NewV4: non-random uuid

[josselin-c]

I'm running this on my macbook pro.

I'm using uuid.Must(uuid.NewV4()).String() to generate random identifiers. I'm generating theses identifiers at a very low rate of a few dozen per days

Here are some non-random UUID that I got in the last weeks:

02524e6f-65a7-4cf5-8000-0000000000002
6e3ef1c8-0000-4000-8000-0000000000001
fa07f1e9-a8a0-427f-8103-e34e000000000
2cfa392b-71d5-4000-8000-000000000000
5f16db16-56ce-4000-8000-000000000000
bac73b37-aab3-498b-8000-000000000000
da5bd82d-4f98-4b51-8000-000000000000
eb245e52-535f-4274-8350-3a7200000000
f7510000-0000-4000-8000-000000000000
5936f955-286e-47ac-8000-000000000000
0a14da92-0000-4000-8000-000000000000
80730000-0000-4000-8000-000000000000
cd2b88a5-0000-4000-8000-000000000000

I think I shouldn't have theses.

[domino14]

humans are bad at telling randomness. While it is extraordinarily, astronomically unlikely that you would have so many uuids with that many zeros, it is still theoretically possible.

[josselin-c]

In an IDs such as 80730000-0000-4000-8000-000000000000 there are ~14 consecutive bytes of zeros.. While a probability of something like 1/2**(14*8) isn't nil, it's still too low for my taste. Drawing theses IDs 5 times out of ~1000 isn't just bad luck. In real life, if it was possible, I would start mining bitcoins with my super powers!

6e3ef1c8-0000-4000-8000-0000000000001
f7510000-0000-4000-8000-000000000000
0a14da92-0000-4000-8000-000000000000
80730000-0000-4000-8000-000000000000
cd2b88a5-0000-4000-8000-000000000000

I don't know the root cause of this.

[peterstace]

These UUIDs seem suspiciously non-random to me as well.

NewV4 gets its randomness from crypto/rand.Reader (and is basically a passthrough to populate a UUID). The code in NewV4 is really simple, I don't see how it could go wrong.

The only possibilities I can think of:

• There is a bug in crypto/rand (seems unlikely).
• Since crypto/rand.Reader is just a variable, some malicious/accidental code could be swapping it out for a less random io.Reader implementation.

[josselin-c]

Okay, I can reproduce it!
On my machine I have two separate processes: one webserver (ServerA) that calls uuid.Must(uuid.NewV4()).String() while simultaneously doing http requests to another webserver (ServerB) used for image resizing (imgproxy).

I instrumented ServerA with this code:

	go func() {
		for {
			time.Sleep(1 * time.Second)
			for i := 0; i < 1000; i++ {
				u := uuid.Must(uuid.NewV4()).String()
				if strings.Contains(u, "000000000") {
					log.Fatal(u)
				}
			}
		}
	}()

Which runs without crashing if ServerA idle or handling standard requests. If ServerA is calling ServerB (to resize images), the ServerA immediately crash with a bad uuid.
I'll try to make a self-contained reproducer

[josselin-c]

Okay, I found the problem.
Here is the current NewV4 code:

// NewV4 returns random generated UUID.
func (g *rfc4122Generator) NewV4() (UUID, error) {
	var err error
	var c int
	u := UUID{}
	if _, err = g.rand.Read(u[:]); err != nil {
		return Nil, err
	}
	u.SetVersion(V4)
	u.SetVariant(VariantRFC4122)

	return u, nil
}

Read() gives no guaranties about the number of bytes actually read. From https://golang.org/pkg/io/#Reader:

Read reads up to len(p) bytes into p. It returns the number of bytes read (0 <= n <= len(p)) and any error encountered

Turns out, on some occasions Read would read less than len(u) bytes. I thinks uuid should use ReadFull instead of plain Read().

[domino14]

@josselin-c Could you make a patch and see if you can reproduce the error with your script and ReadFull? I wonder if using ReadFull would outright return an error in this case, because there aren't enough bytes to read.

[domino14]

One more thing that is interesting to me is the fact that the UUIDs you posted are not of equal length.

[mikesun]

So it looks like this commit introduced the bug: 0ef6afb

Prior to that commit, rand.Read() was used, which is a wrapper that uses io.ReadFull(): https://golang.org/pkg/crypto/rand/#Read

[cameracker]

How is a critical bug like this still unresolved? Especially given that someone already did the work to fix it.