Currently the shim.authenticate does not pass a &mut self and so it is not possible to save the username for later use.
Also the on_init does not pass username or the entire handshake but rather just the db.
Can we add the username to the on_init ?

I am trying to get greptimedb running on a Raspberry PI Zero 2W (512MB RAM, 64 bit processor but 32 bit PI OS seems to be recommended for stability).

Greptimedb depends on this library, which currently does not build on the PI, I think because it is a 32 bit system. The compilation error is:

  Compiling opensrv-mysql v0.7.0 (/home/gareth/opensrv/mysql)
error[E0277]: the trait bound `u64: ToUsize` is not satisfied
   --> mysql/src/commands.rs:76:44
    |
76  |                 nom::bytes::complete::take(size)(i)?
    |                 -------------------------- ^^^^ the trait `ToUsize` is not implemented for `u64`
    |                 |
    |                 required by a bound introduced by this call
    |
    = help: the following other types implement trait `ToUsize`:
              usize
              u8
              u16
              u32
note: required by a bound in `nom::bytes::complete::take`
   --> /home/gareth/.cargo/registry/src/index.crates.io-1cd66030c949c28d/nom-7.1.3/src/bytes/complete.rs:408:6
    |
403 | pub fn take<C, Input, Error: ParseError<Input>>(
    |        ---- required by a bound in this function
...
408 |   C: ToUsize,
    |      ^^^^^^^ required by this bound in `take`

Is there any chance of 32 bit support for this project?

We need to add the following information to ease collaboration

Basic documentation

• README
• LICENSE
• history & credits

Collaboration

• toolchain / other useful settings
• CONTRIBUTING
• PULL_REQUEST_TEMPLATE
• Workflows

Starting from MySQL 8.0, caching_sha2_password authentication is used by default, and mysql_native_password authentication is deprecated in MySQL 9.0. It may be better to support caching_sha2_password in opensrv.

Refer to: https://dev.mysql.com/doc/relnotes/mysql/9.0/en/news-9-0-0.html

ClickHouseSession is too complex and contains many small functions:

fn timezone(&self) -> &str {
    "UTC"
}

fn server_display_name(&self) -> &str {
    "clickhouse-server"
}

fn dbms_version_patch(&self) -> u64 {
    1
}

As a result, it's hard to use and implement one: our simple clickhouse contains nearly 200 lines.

How about returning a ClickHouseMetadata instead?

msql-srv is the source of opensrv-mysql, which is a synchronous implementation.

I think we don't need to keep maintaining the sync methods in opensrv-mysql. Let's focus on advancing the asynchronous implementation.

I found in the code the change of the truncate method in order to improve performance.
but the change cause an issue in case the rest field is greater than 0.

if rest greater than 0 the truncate method will truncate from self.bytes the first remaining bytes number which included some of the parsed bytes.

that scenario cause the next iteration to parse the same packet instead of the remaining packet it should be read from the beginning.

one optional fix - maybe instead of truncate the bytes vector you can set the start value to be the correct one to start from for next iteration.

but I want to consult with you guys before I change some low level code like that.

thanks.

cc @PsiACE

This will be available as an optional feature. We can track jonhoo/msql-srv#23 for a reference implementation.

mysql_common has released new version 0.31.0, with the unsound dependency lexical removed, see here: blackbeam/rust_mysql_common#102

Can opensrv-mysql upgrade its mysql_common dependency to 0.31.0?

In mysql/src/packet_reader.rs, the following code has a memory safety problem:

let bytes = {
    // NOTE: this is all sorts of unfortunate. what we really want to do is to give
    // &self.bytes[self.start..] to `packet()`, and the lifetimes should all work
    // out. however, without NLL, borrowck doesn't realize that self.bytes is no
    // longer borrowed after the match, and so can be mutated.
    let bytes = &self.bytes[self.start..];
    unsafe { ::std::slice::from_raw_parts(bytes.as_ptr(), self.remaining) }
};
match packet(bytes) {
    Ok((rest, p)) => {
        self.remaining = rest.len();
        if self.remaining > 0 {
            self.bytes = rest.to_vec();
            self.start = 0;
        }
        return Ok(Some(p));
    }
    Err(nom::Err::Incomplete(_)) | Err(nom::Err::Error(_)) => {}
    Err(nom::Err::Failure(ctx)) => {
        self.bytes.truncate(self.remaining);
        return Err(io::Error::new(
            io::ErrorKind::InvalidData,
            format!("{:?}", ctx),
        ));
    }
}

We can see that bytes borrows self.bytes, then p in the Ok match arm also borrows self.bytes (the signature of packet function is pub(crate) fn packet(i: &[u8]) -> nom::IResult<&[u8], (u8, Packet<'_>)>. If self.remaining > 0 holds, then the old self.bytes object will be dropped by self.bytes = rest.to_vec();, then p borrowing that object will be invalidated, but p will still be returned.

Another problem relating to pointer aliasing rule violation may exist: the statement self.bytes.truncate(self.remaining); creates a mutable borrow of self.bytes, while another immutable borrow of self.bytes is live (consider the variable ctx). By language reference, breaking the pointer aliasing rules will cause undefined behavior. This can be solved by rewriting the corresponding match arm to:

Err(nom::Err::Failure(ctx)) => {
    let err = Err(io::Error::new(
        io::ErrorKind::InvalidData,
        format!("{:?}", ctx),
    ));
    self.bytes.truncate(self.remaining);
    return err;
}

In decode.rs:

The code asserts that v must have length of either 7 or 11.

However, it seems that 0 and 4 are also valid:

from https://mariadb.com/kb/en/resultset-row/#timestamp-binary-encoding

Any specific considerations why this assertion only makes 7 and 11 valid here?