|
Re: [dhcwg] TSVDIR review of draft-ietf-dhc-dhcpv4-bulk-leasequery
2012-02-24 18:29:34
Hi, Kim,
On 2/23/2012 2:00 PM, Kim Kinnear wrote:
Joe,
I've edited the the email trail to try to concentrate on
the remaining open issues. I've also removed the two
issues that were included in the email I sent you an hour
or so ago, since I expect that they will be covered in
continuations of that thread.
To summarize, there are 3 issues remaining that are
not included below:
1. The NAGLE issue you raised, which has its own thread.
2. The two issues:
a) Interleaving
b) Possible security vulnerability if data
is available that wasn't already in the data
path.
3. The ? issues remaining, below. I think that
the number could be 0, but probably not.
Regards -- Kim
On Feb 17, 2012, at 5:47 PM, Joe Touch wrote:
Hi, Kim,
First, thank you for your detailed response to my quite lengthy review.
Some further clarifications and confirmations appear below.
Joe
On 2/17/2012 12:09 PM, Kim Kinnear wrote:
Joe,
Thank you for your review.
My responses are indented, below...
On Feb 13, 2012, at 5:00 PM, Joe Touch wrote:
Hi, all,
I've reviewed this document as part of the transport area
directorate's ongoing effort to review key IETF documents. These
comments were written primarily for the transport area directors,
but are copied to the document's authors for their information and
to allow them to address any issues raised. The authors should
consider this review together with any other last-call comments
they receive. Please always CC tsv-dir(_at_)ietf(_dot_)org if you reply to or
forward this review.
This request was received Feb. 2, 2012.
This document describes an extension to DHCPv4 for the bulk query
andtransfer of current lease status over TCP.
The following transport issues were noted, and are significant:
CONNECTION MANAGEMENT- The document describes the use of TCP
connections for bulk transfer, but needs to be more specific about which
side (relay client or server) closes the connection, under what
circumstances, and with what mechanism (e.g., graceful CLOSE vs. ABORT,
as per RFC 793)
Here is what we have now, 7.8 is for clients, 8.5 is for
servers:
7.8. Closing Connections [Client Section]
The requestor SHOULD close the connection after the
DHCPLEASEQUERYDONE message is received for the last outstanding query
that it has sent, if it has no more queries to send.
Under what conditions is it useful to not close at this time, and if so,
when/why is it eventually closed?
In this case it's also useful to explain what "no more queries to send" means -
are clients encouraged to keep the connection open in case there are more to send, or do
you expect that the client knows when it opens the connection whether it has more to send
- and if so, shouldn't it have already sent them?
These are not uninteresting questions, but they are over
the line from protocol specification and into
implementation in my view.
While I agree, were that true then the doc would not have *any* advice
on what to do in these cases. I'm suggesting that the advice should be
consistent. Omitting it all would be fine too.
I can certainly imagine that
the requestor has multiple queries to send but has chosen
to send them serially instead of in parallel (on the same
connection or on different connections). That would be
way that the requestor might know that it had more
queries to send. It might have a queue of them.
I will add this to the document:
"If a requestor doesn't know if it has more queries to
send or not, then it SHOULD close the connection."
That works.
8.5. Closing Connections [Server Section]
The DHCPv4 server SHOULD start a timer for BULK_LQ_DATA_TIMEOUT
seconds for a particular connection after it sends a
DHCPLEASEQUERYDONE message over that connection and if there is no
current query outstanding for that connection. It should clear this
Is this a SHOULD?
Yes, I'll change it.
And should the timer be cleared or restarted?
I was thinking of them as the same thing, but restarted
is probably better language. I'll change it.
AOK. "Cleared" could mean that there is no longer a timer running.
Under what conditions is it useful to not close at this time, and if so,
when/why is it closed?
All of these SHOULDs derive from the fundamental SHOULD
about starting a timer. You don't *have* to have a timer
to implement this protocol. You SHOULD use a timer, but
you don't *have* to. I can imagine other ways to dealing
with closing dead connections which don't involve timers.
No, I really don't want to put them into the document.
But the initial SHOULD says that if you don't use a
timer, you can still say that you implement this
capability.
IMO, you're using SHOULD above in non-2119 ways. SHOULD implies there's
a significant issue with either the protocol or implementation that has
exceptions. You're just saying "implementations will save resources by
using a timer and conserving the open connections". IMO, if that's the
goal, then it'd be preferable to make the recommendation that way,
rather than using 2119 language.
timer if a query arrives over that connection. If the timer expires,
the DHCPv4 server should close the connection.
Is this a SHOULD, as well?
Yes, thanks, I'll change it.
Again, under what conditions is it useful to not close at this time, and if so,
when/why is it closed?
Again, this is a SHOULD because you don't even have to
have a timer. But even if you have a timer, you don't
have to close it when the timer goes off. You might want
to be "smarter" about something, or use some external
information not represented here to make what you
consider a better decision. The SHOULD allows you to do
that without causing you to wonder if you really support
the document.
Again, this is good reason to reword the timer issue as "try to conserve
connections; try to reuse them for multiple requests from the same
relay, and try to close them as soon as all pending queries are
complete; the method for doing so is left to the implementer".
Otherwise IMO you'll have people deriving correctness from whether the
connection stays open or not.
The server MUST close its end of the TCP connection if it encounters
an error sending data on the connection. The server MUST close its
end of the TCP connection if it finds that it has to abort an in-
process request. A server aborting an in-process request SHOULD
attempt to signal that to its requestors by using the QueryTerminated
status code in the status-code option in a DHCPLEASEQUERYDONE
message, including a message string indicating details of the reason
for the abort. If the server detects that the requesting end of the
connection has been closed, the server MUST close its end of the
connection.
OK. I found it a bit confusing to have this in separate places, but I viewed
this from the transport perspective. This is fine.
All connections are closed with a graceful CLOSE, and not
with an ABORT. I'll add that to the document.
I'm at a loss to know how to be more specific about which
side of the connection closes the connection. There are
rules for the client in 7.8, and rules for the server in
8.5.
Somewhere you should state that the protocol exchange typically involves the
client closing the connection, and that the server closes only under exceptions.
Ok. I'll do that.
sec 7.3 indicates some conditions for terminating connections; this
section should indicate which side performs this, and by what
method(CLOSE, ABORT)
Section 7.3 discusses in two places where the "requestor" may
drop the connection. The term "requestor" is used to
indicate which "side" performs the actions. All of
Section 7 is about "Requestor Behavior", and the specific
details of section 7.3 in all cases indicate that the
instructions for dealing with the connection concern the
"requestor" explicitly.
I'll enhance the explanation to include that CLOSE should
be used.
If there is a specific set of instructions that are
unclear, I don't see them, so it would help me if you
could give me more detailed guidance than "Section 7.3".
Thanks.
It's probably sufficient to replace "requestor" with "client" to be consistent.
Requestor is the entity sending in the DHCPv4 Bulk Leasequery.
Then you mean relay?
Client is the entity that has a DHCPv4 address and whose
information shows up in the data part of the DHCPv4 Bulk
Leasequery requests. Client would not be consistent.
AOK. (see above; then use "relay" here).
However, this section then supercedes the description in 7.8 and 8.5 above -
i.e., you are handling exceptions on the client side as well.
I will enhance section 7.8 to indicate that requestors
have to handle exception conditions and their intended
actions.
AOK.
I.e., 7.8 needs to include the exception cases listed in 7.3, or they should
just be moved there.
Yes.
this sec also allows connections to stay open after all pending
transactions are complete (MAY); the rationale for this should be
given, or the connection MUST be closed.
I will add some words which clarify the reasons why you
might want to leave the connection open after all pending
transactions are complete.
FWIW, again this contradicts 7.8 which says the client SHOULD close after the
response is received if it has no more queries to send.
The rationale is that you might know from information not
part of this document that you had more queries to send
soon, even though you don't have them now.
See discussion above about rewording this all to "conserve resources
where possible", rather than making specific recommendations as to how
to do so.
This says "MAY" leave open. That's not the complement to SHOULD close.
We don't really care if you keep it open or not. Really.
If you think that you will be happier, keep it open. If
you think it is simpler to close it (as I do), then close
it. Different people (different authors even!) have different
ideas about how they would structure the code for this. That's
ok. The point about all of this open/closed/multiple-operations
stuff it to leave it up to the implementor. Period. We are
not trying to tell them exactly how to implement this. We
are trying to focus on the protocol aspects, and leave understood.
Yet further rationale to reword as above.
This section should more carefully explain behavior when a
connection is dropped and the reason - e.g., timeout, abort,
close.
The application employing Bulk Leasequery doesn't
care about *why* the connection went away. It
doesn't have to know if it was abort, close, or
timeout. I will say that in this section.
Hmm. I would think that this information would be passed up to the application.
It might, but we don't care. Do you think we should?
If the connection is gone, it is gone. I don't think
I care which syscall caused it
I think that this info should be passed up to the app, as per RFC 1122.
I do think that apps would do different things - log the error
differently at least, or retry in some cases but not others.
There were some other issues noted in this document. These
comments appear below, and although not focused on transport
issues, they represent significant issues that IMO should be
addressed as well.
NOTE - regarding some terminology issues, I did not survey current
DHCP RFCs for consistency, but IMO these terms should be corrected
even if they are then inconsistent with existing specs.
Joe
-----------
Major non-transport issues:
- "VPN" is used throughout to refer to "private" addresses (RFC 1918); a VPN is
not just private addressing.
Hmm. Actually, VPN in this document does not refer to RFC 1918
private addressing. If this is important to you, could you
be more specific as to your concern?
OK. I don't understand how it's being used, but if it's consistent with the
rest of DHCP, that's fine.
However, I got thrown off by your use of 10.x.x.x.
There are appropriate addresses to use as examples in RFCs - see RFC 5737, and
please change this example accordingly.
I will replace the reference to 10.x.x.x with a reference
to RFC 1918 addresses. Which is the point that I was trying
to make, not about some example address.
AOK.
- start-time-of-state is expressed as an offset from base-time;
this appears to be the only time indicated as an offset rather than
as an absolute. That inconsistency invites implementation errors;
IMO, this should be absolute too.
There are two kinds of time options here, input options
and output options. The input options are all absolute
times. The output options are all offset times from the
base time (which by design is the only absolute time in
the output packets).
That'd be very useful to explain this way. Is this typical for DHCP? It's quite
confusing IMO.
DHCP times are traditionally in duration from "now", to not have
to deal with clock skew issues (other than ongoing drift/skew).
We introduced the base-time to work with that, and we kept
the query times in absolute time. I'll add words to explain
that.
AOK.
...There are several input options
defined in this document. The start-time-of-state is the
only output option defined in this draft. It is defined
as a offset to the base time in large part because the
*other* output options that are used in the DHCP packet
are *also* offsets to the base time. These options are
not new, but are options that already exist in other
RFC's. So, while it looks like the state-time-of-state
is the odd man out at first glance, it actually is an
offset time because the *other* options with which it
will appear are also offset times.
- option lengths: throughout, the doc refers to option lengths as
being "an octet"; they are *indicated* in one octet. Some are constant
(e.g., DCHP-STATE), some allow the contents of the octet to vary. Again,
this is an *unsigned integer* octet.
Yes, the size of an option is stored in
one octet. So, yes, the length is *indicated*
by the value in that one octet.
But the length may be 1 octet, it may be
5 octets, it may vary between 8 and 10 octets.
I don't understand your concern. Are you
saying that the length should not be
stated to be "1 octet" or "5 octets"?
The length should be "indicated in 1 octet". I'm just suggesting a change to the way it's
described. When you say "the length is an octet", that is ambiguous and can mean that the
length value is 1, not just that this value is stored in one octet.
Ok, I'll fix that.
I will fix the integer to be specifically
unsigned (as previously indicated).
- some of the information provided (in DHCP-STATE) goes beyond
that of DHCP. It would be useful to motivate the need for this
information in a bulk query, and why it is not similarly available
for nodes using UDP (e.g., as an extension to DHCPv4, not just to
the bulk transfer command). again, absence of state information
should not indicate state. State should always be expressed
explicitly. these states are further meaningless without a state
diagram explaining them. if such a diagram is not possible (as
noted at end of sec 6.2.7), then the states are irrelevant and the
option should not be included.
The state transition diagram can certainly be drawn, but
as it is fully connected, such a picture conveys no
useful information.
It's sufficient to indicate this then.
... The transitions are not themselves
useful in any case, as the the Bulk Leasequery is a
snapshot of information at one instant of time. There is
every reason to assume that as the state moves from one
value to another that an external entity would not see
all (or even many) of these transitions. This is a
snapshot of the DHCP state, and could be nothing else in
the environment of Bulk Leasequery.
The reason why the state transition graph is fully
connected is due to the common practice in DHCPv4 of
having two servers running in a failover relationship.
There are then two state machines for the lease state,
and at times one "leads" and the other "follows", and at
other times the other "leads" and its partner "follows".
So, examined individually, the state machines for each
individual failover partner is fully connected because at
any time it may receive an update from its partner that
it decides to accept that will simply "set" the state to
the value supplied by the partner.
All of this is orthogonal to the Bulk Leasequery protocol
-- except that we want the user of the Bulk Leasequery
protocol to be able to make sense out of information
supplied by each of the failover partners. Or any
partners running any increased availability solution.
So, that is the reason why the dhcp-state option exists
at all -- to allow the user of Bulk Leasequery to make
sense out of the results of sending a Bulk Leasequery to
two different (but cooperating) DHCPv4 servers.
The DHC WG felt that this information was worthwhile to
have in this document, and I agree (or we wouldn't have
put it in).
Using the failover case as a justification would be helpful.
Failover isn't an IETF standard, and so we try to step
carefully around it possible existence. I don't think
I could get this document through back through the DHC WG
if I mentioned failover, so I don't think this is going
to work.
FWIW, I don't think that creating this level of detail about state
feedback achieves the goal; it codifies support for a non-standard
mechanism.
IMO, that issue deserves some light, not a shadowy corner. I'll leave it
to the WG to decide, though - this isn't a transport issue.
- sec 7.4 states that the clock skew of zero is desired; this
assumes E2E delays under 1sec. An explanation of why this is
desired should be given, as well as the consequences of it not.
I have given specific instructions on how to handle non-zero
clock skews. There are no "consequences". You code it
so that it works with non-zero clock skew. Period.
It is desired because it is less confusing to humans.
I would argue that this then irrelevant, and there should be no statement about
a skew of zero being preferred, since it could drive implementations that way
unnecessarily.
I believe that the other reason we like zero skew is
that it tends to expose fewer bugs in people's code,
so keeping the admonishment that it is a better
answer seems to add some value.
That's a good reason to avoid the admonishment, IMO. If the code breaks,
it should break early and often. "Expecting" a particular skew
encourages bad programming IMO.
Other non-transport issues:
- The document includes definitions and references to irrelevant
deployment and implementation issues, notably DSLAMs, concentrators, and
access concentrators. These should not appear formally; they should be
used only to usefully illustrate currently intended uses.
These appear in the text in order to give
the otherwise abstract design goals some concrete
examples. We already removed a lot of the design
goals, are you suggesting that they all be removed?
Actually, yes, except as examples in the intro/motivation. Elsewhere they are
IMO distracting.
DSLAM, concentrator, and access concentrator appear only
in Section 3, Design Goals -- which is precisely the intro/motivation
for this document. They don't appear (at least to my text search
engine) elsewhere. So I think we have done what you want.
Design goals is Sec 3; the intro is Sec 1. I'm saying using this
terminology in the intro (sec 1) is fine, but in the design goals (sec
3) is distracting.
- sec 3 should refer to relays, and returning the entire set or
individual bindings; there is no reason to explain the goals in terms of
access concentrators.
I don't know about this. I could certainly change it,
but is this really important to you? There was quite a
bit of negotiation in the WG about these examples. I
don't want to restart that discussion and have it all
over again. Does this really matter that much to you?
The examples should focus on client/server/relay terminology. If other example
terms are used, they should be parenthetical examples:
relay (e.g., at a concentrator)
The point is that this is relevant without the concentrator. A concentrator is
just one current example of where a relay exists, not the only one.
True, and I will say that. There was too much discussion
of these examples to remove them, but while they are
important to the process of getting through the DHC WG
review, they are not of significant importance technically.
So, I will add a sentence discussing the fact that a
concentrator is just an example.
AOK.
- sec 3.2 appears to provide contradictory advice - caching is required, but
should be avoided? it would be useful to resolve this inconsistency.
Use of Bulk Leasequery is designed to help you avoid
caching.
OK - it'd be useful to be clear on that in sec 3.2.
I will say that in Section 3.2.
AOK
- sec 3.3 refers to 'fast path', but this term doesn't make sense in this
context because fast-path is a forwarding issue. it would be useful to explain
what you mean, and pick a different term
Yes, it is a forwarding issue, and many devices
will not forward unless they know that they are forwarding
to a device that has received a valid DHCPv4 address.
The point (well, one point) of Bulk Leasequery is to
allow these forwarders to gather data so that they
can make the correct decision in the fast path.
Per-IP validation of that sort of permission isn't a fast-path decision.
Fast-path involves forwarding and policy decisions (which are not typically /32
decisions).
It would be useful to say this information is useful so such routers can make
the appropriate decision, but the path (fast or slow) is irrelevant IMO.
Yes, fast-path is a forwarding issue, and I believe
the intent of this section was to mention that the
information returned by Leasequery is useful when
deciding to forward a packet -- and that Bulk Leasequery
gives you the information you need to decide to forward
or not in advance of seeing the packet (when it is too
late to deal with it in the fast path).
The issue is that fast-path contrasts to slow-path. Filtering is always
fast-path where supported in the hardware, and slow-path where not.
The point of the bulk leasequery is to allow forwarders to filter.
*where* they filter is irrelevant, and not made slower or faster solely
by whether the bulk leasequery is there. "fast path" isn't relevant to
the issue.
I am not convinced that this should be changed.
Regards -- Kim
------------
Regards -- Kim
_______________________________________________
dhcwg mailing list
dhcwg(_at_)ietf(_dot_)org
https://www.ietf.org/mailman/listinfo/dhcwg
_______________________________________________
Ietf mailing list
Ietf(_at_)ietf(_dot_)org
https://www.ietf.org/mailman/listinfo/ietf
|
|