147 lines
4.4 KiB
JavaScript
147 lines
4.4 KiB
JavaScript
'use strict';
|
|
|
|
//
|
|
// A dumbed-down, minimal ASN.1 parser / packer combo
|
|
//
|
|
// Note: generally I like to write congruent code
|
|
// (i.e. output can be used as input and vice-versa)
|
|
// However, this seemed to be more readable and easier
|
|
// to use written as-is, asymmetrically.
|
|
// (I also generally prefer to export objects rather
|
|
// functions but, yet again, asthetics one in this case)
|
|
|
|
var Enc = require('./encoding.js');
|
|
|
|
//
|
|
// Packer
|
|
//
|
|
|
|
// Almost every ASN.1 type that's important for CSR
|
|
// can be represented generically with only a few rules.
|
|
var ASN1 = module.exports = function ASN1(/*type, hexstrings...*/) {
|
|
var args = Array.prototype.slice.call(arguments);
|
|
var typ = args.shift();
|
|
var str = args.join('').replace(/\s+/g, '').toLowerCase();
|
|
var len = (str.length/2);
|
|
var lenlen = 0;
|
|
var hex = typ;
|
|
|
|
// We can't have an odd number of hex chars
|
|
if (len !== Math.round(len)) {
|
|
throw new Error("invalid hex");
|
|
}
|
|
|
|
// The first byte of any ASN.1 sequence is the type (Sequence, Integer, etc)
|
|
// The second byte is either the size of the value, or the size of its size
|
|
|
|
// 1. If the second byte is < 0x80 (128) it is considered the size
|
|
// 2. If it is > 0x80 then it describes the number of bytes of the size
|
|
// ex: 0x82 means the next 2 bytes describe the size of the value
|
|
// 3. The special case of exactly 0x80 is "indefinite" length (to end-of-file)
|
|
|
|
if (len > 127) {
|
|
lenlen += 1;
|
|
while (len > 255) {
|
|
lenlen += 1;
|
|
len = len >> 8;
|
|
}
|
|
}
|
|
|
|
if (lenlen) { hex += Enc.numToHex(0x80 + lenlen); }
|
|
return hex + Enc.numToHex(str.length/2) + str;
|
|
};
|
|
|
|
// The Integer type has some special rules
|
|
ASN1.UInt = function UINT() {
|
|
var str = Array.prototype.slice.call(arguments).join('');
|
|
var first = parseInt(str.slice(0, 2), 16);
|
|
|
|
// If the first byte is 0x80 or greater, the number is considered negative
|
|
// Therefore we add a '00' prefix if the 0x80 bit is set
|
|
if (0x80 & first) { str = '00' + str; }
|
|
|
|
return ASN1('02', str);
|
|
};
|
|
|
|
// The Bit String type also has a special rule
|
|
ASN1.BitStr = function BITSTR() {
|
|
var str = Array.prototype.slice.call(arguments).join('');
|
|
// '00' is a mask of how many bits of the next byte to ignore
|
|
return ASN1('03', '00' + str);
|
|
};
|
|
|
|
|
|
//
|
|
// Parser
|
|
//
|
|
|
|
ASN1.ELOOP = "uASN1.js Error: iterated over 15+ elements (probably a malformed file)";
|
|
ASN1.EDEEP = "uASN1.js Error: element nested 10+ layers deep (probably a malformed file)";
|
|
// Container Types are Sequence 0x30, Octect String 0x04, Array? (0xA0, 0xA1)
|
|
// Value Types are Integer 0x02, Bit String 0x03, Null 0x05, Object ID 0x06,
|
|
// Sometimes Bit String is used as a container (RSA Pub Spki)
|
|
ASN1.VTYPES = [ 0x02, 0x03, 0x05, 0x06 ];
|
|
ASN1.parse = function parseAsn1(buf, depth) {
|
|
if (depth >= 10) { throw new Error(ASN1.EDEEP); }
|
|
|
|
var index = 2; // we know, at minimum, data starts after type (0) and lengthSize (1)
|
|
var asn1 = { type: buf[0], lengthSize: 0, length: buf[1] };
|
|
var child;
|
|
var iters = 0;
|
|
var adjust = 0;
|
|
|
|
// Determine how many bytes the length uses, and what it is
|
|
if (0x80 & asn1.length) {
|
|
asn1.lengthSize = 0x7f & asn1.length;
|
|
// I think that buf->hex->int solves the problem of Endianness... not sure
|
|
asn1.length = parseInt(buf.slice(index, index + asn1.lengthSize).toString('hex'), 16);
|
|
index += asn1.lengthSize;
|
|
}
|
|
|
|
// High-order bit Integers have a leading 0x00 to signify that they are positive.
|
|
// Bit Streams use the first byte to signify padding, which x.509 doesn't use.
|
|
if (0x00 === buf[index] && (0x02 === asn1.type || 0x03 === asn1.type)) {
|
|
index += 1;
|
|
adjust = -1;
|
|
}
|
|
|
|
// this is a primitive value type
|
|
if (-1 !== ASN1.VTYPES.indexOf(asn1.type)) {
|
|
asn1.value = buf.slice(index, index + asn1.length + adjust);
|
|
return asn1;
|
|
}
|
|
|
|
asn1.children = [];
|
|
while (iters < 15 && index < buf.byteLength) {
|
|
iters += 1;
|
|
child = ASN1.parse(buf.slice(index, index + asn1.length), (depth || 0) + 1);
|
|
index += (2 + child.lengthSize + child.length);
|
|
asn1.children.push(child);
|
|
}
|
|
if (iters >= 15) { throw new Error(ASN1.ELOOP); }
|
|
|
|
return asn1;
|
|
};
|
|
|
|
/*
|
|
ASN1._stringify = function(asn1) {
|
|
//console.log(JSON.stringify(asn1, null, 2));
|
|
//console.log(asn1);
|
|
var ws = '';
|
|
|
|
function write(asn1) {
|
|
console.log(ws, 'ch', Buffer.from([asn1.type]).toString('hex'), asn1.length);
|
|
if (!asn1.children) {
|
|
return;
|
|
}
|
|
asn1.children.forEach(function (a) {
|
|
ws += '\t';
|
|
write(a);
|
|
ws = ws.slice(1);
|
|
});
|
|
}
|
|
write(asn1);
|
|
};
|
|
*/
|
|
|
|
module.exports = ASN1;
|