diff options
Diffstat (limited to 'src/vnet/fib/fib.h')
| -rw-r--r-- | src/vnet/fib/fib.h | 650 |
1 files changed, 650 insertions, 0 deletions
diff --git a/src/vnet/fib/fib.h b/src/vnet/fib/fib.h new file mode 100644 index 00000000..ec97c565 --- /dev/null +++ b/src/vnet/fib/fib.h @@ -0,0 +1,650 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +/** + * \brief + * A IP v4/6 independent FIB. + * + * The main functions provided by the FIB are as follows; + * + * - source priorities + * + * A route can be added to the FIB by more than entity or source. Sources + * include, but are not limited to, API, CLI, LISP, MAP, etc (for the full list + * see fib_entry.h). Each source provides the forwarding information (FI) that + * is has determined as required for that route. Since each source determines the + * FI using different best path and loop prevention algorithms, it is not + * correct for the FI of multiple sources to be combined. Instead the FIB must + * choose to use the FI from only one source. This choose is based on a static + * priority assignment. For example; + * IF a prefix is added as a result of interface configuration: + * set interface address 192.168.1.1/24 GigE0 + * and then it is also added from the CLI + * ip route 192.168.1.1/32 via 2.2.2.2/32 + * then the 'interface' source will prevail, and the route will remain as + * 'local'. + * The requirement of the FIB is to always install the FI from the winning + * source and thus to maintain the FI added by losing sources so it can be + * installed should the winning source be withdrawn. + * + * - adj-fib maintenance + * + * When ARP or ND discover a neighbour on a link an adjacency forms for the + * address of that neighbour. It is also required to insert a route in the + * appropriate FIB table, corresponding to the VRF for the link, an entry for + * that neighbour. This entry is often referred to as an adj-fib. Adj-fibs + * have a dedicated source; 'ADJ'. + * The priority of the ADJ source is lower than most. This is so the following + * config; + * set interface address 192.168.1.1/32 GigE0 + * ip arp 192.168.1.2 GigE0 dead.dead.dead + * ip route add 192.168.1.2 via 10.10.10.10 GigE1 + * will forward traffic for 192.168.1.2 via GigE1. That is the route added + * by the control plane is favoured over the adjacency discovered by ARP. + * The control plane, with its associated authentication, is considered the + * authoritative source. + * To counter the nefarious addition of adj-fib, through the nefarious injection + * of adjacencies, the FIB is also required to ensure that only adj-fibs whose + * less specific covering prefix is connected are installed in forwarding. This + * requires the use of 'cover tracking', where a route maintains a dependency + * relationship with the route that is its less specific cover. When this cover + * changes (i.e. there is a new covering route) or the forwarding information + * of the cover changes, then the covered route is notified. + * + * Overlapping sub-nets are not supported, so no adj-fib has multiple paths. + * The control plane is expected to remove a prefix configured for an interface + * before the interface changes VRF. + * So while the following config is accepted: + * set interface address 192.168.1.1/32 GigE0 + * ip arp 192.168.1.2 GigE0 dead.dead.dead + * set interface ip table GigE0 2 + * it does not result in the desired behaviour. + * + * - attached export. + * + * Further to adj-fib maintenance above consider the following config: + * set interface address 192.168.1.1/24 GigE0 + * ip route add table 2 192.168.1.0/24 GigE0 + * Traffic destined for 192.168.1.2 in table 2 will generate an ARP request + * on GigE0. However, since GigE0 is in table 0, all adj-fibs will be added in + * FIB 0. Hence all hosts in the sub-net are unreachable from table 2. To resolve + * this, all adj-fib and local prefixes are exported (i.e. copied) from the + * 'export' table 0, to the 'import' table 2. There can be many import tables + * for a single export table. + * + * - recursive route resolution + * + * A recursive route is of the form: + * 1.1.1.1/32 via 10.10.10.10 + * i.e. a route for which no egress interface is provided. In order to forward + * traffic to 1.1.1.1/32 the FIB must therefore first determine how to forward + * traffic to 10.10.10.10/32. This is recursive resolution. + * Recursive resolution, just like normal resolution, proceeds via a longest + * prefix match for the 'via-address' 10.10.10.10. Note it is only possible + * to add routes via an address (i.e. a /32 or /128) not via a shorter mask + * prefix. There is no use case for the latter. + * Since recursive resolution proceeds via a longest prefix match, the entry + * in the FIB that will resolve the recursive route, termed the via-entry, may + * change as other routes are added to the FIB. Consider the recursive + * route shown above, and this non-recursive route: + * 10.10.10.0/24 via 192.168.16.1 GigE0 + * The entry for 10.10.10.0/24 is thus the resolving via-entry. If this entry is + * modified, to say; + * 10.10.10.0/24 via 192.16.1.3 GigE0 + * Then packet for 1.1.1.1/32 must also be sent to the new next-hop. + * Now consider the addition of; + * 10.10.10.0/28 via 192.168.16.2 GigE0 + * The more specific /28 is a better longest prefix match and thus becomes the + * via-entry. Removal of the /28 means the resolution will revert to the /24. + * The tracking to the changes in recursive resolution is the requirement of + * the FIB. When the forwarding information of the via-entry changes a back-walk + * is used to update dependent recursive routes. When new routes are added to + * the table the cover tracking feature provides the necessary notifications to + * the via-entry routes. + * The adjacency constructed for 1.1.1.1/32 will be a recursive adjacency + * whose next adjacency will be contributed from the via-entry. Maintaining + * the validity of this recursive adjacency is a requirement of the FIB. + * + * - recursive loop avoidance + * + * Consider this set of routes: + * 1.1.1.1/32 via 2.2.2.2 + * 2.2.2.2/32 via 3.3.3.3 + * 3.3.3.3/32 via 1.1.1.1 + * this is termed a recursion loop - all of the routes in the loop are + * unresolved in so far as they do not have a resolving adjacency, but each + * is resolved because the via-entry is known. It is important here to note + * the distinction between the control-plane objects and the data-plane objects + * (more details in the implementation section). The control plane objects must + * allow the loop to form (i.e. the graph becomes cyclic), however, the + * data-plane absolutely must not allow the loop to form, otherwise the packet + * would loop indefinitely and never egress the device - meltdown would follow. + * The control plane must allow the loop to form, because when the loop breaks, + * all members of the loop need to be updated. Forming the loop allows the + * dependencies to be correctly setup to allow this to happen. + * There is no limit to the depth of recursion supported by VPP so: + * 9.9.9.100/32 via 9.9.9.99 + * 9.9.9.99/32 via 9.9.9.98 + * 9.9.9.98/32 via 9.9.9.97 + * ... turtles, turtles, turtles ... + * 9.9.9.1/32 via 10.10.10.10 Gig0 + * is supported to as many layers of turtles is desired, however, when + * back-walking a graph (in this case from 9.9.9.1/32 up toward 9.9.9.100/32) + * a FIB needs to differentiate the case where the recursion is deep versus + * the case where the recursion is looped. A simple method, employed by VPP FIB, + * is to limit the number of steps. VPP FIB limit is 16. Typical BGP scenarios + * in the wild do not exceed 3 (BGP Inter-AS option C). + * + * - Fast Convergence + * + * After a network topology change, the 'convergence' time, is the time taken + * for the router to complete a transition to forward traffic using the new + * topology. The convergence time is therefore a summation of the time to; + * - detect the failure. + * - calculate the new 'best path' information + * - download the new best paths to the data-plane. + * - install those best best in data-plane forwarding. + * The last two points are of relevance to VPP architecture. The download API is + * binary and batch, details are not discussed here. There is no HW component to + * programme, installation time is bounded by the memory allocation and table + * lookup and insert access times. + * + * 'Fast' convergence refers to a set of technologies that a FIB can employ to + * completely or partially restore forwarding whilst the convergence actions + * listed above are ongoing. Fast convergence technologies are further + * sub-divided into Prefix Independent Convergence (PIC) and Loop Free + * Alternate path Fast re-route (LFA-FRR or sometimes called IP-FRR) which + * affect recursive and non-recursive routes respectively. + * + * LFA-FRR + * + * Consider the network topology below: + * + * C + * / \ + * X -- A --- B - Y + * | | + * D F + * \ / + * E + * + * all links are equal cost, traffic is passing from X to Y. the best path is + * X-A-B-Y. There are two alternative paths, one via C and one via E. An + * alternate path is considered to be loop free if no other router on that path + * would forward the traffic back to the sender. Consider router C, its best + * path to Y is via B, so if A were to send traffic destined to Y to C, then C + * would forward that traffic to B - this is a loop-free alternate path. In + * contrast consider router D. D's shortest path to Y is via A, so if A were to + * send traffic destined to Y via D, then D would send it back to A; this is + * not a loop-free alternate path. There are several points of note; + * - we are considering the pre-failure routing topology + * - any equal-cost multi-path between A and B is also a LFA path. + * - in order for A to calculate LFA paths it must be aware of the best-path + * to Y from the perspective of D. These calculations are thus limited to + * routing protocols that have a full view of the network topology, i.e. + * link-state DB protocols like OSPF or an SDN controller. LFA protected + * prefixes are thus non-recursive. + * + * LFA is specified as a 1 to 1 redundancy; a primary path has only one LFA + * (a.k.a. backup) path. To my knowledge this limitation is one of complexity + * in the calculation of and capacity planning using a 1-n redundancy. + * + * In the event that the link A-B fails, the alternate path via C can be used. + * In order to provide 'fast' failover in the event of a failure, the control + * plane will download both the primary and the backup path to the FIB. It is + * then a requirement of the FIB to perform the failover (a.k.a cutover) from + * the primary to the backup path as quickly as possible, and particularly + * without any other control-plane intervention. The expectation is cutover is + * less than 50 milli-seconds - a value allegedly from the VOIP QoS. Note that + * cutover time still includes the fault detection time, which in a vitalised + * environment could be the dominant factor. Failure detection can be either a + * link down, which will affect multiple paths on a multi-access interface, or + * via a specific path heartbeat (i.e. BFD). + * At this time VPP does not support LFA, that is it does not support the + * installation of a primary and backup path[s] for a route. However, it does + * support ECMP, and VPP FIB is designed to quickly remove failed paths from + * the ECMP set, however, it does not insert shared objects specific to the + * protected resource into the forwarding object graph, since this would incur + * a forwarding/performance cost. Failover time is thus route number dependent. + * Details are provided in the implementation section below. + * + * PIC + * + * PIC refers to the concept that the converge time should be independent of + * the number of prefixes/routes that are affected by the failure. PIC is + * therefore most appropriate when considering networks with large number of + * prefixes, i.e. BGP networks and thus recursive prefixes. There are several + * flavours of PIC covering different locations of protection and failure + * scenarios. An outline is given below, see the literature for more details: + * + * Y/16 - CE1 -- PE1---\ + * | \ P1---\ + * | \ PE3 -- CE3 - X/16 + * | - P2---/ + * Y/16 - CE2 -- PE2---/ + * + * CE = customer edge, PE = provider edge. external-BGP runs between customer + * and provider, internal-BGP runs between provider and provider. + * + * 1) iBGP PIC-core: consider traffic from CE1 to X/16 via CE3. On PE1 there is + * are routes; + * X/16 (and hundreds of thousands of others like it) + * via PE3 + * and + * PE3/32 (its loopback address) + * via 10.0.0.1 Link0 (this is P1) + * via 10.1.1.1 Link1 (this is P2) + * the failure is the loss of link0 or link1 + * As in all PIC scenarios, in order to provide prefix independent convergence + * it must be that the route for X/16 (and all other routes via PE3) do not + * need to be updated in the FIB. The FIB therefore needs to update a single + * object that is shared by all routes - once this shared object is updated, + * then all routes using it will be instantly updated to use the new forwarding + * information. In this case the shared object is the resolving route via PE3. + * Once the route via PE3 is updated via IGP (OSPF) convergence, then all + * recursive routes that resolve through it are also updated. VPP FIB + * implements this scenario via a recursive-adjacency. the X/16 and it sibling + * routes share a recursive-adjacency that links to/points at/stacks on the + * normal adjacency contributed by the route for PE3. Once this shared + * recursive adj is re-linked then all routes are switched to using the new + * forwarding information. This is shown below; + * + * pre-failure; + * X/16 --> R-ADJ-1 --> ADJ-1-PE3 (multi-path via P1 and P2) + * + * post-failure: + * X/16 --> R-ADJ-1 --> ADJ-2-PE3 (single path via P1) + * + * note that R-ADJ-1 (the recursive adj) remains in the forwarding graph, + * therefore X/16 (and all its siblings) is not updated. + * X/16 and its siblings share the recursive adj since they share the same + * path-list. It is the path-list object that contributes the recursive-adj + * (see next section for more details) + * + * + * 2) iBGP PIC-edge; Traffic from CE3 to Y/16. On PE3 there is are routes; + * Y/16 (and hundreds of thousands of others like it) + * via PE1 + * via PE2 + * and + * PE1/32 (PE1's loopback address) + * via 10.0.2.2 Link0 (this is P1) + * PE2/32 (PE2's loopback address) + * via 10.0.3.3 Link1 (this is P2) + * + * the failure is the loss of reachability to PE2. this could be either the + * loss of the link P2-PE2 or the loss of the node PE2. This is detected either + * by the withdrawal of the PE2's loopback route or by some form of failure + * detection (i.e. BFD). + * VPP FIB again provides PIC via the use of the shared recursive-adj. Y/16 and + * its siblings will again share a path-list for the list {PE1,PE2}, this + * path-list will contribute a multi-path-recursive-adj, i.e. a multi-path-adj + * with each choice therein being another adj; + * + * Y/16 -> RM-ADJ --> ADJ1 (for PE1) + * --> ADJ2 (for PE2) + * + * when the route for PE1 is withdrawn then the multi-path-recursive-adjacency + * is updated to be; + * + * Y/16 --> RM-ADJ --> ADJ1 (for PE1) + * --> ADJ1 (for PE1) + * + * that is both choices in the ECMP set are the same and thus all traffic is + * forwarded to PE1. Eventually the control plane will download a route update + * for Y/16 to be via PE1 only. At that time the situation will be: + * + * Y/16 -> R-ADJ --> ADJ1 (for PE1) + * + * In the scenario above we assumed that PE1 and PE2 are ECMP for Y/16. eBGP + * PIC core is also specified for the case were one PE is primary and the other + * backup - VPP FIB does not support that case at this time. + * + * 3) eBGP PIC Edge; Traffic from CE3 to Y/16. On PE1 there is are routes; + * Y/16 (and hundreds of thousands of others like it) + * via CE1 (primary) + * via PE2 (backup) + * and + * CE1 (this is an adj-fib) + * via 11.0.0.1 Link0 (this is CE1) << this is an adj-fib + * PE2 (PE2's loopback address) + * via 10.0.5.5 Link1 (this is link PE1-PE2) + * the failure is the loss of link0 to CE1. The failure can be detected by FIB + * either as a link down event or by the control plane withdrawing the connected + * prefix on the link0 (say 10.0.5.4/30). The latter works because the resolving + * entry is an adj-fib, so removing the connected will withdraw the adj-fib, and + * hence the recursive path becomes unresolved. The former is faster, + * particularly in the case of Inter-AS option A where there are many VLAN + * sub-interfaces on the PE-CE link, one for each VRF, and so the control plane + * must remove the connected prefix for each sub-interface to trigger PIC in + * each VRF. Note though that total PIC cutover time will depend on VRF scale + * with either trigger. + * Primary and backup paths in this eBGP PIC-edge scenario are calculated by + * BGP. Each peer is configured to always advertise its best external path to + * its iBGP peers. Backup paths therefore send traffic from the PE back into the + * core to an alternate PE. A PE may have multiple external paths, i.e. multiple + * directly connected CEs, it may also have multiple backup PEs, however there + * is no correlation between the two, so unlike LFA-FRR, the redundancy model is + * N-M; N primary paths are backed-up by M backup paths - only when all primary + * paths fail, then the cutover is performed onto the M backup paths. Note that + * PE2 must be suitably configured to forward traffic on its external path that + * was received from PE1. VPP FIB does not support external-internal-BGP (eiBGP) + * load-balancing. + * + * As with LFA-FRR the use of primary and backup paths is not currently + * supported, however, the use of a recursive-multi-path-adj, and a suitably + * constrained hashing algorithm to choose from the primary or backup path sets, + * would again provide the necessary shared object and hence the prefix scale + * independent cutover. + * + * Astute readers will recognise that both of the eBGP PIC scenarios refer only + * to a BGP free core. + * + * Fast convergence implementation options come in two flavours: + * 1) Insert switches into the data-path. The switch represents the protected + * resource. If the switch is 'on' the primary path is taken, otherwise + * the backup path is taken. Testing the switch in the data-path comes with + * an associated performance cost. A given packet may encounter more than + * one protected resource as it is forwarded. This approach minimises + * cutover times as packets will be forwarded on the backup path as soon + * as the protected resource is detected to be down and the single switch + * is tripped. However, it comes at a performance cost, which increases + * with each shared resource a packet encounters in the data-path. + * This approach is thus best suited to LFA-FRR where the protected routes + * are non-recursive (i.e. encounter few shared resources) and the + * expectation on cutover times is more stringent (<50msecs). + * 2) Update shared objects. Identify objects in the data-path, that are + * required to be present whether or not fast convergence is required (i.e. + * adjacencies) that can be shared by multiple routes. Create a dependency + * between these objects at the protected resource. When the protected + * resource fails, each of the shared objects is updated in a way that all + * users of it see a consistent change. This approach incurs no performance + * penalty as the data-path structure is unchanged, however, the cutover + * times are longer as more work is required when the resource fails. This + * scheme is thus more appropriate to recursive prefixes (where the packet + * will encounter multiple protected resources) and to fast-convergence + * technologies where the cutover times are less stringent (i.e. PIC). + * + * Implementation: + * --------------- + * + * Due to the requirements outlined above, not all routes known to FIB + * (e.g. adj-fibs) are installed in forwarding. However, should circumstances + * change, those routes will need to be added. This adds the requirement that + * a FIB maintains two tables per-VRF, per-AF (where a 'table' is indexed by + * prefix); the forwarding and non-forwarding tables. + * + * For DP speed in VPP we want the lookup in the forwarding table to directly + * result in the ADJ. So the two tables; one contains all the routes (a + * lookup therein yields a fib_entry_t), the other contains only the forwarding + * routes (a lookup therein yields an ip_adjacency_t). The latter is used by the + * DP. + * This trades memory for forwarding performance. A good trade-off in VPP's + * expected operating environments. + * + * Note these tables are keyed only by the prefix (and since there 2 two + * per-VRF, implicitly by the VRF too). The key for an adjacency is the + * tuple:{next-hop, address (and it's AF), interface, link/ether-type}. + * consider this curious, but allowed, config; + * + * set int ip addr 10.0.0.1/24 Gig0 + * set ip arp Gig0 10.0.0.2 dead.dead.dead + * # a host in that sub-net is routed via a better next hop (say it avoids a + * # big L2 domain) + * ip route add 10.0.0.2 Gig1 192.168.1.1 + * # this recursive should go via Gig1 + * ip route add 1.1.1.1/32 via 10.0.0.2 + * # this non-recursive should go via Gig0 + * ip route add 2.2.2.2/32 via Gig0 10.0.0.2 + * + * for the last route, the lookup for the path (via {Gig0, 10.0.0.2}) in the + * prefix table would not yield the correct result. To fix this we need a + * separate table for the adjacencies. + * + * - FIB data structures; + * + * fib_entry_t: + * - a representation of a route. + * - has a prefix. + * - it maintains an array of path-lists that have been contributed by the + * different sources + * - install an adjacency in the forwarding table contributed by the best + * source's path-list. + * + * fib_path_list_t: + * - a list of paths + * - path-lists may be shared between FIB entries. The path-lists are thus + * kept in a DB. The key is the combined description of the paths. We share + * path-lists when it will aid convergence to do so. Adding path-lists to + * this DB that are never shared, or are not shared by prefixes that are + * not subject to PIC, will increase the size of the DB unnecessarily and + * may lead to increased search times due to hash collisions. + * - the path-list contributes the appropriate adj for the entry in the + * forwarding table. The adj can be 'normal', multi-path or recursive, + * depending on the number of paths and their types. + * - since path-lists are shared there is only one instance of the multi-path + * adj that they [may] create. As such multi-path adjacencies do not need a + * separate DB. + * The path-list with recursive paths and the recursive adjacency that it + * contributes forms the backbone of the fast convergence architecture (as + * described previously). + * + * fib_path_t: + * - a description of how to forward the traffic (i.e. via {Gig1, K}). + * - the path describes the intent on how to forward. This differs from how + * the path resolves. I.e. it might not be resolved at all (since the + * interface is deleted or down). + * - paths have different types, most notably recursive or non-recursive. + * - a fib_path_t will contribute the appropriate adjacency object. It is from + * these contributions that the DP graph/chain for the route is built. + * - if the path is recursive and a recursion loop is detected, then the path + * will contribute the special DROP adjacency. This way, whilst the control + * plane graph is looped, the data-plane graph does not. + * + * we build a graph of these objects; + * + * fib_entry_t -> fib_path_list_t -> fib_path_t -> ... + * + * for recursive paths: + * + * fib_path_t -> fib_entry_t -> .... + * + * for non-recursive paths + * + * fib_path_t -> ip_adjacency_t -> interface + * + * These objects, which constitute the 'control plane' part of the FIB are used + * to represent the resolution of a route. As a whole this is referred to as the + * control plane graph. There is a separate DP graph to represent the forwarding + * of a packet. In the DP graph each object represents an action that is applied + * to a packet as it traverses the graph. For example, a lookup of a IP address + * in the forwarding table could result in the following graph: + * + * recursive-adj --> multi-path-adj --> interface_A + * --> interface_B + * + * A packet traversing this FIB DP graph would thus also traverse a VPP node + * graph of: + * + * ipX_recursive --> ipX_rewrite --> interface_A_tx --> etc + * + * The taxonomy of objects in a FIB graph is as follows, consider; + * + * A --> + * B --> D + * C --> + * + * Where A,B and C are (for example) routes that resolve through D. + * parent; D is the parent of A, B, and C. + * children: A, B, and C are children of D. + * sibling: A, B and C are siblings of one another. + * + * All shared objects in the FIB are reference counted. Users of these objects + * are thus expected to use the add_lock/unlock semantics (as one would + * normally use malloc/free). + * + * WALKS + * + * It is necessary to walk/traverse the graph forwards (entry to interface) to + * perform a collapse or build a recursive adj and backwards (interface + * to entry) to perform updates, i.e. when interface state changes or when + * recursive route resolution updates occur. + * A forward walk follows simply by navigating an object's parent pointer to + * access its parent object. For objects with multiple parents (e.g. a + * path-list), each parent is walked in turn. + * To support back-walks direct dependencies are maintained between objects, + * i.e. in the relationship, {A, B, C} --> D, then object D will maintain a list + * of 'pointers' to its children {A, B, C}. Bare C-language pointers are not + * allowed, so a pointer is described in terms of an object type (i.e. entry, + * path-list, etc) and index - this allows the object to be retrieved from the + * appropriate pool. A list is maintained to achieve fast convergence at scale. + * When there are millions or recursive prefixes, it is very inefficient to + * blindly walk the tables looking for entries that were affected by a given + * topology change. The lowest hanging fruit when optimising is to remove + * actions that are not required, so all back-walks only traverse objects that + * are directly affected by the change. + * + * PIC Core and fast-reroute rely on FIB reacting quickly to an interface + * state change to update the multi-path-adjacencies that use this interface. + * An example graph is shown below: + * + * E_a --> + * E_b --> PL_2 --> P_a --> Interface_A + * ... --> P_c -\ + * E_k --> \ + * Interface_K + * / + * E_l --> / + * E_m --> PL_1 --> P_d -/ + * ... --> P_f --> Interface_F + * E_z --> + * + * E = fib_entry_t + * PL = fib_path_list_t + * P = fib_path_t + * The subscripts are arbitrary and serve only to distinguish object instances. + * This CP graph result in the following DP graph: + * + * M-ADJ-2 --> Interface_A + * \ + * -> Interface_K + * / + * M-ADJ-1 --> Interface_F + * + * M-ADJ = multi-path-adjacency. + * + * When interface K goes down a back-walk is started over its dependants in the + * control plane graph. This back-walk will reach PL_1 and PL_2 and result in + * the calculation of new adjacencies that have interface K removed. The walk + * will continue to the entry objects and thus the forwarding table is updated + * for each prefix with the new adjacency. The DP graph then becomes: + * + * ADJ-3 --> Interface_A + * + * ADJ-4 --> Interface_F + * + * The eBGP PIC scenarios described above relied on the update of a path-list's + * recursive-adjacency to provide the shared point of cutover. This is shown + * below + * + * E_a --> + * E_b --> PL_2 --> P_a --> E_44 --> PL_a --> P_b --> Interface_A + * ... --> P_c -\ + * E_k --> \ + * \ + * E_1 --> PL_k -> P_k --> Interface_K + * / + * E_l --> / + * E_m --> PL_1 --> P_d -/ + * ... --> P_f --> E_55 --> PL_e --> P_e --> Interface_E + * E_z --> + * + * The failure scenario is the removal of entry E_1 and thus the paths P_c and + * P_d become unresolved. To achieve PIC the two shared recursive path-lists, + * PL_1 and PL_2 must be updated to remove E_1 from the recursive-multi-path- + * adjacencies that they contribute, before any entry E_a to E_z is updated. + * This means that as the update propagates backwards (right to left) in the + * graph it must do so breadth first not depth first. Note this approach leads + * to convergence times that are dependent on the number of path-list and so + * the number of combinations of egress PEs - this is desirable as this + * scale is considerably lower than the number of prefixes. + * + * If we consider another section of the graph that is similar to the one + * shown above where there is another prefix E_2 in a similar position to E_1 + * and so also has many dependent children. It is reasonable to expect that a + * particular network failure may simultaneously render E_1 and E_2 unreachable. + * This means that the update to withdraw E_2 is download immediately after the + * update to withdraw E_1. It is a requirement on the FIB to not spend large + * amounts of time in a back-walk whilst processing the update for E_1, i.e. the + * back-walk must not reach as far as E_a and its siblings. Therefore, after the + * back-walk has traversed one generation (breadth first) to update all the + * path-lists it should be suspended/back-ground and further updates allowed + * to be handled. Once the update queue is empty, the suspended walks can be + * resumed. Note that in the case that multiple updates affect the same entry + * (say E_1) then this will trigger multiple similar walks, these are merged, + * so each child is updated only once. + * In the presence of more layers of recursion PIC is still a desirable + * feature. Consider an extension to the diagram above, where more recursive + * routes (E_100 -> E_200) are added as children of E_a: + * + * E_100 --> + * E_101 --> PL_3 --> P_j-\ + * ... \ + * E_199 --> E_a --> + * E_b --> PL_2 --> P_a --> E_44 --> ...etc.. + * ... --> P_c -\ + * E_k \ + * E_1 --> ...etc.. + * / + * E_l --> / + * E_m --> PL_1 --> P_d -/ + * ... --> P_e --> E_55 --> ...etc.. + * E_z --> + * + * To achieve PIC for the routes E_100->E_199, PL_3 needs to be updated before + * E_b -> E_z, a breadth first traversal at each level would not achieve this. + * Instead the walk must proceed intelligently. Children on PL_2 are sorted so + * those Entry objects that themselves have children appear first in the list, + * those without later. When an entry object is walked that has children, a + * walk of its children is pushed to the front background queue. The back + * ground queue is a priority queue. As the breadth first traversal proceeds + * across the dependent entry object E_a to E_k, when the first entry that does + * not have children is reached (E_b), the walk is suspended and placed at the + * back of the queue. Following this prioritisation method shared path-list + * updates are performed before all non-resolving entry objects. + * The CPU/core/thread that handles the updates is the same thread that handles + * the back-walks. Handling updates has a higher priority than making walk + * progress, so a walk is required to be interruptable/suspendable when new + * updates are available. + * !!! TODO - this section describes how walks should be not how they are !!! + * + * In the diagram above E_100 is an IP route, however, VPP has no restrictions + * on the type of object that can be a dependent of a FIB entry. Children of + * a FIB entry can be (and are) GRE & VXLAN tunnels endpoints, L2VPN LSPs etc. + * By including all object types into the graph and extending the back-walk, we + * can thus deliver fast convergence to technologies that overlay on an IP + * network. + * + * If having read all the above carefully you are still thinking; 'i don't need + * all this %&$* i have a route only I know about and I just need to jam it in', + * then fib_table_entry_special_add() is your only friend. + */ + +#ifndef __FIB_H__ +#define __FIB_H__ + +#include <vnet/fib/fib_table.h> +#include <vnet/fib/fib_entry.h> + +#endif |