Quoted fields in CSV policy files keep their quote characters, so quoted policy rules never match

從 casbin/persist/adapter.cpp 中的 LoadPolicyLine 開始，將其以逗號分割的行為與文件中記載的 CSV 引號處理方式進行比對。用一條包含帶引號的 admin 值的策略行，以及另一條包含帶引號的逗號的策略行重現問題，然後執行專案的 GoogleTest 測試套件，必要時新增一個針對性的測試案例。當帶引號的欄位載入後不含引號，而不帶引號的行仍維持原本的行為時，即告完成。

Location (apache/casbin-cpp @ default branch 0ad64a7):

• casbin/persist/adapter.cpp:28-43 — LoadPolicyLine, used by the file, filtered-file and string adapters

void LoadPolicyLine(const std::string& line, const std::shared_ptr<Model>& model) {
    if (line.empty() || line.find('#') == 0)
        return;

    std::vector<std::string> tokens = Split(line, ",", -1);
    for (int i = 0; i < tokens.size(); i++)
        tokens[i] = Trim(tokens[i]);

    std::string key = tokens[0];
    std::string sec = key.substr(0, 1);
    std::vector<std::string> new_tokens(tokens.begin() + 1, tokens.end());

Problem
The Casbin Policy Storage docs describe quoted CSV fields:

If a field contains commas, wrap it in double quotes … If a field contains both commas and double quotes, wrap it in double quotes and escape internal quotes by doubling them

LoadPolicyLine does Split(line, ",") and then Trim on each token. Trim only removes whitespace, so "admin" keeps its quotes and the quote characters are stored as part of the value. A request for admin therefore doesn't match a policy line written as "admin", and Enforce returns false. For the same reason a quoted comma ("data1,data2") is split into two fields (from reading the code; not run).

Expected: a quoted field is unquoted when loading: "admin" → admin, "data1,data2" → one field data1,data2, "" inside a quoted field → ". Spaces inside the quotes stay part of the value.

Reproduction

quoted_policy.csv:

p, admin, /, GET
p,    "admin",  /,  POST

#include <iostream>
#include <casbin/casbin.h>

int main() {
    casbin::Enforcer e("examples/basic_model.conf", "quoted_policy.csv");
    std::cout << std::boolalpha
              << e.Enforce({"admin", "/", "GET"}) << "\n"    // true
              << e.Enforce({"admin", "/", "POST"}) << "\n";  // false, expected true
}

Observed with a gtest:

Expected equality of these values:
  e.Enforce({"admin", "/", "POST"})
    Which is: false
  true

The unquoted GET line in the same file is enforced as expected.

Environment:

• Commit: 0ad64a7 (master)
• Toolchain: GCC 11.5.0, Linux x86_64, default CMake build

Already fixed in apache/casbin-node-casbin
PR: https://github.com/apache/casbin-node-casbin/pull/494 — "feat: add BracketAwareCsvParser to parse CSV file better" (merged 2024-11-22T10:12:57Z, merge commit 4c738831ebb). node-casbin's loadPolicyLine now strips the outer double quotes from the section key and from every policy value and turns "" into ", so a quoted rule loads as the plain value. The PR also adds bracket-aware token joining for policy expressions; the unquoting in PolicyLoader.loadPolicyLine is the part that addresses this behaviour. (node-casbin also trims whitespace inside the quotes; the scope proposed below keeps it.)

How node-casbin fixed it:

--- a/src/persist/helper.ts
+++ b/src/persist/helper.ts
@@ -1,23 +1,83 @@
...
+  loadPolicyLine(line: string, model: Model): void {
+    const tokens = this.parser.parse(line);
     if (!tokens || !tokens[0]) {
       return;
     }
 
-    const key = tokens[0][0];
+    let key = tokens[0][0].trim();
+    if (key.startsWith('"') && key.endsWith('"')) {
+      key = key.slice(1, -1);
+    }
+
...
@@ -28,6 +88,22 @@ export class Helper {
     if (!policy) {
       return;
     }
-    policy.policy.push(tokens[0].slice(1));
+
+    const values = tokens[0].slice(1).map((v) => {
+      if (v.startsWith('"') && v.endsWith('"')) {
+        v = v.slice(1, -1);
+      }
+      return v.replace(/""/g, '"').trim();
+    });
+
+    policy.policy.push(values);
+  }
...

Suggested fix
Replace the Split + Trim tokenizer in LoadPolicyLine with a small CSV-aware splitter: a field wrapped in double quotes may contain commas, "" inside it becomes ", whitespace inside the quotes is kept, and unquoted fields are trimmed as before:

diff --git a/casbin/persist/adapter.cpp b/casbin/persist/adapter.cpp
--- a/casbin/persist/adapter.cpp
+++ b/casbin/persist/adapter.cpp
@@ -24,14 +24,44 @@
 
 namespace casbin {
 
+namespace {
+    // SplitPolicyLine splits a CSV policy line into fields. Unquoted fields are trimmed. A field wrapped in double
+    // quotes may contain commas, "" inside it stands for one double quote, and whitespace inside the quotes is kept.
+    std::vector<std::string> SplitPolicyLine(const std::string& line) {
+        std::vector<std::string> tokens;
+        std::string token;
+        bool in_quotes = false;
+        bool quoted = false;
+        for (size_t i = 0; i < line.size(); i++) {
+            char c = line[i];
+            if (in_quotes) {
+                if (c != '"')
+                    token += c;
+                else if (i + 1 < line.size() && line[i + 1] == '"')
+                    token += line[++i];
+                else
+                    in_quotes = false;
+            } else if (c == ',') {
+                tokens.push_back(quoted ? token : Trim(token));
+                token.clear();
+                quoted = false;
+            } else if (c == '"' && !quoted && Trim(token).empty()) {
+                in_quotes = quoted = true;
+            } else if (!quoted || !std::isspace(static_cast<unsigned char>(c))) {
+                token += c;
+            }
+        }
+        tokens.push_back(quoted ? token : Trim(token));
+        return tokens;
+    }
+} // namespace
+
 // LoadPolicyLine loads a text line as a policy rule to model.
 void LoadPolicyLine(const std::string& line, const std::shared_ptr<Model>& model) {
     if (line.empty() || line.find('#') == 0)
         return;
 
-    std::vector<std::string> tokens = Split(line, ",", -1);
-    for (int i = 0; i < tokens.size(); i++)
-        tokens[i] = Trim(tokens[i]);
+    std::vector<std::string> tokens = SplitPolicyLine(line);
 
     std::string key = tokens[0];
     std::string sec = key.substr(0, 1);

Verified on 0ad64a7 (GCC 11.4.0, GoogleTest 1.14.0): the test above, extended with a quoted-comma line (p, "admin", "/a,b", PUT) and a doubled-quote line, fails without this patch and passes with it; no other test results change.

I can submit a PR for this.

Related: searched issues/PRs for LoadPolicyLine, quotes/quoted CSV, double quotes; the only hit, #132, is about RBAC role links after LoadPolicyLine, not quoting.