Reported: a node at height N+10 on its own short fork, with the peer at N+500,
could never sync. It located the fork point correctly, pooled the branch, and
then refused it forever. At 90s blocks a node is 960 blocks behind after a day
offline, so this is the normal case, not an edge case.
== Why it could never recover ==
The reorg delay was served by LOCAL chain growth alone:
elapsed = tipHeight - observedTip
That only makes sense for a node whose tip is advancing. A node that is behind
has a frozen tip precisely because it is rejecting the branch, so elapsed stays
0 forever while penalty(5) = 42. Both escape hatches also fail: mining out of it
means extending a fork nobody accepts, and the IBD exemption never fires for a
node that is mining, because mining keeps its tip fresh.
Being 490 blocks behind is being behind, not a reorg contest. Refusing to adopt
protects nothing while the node falls a further 40 blocks behind every hour.
The delay is now satisfiable by EITHER side making progress:
localGrowth = tipHeight - observedTip
branchLead = candidateTip - tipHeight (0 if not ahead)
elapsed = max(localGrowth, branchLead)
A branch already extending `penalty` blocks past our tip has demonstrated
exactly what the delay asks for, and every one of those blocks carries PoW we
validated ourselves. Requiring us to independently produce the same amount is
demanding the same proof twice. Only VALIDATED blocks count -- a peer's
advertised height is not evidence and never reaches this code.
It degrades correctly in both directions: a mining node's tip advances, so an
attacker must outpace it by penalty blocks; a node that is only observing
follows the heaviest chain, which is what an observer should do.
== The window has to keep pulling the branch ==
Every forked delivery reset nextReq back to our own tip, so the window
re-requested the same eight heights forever and the pool never grew past the
window size -- branchLead could not rise even in principle. The backward walk
now runs ONCE to establish linkage, then the window marches forward pooling the
branch, retrying adoption per window-full with a final attempt when it drains.
== sync force ==
Operator override that skips the delay for one sync, for a node whose chain is
known to be the wrong one. Threaded explicitly (Chain_ReplaceBranch gains
bypassPenalty, OrphanPool_AttemptAttachForced) rather than through a global, so
nothing a peer sends can reach it. Linkage, work comparison and atomicity still
apply -- it waives only the waiting, and says so loudly in the log.
== Bug found in testing: PoW is branch-relative ==
A block's epoch seed is the last block of the previous epoch ON ITS OWN BRANCH.
Validating a competing branch's block against OUR epoch seed does not merely
fail to resolve: when the chains diverge before the boundary it resolves to the
WRONG seed and rejects a perfectly valid block. Any fork spanning an epoch
boundary was therefore impossible to assemble -- the branch could not grow past
the boundary block, so branchLead stalled one short of it.
The receive path now does self-contained checks only (Block_HasValidStructure:
merkle, transactions, vote, non-empty). Proof of work moved to
Chain_AddBlockLocked, at the point a block joins the chain, where the branch
context is real -- the rollback has put its ancestors in place by then. That is
where the invariant belongs and it removes a duplicate check rather than adding
one. Needs Chain_DagParamsForHeightLocked, because Chain_AddBlockLocked already
holds chainLock for writing and the lock is not recursive.
Consequence worth knowing: the orphan pool can now hold blocks whose work has
not been verified, bounded by MAX_ORPHAN_BLOCKS (512). Each still had to pass
merkle and full transaction/signature validation, and none can reach the chain
unverified.
This bug also affected plain forward sync across block 350000; it was masked
because appending keeps the boundary block in the chain.
== Bug found in testing: stale DAG accepted as current ==
Block_PowHashHeavy matched on epoch index and size but not the seed. A DAG's
content is a function of (seed, size); the epoch index is a label for it. A
reorg is exactly the event that changes the seed while leaving index and size
untouched, so mid-apply the miner's context still held a DAG built from the
PRE-reorg seed, the guard passed, and a valid block was hashed against the wrong
lanes. g_dagSeed now records what each DAG was generated from and both
Block_EnsureAutolykos2Dag and Block_PowHashHeavy compare it.
== Bug found in testing: double free on the failed-apply path ==
SIGABRT in the allocator: free_tiny_botch -> DynArr_destroy -> Block_Destroy ->
Chain_FreeBlockArray -> Chain_ReplaceBranch.
DynArr_push_back stores the struct BY VALUE, so the chain's element and the
caller's block_t share one transactions pointer. Three places free that array
through the chain's copy -- Chain_ClearBlocks, Chain_RollbackToHeightLocked and
Chain_SaveToFile -- and each NULLs only the chain's side, leaving any caller
wrapper dangling. Whether a caller then had to use free() or Block_Destroy() was
a convention carried in comments at every call site plus a consumed-count passed
into Chain_FreeBlockArray. Chain_ReplaceBranch reset that count to 0 after
rolling back a failed apply, which told the cleanup to Block_Destroy exactly the
blocks whose arrays the rollback had just freed.
Rather than fix the count, the aliasing is now safe by construction:
Chain_AddBlockLocked clears the CALLER's transactions pointer immediately after
the push. Since DynArr_destroy(NULL) is a no-op, free(wrapper) and
Block_Destroy(wrapper) become equivalent and both safe regardless of what later
frees the chain's copy. The consumed-count parameter and both counters are gone
-- the thing that could be got wrong no longer exists -- and all call sites are
unified on Block_Destroy.
Placement is deliberate: immediately after the push, not at the end on success.
The ledger pass can fail with the block already in the chain, returning false to
a caller that destroys its wrapper on failure -- OrphanPool_ExtendTip does
exactly that, a third live instance not yet triggered.
Two follow-ons the refactor forced, both improvements anyway: MineAndAppendBlock
read the coinbase for its log line after the add (hoisted above it), and the
success log printed the caller's block rather than the chain's copy.
== Also ==
The deferral line is rate-limited. The maintenance thread retries pooled
branches once a second and elapsed only changes when something moves, so it
printed an identical line every second -- forever, on a node that is not mining.
It now reports each distinct situation once.
== Verification ==
Synthetic fork, node A 5 deep, node B ~60 ahead, EPOCH_LENGTH=8 so the branch
crosses three epoch boundaries:
branchLead climbs 8 -> 16 -> 28 -> 34, crosses penalty(5)=42
Adopted competing branch of 50 block(s) at fork height 20
sync complete: localHeight=70
Chain OK
Repeated under AddressSanitizer: adopted 47 blocks, 0 ASan errors on both nodes.
This matters because the refactor rewrites the exact cleanup path the SIGABRT
came from, and a double free that no longer aborts would otherwise pass silently.
Unit suites pass, including a new assertion "heavy path refuses a DAG built from
a different seed" -- the direct regression for the stale-DAG bug.
Harness note: each node needs its OWN wallet. With a shared one both pay the
same coinbase address, produce identical merkle roots, and at easy difficulty
mine byte-identical blocks -- the fork test silently became a catch-up test.
== Still untested ==
The restore-after-failed-apply path is no longer naturally reachable now that
the two bugs above are fixed, so it needs deliberate corruption to exercise.
Test B (branch only slightly ahead must still DEFER), test C (sync force), and a
TSan pass over the changed paths are outstanding.
Reported symptom: node A (4656 blocks, last 10 mined on its own fork) syncing
from node B (5646 blocks) crawled for minutes while B re-served the whole chain
several times over. A's log was thousands of "Queued orphan BLOCK_DATA" lines
interleaved with "timed out fetching block N, giving up", sliding forward 8
blocks at a time and never reorging.
== Root cause ==
The sync loop had no channel for FETCH_BLOCK replies. It inferred that a
requested block had arrived purely from the chain growing:
if ((uint64_t)Chain_Size(chain) > h) { /* block h was applied */ }
But the reply is handled on the peer's io thread by Node_ParseAndAcceptBlock,
which correctly identifies a block whose prevHash points at the peer's branch
and files it in the orphan pool -- the chain does not grow. So a block that
ARRIVED AND WAS CORRECTLY ORPHANED was indistinguishable from a lost packet.
Every one of the 990 blocks therefore cost MAX_SYNC_RETRIES re-requests plus a
SYNC_REQUEST_TIMEOUT_MS stall before the window slid on and did it again. That
is both the crawl and the repeated serving on the peer.
Neither existing fork probe could rescue it:
* The divergence check sits downstream of a SUCCESSFUL append -- it only runs
after Chain_GetBlockCopy(chain, h) succeeds. When the fork begins at the
very first requested height nothing ever lands, so RequestForkWindow was
never reached from there.
* The no-progress fallback only runs after the entire window range drains,
i.e. after timing out through all 990 blocks, and is capped at 3 rounds.
== Delivery receipts ==
Node_NoteBlockDelivered / Node_TakeBlockDelivery / Node_ResetBlockDeliveries,
backed by a 512-slot ring with its own mutex. Recorded from the BLOCK_DATA
handler only -- BLOCK_DATA is sent solely in reply to a FETCH_BLOCK, so a
receipt means "the peer answered", independently of whether the block could
join our chain. A receipt for a height already present is refreshed in place,
so a retried request cannot leave a stale one for the next window to consume.
Receipts carry a status rather than a bool:
NODE_DELIVERY_APPENDED joined our chain
NODE_DELIVERY_DUPLICATE we already held this exact block -- common ground
NODE_DELIVERY_ORPHANED competing branch; now pooled
NODE_DELIVERY_REJECTED failed validation
DUPLICATE is what lets a backwards fork walk terminate; a plain "appended"
boolean cannot tell "we already have this" from "this is on their branch",
which is precisely the signal the walk needs.
The sync loop now treats any non-APPENDED delivery as the divergence signal and
probes immediately, instead of retrying something retrying cannot fix. Bounded
by MAX_FORK_PROBE_ROUNDS -- without that the window resets to the same height
and re-probes forever -- and the counter resets on real progress so a branch
that forks again further on can still be followed.
== RequestForkWindow is now an actual walk ==
It previously fired REORG_FETCH_DEPTH (128) requests and slept a flat
SYNC_REQUEST_TIMEOUT_MS. Two problems:
* The sleep was a race against the peer's serving rate. At the ~10 blocks/s
observed in testing, a 128-block window cannot land in 5s, so the
OrphanPool_AttemptAttach that followed ran against a half-filled pool and
reported a failure that was not real.
* It always asked for the full depth. On a shallow fork the overwhelming
majority came back as duplicates that were discarded without even entering
the pool -- a wasted full block send each.
It now descends in batches of MAX_PARALLEL_FETCHES, waits on that batch's
receipts rather than guessing, and stops at the first height the peer returns
that we already hold: that block is the fork point and everything below it is
shared. It returns whether common ground was found, so callers no longer
attempt an attach that cannot possibly link.
Measured on the reported scenario: 16 requests instead of 129, locating the
fork point at height 4645 (fork = blocks 4646..4655, exactly the 10 mined).
== IBD_TIP_AGE_BLOCKS 500 -> 20 ==
The reorg penalty is served by LOCAL chain growth, so a node that does not mine
can never serve it -- its tip does not move, and the only way it could grow is
by adopting the branch the penalty is gating. The IBD exemption is that node's
only route back, so at 500 block times it had to sit stalled for ~12.5 hours.
20 block times is ~30 minutes at a 90s target, far beyond normal Poisson block
spacing (a gap that long has probability ~e^-20), so a node genuinely following
the tip will not trip it.
This is what flipped the live test from "initialSync=no" to "initialSync=yes"
and let the reorg be adopted at all. Noted in the constant's comment: the
exemption is all-or-nothing, so lowering it further widens that hole.
== Reorg failures are now legible ==
Chain_ReplaceBranch broke silently for fork-point-beyond-tip, work computation
failure, insufficient work, snapshot failure and candidate copy failure alike,
so "not adopted" could not be diagnosed from a log. Each now says which. The
not-heavier case reports candidate vs incumbent block counts, because the
common cause is a branch that is still arriving -- a partial branch is
genuinely lighter -- and that reads very differently from a peer actually on a
weaker chain.
The sync loop's "Reorg candidate not adopted (lighter branch, or still serving
its reorg penalty)" is replaced by "Reorg not completed on this pass; branch
stays pooled for retry". The old wording was actively misleading: live testing
showed it firing on reorgs that the 1Hz maintenance thread completed a second
later.
== Verification ==
Against a real peer at 5646 blocks, from a real local chain of 4656 with a
10-block fork:
reported receipts only all changes
wall-clock minutes 123s 116s
timed-out fetches 1 per block 0 0
fork-walk requests n/a 129 16
reorg completed never maintenance first pass
final height 4656 5646 5646
fullverify - Chain OK Chain OK
The 116s is not a meaningful speedup over 123s and should not be read as one:
990 blocks at ~10 blocks/s is ~100s, so both are bound by the peer's serving
throughput rather than by anything in the sync loop. The substantive wins are
the 129 -> 16 fork-walk reduction, the reorg completing deterministically
instead of by luck, and logs that no longer report failure on success.
Also exercised: a synthetic partitioned fork (10 vs 60 blocks) confirming zero
timeouts and bounded probe rounds when the branch is correctly refused by the
penalty.
== Known limits, unchanged ==
Chain_ReplaceBranch adopts whatever is pooled at attach time (18 blocks in the
live run); the remainder arrives through normal windowed sync afterwards. That
bounds a single reorg by MAX_ORPHAN_BLOCKS.
Two consensus-parameter bugs, both latent today and both guaranteed to become
live splits later. Neither is a regression; both are cheap to fix now and
expensive to fix after launch.
== 1. The reorg penalty scaled with block time the wrong way ==
FetchScheduler_ComputeReorgPenaltyBlocks had TARGET_BLOCK_TIME in the numerator
and REORG_PENALTY_REF_BLOCK_TIME in the denominator. The penalty is counted in
BLOCKS, so the wall-clock protection it actually buys was
penalty(d) * TARGET_BLOCK_TIME ~ d^2 * T^2 / REF
i.e. quadratic in block time, when the stated intent in constants.h is
wall-clock equivalence to the reference scheme defined at REF_BLOCK_TIME = 150.
At T = 90 the chain got 54*d^2 seconds of protection instead of 150*d^2 --
2.78x weaker than intended -- and any future block-time reduction would have
weakened it further, silently.
Swapping the two makes T cancel out of the wall-clock figure:
penalty(d) = ceil(d^2 * REF / T) -> protection ~ d^2 * REF seconds
Penalty in blocks, depth -> before/after: 4 -> 10/27, 8 -> 39/107,
10 -> 60/167, 50 -> 1500/4167, 100 -> 6000/16667.
Verified by rebuilding at TARGET_BLOCK_TIME = 60: wall-clock protection is now
identical at both block times (2400s, 9600s, 15000s, ...), ratio 1.0000. Under
the old formula 60s blocks would have been 56% weaker.
Still integer-only: the ceiling division and saturation guards are unchanged,
and numeratorScale only moves 90 -> 150 while `raised` stays bounded by
REORG_PENALTY_MAX_DEPTH^2 = 1e6, nowhere near overflow.
This is fork choice, not block validity, so mixed-version nodes converge once
the longer penalty expires and no block ever becomes invalid.
== 2. DAG sizing was a fail-open consensus split waiting on block 350000 ==
Fixed, in rough order of severity:
* FAIL-OPEN PoW. Block_CalculateAutolykos2Hash memset the hash to zero on any
failure, and zero compares below every target -- so a DAG that was missing,
mis-sized or had failed to build made *every* block pass PoW validation
instead of rejecting it. PoW checks now fail closed.
* TWO DIVERGENT IMPLEMENTATIONS of the same rule: CalculateTargetDAGSize in
constants.h (mining) and ComputeEpochDagBytesForHeightFromChain in main.c
(verification). They already disagreed at exactly height == EPOCH_LENGTH,
where the constants.h copy underflowed size_t computing
`Chain_Size - 1 - EPOCH_LENGTH` and returned 0.
* A MEANINGLESS DIFFICULTY COUPLING. difficultyTarget is compact-encoded
[1B exponent][3B coefficient]; subtracting two compact values mixes
exponent and mantissa. The first real retarget yields a delta of 900649,
which times DAG_BASE_GROWTH is ~967 TB, so the result was always clamped
and the "proportional" term degenerated to a binary base+maxUp/base-maxDown
switch. The size also never accumulated -- it was always recomputed from
the constant base -- so the documented 1 GB/epoch growth could not happen.
* A NON-EPOCH-ALIGNED SEED. GetNextDAGSeed returned hash(tip) while the
verifier correctly used hash(block[epochIndex * EPOCH_LENGTH - 1]). The
`chainSize % EPOCH_LENGTH == 0` rebuild guard masked this while running,
but startup called it at an arbitrary height -- so a node restarted
mid-epoch built a DAG from a different seed than one that had run straight
through the boundary, and its blocks were rejected.
* UNSIGNED PROMOTION. `DAG_BASE_GROWTH * difficultyDelta` promoted the signed
delta to unsigned long long and wrapped before being assigned back to
int64_t; the main.c copy cast first, so the two also differed in overflow
behaviour. Both are gone.
* A LEAK on the `targetSize <= 0` path in main.c (both block copies).
* A USE-AFTER-FREE at every epoch boundary: Block_RebuildAutolykos2Dag ran
DagClear -> DagAllocate -> DagGenerate from whichever thread advanced the
tip, freeing ctx->dag.buf while miners read it.
* GetAutolykos2Ctx called DagAllocate without DagGenerate, leaving
dag.len == 0 so every heavy hash failed -- which, combined with the
fail-open above, meant "everything is valid". It also memset 1 GiB that
DagGenerate immediately overwrote.
* TOCTOU / lock reentrancy: CalculateTargetDAGSize called Chain_Size 4x and
Chain_GetBlockCopy 2x, each taking chainLock for reading, making it unsafe
to call from a write-locked section.
--- New sizing rule: default-grow inside a hard band ---
DAG size now follows a recurrence gated by a miner signal in the block header,
clamped to [DAG_MIN_SIZE, DAG_MAX_SIZE]:
brake = (hold + down) * DAG_BRAKE_DEN > EPOCH_LENGTH * DAG_BRAKE_NUM
downQ = down * DAG_DOWN_DEN > EPOCH_LENGTH * DAG_DOWN_NUM
downQ(k) && downQ(k-1) -> size -= DAG_EPOCH_STEP (floored at DAG_MIN_SIZE)
brake -> size unchanged
otherwise -> size += DAG_EPOCH_STEP (capped at DAG_MAX_SIZE)
Growth is the default and there is deliberately NO up-vote: every signal a
miner can express only slows the walk or reverses it. Under stratum-style
pooled mining the pool builds the header and therefore controls its share of
the vote, so the mechanism has to be safe under pool capture -- and it is,
because the lever a pool would want (grow the DAG to price out smaller miners)
does not exist. This is NOT because upward capture would be self-defeating: it
would in fact be profitable, since the fixed block reward redistributes to
whoever survives and difficulty retargets down. The protection is the absence
of the lever. The whole upward trajectory is therefore governance
(DAG_EPOCH_STEP, DAG_MAX_SIZE), not signalling.
Braking keeps the DAG small, which helps old hardware and only costs ASIC
resistance -- bounded by DAG_MIN_SIZE, which is the constant that actually
secures the property. Shrinking needs a supermajority sustained across two
consecutive epochs; that gates the onset only, so miners genuinely being
squeezed get relief every epoch rather than every other one.
Thresholds are cross-multiplied rather than divided, so there is no rounding
for nodes to disagree on, and the denominator is the constant epoch length
rather than blocks-observed, so a partial epoch cannot read as a stronger
signal than it is. No floating point anywhere on this path.
--- Header vote field ---
reserved[0] carries the vote: 0 = grow (default), 1 = hold, 2 = down.
reserved[1..2] must be zero. All three already sat inside the packed, hashed
header, so the vote is committed to by both the canonical hash and the PoW hash
and cannot be altered after mining -- no wire-format or hash-layout change.
0 must mean grow, because the point of this shape is that inaction produces
growth; it also means a miner that knows nothing about the vote contributes to
the intended default rather than silently freezing the schedule.
Rejecting unrecognised vote values and non-zero spare bytes is a new validity
rule. It closes 24 bits of undefined-meaning malleable header space.
--- Validation moves to the light path ---
Autolykos2_DagGenerate fills lane i with exactly what ReadDagLaneFromSeed
recomputes for lane i -- Blake2b(seed || (i/2)_LE64), half i&1 -- so the DAG is
a pure cache and the two hashing paths are bit-for-bit equivalent. Validation
therefore uses the light path: no allocation, correct for any epoch rather than
only whichever one the global DAG happens to hold, and the DAG band becomes a
miner requirement rather than a full-node memory requirement.
Chain_OnTipAdvanced no longer rebuilds the DAG at all. MineBlock builds it on
demand for the height it is working on, so a node that does not mine never
allocates one, generation stays off the tip-advance path, and the buffer is
only ever touched by the miner (the ctx mutex remains as a backstop).
--- API changes ---
+ Chain_DagParamsForHeight(chain, height, &dagBytes, seed) -- single source
of truth for both the size and the epoch seed, backed by a memoised
per-epoch table on blockchain_t. The table is a pure cache of a function of
the headers, extended lazily and dropped whenever anything at or below the
tip changes; it lives on the chain rather than in a global because a
second, header-only chain is built to re-verify historical PoW. Guarded by
a per-chain mutex, always taken after chainLock.
+ Block_EnsureAutolykos2Dag / Block_PowHashHeavy / Block_PowHashLight.
The heavy variant verifies the epoch and size itself, so it can never
answer from a DAG built for another epoch.
+ Block_HasValidProofOfWorkWithParams -- resolve-once form. MineBlock called
Block_HasValidProofOfWork inside its nonce loop, so resolving from the
chain per attempt would have taken chainLock millions of times per block.
+ Block_HasValidVote.
~ Block_HasValidProofOfWork / Block_IsFullyValid now take the chain, because
PoW validity genuinely is chain-relative. blockchain_t gained a struct tag
so block.h can forward-declare it.
- CalculateTargetDAGSize, GetNextDAGSeed, ComputeEpochDagBytesForHeightFromChain,
ComputeEpochSeedForHeightFromChain, Block_CalculateAutolykos2Hash,
Block_RebuildAutolykos2Dag, Autolykos2_LightHash (dead).
- DAG_BASE_GROWTH and the five DAG_MAX_*_SWING_* / DAG_SWING_PERCENT_DEN
macros, all now unreachable.
Also: Block_CalculateAutolykos2Hash truncated the height to uint32 while the
light path takes uint64, so the two would have diverged above block 2^32; the
full width is now passed. Added a `dagvote <grow|hold|down>` REPL command and a
progress line during DAG generation, which is tens of seconds at production
sizes.
static_asserts now enforce DAG_MIN_SIZE <= DAG_BASE_SIZE <= DAG_MAX_SIZE and
32-byte alignment, so a misconfigured band fails the build instead of being
silently clamped.
NOTE: DAG_MIN_SIZE (2 GiB), DAG_BASE_SIZE (2 GiB), DAG_MAX_SIZE (8 GiB) and
DAG_EPOCH_STEP (1 GiB) are economic judgements, not derivations, and since the
vote cannot accelerate growth they are the entire upward story. Sanity-check
them before launch.
== Verification ==
* Penalty: exact table match at d = 4/8/10/50/100/1000, zero within grace,
saturation at MAX_DEPTH, and identical wall-clock protection when rebuilt
at TARGET_BLOCK_TIME = 60.
* DAG recurrence (30 assertions against the real objects): default growth,
legacy all-zero headers read as grow, strict inequality at exactly 1/2 and
exactly 7/8, one qualifying epoch freezes but does not shrink, two
consecutive shrink, sustained shrink repeats, both clamps saturate.
* Epoch seed: constant across an epoch, equal to hash(block[k*EL - 1]),
genesis seed in epoch 0, and resolvable at exactly height == EPOCH_LENGTH.
* Fail-closed: unresolvable params and a zero-byte DAG both reject.
* Heavy/light equivalence across 3 epochs; heavy refuses a DAG built for the
wrong epoch or size.
* Two nodes across 3 epoch boundaries: B reached height 30 purely by
receiving, then mined blocks A accepted; both fullverify Chain OK, zero
rejections.
* Restart mid-epoch: B restarted at height 28 derived epoch 3's seed as
hash(block[23]) -- the boundary block, not the tip -- and kept producing
blocks A accepted. This fails before the change.
* Divergent votes: A voting hold and B voting grow computed identical size
and seed, confirming the tally is chain-derived, not config-derived.
* ThreadSanitizer across epoch rollovers under load: no new races.
Found by running two nodes under -fsanitize=thread through connect -> mine -> broadcast -> clean
exit. None are in consensus code; all are connection setup/teardown.
Stop flags were plain or volatile ints written by one thread and read as a loop condition by
another. volatile stops the compiler hoisting the load but provides neither atomicity nor
ordering, which on arm64 is a real visibility gap, not just a sanitizer complaint. Now _Atomic:
- net_node_t.maintenanceRunning (Node_Destroy vs Node_MaintenanceThread)
- tcp_server_t.isRunning (TcpServer_Stop vs both accept threads) -- was a bare int, so
the accept loop could legally be hoisted and never see the stop
- udp_node_t.isRunning (UdpNode_Stop vs the recv and retry threads)
TcpServer_Stop had a use-after-free, not merely a race: it took clientsMutex only long enough to
read the array pointer, then walked the slots unlocked. An exiting client thread clears its own
slot under that mutex and immediately destroys and frees the connection, so Stop could
RequestClose and pthread_join a freed pointer on any shutdown with an active peer.
- Stop now requests the close and copies each pthread_t under the mutex, then joins from the
copied handles, so the connection is never dereferenced outside the lock
- the client thread's TcpConnection_Destroy/free moved inside the same critical section, which
closes the window entirely
Client threads that disconnected normally were never joined and leaked their thread resources:
Stop only joins clients still present in the array, and a normal exit removes itself first. Stop
now claims each slot as it copies the handle, so the client thread can tell who owns its join --
it detaches itself if it successfully removed its own slot, and stays joinable if Stop already
claimed it. Both decisions happen under clientsMutex so the cases cannot interleave.
Node_Destroy cleared outbound slots with no lock (via TcpClient_Disconnect) while live inbound
client threads read the same field correctly under outboundLock in Node_HasLiveConnectionTo. The
lock cannot just be held across the destroy, because that path joins an io thread whose
on_disconnect callback takes outboundLock itself. Reworked to detach the connections from their
slots under the lock and tear them down outside it -- the pattern Node_ReapDeadOutbound already
uses in this file. This one only surfaced once the other five were fixed.
Verified: TSan clean over the same run; nodes still converge (height 25, identical tip); shallow
fork still adopts, depth-8 fork still defers on the reorg penalty, and the forced-orphan
regression still reaches full height with zero coinbase rejections.
(This is a big one, get ready - I told Claude to write the commit message cause I couldn't be bothered)
Root cause: reorg was broken at every layer and the failures compounded. Verified with the node's own
SKALACOIN_FORCE_ORPHAN_REORG debug mode, which stalled permanently at height 2 on 114 consecutive
coinbase-validation failures. A binary built at HEAD behaves identically, so none of this is a regression —
the reorg path had simply never worked.
Rollback (the keystone):
- Chain_RollbackToHeight always returned false on any node that had ever saved or loaded its chain, and
only after it had already truncated the chain and destroyed the balance sheet. Chain_RecomputeRuntimeState
bails on any header-only block, and Chain_SaveToFile nulls transactions on every in-memory block once
persisted, so the failure was universal in practice. Both callers treated the false as "nothing happened"
- Added Chain_BorrowBlockTransactions / Chain_ReturnBlockTransactions, which fall back to the on-disk copy
when the in-memory block has been compacted to headers
- Supply is now accumulated in the rollback's existing balance-sheet replay pass instead of a second
Chain_RecomputeRuntimeState pass
- Gave Chain_RecomputeRuntimeState the same disk fallback: it had been failing on every restart with an
existing chain, silently leaving currentSupply/currentReward at whatever came out of chain.meta
Fork choice is now cumulative work, not height:
- Added Chain_ComputeBlockWork / Chain_ComputeWorkRange / Chain_ComputeBranchWork, computing
2^256 / (target + 1) per block and summing over a range. Derived on demand from headers — no header,
chain.meta or wire-format change
- uint256 only had add/sub/cmp, so added uint256_divide (restoring binary long division),
uint256_from_be_bytes, uint256_bitwise_not and uint256_is_zero
- Comparison is strictly greater, so tied tips do not cause two nodes to keep swapping
- Height-based choice was wrong now that difficulty actually varies: a long low-difficulty branch beat a
short high-difficulty one
Atomic branch replacement:
- Added Chain_ReplaceBranch: validate linkage, apply the reorg penalty, compare work, snapshot the outgoing
blocks, roll back, apply. On any failure the original chain, balance sheet, supply, reward and difficulty
target are restored. The caller keeps ownership of its blocks in every case — the chain applies copies
- Split Chain_AddBlock and Chain_RollbackToHeight into locked public wrappers over unlocked internals, so a
whole branch swap happens under one lock acquisition and Chain_OnTipAdvanced runs once per reorg rather
than once per block
- Chain_AddBlock now validates header.prevHash against the tip. It never did — that check lived only in
Chain_IsValid and the network path, which is exactly why a rollback-then-reapply could splice blocks from
two different forks into a chain that no longer links up
- Moved the currentSupply/currentReward update into Chain_AddBlock. Each caller used to do it separately, so
the orphan-attach and maintenance-thread paths never did, and the next block's coinbase was then validated
against a stale currentReward and rejected forever. This was the height-2 stall
Reorg penalty (Horizen-style delayed block submission):
- The penalty was only ever reachable from the manual sync command. The P2P broadcast -> orphan pool ->
branch adoption path, which is the path an attacker actually uses, had none at all and picked the winner
by raw height. It is now enforced inside Chain_ReplaceBranch, the single choke point every adoption
passes through
- Removed its application to the sync fetch window. The height gap to a peer is not a reorg depth;
penalising it only throttled honest catch-up, and for gaps of 4-50 it collapsed the window to one block
per pass, defeating MAX_PARALLEL_FETCHES
- The depth is stamped once when a branch is first observed (orphan_entry_t.observedAtTipHeight) and never
recomputed. Re-deriving it from a moving tip never converges: depth and elapsed both grow by one per block
while penalty(depth) grows faster, so a penalized branch could never be adopted at all
- The initial-sync exemption now comes from Chain_IsInitialBlockDownload, which uses the local median time
past over MEDIAN_TIME_SPAN blocks. It used to key off the peer's advertised height, so any peer claiming
localHeight + INITIAL_SYNC_HEIGHT_DIFF could switch reorg handling off for the whole session. A median
rather than the tip alone means one backdated block cannot fake it either
Orphan pool (largely rewritten):
- Added a pool mutex. It had no synchronisation whatsoever while being mutated from the 1 Hz maintenance
thread, every per-peer TCP thread and the REPL thread; a concurrent insert could realloc the array while a
scan held a raw element pointer. The lock is never held across a call into chain.c
- Dedup by block hash, a MAX_ORPHAN_BLOCKS cap with oldest-first eviction, and pruning of entries that can
no longer apply. Nothing was ever reaped before, and orphans are reachable before the chain-derived
difficulty check, so this is also the memory-exhaustion fix
- Candidate branches are now assembled by following prevHash from the fork point. Taking the first orphan
found at each successive height could interleave blocks from two competing forks into one incoherent branch
- Fixed rollbackHeight = forkHeight - 1. Chain_RollbackToHeight is exclusive, so every non-genesis adoption
amputated one block too many and then failed Chain_AddBlock's index check
- Fixed a block_t wrapper leak on every successful attach (free the wrapper, not Block_Destroy — the chain
owns the transactions after a shallow copy)
- Permanently invalid orphans are dropped instead of being retried on every maintenance tick forever
Forks below the tip are now discoverable:
- A block at blockNumber < chainSize was rejected and freed, so the fork point and the lower half of any
competing branch were always thrown away and a sub-tip fork could never be learned. Now the hash is
compared: identical means a duplicate and is ignored, different means it goes to the orphan pool
- The sync loop probes downwards (RequestForkWindow, bounded by REORG_FETCH_DEPTH and MAX_FORK_PROBE_ROUNDS)
when it makes no progress while the peer is ahead. That is the only trigger that fires for a genuine
sub-tip fork, because the old divergence check could only see blocks that had already entered our chain.
FETCH_BLOCK already answers from the peer's own chain, so no protocol change was needed
- Removed the rollback-to-height-0 path. "Could not find the parent" used to wipe the entire local chain,
genesis included, and any peer could trigger it with a single unlinked block
Floating point removed from consensus and network math:
- Chain_ComputeTargetAtHeight (the difficulty retarget) used double ratio arithmetic, and
FetchScheduler_ComputeReorgPenaltyBlocks used double/pow/ceil. Both are consensus-critical and are now
integer only; float results are not reproducible across platforms and compilers, and a single last-digit
difference in a target or a penalty splits the network
- The penalty constants became integer rationals (REORG_PENALTY_FACTOR_NUM/DEN, integer EXPONENT and
REF_BLOCK_TIME) with saturating exponentiation and explicit ceiling division. Output is unchanged:
penalty(4)=10, penalty(8)=39, penalty(10)=60, penalty(50)=1500, penalty(100)=6000
- Removed the unused float macros DAG_MAX_UP/DOWN_SWING_PERCENTAGE and the now-dead math.h include from
constants.h. Both were latent: DAG size feeds PoW verification. Replaced with integer numerator/denominator
constants and used them at both clamp sites (values verified identical)
Other fixes that were blocking fork propagation:
- madeProgressOverall was set but never reset, so after one productive pass the "no progress -> stop" guard
could never fire again and the sync loop could spin forever holding the REPL
- seenBlocks was inserted before/regardless of a successful send, so a block broadcast while no peer was
connected was never offered again. It is now recorded only after the block actually goes out
- Broadcasts relayed to outbound connections only, so in a two-node setup the dialled node never pushed
anything back and the dialer learned of new blocks only via a manual sync. Inbound peers are now relayed to
Verified with two-node harnesses at a shortened adjustment interval:
- forced-orphan regression: was height 2 with 114 coinbase rejections, now reaches the peer's height with
zero rejections and both nodes report Chain OK
- sub-tip fork at depth 3: the node discovers the fork below its own tip, discards its three blocks, adopts
the heavier five, and both nodes converge on an identical tip hash
- deep fork at depth 8: the strictly heavier branch is correctly refused with depth=8 penalty=39 elapsed=0
- uint256 work arithmetic covered by a standalone test (division, big-endian conversion, monotonicity,
halved target doubles work)
One issue found along the way: the chain in build/chain_data does not pass fullverify. Block 7683 reverts to
INITIAL_DIFFICULTY where it should carry 0x1f06df14, i.e. it contains blocks mined before 1288a64 landed. A
binary built at HEAD rejects it too, so this is stale data rather than a regression — it needs a wipechain
and a re-mine.
Root cause: peers were keyed only by (IP, listen port). An IPv6 host holds several addresses at once, so one node appeared as several peers and could not recognise its own addresses.
Protocol:
- Added a random per-run node identity (localNodeId), advertised as a length-guarded trailing field in HELLO and ACK_HELLO — backwards compatible with peers that omit it
- Added peerNodeId to tcp_connection_t; peers are now identified by this rather than by an endpoint
- ACK_HELLO is now sent before any decision to drop the connection, so a rejected dialer learns whose address it reached instead of retrying forever
Self-connection:
- Identity match → close the connection and record that endpoint permanently as our own
- Self-endpoint set preemptively seeded from getifaddrs() at startup, so a node knows its own addresses before ever dialing one
- Node_ConnectPeer refuses self endpoints, which also kills the echo-back chain that could exhaust connection slots
Duplicate connections and churn:
- Dedup by identity per direction — one inbound and one outbound per physical peer, regardless of how many addresses it has
- Moved dial history out of the peer table, so striking a peer no longer resets its connect-retry cooldown (this was the loop engine: strike → re-learn via gossip → redial next tick)
- Node_HasOtherInboundFrom / Node_HasLiveConnectionTo now ignore connections that are already tearing down, and match on identity as well as endpoint
Gossip hygiene:
- Never hand a peer its own other addresses in a PEERS reply (matched on identity, not just the socket address)
- Reject unusable endpoints: link-local without scope id, unspecified, multicast, site-local (loopback stays allowed for local testing)
- Normalise IPv4-mapped IPv6 so one host cannot occupy two entries
Two bugs found along the way:
- All nodes drew the same identity — random_eight_byte() comes from srand(time(NULL)), so processes started in the same second produced identical values. Added random_secure_eight_byte() (/dev/urandom) for the identity, and mixed the pid into the seed so connection IDs stop colliding too
- Identity dedup initially left inbound-only nodes mute — broadcasts traverse outbound connections only, so suppressing a dial-back because an inbound existed would have silenced such a node. Corrected to per-direction
Other:
- peers output now shows each entry's node identity and the node's own endpoints
- Added _DEFAULT_SOURCE to the build so getifaddrs() stays visible on glibc
Reclaim outbound slots on peer disconnect:
- TcpClient_ThreadProc fired on_disconnect but never cleared the outbound
slot, closed the fd, joined the io thread, or freed the connection, so a
dead peer permanently held its outboundClients[] slot. After MAX_CONS (32)
churned connections the node could make no new outbound connections, and
leaked fds/threads/memory. (The inbound side already self-reclaimed.)
- Add a reaper (Node_ReapDeadOutbound) on the maintenance thread: under
outboundLock it detaches dead (disconnect-notified) slots, then joins the
io thread and destroys/frees each connection outside the lock.
- Guard against use-after-free with a pin count on tcp_connection_t
(TcpConnection_Pin/Unpin). The only cross-thread consumer holding a raw
connection pointer across a blocking op is the `sync` command (via
Node_GetBestOutboundPeer); it now pins the peer and unpins when done, and
the reaper skips pinned connections. Discovery's snapshots run on the
reaper's own thread, so they need no pin.
- Node_GetBestOutboundPeer/GetClientList/GetPeerEndpoints skip
disconnect-notified connections so a dead peer is never handed out.
- Node_Destroy stops+joins the maintenance thread before tearing down
outbound clients, so the reaper can't race shutdown.
Strike disconnected peers from the discovery peer list:
- Add NodeDiscovery_RemovePeer + Node_HandlePeerDisconnect: on disconnect,
remove the peer from the known-peer table once no live connection (inbound
or outbound) to its listen endpoint remains (Node_HasLiveConnectionTo;
disconnect-notified conns don't count, so both directions dropping at once
is handled). Wired into Node_Server_OnDisconnect and Node_Client_OnDisconnect.
- Implement NodeDiscovery engine: known-peer table (DynArr) with a
per-tick seed/ping/query/connect state machine driven by the node
maintenance thread; bounded multi-hop crawl (FANOUT peers per node,
hop-capped) that connects to reachable peers lowest-ping-first
- Add GET_PEERS/PEERS TCP opcodes for peer-list exchange, handled on
both inbound and outbound connections
- Measure UDP round-trip time and pass it to the on_pong callback
(previously the send timestamp was only used for retries)
- Advertise each node's listen port in HELLO/ACK_HELLO and store it
per-connection, so inbound-only peers and non-default ports are
discoverable (length-guarded parse; wire-compatible with old peers)
- Wire a udp_node_t + node_discovery_t into net_node_t: init/start in
Node_Create, tick in the maintenance loop, teardown in Node_Destroy
(stop UDP before destroying discovery to avoid callback races)
- Add Node_ConnListenEndpoint / Node_GetPeerEndpoints helpers to derive
peers' listen endpoints (outbound: dialed port; inbound: advertised),
with IPv4-mapped-IPv6 normalization and IP+port dedup
- Match ping pong/timeout callbacks by peer address (the UDP layer owns
the nonce), fixing discovered peers stuck UNREACHABLE
- Ping the peer's listen port instead of the ephemeral TCP source port
(fixes the original stub so pongs actually return)
- Add `peers` CLI command to dump the discovery table (endpoint/hop/
state/ping)
- Add discovery tunables to constants.h (fanout, max hops, target
connections, timeouts, caps)
Mining: blocks now include mempool txs, select spendable txs by fee, and pay coinbase as base reward + fees in main.c.
- Consensus: block validation now enforces coinbase accounting and rejects invalid coinbase placement, including coinbase on amount2, in block.c and transaction.c.
- Chain state: rollback now rebuilds currentSupply/currentReward, and block addition preflights spendability before mutating balances in chain.c.
- Orphans/reorgs: orphan retry is safer, rollback-triggered sync reattaches orphans immediately, and transient orphan failures no longer drop blocks in orphan_pool.c and main.c.
- Networking/mempool: node lifecycle now initializes the mempool, broadcasts can exclude one peer, and mempool snapshotting supports mining selection in net_node.c and txmempool.c.
- Ledger simulation: added non-mutating spendable-transaction selection for block assembly in balance_sheet.c.
dcrubro