Files
ModelRouter/internal/lua/kernel_perms_persist.go
JianFeeeee a07449d32e fix(plugins): 出站主机授权两处真缺陷 —— 已启动 worker 看不到新授权、授权不落盘
生产实证:channels 插件的 auth/state 在主机已授权后仍被拒
(auth/state 失败: host "copilot.tencent.com" is not an allowed outbound destination),
而 /api/plugins/channels/kernel 明明显示 hosts_granted=26。

缺陷一(授权对已启动 worker 无效):
  httpAllowlist 是不可变快照;bootPluginState 在 worker 启动时把指针绑进 Lua
  闭包,闭包与 worker 同寿。GrantExtraHosts 之后 ps.allowHTTP 被**整个换新**
  (newHTTPAllowlist 返回新实例),已启动的 worker 仍持旧指针 → 授权接口返回
  ok、设置页显示已批,请求却继续被拒。分片 VM(plugin_shard.go)同样中招。
  修法:allowlist 改为带 RWMutex 的**可变共享对象**,GrantExtraHosts 与
  rebuildOutboundAllowlistLocked 原地 add —— 所有闭包立即看到新授权,worker
  无需重建。rebuild 从"替换实例"改为"原地折叠",并在注释里写明为什么。

缺陷二(授权不持久化):
  .kernel-perms.json 只写 actions;hosts 授权只存在运行态 map,重启即失。
  设置页看不出异常(actions 有持久化,读起来像"都持久化了")。
  修法:文件行改为 jsonPluginPolicy{actions,all_read,hosts}(旧文件向后兼容,
  纯 actions 行照常解析);Load 恢复 extraHosts;Save 把每个插件的 host 授权
  与 actions 一起写入(支持"只批了主机、没批动作"的行)。

回归测试(internal/lua/plugin_http_test.go):
- TestGrantExtraHostsVisibleToAlreadyBootedWorker:按事故顺序构造(先启动
  worker → 再授权),断言同一批 worker 立即可用新授权 + 旧引用不再失效。
- TestHostGrantsPersistAcrossRestart:授权落盘 + 新 registry 读回 + 插件重载
  后授权保留。

go test ./internal/lua/ 全绿;build/vet 通过。
2026-10-13 06:57:28 +08:00

127 lines
4.2 KiB
Go

package lua
import (
"encoding/json"
"os"
)
// Kernel action grants are OPERATOR policy, persisted separately from both
// plugin state and the endpoint/visibility overrides.
//
// Why a separate file rather than reusing .access.json:
//
// - .access.json is a flat map[string]string ("name" or "name#path" -> spec).
// Kernel grants are a list plus a boolean, so folding them in would either
// change the shape of a file an older binary still reads, or force an
// encoding that cannot round-trip. An upgrade must not corrupt a running
// deployment's restrictions.
// - The two answer different questions — "who may call this endpoint" vs
// "what may this plugin do to the kernel" — and an operator looking at one
// should not have to parse the other.
//
// Shape: { "<plugin>": { "actions": ["reload", ...], "all_read": true } }
const kernelPermsFileName = ".kernel-perms.json"
func (ps *Plugins) kernelPermsFile() string {
if ps.dir == "" {
return ""
}
return ps.dir + string(os.PathSeparator) + kernelPermsFileName
}
// LoadKernelPermissions reads the persisted grants. A missing file is the normal
// fresh-install state (nothing granted); a corrupt file is logged and ignored
// rather than fatal — losing a grant is recoverable, refusing to boot is not.
func (ps *Plugins) LoadKernelPermissions() error {
path := ps.kernelPermsFile()
if path == "" {
return nil
}
b, err := os.ReadFile(path)
if os.IsNotExist(err) {
return nil
}
if err != nil {
return err
}
// The file holds BOTH kernel-action grants (map[string]jsonPermission, the
// historical shape) AND, for plugins that asked for them, the operator's
// extra-host grants (map[string]jsonPluginPolicy.Hosts). A bare
// {actions:[...]} row still parses into jsonPluginPolicy with empty Hosts,
// so old files load unchanged.
raw := map[string]jsonPluginPolicy{}
if err := json.Unmarshal(b, &raw); err != nil {
return err
}
ps.mu.Lock()
defer ps.mu.Unlock()
if ps.kernelPerms == nil {
ps.kernelPerms = kernelPermissionSet{}
}
if ps.extraHosts == nil {
ps.extraHosts = map[string][]string{}
}
for name, jp := range raw {
ps.kernelPerms[name] = KernelPermission{Actions: jp.Actions, AllRead: jp.AllRead}
if len(jp.Hosts) > 0 {
// Load the persisted grant into extraHosts WITHOUT folding it into the
// allowlist yet — sources are loaded later, and the caller owns that
// fold. GrantExtraHosts rebuilds the allowlist when sources are ready.
ps.extraHosts[name] = append([]string(nil), jp.Hosts...)
}
}
return nil
}
// saveKernelPermissionsLocked flushes the grant map. Caller holds ps.mu.
//
// Grants drop unknown action names on write. If an operator authorized an action
// that a later release removed, keeping the name would leave a stale entry that
// nothing can ever satisfy — and it would reappear in the UI as if it were live.
//
// Host grants travel in the same file (jsonPluginPolicy.Hosts): a granted host is
// the operator's explicit decision to let a plugin egress somewhere, and losing
// it across a restart (the previous behaviour) made grants silently vanish.
func (ps *Plugins) saveKernelPermissionsLocked() {
path := ps.kernelPermsFile()
if path == "" {
return
}
out := map[string]jsonPluginPolicy{}
something := false
for name, perm := range ps.kernelPerms {
valid := make([]string, 0, len(perm.Actions))
for _, a := range perm.Actions {
if _, ok := LookupKernelAction(a); ok {
valid = append(valid, a)
}
}
row := policyFromActions(KernelPermission{Actions: valid, AllRead: perm.AllRead})
if hs := ps.extraHosts[name]; len(hs) > 0 {
row.Hosts = append([]string(nil), hs...)
}
if len(valid) == 0 && !perm.AllRead && len(row.Hosts) == 0 {
continue // nothing granted: presence would be a lie
}
out[name] = row
something = true
}
// A plugin may hold host grants with NO kernel actions granted (operators
// approve hosts and actions independently). Emit those too.
for name, hs := range ps.extraHosts {
if len(hs) == 0 {
continue
}
if _, exists := out[name]; !exists {
out[name] = jsonPluginPolicy{Hosts: append([]string(nil), hs...)}
something = true
}
}
if !something {
_ = os.Remove(path)
return
}
writeJSONAtomic(path, out)
}