markdown-it: Quadratic-time rendering of a single multi-line link reference definition with default options (regression since 14.0)

Summary

A link reference definition whose label or title is never closed is O(n²) in markdown-it 15.0.2 with default options. Continuation lines are pulled lazily with str += lineContent in src/rules_block/reference.ts, and the subsequent charCodeAt and slice calls flatten the V8 cons-string, copying everything collected so far; this runs before the rule decides the definition is invalid. It is a regression introduced by PR #1004 in 14.0.0: 13.0.2 processed a 313 KB input in 125 ms, while every version since takes about 11 seconds on the same input. At the sizes below, 200 KB takes 3.6 seconds, 400 KB takes 18.1 seconds, and 800 KB takes 74.0 seconds, which is roughly 4–5× per doubling. No options are required to trigger it.

Details

src/rules_block/reference.ts appends continuation lines with str += lineContent at lines 74, 83, 100, 133, and 150, then keeps indexing the accumulated string via str.charCodeAt(pos) in the label and whitespace loops and str.slice(...) in parseLinkTitle. V8 represents += as a cons-string, and the next charCodeAt or slice flattens it, copying everything collected so far, so one definition spanning n lines costs O(n²). Isolation confirms the rule is responsible: block-only parsing with inline disabled is still quadratic, md.disable('reference') makes the same input linear, and the same text without the leading [ is linear. The fix is to track positions in state.src instead of building str, or to collect line slices in an array and parse across them. At minimum, stop scanning the label after 999 characters (CommonMark §6.3) before pulling more lines, and cap title continuation. Add a pathological test with a 100k-line unclosed title and label. Open PR #1215 edits the same file but its label check runs after the label loop, so it does not bound either variant and the fix can land together with that PR.

The observed output from PoC was: unclosed title n=100000 200009 B 3585.5 ms unclosed title n=200000 400009 B 18107.1 ms (5.05x) unclosed title n=400000 800009 B 73954.4 ms (4.08x)

Regression comparison at 313 KB: 13.0.2 → 125 ms, 14.1.0 → 11.4 s, 15.0.0 → 10.6 s, 15.0.2 → 10.7 s.

Proof of concept

How to execute the PoC:

1. mkdir -p markdown-it-poc && cd markdown-it-poc
2. npm init -y
3. npm i [email protected]
4. node poc_markdown_it.js

PoC below is:

// poc_markdown_it.js
// Requires: npm install [email protected]
const MarkdownIt = require('markdown-it');
const md = new MarkdownIt();

function bench(label, src) {
  const t0 = process.hrtime.bigint();
  md.render(src);
  const t1 = process.hrtime.bigint();
  console.log(
    `${label.padEnd(30)} ${String(src.length).padStart(9)} B ${(Number(t1 - t0) / 1e6).toFixed(1).padStart(9)} ms`
  );
}

// Unclosed title: one reference definition spread over many lines
for (const n of [100000, 200000, 400000]) {
  bench(`unclosed title n=${n}`, '[a]: /u "' + 'x\n'.repeat(n));
}

// Unclosed label
for (const n of [, , ]) {
  (,  + .(n));
}

Impact

This is an algorithmic-complexity denial of service (CWE-407). Any server that renders user Markdown with markdown-it is affected with default options, including comments, issues, chat, wikis, docs previews, and LLM chat UIs. A ~400 KB request body blocks the Node.js event loop for about 18 seconds and an ~800 KB body for about 74 seconds. This is the same class and attacker model as GHSA-253c, but it is reachable without any opt-in option, since the affected path is a core block rule.