| // Copyright 2017 The Chromium Authors. All rights reserved. |
| // Use of this source code is governed by a BSD-style license that can be |
| // found in the LICENSE file. |
| |
| #ifndef SERVICES_RESOURCE_COORDINATOR_MEMORY_INSTRUMENTATION_GRAPH_PROCESSOR_H_ |
| #define SERVICES_RESOURCE_COORDINATOR_MEMORY_INSTRUMENTATION_GRAPH_PROCESSOR_H_ |
| |
| #include <memory> |
| |
| #include "base/process/process_handle.h" |
| #include "base/trace_event/process_memory_dump.h" |
| #include "services/resource_coordinator/memory_instrumentation/graph.h" |
| |
| namespace memory_instrumentation { |
| |
| class GraphProcessor { |
| public: |
| // This map does not own the pointers inside. |
| using MemoryDumpMap = |
| std::map<base::ProcessId, const base::trace_event::ProcessMemoryDump*>; |
| |
| static std::unique_ptr<GlobalDumpGraph> CreateMemoryGraph( |
| const MemoryDumpMap& process_dumps); |
| |
| static void RemoveWeakNodesFromGraph(GlobalDumpGraph* global_graph); |
| |
| static void AddOverheadsAndPropogateEntries(GlobalDumpGraph* global_graph); |
| |
| static void CalculateSizesForGraph(GlobalDumpGraph* global_graph); |
| |
| static std::map<base::ProcessId, uint64_t> ComputeSharedFootprintFromGraph( |
| const GlobalDumpGraph& global_graph); |
| |
| private: |
| friend class GraphProcessorTest; |
| |
| static void CollectAllocatorDumps( |
| const base::trace_event::ProcessMemoryDump& source, |
| GlobalDumpGraph* global_graph, |
| GlobalDumpGraph::Process* process_graph); |
| |
| static void AddEdges(const base::trace_event::ProcessMemoryDump& source, |
| GlobalDumpGraph* global_graph); |
| |
| static void MarkImplicitWeakParentsRecursively(GlobalDumpGraph::Node* node); |
| |
| static void MarkWeakOwnersAndChildrenRecursively( |
| GlobalDumpGraph::Node* node, |
| std::set<const GlobalDumpGraph::Node*>* nodes); |
| |
| static void RemoveWeakNodesRecursively(GlobalDumpGraph::Node* parent); |
| |
| static void AssignTracingOverhead(base::StringPiece allocator, |
| GlobalDumpGraph* global_graph, |
| GlobalDumpGraph::Process* process); |
| |
| static GlobalDumpGraph::Node::Entry AggregateNumericWithNameForNode( |
| GlobalDumpGraph::Node* node, |
| base::StringPiece name); |
| |
| static void AggregateNumericsRecursively(GlobalDumpGraph::Node* node); |
| |
| static void PropagateNumericsAndDiagnosticsRecursively( |
| GlobalDumpGraph::Node* node); |
| |
| static base::Optional<uint64_t> AggregateSizeForDescendantNode( |
| GlobalDumpGraph::Node* root, |
| GlobalDumpGraph::Node* descendant); |
| |
| static void CalculateSizeForNode(GlobalDumpGraph::Node* node); |
| |
| /** |
| * Calculate not-owned and not-owning sub-sizes of a memory allocator dump |
| * from its children's (sub-)sizes. |
| * |
| * Not-owned sub-size refers to the aggregated memory of all children which |
| * is not owned by other MADs. Conversely, not-owning sub-size is the |
| * aggregated memory of all children which do not own another MAD. The |
| * diagram below illustrates these two concepts: |
| * |
| * ROOT 1 ROOT 2 |
| * size: 4 size: 5 |
| * not-owned sub-size: 4 not-owned sub-size: 1 (!) |
| * not-owning sub-size: 0 (!) not-owning sub-size: 5 |
| * |
| * ^ ^ |
| * | | |
| * |
| * PARENT 1 ===== owns =====> PARENT 2 |
| * size: 4 size: 5 |
| * not-owned sub-size: 4 not-owned sub-size: 5 |
| * not-owning sub-size: 4 not-owning sub-size: 5 |
| * |
| * ^ ^ |
| * | | |
| * |
| * CHILD 1 CHILD 2 |
| * size [given]: 4 size [given]: 5 |
| * not-owned sub-size: 4 not-owned sub-size: 5 |
| * not-owning sub-size: 4 not-owning sub-size: 5 |
| * |
| * This method assumes that (1) the size of the dump, its children, and its |
| * owners [see calculateSizes()] and (2) the not-owned and not-owning |
| * sub-sizes of both the children and owners of the dump have already been |
| * calculated [depth-first post-order traversal]. |
| */ |
| static void CalculateDumpSubSizes(GlobalDumpGraph::Node* node); |
| |
| /** |
| * Calculate owned and owning coefficients of a memory allocator dump and |
| * its owners. |
| * |
| * The owning coefficient refers to the proportion of a dump's not-owning |
| * sub-size which is attributed to the dump (only relevant to owning MADs). |
| * Conversely, the owned coefficient is the proportion of a dump's |
| * not-owned sub-size, which is attributed to it (only relevant to owned |
| * MADs). |
| * |
| * The not-owned size of the owned dump is split among its owners in the |
| * order of the ownership importance as demonstrated by the following |
| * example: |
| * |
| * memory allocator dumps |
| * OWNED OWNER1 OWNER2 OWNER3 OWNER4 |
| * not-owned sub-size [given] 10 - - - - |
| * not-owning sub-size [given] - 6 7 5 8 |
| * importance [given] - 2 2 1 0 |
| * attributed not-owned sub-size 2 - - - - |
| * attributed not-owning sub-size - 3 4 0 1 |
| * owned coefficient 2/10 - - - - |
| * owning coefficient - 3/6 4/7 0/5 1/8 |
| * |
| * Explanation: Firstly, 6 bytes are split equally among OWNER1 and OWNER2 |
| * (highest importance). OWNER2 owns one more byte, so its attributed |
| * not-owning sub-size is 6/2 + 1 = 4 bytes. OWNER3 is attributed no size |
| * because it is smaller than the owners with higher priority. However, |
| * OWNER4 is larger, so it's attributed the difference 8 - 7 = 1 byte. |
| * Finally, 2 bytes remain unattributed and are hence kept in the OWNED |
| * dump as attributed not-owned sub-size. The coefficients are then |
| * directly calculated as fractions of the sub-sizes and corresponding |
| * attributed sub-sizes. |
| * |
| * Note that we always assume that all ownerships of a dump overlap (e.g. |
| * OWNER3 is subsumed by both OWNER1 and OWNER2). Hence, the table could |
| * be alternatively represented as follows: |
| * |
| * owned memory range |
| * 0 1 2 3 4 5 6 7 8 9 10 |
| * Priority 2 | OWNER1 + OWNER2 (split) | OWNER2 | |
| * Priority 1 | (already attributed) | |
| * Priority 0 | - - - (already attributed) - - - | OWNER4 | |
| * Remainder | - - - - - (already attributed) - - - - - - | OWNED | |
| * |
| * This method assumes that (1) the size of the dump [see calculateSizes()] |
| * and (2) the not-owned size of the dump and not-owning sub-sizes of its |
| * owners [see the first step of calculateEffectiveSizes()] have already |
| * been calculated. Note that the method doesn't make any assumptions about |
| * the order in which dumps are visited. |
| */ |
| static void CalculateDumpOwnershipCoefficient(GlobalDumpGraph::Node* node); |
| |
| /** |
| * Calculate cumulative owned and owning coefficients of a memory allocator |
| * dump from its (non-cumulative) owned and owning coefficients and the |
| * cumulative coefficients of its parent and/or owned dump. |
| * |
| * The cumulative coefficients represent the total effect of all |
| * (non-strict) ancestor ownerships on a memory allocator dump. The |
| * cumulative owned coefficient of a MAD can be calculated simply as: |
| * |
| * cumulativeOwnedC(M) = ownedC(M) * cumulativeOwnedC(parent(M)) |
| * |
| * This reflects the assumption that if a parent of a child MAD is |
| * (partially) owned, then the parent's owner also indirectly owns (a part |
| * of) the child MAD. |
| * |
| * The cumulative owning coefficient of a MAD depends on whether the MAD |
| * owns another dump: |
| * |
| * [if M doesn't own another MAD] |
| * / cumulativeOwningC(parent(M)) |
| * cumulativeOwningC(M) = |
| * \ [if M owns another MAD] |
| * owningC(M) * cumulativeOwningC(owned(M)) |
| * |
| * The reasoning behind the first case is similar to the one for cumulative |
| * owned coefficient above. The only difference is that we don't need to |
| * include the dump's (non-cumulative) owning coefficient because it is |
| * implicitly 1. |
| * |
| * The formula for the second case is derived as follows: Since the MAD |
| * owns another dump, its memory is not included in its parent's not-owning |
| * sub-size and hence shouldn't be affected by the parent's corresponding |
| * cumulative coefficient. Instead, the MAD indirectly owns everything |
| * owned by its owned dump (and so it should be affected by the |
| * corresponding coefficient). |
| * |
| * Note that undefined coefficients (and coefficients of non-existent |
| * dumps) are implicitly assumed to be 1. |
| * |
| * This method assumes that (1) the size of the dump [see calculateSizes()], |
| * (2) the (non-cumulative) owned and owning coefficients of the dump [see |
| * the second step of calculateEffectiveSizes()], and (3) the cumulative |
| * coefficients of the dump's parent and owned MADs (if present) |
| * [depth-first pre-order traversal] have already been calculated. |
| */ |
| static void CalculateDumpCumulativeOwnershipCoefficient( |
| GlobalDumpGraph::Node* node); |
| |
| /** |
| * Calculate the effective size of a memory allocator dump. |
| * |
| * In order to simplify the (already complex) calculation, we use the fact |
| * that effective size is cumulative (unlike regular size), i.e. the |
| * effective size of a non-leaf node is equal to the sum of effective sizes |
| * of its children. The effective size of a leaf MAD is calculated as: |
| * |
| * effectiveSize(M) = size(M) * cumulativeOwningC(M) * cumulativeOwnedC(M) |
| * |
| * This method assumes that (1) the size of the dump and its children [see |
| * calculateSizes()] and (2) the cumulative owning and owned coefficients |
| * of the dump (if it's a leaf node) [see the third step of |
| * calculateEffectiveSizes()] or the effective sizes of its children (if |
| * it's a non-leaf node) [depth-first post-order traversal] have already |
| * been calculated. |
| */ |
| static void CalculateDumpEffectiveSize(GlobalDumpGraph::Node* node); |
| }; |
| |
| } // namespace memory_instrumentation |
| #endif |