| 1 | /* |
| 2 | * Copyright 2011-present Facebook, Inc. |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | // @author: Xin Liu <xliux@fb.com> |
| 18 | // |
| 19 | // A concurrent skip list (CSL) implementation. |
| 20 | // Ref: http://www.cs.tau.ac.il/~shanir/nir-pubs-web/Papers/OPODIS2006-BA.pdf |
| 21 | |
| 22 | /* |
| 23 | |
| 24 | This implements a sorted associative container that supports only |
| 25 | unique keys. (Similar to std::set.) |
| 26 | |
| 27 | Features: |
| 28 | |
| 29 | 1. Small memory overhead: ~40% less memory overhead compared with |
| 30 | std::set (1.6 words per node versus 3). It has an minimum of 4 |
| 31 | words (7 words if there nodes got deleted) per-list overhead |
| 32 | though. |
| 33 | |
| 34 | 2. Read accesses (count, find iterator, skipper) are lock-free and |
| 35 | mostly wait-free (the only wait a reader may need to do is when |
| 36 | the node it is visiting is in a pending stage, i.e. deleting, |
| 37 | adding and not fully linked). Write accesses (remove, add) need |
| 38 | to acquire locks, but locks are local to the predecessor nodes |
| 39 | and/or successor nodes. |
| 40 | |
| 41 | 3. Good high contention performance, comparable single-thread |
| 42 | performance. In the multithreaded case (12 workers), CSL tested |
| 43 | 10x faster than a RWSpinLocked std::set for an averaged sized |
| 44 | list (1K - 1M nodes). |
| 45 | |
| 46 | Comparable read performance to std::set when single threaded, |
| 47 | especially when the list size is large, and scales better to |
| 48 | larger lists: when the size is small, CSL can be 20-50% slower on |
| 49 | find()/contains(). As the size gets large (> 1M elements), |
| 50 | find()/contains() can be 30% faster. |
| 51 | |
| 52 | Iterating through a skiplist is similar to iterating through a |
| 53 | linked list, thus is much (2-6x) faster than on a std::set |
| 54 | (tree-based). This is especially true for short lists due to |
| 55 | better cache locality. Based on that, it's also faster to |
| 56 | intersect two skiplists. |
| 57 | |
| 58 | 4. Lazy removal with GC support. The removed nodes get deleted when |
| 59 | the last Accessor to the skiplist is destroyed. |
| 60 | |
| 61 | Caveats: |
| 62 | |
| 63 | 1. Write operations are usually 30% slower than std::set in a single |
| 64 | threaded environment. |
| 65 | |
| 66 | 2. Need to have a head node for each list, which has a 4 word |
| 67 | overhead. |
| 68 | |
| 69 | 3. When the list is quite small (< 1000 elements), single threaded |
| 70 | benchmarks show CSL can be 10x slower than std:set. |
| 71 | |
| 72 | 4. The interface requires using an Accessor to access the skiplist. |
| 73 | (See below.) |
| 74 | |
| 75 | 5. Currently x64 only, due to use of MicroSpinLock. |
| 76 | |
| 77 | 6. Freed nodes will not be reclaimed as long as there are ongoing |
| 78 | uses of the list. |
| 79 | |
| 80 | Sample usage: |
| 81 | |
| 82 | typedef ConcurrentSkipList<int> SkipListT; |
| 83 | shared_ptr<SkipListT> sl(SkipListT::createInstance(init_head_height); |
| 84 | { |
| 85 | // It's usually good practice to hold an accessor only during |
| 86 | // its necessary life cycle (but not in a tight loop as |
| 87 | // Accessor creation incurs ref-counting overhead). |
| 88 | // |
| 89 | // Holding it longer delays garbage-collecting the deleted |
| 90 | // nodes in the list. |
| 91 | SkipListT::Accessor accessor(sl); |
| 92 | accessor.insert(23); |
| 93 | accessor.erase(2); |
| 94 | for (auto &elem : accessor) { |
| 95 | // use elem to access data |
| 96 | } |
| 97 | ... ... |
| 98 | } |
| 99 | |
| 100 | Another useful type is the Skipper accessor. This is useful if you |
| 101 | want to skip to locations in the way std::lower_bound() works, |
| 102 | i.e. it can be used for going through the list by skipping to the |
| 103 | node no less than a specified key. The Skipper keeps its location as |
| 104 | state, which makes it convenient for things like implementing |
| 105 | intersection of two sets efficiently, as it can start from the last |
| 106 | visited position. |
| 107 | |
| 108 | { |
| 109 | SkipListT::Accessor accessor(sl); |
| 110 | SkipListT::Skipper skipper(accessor); |
| 111 | skipper.to(30); |
| 112 | if (skipper) { |
| 113 | CHECK_LE(30, *skipper); |
| 114 | } |
| 115 | ... ... |
| 116 | // GC may happen when the accessor gets destructed. |
| 117 | } |
| 118 | */ |
| 119 | |
| 120 | #pragma once |
| 121 | |
| 122 | #include <algorithm> |
| 123 | #include <atomic> |
| 124 | #include <limits> |
| 125 | #include <memory> |
| 126 | #include <type_traits> |
| 127 | |
| 128 | #include <boost/iterator/iterator_facade.hpp> |
| 129 | #include <glog/logging.h> |
| 130 | |
| 131 | #include <folly/ConcurrentSkipList-inl.h> |
| 132 | #include <folly/Likely.h> |
| 133 | #include <folly/Memory.h> |
| 134 | #include <folly/synchronization/MicroSpinLock.h> |
| 135 | |
| 136 | namespace folly { |
| 137 | |
| 138 | template < |
| 139 | typename T, |
| 140 | typename Comp = std::less<T>, |
| 141 | // All nodes are allocated using provided SysAllocator, |
| 142 | // it should be thread-safe. |
| 143 | typename NodeAlloc = SysAllocator<void>, |
| 144 | int MAX_HEIGHT = 24> |
| 145 | class ConcurrentSkipList { |
| 146 | // MAX_HEIGHT needs to be at least 2 to suppress compiler |
| 147 | // warnings/errors (Werror=uninitialized tiggered due to preds_[1] |
| 148 | // being treated as a scalar in the compiler). |
| 149 | static_assert( |
| 150 | MAX_HEIGHT >= 2 && MAX_HEIGHT < 64, |
| 151 | "MAX_HEIGHT can only be in the range of [2, 64)" ); |
| 152 | typedef std::unique_lock<folly::MicroSpinLock> ScopedLocker; |
| 153 | typedef ConcurrentSkipList<T, Comp, NodeAlloc, MAX_HEIGHT> SkipListType; |
| 154 | |
| 155 | public: |
| 156 | typedef detail::SkipListNode<T> NodeType; |
| 157 | typedef T value_type; |
| 158 | typedef T key_type; |
| 159 | |
| 160 | typedef detail::csl_iterator<value_type, NodeType> iterator; |
| 161 | typedef detail::csl_iterator<const value_type, const NodeType> const_iterator; |
| 162 | |
| 163 | class Accessor; |
| 164 | class Skipper; |
| 165 | |
| 166 | explicit ConcurrentSkipList(int height, const NodeAlloc& alloc) |
| 167 | : recycler_(alloc), |
| 168 | head_(NodeType::create(recycler_.alloc(), height, value_type(), true)), |
| 169 | size_(0) {} |
| 170 | |
| 171 | explicit ConcurrentSkipList(int height) |
| 172 | : recycler_(), |
| 173 | head_(NodeType::create(recycler_.alloc(), height, value_type(), true)), |
| 174 | size_(0) {} |
| 175 | |
| 176 | // Convenient function to get an Accessor to a new instance. |
| 177 | static Accessor create(int height, const NodeAlloc& alloc) { |
| 178 | return Accessor(createInstance(height, alloc)); |
| 179 | } |
| 180 | |
| 181 | static Accessor create(int height = 1) { |
| 182 | return Accessor(createInstance(height)); |
| 183 | } |
| 184 | |
| 185 | // Create a shared_ptr skiplist object with initial head height. |
| 186 | static std::shared_ptr<SkipListType> createInstance( |
| 187 | int height, |
| 188 | const NodeAlloc& alloc) { |
| 189 | return std::make_shared<ConcurrentSkipList>(height, alloc); |
| 190 | } |
| 191 | |
| 192 | static std::shared_ptr<SkipListType> createInstance(int height = 1) { |
| 193 | return std::make_shared<ConcurrentSkipList>(height); |
| 194 | } |
| 195 | |
| 196 | //=================================================================== |
| 197 | // Below are implementation details. |
| 198 | // Please see ConcurrentSkipList::Accessor for stdlib-like APIs. |
| 199 | //=================================================================== |
| 200 | |
| 201 | ~ConcurrentSkipList() { |
| 202 | if /* constexpr */ (NodeType::template DestroyIsNoOp<NodeAlloc>::value) { |
| 203 | // Avoid traversing the list if using arena allocator. |
| 204 | return; |
| 205 | } |
| 206 | for (NodeType* current = head_.load(std::memory_order_relaxed); current;) { |
| 207 | NodeType* tmp = current->skip(0); |
| 208 | NodeType::destroy(recycler_.alloc(), current); |
| 209 | current = tmp; |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | private: |
| 214 | static bool greater(const value_type& data, const NodeType* node) { |
| 215 | return node && Comp()(node->data(), data); |
| 216 | } |
| 217 | |
| 218 | static bool less(const value_type& data, const NodeType* node) { |
| 219 | return (node == nullptr) || Comp()(data, node->data()); |
| 220 | } |
| 221 | |
| 222 | static int findInsertionPoint( |
| 223 | NodeType* cur, |
| 224 | int cur_layer, |
| 225 | const value_type& data, |
| 226 | NodeType* preds[], |
| 227 | NodeType* succs[]) { |
| 228 | int foundLayer = -1; |
| 229 | NodeType* pred = cur; |
| 230 | NodeType* foundNode = nullptr; |
| 231 | for (int layer = cur_layer; layer >= 0; --layer) { |
| 232 | NodeType* node = pred->skip(layer); |
| 233 | while (greater(data, node)) { |
| 234 | pred = node; |
| 235 | node = node->skip(layer); |
| 236 | } |
| 237 | if (foundLayer == -1 && !less(data, node)) { // the two keys equal |
| 238 | foundLayer = layer; |
| 239 | foundNode = node; |
| 240 | } |
| 241 | preds[layer] = pred; |
| 242 | |
| 243 | // if found, succs[0..foundLayer] need to point to the cached foundNode, |
| 244 | // as foundNode might be deleted at the same time thus pred->skip() can |
| 245 | // return nullptr or another node. |
| 246 | succs[layer] = foundNode ? foundNode : node; |
| 247 | } |
| 248 | return foundLayer; |
| 249 | } |
| 250 | |
| 251 | size_t size() const { |
| 252 | return size_.load(std::memory_order_relaxed); |
| 253 | } |
| 254 | |
| 255 | int height() const { |
| 256 | return head_.load(std::memory_order_consume)->height(); |
| 257 | } |
| 258 | |
| 259 | int maxLayer() const { |
| 260 | return height() - 1; |
| 261 | } |
| 262 | |
| 263 | size_t incrementSize(int delta) { |
| 264 | return size_.fetch_add(delta, std::memory_order_relaxed) + delta; |
| 265 | } |
| 266 | |
| 267 | // Returns the node if found, nullptr otherwise. |
| 268 | NodeType* find(const value_type& data) { |
| 269 | auto ret = findNode(data); |
| 270 | if (ret.second && !ret.first->markedForRemoval()) { |
| 271 | return ret.first; |
| 272 | } |
| 273 | return nullptr; |
| 274 | } |
| 275 | |
| 276 | // lock all the necessary nodes for changing (adding or removing) the list. |
| 277 | // returns true if all the lock acquried successfully and the related nodes |
| 278 | // are all validate (not in certain pending states), false otherwise. |
| 279 | bool lockNodesForChange( |
| 280 | int nodeHeight, |
| 281 | ScopedLocker guards[MAX_HEIGHT], |
| 282 | NodeType* preds[MAX_HEIGHT], |
| 283 | NodeType* succs[MAX_HEIGHT], |
| 284 | bool adding = true) { |
| 285 | NodeType *pred, *succ, *prevPred = nullptr; |
| 286 | bool valid = true; |
| 287 | for (int layer = 0; valid && layer < nodeHeight; ++layer) { |
| 288 | pred = preds[layer]; |
| 289 | DCHECK(pred != nullptr) << "layer=" << layer << " height=" << height() |
| 290 | << " nodeheight=" << nodeHeight; |
| 291 | succ = succs[layer]; |
| 292 | if (pred != prevPred) { |
| 293 | guards[layer] = pred->acquireGuard(); |
| 294 | prevPred = pred; |
| 295 | } |
| 296 | valid = !pred->markedForRemoval() && |
| 297 | pred->skip(layer) == succ; // check again after locking |
| 298 | |
| 299 | if (adding) { // when adding a node, the succ shouldn't be going away |
| 300 | valid = valid && (succ == nullptr || !succ->markedForRemoval()); |
| 301 | } |
| 302 | } |
| 303 | |
| 304 | return valid; |
| 305 | } |
| 306 | |
| 307 | // Returns a paired value: |
| 308 | // pair.first always stores the pointer to the node with the same input key. |
| 309 | // It could be either the newly added data, or the existed data in the |
| 310 | // list with the same key. |
| 311 | // pair.second stores whether the data is added successfully: |
| 312 | // 0 means not added, otherwise reutrns the new size. |
| 313 | template <typename U> |
| 314 | std::pair<NodeType*, size_t> addOrGetData(U&& data) { |
| 315 | NodeType *preds[MAX_HEIGHT], *succs[MAX_HEIGHT]; |
| 316 | NodeType* newNode; |
| 317 | size_t newSize; |
| 318 | while (true) { |
| 319 | int max_layer = 0; |
| 320 | int layer = findInsertionPointGetMaxLayer(data, preds, succs, &max_layer); |
| 321 | |
| 322 | if (layer >= 0) { |
| 323 | NodeType* nodeFound = succs[layer]; |
| 324 | DCHECK(nodeFound != nullptr); |
| 325 | if (nodeFound->markedForRemoval()) { |
| 326 | continue; // if it's getting deleted retry finding node. |
| 327 | } |
| 328 | // wait until fully linked. |
| 329 | while (UNLIKELY(!nodeFound->fullyLinked())) { |
| 330 | } |
| 331 | return std::make_pair(nodeFound, 0); |
| 332 | } |
| 333 | |
| 334 | // need to capped at the original height -- the real height may have grown |
| 335 | int nodeHeight = |
| 336 | detail::SkipListRandomHeight::instance()->getHeight(max_layer + 1); |
| 337 | |
| 338 | ScopedLocker guards[MAX_HEIGHT]; |
| 339 | if (!lockNodesForChange(nodeHeight, guards, preds, succs)) { |
| 340 | continue; // give up the locks and retry until all valid |
| 341 | } |
| 342 | |
| 343 | // locks acquired and all valid, need to modify the links under the locks. |
| 344 | newNode = NodeType::create( |
| 345 | recycler_.alloc(), nodeHeight, std::forward<U>(data)); |
| 346 | for (int k = 0; k < nodeHeight; ++k) { |
| 347 | newNode->setSkip(k, succs[k]); |
| 348 | preds[k]->setSkip(k, newNode); |
| 349 | } |
| 350 | |
| 351 | newNode->setFullyLinked(); |
| 352 | newSize = incrementSize(1); |
| 353 | break; |
| 354 | } |
| 355 | |
| 356 | int hgt = height(); |
| 357 | size_t sizeLimit = |
| 358 | detail::SkipListRandomHeight::instance()->getSizeLimit(hgt); |
| 359 | |
| 360 | if (hgt < MAX_HEIGHT && newSize > sizeLimit) { |
| 361 | growHeight(hgt + 1); |
| 362 | } |
| 363 | CHECK_GT(newSize, 0); |
| 364 | return std::make_pair(newNode, newSize); |
| 365 | } |
| 366 | |
| 367 | bool remove(const value_type& data) { |
| 368 | NodeType* nodeToDelete = nullptr; |
| 369 | ScopedLocker nodeGuard; |
| 370 | bool isMarked = false; |
| 371 | int nodeHeight = 0; |
| 372 | NodeType *preds[MAX_HEIGHT], *succs[MAX_HEIGHT]; |
| 373 | |
| 374 | while (true) { |
| 375 | int max_layer = 0; |
| 376 | int layer = findInsertionPointGetMaxLayer(data, preds, succs, &max_layer); |
| 377 | if (!isMarked && (layer < 0 || !okToDelete(succs[layer], layer))) { |
| 378 | return false; |
| 379 | } |
| 380 | |
| 381 | if (!isMarked) { |
| 382 | nodeToDelete = succs[layer]; |
| 383 | nodeHeight = nodeToDelete->height(); |
| 384 | nodeGuard = nodeToDelete->acquireGuard(); |
| 385 | if (nodeToDelete->markedForRemoval()) { |
| 386 | return false; |
| 387 | } |
| 388 | nodeToDelete->setMarkedForRemoval(); |
| 389 | isMarked = true; |
| 390 | } |
| 391 | |
| 392 | // acquire pred locks from bottom layer up |
| 393 | ScopedLocker guards[MAX_HEIGHT]; |
| 394 | if (!lockNodesForChange(nodeHeight, guards, preds, succs, false)) { |
| 395 | continue; // this will unlock all the locks |
| 396 | } |
| 397 | |
| 398 | for (int k = nodeHeight - 1; k >= 0; --k) { |
| 399 | preds[k]->setSkip(k, nodeToDelete->skip(k)); |
| 400 | } |
| 401 | |
| 402 | incrementSize(-1); |
| 403 | break; |
| 404 | } |
| 405 | recycle(nodeToDelete); |
| 406 | return true; |
| 407 | } |
| 408 | |
| 409 | const value_type* first() const { |
| 410 | auto node = head_.load(std::memory_order_consume)->skip(0); |
| 411 | return node ? &node->data() : nullptr; |
| 412 | } |
| 413 | |
| 414 | const value_type* last() const { |
| 415 | NodeType* pred = head_.load(std::memory_order_consume); |
| 416 | NodeType* node = nullptr; |
| 417 | for (int layer = maxLayer(); layer >= 0; --layer) { |
| 418 | do { |
| 419 | node = pred->skip(layer); |
| 420 | if (node) { |
| 421 | pred = node; |
| 422 | } |
| 423 | } while (node != nullptr); |
| 424 | } |
| 425 | return pred == head_.load(std::memory_order_relaxed) ? nullptr |
| 426 | : &pred->data(); |
| 427 | } |
| 428 | |
| 429 | static bool okToDelete(NodeType* candidate, int layer) { |
| 430 | DCHECK(candidate != nullptr); |
| 431 | return candidate->fullyLinked() && candidate->maxLayer() == layer && |
| 432 | !candidate->markedForRemoval(); |
| 433 | } |
| 434 | |
| 435 | // find node for insertion/deleting |
| 436 | int findInsertionPointGetMaxLayer( |
| 437 | const value_type& data, |
| 438 | NodeType* preds[], |
| 439 | NodeType* succs[], |
| 440 | int* max_layer) const { |
| 441 | *max_layer = maxLayer(); |
| 442 | return findInsertionPoint( |
| 443 | head_.load(std::memory_order_consume), *max_layer, data, preds, succs); |
| 444 | } |
| 445 | |
| 446 | // Find node for access. Returns a paired values: |
| 447 | // pair.first = the first node that no-less than data value |
| 448 | // pair.second = 1 when the data value is founded, or 0 otherwise. |
| 449 | // This is like lower_bound, but not exact: we could have the node marked for |
| 450 | // removal so still need to check that. |
| 451 | std::pair<NodeType*, int> findNode(const value_type& data) const { |
| 452 | return findNodeDownRight(data); |
| 453 | } |
| 454 | |
| 455 | // Find node by first stepping down then stepping right. Based on benchmark |
| 456 | // results, this is slightly faster than findNodeRightDown for better |
| 457 | // localality on the skipping pointers. |
| 458 | std::pair<NodeType*, int> findNodeDownRight(const value_type& data) const { |
| 459 | NodeType* pred = head_.load(std::memory_order_consume); |
| 460 | int ht = pred->height(); |
| 461 | NodeType* node = nullptr; |
| 462 | |
| 463 | bool found = false; |
| 464 | while (!found) { |
| 465 | // stepping down |
| 466 | for (; ht > 0 && less(data, node = pred->skip(ht - 1)); --ht) { |
| 467 | } |
| 468 | if (ht == 0) { |
| 469 | return std::make_pair(node, 0); // not found |
| 470 | } |
| 471 | // node <= data now, but we need to fix up ht |
| 472 | --ht; |
| 473 | |
| 474 | // stepping right |
| 475 | while (greater(data, node)) { |
| 476 | pred = node; |
| 477 | node = node->skip(ht); |
| 478 | } |
| 479 | found = !less(data, node); |
| 480 | } |
| 481 | return std::make_pair(node, found); |
| 482 | } |
| 483 | |
| 484 | // find node by first stepping right then stepping down. |
| 485 | // We still keep this for reference purposes. |
| 486 | std::pair<NodeType*, int> findNodeRightDown(const value_type& data) const { |
| 487 | NodeType* pred = head_.load(std::memory_order_consume); |
| 488 | NodeType* node = nullptr; |
| 489 | auto top = maxLayer(); |
| 490 | int found = 0; |
| 491 | for (int layer = top; !found && layer >= 0; --layer) { |
| 492 | node = pred->skip(layer); |
| 493 | while (greater(data, node)) { |
| 494 | pred = node; |
| 495 | node = node->skip(layer); |
| 496 | } |
| 497 | found = !less(data, node); |
| 498 | } |
| 499 | return std::make_pair(node, found); |
| 500 | } |
| 501 | |
| 502 | NodeType* lower_bound(const value_type& data) const { |
| 503 | auto node = findNode(data).first; |
| 504 | while (node != nullptr && node->markedForRemoval()) { |
| 505 | node = node->skip(0); |
| 506 | } |
| 507 | return node; |
| 508 | } |
| 509 | |
| 510 | void growHeight(int height) { |
| 511 | NodeType* oldHead = head_.load(std::memory_order_consume); |
| 512 | if (oldHead->height() >= height) { // someone else already did this |
| 513 | return; |
| 514 | } |
| 515 | |
| 516 | NodeType* newHead = |
| 517 | NodeType::create(recycler_.alloc(), height, value_type(), true); |
| 518 | |
| 519 | { // need to guard the head node in case others are adding/removing |
| 520 | // nodes linked to the head. |
| 521 | ScopedLocker g = oldHead->acquireGuard(); |
| 522 | newHead->copyHead(oldHead); |
| 523 | NodeType* expected = oldHead; |
| 524 | if (!head_.compare_exchange_strong( |
| 525 | expected, newHead, std::memory_order_release)) { |
| 526 | // if someone has already done the swap, just return. |
| 527 | NodeType::destroy(recycler_.alloc(), newHead); |
| 528 | return; |
| 529 | } |
| 530 | oldHead->setMarkedForRemoval(); |
| 531 | } |
| 532 | recycle(oldHead); |
| 533 | } |
| 534 | |
| 535 | void recycle(NodeType* node) { |
| 536 | recycler_.add(node); |
| 537 | } |
| 538 | |
| 539 | detail::NodeRecycler<NodeType, NodeAlloc> recycler_; |
| 540 | std::atomic<NodeType*> head_; |
| 541 | std::atomic<size_t> size_; |
| 542 | }; |
| 543 | |
| 544 | template <typename T, typename Comp, typename NodeAlloc, int MAX_HEIGHT> |
| 545 | class ConcurrentSkipList<T, Comp, NodeAlloc, MAX_HEIGHT>::Accessor { |
| 546 | typedef detail::SkipListNode<T> NodeType; |
| 547 | typedef ConcurrentSkipList<T, Comp, NodeAlloc, MAX_HEIGHT> SkipListType; |
| 548 | |
| 549 | public: |
| 550 | typedef T value_type; |
| 551 | typedef T key_type; |
| 552 | typedef T& reference; |
| 553 | typedef T* pointer; |
| 554 | typedef const T& const_reference; |
| 555 | typedef const T* const_pointer; |
| 556 | typedef size_t size_type; |
| 557 | typedef Comp key_compare; |
| 558 | typedef Comp value_compare; |
| 559 | |
| 560 | typedef typename SkipListType::iterator iterator; |
| 561 | typedef typename SkipListType::const_iterator const_iterator; |
| 562 | typedef typename SkipListType::Skipper Skipper; |
| 563 | |
| 564 | explicit Accessor(std::shared_ptr<ConcurrentSkipList> skip_list) |
| 565 | : slHolder_(std::move(skip_list)) { |
| 566 | sl_ = slHolder_.get(); |
| 567 | DCHECK(sl_ != nullptr); |
| 568 | sl_->recycler_.addRef(); |
| 569 | } |
| 570 | |
| 571 | // Unsafe initializer: the caller assumes the responsibility to keep |
| 572 | // skip_list valid during the whole life cycle of the Acessor. |
| 573 | explicit Accessor(ConcurrentSkipList* skip_list) : sl_(skip_list) { |
| 574 | DCHECK(sl_ != nullptr); |
| 575 | sl_->recycler_.addRef(); |
| 576 | } |
| 577 | |
| 578 | Accessor(const Accessor& accessor) |
| 579 | : sl_(accessor.sl_), slHolder_(accessor.slHolder_) { |
| 580 | sl_->recycler_.addRef(); |
| 581 | } |
| 582 | |
| 583 | Accessor& operator=(const Accessor& accessor) { |
| 584 | if (this != &accessor) { |
| 585 | slHolder_ = accessor.slHolder_; |
| 586 | sl_->recycler_.releaseRef(); |
| 587 | sl_ = accessor.sl_; |
| 588 | sl_->recycler_.addRef(); |
| 589 | } |
| 590 | return *this; |
| 591 | } |
| 592 | |
| 593 | ~Accessor() { |
| 594 | sl_->recycler_.releaseRef(); |
| 595 | } |
| 596 | |
| 597 | bool empty() const { |
| 598 | return sl_->size() == 0; |
| 599 | } |
| 600 | size_t size() const { |
| 601 | return sl_->size(); |
| 602 | } |
| 603 | size_type max_size() const { |
| 604 | return std::numeric_limits<size_type>::max(); |
| 605 | } |
| 606 | |
| 607 | // returns end() if the value is not in the list, otherwise returns an |
| 608 | // iterator pointing to the data, and it's guaranteed that the data is valid |
| 609 | // as far as the Accessor is hold. |
| 610 | iterator find(const key_type& value) { |
| 611 | return iterator(sl_->find(value)); |
| 612 | } |
| 613 | const_iterator find(const key_type& value) const { |
| 614 | return iterator(sl_->find(value)); |
| 615 | } |
| 616 | size_type count(const key_type& data) const { |
| 617 | return contains(data); |
| 618 | } |
| 619 | |
| 620 | iterator begin() const { |
| 621 | NodeType* head = sl_->head_.load(std::memory_order_consume); |
| 622 | return iterator(head->next()); |
| 623 | } |
| 624 | iterator end() const { |
| 625 | return iterator(nullptr); |
| 626 | } |
| 627 | const_iterator cbegin() const { |
| 628 | return begin(); |
| 629 | } |
| 630 | const_iterator cend() const { |
| 631 | return end(); |
| 632 | } |
| 633 | |
| 634 | template < |
| 635 | typename U, |
| 636 | typename = |
| 637 | typename std::enable_if<std::is_convertible<U, T>::value>::type> |
| 638 | std::pair<iterator, bool> insert(U&& data) { |
| 639 | auto ret = sl_->addOrGetData(std::forward<U>(data)); |
| 640 | return std::make_pair(iterator(ret.first), ret.second); |
| 641 | } |
| 642 | size_t erase(const key_type& data) { |
| 643 | return remove(data); |
| 644 | } |
| 645 | |
| 646 | iterator lower_bound(const key_type& data) const { |
| 647 | return iterator(sl_->lower_bound(data)); |
| 648 | } |
| 649 | |
| 650 | size_t height() const { |
| 651 | return sl_->height(); |
| 652 | } |
| 653 | |
| 654 | // first() returns pointer to the first element in the skiplist, or |
| 655 | // nullptr if empty. |
| 656 | // |
| 657 | // last() returns the pointer to the last element in the skiplist, |
| 658 | // nullptr if list is empty. |
| 659 | // |
| 660 | // Note: As concurrent writing can happen, first() is not |
| 661 | // guaranteed to be the min_element() in the list. Similarly |
| 662 | // last() is not guaranteed to be the max_element(), and both of them can |
| 663 | // be invalid (i.e. nullptr), so we name them differently from front() and |
| 664 | // tail() here. |
| 665 | const key_type* first() const { |
| 666 | return sl_->first(); |
| 667 | } |
| 668 | const key_type* last() const { |
| 669 | return sl_->last(); |
| 670 | } |
| 671 | |
| 672 | // Try to remove the last element in the skip list. |
| 673 | // |
| 674 | // Returns true if we removed it, false if either the list is empty |
| 675 | // or a race condition happened (i.e. the used-to-be last element |
| 676 | // was already removed by another thread). |
| 677 | bool pop_back() { |
| 678 | auto last = sl_->last(); |
| 679 | return last ? sl_->remove(*last) : false; |
| 680 | } |
| 681 | |
| 682 | std::pair<key_type*, bool> addOrGetData(const key_type& data) { |
| 683 | auto ret = sl_->addOrGetData(data); |
| 684 | return std::make_pair(&ret.first->data(), ret.second); |
| 685 | } |
| 686 | |
| 687 | SkipListType* skiplist() const { |
| 688 | return sl_; |
| 689 | } |
| 690 | |
| 691 | // legacy interfaces |
| 692 | // TODO:(xliu) remove these. |
| 693 | // Returns true if the node is added successfully, false if not, i.e. the |
| 694 | // node with the same key already existed in the list. |
| 695 | bool contains(const key_type& data) const { |
| 696 | return sl_->find(data); |
| 697 | } |
| 698 | bool add(const key_type& data) { |
| 699 | return sl_->addOrGetData(data).second; |
| 700 | } |
| 701 | bool remove(const key_type& data) { |
| 702 | return sl_->remove(data); |
| 703 | } |
| 704 | |
| 705 | private: |
| 706 | SkipListType* sl_; |
| 707 | std::shared_ptr<SkipListType> slHolder_; |
| 708 | }; |
| 709 | |
| 710 | // implements forward iterator concept. |
| 711 | template <typename ValT, typename NodeT> |
| 712 | class detail::csl_iterator : public boost::iterator_facade< |
| 713 | csl_iterator<ValT, NodeT>, |
| 714 | ValT, |
| 715 | boost::forward_traversal_tag> { |
| 716 | public: |
| 717 | typedef ValT value_type; |
| 718 | typedef value_type& reference; |
| 719 | typedef value_type* pointer; |
| 720 | typedef ptrdiff_t difference_type; |
| 721 | |
| 722 | explicit csl_iterator(NodeT* node = nullptr) : node_(node) {} |
| 723 | |
| 724 | template <typename OtherVal, typename OtherNode> |
| 725 | csl_iterator( |
| 726 | const csl_iterator<OtherVal, OtherNode>& other, |
| 727 | typename std::enable_if< |
| 728 | std::is_convertible<OtherVal, ValT>::value>::type* = nullptr) |
| 729 | : node_(other.node_) {} |
| 730 | |
| 731 | size_t nodeSize() const { |
| 732 | return node_ == nullptr ? 0 |
| 733 | : node_->height() * sizeof(NodeT*) + sizeof(*this); |
| 734 | } |
| 735 | |
| 736 | bool good() const { |
| 737 | return node_ != nullptr; |
| 738 | } |
| 739 | |
| 740 | private: |
| 741 | friend class boost::iterator_core_access; |
| 742 | template <class, class> |
| 743 | friend class csl_iterator; |
| 744 | |
| 745 | void increment() { |
| 746 | node_ = node_->next(); |
| 747 | } |
| 748 | bool equal(const csl_iterator& other) const { |
| 749 | return node_ == other.node_; |
| 750 | } |
| 751 | value_type& dereference() const { |
| 752 | return node_->data(); |
| 753 | } |
| 754 | |
| 755 | NodeT* node_; |
| 756 | }; |
| 757 | |
| 758 | // Skipper interface |
| 759 | template <typename T, typename Comp, typename NodeAlloc, int MAX_HEIGHT> |
| 760 | class ConcurrentSkipList<T, Comp, NodeAlloc, MAX_HEIGHT>::Skipper { |
| 761 | typedef detail::SkipListNode<T> NodeType; |
| 762 | typedef ConcurrentSkipList<T, Comp, NodeAlloc, MAX_HEIGHT> SkipListType; |
| 763 | typedef typename SkipListType::Accessor Accessor; |
| 764 | |
| 765 | public: |
| 766 | typedef T value_type; |
| 767 | typedef T& reference; |
| 768 | typedef T* pointer; |
| 769 | typedef ptrdiff_t difference_type; |
| 770 | |
| 771 | Skipper(const std::shared_ptr<SkipListType>& skipList) : accessor_(skipList) { |
| 772 | init(); |
| 773 | } |
| 774 | |
| 775 | Skipper(const Accessor& accessor) : accessor_(accessor) { |
| 776 | init(); |
| 777 | } |
| 778 | |
| 779 | void init() { |
| 780 | // need to cache the head node |
| 781 | NodeType* head_node = head(); |
| 782 | headHeight_ = head_node->height(); |
| 783 | for (int i = 0; i < headHeight_; ++i) { |
| 784 | preds_[i] = head_node; |
| 785 | succs_[i] = head_node->skip(i); |
| 786 | } |
| 787 | int max_layer = maxLayer(); |
| 788 | for (int i = 0; i < max_layer; ++i) { |
| 789 | hints_[i] = uint8_t(i + 1); |
| 790 | } |
| 791 | hints_[max_layer] = max_layer; |
| 792 | } |
| 793 | |
| 794 | // advance to the next node in the list. |
| 795 | Skipper& operator++() { |
| 796 | preds_[0] = succs_[0]; |
| 797 | succs_[0] = preds_[0]->skip(0); |
| 798 | int height = curHeight(); |
| 799 | for (int i = 1; i < height && preds_[0] == succs_[i]; ++i) { |
| 800 | preds_[i] = succs_[i]; |
| 801 | succs_[i] = preds_[i]->skip(i); |
| 802 | } |
| 803 | return *this; |
| 804 | } |
| 805 | |
| 806 | bool good() const { |
| 807 | return succs_[0] != nullptr; |
| 808 | } |
| 809 | |
| 810 | int maxLayer() const { |
| 811 | return headHeight_ - 1; |
| 812 | } |
| 813 | |
| 814 | int curHeight() const { |
| 815 | // need to cap the height to the cached head height, as the current node |
| 816 | // might be some newly inserted node and also during the time period the |
| 817 | // head height may have grown. |
| 818 | return succs_[0] ? std::min(headHeight_, succs_[0]->height()) : 0; |
| 819 | } |
| 820 | |
| 821 | const value_type& data() const { |
| 822 | DCHECK(succs_[0] != nullptr); |
| 823 | return succs_[0]->data(); |
| 824 | } |
| 825 | |
| 826 | value_type& operator*() const { |
| 827 | DCHECK(succs_[0] != nullptr); |
| 828 | return succs_[0]->data(); |
| 829 | } |
| 830 | |
| 831 | value_type* operator->() { |
| 832 | DCHECK(succs_[0] != nullptr); |
| 833 | return &succs_[0]->data(); |
| 834 | } |
| 835 | |
| 836 | /* |
| 837 | * Skip to the position whose data is no less than the parameter. |
| 838 | * (I.e. the lower_bound). |
| 839 | * |
| 840 | * Returns true if the data is found, false otherwise. |
| 841 | */ |
| 842 | bool to(const value_type& data) { |
| 843 | int layer = curHeight() - 1; |
| 844 | if (layer < 0) { |
| 845 | return false; // reaches the end of the list |
| 846 | } |
| 847 | |
| 848 | int lyr = hints_[layer]; |
| 849 | int max_layer = maxLayer(); |
| 850 | while (SkipListType::greater(data, succs_[lyr]) && lyr < max_layer) { |
| 851 | ++lyr; |
| 852 | } |
| 853 | hints_[layer] = lyr; // update the hint |
| 854 | |
| 855 | int foundLayer = SkipListType::findInsertionPoint( |
| 856 | preds_[lyr], lyr, data, preds_, succs_); |
| 857 | if (foundLayer < 0) { |
| 858 | return false; |
| 859 | } |
| 860 | |
| 861 | DCHECK(succs_[0] != nullptr) |
| 862 | << "lyr=" << lyr << "; max_layer=" << max_layer; |
| 863 | return !succs_[0]->markedForRemoval(); |
| 864 | } |
| 865 | |
| 866 | private: |
| 867 | NodeType* head() const { |
| 868 | return accessor_.skiplist()->head_.load(std::memory_order_consume); |
| 869 | } |
| 870 | |
| 871 | Accessor accessor_; |
| 872 | int headHeight_; |
| 873 | NodeType *succs_[MAX_HEIGHT], *preds_[MAX_HEIGHT]; |
| 874 | uint8_t hints_[MAX_HEIGHT]; |
| 875 | }; |
| 876 | |
| 877 | } // namespace folly |
| 878 | |