dfcef536d0 blockstorage: do not flush block to disk if it is already there (Matthew Zipkin)
Pull request description:
Closes https://github.com/bitcoin/bitcoin/issues/2039
When reindexing from flat-file block storage there is no need to write anything back to disk, since the block data is already there. This PR skips flushing to disk those blocks that already have a known position in the datastore. Skipping this means that users can write-protect the `blk` files on disk which may be useful for security or even safely sharing that data between multiple bitcoind instances.
`FindBlockPos()` may also flush the undo data file, but again this is skipped if the corresponding block position is known, like during the initial stage of a reindex when block data is being indexed. Once the block index is complete the validation mechanism will call `ConnectBlock()` which will save undo data at that time.
The call stack looks like this:
```
init()
ThreadImport() <-- process fReindex flag
LoadExternalBlockFile()
AcceptBlock()
SaveBlockToDisk()
FindBlockPos()
FlushBlockFile() <-- unnecessary if block is already on disk
```
A larger refactor of this part of the code was started by mzumsande here: https://github.com/mzumsande/bitcoin/tree/202207_refactor_findblockpos including this fix, reviewers can let me know if the changes should be combined.
ACKs for top commit:
sipa:
utACK dfcef536d0
mzumsande:
re-ACK dfcef536d0
achow101:
ACK dfcef536d0
furszy:
Rebase diff ACK dfcef53.
Tree-SHA512: 385c5ac1288b325135398d0ddd3ab788fa98cc0ca19bd2474c74039f2ce70d5088c1d1c9d4dd10aefcbd4c757767ec5805d07ba8cee9289a66f96e6f9eaa5279
ReadRawBlockFromDisk assumes a non-null pos that has an nPos >= 8.
This simple check makes the function safer to call in the future,
so callers don't need to worry about causing UB if the pos is null.
This fixes the log output when -logsourcelocations is used.
Also, instead of 'ERROR:', the log will now say '[error]', like other
errors logged with LogError.
-BEGIN VERIFY SCRIPT-
sed -i --regexp-extended 's! error\("([^"]+)"! LogError("\1\\n"!g' $( git grep -l ' error(' ./src/ )
-END VERIFY SCRIPT-
This is needed for the next commit.
-BEGIN VERIFY SCRIPT-
# Separate sed invocations to replace one-line, and two-line error(...) calls
sed -i --regexp-extended 's!( +)return (error\(.*\);)!\1\2\n\1return false;!g' $( git grep -l 'return error(' )
sed -i --null-data --regexp-extended 's!( +)return (error\([^\n]*\n[^\n]*\);)!\1\2\n\1return false;!g' $( git grep -l 'return error(' )
-END VERIFY SCRIPT-
d298ff8b62 During IBD, prune as much as possible until we get close to where we will eventually keep blocks (Luke Dashjr)
Pull request description:
This should reduce pruning flushes even more, speeding up IBD with pruning on systems that have a sufficient dbcache.
Assumes 1 MB per block between tip and best header chain. Simply adds this to the buffer pruning is trying to leave available, which results in pruning almost everything up until we get close to where we need to be keeping blocks.
ACKs for top commit:
andrewtoth:
ACK d298ff8b62
fjahr:
utACK d298ff8b62
achow101:
ACK d298ff8b62
Tree-SHA512: 2a482376bfb177e2ba7c2f0bb0b58b02efdb38b34755a18d1fc3e869df5959c85b6f1009e1386fa8b89c4f90d520383e36bd3e21dec221042315134efb1a455b
d27b9a2248 test: fix feature_init.py file perturbation (Martin Zumsande)
ad66ca1e47 init: abort loading of blockindex in case of missing height. (Martin Zumsande)
Pull request description:
When the block index database is non-contiguous due to file corruption (i.e. it contains indexes of height `x-1` and `x+1`, but not `x`), bitcoind can currently crash with an assert in `BuildSkip()` / `GetAncestor()` during `BlockManager::LoadBlockIndex()`:
```
bitcoind: chain.cpp:112: const CBlockIndex* CBlockIndex::GetAncestor(int) const: Assertion `pindexWalk->pprev' failed.
```
This PR changes it such that we instead return an `InitError` to the user.
I stumbled upon this because I noticed that the file perturbation in `feature_init.py` wasn't working as intended, which is fixed in the second commit:
* Opening the file twice in one `with` statement would lead to `tf_read` being empty, so the test wouldn't perturb anything but replace the file with a new one. Fixed by first opening for read, then for write.
* We need to restore the previous state after perturbations, so that only the current perturbation is active and not a mix of the current and previous ones.
* I also added `checkblocks=200` to the startup parameters so that corruption in earlier blocks of `blk00000.dat` is detected during init verification and not ignored.
After fixing `feature_init.py` like that I'd run into the `assert` mentioned above (so running the testfix from the second commit without the first one is a way to reproduce it).
ACKs for top commit:
achow101:
ACK d27b9a2248
furszy:
Code ACK d27b9a224
fjahr:
Code review ACK d27b9a2248
Tree-SHA512: 2e54da6030c5813c86bd58f816401e090bb43c5b834764a5e3c0e55dbfe09e423f88042cab823db3742088204b274d4ad2abf58a3832a4b18328b11a30bf7094
When using an assumedvalid (snapshot) chainstate along with a background
chainstate, we are syncing two very different regions of the chain
simultaneously. If we use the same blockfile space for both of these
syncs, wildly different height blocks will be stored alongside one
another, making pruning ineffective.
This change implements a separate blockfile cursor for the assumedvalid
chainstate when one is in use.
Use the expected AssumeutxoData in order to bootstrap nChainTx values
for assumedvalid blockindex entries in the snapshot chainstate. This
is necessary because nChainTx is normally built up from nTx values,
which are populated using blockdata which the snapshot chainstate
does not yet have.
Introduces ChainstateManager::GetPruneRange().
The prune budget is split evenly between the number of chainstates,
however the prune budget may be exceeded if the resulting shares are
beneath `MIN_DISK_SPACE_FOR_BLOCK_FILES`.
d8041d4e04 blockstorage: Return on fatal undo file flush error (TheCharlatan)
f0207e0030 blockstorage: Return on fatal block file flush error (TheCharlatan)
5671c15f45 blockstorage: Mark FindBlockPos as nodiscard (TheCharlatan)
Pull request description:
The goal of this PR is to establish that fatal blockstorage flush errors should be treated as errors at their call site.
Prior to this patch `FlushBlockFile` may have failed without returning in `Chainstate::FlushStateToDisk`, leading to a potential write from `WriteBlockIndexDB` that may refer to a block that is not fully flushed to disk yet. By returning if either `FlushUndoFile` or `FlushBlockFile` fail, we ensure that no further write operations take place that may lead to an inconsistent database when crashing. Add `[[nodiscard]]` annotations to them such that they are not ignored in future.
Functions that call either `FlushUndoFile` or `FlushBlockFile`, need to handle these extra abort cases properly. Since `Chainstate::FlushStateToDisk` already produces an abort error in case of `WriteBlockIndexDB` failing, no extra logic for functions calling `Chainstate::FlushStateToDisk` is required.
Besides `Chainstate::FlushStateToDisk`, `FlushBlockFile` is also called by `FindBlockPos`, while `FlushUndoFile` is only called by `FlushBlockFile` and `WriteUndoDataForBlock`. For both these cases, the flush error is not further bubbled up. Instead, the error is logged and a comment is provided why bubbling up an error would be less desirable in these cases.
---
This pull request is part of a larger effort towards improving the shutdown / abort / fatal error handling in validation code. It is a first step towards implementing proper fatal error return type enforcement similar as proposed by theuni in this pull request [comment](https://github.com/bitcoin/bitcoin/pull/27711#issuecomment-1563561502). For ease of review of these critical changes, a first step would be checking that `AbortNode` leads to early and error-conveying returns at its call site. Further work for enforcing returns when `AbortNode` is called is done in https://github.com/bitcoin/bitcoin/pull/27862.
ACKs for top commit:
stickies-v:
re-ACK d8041d4
ryanofsky:
Code review ACK d8041d4e04
Tree-SHA512: 47ade9b873b15e567c8f60ca538d5a0daf32163e1031be3212a3a45eb492b866664b225f2787c9e40f3e0c089140157d8fd1039abc00c7bdfeec1b52ecd7e219
This was only explicitly used in the tests, where it can be replaced by
wrapping the original raw file pointer into a CAutoFile on creation and
then calling CAutoFile::fclose().
Also, it was used in LoadExternalBlockFile(), where it can also be
replaced by the (implicit call to the) CAutoFile destructor after
wrapping the original raw file pointer in a CAutoFile.
d506765199 [refactor] Remove compat.h from kernel headers (TheCharlatan)
36193af47c [refactor] Remove netaddress.h from kernel headers (TheCharlatan)
2b08c55f01 [refactor] Add CChainParams member to CConnman (TheCharlatan)
f0d1d8b35c [refactor] Add missing includes for next commit (TheCharlatan)
534b314a74 kernel: Move MessageStartChars to its own file (TheCharlatan)
9be330b654 [refactor] Define MessageStartChars as std::array (TheCharlatan)
37e2b01113 [refactor] Allow std::array<std::byte, N> in serialize.h (MarcoFalke)
Pull request description:
This removes the non-consensus critical `protocol.h` and `netaddress.h` headers from the kernel headers. With this patch, they are no longer required to include in order to use the libbitcoinkernel library. This also allows for the removal of the `compat.h` header from the kernel headers.
As an added future benefit it also reduces the number of of kernel headers that include the platform specific `bitcoin-config.h`.
For those interested, the currently required kernel headers can be inspected visually with the [sourcetrail](https://github.com/CoatiSoftware/Sourcetrail) tool by looking at the required includes of `bitcoin-chainstate.cpp`.
---
This is part of the [libbitcoinkernel project](https://github.com/bitcoin/bitcoin/issues/27587), namely its stage 1 step 3: Decouple most non-consensus headers from libbitcoinkernel.
ACKs for top commit:
stickies-v:
re-ACK d506765
hebasto:
ACK d506765199.
ajtowns:
utACK d506765199
MarcoFalke:
lgtm ACK d506765199🍛
Tree-SHA512: 6f90ea510a302c2927e84d16900e89997c39b8ff3ce9d4effeb8a134bd29cc52bd9e81e51aaa11f7496bad00025b78a58b88c5a9e0bb3f4ebbe9a76309215fb7
The protocol.h file contains many non-consensus related definitions and
should thus not be part of the libbitcoinkernel. This commit makes
protocol.h no longer a required include for users of the
libbitcoinkernel.
This commit is part of the libbitcoinkernel project, namely its stage 1
step 3: Decouple most non-consensus headers from libbitcoinkernel.
Co-Authored-By: Cory Fields <cory-nospam-@coryfields.com>
fae405556d scripted-diff: Rename CBlockTreeDB -> BlockTreeDB (MarcoFalke)
faf63039cc Fixup style of moved code (MarcoFalke)
fa65111b99 move-only: Move CBlockTreeDB to node/blockstorage (MarcoFalke)
fa8685597e index: Drop legacy -txindex check (MarcoFalke)
fa69148a0a scripted-diff: Use blocks_path where possible (MarcoFalke)
Pull request description:
The only reason for the check was to print a warning about an increase in storage use. Now that 22.x is EOL and everyone should have migrated (or decided to not care about storage use), remove the check.
Also, a move-only commit is included. (Rebased from https://github.com/bitcoin/bitcoin/pull/22242)
ACKs for top commit:
TheCharlatan:
ACK fae405556d, though I lack historical context to really judge the second commit fa8685597e.
stickies-v:
ACK fae405556d
Tree-SHA512: 9da8f48767ae52d8e8e21c09a40c949cc0838794f1856cc5f58a91acd3f00a3bca818c8082242b3fdc9ca5badb09059570bb3870850d3807b75a8e23b5222da1
By returning an error code if either `FlushUndoFile` or `FlushBlockFile`
fail, the caller now has to explicitly handle block undo file flushing
errors. Before this change such errors were non-explicitly ignored
without a clear rationale.
Besides the call to `FlushUndoFile` in `FlushBlockFile`, ignore its
return code at its call site in `WriteUndoDataForBlock`. There, a failed
flush of the undo data should not be indicative of a failed write.
Add [[nodiscard]] annotations to `FlushUndoFile` such that its return
value is not just ignored in the future.
By returning an error code if `FlushBlockFile` fails, the caller now has
to explicitly handle block file flushing errors. Before this change
such errors were non-explicitly ignored without a clear rationale.
Prior to this patch `FlushBlockFile` may have failed silently in
`Chainstate::FlushStateToDisk`. Improve this with a log line. Also add a
TODO comment to flesh out whether returning early in the case of an
error is appropriate or not. Returning early might be appropriate to
prohibit `WriteBlockIndexDB` from writing a block index entry that does
not refer to a fully flushed block.
Besides `Chainstate::FlushStateToDisk`, `FlushBlockFile` is also called
by `FindBlockPos`. Don't change the abort behavior there, since we don't
want to fail the function if the flushing of already written blocks
fails. Instead, just document it.
fa69e3a95c Remove unused MessageStartChars parameters from BlockManager methods (MarcoFalke)
Pull request description:
Seems odd to expose these for mocking, when it is not needed.
Fix this by removing the the unused parameters and use the already existing member field instead.
ACKs for top commit:
Empact:
utACK fa69e3a95c
dergoegge:
utACK fa69e3a95c
Tree-SHA512: 7814e9560abba8d9c0926bcffc70f92e502d22f543af43671248f6fcd1433f35238553c0f05123fde6d8e0f80261af0ab0500927548115153bd68d57fe2da746
The block index (CBlockTreeDB) is required to write and read blocks, so
move it to blockstorage. This allows to drop the txdb.h include from
`node/blockstorage.h`.
Can be reviewed with:
--color-moved=dimmed-zebra --color-moved-ws=ignore-all-space
Separate the notion of which blocks are stored on disk, and what data is in our
block index, from what tip a chainstate might be able to get to. We can use
chainstate-agnostic data to determine when to store a block on disk (primarily,
an anti-DoS set of criteria) and let the chainstates figure out for themselves
when a block is of interest for being a candidate tip.
Note: some of the invariants in CheckBlockIndex are modified, but more work is
needed (ie to move CheckBlockIndex to ChainstateManager, as most of what
CheckBlockIndex is doing is checking the consistency of the block index, which
is outside of Chainstate).
If a height is missing we are facing a non-contiguous block index db, and could previously
hit an assert in GetAncestor() called from BuildSkip() instead of returning an error.
When writing a new block to disk, if we have filled up the current block file,
then we flush and truncate that block file (to free allocated but unused
space) before advancing to the next one. When this happens, we have to
determine whether to also flush and truncate the corresponding undo file.
Undo data is only written when blocks are connected, not when blocks are
received. Thus it's possible that the corresponding undo file already has all
the data it will ever have, and we should flush/truncate it as we advance
files; or it's possible that there is more data we expect to write, and should
therefore defer flush/truncation until undo data is later written.
Prior to this commit, we made the determination of whether the undo file was
full of all requisite data by comparing against the chain tip. This patch
replaces that dependence on validation data structures by instead just tracking
the highest height of any block written in the undo file as we go.
This has the benefit of moving the StartShutdown call out of the
blockstorage file and thus out of the kernel's responsibility. The user
can now decide if he wants to start shutdown / interrupt after a block
import or not.
At present, during init, we traverse the chain (once per index)
to confirm that all necessary blocks to sync each index up to
the current tip are present.
To make the process more efficient, we can fetch the oldest block
from the indexers and perform the chain data existence check from
that point only once.
This also moves the pruning violation check to the end of the
'loadinit' thread, which is where the reindex, block loading and
chain activation processes happen.
Making the node's startup process faster, allowing us to remove
the global g_indexes_ready_to_sync flag, and enabling the
execution of the pruning violation verification even when the
reindex or reindex-chainstate flags are enabled (which has being
skipped so far).
By generalizing 'GetFirstStoredBlock' and implementing
'CheckBlockDataAvailability' we can dedup code and
avoid repeating work when multiple indexes are enabled.
E.g. get the oldest block across all indexes and
perform the pruning violation check from that point
up to the tip only once (this feature is being introduced
in a follow-up commit).
This commit shouldn't change behavior in any way.
Co-authored-by: Ryan Ofsky <ryan@ofsky.org>
And transfer the responsibility of verifying whether 'start_block'
has data or not to the caller.
This is because the 'GetFirstStoredBlock' function responsibility
is to return the first block containing data. And the current
implementation can return 'start_block' when it has no data!. Which
is misleading at least.
Edge case behavior change:
Previously, if the block tip lacked data but all preceding blocks
contained data, there was no prune violation. And now, such
scenario will result in a prune violation.
The thread does not only load blocks, it loads the mempool and,
in a future commit, will start the indexes as well.
Also, renamed the 'ThreadImport' function to 'ImportBlocks'
And the 'm_load_block' class member to 'm_thread_load'.
-BEGIN VERIFY SCRIPT-
sed -i "s/ThreadImport/ImportBlocks/g" $(git grep -l ThreadImport -- ':!/doc/')
sed -i "s/loadblk/initload/g" $(git grep -l loadblk -- ':!/doc/release-notes/')
sed -i "s/m_load_block/m_thread_load/g" $(git grep -l m_load_block)
-END VERIFY SCRIPT-
The mempool load can take a while, and it is not
needed for the indexes' synchronization.
Also, having the mempool load function call
inside 'blockstorage.cpp' wasn't structurally
correct.
FatalError replaces what previously was the AbortNode function in
shutdown.cpp.
This commit is part of the libbitcoinkernel project and further removes
the shutdown's and, more generally, the kernel library's dependency on
interface_ui with a kernel notification method. By removing interface_ui
from the kernel library, its dependency on boost is reduced to just
boost::multi_index. At the same time it also takes a step towards
de-globalising the interrupt infrastructure.
Co-authored-by: Russell Yanofsky <russ@yanofsky.org>
Co-authored-by: TheCharlatan <seb.kung@gmail.com>
This is done in addition with the following commit. Both have the goal
of getting rid of direct calls to AbortNode from kernel code. This extra
flushError method is added to notify specifically about errors that
arrise when flushing (syncing) block data to disk. Unlike other
instances, the current calls to AbortNode in the blockstorage flush
functions do not report an error to their callers.
This commit is part of the libbitcoinkernel project and further removes
the shutdown's and, more generally, the kernel library's dependency on
interface_ui with a kernel notification method. By removing interface_ui
from the kernel library, its dependency on boost is reduced to just
boost::multi_index. At the same time it also takes a step towards
de-globalising the interrupt infrastructure.
