Network Working Group J. Moy Internet Draft Sycamore Networks, Inc. Expiration Date: July 2001 February 2001 File name: draft-ietf-ospf-hitless-restart-00.txt Hitless OSPF Restart draft-ietf-ospf-hitless-restart-00.txt Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC2026. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet- Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet- Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract This memo documents an enhancement to the OSPF routing protocol, whereby an OSPF router can stay on the forwarding path even as its OSPF software is restarted. This is called "hitless restart" or "non-stop forwarding". A restarting router may not be capable of adjusting its forwarding in a timely manner when the network topology changes. In order to avoid the possible resulting routing loops the procedure in this memo automatically terminates when such a topology change is detected. The restart procedure is also backward-compatible, reverting to standard OSPF processing when one or more of the restarting router's neighbors do not support the enhancements in this memo. Proper network operation during a hitless restart makes assumptions upon the operating environment of the restarting router; these assumptions are also documented. Moy [Page 1] Internet Draft Hitless OSPF Restart February 2001 Table of Contents 1 Overview ............................................... 2 2 Operation of restarting router ......................... 3 2.1 Entering hitless restart ............................... 3 2.2 Exiting hitless restart ................................ 5 3 Operation of helper neighbor ........................... 6 3.1 Entering helper mode ................................... 6 3.2 Exiting helper mode .................................... 7 4 Backward compatibility ................................. 7 5 Notes .................................................. 7 References ............................................. 8 A Grace-LSA format ....................................... 9 Security Considerations ............................... 10 Authors' Addresses .................................... 10 1. Overview Today many Internet routers implement a separation of control and forwarding functions. Certain processors are dedicated to control and management tasks such as OSPF routing, while other processors perform the data forwarding tasks. This separation creates the possibility of maintaining a router's data forwarding capability while the router's control software is restarted/reloaded. We call such a possibility "hitless restart" or "non-stop forwarding". The problem that the OSPF protocol presents to hitless restart is that, under normal operation, OSPF intentionally routes around a restarting router while it rebuilds its link-state database. OSPF avoids the restarting router to minimize the possibility of routing loops and/or black holes caused by lack of database synchronization. Avoidance is accomplished by have the router's neighbors reissue their LSAs, omitting links to the restarting router. However, if (a) the network topology remains stable and (b) the restarting router is able to keep its forwarding table(s) across the restart, it would be safe to keep the restarting router on the forwarding path. This memo documents an enhancement to OSPF that makes such hitless restart possible, and one that automatically reverts back to standard OSPF for safety when network topology changes are detected. In a nutshell, the OSPF enhancements for hitless restart are as follows. The router attempting a hitless restart originates link- local Opaque-LSAs, herein called Grace-LSAs, announcing the intention to perform a hitless restart, and asking for a "grace period". During the grace period its neighbors continue to announce the restarting router in their LSAs as if it were fully adjacent Moy [Page 2] Internet Draft Hitless OSPF Restart February 2001 (i.e., OSPF neighbor state Full), but only if the network topology remains static (i.e, the contents of the LSAs in the link-state database having LS types 1-5,7 remain unchanged; simple refreshes are allowed). There are two roles being played by OSPF routers during hitless restart. First there is the router that is being restarted. The operation of this router during hitless restart, including how the router enters and leaves hitless restart, is the subject of Section 2. Then there are the router's neighbors, which must cooperate in order for the restart to be hitless. During hitless restart we say that the neighbors are executing in "helper mode". Section 3 covers the responsibilities of a router executing in helper mode, including entering and leaving helper mode. 2. Operation of restarting router After the router restarts/reloads, it must change its OSPF processing somewhat until it re-establishes full adjacencies with all its previously fully-adjacent neighbors. This time period, between the restart/reload and the reestablishment of adjacencies, is called "hitless restart". During hitless restart: (1) The restarting router does not originate LSAs with LS types 1-5,7. Instead, the restarting router wants the other routers in the OSPF domain to calculate routes using the LSAs that it had originated prior to its restart, in order to maintain forwarding through the restart. (2) The restarting router doesn't run its OSPF routing calculations, instead using the forwarding table(s) that it had built prior to the restart. Otherwise, the restarting router operates the same as any other OSPF router. It discovers neighbors using OSPF's Hello protocol, elects Designated and Backup Designated Routers, performs the Database Exchange procedure to initially synchronize link-state databases with its neighbors, and maintains this synchronization through flooding. The processes of entering hitless restart, and of exiting hitless restart (either successfully or not) are covered in the following sections. 2.1. Entering hitless restart The router (call it Router X) is informed of the desire for its hitless restart when an appropriate command is issued by the Moy [Page 3] Internet Draft Hitless OSPF Restart February 2001 network operator. The network operator may also specify the length of the grace period, or the necessary grace period may be calculated by the router's OSPF software. In preparation for the hitless restart, Router X must perform the following actions before its software is restarted/reloaded. Note that common OSPF shutdown procedures are *not* performed, since we want the other OSPF routers to act as if Router X remains in continuous service. For example, Router X does not flush its locally originated LSAs, since we want them to remain in other routers' link-state databases throughout the restart period. (1) Router X must ensure that its forwarding table(s) is/are up-to-date and will remain in place across the restart. (2) Router X must resign any Designated Router (DR) or Backup Designated Router duties that it currently has. It does this by sending Hellos with Designated Router Priority set to 0. Resigning DR duties ensures that flooding works unimpeded across restarts, and that the DR/Backup will not change *after* the Grace-LSA is generated, which would be interpreted as a topology change and would terminate the hitless restart procedure prematurely. (3) The router must note in non-volatile storage the cryptographic sequence numbers being used for each interface. Otherwise it will take up to RouterDeadInterval seconds after the restart before it can start to reestablish its adjacencies, which would force the grace period to be lengthened severely. Router X then originates the grace-LSAs. These are link-local Opaque-LSAs (see Appendix A). Their LS Age field is set to 0, and the requested grace period (in seconds) is inserted into the body of the grace-LSA. A grace-LSA is originated for each of the router's OSPF interfaces. However, a grace-LSA need not be originated for an interface if either a) the interface has no fully adjacent neighbors or b) the interface is of type point- to-point and a grace-LSA has already been sent to the attached neighbor on another interface. If Router X wants to ensure that its neighbors receive the grace-LSAs, it should retransmit the grace-LSAs until they are acknowledged (i.e, perform standard OSPF reliable flooding of the grace-LSAs). If one or more fully adjacent neighbors do not receive grace-LSAs, they will more than likely cause premature termination of the hitless restart procedure (see Section 4). Moy [Page 4] Internet Draft Hitless OSPF Restart February 2001 After the grace-LSAs have been sent, the router should store the fact that it is performing hitless restart along with the length of the requested grace period in non-volatile storage. The OSPF software should then be restarted/reloaded, and when the reloaded software starts executing the hitless restart modifications in Section 2 above are followed. 2.2. Exiting hitless restart On exiting "hitless restart", the reloaded router reverts back to completely normal OSPF operation, reoriginating LSAs based on the router's current state and recalculating its forwarding table(s) based on the current contents of the link-state database. The router exits hitless restart when any of the following occurs: (1) Router X has reestablished all its adjacencies. Router X can determine this by building (but not installing or flooding) its router-LSA, based on the current router state, and comparing it to the router-LSA that it had last originated before the restart (called the "pre- restart router-LSA"). If the contents are the same, all adjacencies have been reestablished. (2) Router X receives an LSA that is inconsistent with its pre-restart router-LSA. For example, X receives a router- LSA originated by router Y that does not contain a link to X, even though X's pre-start router-LSA did contain a link to Y. This indicates that either a) Y does not support hitless restart, b) Y never received the grace- LSA or c) Y has terminated its helper mode for some reason (Section 3.2). (3) The grace period expires. (4) Router X gets a valid hitless restart request (grace-LSA) from another router. A router cannot both simultaneously attempt hitless restart and help a neighboring router undergo hitless restart, because the neighboring router must be monitoring the network state for changes throughout the entire restart period. When it exits hitless restart, the reloaded router should flush any grace-LSAs that it had originated. Moy [Page 5] Internet Draft Hitless OSPF Restart February 2001 3. Operation of helper neighbor As a "helper neighbor" for a router X undergoing hitless restart, router Y has two duties. It monitors the network for topology changes, and as long as there are none, continues to its advertise its LSAs as if X had remained in continuous OSPF operation. This means that Y's LSAs continue to list all adjacencies to X that were full (OSPF neighbor state Full) when the grace-LSA was first received, regardless of their current sycnchronization state. This logic affects the contents of both router-LSAs and network-LSAs, and also depends on the type of interface associated with the (possibly former) adjacency (see Sections 12.4.1.1 through 12.4.1.5 and Section 12.4.2 of [Ref1]). 3.1. Entering helper mode When a router Y receives a grace-LSA from router X, it enters helper mode for X as long as all the following checks pass: (1) There have been no changes in content to the link-state database (LS types 1-5,7) since the beginning of the grace period specified by the grace-LSA. The grace period began N seconds ago, where N is the current LS age of the grace-LSA. (2) The grace period has not yet expired. This means that the LS age of the grace-LSA is less than the grace period specified in the body of the grace-LSA (Appendix A). (3) Local policy allows Y to act as the helper for X. Examples of configured policies might be a) never act as helper, b) never allow the grace period to exceed a Time T, or c) never act as a helper for certain specific routers (specified by OSPF Router ID). Note that Router Y only needs to receive a single grace-LSA from X, even if X and Y attach to multiple common segments. The data in the first valid grace-LSA received is used to indicate the beginning and the end of the grace period -- all subsequent grace-LSAs received from X are ignored. This first grace-LSA is referred to below as simply "the grace-LSA from X". A single router is allowed to simultaneously serve as a helper for multiple restarting neighbors. Moy [Page 6] Internet Draft Hitless OSPF Restart February 2001 3.2. Exiting helper mode Router Y ceases to perform the helper function for its neighbor Router X when one of the following events occurs. (1) The grace-LSA originated by X is flushed. This is the successful termination of hitless restart. (2) The grace period expires. (3) Router Y receives an LSA with LS types 1-5,7 and whose contents have changed. This includes LSAs with no previous link-state database instance and the flushing of LSAs from the database, but excludes simple LSA refreshes. A change in LSA contents indicates a network topology change, which forces termination of a hitless restart. When router Y exits helper mode for X, Y reoriginates its LSAs based on the current state of its Router X adjacencies. 4. Backward compatibility Backward-compatibility with unmodified OSPF routers is an automatic consequence of the functionality documented above. If one or more neighbors of a router requesting hitless restart are unmodified, or if they do not received the grace-LSA, the hitless restart is prematurely aborted. The unmodified routers will start routing around the restarted router X as it performs initial database synchronization, by reissuing their LSAs with links to X omitted. These LSAs will be interpreted by helper neighbors as a topology change, and by X as an LSA inconsistency, in either case aborting hitless restart and resuming normal OSPF operation. 5. Notes Note the following details concerning the hitless OSPF restart mechanism described in this memo. o DoNotAge is never set in a grace-LSA, even if the grace-LSA is flooded over a demand circuit. This is because the grace-LSA's LS age field is used to calculate the extent of the grace period (see Appendix A). o Grace-LSAs have link-local scope because a) they only need to be seen by the router's direct neighbors and b) restricting them to Moy [Page 7] Internet Draft Hitless OSPF Restart February 2001 link-local scope makes it easy to detect the illegal configuration of two restarting routers being asked to help each other (Section 2.2). o It may be noted that the hitless restart mechanisms in this memo can also be used for unplanned outages. For example, after a crash of its control software, the router may come up and send grace-LSAs in an attempt to remain on the forwarding path while it regains its control state. This may not be a good idea, as it seems unlikely that such a router could guarantee the sanity of its forwarding table(s). However, if the router does attempt a hitless restart from an unplanned outage, it should at the least (a) allow the network operator to turn this feature off and (b) attempt to determine when its forwarding tables were last updated, setting the beginning of the grace period accordingly (this means originating the grace-LSA with LS age equal to the time that the forwarding tables were last updated). References [Ref1] Moy, J., "OSPF Version 2", RFC 2328, April 1998. [Ref2] Coltun, R., "The OSPF Opaque LSA Option", RFC 2370, July 1998. [Ref3] Murphy, S., M. Badger and B. Wellington, "OSPF with Digital Signatures", RFC 2154, June 1997. Moy [Page 8] Internet Draft Hitless OSPF Restart February 2001 A. Grace-LSA format The grace-LSA is a link-local scoped Opaque-LSA [Ref2] having Opaque Type of TBD1 and Opaque ID equal to TBD2. The grace-LSA is originated by a router that wishes to execute a hitless restart of its OSPF software. The grace-LSA requests that the router's neighbors aid it in its hitless restart by continuing to advertise the router as fully adjacent during a specified grace period. It is assumed that the grace-LSA has LS age field set to 0 when the LSA is first originated; the current value of LS age then indicates how long ago the restarting router made its request. The body of the LSA contains the length of the grace period in seconds. 0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | LS age | Options | 9 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Opaque Type | Opaque ID | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Advertising Router | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | LS sequence number | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | LS checksum | length | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Grace Period | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ Grace Period The number of seconds that the router's neighbors should continue to advertise the router as fully adjacent, regardless of the the state of database synchronization between the router and its neighbors. Since this time period began when grace-LSA's LS age was equal to 0, the grace period terminates when either a) the LS age of the grace-LSA exceeds the value of Grace Period or b) the grace-LSA is flushed. See Section 3.2 for other conditions which terminate the grace period. Moy [Page 9] Internet Draft Hitless OSPF Restart February 2001 Security Considerations One of the ways to attack a link-state protocol such as OSPF is to inject false LSAs into, or corrupt existing LSAs in, the link-state database. Injecting a false grace-LSA would allow an attacker to spoof a router that, in reality, has been withdrawn from service. The standard way to prevent such corruption of the link-state database is to secure OSPF protocol exchanges using the Crytographic authentication specified in [Ref1]. An even stronger way of securing link-state database contents has been proposed in [Ref3]. Authors' Addresses J. Moy Sycamore Networks, Inc. 150 Apollo Drive Chelmsford, MA 01824 Phone: (978) 367-2505 Fax: (978) 256-4203 email: jmoy@sycamorenet.com Moy [Page 10]