crypto/rand: cap getrandom reads to ~32 MB per call on Linux

[Yawning]

What version of Go are you using (go version)?

The problem is obvious from inspection of the source code and the documented Linux getrandom semantics, I looked at master.

What operating system and processor architecture are you using (go env)?

This affects all Linux platforms with support for the getrandom() system call.

What did you do?

Looked at the code to see if crypto/rand was correct or not.

What did you expect to see?

getrandom() used correctly.

What did you see instead?

From the getrandom documentation:

When reading from the urandom source, a maximum of 33554431 bytes is returned by a single call to getrandom() on systems where int has a size of 32 bits.

From the system call implementation in drivers/char/random.c:

	if (count > INT_MAX)
		count = INT_MAX;

src/crypto/rand/rand_linux.go:getRandomLinux() will return false due to the truncated getrandom() output, and the caller will fall back to servicing the request by opening and reading from /dev/urandom.

While I would be inclined to agree that a Read that is 32 MiB - 1 bytes or larger is excessive and out of the ordinary, this still should be handled correctly or documented.

[bradfitz]

Thanks for the report.

[mmcloughlin]

I just attended the Gophercon Contributor Workshop and thought this might be a manageable fix to take on.

I was looking around drivers/char/random.c as @Yawning pointed out, but I don't think the offending line is if (count > INT_MAX) ... because INT_MAX is probably 2^31-1. That line is here:

https://github.com/torvalds/linux/blob/486088bc4689f826b80aa317b45ac9e42e8b25ee/drivers/char/random.c#L1916

The syscall definition delegates to urandom_read which contains the line

nbytes = min_t(size_t, nbytes, INT_MAX >> (ENTROPY_SHIFT + 3));

https://github.com/torvalds/linux/blob/486088bc4689f826b80aa317b45ac9e42e8b25ee/drivers/char/random.c#L1782

And ENTROPY_SHIFT is 3 (defined near the top of the file). That explains why the limit is 2^25-1 rather than 2^31-1.

[mmcloughlin]

I have been working on a patch. I'll put it out tomorrow for review.

[gopherbot]

CL https://golang.org/cl/49170 mentions this issue.