Do not use crypto rand for nonce generation.

[derekcollison]

Crypto rand is not needed for nonce generation and could drain entropy.

Signed-off-by: Derek Collison derek@nats.io

/cc @nats-io/core

[ripienaar]

more details on slack but calls to:

        rbek.XORKeyStream(buf, buf)

takes around 40ms each

The code here, which includes a call to XorKeyStream(), takes about 60 to 70ms

nats-server/server/filestore.go

Lines 538 to 560 in be17afb

	if createdKeys {
	buf, err := mb.loadBlock(nil)
	if err != nil {
	return nil, err
	}
	if err := mb.indexCacheBuf(buf); err != nil {
	// This likely indicates this was already encrypted or corrupt.
	mb.cache = nil
	return nil, err
	}
	// Undo cache from above for later.
	mb.cache = nil
	wbek, err := chacha20.NewUnauthenticatedCipher(mb.seed, mb.nonce)
	if err != nil {
	return nil, err
	}
	wbek.XORKeyStream(buf, buf)
	if err := ioutil.WriteFile(mb.mfn, buf, defaultFilePerms); err != nil {
	return nil, err
	}
	// Remove the index file here since it will be in plaintext as well so we just rebuild.
	os.Remove(mb.ifn)
	}

This has knock on effects that deleteFirstMsg() takes about 90ms per message, this is called for every message that needs to be expired - in my test I ensure I have 500 message that needs expiring to trigger these timings.

[derekcollison]

ok the expiration I guess was the missing piece to get to see the results you were seeing.

[derekcollison]

ok got a fix in, the expired was the key to triggering the bad behavior.

[kozlovic]

LGTM but a question about the test