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
106acdc3ac Add fuzzing for Witness struct (Riccardo Casatta)
2fd0125bfa Introduce Witness struct mainly to improve ser/de performance while keeping most usability. (Riccardo Casatta)
Pull request description:
At the moment the Witness struct is `Vec<Vec<u8>>`, the vec inside a vec cause a lot of allocations, specifically:
- empty witness -> 1 allocation, while an empty vec doesn't allocate, the outer vec is not empty
- witness with n elements -> n+1 allocations
The proposed Witness struct contains the serialized format of the witness. This reduces the allocations to:
- empty witness -> 0 allocations
- witness with n elements -> 1 allocation for most common cases (you don't know how many bytes is long the entire witness beforehand, thus you need to estimate a good value, not too big to avoid wasting space and not too low to avoid vector reallocation, I used 128 since it covers about 80% of cases on mainnet)
The inconvenience is having slightly less comfortable access to the witness, but the iterator is efficient (no allocations) and you can always collect the iteration to have a Vec of slices. If you collect the iteration you end up doing allocation anyway, but the rationale is that it is an operation you need to do rarely while ser/de is done much more often.
I had to add a bigger block to better see the improvement (ae860247e191e2136d7c87382f78c96e0908d700), these are the results of the benches on my machine:
```
RCasatta/master_with_block
test blockdata::block::benches::bench_block_deserialize ... bench: 5,496,821 ns/iter (+/- 298,859)
test blockdata::block::benches::bench_block_serialize ... bench: 437,389 ns/iter (+/- 31,576)
test blockdata::block::benches::bench_block_serialize_logic ... bench: 108,759 ns/iter (+/- 5,807)
test blockdata::transaction::benches::bench_transaction_deserialize ... bench: 670 ns/iter (+/- 49)
test blockdata::transaction::benches::bench_transaction_get_size ... bench: 7 ns/iter (+/- 0)
test blockdata::transaction::benches::bench_transaction_serialize ... bench: 51 ns/iter (+/- 5)
test blockdata::transaction::benches::bench_transaction_serialize_logic ... bench: 13 ns/iter (+/- 0)
branch witness_with_block (this one)
test blockdata::block::benches::bench_block_deserialize ... bench: 4,302,788 ns/iter (+/- 424,806)
test blockdata::block::benches::bench_block_serialize ... bench: 366,493 ns/iter (+/- 42,216)
test blockdata::block::benches::bench_block_serialize_logic ... bench: 84,646 ns/iter (+/- 7,366)
test blockdata::transaction::benches::bench_transaction_deserialize ... bench: 648 ns/iter (+/- 77)
test blockdata::transaction::benches::bench_transaction_get_size ... bench: 7 ns/iter (+/- 0)
test blockdata::transaction::benches::bench_transaction_serialize ... bench: 50 ns/iter (+/- 5)
test blockdata::transaction::benches::bench_transaction_serialize_logic ... bench: 14 ns/iter (+/- 0)
```
With an increased performance to deserialize a block of about 21% and to serialize a block of about 16% (seems even higher than expected, need to do more tests to confirm, I'll appreciate tests results from reviewers)
ACKs for top commit:
apoelstra:
ACK 106acdc3ac
sanket1729:
ACK 106acdc3ac
dr-orlovsky:
utACK 106acdc3ac
Tree-SHA512: e4f23bdd55075c7ea788bc55846fd9e30f9cb76d5847cb259bddbf72523857715b0d4dbac505be3dfb9d4b1bcae289384ab39885b4887e188f8f1c06caf4049a
Witness struct is in place of the Vec<Vec<u8>> we have before this commit.
from_vec() and to_vec() methods are provided to switch between this type and Vec<Vec<u8>>
Moreover, implementation of Default, Iterator and others allows to have similar behaviour but
using a single Vec prevent many allocations during deserialization which in turns results in
better performance, even 20% better perfomance on recent block.
last() and second_to_last() allows to access respective element without going through costly Vec
transformation
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.
The merkle_root of an empty tree is undefined, this is the only error
case we have for the two `bitcoin_merkle_root*` functions. We can fully
describe this error case by returning an `Option` if args are found to
be empty. We can do the same for the wrapper functions in `block`
module.
While we are at it, refactor out a recursive helper function to make
reading the code between the two functions easier.
Calls to `unwrap` outside of tests are typically unfavourable.
Hash engines do not error when calling `consensus_encode`. Instead of
the current usage of `unwrap` we can use `expect` with a descriptive
string as is done in other parts of the codebase.
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.
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
Validating a block's proof-of-work involves computing the block hash.
Returning it from BlockHeader::validate_pow avoids having callers
recompute the block hash if it is needed.
Taking an external dependency just to convert ints to byte arrays
is somewhat of a waste, especially when Rust isn't very aggressive
about doing cross-crate LTO.
Note that the latest LLVM pattern-matches this, and while I haven't
tested it, that should mean this means no loss of optimization.
The protocol has a bug where a 0u8 is pushed at the end of each
block header on the wire in headers messages. WHy this bug came
about is unrealted and shouldn't impact API design.
Riccardo Casatta
Dr. Maxim Orlovsky
Martin Habovstiak
Tobin Harding
Devrandom
Vis Virial
Jeffrey Czyz
Steven Roose
0xB10C
Elichai Turkel
Braydon Fuller
Alexis Sellier
Andrew Poelstra
Nadav Ivgi
Matt Corallo
Tamas Blummer
Jonas Nick
Matt Corallo
ariard
Carl Dong