You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long. 26KB

Developer Notes

Various coding styles have been used during the history of the codebase, and the result is not very consistent. However, we’re now trying to converge to a single style, which is specified below. When writing patches, favor the new style over attempting to mimic the surrounding style, except for move-only commits.

Do not submit patches solely to modify the style of existing code.

  • Indentation and whitespace rules as specified in src/.clang-format. You can use the provided clang-format-diff script tool to clean up patches automatically before submission.

    • Braces on new lines for namespaces, classes, functions, methods.
    • Braces on the same line for everything else.
    • 4 space indentation (no tabs) for every block except namespaces.
    • No indentation for public/protected/private or for namespace.
    • No extra spaces inside parenthesis; don’t do ( this )
    • No space after function names; one space after if, for and while.
    • If an if only has a single-statement then-clause, it can appear on the same line as the if, without braces. In every other case, braces are required, and the then and else clauses must appear correctly indented on a new line.
  • Symbol naming conventions. These are preferred in new code, but are not required when doing so would need changes to significant pieces of existing code.

    • Variable and namespace names are all lowercase, and may use _ to separate words (snake_case).
    • Class member variables have a m_ prefix.
    • Global variables have a g_ prefix.
    • Constant names are all uppercase, and use _ to separate words.
    • Class names, function names and method names are UpperCamelCase (PascalCase). Do not prefix class names with C.
  • Miscellaneous

    • ++i is preferred over i++.
    • nullptr is preferred over NULL or (void*)0.
    • static_assert is preferred over assert where possible. Generally; compile-time checking is preferred over run-time checking.

Block style example:

int g_count = 0;

namespace foo
class Class
    std::string m_name;

    bool Function(const std::string& s, int n)
        // Comment summarising what this section of code does
        for (int i = 0; i < n; ++i) {
            int total_sum = 0;
            // When something fails, return early
            if (!Something()) return false;
            if (SomethingElse(i)) {
                total_sum += ComputeSomething(g_count);
            } else {
                DoSomething(m_name, total_sum);

        // Success return is usually at the end
        return true;
} // namespace foo

Doxygen comments

To facilitate the generation of documentation, use doxygen-compatible comment blocks for functions, methods and fields.

For example, to describe a function use:

 * ... text ...
 * @param[in] arg1    A description
 * @param[in] arg2    Another argument description
 * @pre Precondition for function...
bool function(int arg1, const char *arg2)

A complete list of @xxx commands can be found at As Doxygen recognizes the comments by the delimiters (/** and */ in this case), you don’t need to provide any commands for a comment to be valid; just a description text is fine.

To describe a class use the same construct above the class definition:

 * Alerts are for notifying old versions if they become too obsolete and
 * need to upgrade. The message is displayed in the status bar.
 * @see GetWarnings()
class CAlert

To describe a member or variable use:

int var; //!< Detailed description after the member


//! Description before the member
int var;

Also OK:

/// ... text ...
bool function2(int arg1, const char *arg2)

Not OK (used plenty in the current source, but not picked up):

// ... text ...

A full list of comment syntaxes picked up by doxygen can be found at, but if possible use one of the above styles.

Development tips and tricks

compiling for debugging

Run configure with the --enable-debug option, then make. Or run configure with CXXFLAGS=“-g -ggdb -O0” or whatever debug flags you need.


If the code is behaving strangely, take a look in the debug.log file in the data directory; error and debugging messages are written there.

The -debug=… command-line option controls debugging; running with just -debug or -debug=1 will turn on all categories (and give you a very large debug.log file).

The Qt code routes qDebug() output to debug.log under category “qt”: run with -debug=qt to see it.

testnet and regtest modes

Run with the -testnet option to run with “play bitcoins” on the test network, if you are testing multi-machine code that needs to operate across the internet.

If you are testing something that can run on one machine, run with the -regtest option. In regression test mode, blocks can be created on-demand; see test/functional/ for tests that run in -regtest mode.


Bitcoin Core is a multithreaded application, and deadlocks or other multithreading bugs can be very difficult to track down. Compiling with -DDEBUG_LOCKORDER (configure CXXFLAGS=“-DDEBUG_LOCKORDER -g”) inserts run-time checks to keep track of which locks are held, and adds warnings to the debug.log file if inconsistencies are detected.

Valgrind suppressions file

Valgrind is a programming tool for memory debugging, memory leak detection, and profiling. The repo contains a Valgrind suppressions file (valgrind.supp) which includes known Valgrind warnings in our dependencies that cannot be fixed in-tree. Example use:

$ valgrind --suppressions=contrib/valgrind.supp src/test/test_bitcoin
$ valgrind --suppressions=contrib/valgrind.supp --leak-check=full \
      --show-leak-kinds=all src/test/test_bitcoin --log_level=test_suite
$ valgrind -v --leak-check=full src/bitcoind -printtoconsole

compiling for test coverage

LCOV can be used to generate a test coverage report based upon make check execution. LCOV must be installed on your system (e.g. the lcov package on Debian/Ubuntu).

To enable LCOV report generation during test runs:

./configure --enable-lcov
make cov

# A coverage report will now be accessible at `./test_bitcoin.coverage/index.html`.

Locking/mutex usage notes

The code is multi-threaded, and uses mutexes and the LOCK/TRY_LOCK macros to protect data structures.

Deadlocks due to inconsistent lock ordering (thread 1 locks cs_main and then cs_wallet, while thread 2 locks them in the opposite order: result, deadlock as each waits for the other to release its lock) are a problem. Compile with -DDEBUG_LOCKORDER to get lock order inconsistencies reported in the debug.log file.

Re-architecting the core code so there are better-defined interfaces between the various components is a goal, with any necessary locking done by the components (e.g. see the self-contained CKeyStore class and its cs_KeyStore lock for example).


  • ThreadScriptCheck : Verifies block scripts.

  • ThreadImport : Loads blocks from blk*.dat files or bootstrap.dat.

  • StartNode : Starts other threads.

  • ThreadDNSAddressSeed : Loads addresses of peers from the DNS.

  • ThreadMapPort : Universal plug-and-play startup/shutdown

  • ThreadSocketHandler : Sends/Receives data from peers on port 8333.

  • ThreadOpenAddedConnections : Opens network connections to added nodes.

  • ThreadOpenConnections : Initiates new connections to peers.

  • ThreadMessageHandler : Higher-level message handling (sending and receiving).

  • DumpAddresses : Dumps IP addresses of nodes to peers.dat.

  • ThreadFlushWalletDB : Close the wallet.dat file if it hasn’t been used in 500ms.

  • ThreadRPCServer : Remote procedure call handler, listens on port 8332 for connections and services them.

  • BitcoinMiner : Generates bitcoins (if wallet is enabled).

  • Shutdown : Does an orderly shutdown of everything.

Ignoring IDE/editor files

In closed-source environments in which everyone uses the same IDE it is common to add temporary files it produces to the project-wide .gitignore file.

However, in open source software such as Bitcoin Core, where everyone uses their own editors/IDE/tools, it is less common. Only you know what files your editor produces and this may change from version to version. The canonical way to do this is thus to create your local gitignore. Add this to ~/.gitconfig:

        excludesfile = /home/.../.gitignore_global

(alternatively, type the command git config --global core.excludesfile ~/.gitignore_global on a terminal)

Then put your favourite tool’s temporary filenames in that file, e.g.

# NetBeans

Another option is to create a per-repository excludes file .git/info/exclude. These are not committed but apply only to one repository.

If a set of tools is used by the build system or scripts the repository (for example, lcov) it is perfectly acceptable to add its files to .gitignore and commit them.

Development guidelines

A few non-style-related recommendations for developers, as well as points to pay attention to for reviewers of Bitcoin Core code.

General Bitcoin Core

  • New features should be exposed on RPC first, then can be made available in the GUI

    • Rationale: RPC allows for better automatic testing. The test suite for the GUI is very limited
  • Make sure pull requests pass Travis CI before merging

    • Rationale: Makes sure that they pass thorough testing, and that the tester will keep passing on the master branch. Otherwise all new pull requests will start failing the tests, resulting in confusion and mayhem

    • Explanation: If the test suite is to be updated for a change, this has to be done first


  • Make sure that no crashes happen with run-time option -disablewallet.

    • Rationale: In RPC code that conditionally uses the wallet (such as validateaddress) it is easy to forget that global pointer pwalletMain can be nullptr. See test/functional/ for functional tests exercising the API with -disablewallet
  • Include db_cxx.h (BerkeleyDB header) only when ENABLE_WALLET is set

    • Rationale: Otherwise compilation of the disable-wallet build will fail in environments without BerkeleyDB

General C++

  • Assertions should not have side-effects

    • Rationale: Even though the source code is set to refuse to compile with assertions disabled, having side-effects in assertions is unexpected and makes the code harder to understand
  • If you use the .h, you must link the .cpp

    • Rationale: Include files define the interface for the code in implementation files. Including one but not linking the other is confusing. Please avoid that. Moving functions from the .h to the .cpp should not result in build errors
  • Use the RAII (Resource Acquisition Is Initialization) paradigm where possible. For example by using unique_ptr for allocations in a function.

    • Rationale: This avoids memory and resource leaks, and ensures exception safety

C++ data structures

  • Never use the std::map [] syntax when reading from a map, but instead use .find()

    • Rationale: [] does an insert (of the default element) if the item doesn’t exist in the map yet. This has resulted in memory leaks in the past, as well as race conditions (expecting read-read behavior). Using [] is fine for writing to a map
  • Do not compare an iterator from one data structure with an iterator of another data structure (even if of the same type)

    • Rationale: Behavior is undefined. In C++ parlor this means “may reformat the universe”, in practice this has resulted in at least one hard-to-debug crash bug
  • Watch out for out-of-bounds vector access. &vch[vch.size()] is illegal, including &vch[0] for an empty vector. Use and + vch.size() instead.

  • Vector bounds checking is only enabled in debug mode. Do not rely on it

  • Make sure that constructors initialize all fields. If this is skipped for a good reason (i.e., optimization on the critical path), add an explicit comment about this

    • Rationale: Ensure determinism by avoiding accidental use of uninitialized values. Also, static analyzers balk about this.
  • By default, declare single-argument constructors explicit.

    • Rationale: This is a precaution to avoid unintended conversions that might arise when single-argument constructors are used as implicit conversion functions.
  • Use explicitly signed or unsigned chars, or even better uint8_t and int8_t. Do not use bare char unless it is to pass to a third-party API. This type can be signed or unsigned depending on the architecture, which can lead to interoperability problems or dangerous conditions such as out-of-bounds array accesses

  • Prefer explicit constructions over implicit ones that rely on ‘magical’ C++ behavior

    • Rationale: Easier to understand what is happening, thus easier to spot mistakes, even for those that are not language lawyers

Strings and formatting

  • Be careful of LogPrint versus LogPrintf. LogPrint takes a category argument, LogPrintf does not.

    • Rationale: Confusion of these can result in runtime exceptions due to formatting mismatch, and it is easy to get wrong because of subtly similar naming
  • Use std::string, avoid C string manipulation functions

    • Rationale: C++ string handling is marginally safer, less scope for buffer overflows and surprises with \0 characters. Also some C string manipulations tend to act differently depending on platform, or even the user locale
  • Use ParseInt32, ParseInt64, ParseUInt32, ParseUInt64, ParseDouble from utilstrencodings.h for number parsing

    • Rationale: These functions do overflow checking, and avoid pesky locale issues
  • For strprintf, LogPrint, LogPrintf formatting characters don’t need size specifiers

    • Rationale: Bitcoin Core uses tinyformat, which is type safe. Leave them out to avoid confusion

Variable names

Although the shadowing warning (-Wshadow) is not enabled by default (it prevents issues rising from using a different variable with the same name), please name variables so that their names do not shadow variables defined in the source code.

E.g. in member initializers, prepend _ to the argument name shadowing the member name:

class AddressBookPage
    Mode mode;

AddressBookPage::AddressBookPage(Mode _mode) :

When using nested cycles, do not name the inner cycle variable the same as in upper cycle etc.

Threads and synchronization

  • Build and run tests with -DDEBUG_LOCKORDER to verify that no potential deadlocks are introduced. As of 0.12, this is defined by default when configuring with --enable-debug

  • When using LOCK/TRY_LOCK be aware that the lock exists in the context of the current scope, so surround the statement and the code that needs the lock with braces


    TRY_LOCK(cs_vNodes, lockNodes);


TRY_LOCK(cs_vNodes, lockNodes);

Source code organization

  • Implementation code should go into the .cpp file and not the .h, unless necessary due to template usage or when performance due to inlining is critical

    • Rationale: Shorter and simpler header files are easier to read, and reduce compile time
  • Every .cpp and .h file should #include every header file it directly uses classes, functions or other definitions from, even if those headers are already included indirectly through other headers. One exception is that a .cpp file does not need to re-include the includes already included in its corresponding .h file.

    • Rationale: Excluding headers because they are already indirectly included results in compilation failures when those indirect dependencies change. Furthermore, it obscures what the real code dependencies are.
  • Don’t import anything into the global namespace (using namespace ...). Use fully specified types such as std::string.

    • Rationale: Avoids symbol conflicts
  • Terminate namespaces with a comment (// namespace mynamespace). The comment should be placed on the same line as the brace closing the namespace, e.g.

namespace mynamespace {
} // namespace mynamespace

namespace {
} // namespace
  • Rationale: Avoids confusion about the namespace context

  • Prefer #include <primitives/transaction.h> bracket syntax instead of #include "primitives/transactions.h" quote syntax when possible.

    • Rationale: Bracket syntax is less ambiguous because the preprocessor searches a fixed list of include directories without taking location of the source file into account. This allows quoted includes to stand out more when the location of the source file actually is relevant.


  • Do not display or manipulate dialogs in model code (classes *Model)

    • Rationale: Model classes pass through events and data from the core, they should not interact with the user. That’s where View classes come in. The converse also holds: try to not directly access core data structures from Views.


Several parts of the repository are subtrees of software maintained elsewhere.

Some of these are maintained by active developers of Bitcoin Core, in which case changes should probably go directly upstream without being PRed directly against the project. They will be merged back in the next subtree merge.

Others are external projects without a tight relationship with our project. Changes to these should also be sent upstream but bugfixes may also be prudent to PR against Bitcoin Core so that they can be integrated quickly. Cosmetic changes should be purely taken upstream.

There is a tool in contrib/devtools/ to check a subtree directory for consistency with its upstream repository.

Current subtrees include:

Git and GitHub tips

  • For resolving merge/rebase conflicts, it can be useful to enable diff3 style using git config merge.conflictstyle diff3. Instead of


you will see


This may make it much clearer what caused the conflict. In this style, you can often just look at what changed between original and theirs, and mechanically apply that to yours (or the other way around).

  • When reviewing patches which change indentation in C++ files, use git diff -w and git show -w. This makes the diff algorithm ignore whitespace changes. This feature is also available on, by adding ?w=1 at the end of any URL which shows a diff.

  • When reviewing patches that change symbol names in many places, use git diff --word-diff. This will instead of showing the patch as deleted/added lines, show deleted/added words.

  • When reviewing patches that move code around, try using git diff --patience commit~:old/file.cpp commit:new/file/name.cpp, and ignoring everything except the moved body of code which should show up as neither + or - lines. In case it was not a pure move, this may even work when combined with the -w or --word-diff options described above.

  • When looking at other’s pull requests, it may make sense to add the following section to your .git/config file:

    [remote "upstream-pull"]
            fetch = +refs/pull/*:refs/remotes/upstream-pull/*
            url =

This will add an upstream-pull remote to your git repository, which can be fetched using git fetch --all or git fetch upstream-pull. Afterwards, you can use upstream-pull/NUMBER/head in arguments to git show, git checkout and anywhere a commit id would be acceptable to see the changes from pull request NUMBER.

Scripted diffs

For reformatting and refactoring commits where the changes can be easily automated using a bash script, we use scripted-diff commits. The bash script is included in the commit message and our Travis CI job checks that the result of the script is identical to the commit. This aids reviewers since they can verify that the script does exactly what it’s supposed to do. It is also helpful for rebasing (since the same script can just be re-run on the new master commit).

To create a scripted-diff:

  • start the commit message with scripted-diff: (and then a description of the diff on the same line)
  • in the commit message include the bash script between lines containing just the following text:

The scripted-diff is verified by the tool contrib/devtools/

Commit bb81e173 is an example of a scripted-diff.

RPC interface guidelines

A few guidelines for introducing and reviewing new RPC interfaces:

  • Method naming: use consecutive lower-case names such as getrawtransaction and submitblock

    • Rationale: Consistency with existing interface.
  • Argument naming: use snake case fee_delta (and not, e.g. camel case feeDelta)

    • Rationale: Consistency with existing interface.
  • Use the JSON parser for parsing, don’t manually parse integers or strings from arguments unless absolutely necessary.

    • Rationale: Introduces hand-rolled string manipulation code at both the caller and callee sites, which is error prone, and it is easy to get things such as escaping wrong. JSON already supports nested data structures, no need to re-invent the wheel.

    • Exception: AmountFromValue can parse amounts as string. This was introduced because many JSON parsers and formatters hard-code handling decimal numbers as floating point values, resulting in potential loss of precision. This is unacceptable for monetary values. Always use AmountFromValue and ValueFromAmount when inputting or outputting monetary values. The only exceptions to this are prioritisetransaction and getblocktemplate because their interface is specified as-is in BIP22.

  • Missing arguments and ‘null’ should be treated the same: as default values. If there is no default value, both cases should fail in the same way. The easiest way to follow this guideline is detect unspecified arguments with params[x].isNull() instead of params.size() <= x. The former returns true if the argument is either null or missing, while the latter returns true if is missing, and false if it is null.

    • Rationale: Avoids surprises when switching to name-based arguments. Missing name-based arguments are passed as ‘null’.
  • Try not to overload methods on argument type. E.g. don’t make getblock(true) and getblock("hash") do different things.

    • Rationale: This is impossible to use with bitcoin-cli, and can be surprising to users.

    • Exception: Some RPC calls can take both an int and bool, most notably when a bool was switched to a multi-value, or due to other historical reasons. Always have false map to 0 and true to 1 in this case.

  • Don’t forget to fill in the argument names correctly in the RPC command table.

    • Rationale: If not, the call can not be used with name-based arguments.
  • Set okSafeMode in the RPC command table to a sensible value: safe mode is when the blockchain is regarded to be in a confused state, and the client deems it unsafe to do anything irreversible such as send. Anything that just queries should be permitted.

    • Rationale: Troubleshooting a node in safe mode is difficult if half the RPCs don’t work.
  • Add every non-string RPC argument (method, idx, name) to the table vRPCConvertParams in rpc/client.cpp.

    • Rationale: bitcoin-cli and the GUI debug console use this table to determine how to convert a plaintext command line to JSON. If the types don’t match, the method can be unusable from there.
  • A RPC method must either be a wallet method or a non-wallet method. Do not introduce new methods such as signrawtransaction that differ in behavior based on presence of a wallet.

    • Rationale: as well as complicating the implementation and interfering with the introduction of multi-wallet, wallet and non-wallet code should be separated to avoid introducing circular dependencies between code units.
  • Try to make the RPC response a JSON object.

    • Rationale: If a RPC response is not a JSON object then it is harder to avoid API breakage if new data in the response is needed.
  • Wallet RPCs call BlockUntilSyncedToCurrentChain to maintain consistency with getblockchaininfo’s state immediately prior to the call’s execution. Wallet RPCs whose behavior does not depend on the current chainstate may omit this call.

    • Rationale: In previous versions of Bitcoin Core, the wallet was always in-sync with the chainstate (by virtue of them all being updated in the same cs_main lock). In order to maintain the behavior that wallet RPCs return results as of at least the highest best-known block an RPC client may be aware of prior to entering a wallet RPC call, we must block until the wallet is caught up to the chainstate as of the RPC call’s entry. This also makes the API much easier for RPC clients to reason about.