Closed preston-evans98 closed 9 months ago
Thanks @preston-evans98 for the changes. Hopefully, @bmwill and I will find time for a proper code review soon.
It looks like CI failed initially due to an outdated clippy allow
directive, and because cargo selected a version of thiserror
which is only compatible with the 2021 rust edition. I removed the lint, and pinned thiserror
to exactly the version specified in the Cargo.toml
. This should be safe to do because thiserror
does not appear in the crate's public API.
Hmm, looks like the MSRV failing due to unrelated changes. Are you opposed to checking in a Cargo.lock
?
@preston-evans98 Thanks for investigating. I fixed the CI on the main branch. Can you rebase your PR?
@preston-evans98 @bmwill This was just released in bcs v0.1.6
Fantastic, thanks for the review @ma2bd @bmwill!
This Pull Request implements BCS deserialization from implementers of
std::io::Read
.Strategy
The PR aims to maintain a minimal diff over the previous implementation. To that end, the following changes have been made.
Deserializer
generic over an inner typeR
BcsReader
, which implements BCS in terms of a few primitive operations. Implementserde::Deserialize<'de> for Deserializer<R> where Deserializer<R>: BcsReader<'de>
BcsReader<'de>
forDeserializer<&'de [u8]>
using existing codeBcsReader
generically for (a version of)Deserializer<R: Read>
.Implementing
BcsReader
genericallyThe existing implementation is very close to supporting deserialization from readers. There are only two places in which its behavior relied on access to a byte slice was in the implementation of
de::MapAccess
.The first case is
map
deserialization.bcs
needs access to the serialized representation of map keys in order to enforce canonicity. When deserializing from a slice, this is straightforward - the implementation can simply deserialize a map key, and then "look back" at the original input slice to determine its serialized representation. When deserializing from a reader, however, this kind of rewinding is not generally possible. To solve this problem, I introduce a theTeeReader
struct, which wraps aRead
er and optionally copies its bytes into acapture_buffer
for later retrieval. Then, I simply implementBcsReader
forDeserializer<TeeReader<...>>
.The second case is the
end
method, which checks that all input bytes have been consumed. To handle this case, I simply attempt to read one extra byte from theRead
er after deserialization and assert that an EOF error is returned from the underlyingRead
er.Testing
All existing unit tests except for
zero_copy_parse
have been modified to test that deserializationfrom_bytes
andfrom_reader
yield the same output. That test is not applicable since readers are not capable of zero-copy deserialization.