Two changes:
First removes a unit test that fails in my development environment
(OSX, compiled -g3 with clang).
sipa says that's not terribly surprising; the CMutableTransaction change
makes signing a little more expensive but verification quicker. The unit
test timed sign+verify-uncached versus verify-cached-five-times.
He also says the test will be invalid when libsec256kp1 is integrated
(because validation is super-optimized over signing).
core.h change fixes a compiler warning (clang -Wall : CMutableTransaction defined
as struct, declared as class in script.h).
Relax the AreInputsStandard() tests for P2SH transactions --
allow any Script in a P2SH transaction to be relayed/mined,
as long as it has 15 or fewer signature operations.
Rationale: https://gist.github.com/gavinandresen/88be40c141bc67acb247
I don't have an easy way to test this, but the code changes are
straightforward and I've updated the AreInputsStandard unit tests.
New RPC methods: return an estimate of the fee (or priority) a
transaction needs to be likely to confirm in a given number of
blocks.
Mike Hearn created the first version of this method for estimating fees.
It works as follows:
For transactions that took 1 to N (I picked N=25) blocks to confirm,
keep N buckets with at most 100 entries in each recording the
fees-per-kilobyte paid by those transactions.
(separate buckets are kept for transactions that confirmed because
they are high-priority)
The buckets are filled as blocks are found, and are saved/restored
in a new fee_estiamtes.dat file in the data directory.
A few variations on Mike's initial scheme:
To estimate the fee needed for a transaction to confirm in X buckets,
all of the samples in all of the buckets are used and a median of
all of the data is used to make the estimate. For example, imagine
25 buckets each containing the full 100 entries. Those 2,500 samples
are sorted, and the estimate of the fee needed to confirm in the very
next block is the 50'th-highest-fee-entry in that sorted list; the
estimate of the fee needed to confirm in the next two blocks is the
150'th-highest-fee-entry, etc.
That algorithm has the nice property that estimates of how much fee
you need to pay to get confirmed in block N will always be greater
than or equal to the estimate for block N+1. It would clearly be wrong
to say "pay 11 uBTC and you'll get confirmed in 3 blocks, but pay
12 uBTC and it will take LONGER".
A single block will not contribute more than 10 entries to any one
bucket, so a single miner and a large block cannot overwhelm
the estimates.
Use CFeeRate instead of an int64_t for quantities that are
fee-per-size.
Helps prevent unit-conversion mismatches between the wallet,
relaying, and mining code.
Taught bitcoind to close the HTTP connection after it gets a 'stop' command,
to make it easier for the regression tests to cleanly stop.
Move bitcoinrpc files to correct location.
Tidied up the python-based regression tests.
Compiling with -DDEBUG_LOCKORDER and running the qa/rpc-test/ regression
tests uncovered a couple of wallet methods that should (but didn't)
acquire the cs_wallet mutext.
I also changed the AssertLockHeld() routine print to stderr and
abort, instead of printing to debug.log and then assert()'ing.
It is annoying to look in debug.log to find out which
AssertLockHeld is failing.
Adds a "walletconflicts" array to transaction info; if
a wallet transaction is mutated, the alternate transaction id
or ids are reported there (usually the array will be empty).
Metadata from the original transaction is copied to the mutant,
so the transaction time and "from" account of the mutant are
reported correctly.
Extend CMerkleTx::GetDepthInMainChain with the concept of
a "conflicted" transaction-- a transaction generated by the wallet
that is not in the main chain or in the mempool, and, therefore,
will likely never be confirmed.
GetDepthInMainChain() now returns -1 for conflicted transactions
(0 for unconfirmed-but-in-the-mempool, and >1 for confirmed).
This makes getbalance, getbalance '*', and listunspent all agree when there are
mutated transactions in the wallet.
Before:
listunspent: one 49BTC output
getbalance: 96 BTC (change counted twice)
getbalance '*': 46 BTC (spends counted twice)
After: all agree, 49 BTC available to spend.
Reworked send.sh, so it works properly on my Mac (killall send.sh
doesn't work, because the process name is 'bash' not 'send.sh').
So now send.sh writes a .send.pid file, and invoking it as
send.sh -STOP (as the bitcoind -walletnotify) signals that PID.
Reworked send.sh, so it works properly on my Mac (killall send.sh
doesn't work, because the process name is 'bash' not 'send.sh').
So now send.sh writes a .send.pid file, and invoking it as
send.sh -STOP (as the bitcoind -walletnotify) signals that PID.
Unit tests would fail if compiled with -DDEBUG_LOCKORDER (AssertLockHeld()
would fail; AssertLockHeld() relies on the DEBUG_LOCKORDER code to keep
track of locks held).
Fixed by LOCK'ing the wallet mutex in the unit tests that manipulate the
wallet.
Running -printtodebugger -debug (or -debug=lock),
compiled with -DDEBUG_LOCKORDER would infinite loop
on Windows because every critical section lock/unlock
triggers a LogPrint.
Solution is to use the raw boost mutex instead of a CCriticalSection.
This dead code can be resurrected from git history if
transaction replacement is ever implemented. Keeping
dead code in the source is a bad idea, because it implies
it was tested and worked at some point, which is not true.
qa/rpc-tests/wallet.sh runs a three-node -regtest network,
generates a fresh blockchain, and then exercises basic wallet
sending/receiving functionality using command-line RPC.
I'm writing some wallet regression tests using -regtest mode, and
need to generate an initial multi-hundred-block chain. Repeatedly
calling setgenerate to generate one block is slow and doesn't
work properly, because block creation happens asynchronously.
This adds two features to setgenerate in -regtest mode:
1) Instead of being interpreted as number of threads to start, the
third argument is the number of blocks to generate.
2) setgenerate will not return until the block creation threads
have created the requested number of blocks.
I regenerated the alert test data; now alerts are tested
against a protocol version way above the current protocol
version.
So we won't have to regenerate them every time we bump
PROTOCOL_VERSION in the future.
Gavin Andresen