idnits 2.17.1 draft-sriram-replay-protection-design-discussion-11.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (October 18, 2018) is 2010 days in the past. Is this intentional? -- Found something which looks like a code comment -- if you have code sections in the document, please surround them with '' and '' lines. Checking references for intended status: Informational ---------------------------------------------------------------------------- No issues found here. Summary: 0 errors (**), 0 flaws (~~), 1 warning (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Secure Inter-Domain Routing K. Sriram 3 Internet-Draft D. Montgomery 4 Intended status: Informational US NIST 5 Expires: April 21, 2019 October 18, 2018 7 Design Discussion and Comparison of Protection Mechanisms for Replay 8 Attack and Withdrawal Suppression in BGPsec 9 draft-sriram-replay-protection-design-discussion-11 11 Abstract 13 In the context of BGPsec, a withdrawal suppression occurs when an 14 adversary AS suppresses a prefix withdrawal with the intension of 15 continuing to attract traffic for that prefix based on a previous 16 (signed and valid) BGPsec announcement that was earlier propagated. 17 Subsequently if the adversary AS had a BGPsec session reset with a 18 neighboring BGPsec speaker and when the session is restored, the AS 19 replays said previous BGPsec announcement (even though it was 20 withdrawn), then such a replay action is called a replay attack. The 21 BGPsec protocol should incorporate a method for protection from 22 Replay Attack and Withdrawal Suppression (RAWS), at least to control 23 the window of exposure. This informational document provides design 24 discussion and comparison of multiple alternative RAWS protection 25 mechanisms weighing their pros and cons. This is meant to be a 26 companion document to the standards track draft-ietf-sidrops-bgpsec- 27 rollover that will specify a method to be used with BGPsec for RAWS 28 protection. 30 Status of This Memo 32 This Internet-Draft is submitted in full conformance with the 33 provisions of BCP 78 and BCP 79. 35 Internet-Drafts are working documents of the Internet Engineering 36 Task Force (IETF). Note that other groups may also distribute 37 working documents as Internet-Drafts. The list of current Internet- 38 Drafts is at https://datatracker.ietf.org/drafts/current/. 40 Internet-Drafts are draft documents valid for a maximum of six months 41 and may be updated, replaced, or obsoleted by other documents at any 42 time. It is inappropriate to use Internet-Drafts as reference 43 material or to cite them other than as "work in progress." 45 This Internet-Draft will expire on April 21, 2019. 47 Copyright Notice 49 Copyright (c) 2018 IETF Trust and the persons identified as the 50 document authors. All rights reserved. 52 This document is subject to BCP 78 and the IETF Trust's Legal 53 Provisions Relating to IETF Documents 54 (https://trustee.ietf.org/license-info) in effect on the date of 55 publication of this document. Please review these documents 56 carefully, as they describe your rights and restrictions with respect 57 to this document. Code Components extracted from this document must 58 include Simplified BSD License text as described in Section 4.e of 59 the Trust Legal Provisions and are provided without warranty as 60 described in the Simplified BSD License. 62 Table of Contents 64 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 65 2. Description and Scenarios of Replay Attacks and Withdrawal 66 Suppression . . . . . . . . . . . . . . . . . . . . . . . . . 3 67 3. Classification of Solutions . . . . . . . . . . . . . . . . . 4 68 4. Expiration Time Method . . . . . . . . . . . . . . . . . . . 5 69 5. Key Rollover Method . . . . . . . . . . . . . . . . . . . . . 6 70 5.1. Periodic Key Rollover Method . . . . . . . . . . . . . . 7 71 5.2. Event-driven Key Rollover Method . . . . . . . . . . . . 9 72 5.2.1. EKR-A: EKR where Update Expiry is Enforced by CRL . . 10 73 5.2.2. EKR-B: EKR where Update Expiry is Enforced by 74 NotAfter Time . . . . . . . . . . . . . . . . . . . . 11 75 5.2.3. EKR with Separate Key for Each Incoming-Outgoing 76 Peering-Pair . . . . . . . . . . . . . . . . . . . . 12 77 6. Summary of Pros and Cons . . . . . . . . . . . . . . . . . . 13 78 7. Summary and Conclusions . . . . . . . . . . . . . . . . . . . 15 79 8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 16 80 9. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 16 81 10. Security Considerations . . . . . . . . . . . . . . . . . . . 16 82 11. Informative References . . . . . . . . . . . . . . . . . . . 16 83 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 17 85 1. Introduction 87 In BGP or BGPsec, prefix or route withdrawals happen, and a 88 withdrawal can be explicit (i.e. route simply withdrawn) or implicit 89 (i.e. a new route announcement replaces the previous). In the 90 context of BGPsec, a withdrawal suppression occurs when an adversary 91 AS suppresses a prefix withdrawal with the intension of continuing to 92 attract traffic for that prefix based on a previous (signed and 93 valid) BGPsec announcement that was earlier propagated. Subsequently 94 if the adversary AS has a BGPsec session reset with a neighboring 95 BGPsec speaker and when the session is restored, the AS replays said 96 previous BGPsec announcement (even though it was withdrawn), then 97 such a replay action is called a replay attack. A method for 98 protection against Replay Attack and Withdrawal Suppression (RAWS) 99 can be used along with the BGPsec protocol [RFC8205], at least to 100 control the window of exposure to such attacks (see Sections 4.3, 4.4 101 of [RFC7353]). 103 In this informational document, we provide design discussion and 104 comparison of various RAWS protection mechanisms that may be used in 105 conjunction with the BGPsec protocol. This is meant to be a 106 companion document to the standards track document 107 [I-D.ietf-sidrops-bgpsec-rollover] that will specify a method to be 108 used with BGPsec for RAWS protection. Here we consider four 109 alternative mechanisms - one based on the explicit Expiration Time 110 approach and three variants based on the Key Rollover approach. We 111 provide a detailed comparison among these mechanisms, weighing their 112 pros and cons. This document is meant to help inform the decision 113 process leading to an exact description for the mechanism to be 114 finalized and formally specified in 115 [I-D.ietf-sidrops-bgpsec-rollover]. 117 2. Description and Scenarios of Replay Attacks and Withdrawal 118 Suppression 120 The following are examples of various forms of replay attack and 121 withdrawal suppression (RAWS): 123 Example 1: AS1 has AS2 and AS3 as eBGPsec peers. At time x, AS1 had 124 announced a prefix (P) to AS2 and AS3. At a later time (x+d), AS1 125 sends a Withdraw for prefix P to AS2. AS2 suppresses the Withdraw 126 (does not send to its peers any explicit or implicit Withdraw). AS2 127 continues to attract some of the data for prefix P by pretending to 128 still have a valid (signed) route for P. In effect, AS2 can conduct 129 a Denial of Service (DOS) attack on a server located at prefix P. 130 (See slide #15 in [RAWS-discussion] for an illustration.) 132 Example 2: AS1 has AS2 and AS3 as eBGPsec peers. AS2 and AS3 are 133 also eBGPsec peers. At time x, AS1 announced a prefix P to AS2 and 134 AS3. AS3 also propagates to AS2 its route (via AS1) for prefix P. 135 At a later time (x+d), AS1 discontinues its peering with AS2. AS2 136 should propagate an alternate longer path via AS3 for prefix P and 137 thus implicitly withdraw the route via AS1. However, AS2 suppresses 138 it. AS2 can thus make some traffic destined for prefix P to flow via 139 itself. This enables AS2 to eavesdrop on the data but not cause a 140 DOS attack. AS2 may also choose to DoS attack hosts in prefix P. 141 (See slide #16 in [RAWS-discussion] for an illustration.) 142 Example 3: AS1 has AS2 and AS3 as eBGPsec peers. AS2 and AS3 are 143 also eBGPsec peers. At time x, AS1 announced a prefix P to AS2 144 without prepending (Update: AS1{pCount=1} P) but announced the same 145 prefix to AS3 with prepending (Update: AS1{pCount=2} P). Thus AS1 146 had preferred its ingress data traffic for prefix P to come in via 147 AS2. At a later time (x+d), AS1 switches ingress data path 148 preference to AS3 over AS2 - announcing prefix P to AS3 without 149 prepending (Update: AS1{pCount=1} P) and to AS2 with prepending 150 (Update: AS1{pCount=2} P). AS2 suppresses the new prepended path 151 announcement (does not send to its peers any new update about P). 152 Thus AS2 continues to attract more of AS1's ingress data traffic and 153 generates more revenue for itself at the expense of AS1. (See slide 154 #17 in [RAWS-discussion] for an illustration.) 156 As illustrated above, the mechanisms and motivations for RAWS may 157 differ. 159 In the context of the examples mentioned above, a requirement for 160 RAWS protection can be stated as follows. An update that AS1 sends 161 to AS2 at time x should expire at time x+w. This capability would 162 allow other ASes to detect actions by AS2 to suppress the Withdraw or 163 replay the update from AS1 for prefix P after time x+w. This limits 164 the RAWS vulnerability window. (Note: If no peering or policy change 165 affecting prefix P occurs during the vulnerability window, then a 166 typical solution would include a method for extending the validity 167 period of the route(s) beyond x+w.) We will later discuss what a 168 reasonable window size, w, should be. 170 The obvious downside of any mechanism that support this capability is 171 that it will require AS1 to send a new update before time x+w, and 172 this update will need to propagate via all the paths that the 173 original update traversed. Thus more update traffic will result than 174 if the RAWS protection mechanism were not employed, and this traffic 175 will require cryptographic processing by all of the routers along the 176 paths. Thus the creation of a mechanism to counter RAWS attacks 177 potentially introduces a new opportunity for DoS attacks against 178 eBGPsec routers. 180 3. Classification of Solutions 182 Mechanisms for RAWS protection can be classified into two broad 183 categories as follows: 185 o Expiration Time (ET) Method: This method uses an explicit 186 Expiration Time field in the BGPsec update. (Note: Explicit 187 Expire Time field was included in an earlier version of the BGPsec 188 protocol specification [draft-ietf-sidr-bgpsec-protocol-01].) 190 o Key Rollover (KR) Method: In this method, the update expiration is 191 enforced by a key rollover. Router transitions to a new 192 certificate with a new pair of keys, and the previous router 193 certificate either expires or is revoked. 195 The Key Rollover method can be further characterized into the 196 following sub categories: 198 o Periodic Key Rollover (PKR): Key rollovers happen at periodic 199 intervals. 201 o Event-driven Key Rollover (EKR): Key rollovers happen only when 202 peering or policy change events occur. 204 * EKR-A: EKR where expiry of previous update is enforced by CRL. 206 * EKR-B: EKR where expiry of previous update is controlled by 207 NotAfter time (router certificate is not revoked at the time 208 when the event happens). 210 In Section 4, Section 5, and Section 6 we describe the various 211 methods listed above, and discuss their pros and cons. 213 4. Expiration Time Method 215 The details of the Expiration Time (ET) method are as follow: 217 o Explicit Expiration Time is used for origin's signature. 219 o Expiration Time field is required in the BGPsec update. 221 o Periodic re-origination (beaconing) of prefixes is performed by 222 origin ASes. The value in the ET field in the update is extended 223 at beaconing time, and thereby the update is refreshed. Every 224 prefix in the Internet is re-originated and propagates through the 225 Internet once every 'beacon' interval. 227 o These beacons are distributed actions by prefix owners and are 228 intended to be jittered in time to reduce burstiness. The beacon 229 interval can be different at each originating AS. 231 o Beacon interval granularity: TBD but preferably in fairly granular 232 units (days). It is important to limit the ability of each AS to 233 specify a short beacon interval, to prevent an AS from using this 234 mechanism to cause BGPsec to thrash. 236 Discussion of Pros and Cons: 238 Pro: This method is easy on transit routers. In the event of peering 239 or policy change, BGPsec with the ET method behaves the same way as 240 BGP-4 in terms of which prefix routes are propagated. That is, the 241 router re-evaluates best paths factoring in peering or policy 242 changes, and propagates only those prefix routes that have a change 243 in best path. In other words, there is no necessity for a transit 244 BGPsec router to re-propagate and refresh prefixes on all peering 245 links. This is because prefix updates are refreshed anyway once 246 every beacon interval by all prefix originators. There is low 247 steady-state traffic associated with beaconing (see Figure on slide 248 #8 in [RAWS-discussion]), but there are no huge bursts or spikes in 249 workload due to peering or policy change events at transit routers. 251 Con: Equipment vendor can potentially facilitate unnecessary frequent 252 beaconing if ISP urges and pays (dollar attack!). This possibility 253 is mitigated by having a well thought-out granularity for ET, for 254 example, setting the unit for advertising ET to one day (rather than 255 one minute). 257 Con: A change in on-the-wire BGPsec protocol would be needed in case 258 the unit of the ET field (granularity) needs to be changed. 260 5. Key Rollover Method 262 Key Rollover (KR) method has three variations as outlined in 263 Section 3. Those will be discussed later in this section. The 264 following features are common to all variants of the KR method: 266 o In the KR method, it is best if the BGPsec router has two pairs of 267 certificates as follows: A pair of origination certificates 268 (current and next) for signing prefixes being originated by the AS 269 of the router, and a pair of transit certificates (current and 270 next) for signing transit prefixes. 272 o Note: If a BGPsec router only originates prefixes (i.e. has no 273 transit prefixes), then it needs to maintain only a pair of 274 origination certificates and need not maintain the extra pair of 275 transit certificates. (This would be the case for the vast 276 majority of ASes, since most are stubs.) 278 o The three KR methods differ in how the rollover of certificates 279 (or keys) is done: 281 * Certificate rollovers are Periodic vs. Event-driven. 283 * In the Event-driven method, the expiry of old update is (A) 284 Enforced by CRL vs. (B) Controlled by NotAfter time. 286 * In (A), certificate's NotAfter field is set to a very large 287 value and CRL is issued to revoke the certificate when 288 necessary. In (B), NotAfter field set to a permissible 289 vulnerability window time, and CRL to revoke certificate is not 290 required. 292 Discussion of Pros and Cons (common to all Key Rollover methods): 294 Pro: The KR method functions by manipulating the RPKI objects 295 (certificates, keys, NotAfter field in certificate, etc.) to refresh 296 updates or to cause expiry of previously propagated updates. Unlike 297 the ET method, it does not rely on any explicit field in the update. 298 Hence, an advantage of the KR method over the ET method is that in 299 case any parameters need to change or if the method itself is 300 modified, then there is no impact on the BGPsec protocol on the wire. 302 Con: The KR method increases the number of objects in the RPKI 303 repository system, by requiring at least two certificates for every 304 transit AS. It also introduces additional churn in the global RPKI 305 as these certificates expire (or are revoked) and are replaced. 307 Con: There is also added update churn. The amount of update churn 308 varies depending on the type of KR method used (see Section 5.1 and 309 Section 5.2). 311 We will now describe and discuss in detail the variants of the KR 312 method. 314 5.1. Periodic Key Rollover Method 316 The details of the Periodic Key Rollover (PKR) method are as follow. 318 o Router's origination certificate's NotAfter time is used 319 effectively as expiration time for origin's signature. 321 o Each origination router re-originates (i.e. beacons) before 322 NotAfter time of the current origination certificate. Beaconing 323 is periodic re-origination of prefixes by origin ASes. 325 o At beaconing time, the next origination certificate becomes the 326 new current certificate, and the new update is signed with the 327 private key of this new current certificate and re-originated. 329 o A new 'next' origination certificate is created and propagated at 330 or before beaconing time. This can also be done with a good lead 331 time. In practice, multiple 'next' certificates for each router 332 could be propagated and kept in the in the RPKI repositories. 334 They must have contiguous or slightly overlapping validity 335 periods. 337 o Every prefix in the Internet is re-originated and propagates 338 through the Internet once every 'beacon' interval. 340 o The re-originations or beacons are distributed actions by prefix 341 owners and jittered in time by design to reduce burstiness. The 342 beacon interval can be different at different originating ASes. 344 o Beacon (or re-origination) interval granularity: TBD but 345 preferably in fairly granular units (days). 347 o Transit certificates can have large NotAfter time (e.g., whatever 348 duration is required normally for key maintenance). 350 o When a peering or policy change event occurs at a transit router, 351 the router does not perform any reactive key rollover. The router 352 re-evaluates best paths factoring in peering or policy changes, 353 and propagates only those prefix routes that have a change in best 354 path (similar to BGP-4). There is no necessity for the BGPsec 355 router to re-propagate and refresh prefixes on all peering links. 356 This is because prefix updates are refreshed anyway once every re- 357 origination (i.e. beaconing) interval by all prefix originators. 359 Discussion of Pros and Cons: 361 Several of the same pros/cons of the Expiration Time method also 362 apply here for the PKR method. 364 Pro: The main pro for the PKR method is the same as that for the 365 Expiration Time (ET) method. That is, being easy on transit routers 366 as discussed in Section 4. Just as in the ET method, there is low 367 steady-state traffic associated with periodic re-originations (i.e. 368 beaconing) (see Figure on slide #8 in [RAWS-discussion]), but there 369 are no huge bursts or spikes in workload due to peering or policy 370 change events at transit routers. (See comparisons with the EKR 371 methods in Section 5.2.) 373 Pro: The common pro discussed previously for all KR methods, namely, 374 not requiring change of protocol on the wire when a parameter change 375 occurs (e.g., change of beacon interval units) is naturally 376 applicable here. 378 Con: Churn in the RPKI is of concern. Every BGPsec router renews and 379 propagates its 'next' origination certificate once in every beacon 380 (i.e. re-origination) interval. 382 5.2. Event-driven Key Rollover Method 384 The common details of the Event-driven Key Rollover (EKR) methods are 385 as follow. 387 o Key rollover is reactive to events (not periodic). 389 o If a peering or policy change event involves only prefixes being 390 originated at the AS of the router, then the router rolls only the 391 origination key. 393 o If a peering change event involves transit prefixes at the AS of 394 the router, then the router rolls its transit key as well as the 395 origination key. Both keys are rolled because any peering 396 relationship change also requires refresh of prefixes originated 397 by the router. 399 o If a key rollover takes place, then a corresponding (origination 400 or transit) new 'next' certificate is propagated in RPKI. 402 Discussion of Pros and Cons: 404 Pro: As long as no triggering events occur, there is no added update 405 churn in BGPsec. 407 Con: Whenever the transit key is rolled, there is a storm of BGPsec 408 updates at routers in transit ASes. For example, consider BGPsec 409 capable transit AS5 that is connected to four BGPsec non-stub 410 customers (AS1, AS2, AS3, AS4). Assume each AS has a single BGPsec 411 router in it. AS1 through AS4 each receives almost full table 412 (approximately 600K signed prefix updates) from AS5. Assume also 413 that AS1 and its customers together originate 100 prefixes in total; 414 likewise for AS2, AS3 and AS4. Now consider that an event occurs 415 whereby the peering between AS1 and AS5 is discontinued. As a result 416 of this event, in the EKR method, the AS5 router signs and re- 417 propagates approximately 3x600K = 1.8 Million signed prefix updates 418 to AS2, AS3 and AS4 combined. In addition, it also sends 4x100 = 400 419 Withdraws, which are negligible. In comparison, in the PKR method, 420 reacting to the same event, the BGPsec router at AS5 sends only 4x100 421 = 400 Withdraws and signs/re-propagates ZERO prefix updates. (An 422 illustration can be found in slide #9 in [RAWS-discussion]. Also, 423 additional peering change scenarios and quantitative comparisons can 424 be found in slides #10 and #11 in [RAWS-discussion].) 426 It remains to be seen through measurement and modeling how the impact 427 of such large bursts of workload in the EKR method at the time of 428 event occurrence can be managed in route processors, e.g., by 429 jittering and throttling the workload. 431 5.2.1. EKR-A: EKR where Update Expiry is Enforced by CRL 433 EKR-A builds on the common principles as described for EKR above in 434 Section 5.2. The additional details of EKR-A operation are as 435 follow: 437 o NotAfter time of origination and transit certificates is set to a 438 large value (e.g., one year or whatever period needed for normal 439 key maintenance). 441 o Whenever key rollover (for origination or transit) occurs, then a 442 CRL is propagated for the certificate that was used until that 443 time. So the old update expires (due to invalid state) only when 444 the CRL propagates and reaches each relying router. 446 o This method relies on end-to-end CRL propagation through the RPKI 447 system to enforce expiry of a previous update whenever the need 448 arises. 450 o The CRL either propagates all the way to the relying router, or 451 the RPKI cache server of the router receives the CRL and then 452 sends a withdrawal of the {AS, SKI, Pub Key} tuple to the router. 453 Either way, the CRL must in effect propagate all the way to the 454 relying router. 456 o Thus the attack vulnerability window with the EKR-A method is 457 governed by the end-to-end CRL propagation time. 459 Discussion of Pros and Cons: 461 The following pro and con for the EKR-A method are in addition to the 462 common pros and cons listed above for the KR and EKR methods 463 (Section 5 and Section 5.2). 465 Pro: EKR-A has much less RPKI churn than PKR or EKR-B (see 466 Section 5.2.2). 468 Con: Router needs to receive a CRL or a withdraw of {AS, SKI, Pub 469 Key} tuple in order to know an update has expired. Hence, the RAWS 470 vulnerability window is determined by the CRL propagation time which 471 can vary widely from one relying router to another router that may be 472 in different regions. It is anticipated that this would be no worse 473 than 24 hours, but needs to be confirmed by measurements in an 474 operational or emulated RPKI systems [rpki-delay]. 476 5.2.2. EKR-B: EKR where Update Expiry is Enforced by NotAfter Time 478 EKR-B builds on the common principles as described for EKR above in 479 Section 5.2. The additional details of EKR-B operation are as 480 follow: 482 o NotAfter time of current origination and transit certificates is 483 set to a value determined by the desired vulnerability window 484 (~day). 486 o Update expiry is controlled by NotAfter time (router certificate 487 is not revoked at the time when the event happens). 489 o If no triggering event occurs to cause origination key rollover 490 within a pre-set time (NotAfter), then new origination (current 491 and next) certificates are issued only to extend the NotAfter time 492 but the corresponding key pairs and SKIs remain unchanged. 494 o Do likewise (i.e. similar to what the above bullet says) for the 495 transit (current and next) certificates and keys. 497 o A previous update automatically becomes invalid at the earliest 498 NotAfter time of the certificates used in the signatures unless 499 each of those certificates' NotAfter time has been extended. 501 o Changes in certificates to extend their NotAfter time need not 502 propagate end-to-end (all the way to the relying routers); they 503 may propagate only up to the RPKI cache server of the relying 504 router. RPKI cache server would send a withdraw for an {AS, SKI, 505 Pub Key} tuple to a relying router if the NotAfter time of the 506 certificate has passed. 508 o Changes in certificates to advance NotAfter time can be scheduled 509 and propagated (in RPKI) reasonably well in advance. 511 Discussion of Pros and Cons: 513 The following pro and con for EKR-B are in addition to the common 514 pros and cons listed above for the KR and EKR methods (Section 5 and 515 Section 5.2). 517 Pro: Update expiration is automatic in case the NotAfter time of any 518 of the certificates used to validate the update has not been 519 extended. So the RAWS vulnerability window is predictable and not 520 influenced by the RPKI end-to-end propagation time. 522 Pro: Routers do not get any RPKI updates from the RPKI cache server 523 when a certificate changes but the corresponding key pair and SKI 524 remain unchanged. Routers do not receive NotAfter time from their 525 RPKI cache server. There is no need for it. Instead, the RPKI cache 526 server keeps track of NotAfter time, and provides to routers only 527 valid {AS, SKI, Pub Key} tuples. This saves some RPKI state 528 maintenance workload at the routers. 530 Con: EKR-B has much more RPKI churn than EKR-A because both 531 origination and transit certificates need to be reissued periodically 532 to extend their validity time (even in the absence of any peering or 533 policy change events). 535 5.2.3. EKR with Separate Key for Each Incoming-Outgoing Peering-Pair 537 This is a place holder section where we mention another variant of 538 the EKR method. This idea has not been considered or vetted by the 539 SIDR WG yet. So we only mention it here briefly. 541 As noted earlier, the EKR methods considered so far generate a huge 542 spike in workload whenever the transit key rollover takes place. One 543 way to reduce that workload is to have a separate signing key for 544 each incoming-outgoing peering pair. For example, consider a BGPsec 545 router in AS4 that has peers in AS1, AS2, and AS3. The router will 546 hold six signing keys, one each corresponding to (AS1, AS2), (AS2, 547 AS1), (AS1, AS3), (AS3, AS1), (AS2, AS3), and (AS3, AS2) peering- 548 pairs. Note that the directionality of peering is included here and 549 is necessary. The key corresponding to (AS-i, AS-j) would only be 550 used to sign updates received from AS-i and being forwarded to AS-j. 551 In the general case, when the BGPsec router has n peers, the number 552 of transit keys will be n(n-1). Since there would be a Current and a 553 Next key (for rollover), the number of transit keys held in the 554 router for signing will be actually 2n(n-1). When a peering or 555 policy change occurs, the router would rollover only those specific 556 keys that correspond to the peering-pairs over which the prefix 557 updates are affected. In the above example, suppose a policy change 558 between AS4 and AS1 causes AS4 to prepend prefixes sent to AS1 559 (pCount changed from 1 to 2). Then AS4 would do key rollover only 560 for (AS2, AS1) and (AS3, AS1) peering-pairs, and not for any of the 561 others. This would substantially reduce the quantity of prefix 562 updates that are signed and re-propagated. In general, when peering 563 or policy changes occur, this method will reduce the number of prefix 564 updates to be re-propagated to exactly the same as that with normal 565 BGP. That means that this method would also be on par with the ET 566 and PKR methods in terms of update churn when a peering or policy 567 change takes place. The downside of this method is that the router 568 needs to maintain 2n(n-1) key pairs if it has n BGPsec peers. 570 Detailed discussion and comparison of this method with other methods 571 can be provided in a later version of this document if the idea picks 572 up interest in the WG. 574 6. Summary of Pros and Cons 576 Table 1 below summarizes the pros and cons for the various RAWS 577 protection methods. This summary follows from the discussion above 578 in Section 4 and Section 5. 580 +----------+---------------------------+----------------------------+ 581 | Method | Pros | Cons | 582 +----------+---------------------------+----------------------------+ 583 | Expirati | 1. The background load | 1. Prefix owner can abuse | 584 | on Time | due to beaconing is low | by beaconing too | 585 | (ET) | and not bursty. | frequently. | 586 | | --- | --- | 587 | | 2. Transit AS does NOT | 2. Any change to the units | 588 | | have a huge spike in | (granularity) of ET field | 589 | | workload even when a | entails a change to on- | 590 | | peering or policy change | the-wire BGPsec protocol. | 591 | | happens at that AS. | | 592 | | Beaconing facilitates | | 593 | | this. | | 594 | | --- | --- | 595 | | 3. Does not add to RPKI | | 596 | | churn. | | 597 | -------- | ------------------------- | -------------------------- | 598 | Periodic | 1. The background load | 1. Prefix owner can abuse | 599 | Key | due to beaconing is low | by beaconing (i.e. re- | 600 | Rollover | and not bursty. | originating) too | 601 | (PKR) | | frequently. | 602 | | --- | --- | 603 | | 2. Transit AS does NOT | 2. Adds to RPKI churn. A | 604 | | have a huge spike in | pair of certificates | 605 | | workload even when a | (current and next) for | 606 | | peering change happens at | each origination router | 607 | | that AS. Beaconing (i.e. | are rolled once every | 608 | | periodic re-origination) | beacon (i.e. re- | 609 | | facilitates this. | origination) interval. | 610 | | | Significantly more RPKI | 611 | | | churn than that with EKR-A | 612 | | | or EKR-B methods. | 613 | | --- | --- | 614 | | 3. If the periodic re- | | 615 | | origination (i.e. | | 616 | | beaconing) interval units | | 617 | | change, BGPsec protocol | | 618 | | on the wire remains | | 619 | | unaffected. | | 620 | | --- | --- | 621 | | 4. Changes in the method | | 622 | | (while still based on Key | | 623 | | Rollover) can be | | 624 | | accommodated without | | 625 | | requiring any change to | | 626 | | on-the-wire BGPsec | | 627 | | protocol. | | 628 | -------- | ------------------------- | -------------------------- | 629 | Event | 1. No update churn for | 1. Whenever the transit | 630 | driven | long periods when no | key is rolled (in response | 631 | Key | peering or policy changes | to a peering or policy | 632 | Rollover | occur. | change event), there is a | 633 | Type A | | storm of BGPsec updates, | 634 | (EKR-A) | | especially at routers in | 635 | | | large transit ASes. | 636 | | --- | --- | 637 | | 2. The added churn in | 2. The RAWS vulnerability | 638 | | RPKI is much lower than | window is dependent on | 639 | | that in the EKR-B method. | end-to-end CRL | 640 | | | propagation. It may vary | 641 | | | significantly from one | 642 | | | relying router to another | 643 | | | that may be in different | 644 | | | regions. | 645 | | --- | --- | 646 | | 3. Same as Pro #4 for the | | 647 | | PKR method. | | 648 | -------- | ------------------------- | -------------------------- | 649 | Event | 1. Same as Pro #1 for the | 1. Same as Con #1 for the | 650 | driven | EKR-A method. | EKR-A method. | 651 | Key | | | 652 | Rollover | | | 653 | Type B | | | 654 | (EKR-B) | | | 655 | | --- | --- | 656 | | 2. The RAWS vulnerability | 2. The added churn in RPKI | 657 | | window is enforced by | is much higher than that | 658 | | NotAfter time in | in the EKR-A method. | 659 | | certificates and is | | 660 | | therefore predictable. | | 661 | | --- | --- | 662 | | 3. Same as Pro #4 for the | | 663 | | PKR method. | | 664 +----------+---------------------------+----------------------------+ 665 Table 1: Table with Summary of Pros and Cons 667 7. Summary and Conclusions 669 We have attempted to provide insights into the operation of multiple 670 alternative methods for RAWS protection. It is hoped that the SIDR 671 WG will utilize the analysis presented here as input for deciding on 672 the choice of a mechanism for protection from RAWS. Once that 673 decision is made, the chosen mechanism would be included in the 674 standards track document [I-D.ietf-sidrops-bgpsec-rollover]. 676 Some important considerations for the decision making can be possibly 677 listed as follow: 679 1. The Expiration Time (ET) method is best (on par with the PKR 680 method) in terms of preventing huge update workloads during 681 peering and policy change events at transit routers with several 682 peers. It has no added RPKI churn. But the ET method has the 683 disadvantage of requiring on-the-wire protocol change if some 684 parameters (e.g., the units of beacon interval) change. 686 2. The Periodic Key Rollover (PKR) method operates the same way as 687 the ET method for preventing huge update workloads during peering 688 and policy change events at transit routers with several peers. 689 It does not have the disadvantage of requiring on-the-wire 690 protocol change if some parameters (e.g., the units of beaconing/ 691 re-origination periodicity) change. But it has the downside of 692 added RPKI churn. 694 3. The Event-driven Key Roll (EKR-A and EKR-B) methods have 695 significantly less RPKI churn than the PKR method. They also 696 have no BGPsec update churn during long quiet periods when no 697 peering or policy change events occur. But they suffer the 698 drawback of creating huge update workloads during peering and 699 policy change events at transit routers with several peers. Can 700 this workload be jittered or flow controlled to spread it over 701 time without convergence delay concerns? May be - needs further 702 study. 704 4. The EKR-A method relies on end-to-end CRL propagation through the 705 RPKI system to enforce expiry of a previous update when needed. 706 By contrast, in the EKR-B method the update expiry is controlled 707 by NotAfter time of the certificates used in update signatures. 708 In EKR-B method, previous update automatically becomes invalid at 709 the earliest NotAfter time of the certificates used in the 710 signatures unless each of those certificates' NotAfter time has 711 been extended. Also, in EKR-B method, changes in certificates to 712 extend their NotAfter time need not propagate end-to-end (all the 713 way to the relying routers); they may propagate only up to the 714 RPKI cache server of the relying router (see Section 5.2.2). The 715 changes in certificates to advance NotAfter time can be scheduled 716 and propagated (in RPKI) reasonably well in advance. 718 5. Besides being out-of-band relative to the BGPsec protocol on the 719 wire, the other good thing about the Key Rollover method is that 720 once the basics of the mechanism are implemented, there may be 721 flexibility to implement PKR, EKR-A or EKR-B on top of it. It 722 may also be possible to switch from one method to another (within 723 this class) if necessary based on operational experience; this 724 transition would not require any change to on-the-wire BGPsec 725 protocol. 727 8. Acknowledgements 729 The authors would like to thank Steve Kent for extensive review and 730 many useful suggestions on an earlier version of this document. 731 Thanks are also due to Roque Gagliano and Brian Weis for helpful 732 discussions. Further, we are thankful to Oliver Borchert and Okhee 733 Kim for comments and suggestions. 735 9. IANA Considerations 737 This memo includes no request to IANA. 739 10. Security Considerations 741 This memo requires no security considerations of its own since it is 742 targeted to be an informational RFC in support of 743 [I-D.ietf-sidrops-bgpsec-rollover] and [RFC8205]. The reader is 744 therefore directed to the security considerations provided in those 745 documents. 747 11. Informative References 749 [I-D.ietf-sidrops-bgpsec-rollover] 750 Weis, B., Gagliano, R., and K. Patel, "BGPsec Router 751 Certificate Rollover", draft-ietf-sidrops-bgpsec- 752 rollover-04 (work in progress), December 2017. 754 [RAWS-discussion] 755 Sriram, K. and D. Montgomery, "Discussion of Key Rollover 756 Mechanisms for Replay-Attack Protection", Presented 757 at IETF-85 SIDR WG Meeting, November 2012, 758 . 761 [RFC7353] Bellovin, S., Bush, R., and D. Ward, "Security 762 Requirements for BGP Path Validation", RFC 7353, 763 DOI 10.17487/RFC7353, August 2014, 764 . 766 [RFC8205] Lepinski, M., Ed. and K. Sriram, Ed., "BGPsec Protocol 767 Specification", RFC 8205, DOI 10.17487/RFC8205, September 768 2017, . 770 [rpki-delay] 771 Kent, S. and K. Sriram, "RPKI rsync Download Delay 772 Modeling", Presented at IETF-86 SIDR WG Meeting, March 773 2013, . 776 Authors' Addresses 778 Kotikalapudi Sriram 779 US NIST 781 Email: ksriram@nist.gov 783 Doug Montgomery 784 US NIST 786 Email: dougm@nist.gov