- plugin-llm 新 crate:llm/upstreams/usage_filter/llm_audit/stats/handlers/api/config/models/store - 下行契约(downstream_key/identity/check_quota/fail)与记账原语(usage::bump/read/bump_day)已下沉 ipass-core - usage_filter 改走 loc::enabled(Switch::Llm) + core::plugin::ctx_index(解除对 gateway 私有 ModuleConfig 的依赖) - gateway 收缩为壳:LlmHandler/LlmHealthHandler/AuditHandler 随 LlmPlugin 装配; /api/admin/trend 留 gateway(需 users+upstreams+apix 三方名册,组合根职责) - 行为等价:262 项回归全绿(gateway 102 + llm 160)
325 lines
13 KiB
Rust
325 lines
13 KiB
Rust
//! LLM 插件的管理 API:上游 CRUD / 连通性测试 / 用量查询。
|
||
//!
|
||
//! 趋势接口 `/api/admin/trend` **不在本表**——它需要 users + upstreams +
|
||
//! apix 三方名册(日桶键带日窗 id,按前缀扫不出来),属于组合逻辑,
|
||
//! 由 gateway 作为装配根实现(对方案的有意偏离,见 PLUGINIZATION_PLAN)。
|
||
//!
|
||
//! 鉴权语义与内核共用一套:`ipass_core::api::{Guard, Actor, run_guarded}`。
|
||
|
||
use ngx::core::{Pool, Status};
|
||
use ngx::ffi::ngx_http_request_t;
|
||
use ngx::http::HTTPStatus;
|
||
use serde_json::{Value, json};
|
||
|
||
use ipass_core::api::{Actor, Guard, run_guarded, user_or_401};
|
||
use ipass_core::{http, util};
|
||
|
||
use crate::llm;
|
||
use crate::stats;
|
||
use crate::upstreams;
|
||
|
||
// ===========================================================================
|
||
// 路由
|
||
// ===========================================================================
|
||
|
||
type Handler = fn(*mut ngx_http_request_t, Actor) -> Status;
|
||
|
||
/// 路由表。刻意用精确匹配而非正则——接口路径是固定的,
|
||
/// 正则只会引入「写错锚点导致前缀误匹配」的风险。
|
||
fn route(method: &str, uri: &str) -> Option<(Guard, Handler)> {
|
||
let entry: (Guard, Handler) = match (method, uri) {
|
||
("GET", "/api/admin/upstreams") => (Guard::Admin, h_admin_upstreams),
|
||
("POST", "/api/admin/upstreams/save") => (Guard::Admin, h_admin_upstream_save),
|
||
("POST", "/api/admin/upstreams/test") => (Guard::Admin, h_admin_upstream_test),
|
||
("POST", "/api/admin/upstreams/delete") => (Guard::Admin, h_admin_upstream_delete),
|
||
("GET", "/api/admin/usage") => (Guard::Admin, h_admin_usage),
|
||
("GET", "/api/v1/models") => (Guard::User, h_my_models),
|
||
("GET", "/api/v1/usage") => (Guard::User, h_my_usage),
|
||
("GET", "/api/v1/upstreams") => (Guard::User, h_my_upstreams),
|
||
_ => return None,
|
||
};
|
||
Some(entry)
|
||
}
|
||
|
||
/// 本表分发:认领则返回响应状态,不认领返回 `None`(交给下一个插件)。
|
||
pub fn dispatch(r: *mut ngx_http_request_t) -> Option<Status> {
|
||
let method = http::method_str(r);
|
||
let uri = http::uri(r);
|
||
|
||
let Some((guard, handler)) = route(&method, &uri) else {
|
||
return None;
|
||
};
|
||
|
||
Some(run_guarded(r, guard, handler))
|
||
}
|
||
|
||
// ===========================================================================
|
||
// 管理员:LLM 上游
|
||
// ===========================================================================
|
||
|
||
fn h_admin_upstreams(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||
http::ok(
|
||
r,
|
||
json!({ "items": upstreams::view_all(&upstreams::all(), &http::host_port(r)) }),
|
||
"ok",
|
||
)
|
||
}
|
||
|
||
fn h_admin_upstream_save(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||
let p = http::params(r);
|
||
|
||
// 表单没提交 enabled 字段时视为「启用」——否则改个备注就会把上游改停,
|
||
// 这种误伤在管理台上极难自查
|
||
let enabled = match p.get("enabled") {
|
||
None => true,
|
||
Some(v) => util::to_bool(Some(v)),
|
||
};
|
||
|
||
// models 缺省 = 保持原清单;显式传(哪怕是空串)= 覆盖。
|
||
// 分隔符同时接受换行/逗号/顿号/分号——粘一份厂商文档里的模型列表是常见操作。
|
||
let models = p.get("models").map(|raw| parse_model_list(raw));
|
||
|
||
let slot_raw = p.get("slot").map(String::as_str).unwrap_or("");
|
||
|
||
match upstreams::upsert(
|
||
slot_raw,
|
||
p.get("base").map(String::as_str).unwrap_or(""),
|
||
models,
|
||
p.get("auth"),
|
||
enabled,
|
||
p.get("note"),
|
||
) {
|
||
Ok(up) => {
|
||
// 用校验后的 slot 回显;直接拿原始入参会把大小写差异带进响应
|
||
let slot = upstreams::validate_slot(slot_raw).unwrap_or_default();
|
||
util::log_notice(&format!(
|
||
"[api] 保存 LLM 上游 {slot} enabled={enabled} models={}",
|
||
up.models.len()
|
||
));
|
||
// base_url 带「调用方正在用的端口」:见 upstreams::base_url 的说明
|
||
http::ok(r, upstreams::view(&slot, &up, &http::host_port(r)), "已保存")
|
||
}
|
||
Err(e) => http::fail(r, HTTPStatus::BAD_REQUEST, &e),
|
||
}
|
||
}
|
||
|
||
/// 把管理台/接口传入的模型列表串拆成模型名数组。
|
||
///
|
||
/// 允许的写法(混用也认):`a,b` / `a、b` / `a;b` / 一行一个。
|
||
pub fn parse_model_list(raw: &str) -> Vec<String> {
|
||
raw.split(|c: char| c == ',' || c == '、' || c == ';' || c == '\n' || c == '\r' || c == '\t')
|
||
.map(|s| util::trim(s).to_owned())
|
||
.filter(|s| !s.is_empty())
|
||
.collect()
|
||
}
|
||
|
||
fn h_admin_upstream_delete(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||
let p = http::params(r);
|
||
match upstreams::remove(p.get("slot").map(String::as_str).unwrap_or("")) {
|
||
Ok(slot) => {
|
||
util::log_notice(&format!("[api] 删除 LLM 上游 {slot}"));
|
||
http::ok(r, json!({ "slot": slot }), "已删除")
|
||
}
|
||
Err(e) => http::fail(r, HTTPStatus::BAD_REQUEST, &e),
|
||
}
|
||
}
|
||
|
||
/// 管理员:上游连通性测试(管理台「测试」按钮)。
|
||
///
|
||
/// 面板把 slot 放在 query(`?slot=x`),把**要发给上游的探测 JSON**(真实
|
||
/// chat/completions 请求,max_tokens=1)放在请求体——模块对体零改写。
|
||
/// 实现上校验 slot 存在后 `ngx_http_internal_redirect` 到管理 server 的
|
||
/// 内部 location `/_upstream_test/<slot>/v1/chat/completions`:
|
||
///
|
||
/// * `proxy_pass $llm_target` 原样复用 LLM 链路的取 slot / 拼路径 / 换真 Key
|
||
/// 逻辑(见 `llm::slot_of` 对 `/_upstream_test` 前缀的识别);
|
||
/// * 请求体在 API 两段式读取里已经进 `r->request_body`,内部重定向后由
|
||
/// proxy 原样转发;
|
||
/// * 上游的真实状态码与响应体直接透传给面板——通不通、错在哪个厂商,
|
||
/// 面板看到的就是厂商原话。
|
||
///
|
||
/// 为什么不能在浏览器里直测:上游真 Key 只存在服务端(面板只有掩码),
|
||
/// 且 lm 域名与 mdm 不同源,fetch 会被 CORS 拦住——测试必须在服务端发起。
|
||
fn h_admin_upstream_test(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||
let slot_raw = http::params(r).get("slot").cloned().unwrap_or_default();
|
||
let slot = match upstreams::validate_slot(&slot_raw) {
|
||
Ok(s) => s,
|
||
Err(e) => return http::fail(r, HTTPStatus::BAD_REQUEST, &e),
|
||
};
|
||
if !llm::upstream_live(&slot) {
|
||
return http::fail(r, HTTPStatus::NOT_FOUND, &format!("上游 {slot} 不存在"));
|
||
}
|
||
|
||
let uri = format!("/_upstream_test/{slot}/v1/chat/completions");
|
||
util::log_notice(&format!("[api] 测试 LLM 上游 {slot}"));
|
||
|
||
// ⚠️ nginx 的 internal_redirect 只**浅拷贝 ngx_str_t 结构体**(data 指针 + len),
|
||
// 不拷贝字节内容。若把栈上 String 的指针直接递过去,函数一返回堆内存就被回收,
|
||
// 之后 log 阶段读 r->uri 全是 NUL——健康标记、slot 解析全部静默失效(实测踩过)。
|
||
// 必须把 URI 字节拷进**请求池**,让它活到请求结束。
|
||
//
|
||
// SAFETY: r 有效;池内存在整个请求生命周期内有效。
|
||
unsafe {
|
||
let pool = Pool::from_ngx_pool((*r).pool);
|
||
let bytes = uri.as_bytes();
|
||
let data = pool.alloc(bytes.len()) as *mut u8;
|
||
if data.is_null() {
|
||
return http::fail(r, HTTPStatus::INTERNAL_SERVER_ERROR, "内存分配失败");
|
||
}
|
||
core::ptr::copy_nonoverlapping(bytes.as_ptr(), data, bytes.len());
|
||
|
||
let mut u = nginx_sys::ngx_str_t {
|
||
len: bytes.len(),
|
||
data,
|
||
};
|
||
let mut args = nginx_sys::ngx_str_t {
|
||
len: 0,
|
||
data: std::ptr::null_mut(),
|
||
};
|
||
nginx_sys::ngx_http_internal_redirect(r, &mut u, &mut args);
|
||
}
|
||
|
||
Status::NGX_DONE
|
||
}
|
||
|
||
/// 管理员:LLM 用量总览(每人 / 每上游 / 合计)。
|
||
///
|
||
/// 口径是**请求数**,且只统计真正转发出去的请求(见 `stats` 模块头注释)。
|
||
fn h_admin_usage(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||
http::ok(r, stats::overview(util::now()), "ok")
|
||
}
|
||
|
||
// ===========================================================================
|
||
// 账号态
|
||
// ===========================================================================
|
||
|
||
/// 账号态:看看自己能用哪些上游。
|
||
///
|
||
/// 刻意只回显 slot / base_url / **模型清单**——**不回显任何密钥**(连脱敏都不给,
|
||
/// 上游密钥与调用方无关,没有知道的必要)。
|
||
fn h_my_upstreams(r: *mut ngx_http_request_t, actor: Actor) -> Status {
|
||
// 回显的 base_url 必须带「调用方正在用的端口」,否则客户端照着拼会打不通
|
||
let port = http::host_port(r);
|
||
let user = match user_or_401(r, actor) {
|
||
Ok(u) => u,
|
||
Err(e) => return e,
|
||
};
|
||
|
||
let items: Vec<Value> = upstreams::available(&user.allowed_models)
|
||
.into_iter()
|
||
.map(|(slot, up)| {
|
||
let (status, code, checked_at) = match upstreams::health(&slot) {
|
||
Some((ok, c, ts)) => (if ok { "ok" } else { "failed" }, c, ts),
|
||
None => ("unknown", 0, 0),
|
||
};
|
||
json!({
|
||
"slot": slot,
|
||
"base_url": upstreams::base_url(&slot, &up, &port),
|
||
"anthropic_base_url": upstreams::anthropic_base_url(&slot, &up, &port),
|
||
"models": up.models,
|
||
"note": up.note,
|
||
"status": status,
|
||
"status_code": if code > 0 { json!(code) } else { Value::Null },
|
||
"checked_at": checked_at,
|
||
})
|
||
})
|
||
.collect();
|
||
|
||
http::ok(r, json!({ "items": items }), "ok")
|
||
}
|
||
|
||
/// 账号态:列出自己**可用的模型**(上游登记的模型清单,按白名单过滤)。
|
||
///
|
||
/// 与 LLM 域名下的 `GET /v1/models` 是同一份数据:那边给 OpenAI SDK 用
|
||
/// (形状 `{object:"list",data:[{id,...}]}`),这边给管理/自研客户端用
|
||
/// (形状 `{code,msg,data}`,带 base_url 便于直接拼请求)。
|
||
fn h_my_models(r: *mut ngx_http_request_t, actor: Actor) -> Status {
|
||
// 同 h_my_upstreams:base_url 跟随调用方所用的端口
|
||
let port = http::host_port(r);
|
||
let user = match user_or_401(r, actor) {
|
||
Ok(u) => u,
|
||
Err(e) => return e,
|
||
};
|
||
|
||
let available = upstreams::available(&user.allowed_models);
|
||
let mut items: Vec<Value> = Vec::new();
|
||
for (slot, up) in &available {
|
||
for model in &up.models {
|
||
items.push(json!({
|
||
"model": model,
|
||
"slot": slot,
|
||
"base_url": upstreams::base_url(slot, up, &port),
|
||
"note": up.note,
|
||
}));
|
||
}
|
||
}
|
||
|
||
http::ok(
|
||
r,
|
||
json!({
|
||
"items": items,
|
||
"upstreams": available
|
||
.iter()
|
||
.map(|(slot, up)| {
|
||
let (status, code, checked_at) = match upstreams::health(slot) {
|
||
Some((ok, c, ts)) => (if ok { "ok" } else { "failed" }, c, ts),
|
||
None => ("unknown", 0, 0),
|
||
};
|
||
json!({
|
||
"slot": slot,
|
||
"base_url": upstreams::base_url(slot, up, &port),
|
||
"anthropic_base_url": upstreams::anthropic_base_url(slot, up, &port),
|
||
"models": up.models,
|
||
"status": status,
|
||
"status_code": if code > 0 { json!(code) } else { Value::Null },
|
||
"checked_at": checked_at,
|
||
})
|
||
})
|
||
.collect::<Vec<_>>(),
|
||
}),
|
||
"ok",
|
||
)
|
||
}
|
||
|
||
/// 账号态:查自己的 LLM 用量(今日 / 累计 / 最近调用 + 分上游明细)。
|
||
fn h_my_usage(r: *mut ngx_http_request_t, actor: Actor) -> Status {
|
||
let user = match user_or_401(r, actor) {
|
||
Ok(u) => u,
|
||
Err(e) => return e,
|
||
};
|
||
|
||
let now = util::now();
|
||
let usage = stats::user(&user.username, now);
|
||
// token 计量(body filter 从上游 usage 抓的;抓不到 = 0,不补估)
|
||
let tin = stats::tokens_user(&user.username, now);
|
||
let models: Vec<Value> = stats::user_models(&user.username, now)
|
||
.into_iter()
|
||
.map(|(slot, u)| {
|
||
json!({
|
||
"slot": slot,
|
||
"today": u.today,
|
||
"total": u.total,
|
||
"last_ts": u.last_ts,
|
||
"last_ts_text": if u.last_ts > 0 { util::fmt_time(u.last_ts) } else { String::new() },
|
||
})
|
||
})
|
||
.collect();
|
||
|
||
http::ok(
|
||
r,
|
||
json!({
|
||
"username": user.username,
|
||
"display_name": user.display_name,
|
||
"today": usage.today,
|
||
"total": usage.total,
|
||
"last_ts": usage.last_ts,
|
||
"last_ts_text": if usage.last_ts > 0 { util::fmt_time(usage.last_ts) } else { String::new() },
|
||
"tokens": tin.json(),
|
||
"models": models,
|
||
// 限额一起回显:客户端可据此自我节流,不必等 429
|
||
"limits": { "rpm": user.llm_rpm, "daily": user.llm_daily },
|
||
}),
|
||
"ok",
|
||
)
|
||
}
|