This commit modernizes the `AcceptedLedger` and `AcceptedLedgerTx`
classes, reduces their memory footprint and reduces unnecessary
dynamic memory allocations.
This commit optimizes the way asynchronous nodestore operations are
processed both by reducing the amount of time locks are held and by
minimizing the number of memory allocations and data copying.
Adds support to TaggedCache to support smart replacement
(Needed to avoid race conditions with negative caching.)
Create a "hotDUMMY" object that represents the absence
of an object.
Allow DatabaseNodeImp::asyncFetch to complete immediately
if object is in cache (positive or negative).
Fix a bug in asyncFetch where the object returned may not
be the correct canonical version because we stash the
object in the results array before we canonicalize it.
When fetching ledgers, the existing code would isolate the peer
that sent the most useful responses and issue follow up queries
only to that peer.
This commit increases the query aggressiveness, and changes the
mechanism used to select which peers to issue follow-up queries
to so as to more evenly spread the load along those peers which
provided useful responses.
The `SHAMapInnerNode` class had a global mutex to protect the
array of node children. Profiling suggested that around 4% of
all attempts to lock the global would block.
This commit removes that global mutex, and replaces it with a
new per-node 16-way spinlock (implemented so as not to effect
the size of an inner node objet), effectively eliminating the
lock contention.
* Txs with the same fee level will sort by TxID XORed with the parent
ledger hash.
* The TxQ is re-sorted after every ledger.
* Attempt to future-proof the TxQ tie breaking test
* Abort background path finding when closed or disconnected
* Exit pathfinding job thread if there are no requests left
* Don't bother creating the path find job if there are no requests
* Refactor to remove circular dependency between InfoSub and PathRequest
The existing trust line caching code was suboptimal in that it stored
redundant information, pinned SLEs into memory and required multiple
memory allocations per cached object.
This commit eliminates redundant data, reducing the size of cached
objects and unpinning SLEs from memory, and uses value_types to
avoid the need for `std::shared_ptr`. As a result of these changes, the
effective size of a cached object, includes the overhead of the memory
allocator and the `std::shared_ptr` should be reduced by at least 64
bytes. This is significant, as there can easily be tens of millions
of these objects.
This commit combines the `apply_mutex` and `read_mutex` into a single `mutex_`
var. This new `mutex_` var is a `shared_mutex`, and most operations only need to
lock it with a `shared_lock`. The only exception is `applyMutex`, which may need
a `unique_lock`.
One consequence of removing the `apply_mutex` is more than one `applyMutex`
function can run at the same time. To help reduce lock contention that a
`unique_lock` would cause, checks that only require reading data are run a
`shared_lock` (call these the "prewriteChecks"), then the lock is released, then
a `unique_lock` is acquired. Since a currently running `applyManifest` may write
data between the time a `shared_lock` is released and the `write_lock` is
acquired, the "prewriteChecks" need to be rerun. Duplicating this work isn't
ideal, but the "prewirteChecks" are relatively inexpensive.
A couple of other designs were considered. We could restrict more than one
`applyMutex` function from running concurrently - either with a `applyMutex` or
my setting the max number of manifest jobs on the job queue to one. The biggest
issue with this is if any other function ever adds a write lock for any reason,
`applyManifest` would not be broken - data could be written between the release
of the `shared_lock` and the acquisition of the `unique_lock`. Note: it is
tempting to solve this problem by not releasing the `shared_mutex` and simply
upgrading the lock. In the presence of concurrently running `applyManifest`
functions, this will deadlock (both function need to wait for the other to
release their read locks before they can acquire a write lock).
In order to preserve the Hooks ABI, it is important that field
values used for hooks be stable going forward.
This commit reserves the required codes so that they will not
be repurposed before Hooks can be proposed for inclusion in
the codebase.
* Remove Application & Database dependency in PerfLog. Replace it with
a callback passed into the constructor.
* Fixes the circular dependency between ripple/nodestore and ripple/basics
If fast loading is enabled but the last persisted ledger is not
entirely on disk, the server would fail to start without manual
intervention by the server operator.
This commit allows the server to detect this scenario and attempt
to automatically recover.
This is a refactor aimed at cleaning up and simplifying the existing
job queue.
As of now, all jobs are cancelled at the same time and in the same
way, so this commit removes the per-job cancellation token. If the
need for such support is demonstrated, support can be re-added.
* Revise documentation for ClosureCounter and Workers.
* Simplify code, removing unnecessary function arguments and
deduplicating expressions
* Restructure job handlers to no longer need to pass a job's
handle to the job.
