Includes a simple test harness for the merge interop M1 milestone
This aims to enable connecting nimbus-eth2 to nimbus-eth1 within
the testing protocol described here:
https://github.com/status-im/nimbus-eth2/blob/amphora-merge-interop/docs/interop_merge.md
To execute the work-in-progress test, please run:
In terminal 1:
tests/amphora/launch-nimbus.sh
In terminal 2:
tests/amphora/check-merge-test-vectors.sh
why:
Previously, the function 'snapshot_desc.loadSnapshot()' contained the
equivalent of 'eth.decode(@[],SnapshotData)' for some type 'SnapshotData'
which should result in an exception of type 'RlpTypeMismatch'.
Before mid October, this worked for all systems on the Github CI. Since
then, a segfault message in the Github CI can be reproduced on all 64bit
Windows wuns when running 'build/all_tests <id-of-test_txpool>' after the
failed 'make test' directive (the latter one needs to be extended by
'|| true'.) This error cannot be reproduced on my local Win7/64 system
with the same MSYS2 and gcc 11.2.0 compiler.
The fix is, rather than catching an exception, to explicitly check the
first argument of 'eth.decode(@[],SnapshotData)' and act if it is empty.
also:
removed some obsolete {.inline.} annotations.
* Redesign of BaseVMState descriptor
why:
BaseVMState provides an environment for executing transactions. The
current descriptor also provides data that cannot generally be known
within the execution environment, e.g. the total gasUsed which is
available not before after all transactions have finished.
Also, the BaseVMState constructor has been replaced by a constructor
that does not need pre-initialised input of the account database.
also:
Previous constructor and some fields are provided with a deprecated
annotation (producing a lot of noise.)
* Replace legacy directives in production sources
* Replace legacy directives in unit test sources
* fix CI (missing premix update)
* Remove legacy directives
* chase CI problem
* rebased
* Re-introduce 'AccountsCache' constructor optimisation for 'BaseVmState' re-initialisation
why:
Constructing a new 'AccountsCache' descriptor can be avoided sometimes
when the current state root is properly positioned already. Such a
feature existed already as the update function 'initStateDB()' for the
'BaseChanDB' where the accounts cache was linked into this desctiptor.
The function 'initStateDB()' was removed and re-implemented into the
'BaseVmState' constructor without optimisation. The old version was of
restricted use as a wrong accounts cache state would unconditionally
throw an exception rather than conceptually ask for a remedy.
The optimised 'BaseVmState' re-initialisation has been implemented for
the 'persistBlocks()' function.
also:
moved some test helpers to 'test/replay' folder
* Remove unused & undocumented fields from Chain descriptor
why:
Reduces attack surface in general & improves reading the code.
* PoW wrapper for verification & mining
why:
It eases data management of per-Epoch lookup tables. Also some unit
tests show limits of usefulness on non-specialised machines for
mining besides developing tests.
details:
For PoW verification, this patch provides a pretty wrapper hiding the
details of the ethash/Hashimoto lookup cache management.
For mining on my development system without special hardware, the
underlying ethash functions are prohibitively slow. It takes
* ~20 minutes to prepare the full ethash/Hashimoto lookup dataset
* a second to run ~25k nonce tests (in the mining loop)
The mining part might be of some use for generating test data for
the tx-pool, though.
* Using PowRef as replacement for EpochHashCache + hashimotoLight()
* Fix typo (CI failed)
why:
was below log level when testing locally
* fix canonical naming
detected when running hive consensus simulator.
when processing an invalid block header and then
a new valid block header with the same block number,
the state root of the stateDB object should be updated
or reverted to parent stateRoot.
using intermediate stateRoot will trigger the hexary trie assertion.
previously, every time the VMState was created, it will also create
new stateDB, and this action will nullify the advantages of cached accounts.
the new changes will conserve the accounts cache if the executed blocks
are contiguous. if not the stateDB need to be reinited.
this changes also allow rpcCallEvm and rpcEstimateGas executed properly
using current stateDB instead of creating new one each time they are called.
* Provide API
details:
API is bundled via clique.nim.
* Set extraValidation as default for PoA chains
why:
This triggers consensus verification and an update of the list
of authorised signers. These signers are integral part of the
PoA block chain.
todo:
Option argument to control validation for the nimbus binary.
* Fix snapshot state block number
why:
Using sub-sequence here, so the len() function was wrong.
* Optional start where block verification begins
why:
Can speed up time building loading initial parts of block chain. For
PoA, this allows to prove & test that authorised signers can be
(correctly) calculated starting at any point on the block chain.
todo:
On Goerli around blocks #193537..#197568, processing time increases
disproportionally -- needs to be understand
* For Clique test, get old grouping back (7 transactions per log entry)
why:
Forgot to change back after troubleshooting
* Fix field/function/module-name misunderstanding
why:
Make compilation work
* Use eth_types.blockHash() rather than utils.hash() in Clique modules
why:
Prefer lib module
* Dissolve snapshot_misc.nim
details:
.. into clique_verify.nim (the other source file clique_unused.nim
is inactive)
* Hide unused AsyncLock in Clique descriptor
details:
Unused here but was part of the Go reference implementation
* Remove fakeDiff flag from Clique descriptor
details:
This flag was a kludge in the Go reference implementation used for the
canonical tests. The tests have been adapted so there is no need for
the fakeDiff flag and its implementation.
* Not observing minimum distance from epoch sync point
why:
For compiling PoA state, the go implementation will walk back to the
epoch header with at least 90000 blocks apart from the current header
in the absence of other synchronisation points.
Here just the nearest epoch header is used. The assumption is that all
the checkpoints before have been vetted already regardless of the
current branch.
details:
The behaviour of using the nearest vs the minimum distance epoch is
controlled by a flag and can be changed at run time.
* Analysing processing time (patch adds some debugging/visualisation support)
why:
At the first half million blocks of the Goerli replay, blocks on the
interval #194854..#196224 take exceptionally long to process, but not
due to PoA processing.
details:
It turns out that much time is spent in p2p/excecutor.processBlock()
where the elapsed transaction execution time is significantly greater
for many of these blocks.
Between the 1371 blocks #194854..#196224 there are 223 blocks with more
than 1/2 seconds execution time whereas there are only 4 such blocks
before and 13 such after this range up to #504192.
* fix debugging symbol in clique_desc (causes CI failing)
* Fixing canonical reference tests
why:
Two errors were introduced earlier but ovelooked:
1. "Remove fakeDiff flag .." patch was incomplete
2. "Not observing minimum distance .." introduced problem w/tests 23/24
details:
Fixing 2. needed to revert the behaviour by setting the
applySnapsMinBacklog flag for the Clique descriptor. Also a new
test was added to lock the new behaviour.
* Remove cruft
why:
Clique/PoA processing was intended to take place somewhere in
executor/process_block.processBlock() but was decided later to run
from chain/persist_block.persistBlock() instead.
* Update API comment
* ditto
* Renamed source file clique_utils => clique_helpers
why:
New name is more in line with other modules where local libraries
are named similarly.
* re-implemented PoA verification module as clique_verify.nim
details:
The verification code was ported from the go sources and provisionally
stored in the clique_misc.nim source file.
todo:
Bring it to life.
* re-design Snapshot descriptor as: ref object
why:
Avoids some copying descriptor objects
details:
The snapshot management in clique_snapshot.nim has been cleaned up.
todo:
There is a lot of unnecessary copying & sub-list manipulation of
seq[BlockHeader] lists which needs to be simplified by managing
index intervals.
* optimised sequence handling for Clique/PoA
why:
To much ado about nothing
details:
* Working with shallow sequences inside PoA processing avoids
unnecessary copying.
* Using degenerate lists in the cliqueVerify() batch where only the
parent (and no other ancestor) is needed.
todo:
Expose only functions that are needed, shallow sequences should be
handles with care.
* fix var-parameter function argument
* Activate PoA engine -- currently proof of concept
details:
PoA engine is activated with newChain(extraValidation = true) applied
to a PoA network.
status and todo:
The extraValidation flag on the Chain object can be set at a later
state which allows to pre-load parts of the block chain without
verification. Setting it later will only go back the block chain to
the latest epoch checkpoint. This is inherent to the Clique protocol,
needs testing though.
PoA engine works in fine weather mode on Goerli replay. With the
canonical eip-225 tests, there are quite a few fringe conditions
that fail. These can easily fudged over to make things work but need
some more work to understand and correct properly.
* Make the last offending verification header available
why:
Makes some fringe case tests work.
details:
Within a failed transaction comprising several blocks, this
feature help to identify the offending block if there was a
PoA verification error.
* Make PoA header verifier store the final snapshot
why:
The last snapshot needed by the verifier is the one of the parent but
the list of authorised signer is derived from the current snapshot. So
updating to the latest snapshot provides the latest signers list.
details:
Also, PoA processing has been implemented as transaction in
persistBlocks() with Clique state rollback.
Clique tests succeed now.
* Avoiding double yields in iterator => replaced by template
why:
Tanks to Andri who observed it (see #762)
* Calibrate logging interval and fix logging event detection
why:
Logging interval as copied from Go implementation was too large and
needed re-calibration. Elapsed time calculation was bonkers, negative
the wrong way round.
* re-shuffled Clique functions
why:
Due to the port from the go-sources, the interface logic is not optimal
for nimbus. The main visible function is currently snapshot() and most
of the _procurement_ of this function result has been moved to a
sub-directory.
* run eip-225 Clique test against p2p/chain.persistBlocks()
why:
Previously, loading the test block chains was fugdged with the purpose
only to fill the database. As it is now clear how nimbus works on
Goerli, the same can be achieved with a more realistic scenario.
details:
Eventually these tests will be pre-cursor to the reply tests for the
Goerli chain supporting TDD approach with more simple cases.
* fix exception annotations for executor module
why:
needed for exception tracking
details:
main annoyance are vmState methods (in state.nim) which potentially
throw a base level Exception (a proc would only throws CatchableError)
* split p2p/chain into sub-modules and fix exception annotations
why:
make space for implementing PoA stuff
* provide over-loadable Clique PRNG
why:
There is a PRNG provided for generating reproducible number sequences.
The functions which employ the PRNG to generate time slots were ported
ported from the go-implementation. They are currently unused.
* implement trusted signer assembly in p2p/chain.persistBlocks()
details:
* PoA processing moved there at the end of a transaction. Currently,
there is no action (eg. transaction rollback) if this fails.
* The unit tests with staged blocks work ok. In particular, there should
be tests with to-be-rejected blocks.
* TODO: 1.Optimise throughput/cache handling; 2.Verify headers
* fix statement cast in pool.nim
* added table features to LRU cache
why:
Clique uses the LRU cache using a mixture of volatile online items
from the LRU cache and database checkpoints for hard synchronisation.
For performance, Clique needs more table like features.
details:
First, last, and query key added, as well as efficient random delete
added. Also key-item pair iterator added for debugging.
* re-factored LRU snapshot caching
why:
Caching was sub-optimal (aka. bonkers) in that it skipped over memory
caches in many cases and so mostly rebuild the snapshot from the
last on-disk checkpoint.
details;
The LRU snapshot toValue() handler has been moved into the module
clique_snapshot. This is for the fact that toValue() is not supposed
to see the whole LRU cache database. So there must be a higher layer
working with the the whole LRU cache and the on-disk checkpoint
database.
also:
some clean up
todo:
The code still assumes that the block headers are valid in itself. This
is particular important when an epoch header (aka re-sync header) is
processed as it must contain the PoA result of all previous headers.
So blocks need to be verified when they come in before used for PoA
processing.
* fix some snapshot cache fringe cases
why:
Must not index empty sequences in clique_snapshot module
jangko
Zahary Karadjov
Jordan Hrycaj