Dean Michael Berris | 980d93d | 2018-05-15 00:42:36 +0000 | [diff] [blame] | 1 | //===-- xray_function_call_trie.h ------------------------------*- C++ -*-===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This file is a part of XRay, a dynamic runtime instrumentation system. |
| 11 | // |
| 12 | // This file defines the interface for a function call trie. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | #ifndef XRAY_FUNCTION_CALL_TRIE_H |
| 16 | #define XRAY_FUNCTION_CALL_TRIE_H |
| 17 | |
| 18 | #include "xray_profiler_flags.h" |
| 19 | #include "xray_segmented_array.h" |
| 20 | #include <utility> |
Dean Michael Berris | 1eb8c20 | 2018-05-31 04:33:52 +0000 | [diff] [blame] | 21 | #include <memory> // For placement new. |
Dean Michael Berris | 980d93d | 2018-05-15 00:42:36 +0000 | [diff] [blame] | 22 | |
| 23 | namespace __xray { |
| 24 | |
| 25 | /// A FunctionCallTrie represents the stack traces of XRay instrumented |
| 26 | /// functions that we've encountered, where a node corresponds to a function and |
| 27 | /// the path from the root to the node its stack trace. Each node in the trie |
| 28 | /// will contain some useful values, including: |
| 29 | /// |
| 30 | /// * The cumulative amount of time spent in this particular node/stack. |
| 31 | /// * The number of times this stack has appeared. |
| 32 | /// * A histogram of latencies for that particular node. |
| 33 | /// |
| 34 | /// Each node in the trie will also contain a list of callees, represented using |
| 35 | /// a Array<NodeIdPair> -- each NodeIdPair instance will contain the function |
| 36 | /// ID of the callee, and a pointer to the node. |
| 37 | /// |
| 38 | /// If we visualise this data structure, we'll find the following potential |
| 39 | /// representation: |
| 40 | /// |
| 41 | /// [function id node] -> [callees] [cumulative time] |
| 42 | /// [call counter] [latency histogram] |
| 43 | /// |
| 44 | /// As an example, when we have a function in this pseudocode: |
| 45 | /// |
| 46 | /// func f(N) { |
| 47 | /// g() |
| 48 | /// h() |
| 49 | /// for i := 1..N { j() } |
| 50 | /// } |
| 51 | /// |
| 52 | /// We may end up with a trie of the following form: |
| 53 | /// |
| 54 | /// f -> [ g, h, j ] [...] [1] [...] |
| 55 | /// g -> [ ... ] [...] [1] [...] |
| 56 | /// h -> [ ... ] [...] [1] [...] |
| 57 | /// j -> [ ... ] [...] [N] [...] |
| 58 | /// |
| 59 | /// If for instance the function g() called j() like so: |
| 60 | /// |
| 61 | /// func g() { |
| 62 | /// for i := 1..10 { j() } |
| 63 | /// } |
| 64 | /// |
| 65 | /// We'll find the following updated trie: |
| 66 | /// |
| 67 | /// f -> [ g, h, j ] [...] [1] [...] |
| 68 | /// g -> [ j' ] [...] [1] [...] |
| 69 | /// h -> [ ... ] [...] [1] [...] |
| 70 | /// j -> [ ... ] [...] [N] [...] |
| 71 | /// j' -> [ ... ] [...] [10] [...] |
| 72 | /// |
| 73 | /// Note that we'll have a new node representing the path `f -> g -> j'` with |
| 74 | /// isolated data. This isolation gives us a means of representing the stack |
| 75 | /// traces as a path, as opposed to a key in a table. The alternative |
| 76 | /// implementation here would be to use a separate table for the path, and use |
| 77 | /// hashes of the path as an identifier to accumulate the information. We've |
| 78 | /// moved away from this approach as it takes a lot of time to compute the hash |
| 79 | /// every time we need to update a function's call information as we're handling |
| 80 | /// the entry and exit events. |
| 81 | /// |
| 82 | /// This approach allows us to maintain a shadow stack, which represents the |
| 83 | /// currently executing path, and on function exits quickly compute the amount |
| 84 | /// of time elapsed from the entry, then update the counters for the node |
| 85 | /// already represented in the trie. This necessitates an efficient |
| 86 | /// representation of the various data structures (the list of callees must be |
| 87 | /// cache-aware and efficient to look up, and the histogram must be compact and |
| 88 | /// quick to update) to enable us to keep the overheads of this implementation |
| 89 | /// to the minimum. |
| 90 | class FunctionCallTrie { |
| 91 | public: |
| 92 | struct Node; |
| 93 | |
| 94 | // We use a NodeIdPair type instead of a std::pair<...> to not rely on the |
| 95 | // standard library types in this header. |
| 96 | struct NodeIdPair { |
| 97 | Node *NodePtr; |
| 98 | int32_t FId; |
| 99 | |
| 100 | // Constructor for inplace-construction. |
| 101 | NodeIdPair(Node *N, int32_t F) : NodePtr(N), FId(F) {} |
| 102 | }; |
| 103 | |
| 104 | using NodeIdPairArray = Array<NodeIdPair>; |
| 105 | using NodeIdPairAllocatorType = NodeIdPairArray::AllocatorType; |
| 106 | |
| 107 | // A Node in the FunctionCallTrie gives us a list of callees, the cumulative |
| 108 | // number of times this node actually appeared, the cumulative amount of time |
| 109 | // for this particular node including its children call times, and just the |
| 110 | // local time spent on this node. Each Node will have the ID of the XRay |
| 111 | // instrumented function that it is associated to. |
| 112 | struct Node { |
| 113 | Node *Parent; |
| 114 | NodeIdPairArray Callees; |
| 115 | int64_t CallCount; |
| 116 | int64_t CumulativeLocalTime; // Typically in TSC deltas, not wall-time. |
| 117 | int32_t FId; |
| 118 | |
| 119 | // We add a constructor here to allow us to inplace-construct through |
| 120 | // Array<...>'s AppendEmplace. |
| 121 | Node(Node *P, NodeIdPairAllocatorType &A, int64_t CC, int64_t CLT, |
| 122 | int32_t F) |
| 123 | : Parent(P), Callees(A), CallCount(CC), CumulativeLocalTime(CLT), |
| 124 | FId(F) {} |
| 125 | |
| 126 | // TODO: Include the compact histogram. |
| 127 | }; |
| 128 | |
| 129 | private: |
| 130 | struct ShadowStackEntry { |
| 131 | int32_t FId; // We're copying the function ID into the stack to avoid having |
| 132 | // to reach into the node just to get the function ID. |
| 133 | uint64_t EntryTSC; |
| 134 | Node *NodePtr; |
| 135 | |
| 136 | // We add a constructor here to allow us to inplace-construct through |
| 137 | // Array<...>'s AppendEmplace. |
| 138 | ShadowStackEntry(int32_t F, uint64_t T, Node *N) |
| 139 | : FId(F), EntryTSC(T), NodePtr(N) {} |
| 140 | }; |
| 141 | |
| 142 | using NodeArray = Array<Node>; |
| 143 | using RootArray = Array<Node *>; |
| 144 | using ShadowStackArray = Array<ShadowStackEntry>; |
| 145 | |
| 146 | public: |
| 147 | // We collate the allocators we need into a single struct, as a convenience to |
| 148 | // allow us to initialize these as a group. |
| 149 | struct Allocators { |
| 150 | using NodeAllocatorType = NodeArray::AllocatorType; |
| 151 | using RootAllocatorType = RootArray::AllocatorType; |
| 152 | using ShadowStackAllocatorType = ShadowStackArray::AllocatorType; |
Dean Michael Berris | 980d93d | 2018-05-15 00:42:36 +0000 | [diff] [blame] | 153 | |
| 154 | NodeAllocatorType *NodeAllocator = nullptr; |
| 155 | RootAllocatorType *RootAllocator = nullptr; |
| 156 | ShadowStackAllocatorType *ShadowStackAllocator = nullptr; |
| 157 | NodeIdPairAllocatorType *NodeIdPairAllocator = nullptr; |
| 158 | |
| 159 | Allocators() {} |
| 160 | Allocators(const Allocators &) = delete; |
| 161 | Allocators &operator=(const Allocators &) = delete; |
| 162 | |
| 163 | Allocators(Allocators &&O) |
| 164 | : NodeAllocator(O.NodeAllocator), RootAllocator(O.RootAllocator), |
| 165 | ShadowStackAllocator(O.ShadowStackAllocator), |
| 166 | NodeIdPairAllocator(O.NodeIdPairAllocator) { |
| 167 | O.NodeAllocator = nullptr; |
| 168 | O.RootAllocator = nullptr; |
| 169 | O.ShadowStackAllocator = nullptr; |
| 170 | O.NodeIdPairAllocator = nullptr; |
| 171 | } |
| 172 | |
| 173 | Allocators &operator=(Allocators &&O) { |
| 174 | { |
| 175 | auto Tmp = O.NodeAllocator; |
| 176 | O.NodeAllocator = this->NodeAllocator; |
| 177 | this->NodeAllocator = Tmp; |
| 178 | } |
| 179 | { |
| 180 | auto Tmp = O.RootAllocator; |
| 181 | O.RootAllocator = this->RootAllocator; |
| 182 | this->RootAllocator = Tmp; |
| 183 | } |
| 184 | { |
| 185 | auto Tmp = O.ShadowStackAllocator; |
| 186 | O.ShadowStackAllocator = this->ShadowStackAllocator; |
| 187 | this->ShadowStackAllocator = Tmp; |
| 188 | } |
| 189 | { |
| 190 | auto Tmp = O.NodeIdPairAllocator; |
| 191 | O.NodeIdPairAllocator = this->NodeIdPairAllocator; |
| 192 | this->NodeIdPairAllocator = Tmp; |
| 193 | } |
| 194 | return *this; |
| 195 | } |
| 196 | |
| 197 | ~Allocators() { |
| 198 | // Note that we cannot use delete on these pointers, as they need to be |
| 199 | // returned to the sanitizer_common library's internal memory tracking |
| 200 | // system. |
| 201 | if (NodeAllocator != nullptr) { |
| 202 | NodeAllocator->~NodeAllocatorType(); |
| 203 | InternalFree(NodeAllocator); |
| 204 | } |
| 205 | if (RootAllocator != nullptr) { |
| 206 | RootAllocator->~RootAllocatorType(); |
| 207 | InternalFree(RootAllocator); |
| 208 | } |
| 209 | if (ShadowStackAllocator != nullptr) { |
| 210 | ShadowStackAllocator->~ShadowStackAllocatorType(); |
| 211 | InternalFree(ShadowStackAllocator); |
| 212 | } |
| 213 | if (NodeIdPairAllocator != nullptr) { |
| 214 | NodeIdPairAllocator->~NodeIdPairAllocatorType(); |
| 215 | InternalFree(NodeIdPairAllocator); |
| 216 | } |
| 217 | } |
| 218 | }; |
| 219 | |
| 220 | // TODO: Support configuration of options through the arguments. |
| 221 | static Allocators InitAllocators() { |
| 222 | Allocators A; |
| 223 | auto NodeAllocator = reinterpret_cast<Allocators::NodeAllocatorType *>( |
| 224 | InternalAlloc(sizeof(Allocators::NodeAllocatorType))); |
| 225 | new (NodeAllocator) Allocators::NodeAllocatorType( |
| 226 | profilerFlags()->per_thread_allocator_max, 0); |
| 227 | A.NodeAllocator = NodeAllocator; |
| 228 | |
| 229 | auto RootAllocator = reinterpret_cast<Allocators::RootAllocatorType *>( |
| 230 | InternalAlloc(sizeof(Allocators::RootAllocatorType))); |
| 231 | new (RootAllocator) Allocators::RootAllocatorType( |
| 232 | profilerFlags()->per_thread_allocator_max, 0); |
| 233 | A.RootAllocator = RootAllocator; |
| 234 | |
| 235 | auto ShadowStackAllocator = |
| 236 | reinterpret_cast<Allocators::ShadowStackAllocatorType *>( |
| 237 | InternalAlloc(sizeof(Allocators::ShadowStackAllocatorType))); |
| 238 | new (ShadowStackAllocator) Allocators::ShadowStackAllocatorType( |
| 239 | profilerFlags()->per_thread_allocator_max, 0); |
| 240 | A.ShadowStackAllocator = ShadowStackAllocator; |
| 241 | |
Dean Michael Berris | d1fe506 | 2018-05-31 05:25:47 +0000 | [diff] [blame^] | 242 | auto NodeIdPairAllocator = reinterpret_cast<NodeIdPairAllocatorType *>( |
| 243 | InternalAlloc(sizeof(NodeIdPairAllocatorType))); |
| 244 | new (NodeIdPairAllocator) |
| 245 | NodeIdPairAllocatorType(profilerFlags()->per_thread_allocator_max, 0); |
Dean Michael Berris | 980d93d | 2018-05-15 00:42:36 +0000 | [diff] [blame] | 246 | A.NodeIdPairAllocator = NodeIdPairAllocator; |
| 247 | return A; |
| 248 | } |
| 249 | |
| 250 | private: |
| 251 | NodeArray Nodes; |
| 252 | RootArray Roots; |
| 253 | ShadowStackArray ShadowStack; |
| 254 | NodeIdPairAllocatorType *NodeIdPairAllocator = nullptr; |
| 255 | |
| 256 | const Allocators &GetGlobalAllocators() { |
| 257 | static const Allocators A = [] { return InitAllocators(); }(); |
| 258 | return A; |
| 259 | } |
| 260 | |
| 261 | public: |
| 262 | explicit FunctionCallTrie(const Allocators &A) |
| 263 | : Nodes(*A.NodeAllocator), Roots(*A.RootAllocator), |
| 264 | ShadowStack(*A.ShadowStackAllocator), |
| 265 | NodeIdPairAllocator(A.NodeIdPairAllocator) {} |
| 266 | |
| 267 | FunctionCallTrie() : FunctionCallTrie(GetGlobalAllocators()) {} |
| 268 | |
| 269 | void enterFunction(int32_t FId, uint64_t TSC) { |
| 270 | // This function primarily deals with ensuring that the ShadowStack is |
| 271 | // consistent and ready for when an exit event is encountered. |
| 272 | if (UNLIKELY(ShadowStack.empty())) { |
| 273 | auto NewRoot = |
| 274 | Nodes.AppendEmplace(nullptr, *NodeIdPairAllocator, 0, 0, FId); |
| 275 | if (UNLIKELY(NewRoot == nullptr)) |
| 276 | return; |
| 277 | Roots.Append(NewRoot); |
| 278 | ShadowStack.AppendEmplace(FId, TSC, NewRoot); |
| 279 | return; |
| 280 | } |
| 281 | |
| 282 | auto &Top = ShadowStack.back(); |
| 283 | auto TopNode = Top.NodePtr; |
| 284 | |
| 285 | // If we've seen this callee before, then we just access that node and place |
| 286 | // that on the top of the stack. |
| 287 | auto Callee = TopNode->Callees.find_element( |
| 288 | [FId](const NodeIdPair &NR) { return NR.FId == FId; }); |
| 289 | if (Callee != nullptr) { |
| 290 | CHECK_NE(Callee->NodePtr, nullptr); |
| 291 | ShadowStack.AppendEmplace(FId, TSC, Callee->NodePtr); |
| 292 | return; |
| 293 | } |
| 294 | |
| 295 | // This means we've never seen this stack before, create a new node here. |
| 296 | auto NewNode = |
| 297 | Nodes.AppendEmplace(TopNode, *NodeIdPairAllocator, 0, 0, FId); |
| 298 | if (UNLIKELY(NewNode == nullptr)) |
| 299 | return; |
| 300 | TopNode->Callees.AppendEmplace(NewNode, FId); |
| 301 | ShadowStack.AppendEmplace(FId, TSC, NewNode); |
| 302 | return; |
| 303 | } |
| 304 | |
| 305 | void exitFunction(int32_t FId, uint64_t TSC) { |
| 306 | // When we exit a function, we look up the ShadowStack to see whether we've |
| 307 | // entered this function before. We do as little processing here as we can, |
| 308 | // since most of the hard work would have already been done at function |
| 309 | // entry. |
| 310 | if (UNLIKELY(ShadowStack.empty())) |
| 311 | return; |
| 312 | |
| 313 | uint64_t CumulativeTreeTime = 0; |
| 314 | while (!ShadowStack.empty()) { |
| 315 | auto &Top = ShadowStack.back(); |
| 316 | auto TopNode = Top.NodePtr; |
| 317 | auto TopFId = TopNode->FId; |
| 318 | auto LocalTime = TSC - Top.EntryTSC; |
| 319 | TopNode->CallCount++; |
| 320 | TopNode->CumulativeLocalTime += LocalTime - CumulativeTreeTime; |
| 321 | CumulativeTreeTime += LocalTime; |
| 322 | ShadowStack.trim(1); |
| 323 | |
| 324 | // TODO: Update the histogram for the node. |
| 325 | if (TopFId == FId) |
| 326 | break; |
| 327 | } |
| 328 | } |
| 329 | |
| 330 | const RootArray &getRoots() const { return Roots; } |
| 331 | |
| 332 | // The deepCopyInto operation will update the provided FunctionCallTrie by |
| 333 | // re-creating the contents of this particular FunctionCallTrie in the other |
| 334 | // FunctionCallTrie. It will do this using a Depth First Traversal from the |
| 335 | // roots, and while doing so recreating the traversal in the provided |
| 336 | // FunctionCallTrie. |
| 337 | // |
| 338 | // This operation will *not* destroy the state in `O`, and thus may cause some |
| 339 | // duplicate entries in `O` if it is not empty. |
| 340 | // |
| 341 | // This function is *not* thread-safe, and may require external |
| 342 | // synchronisation of both "this" and |O|. |
| 343 | // |
| 344 | // This function must *not* be called with a non-empty FunctionCallTrie |O|. |
| 345 | void deepCopyInto(FunctionCallTrie &O) const { |
| 346 | DCHECK(O.getRoots().empty()); |
| 347 | for (const auto Root : getRoots()) { |
| 348 | // Add a node in O for this root. |
| 349 | auto NewRoot = O.Nodes.AppendEmplace( |
| 350 | nullptr, *O.NodeIdPairAllocator, Root->CallCount, |
| 351 | Root->CumulativeLocalTime, Root->FId); |
| 352 | O.Roots.Append(NewRoot); |
| 353 | |
| 354 | // We then push the root into a stack, to use as the parent marker for new |
| 355 | // nodes we push in as we're traversing depth-first down the call tree. |
| 356 | struct NodeAndParent { |
| 357 | FunctionCallTrie::Node *Node; |
| 358 | FunctionCallTrie::Node *NewNode; |
| 359 | }; |
| 360 | using Stack = Array<NodeAndParent>; |
| 361 | |
| 362 | typename Stack::AllocatorType StackAllocator( |
| 363 | profilerFlags()->stack_allocator_max, 0); |
| 364 | Stack DFSStack(StackAllocator); |
| 365 | |
| 366 | // TODO: Figure out what to do if we fail to allocate any more stack |
| 367 | // space. Maybe warn or report once? |
| 368 | DFSStack.Append(NodeAndParent{Root, NewRoot}); |
| 369 | while (!DFSStack.empty()) { |
| 370 | NodeAndParent NP = DFSStack.back(); |
| 371 | DCHECK_NE(NP.Node, nullptr); |
| 372 | DCHECK_NE(NP.NewNode, nullptr); |
| 373 | DFSStack.trim(1); |
| 374 | for (const auto Callee : NP.Node->Callees) { |
| 375 | auto NewNode = O.Nodes.AppendEmplace( |
| 376 | NP.NewNode, *O.NodeIdPairAllocator, Callee.NodePtr->CallCount, |
| 377 | Callee.NodePtr->CumulativeLocalTime, Callee.FId); |
| 378 | DCHECK_NE(NewNode, nullptr); |
| 379 | NP.NewNode->Callees.AppendEmplace(NewNode, Callee.FId); |
| 380 | DFSStack.Append(NodeAndParent{Callee.NodePtr, NewNode}); |
| 381 | } |
| 382 | } |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | // The mergeInto operation will update the provided FunctionCallTrie by |
| 387 | // traversing the current trie's roots and updating (i.e. merging) the data in |
| 388 | // the nodes with the data in the target's nodes. If the node doesn't exist in |
| 389 | // the provided trie, we add a new one in the right position, and inherit the |
| 390 | // data from the original (current) trie, along with all its callees. |
| 391 | // |
| 392 | // This function is *not* thread-safe, and may require external |
| 393 | // synchronisation of both "this" and |O|. |
| 394 | void mergeInto(FunctionCallTrie &O) const { |
| 395 | struct NodeAndTarget { |
| 396 | FunctionCallTrie::Node *OrigNode; |
| 397 | FunctionCallTrie::Node *TargetNode; |
| 398 | }; |
| 399 | using Stack = Array<NodeAndTarget>; |
| 400 | typename Stack::AllocatorType StackAllocator( |
| 401 | profilerFlags()->stack_allocator_max, 0); |
| 402 | Stack DFSStack(StackAllocator); |
| 403 | |
| 404 | for (const auto Root : getRoots()) { |
| 405 | Node *TargetRoot = nullptr; |
| 406 | auto R = O.Roots.find_element( |
| 407 | [&](const Node *Node) { return Node->FId == Root->FId; }); |
| 408 | if (R == nullptr) { |
| 409 | TargetRoot = O.Nodes.AppendEmplace(nullptr, *O.NodeIdPairAllocator, 0, |
| 410 | 0, Root->FId); |
| 411 | O.Roots.Append(TargetRoot); |
| 412 | } else { |
| 413 | TargetRoot = *R; |
| 414 | } |
| 415 | |
| 416 | DFSStack.Append(NodeAndTarget{Root, TargetRoot}); |
| 417 | while (!DFSStack.empty()) { |
| 418 | NodeAndTarget NT = DFSStack.back(); |
| 419 | DCHECK_NE(NT.OrigNode, nullptr); |
| 420 | DCHECK_NE(NT.TargetNode, nullptr); |
| 421 | DFSStack.trim(1); |
| 422 | // TODO: Update the histogram as well when we have it ready. |
| 423 | NT.TargetNode->CallCount += NT.OrigNode->CallCount; |
| 424 | NT.TargetNode->CumulativeLocalTime += NT.OrigNode->CumulativeLocalTime; |
| 425 | for (const auto Callee : NT.OrigNode->Callees) { |
| 426 | auto TargetCallee = NT.TargetNode->Callees.find_element( |
| 427 | [&](const FunctionCallTrie::NodeIdPair &C) { |
| 428 | return C.FId == Callee.FId; |
| 429 | }); |
| 430 | if (TargetCallee == nullptr) { |
| 431 | auto NewTargetNode = O.Nodes.AppendEmplace( |
| 432 | NT.TargetNode, *O.NodeIdPairAllocator, 0, 0, Callee.FId); |
| 433 | TargetCallee = |
| 434 | NT.TargetNode->Callees.AppendEmplace(NewTargetNode, Callee.FId); |
| 435 | } |
| 436 | DFSStack.Append(NodeAndTarget{Callee.NodePtr, TargetCallee->NodePtr}); |
| 437 | } |
| 438 | } |
| 439 | } |
| 440 | } |
| 441 | }; |
| 442 | |
| 443 | } // namespace __xray |
| 444 | |
| 445 | #endif // XRAY_FUNCTION_CALL_TRIE_H |