At Oasis we have spend some time writing a new Ed25519/X25519/sr25519 implementation called curve25519-voi. This PR switches the import from ed25519consensus/go-schnorrkel, which should lead to performance gains on most systems.
Summary of changes:
* curve25519-voi is now used for Ed25519 operations, following the existing ZIP-215 semantics.
* curve25519-voi's public key cache is enabled (hardcoded size of 4096 entries, should be tuned, see the code comment) to accelerate repeated Ed25519 verification with the same public key(s).
* (BREAKING) curve25519-voi is now used for sr25519 operations. This is a breaking change as the current sr25519 support does something decidedly non-standard when going from a MiniSecretKey to a SecretKey and or PublicKey (The expansion routine is called twice). While I believe the new behavior (that expands once and only once) to be more "correct", this changes the semantics as implemented.
* curve25519-voi is now used for merlin since the included STROBE implementation produces much less garbage on the heap.
Side issues fixed:
* The version of go-schnorrkel that is currently imported by tendermint has a badly broken batch verification implementation. Upstream has fixed the issue after I reported it, so the version should be bumped in the interim.
Open design questions/issues:
* As noted, the public key cache size should be tuned. It is currently backed by a trivial thread-safe LRU cache, which is not scan-resistant, but replacing it with something better is a matter of implementing an interface.
* As far as I can tell, the only reason why serial verification on batch failure is necessary is to provide more detailed error messages (that are only used in some unit tests). If you trust the batch verification to be consistent with serial verification then the fallback can be eliminated entirely (the BatchVerifier provided by the new library supports an option that omits the fallback if this is chosen as the way forward).
* curve25519-voi's sr25519 support could use more optimization and more eyes on the code. The algorithm unfortunately is woefully under-specified, and the implementation was done primarily because I got really sad when I actually looked at go-schnorrkel, and we do not use the algorithm at this time.
This PR make some tweaks to backfill after running e2e tests:
- Separates sync and backfill as two distinct processes that the node calls. The reason is because if sync fails then the node should fail but if backfill fails it is still possible to proceed.
- Removes peers who don't have the block at a height from the local peer list. As the process goes backwards if a node doesn't have a block at a height they're likely pruning blocks and thus they won't have any prior ones either.
- Sleep when we've run out of peers, then try again.
There is a possible data race/panic between processBlockchainCh and
processPeerUpdates, since when we send to blockchainCh.Out in one
goroutine and close the channel in the other. The race is seen in some
Github Action runs.
This commit fix the race, by adding a peerUpdatesCh as a bridge between
processPeerUpdates and processBlockchainCh, so the former will send to
this channel, the later will listen and forward the message to
blockchainCh.Out channel.
Updates #6516
## Description
Change block_size gauge to a histogram to observe block size overtime
This will help will see which chains have full blocks vs empty.
closes#5752
## Description
Internalize some libs. This reduces the amount ot public API tendermint is supporting. The moved libraries are mainly ones that are used within Tendermint-core.
Marko
JayT106
Yawning Angel
Sam Kleinman
Callum Waters
Aleksandr Bezobchuk
Cuong Manh Le