* Use static_assert where appropriate
* Use std::min and std::max where appropriate
* Simplify RippleD error reporting
* Remove use of beast::RandomAccessFile
Beast includes a lot of code for encapsulating cross-platform differences
which are not used or needed by rippled. Additionally, a lot of that code
implements functionality that is available from the standard library.
This moves away from custom implementations of features that the standard
library provides and reduces the number of platform-specific interfaces
andfeatures that Beast makes available.
Highlights include:
* Use std:: instead of beast implementations when possible
* Reduce the use of beast::String in public interfaces
* Remove Windows-specific COM and Registry code
* Reduce the public interface of beast::File
* Reduce the public interface of beast::SystemStats
* Remove unused sysctl/getsysinfo functions
* Remove beast::Logger
This is a cleanup to the structure of the sources.
* Rename to ServerHandler
* Move private implementation declaration to separate header
* De-inline function definitions in the class declaration.
Many classes required to support type-erasure of handlers and boost::asio
types are now obsolete, so these classes and files are removed:
HTTPClientType, FixedInputBuffer, PeerRole, socket_wrapper,
client_session, basic_url, abstract_socket, buffer_sequence, memory_buffer,
enable_wait_for_async, shared_handler, wrap_handler, streambuf,
ContentBodyBuffer, SSLContext, completion-handler based handshake detectors.
These structural changes are made:
* Some missing includes added to headers
* asio module directory flattened
* Removed MultiSocket. Code that previously used the MultiSocket now uses
a combination of boost::asio coroutines and CRTP.
* Sitefiles headers rolled up and directory flattened.
* Disabled Sitefiles use of deprecated HTTPClient.
* Validators headers tidied up.
* Disabled Validators use of deprecated HTTPClient.
On Application exit, Overlay was calling PeerImp::close for each peer.
The implementation of PeerImp::close only canceled all pending I/O and did not
call functions necessary for proper transition of Peer state during socket
closure. The correct transition is ensured by calling PeerImp::detach. This
changes PeerImp::close to call PeerImp::detach instead, ensuring that Overlay
invariants are maintained. Specifically, that reference counts for pending I/O
on peers will be correctly unwound by canceling operations and that the Peer
object will be destroyed, thus allowing the Overlay to stop correctly.
This configuration section uses the new BasicConfig interface that supports
key-value pairs in the section. Some exposition is added to the example cfg
file. The new settings for overlay are related to the Hub and Spoke feature
which is currently in development. Production servers should not set
these configuration options, they are clearly marked experimental in the
example cfg file.
Conflicts:
src/ripple/overlay/impl/OverlayImpl.cpp
src/ripple/overlay/impl/OverlayImpl.h
src/ripple/overlay/impl/PeerImp.cpp
src/ripple/overlay/impl/PeerImp.h
The MultiSocket is obsolete technology which is superceded by a more
straightforward, template based implementation that is compatible with
boost::asio::coroutines. This removes support for the unused PROXY handshake
feature. After this change a large number of classes and source files may be
removed.
When JSON-RPC and Websocket responses are calculated, the result is stored
in intermediate Json::Value objects and later composed in a single linear
memory buffer before being sent to the socket. These classes support a
new model for building responses that supports incremental construction
of JSON replies in constant time and removes the requirement that all
data returned be located in continuguous memory.
* New JsonWriter incrementally writes JSON with O(1) granularity and memory.
* Array, Object are RAII wrappers for the O(1) JsonWriter.
This class was used to allow stream style operator<< to write to the
HTTP::Session. This is being superceded by a more robust object-based model
that supports coroutines.
This change to BasicConfig stores all appended lines which are not key/value
pairs in a separate values vector which can be retrieved later. This is to
support sections containing both key/value pairs and a list of values.
This changes the HTTP parser interface to return an error_code instead
of a bool. This eliminates the need for the error() member function and
simplifies calling code.
This works around the limitation that 1.56 boost::asio::ssl::stream objects
do not support r-value move or construction. It is required when the stream
does not own the socket.
If beast::Time::currentTimeMillis is first called from a coroutine launched
using boost::asio::spawn, Win32 throws an exception. This workaround calls
getCurrentTime once in main to prevent the exception.
Reference:
https://svn.boost.org/trac/boost/ticket/10657
The MultiSocket class implements a socket that handshakes in multiple
protocols including SSL and PROXY. Unfortunately the way it type-erases the
handlers and buffers is incompatible with boost::asio coroutines. To pave the
way for coroutines this is part of a larger set of changes that roll back the
usage of MultiSocket to older code, and some custom implementations that use
templates. The custom implementations are more simple since they use
coroutines. Removing MultiSocket will make many other classes and source files
unused, a big win for trimming down the codebase size.
Empirical evidence shows a database access pattern with few hits
and many misses (objects that don't exist). This changes the timing
tests so they more accurately reflect rippled's actual usage:
* Add read missing keys test
* Increase numObjectsToTest to 1,000,000
* Alter PredictableObjectFactory to seed RNG once only
* Make NodeStoreTiming a manual test
If we receive a deferred transaction from a server in our
cluster, treat it as if it wasn't received from a server
in our cluster.
This currently has no effect but is needed for server to
interoperate with future code that will relay deferred
transactions.
The implementation of multi-sign has a SigningAccounts array as a
member of the outermost object. This array could not be parsed
by the previous implementation of STParsedJSON, which only knew
how to parse objects. This refactor supports the required parsing.
The refactor divides the parsing into three separate functions:
o parseNoRecurse() which parses most rippled data types.
o parseObject() which parses object types that may contain
arbitrary other types.
o parseArray() which parses object types that may contain
arbitrary other types.
The change is required by the multi-sign implementation, but is
independent. So the parsing change is going in as a separate
commit.
The parsing is still far from perfect. But this was as much as
needs doing to accomplish the ends and mitigate risk of breaking
the parser.
This reworks the way SHAMaps are stored, flushed, backed, and
traversed. Rather than storing the linkages in the SHAMap itself,
that information is now stored in the nodes. This makes
snapshotting much cheaper and also allows traverse work done on
behalf of one SHAMap to be used by other SHAMaps that share inner
nodes with that SHAMap.
When a SHAMap is modified, nodes are modified all the way up to the
root. This means that the modified nodes in a SHAMap can easily be
traversed for flushing. So they don't need to be separately tracked.
Summary
* Remove mTNByID
* Remove mDirtyNodes
* Much faster traverses
* Much Faster snapshots
* New algorithm for flushing
* New vistNodes/visitLeaves
* Avoid I/O if a map is unbacked
* Remove all calls to setlocale to ensure that the global
locale is always C.
* Also replace beast::SystemStats::getNumCpus() with
std:🧵:hardware_concurrency()
OverlayImpl::onStart calls into PeerFinder before PeerFinder::Manager::onStart,
causing tests to sometimes fail and the application to intermittently not start.
The order of calls to Stoppable::onStart is implementation defined and not
predictable.
This changes PeerFinder to load the database in Stoppable::onPrepare, before
threads are launched. In general, creation and initialization of resources that
are shared between classes should happen in onPrepare rather than onStart,
to solve this problem.
Previously, the PeerFinder manager constructed with a Callback object
provided by the owner which was used to perform operations like connecting,
disconnecting, and sending messages. This made it difficult to change the
overlay code because a single call into the PeerFinder could cause both
OverlayImpl and PeerImp to be re-entered one or more times, sometimes while
holding a recursive mutex. This change eliminates the callback by changing
PeerFinder functions to return values indicating the action the caller should
take.
As a result of this change the PeerFinder no longer needs its own dedicated
thread. OverlayImpl is changed to call into PeerFinder on a timer to perform
periodic activities. Furthermore the Checker class used to perform connectivity
checks has been refactored. It no longer uses an abstract base class, in order
to not type-erase the handler passed to async_connect (ensuring compatibility
with coroutines). To allow unit tests that don't need a network, the Logic
class is now templated on the Checker type. Currently the Manager provides its
own io_service. However, this can easily be changed so that the io_service is
provided upon construction.
Summary
* Remove unused SiteFiles dependency injection
* Remove Callback and update signatures for public APIs
* Remove obsolete functions
* Move timer to overlay
* Steps toward a shared io_service
* Templated, simplified Checker
* Tidy up Checker declaration
