feat(deploy): 新增 .11 一键更新脚本 + 上游多模型/模型发现/用量统计
功能(本轮 5 项需求的落地):
- 上游可挂多模型:llm-upstreams.json 每上游新增 models 清单(仅用于模型发现,不参与路由)
- 模型发现:GET /v1/models(OpenAI {object:list} 形状,网关本地应答)+ /api/v1/models + /api/v1/upstreams
- 用量统计:/api/admin/usage(管理台「用量统计」Tab)+ /api/v1/usage + /me 的 llm_usage 摘要
口径:只统计真正转发出去的请求,401/403/404/429 不计;与配额计数 st: 刻意分开
- 账号展示名 display_name(纯展示,不进 ukey:)
- 修复 /v1 版本段重复拼接(llm::join_target)→ 上游 base 自带 /v1 时不再 404 空体
- 修复删账号未清用量键(stats::forget)
部署脚本 deploy/update-ipass.sh:
- 拉镜像 → tar 备份数据 → 删旧容器 → 按旧容器复刻端口/挂载/网络起新容器 → 自检
- 参数从 docker inspect 静态配置读取(docker port 对停止态容器返回空,会让端口静默回落)
- 端口按 host:container 去重(inspect 同一映射含 IPv4/IPv6 两项,重复 -p 会 address already in use)
- --bg 用 setsid/start_new_session 真正脱离会话(仅 nohup & 会被父进程组连带杀掉)
- 旧镜像自动打 rollback-<时间> 标签,收尾打印回滚命令
回归:bash tests/run.sh all → 网关 78 + LLM 75 = 153 条断言全绿。
文档:readme §12.4 / §15 / §16、CLAUDE.md 常用操作 + 排障速查已同步。
This commit is contained in:
1 parent
94a48bef6e
commit
e38237dc4a
22 files changed
+1719
-93
No files matched your search
+165
-25
@@ -24,6 +24,7 @@ use crate::http;
|
||||
use crate::models::User;
|
||||
use crate::session;
|
||||
use crate::state;
|
||||
use crate::stats;
|
||||
use crate::upstreams;
|
||||
use crate::users;
|
||||
use crate::util;
|
||||
@@ -77,6 +78,7 @@ fn route(method: &str, uri: &str) -> Option<(Guard, Handler)> {
|
||||
("POST", "/api/admin/users/delete") => (Guard::Admin, h_admin_user_delete),
|
||||
("POST", "/api/admin/users/models") => (Guard::Admin, h_admin_user_models),
|
||||
("POST", "/api/admin/users/limits") => (Guard::Admin, h_admin_user_limits),
|
||||
("POST", "/api/admin/users/display") => (Guard::Admin, h_admin_user_display),
|
||||
("GET", "/api/admin/domains") => (Guard::Admin, h_admin_domains),
|
||||
("POST", "/api/admin/domains/review") => (Guard::Admin, h_admin_domain_review),
|
||||
("POST", "/api/admin/domains/create") => (Guard::Admin, h_admin_domain_create),
|
||||
@@ -85,7 +87,10 @@ fn route(method: &str, uri: &str) -> Option<(Guard, Handler)> {
|
||||
("GET", "/api/admin/upstreams") => (Guard::Admin, h_admin_upstreams),
|
||||
("POST", "/api/admin/upstreams/save") => (Guard::Admin, h_admin_upstream_save),
|
||||
("POST", "/api/admin/upstreams/delete") => (Guard::Admin, h_admin_upstream_delete),
|
||||
("GET", "/api/admin/usage") => (Guard::Admin, h_admin_usage),
|
||||
("GET", "/api/v1/me") => (Guard::User, h_me),
|
||||
("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),
|
||||
("GET", "/api/v1/domains") => (Guard::User, h_my_domains),
|
||||
("POST", "/api/v1/domains/apply") => (Guard::User, h_apply),
|
||||
@@ -217,24 +222,31 @@ fn h_admin_users(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
*counts.entry(d.owner).or_insert(0) += 1;
|
||||
}
|
||||
|
||||
let now = util::now();
|
||||
// 一次扫表拿到所有人的 LLM 用量(含各自的上游明细),避免逐账号查共享内存
|
||||
let usage: HashMap<String, stats::UserUsage> =
|
||||
stats::all_users(now).into_iter().map(|u| (u.username.clone(), u)).collect();
|
||||
|
||||
let items: Vec<Value> = users::all()
|
||||
.iter()
|
||||
.map(|u| {
|
||||
let mut v = users::view(u);
|
||||
let mut usage_json = json!({ "today": 0, "total": 0, "last_ts": 0, "last_ts_text": "" });
|
||||
if let Some(uu) = usage.get(&u.username) {
|
||||
usage_json = uu.usage.json();
|
||||
}
|
||||
if let Value::Object(ref mut m) = v {
|
||||
m.insert(
|
||||
"domain_count".to_owned(),
|
||||
json!(counts.get(&u.username).copied().unwrap_or(0)),
|
||||
);
|
||||
// 今日 LLM 用量:读 shm 的日窗计数(与热路径同一数据源)
|
||||
// 今日 LLM 用量:与热路径同一数据源(共享内存)——注意口径是
|
||||
// **已转发**的请求,不含被限流/越权拒掉的那些
|
||||
m.insert(
|
||||
"llm_today".to_owned(),
|
||||
json!(state::window_count(
|
||||
config::KEY_ST_PREFIX,
|
||||
&u.username,
|
||||
util::day_id(util::now())
|
||||
)),
|
||||
json!(usage.get(&u.username).map(|x| x.usage.today).unwrap_or(0)),
|
||||
);
|
||||
m.insert("llm_usage".to_owned(), usage_json);
|
||||
}
|
||||
v
|
||||
})
|
||||
@@ -249,14 +261,15 @@ fn h_admin_user_create(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
|
||||
match users::create(
|
||||
p.get("username").map(String::as_str).unwrap_or(""),
|
||||
p.get("display_name").map(String::as_str).unwrap_or(""),
|
||||
p.get("password").map(String::as_str).unwrap_or(""),
|
||||
p.get("role").map(String::as_str).unwrap_or("user"),
|
||||
grant_temp,
|
||||
) {
|
||||
Ok(user) => {
|
||||
util::log_notice(&format!(
|
||||
"[api] 创建账号 {} role={} temp_grant={grant_temp}",
|
||||
user.username, user.role
|
||||
"[api] 创建账号 {} name={} role={} temp_grant={grant_temp}",
|
||||
user.username, user.display_name, user.role
|
||||
));
|
||||
http::ok(r, users::view(&user), "账号创建成功")
|
||||
}
|
||||
@@ -264,6 +277,17 @@ fn h_admin_user_create(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
}
|
||||
}
|
||||
|
||||
/// 设置展示名(中文姓名等)。传空串 = 清空,前端回落到账号名。
|
||||
fn h_admin_user_display(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
let p = http::params(r);
|
||||
let username = p.get("username").map(|s| util::trim(s).to_owned()).unwrap_or_default();
|
||||
|
||||
match users::set_display_name(&username, p.get("display_name").map(String::as_str).unwrap_or("")) {
|
||||
Ok(u) => http::ok(r, users::view(&u), "已更新展示名"),
|
||||
Err(e) => http::fail(r, HTTPStatus::BAD_REQUEST, &e),
|
||||
}
|
||||
}
|
||||
|
||||
fn h_admin_user_key(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
let p = http::params(r);
|
||||
let username = p.get("username").map(|s| util::trim(s).to_owned()).unwrap_or_default();
|
||||
@@ -399,6 +423,8 @@ fn h_me(r: *mut ngx_http_request_t, actor: Actor) -> Status {
|
||||
"domains".to_owned(),
|
||||
json!(domains::view_all(&domains::by_owner(&user.username))),
|
||||
);
|
||||
// 用量摘要(明细走 /api/v1/usage)——让客户端一个请求就能决定要不要自我节流
|
||||
m.insert("llm_usage".to_owned(), stats::user(&user.username, util::now()).json());
|
||||
m.insert(
|
||||
"policy".to_owned(),
|
||||
json!({
|
||||
@@ -493,11 +519,16 @@ fn h_admin_upstream_save(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
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"),
|
||||
@@ -505,13 +536,26 @@ fn h_admin_upstream_save(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
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}"));
|
||||
util::log_notice(&format!(
|
||||
"[api] 保存 LLM 上游 {slot} enabled={enabled} models={}",
|
||||
up.models.len()
|
||||
));
|
||||
http::ok(r, upstreams::view(&slot, &up), "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("")) {
|
||||
@@ -523,7 +567,18 @@ fn h_admin_upstream_delete(r: *mut ngx_http_request_t, _actor: Actor) -> Status
|
||||
}
|
||||
}
|
||||
|
||||
/// 设置某账号可用的 LLM 上游白名单。`models` 传空串 = 解除限制。
|
||||
/// 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)。
|
||||
/// 想看 token 级用量得改架构(自建 upstream 拦响应体),不在本项目范围。
|
||||
fn h_admin_usage(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
http::ok(r, stats::overview(util::now()), "ok")
|
||||
}
|
||||
|
||||
/// 设置某账号可用的 LLM 上游白名单(`models` 字段 = 上游 slot 列表,传空串 = 解除限制)。
|
||||
///
|
||||
/// 注意与**上游登记模型清单**(`Upstream.models`)是两件事:这里管"这个账号能用哪几家
|
||||
/// 上游",那里管"这家上游提供哪些模型"。白名单按 slot 生效,模型发现再按白名单过滤。
|
||||
fn h_admin_user_models(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
let p = http::params(r);
|
||||
let username = p.get("username").map(|s| util::trim(s).to_owned()).unwrap_or_default();
|
||||
@@ -589,30 +644,35 @@ fn h_admin_user_limits(r: *mut ngx_http_request_t, _actor: Actor) -> Status {
|
||||
}
|
||||
}
|
||||
|
||||
/// 账号态:看看自己能用哪些上游。
|
||||
///
|
||||
/// 刻意只回显 slot 与拼接好的 base_url——**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 {
|
||||
let Actor::User(user) = actor else {
|
||||
return http::fail(
|
||||
/// 账号态接口的公共前置:取账号身份,取不到就回 401。
|
||||
fn user_or_401(r: *mut ngx_http_request_t, actor: Actor) -> Result<User, Status> {
|
||||
match actor {
|
||||
Actor::User(u) => Ok(u),
|
||||
_ => Err(http::fail(
|
||||
r,
|
||||
HTTPStatus::UNAUTHORIZED,
|
||||
"API Key 无效或缺失,请在请求头携带 X-API-Key",
|
||||
);
|
||||
)),
|
||||
}
|
||||
}
|
||||
|
||||
/// 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 {
|
||||
let user = match user_or_401(r, actor) {
|
||||
Ok(u) => u,
|
||||
Err(e) => return e,
|
||||
};
|
||||
|
||||
let items: Vec<Value> = upstreams::all()
|
||||
let items: Vec<Value> = upstreams::available(&user.allowed_models)
|
||||
.into_iter()
|
||||
.filter(|(slot, up)| {
|
||||
up.enabled
|
||||
&& (user.allowed_models.is_empty()
|
||||
|| user.allowed_models.iter().any(|s| s == slot))
|
||||
})
|
||||
.map(|(slot, up)| {
|
||||
json!({
|
||||
"slot": slot,
|
||||
"base_url": format!("https://llm.yourdomain.com/{slot}"),
|
||||
"base_url": upstreams::base_url(&slot, &up),
|
||||
"models": up.models,
|
||||
"note": up.note,
|
||||
})
|
||||
})
|
||||
@@ -620,3 +680,83 @@ fn h_my_upstreams(r: *mut ngx_http_request_t, actor: Actor) -> Status {
|
||||
|
||||
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 {
|
||||
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),
|
||||
"note": up.note,
|
||||
}));
|
||||
}
|
||||
}
|
||||
|
||||
http::ok(
|
||||
r,
|
||||
json!({
|
||||
"items": items,
|
||||
"upstreams": available
|
||||
.iter()
|
||||
.map(|(slot, up)| json!({
|
||||
"slot": slot,
|
||||
"base_url": upstreams::base_url(slot, up),
|
||||
"models": up.models,
|
||||
}))
|
||||
.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);
|
||||
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() },
|
||||
"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",
|
||||
)
|
||||
}
|
||||
@@ -27,6 +27,8 @@ pub const UPSTREAM_SLOT_MAX_LEN: usize = 32;
|
||||
pub const UPSTREAM_BASE_MAX_LEN: usize = 200;
|
||||
pub const USERNAME_MIN_LEN: usize = 2;
|
||||
pub const USERNAME_MAX_LEN: usize = 32;
|
||||
/// 展示名(中文姓名等)长度上限。按**字符数**算,一个汉字算 1。
|
||||
pub const DISPLAY_NAME_MAX_LEN: usize = 32;
|
||||
/// 子域名总长上限:首字符 + 最多 61 个后续字符
|
||||
pub const SUBDOMAIN_MAX_LEN: usize = 62;
|
||||
|
||||
@@ -86,6 +88,25 @@ pub const KEY_RL_PREFIX: &[u8] = b"rl:";
|
||||
/// 键前缀:LLM 每日配额计数。键 = `st:<username>`,value = `<日窗id>:<计数>`。
|
||||
/// 同时是管理台「今日用量」的数据源。
|
||||
pub const KEY_ST_PREFIX: &[u8] = b"st:";
|
||||
/// 键前缀:LLM **每用户**用量统计(口径 = 已放行的请求)。
|
||||
///
|
||||
/// 键 = `ust:<username>`,value = `<日窗id>:<今日数>:<累计数>:<最近调用时间戳>`。
|
||||
/// 与 `st:` 的区别:`st:` 是**配额执行**用的计数(先增后判,被拒的请求也计入),
|
||||
/// 这里统计的是**真的转发出去**的请求——两者在超限时会差几个数,属预期。
|
||||
pub const KEY_UST_PREFIX: &[u8] = b"ust:";
|
||||
/// 键前缀:LLM **每用户 × 每上游**用量统计。
|
||||
///
|
||||
/// 键 = `ums:<username>:<slot>`(username 与 slot 都不含 `:`,切分安全),
|
||||
/// value 同 `ust:`。管理台的「每个模型用量」就是把这些键按 slot 聚合出来的。
|
||||
pub const KEY_UMS_PREFIX: &[u8] = b"ums:";
|
||||
|
||||
/// LLM 代理对外域名,用于账号态 API 回显 `base_url`。
|
||||
/// 改域名只改这里(`conf/nginx.conf` 的 LLM server_name 要同步改)。
|
||||
pub const LLM_PUBLIC_HOST: &str = "lm.mayishangshu.cn";
|
||||
/// 单个模型名长度上限
|
||||
pub const MODEL_NAME_MAX_LEN: usize = 128;
|
||||
/// 单个上游可登记的模型数量上限(挡住把整个模型库灌进 shm/JSON 的手滑)
|
||||
pub const UPSTREAM_MODELS_MAX: usize = 200;
|
||||
/// 共享内存 value 内部分隔符(ASCII US)。
|
||||
/// 后端地址只含 `[A-Za-z0-9._-]` 与 `:`,用户名只含 `[A-Za-z0-9_-]`,
|
||||
/// 都不可能包含该字节,所以切分是安全的。
|
||||
|
||||
@@ -45,6 +45,7 @@ mod llm;
|
||||
mod models;
|
||||
mod session;
|
||||
mod state;
|
||||
mod stats;
|
||||
mod store;
|
||||
mod upstreams;
|
||||
mod util;
|
||||
@@ -268,6 +269,12 @@ impl HttpRequestHandler for LlmHandler {
|
||||
);
|
||||
};
|
||||
|
||||
// 「列出可用模型」是网关自己的元信息接口(OpenAI 兼容形状),
|
||||
// 在查上游之前就地应答——否则 `/v1/models` 会被当成 slot=v1 → 404。
|
||||
if llm::is_models_probe(r) {
|
||||
return llm::respond_models(r, &allowed);
|
||||
}
|
||||
|
||||
if state::get_upstream(&slot).is_none() {
|
||||
return llm::fail(
|
||||
r,
|
||||
@@ -295,6 +302,11 @@ impl HttpRequestHandler for LlmHandler {
|
||||
return resp;
|
||||
}
|
||||
|
||||
// 用量统计:走到这里说明请求**真的会被转发**,才计入。
|
||||
// 被拒的请求(401/403/404/429)不进统计——那是给人看的数字,
|
||||
// 把拒绝也算进"用量"只会误导(与配额计数 st: 的语义刻意不同)。
|
||||
stats::record(&username, &slot, util::now());
|
||||
|
||||
Status::NGX_DECLINED
|
||||
}
|
||||
}
|
||||
|
||||
+79
-1
@@ -157,6 +157,46 @@ fn host_of(base: &str) -> &str {
|
||||
}
|
||||
}
|
||||
|
||||
/// 拆 base:`https://api.deepseek.com/v1` → `("https://api.deepseek.com", "/v1")`
|
||||
fn split_origin(base: &str) -> (&str, &str) {
|
||||
let Some(i) = base.find("://") else {
|
||||
return (base, "");
|
||||
};
|
||||
let after = &base[i + 3..];
|
||||
match after.find('/') {
|
||||
Some(j) => (&base[..i + 3 + j], &after[j..]),
|
||||
None => (base, ""),
|
||||
}
|
||||
}
|
||||
|
||||
/// 拼上游完整 URL:`base + 客户端路径(已剥掉 slot)`。
|
||||
///
|
||||
/// **唯一的例外是版本段重复**:上游 base 自带路径(如 DeepSeek 的
|
||||
/// `https://api.deepseek.com/v1`)而客户端路径又正好以同一段开头时,
|
||||
/// 只保留一份。否则会拼出 `/v1/v1/chat/completions`。
|
||||
///
|
||||
/// ```text
|
||||
/// base .../v1 + /v1/chat/completions → .../v1/chat/completions (去重)
|
||||
/// base .../v1 + /chat/completions → .../v1/chat/completions (照常前缀)
|
||||
/// base .../llm + /v1/chat/completions → .../llm/v1/chat/completions
|
||||
/// ```
|
||||
///
|
||||
/// 这条规则是**实测踩坑换来的**:厂商文档里的 base_url 通常自带 `/v1`
|
||||
/// (`https://api.deepseek.com/v1`),管理员照抄进 `llm-upstreams.json`,
|
||||
/// 客户端又按 OpenAI 习惯把 `/v1` 带上 → 上游 404 且**响应体为空**,
|
||||
/// 从网关侧完全看不出是拼接问题。去重后两种写法都通,
|
||||
/// base_url 也就能做到「把厂商文档里的 host 换成网关域名」的逐个字符替换。
|
||||
fn join_target(base: &str, path: &str) -> String {
|
||||
let (origin, base_path) = split_origin(base);
|
||||
if !base_path.is_empty() {
|
||||
let with_slash = format!("{base_path}/");
|
||||
if path == base_path || path.starts_with(&with_slash) {
|
||||
return format!("{origin}{path}");
|
||||
}
|
||||
}
|
||||
format!("{base}{path}")
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 响应
|
||||
// ===========================================================================
|
||||
@@ -253,6 +293,44 @@ pub unsafe fn check_quota(
|
||||
None
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 模型发现(本地应答,不去上游)
|
||||
// ===========================================================================
|
||||
|
||||
/// 是不是"列出模型"请求:`GET /v1/models`(OpenAI SDK 的 `client.models.list()`)。
|
||||
///
|
||||
/// **必须在查 slot 之前判断**:URL 第一段被当作 slot,`/v1/models` 天然会被
|
||||
/// 解析成 `slot=v1` → 「上游不存在」404。这两个路径是本网关自己的元信息接口,
|
||||
/// 与"某个上游的模型清单"(`GET /<slot>/v1/models`,照常透传给厂商)互不干扰。
|
||||
pub fn is_models_probe(r: *mut ngx_http_request_t) -> bool {
|
||||
if !crate::http::method_str(r).eq_ignore_ascii_case("GET") {
|
||||
return false;
|
||||
}
|
||||
let uri = crate::http::uri(r);
|
||||
let uri = uri.trim_end_matches('/');
|
||||
uri == "/v1/models" || uri == "/models"
|
||||
}
|
||||
|
||||
/// 本地应答模型列表,形状与 OpenAI 的 `/v1/models` 一致。
|
||||
///
|
||||
/// 客户端 SDK 只认这个形状,返回项目自己的 `{code,msg,data}` 会让 SDK
|
||||
/// 报"无法解析的响应"——与错误体必须 OpenAI 兼容是同一个理由。
|
||||
pub fn respond_models(r: *mut ngx_http_request_t, allowed: &[String]) -> Status {
|
||||
let data: Vec<serde_json::Value> = upstreams::models_flat(allowed)
|
||||
.into_iter()
|
||||
.map(|(model, slot)| {
|
||||
json!({ "id": model, "object": "model", "created": 0, "owned_by": slot })
|
||||
})
|
||||
.collect();
|
||||
|
||||
let body = json!({ "object": "list", "data": data }).to_string();
|
||||
// 与错误路径同一手法:自己把响应发出去,再把状态码交回 nginx。
|
||||
// 注意**不能**返回 NGX_OK——access 阶段的 OK 语义是"放行、继续走 content",
|
||||
// 那样 `proxy_pass` 照样会执行,客户端拿到模型列表的同时上游被真调一次。
|
||||
let _ = crate::http::send(r, HTTPStatus::OK, &body);
|
||||
Status(HTTPStatus::OK.0 as ngx_int_t)
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 变量 get_handler
|
||||
// ===========================================================================
|
||||
@@ -299,7 +377,7 @@ pub extern "C" fn var_get(
|
||||
let payload = state::get_upstream(&slot)?;
|
||||
let (base, auth) = upstreams::split_payload(&payload);
|
||||
Some(match data {
|
||||
VAR_TARGET => format!("{}{}", base, upstream_uri(r)),
|
||||
VAR_TARGET => join_target(base, &upstream_uri(r)),
|
||||
VAR_HOST => host_of(base).to_owned(),
|
||||
VAR_AUTH => auth.to_owned(),
|
||||
_ => return None,
|
||||
|
||||
@@ -16,6 +16,12 @@ pub type UserTable = BTreeMap<String, User>;
|
||||
#[derive(Debug, Clone, Serialize, Deserialize)]
|
||||
pub struct User {
|
||||
pub username: String,
|
||||
/// 中文姓名等展示名。**只用于展示**,不参与鉴权、不参与热路径查询。
|
||||
///
|
||||
/// `username` 规则是 ASCII(`valid_username`),中文名塞不进账号名,
|
||||
/// 所以「吴桂荣」这类实名单独放这里;留空时前端回落到 `username`。
|
||||
#[serde(default)]
|
||||
pub display_name: String,
|
||||
/// `admin` | `user`
|
||||
pub role: String,
|
||||
pub salt: String,
|
||||
@@ -67,6 +73,14 @@ pub type UpstreamTable = BTreeMap<String, Upstream>;
|
||||
pub struct Upstream {
|
||||
/// 上游基址,含 scheme,如 `https://api.openai.com`(可带端口)
|
||||
pub base: String,
|
||||
/// 该上游**挂载的模型名清单**(如 `deepseek-chat` / `deepseek-reasoner`)。
|
||||
///
|
||||
/// 一个上游(一家厂商)通常挂多个模型,客户端要能先"列出有什么"再选。
|
||||
/// 这里只是网关侧的**清单**,用于 `GET /v1/models` 之类的模型发现;
|
||||
/// 转发本身仍按 URL 里的 slot 走,**不校验 model 字段**(access 阶段读不到 body)。
|
||||
/// 留空表示"清单未知"——模型发现接口就不列该上游的模型,但转发照常可用。
|
||||
#[serde(default)]
|
||||
pub models: Vec<String>,
|
||||
/// 完整的 Authorization 头值,如 `Bearer sk-xxx`。
|
||||
/// 留空表示不透传鉴权头(本地免鉴权后端,如 vLLM / Ollama)
|
||||
#[serde(default)]
|
||||
|
||||
@@ -219,6 +219,29 @@ pub fn clear_prefix(prefix: &[u8]) {
|
||||
}
|
||||
}
|
||||
|
||||
/// 遍历某前缀下的全部条目,返回 `(去掉前缀的键名, payload)`。
|
||||
///
|
||||
/// 这是**低频读**接口(管理台聚合用量统计),不是热路径——每次都会克隆一份
|
||||
/// 字符串出来,别拿它做每请求调用。`expires_at` 刻意不返回:统计键永不过期,
|
||||
/// 调用方用 payload 里的日窗 id 自行判断"今日"。
|
||||
pub fn iter_prefix(prefix: &[u8]) -> Vec<(String, String)> {
|
||||
let Some(root) = root() else {
|
||||
return Vec::new();
|
||||
};
|
||||
let guard = root.map.read();
|
||||
guard
|
||||
.iter()
|
||||
.filter_map(|(k, v)| {
|
||||
let name = k.as_bytes().strip_prefix(prefix)?;
|
||||
let (payload, _) = split_value(v.as_bytes())?;
|
||||
Some((
|
||||
String::from_utf8_lossy(name).into_owned(),
|
||||
String::from_utf8_lossy(payload).into_owned(),
|
||||
))
|
||||
})
|
||||
.collect()
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 业务语义封装
|
||||
// ===========================================================================
|
||||
|
||||
@@ -0,0 +1,282 @@
|
||||
//! LLM 用量统计:按**账号**与按**账号 × 上游**两个维度计数。
|
||||
//!
|
||||
//! 数据全部躺在共享内存里(`ust:` / `ums:` 两个键前缀),没有新文件、
|
||||
//! 没有新进程——这是项目「单容器零依赖」红线下的必然选择。
|
||||
//!
|
||||
//! ```text
|
||||
//! ust:<username> → <日窗id>:<今日数>:<累计数>:<最近调用时间戳>
|
||||
//! ums:<username>:<slot> → 同 上
|
||||
//! ```
|
||||
//!
|
||||
//! ## 口径:只统计"真的转发出去"的请求
|
||||
//!
|
||||
//! 计数点在 `LlmHandler` 的**鉴权 / 白名单 / 限流全部通过之后**,
|
||||
//! 所以 401 / 403 / 404 / 429 都不计入。这与配额计数 `st:` 刻意不同——
|
||||
//! 后者是"先增后判"(含被拒),要的是"超限后计数继续涨"的执行语义;
|
||||
//! 而用量统计是给人看的,把被拒的请求算进"用量"只会让人误解。
|
||||
//!
|
||||
//! ## 为什么不做 token 级
|
||||
//!
|
||||
//! token 数在上游响应体里,而 SSE 必须关缓冲透传(`proxy_buffering off`),
|
||||
//! 网关不缓冲就解析不到 `usage` 字段。要做 token 级计费得自建 upstream
|
||||
//! 把响应拦下来重组流,与「不引入额外常驻进程 / 不改 nginx 原生转发」冲突。
|
||||
//! 因此口径统一为**请求数**,并在文档里写明。
|
||||
|
||||
use std::collections::BTreeMap;
|
||||
|
||||
use serde_json::{Value, json};
|
||||
|
||||
use crate::config::{KEY_UMS_PREFIX, KEY_UST_PREFIX};
|
||||
use crate::state;
|
||||
use crate::util;
|
||||
|
||||
/// 一条统计记录的读数。`today` 已按调用方给的日窗折算过(跨天即为 0)。
|
||||
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
|
||||
pub struct Usage {
|
||||
/// 今日(本地自然日)请求数
|
||||
pub today: i64,
|
||||
/// 累计请求数(自该计数键建立以来,不随跨天归零)
|
||||
pub total: i64,
|
||||
/// 最近一次调用时间戳;0 = 从未调用
|
||||
pub last_ts: i64,
|
||||
}
|
||||
|
||||
impl Usage {
|
||||
/// 累加(聚合用)。`last_ts` 取最大。
|
||||
fn merge(&mut self, other: &Usage) {
|
||||
self.today += other.today;
|
||||
self.total += other.total;
|
||||
if other.last_ts > self.last_ts {
|
||||
self.last_ts = other.last_ts;
|
||||
}
|
||||
}
|
||||
|
||||
pub fn json(&self) -> Value {
|
||||
json!({
|
||||
"today": self.today,
|
||||
"total": self.total,
|
||||
"last_ts": self.last_ts,
|
||||
"last_ts_text": if self.last_ts > 0 { util::fmt_time(self.last_ts) } else { String::new() },
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
/// 一个账号的用量 + 其下每个上游的明细
|
||||
pub struct UserUsage {
|
||||
pub username: String,
|
||||
pub usage: Usage,
|
||||
pub models: Vec<(String, Usage)>,
|
||||
}
|
||||
|
||||
/// 一个上游的用量 + 用它的每个账号的明细
|
||||
pub struct ModelUsage {
|
||||
pub slot: String,
|
||||
pub usage: Usage,
|
||||
pub users: Vec<(String, Usage)>,
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 编解码
|
||||
// ===========================================================================
|
||||
|
||||
/// `<日窗id>:<今日数>:<累计数>:<最近调用时间戳>`。
|
||||
/// 字段全是数字,用 `:` 分隔不会与键名里的字符冲突(键不参与这段编码)。
|
||||
fn encode(day: i64, u: &Usage) -> String {
|
||||
format!("{}:{}:{}:{}", day, u.today, u.total, u.last_ts)
|
||||
}
|
||||
|
||||
/// 解码并把"今日数"按当前日窗折算——跨天时旧值归零,累计值保留。
|
||||
fn decode(payload: &str, day: i64) -> Usage {
|
||||
let mut it = payload.split(':');
|
||||
let stored_day = it.next().and_then(|s| s.parse::<i64>().ok()).unwrap_or(0);
|
||||
let today = it.next().and_then(|s| s.parse::<i64>().ok()).unwrap_or(0);
|
||||
let total = it.next().and_then(|s| s.parse::<i64>().ok()).unwrap_or(0);
|
||||
let last_ts = it.next().and_then(|s| s.parse::<i64>().ok()).unwrap_or(0);
|
||||
Usage { today: if stored_day == day { today } else { 0 }, total, last_ts }
|
||||
}
|
||||
|
||||
fn bump(prefix: &[u8], name: &str, day: i64, now: i64) {
|
||||
let current = state::get(prefix, name.as_bytes())
|
||||
.map(|(payload, _)| decode(&payload, day))
|
||||
.unwrap_or_default();
|
||||
|
||||
let next = Usage {
|
||||
today: current.today + 1,
|
||||
total: current.total + 1,
|
||||
last_ts: now,
|
||||
};
|
||||
// 与限流的窗口计数同款近似语义:读-增-写不加写锁(并发下偶发少计一次),
|
||||
// 换来的是 LLM 热路径不为统计去抢锁。统计用途下这点误差无影响。
|
||||
state::put(prefix, name.as_bytes(), encode(day, &next).as_bytes(), 0);
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 写
|
||||
// ===========================================================================
|
||||
|
||||
/// 记录一次已放行的 LLM 请求。**在鉴权/白名单/限流全通过之后调用**。
|
||||
pub fn record(username: &str, slot: &str, now: i64) {
|
||||
if username.is_empty() || slot.is_empty() {
|
||||
return;
|
||||
}
|
||||
let day = util::day_id(now);
|
||||
bump(KEY_UST_PREFIX, username, day, now);
|
||||
bump(KEY_UMS_PREFIX, &format!("{username}:{slot}"), day, now);
|
||||
}
|
||||
|
||||
/// 删除账号时清理其全部用量键(`ust:<user>` 与 `ums:<user>:*`)。
|
||||
///
|
||||
/// 用量键是**按前缀扫出来**做报表的(`all_users` / `all_models`),不清就会让
|
||||
/// 已删除的账号继续挂在用量榜上。这与「删账号必须摘 `ukey:` 索引」是同类的
|
||||
/// 正确性问题(只是影响面在报表而非鉴权,因此更容易被漏掉)。
|
||||
pub fn forget(username: &str) {
|
||||
if username.is_empty() {
|
||||
return;
|
||||
}
|
||||
state::remove(KEY_UST_PREFIX, username.as_bytes());
|
||||
|
||||
// 先物化键名再删,避免"遍历中改表"
|
||||
let prefix = format!("{username}:");
|
||||
let keys: Vec<String> = state::iter_prefix(KEY_UMS_PREFIX)
|
||||
.into_iter()
|
||||
.map(|(name, _)| name)
|
||||
.filter(|name| name.starts_with(&prefix))
|
||||
.collect();
|
||||
for name in keys {
|
||||
state::remove(KEY_UMS_PREFIX, name.as_bytes());
|
||||
}
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 读
|
||||
// ===========================================================================
|
||||
|
||||
/// 单账号读数(不含模型明细)
|
||||
pub fn user(username: &str, now: i64) -> Usage {
|
||||
let day = util::day_id(now);
|
||||
state::get(KEY_UST_PREFIX, username.as_bytes())
|
||||
.map(|(payload, _)| decode(&payload, day))
|
||||
.unwrap_or_default()
|
||||
}
|
||||
|
||||
/// 单账号 × 上游明细,按 slot 升序
|
||||
pub fn user_models(username: &str, now: i64) -> Vec<(String, Usage)> {
|
||||
let day = util::day_id(now);
|
||||
let prefix = format!("{username}:");
|
||||
let mut out: Vec<(String, Usage)> = state::iter_prefix(KEY_UMS_PREFIX)
|
||||
.into_iter()
|
||||
.filter_map(|(name, payload)| {
|
||||
let slot = name.strip_prefix(&prefix)?.to_owned();
|
||||
Some((slot, decode(&payload, day)))
|
||||
})
|
||||
.collect();
|
||||
out.sort_by(|a, b| a.0.cmp(&b.0));
|
||||
out
|
||||
}
|
||||
|
||||
/// 全部账号的用量(含各自的模型明细),按账号名升序。
|
||||
///
|
||||
/// 只扫一次 `ums:` 前缀再分组,避免"每个账号遍历一遍全表"。
|
||||
pub fn all_users(now: i64) -> Vec<UserUsage> {
|
||||
let day = util::day_id(now);
|
||||
|
||||
let mut by_user: BTreeMap<String, Vec<(String, Usage)>> = BTreeMap::new();
|
||||
for (name, payload) in state::iter_prefix(KEY_UMS_PREFIX) {
|
||||
let Some((username, slot)) = name.split_once(':') else {
|
||||
continue;
|
||||
};
|
||||
by_user
|
||||
.entry(username.to_owned())
|
||||
.or_default()
|
||||
.push((slot.to_owned(), decode(&payload, day)));
|
||||
}
|
||||
for models in by_user.values_mut() {
|
||||
models.sort_by(|a, b| a.0.cmp(&b.0));
|
||||
}
|
||||
|
||||
state::iter_prefix(KEY_UST_PREFIX)
|
||||
.into_iter()
|
||||
.map(|(username, payload)| {
|
||||
let usage = decode(&payload, day);
|
||||
let models = by_user.remove(&username).unwrap_or_default();
|
||||
UserUsage { username, usage, models }
|
||||
})
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// 全部上游的用量(含用到它的账号明细),按 slot 升序。
|
||||
pub fn all_models(now: i64) -> Vec<ModelUsage> {
|
||||
let day = util::day_id(now);
|
||||
|
||||
// slot → (总量, 各账号)
|
||||
let mut by_slot: BTreeMap<String, (Usage, Vec<(String, Usage)>)> = BTreeMap::new();
|
||||
for (name, payload) in state::iter_prefix(KEY_UMS_PREFIX) {
|
||||
let Some((username, slot)) = name.split_once(':') else {
|
||||
continue;
|
||||
};
|
||||
let u = decode(&payload, day);
|
||||
let entry = by_slot.entry(slot.to_owned()).or_default();
|
||||
entry.0.merge(&u);
|
||||
entry.1.push((username.to_owned(), u));
|
||||
}
|
||||
|
||||
by_slot
|
||||
.into_iter()
|
||||
.map(|(slot, (usage, mut users))| {
|
||||
users.sort_by(|a, b| a.0.cmp(&b.0));
|
||||
ModelUsage { slot, usage, users }
|
||||
})
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// 全局合计
|
||||
pub fn totals(now: i64) -> Usage {
|
||||
let day = util::day_id(now);
|
||||
let mut sum = Usage::default();
|
||||
for (_, payload) in state::iter_prefix(KEY_UST_PREFIX) {
|
||||
sum.merge(&decode(&payload, day));
|
||||
}
|
||||
sum
|
||||
}
|
||||
|
||||
/// 管理台用量页的完整投影
|
||||
pub fn overview(now: i64) -> Value {
|
||||
let users: Vec<Value> = all_users(now)
|
||||
.into_iter()
|
||||
.map(|u| {
|
||||
json!({
|
||||
"username": u.username,
|
||||
"today": u.usage.today,
|
||||
"total": u.usage.total,
|
||||
"last_ts": u.usage.last_ts,
|
||||
"last_ts_text": if u.usage.last_ts > 0 { util::fmt_time(u.usage.last_ts) } else { String::new() },
|
||||
"models": u.models.iter().map(|(slot, usage)| {
|
||||
json!({ "slot": slot, "today": usage.today, "total": usage.total })
|
||||
}).collect::<Vec<_>>(),
|
||||
})
|
||||
})
|
||||
.collect();
|
||||
|
||||
let models: Vec<Value> = all_models(now)
|
||||
.into_iter()
|
||||
.map(|m| {
|
||||
json!({
|
||||
"slot": m.slot,
|
||||
"today": m.usage.today,
|
||||
"total": m.usage.total,
|
||||
"last_ts": m.usage.last_ts,
|
||||
"last_ts_text": if m.usage.last_ts > 0 { util::fmt_time(m.usage.last_ts) } else { String::new() },
|
||||
"users": m.users.iter().map(|(name, usage)| {
|
||||
json!({ "username": name, "today": usage.today, "total": usage.total })
|
||||
}).collect::<Vec<_>>(),
|
||||
})
|
||||
})
|
||||
.collect();
|
||||
|
||||
let sum = totals(now);
|
||||
json!({
|
||||
"total": sum.json(),
|
||||
"users": users,
|
||||
"models": models,
|
||||
})
|
||||
}
|
||||
+96
-3
@@ -15,7 +15,10 @@
|
||||
|
||||
use serde_json::{Value, json};
|
||||
|
||||
use crate::config::{KEY_UP_PREFIX, UPSTREAM_BASE_MAX_LEN, UPSTREAM_SLOT_MAX_LEN};
|
||||
use crate::config::{
|
||||
KEY_UP_PREFIX, LLM_PUBLIC_HOST, MODEL_NAME_MAX_LEN, UPSTREAM_BASE_MAX_LEN,
|
||||
UPSTREAM_MODELS_MAX, UPSTREAM_SLOT_MAX_LEN,
|
||||
};
|
||||
use crate::models::Upstream;
|
||||
use crate::state;
|
||||
use crate::store;
|
||||
@@ -51,6 +54,34 @@ pub fn validate_slot(raw: &str) -> Result<String, String> {
|
||||
Ok(s)
|
||||
}
|
||||
|
||||
/// 模型名清单:去空白、去空项、去重、限长。
|
||||
///
|
||||
/// 刻意**允许** `/` `.` `:` 等字符——`Qwen/Qwen2.5-7B`(HF 风格)、
|
||||
/// `llama3:8b`(Ollama 标签)都是合法模型名。只挡空白与控制字符,
|
||||
/// 因为模型名要出现在 URL/JSON/HTML 里,空白只会带来歧义。
|
||||
fn check_models(raw: &[String]) -> Result<Vec<String>, String> {
|
||||
let mut out: Vec<String> = Vec::new();
|
||||
for item in raw {
|
||||
let m = util::trim(item);
|
||||
if m.is_empty() {
|
||||
continue;
|
||||
}
|
||||
if m.chars().count() > MODEL_NAME_MAX_LEN {
|
||||
return Err(format!("模型名「{m}」超过 {MODEL_NAME_MAX_LEN} 个字符"));
|
||||
}
|
||||
if m.chars().any(char::is_whitespace) || m.chars().any(char::is_control) {
|
||||
return Err(format!("模型名「{m}」不能包含空格或控制字符"));
|
||||
}
|
||||
if !out.iter().any(|x| x == m) {
|
||||
out.push(m.to_owned());
|
||||
}
|
||||
}
|
||||
if out.len() > UPSTREAM_MODELS_MAX {
|
||||
return Err(format!("单个上游最多登记 {UPSTREAM_MODELS_MAX} 个模型"));
|
||||
}
|
||||
Ok(out)
|
||||
}
|
||||
|
||||
/// 上游基址:必须显式带 `http://` 或 `https://`。
|
||||
///
|
||||
/// 比子域名的 `backend`(纯 host:port)宽——LLM 上游常挂在网关路径下
|
||||
@@ -139,6 +170,55 @@ pub fn get(slot: &str) -> Option<Upstream> {
|
||||
store::read_upstreams().remove(slot)
|
||||
}
|
||||
|
||||
/// 某账号**可用**的上游(已启用,且在白名单内;`allowed` 为空 = 不限)。
|
||||
///
|
||||
/// 账号态接口与模型发现都用它,避免"能看见但不能用"或反之的两套判定。
|
||||
pub fn available(allowed: &[String]) -> Vec<(String, Upstream)> {
|
||||
all()
|
||||
.into_iter()
|
||||
.filter(|(slot, up)| {
|
||||
up.enabled && (allowed.is_empty() || allowed.iter().any(|s| s == slot))
|
||||
})
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// 客户端应当使用的 base_url。
|
||||
///
|
||||
/// 规则:**把上游 base 的 scheme+host 换成网关域名,路径原样保留**——
|
||||
/// 这样它才真的能和厂商文档里的 base_url 逐个字符替换:
|
||||
///
|
||||
/// ```text
|
||||
/// https://api.deepseek.com/v1 → https://lm.mayishangshu.cn/deepseek/v1
|
||||
/// https://api.openai.com → https://lm.mayishangshu.cn/openai
|
||||
/// ```
|
||||
///
|
||||
/// 反过来(不管上游 base 的路径、一律拼 `/v1`)会让「base 自带 /v1」的上游
|
||||
/// 变成 `/v1/v1/chat/completions` → 上游 404,且客户端拿到的是空响应体,
|
||||
/// 排查起来完全看不出是网关拼错了。**实测踩过,别再改回去。**
|
||||
pub fn base_url(slot: &str, up: &Upstream) -> String {
|
||||
let rest = up.base.split("://").nth(1).unwrap_or("");
|
||||
let path = match rest.find('/') {
|
||||
Some(i) => &rest[i..],
|
||||
None => "",
|
||||
};
|
||||
format!("https://{LLM_PUBLIC_HOST}/{slot}{path}")
|
||||
}
|
||||
|
||||
/// 模型发现:把可用上游挂载的模型摊平成 `(模型名, slot)` 列表。
|
||||
///
|
||||
/// 没登记 `models` 的上游不产出任何条目(清单未知 ≠ 有 0 个模型),
|
||||
/// 但它的转发能力不受影响——模型发现只是"告诉你有什么",
|
||||
/// 不是"允许你用什么"。
|
||||
pub fn models_flat(allowed: &[String]) -> Vec<(String, String)> {
|
||||
let mut out = Vec::new();
|
||||
for (slot, up) in available(allowed) {
|
||||
for model in &up.models {
|
||||
out.push((model.clone(), slot.clone()));
|
||||
}
|
||||
}
|
||||
out
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 写入
|
||||
// ===========================================================================
|
||||
@@ -146,16 +226,22 @@ pub fn get(slot: &str) -> Option<Upstream> {
|
||||
/// 新增或覆盖一个上游。
|
||||
///
|
||||
/// `auth = None` 表示**保持原值**(表单只改 base 时不必回填密钥);
|
||||
/// `Some("")` 表示显式清空。`enabled = false` 会立刻从共享内存移除。
|
||||
/// `Some("")` 表示显式清空。`models = None` 同款语义(保持原清单)。
|
||||
/// `enabled = false` 会立刻从共享内存移除。
|
||||
pub fn upsert(
|
||||
slot_raw: &str,
|
||||
base_raw: &str,
|
||||
models: Option<Vec<String>>,
|
||||
auth_raw: Option<&String>,
|
||||
enabled: bool,
|
||||
note: Option<&String>,
|
||||
) -> Result<Upstream, String> {
|
||||
let slot = validate_slot(slot_raw)?;
|
||||
let base = validate_base(base_raw)?;
|
||||
let models = match models {
|
||||
Some(list) => Some(check_models(&list)?),
|
||||
None => None,
|
||||
};
|
||||
|
||||
let key = slot.clone();
|
||||
let up = store::update_upstreams(move |table| {
|
||||
@@ -170,13 +256,17 @@ pub fn upsert(
|
||||
Some(n) => util::trim(n).to_owned(),
|
||||
None => existing.as_ref().map(|e| e.note.clone()).unwrap_or_default(),
|
||||
};
|
||||
let models = match models {
|
||||
Some(m) => m,
|
||||
None => existing.as_ref().map(|e| e.models.clone()).unwrap_or_default(),
|
||||
};
|
||||
let created_at = existing
|
||||
.as_ref()
|
||||
.map(|e| e.created_at)
|
||||
.filter(|&t| t > 0)
|
||||
.unwrap_or_else(util::now);
|
||||
|
||||
let item = Upstream { base, auth, enabled, note, created_at };
|
||||
let item = Upstream { base, models, auth, enabled, note, created_at };
|
||||
table.insert(key.clone(), item.clone());
|
||||
Ok(item)
|
||||
})?;
|
||||
@@ -213,6 +303,9 @@ pub fn view(slot: &str, up: &Upstream) -> Value {
|
||||
json!({
|
||||
"slot": slot,
|
||||
"base": up.base,
|
||||
"base_url": base_url(slot, up),
|
||||
"models": up.models,
|
||||
"model_count": up.models.len(),
|
||||
"auth_set": !up.auth.is_empty(),
|
||||
"auth_mask": if up.auth.is_empty() { "-".to_owned() } else { util::mask_key(&up.auth) },
|
||||
"enabled": up.enabled,
|
||||
|
||||
+46
-5
@@ -15,11 +15,12 @@
|
||||
use serde_json::{Value, json};
|
||||
|
||||
use crate::config::{
|
||||
INIT_ADMIN_PASSWORD_DEFAULT, INIT_ADMIN_PASSWORD_ENV, INIT_ADMIN_USER, KEY_UKEY_PREFIX,
|
||||
PASSWORD_MIN_LEN, TEMP_GRANT_TTL, USERNAME_MAX_LEN, USERNAME_MIN_LEN,
|
||||
DISPLAY_NAME_MAX_LEN, INIT_ADMIN_PASSWORD_DEFAULT, INIT_ADMIN_PASSWORD_ENV, INIT_ADMIN_USER,
|
||||
KEY_UKEY_PREFIX, PASSWORD_MIN_LEN, TEMP_GRANT_TTL, USERNAME_MAX_LEN, USERNAME_MIN_LEN,
|
||||
};
|
||||
use crate::models::{User, UserTable};
|
||||
use crate::state;
|
||||
use crate::stats;
|
||||
use crate::store;
|
||||
use crate::util;
|
||||
|
||||
@@ -36,6 +37,19 @@ fn valid_username(name: &str) -> bool {
|
||||
name.chars().all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
|
||||
}
|
||||
|
||||
/// 展示名:允许中文/空格,但**不允许换行或制表等控制字符**——
|
||||
/// 它会被塞进管理台 HTML(已转义)与 JSON,控制字符只会造成排版错乱。
|
||||
fn check_display_name(raw: &str) -> Result<String, String> {
|
||||
let v = util::trim(raw).to_owned();
|
||||
if v.chars().count() > DISPLAY_NAME_MAX_LEN {
|
||||
return Err(format!("展示名不能超过 {DISPLAY_NAME_MAX_LEN} 个字"));
|
||||
}
|
||||
if v.chars().any(|c| c.is_control()) {
|
||||
return Err("展示名不能包含换行或控制字符".to_owned());
|
||||
}
|
||||
Ok(v)
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// 共享内存同步(Key 索引)
|
||||
// ===========================================================================
|
||||
@@ -108,11 +122,18 @@ pub fn find_by_key(key: &str) -> Option<User> {
|
||||
// 写入
|
||||
// ===========================================================================
|
||||
|
||||
fn build_user(username: &str, password: &str, role: &str, grant_temp: bool) -> User {
|
||||
fn build_user(
|
||||
username: &str,
|
||||
display_name: &str,
|
||||
password: &str,
|
||||
role: &str,
|
||||
grant_temp: bool,
|
||||
) -> User {
|
||||
let salt = util::gen_salt();
|
||||
let now = util::now();
|
||||
User {
|
||||
username: username.to_owned(),
|
||||
display_name: display_name.to_owned(),
|
||||
role: role.to_owned(),
|
||||
pwd_hash: util::hash_password(password, &salt),
|
||||
salt,
|
||||
@@ -137,7 +158,7 @@ pub fn seed_admin() -> Result<bool, String> {
|
||||
if !users.is_empty() {
|
||||
return Ok(());
|
||||
}
|
||||
let admin = build_user(INIT_ADMIN_USER, &password, "admin", false);
|
||||
let admin = build_user(INIT_ADMIN_USER, "", &password, "admin", false);
|
||||
users.insert(admin.username.clone(), admin.clone());
|
||||
created = Some(admin);
|
||||
Ok(())
|
||||
@@ -157,6 +178,7 @@ pub fn seed_admin() -> Result<bool, String> {
|
||||
|
||||
pub fn create(
|
||||
username: &str,
|
||||
display_name: &str,
|
||||
password: &str,
|
||||
role: &str,
|
||||
grant_temp: bool,
|
||||
@@ -167,6 +189,7 @@ pub fn create(
|
||||
"账号名只能是 {USERNAME_MIN_LEN}-{USERNAME_MAX_LEN} 位字母、数字、下划线或短横线"
|
||||
));
|
||||
}
|
||||
let display_name = check_display_name(display_name)?;
|
||||
if password.len() < PASSWORD_MIN_LEN {
|
||||
return Err(format!("密码至少 {PASSWORD_MIN_LEN} 位"));
|
||||
}
|
||||
@@ -177,7 +200,7 @@ pub fn create(
|
||||
return Err(format!("账号 {username} 已存在"));
|
||||
}
|
||||
|
||||
let user = build_user(&username, password, role, grant_temp);
|
||||
let user = build_user(&username, &display_name, password, role, grant_temp);
|
||||
let created = user.clone();
|
||||
store::update_users(move |users| {
|
||||
if users.contains_key(&username) {
|
||||
@@ -191,6 +214,18 @@ pub fn create(
|
||||
Ok(created)
|
||||
}
|
||||
|
||||
/// 设置展示名(中文姓名等)。只影响管理台显示,不碰 Key / 口令 / 白名单。
|
||||
pub fn set_display_name(username: &str, display_name: &str) -> Result<User, String> {
|
||||
let display_name = check_display_name(display_name)?;
|
||||
let updated = store::update_users(move |users: &mut UserTable| {
|
||||
let user = users.get_mut(username).ok_or("账号不存在")?;
|
||||
user.display_name = display_name;
|
||||
Ok(user.clone())
|
||||
})?;
|
||||
// 展示名不在 ukey payload 里(热路径不需要它),无需同步共享内存
|
||||
Ok(updated)
|
||||
}
|
||||
|
||||
/// 重置 API Key,返回新 Key。
|
||||
///
|
||||
/// **必须同时摘掉旧索引**:LLM 热路径查的是 `ukey:`,若只加新的不删旧的,
|
||||
@@ -297,6 +332,7 @@ pub fn delete(username: &str, operator: &str) -> Result<(), String> {
|
||||
return Err("不能删除当前登录的账号".to_owned());
|
||||
}
|
||||
let username = username.to_owned();
|
||||
let stats_name = username.clone();
|
||||
|
||||
let removed_key = store::update_users(move |users| {
|
||||
let user = users.get(&username).ok_or("账号不存在")?;
|
||||
@@ -312,6 +348,8 @@ pub fn delete(username: &str, operator: &str) -> Result<(), String> {
|
||||
if !removed_key.is_empty() {
|
||||
state::remove_ukey(&removed_key);
|
||||
}
|
||||
// 用量键是按前缀扫出来出报表的,不清就会让已删账号继续挂在用量榜上
|
||||
stats::forget(&stats_name);
|
||||
Ok(())
|
||||
}
|
||||
|
||||
@@ -327,6 +365,9 @@ pub fn view(user: &User) -> Value {
|
||||
|
||||
json!({
|
||||
"username": user.username,
|
||||
"display_name": user.display_name,
|
||||
// 前端统一用 name 渲染——展示名优先,没有就回落到账号名
|
||||
"name": if user.display_name.is_empty() { user.username.clone() } else { user.display_name.clone() },
|
||||
"role": user.role,
|
||||
"api_key": user.api_key,
|
||||
"api_key_masked": util::mask_key(&user.api_key),
|
||||
|
||||
Reference in new issue
Block a user