7f7bd3111c
57ce20307e604530f78ef4f0f8d9fb94f80ca81b fuzz: allow lower number of sources (Martin Zumsande)
acf656d540a82e6fc30421590305cfe295eabbb5 fuzz: Use public interface to fill addrman tried tables (Martin Zumsande)
eb2e113df13c7b1ede279878f5cbad877af49f8e addrman: Improve performance of Good (Martin Zumsande)
Pull request description:
Currently, `CAddrman::Good()` is rather slow because the process of moving an addr from new to tried involves looping over the new tables twice:
1) In `Good_()`, there is a loop searching for a new bucket the addr is currently in, but this information is never used except for aborting if it is not found anywhere (since [this commit](e6b343d880 (diff-49d1faa58beca1ee1509a247e0331bb91f8604e30a483a7b2dea813e6cea02e2R263)) it is no longer passed to `MakeTried`)
This is unnecessary because in a non-corrupted addrman, an address that is not in New must be either in Tried or not at all in addrman, both cases in which we'd return early in `Good_()` and never get to this point.
I removed this loop (and left a check for `nRefCount` as a belt-and-suspenders check).
2) In `MakeTried()`, which is called from `Good_()`, another loop removes all instances of this address from new. This can be spedup by stopping the search at `nRefCount==0`. Further reductions in `nRefCount` would only lead to an assert anyway.
Moreover, the search can be started at the bucket determined by the source of the addr for which `Good` was called, so that if it is present just once in New, no further buckets need to be checked.
While calls to `Good()` are not that frequent normally, the performance gain is clearly seen in the fuzz target `addman_serdeser`, where, because of the slowness in creating a decently filled addrman, a shortcut was created that would directly populate the tried tables by reaching into addrman's internals, bypassing `Good()` (#21129).
I removed this workaround in the second commit: Using `Good()` is still slower by a factor of 2 (down from a factor of ~60 before), but I think that this compensated by the advantages of not having to reach into the internal structures of addrman (see https://github.com/jnewbery/bitcoin/pull/18#issuecomment-775218676).
[Edit]: For benchmark results see https://github.com/bitcoin/bitcoin/pull/22974#issuecomment-919435266 and https://github.com/bitcoin/bitcoin/pull/22974#issuecomment-920445700 - the benchmark `AddrManGood` shows a significant speedup by a factor >100.
ACKs for top commit:
naumenkogs:
ACK 57ce20307e604530f78ef4f0f8d9fb94f80ca81b
jnewbery:
ACK 57ce20307e
laanwj:
Code review ACK 57ce20307e604530f78ef4f0f8d9fb94f80ca81b
theStack:
ACK 57ce20307e604530f78ef4f0f8d9fb94f80ca81b
vasild:
ACK 57ce20307e604530f78ef4f0f8d9fb94f80ca81b
Tree-SHA512: fb6dfc198f2e28bdbb41cef9709828f22d83b4be0e640a3155ca42e771b6f58466de1468f54d773e794f780a79113f9f7d522032e87fdd75bdc4d99330445198
Unit tests
The sources in this directory are unit test cases. Boost includes a unit testing framework, and since Bitcoin Core already uses Boost, it makes sense to simply use this framework rather than require developers to configure some other framework (we want as few impediments to creating unit tests as possible).
The build system is set up to compile an executable called test_bitcoin
that runs all of the unit tests. The main source file for the test library is found in
util/setup_common.cpp.
Compiling/running unit tests
Unit tests will be automatically compiled if dependencies were met in ./configure
and tests weren't explicitly disabled.
After configuring, they can be run with make check.
To run the unit tests manually, launch src/test/test_bitcoin. To recompile
after a test file was modified, run make and then run the test again. If you
modify a non-test file, use make -C src/test to recompile only what's needed
to run the unit tests.
To add more unit tests, add BOOST_AUTO_TEST_CASE functions to the existing
.cpp files in the test/ directory or add new .cpp files that
implement new BOOST_AUTO_TEST_SUITE sections.
To run the GUI unit tests manually, launch src/qt/test/test_bitcoin-qt
To add more GUI unit tests, add them to the src/qt/test/ directory and
the src/qt/test/test_main.cpp file.
Running individual tests
test_bitcoin has some built-in command-line arguments; for
example, to run just the getarg_tests verbosely:
test_bitcoin --log_level=all --run_test=getarg_tests -- DEBUG_LOG_OUT
log_level controls the verbosity of the test framework, which logs when a
test case is entered, for example. The DEBUG_LOG_OUT after the two dashes
redirects the debug log, which would normally go to a file in the test datadir
(BasicTestingSetup::m_path_root), to the standard terminal output.
... or to run just the doubledash test:
test_bitcoin --run_test=getarg_tests/doubledash
Run test_bitcoin --help for the full list.
Adding test cases
To add a new unit test file to our test suite you need
to add the file to src/Makefile.test.include. The pattern is to create
one test file for each class or source file for which you want to create
unit tests. The file naming convention is <source_filename>_tests.cpp
and such files should wrap their tests in a test suite
called <source_filename>_tests. For an example of this pattern,
see uint256_tests.cpp.
Logging and debugging in unit tests
make check will write to a log file foo_tests.cpp.log and display this file
on failure. For running individual tests verbosely, refer to the section
above.
To write to logs from unit tests you need to use specific message methods
provided by Boost. The simplest is BOOST_TEST_MESSAGE.
For debugging you can launch the test_bitcoin executable with gdbor lldb and
start debugging, just like you would with any other program:
gdb src/test/test_bitcoin
Segmentation faults
If you hit a segmentation fault during a test run, you can diagnose where the fault
is happening by running gdb ./src/test/test_bitcoin and then using the bt command
within gdb.
Another tool that can be used to resolve segmentation faults is valgrind.
If for whatever reason you want to produce a core dump file for this fault, you can do
that as well. By default, the boost test runner will intercept system errors and not
produce a core file. To bypass this, add --catch_system_errors=no to the
test_bitcoin arguments and ensure that your ulimits are set properly (e.g. ulimit -c unlimited).
Running the tests and hitting a segmentation fault should now produce a file called core
(on Linux platforms, the file name will likely depend on the contents of
/proc/sys/kernel/core_pattern).
You can then explore the core dump using
gdb src/test/test_bitcoin core
(gbd) bt # produce a backtrace for where a segfault occurred