Skip to content

[Nexthop][fboss2-dev] fboss2 bgp policy routing-policy term match commands - #1491

Draft
hillol-nexthop wants to merge 18 commits into
facebook:mainfrom
nexthop-ai:bgp-policy-routing-policy-term-match
Draft

[Nexthop][fboss2-dev] fboss2 bgp policy routing-policy term match commands#1491
hillol-nexthop wants to merge 18 commits into
facebook:mainfrom
nexthop-ai:bgp-policy-routing-policy-term-match

Conversation

@hillol-nexthop

Copy link
Copy Markdown
Contributor

Adds the match level of a routing-policy term:

... term <seq-num> match from as-path-list <name>
... term <seq-num> match from origin <IGP|EGP|INCOMPLETE>
... term <seq-num> match from prefix-list <name>

Scoped to these three deliberately. The match conditions are written to policy_match_entries (the deprecated singular container) with the referenced list named in *_list_names, because that is what bgpd reads: the newer policy_matches field is not consumed anywhere in the daemon, so writing it would stage config that is silently ignored. The remaining documented match attributes (med, next-hop, local-preference) have no support in bgpd and are left out rather than staged as dead config.

Stacked on top of the term action PR.

Test

Built //fboss/cli/fboss2:fboss2-dev plus the config and integration test targets; cmd_config_test passes, including CmdConfigBgpPolicyRoutingPolicyTermMatchTest. The integration tests create the referenced as-path-list / prefix-list, commit, and verify the match lands in bgpd's running config via the getRunningConfig RPC.

🤖 Generated with Claude Code

hillol-nexthop and others added 18 commits July 31, 2026 10:16
Follow-up cleanups to the BGP-aware config session infra (facebook#1344):

- Collapse the BGP_RESTART action level into AGENT_WARMBOOT (bgpd has no hitless
  reload; its restart already runs the agent-warmboot code path).
- Introduce a single ConfigDomain descriptor + configDomains() and shared
  per-domain helpers so commit(), rollback() and `config session diff` handle
  the agent and BGP domains uniformly (private DiffDomain removed).
- Make the agent skip-when-unchanged like BGP: a commit whose staged config
  equals what is already promoted is a true no-op (no git revision, no symlink
  churn, no reloadConfig()/bgpd restart). Change detection is semantic (compare
  the deserialized thrift structs), so formatting-only diffs don't count.
- Consolidate `config session clear` onto a static stagedSessionFilePaths()
  and reuse ConfigSession::readStagedContent() in diff.
- Make ConfigSession::saveConfig(service, level) generic over the service and
  reduce saveBgpConfig() to a thin wrapper.
- Keep the heavy generated thrift headers out of ConfigSession.h: use the
  *_types_fwd.h forward-declaration headers, hold agentConfig_/bgpConfig_ by
  std::unique_ptr, and drop the configLoaded_/bgpConfigLoaded_ bools
  (null == not loaded).
- clang-tidy: use auto for the SimpleJSONSerializer template-cast results.

Built fboss2-dev + the config unit tests; config-session/commit/diff/BGP/clear
tests pass. Verified the agent no-op behaviour live on test switches.
Removed comments about the destructor definition in ConfigSession.h.
…thInterface call

main renamed findFirstEthInterface() to getRandomInterfacePortName()
(virtual-management-port fix); convert the branch-added no-op-commit test
to the new helper. Drop the three includes ConfigSession.cpp no longer
uses directly (misc-include-cleaner runs as errors in CI).
Add `fboss2-dev config protocol bgp global <attr> <value>` on top of the
BGP-aware ConfigSession (base PR). Edits the typed bgp::thrift::BgpConfig via
ConfigSession::getBgpConfig()/saveBgpConfig() -- the whole-config,
scope-agnostic typed API (no global/peer/peer-group special-casing in
ConfigSession).

- Collapse the 10 per-attribute global command classes into one dispatcher;
  reject cluster-id (no BgpConfig field) instead of writing dead config; bound
  switch-limit / max_golden_vips so out-of-range values aren't truncated.
- Integration tests: ConfigBgpGlobalTest (each attr set+commit, verified in the
  promoted /etc/coop/bgpcpp/bgpcpp.conf) and ConfigBgpSessionTest
  (clear/diff/commit/rollback + a no-op-restart regression), sharing
  ConfigBgpTestBase.

Test Plan:
- bazel test //fboss/cli/fboss2/test/config:cmd_config_test
- bazel build //fboss/cli/fboss2/test/integration_test:fboss2_integration_test
- fboss2_integration_test on a DUT with bgp_pp active: ConfigBgpSessionTest
  6/6 pass (clear, diff, commit-restarts-bgp_pp, rollback-restores-config, and
  the unchanged-config does-not-restart case); agent-session regression
  (ConfigInterfaceMtuTest) passes.
- Extract the value parsers (parseBool/parseInt/parseNonNegInt32, handler
  Result) into a shared BgpCliValueParsers.h so sibling BGP dispatchers can
  reuse them, and add a bounded parseAsn4Byte: local-asn/confed-asn accepted
  any uint64 and silently persisted out-of-range ASNs (>= 2^32 wrap the i64
  field negative).
- positionals_at_end() on the global command: CLI11's parent-chain subcommand
  fallthrough steals value tokens that match a sibling command name (e.g. a
  policy named "peer-group") and misparses the command.
- Integration test base: probe the bgpd unit, and pass
  -c safe.directory=/etc/coop on the raw git invocations (gitHead /
  bgpTrackedAtRevision), mirroring the CLI's Git class -- /etc/coop is owned
  by another user (e.g. coop) on provisioned devices, which git otherwise
  rejects as dubious ownership and the helpers silently return empty results.
The commit-path global-attribute tests asserted on the promoted
/etc/coop/bgpcpp/bgpcpp.conf, which only proves what the CLI wrote to disk.
Route ConfigBgpTestBase::setAndCommit() through a new
readRunningBgpConfigViaRpc() -- TBgpService::getRunningConfig against the local
bgpd -- so every positive ConfigBgpGlobalTest asserts its attribute in the
daemon's own view of its config, proving bgpd parsed and adopted the promoted
file after the commit-triggered restart.

The helper retries briefly on connection errors: systemd reports bgpd active as
soon as the process starts, but its thrift server binds the port a few seconds
later, so an RPC issued right after a restart races that window.

Test Plan:
- bazel build //fboss/cli/fboss2/test/integration_test:fboss2_integration_test
- fboss2_integration_test on a DUT with bgpd active: ConfigBgpGlobalTest 7/7
  pass (each attribute set+commit verified in bgpd's getRunningConfig view,
  plus the invalid-bool and negative-graceful-restart-time reject paths).
Adds `config protocol bgp neighbor <ip-address> [<attribute> <value> ...]`
and `delete protocol bgp neighbor <ip-address>` — the first per-peer BGP
config family, beneath the `protocol bgp` grouping node.

- neighbor dispatcher keyed by peer address, writing bgp_config.BgpPeer
  through the typed ConfigSession. Attributes cover the peer identity
  (remote/local ASN, description, peer-tag, peer-group), policy bindings
  (ingress/egress policy), session tunables (timers, route limits,
  next-hop, add-path, RR client, and the rest of the documented set), and
  the sheet-documented attributes bgpd does not support, which are
  rejected with an explanatory message rather than staged.
- delete removes the peer by address.
- ConfigBgpTestBase restores the committed BGP config after every test, so
  a failed run cannot leave a device carrying test config; commit-path
  integration tests verify each value through bgpd's own getRunningConfig
  RPC rather than only the staged file.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…y-list

Adds `config protocol bgp policy community-list <name> [community <name>]
[<attribute> <value> ...]` and `delete protocol bgp policy community-list
<name>` — the second BGP policy object-type command family, beneath the
`policy` grouping node alongside as-path-list.

- community-list dispatcher with a two-level key (list name + inline
  community member name) writing bgp_policy.BgpPolicies.community_lists[]
  through the typed ConfigSession; boolean-operator maps to
  routing_policy.BooleanOperator, exact-match to the optional bool, and
  the member attributes (description/type/value) set the CommunityRefType
  union's inline Community arm, keyed by Community.name.
- new boolAttr factory in the shared BgpCliAttrHandlers.h (generalizing
  the neighbor dispatcher's), reused by upcoming prefix-list.
- fix a latent dangling-string_view in the shared enumAttr factory: the
  lambda captured a string_view over the caller's fmt::format temporary,
  so every invalid-enum-value rejection printed garbage after
  "expected". valueDesc is now taken (and captured) by value; the
  as-path-list and community-list rejection tests pin the full message.
- unit tests (20) + commit-path integration tests mirroring
  ConfigBgpPolicyAsPathListTest (not yet run on a DUT).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds `config protocol bgp policy prefix-list <name> [entry <seq-num>]
[<attribute> <value> ...]` and `delete protocol bgp policy prefix-list
<name> [entry <seq-num>]` — the third BGP policy object-type command
family, beneath the `policy` grouping node alongside as-path-list and
community-list.

- prefix-list dispatcher with a two-level key (list name + entry
  seq-num) writing bgp_policy.BgpPolicies.prefix_lists[] through the
  typed ConfigSession. List level: boolean-operator, compare-operator
  (EQ|GE|LE|NE|GT|LT), description, and ip-version <v4|v6> (stored as
  the numeric routing_policy.PrefixList.version, the field the sheet
  documents). Entry level (routing_policy.PrefixListEntry keyed by
  seq_num in prefixes[]): base-prefix (validated as <prefix/len>, an
  explicit /len is required), communities (accumulates into the
  optional set), description, match-logic, max-allowed-subnet-count
  (-> max_allowed_golden_prefix_subnet_count), regex, and the
  prefix-len-range compare-operator|value pair targeting the single
  CompareNumericValue the CLI maintains at prefix_len_ranges[0]
  (compare-operator additionally accepts RG; value is bounded 0-128).
- delete mirrors community-list's two levels: the whole list by name,
  or a single entry by `entry <seq-num>`.
- new intAttr factory in the shared BgpCliAttrHandlers.h (bounded
  int32, valueDesc taken by value like enumAttr's), used by the two
  numeric entry attributes.
- unit tests (30) + commit-path integration tests mirroring
  ConfigBgpPolicyCommunityListTest, run on a DUT (all 4 pass).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…policy

Adds `config protocol bgp policy routing-policy <name>
[<attribute> <value> ...]` and `delete protocol bgp policy routing-policy
<name>` — the fourth BGP policy object-type command family, beneath the
`policy` grouping node alongside as-path-list, community-list, and
prefix-list.

- routing-policy dispatcher keyed by policy name writing
  bgp_policy.BgpPolicies.bgp_policy_statements[] (BgpPolicyStatement)
  through the typed ConfigSession. Policy level: description (joined
  multi-token string, NOS-6671); bare create (NOS-6689). The
  seq-num-keyed `term <seq-num>` level and its match/action attributes
  land as follow-ups.
- delete removes the whole policy statement by name; extra tokens are
  rejected.
- unit tests (9) + commit-path integration tests mirroring
  ConfigBgpPolicyCommunityListTest, verifying the description lands in
  bgpd's running config and that bgpd adopts a policy statement with no
  terms.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds `config protocol bgp policy routing-policy <name> term <seq-num>
[description <string>]` and `delete protocol bgp policy routing-policy
<name> term <seq-num>` — the seq-num-keyed term level of the routing-policy
family. The term's action and match levels land as follow-up subcommands.

Unlike the sibling policy families (which parse their nested level inside
the parent's ObjectArgType), term is a real CLI11 subcommand with its own
handler, mirroring the `config interface ... switchport` chain: the policy
name arrives through the ancestor-args tuple. This required dropping
positionals_at_end() from the routing-policy parent so CLI11 can classify
`term` after the policy name; the trade-off (a policy-level attribute value
spelling `term` is stolen by subcommand matching) is documented in the
traits. The term subcommand keeps positionals_at_end() so its own value
tokens are safe. A policy-level attribute mixed with a term command
(`routing-policy X description y term 10`) is rejected instead of silently
dropping the description, since only the leaf handler runs.

- term dispatcher keyed by (policy name from parent) + seq-num, writing
  bgp_policy.BgpPolicyTerm in policy_entries[]. Term level: description
  (joined multi-token string, NOS-6680). Bare create (NOS-6688) seeds the
  term name from the seq-num (next_term_id references terms by name).
- shared BgpRoutingPolicyCliUtils.h lookup/create helpers, used by the
  policy-level dispatcher, the term subcommand, and delete.
- delete removes a single term by seq-num; the policy and other terms stay.
- unit tests (14) + commit-path integration tests that exercise the CLI11
  subcommand dispatch through the full parse and verify terms in bgpd's
  running config.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ands

Adds `config protocol bgp policy routing-policy <name> term <seq-num>
action result <ACCEPT|REJECT|CONTINUE>` and `... action set
<attribute> <value> ...` — the action level of a routing-policy term.
`action` is a pure grouping node; `result` and `set` are its subcommands,
each with its own handler (the policy and term args arrive through the
ancestor-args tuple). The term traits drop positionals_at_end() so CLI11
can classify `action` (and later `match`) after the seq-num; ancestor
attributes mixed with an action command are rejected since only the leaf
runs.

`set` MUST be a real subcommand rather than a parsed token: CLI11's
_valid_subcommand walks the parent chain unconditionally, so a bare `set`
arg token inside `action` was classified as the top-level `set` VERB and
stolen ("The following arguments were not expected"). A local subcommand
named `set` shadows the verb because _find_subcommand checks the local
scope first.

- `result` (NOS-6672) maps ACCEPT/REJECT/CONTINUE onto FlowControlAction
  ACCEPT/DENY/NEXT_TERM in term_miss_action. GOTO-TERM stays deferred: no
  FlowControlAction arm, and the per-action-entry next_term_id target has
  no slot in the documented grammar.
- `set <attr>` writes one bgp_policy.BgpPolicyAction entry per action kind
  in policy_entries[].policy_action_entries[] (keyed by which payload field
  is set; re-issuing a kind updates its entry):
  - as-path prepend <asn> [<asn> ...] (NOS-6673) -> SetAsPathPrepend{asn,
    repeat_times}; the documented <asn-list> must be uniform since the
    thrift models one ASN repeated N times
  - community <community-string> [additive] (NOS-6674) -> inline
    single-member CommunityList + route_action COMMUNITY_LIST_ADD/SET
  - local-pref <0-4294967295> (NOS-6675) -> LocalPreference.local_pref
  - med <0-4294967295> (NOS-6676) -> MedAction{med_value, SET}
  - next-hop <ip-address> (NOS-6677) -> SetNextHop; the sheet's `self`
    (bgpd rejects set_self) and `peer-address` (no thrift arm) are deferred
  - origin <IGP|EGP|INCOMPLETE> (NOS-6678) -> set_origin
  - weight <0-65535> (NOS-6679) -> WeightAction{weight_value, SET}
- unit tests (12) + a commit-path integration test staging every action
  kind and verifying each entry in bgpd's running config.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…atch

Adds `config protocol bgp policy routing-policy <name> term <seq-num>
match from <attribute> <value>` — the match level of a routing-policy term,
as a CLI11 subcommand with its own handler alongside `action`; the policy
and term args arrive through the ancestor-args tuple, and ancestor
attributes mixed with a match command are rejected since only the leaf
runs.

Scoped to the four match kinds bgpd actually implements. Reading
private-BGP's PolicyTerm.cpp turned up two things the sheet's Thrift
reference column does not convey, but its JSON field path column does:

1. bgpd reads BgpPolicyTerm.policy_match_entries — a single
   BgpPolicyMatch, marked @thrift.Deprecated. Its match loop is guarded by
   `if (term.policy_match_entries())`, and NOTHING in private-BGP reads the
   newer `policy_matches` list, so matches written there are silently
   ignored: the config round-trips through getRunningConfig while bgpd
   applies no match at all. Same for the atomic payloads — bgpd reads the
   deprecated as_path_filters/communities_filter/prefix_filters, not the
   newer as_path_list/community_list/prefix_list.
2. createPolicyAttributeMatchItem has no LOCAL_PREFERENCE, MED or NEXT_HOP
   case, so those atomic types hit `default:` and throw
   "BgpPolicyAtomicMatch Config input error for type", crash-looping the
   daemon. `from local-pref`, `from med` and `from next-hop`
   (NOS-6683/6684/6685) are therefore rejected at parse time rather than
   offered; they need a private-BGP change first, and a unit test pins
   that they stay rejected.

- `from <attr>` writes one BgpPolicyAtomicMatch per match kind (keyed by
  BgpPolicyAtomicMatchType; re-issuing a kind updates its entry) into
  policy_match_entries, composing under its default AND — which is also
  the only operator bgpd accepts for more than one entry:
  - as-path-list <name> (NOS-6681) -> AS_PATH, as_path_filters
  - community-list <name> (NOS-6682) -> COMMUNITY_LIST, communities_filter
  - origin <IGP|EGP|INCOMPLETE> (NOS-6686) -> ORIGIN, origin
  - prefix-list <name> (NOS-6687) -> PREFIX_LIST, prefix_filters
  The three list-typed matches take a full inline object with no name-only
  union arm, and bgpd treats that object's `name` as the reference: the
  match class keeps the name, then PolicyManager::PopulateReferences swaps
  in the real list from its by-name map. A name-only object is exactly how
  a by-name match is expressed.
- setters are named functions, so the registry reads as a table: one line
  per attribute giving its dispatch key, value shape, and setter.
- unit tests (8) + a commit-path integration test. The integration
  assertions are load-bearing here in a way staged-config ones are not:
  bgpd constructs a Policy from every statement, so a wrong encoding
  crash-loops it and fails the test rather than sitting inert.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant