Re: [savi] Some findings in the darft-ietf-savi-dhcp-20

["Leaf Yeh" <[e-mail]>]

1. Guang - R2.2: ...Indeed, there is no problem if letting the Relay or Server B outside the perimeter. ..., in recent versions, we decide to include them into the perimeter.

Leaf - The above decision sounds you make that Relay is the for the connection with the upstream network.

Guang - Exactly. 

 

 

The basic questions here sounds : Do we need Server B/Relay work outside the perimeter? Per the draft-(ver.)21 of today, I guess the answer is 'no'. J

 

 

2. Guang - Trust: both data packet and control packet are trusted;

DHCP-Trust: only control packet is trusted.

Actually it is proper to configure Trust on the port of Relay and server B, if we also trust the data packet from them. In the scenario, the relay and server B are supposed to only send control packets, thus I think it makes no difference whether DHCP-Trust or trust is used. (Maybe we should specify this point in the doc?) 

However, I believe it is necessary to distinguish the two types of trusts.  

 

 

For the thinking of additional modular for the switch, DHCP-Trust attribute is designed for anti-bogus DHCP server/Relay. It is only used for the access of DHCP server/Relay, just because the default (No attribute) is blocking the messages from the Server/Relay on all ports (including validating ports) on the SAVI-switch, right?

 

Can we just use DHCP-Trust take care the DHCP (control) messages from the Server/Relay, and not care the data (plane) packet? I mean it might be enough for the defense against bogus DHCP server/Relay, right?

 

 

Guang - I think the perimeter is only a deployment suggestion. In practice it is not always possible to enforce savi as the diagram.  

 

 

Then I suggested to make the Fig.1 in the draft more popular for the real practice (or networking deployment). In my mind, the basic application scenario looks like the following:

 



 

 

Guang - For example, we may choose to configure DHCP-trust and validating on the same port, which is attached by a DHCP server and some hosts through a hub. In such cases, the DHCP server is actually outside the perimeter. The DHCP messages are not fully trustable but we have to make use of them for there are no other ways.

 

 

That sounds the DHCP server can be outside SAVI-perimeter and DHCP-Trust can be used on the perimeter with Validating port.

 

 

Guang - Besides, I think where to draw the perimeter is actually not a critical problem. It makes no difference whether placing the relay and server inside or outside the perimeter. The savi device works based on port configuration, without knowing the perimeter.

 

 

The concept of 'perimeter' is defined in section 4 of RFC7039 (SAVI arch.) and section 2.5 of RFC6620 (SAVI-SLAAC). It looks to me the Validating port of SAVI-Switch (for SLAAC) and the Validating attribute of the SAVI devices (for DHCP) forms the perimeter. In another word, the binding filters of SAVI form the perimeter, though draft-SAVI-DHCP adds more discussion 'On the Placement of DHCP Server/Relay' at section 4.3.3.

 

 

3. Guang - Besides, to confirm with other solutions, perimeter is the place to perform data packet filtering. Thus, only the attach points with validating attribute are on the perimeter.

...

Guang - It's my fault: the upstream port is also on the perimeter.

We choose not to fully trust the traffic from upstream network: both data packet and control packet. Thus, I think it is necessary to keep it outside the perimeter. 

However, if the operator chooses to fully trust packet from upstream networks, he can use a Trust on the corresponding port. Then in concept the upstream network is inside the perimeter.

 

 

That is really my concern:

a.  I don't think we really need SAVI switch (and its validating attributes) to form the perimeter against the upstream network in practice.

b .  If you don't trust the traffic from the upstream network, then the methods, such as ingress filtering or uRPF, other than SAVI might be employed.

c.  I guess we have not a clear definition for the upstream network. In my mind, the upstream network mention here sounds in the same administrative domain of the operator, which we could trust.

 

 

5. Leaf -4. Question of clarification on the text in section 4.3.2, <quote>(6) Optional: configure filters on the upstream links to filter out spoofing of local addresses from other networks.</quote>...

Guang - R4:...We cannot prevent other networks from spoofing each other. Thus, we can just requre local addresses (addresses assigned to local network) will not be spoofed by the other networks.

Leaf - Thanks for your explanation here, but that method sounds ingress filtering (or uRPF) to me. Meanwhile, the 'other network' in the definition of 'Upstream link' seems a little vague for me. I guess it is usually we don’t use SAVI-switch to connect to the upstream network, we use a gateway or router. So the (6) of SAVI-DHCP Perimeter Configuration Guideline seems unnecessary. I mean we don't usually use SAVI-switch to form the SAVI-perimeter against the upstream link or network.

Guang: ... We do not require the savi switch to connect to upstream network. Actually, they are connected to the gateway or router. This configuration just specifies how to process the traffic from the gateway or router.

 

 

The draft sounds only about SAVI-switch. The configuration guideline & state machines only work for switches (i.e. SAVI devices), right? 

 

 

6. Guang: ...For SAVI-DHCP, a critical problem is that the DHCP Server-Client messages should be trustable. However, there can be bogus DHCP servers in the upstream network. We should not make assumption that the upstream network is always trustable. Thus, they should not be in the perimeter.

 

 

I suppose the SAVI protection perimeter is not suitable against the upstream network, for at least 2 reasons:

a.  we don't use SAVI device to connect the upstream network;

b.  we don't use Validating attribute for the upstream network;

 

I guess your perimeter (SAVI-DHCP perimeter) here sounds not the same as SAVI protection perimeter defined in RFC7039 (SAVI arch.) and RFC6620 (SAVI-SLAAC). I can't find words in the above 2 RFCs on that we can't trust the data or even control packets from the upstream network, or the statement of ' We should not make assumption that the upstream network is always trustable '.

 

If you doubt there is a bogus DHCP server in the upstream network, you might employ something like 'DHCP-trust attribute' configure on the router (or gateway) to block DHCP server-client messages from the upstream network, then you could only use the local trustable DHCP server between the SAVI-perimeter and the router, or use DHCP relay to connect the remote trustable DHCP server. But the date (plane) packet need pass through, from the upstream network, access network, SAVI perimeter to the end-user hosts (i.e. DHCP clients).

 

But this sounds not a feature of SAVI-switch, and might be out-of-scope of the draft. I prefer 'DHCP-trust attribute' could only mention for the SAVI device in this draft.

 

 

8. Guang:'Creation Time' field only present in the stored table rather than BST?

 

 

I believe the solution in draft-SAVI-DHCP-21 can work, but the solution in RFC6620 sounds better. If one want to implement both SAVI-SLAAC & SAVI-DHCP, he might want the same binding table. Right?

 

 

9. Leaf- I guess the text in the draft will not cause any misunderstanding if you delete the prefix of 'EVE_'. Right?

Guang: Thank you for this comment. "EVE_" means it is an event instead of a packet. For example, not all " DHCP_OPTION_RC'" will trigger an event.

 

For example, you can’t find ‘EVE_’ for those messages which act as the event to trigger the state-change described in RFC6620, you still can understand it.

 

 

10. Leaf - In Section 7.1, <quote> Data packets without matching binding entry may trigger this process to set up bindings.</quote>

<quote>This process is not intended to set up a binding whenever a data packet without matched binding entry is received.</quote>

Does the above 2 sentence sound a little conflict?

Guang - R10: ... "Data packets without matching binding entry _may_ trigger this process to set up bindings." We use a "may" to mean this is probable.

Leaf - Can I say ' This process is intended to set up a binding whenever a data packet without matched binding entry is received.' ?

Guang: … Strictly, it is not quite proper. Not all the mismatched packet will trigger the process.

 

 

Can I say ‘This process may not intended to set up a binding whenever a data packet without matched binding entry is received.’ or ‘This process may intended to set up a binding whenever a data packet without matched binding entry is received.’? I am just not sure about the wording in ‘This process is not intended to set up a binding whenever a data packet without matched binding entry is received.’. Anyway, it makes me a little confused. In my mind, the process does intend to do something.

 

 

11. Leaf - …if you keep the text in section 7.5.2, <quote>The messages MUST NOT be sent to the attachment from which the triggering packet is received.</quote>. The DAD process after EVE_DATA_LEASEQUERY looks that 'The Neighbor Solicitation is only sent to the attachment which triggers the binding.' I personally prefer to combine these 2 DAD process, which means to sent DAD_NS to all the ports of SAVI-switch.

Guang: …As I can understand, you mean sending DAD to the attachment which triggers the binding after receiving the DHCP leasequery-reply? What is the purpose of this step? And I'm wondering how to combine the 2 DAD processes since they are performed at different stages? May you please give a further explanation?

 

 

Draft-(ver.)20 has the 2nd DAD process after EVE_DATA_LEASEQUERY (in the state of RECOVERY). Draft-(ver.)21 also have the 2nd ARP process for IPv4 address in same section 7.5.3.2. 

 

I guess the purpose of this process is double-check that one client (or host) assigned the address (specified in the Lease-query) with the valid lease is actively attached on the port which received a data without matched binding entry. 

 

I suppose the target address in the 2nd DAD process is the same as that in the 1st DAD process, which is the source address of the data packet received in the EVE_DATA_UNMATCH, right? If yes, that make possible to combine these 2 processes.

 

 

Best Regards,

Leaf

 

 

 

-----Original Message-----
From: Guang Yao [mailto:[e-mail]] 
Sent: Tuesday, April 01, 2014 3:23 PM
To: 'Leaf Yeh'; 'Jun Bi'; 'Jun Bi'
Cc: 'Ted Lemon'; [e-mail]
Subject: RE: Some findings in the darft-ietf-savi-dhcp-20

 

Dear Leaf,

 

Thank you very much for the quick review!

 

Our responses are as follows:

 

1. Leaf - As to Fig.1, question for clarification : It sounds to set up a filter (as the requirement of DHCP-Trust Attribute) inside the SAVI-perimeter. I guess if the Relay or Server B is directly connected to SAVI Device and outside the SAVI-perimeter, we still need to set up DHCP-Trust Attribute for the port of SAVI Device A and B, right? If yes, does it means the Fig.1 can let the Relay or Server B outside the SAVI-perimeter?

Guang - R2.2: ...Indeed, there is no problem if letting the Relay or Server B outside the perimeter. In fact, in the ealier versions, they are outside the perimeter. However, in recent versions, we decide to include them into the perimeter. 

 

The above decision sounds you make that Relay is the for the connection with the upstream network.

 

Guang:

Thank you for this comment.

Exactly. The strength of the binding based security is actually very weak. Especially, if the DHCP messages are untrustable, the bindings will be meaningless. Thus, We prefer the security mechanism enforced between DHCP relay and server, if we cannot completely trust traffic from upstream.

 

 

2. Guang - On one hand, it is used to distinguish the DHCP server A(which is outside the perimeter) and the DHCP server B/Relay. On the other hand, we think the perimeter should be used to separating the trusted area and the untrusted area. The Relay and Server B should be trusted, and they are actually the sources of trust. 

 

 

But you still need DHCP-Trust Attribute for them. It seem DHCP-Trust Attribute make them to be trust. How about if we just trust the Relay and Server B in the SAVI-perimeter without the additional permission, because they are connected to the Trust Port, (which was introduced by RFC662.)

 

Guang: 

Thank you for this comment.

 

Trust: both data packet and control packet are trusted;

DHCP-Trust: only control packet is trusted.

 

Actually it is proper to configure Trust on the port of Relay and server B, if we also trust the data packet from them. In the scenario, the relay and server B are supposed to only send control packets, thus I think it makes no difference whether DHCP-Trust or trust is used. (Maybe we should specify this point in the doc?)

 

However, I believe it is necessary to distinguish the two types of trusts.  I think the perimeter is only a deployment suggestion. In practice it is not always possible to enforce savi as the diagram.  For example, we may choose to configure DHCP-trust and validating on the same port, which is attached by a DHCP server and some hosts through a hub. In such cases, the DHCP server is actually outside the perimeter. The DHCP messages are not fully trustable but we have to make use of them for there are no other ways. 

 

Besides, I think where to draw the perimeter is actually not a critical problem. It makes no difference whether placing the relay and server inside or outside the perimeter. The savi device works based on port configuration, without knowing the perimeter.

 

 

3. Guang - Besides, to confirm with other solutions, perimeter is the place to perform data packet filtering. Thus, only the attach points with validating attribute are on the perimeter.

 

 

It make me not quite sure you want to configure Validating attribute for the upstream network, which looks outside of the SAVI-perimeter in Fig.1. 

 

Per the description in section 4 of RFC7039, <quote>The IP source addresses in packets passing the protection perimeter are validated by the ingress SAVI instance, but no further validation takes place as long as the packets remain within or leave the protection perimeter.</quote>, I suppose it is not necessary to let upstream network outside the SAVI-perimeter.

 

Guang:

Thank you for this comment.

It's my fault: the upstream port is also on the perimeter.

We choose not to fully trust the traffic from upstream network: both data packet and control packet. Thus, I think it is necessary to keep it outside the perimeter. 

However, if the operator chooses to fully trust packet from upstream networks, he can use a Trust on the corresponding port. Then in concept the upstream network is inside the perimeter.

 

 

4. Leaf - What does “SAVI device B is still protected from spoofing from client A” exactly mean? How can SAVI device B protects from spoofing from client A? It sounds a little confused.

Guang - R3:...It means, "SAVI device B can avoid receiving spoofing traffic from client A.".

 

 

The above statement sounds SAVI device A can avoid receiving spoofing traffic from client A, then 'SAVI device B can avoid receiving spoofing traffic from client A' because of the SAVI perimeter. I doubt the necessary text of 'SAVI device B is still protected from spoofing from client A' here.

 

Guang:

Thank you for this comment.

We will revise this sentence.

 

 

5. Leaf -4. Question of clarification on the text in section 4.3.2, <quote>(6) Optional: configure filters on the upstream links to filter out spoofing of local addresses from other networks.</quote>...

Guang - R4:...We cannot prevent other networks from spoofing each other. Thus, we can just requre local addresses (addresses assigned to local network) will not be spoofed by the other networks.

 

 

Thanks for your explanation here, but that method sounds ingress filtering (or uRPF) to me. Meanwhile, the 'other network' in the definition of 'Upstream link' seems a little vague for me. I guess it is usually we don’t use SAVI-switch to connect to the upstream network, we use a gateway or router. So the (6) of SAVI-DHCP Perimeter Configuration Guideline seems unnecessary. I mean we don't usually use SAVI-switch to form the SAVI-perimeter against the upstream link or network.

 

 

Guang:

Thank you for this comment.

We do not require the savi switch to connect to upstream network. Actually, they are connected to the gateway or router. This configuration just specifies how to process the traffic from the gateway or router.

 

 

6. Leaf - 5. Question of clarification on the text in the last paragraph of section 4.3.2, <quote> In this way, the points of attachments with Validating attribute (and generally together with attachments of upstream devices) on SAVI devices can form a perimeter separating DHCP clients and trusted devices.</quote> The above sentence sounds SAVI devices always form perimeter for upstream devices. But in general thinking, we always trust devices in the upstream network...

Guang - R5:...We should not trust devices in the upstream network by default unless some other securty mechanisms are enforced.  

 

 

Why not if the upstream network is in the SAVI-perimeter?

 

Guang:

Thank you for this comment.

For SAVI-DHCP, a critical problem is that the DHCP Server-Client messages should be trustable. However, there can be bogus DHCP servers in the upstream network. We should not make assumption that the upstream network is always trustable. Thus, they should not be in the perimeter.

 

 

7. Guang - R7.1: ...'A' and 'B' do not stand for any element in the diagram. They just denote some anchor of the clients.

 

 

How about replace A to be 'Port_1' in the instance of Fig.3?

 

Guang:

Thank you for this comment. This is an excellent advice. We will revise the doc accordingly.

 

 

8. Leaf - On the other hand, I guess the ‘Creation Time’ of the entry looks like a field in this table per the usage described in the section 9. For the reason of consistency, we also have the field of ‘Creation Time’ in the FCFS-SAVI DataBase (i.e. SAVI-SLAAC DB, section 3.1 of RFC6620).

Guang - R7.2:...It is a good suggestion. ... For the usage in section 9, the system time of the store operation is enough. There is no need to recording the "Creation Time" in binding set up. Besides, adding a new field in the table will introduce too many modifications...

 

 

Per the description in section 9, <quote> Binding entries MAY be saved into non-volatile storage whenever a new binding entry changes to BOUND state. ... The time when each binding entry is established is also saved.</quote>, I suppose 'Creation Time' could support this action.

 

Guang:

Thank you for this comment. Maybe can the  'Creation Time' field only present in the stored table rather than BST?

 

 

9. Leaf...though I don’t really like the prefix of ‘EVE_’ here. ☺ The prefix of ‘EVE_’ must stand for Event here, supposed it really means the additional valid check defined in section 6.3.2.

Guang - R8: ...We have changed ‘EVE_DHCP_SOLICIT_RC’ to ‘EVE_DHCP_OPTION_RC'. Yes, ‘EVE_’ stands for Event. We need some terms to denote events to be handled.

 

 

I guess the text in the draft will not cause any misunderstanding if you delete the prefix of 'EVE_'. Right?

 

Guang:

Thank you for this comment. "EVE_" means it is an event instead of a packet. For example, not all " DHCP_OPTION_RC'" will trigger an event.

 

 

10. Leaf - In Section 7.1, <quote> Data packets without matching binding entry may trigger this process to set up bindings.</quote> <quote>This process is not intended to set up a binding whenever a data packet without matched binding entry is received.</quote> Does the above 2 sentence sound a little conflict?

Guang - R10: ... "Data packets without matching binding entry _may_ trigger this process to set up bindings." We use a "may" to mean this is probable.

 

 

Can I say ' This process is intended to set up a binding whenever a data packet without matched binding entry is received.' ?

 

 

Guang:

Thank you for this comment. Strictly, it is not quite proper. Not all the mismatched packet will trigger the process.

 

 

11. 

Leaf -13. In Section 7.5.3.2, about ‘IPv6 address’ on page 32, <quote> Send a Neighbor Solicitation message with the target address set to the IP address in the corresponding entry. The Neighbor Solicitation is only sent to the attachment which triggers the binding. If there is no response after DAD_TIMEOUT, send another Neighbor Solicitation.

</quote> Could this additional DAD process be added in Fig.13?

Guang - R12: ... But it is there:...

 

 

I am not quite sure it is a good decision to delete this DAD process after EVE_DATA_LEASEQUERY (in the state of RECOVERY). It looks not the same as that DAD after EVE_DATA_UNMATCH (in the state of DETECTION), if you keep the text in section 7.5.2, <quote>The messages MUST NOT be sent to the attachment from which the triggering packet is received.</quote>. The DAD process after EVE_DATA_LEASEQUERY looks that 'The Neighbor Solicitation is only sent to the attachment which triggers the binding.' I personally prefer to combine these 2 DAD process, which means to sent DAD_NS to all the ports of SAVI-switch.

 

Guang:

Thank you for this comment.

As I can understand, you mean sending DAD to the attachment which triggers the binding after receiving the DHCP leasequery-reply? What is the purpose of this step?

And I'm wondering how to combine the 2 DAD processes since they are performed at different stages?

May you please give a further explanation? Thank you very much!

 

Best regards,

Guang