Salt Generation and open source software

As I understand it, the best practice for generating salts is to use some cryptic formula (or even magic constant) stored in your source code.

I'm working on a project that we plan on releasing as open source, but the problem is that with the source comes the secret formula for generating salts, and therefore the ability to run rainbow table attacks on our site.

I figure that lots of people have contemplated this problem before me, and I'm wondering what the best practice is. It seems to me that there is no point having a salt at all if the code is open source, because salts can be easily reverse-engineered.

Thoughts?

Since questions about salting hashes come along on a quite regular basis and there seems to be quite some confusion about the subject, I extended this answer.

What is a salt?

A salt is a random set of bytes of a fixed length that is added to the input of a hash algorithm.

Why is salting (or seeding) a hash useful?

Adding a random salt to a hash ensures that the same password will produce many different hashes. The salt is usually stored in the database, together with the result of the hash function. Salting a hash is good for a number of reasons:

1. Salting greatly increases the difficulty/cost of precomputated attacks (including rainbow tables)
2. Salting makes sure that the same password does not result in the same hash. This makes sure you cannot determine if two users have the same password. And, even more important, you cannot determine if the same person uses the same password across different systems.
3. Salting increases the complexity of passwords, thereby greatly decreasing the effectiveness of both Dictionary- and Birthday attacks. (This is only true if the salt is stored separate from the hash).
4. Proper salting greatly increases the storage need for precomputation attacks, up to the point where they are no longer practical. (8 character case-sensitive alpha-numeric passwords with 16 bit salt, hashed to a 128 bit value, would take up just under 200 exabytes without rainbow reduction).

There is no need for the salt to be secret.

A salt is not a secret key, instead a salt 'works' by making the hash function specific to each instance. With salted hash, there is not one hash function, but one for every possible salt value. This prevent the attacker from attacking N hashed passwords for less than N times the cost of attacking one password. This is the point of the salt.
A "secret salt" is not a salt, it is called a "key", and it means that you are no longer computing a hash, but a Message Authentication Code (MAC). Computing MAC is tricky business (much trickier than simply slapping together a key and a value into a hash function) and it is a very different subject altogether.

The salt must be random for every instance in which it is used. This ensures that an attacker has to attack every salted hash separately.
If you rely on your salt (or salting algorithm) being secret, you enter the realms of Security Through Obscurity (won't work). Most probably, you do not get additional security from the salt secrecy; you just get the warm fuzzy feeling of security. So instead of making your system more secure, it just distracts you from reality.

So, why does the salt have to be random?

Technically, the salt should be unique. The point of the salt is to be distinct for each hashed password. This is meant worldwide. Since there is no central organization which distributes unique salts on demand, we have to rely on the next best thing, which is random selection with an unpredictable random generator, preferably within a salt space large enough to make collisions improbable (two instances using the same salt value).

It is tempting to try to derive a salt from some data which is "presumably unique", such as the user ID, but such schemes often fail due to some nasty details:

1. If you use for example the user ID, some bad guys, attacking distinct systems, may just pool their resources and create precomputed tables for user IDs 1 to 50. A user ID is unique system-wide but not worldwide.

2. The same applies to the username: there is one "root" per Unix system, but there are many roots in the world. A rainbow table for "root" would be worth the effort, since it could be applied to millions of systems. Worse yet, there are also many "bob" out there, and many do not have sysadmin training: their passwords could be quite weak.

3. Uniqueness is also temporal. Sometimes, users change their password. For each new password, a new salt must be selected. Otherwise, an attacker obtained the hash of the old password and the hash of the new could try to attack both simultaneously.

Using a random salt obtained from a cryptographically secure, unpredictable PRNG may be some kind of overkill, but at least it provably protects you against all those hazards. It's not about preventing the attacker from knowing what an individual salt is, it's about not giving them the big, fat target that will be used on a substantial number of potential targets. Random selection makes the targets as thin as is practical.

In conclusion:

Use a random, evenly distributed, high entropy salt. Use a new salt whenever you create a new password or change a password. Store the salt along with the hashed password. Favor big salts (at least 10 bytes, preferably 16 or more).

A salt does not turn a bad password into a good password. It just makes sure that the attacker will at least pay the dictionary attack price for each bad password he breaks.

Usefull sources:
stackoverflow.com: Non-random salt for password hashes
Bruce Schneier: Practical Cryptography (book)
Matasano Security: Enough with the Rainbow Tables
usenix.org: Unix crypt used salt since 1976
owasp.org: Why add salt
openwall.com: Salts

Disclaimer:
I'm not a security expert. (Although this answer was reviewed by Thomas Pornin)
If any of the security professionals out there find something wrong, please do comment or edit this wiki answer.

Really salts just need to be unique for each entry. Even if the attacker can calculate what the salt is, it makes the rainbow table extremely difficult to create. This is because the salt is added to the password before it is hashed, so it effectively adds to the total number of entries the rainbow table must contain to have a list of all possible values for a password field.

Since Unix became popular, the right way to store a password has been to append a random value (the salt) and hash it. Save the salt away where you can get to it later, but where you hope the bad guys won't get it.

This has some good effects. First, the bad guys can't just make a list of expected passwords like "Password1", hash them into a rainbow table, and go through your password file looking for matches. If you've got a good two-byte salt, they have to generate 65,536 values for each expected password, and that makes the rainbow table a lot less practical. Second, if you can keep the salt from the bad guys who are looking at your password file, you've made it much harder to calculate possible values. Third, you've made it impossible for the bad guys to determine if a given person uses the same password on different sites.

In order to do this, you generate a random salt. This should generate every number in the desired range with uniform probability. This isn't difficult; a simple linear congruential random number generator will do nicely.

If you've got complicated calculations to make the salt, you're doing it wrong. If you calculate it based on the password, you're doing it WAY wrong. In that case, all you're doing is complicating the hash, and not functionally adding any salt.

Nobody good at security would rely on concealing an algorithm. Modern cryptography is based on algorithms that have been extensively tested, and in order to be extensively tested they have to be well known. Generally, it's been found to be safer to use standard algorithms rather than rolling one's own and hoping it's good. It doesn't matter if the code is open source or not, it's still often possible for the bad guys to analyze what a program does.