> *info)
>> +{
>> + struct datapath *dp;
>> + struct ovs_header *ovs_header = info->userhdr;
>> + struct sk_buff *reply;
>> + struct ovs_header *ovs_reply_header;
>> + struct nlattr *nla, *band_nla;
>> + int err;
>> +
>> + /* Check that the datapath exists */
>> + ovs_lock();
>> + dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
>> + ovs_unlock();
>> + if (!dp)
>> + return -ENODEV;
>> +
> why dp check is required for this API?
Is it possible for another core delete the dp, before ovs_lock()
returns? Then, in theory, get_dp() can
return NULL, no?
>
>> + reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
>> + &ovs_reply_header);
>> + if (!reply)
>> + return PTR_ERR(reply);
>> +
>> + if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS, U32_MAX) ||
>> + nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
>> + goto nla_put_failure;
>> +
>> + nla = nla_nest_start(reply, OVS_METER_ATTR_BANDS);
>> + if (!nla)
>> + goto nla_put_failure;
>> +
>> + band_nla = nla_nest_start(reply, OVS_BAND_ATTR_UNSPEC);
>> + if (!band_nla)
>> + goto nla_put_failure;
>> + /* Currently only DROP band type is supported. */
>> + if (nla_put_u32(reply, OVS_BAND_ATTR_TYPE, OVS_METER_BAND_TYPE_DROP))
>> + goto nla_put_failure;
>> + nla_nest_end(reply, band_nla);
>> + nla_nest_end(reply, nla);
>> +
>> + genlmsg_end(reply, ovs_reply_header);
>> + return genlmsg_reply(reply, info);
>> +
>> +nla_put_failure:
>> + nlmsg_free(reply);
>> + err = -EMSGSIZE;
>> + return err;
>> +}
>> +
> ....
>
>> +static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
>> +{
>> + struct nlattr **a = info->attrs;
>> + struct dp_meter *meter, *old_meter;
>> + struct sk_buff *reply;
>> + struct ovs_header *ovs_reply_header;
>> + struct ovs_header *ovs_header = info->userhdr;
>> + struct datapath *dp;
>> + int err;
>> + u32 meter_id;
>> + bool failed;
>> +
>> + meter = dp_meter_create(a);
>> + if (IS_ERR_OR_NULL(meter))
>> + return PTR_ERR(meter);
>> +
>> + reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
>> + &ovs_reply_header);
>> + if (IS_ERR(reply)) {
>> + err = PTR_ERR(reply);
>> + goto exit_free_meter;
>> + }
>> +
>> + ovs_lock();
>> + dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
>> + if (!dp) {
>> + err = -ENODEV;
>> + goto exit_unlock;
>> + }
>> +
>> + if (!a[OVS_METER_ATTR_ID]) {
>> + err = -ENODEV;
>> + goto exit_unlock;
>> + }
>> +
>> + meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
>> +
>> + /* Cannot fail after this. */
>> + old_meter = lookup_meter(dp, meter_id);
>> + attach_meter(dp, meter);
>> + ovs_unlock();
>> +
> After the unlock, it is not safe to keep the ref to old_meter. better
> to release lock at the end. we could optimize it later if required.
I see a problem here: the old_meter has not been removed from the list before
unlock. The code should have been:
old_meter = lookup_meter(dp, meter_id);
detch_meter(dp, old_meter);
attach_meter(dp, meter);
ovs_unlock();
Do you still see a problem w.r.t. unlock() here? Once detch_meter() is called,
another thread should not have access to 'old_meter' any more. right?
>
>> + /* Build response with the meter_id and stats from
>> + * the old meter, if any.
>> + */
>> + failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
>> + WARN_ON(failed);
>> + if (old_meter) {
>> + spin_lock_bh(&old_meter->lock);
>> + if (old_meter->keep_stats) {
>> + err = ovs_meter_cmd_reply_stats(reply, meter_id,
>> + old_meter);
>> + WARN_ON(err);
>> + }
>> + spin_unlock_bh(&old_meter->lock);
>> + ovs_meter_free(old_meter);
>> + }
>> +
>> + genlmsg_end(reply, ovs_reply_header);
>> + return genlmsg_reply(reply, info);
>> +
>> +exit_unlock:
>> + ovs_unlock();
>> + nlmsg_free(reply);
>> +exit_free_meter:
>> + kfree(meter);
>> + return err;
>> +}
>> +
> ....
>
>> +bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
>> + struct sw_flow_key *key, u32 meter_id)
>> +{
>> + struct dp_meter *meter;
>> + struct dp_meter_band *band;
>> + long long int now_ms = ktime_get_ns() / 1000 / 1000;
>> + long long int long_delta_ms;
>> + u32 delta_ms;
>> + u32 cost;
>> + int i, band_exceeded_max = -1;
>> + u32 band_exceeded_rate = 0;
>> +
>> + meter = lookup_meter(dp, meter_id);
>> + /* Do not drop the packet when there is no meter. */
>> + if (!meter)
>> + return false;
>> +
>> + /* Lock the meter while using it. */
>> + spin_lock(&meter->lock);
>> +
>> + long_delta_ms = (now_ms - meter->used); /* ms */
>> +
>> + /* Make sure delta_ms will not be too large, so that bucket will not
>> + * wrap around below.
>> + */
>> + delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
>> + ? meter->max_delta_t : (u32)long_delta_ms;
>> +
>> + /* Update meter statistics.
>> + */
>> + meter->used = now_ms;
>> + meter->stats.n_packets += 1;
>> + meter->stats.n_bytes += skb->len;
>> +
>> + /* Bucket rate is either in kilobits per second, or in packets per
>> + * second. We maintain the bucket in the units of either bits or
>> + * 1/1000th of a packet, correspondingly.
>> + * Then, when rate is multiplied with milliseconds, we get the
>> + * bucket units:
>> + * msec * kbps = bits, and
>> + * msec * packets/sec = 1/1000 packets.
>> + *
>> + * 'cost' is the number of bucket units in this packet.
>> + */
>> + cost = (meter->kbps) ? skb->len * 8 : 1000;
>> +
>> + /* Update all bands and find the one hit with the highest rate. */
>> + for (i = 0; i < meter->n_bands; ++i) {
>> + long long int max_bucket_size;
>> +
>> + band = &meter->bands[i];
>> + max_bucket_size = (band->burst_size + band->rate) * 1000;
>> +
>> + band->bucket += delta_ms * band->rate;
>> + if (band->bucket > max_bucket_size)
>> + band->bucket = max_bucket_size;
>> +
>> + if (band->bucket >= cost) {
>> + band->bucket -= cost;
>> + } else if (band->rate > band_exceeded_rate) {
>> + band_exceeded_rate = band->rate;
>> + band_exceeded_max = i;
>> + }
>> + }
>> +
>> + spin_unlock(&meter->lock);
>> +
>> + if (band_exceeded_max >= 0) {
>> + /* Update band statistics. */
>> + band = &meter->bands[band_exceeded_max];
>> + band->stats.n_packets += 1;
>> + band->stats.n_bytes += skb->len;
>> +
> Is it safe to do outside of the sipinlock?
Good catch, those should be covered by the same lock. Will fix.
>
>> + /* Drop band triggered, let the caller drop the 'skb'. */
>> + if (band->type == OVS_METER_BAND_TYPE_DROP)
>> + return true;
>> + }
>> +
>> + return false;
>> +}
>> +
