Per last paragraph of RFC 8446, Section 5.2, the length of the inner content of an encrypted record must not exceed 2^14 + 1, while that of the whole encrypted record must not exceed 2^14 + 256.
Client for tls was using a function that wasn't declared on the
interface for it. The issue wasn't apparent because net stream
implemented that function.
I changed it to keep the interface promise of what's required to be
compatible with the tls client functionality.
Use inline to vastly simplify the exposed API. This allows a
comptime-known endian parameter to be propogated, making extra functions
for a specific endianness completely unnecessary.
Most of this migration was performed automatically with `zig fmt`. There
were a few exceptions which I had to manually fix:
* `@alignCast` and `@addrSpaceCast` cannot be automatically rewritten
* `@truncate`'s fixup is incorrect for vectors
* Test cases are not formatted, and their error locations change
Individual max buffer sizes are well known, now that arithmetic doesn't
require allocations any more.
Also bump `main_cert_pub_key_buf`, so that e.g. `nodejs.org` public
keys can fit.
A minimal set of simple, safe functions for Montgomery arithmetic,
designed for cryptographic primitives.
Also update the current RSA cert validation to use it, getting rid
of the FixedBuffer hack and the previous limitations.
Make the check of the RSA public key a little bit more strict by
the way.
* When there is buffered cleartext, return it without calling the
underlying read function. This prevents buffer overflow due to space
used up by cleartext.
* Avoid clearing the buffer when the buffered cleartext could not be
completely given to the result read buffer, and there is some
buffered ciphertext left.
* Instead of rounding up the amount of bytes to ask for to the nearest
TLS record size, round down, with a minimum of 1. This prevents the
code path from being taken which requires extra memory copies.
* Avoid calling `@memcpy` with overlapping arguments.
closes#15590
Now they use slices or array pointers with any element type instead of
requiring byte pointers.
This is a breaking enhancement to the language.
The safety check for overlapping pointers will be implemented in a
future commit.
closes#14040
On CPUs without AES support, ChaCha is always faster and safer than
software AES.
Add `crypto.core.aes.has_hardware_support` to represent whether
AES acceleration is available or not, and in `tls.Client`, favor
AES-based ciphers only if hardware support is available.
This matches what BoringSSL is doing.
Although RFC 8446 states:
> Each party MUST send a "close_notify" alert before closing its write
> side of the connection
In practice many servers do not do this. Also in practice, truncation
attacks are thwarted at the application layer by comparing the amount of
bytes received with the amount expected via the HTTP headers.
Previously, the code only checked Common Name, leading to unable to
validate valid certificates which relied on the subject_alt_name
extension for host name verification.
This commit also adds rsa_pss_rsae_* back to the signature algorithms
list in the ClientHello.
This commit adds `writeEnd` and `writeAllEnd` in order to send data and
also notify the server that there will be no more data written.
Unfortunately, it seems most TLS implementations in the wild get this
wrong and immediately close the socket when they see a close_notify,
rather than only ending the data stream on the application layer.
This commit introduces tls.Decoder and then uses it in tls.Client. The
purpose is to make it difficult to introduce vulnerabilities in the
parsing code. With this abstraction in place, bugs in the TLS
implementation will trip checks in the decoder, regardless of the actual
length of packets sent by the other party, so that we can have
confidence when using ReleaseFast builds.
The code we are borrowing from https://github.com/shiguredo/tls13-zig
requires an Allocator for doing RSA certificate verification. As a
stopgap measure, this commit uses a FixedBufferAllocator to avoid heap
allocation for these functions.
Thank you to @naoki9911 for providing this great resource which has been
extremely helpful for me when working on this standard library TLS
implementation. Until Zig has std.crypto.rsa officially, we will borrow
this implementation of RSA. 🙏
Here's what I landed on for the TLS client. It's 16896 bytes
(max_ciphertext_record_len is 16640). I believe this is the theoretical
minimum size, give or take a few bytes.
These constraints are satisfied:
* a call to the readvAdvanced() function makes at most one call to the
underlying readv function
* iovecs are provided by the API, and used by the implementation for
underlying readv() calls to the socket
* the theoretical minimum number of memcpy() calls are issued in all
circumstances
* decryption is only performed once for any given TLS record
* large read buffers are fully exploited
This is accomplished by using the partial read buffer to storing both
cleartext and ciphertext.
The read function has been renamed to readAdvanced since it has slightly
different semantics than typical read functions, specifically regarding
the end-of-file. A higher level read function is implemented on top.
Now, API users may pass small buffers to the read function and
everything will work fine. This is done by re-decrypting the same
ciphertext record with each call to read() until the record is finished
being transmitted.
If the buffer supplied to read() is large enough, then any given
ciphertext record will only be decrypted once, since it decrypts
directly to the read() buffer and therefore does not need any memcpy. On
the other hand, if the buffer supplied to read() is small, then the
ciphertext is decrypted into a stack buffer, a subset is copied to the
read() buffer, and then the entire ciphertext record is saved for the
next call to read().
