2b373fe49d docs: update assumeutxo.md (James O'Beirne)
87a1108c81 test: add snapshot completion unittests (James O'Beirne)
d70919a88f refactor: make MempoolMutex() public (James O'Beirne)
7300ced9de log: add LoadBlockIndex() message for assumedvalid blocks (James O'Beirne)
d96c59cc5c validation: add ChainMan logic for completing UTXO snapshot validation (James O'Beirne)
f2a4f3376f move-only-ish: init: factor out chainstate initialization (James O'Beirne)
637a90b973 add Chainstate::HasCoinsViews() (James O'Beirne)
c29f26b47b validation: add CChainState::m_disabled and ChainMan::isUsable (James O'Beirne)
5ee22cdafd add ChainstateManager.GetSnapshot{BaseHeight,BaseBlock}() (James O'Beirne)
Pull request description:
This is part of the [assumeutxo project](https://github.com/bitcoin/bitcoin/projects/11) (parent PR: https://github.com/bitcoin/bitcoin/pull/15606)
Part two of replacing https://github.com/bitcoin/bitcoin/pull/24232.
---
When a user activates a snapshot, the serialized UTXO set data is used to create an "assumed-valid" chainstate, which becomes active in an attempt to get the node to network tip as quickly as possible. Simultaneously in the background, the already-existing chainstate continues "conventional" IBD to both accumulate full block data and serve as a belt-and-suspenders to validate the assumed-valid chainstate.
Once the background chainstate's tip reaches the base block of the snapshot used, we set `m_stop_use` on that chainstate and immediately take the hash of its UTXO set; we verify that this matches the assumeutxo value in the source code. Note that while we ultimately want to remove this background chainstate, we don't do so until the following initialization process, when we again check the UTXO set hash of the background chainstate, and if it continues to match, we remove the (now unnecessary) background chainstate, and move the (previously) assumed-valid chainstate into its place. We then reinitialize the chainstate in the normal way.
As noted in previous comments, we could do the filesystem operations "inline" immediately when the background validation completes, but that's basically just an optimization that saves disk space until the next restart. It didn't strike me as worth the risk of moving chainstate data around on disk during runtime of the node, though maybe my concerns are overblown.
The final result of this completion process is a fully-validated chain, where the only evidence that the user synced using assumeutxo is the existence of a `base_blockhash` file in the `chainstate` directory.
ACKs for top commit:
achow101:
ACK 2b373fe49d
Tree-SHA512: a204e1d6e6932dd83c799af3606b01a9faf893f04e9ee1a36d63f2f1ccfa9118bdc1c107d86976aa0312814267e6a42074bf3e2bf1dead4b2513efc6d955e13d
Trigger completion when a background validation chainstate reaches the
same height as a UTXO snapshot, and handle cleaning up the chainstate
on subsequent startup.
0af16e7134 doc: add release note for #25574 (Martin Zumsande)
57ef2a4812 validation: report if pruning prevents completion of verification (Martin Zumsande)
0c7785bb25 init, validation: Improve handling if VerifyDB() fails due to insufficient dbcache (Martin Zumsande)
d6f781f1cf validation: return VerifyDBResult::INTERRUPTED if verification was interrupted (Martin Zumsande)
6360b5302d validation: Change return value of VerifyDB to enum type (Martin Zumsande)
Pull request description:
`VerifyDB()` can fail to complete due to insufficient dbcache at the level 3 checks. This PR improves the error handling in this case in the following ways:
- The rpc `-verifychain` now returns false if the check can't be completed due to insufficient cache
- During init, we only log a warning if the default values for `-checkblocks` and `-checklevel` are taken and the check doesn't complete. However, if the user actively specifies one of these args, we return with an InitError if we can't complete the check.
This PR also changes `-verifychain` RPC to return `false` if the verification didn't finish due to missing block data (pruning) or due to being interrupted by the node being shutdown.
Previously, this PR also included a fix for a possible assert during verification - this was done in #27009 (now merged).
ACKs for top commit:
achow101:
ACK 0af16e7134
ryanofsky:
Code review ACK 0af16e7134. Only small suggested changes since the last review, like renaming some of the enum values. I did leave more suggestions, but they are not very important and could be followups
john-moffett:
ACK 0af16e7134
MarcoFalke:
lgtm re-ACK 0af16e7134 🎚
Tree-SHA512: 84b4f767cf9bfbafef362312757c9bf765b41ae3977f4ece840e40c52a2266b1457832df0cdf70440be0aac2168d9b58fc817238630b0b6812f3836ca950bc0e
Moves chainstate initialization into its own function. This is
necessary to later support a more readable way of handling
background-validation chainstate cleanup during init, since the
chainstate initialization functions may need to be repeated after
moving leveldb filesystem content around.
This commit isn't strictly necessary, but the alternative is to (ab)use
the `while` loop in init.cpp with a `continue` on the basis of a
specific ChainstateLoadingError return value from LoadChainstate. Not
only is this harder to read, but it can't be unittested.
The approach here lets us consolidate background-validation cleanup to
LoadChainstate, and therefore exercise it within tests.
This commit is most easily reviewed with
git diff --color-moved=dimmed-zebra
--color-moved-ws=ignore-space-change
The rpc command verifychain now fails if the dbcache was not sufficient
to complete the verification at the specified level and depth.
In the same situation, the VerifyDB check during Init will now fail (and lead to
an early shutdown) if the user has explicitly specified -checkblocks or
-checklevel but the check couldn't be executed because of the limited
cache. If the user didn't change any of the two and is using the defaults, log a warning
but don't prevent the node from starting up.
This does not change behavior. It is in preparation for
special handling of the case where VerifyDB doesn't finish
for various reasons, but doesn't fail.
Use DBParams struct to remove ArgsManager uses from txdb.
To reduce size of this commit, this moves references to gArgs variable out of
txdb.cpp to calling code in chainstate.cpp. But these moves are temporary. The
gArgs references in chainstate.cpp are moved out to calling code in init.cpp in
later commits.
This commit does not change behavior.
This changes the minimum chain work for the bitcoin-chainstate
executable. Previously it was uint256{}, now it is the chain's default
minimum chain work.
Add functionality for activating a snapshot-based chainstate if one is
detected on-disk.
Also cautiously initialize chainstate cache usages so that we don't
somehow blow past our cache allowances during initialization, then
rebalance at the end of init.
Co-authored-by: Russell Yanofsky <russ@yanofsky.org>
Fixes https://github.com/bitcoin/bitcoin/issues/22964
Previously, we used UnloadBlockIndex() in order to reset node.mempool
and node.chainman. However, that has proven to be fragile (see
https://github.com/bitcoin/bitcoin/issues/22964), and requires
UnloadBlockIndex and its callees to be updated manually for each member
that's introduced to the mempool and chainman classes.
In this commit, we stop using the UnloadBlockIndex function and we
simply reconstruct node.mempool and node.chainman.
Since PeerManager needs a valid reference to both node.mempool and
node.chainman, we also move PeerManager's construction via `::make` to
after the chainstate activation sequence is complete.
There are no more callers to UnloadBlockIndex after this commit, so it
and its sole callees can be pruned.
...instead pass in a std::function<int64_t()>
Note that the static_cast is needed (apparently) for the compiler to
know which overloaded GetTime to choose.
...by moving the try/catch out of LoadChainstate
I strongly recommend reviewing with the following git-diff flags:
--color-moved=dimmed_zebra --color-moved-ws=allow-indentation-change
...instead allow the caller to optionally pass in callbacks which are
triggered for certain events.
Behaviour change: The string "Verifying blocks..." was previously
printed for each chainstate in chainman which did not have an
effectively empty coinsview, now it will be printed once unconditionally
before we call VerifyLoadedChain.
