This contains my bachelors thesis and associated tex files, code snippets and maybe more. Topic: Data Movement in Heterogeneous Memories with Intel Data Streaming Accelerator
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

801 lines
30 KiB

  1. #pragma once
  2. #include <iostream>
  3. #include <unordered_map>
  4. #include <shared_mutex>
  5. #include <mutex>
  6. #include <memory>
  7. #include <sched.h>
  8. #include <numa.h>
  9. #include <numaif.h>
  10. #include <dml/dml.hpp>
  11. namespace dml {
  12. inline const std::string StatusCodeToString(const dml::status_code code) {
  13. switch (code) {
  14. case dml::status_code::ok: return "ok";
  15. case dml::status_code::false_predicate: return "false predicate";
  16. case dml::status_code::partial_completion: return "partial completion";
  17. case dml::status_code::nullptr_error: return "nullptr error";
  18. case dml::status_code::bad_size: return "bad size";
  19. case dml::status_code::bad_length: return "bad length";
  20. case dml::status_code::inconsistent_size: return "inconsistent size";
  21. case dml::status_code::dualcast_bad_padding: return "dualcast bad padding";
  22. case dml::status_code::bad_alignment: return "bad alignment";
  23. case dml::status_code::buffers_overlapping: return "buffers overlapping";
  24. case dml::status_code::delta_delta_empty: return "delta delta empty";
  25. case dml::status_code::batch_overflow: return "batch overflow";
  26. case dml::status_code::execution_failed: return "execution failed";
  27. case dml::status_code::unsupported_operation: return "unsupported operation";
  28. case dml::status_code::queue_busy: return "queue busy";
  29. case dml::status_code::error: return "unknown error";
  30. case dml::status_code::config_error: return "config error";
  31. default: return "unhandled error";
  32. }
  33. }
  34. }
  35. namespace dsacache {
  36. inline bool CheckFlag(const uint64_t value, const uint64_t flag) {
  37. return (value & flag) != 0;
  38. }
  39. inline uint64_t UnsetFlag(const uint64_t value, const uint64_t flag) {
  40. return value & (~flag);
  41. }
  42. inline uint64_t SetFlag(const uint64_t value, const uint64_t flag) {
  43. return value | flag;
  44. }
  45. constexpr uint64_t FLAG_WAIT_WEAK = 0b1ULL << 63;
  46. constexpr uint64_t FLAG_HANDLE_PF = 0b1ULL << 62;
  47. constexpr uint64_t FLAG_ACCESS_WEAK = 0b1ULL << 61;
  48. constexpr uint64_t FLAG_FORCE_MAP_PAGES = 0b1ULL << 60;
  49. constexpr uint64_t FLAG_DEFAULT = 0ULL;
  50. class Cache;
  51. // cache policy is defined as a type here to allow flexible usage of the cacher
  52. // given a numa destination node (where the data will be needed), the numa source
  53. // node (current location of the data) and the data size, this function should
  54. // return optimal cache placement
  55. // dst node and returned value can differ if the system, for example, has HBM
  56. // attached accessible directly to node n under a different node id m
  57. typedef int (CachePolicy)(const int numa_dst_node, const int numa_src_node, const size_t data_size);
  58. // copy policy specifies the copy-executing nodes for a given task
  59. // which allows flexibility in assignment for optimizing raw throughput
  60. // or choosing a conservative usage policy
  61. typedef std::vector<int> (CopyPolicy)(const int numa_dst_node, const int numa_src_node, const size_t data_size);
  62. // memory allocation is a complex topic but can have big performance
  63. // impact, we therefore do not handle it in the cache. must return
  64. // pointer to a block of at least the given size, cache will also
  65. // not handle deallocation but signal that the block is free
  66. typedef uint8_t* (MemoryAllocator_Allocate)(const int numa_node, const size_t size);
  67. // signals that the given memory block will not be used by cache anymore
  68. typedef void (MemoryAllocator_Free)(uint8_t* pointer, const size_t size);
  69. /*
  70. * Class Description:
  71. * Holds all required information on one cache entry and is used
  72. * both internally by the Cache and externally by the user.
  73. *
  74. * Important Usage Notes:
  75. * The pointer is only updated in WaitOnCompletion() which
  76. * therefore must be called by the user at some point in order
  77. * to use the cached data. Using this class as T for
  78. * std::shared_ptr<T> is not recommended as references are
  79. * already counted internally.
  80. *
  81. * Cache Lifetime:
  82. * As long as the instance is referenced, the pointer it stores
  83. * is guaranteed to be either nullptr or pointing to a valid copy.
  84. *
  85. * Implementation Detail:
  86. * Performs self-reference counting with a shared atomic integer.
  87. * Therefore on creating a copy the reference count is increased
  88. * and with the destructor it is deacresed. If the last copy is
  89. * destroyed the actual underlying data is freed and all shared
  90. * variables deleted.
  91. *
  92. * Notes on Thread Safety:
  93. * Class is thread safe in any possible state and performs
  94. * reference counting and deallocation itself entirely atomically.
  95. */
  96. class CacheData {
  97. public:
  98. using dml_handler = dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>>;
  99. private:
  100. // set to false if we do not own the cache pointer
  101. bool delete_ = false;
  102. MemoryAllocator_Free* memory_free_function_;
  103. // data source and size of the block
  104. uint8_t* src_;
  105. size_t size_;
  106. // global reference counting object
  107. std::atomic<int32_t>* active_;
  108. // global cache-location pointer
  109. std::atomic<uint8_t*>* cache_;
  110. // object-local incomplete cache location pointer
  111. // contract: only access when being in sole posession of handlers
  112. uint8_t** incomplete_cache_;
  113. // flags inherited from parent cache
  114. uint64_t flags_ = 0;
  115. // dml handler vector pointer which is used
  116. // to wait on caching task completion
  117. std::atomic<std::vector<dml_handler>*>* handlers_;
  118. // invalid handlers pointer as we need a secondary
  119. // invalid state due to issues with waiting
  120. std::vector<dml_handler>* invalid_handlers_;
  121. // deallocates the global cache-location
  122. // and invalidates it
  123. void Deallocate();
  124. size_t GetSize() const { return size_; }
  125. uint8_t* GetSource() const { return src_; }
  126. int32_t GetRefCount() const { return active_->load(); }
  127. void SetCacheToSource() { cache_->store(src_); delete_ = false; }
  128. void SetTaskHandlersAndCache(uint8_t* cache, std::vector<dml_handler>* handlers);
  129. // initializes the class after which it is thread safe
  130. // but may only be destroyed safely after setting handlers
  131. void Init(std::vector<dml_handler>* invalid_handlers);
  132. friend Cache;
  133. public:
  134. CacheData(uint8_t* data, const size_t size, MemoryAllocator_Free* free);
  135. CacheData(const CacheData& other);
  136. ~CacheData();
  137. // waits on completion of caching operations
  138. // for this task and is safe to be called in
  139. // any state of the object, if the flag
  140. // FLAG_WAIT_WEAK is set for this instance
  141. // (can also be inherited from the creating
  142. // Cache-Instance or on copy from another
  143. // CacheData-Instance), WaitOnCompletion
  144. // provides no validity guarantees to the
  145. // cache pointer (GetDataLocation() may
  146. // return nullptr even after return
  147. // of the wait function). On error this
  148. // function will set the cache pointer
  149. // to the source to provide validity
  150. // guarantees after returning.
  151. void WaitOnCompletion();
  152. // returns the cache data location for this
  153. // instance which is valid as long as the
  154. // instance is alive
  155. // !!! this may also return a nullptr !!!
  156. // see WaitOnCompletion() for how to achieve
  157. // validity guarantees if required
  158. uint8_t* GetDataLocation() const { return cache_->load(); }
  159. void SetFlags(const uint64_t flags) { flags_ = flags; }
  160. uint64_t GetFlags() const { return flags_; }
  161. };
  162. /*
  163. * Class Description:
  164. * Class will handle access to data through internal copies.
  165. * These are obtained via work submission to the Intel DSA which takes
  166. * care of asynchronously duplicating the data. The user will define
  167. * where these copies lie and which system nodes will perform the copy.
  168. * This is done through policy functions set during initialization.
  169. *
  170. * Placement Policy:
  171. * The Placement Policy Function decides on which node a particular
  172. * entry is to be placed, given the current executing node and the
  173. * data source node and data size. This in turn means that for one
  174. * datum, multiple cached copies may exist at one time.
  175. *
  176. * Cache Lifetime:
  177. * When accessing the cache, a CacheData-object will be returned.
  178. * As long as this object lives, the pointer which it holds is
  179. * guaranteed to be either nullptr or a valid copy. When destroyed
  180. * the entry is marked for deletion which is only carried out
  181. * when system memory pressure drives an automated cache flush.
  182. *
  183. * Restrictions:
  184. * - Overlapping Pointers may lead to undefined behaviour during
  185. * manual cache invalidation which should not be used if you
  186. * intend to have these types of pointers
  187. * - Cache Invalidation may only be performed manually and gives
  188. * no ordering guarantees. Therefore, it is the users responsibility
  189. * to ensure that results after invalidation have been generated
  190. * using the latest state of data. The cache is best suited
  191. * to static data.
  192. *
  193. * Notes on Thread Safety:
  194. * - Cache is completely thread-safe after initialization
  195. * - CacheData-class will handle deallocation of data itself by
  196. * performing self-reference-counting atomically and only
  197. * deallocating if the last reference is destroyed
  198. * - The internal cache state has one lock which is either
  199. * acquired shared for reading the state (upon accessing an already
  200. * cached element) or unique (accessing a new element, flushing, invalidating)
  201. * - Waiting on copy completion is done over an atomic-wait in copies
  202. * of the original CacheData-instance
  203. * - Overall this class may experience performance issues due to the use
  204. * of locking (in any configuration), lock contention (worsens with higher
  205. * core count, node count and utilization) and atomics (worse in the same
  206. * situations as lock contention)
  207. *
  208. * Improving Performance:
  209. * When data is never shared between threads or memory size for the cache is
  210. * not an issue you may consider having one Cache-instance per thread and removing
  211. * the lock in Cache and modifying the reference counting and waiting mechanisms
  212. * of CacheData accordingly (although this is high effort and will yield little due
  213. * to the atomics not being shared among cores/nodes).
  214. * Otherwise, one Cache-instance per node could also be considered. This will allow
  215. * the placement policy function to be barebones and reduces the lock contention and
  216. * synchronization impact of the atomic variables.
  217. */
  218. class Cache {
  219. private:
  220. // flags to store options duh
  221. uint64_t flags_ = 0;
  222. // secondary invalid handlers vector
  223. // needed due to wake-up issues in CacheData::WaitOnCompletion
  224. std::vector<CacheData::dml_handler> invalid_handlers_;
  225. // map from [dst-numa-node,map2]
  226. // map2 from [data-ptr,cache-structure]
  227. struct LockedNodeCacheState {
  228. std::shared_mutex cache_mutex_;
  229. std::unordered_map<uint8_t*, CacheData> node_cache_state_;
  230. };
  231. std::unordered_map<uint8_t, LockedNodeCacheState*> cache_state_;
  232. CachePolicy* cache_policy_function_ = nullptr;
  233. CopyPolicy* copy_policy_function_ = nullptr;
  234. MemoryAllocator_Allocate* memory_allocate_function_ = nullptr;
  235. MemoryAllocator_Free* memory_free_function_ = nullptr;
  236. // function used to submit a copy task on a specific node to the dml
  237. // engine on that node - will change the current threads node assignment
  238. // to achieve this so take care to restore this
  239. dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>> ExecuteCopy(
  240. const uint8_t* src, uint8_t* dst, const size_t size, const int node
  241. ) const;
  242. // allocates the required memory on the destination node
  243. // and then submits task to the dml library for processing
  244. // and attaches the handlers to the cache data structure
  245. void SubmitTask(CacheData* task, const int dst_node, const int src_node);
  246. // querries the policy functions for the given data and size
  247. // to obtain destination cache node, also returns the datas
  248. // source node for further usage
  249. // output may depend on the calling threads node assignment
  250. // as this is set as the "optimal placement" node
  251. void GetCacheNode(uint8_t* src, const size_t size, int* OUT_DST_NODE, int* OUT_SRC_NODE) const;
  252. // checks whether the cache contains an entry for
  253. // the given data in the given memory node and
  254. // returns it, otherwise returns nullptr
  255. std::unique_ptr<CacheData> GetFromCache(uint8_t* src, const size_t size, const int dst_node);
  256. public:
  257. ~Cache();
  258. Cache() = default;
  259. Cache(const Cache& other) = delete;
  260. // initializes the cache with the two policy functions
  261. // only after this is it safe to use in a threaded environment
  262. void Init(
  263. CachePolicy* cache_policy_function, CopyPolicy* copy_policy_function,
  264. MemoryAllocator_Allocate* memory_allocate_function,
  265. MemoryAllocator_Free* memory_free_function
  266. );
  267. // function to perform data access through the cache, behaviour depends
  268. // on flags, by default will also perform prefetch, otherwise with
  269. // FLAG_ACCESS_WEAK set will not perform prefetch and instead return
  270. // a cache entry with the data source as cache location on cache miss,
  271. // this flag must be set for each invocation, the flags set for the
  272. // entire cache will not be evaluated for this
  273. std::unique_ptr<CacheData> Access(uint8_t* data, const size_t size, const uint64_t flags = FLAG_DEFAULT);
  274. // flushes the cache of inactive entries
  275. // if node is -1 then the whole cache is
  276. // checked and otherwise the specified
  277. // node - no checks on node validity
  278. void Flush(const int node = -1);
  279. // forces out all entries from the
  280. // cache and therefore will also "forget"
  281. // still-in-use entries, these will still
  282. // be properly deleted, but the cache
  283. // will be fresh - use for testing
  284. void Clear();
  285. void Invalidate(uint8_t* data);
  286. void SetFlags(const uint64_t flags) { flags_ = flags; }
  287. uint64_t GetFlags() { return flags_; }
  288. };
  289. }
  290. inline void dsacache::Cache::Clear() {
  291. for (auto& nc : cache_state_) {
  292. std::unique_lock<std::shared_mutex> lock(nc.second->cache_mutex_);
  293. nc.second->node_cache_state_.clear();
  294. }
  295. }
  296. inline void dsacache::Cache::Init(
  297. CachePolicy* cache_policy_function, CopyPolicy* copy_policy_function,
  298. MemoryAllocator_Allocate* memory_allocate_function,
  299. MemoryAllocator_Free* memory_free_function
  300. ) {
  301. cache_policy_function_ = cache_policy_function;
  302. copy_policy_function_ = copy_policy_function;
  303. memory_allocate_function_ = memory_allocate_function;
  304. memory_free_function_ = memory_free_function;
  305. // initialize numa library
  306. numa_available();
  307. // obtain all available nodes
  308. // and those we may allocate
  309. // memory on
  310. const int nodes_max = numa_num_configured_nodes();
  311. const bitmask* valid_nodes = numa_get_mems_allowed();
  312. // prepare the cache state with entries
  313. // for all given nodes
  314. for (int node = 0; node < nodes_max; node++) {
  315. if (numa_bitmask_isbitset(valid_nodes, node)) {
  316. void* block = numa_alloc_onnode(sizeof(LockedNodeCacheState), node);
  317. auto* state = new(block)LockedNodeCacheState;
  318. cache_state_.insert({node,state});
  319. }
  320. }
  321. }
  322. inline std::unique_ptr<dsacache::CacheData> dsacache::Cache::Access(uint8_t* data, const size_t size, const uint64_t flags) {
  323. // get destination numa node for the cache
  324. int dst_node = -1;
  325. int src_node = -1;
  326. GetCacheNode(data, size, &dst_node, &src_node);
  327. // check whether the data is already cached
  328. std::unique_ptr<CacheData> task = GetFromCache(data, size, dst_node);
  329. if (task != nullptr) {
  330. return std::move(task);
  331. }
  332. // at this point the requested data is not present in cache
  333. // and we create a caching task for it, copying our current flags
  334. task = std::make_unique<CacheData>(data, size, memory_free_function_);
  335. task->SetFlags(flags_);
  336. // when the ACCESS_WEAK flag is set for the flags parameter (!)
  337. // and we have reached this point, there was no cache entry
  338. // present for the requested data and therefore we abort
  339. // but to keep validity, we return the previously created
  340. // CacheData struct, setting the cache variable to the
  341. // data source location
  342. if (CheckFlag(flags, FLAG_ACCESS_WEAK)) {
  343. task->SetCacheToSource();
  344. return std::move(task);
  345. }
  346. // the following operation adds the task to the cache state
  347. // which requires unique locking of the current nodes entry
  348. {
  349. LockedNodeCacheState* local_cache_state = cache_state_[dst_node];
  350. std::unique_lock<std::shared_mutex> lock(local_cache_state->cache_mutex_);
  351. const auto state = local_cache_state->node_cache_state_.emplace(task->GetSource(), *task);
  352. // if state.second is false then no insertion took place
  353. // which means that concurrently whith this thread
  354. // some other thread must have accessed the same
  355. // resource in which case we return the other
  356. // threads data cache structure
  357. if (!state.second) {
  358. return std::move(std::make_unique<CacheData>(state.first->second));
  359. }
  360. // initialize the task now for thread safety
  361. // as we are now sure that we will submit work
  362. // to it and will not delete it beforehand
  363. // of the one in cache state - must be
  364. // performed for the local and cache-state
  365. // instance as Init will modify values that
  366. // are not shared but copied on copy-construct
  367. state.first->second.Init(&invalid_handlers_);
  368. task->Init(&invalid_handlers_);
  369. }
  370. SubmitTask(task.get(), dst_node, src_node);
  371. return std::move(task);
  372. }
  373. inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, const int src_node) {
  374. uint8_t* dst = memory_allocate_function_(dst_node, task->GetSize());
  375. if (dst == nullptr) {
  376. return;
  377. }
  378. // querry copy policy function for the nodes to use for the copy
  379. const std::vector<int> executing_nodes = copy_policy_function_(dst_node, src_node, task->GetSize());
  380. const size_t task_count = executing_nodes.size();
  381. // each task will copy one fair part of the total size
  382. // and in case the total size is not a factor of the
  383. // given task count the last node must copy the remainder
  384. const size_t size = task->GetSize() / task_count;
  385. const size_t last_size = size + task->GetSize() % task_count;
  386. // save the current numa node mask to restore later
  387. // as executing the copy task will place this thread
  388. // on a different node
  389. auto handlers = new std::vector<CacheData::dml_handler>();
  390. for (uint32_t i = 0; i < task_count; i++) {
  391. const size_t local_size = i + 1 == task_count ? size : last_size;
  392. const size_t local_offset = i * size;
  393. const uint8_t* local_src = task->GetSource() + local_offset;
  394. uint8_t* local_dst = dst + local_offset;
  395. handlers->emplace_back(ExecuteCopy(local_src, local_dst, local_size, executing_nodes[i]));
  396. }
  397. task->SetTaskHandlersAndCache(dst, handlers);
  398. }
  399. inline dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>> dsacache::Cache::ExecuteCopy(
  400. const uint8_t* src, uint8_t* dst, const size_t size, const int node
  401. ) const {
  402. dml::const_data_view srcv = dml::make_view(src, size);
  403. dml::data_view dstv = dml::make_view(dst, size);
  404. if (CheckFlag(flags_, FLAG_HANDLE_PF)) {
  405. return dml::submit<dml::hardware>(
  406. dml::mem_copy.block_on_fault(), srcv, dstv,
  407. dml::execution_interface<dml::hardware,std::allocator<uint8_t>>(), node
  408. );
  409. }
  410. else {
  411. return dml::submit<dml::hardware>(
  412. dml::mem_copy, srcv, dstv,
  413. dml::execution_interface<dml::hardware,std::allocator<uint8_t>>(), node
  414. );
  415. }
  416. }
  417. inline void dsacache::Cache::GetCacheNode(uint8_t* src, const size_t size, int* OUT_DST_NODE, int* OUT_SRC_NODE) const {
  418. // obtain numa node of current thread to determine where the data is needed
  419. const int current_cpu = sched_getcpu();
  420. const int current_node = numa_node_of_cpu(current_cpu);
  421. // obtain node that the given data pointer is allocated on
  422. *OUT_SRC_NODE = -1;
  423. get_mempolicy(OUT_SRC_NODE, NULL, 0, (void*)src, MPOL_F_NODE | MPOL_F_ADDR);
  424. // querry cache policy function for the destination numa node
  425. *OUT_DST_NODE = cache_policy_function_(current_node, *OUT_SRC_NODE, size);
  426. }
  427. inline void dsacache::Cache::Flush(const int node) {
  428. // this lambda is used because below we have two code paths that
  429. // flush nodes, either one single or all successively
  430. const auto FlushNode = [](std::unordered_map<uint8_t*,CacheData>& map) {
  431. // begin at the front of the map
  432. auto it = map.begin();
  433. // loop until we reach the end of the map
  434. while (it != map.end()) {
  435. // if the iterator points to an inactive element
  436. // then we may erase it
  437. if (it->second.GetRefCount() <= 1) {
  438. // erase the iterator from the map
  439. map.erase(it);
  440. // as the erasure invalidated out iterator
  441. // we must start at the beginning again
  442. it = map.begin();
  443. }
  444. else {
  445. // if element is active just move over to the next one
  446. it++;
  447. }
  448. }
  449. };
  450. // we require exclusive lock as we modify the cache state
  451. // node == -1 means that cache on all nodes should be flushed
  452. if (node == -1) {
  453. for (auto& nc : cache_state_) {
  454. std::unique_lock<std::shared_mutex> lock(nc.second->cache_mutex_);
  455. FlushNode(nc.second->node_cache_state_);
  456. }
  457. }
  458. else {
  459. std::unique_lock<std::shared_mutex> lock(cache_state_[node]->cache_mutex_);
  460. FlushNode(cache_state_[node]->node_cache_state_);
  461. }
  462. }
  463. inline std::unique_ptr<dsacache::CacheData> dsacache::Cache::GetFromCache(uint8_t* src, const size_t size, const int dst_node) {
  464. // the best situation is if this data is already cached
  465. // which we check in an unnamed block in which the cache
  466. // is locked for reading to prevent another thread
  467. // from marking the element we may find as unused and
  468. // clearing it
  469. LockedNodeCacheState* local_cache_state = cache_state_[dst_node];
  470. // lock the cache state in shared-mode because we read
  471. std::shared_lock<std::shared_mutex> lock(local_cache_state->cache_mutex_);
  472. // search for the data in our cache state structure at the given node
  473. const auto search = local_cache_state->node_cache_state_.find(src);
  474. // if the data is in our structure we continue
  475. if (search != local_cache_state->node_cache_state_.end()) {
  476. // now check whether the sizes match
  477. if (search->second.GetSize() >= size) {
  478. // return a unique copy of the entry which uses the object
  479. // lifetime and destructor to safely handle deallocation
  480. return std::move(std::make_unique<CacheData>(search->second));
  481. }
  482. else {
  483. // if the sizes missmatch then we clear the current entry from cache
  484. // which will cause its deletion only after the last possible outside
  485. // reference is also destroyed
  486. local_cache_state->node_cache_state_.erase(search);
  487. }
  488. }
  489. return nullptr;
  490. }
  491. void dsacache::Cache::Invalidate(uint8_t* data) {
  492. // as the cache is modified we must obtain a unique writers lock
  493. // loop through all per-node-caches available
  494. for (auto node : cache_state_) {
  495. std::unique_lock<std::shared_mutex> lock(node.second->cache_mutex_);
  496. // search for an entry for the given data pointer
  497. auto search = node.second->node_cache_state_.find(data);
  498. if (search != node.second->node_cache_state_.end()) {
  499. // if the data is represented in-cache
  500. // then it will be erased to re-trigger
  501. // caching on next access
  502. node.second->node_cache_state_.erase(search);
  503. }
  504. }
  505. }
  506. inline dsacache::Cache::~Cache() {
  507. for (auto node : cache_state_) {
  508. node.second->~LockedNodeCacheState();
  509. numa_free(reinterpret_cast<void*>(node.second), sizeof(LockedNodeCacheState));
  510. }
  511. }
  512. inline dsacache::CacheData::CacheData(uint8_t* data, const size_t size, MemoryAllocator_Free* free) {
  513. src_ = data;
  514. size_ = size;
  515. delete_ = false;
  516. memory_free_function_ = free;
  517. active_ = new std::atomic<int32_t>(1);
  518. cache_ = new std::atomic<uint8_t*>(data);
  519. handlers_ = new std::atomic<std::vector<dml_handler>*>();
  520. incomplete_cache_ = new uint8_t*(nullptr);
  521. }
  522. inline dsacache::CacheData::CacheData(const dsacache::CacheData& other) {
  523. // we copy the ptr to the global atomic reference counter
  524. // and increase the amount of active references
  525. active_ = other.active_;
  526. const int current_active = active_->fetch_add(1);
  527. src_ = other.src_;
  528. size_ = other.size_;
  529. cache_ = other.cache_;
  530. flags_ = other.flags_;
  531. memory_free_function_ = other.memory_free_function_;
  532. incomplete_cache_ = other.incomplete_cache_;
  533. handlers_ = other.handlers_;
  534. invalid_handlers_ = other.invalid_handlers_;
  535. }
  536. inline dsacache::CacheData::~CacheData() {
  537. // due to fetch_sub returning the preivously held value
  538. // we must subtract one locally to get the current value
  539. const int32_t v = active_->fetch_sub(1) - 1;
  540. // if the returned value is zero or lower
  541. // then we must execute proper deletion
  542. // as this was the last reference
  543. if (v == 0) {
  544. // on deletion we must ensure that all offloaded
  545. // operations have completed successfully
  546. // for this we must unset the possibly active
  547. // flag for weak waiting as we wish completion
  548. // guarantees afterwards
  549. flags_ = UnsetFlag(flags_, FLAG_WAIT_WEAK);
  550. WaitOnCompletion();
  551. // only then can we deallocate the memory
  552. Deallocate();
  553. delete active_;
  554. delete cache_;
  555. delete handlers_;
  556. delete incomplete_cache_;
  557. }
  558. }
  559. inline void dsacache::CacheData::Deallocate() {
  560. // although deallocate should only be called from
  561. // a safe context to do so, it can not hurt to
  562. // defensively perform the operation atomically
  563. // and check for incomplete cache if no deallocation
  564. // takes place for the retrieved local cache
  565. uint8_t* cache_local = cache_->exchange(nullptr);
  566. if (cache_local != nullptr && delete_) memory_free_function_(cache_local, size_);
  567. else if (*incomplete_cache_ != nullptr) memory_free_function_(*incomplete_cache_, size_);
  568. else;
  569. }
  570. inline void dsacache::CacheData::WaitOnCompletion() {
  571. // first check if waiting is even neccessary as a valid
  572. // cache pointer signals that no waiting is to be performed
  573. if (cache_->load() != nullptr) {
  574. return;
  575. }
  576. // then check if the handlers are available
  577. handlers_->wait(nullptr);
  578. // exchange the global handlers pointer with nullptr to have a local
  579. // copy - this signals that this thread is the sole owner and therefore
  580. // responsible for waiting for them. we can not set to nullptr here but
  581. // set to secondary invalid value in order to prevent deadlocks from
  582. // the above waiting construct, where threads may miss the short period
  583. // in which the handlers are not nullptr. we could use double width cas
  584. // which however is more expensive and therefore introduce the second
  585. // invalid state to solve the aba occurring here
  586. // see https://en.wikipedia.org/wiki/ABA_problem for more info
  587. std::vector<dml_handler>* local_handlers = handlers_->exchange(invalid_handlers_);
  588. // ensure that no other thread snatched the handlers before us
  589. // and in case one did, wait again and then return
  590. if (local_handlers == invalid_handlers_) {
  591. cache_->wait(nullptr);
  592. return;
  593. }
  594. // at this point we are responsible for waiting for the handlers
  595. // and handling any error that comes through them gracefully
  596. bool error = false;
  597. for (auto& handler : *local_handlers) {
  598. if (CheckFlag(flags_, FLAG_WAIT_WEAK) && !handler.is_finished()) {
  599. handlers_->store(local_handlers);
  600. return;
  601. }
  602. auto result = handler.get();
  603. if (result.status != dml::status_code::ok) {
  604. std::cerr << "[x] ERROR (" << dml::StatusCodeToString(result.status) << ") FOUND FOR TASK IN WAIT!" << std::endl;
  605. error = true;
  606. }
  607. }
  608. // at this point all handlers have been waited for
  609. // and therefore may be decomissioned
  610. delete local_handlers;
  611. // handle errors now by aborting the cache
  612. if (error) {
  613. cache_->store(src_);
  614. numa_free(*incomplete_cache_, size_);
  615. delete_ = false;
  616. *incomplete_cache_ = nullptr;
  617. }
  618. else {
  619. cache_->store(*incomplete_cache_);
  620. }
  621. // notify all waiting threads so they wake up quickly
  622. cache_->notify_all();
  623. handlers_->notify_all();
  624. }
  625. void dsacache::CacheData::SetTaskHandlersAndCache(uint8_t* cache, std::vector<dml_handler>* handlers) {
  626. *incomplete_cache_ = cache;
  627. handlers_->store(handlers);
  628. handlers_->notify_one();
  629. }
  630. void dsacache::CacheData::Init(std::vector<dml_handler>* invalid_handlers) {
  631. cache_->store(nullptr);
  632. delete_ = true;
  633. invalid_handlers_ = invalid_handlers;
  634. }