|
|
/*
Copyright (c) 2000 Microsoft Corporation
File name:
heaplowf.c Author: adrmarin Thu Nov 30 2000 Low Fragmentation Heap implementation The file implements a bucket oriented heap. This approach provides in general a significant bounded low fragmentation for large heap usages. Generally, applications tend to use only a few sizes for allocations. LFH contains a number of 128 buckets to allocate blocks up to 16K. The allocation granularity grows with the block size, keeping though a reasonable internal fragmentation (~6% for the worst case). The size and the granularity within each range is shown in the table below:
Size Range Granularity Buckets 0 256 8 32 257 512 16 16 513 1024 32 16 1025 2048 64 16 2049 4096 128 16 4097 8192 256 16 8193 16384 512 16
Regardless how randomly the allocation pattern is, the LFH will handle only 128 different sizes, choosing the smallest block large enough to complete the request. Each bucket places individual allocations into bigger blocks (sub-segments), which contain several other blocks with the same size. The allocations and free operations within sub-segments are lock free, the algorithm being similar with allocations from lookasides (interlocked S-Lists). This is the fastest path of the heap allocator, and provides performance similar with lookasides; it also keeps all these blocks together, avoiding fragmentation. Depending upon the heap usage, each bucket can have several sub-segments allocated to satisfy all requests, but only one of these is currently in use for allocations (it is active). When the active sub-segments has no sub-blocks available, another sub-segment will become active to satisfy the allocation requests. If the bucket does not have any available sub-segments, it will allocate a new one from the NT heap. Also, if a sub-segment does not contain any busy sub-block, the whole amount of memory will be returned to the NT heap. Unlike the allocations, which are done from the active sub-segments, the free operations can be done to every sub-segments, either active or passive. There are no constrains regarding the number of blocks within a sub-segment. LFH is more concerned with the sub-segments sizes that are allocated from the NT heap. Since the best-fit policy is good if we keep a relatively small number of sizes and blocks, the LFH will allocate sub-segments in size power of two. In practice, only about 9 different sizes will be requested from the NT heap (from 1K to 512K). In this way, depending upon the size in the current bucket will result a number of blocks. When the sub-segment is destroyed, that big chunk is returned to the NT heap, making it possible to be reused later for other buckets. Note that for some app scenarios, with low heap usage, random distributed, LFH is not the best choice. To achieve a good SMP scalability, all operations here are non-blocking. The only situation where we aquire a critical section is when wi allocate an array of sub-segment descriptors. This is a very rare case, even for an intensive MP usage. */
#include "ntrtlp.h"
#include "heap.h"
#include "heappriv.h"
//#define _HEAP_DEBUG
#define PrintMsg DbgPrint
#define HeapAlloc RtlAllocateHeap
#define HeapFree RtlFreeHeap
//
// The conversion code needs the Lock/Unlock APIs
//
#define HeapLock RtlLockHeap
#define HeapUnlock RtlUnlockHeap
#define HeapSize RtlSizeHeap
#ifdef _HEAP_DEBUG
#define HeapValidate RtlValidateHeap
#endif //_HEAP_DEBUG
PSINGLE_LIST_ENTRY FASTCALL RtlpInterlockedPopEntrySList ( IN PSLIST_HEADER ListHead );
PSINGLE_LIST_ENTRY FASTCALL RtlpInterlockedPushEntrySList ( IN PSLIST_HEADER ListHead, IN PSINGLE_LIST_ENTRY ListEntry );
#define RtlpSubSegmentPop RtlpInterlockedPopEntrySList
#define RtlpSubSegmentPush(SList,Block) \
RtlpInterlockedPushEntrySList((SList),(PSINGLE_LIST_ENTRY)(Block))
#define TEBDesiredAffinity (NtCurrentTeb()->HeapVirtualAffinity)
//
// On x86 is not available the interlockCompareExchange64. We need to implement it localy.
// Also the macros below will take care of the inconsistency between definition
// of this function on X86 and 64-bit platforms
//
#if !defined(_WIN64)
LONGLONG FASTCALL RtlInterlockedCompareExchange64 ( IN OUT PLONGLONG Destination, IN PLONGLONG Exchange, IN PLONGLONG Comperand );
#define LOCKCOMP64(l,n,c) \
(RtlInterlockedCompareExchange64((PLONGLONG)(l), (PLONGLONG)(&n), (PLONGLONG)(&c)) == (*((PLONGLONG)(&c))))
#else //#if defined(_WIN64)
//
// 64 bit specific definitions
//
#define LOCKCOMP64(l,n,c) \
(_InterlockedCompareExchange64((PLONGLONG)(l), *((PLONGLONG)(&n)), *((PLONGLONG)(&c))) == (*((PLONGLONG)(&c))))
#endif // #if defined(_WIN64)
ULONG FORCEINLINE RtlpGetFirstBitSet64( LONGLONG Mask ) { if ((ULONG)Mask) {
return RtlFindFirstSetRightMember((ULONG)Mask); }
return 32 + RtlFindFirstSetRightMember((ULONG)(Mask >> 32)); }
#define HEAP_AFFINITY_LIMIT 64 // N.B. This cannot be larger than 64
// (the number of bits in LONGLONG data type)
typedef struct _AFFINITY_STATE{
LONGLONG FreeEntries; LONGLONG UsedEntries;
ULONG Limit; LONG CrtLimit;
ULONG_PTR OwnerTID[ HEAP_AFFINITY_LIMIT ];
//
// The counters below are not absolutely necessary for the affinity manager.
// But these help understanding the frequence of affinity changes. In general
// accessing of these fields should be rare, even for many threads (like hundreds)
// Benchmarks didn't show any visible difference with all these removed
//
ULONG AffinitySwaps; ULONG AffinityResets; ULONG AffinityLoops; ULONG AffinityAllocs;
} AFFINITY_STATE, *PAFFINITY_STATE;
#define GetCrtThreadId() ((ULONG_PTR)NtCurrentTeb()->ClientId.UniqueThread)
AFFINITY_STATE RtlpAffinityState;
VOID RtlpInitializeAffinityManager( UCHAR Size )
/*++
Routine Description: This routine initialize the affinity manager. This should be done only ones into that process, before any other affinity function is invoked Arguments:
Size - The number of virtual affinity entries Return Value: None --*/
{
//
// The size of the affinity bitmap is limited to the number of bits from
// an LONGLONG data type.
//
if (Size > HEAP_AFFINITY_LIMIT) {
PrintMsg( "HEAP: Invalid size %ld for the affinity mask. Using %ld instead\n", Size, HEAP_AFFINITY_LIMIT ); Size = HEAP_AFFINITY_LIMIT; }
RtlpAffinityState.FreeEntries = 0; RtlpAffinityState.UsedEntries = 0; RtlpAffinityState.Limit = Size;
RtlpAffinityState.CrtLimit = -1; RtlpAffinityState.AffinitySwaps = 0; RtlpAffinityState.AffinityResets = 0; RtlpAffinityState.AffinityAllocs = 0; }
ULONG FASTCALL RtlpAllocateAffinityIndex( )
/*++
Routine Description: The function allocates a new index into the virtual affinity array Arguments: Return Value: Return the index, which the current thread can use further. --*/
{ ULONGLONG CapturedMask;
InterlockedIncrement(&RtlpAffinityState.AffinityAllocs);
RETRY:
//
// Check first whether we have at least a free entry in the affinity mask
//
if (CapturedMask = RtlpAffinityState.FreeEntries) {
ULONGLONG AvailableMask;
AvailableMask = CapturedMask & RtlpAffinityState.UsedEntries;
if (AvailableMask) {
ULONG Index = RtlpGetFirstBitSet64(AvailableMask); LONGLONG NewMask = CapturedMask & ~((LONGLONG)1 << Index); if (!LOCKCOMP64(&RtlpAffinityState.FreeEntries, NewMask, CapturedMask)) {
goto RETRY; }
RtlpAffinityState.OwnerTID[ Index ] = GetCrtThreadId();
return Index; }
} //
// Nothing available. We need to allocate a new entry. We won't do this
// unless it's absolutely necessary
//
if (RtlpAffinityState.CrtLimit < (LONG)(RtlpAffinityState.Limit - 1)) {
ULONG NewLimit = InterlockedIncrement(&RtlpAffinityState.CrtLimit);
//
// We already postponed growing the size. We have to do now
//
if ( NewLimit < RtlpAffinityState.Limit) {
LONGLONG CapturedUsed; LONGLONG NewMask;
do {
CapturedUsed = RtlpAffinityState.UsedEntries; NewMask = CapturedUsed | ((LONGLONG)1 << NewLimit);
} while ( !LOCKCOMP64(&RtlpAffinityState.UsedEntries, NewMask, CapturedUsed) );
RtlpAffinityState.FreeEntries = ~((LONGLONG)1 << NewLimit);
RtlpAffinityState.OwnerTID[ NewLimit ] = GetCrtThreadId();
return NewLimit;
} else {
InterlockedDecrement(&RtlpAffinityState.CrtLimit); } } if ((RtlpAffinityState.FreeEntries & RtlpAffinityState.UsedEntries) == 0) {
RtlpAffinityState.FreeEntries = (LONGLONG)-1;
InterlockedIncrement( &RtlpAffinityState.AffinityResets ); }
InterlockedIncrement( &RtlpAffinityState.AffinityLoops );
goto RETRY;
//
// return something to make the compiler happy
//
return 0; }
ULONG FORCEINLINE RtlpGetThreadAffinity( )
/*++
Routine Description: The function returns the affinity which the current thread can use. This function is designed to be called pretty often. It has a quick path, which test whether the last affinity assigned did not expired. If the number of threads is less than the number of processors, the threads will never get moved from an index to another. Arguments: None Return Value: Return the index, which the current thread can use further. --*/
{ LONG NewAffinity; LONG CapturedAffinity = TEBDesiredAffinity - 1;
if (CapturedAffinity >= 0) {
if (RtlpAffinityState.OwnerTID[CapturedAffinity] == GetCrtThreadId()) { if (RtlpAffinityState.FreeEntries & ((LONGLONG)1 << CapturedAffinity)) {
LONGLONG NewMask = RtlpAffinityState.FreeEntries & ~(((LONGLONG)1 << CapturedAffinity));
LOCKCOMP64(&RtlpAffinityState.FreeEntries, NewMask, RtlpAffinityState.FreeEntries); }
return CapturedAffinity; }
} else {
//
// A new thread came up. Reset the affinity
//
RtlpAffinityState.FreeEntries = (LONGLONG) -1; }
NewAffinity = RtlpAllocateAffinityIndex();
if ((NewAffinity + 1) != TEBDesiredAffinity) {
InterlockedIncrement( &RtlpAffinityState.AffinitySwaps ); }
TEBDesiredAffinity = NewAffinity + 1;
return NewAffinity; }
//
// Low fragmentation heap tunning constants
//
#define LOCALPROC FORCEINLINE
//
// The total number of buckets. The default is 128 which coveres
// blocks up to 16 K
//
// N.B. HEAP_BUCKETS_COUNT must be > 32 and multiple of 16
//
#define HEAP_BUCKETS_COUNT 128
//
// Defining the limits for the number of blocks that can exist in sub-segment
// Number of blocks >= 2^HEAP_MIN_BLOCK_CLASS
// &&
// Number of blocks <= 2^HEAP_MAX_BLOCK_CLASS
// &&
// sub-segment size <= HEAP_MAX_SUBSEGMENT_SIZE
//
#define HEAP_MIN_BLOCK_CLASS 4
#define HEAP_MAX_BLOCK_CLASS 10
#define HEAP_MAX_SUBSEGMENT_SIZE (0x0000F000 << HEAP_GRANULARITY_SHIFT) // must be smaller than HEAP_MAXIMUM_BLOCK_SIZE
//
// If a size become very popular, LFH increases the number of blocks
// that could be placed into the subsegments, with the formula below.
//
#define RtlpGetDesiredBlockNumber(Aff,T) \
((Aff) ? (((T) >> 4) / (RtlpHeapMaxAffinity)) : ((T) >> 4))
//
// LFH uses only a few different sizes for subsegments. These have sizes
// of power of two between
// 2^HEAP_LOWEST_USER_SIZE_INDEX and 2^HEAP_HIGHEST_USER_SIZE_INDEX
//
#define HEAP_LOWEST_USER_SIZE_INDEX 7
#define HEAP_HIGHEST_USER_SIZE_INDEX 18
//
// The sub-segments descriptors are allocated in zones, based on
// processor affinity. These descriptors are in general small structures,
// so ignoring the affinity impacts the performance on MP machines with
// caches > sizeof( HEAP_SUBSEGMENT ).
// Also it significantly reduces the number of calls to the NT heap
//
#define HEAP_DEFAULT_ZONE_SIZE (1024 - sizeof(HEAP_ENTRY)) // allocate 1 K at ones
//
// Each bucket holds a number of subsegments into a cache, in order
// to find the emptiest one for reusage. FREE_CACHE_SIZE defines the
// number of sub-segments that will be searched
//
#define FREE_CACHE_SIZE 16
//
// On low memory, the subsegments that are almost free can be converted to
// the regular NT heap. HEAP_CONVERT_LIMIT gives the maximum space that can be
// converted at ones
//
#define HEAP_CONVERT_LIMIT 0x1000000 // Do not convert more than 16 MBytes at ones
//
// Cache tunning constants.
// LFH keeps the subsegments into a cache to be easy reused for different allocations
// This significantly reduces the number of calls to the NT heap with huge impact
// in scalability, performance and footprint for the most common cases.
// The only problem is in a shrinking phase, when the app frees the most part
// of the memory, and we want to reduce the commited space for the heap.
// We handle this case with the following two constants:
// - HEAP_CACHE_FREE_THRESHOLD
// - HEAP_CACHE_SHIFT_THRESHOLD
//
// The heap will free a block to the NT heap only if these conditions are TRUE:
// - The number of blocks in cache for that size > HEAP_CACHE_FREE_THRESHOLD
// - The number of blocks in cache for that size >
// (Total number of blocks of that size) >> HEAP_CACHE_SHIFT_THRESHOLD
//
//
#define HEAP_CACHE_FREE_THRESHOLD 8
#define HEAP_CACHE_SHIFT_THRESHOLD 2
//
// Other definitions
//
#define NO_MORE_ENTRIES 0xFFFF
//
// Locking constants
//
#define HEAP_USERDATA_LOCK 1
#define HEAP_PUBLIC_LOCK 2
#define HEAP_ACTIVE_LOCK 4
#define HEAP_CONVERT_LOCK 8
#define HEAP_FREE_BLOCK_SUCCESS 1
#define HEAP_FREE_BLOCK_CONVERTED 2
#define HEAP_FREE_SEGMENT_EMPTY 3
//
// Low fragmentation heap data structures
//
typedef union _HEAP_BUCKET_COUNTERS{
struct { volatile ULONG TotalBlocks; volatile ULONG SubSegmentCounts; };
volatile LONGLONG Aggregate64;
} HEAP_BUCKET_COUNTERS, *PHEAP_BUCKET_COUNTERS;
//
// The HEAP_BUCKET structure handles same size allocations
//
typedef struct _HEAP_BUCKET {
HEAP_BUCKET_COUNTERS Counters;
USHORT BlockUnits; UCHAR SizeIndex; UCHAR UseAffinity; LONG Conversions;
} HEAP_BUCKET, *PHEAP_BUCKET;
//
// LFH heap uses zones to allocate sub-segment descriptors. This will preallocate
// a large block and then for each individual sub-segment request will move the
// water mark pointer with a non-blocking operation
//
typedef struct _LFH_BLOCK_ZONE {
LIST_ENTRY ListEntry; PVOID FreePointer; PVOID Limit;
} LFH_BLOCK_ZONE, *PLFH_BLOCK_ZONE;
typedef struct _HEAP_LOCAL_SEGMENT_INFO {
PHEAP_SUBSEGMENT Hint; PHEAP_SUBSEGMENT ActiveSubsegment;
PHEAP_SUBSEGMENT CachedItems[ FREE_CACHE_SIZE ]; SLIST_HEADER SListHeader;
SIZE_T BusyEntries; SIZE_T LastUsed;
} HEAP_LOCAL_SEGMENT_INFO, *PHEAP_LOCAL_SEGMENT_INFO;
typedef struct _HEAP_LOCAL_DATA { //
// We reserve the 128 bytes below to avoid sharing memory
// into the same cacheline on MP machines
//
UCHAR Reserved[128];
volatile PLFH_BLOCK_ZONE CrtZone; struct _LFH_HEAP * LowFragHeap;
HEAP_LOCAL_SEGMENT_INFO SegmentInfo[HEAP_BUCKETS_COUNT]; SLIST_HEADER DeletedSubSegments;
ULONG Affinity; ULONG Reserved1;
} HEAP_LOCAL_DATA, *PHEAP_LOCAL_DATA;
//
// Fixed size large block cache data structures & definitions
// This holds in S-Lists the blocks that can be free, but it
// delay the free until no other thread is doing a heap operation
// This helps reducing the contention on the heap lock,
// improve the scalability with a relatively low memory footprint
//
#define HEAP_USER_ENTRIES (HEAP_HIGHEST_USER_SIZE_INDEX - HEAP_LOWEST_USER_SIZE_INDEX + 1)
typedef struct _USER_MEMORY_CACHE {
SLIST_HEADER UserBlocks[ HEAP_USER_ENTRIES ];
ULONG FreeBlocks; ULONG Sequence;
ULONG MinDepth[ HEAP_USER_ENTRIES ]; ULONG AvailableBlocks[ HEAP_USER_ENTRIES ]; } USER_MEMORY_CACHE, *PUSER_MEMORY_CACHE;
typedef struct _LFH_HEAP { RTL_CRITICAL_SECTION Lock;
LIST_ENTRY SubSegmentZones; SIZE_T ZoneBlockSize; HANDLE Heap; LONG Conversions; LONG ConvertedSpace;
ULONG SegmentChange; //
ULONG SegmentCreate; // Various counters (optional)
ULONG SegmentInsertInFree; //
ULONG SegmentDelete; //
USER_MEMORY_CACHE UserBlockCache;
//
// Bucket data
//
HEAP_BUCKET Buckets[HEAP_BUCKETS_COUNT];
//
// The LocalData array must be the last field in LFH structures
// The sizes of the array is choosen depending upon the
// number of processors.
//
HEAP_LOCAL_DATA LocalData[1];
} LFH_HEAP, *PLFH_HEAP;
//
// Debugging macros.
//
#ifdef _HEAP_DEBUG
LONG RtlpColissionCounter = 0;
#define LFHEAPASSERT(exp) \
if (!(exp)) { \ PrintMsg( "\nERROR: %s\n\tSource File: %s, line %ld\n", #exp, __FILE__, __LINE__);\ DbgBreakPoint(); \ }
#define LFHEAPWARN(exp) \
if (!(exp)) PrintMsg( "\nWARNING: %s\n\tSource File: %s, line %ld\n", #exp, __FILE__, __LINE__);
#define LFH_DECLARE_COUNTER ULONG __Counter = 0;
#define LFH_UPDATE_COUNTER \
if ((++__Counter) > 1) { \ InterlockedIncrement(&RtlpColissionCounter); \ }
#else
#define LFHEAPASSERT(exp)
#define LFHEAPWARN(exp)
#define LFH_DECLARE_COUNTER
#define LFH_UPDATE_COUNTER
#endif
BOOLEAN FORCEINLINE RtlpLockSubSegment( PHEAP_SUBSEGMENT SubSegment, ULONG LockMask );
BOOLEAN LOCALPROC RtlpUnlockSubSegment( PHEAP_LOCAL_DATA LocalData, PHEAP_SUBSEGMENT SubSegment, ULONG LockMask );
BOOLEAN FASTCALL RtlpConvertSegmentToHeap ( PLFH_HEAP LowFragHeap, PHEAP_SUBSEGMENT SubSegment );
ULONG RtlpFlushLFHeapCache ( PLFH_HEAP LowFragHeap );
//
// Heap manager globals
//
SIZE_T RtlpBucketBlockSizes[HEAP_BUCKETS_COUNT]; ULONG RtlpHeapMaxAffinity = 0;
//
// User block management private functions
//
SIZE_T FORCEINLINE RtlpConvertSizeIndexToSize( UCHAR SizeIndex )
/*++
Routine Description: The function converts a size index into a memory block size. LFH requests only these particular sizes from the NT heap Arguments: SizeIndex - The size category Return Value: The size in bytes to be requested from the NT heap
--*/
{ SIZE_T Size = 1 << SizeIndex;
LFHEAPASSERT( SizeIndex >= HEAP_LOWEST_USER_SIZE_INDEX ); LFHEAPASSERT( SizeIndex <= HEAP_HIGHEST_USER_SIZE_INDEX );
if (Size > HEAP_MAX_SUBSEGMENT_SIZE) {
Size = HEAP_MAX_SUBSEGMENT_SIZE; }
return Size - sizeof(HEAP_ENTRY); }
PVOID FASTCALL RtlpAllocateUserBlock( PLFH_HEAP LowFragHeap, UCHAR SizeIndex )
/*++
Routine Description: The function allocates a large block for the sub-segment user data It tries first to allocate from the cache. So it makes an NT heap call only if the first one fails. The blocks allocated with this routine can only have power of 2 sizes (from 256, 512, ....) Arguments: LowFragHeap - The pointer to the LF heap SizeIndex - The category size to be allocated Return Value: Returns a pointer to the new allocated block, or NULL if the operation fails
--*/
{ PVOID ListEntry; PHEAP_USERDATA_HEADER UserBlock = NULL;
LFHEAPASSERT(SizeIndex >= HEAP_LOWEST_USER_SIZE_INDEX); LFHEAPASSERT(SizeIndex <= HEAP_HIGHEST_USER_SIZE_INDEX);
//
// Allocates first from the slist cache
//
__try {
//
// Search first into the indicated index
//
if (ListEntry = RtlpSubSegmentPop(&LowFragHeap->UserBlockCache.UserBlocks[SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX])) { UserBlock = CONTAINING_RECORD(ListEntry, HEAP_USERDATA_HEADER, SFreeListEntry);
leave; }
//
// Look for a smaller size
//
if (SizeIndex > HEAP_LOWEST_USER_SIZE_INDEX) {
if (ListEntry = RtlpSubSegmentPop(&LowFragHeap->UserBlockCache.UserBlocks[SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX - 1])) { UserBlock = CONTAINING_RECORD(ListEntry, HEAP_USERDATA_HEADER, SFreeListEntry);
leave; } } } __except (EXCEPTION_EXECUTE_HANDLER) {
//
// Nothing to do
//
}
if (UserBlock == NULL) { //
// There is no available blocks into the cache. We need to
// allocate the subsegment from the NT heap
//
InterlockedIncrement(&LowFragHeap->UserBlockCache.AvailableBlocks[ SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX ]);
UserBlock = HeapAlloc(LowFragHeap->Heap, HEAP_NO_CACHE_BLOCK, RtlpConvertSizeIndexToSize(SizeIndex));
if (UserBlock) {
UserBlock->SizeIndex = SizeIndex; } }
return UserBlock; }
VOID FASTCALL RtlpFreeUserBlock( PLFH_HEAP LowFragHeap, PHEAP_USERDATA_HEADER UserBlock )
/*++
Routine Description:
Free a block previously allocated with RtlpAllocateUserBlock. Arguments: LowFragHeap - The pointer to the LF heap UserBlock - The block to be freed Return Value: None.
--*/
{ ULONG Depth; ULONG SizeIndex = (ULONG)UserBlock->SizeIndex; PSLIST_HEADER ListHeader = &LowFragHeap->UserBlockCache.UserBlocks[UserBlock->SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX]; if (UserBlock->SizeIndex == 0) {
//
// This block was converted before to NT heap block
//
HeapFree(LowFragHeap->Heap, 0, UserBlock);
return; }
LFHEAPASSERT(UserBlock->SizeIndex >= HEAP_LOWEST_USER_SIZE_INDEX); LFHEAPASSERT(UserBlock->SizeIndex <= HEAP_HIGHEST_USER_SIZE_INDEX);
LFHEAPASSERT( RtlpConvertSizeIndexToSize((UCHAR)UserBlock->SizeIndex) == HeapSize(LowFragHeap->Heap, 0, UserBlock) );
Depth = QueryDepthSList(&LowFragHeap->UserBlockCache.UserBlocks[SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX]);
if ((Depth > HEAP_CACHE_FREE_THRESHOLD) && (Depth > (LowFragHeap->UserBlockCache.AvailableBlocks[ SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX ] >> HEAP_CACHE_SHIFT_THRESHOLD))) { PVOID ListEntry; HeapFree(LowFragHeap->Heap, 0, UserBlock);
ListEntry = NULL; __try {
ListEntry = RtlpSubSegmentPop(&LowFragHeap->UserBlockCache.UserBlocks[SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX]);
} __except (EXCEPTION_EXECUTE_HANDLER) { } if (ListEntry != NULL) { UserBlock = CONTAINING_RECORD(ListEntry, HEAP_USERDATA_HEADER, SFreeListEntry); HeapFree(LowFragHeap->Heap, 0, UserBlock); InterlockedDecrement(&LowFragHeap->UserBlockCache.AvailableBlocks[ SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX ]); }
InterlockedDecrement(&LowFragHeap->UserBlockCache.AvailableBlocks[ SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX ]);
} else { RtlpSubSegmentPush( ListHeader, &UserBlock->SFreeListEntry); } }
VOID FORCEINLINE RtlpMarkLFHBlockBusy ( PHEAP_ENTRY Block )
/*++
Routine Description: This function marks a LFH block as busy. Because the convert routine can be invoked any time, the LFH cannot use the same flag as the regular heap (LFH access any fields unsynchronized, but the block flags are supposed to be accessed while holding the heap lock) Arguments: Block - The block being marked as busy Return Value: None
--*/
{ Block->SmallTagIndex = 1; }
VOID FORCEINLINE RtlpMarkLFHBlockFree ( PHEAP_ENTRY Block )
/*++
Routine Description: This function marks a LFH block as free. Because the convert routine can be invoked any time, the LFH cannot use the same flag as the regular heap (LFH access any fields unsynchronized, but the block flags are supposed to be accessed while holding the heap lock) Arguments: Block - The block to be marked free Return Value: None
--*/
{ Block->SmallTagIndex = 0; }
BOOLEAN FORCEINLINE RtlpIsLFHBlockBusy ( PHEAP_ENTRY Block )
/*++
Routine Description: The function returns whether the block is busy or free Arguments: Block - The heap block tested Return Value: Return TRUE if the block is busy
--*/
{ return (Block->SmallTagIndex == 1); }
VOID FORCEINLINE RtlpUpdateLastEntry ( PHEAP Heap, PHEAP_ENTRY Block )
/*++
Routine Description: The function updates the last entry in segment. This is mandatory each time a new block become the last.
Arguments:
Heap - The NT heap structure Block - The block being tested for LAST_ENTRY flag Return Value: None
--*/
{ if (Block->Flags & HEAP_ENTRY_LAST_ENTRY) {
PHEAP_SEGMENT Segment;
Segment = Heap->Segments[Block->SegmentIndex]; Segment->LastEntryInSegment = Block; } }
BOOLEAN FORCEINLINE RtlpIsSubSegmentEmpty( PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
This function tests whether the subsegment does contain available sub-blocks
Arguments: SubSegment - The subsegment being tested
Return Value:
TRUE if no blocks are available.
--*/
{ return SubSegment->AggregateExchg.OffsetAndDepth == (NO_MORE_ENTRIES << 16); }
VOID FORCEINLINE RtlpUpdateBucketCounters ( PHEAP_BUCKET Bucket, LONG TotalBlocks )
/*++
Routine Description: The function updates the total number of blocks from a bucket and the number of sub-segments with a single interlocked operation. This function should be called each time a new segment is allocated / deleted to/from this bucket
Arguments:
Bucket - The heap bucket that needs to be updated TotalBlocks - The number of blocks added / subtracted from the bucket. A positive value means the bucket increased with a new segment, and a positive value means a sub-segment with that many blocks was deleted.
Return Value: None
--*/
{ HEAP_BUCKET_COUNTERS CapturedValue, NewValue; LFH_DECLARE_COUNTER;
do {
//
// Capture the current value for counters
//
CapturedValue.Aggregate64 = Bucket->Counters.Aggregate64;
//
// Calculate the new value depending upon the captured state
//
NewValue.TotalBlocks = CapturedValue.TotalBlocks + TotalBlocks;
if (TotalBlocks > 0) {
NewValue.SubSegmentCounts = CapturedValue.SubSegmentCounts + 1;
} else { NewValue.SubSegmentCounts = CapturedValue.SubSegmentCounts - 1; }
LFH_UPDATE_COUNTER;
//
// try to replace the original value with the current one. If the
// lockcomp below fails, retry all the ops above
//
} while ( !LOCKCOMP64(&Bucket->Counters.Aggregate64, NewValue.Aggregate64, CapturedValue.Aggregate64) );
//
// It's invalid to have negative numbers of blocks or sub-segments
//
LFHEAPASSERT(((LONG)NewValue.SubSegmentCounts) >= 0); LFHEAPASSERT(((LONG)NewValue.TotalBlocks) >= 0); }
ULONG FORCEINLINE RtlpGetThreadAffinityIndex( PHEAP_BUCKET HeapBucket )
/*++
Routine Description:
The affinity is managed independenlty on each bucket. This will spin up the number of sub-segments that can be accessed simultanely only for most used buckets. The routinw will hash the thread ID givving the right affinity index depending the affinity size for that bucket.
Arguments:
Bucket - The heap bucket queried
Return Value:
The affinity the current thread should use for allocation from this bucket
--*/
{ if (HeapBucket->UseAffinity) {
return 1 + RtlpGetThreadAffinity(); } return 0; }
BOOLEAN FORCEINLINE RtlpIsSubSegmentLocked( PHEAP_SUBSEGMENT SubSegment, ULONG LockMask )
/*++
Routine Description:
This function tests whether the subsegment has the given lock bits set
Arguments:
SubSegment - The sub-segment tested LockMask - contains the bits to be tested
Return Value:
It returns false if any bit from the mask is not set
--*/
{ return ((SubSegment->Lock & LockMask) == LockMask); }
BOOLEAN LOCALPROC RtlpAddToSegmentInfo( PHEAP_LOCAL_DATA LocalData, IN PHEAP_LOCAL_SEGMENT_INFO SegmentInfo, IN PHEAP_SUBSEGMENT NewItem ) {
ULONG Index; for (Index = 0; Index < FREE_CACHE_SIZE; Index++) {
ULONG i = (Index + (ULONG)SegmentInfo->LastUsed) & (FREE_CACHE_SIZE - 1);
PHEAP_SUBSEGMENT CrtSubSegment = SegmentInfo->CachedItems[i];
if (CrtSubSegment == NULL ) { if (InterlockedCompareExchangePointer( &SegmentInfo->CachedItems[i], NewItem, NULL) == NULL) {
SegmentInfo->BusyEntries += 1;
return TRUE; }
} else {
if (!RtlpIsSubSegmentLocked(CrtSubSegment, HEAP_USERDATA_LOCK)) {
if (InterlockedCompareExchangePointer( &SegmentInfo->CachedItems[i], NewItem, CrtSubSegment) == CrtSubSegment) {
RtlpUnlockSubSegment(LocalData, CrtSubSegment, HEAP_PUBLIC_LOCK); return TRUE; } } } }
return FALSE; }
PHEAP_SUBSEGMENT LOCALPROC RtlpRemoveFromSegmentInfo( PHEAP_LOCAL_DATA LocalData, IN PHEAP_LOCAL_SEGMENT_INFO SegmentInfo ) {
ULONG i; PHEAP_SUBSEGMENT * Location = NULL; ULONG LargestDepth = 0; PHEAP_SUBSEGMENT CapturedSegment;
RETRY: for (i = 0; i < FREE_CACHE_SIZE; i++) {
ULONG Depth; PHEAP_SUBSEGMENT CrtSubsegment = SegmentInfo->CachedItems[i];
if ( CrtSubsegment && (Depth = CrtSubsegment->AggregateExchg.Depth) > LargestDepth) {
CapturedSegment = CrtSubsegment; LargestDepth = Depth; Location = &SegmentInfo->CachedItems[i]; } }
if (Location) {
PHEAP_SUBSEGMENT NextEntry; while (NextEntry = (PHEAP_SUBSEGMENT)RtlpSubSegmentPop(&SegmentInfo->SListHeader)) {
NextEntry = CONTAINING_RECORD(NextEntry, HEAP_SUBSEGMENT, SFreeListEntry);
#ifdef _HEAP_DEBUG
NextEntry->SFreeListEntry.Next = NULL; #endif
if (RtlpIsSubSegmentLocked(NextEntry, HEAP_USERDATA_LOCK)) {
break; }
RtlpUnlockSubSegment(LocalData, NextEntry, HEAP_PUBLIC_LOCK); }
if (InterlockedCompareExchangePointer( Location, NextEntry, CapturedSegment) == CapturedSegment) {
if (NextEntry == NULL) { SegmentInfo->BusyEntries -= 1; SegmentInfo->LastUsed = Location - &SegmentInfo->CachedItems[0];
LFHEAPASSERT(SegmentInfo->LastUsed < FREE_CACHE_SIZE); }
return CapturedSegment;
} else if (NextEntry){
RtlpSubSegmentPush( &SegmentInfo->SListHeader, &NextEntry->SFreeListEntry); }
Location = NULL; LargestDepth = 0;
goto RETRY; }
return NULL; }
PHEAP_SUBSEGMENT LOCALPROC RtlpRemoveFreeSubSegment( PHEAP_LOCAL_DATA LocalData, ULONG SizeIndex )
/*++
Routine Description:
This function remove a sub-segments that has free sub-blocks from the free list.
Arguments:
LowFragHeap - The Low Fragmentation Heap handle HeapBucket - The bucket where we need to reuse a free block
Return Value:
A subsegment that contains free blocks
--*/
{ PVOID Entry; LONG Depth; PHEAP_LOCAL_SEGMENT_INFO FreeSList; PHEAP_SUBSEGMENT SubSegment;
SubSegment = RtlpRemoveFromSegmentInfo(LocalData, &LocalData->SegmentInfo[SizeIndex]);
if (SubSegment) {
if ( RtlpUnlockSubSegment(LocalData, SubSegment, HEAP_PUBLIC_LOCK)){
return SubSegment; } } FreeSList = &LocalData->SegmentInfo[SizeIndex];
while (Entry = RtlpSubSegmentPop(&FreeSList->SListHeader) ) {
SubSegment = CONTAINING_RECORD(Entry, HEAP_SUBSEGMENT, SFreeListEntry);
#ifdef _HEAP_DEBUG
SubSegment->SFreeListEntry.Next = NULL; #endif
LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK) ); LFHEAPASSERT( SizeIndex == SubSegment->SizeIndex ); //
// If we have a non-empty subsegments we'll return it
//
if ( RtlpUnlockSubSegment(LocalData, SubSegment, HEAP_PUBLIC_LOCK) && (SubSegment->AggregateExchg.Depth != 0)) { return SubSegment; } } return NULL; }
BOOLEAN LOCALPROC RtlpInsertFreeSubSegment( PHEAP_LOCAL_DATA LocalData, PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
The function inserts a Subsegments that has certain number of free blocks into the list with free segments. The insertion is done according with the current thread affinity.
Arguments:
SubSegment - The sub-segment being inserted into the bucket's free list
Return Value:
TRUE if succeeds. False if someone else inserted the segment meanwhile, of freed it
--*/
{ if ( RtlpLockSubSegment(SubSegment, HEAP_PUBLIC_LOCK) ) {
PHEAP_LOCAL_SEGMENT_INFO FreeSList; if (RtlpAddToSegmentInfo(LocalData, &LocalData->SegmentInfo[SubSegment->SizeIndex], SubSegment)) {
return TRUE; } FreeSList = &LocalData->SegmentInfo[SubSegment->SizeIndex];
#ifdef _HEAP_DEBUG
InterlockedIncrement(&LocalData->LowFragHeap->SegmentInsertInFree); #endif
LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK) ); LFHEAPASSERT( SubSegment->SFreeListEntry.Next == NULL ); RtlpSubSegmentPush( &FreeSList->SListHeader, &SubSegment->SFreeListEntry);
return TRUE; }
return FALSE; }
BOOLEAN LOCALPROC RtlpTrySetActiveSubSegment ( PHEAP_LOCAL_DATA LocalData, PHEAP_BUCKET HeapBucket, PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
This function tries to elevate the active segment into an active state. An active state is defined here the state where a segment is used for allocations. There is no guarantee a segment can be set into an active state because of non-blocking algorithms. An other thread can free it meanwhile or set it before. In these cases the function will fail. Arguments:
LowFragHeap - The LFH pointer HeapBucket - The bucket containing the active sub-segment Affinity - the required affinity for that segment SubSegment - The subsegment being activated. If NULL, this function will remove only the current active segment.
Return Value:
TRUE if succeeds.
--*/
{ PHEAP_SUBSEGMENT PreviousSubSegment;
if (SubSegment) {
//
// If we received a sub-segment we need to lock it exclusively in order
// to protect against other threads trying to do the same thing
// at the same time
//
if ( !RtlpLockSubSegment(SubSegment, HEAP_ACTIVE_LOCK | HEAP_PUBLIC_LOCK) ) {
return FALSE; }
//
// We have granted exclusive access to this sub-segment at this point.
// We need to test whether this subsegment wasn't freed meanwhile and reused
// for other allocations (for a different bucket)
//
if (SubSegment->Bucket != HeapBucket) {
//
// Someone freed it before and reuse it. We need to back out
// whatever we've done before. We cannot insert it into this bucket,
// since it contains different block sizes
//
if (RtlpUnlockSubSegment(LocalData, SubSegment, HEAP_ACTIVE_LOCK | HEAP_PUBLIC_LOCK)) { if (SubSegment->AggregateExchg.Depth) {
RtlpInsertFreeSubSegment(LocalData, SubSegment); } }
return FALSE; }
LFHEAPASSERT( SubSegment->SFreeListEntry.Next == NULL ); LFHEAPASSERT( HeapBucket == SubSegment->Bucket); LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK));
#ifdef _HEAP_DEBUG
SubSegment->SFreeListEntry.Next = (PSINGLE_LIST_ENTRY)(ULONG_PTR)0xEEEEEEEE;
#endif
LFHEAPASSERT( SubSegment->AffinityIndex == (UCHAR)LocalData->Affinity ); }
//
// Try to set this sub-segment as active an capture the previous active segment
//
do {
PreviousSubSegment = *((PHEAP_SUBSEGMENT volatile *)&LocalData->SegmentInfo[HeapBucket->SizeIndex].ActiveSubsegment);
} while ( InterlockedCompareExchangePointer( &LocalData->SegmentInfo[HeapBucket->SizeIndex].ActiveSubsegment, SubSegment, PreviousSubSegment) != PreviousSubSegment );
if ( PreviousSubSegment ) {
//
// We had a previous active segment. We need to unlock it, and if it has enough
// free space we'll mark it ready for reuse
//
LFHEAPASSERT( HeapBucket == PreviousSubSegment->Bucket ); LFHEAPASSERT( RtlpIsSubSegmentLocked(PreviousSubSegment, HEAP_PUBLIC_LOCK) ); LFHEAPASSERT( PreviousSubSegment->SFreeListEntry.Next == ((PSINGLE_LIST_ENTRY)(ULONG_PTR)0xEEEEEEEE) );
#ifdef _HEAP_DEBUG
PreviousSubSegment->SFreeListEntry.Next = 0; #endif
if (RtlpUnlockSubSegment(LocalData, PreviousSubSegment, HEAP_ACTIVE_LOCK | HEAP_PUBLIC_LOCK)) {
//
// That was not the last lock reference for that sub-segment.
//
if (PreviousSubSegment->AggregateExchg.Depth) {
RtlpInsertFreeSubSegment(LocalData, PreviousSubSegment); } } }
#ifdef _HEAP_DEBUG
LocalData->LowFragHeap->SegmentChange++; #endif
return TRUE; }
VOID FASTCALL RtlpSubSegmentInitialize ( IN PLFH_HEAP LowFragHeap, IN PHEAP_SUBSEGMENT SubSegment, IN PHEAP_USERDATA_HEADER UserBuffer, IN SIZE_T BlockSize, IN SIZE_T AllocatedSize, IN PVOID Bucket )
/*++
Routine Description:
The routine initialize a sub-segment descriptor. N.B. The Sub-Segment structure can be accessed simultanely by some other threads that captured it to allocate, but they were suspended before alloc completed. If meanwhile the sub-segment was deleted, the descriptor can be reused for a new subblock.
Arguments:
SubSegment - The sub-segment structure being initialized UserBuffer - The block allocated for the user data BlockSize - The size of each sub-block AllocatedSize - the size of the allocated buffer Bucket - The bucket that will own this heap sub-segment
Return Value:
None
--*/
{ ULONG i, NumItems; PVOID Buffer = UserBuffer + 1; PBLOCK_ENTRY BlockEntry; USHORT BlockUnits; USHORT CrtBlockOffset = 0; INTERLOCK_SEQ CapturedValue, NewValue;
CapturedValue.Exchg = SubSegment->AggregateExchg.Exchg;
//
// Add the block header overhead
//
BlockSize += sizeof(HEAP_ENTRY); BlockUnits = (USHORT)(BlockSize >> HEAP_GRANULARITY_SHIFT);
//
// The debug version will check the state for the subsegment
// testing whether the state was modified
//
LFHEAPASSERT(((PHEAP_BUCKET)Bucket)->BlockUnits == BlockUnits); LFHEAPASSERT(SubSegment->Lock == 0); LFHEAPASSERT(CapturedValue.OffsetAndDepth == (NO_MORE_ENTRIES << 16)); LFHEAPASSERT(SubSegment->UserBlocks == NULL); LFHEAPASSERT(SubSegment->SFreeListEntry.Next == 0);
//
// Initialize the user segment. Note that we don't touch the
// sub-segment descriptor, as some other threads can still use it.
//
UserBuffer->SubSegment = SubSegment; UserBuffer->HeapHandle = LowFragHeap->Heap; NumItems = (ULONG)((AllocatedSize - sizeof(HEAP_USERDATA_HEADER)) / BlockSize);
CrtBlockOffset = sizeof(HEAP_USERDATA_HEADER) >> HEAP_GRANULARITY_SHIFT; NewValue.FreeEntryOffset = CrtBlockOffset;
for (i = 0; i < NumItems; i++) {
BlockEntry = (PBLOCK_ENTRY) Buffer; //
// Initialize the block
//
BlockEntry->SubSegment = SubSegment; BlockEntry->SegmentIndex = HEAP_LFH_INDEX;
//
// Points to the next free block
//
CrtBlockOffset += BlockUnits; Buffer = (PCHAR)Buffer + BlockSize;
BlockEntry->LinkOffset = CrtBlockOffset; BlockEntry->Flags = HEAP_ENTRY_BUSY; BlockEntry->UnusedBytes = sizeof(HEAP_ENTRY); RtlpMarkLFHBlockFree( (PHEAP_ENTRY)BlockEntry );
#if defined(_WIN64)
BlockEntry->SubSegment = SubSegment; #endif
#ifdef _HEAP_DEBUG
BlockEntry->Reserved2 = 0xFEFE; #endif
}
//
// Mark the last block from the list
//
BlockEntry->LinkOffset = NO_MORE_ENTRIES; SubSegment->BlockSize = BlockUnits; SubSegment->BlockCount = (USHORT)NumItems; SubSegment->Bucket = Bucket; SubSegment->SizeIndex = ((PHEAP_BUCKET)Bucket)->SizeIndex;
//
// Determine the thresholds depending upon the total number of blocks
//
SubSegment->UserBlocks = UserBuffer; RtlpUpdateBucketCounters(Bucket, NumItems); NewValue.Depth = (USHORT)NumItems; NewValue.Sequence = CapturedValue.Sequence + 1; SubSegment->Lock = HEAP_USERDATA_LOCK; //
// At this point everything is set, so we can with an interlocked operation set the
// entire slist to the segment.
//
if (!LOCKCOMP64(&SubSegment->AggregateExchg.Exchg, NewValue, CapturedValue)) {
//
// Someone changed the state for the heap structure, so the
// initialization failed. We make noise in the debug version.
// (This should never happen)
//
LFHEAPASSERT( FALSE ); } }
VOID LOCALPROC RtlpFreeUserBuffer( PLFH_HEAP LowFragHeap, PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
When all blocks within segment are free we can go ahead and free the whole user buffer The caller should receive a notification from the last free call that the sub-segment does not have any allocated block
Arguments:
LowFragHeap - The LFH SubSegment - The sub-segment being released
Return Value:
None.
--*/
{ PHEAP_BUCKET HeapBucket; SIZE_T UserBlockSize;
HeapBucket = (PHEAP_BUCKET)SubSegment->Bucket; LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_USERDATA_LOCK) );
#ifdef _HEAP_DEBUG
UserBlockSize = HeapSize(LowFragHeap->Heap, 0, (PVOID)SubSegment->UserBlocks); LFHEAPASSERT((LONG_PTR)UserBlockSize > 0); #endif
SubSegment->UserBlocks->Signature = 0;
RtlpFreeUserBlock(LowFragHeap, (PHEAP_USERDATA_HEADER)SubSegment->UserBlocks);
//
// Update the counters
//
RtlpUpdateBucketCounters (HeapBucket, -SubSegment->BlockCount);
SubSegment->UserBlocks = NULL;
LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_USERDATA_LOCK) );
//
// This is a slow path any way. It doesn't harm a rare global interlocked
// in order to estimate the frequency of slower calls
//
InterlockedIncrement(&LowFragHeap->SegmentDelete); }
BOOLEAN FORCEINLINE RtlpLockSubSegment( PHEAP_SUBSEGMENT SubSegment, ULONG LockMask )
/*++
Routine Description:
The function locks a given set of bits to that segment, with a single atomic operation. If any of the bits is already locked the function will fail. If the sub-segment is deleted it will fail too.
Arguments:
SubSegment - The SubSegemnt being locked LockMask - An ULONG value specifying the bits needed locked
Return Value:
TRUE if succeeds.
--*/
{ ULONG CapturedLock; do {
CapturedLock = *((ULONG volatile *)&SubSegment->Lock);
if ((CapturedLock == 0) || (CapturedLock & LockMask)) {
return FALSE; }
} while ( InterlockedCompareExchange((PLONG)&SubSegment->Lock, CapturedLock | LockMask, CapturedLock) != CapturedLock );
return TRUE; }
BOOLEAN LOCALPROC RtlpUnlockSubSegment( PHEAP_LOCAL_DATA LocalData, PHEAP_SUBSEGMENT SubSegment, ULONG LockMask )
/*++
Routine Description:
The function unlocks the given sub-segment. If the last lock went away, the segment descriptor will be deleted and inserted into the recycle queue to be reused further for other allocations.
Arguments:
LowFragHeap - The LFH SubSegment - The sub-segment being unlocked LockMask - the bits that will be released
Return Value:
Returns false if unlocking the segment caused deletion. TRUE if there are still other locks keeping this sub-segment descriptor alive.
--*/
{ ULONG CapturedLock;
do {
CapturedLock = *((ULONG volatile *)&SubSegment->Lock);
//
// Unlock can be called exclusively, ONLY the if lock operation succeded
// It's an invalid state to have the segment already unlocked
// We assert this in debug version
//
LFHEAPASSERT((CapturedLock & LockMask) == LockMask);
} while ( InterlockedCompareExchange((PLONG)&SubSegment->Lock, CapturedLock & ~LockMask, CapturedLock) != CapturedLock );
//
// If That was the last lock released, we go ahead and
// free the sub-segment to the SLists
//
if (CapturedLock == LockMask) {
SubSegment->Bucket = NULL; SubSegment->AggregateExchg.Sequence += 1;
LFHEAPASSERT( RtlpIsSubSegmentEmpty(SubSegment) ); LFHEAPASSERT(SubSegment->Lock == 0); LFHEAPASSERT(SubSegment->SFreeListEntry.Next == 0); RtlpSubSegmentPush(&LocalData->DeletedSubSegments, &SubSegment->SFreeListEntry);
return FALSE; }
return TRUE; }
PVOID LOCALPROC RtlpSubSegmentAllocate ( PHEAP_BUCKET HeapBucket, PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
The function allocates a block from a sub-segment with a interlocked instruction. N.B. Since the access to this sub-segment is done unsynchronized, a tread can play with reading some the the sub-segment fields while another thread deleted it. For this reason the sub-segment descriptors are always allocated, so reading the interlocked counter is consistent over with the states that produced deletion. Every delete or init will increment the sequence counter, so the alloc will simple fail if it ends up using a different sub-segment. This function also handles the contention (interlocked operation failed) on this bucket. If we have too mani concurent access on this bucket, it will spin up the affinity limit on an MP machine.
Arguments:
HeapBucket - The bucket from which we allocate a blocks SubSegment - The subsegment currently in use
Return Value:
The allocated block pointer, if succeeds.
--*/
{ ULONGLONG CapturedValue, NewValue; PBLOCK_ENTRY BlockEntry; PHEAP_USERDATA_HEADER UserBlocks; SHORT Depth; LFH_DECLARE_COUNTER;
RETRY:
CapturedValue = SubSegment->AggregateExchg.Exchg; //
// We need the memory barrier because we are accessing
// another shared data below : UserBlocks
// This has to be fetched in the same order
// We declared these volatile, and on IA64 (MP) we need the
// memory barrier as well
//
RtlMemoryBarrier(); if ((Depth = (USHORT)CapturedValue) && (UserBlocks = (PHEAP_USERDATA_HEADER)SubSegment->UserBlocks) && (SubSegment->Bucket == HeapBucket) && !RtlpIsSubSegmentLocked(SubSegment, HEAP_CONVERT_LOCK)) {
BlockEntry = (PBLOCK_ENTRY)((PCHAR)UserBlocks + ((((ULONG)CapturedValue) >> 16) << HEAP_GRANULARITY_SHIFT));
//
// Accessing BlockEntry->LinkOffset can produce an AV if another thread freed the buffer
// meanwhile and the memory was decommitted. The caller of this function should
// have a try - except around this call. If the memory was used for other blocks
// the interlockedcompare should fail because the sequence number was incremented
//
LFHEAPASSERT(!(((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence) && (BlockEntry->LinkOffset != NO_MORE_ENTRIES) && (BlockEntry->LinkOffset > (SubSegment->BlockCount * SubSegment->BlockSize)) && ((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence)));
LFHEAPASSERT(!(((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence) && (BlockEntry->LinkOffset == NO_MORE_ENTRIES) && (Depth != 0) && ((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence)));
LFHEAPASSERT(!(((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence) && (SubSegment->Bucket != HeapBucket) && ((CapturedValue >> 16) == SubSegment->AggregateExchg.Sequence)));
NewValue = ((CapturedValue - 1) & (~(ULONGLONG)0xFFFF0000)) | ((ULONG)(BlockEntry->LinkOffset) << 16);
if (LOCKCOMP64(&SubSegment->AggregateExchg.Exchg, NewValue, CapturedValue)) {
//
// if the segment has been converted, the bucket will be invalid.
//
// LFHEAPASSERT(SubSegment->Bucket == HeapBucket);
// LFHEAPASSERT(RtlpIsSubSegmentLocked(SubSegment, HEAP_USERDATA_LOCK));
LFHEAPASSERT( !RtlpIsLFHBlockBusy( (PHEAP_ENTRY)BlockEntry ) );
LFHEAPASSERT(((NewValue >> 24) != NO_MORE_ENTRIES) || ((USHORT)NewValue == 0));
#ifdef _HEAP_DEBUG
LFHEAPASSERT((BlockEntry->Reserved2 == 0xFFFC) || (BlockEntry->Reserved2 == 0xFEFE));
//
// In the debug version write something there
//
BlockEntry->LinkOffset = 0xFFFA; BlockEntry->Reserved2 = 0xFFFB;
#endif
RtlpMarkLFHBlockBusy( (PHEAP_ENTRY)BlockEntry );
//
// If we had an interlocked compare failure, we must have another thread playing
// with the same subsegment at the same time. If this happens to often
// we need to increase the affinity limit on this bucket.
//
return ((PHEAP_ENTRY)BlockEntry + 1); }
} else {
return NULL; } if (!HeapBucket->UseAffinity) {
HeapBucket->UseAffinity = 1; } LFH_UPDATE_COUNTER;
goto RETRY;
return NULL; }
ULONG LOCALPROC RtlpSubSegmentFree ( PLFH_HEAP LowfHeap, PHEAP_SUBSEGMENT SubSegment, PBLOCK_ENTRY BlockEntry )
/*++
Routine Description:
This function frees a block from a sub-segment. Because a sub-segment lives as long as there is at least an allocated block inside, we don't have the problem we have for alloc. If the block being freed is happening to be the last one, we'll mark with an interlocked instruction the whole sub-segment as being free. The caller needs then to release the descriptor structure
Arguments:
SubSegment - The subsegment that ownd the block BlockEntry - The block being free.
Return Value:
TRUE if that was the last block, and it's safe now to recycle the descriptor. FALSE otherwise.
--*/
{ ULONGLONG CapturedValue, NewValue; ULONG ReturnStatus; ULONG_PTR UserBlocksRef = (ULONG_PTR)SubSegment->UserBlocks; LFH_DECLARE_COUNTER;
LFHEAPASSERT( RtlpIsLFHBlockBusy((PHEAP_ENTRY)BlockEntry) );
RtlpMarkLFHBlockFree((PHEAP_ENTRY)BlockEntry);
do { LFH_UPDATE_COUNTER; //
// We need to capture the sequence at the first step
// Then we'll capture the other fields from the segment
// If interlock operation below succeeds, means that none
// of the sub-segment fields (UserBlocks, Bucket ....)
// we changed. So the new state was built upon a consistent state
//
CapturedValue = SubSegment->AggregateExchg.Exchg; RtlMemoryBarrier();
NewValue = (CapturedValue + 0x100000001) & (~(ULONGLONG)0xFFFF0000); if (RtlpIsSubSegmentLocked(SubSegment, HEAP_CONVERT_LOCK) || !RtlpIsSubSegmentLocked(SubSegment, HEAP_USERDATA_LOCK) || (BlockEntry->SegmentIndex != HEAP_LFH_INDEX)) {
return HEAP_FREE_BLOCK_CONVERTED; }
//
// Depth and FreeEntryOffset are fetched at at the same time. They need
// to be consistent
//
LFHEAPASSERT(!(((USHORT)CapturedValue > 1) && (((ULONG)(NewValue >> 16)) == NO_MORE_ENTRIES)));
if ((((USHORT)NewValue) != SubSegment->BlockCount)) { ReturnStatus = HEAP_FREE_BLOCK_SUCCESS; BlockEntry->LinkOffset = (USHORT)(CapturedValue >> 16); NewValue |= ((((ULONG_PTR)BlockEntry - UserBlocksRef) >> HEAP_GRANULARITY_SHIFT) << 16); } else {
//
// This was the last block. Instead pushing it into the list
// we'll take the all blocks from the sub-segment to allow releasing the
// subsegment
//
ReturnStatus = HEAP_FREE_SEGMENT_EMPTY; NewValue = (NewValue & 0xFFFFFFFF00000000) | 0xFFFF0000; }
} while ( !LOCKCOMP64(&SubSegment->AggregateExchg.Exchg, NewValue, CapturedValue) );
if (!(USHORT)CapturedValue/*
&& !RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK)*/) {
RtlpInsertFreeSubSegment(&LowfHeap->LocalData[SubSegment->AffinityIndex], SubSegment); }
return ReturnStatus; }
PHEAP_BUCKET FORCEINLINE RtlpGetBucket( PLFH_HEAP LowFragHeap, SIZE_T Index )
/*++
Routine Description: The function simple returns the appropriate bucket for the given allocation index. The index should be < HEAP_BUCKETS_COUNT. This routine does not perform any range checking, it is supposed to be called with appropriate parameters
Arguments:
LowFragHeap - The LFH Index - The allocation index
Return Value:
The bucket that should be used for that allocation index.
--*/
{ return &LowFragHeap->Buckets[Index]; }
HANDLE FASTCALL RtlpCreateLowFragHeap( HANDLE Heap )
/*++
Routine Description:
The function creates a Low fragmentation heap, using for the allocations the heap handle passed in.
Arguments:
Heap - The NT heap handle
Return Value:
Returns a handle to a Lof Fragmentation Heap.
--*/
{ PLFH_HEAP LowFragHeap; ULONG i; PUCHAR Buffer;
SIZE_T TotalSize;
//
// Determine the size of the LFH structure based upon the current affinity limit.
//
TotalSize = sizeof(LFH_HEAP) + sizeof(HEAP_LOCAL_DATA) * RtlpHeapMaxAffinity;
LowFragHeap = HeapAlloc(Heap, HEAP_NO_CACHE_BLOCK, TotalSize);
if (LowFragHeap) {
memset(LowFragHeap, 0, TotalSize); RtlInitializeCriticalSection( &LowFragHeap->Lock ); //
// Initialize the heap zones.
//
InitializeListHead(&LowFragHeap->SubSegmentZones); LowFragHeap->ZoneBlockSize = ROUND_UP_TO_POWER2(sizeof(HEAP_SUBSEGMENT), HEAP_GRANULARITY);
LowFragHeap->Heap = Heap;
//
// Initialize the heap buckets
//
for (i = 0; i < HEAP_BUCKETS_COUNT; i++) {
LowFragHeap->Buckets[i].UseAffinity = 0; LowFragHeap->Buckets[i].SizeIndex = (UCHAR)i; LowFragHeap->Buckets[i].BlockUnits = (USHORT)(RtlpBucketBlockSizes[i] >> HEAP_GRANULARITY_SHIFT) + 1; } for (i = 0; i <= RtlpHeapMaxAffinity; i++) {
LowFragHeap->LocalData[i].LowFragHeap = LowFragHeap; LowFragHeap->LocalData[i].Affinity = i; } }
return LowFragHeap; }
VOID FASTCALL RtlpDestroyLowFragHeap( HANDLE LowFragHeapHandle )
/*++
Routine Description: The function should be called to destroy the LFH. This function should be called only when the NT heap goes away. We cannot rolback everything we've done with this heap. The NT heap is supposed to release all memory this heap allocated.
Arguments:
LowFragHeapHandle - The low fragmentation heap
Return Value:
None.
--*/
{ //
// This cannot be called unless the entire heap will go away
// It only delete the critical section, the all blocks allocated here will
// be deleted by RltDestroyHeap when it destroys the segments.
//
RtlDeleteCriticalSection(&((PLFH_HEAP)LowFragHeapHandle)->Lock); }
PVOID FASTCALL RtlpLowFragHeapAllocateFromZone( PLFH_HEAP LowFragHeap, ULONG Affinity )
/*++
Routine Description:
This function allocates a sub-segment descriptor structure from the heap zone
Arguments:
LowFragHeap - the LFH
Return Value:
The pointer to a new sub-segment descriptor structure.
--*/
{ PLFH_BLOCK_ZONE CrtZone;
RETRY_ALLOC:
CrtZone = LowFragHeap->LocalData[Affinity].CrtZone; if (CrtZone) {
PVOID CapturedFreePointer = CrtZone->FreePointer; PVOID NextFreePointer = (PCHAR)CapturedFreePointer + LowFragHeap->ZoneBlockSize;
//
// See if we have that sub-segment already preallocated
//
if (NextFreePointer < CrtZone->Limit) {
if ( InterlockedCompareExchangePointer( &CrtZone->FreePointer, NextFreePointer, CapturedFreePointer) == CapturedFreePointer) {
//
// The allocation succeeded, we can return that pointer
//
return CapturedFreePointer; }
goto RETRY_ALLOC; } } //
// we need to grow the heap zone. We acquire a lock here to avoid more threads doing the
// same thing
//
RtlEnterCriticalSection(&LowFragHeap->Lock);
//
// Test whether meanwhile another thread already increased the zone
//
if (CrtZone == LowFragHeap->LocalData[Affinity].CrtZone) {
CrtZone = HeapAlloc(LowFragHeap->Heap, HEAP_NO_CACHE_BLOCK, HEAP_DEFAULT_ZONE_SIZE);
if (CrtZone == NULL) {
RtlpFlushLFHeapCache(LowFragHeap); CrtZone = HeapAlloc(LowFragHeap->Heap, HEAP_NO_CACHE_BLOCK, HEAP_DEFAULT_ZONE_SIZE);
RtlLeaveCriticalSection(&LowFragHeap->Lock); return NULL; }
InsertTailList(&LowFragHeap->SubSegmentZones, &CrtZone->ListEntry);
CrtZone->Limit = (PCHAR)CrtZone + HEAP_DEFAULT_ZONE_SIZE; CrtZone->FreePointer = CrtZone + 1;
CrtZone->FreePointer = (PVOID)ROUND_UP_TO_POWER2((ULONG_PTR)CrtZone->FreePointer, HEAP_GRANULARITY);
//
// Everything is set. We can go ahead and set this as the default zone
//
LowFragHeap->LocalData[Affinity].CrtZone = CrtZone; } RtlLeaveCriticalSection(&LowFragHeap->Lock);
goto RETRY_ALLOC; }
SIZE_T FORCEINLINE RtlpSubSegmentGetIndex( SIZE_T BlockUnits )
/*++
Routine Description:
This routine converts the block size (in block units >> HEAP_GRANULARITY_SHIFT) into heap bucket index.
Arguments:
BlockUnits - the block size >> HEAP_GRANULARITY_SHIFT
Return Value:
The index for the bucket that should handle these sizes.
--*/
{ SIZE_T SizeClass; SIZE_T Bucket; if (BlockUnits <= 32) {
return BlockUnits - 1; }
SizeClass = 5; // Add 1 << 5 == 32
while (BlockUnits >> SizeClass) { SizeClass += 1; }
SizeClass -= 5;
BlockUnits = ROUND_UP_TO_POWER2(BlockUnits, (1 << SizeClass));
Bucket = ((SizeClass << 4) + (BlockUnits >> SizeClass) - 1); return Bucket; }
SIZE_T FORCEINLINE RtlpGetSubSegmentSizeIndex( PLFH_HEAP LowFragHeap, SIZE_T BlockSize, ULONG NumBlocks, CHAR AffinityCorrection )
/*++
Routine Description:
This function calculate the appropriate size for a sub-segment depending upon the block size and the minimal number of blocks that should be there.
Arguments:
BlockSize - The size of the block, in bytes NumBlocks - the minimal number of the blocks.
Return Value:
Returns the next power of 2 size that can satisfy the request
--*/
{ SIZE_T MinSize; ULONG SizeShift = HEAP_LOWEST_USER_SIZE_INDEX; SIZE_T ReturnSize;
LFHEAPASSERT(AffinityCorrection < HEAP_MIN_BLOCK_CLASS);
if (BlockSize < 256) {
AffinityCorrection -= 1; }
if (RtlpAffinityState.CrtLimit > (LONG)(RtlpHeapMaxAffinity >> 1)) {
AffinityCorrection += 1; }
if (NumBlocks < ((ULONG)1 << (HEAP_MIN_BLOCK_CLASS - AffinityCorrection))) {
NumBlocks = 1 << (HEAP_MIN_BLOCK_CLASS - AffinityCorrection); } if (LowFragHeap->Conversions) {
NumBlocks = HEAP_MIN_BLOCK_CLASS; } if (NumBlocks > (1 << HEAP_MAX_BLOCK_CLASS)) {
NumBlocks = 1 << HEAP_MAX_BLOCK_CLASS; }
MinSize = ((BlockSize + sizeof(HEAP_ENTRY) ) * NumBlocks) + sizeof(HEAP_USERDATA_HEADER) + sizeof(HEAP_ENTRY);
if (MinSize > HEAP_MAX_SUBSEGMENT_SIZE) {
MinSize = HEAP_MAX_SUBSEGMENT_SIZE; }
while (MinSize >> SizeShift) {
SizeShift += 1; }
if (SizeShift > HEAP_HIGHEST_USER_SIZE_INDEX) {
SizeShift = HEAP_HIGHEST_USER_SIZE_INDEX; } return SizeShift; }
PVOID FASTCALL RtlpLowFragHeapAlloc( HANDLE LowFragHeapHandle, SIZE_T BlockSize )
/*++
Routine Description:
This function allocates a block from the LFH.
Arguments:
Heap - the NT heap handle LowFragHeapHandle - The LFH heap handle BlockSize - the requested size, in bytes
Return Value:
A pointer to a new allocated block if succeeds. If the requested size is > 16K this function will fail too.
--*/
{ SIZE_T BlockUnits; SIZE_T Bucket; PLFH_HEAP LowFragHeap = (PLFH_HEAP)LowFragHeapHandle; PVOID Block; PHEAP_LOCAL_DATA LocalData;
//
// Get the appropriate bucket depending upon the requested size
//
BlockUnits = (BlockSize + HEAP_GRANULARITY - 1) >> HEAP_GRANULARITY_SHIFT; Bucket = RtlpSubSegmentGetIndex( BlockUnits );
if (Bucket < HEAP_BUCKETS_COUNT) {
PHEAP_BUCKET HeapBucket = RtlpGetBucket(LowFragHeap, Bucket); SIZE_T SubSegmentSize; SIZE_T SubSegmentSizeIndex; PHEAP_SUBSEGMENT SubSegment, NewSubSegment; PHEAP_USERDATA_HEADER UserData; PHEAP_LOCAL_SEGMENT_INFO SegmentInfo;
LocalData = &LowFragHeap->LocalData[ RtlpGetThreadAffinityIndex(HeapBucket) ]; SegmentInfo = &LocalData->SegmentInfo[Bucket];
//
// If we have some memory converted to the NT heap
// we need to allocate from NT heap first in order to reuse that memory
//
//
// Try first to allocate from the last segment used for free.
// This will provide a better performance because the data is likely to
// be still in the processor cache
//
if (SubSegment = SegmentInfo->Hint) {
//
// Accessing the user data can generate an exception if another thread freed
// the subsegment meanwhile.
//
LFHEAPASSERT( LocalData->Affinity == SubSegment->AffinityIndex );
__try {
Block = RtlpSubSegmentAllocate(HeapBucket, SubSegment);
} __except (EXCEPTION_EXECUTE_HANDLER) {
Block = NULL; }
if (Block) { RtlpSetUnusedBytes(LowFragHeap->Heap, ((PHEAP_ENTRY)Block - 1), ( ((SIZE_T)HeapBucket->BlockUnits) << HEAP_GRANULARITY_SHIFT) - BlockSize);
return Block; }
SegmentInfo->Hint = NULL; }
RETRY_ALLOC:
//
// Try to allocate from the current active sub-segment
//
if (SubSegment = SegmentInfo->ActiveSubsegment) {
//
// Accessing the user data can generate an exception if another thread freed
// the subsegment meanwhile.
//
LFHEAPASSERT( LocalData->Affinity == SubSegment->AffinityIndex );
__try {
Block = RtlpSubSegmentAllocate(HeapBucket, SubSegment);
} __except (EXCEPTION_EXECUTE_HANDLER) {
Block = NULL; }
if (Block) {
RtlpSetUnusedBytes(LowFragHeap->Heap, ((PHEAP_ENTRY)Block - 1), ( ((SIZE_T)HeapBucket->BlockUnits) << HEAP_GRANULARITY_SHIFT) - BlockSize);
return Block; } }
if (NewSubSegment = RtlpRemoveFreeSubSegment(LocalData, (LONG)Bucket)) { RtlpTrySetActiveSubSegment(LocalData, HeapBucket, NewSubSegment);
goto RETRY_ALLOC; } if (LowFragHeap->ConvertedSpace) {
InterlockedExchangeAdd((PLONG)&LowFragHeap->ConvertedSpace, -(LONG)(BlockSize >> 1));
if ((LONG)LowFragHeap->ConvertedSpace < 0) {
LowFragHeap->ConvertedSpace = 0; } return NULL; }
//
// At this point we don't have any sub-segment we can use to allocate this
// size. We need to create a new one.
//
SubSegmentSizeIndex = RtlpGetSubSegmentSizeIndex( LowFragHeap, RtlpBucketBlockSizes[Bucket], RtlpGetDesiredBlockNumber( HeapBucket->UseAffinity, HeapBucket->Counters.TotalBlocks), HeapBucket->UseAffinity );
UserData = RtlpAllocateUserBlock( LowFragHeap, (UCHAR)SubSegmentSizeIndex );
if (UserData == NULL) {
//
// Low memory condition. Flush the caches and retry the allocation.
// On low memory, the heap will use smaller sizes for sub-segments.
//
RtlpFlushLFHeapCache(LowFragHeap); SubSegmentSizeIndex = RtlpGetSubSegmentSizeIndex( LowFragHeap, RtlpBucketBlockSizes[Bucket], RtlpGetDesiredBlockNumber(HeapBucket->UseAffinity, HeapBucket->Counters.TotalBlocks), HeapBucket->UseAffinity );
UserData = RtlpAllocateUserBlock( LowFragHeap, (UCHAR)SubSegmentSizeIndex ); }
if (UserData) {
PVOID Entry;
SubSegmentSize = RtlpConvertSizeIndexToSize((UCHAR)UserData->SizeIndex);
LFHEAPASSERT( SubSegmentSize == HeapSize(LowFragHeap->Heap, 0, UserData) );
//
// This is a slow path any way, and it is exercised just in rare cases,
// when a bigger sub-segment is allocated. It doesn't hurt if we have an
// extra interlocked-increment.
//
InterlockedIncrement(&LowFragHeap->SegmentCreate); //
// Allocate a sub-segment descriptor structiure. If there isn't any in the
// recycle list we allocate one from the zones.
//
Entry = RtlpSubSegmentPop(&LocalData->DeletedSubSegments);
if (Entry == NULL) {
NewSubSegment = RtlpLowFragHeapAllocateFromZone(LowFragHeap, LocalData->Affinity);
#ifdef _HEAP_DEBUG
//
// We need to do some more extra initializations for
// the debug version, to verify the state of the subsegment
// in the next RtlpSubSegmentInitialize call
//
NewSubSegment->Lock = 0; NewSubSegment->AggregateExchg.OffsetAndDepth = NO_MORE_ENTRIES << 16; NewSubSegment->UserBlocks = NULL; #endif
} else { NewSubSegment = CONTAINING_RECORD(Entry, HEAP_SUBSEGMENT, SFreeListEntry); } if (NewSubSegment) { UserData->Signature = HEAP_LFH_USER_SIGNATURE; NewSubSegment->AffinityIndex = (UCHAR)LocalData->Affinity;
#ifdef _HEAP_DEBUG
//
// We need to do some more extra initializations for
// the debug version, to verify the state of the subsegment
// in the next RtlpSubSegmentInitialize call
//
NewSubSegment->SFreeListEntry.Next = 0; #endif
RtlpSubSegmentInitialize( LowFragHeap, NewSubSegment, UserData, RtlpBucketBlockSizes[Bucket], SubSegmentSize, HeapBucket );
//
// When the segment initialization was completed some other threads
// can access this subsegment (because they captured the pointer before
// if the subsegment was recycled).
// This can change the state for this segment, even it can delete.
// This should be very rare cases, so we'll print a message in
// debugger. However. If this happens too often it's an indication of
// a possible bug in LFH code, or a corruption.
//
LFHEAPWARN( NewSubSegment->Lock == HEAP_USERDATA_LOCK ); LFHEAPWARN( NewSubSegment->UserBlocks ); LFHEAPWARN( NewSubSegment->BlockSize == HeapBucket->BlockUnits ); if (!RtlpTrySetActiveSubSegment(LocalData, HeapBucket, NewSubSegment)) { RtlpInsertFreeSubSegment(LocalData, NewSubSegment); } goto RETRY_ALLOC;
} else {
HeapFree(LowFragHeap->Heap, 0, UserData); } } }
return NULL; }
BOOLEAN FASTCALL RtlpLowFragHeapFree( HANDLE LowFragHeapHandle, PVOID p )
/*++
Routine Description:
The function free a block allocated with RtlpLowFragHeapAlloc.
Arguments:
Heap - the NT heap handle LowFragHeapHandle - The LFH heap handle
Flags - Free flags p - The pointer to the block to be freed Return Value:
TRUE if succeeds.
--*/
{ PLFH_HEAP LowFragHeap = (PLFH_HEAP)LowFragHeapHandle; PBLOCK_ENTRY Block = (PBLOCK_ENTRY)((PHEAP_ENTRY)p - 1); PHEAP_SUBSEGMENT SubSegment; PHEAP_BUCKET HeapBucket; ULONG FreeStatus; SubSegment = Block->SubSegment;
RtlMemoryBarrier();
RETRY_FREE:
//
// Test whether the block belongs to the LFH
//
if (Block->SegmentIndex != HEAP_LFH_INDEX) {
if ( Block->SegmentIndex == HEAP_LFH_IN_CONVERSION ) {
//
// This should happen rarely, at high memory preasure.
// The subsegment was converted meanwhile from another thread
// We need to return FALSE and let the NT heap finish the free
//
// Acquire the heap lock and release it. This way we make sure that
// the conversion completed.
//
HeapLock(LowFragHeap->Heap); HeapUnlock(LowFragHeap->Heap); }
return FALSE; }
#ifdef _HEAP_DEBUG
Block->Reserved2 = 0xFFFC; #endif // _HEAP_DEBUG
//
// Free the block to the appropriate sub-segment
//
FreeStatus = RtlpSubSegmentFree(LowFragHeap, SubSegment, Block);
switch (FreeStatus) { case HEAP_FREE_SEGMENT_EMPTY: {
PHEAP_LOCAL_DATA LocalData = &LowFragHeap->LocalData[SubSegment->AffinityIndex];
//
// The free call above returned TRUE, meanning that the sub-segment can be deleted
// Remove it from the active state (to prevent other threads using it)
//
RtlpTrySetActiveSubSegment(LocalData, SubSegment->Bucket, NULL);
//
// Free thye user buffer
//
RtlpFreeUserBuffer(LowFragHeap, SubSegment);
//
// Unlock the sub-segment structure. This will actually recycle the descriptor
// if that was the last lock.
//
RtlpUnlockSubSegment(LocalData, SubSegment, HEAP_USERDATA_LOCK); }
break;
case HEAP_FREE_BLOCK_SUCCESS:
{ PHEAP_LOCAL_DATA LocalData = &LowFragHeap->LocalData[SubSegment->AffinityIndex];
LocalData->SegmentInfo[SubSegment->SizeIndex].Hint = SubSegment; }
break;
case HEAP_FREE_BLOCK_CONVERTED:
//
// In some rare cases the segment is locked for conversion, but the conversion
// process is abandoned for some reasons. In that case we need to retry
// freeing to the LFH
//
//
// Acquire the heap lock and release it. This way we make sure that
// the conversion completed.
//
HeapLock(LowFragHeap->Heap); HeapUnlock(LowFragHeap->Heap);
goto RETRY_FREE; }
return TRUE; }
BOOLEAN FASTCALL RtlpConvertSegmentToHeap ( PLFH_HEAP LowFragHeap, PHEAP_SUBSEGMENT SubSegment )
/*++
Routine Description:
The function converts a sub-segment to regular heap blocks This function should be invocked with the PUBLIC lock held for this subsegment
Arguments:
LowFragHeap - The Low Fragmentation Heap pointer SubSegment - The subsegment to be converted
Return Value:
Returns TRUE if succeeds
--*/
{ INTERLOCK_SEQ CapturedValue, NewValue; BOOLEAN SubSegmentEmpty; SHORT Depth; PHEAP_USERDATA_HEADER UserBlocks; PHEAP_ENTRY UserBlockHeapEntry; UCHAR SegmentIndex; SIZE_T FreeSize = 0; PVOID EndingFreeBlock = NULL; SIZE_T SegmentUnitsUsed;
LFHEAPASSERT(RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK));
//
// Since this function directly access the NT heap data
// we need first to lock the heap
//
if ( !HeapLock(LowFragHeap->Heap) ) {
return FALSE; }
//
// Mark the segment as being converted. This will block other
// threads to allocate/free from this segment further. These threads will
// make the NT call, which can succeeds only after the conversions is done
// (because the heap lock is held)
//
if (!RtlpLockSubSegment(SubSegment, HEAP_CONVERT_LOCK)) {
HeapUnlock(LowFragHeap->Heap); return FALSE; } SegmentUnitsUsed = SubSegment->BlockCount * SubSegment->BlockSize + ((sizeof( HEAP_ENTRY ) + sizeof(HEAP_USERDATA_HEADER)) >> HEAP_GRANULARITY_SHIFT); do {
//
// We need to capture the sequence at the first step
// Then we'll capture the other fields from the segment
// If interlock operation below succeeds, means that none
// of the sub-segment fields (UserBlocks, Bucket ....)
// we changed. So the new state was built upon a consistent state
//
CapturedValue.Sequence = SubSegment->AggregateExchg.Sequence; CapturedValue.OffsetAndDepth = SubSegment->AggregateExchg.OffsetAndDepth; Depth = (SHORT)CapturedValue.OffsetAndDepth; UserBlocks = (PHEAP_USERDATA_HEADER)SubSegment->UserBlocks;
//
// Test whether the sub-segment was not deleted meanwhile
//
if ((Depth == SubSegment->BlockCount) || (UserBlocks == NULL) || RtlpIsSubSegmentEmpty(SubSegment)) {
RtlpUnlockSubSegment(&LowFragHeap->LocalData[SubSegment->AffinityIndex], SubSegment, HEAP_CONVERT_LOCK); HeapUnlock(LowFragHeap->Heap); return FALSE; }
//
// Do not convert sub-segments which have a trailing block of
// one unit size.
//
UserBlockHeapEntry = (PHEAP_ENTRY)UserBlocks - 1;
if ( (UserBlockHeapEntry->Size - SegmentUnitsUsed) == 1) {
RtlpUnlockSubSegment(&LowFragHeap->LocalData[SubSegment->AffinityIndex], SubSegment, HEAP_CONVERT_LOCK); HeapUnlock(LowFragHeap->Heap);
return FALSE; }
//
// We have some blocks in this segment. We try to take them
// at ones and free to the heap like regular blocks
//
NewValue.Sequence = CapturedValue.Sequence + 1; NewValue.OffsetAndDepth = NO_MORE_ENTRIES << 16;
} while ( !LOCKCOMP64(&SubSegment->AggregateExchg.Exchg, NewValue, CapturedValue) );
//
// We suceeded to take all blocks from the s-list. From now on
// the convert cannot fail
//
InterlockedDecrement(&LowFragHeap->UserBlockCache.AvailableBlocks[ UserBlocks->SizeIndex - HEAP_LOWEST_USER_SIZE_INDEX ]); UserBlocks->SizeIndex = 0;
if (InterlockedIncrement(&(((PHEAP_BUCKET)SubSegment->Bucket))->Conversions) == 1) { InterlockedIncrement(&LowFragHeap->Conversions); }
SegmentIndex = UserBlockHeapEntry->SegmentIndex;
//
// Convert the heap block headers to NT heap style
//
{ LONG i; LONG TotalSize; PVOID Buffer = UserBlocks + 1; PHEAP_ENTRY BlockEntry; PHEAP_ENTRY PreviousBlockEntry = (PHEAP_ENTRY)UserBlocks - 1; PHEAP_ENTRY NextBlock = NULL; TotalSize = PreviousBlockEntry->Size;
InterlockedExchangeAdd(&LowFragHeap->ConvertedSpace, (TotalSize << HEAP_GRANULARITY_SHIFT));
if (!(PreviousBlockEntry->Flags & HEAP_ENTRY_LAST_ENTRY)) {
NextBlock = PreviousBlockEntry + PreviousBlockEntry->Size; }
PreviousBlockEntry->Size = (sizeof(HEAP_ENTRY) + sizeof(HEAP_USERDATA_HEADER)) >> HEAP_GRANULARITY_SHIFT; PreviousBlockEntry->UnusedBytes = sizeof(HEAP_ENTRY);
TotalSize -= PreviousBlockEntry->Size;
//
// Walk the heap sub-segment, and set the attributes for each block
// according with the NT heap rules
//
for (i = 0; i < SubSegment->BlockCount; i++) {
HEAP_ENTRY TempEntry;
BlockEntry = (PHEAP_ENTRY) Buffer; //
// Initialize the block
//
BlockEntry->SegmentIndex = HEAP_LFH_IN_CONVERSION;
BlockEntry->Size = *((volatile USHORT *)&SubSegment->BlockSize); BlockEntry->PreviousSize = *((volatile USHORT *)&PreviousBlockEntry->Size); //
// Restore the index since we are done with initialization
//
BlockEntry->SegmentIndex = SegmentIndex; TotalSize -= SubSegment->BlockSize;
//
// Points to the next free block
//
Buffer = (PCHAR)Buffer + ((SIZE_T)SubSegment->BlockSize << HEAP_GRANULARITY_SHIFT);
PreviousBlockEntry = BlockEntry;
#if defined(_WIN64)
BlockEntry->SubSegment = NULL; #endif
}
LFHEAPASSERT(TotalSize >= 0);
//
// The last block into the segment is a special case. In general it can be smaller than
// other blocks. If the size is less than 2 heap units
// (a block header + a small minimal block) we attach it to the last block
// Otherwise, we create a separate block and initialize it properly
//
if (TotalSize >= 2) { // TotalSize in heap units
BlockEntry = BlockEntry + BlockEntry->Size; //
// Initialize the block
//
BlockEntry->SegmentIndex = SegmentIndex; BlockEntry->PreviousSize = PreviousBlockEntry->Size; BlockEntry->Flags = HEAP_ENTRY_BUSY;
BlockEntry->Size = (USHORT)TotalSize;
EndingFreeBlock = BlockEntry + 1; }
//
// If we have a next block, we need to correct the PreviousSize field
//
if (NextBlock) {
NextBlock->PreviousSize = BlockEntry->Size;
} else { //
// Transfer the LAST_ENTRY flag, if any, to the last block
//
BlockEntry->Flags |= HEAP_ENTRY_LAST_ENTRY; UserBlockHeapEntry->Flags &= ~HEAP_ENTRY_LAST_ENTRY;
RtlpUpdateLastEntry((PHEAP)LowFragHeap->Heap, BlockEntry); } }
HeapUnlock(LowFragHeap->Heap);
//
// Now free all blocks from the captured s-list to the NT heap
// We use HEAP_NO_SERIALIZE flag because we're already holding the lock
//
if (Depth){
ULONG CrtOffset = CapturedValue.FreeEntryOffset;
while (CrtOffset != NO_MORE_ENTRIES) {
PBLOCK_ENTRY BlockEntry; PHEAP_ENTRY PreviousBlockEntry;
BlockEntry = (PBLOCK_ENTRY)((PCHAR)UserBlocks + (CrtOffset << HEAP_GRANULARITY_SHIFT));
CrtOffset = BlockEntry->LinkOffset;
FreeSize += BlockEntry->Size;
HeapFree(LowFragHeap->Heap, 0, (PVOID)((PHEAP_ENTRY)BlockEntry + 1)); } }
//
// Free also the last block, if any.
//
if (EndingFreeBlock) { HeapFree(LowFragHeap->Heap, 0, EndingFreeBlock); }
LFHEAPASSERT(HeapValidate(LowFragHeap->Heap, 0, 0));
RtlpFreeUserBuffer(LowFragHeap, SubSegment);
RtlpUnlockSubSegment(&LowFragHeap->LocalData[SubSegment->AffinityIndex], SubSegment, HEAP_CONVERT_LOCK | HEAP_PUBLIC_LOCK | HEAP_USERDATA_LOCK); return TRUE; }
ULONG RtlpConvertHeapBucket ( PLFH_HEAP LowFragHeap, ULONG SizeIndex )
/*++
Routine Description:
The function converts the segments from a given bucket index to regular heap blocks
N.B. This is a operation can take a long time to complete. The heap is calling this if the allocation of a new segment fails.
Arguments:
LowFragHeap - The Low Fragmentation Heap pointer SizeIndex - Bucket index (in range 0 - 127)
Return Value:
The number of segments converted
--*/
{ PHEAP_BUCKET HeapBucket = RtlpGetBucket(LowFragHeap, SizeIndex); ULONG Affinity; ULONG TotalConverted = 0; SINGLE_LIST_ENTRY Non_ConvertedList; PVOID Entry;
Non_ConvertedList.Next = NULL;
for (Affinity = 0; Affinity < RtlpHeapMaxAffinity; Affinity++) {
PHEAP_LOCAL_SEGMENT_INFO FreeSList = &LowFragHeap->LocalData[ Affinity ].SegmentInfo[ SizeIndex ];
while (Entry = RtlpSubSegmentPop(&FreeSList->SListHeader) ) {
PHEAP_SUBSEGMENT SubSegment;
SubSegment = CONTAINING_RECORD(Entry, HEAP_SUBSEGMENT, SFreeListEntry);
#ifdef _HEAP_DEBUG
SubSegment->SFreeListEntry.Next = NULL; #endif
LFHEAPASSERT( RtlpIsSubSegmentLocked(SubSegment, HEAP_PUBLIC_LOCK) ); LFHEAPASSERT( SizeIndex == SubSegment->SizeIndex );
//
// While we still hold the subsegment public lock, we try to convert
// it to regular NT blocks
//
if (RtlpConvertSegmentToHeap( LowFragHeap, SubSegment )) {
TotalConverted += 1;
if (LowFragHeap->ConvertedSpace > HEAP_CONVERT_LIMIT) {
while (TRUE) {
PHEAP_SUBSEGMENT xSubSegment;
Entry = PopEntryList(&Non_ConvertedList);
if (Entry == NULL) {
break; }
xSubSegment = CONTAINING_RECORD(Entry, HEAP_SUBSEGMENT, SFreeListEntry);
RtlpSubSegmentPush(&FreeSList->SListHeader, &xSubSegment->SFreeListEntry);
} return TotalConverted; }
} else {
PushEntryList(&Non_ConvertedList, &SubSegment->SFreeListEntry); } }
while (TRUE) {
PHEAP_SUBSEGMENT SubSegment;
Entry = PopEntryList(&Non_ConvertedList);
if (Entry == NULL) {
break; }
SubSegment = CONTAINING_RECORD(Entry, HEAP_SUBSEGMENT, SFreeListEntry);
RtlpSubSegmentPush(&FreeSList->SListHeader, &SubSegment->SFreeListEntry);
}
if (!HeapBucket->UseAffinity) {
break; } }
return TotalConverted; }
ULONG RtlpFlushLFHeapCache ( PLFH_HEAP LowFragHeap )
/*++
Routine Description:
The function converts the segments from the free lists to regular heap blocks
N.B. This is a operation can take a long time to complete. The heap is calling this if the allocation of a new segment fails.
Arguments:
LowFragHeap - The Low Fragmentation Heap pointer
Return Value:
The number of segments converted
--*/
{ LONG SizeIndex; ULONG TotalSegments = 0;
//
// Convert to regular heap blocks starting with the upper buckets (with large segments)
//
for (SizeIndex = HEAP_BUCKETS_COUNT - 1; SizeIndex >= 0; SizeIndex--) {
TotalSegments += RtlpConvertHeapBucket(LowFragHeap, SizeIndex); if (LowFragHeap->ConvertedSpace > HEAP_CONVERT_LIMIT) { break; } }
#ifdef _HEAP_DEBUG
HeapValidate(LowFragHeap->Heap, 0, 0);
#endif // _HEAP_DEBUG
return TotalSegments; }
SIZE_T FASTCALL RtlpLowFragHeapGetBlockSize( HANDLE HeapHandle, ULONG Flags, PVOID p )
/*++
Routine Description:
The function returns the size of a LFH block
Arguments:
HeapHandle - The handle of the heap (not used yet) Flags - Not used yet p - The pointer being queried
Return Value:
The size of that block.
--*/
{ PBLOCK_ENTRY Block = (PBLOCK_ENTRY)((PHEAP_ENTRY)p - 1); PHEAP_SUBSEGMENT SubSegment = (PHEAP_SUBSEGMENT)Block->SubSegment;
//
// Test whether the block belongs to LFH. We need to capture the
// subsegment before to protect against segment conversions
//
if (Block->SegmentIndex == HEAP_LFH_INDEX) {
return (((SIZE_T)SubSegment->BlockSize) << HEAP_GRANULARITY_SHIFT) - sizeof(HEAP_ENTRY); }
return 0; }
VOID RtlpInitializeLowFragHeapManager()
/*++
Routine Description:
This function initialize the global variables for the low fragmention heap manager.
Arguments:
Return Value:
--*/
{ SIZE_T Granularity = HEAP_GRANULARITY; ULONG i; SIZE_T PreviousSize = 0; SYSTEM_BASIC_INFORMATION SystemInformation; //
// prevent the second initialization
//
if (RtlpHeapMaxAffinity) {
return; }
#ifdef _HEAP_DEBUG
PrintMsg("Debug version\n"); #endif
//
// Query the number of processors
//
if (NT_SUCCESS(NtQuerySystemInformation (SystemBasicInformation, &SystemInformation, sizeof(SystemInformation), NULL))) {
ULONG Shift = 0;
RtlpHeapMaxAffinity = SystemInformation.NumberOfProcessors;
if (RtlpHeapMaxAffinity > 1) {
RtlpHeapMaxAffinity = (RtlpHeapMaxAffinity << 1); }
if (RtlpHeapMaxAffinity > HEAP_AFFINITY_LIMIT) {
RtlpHeapMaxAffinity = HEAP_AFFINITY_LIMIT; } } else {
PrintMsg("NtQuerySystemInformation failed\n");
RtlpHeapMaxAffinity = 1; }
#ifdef _HEAP_DEBUG
if (RtlpHeapMaxAffinity > 1) {
PrintMsg("Affinity enabled at %ld\n", RtlpHeapMaxAffinity); }
#endif
RtlpInitializeAffinityManager( (UCHAR)RtlpHeapMaxAffinity );
//
// Generate the Bucket size table
//
for (i = 0; i < 32; i++) {
PreviousSize = RtlpBucketBlockSizes[i] = PreviousSize + Granularity; } for (i = 32; i < HEAP_BUCKETS_COUNT; i++) {
if ((i % 16) == 0) {
Granularity <<= 1; }
PreviousSize = RtlpBucketBlockSizes[i] = PreviousSize + Granularity; } }
|