github.com/pidato/unsafe@v0.1.4/memory/rpmalloc/rpmalloc.c (about)

     1  /* rpmalloc.c  -  Memory allocator  -  Public Domain  -  2016-2020 Mattias Jansson
     2   *
     3   * This library provides a cross-platform lock free thread caching malloc implementation in C11.
     4   * The latest source code is always available at
     5   *
     6   * https://github.com/mjansson/rpmalloc
     7   *
     8   * This library is put in the public domain; you can redistribute it and/or modify it without any restrictions.
     9   *
    10   */
    11  
    12  #include "rpmalloc.h"
    13  
    14  ////////////
    15  ///
    16  /// Build time configurable limits
    17  ///
    18  //////
    19  
    20  #if defined(__clang__)
    21  #pragma clang diagnostic ignored "-Wunused-macros"
    22  #pragma clang diagnostic ignored "-Wunused-function"
    23  #if __has_warning("-Wreserved-identifier")
    24  #pragma clang diagnostic ignored "-Wreserved-identifier"
    25  #endif
    26  #if __has_warning("-Wstatic-in-inline")
    27  #pragma clang diagnostic ignored "-Wstatic-in-inline"
    28  #endif
    29  #elif defined(__GNUC__)
    30  #pragma GCC diagnostic ignored "-Wunused-macros"
    31  #pragma GCC diagnostic ignored "-Wunused-function"
    32  #endif
    33  
    34  #ifndef HEAP_ARRAY_SIZE
    35  //! Size of heap hashmap
    36  #define HEAP_ARRAY_SIZE           47
    37  #endif
    38  #ifndef ENABLE_THREAD_CACHE
    39  //! Enable per-thread cache
    40  #define ENABLE_THREAD_CACHE       1
    41  #endif
    42  #ifndef ENABLE_GLOBAL_CACHE
    43  //! Enable global cache shared between all threads, requires thread cache
    44  #define ENABLE_GLOBAL_CACHE       1
    45  #endif
    46  #ifndef ENABLE_VALIDATE_ARGS
    47  //! Enable validation of args to public entry points
    48  #define ENABLE_VALIDATE_ARGS      0
    49  #endif
    50  #ifndef ENABLE_STATISTICS
    51  //! Enable statistics collection
    52  #define ENABLE_STATISTICS         0
    53  #endif
    54  #ifndef ENABLE_ASSERTS
    55  //! Enable asserts
    56  #define ENABLE_ASSERTS            0
    57  #endif
    58  #ifndef ENABLE_OVERRIDE
    59  //! Override standard library malloc/free and new/delete entry points
    60  #define ENABLE_OVERRIDE           1
    61  #endif
    62  #ifndef ENABLE_PRELOAD
    63  //! Support preloading
    64  #define ENABLE_PRELOAD            1
    65  #endif
    66  #ifndef DISABLE_UNMAP
    67  //! Disable unmapping memory pages (also enables unlimited cache)
    68  #define DISABLE_UNMAP             0
    69  #endif
    70  #ifndef ENABLE_UNLIMITED_CACHE
    71  //! Enable unlimited global cache (no unmapping until finalization)
    72  #define ENABLE_UNLIMITED_CACHE    0
    73  #endif
    74  #ifndef ENABLE_ADAPTIVE_THREAD_CACHE
    75  //! Enable adaptive thread cache size based on use heuristics
    76  #define ENABLE_ADAPTIVE_THREAD_CACHE 1
    77  #endif
    78  #ifndef DEFAULT_SPAN_MAP_COUNT
    79  //! Default number of spans to map in call to map more virtual memory (default values yield 4MiB here)
    80  #define DEFAULT_SPAN_MAP_COUNT    64
    81  #endif
    82  #ifndef GLOBAL_CACHE_MULTIPLIER
    83  //! Multiplier for global cache
    84  #define GLOBAL_CACHE_MULTIPLIER   8
    85  #endif
    86  
    87  #if DISABLE_UNMAP && !ENABLE_GLOBAL_CACHE
    88  #error Must use global cache if unmap is disabled
    89  #endif
    90  
    91  #if DISABLE_UNMAP
    92  #undef ENABLE_UNLIMITED_CACHE
    93  #define ENABLE_UNLIMITED_CACHE 1
    94  #endif
    95  
    96  #if !ENABLE_GLOBAL_CACHE
    97  #undef ENABLE_UNLIMITED_CACHE
    98  #define ENABLE_UNLIMITED_CACHE 0
    99  #endif
   100  
   101  #if !ENABLE_THREAD_CACHE
   102  #undef ENABLE_ADAPTIVE_THREAD_CACHE
   103  #define ENABLE_ADAPTIVE_THREAD_CACHE 0
   104  #endif
   105  
   106  #if defined(_WIN32) || defined(__WIN32__) || defined(_WIN64)
   107  #  define PLATFORM_WINDOWS 1
   108  #  define PLATFORM_POSIX 0
   109  #else
   110  #  define PLATFORM_WINDOWS 0
   111  #  define PLATFORM_POSIX 1
   112  #endif
   113  
   114  /// Platform and arch specifics
   115  #if defined(_MSC_VER) && !defined(__clang__)
   116  #  pragma warning (disable: 5105)
   117  #  ifndef FORCEINLINE
   118  #    define FORCEINLINE inline __forceinline
   119  #  endif
   120  #  define _Static_assert static_assert
   121  #else
   122  #  ifndef FORCEINLINE
   123  #    define FORCEINLINE inline __attribute__((__always_inline__))
   124  #  endif
   125  #endif
   126  #if PLATFORM_WINDOWS
   127  #  ifndef WIN32_LEAN_AND_MEAN
   128  #    define WIN32_LEAN_AND_MEAN
   129  #  endif
   130  #  include <windows.h>
   131  #  if ENABLE_VALIDATE_ARGS
   132  #    include <intsafe.h>
   133  #  endif
   134  #else
   135  #  include <unistd.h>
   136  #  include <stdio.h>
   137  #  include <stdlib.h>
   138  #  include <time.h>
   139  #  if defined(__linux__) || defined(__ANDROID__)
   140  #    include <sys/prctl.h>
   141  #    if !defined(PR_SET_VMA)
   142  #      define PR_SET_VMA 0x53564d41
   143  #      define PR_SET_VMA_ANON_NAME 0
   144  #    endif
   145  #  endif
   146  #  if defined(__APPLE__)
   147  #    include <TargetConditionals.h>
   148  #    if !TARGET_OS_IPHONE && !TARGET_OS_SIMULATOR
   149  #    include <mach/mach_vm.h>
   150  #    include <mach/vm_statistics.h>
   151  #    endif
   152  #    include <pthread.h>
   153  #  endif
   154  #  if defined(__HAIKU__) || defined(__TINYC__)
   155  #    include <pthread.h>
   156  #  endif
   157  #endif
   158  
   159  #include <stdint.h>
   160  #include <string.h>
   161  #include <errno.h>
   162  
   163  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
   164  #include <fibersapi.h>
   165  static DWORD fls_key;
   166  #endif
   167  
   168  #if PLATFORM_POSIX
   169  #  include <sys/mman.h>
   170  #  include <sched.h>
   171  #  ifdef __FreeBSD__
   172  #    include <sys/sysctl.h>
   173  #    define MAP_HUGETLB MAP_ALIGNED_SUPER
   174  #    ifndef PROT_MAX
   175  #      define PROT_MAX(f) 0
   176  #    endif
   177  #  else
   178  #    define PROT_MAX(f) 0
   179  #  endif
   180  #  ifdef __sun
   181  extern int madvise(caddr_t, size_t, int);
   182  #  endif
   183  #  ifndef MAP_UNINITIALIZED
   184  #    define MAP_UNINITIALIZED 0
   185  #  endif
   186  #endif
   187  #include <errno.h>
   188  
   189  #if ENABLE_ASSERTS
   190  #  undef NDEBUG
   191  #  if defined(_MSC_VER) && !defined(_DEBUG)
   192  #    define _DEBUG
   193  #  endif
   194  #  include <assert.h>
   195  #define RPMALLOC_TOSTRING_M(x) #x
   196  #define RPMALLOC_TOSTRING(x) RPMALLOC_TOSTRING_M(x)
   197  #define rpmalloc_assert(truth, message)                                                                      \
   198  	do {                                                                                                     \
   199  		if (!(truth)) {                                                                                      \
   200  			if (_memory_config.error_callback) {                                                             \
   201  				_memory_config.error_callback(                                                               \
   202  				    message " (" RPMALLOC_TOSTRING(truth) ") at " __FILE__ ":" RPMALLOC_TOSTRING(__LINE__)); \
   203  			} else {                                                                                         \
   204  				assert((truth) && message);                                                                  \
   205  			}                                                                                                \
   206  		}                                                                                                    \
   207  	} while (0)
   208  #else
   209  #  define rpmalloc_assert(truth, message) do {} while(0)
   210  #endif
   211  #if ENABLE_STATISTICS
   212  #  include <stdio.h>
   213  #endif
   214  
   215  //////
   216  ///
   217  /// Atomic access abstraction (since MSVC does not do C11 yet)
   218  ///
   219  //////
   220  
   221  #if defined(_MSC_VER) && !defined(__clang__)
   222  
   223  typedef volatile long      atomic32_t;
   224  typedef volatile long long atomic64_t;
   225  typedef volatile void*     atomicptr_t;
   226  
   227  static FORCEINLINE int32_t atomic_load32(atomic32_t* src) { return *src; }
   228  static FORCEINLINE void    atomic_store32(atomic32_t* dst, int32_t val) { *dst = val; }
   229  static FORCEINLINE int32_t atomic_incr32(atomic32_t* val) { return (int32_t)InterlockedIncrement(val); }
   230  static FORCEINLINE int32_t atomic_decr32(atomic32_t* val) { return (int32_t)InterlockedDecrement(val); }
   231  static FORCEINLINE int32_t atomic_add32(atomic32_t* val, int32_t add) { return (int32_t)InterlockedExchangeAdd(val, add) + add; }
   232  static FORCEINLINE int     atomic_cas32_acquire(atomic32_t* dst, int32_t val, int32_t ref) { return (InterlockedCompareExchange(dst, val, ref) == ref) ? 1 : 0; }
   233  static FORCEINLINE void    atomic_store32_release(atomic32_t* dst, int32_t val) { *dst = val; }
   234  static FORCEINLINE int64_t atomic_load64(atomic64_t* src) { return *src; }
   235  static FORCEINLINE int64_t atomic_add64(atomic64_t* val, int64_t add) { return (int64_t)InterlockedExchangeAdd64(val, add) + add; }
   236  static FORCEINLINE void*   atomic_load_ptr(atomicptr_t* src) { return (void*)*src; }
   237  static FORCEINLINE void    atomic_store_ptr(atomicptr_t* dst, void* val) { *dst = val; }
   238  static FORCEINLINE void    atomic_store_ptr_release(atomicptr_t* dst, void* val) { *dst = val; }
   239  static FORCEINLINE void*   atomic_exchange_ptr_acquire(atomicptr_t* dst, void* val) { return (void*)InterlockedExchangePointer((void* volatile*)dst, val); }
   240  static FORCEINLINE int     atomic_cas_ptr(atomicptr_t* dst, void* val, void* ref) { return (InterlockedCompareExchangePointer((void* volatile*)dst, val, ref) == ref) ? 1 : 0; }
   241  
   242  #define EXPECTED(x) (x)
   243  #define UNEXPECTED(x) (x)
   244  
   245  #else
   246  
   247  #include <stdatomic.h>
   248  
   249  typedef volatile _Atomic(int32_t) atomic32_t;
   250  typedef volatile _Atomic(int64_t) atomic64_t;
   251  typedef volatile _Atomic(void*) atomicptr_t;
   252  
   253  static FORCEINLINE int32_t atomic_load32(atomic32_t* src) { return atomic_load_explicit(src, memory_order_relaxed); }
   254  static FORCEINLINE void    atomic_store32(atomic32_t* dst, int32_t val) { atomic_store_explicit(dst, val, memory_order_relaxed); }
   255  static FORCEINLINE int32_t atomic_incr32(atomic32_t* val) { return atomic_fetch_add_explicit(val, 1, memory_order_relaxed) + 1; }
   256  static FORCEINLINE int32_t atomic_decr32(atomic32_t* val) { return atomic_fetch_add_explicit(val, -1, memory_order_relaxed) - 1; }
   257  static FORCEINLINE int32_t atomic_add32(atomic32_t* val, int32_t add) { return atomic_fetch_add_explicit(val, add, memory_order_relaxed) + add; }
   258  static FORCEINLINE int     atomic_cas32_acquire(atomic32_t* dst, int32_t val, int32_t ref) { return atomic_compare_exchange_weak_explicit(dst, &ref, val, memory_order_acquire, memory_order_relaxed); }
   259  static FORCEINLINE void    atomic_store32_release(atomic32_t* dst, int32_t val) { atomic_store_explicit(dst, val, memory_order_release); }
   260  static FORCEINLINE int64_t atomic_load64(atomic64_t* val) { return atomic_load_explicit(val, memory_order_relaxed); }
   261  static FORCEINLINE int64_t atomic_add64(atomic64_t* val, int64_t add) { return atomic_fetch_add_explicit(val, add, memory_order_relaxed) + add; }
   262  static FORCEINLINE void*   atomic_load_ptr(atomicptr_t* src) { return atomic_load_explicit(src, memory_order_relaxed); }
   263  static FORCEINLINE void    atomic_store_ptr(atomicptr_t* dst, void* val) { atomic_store_explicit(dst, val, memory_order_relaxed); }
   264  static FORCEINLINE void    atomic_store_ptr_release(atomicptr_t* dst, void* val) { atomic_store_explicit(dst, val, memory_order_release); }
   265  static FORCEINLINE void*   atomic_exchange_ptr_acquire(atomicptr_t* dst, void* val) { return atomic_exchange_explicit(dst, val, memory_order_acquire); }
   266  static FORCEINLINE int     atomic_cas_ptr(atomicptr_t* dst, void* val, void* ref) { return atomic_compare_exchange_weak_explicit(dst, &ref, val, memory_order_relaxed, memory_order_relaxed); }
   267  
   268  #define EXPECTED(x) __builtin_expect((x), 1)
   269  #define UNEXPECTED(x) __builtin_expect((x), 0)
   270  
   271  #endif
   272  
   273  ////////////
   274  ///
   275  /// Statistics related functions (evaluate to nothing when statistics not enabled)
   276  ///
   277  //////
   278  
   279  #if ENABLE_STATISTICS
   280  #  define _rpmalloc_stat_inc(counter) atomic_incr32(counter)
   281  #  define _rpmalloc_stat_dec(counter) atomic_decr32(counter)
   282  #  define _rpmalloc_stat_add(counter, value) atomic_add32(counter, (int32_t)(value))
   283  #  define _rpmalloc_stat_add64(counter, value) atomic_add64(counter, (int64_t)(value))
   284  #  define _rpmalloc_stat_add_peak(counter, value, peak) do { int32_t _cur_count = atomic_add32(counter, (int32_t)(value)); if (_cur_count > (peak)) peak = _cur_count; } while (0)
   285  #  define _rpmalloc_stat_sub(counter, value) atomic_add32(counter, -(int32_t)(value))
   286  #  define _rpmalloc_stat_inc_alloc(heap, class_idx) do { \
   287  	int32_t alloc_current = atomic_incr32(&heap->size_class_use[class_idx].alloc_current); \
   288  	if (alloc_current > heap->size_class_use[class_idx].alloc_peak) \
   289  		heap->size_class_use[class_idx].alloc_peak = alloc_current; \
   290  	atomic_incr32(&heap->size_class_use[class_idx].alloc_total); \
   291  } while(0)
   292  #  define _rpmalloc_stat_inc_free(heap, class_idx) do { \
   293  	atomic_decr32(&heap->size_class_use[class_idx].alloc_current); \
   294  	atomic_incr32(&heap->size_class_use[class_idx].free_total); \
   295  } while(0)
   296  #else
   297  #  define _rpmalloc_stat_inc(counter) do {} while(0)
   298  #  define _rpmalloc_stat_dec(counter) do {} while(0)
   299  #  define _rpmalloc_stat_add(counter, value) do {} while(0)
   300  #  define _rpmalloc_stat_add64(counter, value) do {} while(0)
   301  #  define _rpmalloc_stat_add_peak(counter, value, peak) do {} while (0)
   302  #  define _rpmalloc_stat_sub(counter, value) do {} while(0)
   303  #  define _rpmalloc_stat_inc_alloc(heap, class_idx) do {} while(0)
   304  #  define _rpmalloc_stat_inc_free(heap, class_idx) do {} while(0)
   305  #endif
   306  
   307  
   308  ///
   309  /// Preconfigured limits and sizes
   310  ///
   311  
   312  //! Granularity of a small allocation block (must be power of two)
   313  #define SMALL_GRANULARITY         16
   314  //! Small granularity shift count
   315  #define SMALL_GRANULARITY_SHIFT   4
   316  //! Number of small block size classes
   317  #define SMALL_CLASS_COUNT         65
   318  //! Maximum size of a small block
   319  #define SMALL_SIZE_LIMIT          (SMALL_GRANULARITY * (SMALL_CLASS_COUNT - 1))
   320  //! Granularity of a medium allocation block
   321  #define MEDIUM_GRANULARITY        512
   322  //! Medium granularity shift count
   323  #define MEDIUM_GRANULARITY_SHIFT  9
   324  //! Number of medium block size classes
   325  #define MEDIUM_CLASS_COUNT        61
   326  //! Total number of small + medium size classes
   327  #define SIZE_CLASS_COUNT          (SMALL_CLASS_COUNT + MEDIUM_CLASS_COUNT)
   328  //! Number of large block size classes
   329  #define LARGE_CLASS_COUNT         63
   330  //! Maximum size of a medium block
   331  #define MEDIUM_SIZE_LIMIT         (SMALL_SIZE_LIMIT + (MEDIUM_GRANULARITY * MEDIUM_CLASS_COUNT))
   332  //! Maximum size of a large block
   333  #define LARGE_SIZE_LIMIT          ((LARGE_CLASS_COUNT * _memory_span_size) - SPAN_HEADER_SIZE)
   334  //! Size of a span header (must be a multiple of SMALL_GRANULARITY and a power of two)
   335  #define SPAN_HEADER_SIZE          128
   336  //! Number of spans in thread cache
   337  #define MAX_THREAD_SPAN_CACHE     400
   338  //! Number of spans to transfer between thread and global cache
   339  #define THREAD_SPAN_CACHE_TRANSFER 64
   340  //! Number of spans in thread cache for large spans (must be greater than LARGE_CLASS_COUNT / 2)
   341  #define MAX_THREAD_SPAN_LARGE_CACHE 100
   342  //! Number of spans to transfer between thread and global cache for large spans
   343  #define THREAD_SPAN_LARGE_CACHE_TRANSFER 6
   344  
   345  _Static_assert((SMALL_GRANULARITY & (SMALL_GRANULARITY - 1)) == 0, "Small granularity must be power of two");
   346  _Static_assert((SPAN_HEADER_SIZE & (SPAN_HEADER_SIZE - 1)) == 0, "Span header size must be power of two");
   347  
   348  #if ENABLE_VALIDATE_ARGS
   349  //! Maximum allocation size to avoid integer overflow
   350  #undef  MAX_ALLOC_SIZE
   351  #define MAX_ALLOC_SIZE            (((size_t)-1) - _memory_span_size)
   352  #endif
   353  
   354  #define pointer_offset(ptr, ofs) (void*)((char*)(ptr) + (ptrdiff_t)(ofs))
   355  #define pointer_diff(first, second) (ptrdiff_t)((const char*)(first) - (const char*)(second))
   356  
   357  #define INVALID_POINTER ((void*)((uintptr_t)-1))
   358  
   359  #define SIZE_CLASS_LARGE SIZE_CLASS_COUNT
   360  #define SIZE_CLASS_HUGE ((uint32_t)-1)
   361  
   362  ////////////
   363  ///
   364  /// Data types
   365  ///
   366  //////
   367  
   368  //! A memory heap, per thread
   369  typedef struct heap_t heap_t;
   370  //! Span of memory pages
   371  typedef struct span_t span_t;
   372  //! Span list
   373  typedef struct span_list_t span_list_t;
   374  //! Span active data
   375  typedef struct span_active_t span_active_t;
   376  //! Size class definition
   377  typedef struct size_class_t size_class_t;
   378  //! Global cache
   379  typedef struct global_cache_t global_cache_t;
   380  
   381  //! Flag indicating span is the first (master) span of a split superspan
   382  #define SPAN_FLAG_MASTER 1U
   383  //! Flag indicating span is a secondary (sub) span of a split superspan
   384  #define SPAN_FLAG_SUBSPAN 2U
   385  //! Flag indicating span has blocks with increased alignment
   386  #define SPAN_FLAG_ALIGNED_BLOCKS 4U
   387  //! Flag indicating an unmapped master span
   388  #define SPAN_FLAG_UNMAPPED_MASTER 8U
   389  
   390  #if ENABLE_ADAPTIVE_THREAD_CACHE || ENABLE_STATISTICS
   391  struct span_use_t {
   392  	//! Current number of spans used (actually used, not in cache)
   393  	atomic32_t current;
   394  	//! High water mark of spans used
   395  	atomic32_t high;
   396  #if ENABLE_STATISTICS
   397  	//! Number of spans in deferred list
   398  	atomic32_t spans_deferred;
   399  	//! Number of spans transitioned to global cache
   400  	atomic32_t spans_to_global;
   401  	//! Number of spans transitioned from global cache
   402  	atomic32_t spans_from_global;
   403  	//! Number of spans transitioned to thread cache
   404  	atomic32_t spans_to_cache;
   405  	//! Number of spans transitioned from thread cache
   406  	atomic32_t spans_from_cache;
   407  	//! Number of spans transitioned to reserved state
   408  	atomic32_t spans_to_reserved;
   409  	//! Number of spans transitioned from reserved state
   410  	atomic32_t spans_from_reserved;
   411  	//! Number of raw memory map calls
   412  	atomic32_t spans_map_calls;
   413  #endif
   414  };
   415  typedef struct span_use_t span_use_t;
   416  #endif
   417  
   418  #if ENABLE_STATISTICS
   419  struct size_class_use_t {
   420  	//! Current number of allocations
   421  	atomic32_t alloc_current;
   422  	//! Peak number of allocations
   423  	int32_t alloc_peak;
   424  	//! Total number of allocations
   425  	atomic32_t alloc_total;
   426  	//! Total number of frees
   427  	atomic32_t free_total;
   428  	//! Number of spans in use
   429  	atomic32_t spans_current;
   430  	//! Number of spans transitioned to cache
   431  	int32_t spans_peak;
   432  	//! Number of spans transitioned to cache
   433  	atomic32_t spans_to_cache;
   434  	//! Number of spans transitioned from cache
   435  	atomic32_t spans_from_cache;
   436  	//! Number of spans transitioned from reserved state
   437  	atomic32_t spans_from_reserved;
   438  	//! Number of spans mapped
   439  	atomic32_t spans_map_calls;
   440  	int32_t unused;
   441  };
   442  typedef struct size_class_use_t size_class_use_t;
   443  #endif
   444  
   445  // A span can either represent a single span of memory pages with size declared by span_map_count configuration variable,
   446  // or a set of spans in a continuous region, a super span. Any reference to the term "span" usually refers to both a single
   447  // span or a super span. A super span can further be divided into multiple spans (or this, super spans), where the first
   448  // (super)span is the master and subsequent (super)spans are subspans. The master span keeps track of how many subspans
   449  // that are still alive and mapped in virtual memory, and once all subspans and master have been unmapped the entire
   450  // superspan region is released and unmapped (on Windows for example, the entire superspan range has to be released
   451  // in the same call to release the virtual memory range, but individual subranges can be decommitted individually
   452  // to reduce physical memory use).
   453  struct span_t {
   454  	//! Free list
   455  	void*       free_list;
   456  	//! Total block count of size class
   457  	uint32_t    block_count;
   458  	//! Size class
   459  	uint32_t    size_class;
   460  	//! Index of last block initialized in free list
   461  	uint32_t    free_list_limit;
   462  	//! Number of used blocks remaining when in partial state
   463  	uint32_t    used_count;
   464  	//! Deferred free list
   465  	atomicptr_t free_list_deferred;
   466  	//! Size of deferred free list, or list of spans when part of a cache list
   467  	uint32_t    list_size;
   468  	//! Size of a block
   469  	uint32_t    block_size;
   470  	//! Flags and counters
   471  	uint32_t    flags;
   472  	//! Number of spans
   473  	uint32_t    span_count;
   474  	//! Total span counter for master spans
   475  	uint32_t    total_spans;
   476  	//! Offset from master span for subspans
   477  	uint32_t    offset_from_master;
   478  	//! Remaining span counter, for master spans
   479  	atomic32_t  remaining_spans;
   480  	//! Alignment offset
   481  	uint32_t    align_offset;
   482  	//! Owning heap
   483  	heap_t*     heap;
   484  	//! Next span
   485  	span_t*     next;
   486  	//! Previous span
   487  	span_t*     prev;
   488  };
   489  _Static_assert(sizeof(span_t) <= SPAN_HEADER_SIZE, "span size mismatch");
   490  
   491  struct span_cache_t {
   492  	size_t       count;
   493  	span_t*      span[MAX_THREAD_SPAN_CACHE];
   494  };
   495  typedef struct span_cache_t span_cache_t;
   496  
   497  struct span_large_cache_t {
   498  	size_t       count;
   499  	span_t*      span[MAX_THREAD_SPAN_LARGE_CACHE];
   500  };
   501  typedef struct span_large_cache_t span_large_cache_t;
   502  
   503  struct heap_size_class_t {
   504  	//! Free list of active span
   505  	void*        free_list;
   506  	//! Double linked list of partially used spans with free blocks.
   507  	//  Previous span pointer in head points to tail span of list.
   508  	span_t*      partial_span;
   509  	//! Early level cache of fully free spans
   510  	span_t*      cache;
   511  };
   512  typedef struct heap_size_class_t heap_size_class_t;
   513  
   514  // Control structure for a heap, either a thread heap or a first class heap if enabled
   515  struct heap_t {
   516  	//! Owning thread ID
   517  	uintptr_t    owner_thread;
   518  	//! Free lists for each size class
   519  	heap_size_class_t size_class[SIZE_CLASS_COUNT];
   520  #if ENABLE_THREAD_CACHE
   521  	//! Arrays of fully freed spans, single span
   522  	span_cache_t span_cache;
   523  #endif
   524  	//! List of deferred free spans (single linked list)
   525  	atomicptr_t  span_free_deferred;
   526  	//! Number of full spans
   527  	size_t       full_span_count;
   528  	//! Mapped but unused spans
   529  	span_t*      span_reserve;
   530  	//! Master span for mapped but unused spans
   531  	span_t*      span_reserve_master;
   532  	//! Number of mapped but unused spans
   533  	uint32_t     spans_reserved;
   534  	//! Child count
   535  	atomic32_t   child_count;
   536  	//! Next heap in id list
   537  	heap_t*      next_heap;
   538  	//! Next heap in orphan list
   539  	heap_t*      next_orphan;
   540  	//! Heap ID
   541  	int32_t      id;
   542  	//! Finalization state flag
   543  	int          finalize;
   544  	//! Master heap owning the memory pages
   545  	heap_t*      master_heap;
   546  #if ENABLE_THREAD_CACHE
   547  	//! Arrays of fully freed spans, large spans with > 1 span count
   548  	span_large_cache_t span_large_cache[LARGE_CLASS_COUNT - 1];
   549  #endif
   550  #if RPMALLOC_FIRST_CLASS_HEAPS
   551  	//! Double linked list of fully utilized spans with free blocks for each size class.
   552  	//  Previous span pointer in head points to tail span of list.
   553  	span_t*      full_span[SIZE_CLASS_COUNT];
   554  	//! Double linked list of large and huge spans allocated by this heap
   555  	span_t*      large_huge_span;
   556  #endif
   557  #if ENABLE_ADAPTIVE_THREAD_CACHE || ENABLE_STATISTICS
   558  	//! Current and high water mark of spans used per span count
   559  	span_use_t   span_use[LARGE_CLASS_COUNT];
   560  #endif
   561  #if ENABLE_STATISTICS
   562  	//! Allocation stats per size class
   563  	size_class_use_t size_class_use[SIZE_CLASS_COUNT + 1];
   564  	//! Number of bytes transitioned thread -> global
   565  	atomic64_t   thread_to_global;
   566  	//! Number of bytes transitioned global -> thread
   567  	atomic64_t   global_to_thread;
   568  #endif
   569  };
   570  
   571  // Size class for defining a block size bucket
   572  struct size_class_t {
   573  	//! Size of blocks in this class
   574  	uint32_t block_size;
   575  	//! Number of blocks in each chunk
   576  	uint16_t block_count;
   577  	//! Class index this class is merged with
   578  	uint16_t class_idx;
   579  };
   580  _Static_assert(sizeof(size_class_t) == 8, "Size class size mismatch");
   581  
   582  struct global_cache_t {
   583  	//! Cache lock
   584  	atomic32_t lock;
   585  	//! Cache count
   586  	uint32_t count;
   587  #if ENABLE_STATISTICS
   588  	//! Insert count
   589  	size_t insert_count;
   590  	//! Extract count
   591  	size_t extract_count;
   592  #endif
   593  	//! Cached spans
   594  	span_t* span[GLOBAL_CACHE_MULTIPLIER * MAX_THREAD_SPAN_CACHE];
   595  	//! Unlimited cache overflow
   596  	span_t* overflow;
   597  };
   598  
   599  ////////////
   600  ///
   601  /// Global data
   602  ///
   603  //////
   604  
   605  //! Default span size (64KiB)
   606  #define _memory_default_span_size (64 * 1024)
   607  #define _memory_default_span_size_shift 16
   608  #define _memory_default_span_mask (~((uintptr_t)(_memory_span_size - 1)))
   609  
   610  //! Initialized flag
   611  static int _rpmalloc_initialized;
   612  //! Main thread ID
   613  static uintptr_t _rpmalloc_main_thread_id;
   614  //! Configuration
   615  static rpmalloc_config_t _memory_config;
   616  //! Memory page size
   617  static size_t _memory_page_size;
   618  //! Shift to divide by page size
   619  static size_t _memory_page_size_shift;
   620  //! Granularity at which memory pages are mapped by OS
   621  static size_t _memory_map_granularity;
   622  #if RPMALLOC_CONFIGURABLE
   623  //! Size of a span of memory pages
   624  static size_t _memory_span_size;
   625  //! Shift to divide by span size
   626  static size_t _memory_span_size_shift;
   627  //! Mask to get to start of a memory span
   628  static uintptr_t _memory_span_mask;
   629  #else
   630  //! Hardwired span size
   631  #define _memory_span_size _memory_default_span_size
   632  #define _memory_span_size_shift _memory_default_span_size_shift
   633  #define _memory_span_mask _memory_default_span_mask
   634  #endif
   635  //! Number of spans to map in each map call
   636  static size_t _memory_span_map_count;
   637  //! Number of spans to keep reserved in each heap
   638  static size_t _memory_heap_reserve_count;
   639  //! Global size classes
   640  static size_class_t _memory_size_class[SIZE_CLASS_COUNT];
   641  //! Run-time size limit of medium blocks
   642  static size_t _memory_medium_size_limit;
   643  //! Heap ID counter
   644  static atomic32_t _memory_heap_id;
   645  //! Huge page support
   646  static int _memory_huge_pages;
   647  #if ENABLE_GLOBAL_CACHE
   648  //! Global span cache
   649  static global_cache_t _memory_span_cache[LARGE_CLASS_COUNT];
   650  #endif
   651  //! Global reserved spans
   652  static span_t* _memory_global_reserve;
   653  //! Global reserved count
   654  static size_t _memory_global_reserve_count;
   655  //! Global reserved master
   656  static span_t* _memory_global_reserve_master;
   657  //! All heaps
   658  static heap_t* _memory_heaps[HEAP_ARRAY_SIZE];
   659  //! Used to restrict access to mapping memory for huge pages
   660  static atomic32_t _memory_global_lock;
   661  //! Orphaned heaps
   662  static heap_t* _memory_orphan_heaps;
   663  #if RPMALLOC_FIRST_CLASS_HEAPS
   664  //! Orphaned heaps (first class heaps)
   665  static heap_t* _memory_first_class_orphan_heaps;
   666  #endif
   667  #if ENABLE_STATISTICS
   668  //! Allocations counter
   669  static atomic64_t _allocation_counter;
   670  //! Deallocations counter
   671  static atomic64_t _deallocation_counter;
   672  //! Active heap count
   673  static atomic32_t _memory_active_heaps;
   674  //! Number of currently mapped memory pages
   675  static atomic32_t _mapped_pages;
   676  //! Peak number of concurrently mapped memory pages
   677  static int32_t _mapped_pages_peak;
   678  //! Number of mapped master spans
   679  static atomic32_t _master_spans;
   680  //! Number of unmapped dangling master spans
   681  static atomic32_t _unmapped_master_spans;
   682  //! Running counter of total number of mapped memory pages since start
   683  static atomic32_t _mapped_total;
   684  //! Running counter of total number of unmapped memory pages since start
   685  static atomic32_t _unmapped_total;
   686  //! Number of currently mapped memory pages in OS calls
   687  static atomic32_t _mapped_pages_os;
   688  //! Number of currently allocated pages in huge allocations
   689  static atomic32_t _huge_pages_current;
   690  //! Peak number of currently allocated pages in huge allocations
   691  static int32_t _huge_pages_peak;
   692  #endif
   693  
   694  ////////////
   695  ///
   696  /// Thread local heap and ID
   697  ///
   698  //////
   699  
   700  //! Current thread heap
   701  #if ((defined(__APPLE__) || defined(__HAIKU__)) && ENABLE_PRELOAD) || defined(__TINYC__)
   702  static pthread_key_t _memory_thread_heap;
   703  #else
   704  #  ifdef _MSC_VER
   705  #    define _Thread_local __declspec(thread)
   706  #    define TLS_MODEL
   707  #  else
   708  #    ifndef __HAIKU__
   709  #      define TLS_MODEL __attribute__((tls_model("initial-exec")))
   710  #    else
   711  #      define TLS_MODEL
   712  #    endif
   713  #    if !defined(__clang__) && defined(__GNUC__)
   714  #      define _Thread_local __thread
   715  #    endif
   716  #  endif
   717  static _Thread_local heap_t* _memory_thread_heap TLS_MODEL;
   718  #endif
   719  
   720  static inline heap_t*
   721  get_thread_heap_raw(void) {
   722  #if (defined(__APPLE__) || defined(__HAIKU__)) && ENABLE_PRELOAD
   723  	return pthread_getspecific(_memory_thread_heap);
   724  #else
   725  	return _memory_thread_heap;
   726  #endif
   727  }
   728  
   729  //! Get the current thread heap
   730  static inline heap_t*
   731  get_thread_heap(void) {
   732  	heap_t* heap = get_thread_heap_raw();
   733  #if ENABLE_PRELOAD
   734  	if (EXPECTED(heap != 0))
   735  		return heap;
   736  	rpmalloc_initialize();
   737  	return get_thread_heap_raw();
   738  #else
   739  	return heap;
   740  #endif
   741  }
   742  
   743  //! Fast thread ID
   744  static inline uintptr_t
   745  get_thread_id(void) {
   746  #if defined(_WIN32)
   747  	return (uintptr_t)((void*)NtCurrentTeb());
   748  #elif (defined(__GNUC__) || defined(__clang__)) && !defined(__CYGWIN__)
   749  	uintptr_t tid;
   750  #  if defined(__i386__)
   751  	__asm__("movl %%gs:0, %0" : "=r" (tid) : : );
   752  #  elif defined(__x86_64__)
   753  #    if defined(__MACH__)
   754  	__asm__("movq %%gs:0, %0" : "=r" (tid) : : );
   755  #    else
   756  	__asm__("movq %%fs:0, %0" : "=r" (tid) : : );
   757  #    endif
   758  #  elif defined(__arm__)
   759  	__asm__ volatile ("mrc p15, 0, %0, c13, c0, 3" : "=r" (tid));
   760  #  elif defined(__aarch64__)
   761  #    if defined(__MACH__)
   762  	// tpidr_el0 likely unused, always return 0 on iOS
   763  	__asm__ volatile ("mrs %0, tpidrro_el0" : "=r" (tid));
   764  #    else
   765  	__asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tid));
   766  #    endif
   767  #  else
   768  	tid = (uintptr_t)((void*)get_thread_heap_raw());
   769  #  endif
   770  	return tid;
   771  #else
   772  	return (uintptr_t)((void*)get_thread_heap_raw());
   773  #endif
   774  }
   775  
   776  //! Set the current thread heap
   777  static void
   778  set_thread_heap(heap_t* heap) {
   779  #if ((defined(__APPLE__) || defined(__HAIKU__)) && ENABLE_PRELOAD) || defined(__TINYC__)
   780  	pthread_setspecific(_memory_thread_heap, heap);
   781  #else
   782  	_memory_thread_heap = heap;
   783  #endif
   784  	if (heap)
   785  		heap->owner_thread = get_thread_id();
   786  }
   787  
   788  //! Set main thread ID
   789  extern void
   790  rpmalloc_set_main_thread(void);
   791  
   792  void
   793  rpmalloc_set_main_thread(void) {
   794  	_rpmalloc_main_thread_id = get_thread_id();
   795  }
   796  
   797  static void
   798  _rpmalloc_spin(void) {
   799  #if defined(_MSC_VER)
   800  	_mm_pause();
   801  #elif defined(__x86_64__) || defined(__i386__)
   802  	__asm__ volatile("pause" ::: "memory");
   803  #elif defined(__aarch64__) || (defined(__arm__) && __ARM_ARCH >= 7)
   804  	__asm__ volatile("yield" ::: "memory");
   805  #elif defined(__powerpc__) || defined(__powerpc64__)
   806          // No idea if ever been compiled in such archs but ... as precaution
   807  	__asm__ volatile("or 27,27,27");
   808  #elif defined(__sparc__)
   809  	__asm__ volatile("rd %ccr, %g0 \n\trd %ccr, %g0 \n\trd %ccr, %g0");
   810  #else
   811  	struct timespec ts = {0};
   812  	nanosleep(&ts, 0);
   813  #endif
   814  }
   815  
   816  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
   817  static void NTAPI
   818  _rpmalloc_thread_destructor(void* value) {
   819  #if ENABLE_OVERRIDE
   820  	// If this is called on main thread it means rpmalloc_finalize
   821  	// has not been called and shutdown is forced (through _exit) or unclean
   822  	if (get_thread_id() == _rpmalloc_main_thread_id)
   823  		return;
   824  #endif
   825  	if (value)
   826  		rpmalloc_thread_finalize(1);
   827  }
   828  #endif
   829  
   830  
   831  ////////////
   832  ///
   833  /// Low level memory map/unmap
   834  ///
   835  //////
   836  
   837  static void
   838  _rpmalloc_set_name(void* address, size_t size) {
   839  #if defined(__linux__) || defined(__ANDROID__)
   840  	const char *name = _memory_huge_pages ? _memory_config.huge_page_name : _memory_config.page_name;
   841  	if (address == MAP_FAILED || !name)
   842  		return;
   843  	// If the kernel does not support CONFIG_ANON_VMA_NAME or if the call fails
   844  	// (e.g. invalid name) it is a no-op basically.
   845  	(void)prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, (uintptr_t)address, size, (uintptr_t)name);
   846  #else
   847  	(void)sizeof(size);
   848  	(void)sizeof(address);
   849  #endif
   850  }
   851  
   852  
   853  //! Map more virtual memory
   854  //  size is number of bytes to map
   855  //  offset receives the offset in bytes from start of mapped region
   856  //  returns address to start of mapped region to use
   857  static void*
   858  _rpmalloc_mmap(size_t size, size_t* offset) {
   859  	rpmalloc_assert(!(size % _memory_page_size), "Invalid mmap size");
   860  	rpmalloc_assert(size >= _memory_page_size, "Invalid mmap size");
   861  	void* address = _memory_config.memory_map(size, offset);
   862  	if (EXPECTED(address != 0)) {
   863  		_rpmalloc_stat_add_peak(&_mapped_pages, (size >> _memory_page_size_shift), _mapped_pages_peak);
   864  		_rpmalloc_stat_add(&_mapped_total, (size >> _memory_page_size_shift));
   865  	}
   866  	return address;
   867  }
   868  
   869  //! Unmap virtual memory
   870  //  address is the memory address to unmap, as returned from _memory_map
   871  //  size is the number of bytes to unmap, which might be less than full region for a partial unmap
   872  //  offset is the offset in bytes to the actual mapped region, as set by _memory_map
   873  //  release is set to 0 for partial unmap, or size of entire range for a full unmap
   874  static void
   875  _rpmalloc_unmap(void* address, size_t size, size_t offset, size_t release) {
   876  	rpmalloc_assert(!release || (release >= size), "Invalid unmap size");
   877  	rpmalloc_assert(!release || (release >= _memory_page_size), "Invalid unmap size");
   878  	if (release) {
   879  		rpmalloc_assert(!(release % _memory_page_size), "Invalid unmap size");
   880  		_rpmalloc_stat_sub(&_mapped_pages, (release >> _memory_page_size_shift));
   881  		_rpmalloc_stat_add(&_unmapped_total, (release >> _memory_page_size_shift));
   882  	}
   883  	_memory_config.memory_unmap(address, size, offset, release);
   884  }
   885  
   886  //! Default implementation to map new pages to virtual memory
   887  static void*
   888  _rpmalloc_mmap_os(size_t size, size_t* offset) {
   889  	//Either size is a heap (a single page) or a (multiple) span - we only need to align spans, and only if larger than map granularity
   890  	size_t padding = ((size >= _memory_span_size) && (_memory_span_size > _memory_map_granularity)) ? _memory_span_size : 0;
   891  	rpmalloc_assert(size >= _memory_page_size, "Invalid mmap size");
   892  #if PLATFORM_WINDOWS
   893  	//Ok to MEM_COMMIT - according to MSDN, "actual physical pages are not allocated unless/until the virtual addresses are actually accessed"
   894  	void* ptr = VirtualAlloc(0, size + padding, (_memory_huge_pages ? MEM_LARGE_PAGES : 0) | MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
   895  	if (!ptr) {
   896  		if (_memory_config.map_fail_callback) {
   897  			if (_memory_config.map_fail_callback(size + padding))
   898  				return _rpmalloc_mmap_os(size, offset);
   899  		} else {
   900  			rpmalloc_assert(ptr, "Failed to map virtual memory block");
   901  		}
   902  		return 0;
   903  	}
   904  #else
   905  	int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_UNINITIALIZED;
   906  #  if defined(__APPLE__) && !TARGET_OS_IPHONE && !TARGET_OS_SIMULATOR
   907  	int fd = (int)VM_MAKE_TAG(240U);
   908  	if (_memory_huge_pages)
   909  		fd |= VM_FLAGS_SUPERPAGE_SIZE_2MB;
   910  	void* ptr = mmap(0, size + padding, PROT_READ | PROT_WRITE, flags, fd, 0);
   911  #  elif defined(MAP_HUGETLB)
   912  	void* ptr = mmap(0, size + padding, PROT_READ | PROT_WRITE | PROT_MAX(PROT_READ | PROT_WRITE), (_memory_huge_pages ? MAP_HUGETLB : 0) | flags, -1, 0);
   913  #    if defined(MADV_HUGEPAGE)
   914  	// In some configurations, huge pages allocations might fail thus
   915  	// we fallback to normal allocations and promote the region as transparent huge page
   916  	if ((ptr == MAP_FAILED || !ptr) && _memory_huge_pages) {
   917  		ptr = mmap(0, size + padding, PROT_READ | PROT_WRITE, flags, -1, 0);
   918  		if (ptr && ptr != MAP_FAILED) {
   919  			int prm = madvise(ptr, size + padding, MADV_HUGEPAGE);
   920  			(void)prm;
   921  			rpmalloc_assert((prm == 0), "Failed to promote the page to THP");
   922  		}
   923  	}
   924  #    endif
   925  	_rpmalloc_set_name(ptr, size + padding);
   926  #  elif defined(MAP_ALIGNED)
   927  	const size_t align = (sizeof(size_t) * 8) - (size_t)(__builtin_clzl(size - 1));
   928  	void* ptr = mmap(0, size + padding, PROT_READ | PROT_WRITE, (_memory_huge_pages ? MAP_ALIGNED(align) : 0) | flags, -1, 0);
   929  #  elif defined(MAP_ALIGN)
   930  	caddr_t base = (_memory_huge_pages ? (caddr_t)(4 << 20) : 0);
   931  	void* ptr = mmap(base, size + padding, PROT_READ | PROT_WRITE, (_memory_huge_pages ? MAP_ALIGN : 0) | flags, -1, 0);
   932  #  else
   933  	void* ptr = mmap(0, size + padding, PROT_READ | PROT_WRITE, flags, -1, 0);
   934  #  endif
   935  	if ((ptr == MAP_FAILED) || !ptr) {
   936  		if (_memory_config.map_fail_callback) {
   937  			if (_memory_config.map_fail_callback(size + padding))
   938  				return _rpmalloc_mmap_os(size, offset);
   939  		} else if (errno != ENOMEM) {
   940  			rpmalloc_assert((ptr != MAP_FAILED) && ptr, "Failed to map virtual memory block");
   941  		}
   942  		return 0;
   943  	}
   944  #endif
   945  	_rpmalloc_stat_add(&_mapped_pages_os, (int32_t)((size + padding) >> _memory_page_size_shift));
   946  	if (padding) {
   947  		size_t final_padding = padding - ((uintptr_t)ptr & ~_memory_span_mask);
   948  		rpmalloc_assert(final_padding <= _memory_span_size, "Internal failure in padding");
   949  		rpmalloc_assert(final_padding <= padding, "Internal failure in padding");
   950  		rpmalloc_assert(!(final_padding % 8), "Internal failure in padding");
   951  		ptr = pointer_offset(ptr, final_padding);
   952  		*offset = final_padding >> 3;
   953  	}
   954  	rpmalloc_assert((size < _memory_span_size) || !((uintptr_t)ptr & ~_memory_span_mask), "Internal failure in padding");
   955  	return ptr;
   956  }
   957  
   958  //! Default implementation to unmap pages from virtual memory
   959  static void
   960  _rpmalloc_unmap_os(void* address, size_t size, size_t offset, size_t release) {
   961  	rpmalloc_assert(release || (offset == 0), "Invalid unmap size");
   962  	rpmalloc_assert(!release || (release >= _memory_page_size), "Invalid unmap size");
   963  	rpmalloc_assert(size >= _memory_page_size, "Invalid unmap size");
   964  	if (release && offset) {
   965  		offset <<= 3;
   966  		address = pointer_offset(address, -(int32_t)offset);
   967  		if ((release >= _memory_span_size) && (_memory_span_size > _memory_map_granularity)) {
   968  			//Padding is always one span size
   969  			release += _memory_span_size;
   970  		}
   971  	}
   972  #if !DISABLE_UNMAP
   973  #if PLATFORM_WINDOWS
   974  	if (!VirtualFree(address, release ? 0 : size, release ? MEM_RELEASE : MEM_DECOMMIT)) {
   975  		rpmalloc_assert(0, "Failed to unmap virtual memory block");
   976  	}
   977  #else
   978  	if (release) {
   979  		if (munmap(address, release)) {
   980  			rpmalloc_assert(0, "Failed to unmap virtual memory block");
   981  		}
   982  	} else {
   983  #if defined(MADV_FREE_REUSABLE)
   984  		int ret;
   985  		while ((ret = madvise(address, size, MADV_FREE_REUSABLE)) == -1 && (errno == EAGAIN))
   986  			errno = 0;
   987  		if ((ret == -1) && (errno != 0)) {
   988  #elif defined(MADV_DONTNEED)
   989  		if (madvise(address, size, MADV_DONTNEED)) {
   990  #elif defined(MADV_PAGEOUT)
   991  		if (madvise(address, size, MADV_PAGEOUT)) {
   992  #elif defined(MADV_FREE)
   993  		if (madvise(address, size, MADV_FREE)) {
   994  #else
   995  		if (posix_madvise(address, size, POSIX_MADV_DONTNEED)) {
   996  #endif
   997  			rpmalloc_assert(0, "Failed to madvise virtual memory block as free");
   998  		}
   999  	}
  1000  #endif
  1001  #endif
  1002  	if (release)
  1003  		_rpmalloc_stat_sub(&_mapped_pages_os, release >> _memory_page_size_shift);
  1004  }
  1005  
  1006  static void
  1007  _rpmalloc_span_mark_as_subspan_unless_master(span_t* master, span_t* subspan, size_t span_count);
  1008  
  1009  //! Use global reserved spans to fulfill a memory map request (reserve size must be checked by caller)
  1010  static span_t*
  1011  _rpmalloc_global_get_reserved_spans(size_t span_count) {
  1012  	span_t* span = _memory_global_reserve;
  1013  	_rpmalloc_span_mark_as_subspan_unless_master(_memory_global_reserve_master, span, span_count);
  1014  	_memory_global_reserve_count -= span_count;
  1015  	if (_memory_global_reserve_count)
  1016  		_memory_global_reserve = (span_t*)pointer_offset(span, span_count << _memory_span_size_shift);
  1017  	else
  1018  		_memory_global_reserve = 0;
  1019  	return span;
  1020  }
  1021  
  1022  //! Store the given spans as global reserve (must only be called from within new heap allocation, not thread safe)
  1023  static void
  1024  _rpmalloc_global_set_reserved_spans(span_t* master, span_t* reserve, size_t reserve_span_count) {
  1025  	_memory_global_reserve_master = master;
  1026  	_memory_global_reserve_count = reserve_span_count;
  1027  	_memory_global_reserve = reserve;
  1028  }
  1029  
  1030  
  1031  ////////////
  1032  ///
  1033  /// Span linked list management
  1034  ///
  1035  //////
  1036  
  1037  //! Add a span to double linked list at the head
  1038  static void
  1039  _rpmalloc_span_double_link_list_add(span_t** head, span_t* span) {
  1040  	if (*head)
  1041  		(*head)->prev = span;
  1042  	span->next = *head;
  1043  	*head = span;
  1044  }
  1045  
  1046  //! Pop head span from double linked list
  1047  static void
  1048  _rpmalloc_span_double_link_list_pop_head(span_t** head, span_t* span) {
  1049  	rpmalloc_assert(*head == span, "Linked list corrupted");
  1050  	span = *head;
  1051  	*head = span->next;
  1052  }
  1053  
  1054  //! Remove a span from double linked list
  1055  static void
  1056  _rpmalloc_span_double_link_list_remove(span_t** head, span_t* span) {
  1057  	rpmalloc_assert(*head, "Linked list corrupted");
  1058  	if (*head == span) {
  1059  		*head = span->next;
  1060  	} else {
  1061  		span_t* next_span = span->next;
  1062  		span_t* prev_span = span->prev;
  1063  		prev_span->next = next_span;
  1064  		if (EXPECTED(next_span != 0))
  1065  			next_span->prev = prev_span;
  1066  	}
  1067  }
  1068  
  1069  
  1070  ////////////
  1071  ///
  1072  /// Span control
  1073  ///
  1074  //////
  1075  
  1076  static void
  1077  _rpmalloc_heap_cache_insert(heap_t* heap, span_t* span);
  1078  
  1079  static void
  1080  _rpmalloc_heap_finalize(heap_t* heap);
  1081  
  1082  static void
  1083  _rpmalloc_heap_set_reserved_spans(heap_t* heap, span_t* master, span_t* reserve, size_t reserve_span_count);
  1084  
  1085  //! Declare the span to be a subspan and store distance from master span and span count
  1086  static void
  1087  _rpmalloc_span_mark_as_subspan_unless_master(span_t* master, span_t* subspan, size_t span_count) {
  1088  	rpmalloc_assert((subspan != master) || (subspan->flags & SPAN_FLAG_MASTER), "Span master pointer and/or flag mismatch");
  1089  	if (subspan != master) {
  1090  		subspan->flags = SPAN_FLAG_SUBSPAN;
  1091  		subspan->offset_from_master = (uint32_t)((uintptr_t)pointer_diff(subspan, master) >> _memory_span_size_shift);
  1092  		subspan->align_offset = 0;
  1093  	}
  1094  	subspan->span_count = (uint32_t)span_count;
  1095  }
  1096  
  1097  //! Use reserved spans to fulfill a memory map request (reserve size must be checked by caller)
  1098  static span_t*
  1099  _rpmalloc_span_map_from_reserve(heap_t* heap, size_t span_count) {
  1100  	//Update the heap span reserve
  1101  	span_t* span = heap->span_reserve;
  1102  	heap->span_reserve = (span_t*)pointer_offset(span, span_count * _memory_span_size);
  1103  	heap->spans_reserved -= (uint32_t)span_count;
  1104  
  1105  	_rpmalloc_span_mark_as_subspan_unless_master(heap->span_reserve_master, span, span_count);
  1106  	if (span_count <= LARGE_CLASS_COUNT)
  1107  		_rpmalloc_stat_inc(&heap->span_use[span_count - 1].spans_from_reserved);
  1108  
  1109  	return span;
  1110  }
  1111  
  1112  //! Get the aligned number of spans to map in based on wanted count, configured mapping granularity and the page size
  1113  static size_t
  1114  _rpmalloc_span_align_count(size_t span_count) {
  1115  	size_t request_count = (span_count > _memory_span_map_count) ? span_count : _memory_span_map_count;
  1116  	if ((_memory_page_size > _memory_span_size) && ((request_count * _memory_span_size) % _memory_page_size))
  1117  		request_count += _memory_span_map_count - (request_count % _memory_span_map_count);
  1118  	return request_count;
  1119  }
  1120  
  1121  //! Setup a newly mapped span
  1122  static void
  1123  _rpmalloc_span_initialize(span_t* span, size_t total_span_count, size_t span_count, size_t align_offset) {
  1124  	span->total_spans = (uint32_t)total_span_count;
  1125  	span->span_count = (uint32_t)span_count;
  1126  	span->align_offset = (uint32_t)align_offset;
  1127  	span->flags = SPAN_FLAG_MASTER;
  1128  	atomic_store32(&span->remaining_spans, (int32_t)total_span_count);
  1129  }
  1130  
  1131  static void
  1132  _rpmalloc_span_unmap(span_t* span);
  1133  
  1134  //! Map an aligned set of spans, taking configured mapping granularity and the page size into account
  1135  static span_t*
  1136  _rpmalloc_span_map_aligned_count(heap_t* heap, size_t span_count) {
  1137  	//If we already have some, but not enough, reserved spans, release those to heap cache and map a new
  1138  	//full set of spans. Otherwise we would waste memory if page size > span size (huge pages)
  1139  	size_t aligned_span_count = _rpmalloc_span_align_count(span_count);
  1140  	size_t align_offset = 0;
  1141  	span_t* span = (span_t*)_rpmalloc_mmap(aligned_span_count * _memory_span_size, &align_offset);
  1142  	if (!span)
  1143  		return 0;
  1144  	_rpmalloc_span_initialize(span, aligned_span_count, span_count, align_offset);
  1145  	_rpmalloc_stat_inc(&_master_spans);
  1146  	if (span_count <= LARGE_CLASS_COUNT)
  1147  		_rpmalloc_stat_inc(&heap->span_use[span_count - 1].spans_map_calls);
  1148  	if (aligned_span_count > span_count) {
  1149  		span_t* reserved_spans = (span_t*)pointer_offset(span, span_count * _memory_span_size);
  1150  		size_t reserved_count = aligned_span_count - span_count;
  1151  		if (heap->spans_reserved) {
  1152  			_rpmalloc_span_mark_as_subspan_unless_master(heap->span_reserve_master, heap->span_reserve, heap->spans_reserved);
  1153  			_rpmalloc_heap_cache_insert(heap, heap->span_reserve);
  1154  		}
  1155  		if (reserved_count > _memory_heap_reserve_count) {
  1156  			// If huge pages or eager spam map count, the global reserve spin lock is held by caller, _rpmalloc_span_map
  1157  			rpmalloc_assert(atomic_load32(&_memory_global_lock) == 1, "Global spin lock not held as expected");
  1158  			size_t remain_count = reserved_count - _memory_heap_reserve_count;
  1159  			reserved_count = _memory_heap_reserve_count;
  1160  			span_t* remain_span = (span_t*)pointer_offset(reserved_spans, reserved_count * _memory_span_size);
  1161  			if (_memory_global_reserve) {
  1162  				_rpmalloc_span_mark_as_subspan_unless_master(_memory_global_reserve_master, _memory_global_reserve, _memory_global_reserve_count);
  1163  				_rpmalloc_span_unmap(_memory_global_reserve);
  1164  			}
  1165  			_rpmalloc_global_set_reserved_spans(span, remain_span, remain_count);
  1166  		}
  1167  		_rpmalloc_heap_set_reserved_spans(heap, span, reserved_spans, reserved_count);
  1168  	}
  1169  	return span;
  1170  }
  1171  
  1172  //! Map in memory pages for the given number of spans (or use previously reserved pages)
  1173  static span_t*
  1174  _rpmalloc_span_map(heap_t* heap, size_t span_count) {
  1175  	if (span_count <= heap->spans_reserved)
  1176  		return _rpmalloc_span_map_from_reserve(heap, span_count);
  1177  	span_t* span = 0;
  1178  	int use_global_reserve = (_memory_page_size > _memory_span_size) || (_memory_span_map_count > _memory_heap_reserve_count);
  1179  	if (use_global_reserve) {
  1180  		// If huge pages, make sure only one thread maps more memory to avoid bloat
  1181  		while (!atomic_cas32_acquire(&_memory_global_lock, 1, 0))
  1182  			_rpmalloc_spin();
  1183  		if (_memory_global_reserve_count >= span_count) {
  1184  			size_t reserve_count = (!heap->spans_reserved ? _memory_heap_reserve_count : span_count);
  1185  			if (_memory_global_reserve_count < reserve_count)
  1186  				reserve_count = _memory_global_reserve_count;
  1187  			span = _rpmalloc_global_get_reserved_spans(reserve_count);
  1188  			if (span) {
  1189  				if (reserve_count > span_count) {
  1190  					span_t* reserved_span = (span_t*)pointer_offset(span, span_count << _memory_span_size_shift);
  1191  					_rpmalloc_heap_set_reserved_spans(heap, _memory_global_reserve_master, reserved_span, reserve_count - span_count);
  1192  				}
  1193  				// Already marked as subspan in _rpmalloc_global_get_reserved_spans
  1194  				span->span_count = (uint32_t)span_count;
  1195  			}
  1196  		}
  1197  	}
  1198  	if (!span)
  1199  		span = _rpmalloc_span_map_aligned_count(heap, span_count);
  1200  	if (use_global_reserve)
  1201  		atomic_store32_release(&_memory_global_lock, 0);
  1202  	return span;
  1203  }
  1204  
  1205  //! Unmap memory pages for the given number of spans (or mark as unused if no partial unmappings)
  1206  static void
  1207  _rpmalloc_span_unmap(span_t* span) {
  1208  	rpmalloc_assert((span->flags & SPAN_FLAG_MASTER) || (span->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  1209  	rpmalloc_assert(!(span->flags & SPAN_FLAG_MASTER) || !(span->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  1210  
  1211  	int is_master = !!(span->flags & SPAN_FLAG_MASTER);
  1212  	span_t* master = is_master ? span : ((span_t*)pointer_offset(span, -(intptr_t)((uintptr_t)span->offset_from_master * _memory_span_size)));
  1213  	rpmalloc_assert(is_master || (span->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  1214  	rpmalloc_assert(master->flags & SPAN_FLAG_MASTER, "Span flag corrupted");
  1215  
  1216  	size_t span_count = span->span_count;
  1217  	if (!is_master) {
  1218  		//Directly unmap subspans (unless huge pages, in which case we defer and unmap entire page range with master)
  1219  		rpmalloc_assert(span->align_offset == 0, "Span align offset corrupted");
  1220  		if (_memory_span_size >= _memory_page_size)
  1221  			_rpmalloc_unmap(span, span_count * _memory_span_size, 0, 0);
  1222  	} else {
  1223  		//Special double flag to denote an unmapped master
  1224  		//It must be kept in memory since span header must be used
  1225  		span->flags |= SPAN_FLAG_MASTER | SPAN_FLAG_SUBSPAN | SPAN_FLAG_UNMAPPED_MASTER;
  1226  		_rpmalloc_stat_add(&_unmapped_master_spans, 1);
  1227  	}
  1228  
  1229  	if (atomic_add32(&master->remaining_spans, -(int32_t)span_count) <= 0) {
  1230  		//Everything unmapped, unmap the master span with release flag to unmap the entire range of the super span
  1231  		rpmalloc_assert(!!(master->flags & SPAN_FLAG_MASTER) && !!(master->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  1232  		size_t unmap_count = master->span_count;
  1233  		if (_memory_span_size < _memory_page_size)
  1234  			unmap_count = master->total_spans;
  1235  		_rpmalloc_stat_sub(&_master_spans, 1);
  1236  		_rpmalloc_stat_sub(&_unmapped_master_spans, 1);
  1237  		_rpmalloc_unmap(master, unmap_count * _memory_span_size, master->align_offset, (size_t)master->total_spans * _memory_span_size);
  1238  	}
  1239  }
  1240  
  1241  //! Move the span (used for small or medium allocations) to the heap thread cache
  1242  static void
  1243  _rpmalloc_span_release_to_cache(heap_t* heap, span_t* span) {
  1244  	rpmalloc_assert(heap == span->heap, "Span heap pointer corrupted");
  1245  	rpmalloc_assert(span->size_class < SIZE_CLASS_COUNT, "Invalid span size class");
  1246  	rpmalloc_assert(span->span_count == 1, "Invalid span count");
  1247  #if ENABLE_ADAPTIVE_THREAD_CACHE || ENABLE_STATISTICS
  1248  	atomic_decr32(&heap->span_use[0].current);
  1249  #endif
  1250  	_rpmalloc_stat_dec(&heap->size_class_use[span->size_class].spans_current);
  1251  	if (!heap->finalize) {
  1252  		_rpmalloc_stat_inc(&heap->span_use[0].spans_to_cache);
  1253  		_rpmalloc_stat_inc(&heap->size_class_use[span->size_class].spans_to_cache);
  1254  		if (heap->size_class[span->size_class].cache)
  1255  			_rpmalloc_heap_cache_insert(heap, heap->size_class[span->size_class].cache);
  1256  		heap->size_class[span->size_class].cache = span;
  1257  	} else {
  1258  		_rpmalloc_span_unmap(span);
  1259  	}
  1260  }
  1261  
  1262  //! Initialize a (partial) free list up to next system memory page, while reserving the first block
  1263  //! as allocated, returning number of blocks in list
  1264  static uint32_t
  1265  free_list_partial_init(void** list, void** first_block, void* page_start, void* block_start, uint32_t block_count, uint32_t block_size) {
  1266  	rpmalloc_assert(block_count, "Internal failure");
  1267  	*first_block = block_start;
  1268  	if (block_count > 1) {
  1269  		void* free_block = pointer_offset(block_start, block_size);
  1270  		void* block_end = pointer_offset(block_start, (size_t)block_size * block_count);
  1271  		//If block size is less than half a memory page, bound init to next memory page boundary
  1272  		if (block_size < (_memory_page_size >> 1)) {
  1273  			void* page_end = pointer_offset(page_start, _memory_page_size);
  1274  			if (page_end < block_end)
  1275  				block_end = page_end;
  1276  		}
  1277  		*list = free_block;
  1278  		block_count = 2;
  1279  		void* next_block = pointer_offset(free_block, block_size);
  1280  		while (next_block < block_end) {
  1281  			*((void**)free_block) = next_block;
  1282  			free_block = next_block;
  1283  			++block_count;
  1284  			next_block = pointer_offset(next_block, block_size);
  1285  		}
  1286  		*((void**)free_block) = 0;
  1287  	} else {
  1288  		*list = 0;
  1289  	}
  1290  	return block_count;
  1291  }
  1292  
  1293  //! Initialize an unused span (from cache or mapped) to be new active span, putting the initial free list in heap class free list
  1294  static void*
  1295  _rpmalloc_span_initialize_new(heap_t* heap, heap_size_class_t* heap_size_class, span_t* span, uint32_t class_idx) {
  1296  	rpmalloc_assert(span->span_count == 1, "Internal failure");
  1297  	size_class_t* size_class = _memory_size_class + class_idx;
  1298  	span->size_class = class_idx;
  1299  	span->heap = heap;
  1300  	span->flags &= ~SPAN_FLAG_ALIGNED_BLOCKS;
  1301  	span->block_size = size_class->block_size;
  1302  	span->block_count = size_class->block_count;
  1303  	span->free_list = 0;
  1304  	span->list_size = 0;
  1305  	atomic_store_ptr_release(&span->free_list_deferred, 0);
  1306  
  1307  	//Setup free list. Only initialize one system page worth of free blocks in list
  1308  	void* block;
  1309  	span->free_list_limit = free_list_partial_init(&heap_size_class->free_list, &block,
  1310  		span, pointer_offset(span, SPAN_HEADER_SIZE), size_class->block_count, size_class->block_size);
  1311  	//Link span as partial if there remains blocks to be initialized as free list, or full if fully initialized
  1312  	if (span->free_list_limit < span->block_count) {
  1313  		_rpmalloc_span_double_link_list_add(&heap_size_class->partial_span, span);
  1314  		span->used_count = span->free_list_limit;
  1315  	} else {
  1316  #if RPMALLOC_FIRST_CLASS_HEAPS
  1317  		_rpmalloc_span_double_link_list_add(&heap->full_span[class_idx], span);
  1318  #endif
  1319  		++heap->full_span_count;
  1320  		span->used_count = span->block_count;
  1321  	}
  1322  	return block;
  1323  }
  1324  
  1325  static void
  1326  _rpmalloc_span_extract_free_list_deferred(span_t* span) {
  1327  	// We need acquire semantics on the CAS operation since we are interested in the list size
  1328  	// Refer to _rpmalloc_deallocate_defer_small_or_medium for further comments on this dependency
  1329  	do {
  1330  		span->free_list = atomic_exchange_ptr_acquire(&span->free_list_deferred, INVALID_POINTER);
  1331  	} while (span->free_list == INVALID_POINTER);
  1332  	span->used_count -= span->list_size;
  1333  	span->list_size = 0;
  1334  	atomic_store_ptr_release(&span->free_list_deferred, 0);
  1335  }
  1336  
  1337  static int
  1338  _rpmalloc_span_is_fully_utilized(span_t* span) {
  1339  	rpmalloc_assert(span->free_list_limit <= span->block_count, "Span free list corrupted");
  1340  	return !span->free_list && (span->free_list_limit >= span->block_count);
  1341  }
  1342  
  1343  static int
  1344  _rpmalloc_span_finalize(heap_t* heap, size_t iclass, span_t* span, span_t** list_head) {
  1345  	void* free_list = heap->size_class[iclass].free_list;
  1346  	span_t* class_span = (span_t*)((uintptr_t)free_list & _memory_span_mask);
  1347  	if (span == class_span) {
  1348  		// Adopt the heap class free list back into the span free list
  1349  		void* block = span->free_list;
  1350  		void* last_block = 0;
  1351  		while (block) {
  1352  			last_block = block;
  1353  			block = *((void**)block);
  1354  		}
  1355  		uint32_t free_count = 0;
  1356  		block = free_list;
  1357  		while (block) {
  1358  			++free_count;
  1359  			block = *((void**)block);
  1360  		}
  1361  		if (last_block) {
  1362  			*((void**)last_block) = free_list;
  1363  		} else {
  1364  			span->free_list = free_list;
  1365  		}
  1366  		heap->size_class[iclass].free_list = 0;
  1367  		span->used_count -= free_count;
  1368  	}
  1369  	//If this assert triggers you have memory leaks
  1370  	rpmalloc_assert(span->list_size == span->used_count, "Memory leak detected");
  1371  	if (span->list_size == span->used_count) {
  1372  		_rpmalloc_stat_dec(&heap->span_use[0].current);
  1373  		_rpmalloc_stat_dec(&heap->size_class_use[iclass].spans_current);
  1374  		// This function only used for spans in double linked lists
  1375  		if (list_head)
  1376  			_rpmalloc_span_double_link_list_remove(list_head, span);
  1377  		_rpmalloc_span_unmap(span);
  1378  		return 1;
  1379  	}
  1380  	return 0;
  1381  }
  1382  
  1383  
  1384  ////////////
  1385  ///
  1386  /// Global cache
  1387  ///
  1388  //////
  1389  
  1390  #if ENABLE_GLOBAL_CACHE
  1391  
  1392  //! Finalize a global cache
  1393  static void
  1394  _rpmalloc_global_cache_finalize(global_cache_t* cache) {
  1395  	while (!atomic_cas32_acquire(&cache->lock, 1, 0))
  1396  		_rpmalloc_spin();
  1397  
  1398  	for (size_t ispan = 0; ispan < cache->count; ++ispan)
  1399  		_rpmalloc_span_unmap(cache->span[ispan]);
  1400  	cache->count = 0;
  1401  
  1402  	while (cache->overflow) {
  1403  		span_t* span = cache->overflow;
  1404  		cache->overflow = span->next;
  1405  		_rpmalloc_span_unmap(span);
  1406  	}
  1407  
  1408  	atomic_store32_release(&cache->lock, 0);
  1409  }
  1410  
  1411  static void
  1412  _rpmalloc_global_cache_insert_spans(span_t** span, size_t span_count, size_t count) {
  1413  	const size_t cache_limit = (span_count == 1) ?
  1414  		GLOBAL_CACHE_MULTIPLIER * MAX_THREAD_SPAN_CACHE :
  1415  		GLOBAL_CACHE_MULTIPLIER * (MAX_THREAD_SPAN_LARGE_CACHE - (span_count >> 1));
  1416  
  1417  	global_cache_t* cache = &_memory_span_cache[span_count - 1];
  1418  
  1419  	size_t insert_count = count;
  1420  	while (!atomic_cas32_acquire(&cache->lock, 1, 0))
  1421  		_rpmalloc_spin();
  1422  
  1423  #if ENABLE_STATISTICS
  1424  	cache->insert_count += count;
  1425  #endif
  1426  	if ((cache->count + insert_count) > cache_limit)
  1427  		insert_count = cache_limit - cache->count;
  1428  
  1429  	memcpy(cache->span + cache->count, span, sizeof(span_t*) * insert_count);
  1430  	cache->count += (uint32_t)insert_count;
  1431  
  1432  #if ENABLE_UNLIMITED_CACHE
  1433  	while (insert_count < count) {
  1434  #else
  1435  	// Enable unlimited cache if huge pages, or we will leak since it is unlikely that an entire huge page
  1436  	// will be unmapped, and we're unable to partially decommit a huge page
  1437  	while ((_memory_page_size > _memory_span_size) && (insert_count < count)) {
  1438  #endif
  1439  		span_t* current_span = span[insert_count++];
  1440  		current_span->next = cache->overflow;
  1441  		cache->overflow = current_span;
  1442  	}
  1443  	atomic_store32_release(&cache->lock, 0);
  1444  
  1445  	span_t* keep = 0;
  1446  	for (size_t ispan = insert_count; ispan < count; ++ispan) {
  1447  		span_t* current_span = span[ispan];
  1448  		// Keep master spans that has remaining subspans to avoid dangling them
  1449  		if ((current_span->flags & SPAN_FLAG_MASTER) &&
  1450  		    (atomic_load32(&current_span->remaining_spans) > (int32_t)current_span->span_count)) {
  1451  			current_span->next = keep;
  1452  			keep = current_span;
  1453  		} else {
  1454  			_rpmalloc_span_unmap(current_span);
  1455  		}
  1456  	}
  1457  
  1458  	if (keep) {
  1459  		while (!atomic_cas32_acquire(&cache->lock, 1, 0))
  1460  			_rpmalloc_spin();
  1461  
  1462  		size_t islot = 0;
  1463  		while (keep) {
  1464  			for (; islot < cache->count; ++islot) {
  1465  				span_t* current_span = cache->span[islot];
  1466  				if (!(current_span->flags & SPAN_FLAG_MASTER) || ((current_span->flags & SPAN_FLAG_MASTER) &&
  1467  				    (atomic_load32(&current_span->remaining_spans) <= (int32_t)current_span->span_count))) {
  1468  					_rpmalloc_span_unmap(current_span);
  1469  					cache->span[islot] = keep;
  1470  					break;
  1471  				}
  1472  			}
  1473  			if (islot == cache->count)
  1474  				break;
  1475  			keep = keep->next;
  1476  		}
  1477  
  1478  		if (keep) {
  1479  			span_t* tail = keep;
  1480  			while (tail->next)
  1481  				tail = tail->next;
  1482  			tail->next = cache->overflow;
  1483  			cache->overflow = keep;
  1484  		}
  1485  
  1486  		atomic_store32_release(&cache->lock, 0);
  1487  	}
  1488  }
  1489  
  1490  static size_t
  1491  _rpmalloc_global_cache_extract_spans(span_t** span, size_t span_count, size_t count) {
  1492  	global_cache_t* cache = &_memory_span_cache[span_count - 1];
  1493  
  1494  	size_t extract_count = 0;
  1495  	while (!atomic_cas32_acquire(&cache->lock, 1, 0))
  1496  		_rpmalloc_spin();
  1497  
  1498  #if ENABLE_STATISTICS
  1499  	cache->extract_count += count;
  1500  #endif
  1501  	size_t want = count - extract_count;
  1502  	if (want > cache->count)
  1503  		want = cache->count;
  1504  
  1505  	memcpy(span + extract_count, cache->span + (cache->count - want), sizeof(span_t*) * want);
  1506  	cache->count -= (uint32_t)want;
  1507  	extract_count += want;
  1508  
  1509  	while ((extract_count < count) && cache->overflow) {
  1510  		span_t* current_span = cache->overflow;
  1511  		span[extract_count++] = current_span;
  1512  		cache->overflow = current_span->next;
  1513  	}
  1514  
  1515  #if ENABLE_ASSERTS
  1516  	for (size_t ispan = 0; ispan < extract_count; ++ispan) {
  1517  		assert(span[ispan]->span_count == span_count);
  1518  	}
  1519  #endif
  1520  
  1521  	atomic_store32_release(&cache->lock, 0);
  1522  
  1523  	return extract_count;
  1524  }
  1525  
  1526  #endif
  1527  
  1528  ////////////
  1529  ///
  1530  /// Heap control
  1531  ///
  1532  //////
  1533  
  1534  static void _rpmalloc_deallocate_huge(span_t*);
  1535  
  1536  //! Store the given spans as reserve in the given heap
  1537  static void
  1538  _rpmalloc_heap_set_reserved_spans(heap_t* heap, span_t* master, span_t* reserve, size_t reserve_span_count) {
  1539  	heap->span_reserve_master = master;
  1540  	heap->span_reserve = reserve;
  1541  	heap->spans_reserved = (uint32_t)reserve_span_count;
  1542  }
  1543  
  1544  //! Adopt the deferred span cache list, optionally extracting the first single span for immediate re-use
  1545  static void
  1546  _rpmalloc_heap_cache_adopt_deferred(heap_t* heap, span_t** single_span) {
  1547  	span_t* span = (span_t*)((void*)atomic_exchange_ptr_acquire(&heap->span_free_deferred, 0));
  1548  	while (span) {
  1549  		span_t* next_span = (span_t*)span->free_list;
  1550  		rpmalloc_assert(span->heap == heap, "Span heap pointer corrupted");
  1551  		if (EXPECTED(span->size_class < SIZE_CLASS_COUNT)) {
  1552  			rpmalloc_assert(heap->full_span_count, "Heap span counter corrupted");
  1553  			--heap->full_span_count;
  1554  			_rpmalloc_stat_dec(&heap->span_use[0].spans_deferred);
  1555  #if RPMALLOC_FIRST_CLASS_HEAPS
  1556  			_rpmalloc_span_double_link_list_remove(&heap->full_span[span->size_class], span);
  1557  #endif
  1558  			_rpmalloc_stat_dec(&heap->span_use[0].current);
  1559  			_rpmalloc_stat_dec(&heap->size_class_use[span->size_class].spans_current);
  1560  			if (single_span && !*single_span)
  1561  				*single_span = span;
  1562  			else
  1563  				_rpmalloc_heap_cache_insert(heap, span);
  1564  		} else {
  1565  			if (span->size_class == SIZE_CLASS_HUGE) {
  1566  				_rpmalloc_deallocate_huge(span);
  1567  			} else {
  1568  				rpmalloc_assert(span->size_class == SIZE_CLASS_LARGE, "Span size class invalid");
  1569  				rpmalloc_assert(heap->full_span_count, "Heap span counter corrupted");
  1570  				--heap->full_span_count;
  1571  #if RPMALLOC_FIRST_CLASS_HEAPS
  1572  				_rpmalloc_span_double_link_list_remove(&heap->large_huge_span, span);
  1573  #endif
  1574  				uint32_t idx = span->span_count - 1;
  1575  				_rpmalloc_stat_dec(&heap->span_use[idx].spans_deferred);
  1576  				_rpmalloc_stat_dec(&heap->span_use[idx].current);
  1577  				if (!idx && single_span && !*single_span)
  1578  					*single_span = span;
  1579  				else
  1580  					_rpmalloc_heap_cache_insert(heap, span);
  1581  			}
  1582  		}
  1583  		span = next_span;
  1584  	}
  1585  }
  1586  
  1587  static void
  1588  _rpmalloc_heap_unmap(heap_t* heap) {
  1589  	if (!heap->master_heap) {
  1590  		if ((heap->finalize > 1) && !atomic_load32(&heap->child_count)) {
  1591  			span_t* span = (span_t*)((uintptr_t)heap & _memory_span_mask);
  1592  			_rpmalloc_span_unmap(span);
  1593  		}
  1594  	} else {
  1595  		if (atomic_decr32(&heap->master_heap->child_count) == 0) {
  1596  			_rpmalloc_heap_unmap(heap->master_heap);
  1597  		}
  1598  	}
  1599  }
  1600  
  1601  static void
  1602  _rpmalloc_heap_global_finalize(heap_t* heap) {
  1603  	if (heap->finalize++ > 1) {
  1604  		--heap->finalize;
  1605  		return;
  1606  	}
  1607  
  1608  	_rpmalloc_heap_finalize(heap);
  1609  
  1610  #if ENABLE_THREAD_CACHE
  1611  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  1612  		span_cache_t* span_cache;
  1613  		if (!iclass)
  1614  			span_cache = &heap->span_cache;
  1615  		else
  1616  			span_cache = (span_cache_t*)(heap->span_large_cache + (iclass - 1));
  1617  		for (size_t ispan = 0; ispan < span_cache->count; ++ispan)
  1618  			_rpmalloc_span_unmap(span_cache->span[ispan]);
  1619  		span_cache->count = 0;
  1620  	}
  1621  #endif
  1622  
  1623  	if (heap->full_span_count) {
  1624  		--heap->finalize;
  1625  		return;
  1626  	}
  1627  
  1628  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  1629  		if (heap->size_class[iclass].free_list || heap->size_class[iclass].partial_span) {
  1630  			--heap->finalize;
  1631  			return;
  1632  		}
  1633  	}
  1634  	//Heap is now completely free, unmap and remove from heap list
  1635  	size_t list_idx = (size_t)heap->id % HEAP_ARRAY_SIZE;
  1636  	heap_t* list_heap = _memory_heaps[list_idx];
  1637  	if (list_heap == heap) {
  1638  		_memory_heaps[list_idx] = heap->next_heap;
  1639  	} else {
  1640  		while (list_heap->next_heap != heap)
  1641  			list_heap = list_heap->next_heap;
  1642  		list_heap->next_heap = heap->next_heap;
  1643  	}
  1644  
  1645  	_rpmalloc_heap_unmap(heap);
  1646  }
  1647  
  1648  //! Insert a single span into thread heap cache, releasing to global cache if overflow
  1649  static void
  1650  _rpmalloc_heap_cache_insert(heap_t* heap, span_t* span) {
  1651  	if (UNEXPECTED(heap->finalize != 0)) {
  1652  		_rpmalloc_span_unmap(span);
  1653  		_rpmalloc_heap_global_finalize(heap);
  1654  		return;
  1655  	}
  1656  #if ENABLE_THREAD_CACHE
  1657  	size_t span_count = span->span_count;
  1658  	_rpmalloc_stat_inc(&heap->span_use[span_count - 1].spans_to_cache);
  1659  	if (span_count == 1) {
  1660  		span_cache_t* span_cache = &heap->span_cache;
  1661  		span_cache->span[span_cache->count++] = span;
  1662  		if (span_cache->count == MAX_THREAD_SPAN_CACHE) {
  1663  			const size_t remain_count = MAX_THREAD_SPAN_CACHE - THREAD_SPAN_CACHE_TRANSFER;
  1664  #if ENABLE_GLOBAL_CACHE
  1665  			_rpmalloc_stat_add64(&heap->thread_to_global, THREAD_SPAN_CACHE_TRANSFER * _memory_span_size);
  1666  			_rpmalloc_stat_add(&heap->span_use[span_count - 1].spans_to_global, THREAD_SPAN_CACHE_TRANSFER);
  1667  			_rpmalloc_global_cache_insert_spans(span_cache->span + remain_count, span_count, THREAD_SPAN_CACHE_TRANSFER);
  1668  #else
  1669  			for (size_t ispan = 0; ispan < THREAD_SPAN_CACHE_TRANSFER; ++ispan)
  1670  				_rpmalloc_span_unmap(span_cache->span[remain_count + ispan]);
  1671  #endif
  1672  			span_cache->count = remain_count;
  1673  		}
  1674  	} else {
  1675  		size_t cache_idx = span_count - 2;
  1676  		span_large_cache_t* span_cache = heap->span_large_cache + cache_idx;
  1677  		span_cache->span[span_cache->count++] = span;
  1678  		const size_t cache_limit = (MAX_THREAD_SPAN_LARGE_CACHE - (span_count >> 1));
  1679  		if (span_cache->count == cache_limit) {
  1680  			const size_t transfer_limit = 2 + (cache_limit >> 2);
  1681  			const size_t transfer_count = (THREAD_SPAN_LARGE_CACHE_TRANSFER <= transfer_limit ? THREAD_SPAN_LARGE_CACHE_TRANSFER : transfer_limit);
  1682  			const size_t remain_count = cache_limit - transfer_count;
  1683  #if ENABLE_GLOBAL_CACHE
  1684  			_rpmalloc_stat_add64(&heap->thread_to_global, transfer_count * span_count * _memory_span_size);
  1685  			_rpmalloc_stat_add(&heap->span_use[span_count - 1].spans_to_global, transfer_count);
  1686  			_rpmalloc_global_cache_insert_spans(span_cache->span + remain_count, span_count, transfer_count);
  1687  #else
  1688  			for (size_t ispan = 0; ispan < transfer_count; ++ispan)
  1689  				_rpmalloc_span_unmap(span_cache->span[remain_count + ispan]);
  1690  #endif
  1691  			span_cache->count = remain_count;
  1692  		}
  1693  	}
  1694  #else
  1695  	(void)sizeof(heap);
  1696  	_rpmalloc_span_unmap(span);
  1697  #endif
  1698  }
  1699  
  1700  //! Extract the given number of spans from the different cache levels
  1701  static span_t*
  1702  _rpmalloc_heap_thread_cache_extract(heap_t* heap, size_t span_count) {
  1703  	span_t* span = 0;
  1704  #if ENABLE_THREAD_CACHE
  1705  	span_cache_t* span_cache;
  1706  	if (span_count == 1)
  1707  		span_cache = &heap->span_cache;
  1708  	else
  1709  		span_cache = (span_cache_t*)(heap->span_large_cache + (span_count - 2));
  1710  	if (span_cache->count) {
  1711  		_rpmalloc_stat_inc(&heap->span_use[span_count - 1].spans_from_cache);
  1712  		return span_cache->span[--span_cache->count];
  1713  	}
  1714  #endif
  1715  	return span;
  1716  }
  1717  
  1718  static span_t*
  1719  _rpmalloc_heap_thread_cache_deferred_extract(heap_t* heap, size_t span_count) {
  1720  	span_t* span = 0;
  1721  	if (span_count == 1) {
  1722  		_rpmalloc_heap_cache_adopt_deferred(heap, &span);
  1723  	} else {
  1724  		_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  1725  		span = _rpmalloc_heap_thread_cache_extract(heap, span_count);
  1726  	}
  1727  	return span;
  1728  }
  1729  
  1730  static span_t*
  1731  _rpmalloc_heap_reserved_extract(heap_t* heap, size_t span_count) {
  1732  	if (heap->spans_reserved >= span_count)
  1733  		return _rpmalloc_span_map(heap, span_count);
  1734  	return 0;
  1735  }
  1736  
  1737  //! Extract a span from the global cache
  1738  static span_t*
  1739  _rpmalloc_heap_global_cache_extract(heap_t* heap, size_t span_count) {
  1740  #if ENABLE_GLOBAL_CACHE
  1741  #if ENABLE_THREAD_CACHE
  1742  	span_cache_t* span_cache;
  1743  	size_t wanted_count;
  1744  	if (span_count == 1) {
  1745  		span_cache = &heap->span_cache;
  1746  		wanted_count = THREAD_SPAN_CACHE_TRANSFER;
  1747  	} else {
  1748  		span_cache = (span_cache_t*)(heap->span_large_cache + (span_count - 2));
  1749  		wanted_count = THREAD_SPAN_LARGE_CACHE_TRANSFER;
  1750  	}
  1751  	span_cache->count = _rpmalloc_global_cache_extract_spans(span_cache->span, span_count, wanted_count);
  1752  	if (span_cache->count) {
  1753  		_rpmalloc_stat_add64(&heap->global_to_thread, span_count * span_cache->count * _memory_span_size);
  1754  		_rpmalloc_stat_add(&heap->span_use[span_count - 1].spans_from_global, span_cache->count);
  1755  		return span_cache->span[--span_cache->count];
  1756  	}
  1757  #else
  1758  	span_t* span = 0;
  1759  	size_t count = _rpmalloc_global_cache_extract_spans(&span, span_count, 1);
  1760  	if (count) {
  1761  		_rpmalloc_stat_add64(&heap->global_to_thread, span_count * count * _memory_span_size);
  1762  		_rpmalloc_stat_add(&heap->span_use[span_count - 1].spans_from_global, count);
  1763  		return span;
  1764  	}
  1765  #endif
  1766  #endif
  1767  	(void)sizeof(heap);
  1768  	(void)sizeof(span_count);
  1769  	return 0;
  1770  }
  1771  
  1772  static void
  1773  _rpmalloc_inc_span_statistics(heap_t* heap, size_t span_count, uint32_t class_idx) {
  1774  	(void)sizeof(heap);
  1775  	(void)sizeof(span_count);
  1776  	(void)sizeof(class_idx);
  1777  #if ENABLE_ADAPTIVE_THREAD_CACHE || ENABLE_STATISTICS
  1778  	uint32_t idx = (uint32_t)span_count - 1;
  1779  	uint32_t current_count = (uint32_t)atomic_incr32(&heap->span_use[idx].current);
  1780  	if (current_count > (uint32_t)atomic_load32(&heap->span_use[idx].high))
  1781  		atomic_store32(&heap->span_use[idx].high, (int32_t)current_count);
  1782  	_rpmalloc_stat_add_peak(&heap->size_class_use[class_idx].spans_current, 1, heap->size_class_use[class_idx].spans_peak);
  1783  #endif
  1784  }
  1785  
  1786  //! Get a span from one of the cache levels (thread cache, reserved, global cache) or fallback to mapping more memory
  1787  static span_t*
  1788  _rpmalloc_heap_extract_new_span(heap_t* heap, heap_size_class_t* heap_size_class, size_t span_count, uint32_t class_idx) {
  1789  	span_t* span;
  1790  #if ENABLE_THREAD_CACHE
  1791  	if (heap_size_class && heap_size_class->cache) {
  1792  		span = heap_size_class->cache;
  1793  		heap_size_class->cache = (heap->span_cache.count ? heap->span_cache.span[--heap->span_cache.count] : 0);
  1794  		_rpmalloc_inc_span_statistics(heap, span_count, class_idx);
  1795  		return span;
  1796  	}
  1797  #endif
  1798  	(void)sizeof(class_idx);
  1799  	// Allow 50% overhead to increase cache hits
  1800  	size_t base_span_count = span_count;
  1801  	size_t limit_span_count = (span_count > 2) ? (span_count + (span_count >> 1)) : span_count;
  1802  	if (limit_span_count > LARGE_CLASS_COUNT)
  1803  		limit_span_count = LARGE_CLASS_COUNT;
  1804  	do {
  1805  		span = _rpmalloc_heap_thread_cache_extract(heap, span_count);
  1806  		if (EXPECTED(span != 0)) {
  1807  			_rpmalloc_stat_inc(&heap->size_class_use[class_idx].spans_from_cache);
  1808  			_rpmalloc_inc_span_statistics(heap, span_count, class_idx);
  1809  			return span;
  1810  		}
  1811  		span = _rpmalloc_heap_thread_cache_deferred_extract(heap, span_count);
  1812  		if (EXPECTED(span != 0)) {
  1813  			_rpmalloc_stat_inc(&heap->size_class_use[class_idx].spans_from_cache);
  1814  			_rpmalloc_inc_span_statistics(heap, span_count, class_idx);
  1815  			return span;
  1816  		}
  1817  		span = _rpmalloc_heap_reserved_extract(heap, span_count);
  1818  		if (EXPECTED(span != 0)) {
  1819  			_rpmalloc_stat_inc(&heap->size_class_use[class_idx].spans_from_reserved);
  1820  			_rpmalloc_inc_span_statistics(heap, span_count, class_idx);
  1821  			return span;
  1822  		}
  1823  		span = _rpmalloc_heap_global_cache_extract(heap, span_count);
  1824  		if (EXPECTED(span != 0)) {
  1825  			_rpmalloc_stat_inc(&heap->size_class_use[class_idx].spans_from_cache);
  1826  			_rpmalloc_inc_span_statistics(heap, span_count, class_idx);
  1827  			return span;
  1828  		}
  1829  		++span_count;
  1830  	} while (span_count <= limit_span_count);
  1831  	//Final fallback, map in more virtual memory
  1832  	span = _rpmalloc_span_map(heap, base_span_count);
  1833  	_rpmalloc_inc_span_statistics(heap, base_span_count, class_idx);
  1834  	_rpmalloc_stat_inc(&heap->size_class_use[class_idx].spans_map_calls);
  1835  	return span;
  1836  }
  1837  
  1838  static void
  1839  _rpmalloc_heap_initialize(heap_t* heap) {
  1840  	memset(heap, 0, sizeof(heap_t));
  1841  	//Get a new heap ID
  1842  	heap->id = 1 + atomic_incr32(&_memory_heap_id);
  1843  
  1844  	//Link in heap in heap ID map
  1845  	size_t list_idx = (size_t)heap->id % HEAP_ARRAY_SIZE;
  1846  	heap->next_heap = _memory_heaps[list_idx];
  1847  	_memory_heaps[list_idx] = heap;
  1848  }
  1849  
  1850  static void
  1851  _rpmalloc_heap_orphan(heap_t* heap, int first_class) {
  1852  	heap->owner_thread = (uintptr_t)-1;
  1853  #if RPMALLOC_FIRST_CLASS_HEAPS
  1854  	heap_t** heap_list = (first_class ? &_memory_first_class_orphan_heaps : &_memory_orphan_heaps);
  1855  #else
  1856  	(void)sizeof(first_class);
  1857  	heap_t** heap_list = &_memory_orphan_heaps;
  1858  #endif
  1859  	heap->next_orphan = *heap_list;
  1860  	*heap_list = heap;
  1861  }
  1862  
  1863  //! Allocate a new heap from newly mapped memory pages
  1864  static heap_t*
  1865  _rpmalloc_heap_allocate_new(void) {
  1866  	// Map in pages for a 16 heaps. If page size is greater than required size for this, map a page and
  1867  	// use first part for heaps and remaining part for spans for allocations. Adds a lot of complexity,
  1868  	// but saves a lot of memory on systems where page size > 64 spans (4MiB)
  1869  	size_t heap_size = sizeof(heap_t);
  1870  	size_t aligned_heap_size = 16 * ((heap_size + 15) / 16);
  1871  	size_t request_heap_count = 16;
  1872  	size_t heap_span_count = ((aligned_heap_size * request_heap_count) + sizeof(span_t) + _memory_span_size - 1) / _memory_span_size;
  1873  	size_t block_size = _memory_span_size * heap_span_count;
  1874  	size_t span_count = heap_span_count;
  1875  	span_t* span = 0;
  1876  	// If there are global reserved spans, use these first
  1877  	if (_memory_global_reserve_count >= heap_span_count) {
  1878  		span = _rpmalloc_global_get_reserved_spans(heap_span_count);
  1879  	}
  1880  	if (!span) {
  1881  		if (_memory_page_size > block_size) {
  1882  			span_count = _memory_page_size / _memory_span_size;
  1883  			block_size = _memory_page_size;
  1884  			// If using huge pages, make sure to grab enough heaps to avoid reallocating a huge page just to serve new heaps
  1885  			size_t possible_heap_count = (block_size - sizeof(span_t)) / aligned_heap_size;
  1886  			if (possible_heap_count >= (request_heap_count * 16))
  1887  				request_heap_count *= 16;
  1888  			else if (possible_heap_count < request_heap_count)
  1889  				request_heap_count = possible_heap_count;
  1890  			heap_span_count = ((aligned_heap_size * request_heap_count) + sizeof(span_t) + _memory_span_size - 1) / _memory_span_size;
  1891  		}
  1892  
  1893  		size_t align_offset = 0;
  1894  		span = (span_t*)_rpmalloc_mmap(block_size, &align_offset);
  1895  		if (!span)
  1896  			return 0;
  1897  
  1898  		// Master span will contain the heaps
  1899  		_rpmalloc_stat_inc(&_master_spans);
  1900  		_rpmalloc_span_initialize(span, span_count, heap_span_count, align_offset);
  1901  	}
  1902  
  1903  	size_t remain_size = _memory_span_size - sizeof(span_t);
  1904  	heap_t* heap = (heap_t*)pointer_offset(span, sizeof(span_t));
  1905  	_rpmalloc_heap_initialize(heap);
  1906  
  1907  	// Put extra heaps as orphans
  1908  	size_t num_heaps = remain_size / aligned_heap_size;
  1909  	if (num_heaps < request_heap_count)
  1910  		num_heaps = request_heap_count;
  1911  	atomic_store32(&heap->child_count, (int32_t)num_heaps - 1);
  1912  	heap_t* extra_heap = (heap_t*)pointer_offset(heap, aligned_heap_size);
  1913  	while (num_heaps > 1) {
  1914  		_rpmalloc_heap_initialize(extra_heap);
  1915  		extra_heap->master_heap = heap;
  1916  		_rpmalloc_heap_orphan(extra_heap, 1);
  1917  		extra_heap = (heap_t*)pointer_offset(extra_heap, aligned_heap_size);
  1918  		--num_heaps;
  1919  	}
  1920  
  1921  	if (span_count > heap_span_count) {
  1922  		// Cap reserved spans
  1923  		size_t remain_count = span_count - heap_span_count;
  1924  		size_t reserve_count = (remain_count > _memory_heap_reserve_count ? _memory_heap_reserve_count : remain_count);
  1925  		span_t* remain_span = (span_t*)pointer_offset(span, heap_span_count * _memory_span_size);
  1926  		_rpmalloc_heap_set_reserved_spans(heap, span, remain_span, reserve_count);
  1927  
  1928  		if (remain_count > reserve_count) {
  1929  			// Set to global reserved spans
  1930  			remain_span = (span_t*)pointer_offset(remain_span, reserve_count * _memory_span_size);
  1931  			reserve_count = remain_count - reserve_count;
  1932  			_rpmalloc_global_set_reserved_spans(span, remain_span, reserve_count);
  1933  		}
  1934  	}
  1935  
  1936  	return heap;
  1937  }
  1938  
  1939  static heap_t*
  1940  _rpmalloc_heap_extract_orphan(heap_t** heap_list) {
  1941  	heap_t* heap = *heap_list;
  1942  	*heap_list = (heap ? heap->next_orphan : 0);
  1943  	return heap;
  1944  }
  1945  
  1946  //! Allocate a new heap, potentially reusing a previously orphaned heap
  1947  static heap_t*
  1948  _rpmalloc_heap_allocate(int first_class) {
  1949  	heap_t* heap = 0;
  1950  	while (!atomic_cas32_acquire(&_memory_global_lock, 1, 0))
  1951  		_rpmalloc_spin();
  1952  	if (first_class == 0)
  1953  		heap = _rpmalloc_heap_extract_orphan(&_memory_orphan_heaps);
  1954  #if RPMALLOC_FIRST_CLASS_HEAPS
  1955  	if (!heap)
  1956  		heap = _rpmalloc_heap_extract_orphan(&_memory_first_class_orphan_heaps);
  1957  #endif
  1958  	if (!heap)
  1959  		heap = _rpmalloc_heap_allocate_new();
  1960  	atomic_store32_release(&_memory_global_lock, 0);
  1961  	_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  1962  	return heap;
  1963  }
  1964  
  1965  static void
  1966  _rpmalloc_heap_release(void* heapptr, int first_class, int release_cache) {
  1967  	heap_t* heap = (heap_t*)heapptr;
  1968  	if (!heap)
  1969  		return;
  1970  	//Release thread cache spans back to global cache
  1971  	_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  1972  	if (release_cache  || heap->finalize) {
  1973  #if ENABLE_THREAD_CACHE
  1974  		for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  1975  			span_cache_t* span_cache;
  1976  			if (!iclass)
  1977  				span_cache = &heap->span_cache;
  1978  			else
  1979  				span_cache = (span_cache_t*)(heap->span_large_cache + (iclass - 1));
  1980  			if (!span_cache->count)
  1981  				continue;
  1982  #if ENABLE_GLOBAL_CACHE
  1983  			if (heap->finalize) {
  1984  				for (size_t ispan = 0; ispan < span_cache->count; ++ispan)
  1985  					_rpmalloc_span_unmap(span_cache->span[ispan]);
  1986  			} else {
  1987  				_rpmalloc_stat_add64(&heap->thread_to_global, span_cache->count * (iclass + 1) * _memory_span_size);
  1988  				_rpmalloc_stat_add(&heap->span_use[iclass].spans_to_global, span_cache->count);
  1989  				_rpmalloc_global_cache_insert_spans(span_cache->span, iclass + 1, span_cache->count);
  1990  			}
  1991  #else
  1992  			for (size_t ispan = 0; ispan < span_cache->count; ++ispan)
  1993  				_rpmalloc_span_unmap(span_cache->span[ispan]);
  1994  #endif
  1995  			span_cache->count = 0;
  1996  		}
  1997  #endif
  1998  	}
  1999  
  2000  	if (get_thread_heap_raw() == heap)
  2001  		set_thread_heap(0);
  2002  
  2003  #if ENABLE_STATISTICS
  2004  	atomic_decr32(&_memory_active_heaps);
  2005  	rpmalloc_assert(atomic_load32(&_memory_active_heaps) >= 0, "Still active heaps during finalization");
  2006  #endif
  2007  
  2008  	// If we are forcibly terminating with _exit the state of the
  2009  	// lock atomic is unknown and it's best to just go ahead and exit
  2010  	if (get_thread_id() != _rpmalloc_main_thread_id) {
  2011  		while (!atomic_cas32_acquire(&_memory_global_lock, 1, 0))
  2012  			_rpmalloc_spin();
  2013  	}
  2014  	_rpmalloc_heap_orphan(heap, first_class);
  2015  	atomic_store32_release(&_memory_global_lock, 0);
  2016  }
  2017  
  2018  static void
  2019  _rpmalloc_heap_release_raw(void* heapptr, int release_cache) {
  2020  	_rpmalloc_heap_release(heapptr, 0, release_cache);
  2021  }
  2022  
  2023  static void
  2024  _rpmalloc_heap_release_raw_fc(void* heapptr) {
  2025  	_rpmalloc_heap_release_raw(heapptr, 1);
  2026  }
  2027  
  2028  static void
  2029  _rpmalloc_heap_finalize(heap_t* heap) {
  2030  	if (heap->spans_reserved) {
  2031  		span_t* span = _rpmalloc_span_map(heap, heap->spans_reserved);
  2032  		_rpmalloc_span_unmap(span);
  2033  		heap->spans_reserved = 0;
  2034  	}
  2035  
  2036  	_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  2037  
  2038  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  2039  		if (heap->size_class[iclass].cache)
  2040  			_rpmalloc_span_unmap(heap->size_class[iclass].cache);
  2041  		heap->size_class[iclass].cache = 0;
  2042  		span_t* span = heap->size_class[iclass].partial_span;
  2043  		while (span) {
  2044  			span_t* next = span->next;
  2045  			_rpmalloc_span_finalize(heap, iclass, span, &heap->size_class[iclass].partial_span);
  2046  			span = next;
  2047  		}
  2048  		// If class still has a free list it must be a full span
  2049  		if (heap->size_class[iclass].free_list) {
  2050  			span_t* class_span = (span_t*)((uintptr_t)heap->size_class[iclass].free_list & _memory_span_mask);
  2051  			span_t** list = 0;
  2052  #if RPMALLOC_FIRST_CLASS_HEAPS
  2053  			list = &heap->full_span[iclass];
  2054  #endif
  2055  			--heap->full_span_count;
  2056  			if (!_rpmalloc_span_finalize(heap, iclass, class_span, list)) {
  2057  				if (list)
  2058  					_rpmalloc_span_double_link_list_remove(list, class_span);
  2059  				_rpmalloc_span_double_link_list_add(&heap->size_class[iclass].partial_span, class_span);
  2060  			}
  2061  		}
  2062  	}
  2063  
  2064  #if ENABLE_THREAD_CACHE
  2065  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  2066  		span_cache_t* span_cache;
  2067  		if (!iclass)
  2068  			span_cache = &heap->span_cache;
  2069  		else
  2070  			span_cache = (span_cache_t*)(heap->span_large_cache + (iclass - 1));
  2071  		for (size_t ispan = 0; ispan < span_cache->count; ++ispan)
  2072  			_rpmalloc_span_unmap(span_cache->span[ispan]);
  2073  		span_cache->count = 0;
  2074  	}
  2075  #endif
  2076  	rpmalloc_assert(!atomic_load_ptr(&heap->span_free_deferred), "Heaps still active during finalization");
  2077  }
  2078  
  2079  
  2080  ////////////
  2081  ///
  2082  /// Allocation entry points
  2083  ///
  2084  //////
  2085  
  2086  //! Pop first block from a free list
  2087  static void*
  2088  free_list_pop(void** list) {
  2089  	void* block = *list;
  2090  	*list = *((void**)block);
  2091  	return block;
  2092  }
  2093  
  2094  //! Allocate a small/medium sized memory block from the given heap
  2095  static void*
  2096  _rpmalloc_allocate_from_heap_fallback(heap_t* heap, heap_size_class_t* heap_size_class, uint32_t class_idx) {
  2097  	span_t* span = heap_size_class->partial_span;
  2098  	if (EXPECTED(span != 0)) {
  2099  		rpmalloc_assert(span->block_count == _memory_size_class[span->size_class].block_count, "Span block count corrupted");
  2100  		rpmalloc_assert(!_rpmalloc_span_is_fully_utilized(span), "Internal failure");
  2101  		void* block;
  2102  		if (span->free_list) {
  2103  			//Span local free list is not empty, swap to size class free list
  2104  			block = free_list_pop(&span->free_list);
  2105  			heap_size_class->free_list = span->free_list;
  2106  			span->free_list = 0;
  2107  		} else {
  2108  			//If the span did not fully initialize free list, link up another page worth of blocks
  2109  			void* block_start = pointer_offset(span, SPAN_HEADER_SIZE + ((size_t)span->free_list_limit * span->block_size));
  2110  			span->free_list_limit += free_list_partial_init(&heap_size_class->free_list, &block,
  2111  				(void*)((uintptr_t)block_start & ~(_memory_page_size - 1)), block_start,
  2112  				span->block_count - span->free_list_limit, span->block_size);
  2113  		}
  2114  		rpmalloc_assert(span->free_list_limit <= span->block_count, "Span block count corrupted");
  2115  		span->used_count = span->free_list_limit;
  2116  
  2117  		//Swap in deferred free list if present
  2118  		if (atomic_load_ptr(&span->free_list_deferred))
  2119  			_rpmalloc_span_extract_free_list_deferred(span);
  2120  
  2121  		//If span is still not fully utilized keep it in partial list and early return block
  2122  		if (!_rpmalloc_span_is_fully_utilized(span))
  2123  			return block;
  2124  
  2125  		//The span is fully utilized, unlink from partial list and add to fully utilized list
  2126  		_rpmalloc_span_double_link_list_pop_head(&heap_size_class->partial_span, span);
  2127  #if RPMALLOC_FIRST_CLASS_HEAPS
  2128  		_rpmalloc_span_double_link_list_add(&heap->full_span[class_idx], span);
  2129  #endif
  2130  		++heap->full_span_count;
  2131  		return block;
  2132  	}
  2133  
  2134  	//Find a span in one of the cache levels
  2135  	span = _rpmalloc_heap_extract_new_span(heap, heap_size_class, 1, class_idx);
  2136  	if (EXPECTED(span != 0)) {
  2137  		//Mark span as owned by this heap and set base data, return first block
  2138  		return _rpmalloc_span_initialize_new(heap, heap_size_class, span, class_idx);
  2139  	}
  2140  
  2141  	return 0;
  2142  }
  2143  
  2144  //! Allocate a small sized memory block from the given heap
  2145  static void*
  2146  _rpmalloc_allocate_small(heap_t* heap, size_t size) {
  2147  	rpmalloc_assert(heap, "No thread heap");
  2148  	//Small sizes have unique size classes
  2149  	const uint32_t class_idx = (uint32_t)((size + (SMALL_GRANULARITY - 1)) >> SMALL_GRANULARITY_SHIFT);
  2150  	heap_size_class_t* heap_size_class = heap->size_class + class_idx;
  2151  	_rpmalloc_stat_inc_alloc(heap, class_idx);
  2152  	if (EXPECTED(heap_size_class->free_list != 0))
  2153  		return free_list_pop(&heap_size_class->free_list);
  2154  	return _rpmalloc_allocate_from_heap_fallback(heap, heap_size_class, class_idx);
  2155  }
  2156  
  2157  //! Allocate a medium sized memory block from the given heap
  2158  static void*
  2159  _rpmalloc_allocate_medium(heap_t* heap, size_t size) {
  2160  	rpmalloc_assert(heap, "No thread heap");
  2161  	//Calculate the size class index and do a dependent lookup of the final class index (in case of merged classes)
  2162  	const uint32_t base_idx = (uint32_t)(SMALL_CLASS_COUNT + ((size - (SMALL_SIZE_LIMIT + 1)) >> MEDIUM_GRANULARITY_SHIFT));
  2163  	const uint32_t class_idx = _memory_size_class[base_idx].class_idx;
  2164  	heap_size_class_t* heap_size_class = heap->size_class + class_idx;
  2165  	_rpmalloc_stat_inc_alloc(heap, class_idx);
  2166  	if (EXPECTED(heap_size_class->free_list != 0))
  2167  		return free_list_pop(&heap_size_class->free_list);
  2168  	return _rpmalloc_allocate_from_heap_fallback(heap, heap_size_class, class_idx);
  2169  }
  2170  
  2171  //! Allocate a large sized memory block from the given heap
  2172  static void*
  2173  _rpmalloc_allocate_large(heap_t* heap, size_t size) {
  2174  	rpmalloc_assert(heap, "No thread heap");
  2175  	//Calculate number of needed max sized spans (including header)
  2176  	//Since this function is never called if size > LARGE_SIZE_LIMIT
  2177  	//the span_count is guaranteed to be <= LARGE_CLASS_COUNT
  2178  	size += SPAN_HEADER_SIZE;
  2179  	size_t span_count = size >> _memory_span_size_shift;
  2180  	if (size & (_memory_span_size - 1))
  2181  		++span_count;
  2182  
  2183  	//Find a span in one of the cache levels
  2184  	span_t* span = _rpmalloc_heap_extract_new_span(heap, 0, span_count, SIZE_CLASS_LARGE);
  2185  	if (!span)
  2186  		return span;
  2187  
  2188  	//Mark span as owned by this heap and set base data
  2189  	rpmalloc_assert(span->span_count >= span_count, "Internal failure");
  2190  	span->size_class = SIZE_CLASS_LARGE;
  2191  	span->heap = heap;
  2192  
  2193  #if RPMALLOC_FIRST_CLASS_HEAPS
  2194  	_rpmalloc_span_double_link_list_add(&heap->large_huge_span, span);
  2195  #endif
  2196  	++heap->full_span_count;
  2197  
  2198  	return pointer_offset(span, SPAN_HEADER_SIZE);
  2199  }
  2200  
  2201  //! Allocate a huge block by mapping memory pages directly
  2202  static void*
  2203  _rpmalloc_allocate_huge(heap_t* heap, size_t size) {
  2204  	rpmalloc_assert(heap, "No thread heap");
  2205  	_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  2206  	size += SPAN_HEADER_SIZE;
  2207  	size_t num_pages = size >> _memory_page_size_shift;
  2208  	if (size & (_memory_page_size - 1))
  2209  		++num_pages;
  2210  	size_t align_offset = 0;
  2211  	span_t* span = (span_t*)_rpmalloc_mmap(num_pages * _memory_page_size, &align_offset);
  2212  	if (!span)
  2213  		return span;
  2214  
  2215  	//Store page count in span_count
  2216  	span->size_class = SIZE_CLASS_HUGE;
  2217  	span->span_count = (uint32_t)num_pages;
  2218  	span->align_offset = (uint32_t)align_offset;
  2219  	span->heap = heap;
  2220  	_rpmalloc_stat_add_peak(&_huge_pages_current, num_pages, _huge_pages_peak);
  2221  
  2222  #if RPMALLOC_FIRST_CLASS_HEAPS
  2223  	_rpmalloc_span_double_link_list_add(&heap->large_huge_span, span);
  2224  #endif
  2225  	++heap->full_span_count;
  2226  
  2227  	return pointer_offset(span, SPAN_HEADER_SIZE);
  2228  }
  2229  
  2230  //! Allocate a block of the given size
  2231  static void*
  2232  _rpmalloc_allocate(heap_t* heap, size_t size) {
  2233  	_rpmalloc_stat_add64(&_allocation_counter, 1);
  2234  	if (EXPECTED(size <= SMALL_SIZE_LIMIT))
  2235  		return _rpmalloc_allocate_small(heap, size);
  2236  	else if (size <= _memory_medium_size_limit)
  2237  		return _rpmalloc_allocate_medium(heap, size);
  2238  	else if (size <= LARGE_SIZE_LIMIT)
  2239  		return _rpmalloc_allocate_large(heap, size);
  2240  	return _rpmalloc_allocate_huge(heap, size);
  2241  }
  2242  
  2243  static void*
  2244  _rpmalloc_aligned_allocate(heap_t* heap, size_t alignment, size_t size) {
  2245  	if (alignment <= SMALL_GRANULARITY)
  2246  		return _rpmalloc_allocate(heap, size);
  2247  
  2248  #if ENABLE_VALIDATE_ARGS
  2249  	if ((size + alignment) < size) {
  2250  		errno = EINVAL;
  2251  		return 0;
  2252  	}
  2253  	if (alignment & (alignment - 1)) {
  2254  		errno = EINVAL;
  2255  		return 0;
  2256  	}
  2257  #endif
  2258  
  2259  	if ((alignment <= SPAN_HEADER_SIZE) && (size < _memory_medium_size_limit)) {
  2260  		// If alignment is less or equal to span header size (which is power of two),
  2261  		// and size aligned to span header size multiples is less than size + alignment,
  2262  		// then use natural alignment of blocks to provide alignment
  2263  		size_t multiple_size = size ? (size + (SPAN_HEADER_SIZE - 1)) & ~(uintptr_t)(SPAN_HEADER_SIZE - 1) : SPAN_HEADER_SIZE;
  2264  		rpmalloc_assert(!(multiple_size % SPAN_HEADER_SIZE), "Failed alignment calculation");
  2265  		if (multiple_size <= (size + alignment))
  2266  			return _rpmalloc_allocate(heap, multiple_size);
  2267  	}
  2268  
  2269  	void* ptr = 0;
  2270  	size_t align_mask = alignment - 1;
  2271  	if (alignment <= _memory_page_size) {
  2272  		ptr = _rpmalloc_allocate(heap, size + alignment);
  2273  		if ((uintptr_t)ptr & align_mask) {
  2274  			ptr = (void*)(((uintptr_t)ptr & ~(uintptr_t)align_mask) + alignment);
  2275  			//Mark as having aligned blocks
  2276  			span_t* span = (span_t*)((uintptr_t)ptr & _memory_span_mask);
  2277  			span->flags |= SPAN_FLAG_ALIGNED_BLOCKS;
  2278  		}
  2279  		return ptr;
  2280  	}
  2281  
  2282  	// Fallback to mapping new pages for this request. Since pointers passed
  2283  	// to rpfree must be able to reach the start of the span by bitmasking of
  2284  	// the address with the span size, the returned aligned pointer from this
  2285  	// function must be with a span size of the start of the mapped area.
  2286  	// In worst case this requires us to loop and map pages until we get a
  2287  	// suitable memory address. It also means we can never align to span size
  2288  	// or greater, since the span header will push alignment more than one
  2289  	// span size away from span start (thus causing pointer mask to give us
  2290  	// an invalid span start on free)
  2291  	if (alignment & align_mask) {
  2292  		errno = EINVAL;
  2293  		return 0;
  2294  	}
  2295  	if (alignment >= _memory_span_size) {
  2296  		errno = EINVAL;
  2297  		return 0;
  2298  	}
  2299  
  2300  	size_t extra_pages = alignment / _memory_page_size;
  2301  
  2302  	// Since each span has a header, we will at least need one extra memory page
  2303  	size_t num_pages = 1 + (size / _memory_page_size);
  2304  	if (size & (_memory_page_size - 1))
  2305  		++num_pages;
  2306  
  2307  	if (extra_pages > num_pages)
  2308  		num_pages = 1 + extra_pages;
  2309  
  2310  	size_t original_pages = num_pages;
  2311  	size_t limit_pages = (_memory_span_size / _memory_page_size) * 2;
  2312  	if (limit_pages < (original_pages * 2))
  2313  		limit_pages = original_pages * 2;
  2314  
  2315  	size_t mapped_size, align_offset;
  2316  	span_t* span;
  2317  
  2318  retry:
  2319  	align_offset = 0;
  2320  	mapped_size = num_pages * _memory_page_size;
  2321  
  2322  	span = (span_t*)_rpmalloc_mmap(mapped_size, &align_offset);
  2323  	if (!span) {
  2324  		errno = ENOMEM;
  2325  		return 0;
  2326  	}
  2327  	ptr = pointer_offset(span, SPAN_HEADER_SIZE);
  2328  
  2329  	if ((uintptr_t)ptr & align_mask)
  2330  		ptr = (void*)(((uintptr_t)ptr & ~(uintptr_t)align_mask) + alignment);
  2331  
  2332  	if (((size_t)pointer_diff(ptr, span) >= _memory_span_size) ||
  2333  	    (pointer_offset(ptr, size) > pointer_offset(span, mapped_size)) ||
  2334  	    (((uintptr_t)ptr & _memory_span_mask) != (uintptr_t)span)) {
  2335  		_rpmalloc_unmap(span, mapped_size, align_offset, mapped_size);
  2336  		++num_pages;
  2337  		if (num_pages > limit_pages) {
  2338  			errno = EINVAL;
  2339  			return 0;
  2340  		}
  2341  		goto retry;
  2342  	}
  2343  
  2344  	//Store page count in span_count
  2345  	span->size_class = SIZE_CLASS_HUGE;
  2346  	span->span_count = (uint32_t)num_pages;
  2347  	span->align_offset = (uint32_t)align_offset;
  2348  	span->heap = heap;
  2349  	_rpmalloc_stat_add_peak(&_huge_pages_current, num_pages, _huge_pages_peak);
  2350  
  2351  #if RPMALLOC_FIRST_CLASS_HEAPS
  2352  	_rpmalloc_span_double_link_list_add(&heap->large_huge_span, span);
  2353  #endif
  2354  	++heap->full_span_count;
  2355  
  2356  	_rpmalloc_stat_add64(&_allocation_counter, 1);
  2357  
  2358  	return ptr;
  2359  }
  2360  
  2361  
  2362  ////////////
  2363  ///
  2364  /// Deallocation entry points
  2365  ///
  2366  //////
  2367  
  2368  //! Deallocate the given small/medium memory block in the current thread local heap
  2369  static void
  2370  _rpmalloc_deallocate_direct_small_or_medium(span_t* span, void* block) {
  2371  	heap_t* heap = span->heap;
  2372  	rpmalloc_assert(heap->owner_thread == get_thread_id() || !heap->owner_thread || heap->finalize, "Internal failure");
  2373  	//Add block to free list
  2374  	if (UNEXPECTED(_rpmalloc_span_is_fully_utilized(span))) {
  2375  		span->used_count = span->block_count;
  2376  #if RPMALLOC_FIRST_CLASS_HEAPS
  2377  		_rpmalloc_span_double_link_list_remove(&heap->full_span[span->size_class], span);
  2378  #endif
  2379  		_rpmalloc_span_double_link_list_add(&heap->size_class[span->size_class].partial_span, span);
  2380  		--heap->full_span_count;
  2381  	}
  2382  	*((void**)block) = span->free_list;
  2383  	--span->used_count;
  2384  	span->free_list = block;
  2385  	if (UNEXPECTED(span->used_count == span->list_size)) {
  2386  		// If there are no used blocks it is guaranteed that no other external thread is accessing the span
  2387  		if (span->used_count) {
  2388  			// Make sure we have synchronized the deferred list and list size by using acquire semantics
  2389  			// and guarantee that no external thread is accessing span concurrently
  2390  			void* free_list;
  2391  			do {
  2392  				free_list = atomic_exchange_ptr_acquire(&span->free_list_deferred, INVALID_POINTER);
  2393  			} while (free_list == INVALID_POINTER);
  2394  			atomic_store_ptr_release(&span->free_list_deferred, free_list);
  2395  		}
  2396  		_rpmalloc_span_double_link_list_remove(&heap->size_class[span->size_class].partial_span, span);
  2397  		_rpmalloc_span_release_to_cache(heap, span);
  2398  	}
  2399  }
  2400  
  2401  static void
  2402  _rpmalloc_deallocate_defer_free_span(heap_t* heap, span_t* span) {
  2403  	if (span->size_class != SIZE_CLASS_HUGE)
  2404  		_rpmalloc_stat_inc(&heap->span_use[span->span_count - 1].spans_deferred);
  2405  	//This list does not need ABA protection, no mutable side state
  2406  	do {
  2407  		span->free_list = (void*)atomic_load_ptr(&heap->span_free_deferred);
  2408  	} while (!atomic_cas_ptr(&heap->span_free_deferred, span, span->free_list));
  2409  }
  2410  
  2411  //! Put the block in the deferred free list of the owning span
  2412  static void
  2413  _rpmalloc_deallocate_defer_small_or_medium(span_t* span, void* block) {
  2414  	// The memory ordering here is a bit tricky, to avoid having to ABA protect
  2415  	// the deferred free list to avoid desynchronization of list and list size
  2416  	// we need to have acquire semantics on successful CAS of the pointer to
  2417  	// guarantee the list_size variable validity + release semantics on pointer store
  2418  	void* free_list;
  2419  	do {
  2420  		free_list = atomic_exchange_ptr_acquire(&span->free_list_deferred, INVALID_POINTER);
  2421  	} while (free_list == INVALID_POINTER);
  2422  	*((void**)block) = free_list;
  2423  	uint32_t free_count = ++span->list_size;
  2424  	int all_deferred_free = (free_count == span->block_count);
  2425  	atomic_store_ptr_release(&span->free_list_deferred, block);
  2426  	if (all_deferred_free) {
  2427  		// Span was completely freed by this block. Due to the INVALID_POINTER spin lock
  2428  		// no other thread can reach this state simultaneously on this span.
  2429  		// Safe to move to owner heap deferred cache
  2430  		_rpmalloc_deallocate_defer_free_span(span->heap, span);
  2431  	}
  2432  }
  2433  
  2434  static void
  2435  _rpmalloc_deallocate_small_or_medium(span_t* span, void* p) {
  2436  	_rpmalloc_stat_inc_free(span->heap, span->size_class);
  2437  	if (span->flags & SPAN_FLAG_ALIGNED_BLOCKS) {
  2438  		//Realign pointer to block start
  2439  		void* blocks_start = pointer_offset(span, SPAN_HEADER_SIZE);
  2440  		uint32_t block_offset = (uint32_t)pointer_diff(p, blocks_start);
  2441  		p = pointer_offset(p, -(int32_t)(block_offset % span->block_size));
  2442  	}
  2443  	//Check if block belongs to this heap or if deallocation should be deferred
  2444  #if RPMALLOC_FIRST_CLASS_HEAPS
  2445  	int defer = (span->heap->owner_thread && (span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2446  #else
  2447  	int defer = ((span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2448  #endif
  2449  	if (!defer)
  2450  		_rpmalloc_deallocate_direct_small_or_medium(span, p);
  2451  	else
  2452  		_rpmalloc_deallocate_defer_small_or_medium(span, p);
  2453  }
  2454  
  2455  //! Deallocate the given large memory block to the current heap
  2456  static void
  2457  _rpmalloc_deallocate_large(span_t* span) {
  2458  	rpmalloc_assert(span->size_class == SIZE_CLASS_LARGE, "Bad span size class");
  2459  	rpmalloc_assert(!(span->flags & SPAN_FLAG_MASTER) || !(span->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  2460  	rpmalloc_assert((span->flags & SPAN_FLAG_MASTER) || (span->flags & SPAN_FLAG_SUBSPAN), "Span flag corrupted");
  2461  	//We must always defer (unless finalizing) if from another heap since we cannot touch the list or counters of another heap
  2462  #if RPMALLOC_FIRST_CLASS_HEAPS
  2463  	int defer = (span->heap->owner_thread && (span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2464  #else
  2465  	int defer = ((span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2466  #endif
  2467  	if (defer) {
  2468  		_rpmalloc_deallocate_defer_free_span(span->heap, span);
  2469  		return;
  2470  	}
  2471  	rpmalloc_assert(span->heap->full_span_count, "Heap span counter corrupted");
  2472  	--span->heap->full_span_count;
  2473  #if RPMALLOC_FIRST_CLASS_HEAPS
  2474  	_rpmalloc_span_double_link_list_remove(&span->heap->large_huge_span, span);
  2475  #endif
  2476  #if ENABLE_ADAPTIVE_THREAD_CACHE || ENABLE_STATISTICS
  2477  	//Decrease counter
  2478  	size_t idx = span->span_count - 1;
  2479  	atomic_decr32(&span->heap->span_use[idx].current);
  2480  #endif
  2481  	heap_t* heap = span->heap;
  2482  	rpmalloc_assert(heap, "No thread heap");
  2483  #if ENABLE_THREAD_CACHE
  2484  	const int set_as_reserved = ((span->span_count > 1) && (heap->span_cache.count == 0) && !heap->finalize && !heap->spans_reserved);
  2485  #else
  2486  	const int set_as_reserved = ((span->span_count > 1) && !heap->finalize && !heap->spans_reserved);
  2487  #endif
  2488  	if (set_as_reserved) {
  2489  		heap->span_reserve = span;
  2490  		heap->spans_reserved = span->span_count;
  2491  		if (span->flags & SPAN_FLAG_MASTER) {
  2492  			heap->span_reserve_master = span;
  2493  		} else { //SPAN_FLAG_SUBSPAN
  2494  			span_t* master = (span_t*)pointer_offset(span, -(intptr_t)((size_t)span->offset_from_master * _memory_span_size));
  2495  			heap->span_reserve_master = master;
  2496  			rpmalloc_assert(master->flags & SPAN_FLAG_MASTER, "Span flag corrupted");
  2497  			rpmalloc_assert(atomic_load32(&master->remaining_spans) >= (int32_t)span->span_count, "Master span count corrupted");
  2498  		}
  2499  		_rpmalloc_stat_inc(&heap->span_use[idx].spans_to_reserved);
  2500  	} else {
  2501  		//Insert into cache list
  2502  		_rpmalloc_heap_cache_insert(heap, span);
  2503  	}
  2504  }
  2505  
  2506  //! Deallocate the given huge span
  2507  static void
  2508  _rpmalloc_deallocate_huge(span_t* span) {
  2509  	rpmalloc_assert(span->heap, "No span heap");
  2510  #if RPMALLOC_FIRST_CLASS_HEAPS
  2511  	int defer = (span->heap->owner_thread && (span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2512  #else
  2513  	int defer = ((span->heap->owner_thread != get_thread_id()) && !span->heap->finalize);
  2514  #endif
  2515  	if (defer) {
  2516  		_rpmalloc_deallocate_defer_free_span(span->heap, span);
  2517  		return;
  2518  	}
  2519  	rpmalloc_assert(span->heap->full_span_count, "Heap span counter corrupted");
  2520  	--span->heap->full_span_count;
  2521  #if RPMALLOC_FIRST_CLASS_HEAPS
  2522  	_rpmalloc_span_double_link_list_remove(&span->heap->large_huge_span, span);
  2523  #endif
  2524  
  2525  	//Oversized allocation, page count is stored in span_count
  2526  	size_t num_pages = span->span_count;
  2527  	_rpmalloc_unmap(span, num_pages * _memory_page_size, span->align_offset, num_pages * _memory_page_size);
  2528  	_rpmalloc_stat_sub(&_huge_pages_current, num_pages);
  2529  }
  2530  
  2531  //! Deallocate the given block
  2532  static void
  2533  _rpmalloc_deallocate(void* p) {
  2534  	_rpmalloc_stat_add64(&_deallocation_counter, 1);
  2535  	//Grab the span (always at start of span, using span alignment)
  2536  	span_t* span = (span_t*)((uintptr_t)p & _memory_span_mask);
  2537  	if (UNEXPECTED(!span))
  2538  		return;
  2539  	if (EXPECTED(span->size_class < SIZE_CLASS_COUNT))
  2540  		_rpmalloc_deallocate_small_or_medium(span, p);
  2541  	else if (span->size_class == SIZE_CLASS_LARGE)
  2542  		_rpmalloc_deallocate_large(span);
  2543  	else
  2544  		_rpmalloc_deallocate_huge(span);
  2545  }
  2546  
  2547  ////////////
  2548  ///
  2549  /// Reallocation entry points
  2550  ///
  2551  //////
  2552  
  2553  static size_t
  2554  _rpmalloc_usable_size(void* p);
  2555  
  2556  //! Reallocate the given block to the given size
  2557  static void*
  2558  _rpmalloc_reallocate(heap_t* heap, void* p, size_t size, size_t oldsize, unsigned int flags) {
  2559  	if (p) {
  2560  		//Grab the span using guaranteed span alignment
  2561  		span_t* span = (span_t*)((uintptr_t)p & _memory_span_mask);
  2562  		if (EXPECTED(span->size_class < SIZE_CLASS_COUNT)) {
  2563  			//Small/medium sized block
  2564  			rpmalloc_assert(span->span_count == 1, "Span counter corrupted");
  2565  			void* blocks_start = pointer_offset(span, SPAN_HEADER_SIZE);
  2566  			uint32_t block_offset = (uint32_t)pointer_diff(p, blocks_start);
  2567  			uint32_t block_idx = block_offset / span->block_size;
  2568  			void* block = pointer_offset(blocks_start, (size_t)block_idx * span->block_size);
  2569  			if (!oldsize)
  2570  				oldsize = (size_t)((ptrdiff_t)span->block_size - pointer_diff(p, block));
  2571  			if ((size_t)span->block_size >= size) {
  2572  				//Still fits in block, never mind trying to save memory, but preserve data if alignment changed
  2573  				if ((p != block) && !(flags & RPMALLOC_NO_PRESERVE))
  2574  					memmove(block, p, oldsize);
  2575  				return block;
  2576  			}
  2577  		} else if (span->size_class == SIZE_CLASS_LARGE) {
  2578  			//Large block
  2579  			size_t total_size = size + SPAN_HEADER_SIZE;
  2580  			size_t num_spans = total_size >> _memory_span_size_shift;
  2581  			if (total_size & (_memory_span_mask - 1))
  2582  				++num_spans;
  2583  			size_t current_spans = span->span_count;
  2584  			void* block = pointer_offset(span, SPAN_HEADER_SIZE);
  2585  			if (!oldsize)
  2586  				oldsize = (current_spans * _memory_span_size) - (size_t)pointer_diff(p, block) - SPAN_HEADER_SIZE;
  2587  			if ((current_spans >= num_spans) && (total_size >= (oldsize / 2))) {
  2588  				//Still fits in block, never mind trying to save memory, but preserve data if alignment changed
  2589  				if ((p != block) && !(flags & RPMALLOC_NO_PRESERVE))
  2590  					memmove(block, p, oldsize);
  2591  				return block;
  2592  			}
  2593  		} else {
  2594  			//Oversized block
  2595  			size_t total_size = size + SPAN_HEADER_SIZE;
  2596  			size_t num_pages = total_size >> _memory_page_size_shift;
  2597  			if (total_size & (_memory_page_size - 1))
  2598  				++num_pages;
  2599  			//Page count is stored in span_count
  2600  			size_t current_pages = span->span_count;
  2601  			void* block = pointer_offset(span, SPAN_HEADER_SIZE);
  2602  			if (!oldsize)
  2603  				oldsize = (current_pages * _memory_page_size) - (size_t)pointer_diff(p, block) - SPAN_HEADER_SIZE;
  2604  			if ((current_pages >= num_pages) && (num_pages >= (current_pages / 2))) {
  2605  				//Still fits in block, never mind trying to save memory, but preserve data if alignment changed
  2606  				if ((p != block) && !(flags & RPMALLOC_NO_PRESERVE))
  2607  					memmove(block, p, oldsize);
  2608  				return block;
  2609  			}
  2610  		}
  2611  	} else {
  2612  		oldsize = 0;
  2613  	}
  2614  
  2615  	if (!!(flags & RPMALLOC_GROW_OR_FAIL))
  2616  		return 0;
  2617  
  2618  	//Size is greater than block size, need to allocate a new block and deallocate the old
  2619  	//Avoid hysteresis by overallocating if increase is small (below 37%)
  2620  	size_t lower_bound = oldsize + (oldsize >> 2) + (oldsize >> 3);
  2621  	size_t new_size = (size > lower_bound) ? size : ((size > oldsize) ? lower_bound : size);
  2622  	void* block = _rpmalloc_allocate(heap, new_size);
  2623  	if (p && block) {
  2624  		if (!(flags & RPMALLOC_NO_PRESERVE))
  2625  			memcpy(block, p, oldsize < new_size ? oldsize : new_size);
  2626  		_rpmalloc_deallocate(p);
  2627  	}
  2628  
  2629  	return block;
  2630  }
  2631  
  2632  static void*
  2633  _rpmalloc_aligned_reallocate(heap_t* heap, void* ptr, size_t alignment, size_t size, size_t oldsize,
  2634                             unsigned int flags) {
  2635  	if (alignment <= SMALL_GRANULARITY)
  2636  		return _rpmalloc_reallocate(heap, ptr, size, oldsize, flags);
  2637  
  2638  	int no_alloc = !!(flags & RPMALLOC_GROW_OR_FAIL);
  2639  	size_t usablesize = (ptr ? _rpmalloc_usable_size(ptr) : 0);
  2640  	if ((usablesize >= size) && !((uintptr_t)ptr & (alignment - 1))) {
  2641  		if (no_alloc || (size >= (usablesize / 2)))
  2642  			return ptr;
  2643  	}
  2644  	// Aligned alloc marks span as having aligned blocks
  2645  	void* block = (!no_alloc ? _rpmalloc_aligned_allocate(heap, alignment, size) : 0);
  2646  	if (EXPECTED(block != 0)) {
  2647  		if (!(flags & RPMALLOC_NO_PRESERVE) && ptr) {
  2648  			if (!oldsize)
  2649  				oldsize = usablesize;
  2650  			memcpy(block, ptr, oldsize < size ? oldsize : size);
  2651  		}
  2652  		_rpmalloc_deallocate(ptr);
  2653  	}
  2654  	return block;
  2655  }
  2656  
  2657  
  2658  ////////////
  2659  ///
  2660  /// Initialization, finalization and utility
  2661  ///
  2662  //////
  2663  
  2664  //! Get the usable size of the given block
  2665  static size_t
  2666  _rpmalloc_usable_size(void* p) {
  2667  	//Grab the span using guaranteed span alignment
  2668  	span_t* span = (span_t*)((uintptr_t)p & _memory_span_mask);
  2669  	if (span->size_class < SIZE_CLASS_COUNT) {
  2670  		//Small/medium block
  2671  		void* blocks_start = pointer_offset(span, SPAN_HEADER_SIZE);
  2672  		return span->block_size - ((size_t)pointer_diff(p, blocks_start) % span->block_size);
  2673  	}
  2674  	if (span->size_class == SIZE_CLASS_LARGE) {
  2675  		//Large block
  2676  		size_t current_spans = span->span_count;
  2677  		return (current_spans * _memory_span_size) - (size_t)pointer_diff(p, span);
  2678  	}
  2679  	//Oversized block, page count is stored in span_count
  2680  	size_t current_pages = span->span_count;
  2681  	return (current_pages * _memory_page_size) - (size_t)pointer_diff(p, span);
  2682  }
  2683  
  2684  //! Adjust and optimize the size class properties for the given class
  2685  static void
  2686  _rpmalloc_adjust_size_class(size_t iclass) {
  2687  	size_t block_size = _memory_size_class[iclass].block_size;
  2688  	size_t block_count = (_memory_span_size - SPAN_HEADER_SIZE) / block_size;
  2689  
  2690  	_memory_size_class[iclass].block_count = (uint16_t)block_count;
  2691  	_memory_size_class[iclass].class_idx = (uint16_t)iclass;
  2692  
  2693  	//Check if previous size classes can be merged
  2694  	if (iclass >= SMALL_CLASS_COUNT) {
  2695  		size_t prevclass = iclass;
  2696  		while (prevclass > 0) {
  2697  			--prevclass;
  2698  			//A class can be merged if number of pages and number of blocks are equal
  2699  			if (_memory_size_class[prevclass].block_count == _memory_size_class[iclass].block_count)
  2700  				memcpy(_memory_size_class + prevclass, _memory_size_class + iclass, sizeof(_memory_size_class[iclass]));
  2701  			else
  2702  				break;
  2703  		}
  2704  	}
  2705  }
  2706  
  2707  //! Initialize the allocator and setup global data
  2708  extern inline int
  2709  rpmalloc_initialize(void) {
  2710  	if (_rpmalloc_initialized) {
  2711  		rpmalloc_thread_initialize();
  2712  		return 0;
  2713  	}
  2714  	return rpmalloc_initialize_config(0);
  2715  }
  2716  
  2717  int
  2718  rpmalloc_initialize_config(const rpmalloc_config_t* config) {
  2719  	if (_rpmalloc_initialized) {
  2720  		rpmalloc_thread_initialize();
  2721  		return 0;
  2722  	}
  2723  	_rpmalloc_initialized = 1;
  2724  
  2725  	if (config)
  2726  		memcpy(&_memory_config, config, sizeof(rpmalloc_config_t));
  2727  	else
  2728  		memset(&_memory_config, 0, sizeof(rpmalloc_config_t));
  2729  
  2730  	if (!_memory_config.memory_map || !_memory_config.memory_unmap) {
  2731  		_memory_config.memory_map = _rpmalloc_mmap_os;
  2732  		_memory_config.memory_unmap = _rpmalloc_unmap_os;
  2733  	}
  2734  
  2735  #if PLATFORM_WINDOWS
  2736  	SYSTEM_INFO system_info;
  2737  	memset(&system_info, 0, sizeof(system_info));
  2738  	GetSystemInfo(&system_info);
  2739  	_memory_map_granularity = system_info.dwAllocationGranularity;
  2740  #else
  2741  	_memory_map_granularity = (size_t)sysconf(_SC_PAGESIZE);
  2742  #endif
  2743  
  2744  #if RPMALLOC_CONFIGURABLE
  2745  	_memory_page_size = _memory_config.page_size;
  2746  #else
  2747  	_memory_page_size = 0;
  2748  #endif
  2749  	_memory_huge_pages = 0;
  2750  	if (!_memory_page_size) {
  2751  #if PLATFORM_WINDOWS
  2752  		_memory_page_size = system_info.dwPageSize;
  2753  #else
  2754  		_memory_page_size = _memory_map_granularity;
  2755  		if (_memory_config.enable_huge_pages) {
  2756  #if defined(__linux__)
  2757  			size_t huge_page_size = 0;
  2758  			FILE* meminfo = fopen("/proc/meminfo", "r");
  2759  			if (meminfo) {
  2760  				char line[128];
  2761  				while (!huge_page_size && fgets(line, sizeof(line) - 1, meminfo)) {
  2762  					line[sizeof(line) - 1] = 0;
  2763  					if (strstr(line, "Hugepagesize:"))
  2764  						huge_page_size = (size_t)strtol(line + 13, 0, 10) * 1024;
  2765  				}
  2766  				fclose(meminfo);
  2767  			}
  2768  			if (huge_page_size) {
  2769  				_memory_huge_pages = 1;
  2770  				_memory_page_size = huge_page_size;
  2771  				_memory_map_granularity = huge_page_size;
  2772  			}
  2773  #elif defined(__FreeBSD__)
  2774  			int rc;
  2775  			size_t sz = sizeof(rc);
  2776  
  2777  			if (sysctlbyname("vm.pmap.pg_ps_enabled", &rc, &sz, NULL, 0) == 0 && rc == 1) {
  2778  				_memory_huge_pages = 1;
  2779  				_memory_page_size = 2 * 1024 * 1024;
  2780  				_memory_map_granularity = _memory_page_size;
  2781  			}
  2782  #elif defined(__APPLE__) || defined(__NetBSD__)
  2783  			_memory_huge_pages = 1;
  2784  			_memory_page_size = 2 * 1024 * 1024;
  2785  			_memory_map_granularity = _memory_page_size;
  2786  #endif
  2787  		}
  2788  #endif
  2789  	} else {
  2790  		if (_memory_config.enable_huge_pages)
  2791  			_memory_huge_pages = 1;
  2792  	}
  2793  
  2794  #if PLATFORM_WINDOWS
  2795  	if (_memory_config.enable_huge_pages) {
  2796  		HANDLE token = 0;
  2797  		size_t large_page_minimum = GetLargePageMinimum();
  2798  		if (large_page_minimum)
  2799  			OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &token);
  2800  		if (token) {
  2801  			LUID luid;
  2802  			if (LookupPrivilegeValue(0, SE_LOCK_MEMORY_NAME, &luid)) {
  2803  				TOKEN_PRIVILEGES token_privileges;
  2804  				memset(&token_privileges, 0, sizeof(token_privileges));
  2805  				token_privileges.PrivilegeCount = 1;
  2806  				token_privileges.Privileges[0].Luid = luid;
  2807  				token_privileges.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
  2808  				if (AdjustTokenPrivileges(token, FALSE, &token_privileges, 0, 0, 0)) {
  2809  					if (GetLastError() == ERROR_SUCCESS)
  2810  						_memory_huge_pages = 1;
  2811  				}
  2812  			}
  2813  			CloseHandle(token);
  2814  		}
  2815  		if (_memory_huge_pages) {
  2816  			if (large_page_minimum > _memory_page_size)
  2817  				_memory_page_size = large_page_minimum;
  2818  			if (large_page_minimum > _memory_map_granularity)
  2819  				_memory_map_granularity = large_page_minimum;
  2820  		}
  2821  	}
  2822  #endif
  2823  
  2824  	size_t min_span_size = 256;
  2825  	size_t max_page_size;
  2826  #if UINTPTR_MAX > 0xFFFFFFFF
  2827  	max_page_size = 4096ULL * 1024ULL * 1024ULL;
  2828  #else
  2829  	max_page_size = 4 * 1024 * 1024;
  2830  #endif
  2831  	if (_memory_page_size < min_span_size)
  2832  		_memory_page_size = min_span_size;
  2833  	if (_memory_page_size > max_page_size)
  2834  		_memory_page_size = max_page_size;
  2835  	_memory_page_size_shift = 0;
  2836  	size_t page_size_bit = _memory_page_size;
  2837  	while (page_size_bit != 1) {
  2838  		++_memory_page_size_shift;
  2839  		page_size_bit >>= 1;
  2840  	}
  2841  	_memory_page_size = ((size_t)1 << _memory_page_size_shift);
  2842  
  2843  #if RPMALLOC_CONFIGURABLE
  2844  	if (!_memory_config.span_size) {
  2845  		_memory_span_size = _memory_default_span_size;
  2846  		_memory_span_size_shift = _memory_default_span_size_shift;
  2847  		_memory_span_mask = _memory_default_span_mask;
  2848  	} else {
  2849  		size_t span_size = _memory_config.span_size;
  2850  		if (span_size > (256 * 1024))
  2851  			span_size = (256 * 1024);
  2852  		_memory_span_size = 4096;
  2853  		_memory_span_size_shift = 12;
  2854  		while (_memory_span_size < span_size) {
  2855  			_memory_span_size <<= 1;
  2856  			++_memory_span_size_shift;
  2857  		}
  2858  		_memory_span_mask = ~(uintptr_t)(_memory_span_size - 1);
  2859  	}
  2860  #endif
  2861  
  2862  	_memory_span_map_count = ( _memory_config.span_map_count ? _memory_config.span_map_count : DEFAULT_SPAN_MAP_COUNT);
  2863  	if ((_memory_span_size * _memory_span_map_count) < _memory_page_size)
  2864  		_memory_span_map_count = (_memory_page_size / _memory_span_size);
  2865  	if ((_memory_page_size >= _memory_span_size) && ((_memory_span_map_count * _memory_span_size) % _memory_page_size))
  2866  		_memory_span_map_count = (_memory_page_size / _memory_span_size);
  2867  	_memory_heap_reserve_count = (_memory_span_map_count > DEFAULT_SPAN_MAP_COUNT) ? DEFAULT_SPAN_MAP_COUNT : _memory_span_map_count;
  2868  
  2869  	_memory_config.page_size = _memory_page_size;
  2870  	_memory_config.span_size = _memory_span_size;
  2871  	_memory_config.span_map_count = _memory_span_map_count;
  2872  	_memory_config.enable_huge_pages = _memory_huge_pages;
  2873  
  2874  #if ((defined(__APPLE__) || defined(__HAIKU__)) && ENABLE_PRELOAD) || defined(__TINYC__)
  2875  	if (pthread_key_create(&_memory_thread_heap, _rpmalloc_heap_release_raw_fc))
  2876  		return -1;
  2877  #endif
  2878  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
  2879  	fls_key = FlsAlloc(&_rpmalloc_thread_destructor);
  2880  #endif
  2881  
  2882  	//Setup all small and medium size classes
  2883  	size_t iclass = 0;
  2884  	_memory_size_class[iclass].block_size = SMALL_GRANULARITY;
  2885  	_rpmalloc_adjust_size_class(iclass);
  2886  	for (iclass = 1; iclass < SMALL_CLASS_COUNT; ++iclass) {
  2887  		size_t size = iclass * SMALL_GRANULARITY;
  2888  		_memory_size_class[iclass].block_size = (uint32_t)size;
  2889  		_rpmalloc_adjust_size_class(iclass);
  2890  	}
  2891  	//At least two blocks per span, then fall back to large allocations
  2892  	_memory_medium_size_limit = (_memory_span_size - SPAN_HEADER_SIZE) >> 1;
  2893  	if (_memory_medium_size_limit > MEDIUM_SIZE_LIMIT)
  2894  		_memory_medium_size_limit = MEDIUM_SIZE_LIMIT;
  2895  	for (iclass = 0; iclass < MEDIUM_CLASS_COUNT; ++iclass) {
  2896  		size_t size = SMALL_SIZE_LIMIT + ((iclass + 1) * MEDIUM_GRANULARITY);
  2897  		if (size > _memory_medium_size_limit)
  2898  			break;
  2899  		_memory_size_class[SMALL_CLASS_COUNT + iclass].block_size = (uint32_t)size;
  2900  		_rpmalloc_adjust_size_class(SMALL_CLASS_COUNT + iclass);
  2901  	}
  2902  
  2903  	_memory_orphan_heaps = 0;
  2904  #if RPMALLOC_FIRST_CLASS_HEAPS
  2905  	_memory_first_class_orphan_heaps = 0;
  2906  #endif
  2907  #if ENABLE_STATISTICS
  2908  	atomic_store32(&_memory_active_heaps, 0);
  2909  	atomic_store32(&_mapped_pages, 0);
  2910  	_mapped_pages_peak = 0;
  2911  	atomic_store32(&_master_spans, 0);
  2912  	atomic_store32(&_mapped_total, 0);
  2913  	atomic_store32(&_unmapped_total, 0);
  2914  	atomic_store32(&_mapped_pages_os, 0);
  2915  	atomic_store32(&_huge_pages_current, 0);
  2916  	_huge_pages_peak = 0;
  2917  #endif
  2918  	memset(_memory_heaps, 0, sizeof(_memory_heaps));
  2919  	atomic_store32_release(&_memory_global_lock, 0);
  2920  
  2921  	rpmalloc_linker_reference();
  2922  
  2923  	//Initialize this thread
  2924  	rpmalloc_thread_initialize();
  2925  	return 0;
  2926  }
  2927  
  2928  //! Finalize the allocator
  2929  void
  2930  rpmalloc_finalize(void) {
  2931  	rpmalloc_thread_finalize(1);
  2932  	//rpmalloc_dump_statistics(stdout);
  2933  
  2934  	if (_memory_global_reserve) {
  2935  		atomic_add32(&_memory_global_reserve_master->remaining_spans, -(int32_t)_memory_global_reserve_count);
  2936  		_memory_global_reserve_master = 0;
  2937  		_memory_global_reserve_count = 0;
  2938  		_memory_global_reserve = 0;
  2939  	}
  2940  	atomic_store32_release(&_memory_global_lock, 0);
  2941  
  2942  	//Free all thread caches and fully free spans
  2943  	for (size_t list_idx = 0; list_idx < HEAP_ARRAY_SIZE; ++list_idx) {
  2944  		heap_t* heap = _memory_heaps[list_idx];
  2945  		while (heap) {
  2946  			heap_t* next_heap = heap->next_heap;
  2947  			heap->finalize = 1;
  2948  			_rpmalloc_heap_global_finalize(heap);
  2949  			heap = next_heap;
  2950  		}
  2951  	}
  2952  
  2953  #if ENABLE_GLOBAL_CACHE
  2954  	//Free global caches
  2955  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass)
  2956  		_rpmalloc_global_cache_finalize(&_memory_span_cache[iclass]);
  2957  #endif
  2958  
  2959  #if (defined(__APPLE__) || defined(__HAIKU__)) && ENABLE_PRELOAD
  2960  	pthread_key_delete(_memory_thread_heap);
  2961  #endif
  2962  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
  2963  	FlsFree(fls_key);
  2964  	fls_key = 0;
  2965  #endif
  2966  #if ENABLE_STATISTICS
  2967  	//If you hit these asserts you probably have memory leaks (perhaps global scope data doing dynamic allocations) or double frees in your code
  2968  	rpmalloc_assert(atomic_load32(&_mapped_pages) == 0, "Memory leak detected");
  2969  	rpmalloc_assert(atomic_load32(&_mapped_pages_os) == 0, "Memory leak detected");
  2970  #endif
  2971  
  2972  	_rpmalloc_initialized = 0;
  2973  }
  2974  
  2975  //! Initialize thread, assign heap
  2976  extern inline void
  2977  rpmalloc_thread_initialize(void) {
  2978  	if (!get_thread_heap_raw()) {
  2979  		heap_t* heap = _rpmalloc_heap_allocate(0);
  2980  		if (heap) {
  2981  			_rpmalloc_stat_inc(&_memory_active_heaps);
  2982  			set_thread_heap(heap);
  2983  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
  2984  			FlsSetValue(fls_key, heap);
  2985  #endif
  2986  		}
  2987  	}
  2988  }
  2989  
  2990  //! Finalize thread, orphan heap
  2991  void
  2992  rpmalloc_thread_finalize(int release_caches) {
  2993  	heap_t* heap = get_thread_heap_raw();
  2994  	if (heap)
  2995  		_rpmalloc_heap_release_raw(heap, release_caches);
  2996  	set_thread_heap(0);
  2997  #if defined(_WIN32) && (!defined(BUILD_DYNAMIC_LINK) || !BUILD_DYNAMIC_LINK)
  2998  	FlsSetValue(fls_key, 0);
  2999  #endif
  3000  }
  3001  
  3002  int
  3003  rpmalloc_is_thread_initialized(void) {
  3004  	return (get_thread_heap_raw() != 0) ? 1 : 0;
  3005  }
  3006  
  3007  const rpmalloc_config_t*
  3008  rpmalloc_config(void) {
  3009  	return &_memory_config;
  3010  }
  3011  
  3012  // Extern interface
  3013  
  3014  extern inline RPMALLOC_ALLOCATOR void*
  3015  rpmalloc(size_t size) {
  3016  #if ENABLE_VALIDATE_ARGS
  3017  	if (size >= MAX_ALLOC_SIZE) {
  3018  		errno = EINVAL;
  3019  		return 0;
  3020  	}
  3021  #endif
  3022  	heap_t* heap = get_thread_heap();
  3023  	return _rpmalloc_allocate(heap, size);
  3024  }
  3025  
  3026  extern inline void
  3027  rpfree(void* ptr) {
  3028  	_rpmalloc_deallocate(ptr);
  3029  }
  3030  
  3031  extern inline RPMALLOC_ALLOCATOR void*
  3032  rpcalloc(size_t num, size_t size) {
  3033  	size_t total;
  3034  #if ENABLE_VALIDATE_ARGS
  3035  #if PLATFORM_WINDOWS
  3036  	int err = SizeTMult(num, size, &total);
  3037  	if ((err != S_OK) || (total >= MAX_ALLOC_SIZE)) {
  3038  		errno = EINVAL;
  3039  		return 0;
  3040  	}
  3041  #else
  3042  	int err = __builtin_umull_overflow(num, size, &total);
  3043  	if (err || (total >= MAX_ALLOC_SIZE)) {
  3044  		errno = EINVAL;
  3045  		return 0;
  3046  	}
  3047  #endif
  3048  #else
  3049  	total = num * size;
  3050  #endif
  3051  	heap_t* heap = get_thread_heap();
  3052  	void* block = _rpmalloc_allocate(heap, total);
  3053  	if (block)
  3054  		memset(block, 0, total);
  3055  	return block;
  3056  }
  3057  
  3058  extern inline RPMALLOC_ALLOCATOR void*
  3059  rprealloc(void* ptr, size_t size) {
  3060  #if ENABLE_VALIDATE_ARGS
  3061  	if (size >= MAX_ALLOC_SIZE) {
  3062  		errno = EINVAL;
  3063  		return ptr;
  3064  	}
  3065  #endif
  3066  	heap_t* heap = get_thread_heap();
  3067  	return _rpmalloc_reallocate(heap, ptr, size, 0, 0);
  3068  }
  3069  
  3070  extern RPMALLOC_ALLOCATOR void*
  3071  rpaligned_realloc(void* ptr, size_t alignment, size_t size, size_t oldsize,
  3072                    unsigned int flags) {
  3073  #if ENABLE_VALIDATE_ARGS
  3074  	if ((size + alignment < size) || (alignment > _memory_page_size)) {
  3075  		errno = EINVAL;
  3076  		return 0;
  3077  	}
  3078  #endif
  3079  	heap_t* heap = get_thread_heap();
  3080  	return _rpmalloc_aligned_reallocate(heap, ptr, alignment, size, oldsize, flags);
  3081  }
  3082  
  3083  extern RPMALLOC_ALLOCATOR void*
  3084  rpaligned_alloc(size_t alignment, size_t size) {
  3085  	heap_t* heap = get_thread_heap();
  3086  	return _rpmalloc_aligned_allocate(heap, alignment, size);
  3087  }
  3088  
  3089  extern inline RPMALLOC_ALLOCATOR void*
  3090  rpaligned_calloc(size_t alignment, size_t num, size_t size) {
  3091  	size_t total;
  3092  #if ENABLE_VALIDATE_ARGS
  3093  #if PLATFORM_WINDOWS
  3094  	int err = SizeTMult(num, size, &total);
  3095  	if ((err != S_OK) || (total >= MAX_ALLOC_SIZE)) {
  3096  		errno = EINVAL;
  3097  		return 0;
  3098  	}
  3099  #else
  3100  	int err = __builtin_umull_overflow(num, size, &total);
  3101  	if (err || (total >= MAX_ALLOC_SIZE)) {
  3102  		errno = EINVAL;
  3103  		return 0;
  3104  	}
  3105  #endif
  3106  #else
  3107  	total = num * size;
  3108  #endif
  3109  	void* block = rpaligned_alloc(alignment, total);
  3110  	if (block)
  3111  		memset(block, 0, total);
  3112  	return block;
  3113  }
  3114  
  3115  extern inline RPMALLOC_ALLOCATOR void*
  3116  rpmemalign(size_t alignment, size_t size) {
  3117  	return rpaligned_alloc(alignment, size);
  3118  }
  3119  
  3120  extern inline int
  3121  rpposix_memalign(void **memptr, size_t alignment, size_t size) {
  3122  	if (memptr)
  3123  		*memptr = rpaligned_alloc(alignment, size);
  3124  	else
  3125  		return EINVAL;
  3126  	return *memptr ? 0 : ENOMEM;
  3127  }
  3128  
  3129  extern inline size_t
  3130  rpmalloc_usable_size(void* ptr) {
  3131  	return (ptr ? _rpmalloc_usable_size(ptr) : 0);
  3132  }
  3133  
  3134  extern inline void
  3135  rpmalloc_thread_collect(void) {
  3136  }
  3137  
  3138  void
  3139  rpmalloc_thread_statistics(rpmalloc_thread_statistics_t* stats) {
  3140  	memset(stats, 0, sizeof(rpmalloc_thread_statistics_t));
  3141  	heap_t* heap = get_thread_heap_raw();
  3142  	if (!heap)
  3143  		return;
  3144  
  3145  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  3146  		size_class_t* size_class = _memory_size_class + iclass;
  3147  		span_t* span = heap->size_class[iclass].partial_span;
  3148  		while (span) {
  3149  			size_t free_count = span->list_size;
  3150  			size_t block_count = size_class->block_count;
  3151  			if (span->free_list_limit < block_count)
  3152  				block_count = span->free_list_limit;
  3153  			free_count += (block_count - span->used_count);
  3154  			stats->sizecache = free_count * size_class->block_size;
  3155  			span = span->next;
  3156  		}
  3157  	}
  3158  
  3159  #if ENABLE_THREAD_CACHE
  3160  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3161  		span_cache_t* span_cache;
  3162  		if (!iclass)
  3163  			span_cache = &heap->span_cache;
  3164  		else
  3165  			span_cache = (span_cache_t*)(heap->span_large_cache + (iclass - 1));
  3166  		stats->spancache = span_cache->count * (iclass + 1) * _memory_span_size;
  3167  	}
  3168  #endif
  3169  
  3170  	span_t* deferred = (span_t*)atomic_load_ptr(&heap->span_free_deferred);
  3171  	while (deferred) {
  3172  		if (deferred->size_class != SIZE_CLASS_HUGE)
  3173  			stats->spancache = (size_t)deferred->span_count * _memory_span_size;
  3174  		deferred = (span_t*)deferred->free_list;
  3175  	}
  3176  
  3177  #if ENABLE_STATISTICS
  3178  	stats->thread_to_global = (size_t)atomic_load64(&heap->thread_to_global);
  3179  	stats->global_to_thread = (size_t)atomic_load64(&heap->global_to_thread);
  3180  
  3181  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3182  		stats->span_use[iclass].current = (size_t)atomic_load32(&heap->span_use[iclass].current);
  3183  		stats->span_use[iclass].peak = (size_t)atomic_load32(&heap->span_use[iclass].high);
  3184  		stats->span_use[iclass].to_global = (size_t)atomic_load32(&heap->span_use[iclass].spans_to_global);
  3185  		stats->span_use[iclass].from_global = (size_t)atomic_load32(&heap->span_use[iclass].spans_from_global);
  3186  		stats->span_use[iclass].to_cache = (size_t)atomic_load32(&heap->span_use[iclass].spans_to_cache);
  3187  		stats->span_use[iclass].from_cache = (size_t)atomic_load32(&heap->span_use[iclass].spans_from_cache);
  3188  		stats->span_use[iclass].to_reserved = (size_t)atomic_load32(&heap->span_use[iclass].spans_to_reserved);
  3189  		stats->span_use[iclass].from_reserved = (size_t)atomic_load32(&heap->span_use[iclass].spans_from_reserved);
  3190  		stats->span_use[iclass].map_calls = (size_t)atomic_load32(&heap->span_use[iclass].spans_map_calls);
  3191  	}
  3192  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  3193  		stats->size_use[iclass].alloc_current = (size_t)atomic_load32(&heap->size_class_use[iclass].alloc_current);
  3194  		stats->size_use[iclass].alloc_peak = (size_t)heap->size_class_use[iclass].alloc_peak;
  3195  		stats->size_use[iclass].alloc_total = (size_t)atomic_load32(&heap->size_class_use[iclass].alloc_total);
  3196  		stats->size_use[iclass].free_total = (size_t)atomic_load32(&heap->size_class_use[iclass].free_total);
  3197  		stats->size_use[iclass].spans_to_cache = (size_t)atomic_load32(&heap->size_class_use[iclass].spans_to_cache);
  3198  		stats->size_use[iclass].spans_from_cache = (size_t)atomic_load32(&heap->size_class_use[iclass].spans_from_cache);
  3199  		stats->size_use[iclass].spans_from_reserved = (size_t)atomic_load32(&heap->size_class_use[iclass].spans_from_reserved);
  3200  		stats->size_use[iclass].map_calls = (size_t)atomic_load32(&heap->size_class_use[iclass].spans_map_calls);
  3201  	}
  3202  #endif
  3203  }
  3204  
  3205  void
  3206  rpmalloc_global_statistics(rpmalloc_global_statistics_t* stats) {
  3207  	memset(stats, 0, sizeof(rpmalloc_global_statistics_t));
  3208  #if ENABLE_STATISTICS
  3209  	stats->mapped = (size_t)atomic_load32(&_mapped_pages) * _memory_page_size;
  3210  	stats->mapped_peak = (size_t)_mapped_pages_peak * _memory_page_size;
  3211  	stats->mapped_total = (size_t)atomic_load32(&_mapped_total) * _memory_page_size;
  3212  	stats->unmapped_total = (size_t)atomic_load32(&_unmapped_total) * _memory_page_size;
  3213  	stats->huge_alloc = (size_t)atomic_load32(&_huge_pages_current) * _memory_page_size;
  3214  	stats->huge_alloc_peak = (size_t)_huge_pages_peak * _memory_page_size;
  3215  #endif
  3216  #if ENABLE_GLOBAL_CACHE
  3217  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass)
  3218  		stats->cached += _memory_span_cache[iclass].count * (iclass + 1) * _memory_span_size;
  3219  #endif
  3220  }
  3221  
  3222  #if ENABLE_STATISTICS
  3223  
  3224  static void
  3225  _memory_heap_dump_statistics(heap_t* heap, void* file) {
  3226  	fprintf(file, "Heap %d stats:\n", heap->id);
  3227  	fprintf(file, "Class   CurAlloc  PeakAlloc   TotAlloc    TotFree  BlkSize BlkCount SpansCur SpansPeak  PeakAllocMiB  ToCacheMiB FromCacheMiB FromReserveMiB MmapCalls\n");
  3228  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  3229  		if (!atomic_load32(&heap->size_class_use[iclass].alloc_total))
  3230  			continue;
  3231  		fprintf(file, "%3u:  %10u %10u %10u %10u %8u %8u %8d %9d %13zu %11zu %12zu %14zu %9u\n", (uint32_t)iclass,
  3232  			atomic_load32(&heap->size_class_use[iclass].alloc_current),
  3233  			heap->size_class_use[iclass].alloc_peak,
  3234  			atomic_load32(&heap->size_class_use[iclass].alloc_total),
  3235  			atomic_load32(&heap->size_class_use[iclass].free_total),
  3236  			_memory_size_class[iclass].block_size,
  3237  			_memory_size_class[iclass].block_count,
  3238  			atomic_load32(&heap->size_class_use[iclass].spans_current),
  3239  			heap->size_class_use[iclass].spans_peak,
  3240  			((size_t)heap->size_class_use[iclass].alloc_peak * (size_t)_memory_size_class[iclass].block_size) / (size_t)(1024 * 1024),
  3241  			((size_t)atomic_load32(&heap->size_class_use[iclass].spans_to_cache) * _memory_span_size) / (size_t)(1024 * 1024),
  3242  			((size_t)atomic_load32(&heap->size_class_use[iclass].spans_from_cache) * _memory_span_size) / (size_t)(1024 * 1024),
  3243  			((size_t)atomic_load32(&heap->size_class_use[iclass].spans_from_reserved) * _memory_span_size) / (size_t)(1024 * 1024),
  3244  			atomic_load32(&heap->size_class_use[iclass].spans_map_calls));
  3245  	}
  3246  	fprintf(file, "Spans  Current     Peak Deferred  PeakMiB  Cached  ToCacheMiB FromCacheMiB ToReserveMiB FromReserveMiB ToGlobalMiB FromGlobalMiB  MmapCalls\n");
  3247  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3248  		if (!atomic_load32(&heap->span_use[iclass].high) && !atomic_load32(&heap->span_use[iclass].spans_map_calls))
  3249  			continue;
  3250  		fprintf(file, "%4u: %8d %8u %8u %8zu %7u %11zu %12zu %12zu %14zu %11zu %13zu %10u\n", (uint32_t)(iclass + 1),
  3251  			atomic_load32(&heap->span_use[iclass].current),
  3252  			atomic_load32(&heap->span_use[iclass].high),
  3253  			atomic_load32(&heap->span_use[iclass].spans_deferred),
  3254  			((size_t)atomic_load32(&heap->span_use[iclass].high) * (size_t)_memory_span_size * (iclass + 1)) / (size_t)(1024 * 1024),
  3255  #if ENABLE_THREAD_CACHE
  3256  			(unsigned int)(!iclass ? heap->span_cache.count : heap->span_large_cache[iclass - 1].count),
  3257  			((size_t)atomic_load32(&heap->span_use[iclass].spans_to_cache) * (iclass + 1) * _memory_span_size) / (size_t)(1024 * 1024),
  3258  			((size_t)atomic_load32(&heap->span_use[iclass].spans_from_cache) * (iclass + 1) * _memory_span_size) / (size_t)(1024 * 1024),
  3259  #else
  3260  			0, (size_t)0, (size_t)0,
  3261  #endif
  3262  			((size_t)atomic_load32(&heap->span_use[iclass].spans_to_reserved) * (iclass + 1) * _memory_span_size) / (size_t)(1024 * 1024),
  3263  			((size_t)atomic_load32(&heap->span_use[iclass].spans_from_reserved) * (iclass + 1) * _memory_span_size) / (size_t)(1024 * 1024),
  3264  			((size_t)atomic_load32(&heap->span_use[iclass].spans_to_global) * (size_t)_memory_span_size * (iclass + 1)) / (size_t)(1024 * 1024),
  3265  			((size_t)atomic_load32(&heap->span_use[iclass].spans_from_global) * (size_t)_memory_span_size * (iclass + 1)) / (size_t)(1024 * 1024),
  3266  			atomic_load32(&heap->span_use[iclass].spans_map_calls));
  3267  	}
  3268  	fprintf(file, "Full spans: %zu\n", heap->full_span_count);
  3269  	fprintf(file, "ThreadToGlobalMiB GlobalToThreadMiB\n");
  3270  	fprintf(file, "%17zu %17zu\n", (size_t)atomic_load64(&heap->thread_to_global) / (size_t)(1024 * 1024), (size_t)atomic_load64(&heap->global_to_thread) / (size_t)(1024 * 1024));
  3271  }
  3272  
  3273  #endif
  3274  
  3275  void
  3276  rpmalloc_dump_statistics(void* file) {
  3277  #if ENABLE_STATISTICS
  3278  	for (size_t list_idx = 0; list_idx < HEAP_ARRAY_SIZE; ++list_idx) {
  3279  		heap_t* heap = _memory_heaps[list_idx];
  3280  		while (heap) {
  3281  			int need_dump = 0;
  3282  			for (size_t iclass = 0; !need_dump && (iclass < SIZE_CLASS_COUNT); ++iclass) {
  3283  				if (!atomic_load32(&heap->size_class_use[iclass].alloc_total)) {
  3284  					rpmalloc_assert(!atomic_load32(&heap->size_class_use[iclass].free_total), "Heap statistics counter mismatch");
  3285  					rpmalloc_assert(!atomic_load32(&heap->size_class_use[iclass].spans_map_calls), "Heap statistics counter mismatch");
  3286  					continue;
  3287  				}
  3288  				need_dump = 1;
  3289  			}
  3290  			for (size_t iclass = 0; !need_dump && (iclass < LARGE_CLASS_COUNT); ++iclass) {
  3291  				if (!atomic_load32(&heap->span_use[iclass].high) && !atomic_load32(&heap->span_use[iclass].spans_map_calls))
  3292  					continue;
  3293  				need_dump = 1;
  3294  			}
  3295  			if (need_dump)
  3296  				_memory_heap_dump_statistics(heap, file);
  3297  			heap = heap->next_heap;
  3298  		}
  3299  	}
  3300  	fprintf(file, "Global stats:\n");
  3301  	size_t huge_current = (size_t)atomic_load32(&_huge_pages_current) * _memory_page_size;
  3302  	size_t huge_peak = (size_t)_huge_pages_peak * _memory_page_size;
  3303  	fprintf(file, "HugeCurrentMiB HugePeakMiB\n");
  3304  	fprintf(file, "%14zu %11zu\n", huge_current / (size_t)(1024 * 1024), huge_peak / (size_t)(1024 * 1024));
  3305  
  3306  	fprintf(file, "GlobalCacheMiB\n");
  3307  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3308  		global_cache_t* cache = _memory_span_cache + iclass;
  3309  		size_t global_cache = (size_t)cache->count * iclass * _memory_span_size;
  3310  
  3311  		size_t global_overflow_cache = 0;
  3312  		span_t* span = cache->overflow;
  3313  		while (span) {
  3314  			global_overflow_cache += iclass * _memory_span_size;
  3315  			span = span->next;
  3316  		}
  3317  		if (global_cache || global_overflow_cache || cache->insert_count || cache->extract_count)
  3318  			fprintf(file, "%4zu: %8zuMiB (%8zuMiB overflow) %14zu insert %14zu extract\n", iclass + 1, global_cache / (size_t)(1024 * 1024), global_overflow_cache / (size_t)(1024 * 1024), cache->insert_count, cache->extract_count);
  3319  	}
  3320  
  3321  	size_t mapped = (size_t)atomic_load32(&_mapped_pages) * _memory_page_size;
  3322  	size_t mapped_os = (size_t)atomic_load32(&_mapped_pages_os) * _memory_page_size;
  3323  	size_t mapped_peak = (size_t)_mapped_pages_peak * _memory_page_size;
  3324  	size_t mapped_total = (size_t)atomic_load32(&_mapped_total) * _memory_page_size;
  3325  	size_t unmapped_total = (size_t)atomic_load32(&_unmapped_total) * _memory_page_size;
  3326  	fprintf(file, "MappedMiB MappedOSMiB MappedPeakMiB MappedTotalMiB UnmappedTotalMiB\n");
  3327  	fprintf(file, "%9zu %11zu %13zu %14zu %16zu\n",
  3328  		mapped / (size_t)(1024 * 1024),
  3329  		mapped_os / (size_t)(1024 * 1024),
  3330  		mapped_peak / (size_t)(1024 * 1024),
  3331  		mapped_total / (size_t)(1024 * 1024),
  3332  		unmapped_total / (size_t)(1024 * 1024));
  3333  
  3334  	fprintf(file, "\n");
  3335  #if 0
  3336  	int64_t allocated = atomic_load64(&_allocation_counter);
  3337  	int64_t deallocated = atomic_load64(&_deallocation_counter);
  3338  	fprintf(file, "Allocation count: %lli\n", allocated);
  3339  	fprintf(file, "Deallocation count: %lli\n", deallocated);
  3340  	fprintf(file, "Current allocations: %lli\n", (allocated - deallocated));
  3341  	fprintf(file, "Master spans: %d\n", atomic_load32(&_master_spans));
  3342  	fprintf(file, "Dangling master spans: %d\n", atomic_load32(&_unmapped_master_spans));
  3343  #endif
  3344  #endif
  3345  	(void)sizeof(file);
  3346  }
  3347  
  3348  #if RPMALLOC_FIRST_CLASS_HEAPS
  3349  
  3350  extern inline rpmalloc_heap_t*
  3351  rpmalloc_heap_acquire(void) {
  3352  	// Must be a pristine heap from newly mapped memory pages, or else memory blocks
  3353  	// could already be allocated from the heap which would (wrongly) be released when
  3354  	// heap is cleared with rpmalloc_heap_free_all(). Also heaps guaranteed to be
  3355  	// pristine from the dedicated orphan list can be used.
  3356  	heap_t* heap = _rpmalloc_heap_allocate(1);
  3357  	heap->owner_thread = 0;
  3358  	_rpmalloc_stat_inc(&_memory_active_heaps);
  3359  	return heap;
  3360  }
  3361  
  3362  extern inline void
  3363  rpmalloc_heap_release(rpmalloc_heap_t* heap) {
  3364  	if (heap)
  3365  		_rpmalloc_heap_release(heap, 1, 1);
  3366  }
  3367  
  3368  extern inline RPMALLOC_ALLOCATOR void*
  3369  rpmalloc_heap_alloc(rpmalloc_heap_t* heap, size_t size) {
  3370  #if ENABLE_VALIDATE_ARGS
  3371  	if (size >= MAX_ALLOC_SIZE) {
  3372  		errno = EINVAL;
  3373  		return 0;
  3374  	}
  3375  #endif
  3376  	return _rpmalloc_allocate(heap, size);
  3377  }
  3378  
  3379  extern inline RPMALLOC_ALLOCATOR void*
  3380  rpmalloc_heap_aligned_alloc(rpmalloc_heap_t* heap, size_t alignment, size_t size) {
  3381  #if ENABLE_VALIDATE_ARGS
  3382  	if (size >= MAX_ALLOC_SIZE) {
  3383  		errno = EINVAL;
  3384  		return 0;
  3385  	}
  3386  #endif
  3387  	return _rpmalloc_aligned_allocate(heap, alignment, size);
  3388  }
  3389  
  3390  extern inline RPMALLOC_ALLOCATOR void*
  3391  rpmalloc_heap_calloc(rpmalloc_heap_t* heap, size_t num, size_t size) {
  3392  	return rpmalloc_heap_aligned_calloc(heap, 0, num, size);
  3393  }
  3394  
  3395  extern inline RPMALLOC_ALLOCATOR void*
  3396  rpmalloc_heap_aligned_calloc(rpmalloc_heap_t* heap, size_t alignment, size_t num, size_t size) {
  3397  	size_t total;
  3398  #if ENABLE_VALIDATE_ARGS
  3399  #if PLATFORM_WINDOWS
  3400  	int err = SizeTMult(num, size, &total);
  3401  	if ((err != S_OK) || (total >= MAX_ALLOC_SIZE)) {
  3402  		errno = EINVAL;
  3403  		return 0;
  3404  	}
  3405  #else
  3406  	int err = __builtin_umull_overflow(num, size, &total);
  3407  	if (err || (total >= MAX_ALLOC_SIZE)) {
  3408  		errno = EINVAL;
  3409  		return 0;
  3410  	}
  3411  #endif
  3412  #else
  3413  	total = num * size;
  3414  #endif
  3415  	void* block = _rpmalloc_aligned_allocate(heap, alignment, total);
  3416  	if (block)
  3417  		memset(block, 0, total);
  3418  	return block;
  3419  }
  3420  
  3421  extern inline RPMALLOC_ALLOCATOR void*
  3422  rpmalloc_heap_realloc(rpmalloc_heap_t* heap, void* ptr, size_t size, unsigned int flags) {
  3423  #if ENABLE_VALIDATE_ARGS
  3424  	if (size >= MAX_ALLOC_SIZE) {
  3425  		errno = EINVAL;
  3426  		return ptr;
  3427  	}
  3428  #endif
  3429  	return _rpmalloc_reallocate(heap, ptr, size, 0, flags);
  3430  }
  3431  
  3432  extern inline RPMALLOC_ALLOCATOR void*
  3433  rpmalloc_heap_aligned_realloc(rpmalloc_heap_t* heap, void* ptr, size_t alignment, size_t size, unsigned int flags) {
  3434  #if ENABLE_VALIDATE_ARGS
  3435  	if ((size + alignment < size) || (alignment > _memory_page_size)) {
  3436  		errno = EINVAL;
  3437  		return 0;
  3438  	}
  3439  #endif
  3440  	return _rpmalloc_aligned_reallocate(heap, ptr, alignment, size, 0, flags);
  3441  }
  3442  
  3443  extern inline void
  3444  rpmalloc_heap_free(rpmalloc_heap_t* heap, void* ptr) {
  3445  	(void)sizeof(heap);
  3446  	_rpmalloc_deallocate(ptr);
  3447  }
  3448  
  3449  extern inline void
  3450  rpmalloc_heap_free_all(rpmalloc_heap_t* heap) {
  3451  	span_t* span;
  3452  	span_t* next_span;
  3453  
  3454  	_rpmalloc_heap_cache_adopt_deferred(heap, 0);
  3455  
  3456  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  3457  		span = heap->size_class[iclass].partial_span;
  3458  		while (span) {
  3459  			next_span = span->next;
  3460  			_rpmalloc_heap_cache_insert(heap, span);
  3461  			span = next_span;
  3462  		}
  3463  		heap->size_class[iclass].partial_span = 0;
  3464  		span = heap->full_span[iclass];
  3465  		while (span) {
  3466  			next_span = span->next;
  3467  			_rpmalloc_heap_cache_insert(heap, span);
  3468  			span = next_span;
  3469  		}
  3470  	}
  3471  	memset(heap->size_class, 0, sizeof(heap->size_class));
  3472  	memset(heap->full_span, 0, sizeof(heap->full_span));
  3473  
  3474  	span = heap->large_huge_span;
  3475  	while (span) {
  3476  		next_span = span->next;
  3477  		if (UNEXPECTED(span->size_class == SIZE_CLASS_HUGE))
  3478  			_rpmalloc_deallocate_huge(span);
  3479  		else
  3480  			_rpmalloc_heap_cache_insert(heap, span);
  3481  		span = next_span;
  3482  	}
  3483  	heap->large_huge_span = 0;
  3484  	heap->full_span_count = 0;
  3485  
  3486  #if ENABLE_THREAD_CACHE
  3487  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3488  		span_cache_t* span_cache;
  3489  		if (!iclass)
  3490  			span_cache = &heap->span_cache;
  3491  		else
  3492  			span_cache = (span_cache_t*)(heap->span_large_cache + (iclass - 1));
  3493  		if (!span_cache->count)
  3494  			continue;
  3495  #if ENABLE_GLOBAL_CACHE
  3496  		_rpmalloc_stat_add64(&heap->thread_to_global, span_cache->count * (iclass + 1) * _memory_span_size);
  3497  		_rpmalloc_stat_add(&heap->span_use[iclass].spans_to_global, span_cache->count);
  3498  		_rpmalloc_global_cache_insert_spans(span_cache->span, iclass + 1, span_cache->count);
  3499  #else
  3500  		for (size_t ispan = 0; ispan < span_cache->count; ++ispan)
  3501  			_rpmalloc_span_unmap(span_cache->span[ispan]);
  3502  #endif
  3503  		span_cache->count = 0;
  3504  	}
  3505  #endif
  3506  
  3507  #if ENABLE_STATISTICS
  3508  	for (size_t iclass = 0; iclass < SIZE_CLASS_COUNT; ++iclass) {
  3509  		atomic_store32(&heap->size_class_use[iclass].alloc_current, 0);
  3510  		atomic_store32(&heap->size_class_use[iclass].spans_current, 0);
  3511  	}
  3512  	for (size_t iclass = 0; iclass < LARGE_CLASS_COUNT; ++iclass) {
  3513  		atomic_store32(&heap->span_use[iclass].current, 0);
  3514  	}
  3515  #endif
  3516  }
  3517  
  3518  extern inline void
  3519  rpmalloc_heap_thread_set_current(rpmalloc_heap_t* heap) {
  3520  	heap_t* prev_heap = get_thread_heap_raw();
  3521  	if (prev_heap != heap) {
  3522  		set_thread_heap(heap);
  3523  		if (prev_heap)
  3524  			rpmalloc_heap_release(prev_heap);
  3525  	}
  3526  }
  3527  
  3528  #endif
  3529  
  3530  #if ENABLE_PRELOAD || ENABLE_OVERRIDE
  3531  
  3532  #include "malloc.c"
  3533  
  3534  #endif
  3535  
  3536  void
  3537  rpmalloc_linker_reference(void) {
  3538  	(void)sizeof(_rpmalloc_initialized);
  3539  }