backport of https://github.com/decred/dcrd/pull/1273
Notable difference being that btcd mainline currenlty
doesn't have a blockchain/blockindex_test.go file, so
those changes are omitted.
Great work @davecgh :)
This generalizes the reorganizeChain function in the blockchain package
to allow it to rewind the chain in the case no attach nodes are provided
or extend the chain in the case no detach nodes are provided.
It also adds several assertions to ensure the assumptions about the
state hold and cleans up the handling of setting invalid ancestor nodes
in the case of a failed block validation.
In this commit, we update the IndexManager interface to use spent txos
rather than the unspent output set for a particualr block. We do this in
order to improve the performance of the current address index which
requires reconstructing the utxo view from the PoV of that new block. In
practice, this is very slow as we need to perform a series of random
reads in order to reconstruct the utxo set. Instead, we can use the set
of SpentTxOut's for that block as this already contains the previous
output scripts which is what all of the current indexers really need.
In this commit, we publicly export the spentTxOut struct and all its
attributes. This is the first in a set of commits to optimize the
existing address index by using the spend journal rather than manually
re-creating the utxoViewPoint each time.
This modifies the utxoset in the database and related UtxoViewpoint to
store and work with unspent transaction outputs on a per-output basis
instead of at a transaction level. This was inspired by similar recent
changes in Bitcoin Core.
The primary motivation is to simplify the code, pave the way for a
utxo cache, and generally focus on optimizing runtime performance.
The tradeoff is that this approach does somewhat increase the size of
the serialized utxoset since it means that the transaction hash is
duplicated for each output as a part of the key and some additional
details such as whether the containing transaction is a coinbase and the
block height it was a part of are duplicated in each output.
However, in practice, the size difference isn't all that large, disk
space is relatively cheap, certainly cheaper than memory, and it is much
more important to provide more efficient runtime operation since that is
the ultimate purpose of the daemon.
While performing this conversion, it also simplifies the code to remove
the transaction version information from the utxoset as well as the
spend journal. The logic for only serializing it under certain
circumstances is complicated and it isn't actually used anywhere aside
from the gettxout RPC where it also isn't used by anything important
either. Consequently, this also removes the version field of the
gettxout RPC result.
The utxos in the database are automatically migrated to the new format
with this commit and it is possible to interrupt and resume the
migration process.
Finally, it also updates the tests for the new format and adds a new
function to the tests to convert the old test data to the new format for
convenience. The data has already been converted and updated in the
commit.
An overview of the changes are as follows:
- Remove transaction version from both spent and unspent output entries
- Update utxo serialization format to exclude the version
- Modify the spend journal serialization format
- The old version field is now reserved and always stores zero and
ignores it when reading
- This allows old entries to be used by new code without having to
migrate the entire spend journal
- Remove version field from gettxout RPC result
- Convert UtxoEntry to represent a specific utxo instead of a
transaction with all remaining utxos
- Optimize for memory usage with an eye towards a utxo cache
- Combine details such as whether the txout was contained in a
coinbase, is spent, and is modified into a single packed field of
bit flags
- Align entry fields to eliminate extra padding since ultimately
there will be a lot of these in memory
- Introduce a free list for serializing an outpoint to the database
key format to significantly reduce pressure on the GC
- Update all related functions that previously dealt with transaction
hashes to accept outpoints instead
- Update all callers accordingly
- Only add individually requested outputs from the mempool when
constructing a mempool view
- Modify the spend journal to always store the block height and coinbase
information with every spent txout
- Introduce code to handle fetching the missing information from
another utxo from the same transaction in the event an old style
entry is encountered
- Make use of a database cursor with seek to do this much more
efficiently than testing every possible output
- Always decompress data loaded from the database now that a utxo entry
only consists of a specific output
- Introduce upgrade code to migrate the utxo set to the new format
- Store versions of the utxoset and spend journal buckets
- Allow migration process to be interrupted and resumed
- Update all tests to expect the correct encodings, remove tests that no
longer apply, and add new ones for the new expected behavior
- Convert old tests for the legacy utxo format deserialization code to
test the new function that is used during upgrade
- Update the utxostore test data and add function that was used to
convert it
- Introduce a few new functions on UtxoViewpoint
- AddTxOut for adding an individual txout versus all of them
- addTxOut to handle the common code between the new AddTxOut and
existing AddTxOuts
- RemoveEntry for removing an individual txout
- fetchEntryByHash for fetching any remaining utxo for a given
transaction hash
The block index now tracks the set of dirty block nodes with status
changes that haven't been persisted and flushes the changes to the DB
at the appropriate times.
The bucket contains block headers keyed by the block height encoded as
big-endian concatenated with the block hash. This allows block headers
to be fetched from the DB in height order with a cursor.
These method allows safe concurrent access to reading and modifying
block node statuses. When block statuses get persisted in a later
change, the setter methods can be used to mark block nodes as dirty.
Each node in the block index records some flags about its validation
state. This is just stored in memory for now, but can save effort if
attempting to reconnect a block that failed validation or was
disconnected.
This was only used to test block proposals, which has been changed to
instead use CheckConnectBlockTemplate. The flag complicated the
implementation of some chain processing routines and would be
difficult to implement with headers-first syncing.
This propagates the interrupt channel through to blockchain and the
indexers so that it is possible to interrupt long-running operations
such as catching up indexes.
This refactors the code that locates blocks (inventory discovery) out of
server and into blockchain where it can make use of the new much more
efficient chain view and more easily be tested. As an aside, it really
belongs in blockchain anyways since it's purely dealing with the block
index and best chain.
Since the majority of the network has moved to header-based semantics,
this also provides an additional optimization to allow headers to be
located directly versus needing to first discover the hashes and then
fetch the headers.
The new functions are named LocateBlocks and LocateHeaders. The former
returns a slice of located hashes and the latter returns a slice of
located headers.
Finally, it also updates the RPC server getheaders call and related
plumbing to use the new LocateHeaders function.
A comprehensive suite of tests is provided to ensure both functions
behave correctly for both correct and incorrect block locators.
- Remove inMainChain from block nodes since that can now be efficiently
determined by using the chain view
- Track the best chain via a chain view instead of a single block node
- Use the tip of the best chain view everywhere bestNode was used
- Update chain view tip instead of updating best node
- Change reorg logic to use more efficient chain view fork finding logic
- Change block locator code over to use more efficient chain view logic
- Remove now unused block-index-based block locator code
- Move BlockLocator definition to chain.go
- Move BlockLocatorFromHash and LatestBlockLocator to chain.go
- Update both to use more efficient chain view logic
- Rework IsCheckpointCandidate to use block index and chain view
- Optimize MainChainHasBlock to use chain view instead of hitting db
- Move to chain.go since it no longer involves database I/O
- Removed error return since it can no longer fail
- Optimize BlockHeightByHash to use chain view instead of hitting db
- Move to chain.go since it no longer involves database I/O
- Removed error return since it can no longer fail
- Optimize BlockHashByHeight to use chain view instead of hitting db
- Move to chain.go since it no longer involves database I/O
- Removed error return since it can no longer fail
- Optimize HeightRange to use chain view instead of hitting db
- Move to chain.go since it no longer involves database I/O
- Optimize BlockByHeight to use chain view for main chain check
- Optimize BlockByHash to use chain view for main chain check
This removes the DisableVerify function and related state since nothing
uses it anymore since the command line option was removed. It is a
remnant of initial development.
This modifies the code that determines the most recently known
checkpoint to take advantage of recent changes which make the entire
block index available in memory by only storing a reference to the
specific node in the index that represents the latest known checkpoint.
Previously, the entire block was stored and new checkpoints required
loading it from the database.
This completely removes the threshold state database caching code since
it can very quickly be calculated at startup now that the entire block
index is loaded first.
This reworks the block index code such that it loads all of the headers
in the main chain at startup and constructs the full block index
accordingly.
Since the full index from the current best tip all the way back to the
genesis block is now guaranteed to be in memory, this also removes all
code related to dynamically loading the nodes and updates some of the
logic to take advantage of the fact traversing the block index can
longer potentially fail. There are also more optimizations and
simplifications that can be made in the future as a result of this.
Due to removing all of the extra overhead of tracking the dynamic state,
and ensuring the block node structs are aligned to eliminate extra
padding, the end result of a fully populated block index now takes quite
a bit less memory than the previous dynamically loaded version.
The main downside is that it now takes a while to start whereas it was
nearly instant before, however, it is much better to provide more
efficient runtime operation since that is its ultimate purpose and the
benefits far outweigh this downside.
Some benefits are:
- Since every block node is in memory, the recent code which
reconstructs headers from block nodes means that all headers can
always be served from memory which is important since the majority of
the network has moved to header-based semantics
- Several of the error paths can be removed since they are no longer
necessary
- It is no longer expensive to calculate CSV sequence locks or median
times of blocks way in the past
- It will be possible to create much more efficient iteration and
simplified views of the overall index
- The entire threshold state database cache can be removed since it is
cheap to construct it from the full block index as needed
An overview of the logic changes are as follows:
- Move AncestorNode from blockIndex to blockNode and greatly simplify
since it no longer has to deal with the possibility of dynamically
loading nodes and related failures
- Rename RelativeNode to RelativeAncestor, move to blockNode, and
redefine in terms of AncestorNode
- Move CalcPastMedianTime from blockIndex to blockNode and remove no
longer necessary test for nil
- Change calcSequenceLock to use Ancestor instead of RelativeAncestor
since it reads more clearly
This takes care of a few minor nits on the recently merged subscribe
code. In particular:
- Avoid extra unnecessary allocation on notifications slice
- Avoid defer overhead on notification mutex in sendNotifications
- Make test function comment start with the name of the function per
Effective Go guidelines
- Use constant for number of subscribers in test
- Don't exceed column 80 in test print
The BlockChain struct emits notifications for various events, but
it is only possible to register one listener. This changes the
interface and implementations to allow multiple listeners.
This replaces the ErrDoubleSpend and ErrMissingTx error codes with a
single error code named ErrMissingTxOut and updates the relevant errors
and expected test results accordingly.
Once upon a time, the code relied on a transaction index, so it was able
to definitively differentiate between a transaction output that
legitimately did not exist and one that had already been spent.
However, since the code now uses a pruned utxoset, it is no longer
possible to reliably differentiate since once all outputs of a
transaction are spent, it is removed from the utxoset completely.
Consequently, a missing transaction could be either because the
transaction never existed or because it is fully spent.
This commit implements the new block validation rules as defined by
BIP0141. The new rules include the constraints that if a block has
transactions with witness data, then there MUST be a commitment within
the conies transaction to the root of a new merkle tree which commits
to the wtxid of all transactions. Additionally, rather than limiting
the size of a block by size in bytes, blocks are now limited by their
total weight unit. Similarly, a newly define “sig op cost” is now used
to limit the signature validation cost of transactions found within
blocks.
This commit modifies the existing block validation logic to examine the
current version bits state of the CSV soft-fork, enforcing the new
validation rules (BIPs 68, 112, and 113) accordingly based on the
current `ThesholdState`.
This modifies the blockNode and BestState structs in the blockchain
package to store hashes directly instead of pointers to them and updates
callers to deal with the API change in the exported BestState struct.
In general, the preferred approach for hashes moving forward is to store
hash values in complex data structures, particularly those that will be
used for cache entries, and accept pointers to hashes in arguments to
functions.
Some of the reasoning behind making this change is:
- It is generally preferred to avoid storing pointers to data in cache
objects since doing so can easily lead to storing interior pointers
into other structs that then can't be GC'd
- Keeping the hash values directly in the block node provides better
cache locality
This refactors the block index logic into a separate struct and
introduces an individual lock for it so it can be queried independent of
the chain lock.
This modifies the block node structure to include a couple of extra
fields needed to be able to reconstruct the block header from a node,
and exposes a new function from chain to fetch the block headers which
takes advantage of the new functionality to reconstruct the headers from
memory when possible. Finally, it updates both the p2p and RPC servers
to make use of the new function.
This is useful since many of the block header fields need to be kept in
order to form the block index anyways and storing the extra fields means
the database does not have to be consulted when headers are requested if
the associated node is still in memory.
The following timings show representative performance gains as measured
from one system:
new: Time to fetch 100000 headers: 59ms
old: Time to fetch 100000 headers: 4783ms
This removes the CalcPastMedianTime since it is now exposed much more
efficiently via the MedianTime field of the BestState snapshot returned
from the BestSnapshot function.
This modifies the block nodes used in the blockchain package for keeping
track of the block index to use int64 for the timestamps instead of
time.Time.
This is being done because a time.Time takes 24 bytes while an int64
only takes 8 and the plan is to eventually move the entire block index
into memory instead of the current dynamically-loaded version, so
cutting the number of bytes used for the timestamp by a third is highly
desirable.
Also, the consensus code requires working with unix-style timestamps
anyways, so switching over to them in the block node does not seem
unreasonable.
Finally, this does not go so far as to change all of the time.Time
references, particularly those that are in the public API, so it is
purely an internal change.
This modifies the blockchain code to store all blocks that have passed
proof-of-work and contextual validity tests in the database even if they
may ultimately fail to connect.
This eliminates the need to store those blocks in memory, allows them to
be available as orphans later even if they were never part of the main
chain, and helps pave the way toward being able to separate the download
logic from the connection logic.
This contains a bit of cleanup and additional logic to improve the
recently-added ability to specify additional checkpoints via the
--addcheckpoint option.
In particular:
- Improve error messages in the checkpoint parsing
- Correct the mergeCheckpoints function to weed out duplicate height
checkpoints while using the most-recently provided one as described by
its comment
- Add an assertion to blockchain.New that the provided checkpoints are
sorted as required
- Keep comments to 80 columns and use two spaces after periods in them to
be consistent with the rest of the code base
- Make the entry in doc.go match the actual btcd -h output
Now that all softforking is done via BIP0009 versionbits, replace the
old isMajorityVersion deployment mechanism with hard coded historical
block heights at which they became active.
Since the activation heights vary per network, this adds new parameters
to the chaincfg.Params struct for them and sets the correct heights at
which each softfork became active on each chain.
It should be noted that this is a technically hard fork since the
behavior of alternate chain history is different with these hard-coded
activation heights as opposed to the old isMajorityVersion code. In
particular, an alternate chain history could activate one of the soft
forks earlier than these hard-coded heights which means the old code
would reject blocks which violate the new soft fork rules whereas this
new code would not.
However, all of the soft forks this refers to were activated so far in
the chain history there is there is no way a reorg that long could
happen and checkpoints reject alternate chains before the most recent
checkpoint anyways. Furthermore, the same change was made in Bitcoin
Core so this needs to be changed to be consistent anyways.
