Two defects in a single autoshield round, both of the quiet kind.
The size estimate reserved a flat 2000 bytes for "header + sietch outputs", but
every autoshield tx carries three Sapling OutputDescriptions -- the change note
plus the two Sietch dummies -- and the real fixed cost is ~2937 bytes. Measured
across seven live mainnet coinbase shields: 113.5 bytes per input and 2936.6
+/- 0.8 bytes fixed, of which 3 * 948 = 2844 is the output descriptions. The
estimate was therefore short by ~937 bytes before a single input was counted.
Inputs are charged AUTOSHIELD_CTXIN_DUST_SIZE = 148, which is conservative for
the default P2PK coinbase but exact for P2PKH, so with a P2PKH coinbase a
backlog of 1332..1337 utxos passed the estimate and built a tx over
MAX_TX_SIZE_AFTER_SAPLING. CommitTransaction calls AddToWallet before
AcceptToMemoryPool, so a rejected oversize tx leaves its inputs reading as spent.
z_shieldcoinbase caps a manual shield at SHIELD_COINBASE_DEFAULT_LIMIT = 50
utxos; autoshield dropped that cap and relied on the byte estimate alone.
Restore one -- AUTOSHIELD_MAX_INPUTS = 400 -- so the byte arithmetic is no longer
the only thing between a large backlog and an oversize transaction. The
remainder is shielded on the next round.
Second, the proof build deliberately runs without cs_wallet so wallet RPCs are
not stalled, which leaves a multi-second window in which a concurrent
z_shieldcoinbase or z_sendmany can re-select the same coinbase outputs.
AvailableCoins already honours IsLockedCoin and z_shieldcoinbase already
brackets its selection with LockCoin/UnlockCoin; autoshield made zero LockCoin
calls. Take the locks under cs_wallet at selection time and release them via
RAII, since several early returns sit between selection and commit and a leaked
lock would exclude those coins from every future round.
Verified on an isolated regtest chain with a 540-utxo backlog:
round 1 logged "reached per-round input cap (400)" and committed exactly 400
inputs in a 48351-byte tx (estimate 62300, limit 200000)
round 2 took the remaining 151; backlog drained 540 -> 0
listlockunspent showed 400 coins locked mid-round and 0 afterwards
48351 bytes for 400 inputs implies 2937 bytes of fixed overhead, agreeing
with the mainnet measurement to 14 bytes
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Auto-shielding could silently decline to run with no way to ask why. The
destination it resolves was equally invisible: the only evidence was a LogPrintf
emitted once per round, so an operator wanting to know where their mined coinbase
was going had to grep debug.log.
z_autoshieldstatus reports the enable state, whether a round is in flight, the
next height, interval, fee, minimum utxos, and the resolved destination -- plus
the HD seed provenance in both numeric and readable form, whether the seed is
phrase-recoverable, and a disabled_reason explaining why it is off when it is.
That last field is the point. "autoshield": false on its own does not distinguish
an operator who passed -autoshield=0 from a wallet whose seed provenance is not
known-recoverable, and those need different responses.
Mirrors z_sweepstatus in shape and registration.
Verified on all three branches:
fresh wallet -> autoshield true, origin 1 "created on an empty wallet",
seed_recoverable true, disabled_reason ""
-autoshield=0 -> disabled_reason "disabled by -autoshield=0"
upgraded wallet -> autoshield false, origin 4 "predates provenance recording",
seed_recoverable false, disabled_reason "HD seed origin is
not known-recoverable; back the seed up and pass -autoshield=1"
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
hush-gtest failed to link with "undefined reference to
CHDChain::VERSION_HD_MNEMONIC". The version constants are static const int with
in-class initialisers and no definition anywhere, so any ODR use needs one --
and gtest's EXPECT_*/ASSERT_* macros take their arguments by const reference,
which is exactly that. test_mnemonic_compat.cpp:161 passes VERSION_HD_MNEMONIC
to EXPECT_LT.
dragonxd links either way, because nothing in the daemon binds these to a
reference; only the test target exposed it, and the test target was never built
on the branch that introduced the test.
Define all four rather than only the one that failed: VERSION_HD_BASE,
VERSION_HD_TRANSPARENT and CURRENT_VERSION carry the identical latent fault, and
the next EXPECT_EQ against any of them would hit the same wall. Fixing the test
instead would have hidden the problem rather than removed it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Brings the auto-shield-coinbase feature and the wallet seed work onto the release
line so both are exercised from dev rather than a side branch.
From upstream: auto-shield of matured coinbase into a wallet z-address, plus
fixes to two pre-existing scheduler wedges (consolidation dispatching every block
instead of once per interval, and a failing sweep stranding its running flag).
On top of that:
* the auto-shield destination is chosen by re-deriving m/32'/coin'/i' from the
seed and taking the lowest in-gap account the wallet holds a key for, rather
than the first entry in std::set order. Key metadata is NOT evidence of
provenance -- z_importwallet copies hdKeypath and seedFp verbatim from the
import file -- so a crafted import could otherwise claim account 0 and
capture every shielded reward;
* HD seed/chain persistence is hardened: the chain record is written before the
seed, a corrupt hdchain is loud rather than silently reverting derivation to
the raw entropy, hdchain survives -salvagewallet, and the silent BIP39 ->
random seed fallback is gone;
* seed provenance is recorded, and -autoshield defaults ON only where that
provenance says the seed is recoverable;
* mnemonic wallets now store the EXPANDED 64-byte BIP39 seed with the 32-byte
entropy in a separate display-only record. Derivation reads stored bytes
directly on every binary, so key trees are identical and no CHDChain version
bump or minversion fence is needed -- the format stays readable by earlier
releases;
* new wallets are created from a BIP39 phrase by default;
* the plaintext hdseed record is erased when a wallet is encrypted. It was
previously left behind, and CDB::Rewrite copies surviving records verbatim,
so the unencrypted seed persisted on disk forever.
TWO DEFAULTS CHANGE for new wallets: auto-shielding (where the seed provenance is
known) and BIP39 seed generation. Existing wallets are untouched -- seed
generation is reachable only when a wallet has none, and all three key stores
refuse to replace an existing seed.
Testing state, stated plainly: every commit built clean and the branch was
verified on an isolated chain -- destination selection provably ignores an
imported foreign key, both provenance branches behave, a new-format wallet
reopened by a pre-change binary lists identical addresses, and restoring only the
24 words recovers the wallet. NOT yet tested: any of this against a funded wallet
on mainnet, or a soak of these defaults on a real node.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
dev branched from the release line before 1c3523aac and so lost -Wa,-mbig-obj
from util/build-win.sh. Without it the Windows cross-compile fails at link: large
template/boost-heavy translation units exceed the PE/COFF ~32k-section limit, and
GNU ld emits "dangerous relocation" on .pdata and crashes. v1.1.0 therefore could
not produce a win64 binary at all.
The merge brings only that one file back. util/build-win.sh is the only file both
sides touched, and the two edits are ~20 lines apart, so it auto-merges keeping
both fixes: -Wa,-mbig-obj on the configure line and -DARCH=default on the cmake
line. Verified after merging that dev's own work is intact -- the extended
checkpoint table, the guarded RandomX dedup, the 1.1.0 version bump, and
ARCH=default in the other two build scripts.
No C++ changed; the only delta from the pre-merge dev is the shell script.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
New wallets now get an exportable 24-word phrase instead of a random seed that
no phrase can ever reproduce. Only wallets with no seed yet are affected;
GenerateNewSeed is reachable from one place, under !HaveHDSeed(), and all three
key stores refuse to replace an existing seed.
This is safe to default on now that the storage form is backwards compatible: the
expanded 64-byte BIP39 seed is what gets stored, so a binary predating any of
this reads it and derives the same keys.
Verified on an isolated chain before flipping:
- a new-format wallet reopened with the tagged v1.1.0 binary, which has no
knowledge of the entropy record, listed identical addresses;
- restoring only the 24 words into a fresh datadir recovered every address.
Note this changes what a new wallet is, not what an existing one is: the same
entropy yields a different key tree depending on which side of this commit
created the wallet. Nothing migrates, and nothing needs to.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Mnemonic wallets stored the 32-byte BIP39 entropy as the HD seed and relied on
CHDChain.fMnemonicSeed to tell the deriver to expand it first. That flag is
version-gated in the CHDChain serialisation, so a binary that predates it reads
the record, never consumes the trailing byte, and derives from the raw entropy --
a different key tree, silently, with no error.
Store the expanded 64-byte BIP39 seed instead, with fMnemonic = false, and keep
the entropy in its own display-only record. Derivation then reads the stored
bytes directly on every binary, old or new, so key trees are identical and no
CHDChain version bump or minversion fence is needed. The wallet format stays
readable by earlier releases rather than becoming one-way.
Addresses are unchanged: the previous format expanded the entropy on every read
and fed the same 64 bytes to Master(). This is also the form the tree already
round-trips through -- z_exportwallet dumps the expanded seed, and restoring that
hex via -hdseed installs it with fMnemonic = false.
Write order is load-bearing: seed first, entropy second. A crash between them
leaves a wallet with a seed and no phrase, which is merely inconvenient. The
reverse would leave an entropy record with no seed, and the next start would mint
a different seed while the wallet still held a phrase for the old one.
-usemnemonic still defaults to false; only explicit opt-in and -mnemonic restores
take this path.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Wires the key store secret to the database records: load and store paths on
CWallet, the two ReadKeyValue arms, and the export path.
GetMnemonicPhrase now prefers the entropy record and verifies it before printing:
a phrase is only returned if expanding it reproduces the seed derivation actually
uses. It falls back to the existing fMnemonicSeed path, so wallets that store the
entropy AS the seed keep working unchanged.
IsMnemonicSeed() now means "a phrase is available" rather than "the seed is the
entropy", which is what every caller actually wants.
Still a no-op on every existing wallet: nothing creates an entropy record yet, so
GetMnemonicEntropy returns false and the old code path is taken.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds the record pair for the entropy, mirroring "hdseed"/"chdseed": a plaintext
form, an encrypted form that erases its plaintext counterpart the way
WriteCryptedKey does, and an erase.
Both new types are registered in IsKeyType so -salvagewallet preserves them.
Without that, salvage would silently drop the phrase while keeping the wallet
otherwise intact.
The records are defined but nothing writes them yet.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Additive plumbing for storing a BIP39 entropy alongside the HD seed, mirroring
how the seed itself is handled through both key store layers: a plaintext member
on CBasicKeyStore, an encrypted pair on CCryptoKeyStore, encryption during the
unencrypted-to-encrypted conversion in EncryptKeys with the plaintext cleared,
and decryption on unlock.
RawHDSeed and CKeyingMaterial are the same secure_allocator vector type, so the
entropy passes through EncryptSecret/DecryptSecret with no adaptation.
Nothing calls this yet; there is no behaviour change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
CWalletDB::WriteCryptedHDSeed wrote the "chdseed" record and left "hdseed" in
place, unlike WriteCryptedKey which erases "key"/"wkey" after writing "ckey".
No erase of "hdseed" existed anywhere in src/wallet/.
CDB::Rewrite does not save us: EncryptWallet calls it with pszSkip defaulted, so
it copies every surviving record verbatim into the new file. The result is that a
wallet created unencrypted and later encrypted keeps its raw HD seed in cleartext
on disk permanently, and reloads it into memory on every start.
Add CWalletDB::EraseHDSeed and call it from CWallet::SetCryptedHDSeed after the
encrypted record is written, through the same CWalletDB so it shares
EncryptWallet's transaction. Erase returns true on DB_NOTFOUND, so a wallet that
was never written in plaintext is unaffected.
The erase is deliberately best-effort and only logs on failure. A hard failure
here propagates into CCryptoKeyStore::EncryptKeys, which EncryptWallet turns into
assert(false) with half the keys encrypted in memory; a warning is strictly
better than that.
Note this path is only reachable with -developerencryptwallet, which is
experimental and off by default on this chain, so this is a latent fix rather
than a live one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Autoshield moves mined coinbase into a z-address that only this wallet's HD seed
can re-derive. Defaulting that ON is only defensible where the user can actually
restore that seed.
Two cases fail that test. A seedless legacy wallet has a seed minted onto it
silently at first start, so no backup the user already holds contains it. And a
wallet seeded by an earlier build predates provenance recording, so we cannot
tell which case it was. Both are now classified as not-known-recoverable and
autoshield stays off there until the operator backs the seed up and passes
-autoshield=1. An explicit -autoshield=0/1 still wins in either direction.
Wallets this software created on an empty datadir, or restored from a
user-supplied -mnemonic/-hdseed, keep the ON default: in both cases the user
either has the phrase or supplied the seed themselves.
Verified on real wallets: a wallet carrying no origin record is classified
unknown and logs "autoshield left OFF by default: HD seed origin 4"; a wallet
created by the previous commit logs "autoshield enabled" with no
re-classification, confirming the record persists rather than being recomputed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Autoshield sends mined coinbase to a seed-derived z-address, so the records that
decide how the seed derives keys become fund-safety critical. Three of them were
not treated that way.
InstallHDSeed wrote the seed record before the chain record, non-transactionally.
A crash between the two left a wallet holding a seed with no hdchain: on the next
load hdChain silently reverts to defaults, clearing fMnemonicSeed -- which
switches the derivation input from the expanded BIP39 seed to the raw entropy --
and resetting saplingAccountCounter. Write the chain first; the opposite torn
state is harmless and self-heals, because HaveHDSeed() is then false and init
installs again.
The hdchain record was read as a bare deserialise inside a catch-all with strErr
never set, and it is not a key type, so a corrupt record was downgraded to a
non-critical error and the node booted into the wrong key tree. Report it, track
whether it was read, and refuse to load a wallet that holds a seed but no
readable hdchain. Also preserve hdchain through a keys-only salvage, which would
otherwise drop it and produce exactly the state we now refuse.
GenerateNewSeed silently fell back to a random seed when BIP39 generation failed,
producing a wallet that looks mnemonic-capable but whose words can never be
exported and which no seed phrase can restore. A user who asked for -usemnemonic
now gets that or a hard failure.
Finally, record how the seed came to exist -- created on an empty wallet,
restored from -mnemonic/-hdseed, retrofitted onto a pre-existing seedless wallet,
or predating this record. A retrofitted seed is in no backup the user already
holds, so a feature that moves funds into addresses only that seed can re-derive
must not enable itself there by default. Nothing consumes this yet.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
resolveDestination() took the first spendable address in std::set order, which
orders on the raw Sapling diversifier (zcash/Address.hpp:95-98) -- uncorrelated
with anything the operator can see, and unstable across restarts as addresses
are added.
Worse, GetSaplingPaymentAddresses() also returns z_importkey/z_importwallet
addresses, and CKeyMetadata is NOT evidence of provenance: both hdKeypath and
seedFp are copied verbatim out of the import source (rpcdump.cpp:511-516 ->
wallet.cpp:5440-5441) with no verification. A crafted import can therefore claim
this wallet's seedFp and keypath m/32'/coin'/0' and capture every shielded
mining reward into a key no seed restore can reproduce. Filtering on metadata
would not have caught that.
Derive instead. A bare -mnemonic/-hdseed restore pre-derives exactly
-mnemonicsaplinggap sapling accounts from index 0 with saplingAccountCounter
reset (init.cpp:2349-2355), so the only self-recoverable destinations are the
default addresses of m/32'/<coin>'/i' for i below the gap. Walk that window from
the seed and take the lowest index the wallet holds a spending key for. Deriving
is the only authoritative test and cannot be spoofed.
When nothing in the window is held yet, derive the next account -- but only if it
will land inside the window. GenerateNewSaplingZKey does not derive at
saplingAccountCounter: its do/while skips indices already held
(wallet.cpp:150-157), so a bare counter-below-gap test is unsound. Predict the
lowest free index at or above the counter and post-verify the returned address.
If no free account remains below the gap, refuse the round and leave the coinbase
transparent -- transparent funds are still recoverable through the 1000-key
transparent gap, an unfindable note is not.
Refusing is safe: main_impl turns a false return into a clean skip, and main()
always advances nextAutoShield and clears fAutoShieldRunning, so a refused round
cannot latch the feature off.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Not for release. Built solely to exercise auto-shield-coinbase on an isolated
mining chain, since the feature is a no-op on a non-mining node (the canary
seed has no balance, no z-address, no coinbase, and does not mine).
Merge is clean: util/build-win.sh auto-merged, keeping both our ARCH=default
and their -Wa,-mbig-obj. Checkpoints, verify-once guard and the version bump
all survive.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
CLIENT_VERSION 1000350 -> 1010050. Goes to 1.1.0 rather than 1.0.4 because
1.0.3 is already burned and ambiguous: origin/dragonx's debian changelog
already claims 1.0.3 and the daemon bundled in the ObsidianDragon 2.0.1
installer is labelled v1.0.3-dc45e7d90, so a 1.0.4 would sort above builds
that contain less.
Bumped in configure.ac (authoritative) and in the src/clientversion.h fallback
used when HAVE_CONFIG_H is unset, which the header itself asks to be kept in
sync.
DELIBERATELY NOT BUMPED: SPROUT_VALUE_VERSION, SAPLING_VALUE_VERSION and
SAPLING_VALUE_OPTIONAL_VERSION in chain.h. Those are thresholds marking the
CLIENT_VERSION that INTRODUCED each block-index format, not "the current
version". Raising SAPLING_VALUE_OPTIONAL_VERSION to 1010050 would push every
record written by a v1.0.3 node (nVersion 1000350) into the legacy
raw-CAmount branch of the deserializer and misparse it. The stale comment
naming 1000350 as the current CLIENT_VERSION is updated; the constants stand.
1010050 >= 1000350, so this build still writes and reads the optional format.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Brings onto dev the two fixes that until now existed only on release/1.0.4, so
nothing is stranded on a branch we are not shipping from.
1. GUARDED VERIFY-ONCE (main.cpp)
4e67e687d arms the RandomX dedup unconditionally whenever fCheckPOW is set. But
CheckBlockHeader returns early -- BEFORE reaching its RandomX check -- for a
block whose timestamp is >60s in the future (*futureblockp==1), and CheckBlock
deliberately continues on that path. There, hush_checkPOW is the ONLY RandomX
verification the block gets, so suppressing it leaves the block unverified.
Not a chain-acceptance hole: ConnectBlock rejects futureblock != 0, so such a
block never joins the chain. But it silently weakens DoS banning -- an invalid
future block gets rejected for its timestamp instead of for bad PoW, which is a
regression against the un-deduped behaviour it replaced.
ScopedRandomXSkip now takes an `arm` flag and CheckBlock passes fHeaderChecked,
so the dedup applies only where the header check actually completed and did the
verification. Strictly a tightening: it can only cause MORE verification than
before, never less.
2. NO FAKE GIT IDENTITY (clientversion.cpp)
A hardcoded `#define GIT_ARCHIVE 1` with GIT_COMMIT_ID "a86845f3dc", dated Feb
2018, is reached whenever build.h supplies no BUILD_DESC -- i.e. any build
without git metadata, which is exactly the tarball/CI release case. Such
binaries reported themselves as that Komodo commit regardless of content; a
build here did precisely that before this was found. The archive substitution
placeholders are kept, so a real git-archive export still works; a git-less
build now reports "-unk", which is honest and greppable.
Both syntax-clean. Rebuild and re-validation on EPYC follows; the earlier
validated binary (md5 fe83d70fec5b50c38bf65ea6c733ffa9) predates these.
release/1.0.4 is parked, not deleted -- its commit records why the v1.0.2
lineage cannot ship (block-index format incompatibility with dev-written
chainstate).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The RandomX skip in RandomXValidationRequired() has never fired in production.
It skips verification below the last in-index checkpoint, but the DRAGONX
checkpoint table ended at 2,838,000 while ASSETCHAINS_RANDOMX_VALIDATION is
2,838,976 -- the window was empty by 976 blocks. Every block since the RandomX
activation has been fully verified, at ~65ms per hash on the fastest x86 core
available and ~180ms on a typical user machine.
Extends the table by 388 entries at stride 1000, from 2,839,000 to 3,226,000
(tip - ~5,600, far beyond any reorg this chain has produced; max observed depth
is 3-4). Blocks requiring a RandomX verify drop from 391,447 to ~5,500. The
same extension carries the existing script/zk-proof skip (fScriptChecks,
fExpensiveChecks) over the same range.
Checkpoint data verification, before it went anywhere near source:
- generated on a continuously-synced node
- all 388 hashes identical on 4 other full nodes (388/388 on each)
- reverse-verified hash -> height, all on the active chain
- re-extracted from the patched file and diffed against the verified set
- 2,838 pre-existing entries unchanged; all 3,226 ascending and unique
Trailer fields computed from RPC, not util/checkpoints.pl, which greps a
rotating debug.log and assumes 1440 blk/day (DragonX is 2400).
Also switches all three build scripts from -DARCH=native to -DARCH=default.
RandomX's CMakeLists maps ARCH=native to -march=native, tuning the binary to
the build host: a Zen4 build emitted 746 AVX-512 instructions into
librandomx.a, and every seed reports avx512f=no, so that binary would SIGILL
inside RandomX fleet-wide -- and on any user CPU older than the build machine.
build-win.sh had the same flag, so shipped Windows binaries inherited it.
ARCH=default keeps -maes and per-file -mssse3/-mavx2, so the portable baseline
costs essentially nothing. Note build.sh skips cmake entirely when
src/RandomX/build/ exists, so a stale dir silently preserves the old ARCH.
Validated on an isolated datadir on an EPYC seed, bootstrap -> tip:
- below 3,226,000 (RandomX skipped): 91.3 blk/s (22,823 blocks / 250s)
- at/above 3,226,000 (verified): 4.7 blk/s
~19x at the boundary. The 4.7 blk/s baseline matches a same-day restore on
the old binary, corroborating it independently.
- gettxoutsetinfo at height 3,231,951 BYTE-IDENTICAL to a live node
(hash_serialized 4885c2374ef8b84c648b97d560a57cfcc99bb979142dc89c2a5ccbc90a1f1692,
222,635 txs/txouts, 15,826,352 bytes, total 667909.93689180)
- synced through all 388 new checkpoints with zero rejections
- verifychain 4 (500) and 3 (2000) both true
Binary: v1.0.3-bf3c33c53-dirty, stripped md5 fe83d70fec5b50c38bf65ea6c733ffa9
Checkpoints decay at 2,400 blocks/day; regenerating them belongs on the
release checklist.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The zaddr-sweep op cleared fSweepRunning/nextSweep only on the sweepComplete
success path (inside main_impl), so a cancelled or throwing sweep left
fSweepRunning stuck true. Since fSweepRunning now also gates consolidation and
the default-on autoshield, a persistently failing sweep (e.g. a corrupt-witness
note) would wedge all three background ops for the session.
Move the scheduler bookkeeping into main() so it runs on every terminal state
(success/failure/exception/cancel), guarded by op id. Preserve the intended
"keep draining every block until swept" model: on a successful-but-incomplete
round the flag stays set and nextSweep is not advanced; on completion OR on
failure/exception the flag is released and nextSweep backs off one interval, so
a failing sweep no longer retries every block or wedges the other ops.
Also fix RunSaplingSweep to take cs_wallet itself (was AssertLockHeld, a no-op
in release builds) since ChainTip does not hold it there and the driver mutates
scheduler state + enqueues -- matching RunSaplingConsolidation and
RunAutoShieldCoinbase.
sweepComplete is recorded via a new member; saplingSweepOperationId made public.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Sapling auto-consolidation scheduler never advanced nextConsolidation after
init and never set fConsolidationRunning, so once the tip passed the init
threshold `-consolidation` dispatched a consolidation op every block instead of
once per -consolidationinterval, and every guard that reads fConsolidationRunning
(in RunSaplingSweep, and in the new autoshield driver) was dead.
Mirror the intended scheduler model:
- RunSaplingConsolidation sets fConsolidationRunning=true before dispatch and
self-guards with `if (fConsolidationRunning) return;`.
- The consolidation op advances nextConsolidation = consolidationInterval +
tipHeight and clears fConsolidationRunning on every terminal state
(success/failure/exception/cancel), guarded by op id so only the current op
mutates scheduler state.
- saplingConsolidationOperationId moved to public so the op can read it.
Restores the documented once-per-interval cadence and makes the
sweep/consolidation/autoshield mutual-exclusion guards effective.
Note: the sweep op has the same latent wedge (fSweepRunning/nextSweep are
cleared only on the sweepComplete success path, so a cancelled or throwing
sweep leaves fSweepRunning stuck true) - left for a follow-up; this commit is
the template to port.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On this ac_private=1 chain miners accumulate one transparent coinbase UTXO per
block (the only transparent output the chain permits). It had to be shielded
manually via z_shieldcoinbase, and left unshielded it is the sole persistent
metadata leak on the chain and the source of miner UTXO-fragmentation send
failures.
Add AsyncRPCOperation_autoshieldcoinbase: a periodic, default-on wallet op
driven from CWallet::ChainTip alongside sweep/consolidation, draining matured
coinbase into a wallet-owned Sapling z-address in size-bounded batches.
- Enqueue-only driver (RunAutoShieldCoinbase): takes only cs_wallet in the
notify context; all gathering runs on the async worker under LOCK2(cs_main,
cs_wallet).
- Dedicated op (not a reuse of z_shieldcoinbase) so it never toggles mining.
- Default-ON but conditional: silent no-op on -disablewallet, external
-mineraddress, non-mining, or locked wallets (explicit IsLocked() guard).
- Destination is reuse-then-create; -autoshieldaddress override is
spend-key-validated at init so funds cannot be stranded.
- Sweep-model bookkeeping: advances nextAutoShield and clears the running flag
on every terminal state; an op-id guard stops a stale op clobbering scheduler
state; a self-guard stops cancel/re-enqueue churn.
- Sietch-padded output shape matches manual z_shieldcoinbase txns.
Config: -autoshield (default true), -autoshieldinterval, -autoshieldaddress,
-autoshieldfee (range-validated), -autoshieldminutxos.
Incorporates fixes from an 8-angle code review: CAmount fee type with init-time
range validation, op-id-guarded flag bookkeeping, driver self-guard, and a
tipHeight-consistent NU-activation guard.
Note: the fConsolidationRunning / nextConsolidation consolidation-scheduler
wedge is a pre-existing bug, left for a separate change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Large template/boost-heavy TUs (e.g. asyncrpcoperation.cpp) exceed the standard
PE/COFF ~32k-section limit under mingw-w64, which makes GNU ld emit
"dangerous relocation" on .pdata and crash (SIGSEGV) at link. The bigobj COFF
variant lifts that limit; this is the same flag Bitcoin Core sets for its mingw
host. Fixes the Windows daemon cross-compile.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
build-debian-package.sh warns at startup that lintian is optional, but then
called `lintian -i ...` unconditionally at the end. Under `set -e` that aborted
with exit 127 on hosts without lintian — after the .deb had already been built.
Guard the call with `command -v lintian` so the build exits 0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
libsodium's autogen.sh fetches config.sub/config.guess from git.savannah.gnu.org
gitweb, which is frequently down (currently returns 502). curl saved the HTML
error page over config.sub, so libsodium's configure died with
"cannot run /bin/bash ./build-aux/config.sub" and the whole build failed.
autoreconf -ivf (run earlier in autogen.sh) already installs valid config.sub/
config.guess from the build host, so set DO_NOT_UPDATE_CONFIG_SCRIPTS=1 (the
script's own opt-out) to skip the fragile download. Validated: the full build
now completes and produces working dragonxd/dragonx-cli/dragonx-tx.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Follow-up to the doc rebrand, addressing the previously out-of-scope legacy:
- Currency unit: strCurrencyUnits (chainparams.cpp) and CURRENCY_UNIT
(amount.cpp) "HUSH" -> "DRAGONX". Both are display-only (RPC help + metrics);
no logic comparisons, verified.
- depends mirrors: libsodium/boost/utfcpp fetched from git.hush.is/attachments;
repointed to canonical upstream (GitHub release / archives.boost.io / GitHub
tag) with the existing sha256 hashes verified to match those sources.
- Seeds: nodes_main.txt now lists the five node[1-5].dragonx.is IPs (DNS-resolved)
instead of Hush nodes; regenerated src/chainparamsseeds.h (was compiling Hush
seed IPs as the fixed fallback); generate-seeds.py header now says DragonX;
hush_seed_nodes.txt updated to DragonX seeds.
- Deleted Hush-history / wrong-for-DragonX files: contrib/snapshot/ (block-500000
Hush airdrop, ~10MB), notary_seeds.txt (Hush notaries; DragonX isn't notarized),
and the Hush emission scripts hush_supply, hush_supply_old, hush_halvings,
hush_block_subsidy_per_halving (hardcode Hush's 340000/12.5 economics).
Kept: hush_scanner (engine invoked by dragonx_scanner) and the "The Hush
developers" copyright headers (lineage credit).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The docs/packaging were largely un-rebranded Hush3 content, with several
docs stating facts that are wrong for DragonX. This rewrites them against
the verified DragonX source state.
Corrections (not just branding):
- PoW: RandomX (CPU), not Equihash/ASIC — README, overview.md, randomx.md
- Privacy: private from genesis (ac_private=1, Sapling@height1), not "as of
block 340000" — overview.md, payment-api.md
- Removed the false "coinbase must be shielded" consensus claim
(shield-coinbase.md, payment-api.md); coinbase is directly spendable
- Fixed default fee 0.0001 (was 0.0010000, 10x); stratum port 22769 (was 19031)
- datadir ~/.hush/DRAGONX, DRAGONX.conf, dragonxd/dragonx-cli/dragonx-tx,
git.dragonx.is throughout; branch model dev->dragonx
- Softened the inherited dPoW reorg claim (no live DragonX notary infra)
Packaging: fix build-debian-package.sh + gen-manpages.sh to use the dragonx
binaries/manpages; rename bash-completions to dragonx*; drop hush-arrakis-chain
from the package. Keep /usr/share/hush (hardcoded in the binary for params).
Also: README links/logo, ObsidianDragon + SilentDragonXAndroid wallets,
networking/init/dev-process/contrib/util rebrand, and leftover helper scripts.
Delete legacy duplicates (hushd.* init/service, HUSH3.conf examples,
OLD_WALLETS.md, hsc.md) and rename hush-uri.bat -> dragonx-uri.bat.
Out of scope (noted, not changed): historical changelog/copyright, the Hush
mainnet airdrop snapshot, seed data files, depends/ source mirrors, and the
in-code strCurrencyUnits="HUSH".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
An adversarial re-review found the header-accept RandomX check (b9fdc7981 +
7e9b2c661 header-PoW + d2124a303 defer) to be a persistent source of
consensus-liveness bugs: it derives the RandomX key from the ACTIVE chain
(hush_chainactive), the wrong branch for reorg/side-branch/catch-up headers, so
it repeatedly false-rejected validly-mined headers and DoS(100)-hard-banned
honest peers (IBD-tail catch-up and deep-reorg cases); the defer fix and an
extend-tip fix each addressed one case while leaving/creating others (an
extend-tip variant re-opened an unbounded post-IBD side-branch flood). It only
mitigated a low-harm resource DoS -- forged headers bloat mapBlockIndex memory/
disk but are never SELECTED (nMinimumChainWork) and the full RandomX + target
check still runs at block-connect. Revert to fCheckPOW=0 at header-accept
(original behavior). A comment in AcceptBlockHeader records that any re-attempt
must derive the key from the header's OWN ancestry (pindexPrev->GetAncestor),
never the active chain.
Also hardens two issues the same review found:
- #8 IBD header cap now bounds against the VALIDATED chainActive.Height()
(attacker-hard) instead of pindexBestHeader, which a forward-extending flood
advanced in lockstep, defeating the cap.
- opreturn_burn only emits a change output above the dust threshold; a sub-dust
change made the returned tx non-standard/unrelayable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Audit #8. The HEADERS handler accepted unbounded headers per peer with no
cumulative cap; during IBD (fCheckPOW=0) a peer could flood cost-free PoW-less
headers into mapBlockIndex/leveldb (never selected -- nMinimumChainWork gates
that -- but still memory/disk growth). Add a per-peer nHeadersProcessed counter
in CNodeState; while IsInitialBlockDownload(), if one peer exceeds
2*max(pindexBestHeader height, checkpoint height) + 200000 headers,
Misbehaving(100) and drop it. The cap is ~2x the chain length, so honest sync
never approaches it; inert post-IBD (the RandomX header check handles forged
headers there).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-on to the post-IBD header-PoW verification (b9fdc7981 + 7e9b2c661).
CheckRandomXSolution derives the RandomX key from the block at
keyHeight = ((height-lag)/interval)*interval, looked up on the ACTIVE chain
(hush_chainactive), so that block must be CONNECTED. When a post-IBD node's
block tip lags the header tip by more than ~one RandomX interval -- the normal
IBD tail, or any node catching up -- the key block is not connected yet, so
GetRandomXKey returns empty. The old code returned an error, making
CheckBlockHeader DoS(100)-ban the honest peer that sent a perfectly valid tip
header we simply could not verify yet.
Observed live: a node finishing a mainnet reindex banned the pool box + seeds
and stalled ~2000 blocks short of the tip. Fix: on an empty key, DEFER (return
true) instead of error -- the header is fully RandomX-verified at block-connect,
where the key block is always connected (blocks connect in order,
keyHeight <= height-lag < the connected tip). Flood protection is preserved for
synced nodes (key present -> real check) and bounded during catch-up by the
per-peer IBD header cap + nMinimumChainWork.
Validated on the live 3.14M-block chain: the affected node caught up the full
~2135-block gap to the tip with zero peer bans (was stalled + banned before).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
#10 (HIGH) opreturn_burn selected UTXOs for nAmount+txfee but pushed only the
burn vout and returned - so the entire selected-input surplus was silently paid
as miner fee (e.g. a 500-coin UTXO burning 10 lost ~490). Push a change output
for (inputs - nAmount - txfee). Also widen the int32_t txfee (which truncated
large CAmount fees) to CAmount and MoneyRange-validate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
#11 (HIGH) z_validateaddress locked LOCK2(cs_main, pwalletMain->cs_wallet) with
no availability guard; under -disablewallet pwalletMain is NULL, so the member
deref SIGSEGVs the daemon (execute() only catches std::exception). Use the
null-safe LOCK2 idiom already used by sibling RPCs so validation still works
without a wallet.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
nSPV handlers (gated behind non-default -nspv_msg) read request[1]/vopret[1]
before confirming the peer sent >=2 bytes:
#6 (MEDIUM) NSPV_UTXOS/NSPV_TXIDS evaluated request[1] whenever len<69 (incl
len==1); the 4351d5b73 value-clamp left this lower bound open. The TXIDS/MEMPOOL
else-branch debug prints also read request[1] unconditionally. Add len>=2 guards
/ drop request[1] from the prints.
#7 (LOW) NSPV_MEMPOOL_CCEVALCODE read vopret[1] on a possibly-1-byte vector.
Guard with vopret.size()>=2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
header-pow: AcceptBlockHeader passed the caller's reused *ppindex (and a height
derived from it, ==0 for a new header) to CheckBlockHeader instead of the
header's own local pindex + real height. Post-IBD this made
RandomXValidationRequired(0) false, so CheckRandomXSolution returned true WITHOUT
verifying (and the fRandomXVerified short-circuit could fire on an unverified
header) - silently defeating the header-flood PoW gate from b9fdc7981. Resolve
pindexPrev up-front, pass real height (parent+1) and the local (NULL) pindex so
the post-IBD RandomX check actually runs; IBD stays fast (fCheckPOW=0).
Stability-tested: 303 valid headers accepted across a 4-node RandomX net,
0 false rejects / bans.
#9 (MEDIUM) GETBLOCKS/GETHEADERS deserialized an unbounded CBlockLocator.vHave
(~130k hashes) and scanned it linearly under cs_main with no ban - a
message-thread liveness DoS. Add MAX_LOCATOR_SZ=101 + Misbehaving, matching the
adjacent vInv/headers caps.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Defensive-audit findings (adversarially verified + fleet stability-tested):
#4 (CRITICAL) hush_voutupdate trusted an attacker-decoded OP_RETURN length
(opretlen, up to 65535 via OP_PUSHDATA2) with no check against the real script
length, driving up to ~64KB out-of-bounds reads through hush_stateupdate ->
hush_eventadd_opreturn -> hush_kvupdate (persisted to disk, leaked via kvsearch
RPC, reliable crash on block connect). Reject any opret claiming more bytes than
remain in the script, at the single taint source.
#5 (HIGH) notary-ratification loop did memcpy(pubkeys[numvalid++],..) into a
fixed uint8_t[64][33] with no bound; >64 crafted vouts smashed the stack. Clamp
numvalid < 64.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
nMinimumChainWork was defined in chainparams but never checked, and IsInitialBlockDownload
decided "synced" from tip timestamp/height alone -- so an eclipsed or bootstrapping node
could be fed a cheap low-work fake chain with recent timestamps and trust it. Reset the
stale mainnet floor (0x281b32ff3198a1 was ABOVE the live chain, would have bricked mainnet)
to the real chainwork at height ~3,100,000, and hold a node in IBD until its tip reaches the
floor. Gated to the DRAGONX symbol so ephemeral assetchains from the same binary are not
trapped in IBD; the check can only keep a node in IBD, never force it out (no false-sync risk).
Complements the header-flood fix (b9fdc7981): that stops invalid-PoW headers off the real tip;
this stops valid-but-cheap fake chains from a fake genesis.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Sapling note-selection loop stopped once total_value >= nTotalOut, ignoring
the miner fee, so a wallet with notes covering the amount but not amount+fee
selected too few notes and failed later with a spurious "insufficient funds".
Reserve the fee (default or user-supplied) in the selection target.
Leto eb4fc52273.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
AcceptBlockHeader called CheckBlockHeader with fCheckPOW=0, so a synced node
stored any well-formed PoW-less header off the tip into mapBlockIndex without
bound (memory/disk DoS). nMinimumChainWork is defined but unenforced and would
not stop tip-siblings anyway (they inherit the tip's chain work). Verify PoW at
header-accept time when not in IBD: forged headers now fail RandomX and the peer
is DoS-banned. IBD keeps fCheckPOW=0 for fast header sync; the full-block
RandomX/target check at connect is unchanged, so no valid header is rejected
(not a consensus-rule change).
fCheckPOW=0 call site is Leto (6a30b40415).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
hush_nSPV_fullnode.h: bound the REMOTERPC method strcpy and json memcpy to their
fixed buffers (method[64], json[11000]); add lower-length and memcpy-source bounds
to the UTXOS/TXIDS coinaddr[64] copies and the MEMPOOL handler. These paths
deserialize attacker-controlled request bytes -> stack overflow / OOB read. The
nSPV server is opt-in via -nspv_msg (off by default; DragonX uses lightwalletd).
rpc/blockchain.cpp: getchaintxstats null-checks pwalletMain (crash under -disablewallet).
wallet/rpcwallet.cpp: z_sendmany initializes nFee to the default miners fee (was read
uninitialized when no fee param supplied).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
hush_validate_chain() enters its body when hush_getblockindex(srchash) returns
NULL (via || short-circuit) -- srchash comes from an attacker-controlled
notarization OP_RETURN -- then a debug fprintf dereferenced the NULL pindex.
A block carrying one crafted OP_RETURN tx crashed every synced node on connect,
and crash-looped on restart. Guard the deref: pindex ? GetHeight() : -1.
Introduced by Leto commit 4988ce6f2 ("much debug such wow", 2022).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sprout JoinSplit proofs/sigs/nullifiers/anchors are never verified (the verifier
arg to CheckTransaction is unused), yet vpub_new is counted as transparent
value-in -- a forged all-zero JoinSplit mints arbitrary value from nothing.
Reproduced on an isolated ac_private=1 chain: 500,000 minted into a z-addr,
accepted + mined + verifychain=true.
Reject any non-coinbase tx carrying a JoinSplit in ContextualCheckTransaction
(covers both mempool acceptance and ConnectBlock). DragonX is Sapling-only from
genesis with zero JoinSplits in its history (mainnet supply audit), so this is
inert on all legitimate traffic and never invalidates a historical block.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Belt-and-suspenders inflation/counterfeiting guard on top of the per-tx Sapling
binding signature: ConnectBlock rejects any block whose cumulative Sapling value
pool would go negative (bad-sapling-value-pool-negative) -- a block can never
deshield more value than was ever shielded.
Enforced ONLY when the pool is reliably tracked from genesis (pprev's
nChainSaplingValue is engaged), so it can never false-reject a valid block or
split the chain on nodes that don't track the pool -- those stay dormant.
To make "not reliably tracked" propagate safely, nSaplingValue becomes a
boost::optional<CAmount> (was a plain CAmount). A version-gated dual-read in
CDiskBlockIndex reads records written before SAPLING_VALUE_OPTIONAL_VERSION
(1000350 = v1.0.3) as the legacy raw 8-byte CAmount but DISCARDS the value
(reads boost::none). Records written at >= 1000350 use the optional format and
persist, so from-genesis and reindexed v1.0.3 nodes are durably active across
restarts.
Tested on a 5-node RandomX fleet: old-format DB loads dormant (0 corruption);
from-genesis stays active with correct pool accumulation and 0 false-rejects
across shield/deshield cycles; dormant/active/reindexed nodes converge; a crafted
counterfeit block is rejected (guard fires, no crash); active state persists
across restart (verified at CLIENT_VERSION 1000351 with gate 1000350).
*** MANDATORY UPGRADE STEP (v1.0.3 dev/test nodes) ***
CLIENT_VERSION stays 1000350, and pre-turnstile v1.0.3 builds ALSO stamped
records at 1000350 but in the old plain-8-byte format. Those records now route to
the OPTIONAL read branch and MISPARSE: LoadBlockIndexDB throws and the node
ABORTS on startup (looks like block-DB corruption). Therefore any node that ran an
earlier v1.0.3 (1000350) build MUST have its block data wiped or be -reindexed
before running this build -- do NOT upgrade a 1000350 datadir in place.
Production mainnet (v1.0.2 = CLIENT_VERSION 1000250) is UNAFFECTED: those records
take the legacy branch and read correctly (dormant until reindex). v1.0.3 is
unreleased, so only dev/test datadirs are affected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This fork never ported Bitcoin's mempool size-limiting: CTxMemPool had no TrimToSize/
Expire and LimitMempoolSize was commented out. An earlier commit added a blunt
DynamicMemoryUsage admission cap that bounded memory but bluntly REJECTED new txs when
full -- so a high-fee tx could not push out a low-fee one. This implements proper
fee-ordered eviction using the per-tx feerate index that already exists (mapTx index 1,
CompareTxMemPoolEntryByFee), with no new index and no descendant-tracking port.
- CTxMemPool::TrimToSize(sizelimit, pvNoSpendsRemaining): while DynamicMemoryUsage() is
over the limit, evict the lowest-feerate tx (the tail of the feerate index) and its
in-mempool descendants (recursive remove), re-deriving the tail each iteration.
Terminates (pool strictly shrinks) and cleans every secondary index via remove().
- CTxMemPool::Expire(time): age-based sweep (entry time older than `time`), for
LimitMempoolSize's -mempoolexpiry.
- LimitMempoolSize re-enabled (Expire + TrimToSize) and called from ConnectTip on every
block connect. (No pcoinsTip->Uncache -- CCoinsViewCache has none in this fork; it is
only a UTXO-cache perf hint.)
- AcceptToMemoryPool now ADDS the tx then TrimToSizes: a higher-fee tx displaces
lower-fee ones; if this tx was itself the lowest-feerate (evicted), it is rejected
("mempool full"). Replaces the blunt reject-when-full cap.
- DEFAULT_MEMPOOL_EXPIRY 1 -> 72 hours (age-Expire is now live; 1h was too aggressive).
Known simplification (documented in code): per-tx feerate, not descendant-aggregate
(CPFP) scoring, and no rollingMinimumFeeRate anti-thrash. Adversarially reviewed
(termination, iterator safety, recursive-lock safety, index cleanup all confirmed) and
runtime-tested on the fleet: pool stays bounded under a 1600-tx flood, verifychain ok,
no hang/crash.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 24-word seed restores the same wallet in SilentDragonXLite only on mainnet;
testnet/regtest derive a different HD coin_type (per BIP44), so a phrase does not
round-trip across wallets there. Document that in the -mnemonic help so it is not
mistaken for a bug.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The -rpcport help string in bitcoin-cli.cpp hardcoded hush's 18030; the actual default (BaseParams().RPCPort()) is DragonX's 21769, so this was misleading help text only (the CLI already connects to 21769). Set to 21769 and regenerated doc/man/dragonx-cli.1 from the rebuilt binary. NOTE: a separate hush 18030 leftover remains in src/rpc/net.cpp:357 (getpeerinfo help example address) - daemon RPC help, out of scope here. Staged on 176; not pushed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Legal: correct GPLv3 LICENSE (fixes garbled 'GENERAL GENERAL'), AUTHORS DragonX attribution, COPYING. Packaging: man pages REGENERATED from the 1.0.3 binaries via help2man (dragonxd/dragonx-cli/dragonx-tx.1 -> v1.0.3, correct dates), wired into doc/man/Makefile.am (dist_man1_MANS), orphaned hush*.1 removed. Init/openrc/systemd scripts, Debian packaging (control/changelog/copyright rebranded hush->dragonx + install stubs), example confs taken from origin/compliant-rebrand (c05134e77). REMAINING follow-ups: (1) debian/changelog still tops at 1.0.0 - add a 1.0.3 entry; (2) dragonx-cli --help hardcodes rpcport default 18030 (hush) - fix the HelpMessage string in source then regen. Staged on 176 for review; not pushed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
nMaxDbCache capped the adaptive UTXO/db cache (and a manual -dbcache) at 16 GiB, so the
help's "uses most of free RAM" was false above ~20 GB of RAM. Raise the 64-bit ceiling to
64 GiB. The adaptive controller + its RAM reserve still bound actual usage and shrink under
memory pressure, and small hosts are unaffected -- the ceiling only binds once RAM-minus-
reserve exceeds it. The coins cache grows lazily to the target, so nothing is pre-allocated.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Three small cleanups to the parallel RandomX pre-verify + P2P-window features:
- Call RandomXValidatorShutdown() in Shutdown() to release the ~256MB shared RandomX
verify cache. It was allocated on first use but never freed, leaking at every exit.
Safe here: threadGroup.interrupt_all() (earlier in Shutdown) stops the pre-verify
worker, and the release takes g_rxvMutex so it can't race a mid-flight verify.
- Clarify the -randomxverifythreads help: the pool only helps NETWORK sync, not reindex
(reindex runs with a window of 1, so the pool does nothing there).
- Clamp -maxblocksintransit to the real BLOCK_DOWNLOAD_WINDOW (1024) ceiling instead of a
misleading 4096. Values above the window are a silent no-op (FindNextBlocksToDownload
never fetches beyond pindexLastCommonBlock + BLOCK_DOWNLOAD_WINDOW); log when clamping.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
