e860333bf3 Fix typos (Riccardo Casatta)
9189539715 Use BufReader internally in StreamReader to avoid performance regression on existing callers (Riccardo Casatta)
5dfb93df71 Deprecate StreamReader (Riccardo Casatta)
9ca6c75b18 Bench StreamReader (Riccardo Casatta)
Pull request description:
`StreamReader` performance is extremely poor in case the object decoded is "big enough" for example a full Block.
In the common case, the buffer is 64k, so to successfully parse a 1MB block 16 decode attempts are made.
Even if a user increases the buffer size, `read` is not going to necessarily fill the buffer, as stated in the doc https://doc.rust-lang.org/stable/std/io/trait.Read.html#tymethod.read. In my tests, the reads are 64kB even with a 1MB buffer.
I think this is the root issue of the performance issue found in electrs in https://github.com/romanz/electrs/issues/547 and they now have decided to decode the TCP stream with their own code in cd0531b8b7 and 05e0221b8e.
Using directly `consensus_encode` seems to make more sense (taking care of using `BufRead` if necessary) so the `StreamReader` is deprecated
ACKs for top commit:
Kixunil:
ACK e860333bf3
apoelstra:
ACK e860333bf3
Tree-SHA512: a15a14f3f087be36271da5008d8dfb63866c9ddeb5ceb0e328b4a6d870131132a8b05103f7a3fed231f5bca099865efd07856b4766834d56ce2384b1bcdb889b
StreamReader before this commit is trying to repeatedly parse big object like
blocks at every read, causing useless overhead.
consensus_encode deal with partial data by simply blocking.
After this changes it doesn't look what remain of the StreamReader is really giving
value, so it's deprecated
This is the initial step towards using and maybe enforcing clippy.
It does not fix all lints as some are not applicable. They may be
explicitly ignored later.
Docs can always do with a bit of love.
Clean up the module level (`//!`) rustdocs for all public modules.
I claim uniform is better than any specific method/style. I tried to fit
in with what ever was either most sane of most prevalent, therefore
attaining uniformity without unnecessary code churn (one exception being
the changes to headings described below).
Notes:
* Headings - use heading as a regular sentence for all modules e.g.,
```
//! Bitcoin network messages.
```
as opposed to
```
//! # Bitcoin Network Messages
```
It was not clear which style to use so I picked a 'random' mature
project and copied their style.
* Added 'This module' in _most_ places as the start of the module
description, however I was not religious about this one.
* Fixed line length if necessary since most of our code seems to follow
short (80 char) line lengths for comments anyways.
* Added periods and fixed obvious (and sometimes not so obvious)
grammatically errors.
* Added a trailing `//!` to every block since this was almost universal
already. I don't really like this one but I'm guessing it is Andrew's
preferred style since its on the copyright notices as well.
This documents cargo features in two ways: explictly in text and in code
using `#[doc(cfg(...))]` attribute where possible. Notably, this is
impossible for `serde` derives. The attribute is contitional and only
activated for docs.rs or explicit local builds.
This change also adds `package.metadata.docs.rs` field to `Cargo.toml`
which instructs docs.rs to build with relevant features and with
`docsrs` config activated enabling `#[doc(cfg(...))] attributes.
I also took the opportunity to fix a few missing spaces in nearby code.
Based on the original work by Justin Moon.
*MSRV unchanged from 1.29.0.*
When `std` is off, `no-std` must be on, and we use the [`alloc`](https://doc.rust-lang.org/alloc/) and core2 crates. The `alloc` crate requires the user define a global allocator.
* Import from `core` and `alloc` instead of `std`
* `alloc` only used if `no-std` is on
* Create `std` feature
* Create `no-std` feature which adds a core2 dependency to polyfill `std::io` features. This is an experimental feature and should be
used with caution.
* CI runs tests `no-std`
* MSRV for `no-std` is 1.51 or so
This is instead of encode::Errors because the encoders should
not be allowed to return errors that don't originate in the writer
they are writing into.
This is a part of the method definition that has been relied upon for a
while already.
Currently whenever an unrecognized network message is received, it is never
flushed from the read buffer, meaning that unless the stream is closed and
recreated it will keep returning the same error every time `read_next()` is
called.
This commit adds the length of the message to `UnrecognizedNetworkCommand`,
so that the `StreamReader` can flush those bytes before returning the error
to the caller.
Riccardo Casatta
Martin Habovstiak
Tobin Harding
Ben Carman
Andrew Poelstra
Devrandom
Steven Roose
sanket1729
Alekos Filini
Jake Rawsthorne
Jake Rawsthorne
Elichai Turkel
Karl-Johan Alm
Alexis Sellier
kiminuo
Dr Maxim Orlovsky