feat(agent): node agent — enroll/mTLS, heartbeat, command stream, sing-box 用户表 (tsk__R8M4jEw43JR)
与控制面同仓同 go.mod,新增节点 agent 实现:
- proto/agent/v1/agent.proto + internal/pb/agentv1:冻结的控制面↔agent gRPC 契约
(Enroll/Register/Heartbeat/Subscribe/Ack/ReportUsage)。仓库尚无 protoc 流水线,
暂以手写 Go 类型 + JSON gRPC codec 实现,与 proto 1:1 对应,待 protoc 接入即可替换。
- internal/agentd:
- enroll.go:首启生成 EC 密钥+CSR,持 bootstrap token 调 Enroll 换 90d 节点证书
(CN=node_uuid),落 /etc/pangolin-agent/,此后 mTLS。
- conn.go(agent.go)+creds.go:mTLS 主动拨号 + 指数退避重连;重连携带 last_command_id;
Register 取 ConfigSnapshot 全量配置覆盖本地。
- heartbeat.go:30s 上报 peer/带宽/CPU + config_version;need_full_resync→全量同步。
- command.go:消费 Subscribe,Upsert/Revoke/Rotate/ApplyConfig/Lifecycle 幂等处理后
Ack(at-least-once,按 command_id 去重)。
- singbox.go+render.go:内存用户表 + 落盘 state.json(仅 dp_uuid+expires_at);任何变更
渲染完整 sing-box 配置(REALITY users[uuid,flow] + Hy2 users[派生口令])→ 500ms 去抖
合并 → systemd 重启。
- ttl.go:凭证 TTL 定时移除并上报。
- usage.go:按 dp_uuid 聚合上报,绝无 user_id/email/目的地址。
- derive.go:Hy2 口令 = HMAC-SHA256(key, dp_uuid),与控制面同源派生。
- cmd/agent:入口(flag/env 配置)。
- infra/cloud-init/{node.yaml.tmpl,install-node.sh,README.md}:一段式安装,下载锁定版本
二进制并校验 SHA-256,systemd 拉管,首启即 Enroll/Register。shellcheck -S warning 通过。
测试(bufconn mock 控制面,无需 docker):Enroll→Register→Heartbeat 全流转;Upsert/Revoke
渲染正确;Rotate 宽限期新旧并存到点移除;TTL 自动移除并上报;断流重连 last_command_id
续发不丢不重;need_full_resync 触发重注册;state.json 恢复;去抖合并;扫描确认无身份字段。
go test -race ./internal/agentd/... ./internal/pb/... 通过;go vet ./... 通过。
落实 doc/04 §2 节点无状态化与 doc/06 §3 数据面红线(节点仅见 dp_uuid)。
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,54 @@
|
||||
package agentd
|
||||
|
||||
import (
|
||||
"context"
|
||||
"time"
|
||||
|
||||
agentv1 "github.com/wangjia/pangolin/server/internal/pb/agentv1"
|
||||
)
|
||||
|
||||
// UsageSource yields per-dp_uuid aggregate counters for the elapsed window. It
|
||||
// MUST return only dp_uuids and byte/minute counts — never user ids, addresses or
|
||||
// DNS. The default returns nothing (no stats backend wired); production plugs in a
|
||||
// sing-box stats reader.
|
||||
type UsageSource interface {
|
||||
// Collect drains and returns usage accumulated since the previous call.
|
||||
Collect() []*agentv1.UsageEntry
|
||||
}
|
||||
|
||||
// nopUsageSource reports nothing.
|
||||
type nopUsageSource struct{}
|
||||
|
||||
func (nopUsageSource) Collect() []*agentv1.UsageEntry { return nil }
|
||||
|
||||
// runUsage periodically aggregates and uploads usage. Reporting failures are
|
||||
// logged but do not tear down the session (usage is best-effort, at-least-once).
|
||||
func (a *Agent) runUsage(ctx context.Context, client agentv1.AgentServiceClient) error {
|
||||
ticker := time.NewTicker(a.cfg.UsageInterval)
|
||||
defer ticker.Stop()
|
||||
windowStart := a.clock().Unix()
|
||||
for {
|
||||
select {
|
||||
case <-ctx.Done():
|
||||
return ctx.Err()
|
||||
case <-ticker.C:
|
||||
}
|
||||
entries := a.usage.Collect()
|
||||
if len(entries) == 0 {
|
||||
windowStart = a.clock().Unix()
|
||||
continue
|
||||
}
|
||||
now := a.clock().Unix()
|
||||
_, err := client.ReportUsage(ctx, &agentv1.UsageReport{
|
||||
NodeUUID: a.NodeUUID(),
|
||||
WindowStartUnix: windowStart,
|
||||
WindowEndUnix: now,
|
||||
Entries: entries,
|
||||
})
|
||||
if err != nil {
|
||||
// Connection-level errors should restart the session; surface them.
|
||||
return err
|
||||
}
|
||||
windowStart = now
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user