Major cleanup in our baseXX APIs

[dgoulet]

Our base16/32/64 APIs are quite inconsistent and a timebomb of possible errors and issues. Here is some recommendation with this:

1) We should have for _each_ type of encoding a function that returns the encoded size using a source length. We have such function for baese32 and base64 but they are not consistent:

• base32: base32_encoded_size() returns the size including the NUL byte

• base64: base64_encode_size() has a different name and does NOT add the NUL byte.

They should all return the size with the extra NUL byte because every _encode() function we have requires it in the first place. The other solution is to have a new function baseXX_encoded_string_size() or something that return the NUL byte and the non string function returns the value without it.

Else we end up with this kind of code pattern:

len = base64_encoded_size() + 1
buf = tor_malloc_zero(len)
ret = base64_encode()
tor_assert(ret == len - 1)

or the +1 added explicitly like so which is _not_ good.

base32_encode(serviceid, REND_SERVICE_ID_LEN_BASE32+1,
              circuit->rend_data->rend_pk_digest, REND_SERVICE_ID_LEN);

or even better:

base16_encode(hex, 2*CONDITIONAL_CONSENSUS_FPR_LEN+1,
              ds->v3_identity_digest, CONDITIONAL_CONSENSUS_FPR_LEN);

2) Source length requirement issue. I think though most of them are fixed except base64 API with ticket #17868.

I do see this in the base16_decode which could either be a _hard_ requirement assert or assume that if it gets 9 bytes in srclen, well only 8 are decoded. We have callsites that do NOT check the returned value so...

  if ((srclen % 2) != 0)
    return -1;

3) Every baseXX_encode/decode should return the amount of bytes that has been encoded or decoded. They also should return ssize_t; but that's another ticket I feel like because it's a large refactoring. Here are the missing pieces:

• base16_encode() returns void so it should return int as the encoded length.
• base32_encode() returns void.
• base32_decode() returns 0 on success.

4) We should also have macros for the encoded length computation else we ended up with stuff like this (taken from the prop250 branch).

#define SR_SRV_VALUE_BASE64_LEN \
  (((DIGEST256_LEN - 1) / 3) * 4 + 4)

or flat values

#define REND_DESC_COOKIE_LEN_BASE64 22

This is a static calculation so most of the times that macro would be more useful than the function since it could be used for stack allocation which is a BIG use case of ours.

[asn]

There is also a base64_decode() overflow at #19222 which might be worth fixing.

[dgoulet]

Ok, we are getting to a point we have lots of tickets about the baseXX APIs so I'm making this one the parent ticket for them all. Marking it for 030 because I believe we should fix all those sooner than later.

[teor]

Milestone renamed

[nickm]

Finally admitting that 0.3.??? was a euphemism for Tor: unspecified all along.

[catalyst]

Further complication: in our baseXX APIs, base64 is optionally padded, but base32 never is (even though the RFC4648 default is to pad).

Also, should we reject encodings whose padding bits are non-zero?

[nickm]

Also, should we reject encodings whose padding bits are non-zero?

You mean, ones where the padding isn't made of "="s ? Or something else?

[catalyst]

I mean where a non-= character at the end of the encoding is supposed to contain some padding bits. e.g., for a base64 encoding where the input length is 1 mod 3, the last 4 characters are XY==, where X can be any base64 alphabet character, but there are only 4 valid canonical values for Y (because it encodes the low order 2 bits of the last input byte, along with 4 padding bits that should be zero).

To put it another way, when you're decoding a base64 encoding, you can end up with 4 or 2 extra bits at the end that don't go into an output byte. For a canonical encoding, these bits should be zero, but a lenient decoder might accept nonzero bits there.

[nickm]

Ah. Yes, I think we should probably reject stuff where the bits to discard are nonzero -- that would probably indicate an encoding error.

[catalyst]

REND_DESC_COOKIE_LEN_BASE64 is a special case because it's the length of a truncated base64 encoding (not just an unpadded one). rend_auth_encode_cookie() removes the A character that represents the low bits of the extended auth cookie that are known to be zero. This gives the encoded cookie different properties than a 22-byte unpadded base64 encoding (which would ordinarily decode to 16 bytes).

I think this is not a great situation but it's already deployed, right? We should probably document it better.

[nickm]

Remove an old triaging keyword.