| 1 |
to-do list for libxcrypt |
| 2 |
------------------------ |
| 3 |
|
| 4 |
This list is categorized but not in any kind of priority order. |
| 5 |
It was last updated 20 October 2018. |
| 6 |
|
| 7 |
* Code cleanliness |
| 8 |
* Find and remove any code that still does dodgy things with type punning |
| 9 |
* Factor out all of the repetitive base64 code |
| 10 |
* Factor out the multiple implementations of HMAC and PBKDF |
| 11 |
|
| 12 |
* Testsuite improvements |
| 13 |
* Investigate branch coverage |
| 14 |
* Do some API fuzz testing and add missing cases to the testsuite |
| 15 |
* Many of the `test-crypt-*.c` files repeat more or less the same |
| 16 |
code with different data, consider merging them |
| 17 |
|
| 18 |
* Portability |
| 19 |
* Make sure the symbol versioning macros work with all of the |
| 20 |
compilers that anyone needs (they use GCC extensions that clang |
| 21 |
also supports). |
| 22 |
|
| 23 |
* Hardening |
| 24 |
* bcrypt-like selftest/memory scribble for all hashing methods |
| 25 |
* how do we know the memory scribble is doing its job? |
| 26 |
* build out of the box with compiler hardening features turned on |
| 27 |
* something bespoke for not having to write serialization and |
| 28 |
deserialization logic for hash strings by hand, as this is |
| 29 |
probably the most error-prone part of writing a hashing method |
| 30 |
|
| 31 |
* the most sensitive piece of data handled by this library is a |
| 32 |
cleartext passphrase. OS may have trusted-path facilities for |
| 33 |
prompting the user for a passphrase and feeding it to a KDF |
| 34 |
without its ever being accessible in normal memory. investigate |
| 35 |
whether we can use these. |
| 36 |
|
| 37 |
* Additional hashing methods |
| 38 |
* Argon2 <https://password-hashing.net/> |
| 39 |
* ...? |
| 40 |
|
| 41 |
* Runtime configurability (in progress on the [crypt.conf branch][]) |
| 42 |
* allow installations to enable or disable specific hash methods |
| 43 |
without rebuilding the library |
| 44 |
* make the default cost parameter used by `crypt_gensalt_*` for new |
| 45 |
hashes configurable |
| 46 |
* update the compiled-in defaults used by `crypt_gensalt_*` (not the |
| 47 |
defaults used when no explicit cost parameter is present in a |
| 48 |
hash; those can’t be changed without breaking existing stored hashes) |
| 49 |
* relevant benchmarking at |
| 50 |
<https://pthree.org/2016/06/28/lets-talk-password-hashing/> |
| 51 |
* offer a way to tune cost parameters for a specific installation |
| 52 |
* N.B. Solaris 11 has all of these features but our implementation will |
| 53 |
probably not match them (they have a `crypt.conf` but it’s not the |
| 54 |
same, and their `crypt_gensalt` is API-incompatible anyway). |
| 55 |
|
| 56 |
[crypt.conf branch]: https://github.com/besser82/libxcrypt/tree/zack/crypt.conf |
| 57 |
|
| 58 |
* Potential API enhancements: |
| 59 |
|
| 60 |
* Support for "pepper" (an additional piece of information, _not_ |
| 61 |
stored in the password file, that you need to check a password) |
| 62 |
|
| 63 |
* Reading passphrases from the terminal is finicky and there are |
| 64 |
several competing, poorly portable, questionably sound library |
| 65 |
functions to do it (`getpass`, `readpassphrase`, etc) -- should we |
| 66 |
incorporate one? |
| 67 |
* If we do, should it know how to trigger the trusted-path |
| 68 |
password prompt in modern GUI environments? (probably) |
| 69 |
|
| 70 |
* Make the crypt and crypt_gensalt static state thread-specific? |
| 71 |
* Solaris 11 may have done this (its `crypt(3)` manpage describes |
| 72 |
it as MT-Safe and I don’t see any other way they could have |
| 73 |
accomplished that). |
| 74 |
* if allocated on first use, this would also shave 32kB of |
| 75 |
data segment off the shared library |
| 76 |
* alternatively, add a global lock and *crash the program* if we |
| 77 |
detect concurrent calls |
| 78 |
|
| 79 |
* Allow access to more of yescrypt’s tunable parameters and ROM |
| 80 |
feature, in a way that’s generic enough that we could also use it |
| 81 |
for e.g. Argon2’s tunable parameters |
| 82 |
|
| 83 |
* Other yescrypt-inspired features relevant to using this library to |
| 84 |
back a “dedicated authentication service,” e.g. preallocation of |
| 85 |
large blocks of scratch memory |
| 86 |
* the main obstacles here are that `struct crypt_data` has a fixed |
| 87 |
size which is either too big or too small depending how you look |
| 88 |
at it, and no destructor function |
| 89 |
|
| 90 |
* Permissive relicensing, to encourage use beyond the GNU ecosystem? |
| 91 |
* Replace crypt-md5.c with original md5crypt from FreeBSD? |
| 92 |
* Other files subject to the (L)GPL are crypt.c, crypt-static.c, |
| 93 |
crypt-gensalt-static.c, crypt-obsolete.h, crypt-port.h, |
| 94 |
test-badsalt.c. It is not clear to me how much material originally |
| 95 |
assigned to the FSF remains in these files. |
| 96 |
Several of them are API definitions and trivial wrappers that |
| 97 |
could not be meaningfully changed without breaking them (so are |
| 98 |
arguably uncopyrightable). |
| 99 |
* Most of the test suite lacks any license or even authorship |
| 100 |
information. We would have to track down the original authors. |