Branch data Line data Source code
1 : : // Copyright (c) The Bitcoin Core developers
2 : : // Distributed under the MIT software license, see the accompanying
3 : : // file COPYING or http://www.opensource.org/licenses/mit-license.php.
4 : :
5 : : #include <txgraph.h>
6 : :
7 : : #include <cluster_linearize.h>
8 : : #include <random.h>
9 : : #include <util/bitset.h>
10 : : #include <util/check.h>
11 : : #include <util/feefrac.h>
12 : : #include <util/vector.h>
13 : :
14 : : #include <compare>
15 : : #include <functional>
16 : : #include <memory>
17 : : #include <set>
18 : : #include <span>
19 : : #include <unordered_set>
20 : : #include <utility>
21 : :
22 : : namespace {
23 : :
24 : : using namespace cluster_linearize;
25 : :
26 : : /** The maximum number of levels a TxGraph can have (0 = main, 1 = staging). */
27 : : static constexpr int MAX_LEVELS{2};
28 : :
29 : : // Forward declare the TxGraph implementation class.
30 : : class TxGraphImpl;
31 : :
32 : : /** Position of a DepGraphIndex within a Cluster::m_linearization. */
33 : : using LinearizationIndex = uint32_t;
34 : : /** Position of a Cluster within TxGraphImpl::ClusterSet::m_clusters. */
35 : : using ClusterSetIndex = uint32_t;
36 : :
37 : : /** Quality levels for cached cluster linearizations. */
38 : : enum class QualityLevel
39 : : {
40 : : /** This is a singleton cluster consisting of a transaction that individually exceeds the
41 : : * cluster size limit. It cannot be merged with anything. */
42 : : OVERSIZED_SINGLETON,
43 : : /** This cluster may have multiple disconnected components, which are all NEEDS_RELINEARIZE. */
44 : : NEEDS_SPLIT,
45 : : /** This cluster may have multiple disconnected components, which are all ACCEPTABLE. */
46 : : NEEDS_SPLIT_ACCEPTABLE,
47 : : /** This cluster has undergone changes that warrant re-linearization. */
48 : : NEEDS_RELINEARIZE,
49 : : /** The minimal level of linearization has been performed, but it is not known to be optimal. */
50 : : ACCEPTABLE,
51 : : /** The linearization is known to be optimal. */
52 : : OPTIMAL,
53 : : /** This cluster is not registered in any ClusterSet::m_clusters.
54 : : * This must be the last entry in QualityLevel as ClusterSet::m_clusters is sized using it. */
55 : : NONE,
56 : : };
57 : :
58 : : /** Information about a transaction inside TxGraphImpl::Trim. */
59 : 64217 : struct TrimTxData
60 : : {
61 : : // Fields populated by Cluster::AppendTrimData(). These are immutable after TrimTxData
62 : : // construction.
63 : : /** Chunk feerate for this transaction. */
64 : : FeePerWeight m_chunk_feerate;
65 : : /** GraphIndex of the transaction. */
66 : : TxGraph::GraphIndex m_index;
67 : : /** Size of the transaction. */
68 : : uint32_t m_tx_size;
69 : :
70 : : // Fields only used internally by TxGraphImpl::Trim():
71 : : /** Number of unmet dependencies this transaction has. -1 if the transaction is included. */
72 : : uint32_t m_deps_left;
73 : : /** Number of dependencies that apply to this transaction as child. */
74 : : uint32_t m_parent_count;
75 : : /** Where in deps_by_child those dependencies begin. */
76 : : uint32_t m_parent_offset;
77 : : /** Number of dependencies that apply to this transaction as parent. */
78 : : uint32_t m_children_count;
79 : : /** Where in deps_by_parent those dependencies begin. */
80 : : uint32_t m_children_offset;
81 : :
82 : : // Fields only used internally by TxGraphImpl::Trim()'s union-find implementation, and only for
83 : : // transactions that are definitely included or definitely rejected.
84 : : //
85 : : // As transactions get processed, they get organized into trees which form partitions
86 : : // representing the would-be clusters up to that point. The root of each tree is a
87 : : // representative for that partition. See
88 : : // https://en.wikipedia.org/wiki/Disjoint-set_data_structure.
89 : : //
90 : : /** Pointer to another TrimTxData, towards the root of the tree. If this is a root, m_uf_parent
91 : : * is equal to this itself. */
92 : : TrimTxData* m_uf_parent;
93 : : /** If this is a root, the total number of transactions in the partition. */
94 : : uint32_t m_uf_count;
95 : : /** If this is a root, the total size of transactions in the partition. */
96 : : uint64_t m_uf_size;
97 : : };
98 : :
99 : : /** A grouping of connected transactions inside a TxGraphImpl::ClusterSet. */
100 : : class Cluster
101 : : {
102 : : friend class TxGraphImpl;
103 : : friend class BlockBuilderImpl;
104 : :
105 : : protected:
106 : : using GraphIndex = TxGraph::GraphIndex;
107 : : using SetType = BitSet<MAX_CLUSTER_COUNT_LIMIT>;
108 : : /** The quality level of m_linearization. */
109 : : QualityLevel m_quality{QualityLevel::NONE};
110 : : /** Which position this Cluster has in TxGraphImpl::ClusterSet::m_clusters[m_quality]. */
111 : : ClusterSetIndex m_setindex{ClusterSetIndex(-1)};
112 : : /** Sequence number for this Cluster (for tie-breaking comparison between equal-chunk-feerate
113 : : transactions in distinct clusters). */
114 : : uint64_t m_sequence;
115 : :
116 : 64311 : explicit Cluster(uint64_t sequence) noexcept : m_sequence(sequence) {}
117 : :
118 : : public:
119 : : // Provide virtual destructor, for safe polymorphic usage inside std::unique_ptr.
120 : 0 : virtual ~Cluster() = default;
121 : :
122 : : // Cannot move or copy (would invalidate Cluster* in Locator and ClusterSet). */
123 : : Cluster(const Cluster&) = delete;
124 : : Cluster& operator=(const Cluster&) = delete;
125 : : Cluster(Cluster&&) = delete;
126 : : Cluster& operator=(Cluster&&) = delete;
127 : :
128 : : // Generic helper functions.
129 : :
130 : : /** Whether the linearization of this Cluster can be exposed. */
131 : 257122 : bool IsAcceptable(bool after_split = false) const noexcept
132 : : {
133 [ # # ]: 0 : return m_quality == QualityLevel::ACCEPTABLE || m_quality == QualityLevel::OPTIMAL ||
134 [ # # # # ]: 0 : (after_split && m_quality == QualityLevel::NEEDS_SPLIT_ACCEPTABLE);
135 : : }
136 : : /** Whether the linearization of this Cluster is optimal. */
137 : 192822 : bool IsOptimal() const noexcept
138 : : {
139 : 192822 : return m_quality == QualityLevel::OPTIMAL;
140 : : }
141 : : /** Whether this cluster is oversized. Note that no changes that can cause oversizedness are
142 : : * ever applied, so the only way a materialized Cluster object can be oversized is by being
143 : : * an individually oversized transaction singleton. */
144 : 385644 : bool IsOversized() const noexcept { return m_quality == QualityLevel::OVERSIZED_SINGLETON; }
145 : : /** Whether this cluster requires splitting. */
146 : 192833 : bool NeedsSplitting() const noexcept
147 : : {
148 : 192833 : return m_quality == QualityLevel::NEEDS_SPLIT ||
149 : : m_quality == QualityLevel::NEEDS_SPLIT_ACCEPTABLE;
150 : : }
151 : :
152 : : /** Get the smallest number of transactions this Cluster is intended for. */
153 : : virtual DepGraphIndex GetMinIntendedTxCount() const noexcept = 0;
154 : : /** Get the maximum number of transactions this Cluster supports. */
155 : : virtual DepGraphIndex GetMaxTxCount() const noexcept = 0;
156 : : /** Total memory usage currently for this Cluster, including all its dynamic memory, plus Cluster
157 : : * structure itself, and ClusterSet::m_clusters entry. */
158 : : virtual size_t TotalMemoryUsage() const noexcept = 0;
159 : : /** Determine the range of DepGraphIndexes used by this Cluster. */
160 : : virtual DepGraphIndex GetDepGraphIndexRange() const noexcept = 0;
161 : : /** Get the number of transactions in this Cluster. */
162 : : virtual LinearizationIndex GetTxCount() const noexcept = 0;
163 : : /** Get the total size of the transactions in this Cluster. */
164 : : virtual uint64_t GetTotalTxSize() const noexcept = 0;
165 : : /** Given a DepGraphIndex into this Cluster, find the corresponding GraphIndex. */
166 : : virtual GraphIndex GetClusterEntry(DepGraphIndex index) const noexcept = 0;
167 : : /** Append a transaction with given GraphIndex at the end of this Cluster and its
168 : : * linearization. Return the DepGraphIndex it was placed at. */
169 : : virtual DepGraphIndex AppendTransaction(GraphIndex graph_idx, FeePerWeight feerate) noexcept = 0;
170 : : /** Add dependencies to a given child in this cluster. */
171 : : virtual void AddDependencies(SetType parents, DepGraphIndex child) noexcept = 0;
172 : : /** Invoke visitor_fn for each transaction in the cluster, in linearization order, then wipe this Cluster. */
173 : : virtual void ExtractTransactions(const std::function<void (DepGraphIndex, GraphIndex, FeePerWeight, SetType)>& visit_fn) noexcept = 0;
174 : : /** Figure out what level this Cluster exists at in the graph. In most cases this is known by
175 : : * the caller already (see all "int level" arguments below), but not always. */
176 : : virtual int GetLevel(const TxGraphImpl& graph) const noexcept = 0;
177 : : /** Only called by TxGraphImpl::SwapIndexes. */
178 : : virtual void UpdateMapping(DepGraphIndex cluster_idx, GraphIndex graph_idx) noexcept = 0;
179 : : /** Push changes to Cluster and its linearization to the TxGraphImpl Entry objects. */
180 : : virtual void Updated(TxGraphImpl& graph, int level) noexcept = 0;
181 : : /** Create a copy of this Cluster in staging, returning a pointer to it (used by PullIn). */
182 : : virtual Cluster* CopyToStaging(TxGraphImpl& graph) const noexcept = 0;
183 : : /** Get the list of Clusters in main that conflict with this one (which is assumed to be in staging). */
184 : : virtual void GetConflicts(const TxGraphImpl& graph, std::vector<Cluster*>& out) const noexcept = 0;
185 : : /** Mark all the Entry objects belonging to this staging Cluster as missing. The Cluster must be
186 : : * deleted immediately after. */
187 : : virtual void MakeStagingTransactionsMissing(TxGraphImpl& graph) noexcept = 0;
188 : : /** Remove all transactions from a (non-empty) Cluster. */
189 : : virtual void Clear(TxGraphImpl& graph, int level) noexcept = 0;
190 : : /** Change a Cluster's level from 1 (staging) to 0 (main). */
191 : : virtual void MoveToMain(TxGraphImpl& graph) noexcept = 0;
192 : : /** Minimize this Cluster's memory usage. */
193 : : virtual void Compact() noexcept = 0;
194 : :
195 : : // Functions that implement the Cluster-specific side of internal TxGraphImpl mutations.
196 : :
197 : : /** Apply all removals from the front of to_remove that apply to this Cluster, popping them
198 : : * off. There must be at least one such entry. */
199 : : virtual void ApplyRemovals(TxGraphImpl& graph, int level, std::span<GraphIndex>& to_remove) noexcept = 0;
200 : : /** Split this cluster (must have a NEEDS_SPLIT* quality). Returns whether to delete this
201 : : * Cluster afterwards. */
202 : : [[nodiscard]] virtual bool Split(TxGraphImpl& graph, int level) noexcept = 0;
203 : : /** Move all transactions from cluster to *this (as separate components). */
204 : : virtual void Merge(TxGraphImpl& graph, int level, Cluster& cluster) noexcept = 0;
205 : : /** Given a span of (parent, child) pairs that all belong to this Cluster, apply them. */
206 : : virtual void ApplyDependencies(TxGraphImpl& graph, int level, std::span<std::pair<GraphIndex, GraphIndex>> to_apply) noexcept = 0;
207 : : /** Improve the linearization of this Cluster. Returns how much work was performed and whether
208 : : * the Cluster's QualityLevel improved as a result. */
209 : : virtual std::pair<uint64_t, bool> Relinearize(TxGraphImpl& graph, int level, uint64_t max_iters) noexcept = 0;
210 : : /** For every chunk in the cluster, append its FeeFrac to ret. */
211 : : virtual void AppendChunkFeerates(std::vector<FeeFrac>& ret) const noexcept = 0;
212 : : /** Add a TrimTxData entry (filling m_chunk_feerate, m_index, m_tx_size) for every
213 : : * transaction in the Cluster to ret. Implicit dependencies between consecutive transactions
214 : : * in the linearization are added to deps. Return the Cluster's total transaction size. */
215 : : virtual uint64_t AppendTrimData(std::vector<TrimTxData>& ret, std::vector<std::pair<GraphIndex, GraphIndex>>& deps) const noexcept = 0;
216 : :
217 : : // Functions that implement the Cluster-specific side of public TxGraph functions.
218 : :
219 : : /** Process elements from the front of args that apply to this cluster, and append Refs for the
220 : : * union of their ancestors to output. */
221 : : virtual void GetAncestorRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept = 0;
222 : : /** Process elements from the front of args that apply to this cluster, and append Refs for the
223 : : * union of their descendants to output. */
224 : : virtual void GetDescendantRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept = 0;
225 : : /** Populate range with refs for the transactions in this Cluster's linearization, from
226 : : * position start_pos until start_pos+range.size()-1, inclusive. Returns whether that
227 : : * range includes the last transaction in the linearization. */
228 : : virtual bool GetClusterRefs(TxGraphImpl& graph, std::span<TxGraph::Ref*> range, LinearizationIndex start_pos) noexcept = 0;
229 : : /** Get the individual transaction feerate of a Cluster element. */
230 : : virtual FeePerWeight GetIndividualFeerate(DepGraphIndex idx) noexcept = 0;
231 : : /** Modify the fee of a Cluster element. */
232 : : virtual void SetFee(TxGraphImpl& graph, int level, DepGraphIndex idx, int64_t fee) noexcept = 0;
233 : :
234 : : // Debugging functions.
235 : :
236 : : virtual void SanityCheck(const TxGraphImpl& graph, int level) const = 0;
237 : : };
238 : :
239 : : /** An implementation of Cluster that uses a DepGraph and vectors, to support arbitrary numbers of
240 : : * transactions up to MAX_CLUSTER_COUNT_LIMIT. */
241 : : class GenericClusterImpl final : public Cluster
242 : : {
243 : : friend class TxGraphImpl;
244 : : /** The DepGraph for this cluster, holding all feerates, and ancestors/descendants. */
245 : : DepGraph<SetType> m_depgraph;
246 : : /** m_mapping[i] gives the GraphIndex for the position i transaction in m_depgraph. Values for
247 : : * positions i that do not exist in m_depgraph shouldn't ever be accessed and thus don't
248 : : * matter. m_mapping.size() equals m_depgraph.PositionRange(). */
249 : : std::vector<GraphIndex> m_mapping;
250 : : /** The current linearization of the cluster. m_linearization.size() equals
251 : : * m_depgraph.TxCount(). This is always kept topological. */
252 : : std::vector<DepGraphIndex> m_linearization;
253 : :
254 : : public:
255 : : /** The smallest number of transactions this Cluster implementation is intended for. */
256 : : static constexpr DepGraphIndex MIN_INTENDED_TX_COUNT{2};
257 : : /** The largest number of transactions this Cluster implementation supports. */
258 : : static constexpr DepGraphIndex MAX_TX_COUNT{SetType::Size()};
259 : :
260 : : GenericClusterImpl() noexcept = delete;
261 : : /** Construct an empty GenericClusterImpl. */
262 : : explicit GenericClusterImpl(uint64_t sequence) noexcept;
263 : :
264 : : size_t TotalMemoryUsage() const noexcept final;
265 : 0 : constexpr DepGraphIndex GetMinIntendedTxCount() const noexcept final { return MIN_INTENDED_TX_COUNT; }
266 : 0 : constexpr DepGraphIndex GetMaxTxCount() const noexcept final { return MAX_TX_COUNT; }
267 [ # # ]: 0 : DepGraphIndex GetDepGraphIndexRange() const noexcept final { return m_depgraph.PositionRange(); }
268 [ # # ]: 0 : LinearizationIndex GetTxCount() const noexcept final { return m_linearization.size(); }
269 : : uint64_t GetTotalTxSize() const noexcept final;
270 : 0 : GraphIndex GetClusterEntry(DepGraphIndex index) const noexcept final { return m_mapping[index]; }
271 : : DepGraphIndex AppendTransaction(GraphIndex graph_idx, FeePerWeight feerate) noexcept final;
272 : : void AddDependencies(SetType parents, DepGraphIndex child) noexcept final;
273 : : void ExtractTransactions(const std::function<void (DepGraphIndex, GraphIndex, FeePerWeight, SetType)>& visit_fn) noexcept final;
274 : : int GetLevel(const TxGraphImpl& graph) const noexcept final;
275 : 0 : void UpdateMapping(DepGraphIndex cluster_idx, GraphIndex graph_idx) noexcept final { m_mapping[cluster_idx] = graph_idx; }
276 : : void Updated(TxGraphImpl& graph, int level) noexcept final;
277 : : Cluster* CopyToStaging(TxGraphImpl& graph) const noexcept final;
278 : : void GetConflicts(const TxGraphImpl& graph, std::vector<Cluster*>& out) const noexcept final;
279 : : void MakeStagingTransactionsMissing(TxGraphImpl& graph) noexcept final;
280 : : void Clear(TxGraphImpl& graph, int level) noexcept final;
281 : : void MoveToMain(TxGraphImpl& graph) noexcept final;
282 : : void Compact() noexcept final;
283 : : void ApplyRemovals(TxGraphImpl& graph, int level, std::span<GraphIndex>& to_remove) noexcept final;
284 : : [[nodiscard]] bool Split(TxGraphImpl& graph, int level) noexcept final;
285 : : void Merge(TxGraphImpl& graph, int level, Cluster& cluster) noexcept final;
286 : : void ApplyDependencies(TxGraphImpl& graph, int level, std::span<std::pair<GraphIndex, GraphIndex>> to_apply) noexcept final;
287 : : std::pair<uint64_t, bool> Relinearize(TxGraphImpl& graph, int level, uint64_t max_iters) noexcept final;
288 : : void AppendChunkFeerates(std::vector<FeeFrac>& ret) const noexcept final;
289 : : uint64_t AppendTrimData(std::vector<TrimTxData>& ret, std::vector<std::pair<GraphIndex, GraphIndex>>& deps) const noexcept final;
290 : : void GetAncestorRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept final;
291 : : void GetDescendantRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept final;
292 : : bool GetClusterRefs(TxGraphImpl& graph, std::span<TxGraph::Ref*> range, LinearizationIndex start_pos) noexcept final;
293 : : FeePerWeight GetIndividualFeerate(DepGraphIndex idx) noexcept final;
294 : : void SetFee(TxGraphImpl& graph, int level, DepGraphIndex idx, int64_t fee) noexcept final;
295 : : void SanityCheck(const TxGraphImpl& graph, int level) const final;
296 : : };
297 : :
298 : : /** An implementation of Cluster that only supports 1 transaction. */
299 : 0 : class SingletonClusterImpl final : public Cluster
300 : : {
301 : : friend class TxGraphImpl;
302 : :
303 : : /** The feerate of the (singular) transaction in this Cluster. */
304 : : FeePerWeight m_feerate;
305 : : /** Constant to indicate that this Cluster is empty. */
306 : : static constexpr auto NO_GRAPH_INDEX = GraphIndex(-1);
307 : : /** The GraphIndex of the transaction. NO_GRAPH_INDEX if this Cluster is empty. */
308 : : GraphIndex m_graph_index = NO_GRAPH_INDEX;
309 : :
310 : : public:
311 : : /** The smallest number of transactions this Cluster implementation is intended for. */
312 : : static constexpr DepGraphIndex MIN_INTENDED_TX_COUNT{1};
313 : : /** The largest number of transactions this Cluster implementation supports. */
314 : : static constexpr DepGraphIndex MAX_TX_COUNT{1};
315 : :
316 : : SingletonClusterImpl() noexcept = delete;
317 : : /** Construct an empty SingletonClusterImpl. */
318 : 64311 : explicit SingletonClusterImpl(uint64_t sequence) noexcept : Cluster(sequence) {}
319 : :
320 : : size_t TotalMemoryUsage() const noexcept final;
321 : 192811 : constexpr DepGraphIndex GetMinIntendedTxCount() const noexcept final { return MIN_INTENDED_TX_COUNT; }
322 : 192822 : constexpr DepGraphIndex GetMaxTxCount() const noexcept final { return MAX_TX_COUNT; }
323 : 1927516 : LinearizationIndex GetTxCount() const noexcept final { return m_graph_index != NO_GRAPH_INDEX; }
324 : 0 : DepGraphIndex GetDepGraphIndexRange() const noexcept final { return GetTxCount(); }
325 [ + + ]: 321027 : uint64_t GetTotalTxSize() const noexcept final { return GetTxCount() ? m_feerate.size : 0; }
326 : 192817 : GraphIndex GetClusterEntry(DepGraphIndex index) const noexcept final { Assume(index == 0); Assume(GetTxCount()); return m_graph_index; }
327 : : DepGraphIndex AppendTransaction(GraphIndex graph_idx, FeePerWeight feerate) noexcept final;
328 : : void AddDependencies(SetType parents, DepGraphIndex child) noexcept final;
329 : : void ExtractTransactions(const std::function<void (DepGraphIndex, GraphIndex, FeePerWeight, SetType)>& visit_fn) noexcept final;
330 : : int GetLevel(const TxGraphImpl& graph) const noexcept final;
331 : 1 : void UpdateMapping(DepGraphIndex cluster_idx, GraphIndex graph_idx) noexcept final { Assume(cluster_idx == 0); m_graph_index = graph_idx; }
332 : : void Updated(TxGraphImpl& graph, int level) noexcept final;
333 : : Cluster* CopyToStaging(TxGraphImpl& graph) const noexcept final;
334 : : void GetConflicts(const TxGraphImpl& graph, std::vector<Cluster*>& out) const noexcept final;
335 : : void MakeStagingTransactionsMissing(TxGraphImpl& graph) noexcept final;
336 : : void Clear(TxGraphImpl& graph, int level) noexcept final;
337 : : void MoveToMain(TxGraphImpl& graph) noexcept final;
338 : : void Compact() noexcept final;
339 : : void ApplyRemovals(TxGraphImpl& graph, int level, std::span<GraphIndex>& to_remove) noexcept final;
340 : : [[nodiscard]] bool Split(TxGraphImpl& graph, int level) noexcept final;
341 : : void Merge(TxGraphImpl& graph, int level, Cluster& cluster) noexcept final;
342 : : void ApplyDependencies(TxGraphImpl& graph, int level, std::span<std::pair<GraphIndex, GraphIndex>> to_apply) noexcept final;
343 : : std::pair<uint64_t, bool> Relinearize(TxGraphImpl& graph, int level, uint64_t max_iters) noexcept final;
344 : : void AppendChunkFeerates(std::vector<FeeFrac>& ret) const noexcept final;
345 : : uint64_t AppendTrimData(std::vector<TrimTxData>& ret, std::vector<std::pair<GraphIndex, GraphIndex>>& deps) const noexcept final;
346 : : void GetAncestorRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept final;
347 : : void GetDescendantRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept final;
348 : : bool GetClusterRefs(TxGraphImpl& graph, std::span<TxGraph::Ref*> range, LinearizationIndex start_pos) noexcept final;
349 : : FeePerWeight GetIndividualFeerate(DepGraphIndex idx) noexcept final;
350 : : void SetFee(TxGraphImpl& graph, int level, DepGraphIndex idx, int64_t fee) noexcept final;
351 : : void SanityCheck(const TxGraphImpl& graph, int level) const final;
352 : : };
353 : :
354 : : /** The transaction graph, including staged changes.
355 : : *
356 : : * The overall design of the data structure consists of 3 interlinked representations:
357 : : * - The transactions (held as a vector of TxGraphImpl::Entry inside TxGraphImpl).
358 : : * - The clusters (Cluster objects in per-quality vectors inside TxGraphImpl::ClusterSet).
359 : : * - The Refs (TxGraph::Ref objects, held externally by users of the TxGraph class)
360 : : *
361 : : * The Clusters are kept in one or two ClusterSet objects, one for the "main" graph, and one for
362 : : * the proposed changes ("staging"). If a transaction occurs in both, they share the same Entry,
363 : : * but there will be a separate Cluster per graph.
364 : : *
365 : : * Clusters and Refs contain the index of the Entry objects they refer to, and the Entry objects
366 : : * refer back to the Clusters and Refs the corresponding transaction is contained in.
367 : : *
368 : : * While redundant, this permits moving all of them independently, without invalidating things
369 : : * or costly iteration to fix up everything:
370 : : * - Entry objects can be moved to fill holes left by removed transactions in the Entry vector
371 : : * (see TxGraphImpl::Compact).
372 : : * - Clusters can be rewritten continuously (removals can cause them to split, new dependencies
373 : : * can cause them to be merged).
374 : : * - Ref objects can be held outside the class, while permitting them to be moved around, and
375 : : * inherited from.
376 : : */
377 : : class TxGraphImpl final : public TxGraph
378 : : {
379 : : friend class Cluster;
380 : : friend class SingletonClusterImpl;
381 : : friend class GenericClusterImpl;
382 : : friend class BlockBuilderImpl;
383 : : private:
384 : : /** Internal RNG. */
385 : : FastRandomContext m_rng;
386 : : /** This TxGraphImpl's maximum cluster count limit. */
387 : : const DepGraphIndex m_max_cluster_count;
388 : : /** This TxGraphImpl's maximum cluster size limit. */
389 : : const uint64_t m_max_cluster_size;
390 : : /** The number of linearization improvement steps needed per cluster to be considered
391 : : * acceptable. */
392 : : const uint64_t m_acceptable_iters;
393 : :
394 : : /** Information about one group of Clusters to be merged. */
395 : : struct GroupEntry
396 : : {
397 : : /** Where the clusters to be merged start in m_group_clusters. */
398 : : uint32_t m_cluster_offset;
399 : : /** How many clusters to merge. */
400 : : uint32_t m_cluster_count;
401 : : /** Where the dependencies for this cluster group in m_deps_to_add start. */
402 : : uint32_t m_deps_offset;
403 : : /** How many dependencies to add. */
404 : : uint32_t m_deps_count;
405 : : };
406 : :
407 : : /** Information about all groups of Clusters to be merged. */
408 : 14 : struct GroupData
409 : : {
410 : : /** The groups of Clusters to be merged. */
411 : : std::vector<GroupEntry> m_groups;
412 : : /** Which clusters are to be merged. GroupEntry::m_cluster_offset indexes into this. */
413 : : std::vector<Cluster*> m_group_clusters;
414 : : };
415 : :
416 : : /** The collection of all Clusters in main or staged. */
417 : : struct ClusterSet
418 : : {
419 : : /** The vectors of clusters, one vector per quality level. ClusterSetIndex indexes into each. */
420 : : std::array<std::vector<std::unique_ptr<Cluster>>, int(QualityLevel::NONE)> m_clusters;
421 : : /** Which removals have yet to be applied. */
422 : : std::vector<GraphIndex> m_to_remove;
423 : : /** Which dependencies are to be added ((parent,child) pairs). GroupData::m_deps_offset indexes
424 : : * into this. */
425 : : std::vector<std::pair<GraphIndex, GraphIndex>> m_deps_to_add;
426 : : /** Information about the merges to be performed, if known. */
427 : : std::optional<GroupData> m_group_data = GroupData{};
428 : : /** Which entries were removed in this ClusterSet (so they can be wiped on abort). This
429 : : * includes all entries which have an (R) removed locator at this level (staging only),
430 : : * plus optionally any transaction in m_unlinked. */
431 : : std::vector<GraphIndex> m_removed;
432 : : /** Total number of transactions in this graph (sum of all transaction counts in all
433 : : * Clusters, and for staging also those inherited from the main ClusterSet). */
434 : : GraphIndex m_txcount{0};
435 : : /** Total number of individually oversized transactions in the graph. */
436 : : GraphIndex m_txcount_oversized{0};
437 : : /** Whether this graph is oversized (if known). */
438 : : std::optional<bool> m_oversized{false};
439 : : /** The combined TotalMemoryUsage of all clusters in this level (only Clusters that
440 : : * are materialized; in staging, implicit Clusters from main are not counted), */
441 : : size_t m_cluster_usage{0};
442 : :
443 : 4 : ClusterSet() noexcept = default;
444 : : };
445 : :
446 : : /** The main ClusterSet. */
447 : : ClusterSet m_main_clusterset;
448 : : /** The staging ClusterSet, if any. */
449 : : std::optional<ClusterSet> m_staging_clusterset;
450 : : /** Next sequence number to assign to created Clusters. */
451 : : uint64_t m_next_sequence_counter{0};
452 : :
453 : : /** Information about a chunk in the main graph. */
454 : : struct ChunkData
455 : : {
456 : : /** The Entry which is the last transaction of the chunk. */
457 : : mutable GraphIndex m_graph_index;
458 : : /** How many transactions the chunk contains (-1 = singleton tail of cluster). */
459 : : LinearizationIndex m_chunk_count;
460 : :
461 : 64305 : ChunkData(GraphIndex graph_index, LinearizationIndex chunk_count) noexcept :
462 : 64305 : m_graph_index{graph_index}, m_chunk_count{chunk_count} {}
463 : : };
464 : :
465 : : /** Compare two Cluster* by their m_sequence value (while supporting nullptr). */
466 : 82 : static std::strong_ordering CompareClusters(Cluster* a, Cluster* b) noexcept
467 : : {
468 : : // The nullptr pointer compares before everything else.
469 [ - + ]: 82 : if (a == nullptr || b == nullptr) {
470 [ # # # # ]: 0 : return (a != nullptr) <=> (b != nullptr);
471 : : }
472 : : // If neither pointer is nullptr, compare the Clusters' sequence numbers.
473 : 82 : Assume(a == b || a->m_sequence != b->m_sequence);
474 [ + - + + ]: 82 : return a->m_sequence <=> b->m_sequence;
475 : : }
476 : :
477 : : /** Compare two entries (which must both exist within the main graph). */
478 : 1081681 : std::strong_ordering CompareMainTransactions(GraphIndex a, GraphIndex b) const noexcept
479 : : {
480 [ - + + - ]: 1081681 : Assume(a < m_entries.size() && b < m_entries.size());
481 [ + + ]: 1081681 : const auto& entry_a = m_entries[a];
482 : 1081681 : const auto& entry_b = m_entries[b];
483 : : // Compare chunk feerates, and return result if it differs.
484 : 1081681 : auto feerate_cmp = FeeRateCompare(entry_b.m_main_chunk_feerate, entry_a.m_main_chunk_feerate);
485 [ + + ]: 1081681 : if (feerate_cmp < 0) return std::strong_ordering::less;
486 [ + + ]: 512314 : if (feerate_cmp > 0) return std::strong_ordering::greater;
487 : : // Compare Cluster m_sequence as tie-break for equal chunk feerates.
488 : 52 : const auto& locator_a = entry_a.m_locator[0];
489 : 52 : const auto& locator_b = entry_b.m_locator[0];
490 : 104 : Assume(locator_a.IsPresent() && locator_b.IsPresent());
491 [ + - ]: 52 : if (locator_a.cluster != locator_b.cluster) {
492 : 52 : return CompareClusters(locator_a.cluster, locator_b.cluster);
493 : : }
494 : : // As final tie-break, compare position within cluster linearization.
495 [ # # # # ]: 0 : return entry_a.m_main_lin_index <=> entry_b.m_main_lin_index;
496 : : }
497 : :
498 : : /** Comparator for ChunkData objects in mining order. */
499 : : class ChunkOrder
500 : : {
501 : : const TxGraphImpl* const m_graph;
502 : : public:
503 : 4 : explicit ChunkOrder(const TxGraphImpl* graph) : m_graph(graph) {}
504 : :
505 : 1081681 : bool operator()(const ChunkData& a, const ChunkData& b) const noexcept
506 : : {
507 : 1081681 : return m_graph->CompareMainTransactions(a.m_graph_index, b.m_graph_index) < 0;
508 : : }
509 : : };
510 : :
511 : : /** Definition for the mining index type. */
512 : : using ChunkIndex = std::set<ChunkData, ChunkOrder>;
513 : :
514 : : /** Index of ChunkData objects, indexing the last transaction in each chunk in the main
515 : : * graph. */
516 : : ChunkIndex m_main_chunkindex;
517 : : /** Number of index-observing objects in existence (BlockBuilderImpls). */
518 : : size_t m_main_chunkindex_observers{0};
519 : : /** Cache of discarded ChunkIndex node handles to reuse, avoiding additional allocation. */
520 : : std::vector<ChunkIndex::node_type> m_main_chunkindex_discarded;
521 : :
522 : : /** A Locator that describes whether, where, and in which Cluster an Entry appears.
523 : : * Every Entry has MAX_LEVELS locators, as it may appear in one Cluster per level.
524 : : *
525 : : * Each level of a Locator is in one of three states:
526 : : *
527 : : * - (P)resent: actually occurs in a Cluster at that level.
528 : : *
529 : : * - (M)issing:
530 : : * - In the main graph: the transaction does not exist in main.
531 : : * - In the staging graph: the transaction's existence is the same as in main. If it doesn't
532 : : * exist in main, (M) in staging means it does not exist there
533 : : * either. If it does exist in main, (M) in staging means the
534 : : * cluster it is in has not been modified in staging, and thus the
535 : : * transaction implicitly exists in staging too (without explicit
536 : : * Cluster object; see PullIn() to create it in staging too).
537 : : *
538 : : * - (R)emoved: only possible in staging; it means the transaction exists in main, but is
539 : : * removed in staging.
540 : : *
541 : : * The following combinations are possible:
542 : : * - (M,M): the transaction doesn't exist in either graph.
543 : : * - (P,M): the transaction exists in both, but only exists explicitly in a Cluster object in
544 : : * main. Its existence in staging is inherited from main.
545 : : * - (P,P): the transaction exists in both, and is materialized in both. Thus, the clusters
546 : : * and/or their linearizations may be different in main and staging.
547 : : * - (M,P): the transaction is added in staging, and does not exist in main.
548 : : * - (P,R): the transaction exists in main, but is removed in staging.
549 : : *
550 : : * When staging does not exist, only (M,M) and (P,M) are possible.
551 : : */
552 : : struct Locator
553 : : {
554 : : /** Which Cluster the Entry appears in (nullptr = missing). */
555 : : Cluster* cluster{nullptr};
556 : : /** Where in the Cluster it appears (if cluster == nullptr: 0 = missing, -1 = removed). */
557 : : DepGraphIndex index{0};
558 : :
559 : : /** Mark this Locator as missing (= same as lower level, or non-existing if level 0). */
560 : 0 : void SetMissing() noexcept { cluster = nullptr; index = 0; }
561 : : /** Mark this Locator as removed (not allowed in level 0). */
562 : 0 : void SetRemoved() noexcept { cluster = nullptr; index = DepGraphIndex(-1); }
563 : : /** Mark this Locator as present, in the specified Cluster. */
564 : 64311 : void SetPresent(Cluster* c, DepGraphIndex i) noexcept { cluster = c; index = i; }
565 : : /** Check if this Locator is missing. */
566 [ - + + - ]: 192841 : bool IsMissing() const noexcept { return cluster == nullptr && index == 0; }
567 : : /** Check if this Locator is removed. */
568 [ + - - - : 192833 : bool IsRemoved() const noexcept { return cluster == nullptr && index == DepGraphIndex(-1); }
- - ]
569 : : /** Check if this Locator is present (in some Cluster). */
570 [ - - + - ]: 52 : bool IsPresent() const noexcept { return cluster != nullptr; }
571 : : };
572 : :
573 : : /** Internal information about each transaction in a TxGraphImpl. */
574 : 64311 : struct Entry
575 : : {
576 : : /** Pointer to the corresponding Ref object if any, or nullptr if unlinked. */
577 : : Ref* m_ref{nullptr};
578 : : /** Iterator to the corresponding ChunkData, if any, and m_main_chunkindex.end() otherwise.
579 : : * This is initialized on construction of the Entry, in AddTransaction. */
580 : : ChunkIndex::iterator m_main_chunkindex_iterator;
581 : : /** Which Cluster and position therein this Entry appears in. ([0] = main, [1] = staged). */
582 : : Locator m_locator[MAX_LEVELS];
583 : : /** The chunk feerate of this transaction in main (if present in m_locator[0]). */
584 : : FeePerWeight m_main_chunk_feerate;
585 : : /** The position this transaction has in the main linearization (if present). */
586 : : LinearizationIndex m_main_lin_index;
587 : : };
588 : :
589 : : /** The set of all transactions (in all levels combined). GraphIndex values index into this. */
590 : : std::vector<Entry> m_entries;
591 : :
592 : : /** Set of Entries which have no linked Ref anymore. */
593 : : std::vector<GraphIndex> m_unlinked;
594 : :
595 : : public:
596 : : /** Construct a new TxGraphImpl with the specified limits. */
597 : 4 : explicit TxGraphImpl(DepGraphIndex max_cluster_count, uint64_t max_cluster_size, uint64_t acceptable_iters) noexcept :
598 : 4 : m_max_cluster_count(max_cluster_count),
599 : 4 : m_max_cluster_size(max_cluster_size),
600 : 4 : m_acceptable_iters(acceptable_iters),
601 : 4 : m_main_chunkindex(ChunkOrder(this))
602 : : {
603 : 4 : Assume(max_cluster_count >= 1);
604 : 4 : Assume(max_cluster_count <= MAX_CLUSTER_COUNT_LIMIT);
605 : 4 : }
606 : :
607 : : /** Destructor. */
608 : : ~TxGraphImpl() noexcept;
609 : :
610 : : // Cannot move or copy (would invalidate TxGraphImpl* in Ref, MiningOrder, EvictionOrder).
611 : : TxGraphImpl(const TxGraphImpl&) = delete;
612 : : TxGraphImpl& operator=(const TxGraphImpl&) = delete;
613 : : TxGraphImpl(TxGraphImpl&&) = delete;
614 : : TxGraphImpl& operator=(TxGraphImpl&&) = delete;
615 : :
616 : : // Simple helper functions.
617 : :
618 : : /** Swap the Entry referred to by a and the one referred to by b. */
619 : : void SwapIndexes(GraphIndex a, GraphIndex b) noexcept;
620 : : /** If idx exists in the specified level ClusterSet (explicitly, or in the level below and not
621 : : * removed), return the Cluster it is in. Otherwise, return nullptr. */
622 : 638402 : Cluster* FindCluster(GraphIndex idx, int level) const noexcept { return FindClusterAndLevel(idx, level).first; }
623 : : /** Like FindCluster, but also return what level the match was found in (-1 if not found). */
624 : : std::pair<Cluster*, int> FindClusterAndLevel(GraphIndex idx, int level) const noexcept;
625 : : /** Extract a Cluster from its ClusterSet, and set its quality to QualityLevel::NONE. */
626 : : std::unique_ptr<Cluster> ExtractCluster(int level, QualityLevel quality, ClusterSetIndex setindex) noexcept;
627 : : /** Delete a Cluster. */
628 : : void DeleteCluster(Cluster& cluster, int level) noexcept;
629 : : /** Insert a Cluster into its ClusterSet. */
630 : : ClusterSetIndex InsertCluster(int level, std::unique_ptr<Cluster>&& cluster, QualityLevel quality) noexcept;
631 : : /** Change the QualityLevel of a Cluster (identified by old_quality and old_index). */
632 : : void SetClusterQuality(int level, QualityLevel old_quality, ClusterSetIndex old_index, QualityLevel new_quality) noexcept;
633 : : /** Get the index of the top level ClusterSet (staging if it exists, main otherwise). */
634 [ - + ]: 893158 : int GetTopLevel() const noexcept { return m_staging_clusterset.has_value(); }
635 : : /** Get the specified level (staging if it exists and level is TOP, main otherwise). */
636 [ - - + - : 316 : int GetSpecifiedLevel(Level level) const noexcept { return level == Level::TOP && m_staging_clusterset.has_value(); }
+ - - - -
- - - - -
- - + - ]
637 : : /** Get a reference to the ClusterSet at the specified level (which must exist). */
638 : : ClusterSet& GetClusterSet(int level) noexcept;
639 : : const ClusterSet& GetClusterSet(int level) const noexcept;
640 : : /** Make a transaction not exist at a specified level. It must currently exist there.
641 : : * oversized_tx indicates whether the transaction is an individually-oversized one
642 : : * (OVERSIZED_SINGLETON). */
643 : : void ClearLocator(int level, GraphIndex index, bool oversized_tx) noexcept;
644 : : /** Find which Clusters in main conflict with ones in staging. */
645 : : std::vector<Cluster*> GetConflicts() const noexcept;
646 : : /** Clear an Entry's ChunkData. */
647 : : void ClearChunkData(Entry& entry) noexcept;
648 : : /** Give an Entry a ChunkData object. */
649 : : void CreateChunkData(GraphIndex idx, LinearizationIndex chunk_count) noexcept;
650 : : /** Create an empty GenericClusterImpl object. */
651 : 0 : std::unique_ptr<GenericClusterImpl> CreateEmptyGenericCluster() noexcept
652 : : {
653 : 0 : return std::make_unique<GenericClusterImpl>(m_next_sequence_counter++);
654 : : }
655 : : /** Create an empty SingletonClusterImpl object. */
656 : 64311 : std::unique_ptr<SingletonClusterImpl> CreateEmptySingletonCluster() noexcept
657 : : {
658 : 64311 : return std::make_unique<SingletonClusterImpl>(m_next_sequence_counter++);
659 : : }
660 : : /** Create an empty Cluster of the appropriate implementation for the specified (maximum) tx
661 : : * count. */
662 : 64311 : std::unique_ptr<Cluster> CreateEmptyCluster(DepGraphIndex tx_count) noexcept
663 : : {
664 [ + - ]: 64311 : if (tx_count >= SingletonClusterImpl::MIN_INTENDED_TX_COUNT && tx_count <= SingletonClusterImpl::MAX_TX_COUNT) {
665 : 64311 : return CreateEmptySingletonCluster();
666 : : }
667 [ # # ]: 0 : if (tx_count >= GenericClusterImpl::MIN_INTENDED_TX_COUNT && tx_count <= GenericClusterImpl::MAX_TX_COUNT) {
668 [ # # ]: 0 : return CreateEmptyGenericCluster();
669 : : }
670 : 0 : assert(false);
671 : : return {};
672 : : }
673 : :
674 : : // Functions for handling Refs.
675 : :
676 : : /** Only called by Ref's move constructor/assignment to update Ref locations. */
677 : 129861 : void UpdateRef(GraphIndex idx, Ref& new_location) noexcept final
678 : : {
679 : 129861 : auto& entry = m_entries[idx];
680 : 129861 : Assume(entry.m_ref != nullptr);
681 : 129861 : entry.m_ref = &new_location;
682 : 129861 : }
683 : :
684 : : /** Only called by Ref::~Ref to unlink Refs, and Ref's move assignment. */
685 : 300 : void UnlinkRef(GraphIndex idx) noexcept final
686 : : {
687 [ - + ]: 300 : auto& entry = m_entries[idx];
688 : 300 : Assume(entry.m_ref != nullptr);
689 [ - + ]: 300 : Assume(m_main_chunkindex_observers == 0 || !entry.m_locator[0].IsPresent());
690 : 300 : entry.m_ref = nullptr;
691 : : // Mark the transaction as to be removed in all levels where it explicitly or implicitly
692 : : // exists.
693 : 300 : bool exists_anywhere{false};
694 : 300 : bool exists{false};
695 [ + + ]: 600 : for (int level = 0; level <= GetTopLevel(); ++level) {
696 [ + + ]: 300 : if (entry.m_locator[level].IsPresent()) {
697 : : exists_anywhere = true;
698 : : exists = true;
699 [ + - ]: 8 : } else if (entry.m_locator[level].IsRemoved()) {
700 : : exists = false;
701 : : }
702 [ - + ]: 8 : if (exists) {
703 : 292 : auto& clusterset = GetClusterSet(level);
704 : 292 : clusterset.m_to_remove.push_back(idx);
705 : : // Force recomputation of grouping data.
706 [ - + ]: 292 : clusterset.m_group_data = std::nullopt;
707 : : // Do not wipe the oversized state of main if staging exists. The reason for this
708 : : // is that the alternative would mean that cluster merges may need to be applied to
709 : : // a formerly-oversized main graph while staging exists (to satisfy chunk feerate
710 : : // queries into main, for example), and such merges could conflict with pulls of
711 : : // some of their constituents into staging.
712 [ + - + - ]: 592 : if (level == GetTopLevel() && clusterset.m_oversized == true) {
713 : 0 : clusterset.m_oversized = std::nullopt;
714 : : }
715 : : }
716 : : }
717 : 300 : m_unlinked.push_back(idx);
718 [ + + ]: 300 : if (!exists_anywhere) Compact();
719 : 300 : }
720 : :
721 : : // Functions related to various normalization/application steps.
722 : : /** Get rid of unlinked Entry objects in m_entries, if possible (this changes the GraphIndex
723 : : * values for remaining Entry objects, so this only does something when no to-be-applied
724 : : * operations or staged removals referring to GraphIndexes remain). */
725 : : void Compact() noexcept;
726 : : /** If cluster is not in staging, copy it there, and return a pointer to it.
727 : : * Staging must exist, and this modifies the locators of its
728 : : * transactions from inherited (P,M) to explicit (P,P). */
729 : : Cluster* PullIn(Cluster* cluster, int level) noexcept;
730 : : /** Apply all removals queued up in m_to_remove to the relevant Clusters (which get a
731 : : * NEEDS_SPLIT* QualityLevel) up to the specified level. */
732 : : void ApplyRemovals(int up_to_level) noexcept;
733 : : /** Split an individual cluster. */
734 : : void Split(Cluster& cluster, int level) noexcept;
735 : : /** Split all clusters that need splitting up to the specified level. */
736 : : void SplitAll(int up_to_level) noexcept;
737 : : /** Populate m_group_data based on m_deps_to_add in the specified level. */
738 : : void GroupClusters(int level) noexcept;
739 : : /** Merge the specified clusters. */
740 : : void Merge(std::span<Cluster*> to_merge, int level) noexcept;
741 : : /** Apply all m_deps_to_add to the relevant Clusters in the specified level. */
742 : : void ApplyDependencies(int level) noexcept;
743 : : /** Make a specified Cluster have quality ACCEPTABLE or OPTIMAL. */
744 : : void MakeAcceptable(Cluster& cluster, int level) noexcept;
745 : : /** Make all Clusters at the specified level have quality ACCEPTABLE or OPTIMAL. */
746 : : void MakeAllAcceptable(int level) noexcept;
747 : :
748 : : // Implementations for the public TxGraph interface.
749 : :
750 : : Ref AddTransaction(const FeePerWeight& feerate) noexcept final;
751 : : void RemoveTransaction(const Ref& arg) noexcept final;
752 : : void AddDependency(const Ref& parent, const Ref& child) noexcept final;
753 : : void SetTransactionFee(const Ref&, int64_t fee) noexcept final;
754 : :
755 : : bool DoWork(uint64_t iters) noexcept final;
756 : :
757 : : void StartStaging() noexcept final;
758 : : void CommitStaging() noexcept final;
759 : : void AbortStaging() noexcept final;
760 : 0 : bool HaveStaging() const noexcept final { return m_staging_clusterset.has_value(); }
761 : :
762 : : bool Exists(const Ref& arg, Level level) noexcept final;
763 : : FeePerWeight GetMainChunkFeerate(const Ref& arg) noexcept final;
764 : : FeePerWeight GetIndividualFeerate(const Ref& arg) noexcept final;
765 : : std::vector<Ref*> GetCluster(const Ref& arg, Level level) noexcept final;
766 : : std::vector<Ref*> GetAncestors(const Ref& arg, Level level) noexcept final;
767 : : std::vector<Ref*> GetDescendants(const Ref& arg, Level level) noexcept final;
768 : : std::vector<Ref*> GetAncestorsUnion(std::span<const Ref* const> args, Level level) noexcept final;
769 : : std::vector<Ref*> GetDescendantsUnion(std::span<const Ref* const> args, Level level) noexcept final;
770 : : GraphIndex GetTransactionCount(Level level) noexcept final;
771 : : bool IsOversized(Level level) noexcept final;
772 : : std::strong_ordering CompareMainOrder(const Ref& a, const Ref& b) noexcept final;
773 : : GraphIndex CountDistinctClusters(std::span<const Ref* const> refs, Level level) noexcept final;
774 : : std::pair<std::vector<FeeFrac>, std::vector<FeeFrac>> GetMainStagingDiagrams() noexcept final;
775 : : std::vector<Ref*> Trim() noexcept final;
776 : :
777 : : std::unique_ptr<BlockBuilder> GetBlockBuilder() noexcept final;
778 : : std::pair<std::vector<Ref*>, FeePerWeight> GetWorstMainChunk() noexcept final;
779 : :
780 : : size_t GetMainMemoryUsage() noexcept final;
781 : :
782 : : void SanityCheck() const final;
783 : : };
784 : :
785 : 193538 : TxGraphImpl::ClusterSet& TxGraphImpl::GetClusterSet(int level) noexcept
786 : : {
787 [ + - ]: 193538 : if (level == 0) return m_main_clusterset;
788 : 0 : Assume(level == 1);
789 : 0 : Assume(m_staging_clusterset.has_value());
790 : 0 : return *m_staging_clusterset;
791 : : }
792 : :
793 : 11 : const TxGraphImpl::ClusterSet& TxGraphImpl::GetClusterSet(int level) const noexcept
794 : : {
795 [ + - ]: 11 : if (level == 0) return m_main_clusterset;
796 : 0 : Assume(level == 1);
797 : 0 : Assume(m_staging_clusterset.has_value());
798 : 0 : return *m_staging_clusterset;
799 : : }
800 : :
801 : : /** Implementation of the TxGraph::BlockBuilder interface. */
802 : : class BlockBuilderImpl final : public TxGraph::BlockBuilder
803 : : {
804 : : /** Which TxGraphImpl this object is doing block building for. It will have its
805 : : * m_main_chunkindex_observers incremented as long as this BlockBuilderImpl exists. */
806 : : TxGraphImpl* const m_graph;
807 : : /** Cluster sequence numbers which we're not including further transactions from. */
808 : : std::unordered_set<uint64_t> m_excluded_clusters;
809 : : /** Iterator to the current chunk in the chunk index. end() if nothing further remains. */
810 : : TxGraphImpl::ChunkIndex::const_iterator m_cur_iter;
811 : : /** Which cluster the current chunk belongs to, so we can exclude further transactions from it
812 : : * when that chunk is skipped. */
813 : : Cluster* m_cur_cluster;
814 : : /** Whether we know that m_cur_iter points to the last chunk of m_cur_cluster. */
815 : : bool m_known_end_of_cluster;
816 : :
817 : : // Move m_cur_iter / m_cur_cluster to the next acceptable chunk.
818 : : void Next() noexcept;
819 : :
820 : : public:
821 : : /** Construct a new BlockBuilderImpl to build blocks for the provided graph. */
822 : : BlockBuilderImpl(TxGraphImpl& graph) noexcept;
823 : :
824 : : // Implement the public interface.
825 : : ~BlockBuilderImpl() final;
826 : : std::optional<std::pair<std::vector<TxGraph::Ref*>, FeePerWeight>> GetCurrentChunk() noexcept final;
827 : : void Include() noexcept final;
828 : : void Skip() noexcept final;
829 : : };
830 : :
831 : 64330 : void TxGraphImpl::ClearChunkData(Entry& entry) noexcept
832 : : {
833 [ + + ]: 64330 : if (entry.m_main_chunkindex_iterator != m_main_chunkindex.end()) {
834 : 13 : Assume(m_main_chunkindex_observers == 0);
835 : : // If the Entry has a non-empty m_main_chunkindex_iterator, extract it, and move the handle
836 : : // to the cache of discarded chunkindex entries.
837 : 13 : m_main_chunkindex_discarded.emplace_back(m_main_chunkindex.extract(entry.m_main_chunkindex_iterator));
838 : 13 : entry.m_main_chunkindex_iterator = m_main_chunkindex.end();
839 : : }
840 : 64330 : }
841 : :
842 : 64305 : void TxGraphImpl::CreateChunkData(GraphIndex idx, LinearizationIndex chunk_count) noexcept
843 : : {
844 [ - + ]: 64305 : auto& entry = m_entries[idx];
845 [ - + ]: 64305 : if (!m_main_chunkindex_discarded.empty()) {
846 : : // Reuse an discarded node handle.
847 : 0 : auto& node = m_main_chunkindex_discarded.back().value();
848 : 0 : node.m_graph_index = idx;
849 : 0 : node.m_chunk_count = chunk_count;
850 : 0 : auto insert_result = m_main_chunkindex.insert(std::move(m_main_chunkindex_discarded.back()));
851 : 0 : Assume(insert_result.inserted);
852 : 0 : entry.m_main_chunkindex_iterator = insert_result.position;
853 : 0 : m_main_chunkindex_discarded.pop_back();
854 : 0 : } else {
855 : : // Construct a new entry.
856 : 64305 : auto emplace_result = m_main_chunkindex.emplace(idx, chunk_count);
857 : 64305 : Assume(emplace_result.second);
858 : 64305 : entry.m_main_chunkindex_iterator = emplace_result.first;
859 : : }
860 : 64305 : }
861 : :
862 : 0 : size_t GenericClusterImpl::TotalMemoryUsage() const noexcept
863 : : {
864 : 0 : return // Dynamic memory allocated in this Cluster.
865 [ # # # # ]: 0 : memusage::DynamicUsage(m_mapping) + memusage::DynamicUsage(m_linearization) +
866 : : // Dynamic memory usage inside m_depgraph.
867 [ # # ]: 0 : m_depgraph.DynamicMemoryUsage() +
868 : : // Memory usage of the allocated Cluster itself.
869 : 0 : memusage::MallocUsage(sizeof(GenericClusterImpl)) +
870 : : // Memory usage of the ClusterSet::m_clusters entry.
871 : 0 : sizeof(std::unique_ptr<Cluster>);
872 : : }
873 : :
874 : 257133 : size_t SingletonClusterImpl::TotalMemoryUsage() const noexcept
875 : : {
876 : 257133 : return // Memory usage of the allocated SingletonClusterImpl itself.
877 : 257133 : memusage::MallocUsage(sizeof(SingletonClusterImpl)) +
878 : : // Memory usage of the ClusterSet::m_clusters entry.
879 : 257133 : sizeof(std::unique_ptr<Cluster>);
880 : : }
881 : :
882 : 0 : uint64_t GenericClusterImpl::GetTotalTxSize() const noexcept
883 : : {
884 : 0 : uint64_t ret{0};
885 [ # # ]: 0 : for (auto i : m_linearization) {
886 : 0 : ret += m_depgraph.FeeRate(i).size;
887 : : }
888 : 0 : return ret;
889 : : }
890 : :
891 : 0 : DepGraphIndex GenericClusterImpl::AppendTransaction(GraphIndex graph_idx, FeePerWeight feerate) noexcept
892 : : {
893 : 0 : Assume(graph_idx != GraphIndex(-1));
894 : 0 : auto ret = m_depgraph.AddTransaction(feerate);
895 : 0 : m_mapping.push_back(graph_idx);
896 : 0 : m_linearization.push_back(ret);
897 : 0 : return ret;
898 : : }
899 : :
900 : 64311 : DepGraphIndex SingletonClusterImpl::AppendTransaction(GraphIndex graph_idx, FeePerWeight feerate) noexcept
901 : : {
902 : 64311 : Assume(!GetTxCount());
903 : 64311 : m_graph_index = graph_idx;
904 : 64311 : m_feerate = feerate;
905 : 64311 : return 0;
906 : : }
907 : :
908 : 0 : void GenericClusterImpl::AddDependencies(SetType parents, DepGraphIndex child) noexcept
909 : : {
910 : 0 : m_depgraph.AddDependencies(parents, child);
911 : 0 : }
912 : :
913 : 0 : void SingletonClusterImpl::AddDependencies(SetType parents, DepGraphIndex child) noexcept
914 : : {
915 : : // Singletons cannot have any dependencies.
916 : 0 : Assume(child == 0);
917 [ # # ]: 0 : Assume(parents == SetType{} || parents == SetType::Fill(0));
918 : 0 : }
919 : :
920 : 0 : void GenericClusterImpl::ExtractTransactions(const std::function<void (DepGraphIndex, GraphIndex, FeePerWeight, SetType)>& visit_fn) noexcept
921 : : {
922 [ # # ]: 0 : for (auto pos : m_linearization) {
923 : 0 : visit_fn(pos, m_mapping[pos], FeePerWeight::FromFeeFrac(m_depgraph.FeeRate(pos)), m_depgraph.GetReducedParents(pos));
924 : : }
925 : : // Purge this Cluster, now that everything has been moved.
926 : 0 : m_depgraph = DepGraph<SetType>{};
927 [ # # ]: 0 : m_linearization.clear();
928 [ # # ]: 0 : m_mapping.clear();
929 : 0 : }
930 : :
931 : 0 : void SingletonClusterImpl::ExtractTransactions(const std::function<void (DepGraphIndex, GraphIndex, FeePerWeight, SetType)>& visit_fn) noexcept
932 : : {
933 [ # # ]: 0 : if (GetTxCount()) {
934 : 0 : visit_fn(0, m_graph_index, m_feerate, SetType{});
935 : 0 : m_graph_index = NO_GRAPH_INDEX;
936 : : }
937 : 0 : }
938 : :
939 : 0 : int GenericClusterImpl::GetLevel(const TxGraphImpl& graph) const noexcept
940 : : {
941 : : // GetLevel() does not work for empty Clusters.
942 [ # # ]: 0 : if (!Assume(!m_linearization.empty())) return -1;
943 : :
944 : : // Pick an arbitrary Entry that occurs in this Cluster.
945 : 0 : const auto& entry = graph.m_entries[m_mapping[m_linearization.front()]];
946 : : // See if there is a level whose Locator matches this Cluster, if so return that level.
947 [ # # ]: 0 : for (int level = 0; level < MAX_LEVELS; ++level) {
948 [ # # ]: 0 : if (entry.m_locator[level].cluster == this) return level;
949 : : }
950 : : // Given that we started with an Entry that occurs in this Cluster, one of its Locators must
951 : : // point back to it.
952 : 0 : assert(false);
953 : : return -1;
954 : : }
955 : :
956 : 192811 : int SingletonClusterImpl::GetLevel(const TxGraphImpl& graph) const noexcept
957 : : {
958 : : // GetLevel() does not work for empty Clusters.
959 [ + - ]: 192811 : if (!Assume(GetTxCount())) return -1;
960 : :
961 : : // Get the Entry in this Cluster.
962 : 192811 : const auto& entry = graph.m_entries[m_graph_index];
963 : : // See if there is a level whose Locator matches this Cluster, if so return that level.
964 [ + - ]: 192811 : for (int level = 0; level < MAX_LEVELS; ++level) {
965 [ - + ]: 192811 : if (entry.m_locator[level].cluster == this) return level;
966 : : }
967 : : // Given that we started with an Entry that occurs in this Cluster, one of its Locators must
968 : : // point back to it.
969 : 0 : assert(false);
970 : : return -1;
971 : : }
972 : :
973 : 19 : void TxGraphImpl::ClearLocator(int level, GraphIndex idx, bool oversized_tx) noexcept
974 : : {
975 [ - + ]: 19 : auto& entry = m_entries[idx];
976 : 19 : auto& clusterset = GetClusterSet(level);
977 : 19 : Assume(entry.m_locator[level].IsPresent());
978 : : // Change the locator from Present to Missing or Removed.
979 [ - + - - ]: 19 : if (level == 0 || !entry.m_locator[level - 1].IsPresent()) {
980 : 19 : entry.m_locator[level].SetMissing();
981 : : } else {
982 : 0 : entry.m_locator[level].SetRemoved();
983 : 0 : clusterset.m_removed.push_back(idx);
984 : : }
985 : : // Update the transaction count.
986 : 19 : --clusterset.m_txcount;
987 : 19 : clusterset.m_txcount_oversized -= oversized_tx;
988 : : // If clearing main, adjust the status of Locators of this transaction in staging, if it exists.
989 [ + - - + ]: 19 : if (level == 0 && GetTopLevel() == 1) {
990 [ # # ]: 0 : if (entry.m_locator[1].IsRemoved()) {
991 : 0 : entry.m_locator[1].SetMissing();
992 [ # # ]: 0 : } else if (!entry.m_locator[1].IsPresent()) {
993 : 0 : --m_staging_clusterset->m_txcount;
994 : 0 : m_staging_clusterset->m_txcount_oversized -= oversized_tx;
995 : : }
996 : : }
997 [ + - ]: 19 : if (level == 0) ClearChunkData(entry);
998 : 19 : }
999 : :
1000 : 0 : void GenericClusterImpl::Updated(TxGraphImpl& graph, int level) noexcept
1001 : : {
1002 : : // Update all the Locators for this Cluster's Entry objects.
1003 [ # # ]: 0 : for (DepGraphIndex idx : m_linearization) {
1004 [ # # ]: 0 : auto& entry = graph.m_entries[m_mapping[idx]];
1005 : : // Discard any potential ChunkData prior to modifying the Cluster (as that could
1006 : : // invalidate its ordering).
1007 [ # # ]: 0 : if (level == 0) graph.ClearChunkData(entry);
1008 : 0 : entry.m_locator[level].SetPresent(this, idx);
1009 : : }
1010 : : // If this is for the main graph (level = 0), and the Cluster's quality is ACCEPTABLE or
1011 : : // OPTIMAL, compute its chunking and store its information in the Entry's m_main_lin_index
1012 : : // and m_main_chunk_feerate. These fields are only accessed after making the entire graph
1013 : : // ACCEPTABLE, so it is pointless to compute these if we haven't reached that quality level
1014 : : // yet.
1015 [ # # # # ]: 0 : if (level == 0 && IsAcceptable()) {
1016 [ # # ]: 0 : const LinearizationChunking chunking(m_depgraph, m_linearization);
1017 : 0 : LinearizationIndex lin_idx{0};
1018 : : // Iterate over the chunks.
1019 [ # # # # ]: 0 : for (unsigned chunk_idx = 0; chunk_idx < chunking.NumChunksLeft(); ++chunk_idx) {
1020 : 0 : auto chunk = chunking.GetChunk(chunk_idx);
1021 : 0 : auto chunk_count = chunk.transactions.Count();
1022 : 0 : Assume(chunk_count > 0);
1023 : : // Iterate over the transactions in the linearization, which must match those in chunk.
1024 : 0 : while (true) {
1025 [ # # ]: 0 : DepGraphIndex idx = m_linearization[lin_idx];
1026 : 0 : GraphIndex graph_idx = m_mapping[idx];
1027 : 0 : auto& entry = graph.m_entries[graph_idx];
1028 : 0 : entry.m_main_lin_index = lin_idx++;
1029 [ # # ]: 0 : entry.m_main_chunk_feerate = FeePerWeight::FromFeeFrac(chunk.feerate);
1030 : 0 : Assume(chunk.transactions[idx]);
1031 [ # # ]: 0 : chunk.transactions.Reset(idx);
1032 [ # # ]: 0 : if (chunk.transactions.None()) {
1033 : : // Last transaction in the chunk.
1034 [ # # # # ]: 0 : if (chunk_count == 1 && chunk_idx + 1 == chunking.NumChunksLeft()) {
1035 : : // If this is the final chunk of the cluster, and it contains just a single
1036 : : // transaction (which will always be true for the very common singleton
1037 : : // clusters), store the special value -1 as chunk count.
1038 : : chunk_count = LinearizationIndex(-1);
1039 : : }
1040 : 0 : graph.CreateChunkData(graph_idx, chunk_count);
1041 : 0 : break;
1042 : : }
1043 : : }
1044 : : }
1045 : 0 : }
1046 : 0 : }
1047 : :
1048 : 64311 : void SingletonClusterImpl::Updated(TxGraphImpl& graph, int level) noexcept
1049 : : {
1050 : : // Don't do anything if this is empty.
1051 [ + - ]: 64311 : if (GetTxCount() == 0) return;
1052 : :
1053 [ + - ]: 64311 : auto& entry = graph.m_entries[m_graph_index];
1054 : : // Discard any potential ChunkData prior to modifying the Cluster (as that could
1055 : : // invalidate its ordering).
1056 [ + - ]: 64311 : if (level == 0) graph.ClearChunkData(entry);
1057 : 64311 : entry.m_locator[level].SetPresent(this, 0);
1058 : : // If this is for the main graph (level = 0), compute its chunking and store its information in
1059 : : // the Entry's m_main_lin_index and m_main_chunk_feerate.
1060 [ + - + + ]: 64311 : if (level == 0 && IsAcceptable()) {
1061 : 64305 : entry.m_main_lin_index = 0;
1062 : 64305 : entry.m_main_chunk_feerate = m_feerate;
1063 : : // Always use the special LinearizationIndex(-1), indicating singleton chunk at end of
1064 : : // Cluster, here.
1065 : 64305 : graph.CreateChunkData(m_graph_index, LinearizationIndex(-1));
1066 : : }
1067 : : }
1068 : :
1069 : 0 : void GenericClusterImpl::GetConflicts(const TxGraphImpl& graph, std::vector<Cluster*>& out) const noexcept
1070 : : {
1071 [ # # ]: 0 : for (auto i : m_linearization) {
1072 [ # # ]: 0 : auto& entry = graph.m_entries[m_mapping[i]];
1073 : : // For every transaction Entry in this Cluster, if it also exists in a lower-level Cluster,
1074 : : // then that Cluster conflicts.
1075 [ # # ]: 0 : if (entry.m_locator[0].IsPresent()) {
1076 : 0 : out.push_back(entry.m_locator[0].cluster);
1077 : : }
1078 : : }
1079 : 0 : }
1080 : :
1081 : 0 : void SingletonClusterImpl::GetConflicts(const TxGraphImpl& graph, std::vector<Cluster*>& out) const noexcept
1082 : : {
1083 : : // Empty clusters have no conflicts.
1084 [ # # ]: 0 : if (GetTxCount() == 0) return;
1085 : :
1086 [ # # ]: 0 : auto& entry = graph.m_entries[m_graph_index];
1087 : : // If the transaction in this Cluster also exists in a lower-level Cluster, then that Cluster
1088 : : // conflicts.
1089 [ # # ]: 0 : if (entry.m_locator[0].IsPresent()) {
1090 : 0 : out.push_back(entry.m_locator[0].cluster);
1091 : : }
1092 : : }
1093 : :
1094 : 0 : std::vector<Cluster*> TxGraphImpl::GetConflicts() const noexcept
1095 : : {
1096 : 0 : Assume(GetTopLevel() == 1);
1097 : 0 : auto& clusterset = GetClusterSet(1);
1098 : 0 : std::vector<Cluster*> ret;
1099 : : // All main Clusters containing transactions in m_removed (so (P,R) ones) are conflicts.
1100 [ # # ]: 0 : for (auto i : clusterset.m_removed) {
1101 [ # # ]: 0 : auto& entry = m_entries[i];
1102 [ # # ]: 0 : if (entry.m_locator[0].IsPresent()) {
1103 : 0 : ret.push_back(entry.m_locator[0].cluster);
1104 : : }
1105 : : }
1106 : : // Then go over all Clusters at this level, and find their conflicts (the (P,P) ones).
1107 [ # # ]: 0 : for (int quality = 0; quality < int(QualityLevel::NONE); ++quality) {
1108 : 0 : auto& clusters = clusterset.m_clusters[quality];
1109 [ # # ]: 0 : for (const auto& cluster : clusters) {
1110 : 0 : cluster->GetConflicts(*this, ret);
1111 : : }
1112 : : }
1113 : : // Deduplicate the result (the same Cluster may appear multiple times).
1114 : 0 : std::sort(ret.begin(), ret.end(), [](Cluster* a, Cluster* b) noexcept { return CompareClusters(a, b) < 0; });
1115 : 0 : ret.erase(std::unique(ret.begin(), ret.end()), ret.end());
1116 : 0 : return ret;
1117 : : }
1118 : :
1119 : 0 : Cluster* GenericClusterImpl::CopyToStaging(TxGraphImpl& graph) const noexcept
1120 : : {
1121 : : // Construct an empty Cluster.
1122 : 0 : auto ret = graph.CreateEmptyGenericCluster();
1123 : 0 : auto ptr = ret.get();
1124 : : // Copy depgraph, mapping, and linearization.
1125 : 0 : ptr->m_depgraph = m_depgraph;
1126 : 0 : ptr->m_mapping = m_mapping;
1127 : 0 : ptr->m_linearization = m_linearization;
1128 : : // Insert the new Cluster into the graph.
1129 : 0 : graph.InsertCluster(/*level=*/1, std::move(ret), m_quality);
1130 : : // Update its Locators.
1131 : 0 : ptr->Updated(graph, /*level=*/1);
1132 : : // Update memory usage.
1133 : 0 : graph.GetClusterSet(/*level=*/1).m_cluster_usage += ptr->TotalMemoryUsage();
1134 [ # # ]: 0 : return ptr;
1135 : 0 : }
1136 : :
1137 : 0 : Cluster* SingletonClusterImpl::CopyToStaging(TxGraphImpl& graph) const noexcept
1138 : : {
1139 : : // Construct an empty Cluster.
1140 : 0 : auto ret = graph.CreateEmptySingletonCluster();
1141 : 0 : auto ptr = ret.get();
1142 : : // Copy data.
1143 : 0 : ptr->m_graph_index = m_graph_index;
1144 : 0 : ptr->m_feerate = m_feerate;
1145 : : // Insert the new Cluster into the graph.
1146 : 0 : graph.InsertCluster(/*level=*/1, std::move(ret), m_quality);
1147 : : // Update its Locators.
1148 : 0 : ptr->Updated(graph, /*level=*/1);
1149 : : // Update memory usage.
1150 : 0 : graph.GetClusterSet(/*level=*/1).m_cluster_usage += ptr->TotalMemoryUsage();
1151 : 0 : return ptr;
1152 : 0 : }
1153 : :
1154 : 0 : void GenericClusterImpl::ApplyRemovals(TxGraphImpl& graph, int level, std::span<GraphIndex>& to_remove) noexcept
1155 : : {
1156 : : // Iterate over the prefix of to_remove that applies to this cluster.
1157 : 0 : Assume(!to_remove.empty());
1158 : 0 : SetType todo;
1159 : 0 : graph.GetClusterSet(level).m_cluster_usage -= TotalMemoryUsage();
1160 : 0 : do {
1161 [ # # ]: 0 : GraphIndex idx = to_remove.front();
1162 [ # # ]: 0 : Assume(idx < graph.m_entries.size());
1163 [ # # ]: 0 : auto& entry = graph.m_entries[idx];
1164 : 0 : auto& locator = entry.m_locator[level];
1165 : : // Stop once we hit an entry that applies to another Cluster.
1166 [ # # ]: 0 : if (locator.cluster != this) break;
1167 : : // - Remember it in a set of to-remove DepGraphIndexes.
1168 [ # # ]: 0 : todo.Set(locator.index);
1169 : : // - Remove from m_mapping. This isn't strictly necessary as unused positions in m_mapping
1170 : : // are just never accessed, but set it to -1 here to increase the ability to detect a bug
1171 : : // that causes it to be accessed regardless.
1172 [ # # ]: 0 : m_mapping[locator.index] = GraphIndex(-1);
1173 : : // - Remove its linearization index from the Entry (if in main).
1174 [ # # ]: 0 : if (level == 0) {
1175 : 0 : entry.m_main_lin_index = LinearizationIndex(-1);
1176 : : }
1177 : : // - Mark it as missing/removed in the Entry's locator.
1178 : 0 : graph.ClearLocator(level, idx, m_quality == QualityLevel::OVERSIZED_SINGLETON);
1179 [ # # ]: 0 : to_remove = to_remove.subspan(1);
1180 [ # # ]: 0 : } while(!to_remove.empty());
1181 : :
1182 : 0 : auto quality = m_quality;
1183 : 0 : Assume(todo.Any());
1184 : : // Wipe from the Cluster's DepGraph (this is O(n) regardless of the number of entries
1185 : : // removed, so we benefit from batching all the removals).
1186 : 0 : m_depgraph.RemoveTransactions(todo);
1187 [ # # ]: 0 : m_mapping.resize(m_depgraph.PositionRange());
1188 : :
1189 : : // First remove all removals at the end of the linearization.
1190 [ # # # # ]: 0 : while (!m_linearization.empty() && todo[m_linearization.back()]) {
1191 : 0 : todo.Reset(m_linearization.back());
1192 : 0 : m_linearization.pop_back();
1193 : : }
1194 [ # # ]: 0 : if (todo.None()) {
1195 : : // If no further removals remain, and thus all removals were at the end, we may be able
1196 : : // to leave the cluster at a better quality level.
1197 [ # # ]: 0 : if (IsAcceptable(/*after_split=*/true)) {
1198 : : quality = QualityLevel::NEEDS_SPLIT_ACCEPTABLE;
1199 : : } else {
1200 : : quality = QualityLevel::NEEDS_SPLIT;
1201 : : }
1202 : : } else {
1203 : : // If more removals remain, filter those out of m_linearization.
1204 : 0 : m_linearization.erase(std::remove_if(
1205 : : m_linearization.begin(),
1206 : : m_linearization.end(),
1207 [ # # # # ]: 0 : [&](auto pos) { return todo[pos]; }), m_linearization.end());
1208 : 0 : quality = QualityLevel::NEEDS_SPLIT;
1209 : : }
1210 : 0 : Compact();
1211 : 0 : graph.GetClusterSet(level).m_cluster_usage += TotalMemoryUsage();
1212 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, quality);
1213 : 0 : Updated(graph, level);
1214 : 0 : }
1215 : :
1216 : 19 : void SingletonClusterImpl::ApplyRemovals(TxGraphImpl& graph, int level, std::span<GraphIndex>& to_remove) noexcept
1217 : : {
1218 : : // We can only remove the one transaction this Cluster has.
1219 : 19 : Assume(!to_remove.empty());
1220 : 19 : Assume(GetTxCount());
1221 : 19 : Assume(to_remove.front() == m_graph_index);
1222 : : // Pop all copies of m_graph_index from the front of to_remove (at least one, but there may be
1223 : : // multiple).
1224 : 19 : do {
1225 [ + + ]: 19 : to_remove = to_remove.subspan(1);
1226 [ + + + - : 34 : } while (!to_remove.empty() && to_remove.front() == m_graph_index);
- + ]
1227 : : // Clear this cluster.
1228 : 19 : graph.ClearLocator(level, m_graph_index, m_quality == QualityLevel::OVERSIZED_SINGLETON);
1229 : 19 : m_graph_index = NO_GRAPH_INDEX;
1230 : 19 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::NEEDS_SPLIT);
1231 : : // No need to account for m_cluster_usage changes here, as SingletonClusterImpl has constant
1232 : : // memory usage.
1233 : 19 : }
1234 : :
1235 : 0 : void GenericClusterImpl::Clear(TxGraphImpl& graph, int level) noexcept
1236 : : {
1237 [ # # ]: 0 : Assume(GetTxCount());
1238 : 0 : graph.GetClusterSet(level).m_cluster_usage -= TotalMemoryUsage();
1239 [ # # ]: 0 : for (auto i : m_linearization) {
1240 : 0 : graph.ClearLocator(level, m_mapping[i], m_quality == QualityLevel::OVERSIZED_SINGLETON);
1241 : : }
1242 : 0 : m_depgraph = {};
1243 [ # # ]: 0 : m_linearization.clear();
1244 [ # # ]: 0 : m_mapping.clear();
1245 : 0 : }
1246 : :
1247 : 0 : void SingletonClusterImpl::Clear(TxGraphImpl& graph, int level) noexcept
1248 : : {
1249 : 0 : Assume(GetTxCount());
1250 : 0 : graph.GetClusterSet(level).m_cluster_usage -= TotalMemoryUsage();
1251 : 0 : graph.ClearLocator(level, m_graph_index, m_quality == QualityLevel::OVERSIZED_SINGLETON);
1252 : 0 : m_graph_index = NO_GRAPH_INDEX;
1253 : 0 : }
1254 : :
1255 : 0 : void GenericClusterImpl::MoveToMain(TxGraphImpl& graph) noexcept
1256 : : {
1257 [ # # ]: 0 : for (auto i : m_linearization) {
1258 : 0 : GraphIndex idx = m_mapping[i];
1259 : 0 : auto& entry = graph.m_entries[idx];
1260 : 0 : entry.m_locator[1].SetMissing();
1261 : : }
1262 : 0 : auto quality = m_quality;
1263 : : // Subtract memory usage from staging and add it to main.
1264 : 0 : graph.GetClusterSet(/*level=*/1).m_cluster_usage -= TotalMemoryUsage();
1265 : 0 : graph.GetClusterSet(/*level=*/0).m_cluster_usage += TotalMemoryUsage();
1266 : : // Remove cluster itself from staging and add it to main.
1267 : 0 : auto cluster = graph.ExtractCluster(1, quality, m_setindex);
1268 : 0 : graph.InsertCluster(/*level=*/0, std::move(cluster), quality);
1269 : 0 : Updated(graph, /*level=*/0);
1270 : 0 : }
1271 : :
1272 : 0 : void SingletonClusterImpl::MoveToMain(TxGraphImpl& graph) noexcept
1273 : : {
1274 [ # # ]: 0 : if (GetTxCount()) {
1275 : 0 : auto& entry = graph.m_entries[m_graph_index];
1276 : 0 : entry.m_locator[1].SetMissing();
1277 : : }
1278 : 0 : auto quality = m_quality;
1279 : 0 : graph.GetClusterSet(/*level=*/1).m_cluster_usage -= TotalMemoryUsage();
1280 : 0 : auto cluster = graph.ExtractCluster(/*level=*/1, quality, m_setindex);
1281 : 0 : graph.InsertCluster(/*level=*/0, std::move(cluster), quality);
1282 : 0 : graph.GetClusterSet(/*level=*/0).m_cluster_usage += TotalMemoryUsage();
1283 : 0 : Updated(graph, /*level=*/0);
1284 : 0 : }
1285 : :
1286 : 0 : void GenericClusterImpl::Compact() noexcept
1287 : : {
1288 : 0 : m_linearization.shrink_to_fit();
1289 : 0 : m_mapping.shrink_to_fit();
1290 : 0 : m_depgraph.Compact();
1291 : 0 : }
1292 : :
1293 : 0 : void SingletonClusterImpl::Compact() noexcept
1294 : : {
1295 : : // Nothing to compact; SingletonClusterImpl is constant size.
1296 : 0 : }
1297 : :
1298 : 0 : void GenericClusterImpl::AppendChunkFeerates(std::vector<FeeFrac>& ret) const noexcept
1299 : : {
1300 [ # # ]: 0 : auto chunk_feerates = ChunkLinearization(m_depgraph, m_linearization);
1301 [ # # # # ]: 0 : ret.reserve(ret.size() + chunk_feerates.size());
1302 : 0 : ret.insert(ret.end(), chunk_feerates.begin(), chunk_feerates.end());
1303 : 0 : }
1304 : :
1305 : 0 : void SingletonClusterImpl::AppendChunkFeerates(std::vector<FeeFrac>& ret) const noexcept
1306 : : {
1307 [ # # ]: 0 : if (GetTxCount()) {
1308 : 0 : ret.push_back(m_feerate);
1309 : : }
1310 : 0 : }
1311 : :
1312 : 0 : uint64_t GenericClusterImpl::AppendTrimData(std::vector<TrimTxData>& ret, std::vector<std::pair<GraphIndex, GraphIndex>>& deps) const noexcept
1313 : : {
1314 [ # # ]: 0 : const LinearizationChunking linchunking(m_depgraph, m_linearization);
1315 : 0 : LinearizationIndex pos{0};
1316 : 0 : uint64_t size{0};
1317 : 0 : auto prev_index = GraphIndex(-1);
1318 : : // Iterate over the chunks of this cluster's linearization.
1319 [ # # # # ]: 0 : for (unsigned i = 0; i < linchunking.NumChunksLeft(); ++i) {
1320 : 0 : const auto& [chunk, chunk_feerate] = linchunking.GetChunk(i);
1321 : : // Iterate over the transactions of that chunk, in linearization order.
1322 : 0 : auto chunk_tx_count = chunk.Count();
1323 [ # # ]: 0 : for (unsigned j = 0; j < chunk_tx_count; ++j) {
1324 : 0 : auto cluster_idx = m_linearization[pos];
1325 : : // The transaction must appear in the chunk.
1326 : 0 : Assume(chunk[cluster_idx]);
1327 : : // Construct a new element in ret.
1328 : 0 : auto& entry = ret.emplace_back();
1329 [ # # ]: 0 : entry.m_chunk_feerate = FeePerWeight::FromFeeFrac(chunk_feerate);
1330 [ # # ]: 0 : entry.m_index = m_mapping[cluster_idx];
1331 : : // If this is not the first transaction of the cluster linearization, it has an
1332 : : // implicit dependency on its predecessor.
1333 [ # # ]: 0 : if (pos != 0) deps.emplace_back(prev_index, entry.m_index);
1334 : 0 : prev_index = entry.m_index;
1335 : 0 : entry.m_tx_size = m_depgraph.FeeRate(cluster_idx).size;
1336 : 0 : size += entry.m_tx_size;
1337 : 0 : ++pos;
1338 : : }
1339 : : }
1340 : 0 : return size;
1341 : 0 : }
1342 : :
1343 : 64217 : uint64_t SingletonClusterImpl::AppendTrimData(std::vector<TrimTxData>& ret, std::vector<std::pair<GraphIndex, GraphIndex>>& deps) const noexcept
1344 : : {
1345 [ + - ]: 64217 : if (!GetTxCount()) return 0;
1346 : 64217 : auto& entry = ret.emplace_back();
1347 : 64217 : entry.m_chunk_feerate = m_feerate;
1348 : 64217 : entry.m_index = m_graph_index;
1349 : 64217 : entry.m_tx_size = m_feerate.size;
1350 : 64217 : return m_feerate.size;
1351 : : }
1352 : :
1353 : 0 : bool GenericClusterImpl::Split(TxGraphImpl& graph, int level) noexcept
1354 : : {
1355 : : // This function can only be called when the Cluster needs splitting.
1356 : 0 : Assume(NeedsSplitting());
1357 : : // Determine the new quality the split-off Clusters will have.
1358 [ # # ]: 0 : QualityLevel new_quality = IsAcceptable(/*after_split=*/true) ? QualityLevel::ACCEPTABLE
1359 : : : QualityLevel::NEEDS_RELINEARIZE;
1360 : : // If we're going to produce ACCEPTABLE clusters (i.e., when in NEEDS_SPLIT_ACCEPTABLE), we
1361 : : // need to post-linearize to make sure the split-out versions are all connected (as
1362 : : // connectivity may have changed by removing part of the cluster). This could be done on each
1363 : : // resulting split-out cluster separately, but it is simpler to do it once up front before
1364 : : // splitting. This step is not necessary if the resulting clusters are NEEDS_RELINEARIZE, as
1365 : : // they will be post-linearized anyway in MakeAcceptable().
1366 : : if (new_quality == QualityLevel::ACCEPTABLE) {
1367 [ # # ]: 0 : PostLinearize(m_depgraph, m_linearization);
1368 : : }
1369 : : /** Which positions are still left in this Cluster. */
1370 : 0 : auto todo = m_depgraph.Positions();
1371 : : /** Mapping from transaction positions in this Cluster to the Cluster where it ends up, and
1372 : : * its position therein. */
1373 [ # # ]: 0 : std::vector<std::pair<Cluster*, DepGraphIndex>> remap(m_depgraph.PositionRange());
1374 : 0 : std::vector<Cluster*> new_clusters;
1375 : 0 : bool first{true};
1376 : : // Iterate over the connected components of this Cluster's m_depgraph.
1377 [ # # ]: 0 : while (todo.Any()) {
1378 : 0 : auto component = m_depgraph.FindConnectedComponent(todo);
1379 [ # # ]: 0 : auto component_size = component.Count();
1380 [ # # ]: 0 : auto split_quality = component_size <= 2 ? QualityLevel::OPTIMAL : new_quality;
1381 [ # # # # : 0 : if (first && component == todo && SetType::Fill(component_size) == component && component_size >= MIN_INTENDED_TX_COUNT) {
# # # # ]
1382 : : // The existing Cluster is an entire component, without holes. Leave it be, but update
1383 : : // its quality. If there are holes, we continue, so that the Cluster is reconstructed
1384 : : // without holes, reducing memory usage. If the component's size is below the intended
1385 : : // transaction count for this Cluster implementation, continue so that it can get
1386 : : // converted.
1387 : 0 : Assume(todo == m_depgraph.Positions());
1388 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, split_quality);
1389 : : // If this made the quality ACCEPTABLE or OPTIMAL, we need to compute and cache its
1390 : : // chunking.
1391 : 0 : Updated(graph, level);
1392 : 0 : return false;
1393 : : }
1394 : 0 : first = false;
1395 : : // Construct a new Cluster to hold the found component.
1396 : 0 : auto new_cluster = graph.CreateEmptyCluster(component_size);
1397 : 0 : new_clusters.push_back(new_cluster.get());
1398 : : // Remember that all the component's transactions go to this new Cluster. The positions
1399 : : // will be determined below, so use -1 for now.
1400 [ # # # # ]: 0 : for (auto i : component) {
1401 [ # # ]: 0 : remap[i] = {new_cluster.get(), DepGraphIndex(-1)};
1402 : : }
1403 : 0 : graph.InsertCluster(level, std::move(new_cluster), split_quality);
1404 : 0 : todo -= component;
1405 : 0 : }
1406 : : // We have to split the Cluster up. Remove accounting for the existing one first.
1407 : 0 : graph.GetClusterSet(level).m_cluster_usage -= TotalMemoryUsage();
1408 : : // Redistribute the transactions.
1409 [ # # ]: 0 : for (auto i : m_linearization) {
1410 : : /** The cluster which transaction originally in position i is moved to. */
1411 : 0 : Cluster* new_cluster = remap[i].first;
1412 : : // Copy the transaction to the new cluster's depgraph, and remember the position.
1413 : 0 : remap[i].second = new_cluster->AppendTransaction(m_mapping[i], FeePerWeight::FromFeeFrac(m_depgraph.FeeRate(i)));
1414 : : }
1415 : : // Redistribute the dependencies.
1416 [ # # ]: 0 : for (auto i : m_linearization) {
1417 : : /** The cluster transaction in position i is moved to. */
1418 : 0 : Cluster* new_cluster = remap[i].first;
1419 : : // Copy its parents, translating positions.
1420 : 0 : SetType new_parents;
1421 [ # # # # : 0 : for (auto par : m_depgraph.GetReducedParents(i)) new_parents.Set(remap[par].second);
# # ]
1422 : 0 : new_cluster->AddDependencies(new_parents, remap[i].second);
1423 : : }
1424 : : // Update all the Locators of moved transactions, and memory usage.
1425 [ # # ]: 0 : for (Cluster* new_cluster : new_clusters) {
1426 : 0 : new_cluster->Updated(graph, level);
1427 : 0 : new_cluster->Compact();
1428 : 0 : graph.GetClusterSet(level).m_cluster_usage += new_cluster->TotalMemoryUsage();
1429 : : }
1430 : : // Wipe this Cluster, and return that it needs to be deleted.
1431 : 0 : m_depgraph = DepGraph<SetType>{};
1432 [ # # ]: 0 : m_mapping.clear();
1433 [ # # ]: 0 : m_linearization.clear();
1434 : : return true;
1435 : 0 : }
1436 : :
1437 : 0 : bool SingletonClusterImpl::Split(TxGraphImpl& graph, int level) noexcept
1438 : : {
1439 : 0 : Assume(NeedsSplitting());
1440 [ # # ]: 0 : if (GetTxCount() == 0) {
1441 : : // The cluster is now empty.
1442 : 0 : graph.GetClusterSet(level).m_cluster_usage -= TotalMemoryUsage();
1443 : 0 : return true;
1444 : : } else {
1445 : : // Nothing changed.
1446 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::OPTIMAL);
1447 : 0 : Updated(graph, level);
1448 : 0 : return false;
1449 : : }
1450 : : }
1451 : :
1452 : 0 : void GenericClusterImpl::Merge(TxGraphImpl& graph, int level, Cluster& other) noexcept
1453 : : {
1454 : : /** Vector to store the positions in this Cluster for each position in other. */
1455 : 0 : std::vector<DepGraphIndex> remap(other.GetDepGraphIndexRange());
1456 : : // Iterate over all transactions in the other Cluster (the one being absorbed).
1457 : 0 : other.ExtractTransactions([&](DepGraphIndex pos, GraphIndex idx, FeePerWeight feerate, SetType other_parents) noexcept {
1458 : : // Copy the transaction into this Cluster, and remember its position.
1459 : 0 : auto new_pos = m_depgraph.AddTransaction(feerate);
1460 : : // Since this cluster must have been made hole-free before being merged into, all added
1461 : : // transactions should appear at the end.
1462 [ # # ]: 0 : Assume(new_pos == m_mapping.size());
1463 : 0 : remap[pos] = new_pos;
1464 : 0 : m_mapping.push_back(idx);
1465 : 0 : m_linearization.push_back(new_pos);
1466 : : // Copy the transaction's dependencies, translating them using remap. Note that since
1467 : : // pos iterates in linearization order, which is topological, all parents of pos should
1468 : : // already be in remap.
1469 : 0 : SetType parents;
1470 [ # # # # ]: 0 : for (auto par : other_parents) {
1471 [ # # ]: 0 : parents.Set(remap[par]);
1472 : : }
1473 : 0 : m_depgraph.AddDependencies(parents, remap[pos]);
1474 : : // Update the transaction's Locator. There is no need to call Updated() to update chunk
1475 : : // feerates, as Updated() will be invoked by Cluster::ApplyDependencies on the resulting
1476 : : // merged Cluster later anyway.
1477 [ # # ]: 0 : auto& entry = graph.m_entries[idx];
1478 : : // Discard any potential ChunkData prior to modifying the Cluster (as that could
1479 : : // invalidate its ordering).
1480 [ # # ]: 0 : if (level == 0) graph.ClearChunkData(entry);
1481 : 0 : entry.m_locator[level].SetPresent(this, new_pos);
1482 : 0 : });
1483 : 0 : }
1484 : :
1485 : 0 : void SingletonClusterImpl::Merge(TxGraphImpl& graph, int level, Cluster& other_abstract) noexcept
1486 : : {
1487 : : // Nothing can be merged into a singleton; it should have been converted to GenericClusterImpl
1488 : : // first.
1489 : 0 : Assume(false);
1490 : 0 : }
1491 : :
1492 : 0 : void GenericClusterImpl::ApplyDependencies(TxGraphImpl& graph, int level, std::span<std::pair<GraphIndex, GraphIndex>> to_apply) noexcept
1493 : : {
1494 : : // This function is invoked by TxGraphImpl::ApplyDependencies after merging groups of Clusters
1495 : : // between which dependencies are added, which simply concatenates their linearizations. Invoke
1496 : : // PostLinearize, which has the effect that the linearization becomes a merge-sort of the
1497 : : // constituent linearizations. Do this here rather than in Cluster::Merge, because this
1498 : : // function is only invoked once per merged Cluster, rather than once per constituent one.
1499 : : // This concatenation + post-linearization could be replaced with an explicit merge-sort.
1500 [ # # ]: 0 : PostLinearize(m_depgraph, m_linearization);
1501 : :
1502 : : // Sort the list of dependencies to apply by child, so those can be applied in batch.
1503 [ # # # # : 0 : std::sort(to_apply.begin(), to_apply.end(), [](auto& a, auto& b) { return a.second < b.second; });
# # # # #
# # # # #
# # # # #
# # # #
# ]
1504 : : // Iterate over groups of to-be-added dependencies with the same child.
1505 : 0 : auto it = to_apply.begin();
1506 [ # # ]: 0 : while (it != to_apply.end()) {
1507 : 0 : auto& first_child = graph.m_entries[it->second].m_locator[level];
1508 : 0 : const auto child_idx = first_child.index;
1509 : : // Iterate over all to-be-added dependencies within that same child, gather the relevant
1510 : : // parents.
1511 : 0 : SetType parents;
1512 [ # # ]: 0 : while (it != to_apply.end()) {
1513 [ # # ]: 0 : auto& child = graph.m_entries[it->second].m_locator[level];
1514 : 0 : auto& parent = graph.m_entries[it->first].m_locator[level];
1515 : 0 : Assume(child.cluster == this && parent.cluster == this);
1516 [ # # ]: 0 : if (child.index != child_idx) break;
1517 : 0 : parents.Set(parent.index);
1518 : 0 : ++it;
1519 : : }
1520 : : // Push all dependencies to the underlying DepGraph. Note that this is O(N) in the size of
1521 : : // the cluster, regardless of the number of parents being added, so batching them together
1522 : : // has a performance benefit.
1523 : 0 : m_depgraph.AddDependencies(parents, child_idx);
1524 : : }
1525 : :
1526 : : // Finally fix the linearization, as the new dependencies may have invalidated the
1527 : : // linearization, and post-linearize it to fix up the worst problems with it.
1528 [ # # ]: 0 : FixLinearization(m_depgraph, m_linearization);
1529 [ # # ]: 0 : PostLinearize(m_depgraph, m_linearization);
1530 : 0 : Assume(!NeedsSplitting());
1531 : 0 : Assume(!IsOversized());
1532 [ # # ]: 0 : if (IsAcceptable()) {
1533 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::NEEDS_RELINEARIZE);
1534 : : }
1535 : :
1536 : : // Finally push the changes to graph.m_entries.
1537 : 0 : Updated(graph, level);
1538 : 0 : }
1539 : :
1540 : 0 : void SingletonClusterImpl::ApplyDependencies(TxGraphImpl& graph, int level, std::span<std::pair<GraphIndex, GraphIndex>> to_apply) noexcept
1541 : : {
1542 : : // Nothing can actually be applied.
1543 [ # # ]: 0 : for (auto& [par, chl] : to_apply) {
1544 : 0 : Assume(par == m_graph_index);
1545 : 0 : Assume(chl == m_graph_index);
1546 : : }
1547 : 0 : }
1548 : :
1549 : 8 : TxGraphImpl::~TxGraphImpl() noexcept
1550 : : {
1551 : : // If Refs outlive the TxGraphImpl they refer to, unlink them, so that their destructor does not
1552 : : // try to reach into a non-existing TxGraphImpl anymore.
1553 [ + + ]: 64314 : for (auto& entry : m_entries) {
1554 [ + + ]: 64310 : if (entry.m_ref != nullptr) {
1555 : 64011 : GetRefGraph(*entry.m_ref) = nullptr;
1556 : : }
1557 : : }
1558 : 8 : }
1559 : :
1560 : 19 : std::unique_ptr<Cluster> TxGraphImpl::ExtractCluster(int level, QualityLevel quality, ClusterSetIndex setindex) noexcept
1561 : : {
1562 : 19 : Assume(quality != QualityLevel::NONE);
1563 : :
1564 : 19 : auto& clusterset = GetClusterSet(level);
1565 [ - + ]: 19 : auto& quality_clusters = clusterset.m_clusters[int(quality)];
1566 [ - + ]: 19 : Assume(setindex < quality_clusters.size());
1567 : :
1568 : : // Extract the Cluster-owning unique_ptr.
1569 [ - + ]: 19 : std::unique_ptr<Cluster> ret = std::move(quality_clusters[setindex]);
1570 [ - + ]: 19 : ret->m_quality = QualityLevel::NONE;
1571 : 19 : ret->m_setindex = ClusterSetIndex(-1);
1572 : :
1573 : : // Clean up space in quality_cluster.
1574 [ - + ]: 19 : auto max_setindex = quality_clusters.size() - 1;
1575 [ + + ]: 19 : if (setindex != max_setindex) {
1576 : : // If the cluster was not the last element of quality_clusters, move that to take its place.
1577 : 10 : quality_clusters.back()->m_setindex = setindex;
1578 : 10 : quality_clusters[setindex] = std::move(quality_clusters.back());
1579 : : }
1580 : : // The last element of quality_clusters is now unused; drop it.
1581 : 19 : quality_clusters.pop_back();
1582 : :
1583 : 19 : return ret;
1584 : : }
1585 : :
1586 : 64330 : ClusterSetIndex TxGraphImpl::InsertCluster(int level, std::unique_ptr<Cluster>&& cluster, QualityLevel quality) noexcept
1587 : : {
1588 : : // Cannot insert with quality level NONE (as that would mean not inserted).
1589 : 64330 : Assume(quality != QualityLevel::NONE);
1590 : : // The passed-in Cluster must not currently be in the TxGraphImpl.
1591 [ - + ]: 64330 : Assume(cluster->m_quality == QualityLevel::NONE);
1592 : :
1593 : : // Append it at the end of the relevant TxGraphImpl::m_cluster.
1594 : 64330 : auto& clusterset = GetClusterSet(level);
1595 [ - + ]: 64330 : auto& quality_clusters = clusterset.m_clusters[int(quality)];
1596 [ - + ]: 64330 : ClusterSetIndex ret = quality_clusters.size();
1597 : 64330 : cluster->m_quality = quality;
1598 : 64330 : cluster->m_setindex = ret;
1599 : 64330 : quality_clusters.push_back(std::move(cluster));
1600 : 64330 : return ret;
1601 : : }
1602 : :
1603 : 19 : void TxGraphImpl::SetClusterQuality(int level, QualityLevel old_quality, ClusterSetIndex old_index, QualityLevel new_quality) noexcept
1604 : : {
1605 : 19 : Assume(new_quality != QualityLevel::NONE);
1606 : :
1607 : : // Don't do anything if the quality did not change.
1608 [ + - ]: 19 : if (old_quality == new_quality) return;
1609 : : // Extract the cluster from where it currently resides.
1610 : 19 : auto cluster_ptr = ExtractCluster(level, old_quality, old_index);
1611 : : // And re-insert it where it belongs.
1612 : 19 : InsertCluster(level, std::move(cluster_ptr), new_quality);
1613 : 19 : }
1614 : :
1615 : 0 : void TxGraphImpl::DeleteCluster(Cluster& cluster, int level) noexcept
1616 : : {
1617 : : // Extract the cluster from where it currently resides.
1618 : 0 : auto cluster_ptr = ExtractCluster(level, cluster.m_quality, cluster.m_setindex);
1619 : : // And throw it away.
1620 [ # # ]: 0 : cluster_ptr.reset();
1621 : 0 : }
1622 : :
1623 : 892818 : std::pair<Cluster*, int> TxGraphImpl::FindClusterAndLevel(GraphIndex idx, int level) const noexcept
1624 : : {
1625 [ + - ]: 892818 : Assume(level >= 0 && level <= GetTopLevel());
1626 : 892818 : auto& entry = m_entries[idx];
1627 : : // Search the entry's locators from top to bottom.
1628 [ + + ]: 892826 : for (int l = level; l >= 0; --l) {
1629 : : // If the locator is missing, dig deeper; it may exist at a lower level and therefore be
1630 : : // implicitly existing at this level too.
1631 [ + + ]: 892826 : if (entry.m_locator[l].IsMissing()) continue;
1632 : : // If the locator has the entry marked as explicitly removed, stop.
1633 [ - + ]: 892810 : if (entry.m_locator[l].IsRemoved()) break;
1634 : : // Otherwise, we have found the topmost ClusterSet that contains this entry.
1635 : 892810 : return {entry.m_locator[l].cluster, l};
1636 : : }
1637 : : // If no non-empty locator was found, or an explicitly removed was hit, return nothing.
1638 : 8 : return {nullptr, -1};
1639 : : }
1640 : :
1641 : 0 : Cluster* TxGraphImpl::PullIn(Cluster* cluster, int level) noexcept
1642 : : {
1643 [ # # ]: 0 : int to_level = GetTopLevel();
1644 : 0 : Assume(to_level == 1);
1645 : 0 : Assume(level <= to_level);
1646 : : // Copy the Cluster from main to staging, if it's not already there.
1647 [ # # ]: 0 : if (level == 0) {
1648 : : // Make the Cluster Acceptable before copying. This isn't strictly necessary, but doing it
1649 : : // now avoids doing double work later.
1650 : 0 : MakeAcceptable(*cluster, level);
1651 : 0 : cluster = cluster->CopyToStaging(*this);
1652 : : }
1653 : 0 : return cluster;
1654 : : }
1655 : :
1656 : 316 : void TxGraphImpl::ApplyRemovals(int up_to_level) noexcept
1657 : : {
1658 [ + - ]: 316 : Assume(up_to_level >= 0 && up_to_level <= GetTopLevel());
1659 [ + + ]: 632 : for (int level = 0; level <= up_to_level; ++level) {
1660 : 316 : auto& clusterset = GetClusterSet(level);
1661 : 316 : auto& to_remove = clusterset.m_to_remove;
1662 : : // Skip if there is nothing to remove in this level.
1663 [ + + ]: 316 : if (to_remove.empty()) continue;
1664 : : // Pull in all Clusters that are not in staging.
1665 [ - + ]: 4 : if (level == 1) {
1666 [ # # ]: 0 : for (GraphIndex index : to_remove) {
1667 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(index, level);
1668 [ # # ]: 0 : if (cluster != nullptr) PullIn(cluster, cluster_level);
1669 : : }
1670 : : }
1671 : : // Group the set of to-be-removed entries by Cluster::m_sequence.
1672 : 4 : std::sort(to_remove.begin(), to_remove.end(), [&](GraphIndex a, GraphIndex b) noexcept {
1673 : 30 : Cluster* cluster_a = m_entries[a].m_locator[level].cluster;
1674 : 30 : Cluster* cluster_b = m_entries[b].m_locator[level].cluster;
1675 : 30 : return CompareClusters(cluster_a, cluster_b) < 0;
1676 : : });
1677 : : // Process per Cluster.
1678 [ - + ]: 4 : std::span to_remove_span{to_remove};
1679 [ + + ]: 23 : while (!to_remove_span.empty()) {
1680 [ + - ]: 19 : Cluster* cluster = m_entries[to_remove_span.front()].m_locator[level].cluster;
1681 [ + - ]: 19 : if (cluster != nullptr) {
1682 : : // If the first to_remove_span entry's Cluster exists, hand to_remove_span to it, so it
1683 : : // can pop off whatever applies to it.
1684 : 19 : cluster->ApplyRemovals(*this, level, to_remove_span);
1685 : : } else {
1686 : : // Otherwise, skip this already-removed entry. This may happen when
1687 : : // RemoveTransaction was called twice on the same Ref, for example.
1688 : 0 : to_remove_span = to_remove_span.subspan(1);
1689 : : }
1690 : : }
1691 [ + - ]: 320 : to_remove.clear();
1692 : : }
1693 : 316 : Compact();
1694 : 316 : }
1695 : :
1696 : 1 : void TxGraphImpl::SwapIndexes(GraphIndex a, GraphIndex b) noexcept
1697 : : {
1698 [ - + ]: 1 : Assume(a < m_entries.size());
1699 : 1 : Assume(b < m_entries.size());
1700 : : // Swap the Entry objects.
1701 : 1 : std::swap(m_entries[a], m_entries[b]);
1702 : : // Iterate over both objects.
1703 [ + + ]: 3 : for (int i = 0; i < 2; ++i) {
1704 [ + + ]: 2 : GraphIndex idx = i ? b : a;
1705 [ + + ]: 2 : Entry& entry = m_entries[idx];
1706 : : // Update linked Ref, if any exists.
1707 [ + + ]: 2 : if (entry.m_ref) GetRefIndex(*entry.m_ref) = idx;
1708 : : // Update linked chunk index entries, if any exist.
1709 [ + + ]: 2 : if (entry.m_main_chunkindex_iterator != m_main_chunkindex.end()) {
1710 : 1 : entry.m_main_chunkindex_iterator->m_graph_index = idx;
1711 : : }
1712 : : // Update the locators for both levels. The rest of the Entry information will not change,
1713 : : // so no need to invoke Cluster::Updated().
1714 [ + + ]: 6 : for (int level = 0; level < MAX_LEVELS; ++level) {
1715 : 4 : Locator& locator = entry.m_locator[level];
1716 [ + + ]: 4 : if (locator.IsPresent()) {
1717 : 1 : locator.cluster->UpdateMapping(locator.index, idx);
1718 : : }
1719 : : }
1720 : : }
1721 : 1 : }
1722 : :
1723 : 329 : void TxGraphImpl::Compact() noexcept
1724 : : {
1725 : : // We cannot compact while any to-be-applied operations or staged removals remain as we'd need
1726 : : // to rewrite them. It is easier to delay the compaction until they have been applied.
1727 [ + + ]: 329 : if (!m_main_clusterset.m_deps_to_add.empty()) return;
1728 [ + + ]: 109 : if (!m_main_clusterset.m_to_remove.empty()) return;
1729 : 104 : Assume(m_main_clusterset.m_removed.empty()); // non-staging m_removed is always empty
1730 [ - + ]: 104 : if (m_staging_clusterset.has_value()) {
1731 [ # # ]: 0 : if (!m_staging_clusterset->m_deps_to_add.empty()) return;
1732 [ # # ]: 0 : if (!m_staging_clusterset->m_to_remove.empty()) return;
1733 [ # # ]: 0 : if (!m_staging_clusterset->m_removed.empty()) return;
1734 : : }
1735 : :
1736 : : // Release memory used by discarded ChunkData index entries.
1737 : 104 : ClearShrink(m_main_chunkindex_discarded);
1738 : :
1739 : : // Sort the GraphIndexes that need to be cleaned up. They are sorted in reverse, so the last
1740 : : // ones get processed first. This means earlier-processed GraphIndexes will not cause moving of
1741 : : // later-processed ones during the "swap with end of m_entries" step below (which might
1742 : : // invalidate them).
1743 : 104 : std::sort(m_unlinked.begin(), m_unlinked.end(), std::greater{});
1744 : :
1745 : 104 : auto last = GraphIndex(-1);
1746 [ + + ]: 105 : for (GraphIndex idx : m_unlinked) {
1747 : : // m_unlinked should never contain the same GraphIndex twice (the code below would fail
1748 : : // if so, because GraphIndexes get invalidated by removing them).
1749 : 1 : Assume(idx != last);
1750 : 1 : last = idx;
1751 : :
1752 : : // Make sure the entry is unlinked.
1753 : 1 : Entry& entry = m_entries[idx];
1754 : 1 : Assume(entry.m_ref == nullptr);
1755 : : // Make sure the entry does not occur in the graph.
1756 [ + + ]: 3 : for (int level = 0; level < MAX_LEVELS; ++level) {
1757 : 2 : Assume(!entry.m_locator[level].IsPresent());
1758 : : }
1759 : :
1760 : : // Move the entry to the end.
1761 [ - + + - ]: 1 : if (idx != m_entries.size() - 1) SwapIndexes(idx, m_entries.size() - 1);
1762 : : // Drop the entry for idx, now that it is at the end.
1763 : 1 : m_entries.pop_back();
1764 : : }
1765 [ + + ]: 104 : m_unlinked.clear();
1766 : : }
1767 : :
1768 : 0 : void TxGraphImpl::Split(Cluster& cluster, int level) noexcept
1769 : : {
1770 : : // To split a Cluster, first make sure all removals are applied (as we might need to split
1771 : : // again afterwards otherwise).
1772 : 0 : ApplyRemovals(level);
1773 : 0 : bool del = cluster.Split(*this, level);
1774 [ # # ]: 0 : if (del) {
1775 : : // Cluster::Split reports whether the Cluster is to be deleted.
1776 : 0 : DeleteCluster(cluster, level);
1777 : : }
1778 : 0 : }
1779 : :
1780 : 5 : void TxGraphImpl::SplitAll(int up_to_level) noexcept
1781 : : {
1782 [ + - ]: 5 : Assume(up_to_level >= 0 && up_to_level <= GetTopLevel());
1783 : : // Before splitting all Cluster, first make sure all removals are applied.
1784 : 5 : ApplyRemovals(up_to_level);
1785 [ + + ]: 10 : for (int level = 0; level <= up_to_level; ++level) {
1786 [ + + ]: 15 : for (auto quality : {QualityLevel::NEEDS_SPLIT, QualityLevel::NEEDS_SPLIT_ACCEPTABLE}) {
1787 : 10 : auto& queue = GetClusterSet(level).m_clusters[int(quality)];
1788 [ - + ]: 10 : while (!queue.empty()) {
1789 : 0 : Split(*queue.back().get(), level);
1790 : : }
1791 : : }
1792 : : }
1793 : 5 : }
1794 : :
1795 : 8 : void TxGraphImpl::GroupClusters(int level) noexcept
1796 : : {
1797 : 8 : auto& clusterset = GetClusterSet(level);
1798 : : // If the groupings have been computed already, nothing is left to be done.
1799 [ + + ]: 8 : if (clusterset.m_group_data.has_value()) return;
1800 : :
1801 : : // Before computing which Clusters need to be merged together, first apply all removals and
1802 : : // split the Clusters into connected components. If we would group first, we might end up
1803 : : // with inefficient and/or oversized Clusters which just end up being split again anyway.
1804 : 5 : SplitAll(level);
1805 : :
1806 : : /** Annotated clusters: an entry for each Cluster, together with the sequence number for the
1807 : : * representative for the partition it is in (initially its own, later that of the
1808 : : * to-be-merged group). */
1809 : 5 : std::vector<std::pair<Cluster*, uint64_t>> an_clusters;
1810 : : /** Annotated dependencies: an entry for each m_deps_to_add entry (excluding ones that apply
1811 : : * to removed transactions), together with the sequence number of the representative root of
1812 : : * Clusters it applies to (initially that of the child Cluster, later that of the
1813 : : * to-be-merged group). */
1814 : 5 : std::vector<std::pair<std::pair<GraphIndex, GraphIndex>, uint64_t>> an_deps;
1815 : :
1816 : : // Construct an an_clusters entry for every oversized cluster, including ones from levels below,
1817 : : // as they may be inherited in this one.
1818 [ + + ]: 10 : for (int level_iter = 0; level_iter <= level; ++level_iter) {
1819 [ + + ]: 11 : for (auto& cluster : GetClusterSet(level_iter).m_clusters[int(QualityLevel::OVERSIZED_SINGLETON)]) {
1820 : 6 : auto graph_idx = cluster->GetClusterEntry(0);
1821 : 6 : auto cur_cluster = FindCluster(graph_idx, level);
1822 [ - + ]: 6 : if (cur_cluster == nullptr) continue;
1823 : 6 : an_clusters.emplace_back(cur_cluster, cur_cluster->m_sequence);
1824 : : }
1825 : : }
1826 : :
1827 : : // Construct a an_clusters entry for every parent and child in the to-be-applied dependencies,
1828 : : // and an an_deps entry for each dependency to be applied.
1829 [ - + ]: 5 : an_deps.reserve(clusterset.m_deps_to_add.size());
1830 [ + - + + ]: 127213 : for (const auto& [par, chl] : clusterset.m_deps_to_add) {
1831 : 127208 : auto par_cluster = FindCluster(par, level);
1832 : 127208 : auto chl_cluster = FindCluster(chl, level);
1833 : : // Skip dependencies for which the parent or child transaction is removed.
1834 [ + - - + ]: 127208 : if (par_cluster == nullptr || chl_cluster == nullptr) continue;
1835 : 127208 : an_clusters.emplace_back(par_cluster, par_cluster->m_sequence);
1836 : : // Do not include a duplicate when parent and child are identical, as it'll be removed
1837 : : // below anyway.
1838 [ + - ]: 127208 : if (chl_cluster != par_cluster) an_clusters.emplace_back(chl_cluster, chl_cluster->m_sequence);
1839 : : // Add entry to an_deps, using the child sequence number.
1840 : 127208 : an_deps.emplace_back(std::pair{par, chl}, chl_cluster->m_sequence);
1841 : : }
1842 : : // Sort and deduplicate an_clusters, so we end up with a sorted list of all involved Clusters
1843 : : // to which dependencies apply, or which are oversized.
1844 [ + + + + : 4849133 : std::sort(an_clusters.begin(), an_clusters.end(), [](auto& a, auto& b) noexcept { return a.second < b.second; });
+ + + + +
+ + + + +
+ + + + +
+ - - +
+ ]
1845 : 5 : an_clusters.erase(std::unique(an_clusters.begin(), an_clusters.end()), an_clusters.end());
1846 : : // Sort an_deps by applying the same order to the involved child cluster.
1847 [ - - - - : 2312193 : std::sort(an_deps.begin(), an_deps.end(), [&](auto& a, auto& b) noexcept { return a.second < b.second; });
+ + + + +
+ + + + +
+ + + + +
+ - - +
+ ]
1848 : :
1849 : : // Run the union-find algorithm to to find partitions of the input Clusters which need to be
1850 : : // grouped together. See https://en.wikipedia.org/wiki/Disjoint-set_data_structure.
1851 : 5 : {
1852 : : /** Each PartitionData entry contains information about a single input Cluster. */
1853 : 5 : struct PartitionData
1854 : : {
1855 : : /** The sequence number of the cluster this holds information for. */
1856 : : uint64_t sequence;
1857 : : /** All PartitionData entries belonging to the same partition are organized in a tree.
1858 : : * Each element points to its parent, or to itself if it is the root. The root is then
1859 : : * a representative for the entire tree, and can be found by walking upwards from any
1860 : : * element. */
1861 : : PartitionData* parent;
1862 : : /** (only if this is a root, so when parent == this) An upper bound on the height of
1863 : : * tree for this partition. */
1864 : : unsigned rank;
1865 : : };
1866 : : /** Information about each input Cluster. Sorted by Cluster::m_sequence. */
1867 : 5 : std::vector<PartitionData> partition_data;
1868 : :
1869 : : /** Given a Cluster, find its corresponding PartitionData. */
1870 : 253274 : auto locate_fn = [&](uint64_t sequence) noexcept -> PartitionData* {
1871 : 253269 : auto it = std::lower_bound(partition_data.begin(), partition_data.end(), sequence,
1872 [ + + ]: 4043971 : [](auto& a, uint64_t seq) noexcept { return a.sequence < seq; });
1873 : 253269 : Assume(it != partition_data.end());
1874 : 253269 : Assume(it->sequence == sequence);
1875 : 253269 : return &*it;
1876 : 5 : };
1877 : :
1878 : : /** Given a PartitionData, find the root of the tree it is in (its representative). */
1879 : 255430 : static constexpr auto find_root_fn = [](PartitionData* data) noexcept -> PartitionData* {
1880 [ + + + + ]: 447428 : while (data->parent != data) {
1881 : : // Replace pointers to parents with pointers to grandparents.
1882 : : // See https://en.wikipedia.org/wiki/Disjoint-set_data_structure#Finding_set_representatives.
1883 : 314003 : auto par = data->parent;
1884 : 314003 : data->parent = par->parent;
1885 : 314003 : data = par;
1886 : : }
1887 : 255425 : return data;
1888 : : };
1889 : :
1890 : : /** Given two PartitionDatas, union the partitions they are in, and return their
1891 : : * representative. */
1892 : 127213 : static constexpr auto union_fn = [](PartitionData* arg1, PartitionData* arg2) noexcept {
1893 : : // Find the roots of the trees, and bail out if they are already equal (which would
1894 : : // mean they are in the same partition already).
1895 [ + + ]: 376416 : auto rep1 = find_root_fn(arg1);
1896 : 127208 : auto rep2 = find_root_fn(arg2);
1897 [ + - ]: 127208 : if (rep1 == rep2) return rep1;
1898 : : // Pick the lower-rank root to become a child of the higher-rank one.
1899 : : // See https://en.wikipedia.org/wiki/Disjoint-set_data_structure#Union_by_rank.
1900 [ + + ]: 127208 : if (rep1->rank < rep2->rank) std::swap(rep1, rep2);
1901 : 127208 : rep2->parent = rep1;
1902 : 127208 : rep1->rank += (rep1->rank == rep2->rank);
1903 : 127208 : return rep1;
1904 : : };
1905 : :
1906 : : // Start by initializing every Cluster as its own singleton partition.
1907 [ - + ]: 5 : partition_data.resize(an_clusters.size());
1908 [ - + + + ]: 128222 : for (size_t i = 0; i < an_clusters.size(); ++i) {
1909 : 128217 : partition_data[i].sequence = an_clusters[i].first->m_sequence;
1910 : 128217 : partition_data[i].parent = &partition_data[i];
1911 : 128217 : partition_data[i].rank = 0;
1912 : : }
1913 : :
1914 : : // Run through all parent/child pairs in an_deps, and union the partitions their Clusters
1915 : : // are in.
1916 : 5 : Cluster* last_chl_cluster{nullptr};
1917 : 5 : PartitionData* last_partition{nullptr};
1918 [ - + + + ]: 127213 : for (const auto& [dep, _] : an_deps) {
1919 [ - + ]: 127208 : auto [par, chl] = dep;
1920 : 127208 : auto par_cluster = FindCluster(par, level);
1921 : 127208 : auto chl_cluster = FindCluster(chl, level);
1922 : 127208 : Assume(chl_cluster != nullptr && par_cluster != nullptr);
1923 : : // Nothing to do if parent and child are in the same Cluster.
1924 [ - + ]: 127208 : if (par_cluster == chl_cluster) continue;
1925 : 127208 : Assume(par != chl);
1926 [ + + ]: 127208 : if (chl_cluster == last_chl_cluster) {
1927 : : // If the child Clusters is the same as the previous iteration, union with the
1928 : : // tree they were in, avoiding the need for another lookup. Note that an_deps
1929 : : // is sorted by child Cluster, so batches with the same child are expected.
1930 : 1147 : last_partition = union_fn(locate_fn(par_cluster->m_sequence), last_partition);
1931 : : } else {
1932 : 126061 : last_chl_cluster = chl_cluster;
1933 : 126061 : last_partition = union_fn(locate_fn(par_cluster->m_sequence), locate_fn(chl_cluster->m_sequence));
1934 : : }
1935 : : }
1936 : :
1937 : : // Update the sequence numbers in an_clusters and an_deps to be those of the partition
1938 : : // representative.
1939 : 5 : auto deps_it = an_deps.begin();
1940 [ - + + + ]: 128222 : for (size_t i = 0; i < partition_data.size(); ++i) {
1941 : 128217 : auto& data = partition_data[i];
1942 : : // Find the sequence of the representative of the partition Cluster i is in, and store
1943 : : // it with the Cluster.
1944 : 128217 : auto rep_seq = find_root_fn(&data)->sequence;
1945 : 128217 : an_clusters[i].second = rep_seq;
1946 : : // Find all dependencies whose child Cluster is Cluster i, and annotate them with rep.
1947 [ + + ]: 255425 : while (deps_it != an_deps.end()) {
1948 [ + + ]: 255264 : auto [par, chl] = deps_it->first;
1949 : 255264 : auto chl_cluster = FindCluster(chl, level);
1950 : 255264 : Assume(chl_cluster != nullptr);
1951 [ + + ]: 255264 : if (chl_cluster->m_sequence > data.sequence) break;
1952 : 127208 : deps_it->second = rep_seq;
1953 : 127208 : ++deps_it;
1954 : : }
1955 : : }
1956 : 5 : }
1957 : :
1958 : : // Sort both an_clusters and an_deps by sequence number of the representative of the
1959 : : // partition they are in, grouping all those applying to the same partition together.
1960 [ - - - - : 1767204 : std::sort(an_deps.begin(), an_deps.end(), [](auto& a, auto& b) noexcept { return a.second < b.second; });
+ + + + +
+ + + + +
+ + + + +
+ - - -
+ ]
1961 [ - - - - : 1763821 : std::sort(an_clusters.begin(), an_clusters.end(), [](auto& a, auto& b) noexcept { return a.second < b.second; });
+ + + + +
+ + + + +
+ + + + -
+ - - -
+ ]
1962 : :
1963 : : // Translate the resulting cluster groups to the m_group_data structure, and the dependencies
1964 : : // back to m_deps_to_add.
1965 : 5 : clusterset.m_group_data = GroupData{};
1966 [ - + ]: 5 : clusterset.m_group_data->m_group_clusters.reserve(an_clusters.size());
1967 [ + + ]: 5 : clusterset.m_deps_to_add.clear();
1968 [ - + ]: 5 : clusterset.m_deps_to_add.reserve(an_deps.size());
1969 : 5 : clusterset.m_oversized = false;
1970 : 5 : auto an_deps_it = an_deps.begin();
1971 : 5 : auto an_clusters_it = an_clusters.begin();
1972 [ + + ]: 1014 : while (an_clusters_it != an_clusters.end()) {
1973 : : // Process all clusters/dependencies belonging to the partition with representative rep.
1974 : 1009 : auto rep = an_clusters_it->second;
1975 : : // Create and initialize a new GroupData entry for the partition.
1976 : 1009 : auto& new_entry = clusterset.m_group_data->m_groups.emplace_back();
1977 [ - + ]: 1009 : new_entry.m_cluster_offset = clusterset.m_group_data->m_group_clusters.size();
1978 : 1009 : new_entry.m_cluster_count = 0;
1979 [ - + ]: 1009 : new_entry.m_deps_offset = clusterset.m_deps_to_add.size();
1980 : 1009 : new_entry.m_deps_count = 0;
1981 : 1009 : uint32_t total_count{0};
1982 : 1009 : uint64_t total_size{0};
1983 : : // Add all its clusters to it (copying those from an_clusters to m_group_clusters).
1984 [ + + + + ]: 129226 : while (an_clusters_it != an_clusters.end() && an_clusters_it->second == rep) {
1985 : 128217 : clusterset.m_group_data->m_group_clusters.push_back(an_clusters_it->first);
1986 : 128217 : total_count += an_clusters_it->first->GetTxCount();
1987 : 128217 : total_size += an_clusters_it->first->GetTotalTxSize();
1988 : 128217 : ++an_clusters_it;
1989 : 128217 : ++new_entry.m_cluster_count;
1990 : : }
1991 : : // Add all its dependencies to it (copying those back from an_deps to m_deps_to_add).
1992 [ + + + + ]: 128217 : while (an_deps_it != an_deps.end() && an_deps_it->second == rep) {
1993 : 127208 : clusterset.m_deps_to_add.push_back(an_deps_it->first);
1994 : 127208 : ++an_deps_it;
1995 : 127208 : ++new_entry.m_deps_count;
1996 : : }
1997 : : // Detect oversizedness.
1998 [ + + + + ]: 1009 : if (total_count > m_max_cluster_count || total_size > m_max_cluster_size) {
1999 : 9 : clusterset.m_oversized = true;
2000 : : }
2001 : : }
2002 : 5 : Assume(an_deps_it == an_deps.end());
2003 : 5 : Assume(an_clusters_it == an_clusters.end());
2004 : 5 : Compact();
2005 : 5 : }
2006 : :
2007 : 0 : void TxGraphImpl::Merge(std::span<Cluster*> to_merge, int level) noexcept
2008 : : {
2009 [ # # ]: 0 : Assume(!to_merge.empty());
2010 : : // Nothing to do if a group consists of just a single Cluster.
2011 [ # # ]: 0 : if (to_merge.size() == 1) return;
2012 : :
2013 : : // Move the largest Cluster to the front of to_merge. As all transactions in other to-be-merged
2014 : : // Clusters will be moved to that one, putting the largest one first minimizes the number of
2015 : : // moves.
2016 : 0 : size_t max_size_pos{0};
2017 : 0 : DepGraphIndex max_size = to_merge[max_size_pos]->GetTxCount();
2018 : 0 : GetClusterSet(level).m_cluster_usage -= to_merge[max_size_pos]->TotalMemoryUsage();
2019 : 0 : DepGraphIndex total_size = max_size;
2020 [ # # ]: 0 : for (size_t i = 1; i < to_merge.size(); ++i) {
2021 : 0 : GetClusterSet(level).m_cluster_usage -= to_merge[i]->TotalMemoryUsage();
2022 : 0 : DepGraphIndex size = to_merge[i]->GetTxCount();
2023 : 0 : total_size += size;
2024 [ # # ]: 0 : if (size > max_size) {
2025 : 0 : max_size_pos = i;
2026 : 0 : max_size = size;
2027 : : }
2028 : : }
2029 [ # # ]: 0 : if (max_size_pos != 0) std::swap(to_merge[0], to_merge[max_size_pos]);
2030 : :
2031 : 0 : size_t start_idx = 1;
2032 : 0 : Cluster* into_cluster = to_merge[0];
2033 [ # # ]: 0 : if (total_size > into_cluster->GetMaxTxCount()) {
2034 : : // The into_merge cluster is too small to fit all transactions being merged. Construct a
2035 : : // a new Cluster using an implementation that matches the total size, and merge everything
2036 : : // in there.
2037 : 0 : auto new_cluster = CreateEmptyCluster(total_size);
2038 : 0 : into_cluster = new_cluster.get();
2039 : 0 : InsertCluster(level, std::move(new_cluster), QualityLevel::OPTIMAL);
2040 : 0 : start_idx = 0;
2041 : 0 : }
2042 : :
2043 : : // Merge all further Clusters in the group into the result (first one, or new one), and delete
2044 : : // them.
2045 [ # # ]: 0 : for (size_t i = start_idx; i < to_merge.size(); ++i) {
2046 : 0 : into_cluster->Merge(*this, level, *to_merge[i]);
2047 : 0 : DeleteCluster(*to_merge[i], level);
2048 : : }
2049 : 0 : into_cluster->Compact();
2050 : 0 : GetClusterSet(level).m_cluster_usage += into_cluster->TotalMemoryUsage();
2051 : : }
2052 : :
2053 : 0 : void TxGraphImpl::ApplyDependencies(int level) noexcept
2054 : : {
2055 : 0 : auto& clusterset = GetClusterSet(level);
2056 : : // Do not bother computing groups if we already know the result will be oversized.
2057 [ # # ]: 0 : if (clusterset.m_oversized == true) return;
2058 : : // Compute the groups of to-be-merged Clusters (which also applies all removals, and splits).
2059 : 0 : GroupClusters(level);
2060 [ # # ]: 0 : Assume(clusterset.m_group_data.has_value());
2061 : : // Nothing to do if there are no dependencies to be added.
2062 [ # # ]: 0 : if (clusterset.m_deps_to_add.empty()) return;
2063 : : // Dependencies cannot be applied if it would result in oversized clusters.
2064 [ # # ]: 0 : if (clusterset.m_oversized == true) return;
2065 : :
2066 : : // For each group of to-be-merged Clusters.
2067 [ # # ]: 0 : for (const auto& group_entry : clusterset.m_group_data->m_groups) {
2068 [ # # ]: 0 : auto cluster_span = std::span{clusterset.m_group_data->m_group_clusters}
2069 [ # # ]: 0 : .subspan(group_entry.m_cluster_offset, group_entry.m_cluster_count);
2070 : : // Pull in all the Clusters that contain dependencies.
2071 [ # # ]: 0 : if (level == 1) {
2072 [ # # ]: 0 : for (Cluster*& cluster : cluster_span) {
2073 : 0 : cluster = PullIn(cluster, cluster->GetLevel(*this));
2074 : : }
2075 : : }
2076 : : // Invoke Merge() to merge them into a single Cluster.
2077 : 0 : Merge(cluster_span, level);
2078 : : // Actually apply all to-be-added dependencies (all parents and children from this grouping
2079 : : // belong to the same Cluster at this point because of the merging above).
2080 [ # # ]: 0 : auto deps_span = std::span{clusterset.m_deps_to_add}
2081 : 0 : .subspan(group_entry.m_deps_offset, group_entry.m_deps_count);
2082 : 0 : Assume(!deps_span.empty());
2083 : 0 : const auto& loc = m_entries[deps_span[0].second].m_locator[level];
2084 : 0 : Assume(loc.IsPresent());
2085 : 0 : loc.cluster->ApplyDependencies(*this, level, deps_span);
2086 : : }
2087 : :
2088 : : // Wipe the list of to-be-added dependencies now that they are applied.
2089 [ # # ]: 0 : clusterset.m_deps_to_add.clear();
2090 : 0 : Compact();
2091 : : // Also no further Cluster mergings are needed (note that we clear, but don't set to
2092 : : // std::nullopt, as that would imply the groupings are unknown).
2093 : 0 : clusterset.m_group_data = GroupData{};
2094 : : }
2095 : :
2096 : 0 : std::pair<uint64_t, bool> GenericClusterImpl::Relinearize(TxGraphImpl& graph, int level, uint64_t max_iters) noexcept
2097 : : {
2098 : : // We can only relinearize Clusters that do not need splitting.
2099 : 0 : Assume(!NeedsSplitting());
2100 : : // No work is required for Clusters which are already optimally linearized.
2101 [ # # ]: 0 : if (IsOptimal()) return {0, false};
2102 : : // Invoke the actual linearization algorithm (passing in the existing one).
2103 : 0 : uint64_t rng_seed = graph.m_rng.rand64();
2104 [ # # # # ]: 0 : auto [linearization, optimal, cost] = Linearize(m_depgraph, max_iters, rng_seed, m_linearization);
2105 : : // Postlinearize if the result isn't optimal already. This guarantees (among other things)
2106 : : // that the chunks of the resulting linearization are all connected.
2107 [ # # # # ]: 0 : if (!optimal) PostLinearize(m_depgraph, linearization);
2108 : : // Update the linearization.
2109 : 0 : m_linearization = std::move(linearization);
2110 : : // Update the Cluster's quality.
2111 : 0 : bool improved = false;
2112 [ # # ]: 0 : if (optimal) {
2113 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::OPTIMAL);
2114 : 0 : improved = true;
2115 [ # # # # ]: 0 : } else if (max_iters >= graph.m_acceptable_iters && !IsAcceptable()) {
2116 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::ACCEPTABLE);
2117 : 0 : improved = true;
2118 : : }
2119 : : // Update the Entry objects.
2120 : 0 : Updated(graph, level);
2121 : 0 : return {cost, improved};
2122 : 0 : }
2123 : :
2124 : 0 : std::pair<uint64_t, bool> SingletonClusterImpl::Relinearize(TxGraphImpl& graph, int level, uint64_t max_iters) noexcept
2125 : : {
2126 : : // All singletons are optimal, oversized, or need splitting. Each of these precludes
2127 : : // Relinearize from being called.
2128 : 0 : assert(false);
2129 : : return {0, false};
2130 : : }
2131 : :
2132 : 0 : void TxGraphImpl::MakeAcceptable(Cluster& cluster, int level) noexcept
2133 : : {
2134 : : // Relinearize the Cluster if needed.
2135 [ # # # # ]: 0 : if (!cluster.NeedsSplitting() && !cluster.IsAcceptable() && !cluster.IsOversized()) {
2136 : 0 : cluster.Relinearize(*this, level, m_acceptable_iters);
2137 : : }
2138 : 0 : }
2139 : :
2140 : 0 : void TxGraphImpl::MakeAllAcceptable(int level) noexcept
2141 : : {
2142 : 0 : ApplyDependencies(level);
2143 : 0 : auto& clusterset = GetClusterSet(level);
2144 [ # # ]: 0 : if (clusterset.m_oversized == true) return;
2145 : 0 : auto& queue = clusterset.m_clusters[int(QualityLevel::NEEDS_RELINEARIZE)];
2146 [ # # ]: 0 : while (!queue.empty()) {
2147 : 0 : MakeAcceptable(*queue.back().get(), level);
2148 : : }
2149 : : }
2150 : :
2151 : 0 : GenericClusterImpl::GenericClusterImpl(uint64_t sequence) noexcept : Cluster{sequence} {}
2152 : :
2153 : 64311 : TxGraph::Ref TxGraphImpl::AddTransaction(const FeePerWeight& feerate) noexcept
2154 : : {
2155 [ - + ]: 64311 : Assume(m_main_chunkindex_observers == 0 || GetTopLevel() != 0);
2156 : 64311 : Assume(feerate.size > 0);
2157 : : // Construct a new Ref.
2158 : 64311 : Ref ret;
2159 : : // Construct a new Entry, and link it with the Ref.
2160 [ - + ]: 64311 : auto idx = m_entries.size();
2161 : 64311 : m_entries.emplace_back();
2162 : 64311 : auto& entry = m_entries.back();
2163 : 64311 : entry.m_main_chunkindex_iterator = m_main_chunkindex.end();
2164 : 64311 : entry.m_ref = &ret;
2165 : 64311 : GetRefGraph(ret) = this;
2166 : 64311 : GetRefIndex(ret) = idx;
2167 : : // Construct a new singleton Cluster (which is necessarily optimally linearized).
2168 : 64311 : bool oversized = uint64_t(feerate.size) > m_max_cluster_size;
2169 : 64311 : auto cluster = CreateEmptyCluster(1);
2170 : 64311 : cluster->AppendTransaction(idx, feerate);
2171 [ + + ]: 64311 : auto cluster_ptr = cluster.get();
2172 [ + + ]: 64311 : int level = GetTopLevel();
2173 : 64311 : auto& clusterset = GetClusterSet(level);
2174 [ + + ]: 128616 : InsertCluster(level, std::move(cluster), oversized ? QualityLevel::OVERSIZED_SINGLETON : QualityLevel::OPTIMAL);
2175 : 64311 : cluster_ptr->Updated(*this, level);
2176 : 64311 : clusterset.m_cluster_usage += cluster_ptr->TotalMemoryUsage();
2177 : 64311 : ++clusterset.m_txcount;
2178 : : // Deal with individually oversized transactions.
2179 [ + + ]: 64311 : if (oversized) {
2180 : 6 : ++clusterset.m_txcount_oversized;
2181 : 6 : clusterset.m_oversized = true;
2182 [ + + ]: 6 : clusterset.m_group_data = std::nullopt;
2183 : : }
2184 : : // Return the Ref.
2185 : 128622 : return ret;
2186 : 64311 : }
2187 : :
2188 : 0 : void TxGraphImpl::RemoveTransaction(const Ref& arg) noexcept
2189 : : {
2190 : : // Don't do anything if the Ref is empty (which may be indicative of the transaction already
2191 : : // having been removed).
2192 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return;
2193 : 0 : Assume(GetRefGraph(arg) == this);
2194 [ # # ]: 0 : Assume(m_main_chunkindex_observers == 0 || GetTopLevel() != 0);
2195 : : // Find the Cluster the transaction is in, and stop if it isn't in any.
2196 [ # # ]: 0 : int level = GetTopLevel();
2197 [ # # ]: 0 : auto cluster = FindCluster(GetRefIndex(arg), level);
2198 [ # # ]: 0 : if (cluster == nullptr) return;
2199 : : // Remember that the transaction is to be removed.
2200 : 0 : auto& clusterset = GetClusterSet(level);
2201 : 0 : clusterset.m_to_remove.push_back(GetRefIndex(arg));
2202 : : // Wipe m_group_data (as it will need to be recomputed).
2203 [ # # ]: 0 : clusterset.m_group_data.reset();
2204 [ # # ]: 0 : if (clusterset.m_oversized == true) clusterset.m_oversized = std::nullopt;
2205 : : }
2206 : :
2207 : 64208 : void TxGraphImpl::AddDependency(const Ref& parent, const Ref& child) noexcept
2208 : : {
2209 : : // Don't do anything if either Ref is empty (which may be indicative of it having already been
2210 : : // removed).
2211 [ + - + - ]: 64208 : if (GetRefGraph(parent) == nullptr || GetRefGraph(child) == nullptr) return;
2212 : 64208 : Assume(GetRefGraph(parent) == this && GetRefGraph(child) == this);
2213 [ - + ]: 64208 : Assume(m_main_chunkindex_observers == 0 || GetTopLevel() != 0);
2214 : : // Don't do anything if this is a dependency on self.
2215 [ + - ]: 64208 : if (GetRefIndex(parent) == GetRefIndex(child)) return;
2216 : : // Find the Cluster the parent and child transaction are in, and stop if either appears to be
2217 : : // already removed.
2218 [ + - ]: 64208 : int level = GetTopLevel();
2219 : 64208 : auto par_cluster = FindCluster(GetRefIndex(parent), level);
2220 [ + - ]: 64208 : if (par_cluster == nullptr) return;
2221 : 64208 : auto chl_cluster = FindCluster(GetRefIndex(child), level);
2222 [ + - ]: 64208 : if (chl_cluster == nullptr) return;
2223 : : // Remember that this dependency is to be applied.
2224 : 64208 : auto& clusterset = GetClusterSet(level);
2225 : 64208 : clusterset.m_deps_to_add.emplace_back(GetRefIndex(parent), GetRefIndex(child));
2226 : : // Wipe m_group_data (as it will need to be recomputed).
2227 [ + + ]: 64208 : clusterset.m_group_data.reset();
2228 [ + + ]: 64212 : if (clusterset.m_oversized == false) clusterset.m_oversized = std::nullopt;
2229 : : }
2230 : :
2231 : 300 : bool TxGraphImpl::Exists(const Ref& arg, Level level_select) noexcept
2232 : : {
2233 [ + - ]: 300 : if (GetRefGraph(arg) == nullptr) return false;
2234 : 300 : Assume(GetRefGraph(arg) == this);
2235 [ + - ]: 300 : size_t level = GetSpecifiedLevel(level_select);
2236 : : // Make sure the transaction isn't scheduled for removal.
2237 : 300 : ApplyRemovals(level);
2238 : 300 : auto cluster = FindCluster(GetRefIndex(arg), level);
2239 : 300 : return cluster != nullptr;
2240 : : }
2241 : :
2242 : 0 : void GenericClusterImpl::GetAncestorRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept
2243 : : {
2244 : : /** The union of all ancestors to be returned. */
2245 : 0 : SetType ancestors_union;
2246 : : // Process elements from the front of args, as long as they apply.
2247 [ # # ]: 0 : while (!args.empty()) {
2248 [ # # ]: 0 : if (args.front().first != this) break;
2249 : 0 : ancestors_union |= m_depgraph.Ancestors(args.front().second);
2250 : 0 : args = args.subspan(1);
2251 : : }
2252 [ # # ]: 0 : Assume(ancestors_union.Any());
2253 : : // Translate all ancestors (in arbitrary order) to Refs (if they have any), and return them.
2254 [ # # # # : 0 : for (auto idx : ancestors_union) {
# # ]
2255 : 0 : const auto& entry = graph.m_entries[m_mapping[idx]];
2256 : 0 : Assume(entry.m_ref != nullptr);
2257 : 0 : output.push_back(entry.m_ref);
2258 : : }
2259 : 0 : }
2260 : :
2261 : 0 : void SingletonClusterImpl::GetAncestorRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept
2262 : : {
2263 : 0 : Assume(GetTxCount());
2264 [ # # ]: 0 : while (!args.empty()) {
2265 [ # # ]: 0 : if (args.front().first != this) break;
2266 : 0 : args = args.subspan(1);
2267 : : }
2268 : 0 : const auto& entry = graph.m_entries[m_graph_index];
2269 : 0 : Assume(entry.m_ref != nullptr);
2270 : 0 : output.push_back(entry.m_ref);
2271 : 0 : }
2272 : :
2273 : 0 : void GenericClusterImpl::GetDescendantRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept
2274 : : {
2275 : : /** The union of all descendants to be returned. */
2276 : 0 : SetType descendants_union;
2277 : : // Process elements from the front of args, as long as they apply.
2278 [ # # ]: 0 : while (!args.empty()) {
2279 [ # # ]: 0 : if (args.front().first != this) break;
2280 : 0 : descendants_union |= m_depgraph.Descendants(args.front().second);
2281 : 0 : args = args.subspan(1);
2282 : : }
2283 [ # # ]: 0 : Assume(descendants_union.Any());
2284 : : // Translate all descendants (in arbitrary order) to Refs (if they have any), and return them.
2285 [ # # # # : 0 : for (auto idx : descendants_union) {
# # ]
2286 : 0 : const auto& entry = graph.m_entries[m_mapping[idx]];
2287 : 0 : Assume(entry.m_ref != nullptr);
2288 : 0 : output.push_back(entry.m_ref);
2289 : : }
2290 : 0 : }
2291 : :
2292 : 0 : void SingletonClusterImpl::GetDescendantRefs(const TxGraphImpl& graph, std::span<std::pair<Cluster*, DepGraphIndex>>& args, std::vector<TxGraph::Ref*>& output) noexcept
2293 : : {
2294 : : // In a singleton cluster, the ancestors or descendants are always just the entire cluster.
2295 : 0 : GetAncestorRefs(graph, args, output);
2296 : 0 : }
2297 : :
2298 : 0 : bool GenericClusterImpl::GetClusterRefs(TxGraphImpl& graph, std::span<TxGraph::Ref*> range, LinearizationIndex start_pos) noexcept
2299 : : {
2300 : : // Translate the transactions in the Cluster (in linearization order, starting at start_pos in
2301 : : // the linearization) to Refs, and fill them in range.
2302 [ # # ]: 0 : for (auto& ref : range) {
2303 [ # # ]: 0 : Assume(start_pos < m_linearization.size());
2304 : 0 : const auto& entry = graph.m_entries[m_mapping[m_linearization[start_pos++]]];
2305 : 0 : Assume(entry.m_ref != nullptr);
2306 : 0 : ref = entry.m_ref;
2307 : : }
2308 : : // Return whether start_pos has advanced to the end of the Cluster.
2309 [ # # ]: 0 : return start_pos == m_linearization.size();
2310 : : }
2311 : :
2312 : 0 : bool SingletonClusterImpl::GetClusterRefs(TxGraphImpl& graph, std::span<TxGraph::Ref*> range, LinearizationIndex start_pos) noexcept
2313 : : {
2314 : 0 : Assume(!range.empty());
2315 : 0 : Assume(GetTxCount());
2316 : 0 : Assume(start_pos == 0);
2317 : 0 : const auto& entry = graph.m_entries[m_graph_index];
2318 : 0 : Assume(entry.m_ref != nullptr);
2319 : 0 : range[0] = entry.m_ref;
2320 : 0 : return true;
2321 : : }
2322 : :
2323 : 0 : FeePerWeight GenericClusterImpl::GetIndividualFeerate(DepGraphIndex idx) noexcept
2324 : : {
2325 : 0 : return FeePerWeight::FromFeeFrac(m_depgraph.FeeRate(idx));
2326 : : }
2327 : :
2328 : 0 : FeePerWeight SingletonClusterImpl::GetIndividualFeerate(DepGraphIndex idx) noexcept
2329 : : {
2330 : 0 : Assume(GetTxCount());
2331 : 0 : Assume(idx == 0);
2332 : 0 : return m_feerate;
2333 : : }
2334 : :
2335 : 0 : void GenericClusterImpl::MakeStagingTransactionsMissing(TxGraphImpl& graph) noexcept
2336 : : {
2337 : : // Mark all transactions of a Cluster missing, needed when aborting staging, so that the
2338 : : // corresponding Locators don't retain references into aborted Clusters.
2339 [ # # ]: 0 : for (auto ci : m_linearization) {
2340 : 0 : GraphIndex idx = m_mapping[ci];
2341 : 0 : auto& entry = graph.m_entries[idx];
2342 : 0 : entry.m_locator[1].SetMissing();
2343 : : }
2344 : 0 : }
2345 : :
2346 : 0 : void SingletonClusterImpl::MakeStagingTransactionsMissing(TxGraphImpl& graph) noexcept
2347 : : {
2348 [ # # ]: 0 : if (GetTxCount()) {
2349 : 0 : auto& entry = graph.m_entries[m_graph_index];
2350 : 0 : entry.m_locator[1].SetMissing();
2351 : : }
2352 : 0 : }
2353 : :
2354 : 0 : std::vector<TxGraph::Ref*> TxGraphImpl::GetAncestors(const Ref& arg, Level level_select) noexcept
2355 : : {
2356 : : // Return the empty vector if the Ref is empty.
2357 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return {};
2358 : 0 : Assume(GetRefGraph(arg) == this);
2359 : : // Apply all removals and dependencies, as the result might be incorrect otherwise.
2360 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2361 : 0 : ApplyDependencies(level);
2362 : : // Ancestry cannot be known if unapplied dependencies remain.
2363 : 0 : Assume(GetClusterSet(level).m_deps_to_add.empty());
2364 : : // Find the Cluster the argument is in, and return the empty vector if it isn't in any.
2365 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(arg), level);
2366 [ # # ]: 0 : if (cluster == nullptr) return {};
2367 : : // Dispatch to the Cluster.
2368 : 0 : std::pair<Cluster*, DepGraphIndex> match = {cluster, m_entries[GetRefIndex(arg)].m_locator[cluster_level].index};
2369 : 0 : auto matches = std::span(&match, 1);
2370 : 0 : std::vector<TxGraph::Ref*> ret;
2371 : 0 : cluster->GetAncestorRefs(*this, matches, ret);
2372 : 0 : return ret;
2373 : 0 : }
2374 : :
2375 : 0 : std::vector<TxGraph::Ref*> TxGraphImpl::GetDescendants(const Ref& arg, Level level_select) noexcept
2376 : : {
2377 : : // Return the empty vector if the Ref is empty.
2378 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return {};
2379 : 0 : Assume(GetRefGraph(arg) == this);
2380 : : // Apply all removals and dependencies, as the result might be incorrect otherwise.
2381 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2382 : 0 : ApplyDependencies(level);
2383 : : // Ancestry cannot be known if unapplied dependencies remain.
2384 : 0 : Assume(GetClusterSet(level).m_deps_to_add.empty());
2385 : : // Find the Cluster the argument is in, and return the empty vector if it isn't in any.
2386 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(arg), level);
2387 [ # # ]: 0 : if (cluster == nullptr) return {};
2388 : : // Dispatch to the Cluster.
2389 : 0 : std::pair<Cluster*, DepGraphIndex> match = {cluster, m_entries[GetRefIndex(arg)].m_locator[cluster_level].index};
2390 : 0 : auto matches = std::span(&match, 1);
2391 : 0 : std::vector<TxGraph::Ref*> ret;
2392 : 0 : cluster->GetDescendantRefs(*this, matches, ret);
2393 : 0 : return ret;
2394 : 0 : }
2395 : :
2396 : 0 : std::vector<TxGraph::Ref*> TxGraphImpl::GetAncestorsUnion(std::span<const Ref* const> args, Level level_select) noexcept
2397 : : {
2398 : : // Apply all dependencies, as the result might be incorrect otherwise.
2399 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2400 : 0 : ApplyDependencies(level);
2401 : : // Ancestry cannot be known if unapplied dependencies remain.
2402 : 0 : Assume(GetClusterSet(level).m_deps_to_add.empty());
2403 : :
2404 : : // Translate args to matches.
2405 : 0 : std::vector<std::pair<Cluster*, DepGraphIndex>> matches;
2406 : 0 : matches.reserve(args.size());
2407 [ # # ]: 0 : for (auto arg : args) {
2408 : 0 : Assume(arg);
2409 : : // Skip empty Refs.
2410 [ # # ]: 0 : if (GetRefGraph(*arg) == nullptr) continue;
2411 : 0 : Assume(GetRefGraph(*arg) == this);
2412 : : // Find the Cluster the argument is in, and skip if none is found.
2413 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(*arg), level);
2414 [ # # ]: 0 : if (cluster == nullptr) continue;
2415 : : // Append to matches.
2416 : 0 : matches.emplace_back(cluster, m_entries[GetRefIndex(*arg)].m_locator[cluster_level].index);
2417 : : }
2418 : : // Group by Cluster.
2419 : 0 : std::sort(matches.begin(), matches.end(), [](auto& a, auto& b) noexcept { return CompareClusters(a.first, b.first) < 0; });
2420 : : // Dispatch to the Clusters.
2421 [ # # ]: 0 : std::span match_span(matches);
2422 : 0 : std::vector<TxGraph::Ref*> ret;
2423 [ # # ]: 0 : while (!match_span.empty()) {
2424 : 0 : match_span.front().first->GetAncestorRefs(*this, match_span, ret);
2425 : : }
2426 : 0 : return ret;
2427 : 0 : }
2428 : :
2429 : 0 : std::vector<TxGraph::Ref*> TxGraphImpl::GetDescendantsUnion(std::span<const Ref* const> args, Level level_select) noexcept
2430 : : {
2431 : : // Apply all dependencies, as the result might be incorrect otherwise.
2432 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2433 : 0 : ApplyDependencies(level);
2434 : : // Ancestry cannot be known if unapplied dependencies remain.
2435 : 0 : Assume(GetClusterSet(level).m_deps_to_add.empty());
2436 : :
2437 : : // Translate args to matches.
2438 : 0 : std::vector<std::pair<Cluster*, DepGraphIndex>> matches;
2439 : 0 : matches.reserve(args.size());
2440 [ # # ]: 0 : for (auto arg : args) {
2441 : 0 : Assume(arg);
2442 : : // Skip empty Refs.
2443 [ # # ]: 0 : if (GetRefGraph(*arg) == nullptr) continue;
2444 : 0 : Assume(GetRefGraph(*arg) == this);
2445 : : // Find the Cluster the argument is in, and skip if none is found.
2446 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(*arg), level);
2447 [ # # ]: 0 : if (cluster == nullptr) continue;
2448 : : // Append to matches.
2449 : 0 : matches.emplace_back(cluster, m_entries[GetRefIndex(*arg)].m_locator[cluster_level].index);
2450 : : }
2451 : : // Group by Cluster.
2452 : 0 : std::sort(matches.begin(), matches.end(), [](auto& a, auto& b) noexcept { return CompareClusters(a.first, b.first) < 0; });
2453 : : // Dispatch to the Clusters.
2454 [ # # ]: 0 : std::span match_span(matches);
2455 : 0 : std::vector<TxGraph::Ref*> ret;
2456 [ # # ]: 0 : while (!match_span.empty()) {
2457 : 0 : match_span.front().first->GetDescendantRefs(*this, match_span, ret);
2458 : : }
2459 : 0 : return ret;
2460 : 0 : }
2461 : :
2462 : 0 : std::vector<TxGraph::Ref*> TxGraphImpl::GetCluster(const Ref& arg, Level level_select) noexcept
2463 : : {
2464 : : // Return the empty vector if the Ref is empty (which may be indicative of the transaction
2465 : : // having been removed already.
2466 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return {};
2467 : 0 : Assume(GetRefGraph(arg) == this);
2468 : : // Apply all removals and dependencies, as the result might be incorrect otherwise.
2469 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2470 : 0 : ApplyDependencies(level);
2471 : : // Cluster linearization cannot be known if unapplied dependencies remain.
2472 : 0 : Assume(GetClusterSet(level).m_deps_to_add.empty());
2473 : : // Find the Cluster the argument is in, and return the empty vector if it isn't in any.
2474 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(arg), level);
2475 [ # # ]: 0 : if (cluster == nullptr) return {};
2476 : : // Make sure the Cluster has an acceptable quality level, and then dispatch to it.
2477 : 0 : MakeAcceptable(*cluster, cluster_level);
2478 [ # # ]: 0 : std::vector<TxGraph::Ref*> ret(cluster->GetTxCount());
2479 [ # # ]: 0 : cluster->GetClusterRefs(*this, ret, 0);
2480 : 0 : return ret;
2481 : 0 : }
2482 : :
2483 : 7 : TxGraph::GraphIndex TxGraphImpl::GetTransactionCount(Level level_select) noexcept
2484 : : {
2485 [ + - ]: 7 : size_t level = GetSpecifiedLevel(level_select);
2486 : 7 : ApplyRemovals(level);
2487 : 7 : return GetClusterSet(level).m_txcount;
2488 : : }
2489 : :
2490 : 0 : FeePerWeight TxGraphImpl::GetIndividualFeerate(const Ref& arg) noexcept
2491 : : {
2492 : : // Return the empty FeePerWeight if the passed Ref is empty.
2493 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return {};
2494 : : Assume(GetRefGraph(arg) == this);
2495 : : // Find the cluster the argument is in (the level does not matter as individual feerates will
2496 : : // be identical if it occurs in both), and return the empty FeePerWeight if it isn't in any.
2497 : 0 : Cluster* cluster{nullptr};
2498 : : int level;
2499 [ # # ]: 0 : for (level = 0; level <= GetTopLevel(); ++level) {
2500 : : // Apply removals, so that we can correctly report FeePerWeight{} for non-existing
2501 : : // transactions.
2502 : 0 : ApplyRemovals(level);
2503 [ # # ]: 0 : if (m_entries[GetRefIndex(arg)].m_locator[level].IsPresent()) {
2504 : : cluster = m_entries[GetRefIndex(arg)].m_locator[level].cluster;
2505 : : break;
2506 : : }
2507 : : }
2508 [ # # ]: 0 : if (cluster == nullptr) return {};
2509 : : // Dispatch to the Cluster.
2510 : 0 : return cluster->GetIndividualFeerate(m_entries[GetRefIndex(arg)].m_locator[level].index);
2511 : : }
2512 : :
2513 : 0 : FeePerWeight TxGraphImpl::GetMainChunkFeerate(const Ref& arg) noexcept
2514 : : {
2515 : : // Return the empty FeePerWeight if the passed Ref is empty.
2516 [ # # ]: 0 : if (GetRefGraph(arg) == nullptr) return {};
2517 : 0 : Assume(GetRefGraph(arg) == this);
2518 : : // Apply all removals and dependencies, as the result might be inaccurate otherwise.
2519 : 0 : ApplyDependencies(/*level=*/0);
2520 : : // Chunk feerates cannot be accurately known if unapplied dependencies remain.
2521 : 0 : Assume(m_main_clusterset.m_deps_to_add.empty());
2522 : : // Find the cluster the argument is in, and return the empty FeePerWeight if it isn't in any.
2523 [ # # ]: 0 : auto [cluster, cluster_level] = FindClusterAndLevel(GetRefIndex(arg), /*level=*/0);
2524 [ # # ]: 0 : if (cluster == nullptr) return {};
2525 : : // Make sure the Cluster has an acceptable quality level, and then return the transaction's
2526 : : // chunk feerate.
2527 : 0 : MakeAcceptable(*cluster, cluster_level);
2528 : 0 : const auto& entry = m_entries[GetRefIndex(arg)];
2529 : 0 : return entry.m_main_chunk_feerate;
2530 : : }
2531 : :
2532 : 9 : bool TxGraphImpl::IsOversized(Level level_select) noexcept
2533 : : {
2534 [ + - ]: 9 : size_t level = GetSpecifiedLevel(level_select);
2535 : 9 : auto& clusterset = GetClusterSet(level);
2536 [ + + ]: 9 : if (clusterset.m_oversized.has_value()) {
2537 : : // Return cached value if known.
2538 : 5 : return *clusterset.m_oversized;
2539 : : }
2540 : 4 : ApplyRemovals(level);
2541 [ - + ]: 4 : if (clusterset.m_txcount_oversized > 0) {
2542 : 0 : clusterset.m_oversized = true;
2543 : : } else {
2544 : : // Find which Clusters will need to be merged together, as that is where the oversize
2545 : : // property is assessed.
2546 : 4 : GroupClusters(level);
2547 : : }
2548 : 4 : Assume(clusterset.m_oversized.has_value());
2549 : 4 : return *clusterset.m_oversized;
2550 : : }
2551 : :
2552 : 0 : void TxGraphImpl::StartStaging() noexcept
2553 : : {
2554 : : // Staging cannot already exist.
2555 : 0 : Assume(!m_staging_clusterset.has_value());
2556 : : // Apply all remaining dependencies in main before creating a staging graph. Once staging
2557 : : // exists, we cannot merge Clusters anymore (because of interference with Clusters being
2558 : : // pulled into staging), so to make sure all inspectors are available (if not oversized), do
2559 : : // all merging work now. Call SplitAll() first, so that even if ApplyDependencies does not do
2560 : : // any thing due to knowing the result is oversized, splitting is still performed.
2561 : 0 : SplitAll(0);
2562 : 0 : ApplyDependencies(0);
2563 : : // Construct the staging ClusterSet.
2564 : 0 : m_staging_clusterset.emplace();
2565 : : // Copy statistics, precomputed data, and to-be-applied dependencies (only if oversized) to
2566 : : // the new graph. To-be-applied removals will always be empty at this point.
2567 : 0 : m_staging_clusterset->m_txcount = m_main_clusterset.m_txcount;
2568 : 0 : m_staging_clusterset->m_txcount_oversized = m_main_clusterset.m_txcount_oversized;
2569 : 0 : m_staging_clusterset->m_deps_to_add = m_main_clusterset.m_deps_to_add;
2570 : 0 : m_staging_clusterset->m_group_data = m_main_clusterset.m_group_data;
2571 : 0 : m_staging_clusterset->m_oversized = m_main_clusterset.m_oversized;
2572 : 0 : Assume(m_staging_clusterset->m_oversized.has_value());
2573 : 0 : m_staging_clusterset->m_cluster_usage = 0;
2574 : 0 : }
2575 : :
2576 : 0 : void TxGraphImpl::AbortStaging() noexcept
2577 : : {
2578 : : // Staging must exist.
2579 : 0 : Assume(m_staging_clusterset.has_value());
2580 : : // Mark all removed transactions as Missing (so the staging locator for these transactions
2581 : : // can be reused if another staging is created).
2582 [ # # ]: 0 : for (auto idx : m_staging_clusterset->m_removed) {
2583 : 0 : m_entries[idx].m_locator[1].SetMissing();
2584 : : }
2585 : : // Do the same with the non-removed transactions in staging Clusters.
2586 [ # # ]: 0 : for (int quality = 0; quality < int(QualityLevel::NONE); ++quality) {
2587 [ # # ]: 0 : for (auto& cluster : m_staging_clusterset->m_clusters[quality]) {
2588 : 0 : cluster->MakeStagingTransactionsMissing(*this);
2589 : : }
2590 : : }
2591 : : // Destroy the staging ClusterSet.
2592 : 0 : m_staging_clusterset.reset();
2593 : 0 : Compact();
2594 [ # # ]: 0 : if (!m_main_clusterset.m_group_data.has_value()) {
2595 : : // In case m_oversized in main was kept after a Ref destruction while staging exists, we
2596 : : // need to re-evaluate m_oversized now.
2597 [ # # # # ]: 0 : if (m_main_clusterset.m_to_remove.empty() && m_main_clusterset.m_txcount_oversized > 0) {
2598 : : // It is possible that a Ref destruction caused a removal in main while staging existed.
2599 : : // In this case, m_txcount_oversized may be inaccurate.
2600 : 0 : m_main_clusterset.m_oversized = true;
2601 : : } else {
2602 [ # # ]: 0 : m_main_clusterset.m_oversized = std::nullopt;
2603 : : }
2604 : : }
2605 : 0 : }
2606 : :
2607 : 0 : void TxGraphImpl::CommitStaging() noexcept
2608 : : {
2609 : : // Staging must exist.
2610 : 0 : Assume(m_staging_clusterset.has_value());
2611 : 0 : Assume(m_main_chunkindex_observers == 0);
2612 : : // Delete all conflicting Clusters in main, to make place for moving the staging ones
2613 : : // there. All of these have been copied to staging in PullIn().
2614 : 0 : auto conflicts = GetConflicts();
2615 [ # # ]: 0 : for (Cluster* conflict : conflicts) {
2616 : 0 : conflict->Clear(*this, /*level=*/0);
2617 : 0 : DeleteCluster(*conflict, /*level=*/0);
2618 : : }
2619 : : // Mark the removed transactions as Missing (so the staging locator for these transactions
2620 : : // can be reused if another staging is created).
2621 [ # # ]: 0 : for (auto idx : m_staging_clusterset->m_removed) {
2622 : 0 : m_entries[idx].m_locator[1].SetMissing();
2623 : : }
2624 : : // Then move all Clusters in staging to main.
2625 [ # # ]: 0 : for (int quality = 0; quality < int(QualityLevel::NONE); ++quality) {
2626 : 0 : auto& stage_sets = m_staging_clusterset->m_clusters[quality];
2627 [ # # ]: 0 : while (!stage_sets.empty()) {
2628 : 0 : stage_sets.back()->MoveToMain(*this);
2629 : : }
2630 : : }
2631 : : // Move all statistics, precomputed data, and to-be-applied removals and dependencies.
2632 : 0 : m_main_clusterset.m_deps_to_add = std::move(m_staging_clusterset->m_deps_to_add);
2633 : 0 : m_main_clusterset.m_to_remove = std::move(m_staging_clusterset->m_to_remove);
2634 : 0 : m_main_clusterset.m_group_data = std::move(m_staging_clusterset->m_group_data);
2635 : 0 : m_main_clusterset.m_oversized = std::move(m_staging_clusterset->m_oversized);
2636 : 0 : m_main_clusterset.m_txcount = std::move(m_staging_clusterset->m_txcount);
2637 : 0 : m_main_clusterset.m_txcount_oversized = std::move(m_staging_clusterset->m_txcount_oversized);
2638 : : // Delete the old staging graph, after all its information was moved to main.
2639 : 0 : m_staging_clusterset.reset();
2640 : 0 : Compact();
2641 : 0 : }
2642 : :
2643 : 0 : void GenericClusterImpl::SetFee(TxGraphImpl& graph, int level, DepGraphIndex idx, int64_t fee) noexcept
2644 : : {
2645 : : // Make sure the specified DepGraphIndex exists in this Cluster.
2646 : 0 : Assume(m_depgraph.Positions()[idx]);
2647 : : // Bail out if the fee isn't actually being changed.
2648 [ # # ]: 0 : if (m_depgraph.FeeRate(idx).fee == fee) return;
2649 : : // Update the fee, remember that relinearization will be necessary, and update the Entries
2650 : : // in the same Cluster.
2651 [ # # ]: 0 : m_depgraph.FeeRate(idx).fee = fee;
2652 [ # # ]: 0 : if (m_quality == QualityLevel::OVERSIZED_SINGLETON) {
2653 : : // Nothing to do.
2654 [ # # ]: 0 : } else if (!NeedsSplitting()) {
2655 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::NEEDS_RELINEARIZE);
2656 : : } else {
2657 : 0 : graph.SetClusterQuality(level, m_quality, m_setindex, QualityLevel::NEEDS_SPLIT);
2658 : : }
2659 : 0 : Updated(graph, level);
2660 : : }
2661 : :
2662 : 0 : void SingletonClusterImpl::SetFee(TxGraphImpl& graph, int level, DepGraphIndex idx, int64_t fee) noexcept
2663 : : {
2664 : 0 : Assume(GetTxCount());
2665 : 0 : Assume(idx == 0);
2666 : 0 : m_feerate.fee = fee;
2667 : 0 : Updated(graph, level);
2668 : 0 : }
2669 : :
2670 : 0 : void TxGraphImpl::SetTransactionFee(const Ref& ref, int64_t fee) noexcept
2671 : : {
2672 : : // Don't do anything if the passed Ref is empty.
2673 [ # # ]: 0 : if (GetRefGraph(ref) == nullptr) return;
2674 : 0 : Assume(GetRefGraph(ref) == this);
2675 : 0 : Assume(m_main_chunkindex_observers == 0);
2676 : : // Find the entry, its locator, and inform its Cluster about the new feerate, if any.
2677 : 0 : auto& entry = m_entries[GetRefIndex(ref)];
2678 [ # # ]: 0 : for (int level = 0; level < MAX_LEVELS; ++level) {
2679 : 0 : auto& locator = entry.m_locator[level];
2680 [ # # ]: 0 : if (locator.IsPresent()) {
2681 : 0 : locator.cluster->SetFee(*this, level, locator.index, fee);
2682 : : }
2683 : : }
2684 : : }
2685 : :
2686 : 0 : std::strong_ordering TxGraphImpl::CompareMainOrder(const Ref& a, const Ref& b) noexcept
2687 : : {
2688 : : // The references must not be empty.
2689 : 0 : Assume(GetRefGraph(a) == this);
2690 : 0 : Assume(GetRefGraph(b) == this);
2691 : : // Apply dependencies in main.
2692 : 0 : ApplyDependencies(0);
2693 : 0 : Assume(m_main_clusterset.m_deps_to_add.empty());
2694 : : // Make both involved Clusters acceptable, so chunk feerates are relevant.
2695 : 0 : const auto& entry_a = m_entries[GetRefIndex(a)];
2696 : 0 : const auto& entry_b = m_entries[GetRefIndex(b)];
2697 : 0 : const auto& locator_a = entry_a.m_locator[0];
2698 : 0 : const auto& locator_b = entry_b.m_locator[0];
2699 : 0 : Assume(locator_a.IsPresent());
2700 : 0 : Assume(locator_b.IsPresent());
2701 : 0 : MakeAcceptable(*locator_a.cluster, /*level=*/0);
2702 : 0 : MakeAcceptable(*locator_b.cluster, /*level=*/0);
2703 : : // Invoke comparison logic.
2704 : 0 : return CompareMainTransactions(GetRefIndex(a), GetRefIndex(b));
2705 : : }
2706 : :
2707 : 0 : TxGraph::GraphIndex TxGraphImpl::CountDistinctClusters(std::span<const Ref* const> refs, Level level_select) noexcept
2708 : : {
2709 [ # # ]: 0 : size_t level = GetSpecifiedLevel(level_select);
2710 : 0 : ApplyDependencies(level);
2711 : 0 : auto& clusterset = GetClusterSet(level);
2712 : 0 : Assume(clusterset.m_deps_to_add.empty());
2713 : : // Build a vector of Clusters that the specified Refs occur in.
2714 : 0 : std::vector<Cluster*> clusters;
2715 : 0 : clusters.reserve(refs.size());
2716 [ # # ]: 0 : for (const Ref* ref : refs) {
2717 : 0 : Assume(ref);
2718 [ # # ]: 0 : if (GetRefGraph(*ref) == nullptr) continue;
2719 : 0 : Assume(GetRefGraph(*ref) == this);
2720 [ # # ]: 0 : auto cluster = FindCluster(GetRefIndex(*ref), level);
2721 [ # # ]: 0 : if (cluster != nullptr) clusters.push_back(cluster);
2722 : : }
2723 : : // Count the number of distinct elements in clusters.
2724 : 0 : std::sort(clusters.begin(), clusters.end(), [](Cluster* a, Cluster* b) noexcept { return CompareClusters(a, b) < 0; });
2725 : 0 : Cluster* last{nullptr};
2726 : 0 : GraphIndex ret{0};
2727 [ # # ]: 0 : for (Cluster* cluster : clusters) {
2728 : 0 : ret += (cluster != last);
2729 : 0 : last = cluster;
2730 : : }
2731 : 0 : return ret;
2732 : 0 : }
2733 : :
2734 : 0 : std::pair<std::vector<FeeFrac>, std::vector<FeeFrac>> TxGraphImpl::GetMainStagingDiagrams() noexcept
2735 : : {
2736 : 0 : Assume(m_staging_clusterset.has_value());
2737 : 0 : MakeAllAcceptable(0);
2738 : 0 : Assume(m_main_clusterset.m_deps_to_add.empty()); // can only fail if main is oversized
2739 : 0 : MakeAllAcceptable(1);
2740 : 0 : Assume(m_staging_clusterset->m_deps_to_add.empty()); // can only fail if staging is oversized
2741 : : // For all Clusters in main which conflict with Clusters in staging (i.e., all that are removed
2742 : : // by, or replaced in, staging), gather their chunk feerates.
2743 : 0 : auto main_clusters = GetConflicts();
2744 : 0 : std::vector<FeeFrac> main_feerates, staging_feerates;
2745 [ # # ]: 0 : for (Cluster* cluster : main_clusters) {
2746 : 0 : cluster->AppendChunkFeerates(main_feerates);
2747 : : }
2748 : : // Do the same for the Clusters in staging themselves.
2749 [ # # ]: 0 : for (int quality = 0; quality < int(QualityLevel::NONE); ++quality) {
2750 [ # # ]: 0 : for (const auto& cluster : m_staging_clusterset->m_clusters[quality]) {
2751 : 0 : cluster->AppendChunkFeerates(staging_feerates);
2752 : : }
2753 : : }
2754 : : // Sort both by decreasing feerate to obtain diagrams, and return them.
2755 [ # # # # : 0 : std::sort(main_feerates.begin(), main_feerates.end(), [](auto& a, auto& b) { return a > b; });
# # # # #
# # # # #
# # # # #
# # # #
# ]
2756 [ # # # # : 0 : std::sort(staging_feerates.begin(), staging_feerates.end(), [](auto& a, auto& b) { return a > b; });
# # # # #
# # # # #
# # # # #
# # # #
# ]
2757 : 0 : return std::make_pair(std::move(main_feerates), std::move(staging_feerates));
2758 : 0 : }
2759 : :
2760 : 0 : void GenericClusterImpl::SanityCheck(const TxGraphImpl& graph, int level) const
2761 : : {
2762 : : // There must be an m_mapping for each m_depgraph position (including holes).
2763 [ # # # # : 0 : assert(m_depgraph.PositionRange() == m_mapping.size());
# # ]
2764 : : // The linearization for this Cluster must contain every transaction once.
2765 [ # # # # ]: 0 : assert(m_depgraph.TxCount() == m_linearization.size());
2766 : : // Unless a split is to be applied, the cluster cannot have any holes.
2767 [ # # ]: 0 : if (!NeedsSplitting()) {
2768 [ # # # # ]: 0 : assert(m_depgraph.Positions() == SetType::Fill(m_depgraph.TxCount()));
2769 : : }
2770 : :
2771 : : // Compute the chunking of m_linearization.
2772 : 0 : LinearizationChunking linchunking(m_depgraph, m_linearization);
2773 : :
2774 : : // Verify m_linearization.
2775 : 0 : SetType m_done;
2776 : 0 : LinearizationIndex linindex{0};
2777 : 0 : DepGraphIndex chunk_pos{0}; //!< position within the current chunk
2778 [ # # ]: 0 : assert(m_depgraph.IsAcyclic());
2779 [ # # ]: 0 : for (auto lin_pos : m_linearization) {
2780 [ # # # # ]: 0 : assert(lin_pos < m_mapping.size());
2781 [ # # ]: 0 : const auto& entry = graph.m_entries[m_mapping[lin_pos]];
2782 : : // Check that the linearization is topological.
2783 [ # # ]: 0 : m_done.Set(lin_pos);
2784 [ # # ]: 0 : assert(m_done.IsSupersetOf(m_depgraph.Ancestors(lin_pos)));
2785 : : // Check that the Entry has a locator pointing back to this Cluster & position within it.
2786 [ # # ]: 0 : assert(entry.m_locator[level].cluster == this);
2787 [ # # ]: 0 : assert(entry.m_locator[level].index == lin_pos);
2788 : : // For main-level entries, check linearization position and chunk feerate.
2789 [ # # # # ]: 0 : if (level == 0 && IsAcceptable()) {
2790 [ # # ]: 0 : assert(entry.m_main_lin_index == linindex);
2791 : 0 : ++linindex;
2792 [ # # # # ]: 0 : if (!linchunking.GetChunk(0).transactions[lin_pos]) {
2793 : 0 : linchunking.MarkDone(linchunking.GetChunk(0).transactions);
2794 : 0 : chunk_pos = 0;
2795 : : }
2796 [ # # # # ]: 0 : assert(entry.m_main_chunk_feerate == linchunking.GetChunk(0).feerate);
2797 : : // Verify that an entry in the chunk index exists for every chunk-ending transaction.
2798 : 0 : ++chunk_pos;
2799 [ # # ]: 0 : bool is_chunk_end = (chunk_pos == linchunking.GetChunk(0).transactions.Count());
2800 [ # # ]: 0 : assert((entry.m_main_chunkindex_iterator != graph.m_main_chunkindex.end()) == is_chunk_end);
2801 [ # # ]: 0 : if (is_chunk_end) {
2802 [ # # ]: 0 : auto& chunk_data = *entry.m_main_chunkindex_iterator;
2803 [ # # # # ]: 0 : if (m_done == m_depgraph.Positions() && chunk_pos == 1) {
2804 [ # # ]: 0 : assert(chunk_data.m_chunk_count == LinearizationIndex(-1));
2805 : : } else {
2806 [ # # ]: 0 : assert(chunk_data.m_chunk_count == chunk_pos);
2807 : : }
2808 : : }
2809 : : // If this Cluster has an acceptable quality level, its chunks must be connected.
2810 [ # # ]: 0 : assert(m_depgraph.IsConnected(linchunking.GetChunk(0).transactions));
2811 : : }
2812 : : }
2813 : : // Verify that each element of m_depgraph occurred in m_linearization.
2814 [ # # ]: 0 : assert(m_done == m_depgraph.Positions());
2815 : 0 : }
2816 : :
2817 : 192822 : void SingletonClusterImpl::SanityCheck(const TxGraphImpl& graph, int level) const
2818 : : {
2819 : : // All singletons are optimal, oversized, or need splitting.
2820 [ + + + + ]: 192822 : Assume(IsOptimal() || IsOversized() || NeedsSplitting());
2821 [ + + ]: 192822 : if (GetTxCount()) {
2822 [ - + ]: 192811 : const auto& entry = graph.m_entries[m_graph_index];
2823 : : // Check that the Entry has a locator pointing back to this Cluster & position within it.
2824 [ - + ]: 192811 : assert(entry.m_locator[level].cluster == this);
2825 [ - + ]: 192811 : assert(entry.m_locator[level].index == 0);
2826 : : // For main-level entries, check linearization position and chunk feerate.
2827 [ + - + + ]: 192811 : if (level == 0 && IsAcceptable()) {
2828 [ - + ]: 192799 : assert(entry.m_main_lin_index == 0);
2829 [ + - ]: 192799 : assert(entry.m_main_chunk_feerate == m_feerate);
2830 [ - + ]: 192799 : assert(entry.m_main_chunkindex_iterator != graph.m_main_chunkindex.end());
2831 [ - + ]: 192799 : auto& chunk_data = *entry.m_main_chunkindex_iterator;
2832 [ - + ]: 192799 : assert(chunk_data.m_chunk_count == LinearizationIndex(-1));
2833 : : }
2834 : : }
2835 : 192822 : }
2836 : :
2837 : 11 : void TxGraphImpl::SanityCheck() const
2838 : : {
2839 : : /** Which GraphIndexes ought to occur in m_unlinked, based on m_entries. */
2840 : 11 : std::set<GraphIndex> expected_unlinked;
2841 : : /** Which Clusters ought to occur in ClusterSet::m_clusters, based on m_entries. */
2842 [ + + ]: 55 : std::set<const Cluster*> expected_clusters[MAX_LEVELS];
2843 : : /** Which GraphIndexes ought to occur in ClusterSet::m_removed, based on m_entries. */
2844 [ + + ]: 55 : std::set<GraphIndex> expected_removed[MAX_LEVELS];
2845 : : /** Which Cluster::m_sequence values have been encountered. */
2846 : 11 : std::set<uint64_t> sequences;
2847 : : /** Which GraphIndexes ought to occur in m_main_chunkindex, based on m_entries. */
2848 : 11 : std::set<GraphIndex> expected_chunkindex;
2849 : : /** Whether compaction is possible in the current state. */
2850 : 11 : bool compact_possible{true};
2851 : :
2852 : : // Go over all Entry objects in m_entries.
2853 [ - + + + ]: 192833 : for (GraphIndex idx = 0; idx < m_entries.size(); ++idx) {
2854 [ - + ]: 192822 : const auto& entry = m_entries[idx];
2855 [ - + ]: 192822 : if (entry.m_ref == nullptr) {
2856 : : // Unlinked Entry must have indexes appear in m_unlinked.
2857 [ # # ]: 0 : expected_unlinked.insert(idx);
2858 : : } else {
2859 : : // Every non-unlinked Entry must have a Ref that points back to it.
2860 [ - + ]: 192822 : assert(GetRefGraph(*entry.m_ref) == this);
2861 [ - + ]: 192822 : assert(GetRefIndex(*entry.m_ref) == idx);
2862 : : }
2863 [ + + ]: 192822 : if (entry.m_main_chunkindex_iterator != m_main_chunkindex.end()) {
2864 : : // Remember which entries we see a chunkindex entry for.
2865 [ - + ]: 192799 : assert(entry.m_locator[0].IsPresent());
2866 [ + - ]: 192799 : expected_chunkindex.insert(idx);
2867 : : }
2868 : : // Verify the Entry m_locators.
2869 : : bool was_present{false}, was_removed{false};
2870 [ + + ]: 578466 : for (int level = 0; level < MAX_LEVELS; ++level) {
2871 : 385644 : const auto& locator = entry.m_locator[level];
2872 : : // Every Locator must be in exactly one of these 3 states.
2873 [ + + + + : 1156932 : assert(locator.IsMissing() + locator.IsRemoved() + locator.IsPresent() == 1);
- + ]
2874 [ + + ]: 385644 : if (locator.IsPresent()) {
2875 : : // Once removed, a transaction cannot be revived.
2876 [ - + ]: 192811 : assert(!was_removed);
2877 : : // Verify that the Cluster agrees with where the Locator claims the transaction is.
2878 [ - + ]: 192811 : assert(locator.cluster->GetClusterEntry(locator.index) == idx);
2879 : : // Remember that we expect said Cluster to appear in the ClusterSet::m_clusters.
2880 [ + - ]: 192811 : expected_clusters[level].insert(locator.cluster);
2881 : : was_present = true;
2882 [ - + ]: 385644 : } else if (locator.IsRemoved()) {
2883 : : // Level 0 (main) cannot have IsRemoved locators (IsMissing there means non-existing).
2884 [ # # ]: 0 : assert(level > 0);
2885 : : // A Locator can only be IsRemoved if it was IsPresent before, and only once.
2886 [ # # ]: 0 : assert(was_present && !was_removed);
2887 : : // Remember that we expect this GraphIndex to occur in the ClusterSet::m_removed.
2888 [ # # ]: 0 : expected_removed[level].insert(idx);
2889 : : was_removed = true;
2890 : : }
2891 : : }
2892 : : }
2893 : :
2894 : : // For all levels (0 = main, 1 = staged)...
2895 [ + + ]: 22 : for (int level = 0; level <= GetTopLevel(); ++level) {
2896 : 11 : assert(level < MAX_LEVELS);
2897 : 11 : auto& clusterset = GetClusterSet(level);
2898 : 11 : std::set<const Cluster*> actual_clusters;
2899 : 11 : size_t recomputed_cluster_usage{0};
2900 : :
2901 : : // For all quality levels...
2902 [ + + ]: 77 : for (int qual = 0; qual < int(QualityLevel::NONE); ++qual) {
2903 : 66 : QualityLevel quality{qual};
2904 : 66 : const auto& quality_clusters = clusterset.m_clusters[qual];
2905 : : // ... for all clusters in them ...
2906 [ - + + + ]: 192888 : for (ClusterSetIndex setindex = 0; setindex < quality_clusters.size(); ++setindex) {
2907 : 192822 : const auto& cluster = *quality_clusters[setindex];
2908 : : // The number of transactions in a Cluster cannot exceed m_max_cluster_count.
2909 [ - + ]: 192822 : assert(cluster.GetTxCount() <= m_max_cluster_count);
2910 : : // The level must match the Cluster's own idea of what level it is in (but GetLevel
2911 : : // can only be called for non-empty Clusters).
2912 [ + + - + ]: 192822 : assert(cluster.GetTxCount() == 0 || level == cluster.GetLevel(*this));
2913 : : // The sum of their sizes cannot exceed m_max_cluster_size, unless it is an
2914 : : // individually oversized transaction singleton. Note that groups of to-be-merged
2915 : : // clusters which would exceed this limit are marked oversized, which means they
2916 : : // are never applied.
2917 [ + + - + ]: 192822 : assert(cluster.IsOversized() || cluster.GetTotalTxSize() <= m_max_cluster_size);
2918 : : // OVERSIZED clusters are singletons.
2919 [ + + - + ]: 192822 : assert(!cluster.IsOversized() || cluster.GetTxCount() == 1);
2920 : : // Transaction counts cannot exceed the Cluster implementation's maximum
2921 : : // supported transactions count.
2922 [ - + ]: 192822 : assert(cluster.GetTxCount() <= cluster.GetMaxTxCount());
2923 : : // Unless a Split is yet to be applied, the number of transactions must not be
2924 : : // below the Cluster implementation's intended transaction count.
2925 [ + + ]: 192822 : if (!cluster.NeedsSplitting()) {
2926 [ - + ]: 192811 : assert(cluster.GetTxCount() >= cluster.GetMinIntendedTxCount());
2927 : : }
2928 : :
2929 : : // Check the sequence number.
2930 [ - + ]: 192822 : assert(cluster.m_sequence < m_next_sequence_counter);
2931 [ - + ]: 192822 : assert(sequences.count(cluster.m_sequence) == 0);
2932 [ + - ]: 192822 : sequences.insert(cluster.m_sequence);
2933 : : // Remember we saw this Cluster (only if it is non-empty; empty Clusters aren't
2934 : : // expected to be referenced by the Entry vector).
2935 [ + + ]: 192822 : if (cluster.GetTxCount() != 0) {
2936 [ + - ]: 192811 : actual_clusters.insert(&cluster);
2937 : : }
2938 : : // Sanity check the cluster, according to the Cluster's internal rules.
2939 [ + - ]: 192822 : cluster.SanityCheck(*this, level);
2940 : : // Check that the cluster's quality and setindex matches its position in the quality list.
2941 [ - + ]: 192822 : assert(cluster.m_quality == quality);
2942 [ - + ]: 192822 : assert(cluster.m_setindex == setindex);
2943 : : // Count memory usage.
2944 : 192822 : recomputed_cluster_usage += cluster.TotalMemoryUsage();
2945 : : }
2946 : : }
2947 : :
2948 : : // Verify memory usage.
2949 [ - + ]: 11 : assert(clusterset.m_cluster_usage == recomputed_cluster_usage);
2950 : :
2951 : : // Verify that all to-be-removed transactions have valid identifiers.
2952 [ + + ]: 19 : for (GraphIndex idx : clusterset.m_to_remove) {
2953 [ - + - + ]: 8 : assert(idx < m_entries.size());
2954 : : // We cannot assert that all m_to_remove transactions are still present: ~Ref on a
2955 : : // (P,M) transaction (present in main, inherited in staging) will cause an m_to_remove
2956 : : // addition in both main and staging, but a subsequence ApplyRemovals in main will
2957 : : // cause it to disappear from staging too, leaving the m_to_remove in place.
2958 : : }
2959 : :
2960 : : // Verify that all to-be-added dependencies have valid identifiers.
2961 [ - + + + ]: 191427 : for (auto [par_idx, chl_idx] : clusterset.m_deps_to_add) {
2962 [ - + ]: 191416 : assert(par_idx != chl_idx);
2963 [ - + - + ]: 191416 : assert(par_idx < m_entries.size());
2964 [ - + ]: 191416 : assert(chl_idx < m_entries.size());
2965 : : }
2966 : :
2967 : : // Verify that the actually encountered clusters match the ones occurring in Entry vector.
2968 [ - + ]: 11 : assert(actual_clusters == expected_clusters[level]);
2969 : :
2970 : : // Verify that the contents of m_removed matches what was expected based on the Entry vector.
2971 [ + - ]: 11 : std::set<GraphIndex> actual_removed(clusterset.m_removed.begin(), clusterset.m_removed.end());
2972 [ - + ]: 11 : for (auto i : expected_unlinked) {
2973 : : // If a transaction exists in both main and staging, and is removed from staging (adding
2974 : : // it to m_removed there), and consequently destroyed (wiping the locator completely),
2975 : : // it can remain in m_removed despite not having an IsRemoved() locator. Exclude those
2976 : : // transactions from the comparison here.
2977 : 0 : actual_removed.erase(i);
2978 : 0 : expected_removed[level].erase(i);
2979 : : }
2980 : :
2981 [ - + ]: 11 : assert(actual_removed == expected_removed[level]);
2982 : :
2983 : : // If any GraphIndex entries remain in this ClusterSet, compact is not possible.
2984 [ + + ]: 11 : if (!clusterset.m_deps_to_add.empty()) compact_possible = false;
2985 [ + + ]: 11 : if (!clusterset.m_to_remove.empty()) compact_possible = false;
2986 [ - + ]: 11 : if (!clusterset.m_removed.empty()) compact_possible = false;
2987 : :
2988 : : // For non-top levels, m_oversized must be known (as it cannot change until the level
2989 : : // on top is gone).
2990 [ - + - - ]: 11 : if (level < GetTopLevel()) assert(clusterset.m_oversized.has_value());
2991 : 11 : }
2992 : :
2993 : : // Verify that the contents of m_unlinked matches what was expected based on the Entry vector.
2994 [ + - ]: 11 : std::set<GraphIndex> actual_unlinked(m_unlinked.begin(), m_unlinked.end());
2995 [ - + ]: 11 : assert(actual_unlinked == expected_unlinked);
2996 : :
2997 : : // If compaction was possible, it should have been performed already, and m_unlinked must be
2998 : : // empty (to prevent memory leaks due to an ever-growing m_entries vector).
2999 [ + + ]: 11 : if (compact_possible) {
3000 [ - + ]: 3 : assert(actual_unlinked.empty());
3001 : : }
3002 : :
3003 : : // Finally, check the chunk index.
3004 : 11 : std::set<GraphIndex> actual_chunkindex;
3005 : 11 : FeeFrac last_chunk_feerate;
3006 [ + + ]: 192810 : for (const auto& chunk : m_main_chunkindex) {
3007 : 192799 : GraphIndex idx = chunk.m_graph_index;
3008 [ + - ]: 192799 : actual_chunkindex.insert(idx);
3009 [ + + ]: 192799 : auto chunk_feerate = m_entries[idx].m_main_chunk_feerate;
3010 [ + + ]: 192799 : if (!last_chunk_feerate.IsEmpty()) {
3011 [ - + ]: 192788 : assert(FeeRateCompare(last_chunk_feerate, chunk_feerate) >= 0);
3012 : : }
3013 : 192799 : last_chunk_feerate = chunk_feerate;
3014 : : }
3015 [ - + ]: 11 : assert(actual_chunkindex == expected_chunkindex);
3016 [ + + + + : 77 : }
- - - - ]
3017 : :
3018 : 0 : bool TxGraphImpl::DoWork(uint64_t iters) noexcept
3019 : : {
3020 : 0 : uint64_t iters_done{0};
3021 : : // First linearize everything in NEEDS_RELINEARIZE to an acceptable level. If more budget
3022 : : // remains after that, try to make everything optimal.
3023 [ # # ]: 0 : for (QualityLevel quality : {QualityLevel::NEEDS_RELINEARIZE, QualityLevel::ACCEPTABLE}) {
3024 : : // First linearize staging, if it exists, then main.
3025 [ # # ]: 0 : for (int level = GetTopLevel(); level >= 0; --level) {
3026 : : // Do not modify main if it has any observers.
3027 [ # # # # ]: 0 : if (level == 0 && m_main_chunkindex_observers != 0) continue;
3028 : 0 : ApplyDependencies(level);
3029 : 0 : auto& clusterset = GetClusterSet(level);
3030 : : // Do not modify oversized levels.
3031 [ # # ]: 0 : if (clusterset.m_oversized == true) continue;
3032 : 0 : auto& queue = clusterset.m_clusters[int(quality)];
3033 [ # # ]: 0 : while (!queue.empty()) {
3034 [ # # ]: 0 : if (iters_done >= iters) return false;
3035 : : // Randomize the order in which we process, so that if the first cluster somehow
3036 : : // needs more work than what iters allows, we don't keep spending it on the same
3037 : : // one.
3038 [ # # ]: 0 : auto pos = m_rng.randrange<size_t>(queue.size());
3039 : 0 : auto iters_now = iters - iters_done;
3040 [ # # ]: 0 : if (quality == QualityLevel::NEEDS_RELINEARIZE) {
3041 : : // If we're working with clusters that need relinearization still, only perform
3042 : : // up to m_acceptable_iters iterations. If they become ACCEPTABLE, and we still
3043 : : // have budget after all other clusters are ACCEPTABLE too, we'll spend the
3044 : : // remaining budget on trying to make them OPTIMAL.
3045 [ # # ]: 0 : iters_now = std::min(iters_now, m_acceptable_iters);
3046 : : }
3047 [ # # ]: 0 : auto [cost, improved] = queue[pos].get()->Relinearize(*this, level, iters_now);
3048 : 0 : iters_done += cost;
3049 : : // If no improvement was made to the Cluster, it means we've essentially run out of
3050 : : // budget. Even though it may be the case that iters_done < iters still, the
3051 : : // linearizer decided there wasn't enough budget left to attempt anything with.
3052 : : // To avoid an infinite loop that keeps trying clusters with minuscule budgets,
3053 : : // stop here too.
3054 [ # # ]: 0 : if (!improved) return false;
3055 : : }
3056 : : }
3057 : : }
3058 : : // All possible work has been performed, so we can return true. Note that this does *not* mean
3059 : : // that all clusters are optimally linearized now. It may be that there is nothing to do left
3060 : : // because all non-optimal clusters are in oversized and/or observer-bearing levels.
3061 : : return true;
3062 : : }
3063 : :
3064 : 0 : void BlockBuilderImpl::Next() noexcept
3065 : : {
3066 : : // Don't do anything if we're already done.
3067 [ # # ]: 0 : if (m_cur_iter == m_graph->m_main_chunkindex.end()) return;
3068 : 0 : while (true) {
3069 : : // Advance the pointer, and stop if we reach the end.
3070 [ # # ]: 0 : ++m_cur_iter;
3071 : 0 : m_cur_cluster = nullptr;
3072 [ # # ]: 0 : if (m_cur_iter == m_graph->m_main_chunkindex.end()) break;
3073 : : // Find the cluster pointed to by m_cur_iter.
3074 : 0 : const auto& chunk_data = *m_cur_iter;
3075 : 0 : const auto& chunk_end_entry = m_graph->m_entries[chunk_data.m_graph_index];
3076 : 0 : m_cur_cluster = chunk_end_entry.m_locator[0].cluster;
3077 : 0 : m_known_end_of_cluster = false;
3078 : : // If we previously skipped a chunk from this cluster we cannot include more from it.
3079 [ # # ]: 0 : if (!m_excluded_clusters.contains(m_cur_cluster->m_sequence)) break;
3080 : : }
3081 : : }
3082 : :
3083 : 0 : std::optional<std::pair<std::vector<TxGraph::Ref*>, FeePerWeight>> BlockBuilderImpl::GetCurrentChunk() noexcept
3084 : : {
3085 : 0 : std::optional<std::pair<std::vector<TxGraph::Ref*>, FeePerWeight>> ret;
3086 : : // Populate the return value if we are not done.
3087 [ # # ]: 0 : if (m_cur_iter != m_graph->m_main_chunkindex.end()) {
3088 : 0 : ret.emplace();
3089 [ # # ]: 0 : const auto& chunk_data = *m_cur_iter;
3090 [ # # ]: 0 : const auto& chunk_end_entry = m_graph->m_entries[chunk_data.m_graph_index];
3091 [ # # ]: 0 : if (chunk_data.m_chunk_count == LinearizationIndex(-1)) {
3092 : : // Special case in case just a single transaction remains, avoiding the need to
3093 : : // dispatch to and dereference Cluster.
3094 : 0 : ret->first.resize(1);
3095 : 0 : Assume(chunk_end_entry.m_ref != nullptr);
3096 : 0 : ret->first[0] = chunk_end_entry.m_ref;
3097 : 0 : m_known_end_of_cluster = true;
3098 : : } else {
3099 : 0 : Assume(m_cur_cluster);
3100 : 0 : ret->first.resize(chunk_data.m_chunk_count);
3101 : 0 : auto start_pos = chunk_end_entry.m_main_lin_index + 1 - chunk_data.m_chunk_count;
3102 [ # # ]: 0 : m_known_end_of_cluster = m_cur_cluster->GetClusterRefs(*m_graph, ret->first, start_pos);
3103 : : // If the chunk size was 1 and at end of cluster, then the special case above should
3104 : : // have been used.
3105 : 0 : Assume(!m_known_end_of_cluster || chunk_data.m_chunk_count > 1);
3106 : : }
3107 : 0 : ret->second = chunk_end_entry.m_main_chunk_feerate;
3108 : : }
3109 : 0 : return ret;
3110 : : }
3111 : :
3112 : 0 : BlockBuilderImpl::BlockBuilderImpl(TxGraphImpl& graph) noexcept : m_graph(&graph)
3113 : : {
3114 : : // Make sure all clusters in main are up to date, and acceptable.
3115 : 0 : m_graph->MakeAllAcceptable(0);
3116 : : // There cannot remain any inapplicable dependencies (only possible if main is oversized).
3117 [ # # ]: 0 : Assume(m_graph->m_main_clusterset.m_deps_to_add.empty());
3118 : : // Remember that this object is observing the graph's index, so that we can detect concurrent
3119 : : // modifications.
3120 : 0 : ++m_graph->m_main_chunkindex_observers;
3121 : : // Find the first chunk.
3122 [ # # ]: 0 : m_cur_iter = m_graph->m_main_chunkindex.begin();
3123 : 0 : m_cur_cluster = nullptr;
3124 [ # # ]: 0 : if (m_cur_iter != m_graph->m_main_chunkindex.end()) {
3125 : : // Find the cluster pointed to by m_cur_iter.
3126 : 0 : const auto& chunk_data = *m_cur_iter;
3127 : 0 : const auto& chunk_end_entry = m_graph->m_entries[chunk_data.m_graph_index];
3128 : 0 : m_cur_cluster = chunk_end_entry.m_locator[0].cluster;
3129 : : }
3130 : 0 : }
3131 : :
3132 : 0 : BlockBuilderImpl::~BlockBuilderImpl()
3133 : : {
3134 : 0 : Assume(m_graph->m_main_chunkindex_observers > 0);
3135 : : // Permit modifications to the main graph again after destroying the BlockBuilderImpl.
3136 : 0 : --m_graph->m_main_chunkindex_observers;
3137 : 0 : }
3138 : :
3139 : 0 : void BlockBuilderImpl::Include() noexcept
3140 : : {
3141 : : // The actual inclusion of the chunk is done by the calling code. All we have to do is switch
3142 : : // to the next chunk.
3143 : 0 : Next();
3144 : 0 : }
3145 : :
3146 : 0 : void BlockBuilderImpl::Skip() noexcept
3147 : : {
3148 : : // When skipping a chunk we need to not include anything more of the cluster, as that could make
3149 : : // the result topologically invalid. However, don't do this if the chunk is known to be the last
3150 : : // chunk of the cluster. This may significantly reduce the size of m_excluded_clusters,
3151 : : // especially when many singleton clusters are ignored.
3152 [ # # # # ]: 0 : if (m_cur_cluster != nullptr && !m_known_end_of_cluster) {
3153 : 0 : m_excluded_clusters.insert(m_cur_cluster->m_sequence);
3154 : : }
3155 : 0 : Next();
3156 : 0 : }
3157 : :
3158 : 0 : std::unique_ptr<TxGraph::BlockBuilder> TxGraphImpl::GetBlockBuilder() noexcept
3159 : : {
3160 [ # # ]: 0 : return std::make_unique<BlockBuilderImpl>(*this);
3161 : : }
3162 : :
3163 : 0 : std::pair<std::vector<TxGraph::Ref*>, FeePerWeight> TxGraphImpl::GetWorstMainChunk() noexcept
3164 : : {
3165 : 0 : std::pair<std::vector<Ref*>, FeePerWeight> ret;
3166 : : // Make sure all clusters in main are up to date, and acceptable.
3167 : 0 : MakeAllAcceptable(0);
3168 : 0 : Assume(m_main_clusterset.m_deps_to_add.empty());
3169 : : // If the graph is not empty, populate ret.
3170 [ # # ]: 0 : if (!m_main_chunkindex.empty()) {
3171 [ # # ]: 0 : const auto& chunk_data = *m_main_chunkindex.rbegin();
3172 [ # # ]: 0 : const auto& chunk_end_entry = m_entries[chunk_data.m_graph_index];
3173 : 0 : Cluster* cluster = chunk_end_entry.m_locator[0].cluster;
3174 [ # # ]: 0 : if (chunk_data.m_chunk_count == LinearizationIndex(-1) || chunk_data.m_chunk_count == 1) {
3175 : : // Special case for singletons.
3176 : 0 : ret.first.resize(1);
3177 : 0 : Assume(chunk_end_entry.m_ref != nullptr);
3178 : 0 : ret.first[0] = chunk_end_entry.m_ref;
3179 : : } else {
3180 : 0 : ret.first.resize(chunk_data.m_chunk_count);
3181 : 0 : auto start_pos = chunk_end_entry.m_main_lin_index + 1 - chunk_data.m_chunk_count;
3182 [ # # ]: 0 : cluster->GetClusterRefs(*this, ret.first, start_pos);
3183 : 0 : std::reverse(ret.first.begin(), ret.first.end());
3184 : : }
3185 : 0 : ret.second = chunk_end_entry.m_main_chunk_feerate;
3186 : : }
3187 : 0 : return ret;
3188 : : }
3189 : :
3190 : 4 : std::vector<TxGraph::Ref*> TxGraphImpl::Trim() noexcept
3191 : : {
3192 [ + - ]: 4 : int level = GetTopLevel();
3193 : 4 : Assume(m_main_chunkindex_observers == 0 || level != 0);
3194 : 4 : std::vector<TxGraph::Ref*> ret;
3195 : :
3196 : : // Compute the groups of to-be-merged Clusters (which also applies all removals, and splits).
3197 : 4 : auto& clusterset = GetClusterSet(level);
3198 [ + - ]: 4 : if (clusterset.m_oversized == false) return ret;
3199 : 4 : GroupClusters(level);
3200 [ + - ]: 4 : Assume(clusterset.m_group_data.has_value());
3201 : : // Nothing to do if not oversized.
3202 : 4 : Assume(clusterset.m_oversized.has_value());
3203 [ + - ]: 4 : if (clusterset.m_oversized == false) return ret;
3204 : :
3205 : : // In this function, would-be clusters (as precomputed in m_group_data by GroupClusters) are
3206 : : // trimmed by removing transactions in them such that the resulting clusters satisfy the size
3207 : : // and count limits.
3208 : : //
3209 : : // It works by defining for each would-be cluster a rudimentary linearization: at every point
3210 : : // the highest-chunk-feerate remaining transaction is picked among those with no unmet
3211 : : // dependencies. "Dependency" here means either a to-be-added dependency (m_deps_to_add), or
3212 : : // an implicit dependency added between any two consecutive transaction in their current
3213 : : // cluster linearization. So it can be seen as a "merge sort" of the chunks of the clusters,
3214 : : // but respecting the dependencies being added.
3215 : : //
3216 : : // This rudimentary linearization is computed lazily, by putting all eligible (no unmet
3217 : : // dependencies) transactions in a heap, and popping the highest-feerate one from it. Along the
3218 : : // way, the counts and sizes of the would-be clusters up to that point are tracked (by
3219 : : // partitioning the involved transactions using a union-find structure). Any transaction whose
3220 : : // addition would cause a violation is removed, along with all their descendants.
3221 : : //
3222 : : // A next invocation of GroupClusters (after applying the removals) will compute the new
3223 : : // resulting clusters, and none of them will violate the limits.
3224 : :
3225 : : /** All dependencies (both to be added ones, and implicit ones between consecutive transactions
3226 : : * in existing cluster linearizations), sorted by parent. */
3227 : 4 : std::vector<std::pair<GraphIndex, GraphIndex>> deps_by_parent;
3228 : : /** Same, but sorted by child. */
3229 : 4 : std::vector<std::pair<GraphIndex, GraphIndex>> deps_by_child;
3230 : : /** Information about all transactions involved in a Cluster group to be trimmed, sorted by
3231 : : * GraphIndex. It contains entries both for transactions that have already been included,
3232 : : * and ones that have not yet been. */
3233 : 4 : std::vector<TrimTxData> trim_data;
3234 : : /** Iterators into trim_data, treated as a max heap according to cmp_fn below. Each entry is
3235 : : * a transaction that has not yet been included yet, but all its ancestors have. */
3236 : 4 : std::vector<std::vector<TrimTxData>::iterator> trim_heap;
3237 : : /** The list of representatives of the partitions a given transaction depends on. */
3238 : 4 : std::vector<TrimTxData*> current_deps;
3239 : :
3240 : : /** Function to define the ordering of trim_heap. */
3241 : 1082781 : static constexpr auto cmp_fn = [](auto a, auto b) noexcept {
3242 : : // Sort by increasing chunk feerate, and then by decreasing size.
3243 : : // We do not need to sort by cluster or within clusters, because due to the implicit
3244 : : // dependency between consecutive linearization elements, no two transactions from the
3245 : : // same Cluster will ever simultaneously be in the heap.
3246 : 1082777 : return a->m_chunk_feerate < b->m_chunk_feerate;
3247 : : };
3248 : :
3249 : : /** Given a TrimTxData entry, find the representative of the partition it is in. */
3250 : 127308 : static constexpr auto find_fn = [](TrimTxData* arg) noexcept {
3251 [ + + - + ]: 189351 : while (arg != arg->m_uf_parent) {
3252 : : // Replace pointer to parent with pointer to grandparent (path splitting).
3253 : : // See https://en.wikipedia.org/wiki/Disjoint-set_data_structure#Finding_set_representatives.
3254 : 62047 : auto par = arg->m_uf_parent;
3255 : 62047 : arg->m_uf_parent = par->m_uf_parent;
3256 : 62047 : arg = par;
3257 : : }
3258 : 127304 : return arg;
3259 : : };
3260 : :
3261 : : /** Given two TrimTxData entries, union the partitions they are in, and return the
3262 : : * representative. */
3263 : 63100 : static constexpr auto union_fn = [](TrimTxData* arg1, TrimTxData* arg2) noexcept {
3264 : : // Replace arg1 and arg2 by their representatives.
3265 [ - + ]: 126192 : auto rep1 = find_fn(arg1);
3266 : 63096 : auto rep2 = find_fn(arg2);
3267 : : // Bail out if both representatives are the same, because that means arg1 and arg2 are in
3268 : : // the same partition already.
3269 [ + - ]: 63096 : if (rep1 == rep2) return rep1;
3270 : : // Pick the lower-count root to become a child of the higher-count one.
3271 : : // See https://en.wikipedia.org/wiki/Disjoint-set_data_structure#Union_by_size.
3272 [ + + ]: 63096 : if (rep1->m_uf_count < rep2->m_uf_count) std::swap(rep1, rep2);
3273 : 63096 : rep2->m_uf_parent = rep1;
3274 : : // Add the statistics of arg2 (which is no longer a representative) to those of arg1 (which
3275 : : // is now the representative for both).
3276 : 63096 : rep1->m_uf_size += rep2->m_uf_size;
3277 : 63096 : rep1->m_uf_count += rep2->m_uf_count;
3278 : 63096 : return rep1;
3279 : : };
3280 : :
3281 : : /** Get iterator to TrimTxData entry for a given index. */
3282 : 128420 : auto locate_fn = [&](GraphIndex index) noexcept {
3283 : 128416 : auto it = std::lower_bound(trim_data.begin(), trim_data.end(), index, [](TrimTxData& elem, GraphIndex idx) noexcept {
3284 [ + + ]: 2048000 : return elem.m_index < idx;
3285 : : });
3286 [ + - ]: 128416 : Assume(it != trim_data.end() && it->m_index == index);
3287 : 128416 : return it;
3288 : 4 : };
3289 : :
3290 : : // For each group of to-be-merged Clusters.
3291 [ + + ]: 13 : for (const auto& group_data : clusterset.m_group_data->m_groups) {
3292 [ + + ]: 9 : trim_data.clear();
3293 [ - + ]: 9 : trim_heap.clear();
3294 [ - + ]: 9 : deps_by_child.clear();
3295 [ - + ]: 9 : deps_by_parent.clear();
3296 : :
3297 : : // Gather trim data and implicit dependency data from all involved Clusters.
3298 [ - + ]: 9 : auto cluster_span = std::span{clusterset.m_group_data->m_group_clusters}
3299 : 9 : .subspan(group_data.m_cluster_offset, group_data.m_cluster_count);
3300 : 9 : uint64_t size{0};
3301 [ + + ]: 64226 : for (Cluster* cluster : cluster_span) {
3302 : 64217 : size += cluster->AppendTrimData(trim_data, deps_by_child);
3303 : : }
3304 : : // If this group of Clusters does not violate any limits, continue to the next group.
3305 [ - + + + : 9 : if (trim_data.size() <= m_max_cluster_count && size <= m_max_cluster_size) continue;
+ - ]
3306 : : // Sort the trim data by GraphIndex. In what follows, we will treat this sorted vector as
3307 : : // a map from GraphIndex to TrimTxData via locate_fn, and its ordering will not change
3308 : : // anymore.
3309 [ + + + + : 1315115 : std::sort(trim_data.begin(), trim_data.end(), [](auto& a, auto& b) noexcept { return a.m_index < b.m_index; });
+ + + + +
+ + + + +
+ + + + +
+ - - +
+ ]
3310 : :
3311 : : // Add the explicitly added dependencies to deps_by_child.
3312 : 9 : deps_by_child.insert(deps_by_child.end(),
3313 : 9 : clusterset.m_deps_to_add.begin() + group_data.m_deps_offset,
3314 : 9 : clusterset.m_deps_to_add.begin() + group_data.m_deps_offset + group_data.m_deps_count);
3315 : :
3316 : : // Sort deps_by_child by child transaction GraphIndex. The order will not be changed
3317 : : // anymore after this.
3318 [ + + + + : 1453049 : std::sort(deps_by_child.begin(), deps_by_child.end(), [](auto& a, auto& b) noexcept { return a.second < b.second; });
+ + + + +
+ + + + +
+ + + + +
+ - - +
+ ]
3319 : : // Fill m_parents_count and m_parents_offset in trim_data, as well as m_deps_left, and
3320 : : // initially populate trim_heap. Because of the sort above, all dependencies involving the
3321 : : // same child are grouped together, so a single linear scan suffices.
3322 : 9 : auto deps_it = deps_by_child.begin();
3323 [ + + ]: 64226 : for (auto trim_it = trim_data.begin(); trim_it != trim_data.end(); ++trim_it) {
3324 : 64217 : trim_it->m_parent_offset = deps_it - deps_by_child.begin();
3325 : 64217 : trim_it->m_deps_left = 0;
3326 [ + + + + ]: 128425 : while (deps_it != deps_by_child.end() && deps_it->second == trim_it->m_index) {
3327 : 64208 : ++trim_it->m_deps_left;
3328 : 64208 : ++deps_it;
3329 : : }
3330 : 64217 : trim_it->m_parent_count = trim_it->m_deps_left;
3331 : : // If this transaction has no unmet dependencies, and is not oversized, add it to the
3332 : : // heap (just append for now, the heapification happens below).
3333 [ + + + + ]: 64217 : if (trim_it->m_deps_left == 0 && trim_it->m_tx_size <= m_max_cluster_size) {
3334 : 1150 : trim_heap.push_back(trim_it);
3335 : : }
3336 : : }
3337 : 9 : Assume(deps_it == deps_by_child.end());
3338 : :
3339 : : // Construct deps_by_parent, sorted by parent transaction GraphIndex. The order will not be
3340 : : // changed anymore after this.
3341 : 9 : deps_by_parent = deps_by_child;
3342 [ + + + + : 1787289 : std::sort(deps_by_parent.begin(), deps_by_parent.end(), [](auto& a, auto& b) noexcept { return a.first < b.first; });
+ + + + +
+ + + + +
+ + + + +
+ - - +
+ ]
3343 : : // Fill m_children_offset and m_children_count in trim_data. Because of the sort above, all
3344 : : // dependencies involving the same parent are grouped together, so a single linear scan
3345 : : // suffices.
3346 : 9 : deps_it = deps_by_parent.begin();
3347 [ + + ]: 64226 : for (auto& trim_entry : trim_data) {
3348 : 64217 : trim_entry.m_children_count = 0;
3349 : 64217 : trim_entry.m_children_offset = deps_it - deps_by_parent.begin();
3350 [ + + + + ]: 128425 : while (deps_it != deps_by_parent.end() && deps_it->first == trim_entry.m_index) {
3351 : 64208 : ++trim_entry.m_children_count;
3352 : 64208 : ++deps_it;
3353 : : }
3354 : : }
3355 : 9 : Assume(deps_it == deps_by_parent.end());
3356 : :
3357 : : // Build a heap of all transactions with 0 unmet dependencies.
3358 : 9 : std::make_heap(trim_heap.begin(), trim_heap.end(), cmp_fn);
3359 : :
3360 : : // Iterate over to-be-included transactions, and convert them to included transactions, or
3361 : : // skip them if adding them would violate resource limits of the would-be cluster.
3362 : : //
3363 : : // It is possible that the heap empties without ever hitting either cluster limit, in case
3364 : : // the implied graph (to be added dependencies plus implicit dependency between each
3365 : : // original transaction and its predecessor in the linearization it came from) contains
3366 : : // cycles. Such cycles will be removed entirely, because each of the transactions in the
3367 : : // cycle permanently have unmet dependencies. However, this cannot occur in real scenarios
3368 : : // where Trim() is called to deal with reorganizations that would violate cluster limits,
3369 : : // as all added dependencies are in the same direction (from old mempool transactions to
3370 : : // new from-block transactions); cycles require dependencies in both directions to be
3371 : : // added.
3372 [ + + ]: 64229 : while (!trim_heap.empty()) {
3373 : : // Move the best remaining transaction to the end of trim_heap.
3374 : 64211 : std::pop_heap(trim_heap.begin(), trim_heap.end(), cmp_fn);
3375 : : // Pop it, and find its TrimTxData.
3376 [ + + ]: 64211 : auto& entry = *trim_heap.back();
3377 [ + + ]: 64211 : trim_heap.pop_back();
3378 : :
3379 : : // Initialize it as a singleton partition.
3380 : 64211 : entry.m_uf_parent = &entry;
3381 : 64211 : entry.m_uf_count = 1;
3382 : 64211 : entry.m_uf_size = entry.m_tx_size;
3383 : :
3384 : : // Find the distinct transaction partitions this entry depends on.
3385 [ + + ]: 64211 : current_deps.clear();
3386 [ - + + + ]: 128419 : for (auto& [par, chl] : std::span{deps_by_child}.subspan(entry.m_parent_offset, entry.m_parent_count)) {
3387 : 64208 : Assume(chl == entry.m_index);
3388 : 128416 : current_deps.push_back(find_fn(&*locate_fn(par)));
3389 : : }
3390 : 64211 : std::sort(current_deps.begin(), current_deps.end());
3391 : 64211 : current_deps.erase(std::unique(current_deps.begin(), current_deps.end()), current_deps.end());
3392 : :
3393 : : // Compute resource counts.
3394 : 64211 : uint32_t new_count = 1;
3395 : 64211 : uint64_t new_size = entry.m_tx_size;
3396 [ + + ]: 128419 : for (TrimTxData* ptr : current_deps) {
3397 : 64208 : new_count += ptr->m_uf_count;
3398 : 64208 : new_size += ptr->m_uf_size;
3399 : : }
3400 : : // Skip the entry if this would violate any limit.
3401 [ + + - + ]: 64211 : if (new_count > m_max_cluster_count || new_size > m_max_cluster_size) continue;
3402 : :
3403 : : // Union the partitions this transaction and all its dependencies are in together.
3404 : 64198 : auto rep = &entry;
3405 [ + + ]: 127294 : for (TrimTxData* ptr : current_deps) rep = union_fn(ptr, rep);
3406 : : // Mark the entry as included (so the loop below will not remove the transaction).
3407 : 64198 : entry.m_deps_left = uint32_t(-1);
3408 : : // Mark each to-be-added dependency involving this transaction as parent satisfied.
3409 [ - + + + ]: 128406 : for (auto& [par, chl] : std::span{deps_by_parent}.subspan(entry.m_children_offset, entry.m_children_count)) {
3410 : 64208 : Assume(par == entry.m_index);
3411 : 64208 : auto chl_it = locate_fn(chl);
3412 : : // Reduce the number of unmet dependencies of chl_it, and if that brings the number
3413 : : // to zero, add it to the heap of includable transactions.
3414 : 64208 : Assume(chl_it->m_deps_left > 0);
3415 [ + + ]: 64208 : if (--chl_it->m_deps_left == 0) {
3416 : 63061 : trim_heap.push_back(chl_it);
3417 : 63061 : std::push_heap(trim_heap.begin(), trim_heap.end(), cmp_fn);
3418 : : }
3419 : : }
3420 : : }
3421 : :
3422 : : // Remove all the transactions that were not processed above. Because nothing gets
3423 : : // processed until/unless all its dependencies are met, this automatically guarantees
3424 : : // that if a transaction is removed, all its descendants, or would-be descendants, are
3425 : : // removed as well.
3426 [ + + ]: 64226 : for (const auto& trim_entry : trim_data) {
3427 [ + + ]: 64217 : if (trim_entry.m_deps_left != uint32_t(-1)) {
3428 : 19 : ret.push_back(m_entries[trim_entry.m_index].m_ref);
3429 : 19 : clusterset.m_to_remove.push_back(trim_entry.m_index);
3430 : : }
3431 : : }
3432 : : }
3433 [ + - ]: 4 : clusterset.m_group_data.reset();
3434 : 4 : clusterset.m_oversized = false;
3435 : 4 : Assume(!ret.empty());
3436 : 4 : return ret;
3437 : 4 : }
3438 : :
3439 : 0 : size_t TxGraphImpl::GetMainMemoryUsage() noexcept
3440 : : {
3441 : : // Make sure splits/merges are applied, as memory usage may not be representative otherwise.
3442 : 0 : SplitAll(/*up_to_level=*/0);
3443 : 0 : ApplyDependencies(/*level=*/0);
3444 : : // Compute memory usage
3445 : 0 : size_t usage = /* From clusters */
3446 : 0 : m_main_clusterset.m_cluster_usage +
3447 : : /* From Entry objects. */
3448 : 0 : sizeof(Entry) * m_main_clusterset.m_txcount +
3449 : : /* From the chunk index. */
3450 : 0 : memusage::DynamicUsage(m_main_chunkindex);
3451 : 0 : return usage;
3452 : : }
3453 : :
3454 : : } // namespace
3455 : :
3456 : 194172 : TxGraph::Ref::~Ref()
3457 : : {
3458 [ + + ]: 194172 : if (m_graph) {
3459 : : // Inform the TxGraph about the Ref being destroyed.
3460 : 300 : m_graph->UnlinkRef(m_index);
3461 : 300 : m_graph = nullptr;
3462 : : }
3463 : 194172 : }
3464 : :
3465 : 0 : TxGraph::Ref& TxGraph::Ref::operator=(Ref&& other) noexcept
3466 : : {
3467 : : // Unlink the current graph, if any.
3468 [ # # ]: 0 : if (m_graph) m_graph->UnlinkRef(m_index);
3469 : : // Inform the other's graph about the move, if any.
3470 [ # # ]: 0 : if (other.m_graph) other.m_graph->UpdateRef(other.m_index, *this);
3471 : : // Actually update the contents.
3472 : 0 : m_graph = other.m_graph;
3473 : 0 : m_index = other.m_index;
3474 : 0 : other.m_graph = nullptr;
3475 : 0 : other.m_index = GraphIndex(-1);
3476 : 0 : return *this;
3477 : : }
3478 : :
3479 : 129861 : TxGraph::Ref::Ref(Ref&& other) noexcept
3480 : : {
3481 : : // Inform the TxGraph of other that its Ref is being moved.
3482 [ + - ]: 129861 : if (other.m_graph) other.m_graph->UpdateRef(other.m_index, *this);
3483 : : // Actually move the contents.
3484 : 129861 : std::swap(m_graph, other.m_graph);
3485 : 129861 : std::swap(m_index, other.m_index);
3486 : 129861 : }
3487 : :
3488 : 4 : std::unique_ptr<TxGraph> MakeTxGraph(unsigned max_cluster_count, uint64_t max_cluster_size, uint64_t acceptable_iters) noexcept
3489 : : {
3490 [ - + ]: 4 : return std::make_unique<TxGraphImpl>(max_cluster_count, max_cluster_size, acceptable_iters);
3491 : : }
|
