/*++ Copyright (c) 1998-2000 Microsoft Corporation Module Name : LKRhash.h Abstract: Declares LKRhash: a fast, scalable, cache- and MP-friendly hash table Author: Paul (Per-Ake) Larson, palarson@microsoft.com, July 1997 Murali R. Krishnan (MuraliK) George V. Reilly (GeorgeRe) 06-Jan-1998 Environment: Win32 - User Mode Project: Internet Information Server RunTime Library Revision History: 10/01/1998 - Change name from LKhash to LKRhash --*/ #undef LKR_OLD_SEGMENT #define LKR_SIGS_NODES #define LKR_TEST_SIGNATURE #undef LKR_COUNTDOWN #define LKR_EQUALKEYS #define LKR_DEPRECATED_ITERATORS #define LKR_BITSCRAMBLE_SUBTABLE_INDEX #define LKR_SUBTABLE_BITMASK #define LKR_SUBTABLE_MASK #ifndef __LKRHASH_H__ #define __LKRHASH_H__ //===================================================================== // The class CLKRLinearHashTable defined in this file provides dynamic hash // tables, i.e. tables that grow and shrink dynamically with // the number of records in the table. // The basic method used is linear hashing, as explained in: // // P.-A. Larson, Dynamic Hash Tables, Comm. of the ACM, 31, 4 (1988) // // This version has the following characteristics: // - It is thread-safe and uses spin locks for synchronization. // - It was designed to support very high rates of concurrent // operations (inserts/deletes/lookups). It achieves this by // (a) partitioning a CLKRHashTable into a collection of // CLKRLinearHashTables to reduce contention on the global table lock. // (b) minimizing the hold time on a table lock, preferring to lock // down a bucket chain instead. // - The design is L1 cache-conscious. See CNodeClump. // - It is designed for sets varying in size from a dozen // elements to several million. // // Main classes: // CLKRLinearHashTable: thread-safe linear hash table // CLKRHashTable: collection of CLKRLinearHashTables // CTypedHashTable: typesafe wrapper for CLKRHashTable // // // Paul Larson, palarson@microsoft.com, July 1997 // Original implementation with input from Murali R. Krishnan, // muralik@microsoft.com. // // George V. Reilly, georgere@microsoft.com, Dec 1997-Jan 1998 // Massive cleanup and rewrite. Added templates. //===================================================================== // 1) Linear Hashing // ------------------ // // Linear hash tables grow and shrink dynamically with the number of // records in the table. The growth or shrinkage is smooth: logically, // one bucket at a time but physically in larger increments // (64 buckets). An insertion (deletion) may cause an expansion // (contraction) of the table. This causes relocation of a small number // of records (at most one bucket worth). All operations (insert, // delete, lookup) take constant expected time, regardless of the // current size or the growth of the table. // // 2) LKR extensions to Linear hash table // -------------------------------------- // // Larson-Krishnan-Reilly extensions to Linear hash tables for multiprocessor // scalability and improved cache performance. // // Traditional implementations of linear hash tables use one global lock // to prevent interference between concurrent operations // (insert/delete/lookup) on the table. The single lock easily becomes // the bottleneck in SMP scenarios when multiple threads are used. // // Traditionally, a (hash) bucket is implemented as a chain of // single-item nodes. Every operation results in chasing down a chain // looking for an item. However, pointer chasing is very slow on modern // systems because almost every jump results in a cache miss. L2 (or L3) // cache misses are very expensive in missed CPU cycles and the cost is // increasing (going to 100s of cycles in the future). // // LKR extensions offer // 1) Partitioning (by hashing) of records among multiple subtables. // Each subtable has locks but there is no global lock. Each // subtable receives a much lower rate of operations, resulting in // fewer conflicts. // // 2) Improved cache locality by grouping keys and their hash values // into contigous chunks that fit exactly into one (or a few) // cache lines. // // Specifically the implementation that exists here achieves this using // the following techniques. // // Class CLKRHashTable is the top-level data structure that dynamically // creates m_cSubTables linear hash tables. The CLKRLinearHashTables act as // the subtables to which items and accesses are fanned out. A good // hash function multiplexes requests uniformly to various subtables, // thus minimizing traffic to any single subtable. The implemenation // uses a home-grown version of bounded spinlocks, that is, a thread // does not spin on a lock indefinitely, instead yielding after a // predetermined number of loops. // // Each CLKRLinearHashTable consists of a CDirEntry pointing to segments // each holding m_dwSegSize CBuckets. Each CBucket in turn consists of a // chain of CNodeClumps. Each CNodeClump contains a group of // NODES_PER_CLUMP hash values (aka hash keys or signatures) and // pointers to the associated data items. Keeping the signatures // together increases the cache locality in scans for lookup. // // Traditionally, people store a link-list element right inside the // object that is hashed and use this link-list for the chaining of data // blocks. However, keeping just the pointers to the data object and // not chaining through them limits the need for bringing in the data // object to the cache. We need to access the data object only if the // hash values match. This limits the cache-thrashing behaviour // exhibited by conventional implementations. It has the additional // benefit that the objects themselves do not need to be modified // in order to be collected in the hash table (i.e., it's non-invasive). //-------------------------------------------------------------------- // TODO // * Debugging support for iisprobe? // * Use auto_ptrs. // * Provide ReplaceRecord and DeleteRecord methods on iterators. // * Sloppy iterators // * Provide implementations of the STL collection classes, map, set, // multimap, and multiset. // * Make exception safe. //-------------------------------------------------------------------- #include #include #include #include #ifdef __LKRHASH_NAMESPACE__ namespace LKRHash { #endif // __LKRHASH_NAMESPACE__ enum LK_TABLESIZE { LK_SMALL_TABLESIZE= 1, // < 200 elements LK_MEDIUM_TABLESIZE= 2, // 200...10,000 elements LK_LARGE_TABLESIZE= 3, // 10,000+ elements }; // Default values for the hashtable constructors enum { #ifndef _WIN64 LK_DFLT_MAXLOAD= 6, // Default upperbound on average chain length. #else // _WIN64 LK_DFLT_MAXLOAD= 4, // 64-byte nodes => NODES_PER_CLUMP = 4 #endif // _WIN64 LK_DFLT_INITSIZE=LK_MEDIUM_TABLESIZE, // Default initial size of hash table LK_DFLT_NUM_SUBTBLS= 0, // Use a heuristic to choose #subtables }; // build fix hack // enum { // DFLT_LK_MAXLOAD= LK_DFLT_MAXLOAD, // DFLT_LK_INITSIZE= LK_DFLT_INITSIZE, // DFLT_LK_NUM_SUBTBLS= LK_DFLT_NUM_SUBTBLS, // }; //-------------------------------------------------------------------- // forward declarations class IRTL_DLLEXP CLKRLinearHashTable; class IRTL_DLLEXP CLKRHashTable; template class CTypedHashTable; //-------------------------------------------------------------------- // Possible return codes from public member functions of // CLKRLinearHashTable, CLKRHashTable, and CTypedHashTable enum LK_RETCODE { // severe errors < 0 LK_UNUSABLE = -99, // Table corrupted: all bets are off LK_ALLOC_FAIL, // ran out of memory LK_BAD_ITERATOR, // invalid iterator; e.g., points to another table LK_BAD_RECORD, // invalid record; e.g., NULL for InsertRecord LK_BAD_PARAMETERS, // invalid parameters; e.g., NULL fnptrs to ctor LK_NOT_INITIALIZED, // LKRHashTableInit was not called LK_SUCCESS = 0, // everything's okay LK_KEY_EXISTS, // key already present for InsertRecord(no-overwrite) LK_NO_SUCH_KEY, // key not found LK_NO_MORE_ELEMENTS,// iterator exhausted }; #define LK_SUCCEEDED(lkrc) ((lkrc) >= LK_SUCCESS) #ifdef LKR_DEPRECATED_ITERATORS //-------------------------------------------------------------------- // Return codes from PFnRecordPred. enum LK_PREDICATE { LKP_ABORT = 1, // Stop walking the table immediately LKP_NO_ACTION = 2, // No action, just keep walking LKP_PERFORM = 3, // Perform action and continue walking LKP_PERFORM_STOP = 4, // Perform action, then stop LKP_DELETE = 5, // Delete record and keep walking LKP_DELETE_STOP = 6, // Delete record, then stop }; //-------------------------------------------------------------------- // Return codes from PFnRecordAction. enum LK_ACTION { LKA_ABORT = 1, // Stop walking the table immediately LKA_FAILED = 2, // Action failed; continue walking the table LKA_SUCCEEDED = 3, // Action succeeded; continue walking the table }; //-------------------------------------------------------------------- // Parameter to Apply and ApplyIf. enum LK_LOCKTYPE { LKL_READLOCK = 1, // Lock the table for reading (for constness) LKL_WRITELOCK = 2, // Lock the table for writing }; #endif // LKR_DEPRECATED_ITERATORS //-------------------------------------------------------------------- // Global table lock code. This is only used to measure how much a // slowdown having a global lock on the CLKRHashTable causes. It is *never* // used in production code. // #define LKRHASH_GLOBAL_LOCK CCritSec #ifdef LKRHASH_GLOBAL_LOCK # define LKRHASH_GLOBAL_LOCK_DECLARATIONS() \ typedef LKRHASH_GLOBAL_LOCK GlobalLock; \ mutable GlobalLock m_lkGlobal; # define LKRHASH_GLOBAL_READ_LOCK() m_lkGlobal.ReadLock() # define LKRHASH_GLOBAL_WRITE_LOCK() m_lkGlobal.WriteLock() # define LKRHASH_GLOBAL_READ_UNLOCK() m_lkGlobal.ReadUnlock() # define LKRHASH_GLOBAL_WRITE_UNLOCK() m_lkGlobal.WriteUnlock() #else // !LKRHASH_GLOBAL_LOCK # define LKRHASH_GLOBAL_LOCK_DECLARATIONS() // These ones will be optimized away by the compiler # define LKRHASH_GLOBAL_READ_LOCK() ((void)0) # define LKRHASH_GLOBAL_WRITE_LOCK() ((void)0) # define LKRHASH_GLOBAL_READ_UNLOCK() ((void)0) # define LKRHASH_GLOBAL_WRITE_UNLOCK() ((void)0) #endif // !LKRHASH_GLOBAL_LOCK //-------------------------------------------------------------------- // Statistical information returned by GetStatistics //-------------------------------------------------------------------- #ifdef LOCK_INSTRUMENTATION class IRTL_DLLEXP CAveragedLockStats : public CLockStatistics { public: int m_nItems; CAveragedLockStats() : m_nItems(1) {} }; #endif // LOCK_INSTRUMENTATION class IRTL_DLLEXP CLKRHashTableStats { public: int RecordCount; // number of records in the table int TableSize; // table size in number of slots int DirectorySize; // number of entries in directory int LongestChain; // longest hash chain in the table int EmptySlots; // number of unused hash slots double SplitFactor; // fraction of buckets split double AvgSearchLength; // average length of a successful search double ExpSearchLength; // theoretically expected length double AvgUSearchLength; // average length of an unsuccessful search double ExpUSearchLength; // theoretically expected length int NodeClumpSize; // number of slots in a node clump int CBucketSize; // sizeof(CBucket) #ifdef LOCK_INSTRUMENTATION CAveragedLockStats m_alsTable; // stats for table lock CAveragedLockStats m_alsBucketsAvg; // avg of stats for bucket locks CGlobalLockStatistics m_gls; // global statistics for all locks #endif // LOCK_INSTRUMENTATION enum { MAX_BUCKETS = 40, }; // histogram of bucket lengths LONG m_aBucketLenHistogram[MAX_BUCKETS]; CLKRHashTableStats() : RecordCount(0), TableSize(0), DirectorySize(0), LongestChain(0), EmptySlots(0), SplitFactor(0.0), AvgSearchLength(0.0), ExpSearchLength(0.0), AvgUSearchLength(0.0), ExpUSearchLength(0.0), NodeClumpSize(1), CBucketSize(0) { for (int i = MAX_BUCKETS; --i >= 0; ) m_aBucketLenHistogram[i] = 0; } static const LONG* BucketSizes() { static const LONG s_aBucketSizes[MAX_BUCKETS] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 30, 40, 50, 60, 70, 80, 90, 100, 200, 500, 1000,10000, 100000, LONG_MAX, }; return s_aBucketSizes; } static LONG BucketSize( LONG nBucketIndex) { IRTLASSERT(0 <= nBucketIndex && nBucketIndex < MAX_BUCKETS); return BucketSizes()[nBucketIndex]; } static LONG BucketIndex( LONG nBucketLength) { const LONG* palBucketSizes = BucketSizes(); LONG i = 0; while (palBucketSizes[i] < nBucketLength) ++i; if (i == MAX_BUCKETS || palBucketSizes[i] > nBucketLength) --i; IRTLASSERT(0 <= i && i < MAX_BUCKETS); return i; } }; // Use types defined in recent versions of the Platform SDK. #ifndef _W64 typedef DWORD DWORD_PTR; // integral type big enough to hold a pointer #endif //-------------------------------------------------------------------- // CLKRLinearHashTable deals with void* records. These typedefs // provide prototypes for functions that manipulate instances of // those records. CTypedHashTable and CStringTestHashTable (below) show a // way to encapsulate these in typesafe wrappers. //-------------------------------------------------------------------- // Given a record, return its key. Assumes that the key is embedded in // the record, or at least somehow derivable from the record. For // completely unrelated keys & values, a wrapper class should use // something like STL's pair template to aggregate them // into a record. typedef const DWORD_PTR (WINAPI *PFnExtractKey) (const void* pvRecord); // Given a key, return its hash signature. The hashing functions in // hashfn.h (or something that builds upon them) are suggested. typedef DWORD (WINAPI *PFnCalcKeyHash) (const DWORD_PTR pnKey); // Compare two keys for equality; e.g., _stricmp, memcmp, operator== typedef bool (WINAPI *PFnEqualKeys) (const DWORD_PTR pnKey1, const DWORD_PTR pnKey2); // Increment the reference count of a record before returning it from // FindKey. It's necessary to do it in FindKey itself while the bucket // is still locked, rather than one of the wrappers, to avoid race // conditions. Similarly, the reference count is incremented in // InsertRecord and decremented in DeleteKey. Finally, if an old record // is overwritten in InsertRecord, its reference count is decremented. // // It's up to you to decrement the reference count when you're finished // with it after retrieving it via FindKey and to determine the // semantics of what this means. The hashtable itself has no notion of // reference counts; this is merely to help with the lifetime management // of the record objects. typedef void (WINAPI *PFnAddRefRecord)(const void* pvRecord, int nIncr); #ifdef LKR_DEPRECATED_ITERATORS // ApplyIf() and DeleteIf(): Does the record match the predicate? typedef LK_PREDICATE (WINAPI *PFnRecordPred) (const void* pvRecord, void* pvState); // Apply() et al: Perform action on record. typedef LK_ACTION (WINAPI *PFnRecordAction)(const void* pvRecord, void* pvState); #endif // LKR_DEPRECATED_ITERATORS //-------------------------------------------------------------------- // Custom memory allocators //-------------------------------------------------------------------- #ifndef LKR_NO_ALLOCATORS // # define LKRHASH_ACACHE 1 // # define LKRHASH_MANODEL 1 // # define LKRHASH_MADEL 1 // # define LKRHASH_ROCKALL_FAST 1 // # define LKRHASH_MEM_DEFAULT_ALIGN 32 #endif // !LKR_NO_ALLOCATORS #ifndef LKRHASH_MEM_DEFAULT_ALIGN # define LKRHASH_MEM_DEFAULT_ALIGN 8 #endif // !LKRHASH_MEM_DEFAULT_ALIGN #if defined(LKRHASH_ACACHE) # include typedef ALLOC_CACHE_HANDLER CLKRhashAllocator; # define LKRHASH_ALLOCATOR_NEW(C, N) \ const ALLOC_CACHE_CONFIGURATION acc = { 1, N, sizeof(C) }; \ C::sm_palloc = new ALLOC_CACHE_HANDLER("IISRTL:" #C, &acc); #elif defined(LKRHASH_MANODEL) # include typedef MEMORY_ALLOC_NO_DELETE CLKRhashAllocator; # define LKRHASH_ALLOCATOR_NEW(C, N) \ C::sm_palloc = new MEMORY_ALLOC_NO_DELETE(sizeof(C), \ LKRHASH_MEM_DEFAULT_ALIGN); #elif defined(LKRHASH_MADEL) # include typedef MEMORY_ALLOC_DELETE CLKRhashAllocator; # define LKRHASH_ALLOCATOR_NEW(C, N) \ C::sm_palloc = new MEMORY_ALLOC_DELETE(sizeof(C), \ LKRHASH_MEM_DEFAULT_ALIGN, N); #elif defined(LKRHASH_ROCKALL_FAST) # include class FastHeap : public FAST_HEAP { public: FastHeap( SIZE_T cb) : m_cb(cb) {} LPVOID Alloc() { return New(m_cb, NULL, false); } BOOL Free(LPVOID pvMem) { return Delete(pvMem); } SIZE_T m_cb; }; typedef FastHeap CLKRhashAllocator; # define LKRHASH_ALLOCATOR_NEW(C, N) \ C::sm_palloc = new FastHeap(sizeof(C)); #else // no custom allocator # undef LKRHASH_ALLOCATOR_NEW #endif // no custom allocator // Used to initialize and destroy custom allocators bool LKRHashTableInit(); void LKRHashTableUninit(); #ifdef LKRHASH_ALLOCATOR_NEW // placed inline in the declaration of class C # define LKRHASH_ALLOCATOR_DEFINITIONS(C) \ protected: \ static CLKRhashAllocator* sm_palloc; \ friend bool LKRHashTableInit(); \ friend void LKRHashTableUninit(); \ public: \ static void* operator new(size_t s) \ { \ IRTLASSERT(s == sizeof(C)); \ IRTLASSERT(sm_palloc != NULL); \ return sm_palloc->Alloc(); \ } \ static void operator delete(void* pv) \ { \ IRTLASSERT(pv != NULL); \ IRTLASSERT(sm_palloc != NULL); \ sm_palloc->Free(pv); \ } // used in LKRHashTableInit() # define LKRHASH_ALLOCATOR_INIT(C, N, f) \ { \ if (f) \ { \ IRTLASSERT(C::sm_palloc == NULL); \ LKRHASH_ALLOCATOR_NEW(C, N); \ f = (C::sm_palloc != NULL); \ } \ } // used in LKRHashTableUninit() # define LKRHASH_ALLOCATOR_UNINIT(C) \ { \ if (C::sm_palloc != NULL) \ { \ delete C::sm_palloc; \ C::sm_palloc = NULL; \ } \ } #else // !LKRHASH_ALLOCATOR_NEW # define LKRHASH_ALLOCATOR_DEFINITIONS(C) # define LKRHASH_ALLOCATOR_INIT(C, N, f) # define LKRHASH_ALLOCATOR_UNINIT(C) #endif // !LKRHASH_ALLOCATOR_NEW //-------------------------------------------------------------------- // CLKRLinearHashTable // // A thread-safe linear hash table. //-------------------------------------------------------------------- class IRTL_DLLEXP CLKRLinearHashTable { public: // typedef CSmallSpinLock TableLock; // typedef CFakeLock TableLock; // typedef CSpinLock TableLock; // typedef CSpinLock2 TableLock; // typedef CSpinLock3 TableLock; // typedef CShareLock TableLock; // typedef CReaderWriterLock TableLock; // typedef CReaderWriterLock2 TableLock; typedef CReaderWriterLock3 TableLock; // typedef CSmallSpinLock BucketLock; // typedef CFakeLock BucketLock; // typedef CSpinLock BucketLock; // typedef CSpinLock2 BucketLock; // typedef CSpinLock3 BucketLock; // typedef CShareLock BucketLock; // typedef CReaderWriterLock BucketLock; // typedef CReaderWriterLock2 BucketLock; typedef CReaderWriterLock3 BucketLock; #ifdef LKR_DEPRECATED_ITERATORS class CIterator; friend class CLKRLinearHashTable::CIterator; #endif // LKR_DEPRECATED_ITERATORS private: class CNodeClump; friend class CLKRLinearHashTable::CNodeClump; friend class CLKRHashTable; #ifdef LKRHASH_ALLOCATOR_NEW friend bool LKRHashTableInit(); friend void LKRHashTableUninit(); #endif // LKRHASH_ALLOCATOR_NEW #ifdef LKRHASH_INSTRUMENTATION // TODO #endif // LKRHASH_INSTRUMENTATION #define HASH_RANDOMIZE_BITS #ifdef HASH_RANDOMIZE_BITS enum { // No number in 0..2^31-1 maps to this number after it has been // scrambled by RandomizeBits HASH_INVALID_SIGNATURE = 31678523, }; #else // !HASH_RANDOMIZE_BITS enum { // Given M = A % B, A and B unsigned 32-bit integers greater than zero, // there are no values of A or B which yield M = 2^32-1. Why? Because // M must be less than B. HASH_INVALID_SIGNATURE = ULONG_MAX, }; #endif // !HASH_RANDOMIZE_BITS // Class for nodes on a bucket chain. Instead of a node containing // one (signature, record-pointer, next-tuple-pointer) tuple, it // contains _N_ such tuples. (N-1 next-tuple-pointers are omitted.) // This improves locality of reference greatly; i.e., it's L1 // cache-friendly. It also reduces memory fragmentation and memory // allocator overhead. It does complicate the chain traversal code // slightly, admittedly. // // This theory is beautiful. In practice, however, CNodeClumps // are *not* perfectly aligned on 32-byte boundaries by the memory // allocators. Experimental results indicate that we get a 2-3% // speed improvement by using 32-byte-aligned blocks, but this must // be considered against the average of 16 bytes wasted per block. class CNodeClump { public: // Record slots per chunk - set so a chunk matches (one or // two) cache lines. 3 ==> 32 bytes, 7 ==> 64 bytes // Note: the default max load factor is 4.0, which implies that // there will seldom be more than one node clump in a chain. enum { BUCKET_BYTE_SIZE = 64, BUCKET_OVERHEAD = sizeof(BucketLock) + sizeof(CNodeClump*), NODE_SIZE = sizeof(const void*) + sizeof(DWORD), NODES_PER_CLUMP = (BUCKET_BYTE_SIZE - BUCKET_OVERHEAD) / NODE_SIZE }; #ifdef LKR_SIGS_NODES DWORD m_dwKeySigs[NODES_PER_CLUMP]; // hash values computed from keys CNodeClump* m_pncNext; // next node clump on the chain const void* m_pvNode[NODES_PER_CLUMP];// pointers to records #else // !LKR_SIGS_NODES const void* m_pvNode[NODES_PER_CLUMP];// pointers to records CNodeClump* m_pncNext; // next node clump on the chain DWORD m_dwKeySigs[NODES_PER_CLUMP]; // hash values computed from keys #endif // !LKR_SIGS_NODES CNodeClump() { Clear(); } void Clear() { m_pncNext = NULL; // no dangling pointers for (int i = NODES_PER_CLUMP; --i >= 0; ) { m_dwKeySigs[i] = HASH_INVALID_SIGNATURE; m_pvNode[i] = NULL; } } bool InvalidSignature( DWORD i) const { IRTLASSERT(0 <= i && i < NODES_PER_CLUMP); return (m_dwKeySigs[i] == HASH_INVALID_SIGNATURE); } bool IsEmptyNode( DWORD i) const { IRTLASSERT(0 <= i && i < NODES_PER_CLUMP); return (m_pvNode[i] == NULL); } bool IsEmptyAndInvalid( DWORD i) const { return IsEmptyNode(i) && InvalidSignature(i); } bool IsEmptySlot( DWORD i) const { #ifdef LKR_TEST_SIGNATURE return InvalidSignature(i); #else // !LKR_TEST_SIGNATURE return IsEmptyNode(i); #endif // !LKR_TEST_SIGNATURE } bool IsLastClump() const { return (m_pncNext == NULL); } #ifdef _DEBUG // Don't want overhead of calls to dtor in retail build ~CNodeClump() { IRTLASSERT(IsLastClump()); // no dangling pointers for (int i = NODES_PER_CLUMP; --i >= 0; ) IRTLASSERT(InvalidSignature(i) && IsEmptyNode(i)); } #endif // _DEBUG LKRHASH_ALLOCATOR_DEFINITIONS(CNodeClump); }; // Class for bucket chains of the hash table. Note that the first // nodeclump is actually included in the bucket and not dynamically // allocated, which increases space requirements slightly but does // improve performance. class CBucket { private: mutable BucketLock m_Lock; // lock protecting this bucket #ifdef LOCK_INSTRUMENTATION static LONG sm_cBuckets; static const char* _LockName() { LONG l = ++sm_cBuckets; // possible race condition but we don't care, as this is never // used in production code static char s_szName[CLockStatistics::L_NAMELEN]; wsprintf(s_szName, "B%06x", 0xFFFFFF & l); return s_szName; } #endif // LOCK_INSTRUMENTATION public: CNodeClump m_ncFirst; // first CNodeClump of this bucket #if defined(LOCK_INSTRUMENTATION) || defined(_DEBUG) CBucket() #ifdef LOCK_INSTRUMENTATION : m_Lock(_LockName()) #endif // LOCK_INSTRUMENTATION { #ifdef _DEBUG LOCK_LOCKTYPE lt = BucketLock::LockType(); if (lt == LOCK_SPINLOCK || lt == LOCK_FAKELOCK) IRTLASSERT(sizeof(*this) <= 64); #endif _DEBUG } #endif // LOCK_INSTRUMENTATION || _DEBUG void WriteLock() { m_Lock.WriteLock(); } void ReadLock() const { m_Lock.ReadLock(); } void WriteUnlock() const { m_Lock.WriteUnlock(); } void ReadUnlock() const { m_Lock.ReadUnlock(); } bool IsWriteLocked() const { return m_Lock.IsWriteLocked(); } bool IsReadLocked() const { return m_Lock.IsReadLocked(); } bool IsWriteUnlocked() const { return m_Lock.IsWriteUnlocked(); } bool IsReadUnlocked() const { return m_Lock.IsReadUnlocked(); } void SetSpinCount(WORD wSpins) { m_Lock.SetSpinCount(wSpins); } WORD GetSpinCount() const { return m_Lock.GetSpinCount(); } #ifdef LOCK_INSTRUMENTATION CLockStatistics LockStats() const {return m_Lock.Statistics();} #endif // LOCK_INSTRUMENTATION }; // The hash table space is divided into fixed-size segments (arrays of // CBuckets) and physically grows/shrinks one segment at a time. // We provide small, medium, and large segments to better tune the // overall memory requirements of the hash table according to the // expected usage of an instance. class CSegment { public: #ifdef LKR_OLD_SEGMENT virtual ~CSegment() {}; // link fails if this is pure virtual virtual DWORD Bits() const = 0; virtual DWORD Size() const = 0; virtual DWORD Mask() const = 0; virtual DWORD InitSize() const = 0; virtual CBucket& Slot(DWORD i) = 0; #else // !LKR_OLD_SEGMENT CBucket m_bktSlots; CBucket& Slot(DWORD i) { return static_cast(&m_bktSlots)[i]; } #endif // !LKR_OLD_SEGMENT }; // Small-sized segments contain 2^3 = 8 buckets => ~0.5Kb class CSmallSegment : public CSegment { public: // Maximum table size equals MAX_DIRSIZE * SEGSIZE buckets. enum { SEGBITS = 3,// number of bits extracted from a hash // address for offset within a segment SEGSIZE = (1< ~4Kb class CMediumSegment : public CSegment { public: enum { SEGBITS = 6, SEGSIZE = (1< ~32Kb class CLargeSegment : public CSegment { public: enum { SEGBITS = 9, SEGSIZE = (1<> m_dwSegBits; IRTLASSERT(m_paDirSegs != NULL && iSeg < m_cDirSegs); return m_paDirSegs[iSeg].m_pseg; } // Offset within the segment of the bucketaddress DWORD _SegIndex(DWORD dwBucketAddr) const { return (dwBucketAddr & m_dwSegMask); } // Convert a bucketaddress to a CBucket* inline CBucket* _Bucket(DWORD dwBucketAddr) const { IRTLASSERT(dwBucketAddr < m_cActiveBuckets); CSegment* const pseg = _Segment(dwBucketAddr); IRTLASSERT(pseg != NULL); return &(pseg->Slot(_SegIndex(dwBucketAddr))); } // Extract the key from a record const DWORD_PTR _ExtractKey(const void* pvRecord) const { IRTLASSERT(pvRecord != NULL); IRTLASSERT(m_pfnExtractKey != NULL); return (*m_pfnExtractKey)(pvRecord); } // Hash the key DWORD _CalcKeyHash(const DWORD_PTR pnKey) const { // Note pnKey==0 is acceptable, as the real key type could be an int IRTLASSERT(m_pfnCalcKeyHash != NULL); DWORD dwHash = (*m_pfnCalcKeyHash)(pnKey); // We forcibly scramble the result to help ensure a better distribution #ifdef HASH_RANDOMIZE_BITS dwHash = RandomizeBits(dwHash); #else // !HASH_RANDOMIZE_BITS dwHash = HashScramble(dwHash); #endif // !HASH_RANDOMIZE_BITS IRTLASSERT(dwHash != HASH_INVALID_SIGNATURE); return dwHash; } // Compare two keys for equality bool _EqualKeys(const DWORD_PTR pnKey1, const DWORD_PTR pnKey2) const { IRTLASSERT(m_pfnEqualKeys != NULL); return (*m_pfnEqualKeys)(pnKey1, pnKey2); } // AddRef or Release a record. void _AddRefRecord(const void* pvRecord, int nIncr) const { IRTLASSERT(pvRecord != NULL && (nIncr == -1 || nIncr == +1)); IRTLASSERT(m_pfnAddRefRecord != NULL); (*m_pfnAddRefRecord)(pvRecord, nIncr); } // Find a bucket, given its signature. CBucket* _FindBucket(DWORD dwSignature, bool fLockForWrite) const; // Used by _FindKey so that the thread won't deadlock if the user has // already explicitly called table->WriteLock(). bool _ReadOrWriteLock() const { return m_Lock.ReadOrWriteLock(); } void _ReadOrWriteUnlock(bool fReadLocked) const { m_Lock.ReadOrWriteUnlock(fReadLocked); } // Memory allocation wrappers to allow us to simulate allocation // failures during testing static CDirEntry* const _AllocateSegmentDirectory( size_t n); static bool _FreeSegmentDirectory( CDirEntry* paDirSegs); static CNodeClump* const _AllocateNodeClump(); static bool _FreeNodeClump( CNodeClump* pnc); CSegment* const _AllocateSegment() const; bool _FreeSegment( CSegment* pseg) const; #ifdef LOCK_INSTRUMENTATION static LONG sm_cTables; static const char* _LockName() { LONG l = ++sm_cTables; // possible race condition but we don't care, as this is never // used in production code static char s_szName[CLockStatistics::L_NAMELEN]; wsprintf(s_szName, "LH%05x", 0xFFFFF & l); return s_szName; } // Statistics for the table lock CLockStatistics _LockStats() const { return m_Lock.Statistics(); } #endif // LOCK_INSTRUMENTATION private: // Fields are ordered so as to minimize number of cache lines touched DWORD m_dwSignature; // debugging: id & corruption check CHAR m_szName[NAME_SIZE]; // an identifier for debugging mutable LK_RETCODE m_lkrcState; // Internal state of table mutable TableLock m_Lock; // Lock on entire linear hash table // type-specific function pointers PFnExtractKey m_pfnExtractKey; // Extract key from record PFnCalcKeyHash m_pfnCalcKeyHash; // Calculate hash signature of key PFnEqualKeys m_pfnEqualKeys; // Compare two keys PFnAddRefRecord m_pfnAddRefRecord; // AddRef a record LK_TABLESIZE m_lkts; // "size" of table: small, medium, or large DWORD m_dwSegBits; // C{Small,Medium,Large}Segment::SEGBITS DWORD m_dwSegSize; // C{Small,Medium,Large}Segment::SEGSIZE DWORD m_dwSegMask; // C{Small,Medium,Large}Segment::SEGMASK double m_MaxLoad; // max load factor (average chain length) DWORD m_dwBktAddrMask0; // mask used for address calculation DWORD m_dwBktAddrMask1; // used in _H1 calculation DWORD m_iExpansionIdx; // address of next bucket to be expanded CDirEntry* m_paDirSegs; // directory of table segments DWORD m_nLevel; // number of table doublings performed DWORD m_cDirSegs; // segment directory size: varies between // MIN_DIRSIZE and MAX_DIRSIZE DWORD m_cRecords; // number of records in the table DWORD m_cActiveBuckets; // number of buckets in use (table size) WORD m_wBucketLockSpins;// default spin count for bucket locks const BYTE m_nTableLockType; // for debugging: LOCK_SPINLOCK, etc const BYTE m_nBucketLockType;// for debugging: LOCK_SPINLOCK, etc const CLKRHashTable* const m_phtParent;// Owning table. NULL => standalone #ifndef LKR_NO_GLOBAL_LIST static CLockedDoubleList sm_llGlobalList;// All active CLKRLinearHashTables CListEntry m_leGlobalList; #endif // !LKR_NO_GLOBAL_LIST void _InsertThisIntoGlobalList() { #ifndef LKR_NO_GLOBAL_LIST // Only add standalone CLKRLinearHashTables to global list. // CLKRHashTables have their own global list. if (m_phtParent == NULL) sm_llGlobalList.InsertHead(&m_leGlobalList); #endif // !LKR_NO_GLOBAL_LIST } void _RemoveThisFromGlobalList() { #ifndef LKR_NO_GLOBAL_LIST if (m_phtParent == NULL) sm_llGlobalList.RemoveEntry(&m_leGlobalList); #endif // !LKR_NO_GLOBAL_LIST } // Non-trivial implementation functions LK_RETCODE _InsertRecord(const void* pvRecord, DWORD dwSignature, bool fOverwrite); LK_RETCODE _DeleteKey(const DWORD_PTR pnKey, DWORD dwSignature); LK_RETCODE _DeleteRecord(const void* pvRecord, DWORD dwSignature); bool _DeleteNode(CBucket* pbkt, CNodeClump*& rpnc, CNodeClump*& rpncPrev, int& riNode); LK_RETCODE _FindKey(const DWORD_PTR pnKey, DWORD dwSignature, const void** ppvRecord) const; LK_RETCODE _FindRecord(const void* pvRecord, DWORD dwSignature) const; // returns count of errors in compacted state => 0 is good int _IsNodeCompact(CBucket* const pbkt) const; #ifdef LKR_DEPRECATED_ITERATORS // Predicate functions static LK_PREDICATE WINAPI _PredTrue(const void* /*pvRecord*/, void* /*pvState*/) { return LKP_PERFORM; } DWORD _Apply(PFnRecordAction pfnAction, void* pvState, LK_LOCKTYPE lkl, LK_PREDICATE& rlkp); DWORD _ApplyIf(PFnRecordPred pfnPredicate, PFnRecordAction pfnAction, void* pvState, LK_LOCKTYPE lkl, LK_PREDICATE& rlkp); DWORD _DeleteIf(PFnRecordPred pfnPredicate, void* pvState, LK_PREDICATE& rlkp); #endif // LKR_DEPRECATED_ITERATORS void _Clear(bool fShrinkDirectory); void _SetSegVars(LK_TABLESIZE lkts); LK_RETCODE _Expand(); LK_RETCODE _Contract(); LK_RETCODE _SplitRecordSet(CNodeClump* pncOldTarget, CNodeClump* pncNewTarget, DWORD iExpansionIdx, DWORD dwBktAddrMask, DWORD dwNewBkt, CNodeClump* pncFreeList); LK_RETCODE _MergeRecordSets(CBucket* pbktNewTarget, CNodeClump* pncOldList, CNodeClump* pncFreeList); // Private copy ctor and op= to prevent compiler synthesizing them. // Must provide a (bad) implementation because we export instantiations. // TODO: implement these properly; they could be useful. CLKRLinearHashTable(const CLKRLinearHashTable&) : m_dwSignature(SIGNATURE_FREE) #ifdef LOCK_INSTRUMENTATION , m_Lock(NULL) #endif // LOCK_INSTRUMENTATION , m_nTableLockType(0), m_nBucketLockType(0), m_phtParent(NULL) {*(BYTE*)NULL;} CLKRLinearHashTable& operator=(const CLKRLinearHashTable&) {return *(CLKRLinearHashTable*)NULL;} private: // This ctor is used by CLKRHashTable CLKRLinearHashTable( LPCSTR pszName, // An identifier for debugging PFnExtractKey pfnExtractKey, // Extract key from record PFnCalcKeyHash pfnCalcKeyHash, // Calculate hash signature of key PFnEqualKeys pfnEqualKeys, // Compare two keys PFnAddRefRecord pfnAddRefRecord,// AddRef in FindKey, etc double maxload, // Upperbound on average chain length DWORD initsize, // Initial size of hash table. CLKRHashTable* phtParent // Owning table. ); LK_RETCODE _Initialize( PFnExtractKey pfnExtractKey, PFnCalcKeyHash pfnCalcKeyHash, PFnEqualKeys pfnEqualKeys, PFnAddRefRecord pfnAddRefRecord, LPCSTR pszName, double maxload, DWORD initsize); public: CLKRLinearHashTable( LPCSTR pszName, // An identifier for debugging PFnExtractKey pfnExtractKey, // Extract key from record PFnCalcKeyHash pfnCalcKeyHash, // Calculate hash signature of key PFnEqualKeys pfnEqualKeys, // Compare two keys PFnAddRefRecord pfnAddRefRecord,// AddRef in FindKey, etc double maxload=LK_DFLT_MAXLOAD,// Upperbound on average chain length DWORD initsize=LK_DFLT_INITSIZE, // Initial size of hash table. DWORD num_subtbls=LK_DFLT_NUM_SUBTBLS // for signature compatiblity // with CLKRHashTable ); ~CLKRLinearHashTable(); static const char* ClassName() {return "CLKRLinearHashTable";} int NumSubTables() const {return 1;} static LK_TABLESIZE NumSubTables(DWORD& rinitsize, DWORD& rnum_subtbls); // Insert a new record into hash table. // Returns LK_SUCCESS if all OK, LK_KEY_EXISTS if same key already // exists (unless fOverwrite), LK_ALLOC_FAIL if out of space, // or LK_BAD_RECORD for a bad record. LK_RETCODE InsertRecord(const void* pvRecord, bool fOverwrite=false) { if (!IsUsable()) return m_lkrcState; if (pvRecord == NULL) return LK_BAD_RECORD; return _InsertRecord(pvRecord, _CalcKeyHash(_ExtractKey(pvRecord)), fOverwrite); } // Delete record with the given key. // Returns LK_SUCCESS if all OK, or LK_NO_SUCH_KEY if not found LK_RETCODE DeleteKey(const DWORD_PTR pnKey) { if (!IsUsable()) return m_lkrcState; return _DeleteKey(pnKey, _CalcKeyHash(pnKey)); } // Delete a record from the table, if present. // Returns LK_SUCCESS if all OK, or LK_NO_SUCH_KEY if not found LK_RETCODE DeleteRecord(const void* pvRecord) { if (!IsUsable()) return m_lkrcState; if (pvRecord == NULL) return LK_BAD_RECORD; return _DeleteRecord(pvRecord, _CalcKeyHash(_ExtractKey(pvRecord))); } // Find record with given key. // Returns: LK_SUCCESS, if record found (record is returned in *ppvRecord) // LK_BAD_RECORD, if ppvRecord is invalid // LK_NO_SUCH_KEY, if no record with given key value was found // LK_UNUSABLE, if hash table not in usable state // Note: the record is AddRef'd. You must decrement the reference // count when you are finished with the record (if you're implementing // refcounting semantics). LK_RETCODE FindKey(const DWORD_PTR pnKey, const void** ppvRecord) const { if (!IsUsable()) return m_lkrcState; if (ppvRecord == NULL) return LK_BAD_RECORD; return _FindKey(pnKey, _CalcKeyHash(pnKey), ppvRecord); } // Sees if the record is contained in the table // Returns: LK_SUCCESS, if record found // LK_BAD_RECORD, if pvRecord is invalid // LK_NO_SUCH_KEY, if record is not in the table // LK_UNUSABLE, if hash table not in usable state // Note: the record is *not* AddRef'd. LK_RETCODE FindRecord(const void* pvRecord) const { if (!IsUsable()) return m_lkrcState; if (pvRecord == NULL) return LK_BAD_RECORD; return _FindRecord(pvRecord, _CalcKeyHash(_ExtractKey(pvRecord))); } #ifdef LKR_DEPRECATED_ITERATORS // Walk the hash table, applying pfnAction to all records. // Locks the whole table for the duration with either a (possibly // shared) readlock or a writelock, according to lkl. // Loop is aborted if pfnAction returns LKA_ABORT. // Returns the number of successful applications. DWORD Apply(PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK); // Walk the hash table, applying pfnAction to any records that match // pfnPredicate. Locks the whole table for the duration with either // a (possibly shared) readlock or a writelock, according to lkl. // Loop is aborted if pfnAction returns LKA_ABORT. // Returns the number of successful applications. DWORD ApplyIf(PFnRecordPred pfnPredicate, PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK); // Delete any records that match pfnPredicate. // Locks the table for the duration with a writelock. // Returns the number of deletions. // // Do *not* walk the hash table by hand with an iterator and call // DeleteKey. The iterator will end up pointing to garbage. DWORD DeleteIf(PFnRecordPred pfnPredicate, void* pvState=NULL); #endif // LKR_DEPRECATED_ITERATORS // Check table for consistency. Returns 0 if okay, or the number of // errors otherwise. int CheckTable() const; // Remove all data from the table void Clear() { WriteLock(); _Clear(true); WriteUnlock(); } // Number of elements in the table DWORD Size() const { return m_cRecords; } // Maximum possible number of elements in the table DWORD MaxSize() const { return static_cast(m_MaxLoad * MAX_DIRSIZE * m_dwSegSize); } // Get hash table statistics CLKRHashTableStats GetStatistics() const; // Is the hash table usable? bool IsUsable() const { return (m_lkrcState == LK_SUCCESS); } // Is the hash table consistent and correct? bool IsValid() const { bool f = (m_lkrcState == LK_SUCCESS // serious internal failure? && m_paDirSegs != NULL && (MIN_DIRSIZE & (MIN_DIRSIZE-1)) == 0 // == (1 << N) && (MAX_DIRSIZE & (MAX_DIRSIZE-1)) == 0 && MAX_DIRSIZE > MIN_DIRSIZE && MIN_DIRSIZE <= m_cDirSegs && m_cDirSegs <= MAX_DIRSIZE && (m_cDirSegs & (m_cDirSegs-1)) == 0 && m_pfnExtractKey != NULL && m_pfnCalcKeyHash != NULL && m_pfnEqualKeys != NULL && m_cActiveBuckets > 0 && ValidSignature() ); if (!f) m_lkrcState = LK_UNUSABLE; return f; } // Set the spin count on the table lock void SetTableLockSpinCount(WORD wSpins) { m_Lock.SetSpinCount(wSpins); } // Get the spin count on the table lock WORD GetTableLockSpinCount() { return m_Lock.GetSpinCount(); } // Set/Get the spin count on the bucket locks void SetBucketLockSpinCount(WORD wSpins); WORD GetBucketLockSpinCount(); enum { SIGNATURE = (('L') | ('K' << 8) | ('L' << 16) | ('H' << 24)), SIGNATURE_FREE = (('L') | ('K' << 8) | ('L' << 16) | ('x' << 24)), }; bool ValidSignature() const { return m_dwSignature == SIGNATURE;} // Lock manipulators // Lock the table (exclusively) for writing void WriteLock() { m_Lock.WriteLock(); } // Lock the table (possibly shared) for reading void ReadLock() const { m_Lock.ReadLock(); } // Unlock the table for writing void WriteUnlock() const { m_Lock.WriteUnlock(); } // Unlock the table for reading void ReadUnlock() const { m_Lock.ReadUnlock(); } // Is the table already locked for writing? bool IsWriteLocked() const { return m_Lock.IsWriteLocked(); } // Is the table already locked for reading? bool IsReadLocked() const { return m_Lock.IsReadLocked(); } // Is the table unlocked for writing? bool IsWriteUnlocked() const { return m_Lock.IsWriteUnlocked(); } // Is the table unlocked for reading? bool IsReadUnlocked() const { return m_Lock.IsReadUnlocked(); } // Convert the read lock to a write lock void ConvertSharedToExclusive() const { m_Lock.ConvertSharedToExclusive(); } // Convert the write lock to a read lock void ConvertExclusiveToShared() const { m_Lock.ConvertExclusiveToShared(); } // LKRHASH_ALLOCATOR_DEFINITIONS(CLKRLinearHashTable); public: #ifdef LKR_DEPRECATED_ITERATORS // Iterators can be used to walk the table. To ensure a consistent // view of the data, the iterator locks the whole table. This can // have a negative effect upon performance, because no other thread // can do anything with the table. Use with care. // // You should not use an iterator to walk the table, calling DeleteKey, // as the iterator will end up pointing to garbage. // // Use Apply, ApplyIf, or DeleteIf instead of iterators to safely // walk the tree. // // Note that iterators acquire a reference to the record pointed to // and release that reference as soon as the iterator is incremented. // In other words, this code is safe: // lkrc = ht.IncrementIterator(&iter); // // assume lkrc == LK_SUCCESS for the sake of this example // CMyHashTable::Record* pRec = iter.Record(); // Foo(pRec); // uses pRec but doesn't hang on to it // lkrc = ht.IncrementIterator(&iter); // // But this code is not safe because pRec is used out of the scope of the // iterator that provided it: // lkrc = ht.IncrementIterator(&iter); // CMyHashTable::Record* pRec = iter.Record(); // // Broken code: Should have called ht.AddRefRecord(pRec, +1) here // lkrc = ht.IncrementIterator(&iter); // Foo(pRec); // Unsafe: because no longer have a valid reference // // If record has no reference-counting semantics, then you can ignore // the above remarks about scope. class CIterator { protected: friend class CLKRLinearHashTable; CLKRLinearHashTable* m_plht; // which linear hash table? DWORD m_dwBucketAddr; // bucket index CNodeClump* m_pnc; // a CNodeClump in bucket int m_iNode; // offset within m_pnc LK_LOCKTYPE m_lkl; // readlock or writelock? private: // Private copy ctor and op= to prevent compiler synthesizing them. // Must provide (bad) implementation because we export instantiations. CIterator(const CIterator&) {*(BYTE*)NULL;} CIterator& operator=(const CIterator&) {return *(CIterator*)NULL;} public: CIterator( LK_LOCKTYPE lkl=LKL_WRITELOCK) : m_plht(NULL), m_dwBucketAddr(0), m_pnc(NULL), m_iNode(-1), m_lkl(lkl) {} // Return the record associated with this iterator const void* Record() const { IRTLASSERT(IsValid()); return ((m_pnc != NULL && m_iNode >= 0 && m_iNode < CLKRLinearHashTable::NODES_PER_CLUMP) ? m_pnc->m_pvNode[m_iNode] : NULL); } // Return the key associated with this iterator const DWORD_PTR Key() const { IRTLASSERT(m_plht != NULL); const void* pRec = Record(); return ((pRec != NULL && m_plht != NULL) ? m_plht->_ExtractKey(pRec) : NULL); } bool IsValid() const { return ((m_plht != NULL) && (m_pnc != NULL) && (0 <= m_iNode && m_iNode < CLKRLinearHashTable::NODES_PER_CLUMP) && (!m_pnc->IsEmptyNode(m_iNode))); } // Delete the record that the iterator points to. Does an implicit // IncrementIterator after deletion. LK_RETCODE DeleteRecord(); // Change the record that the iterator points to. The new record // must have the same key as the old one. LK_RETCODE ChangeRecord(const void* pNewRec); }; // Const iterators for readonly access. You must use these with // const CLKRLinearHashTables. class CConstIterator : public CIterator { private: // Private, unimplemented copy ctor and op= to prevent // compiler synthesizing them. CConstIterator(const CConstIterator&); CConstIterator& operator=(const CConstIterator&); public: CConstIterator() : CIterator(LKL_READLOCK) {} }; private: // The public APIs lock the table. The private ones, which are used // directly by CLKRHashTable, don't. LK_RETCODE _InitializeIterator(CIterator* piter); LK_RETCODE _CloseIterator(CIterator* piter); public: // Initialize the iterator to point to the first item in the hash table // Returns LK_SUCCESS, LK_NO_MORE_ELEMENTS, or LK_BAD_ITERATOR. LK_RETCODE InitializeIterator(CIterator* piter) { IRTLASSERT(piter != NULL && piter->m_plht == NULL); if (piter == NULL || piter->m_plht != NULL) return LK_BAD_ITERATOR; if (piter->m_lkl == LKL_WRITELOCK) WriteLock(); else ReadLock(); return _InitializeIterator(piter); } // The const iterator version LK_RETCODE InitializeIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_plht == NULL); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_plht != NULL || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; ReadLock(); return const_cast(this) ->_InitializeIterator(static_cast(piter)); } // Move the iterator on to the next item in the table. // Returns LK_SUCCESS, LK_NO_MORE_ELEMENTS, or LK_BAD_ITERATOR. LK_RETCODE IncrementIterator(CIterator* piter); LK_RETCODE IncrementIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_plht == this); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_plht != this || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; return const_cast(this) ->IncrementIterator(static_cast(piter)); } // Close the iterator. LK_RETCODE CloseIterator(CIterator* piter) { IRTLASSERT(piter != NULL && piter->m_plht == this); if (piter == NULL || piter->m_plht != this) return LK_BAD_ITERATOR; _CloseIterator(piter); if (piter->m_lkl == LKL_WRITELOCK) WriteUnlock(); else ReadUnlock(); return LK_SUCCESS; }; // Close the CConstIterator LK_RETCODE CloseIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_plht == this); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_plht != this || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; const_cast(this) ->_CloseIterator(static_cast(piter)); ReadUnlock(); return LK_SUCCESS; }; #endif // LKR_DEPRECATED_ITERATORS }; //-------------------------------------------------------------------- // CLKRHashTable // // To improve concurrency, a hash table is divided into a number of // (independent) subtables. Each subtable is a linear hash table. The // number of subtables is defined when the table is created and remains // fixed thereafter. Records are assigned to subtables based on their // hashed key. // // For small or low-contention hashtables, you can bypass this // thin wrapper and use CLKRLinearHashTable directly. The methods are // documented in the declarations for CLKRHashTable (above). //-------------------------------------------------------------------- class IRTL_DLLEXP CLKRHashTable { private: typedef CLKRLinearHashTable SubTable; public: typedef SubTable::TableLock TableLock; typedef SubTable::BucketLock BucketLock; #ifdef LKR_DEPRECATED_ITERATORS class CIterator; friend class CLKRHashTable::CIterator; #endif // LKR_DEPRECATED_ITERATORS // aliases for convenience enum { NAME_SIZE = SubTable::NAME_SIZE, HASH_INVALID_SIGNATURE = SubTable::HASH_INVALID_SIGNATURE, NODES_PER_CLUMP = SubTable::NODES_PER_CLUMP, }; private: // Hash table parameters DWORD m_dwSignature; // debugging: id & corruption check CHAR m_szName[NAME_SIZE]; // an identifier for debugging DWORD m_cSubTables; // number of subtables SubTable** m_palhtDir; // array of subtables // type-specific function pointers PFnExtractKey m_pfnExtractKey; PFnCalcKeyHash m_pfnCalcKeyHash; mutable LK_RETCODE m_lkrcState; // Internal state of table #ifdef LKR_SUBTABLE_MASK int m_nSubTableMask; #endif // LKR_SUBTABLE_MASK #ifndef LKR_NO_GLOBAL_LIST static CLockedDoubleList sm_llGlobalList; // All active CLKRHashTables CListEntry m_leGlobalList; #endif // !LKR_NO_GLOBAL_LIST void _InsertThisIntoGlobalList() { #ifndef LKR_NO_GLOBAL_LIST sm_llGlobalList.InsertHead(&m_leGlobalList); #endif // !LKR_NO_GLOBAL_LIST } void _RemoveThisFromGlobalList() { #ifndef LKR_NO_GLOBAL_LIST sm_llGlobalList.RemoveEntry(&m_leGlobalList); #endif // !LKR_NO_GLOBAL_LIST } LKRHASH_GLOBAL_LOCK_DECLARATIONS(); // Private copy ctor and op= to prevent compiler synthesizing them. // Must provide a (bad) implementation because we export instantiations. // TODO: implement these properly; they could be useful. CLKRHashTable(const CLKRHashTable&) {*(BYTE*)NULL;} CLKRHashTable& operator=(const CLKRHashTable&) {return *(CLKRHashTable*)NULL;} // Extract the key from the record const DWORD_PTR _ExtractKey(const void* pvRecord) const { IRTLASSERT(pvRecord != NULL); IRTLASSERT(m_pfnExtractKey != NULL); return (*m_pfnExtractKey)(pvRecord); } // Hash the key DWORD _CalcKeyHash(const DWORD_PTR pnKey) const { // Note pnKey==0 is acceptable, as the real key type could be an int IRTLASSERT(m_pfnCalcKeyHash != NULL); DWORD dwHash = (*m_pfnCalcKeyHash)(pnKey); // We forcibly scramble the result to help ensure a better distribution #ifdef HASH_RANDOMIZE_BITS dwHash = RandomizeBits(dwHash); #else // !HASH_RANDOMIZE_BITS dwHash = HashScramble(dwHash); #endif // !HASH_RANDOMIZE_BITS IRTLASSERT(dwHash != HASH_INVALID_SIGNATURE); return dwHash; } // Use the key's hash signature to multiplex into a subtable SubTable* _SubTable(DWORD dwSignature) const; // Memory allocation wrappers to allow us to simulate allocation // failures during testing static SubTable** const _AllocateSubTableArray( size_t n); static bool _FreeSubTableArray( SubTable** palht); static SubTable* const _AllocateSubTable( LPCSTR pszName, // An identifier for debugging PFnExtractKey pfnExtractKey, // Extract key from record PFnCalcKeyHash pfnCalcKeyHash, // Calculate hash signature of key PFnEqualKeys pfnEqualKeys, // Compare two keys PFnAddRefRecord pfnAddRefRecord,// AddRef in FindKey, etc double maxload, // Upperbound on average chain length DWORD initsize, // Initial size of hash table. CLKRHashTable* phtParent // Owning table. ); static bool _FreeSubTable( SubTable* plht); public: CLKRHashTable( LPCSTR pszName, // An identifier for debugging PFnExtractKey pfnExtractKey, // Extract key from record PFnCalcKeyHash pfnCalcKeyHash, // Calculate hash signature of key PFnEqualKeys pfnEqualKeys, // Compare two keys PFnAddRefRecord pfnAddRefRecord,// AddRef in FindKey, etc double maxload=LK_DFLT_MAXLOAD, // bound on avg chain length DWORD initsize=LK_DFLT_INITSIZE, // Initial size of hash table. DWORD num_subtbls=LK_DFLT_NUM_SUBTBLS // #subordinate hash tables. ); ~CLKRHashTable(); static const char* ClassName() {return "CLKRHashTable";} int NumSubTables() const {return m_cSubTables;} static LK_TABLESIZE NumSubTables(DWORD& rinitsize, DWORD& rnum_subtbls); // Thin wrappers for the corresponding methods in CLKRLinearHashTable LK_RETCODE InsertRecord(const void* pvRecord, bool fOverwrite=false); LK_RETCODE DeleteKey(const DWORD_PTR pnKey); LK_RETCODE DeleteRecord(const void* pvRecord); LK_RETCODE FindKey(const DWORD_PTR pnKey, const void** ppvRecord) const; LK_RETCODE FindRecord(const void* pvRecord) const; #ifdef LKR_DEPRECATED_ITERATORS DWORD Apply(PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK); DWORD ApplyIf(PFnRecordPred pfnPredicate, PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK); DWORD DeleteIf(PFnRecordPred pfnPredicate, void* pvState=NULL); #endif // LKR_DEPRECATED_ITERATORS void Clear(); int CheckTable() const; DWORD Size() const; DWORD MaxSize() const; CLKRHashTableStats GetStatistics() const; bool IsValid() const; void SetTableLockSpinCount(WORD wSpins); WORD GetTableLockSpinCount(); void SetBucketLockSpinCount(WORD wSpins); WORD GetBucketLockSpinCount(); enum { SIGNATURE = (('L') | ('K' << 8) | ('H' << 16) | ('T' << 24)), SIGNATURE_FREE = (('L') | ('K' << 8) | ('H' << 16) | ('x' << 24)), }; bool ValidSignature() const { return m_dwSignature == SIGNATURE;} // Is the hash table usable? bool IsUsable() const { return (m_lkrcState == LK_SUCCESS); } void WriteLock(); void ReadLock() const; void WriteUnlock() const; void ReadUnlock() const; bool IsWriteLocked() const; bool IsReadLocked() const; bool IsWriteUnlocked() const; bool IsReadUnlocked() const; void ConvertSharedToExclusive() const; void ConvertExclusiveToShared() const; // LKRHASH_ALLOCATOR_DEFINITIONS(CLKRHashTable); public: #ifdef LKR_DEPRECATED_ITERATORS typedef SubTable::CIterator CLHTIterator; class CIterator : public CLHTIterator { protected: friend class CLKRHashTable; CLKRHashTable* m_pht; // which hash table? int m_ist; // which subtable private: // Private copy ctor and op= to prevent compiler synthesizing them. // Must provide (bad) implementation because we export instantiations. CIterator(const CIterator&) {*(BYTE*)NULL;} CIterator& operator=(const CIterator&) {return *(CIterator*)NULL;} public: CIterator( LK_LOCKTYPE lkl=LKL_WRITELOCK) : CLHTIterator(lkl), m_pht(NULL), m_ist(-1) {} const void* Record() const { IRTLASSERT(IsValid()); // This is a hack to work around a compiler bug. Calling // CLHTIterator::Record calls this function recursively until // the stack overflows. const CLHTIterator* pBase = static_cast(this); return pBase->Record(); } const DWORD_PTR Key() const { IRTLASSERT(IsValid()); const CLHTIterator* pBase = static_cast(this); return pBase->Key(); } bool IsValid() const { const CLHTIterator* pBase = static_cast(this); return (m_pht != NULL && m_ist >= 0 && pBase->IsValid()); } }; // Const iterators for readonly access class CConstIterator : public CIterator { private: // Private, unimplemented copy ctor and op= to prevent // compiler synthesizing them. CConstIterator(const CConstIterator&); CConstIterator& operator=(const CConstIterator&); public: CConstIterator() : CIterator(LKL_READLOCK) {} }; public: LK_RETCODE InitializeIterator(CIterator* piter); LK_RETCODE IncrementIterator(CIterator* piter); LK_RETCODE CloseIterator(CIterator* piter); LK_RETCODE InitializeIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_pht == NULL); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_pht != NULL || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; return const_cast(this) ->InitializeIterator(static_cast(piter)); } LK_RETCODE IncrementIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_pht == this); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_pht != this || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; return const_cast(this) ->IncrementIterator(static_cast(piter)); } LK_RETCODE CloseIterator(CConstIterator* piter) const { IRTLASSERT(piter != NULL && piter->m_pht == this); IRTLASSERT(piter->m_lkl != LKL_WRITELOCK); if (piter == NULL || piter->m_pht != this || piter->m_lkl == LKL_WRITELOCK) return LK_BAD_ITERATOR; return const_cast(this) ->CloseIterator(static_cast(piter)); } #endif // LKR_DEPRECATED_ITERATORS }; //-------------------------------------------------------------------- // A typesafe wrapper for CLKRHashTable (or CLKRLinearHashTable). // // * _Derived must derive from CTypedHashTable and provide certain member // functions. It's needed for various downcasting operations. See // CStringTestHashTable and CNumberTestHashTable below. // * _Record is the type of the record. C{Linear}HashTable will store // pointers to _Record. // * _Key is the type of the key. _Key is used directly; i.e., it is // not assumed to be a pointer type. C{Linear}HashTable assumes that // the key is stored in the associated record. See the comments // at the declaration of PFnExtractKey for more details. // // (optional parameters): // * _BaseHashTable is the base hash table: CLKRHashTable or /// CLKRLinearHashTable // * _BaseIterator is the iterator type, _BaseHashTable::CIterator // // CTypedHashTable could derive directly from CLKRLinearHashTable, if you // don't need the extra overhead of CLKRHashTable (which is quite low). // // You may need to add the following line to your code to disable // warning messages about truncating extremly long identifiers. // #pragma warning (disable : 4786) //-------------------------------------------------------------------- #define LKRHASH_HACKY_CAST(T, pv) ((T) (UINT_PTR) (pv)) template < class _Derived, class _Record, class _Key, class _BaseHashTable=CLKRHashTable #ifdef LKR_DEPRECATED_ITERATORS , class _BaseIterator=_BaseHashTable::CIterator #endif // LKR_DEPRECATED_ITERATORS > class CTypedHashTable : public _BaseHashTable { public: // convenient aliases typedef _Derived Derived; typedef _Record Record; typedef _Key Key; typedef _BaseHashTable BaseHashTable; typedef CTypedHashTable<_Derived, _Record, _Key, _BaseHashTable #ifdef LKR_DEPRECATED_ITERATORS , _BaseIterator #endif // LKR_DEPRECATED_ITERATORS > HashTable; #ifdef LKR_DEPRECATED_ITERATORS typedef _BaseIterator BaseIterator; #endif // LKR_DEPRECATED_ITERATORS #ifdef LKR_DEPRECATED_ITERATORS // ApplyIf() and DeleteIf(): Does the record match the predicate? // Note: takes a Record*, not a const Record*. You can modify the // record in Pred() or Action(), if you like, but if you do, you // should use LKL_WRITELOCK to lock the table. typedef LK_PREDICATE (WINAPI *PFnRecordPred) (Record* pRec, void* pvState); // Apply() et al: Perform action on record. typedef LK_ACTION (WINAPI *PFnRecordAction)(Record* pRec, void* pvState); #endif // LKR_DEPRECATED_ITERATORS private: // Wrappers for the typesafe methods exposed by the derived class static const DWORD_PTR WINAPI _ExtractKey(const void* pvRecord) { const _Record* pRec = static_cast(pvRecord); _Key key = static_cast<_Key>(_Derived::ExtractKey(pRec)); return (DWORD_PTR) key; } static DWORD WINAPI _CalcKeyHash(const DWORD_PTR pnKey) { _Key key = LKRHASH_HACKY_CAST(_Key, pnKey); return _Derived::CalcKeyHash(key); } static bool WINAPI _EqualKeys(const DWORD_PTR pnKey1, const DWORD_PTR pnKey2) { _Key key1 = LKRHASH_HACKY_CAST(_Key, pnKey1); _Key key2 = LKRHASH_HACKY_CAST(_Key, pnKey2); return _Derived::EqualKeys(key1, key2); } static void WINAPI _AddRefRecord(const void* pvRecord, int nIncr) { _Record* pRec = static_cast<_Record*>(const_cast(pvRecord)); _Derived::AddRefRecord(pRec, nIncr); } #ifdef LKR_DEPRECATED_ITERATORS // Typesafe wrappers for Apply, ApplyIf, and DeleteIf. class CState { public: PFnRecordPred m_pfnPred; PFnRecordAction m_pfnAction; void* m_pvState; CState( PFnRecordPred pfnPred, PFnRecordAction pfnAction, void* pvState) : m_pfnPred(pfnPred), m_pfnAction(pfnAction), m_pvState(pvState) {} }; static LK_PREDICATE WINAPI _Pred(const void* pvRecord, void* pvState) { _Record* pRec = static_cast<_Record*>(const_cast(pvRecord)); CState* pState = static_cast(pvState); return (*pState->m_pfnPred)(pRec, pState->m_pvState); } static LK_ACTION WINAPI _Action(const void* pvRecord, void* pvState) { _Record* pRec = static_cast<_Record*>(const_cast(pvRecord)); CState* pState = static_cast(pvState); return (*pState->m_pfnAction)(pRec, pState->m_pvState); } #endif // LKR_DEPRECATED_ITERATORS public: CTypedHashTable( LPCSTR pszName, // An identifier for debugging double maxload=LK_DFLT_MAXLOAD, // Upperbound on avg chain length DWORD initsize=LK_DFLT_INITSIZE, // Initial size of hash table. DWORD num_subtbls=LK_DFLT_NUM_SUBTBLS// #subordinate hash tables. ) : _BaseHashTable(pszName, _ExtractKey, _CalcKeyHash, _EqualKeys, _AddRefRecord, maxload, initsize, num_subtbls) {} LK_RETCODE InsertRecord(const _Record* pRec, bool fOverwrite=false) { return _BaseHashTable::InsertRecord(pRec, fOverwrite); } LK_RETCODE DeleteKey(const _Key key) { const void* pvKey = UlongToPtr(key); DWORD_PTR pnKey = reinterpret_cast(pvKey); return _BaseHashTable::DeleteKey(pnKey); } //{ return _BaseHashTable::DeleteKey(reinterpret_cast(key));} LK_RETCODE DeleteRecord(const _Record* pRec) { return _BaseHashTable::DeleteRecord(pRec);} // Note: returns a _Record**, not a const Record**. Note that you // can use a const type for the template parameter to ensure constness. LK_RETCODE FindKey(const _Key key, _Record** ppRec) const { if (ppRec == NULL) return LK_BAD_RECORD; *ppRec = NULL; const void* pvRec = NULL; const void* pvKey = UlongToPtr(key); DWORD_PTR pnKey = reinterpret_cast(pvKey); LK_RETCODE lkrc = _BaseHashTable::FindKey(pnKey, &pvRec); *ppRec = static_cast<_Record*>(const_cast(pvRec)); return lkrc; } // Note: returns a _Record**, not a const Record**. Note that you // can use a const type for the template parameter to ensure constness. LK_RETCODE FindKeyBstr(const _Key key, _Record** ppRec) const { if (ppRec == NULL) return LK_BAD_RECORD; *ppRec = NULL; const void* pvRec = NULL; const void* pvKey = key; DWORD_PTR pnKey = reinterpret_cast(pvKey); LK_RETCODE lkrc = _BaseHashTable::FindKey(pnKey, &pvRec); *ppRec = static_cast<_Record*>(const_cast(pvRec)); return lkrc; } LK_RETCODE FindRecord(const _Record* pRec) const { return _BaseHashTable::FindRecord(pRec);} // Other C{Linear}HashTable methods can be exposed without change #ifdef LKR_DEPRECATED_ITERATORS // Typesafe wrappers for Apply et al DWORD Apply(PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK) { IRTLASSERT(pfnAction != NULL); if (pfnAction == NULL) return 0; CState state(NULL, pfnAction, pvState); return _BaseHashTable::Apply(_Action, &state, lkl); } DWORD ApplyIf(PFnRecordPred pfnPredicate, PFnRecordAction pfnAction, void* pvState=NULL, LK_LOCKTYPE lkl=LKL_READLOCK) { IRTLASSERT(pfnPredicate != NULL && pfnAction != NULL); if (pfnPredicate == NULL || pfnAction == NULL) return 0; CState state(pfnPredicate, pfnAction, pvState); return _BaseHashTable::ApplyIf(_Pred, _Action, &state, lkl); } DWORD DeleteIf(PFnRecordPred pfnPredicate, void* pvState=NULL) { IRTLASSERT(pfnPredicate != NULL); if (pfnPredicate == NULL) return 0; CState state(pfnPredicate, NULL, pvState); return _BaseHashTable::DeleteIf(_Pred, &state); } // Typesafe wrappers for iterators class CIterator : public _BaseIterator { private: // Private, unimplemented copy ctor and op= to prevent // compiler synthesizing them. CIterator(const CIterator&); CIterator& operator=(const CIterator&); public: CIterator( LK_LOCKTYPE lkl=LKL_WRITELOCK) : _BaseIterator(lkl) {} _Record* Record() const { const _BaseIterator* pBase = static_cast(this); return reinterpret_cast<_Record*>(const_cast( pBase->Record())); } _Key Key() const { const _BaseIterator* pBase = static_cast(this); return reinterpret_cast<_Key>(reinterpret_cast(pBase->Key())); } }; // readonly iterator class CConstIterator : public CIterator { private: // Private, unimplemented copy ctor and op= to prevent // compiler synthesizing them. CConstIterator(const CConstIterator&); CConstIterator& operator=(const CConstIterator&); public: CConstIterator() : CIterator(LKL_READLOCK) {} const _Record* Record() const { return CIterator::Record(); } const _Key Key() const { return CIterator::Key(); } }; public: LK_RETCODE InitializeIterator(CIterator* piter) { return _BaseHashTable::InitializeIterator(piter); } LK_RETCODE IncrementIterator(CIterator* piter) { return _BaseHashTable::IncrementIterator(piter); } LK_RETCODE CloseIterator(CIterator* piter) { return _BaseHashTable::CloseIterator(piter); } LK_RETCODE InitializeIterator(CConstIterator* piter) const { return const_cast(this) ->InitializeIterator(static_cast(piter)); } LK_RETCODE IncrementIterator(CConstIterator* piter) const { return const_cast(this) ->IncrementIterator(static_cast(piter)); } LK_RETCODE CloseIterator(CConstIterator* piter) const { return const_cast(this) ->CloseIterator(static_cast(piter)); } #endif // LKR_DEPRECATED_ITERATORS }; #ifdef __LKRHASH_NAMESPACE__ } #endif // __LKRHASH_NAMESPACE__ #endif // __LKRHASH_H__