wip: further progress on removing segments

This commit is contained in:
daanx 2024-11-29 10:40:18 -08:00
parent 71cfa45e76
commit 441d4fed9f
21 changed files with 2396 additions and 2492 deletions

View file

@ -100,6 +100,10 @@ typedef int32_t mi_ssize_t;
#define __BMI1__ 1
#endif
// Define big endian if needed
// #define MI_BIG_ENDIAN 1
/* --------------------------------------------------------------------------------
Builtin's
-------------------------------------------------------------------------------- */
@ -310,4 +314,6 @@ static inline size_t mi_rotl(size_t x, size_t r) {
#endif
}
#endif // MI_BITS_H

View file

@ -108,6 +108,7 @@ size_t _mi_os_page_size(void);
size_t _mi_os_good_alloc_size(size_t size);
bool _mi_os_has_overcommit(void);
bool _mi_os_has_virtual_reserve(void);
size_t _mi_os_virtual_address_bits(void);
bool _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats);
bool _mi_os_commit(void* p, size_t size, bool* is_zero, mi_stats_t* stats);
@ -136,12 +137,11 @@ bool _mi_arena_contains(const void* p);
void _mi_arenas_collect(bool force_purge, mi_stats_t* stats);
void _mi_arena_unsafe_destroy_all(mi_stats_t* stats);
bool _mi_arena_segment_clear_abandoned(mi_segment_t* segment);
void _mi_arena_segment_mark_abandoned(mi_segment_t* segment);
void* _mi_arena_meta_zalloc(size_t size, mi_memid_t* memid);
void _mi_arena_meta_free(void* p, mi_memid_t memid, size_t size);
/*
typedef struct mi_arena_field_cursor_s { // abstract struct
size_t os_list_count; // max entries to visit in the OS abandoned list
size_t start; // start arena idx (may need to be wrapped)
@ -154,27 +154,12 @@ typedef struct mi_arena_field_cursor_s { // abstract struct
void _mi_arena_field_cursor_init(mi_heap_t* heap, mi_subproc_t* subproc, bool visit_all, mi_arena_field_cursor_t* current);
mi_segment_t* _mi_arena_segment_clear_abandoned_next(mi_arena_field_cursor_t* previous);
void _mi_arena_field_cursor_done(mi_arena_field_cursor_t* current);
*/
// "segment-map.c"
void _mi_segment_map_allocated_at(const mi_segment_t* segment);
void _mi_segment_map_freed_at(const mi_segment_t* segment);
// "page-map.c"
void _mi_page_map_register(mi_page_t* page);
void _mi_page_map_unregister(mi_page_t* page);
// "segment.c"
mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld);
void _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld);
void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size);
#if MI_HUGE_PAGE_ABANDON
void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
#else
void _mi_segment_huge_page_reset(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
#endif
void _mi_segments_collect(bool force, mi_segments_tld_t* tld);
void _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld);
bool _mi_segment_attempt_reclaim(mi_heap_t* heap, mi_segment_t* segment);
bool _mi_segment_visit_blocks(mi_segment_t* segment, int heap_tag, bool visit_blocks, mi_block_visit_fun* visitor, void* arg);
// "page.c"
void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept mi_attr_malloc;
@ -226,7 +211,7 @@ void* _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, siz
void* _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept;
mi_block_t* _mi_page_ptr_unalign(const mi_page_t* page, const void* p);
bool _mi_free_delayed_block(mi_block_t* block);
void _mi_free_generic(mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept; // for runtime integration
// void _mi_free_generic(mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept; // for runtime integration
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size);
// "libc.c"
@ -338,8 +323,8 @@ static inline uintptr_t _mi_align_up(uintptr_t sz, size_t alignment) {
// Align a pointer upwards
static inline void* mi_align_up_ptr(void* p, size_t alignment) {
return (void*)_mi_align_up((uintptr_t)p, alignment);
static inline uint8_t* _mi_align_up_ptr(void* p, size_t alignment) {
return (uint8_t*)_mi_align_up((uintptr_t)p, alignment);
}
@ -445,68 +430,44 @@ static inline mi_page_t* _mi_heap_get_free_small_page(mi_heap_t* heap, size_t si
return heap->pages_free_direct[idx];
}
// Segment that contains the pointer
// Large aligned blocks may be aligned at N*MI_SEGMENT_SIZE (inside a huge segment > MI_SEGMENT_SIZE),
// and we need align "down" to the segment info which is `MI_SEGMENT_SIZE` bytes before it;
// therefore we align one byte before `p`.
// We check for NULL afterwards on 64-bit systems to improve codegen for `mi_free`.
static inline mi_segment_t* _mi_ptr_segment(const void* p) {
mi_segment_t* const segment = (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
#if MI_INTPTR_SIZE <= 4
return (p==NULL ? NULL : segment);
#else
return ((intptr_t)segment <= 0 ? NULL : segment);
extern signed char* _mi_page_map;
#define MI_PAGE_PTR_INVALID ((mi_page_t*)(1))
static inline mi_page_t* _mi_ptr_page(const void* p) {
const uintptr_t up = ((uintptr_t)p) >> MI_ARENA_BLOCK_SHIFT;
const ptrdiff_t ofs = _mi_page_map[up];
#if MI_DEBUG
if mi_unlikely(ofs==0) return MI_PAGE_PTR_INVALID;
#endif
return (mi_page_t*)((up + ofs - 1) << MI_ARENA_BLOCK_SHIFT);
}
// Segment belonging to a page
static inline mi_segment_t* _mi_page_segment(const mi_page_t* page) {
mi_assert_internal(page!=NULL);
mi_segment_t* segment = _mi_ptr_segment(page);
mi_assert_internal(segment == NULL || page == &segment->pages[page->segment_idx]);
return segment;
}
// used internally
static inline size_t _mi_segment_page_idx_of(const mi_segment_t* segment, const void* p) {
// if (segment->page_size > MI_SEGMENT_SIZE) return &segment->pages[0]; // huge pages
ptrdiff_t diff = (uint8_t*)p - (uint8_t*)segment;
mi_assert_internal(diff >= 0 && (size_t)diff <= MI_SEGMENT_SIZE /* for huge alignment it can be equal */);
size_t idx = (size_t)diff >> segment->page_shift;
mi_assert_internal(idx < segment->capacity);
mi_assert_internal(segment->page_kind <= MI_PAGE_MEDIUM || idx == 0);
return idx;
}
// Get the page containing the pointer
static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const void* p) {
size_t idx = _mi_segment_page_idx_of(segment, p);
return &((mi_segment_t*)segment)->pages[idx];
}
// Quick page start for initialized pages
static inline uint8_t* mi_page_start(const mi_page_t* page) {
mi_assert_internal(page->page_start != NULL);
mi_assert_expensive(_mi_segment_page_start(_mi_page_segment(page),page,NULL) == page->page_start);
return page->page_start;
}
// Get the page containing the pointer
static inline mi_page_t* _mi_ptr_page(void* p) {
mi_assert_internal(p!=NULL);
return _mi_segment_page_of(_mi_ptr_segment(p), p);
}
// Get the block size of a page (special case for huge objects)
// Get the block size of a page
static inline size_t mi_page_block_size(const mi_page_t* page) {
mi_assert_internal(page->block_size > 0);
return page->block_size;
}
static inline bool mi_page_is_huge(const mi_page_t* page) {
mi_assert_internal((page->is_huge && _mi_page_segment(page)->page_kind == MI_PAGE_HUGE) ||
(!page->is_huge && _mi_page_segment(page)->page_kind != MI_PAGE_HUGE));
return page->is_huge;
// Page start
static inline uint8_t* mi_page_start(const mi_page_t* page) {
mi_assert(sizeof(mi_page_t) <= MI_PAGE_INFO_SIZE);
return (uint8_t*)page + MI_PAGE_INFO_SIZE;
}
static inline uint8_t* mi_page_area(const mi_page_t* page, size_t* size) {
if (size) { *size = mi_page_block_size(page) * page->reserved; }
return mi_page_start(page);
}
static inline bool mi_page_is_in_arena(const mi_page_t* page) {
return (page->memid.memkind == MI_MEM_ARENA);
}
static inline bool mi_page_is_singleton(const mi_page_t* page) {
return (page->reserved == 1);
}
// Get the usable block size of a page without fixed padding.
@ -515,11 +476,6 @@ static inline size_t mi_page_usable_block_size(const mi_page_t* page) {
return mi_page_block_size(page) - MI_PADDING_SIZE;
}
// size of a segment
static inline size_t mi_segment_size(mi_segment_t* segment) {
return segment->segment_size;
}
// Thread free access
static inline mi_block_t* mi_page_thread_free(const mi_page_t* page) {
return (mi_block_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xthread_free) & ~3);
@ -534,10 +490,20 @@ static inline mi_heap_t* mi_page_heap(const mi_page_t* page) {
return (mi_heap_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xheap));
}
static inline mi_threadid_t mi_page_thread_id(const mi_page_t* page) {
return mi_atomic_load_relaxed(&page->xthread_id);
}
static inline void mi_page_set_heap(mi_page_t* page, mi_heap_t* heap) {
mi_assert_internal(mi_page_thread_free_flag(page) != MI_DELAYED_FREEING);
mi_atomic_store_release(&page->xheap,(uintptr_t)heap);
if (heap != NULL) { page->heap_tag = heap->tag; }
if (heap != NULL) {
page->heap_tag = heap->tag;
mi_atomic_store_release(&page->xthread_id, heap->thread_id);
}
else {
mi_atomic_store_release(&page->xthread_id,0);
}
}
// Thread free flag helpers
@ -576,6 +542,21 @@ static inline bool mi_page_immediate_available(const mi_page_t* page) {
return (page->free != NULL);
}
// is the page not yet used up to its reserved space?
static inline bool mi_page_is_expandable(const mi_page_t* page) {
mi_assert_internal(page != NULL);
mi_assert_internal(page->capacity <= page->reserved);
return (page->capacity < page->reserved);
}
static inline bool mi_page_is_full(mi_page_t* page) {
bool full = (page->reserved == page->used);
mi_assert_internal(!full || page->free == NULL);
return full;
}
// is more than 7/8th of a page in use?
static inline bool mi_page_mostly_used(const mi_page_t* page) {
if (page==NULL) return true;
@ -583,6 +564,15 @@ static inline bool mi_page_mostly_used(const mi_page_t* page) {
return (page->reserved - page->used <= frac);
}
static inline bool mi_page_is_abandoned(mi_page_t* page) {
return (mi_page_thread_id(page) == 0);
}
static inline bool mi_page_is_huge(mi_page_t* page) {
return (page->block_size > MI_LARGE_MAX_OBJ_SIZE);
}
static inline mi_page_queue_t* mi_page_queue(const mi_heap_t* heap, size_t size) {
return &((mi_heap_t*)heap)->pages[_mi_bin(size)];
}
@ -667,17 +657,8 @@ We also pass a separate `null` value to be used as `NULL` or otherwise
`(k2<<<k1)+k1` would appear (too) often as a sentinel value.
------------------------------------------------------------------- */
static inline bool mi_is_in_same_segment(const void* p, const void* q) {
return (_mi_ptr_segment(p) == _mi_ptr_segment(q));
}
static inline bool mi_is_in_same_page(const void* p, const void* q) {
mi_segment_t* segmentp = _mi_ptr_segment(p);
mi_segment_t* segmentq = _mi_ptr_segment(q);
if (segmentp != segmentq) return false;
size_t idxp = _mi_segment_page_idx_of(segmentp, p);
size_t idxq = _mi_segment_page_idx_of(segmentq, q);
return (idxp == idxq);
return (_mi_ptr_page(p) == _mi_ptr_page(q));
}
static inline void* mi_ptr_decode(const void* null, const mi_encoded_t x, const uintptr_t* keys) {
@ -693,7 +674,7 @@ static inline mi_encoded_t mi_ptr_encode(const void* null, const void* p, const
static inline uint32_t mi_ptr_encode_canary(const void* null, const void* p, const uintptr_t* keys) {
const uint32_t x = (uint32_t)(mi_ptr_encode(null,p,keys));
// make the lowest byte 0 to prevent spurious read overflows which could be a security issue (issue #951)
#ifdef MI_BIG_ENDIAN
#if MI_BIG_ENDIAN
return (x & 0x00FFFFFF);
#else
return (x & 0xFFFFFF00);
@ -749,6 +730,20 @@ static inline void mi_block_set_next(const mi_page_t* page, mi_block_t* block, c
#endif
}
/* -----------------------------------------------------------
arena blocks
----------------------------------------------------------- */
// Blocks needed for a given byte size
static inline size_t mi_block_count_of_size(size_t size) {
return _mi_divide_up(size, MI_ARENA_BLOCK_SIZE);
}
// Byte size of a number of blocks
static inline size_t mi_size_of_blocks(size_t bcount) {
return (bcount * MI_ARENA_BLOCK_SIZE);
}
/* -----------------------------------------------------------
memory id's

View file

@ -111,40 +111,29 @@ terms of the MIT license. A copy of the license can be found in the file
// Main internal data-structures
// ------------------------------------------------------
// Main tuning parameters for segment and page sizes
// Sizes for 64-bit, divide by two for 32-bit
#ifndef MI_SMALL_PAGE_SHIFT
#define MI_SMALL_PAGE_SHIFT (13 + MI_INTPTR_SHIFT) // 64KiB
// Sizes are for 64-bit
#ifndef MI_ARENA_BLOCK_SHIFT
#ifdef MI_SMALL_PAGE_SHIFT // compatibility
#define MI_ARENA_BLOCK_SHIFT MI_SMALL_PAGE_SHIFT
#else
#define MI_ARENA_BLOCK_SHIFT (13 + MI_SIZE_SHIFT) // 64 KiB (32 KiB on 32-bit)
#endif
#ifndef MI_MEDIUM_PAGE_SHIFT
#define MI_MEDIUM_PAGE_SHIFT ( 3 + MI_SMALL_PAGE_SHIFT) // 512KiB
#endif
#ifndef MI_LARGE_PAGE_SHIFT
#define MI_LARGE_PAGE_SHIFT ( 3 + MI_MEDIUM_PAGE_SHIFT) // 4MiB
#endif
#ifndef MI_SEGMENT_SHIFT
#define MI_SEGMENT_SHIFT ( MI_LARGE_PAGE_SHIFT) // 4MiB -- must be equal to `MI_LARGE_PAGE_SHIFT`
#ifndef MI_BITMAP_CHUNK_BITS_SHIFT
#define MI_BITMAP_CHUNK_BITS_SHIFT 8 // optimized for 256 bits per chunk (avx2)
#endif
// Derived constants
#define MI_SEGMENT_SIZE (MI_ZU(1)<<MI_SEGMENT_SHIFT)
#define MI_SEGMENT_ALIGN (MI_SEGMENT_SIZE)
#define MI_SEGMENT_MASK ((uintptr_t)(MI_SEGMENT_ALIGN - 1))
#define MI_ARENA_BLOCK_SIZE (MI_ZU(1) << MI_ARENA_BLOCK_SHIFT)
#define MI_ARENA_BLOCK_ALIGN (MI_ARENA_BLOCK_SIZE)
#define MI_BITMAP_CHUNK_BITS (MI_ZU(1) << MI_BITMAP_CHUNK_BITS_SHIFT)
#define MI_SMALL_PAGE_SIZE (MI_ZU(1)<<MI_SMALL_PAGE_SHIFT)
#define MI_MEDIUM_PAGE_SIZE (MI_ZU(1)<<MI_MEDIUM_PAGE_SHIFT)
#define MI_LARGE_PAGE_SIZE (MI_ZU(1)<<MI_LARGE_PAGE_SHIFT)
#define MI_ARENA_MIN_OBJ_SIZE MI_ARENA_BLOCK_SIZE
#define MI_ARENA_MAX_OBJ_SIZE (MI_BITMAP_CHUNK_BITS * MI_ARENA_BLOCK_SIZE) // for now, cannot cross chunk boundaries
#define MI_SMALL_PAGES_PER_SEGMENT (MI_SEGMENT_SIZE/MI_SMALL_PAGE_SIZE)
#define MI_MEDIUM_PAGES_PER_SEGMENT (MI_SEGMENT_SIZE/MI_MEDIUM_PAGE_SIZE)
#define MI_LARGE_PAGES_PER_SEGMENT (MI_SEGMENT_SIZE/MI_LARGE_PAGE_SIZE)
#define MI_SMALL_PAGE_SIZE MI_ARENA_MIN_OBJ_SIZE
#define MI_MEDIUM_PAGE_SIZE (8*MI_SMALL_PAGE_SIZE) // 512 KiB (=byte in the bitmap)
#define MI_LARGE_PAGE_SIZE (MI_SIZE_SIZE*MI_MEDIUM_PAGE_SIZE) // 4 MiB (=word in the bitmap)
// The max object size are checked to not waste more than 12.5% internally over the page sizes.
// (Except for large pages since huge objects are allocated in 4MiB chunks)
#define MI_SMALL_OBJ_SIZE_MAX (MI_SMALL_PAGE_SIZE/4) // 16KiB
#define MI_MEDIUM_OBJ_SIZE_MAX (MI_MEDIUM_PAGE_SIZE/4) // 128KiB
#define MI_LARGE_OBJ_SIZE_MAX (MI_LARGE_PAGE_SIZE/2) // 2MiB
#define MI_LARGE_OBJ_WSIZE_MAX (MI_LARGE_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
// Maximum number of size classes. (spaced exponentially in 12.5% increments)
#define MI_BIN_HUGE (73U)
@ -152,18 +141,54 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_BIN_COUNT (MI_BIN_FULL+1)
#if (MI_LARGE_OBJ_WSIZE_MAX >= 655360)
#error "mimalloc internal: define more bins"
#endif
// Maximum block size for which blocks are guaranteed to be block size aligned. (see `segment.c:_mi_segment_page_start`)
#define MI_MAX_ALIGN_GUARANTEE (MI_MEDIUM_OBJ_SIZE_MAX)
// Alignments over MI_BLOCK_ALIGNMENT_MAX are allocated in dedicated huge page segments
#define MI_BLOCK_ALIGNMENT_MAX (MI_SEGMENT_SIZE >> 1)
// Alignments over MI_BLOCK_ALIGNMENT_MAX are allocated in dedicated orphan pages
#define MI_BLOCK_ALIGNMENT_MAX (MI_ARENA_BLOCK_ALIGN)
// We never allocate more than PTRDIFF_MAX (see also <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
#define MI_MAX_ALLOC_SIZE PTRDIFF_MAX
#define MI_MAX_ALLOC_SIZE PTRDIFF_MAX
// ---------------------------------------------------------------
// a memory id tracks the provenance of arena/OS allocated memory
// ---------------------------------------------------------------
// Memory can reside in arena's, direct OS allocated, or statically allocated. The memid keeps track of this.
typedef enum mi_memkind_e {
MI_MEM_NONE, // not allocated
MI_MEM_EXTERNAL, // not owned by mimalloc but provided externally (via `mi_manage_os_memory` for example)
MI_MEM_STATIC, // allocated in a static area and should not be freed (for arena meta data for example)
MI_MEM_OS, // allocated from the OS
MI_MEM_OS_HUGE, // allocated as huge OS pages (usually 1GiB, pinned to physical memory)
MI_MEM_OS_REMAP, // allocated in a remapable area (i.e. using `mremap`)
MI_MEM_ARENA // allocated from an arena (the usual case)
} mi_memkind_t;
static inline bool mi_memkind_is_os(mi_memkind_t memkind) {
return (memkind >= MI_MEM_OS && memkind <= MI_MEM_OS_REMAP);
}
typedef struct mi_memid_os_info {
void* base; // actual base address of the block (used for offset aligned allocations)
size_t alignment; // alignment at allocation
} mi_memid_os_info_t;
typedef struct mi_memid_arena_info {
size_t block_index; // index in the arena
mi_arena_id_t id; // arena id (>= 1)
bool is_exclusive; // this arena can only be used for specific arena allocations
} mi_memid_arena_info_t;
typedef struct mi_memid_s {
union {
mi_memid_os_info_t os; // only used for MI_MEM_OS
mi_memid_arena_info_t arena; // only used for MI_MEM_ARENA
} mem;
bool is_pinned; // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large (2Mib) or huge (1GiB) OS pages)
bool initially_committed;// `true` if the memory was originally allocated as committed
bool initially_zero; // `true` if the memory was originally zero initialized
mi_memkind_t memkind;
} mi_memid_t;
// ------------------------------------------------------
// Mimalloc pages contain allocated blocks
@ -223,6 +248,10 @@ typedef union mi_page_flags_s {
// We use the bottom 2 bits of the pointer for mi_delayed_t flags
typedef uintptr_t mi_thread_free_t;
// Sub processes are used to keep memory separate between them (e.g. multiple interpreters in CPython)
typedef struct mi_subproc_s mi_subproc_t;
// A page contains blocks of one specific size (`block_size`).
// Each page has three list of free blocks:
// `free` for blocks that can be allocated,
@ -242,8 +271,6 @@ typedef uintptr_t mi_thread_free_t;
// Notes:
// - Access is optimized for `free.c:mi_free` and `alloc.c:mi_page_alloc`
// - Using `uint16_t` does not seem to slow things down
// - The size is 10 words on 64-bit which helps the page index calculations
// (and 12 words on 32-bit, and encoded free lists add 2 words)
// - `xthread_free` uses the bottom bits as a delayed-free flags to optimize
// concurrent frees where only the first concurrent free adds to the owning
// heap `thread_delayed_free` list (see `free.c:mi_free_block_mt`).
@ -252,15 +279,8 @@ typedef uintptr_t mi_thread_free_t;
// the owning heap `thread_delayed_free` list. This guarantees that pages
// will be freed correctly even if only other threads free blocks.
typedef struct mi_page_s {
// "owned" by the segment
uint8_t segment_idx; // index in the segment `pages` array, `page == &segment->pages[page->segment_idx]`
uint8_t segment_in_use:1; // `true` if the segment allocated this page
uint8_t is_committed:1; // `true` if the page virtual memory is committed
uint8_t is_zero_init:1; // `true` if the page was initially zero initialized
uint8_t is_huge:1; // `true` if the page is in a huge segment
// layout like this to optimize access in `mi_malloc` and `mi_free`
uint16_t capacity; // number of blocks committed, must be the first field, see `segment.c:page_clear`
mi_memid_t memid; // provenance of the page memory
uint16_t capacity; // number of blocks committed (must be the first field for proper zero-initialisation)
uint16_t reserved; // number of blocks reserved in memory
mi_page_flags_t flags; // `in_full` and `has_aligned` flags (8 bits)
uint8_t free_is_zero:1; // `true` if the blocks in the free list are zero initialized
@ -272,120 +292,54 @@ typedef struct mi_page_s {
uint8_t block_size_shift; // if not zero, then `(1 << block_size_shift) == block_size` (only used for fast path in `free.c:_mi_page_ptr_unalign`)
uint8_t heap_tag; // tag of the owning heap, used to separate heaps by object type
// padding
size_t block_size; // size available in each block (always `>0`)
uint8_t* page_start; // start of the page area containing the blocks
size_t block_size; // size available in each block (always `>0`)
#if (MI_ENCODE_FREELIST || MI_PADDING)
uintptr_t keys[2]; // two random keys to encode the free lists (see `_mi_block_next`) or padding canary
#endif
_Atomic(mi_thread_free_t) xthread_free; // list of deferred free blocks freed by other threads
_Atomic(uintptr_t) xheap;
_Atomic(uintptr_t) xheap; // heap this threads belong to.
_Atomic(mi_threadid_t)xthread_id; // thread this page belongs to. (= xheap->thread_id, or 0 if abandoned)
struct mi_page_s* next; // next page owned by the heap with the same `block_size`
struct mi_page_s* prev; // previous page owned by the heap with the same `block_size`
#if MI_INTPTR_SIZE==4 // pad to 12 words on 32-bit
void* padding[1];
#endif
} mi_page_t;
// ------------------------------------------------------
// Object sizes
// ------------------------------------------------------
#define MI_PAGE_ALIGN (64)
#define MI_PAGE_INFO_SIZE (MI_SIZE_SHIFT*MI_PAGE_ALIGN) // should be > sizeof(mi_page_t)
// The max object size are checked to not waste more than 12.5% internally over the page sizes.
// (Except for large pages since huge objects are allocated in 4MiB chunks)
#define MI_SMALL_MAX_OBJ_SIZE ((MI_SMALL_PAGE_SIZE-MI_PAGE_INFO_SIZE)/4) // ~16KiB
#define MI_MEDIUM_MAX_OBJ_SIZE ((MI_MEDIUM_PAGE_SIZE-MI_PAGE_INFO_SIZE)/4) // ~128KiB
#define MI_LARGE_MAX_OBJ_SIZE ((MI_LARGE_PAGE_SIZE-MI_PAGE_INFO_SIZE)/2) // ~2MiB
#define MI_LARGE_MAX_OBJ_WSIZE (MI_LARGE_MAX_OBJ_SIZE/MI_SIZE_SIZE)
#if (MI_LARGE_MAX_OBJ_WSIZE >= 655360)
#error "mimalloc internal: define more bins"
#endif
// ------------------------------------------------------
// Mimalloc segments contain mimalloc pages
// Page kinds
// ------------------------------------------------------
typedef enum mi_page_kind_e {
MI_PAGE_SMALL, // small blocks go into 64KiB pages inside a segment
MI_PAGE_MEDIUM, // medium blocks go into 512KiB pages inside a segment
MI_PAGE_LARGE, // larger blocks go into a single page spanning a whole segment
MI_PAGE_HUGE // a huge page is a single page in a segment of variable size (but still 2MiB aligned)
MI_PAGE_SMALL, // small blocks go into 64KiB pages
MI_PAGE_MEDIUM, // medium blocks go into 512KiB pages
MI_PAGE_LARGE, // larger blocks go into 4MiB pages
MI_PAGE_SINGLETON // page containing a single block.
// used for blocks `> MI_LARGE_OBJ_SIZE_MAX` or an aligment `> MI_BLOCK_ALIGNMENT_MAX`.
} mi_page_kind_t;
// ---------------------------------------------------------------
// a memory id tracks the provenance of arena/OS allocated memory
// ---------------------------------------------------------------
// Memory can reside in arena's, direct OS allocated, or statically allocated. The memid keeps track of this.
typedef enum mi_memkind_e {
MI_MEM_NONE, // not allocated
MI_MEM_EXTERNAL, // not owned by mimalloc but provided externally (via `mi_manage_os_memory` for example)
MI_MEM_STATIC, // allocated in a static area and should not be freed (for arena meta data for example)
MI_MEM_OS, // allocated from the OS
MI_MEM_OS_HUGE, // allocated as huge OS pages (usually 1GiB, pinned to physical memory)
MI_MEM_OS_REMAP, // allocated in a remapable area (i.e. using `mremap`)
MI_MEM_ARENA // allocated from an arena (the usual case)
} mi_memkind_t;
static inline bool mi_memkind_is_os(mi_memkind_t memkind) {
return (memkind >= MI_MEM_OS && memkind <= MI_MEM_OS_REMAP);
}
typedef struct mi_memid_os_info {
void* base; // actual base address of the block (used for offset aligned allocations)
size_t alignment; // alignment at allocation
} mi_memid_os_info_t;
typedef struct mi_memid_arena_info {
size_t block_index; // index in the arena
mi_arena_id_t id; // arena id (>= 1)
bool is_exclusive; // this arena can only be used for specific arena allocations
} mi_memid_arena_info_t;
typedef struct mi_memid_s {
union {
mi_memid_os_info_t os; // only used for MI_MEM_OS
mi_memid_arena_info_t arena; // only used for MI_MEM_ARENA
} mem;
bool is_pinned; // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large (2Mib) or huge (1GiB) OS pages)
bool initially_committed;// `true` if the memory was originally allocated as committed
bool initially_zero; // `true` if the memory was originally zero initialized
mi_memkind_t memkind;
} mi_memid_t;
// ---------------------------------------------------------------
// Segments contain mimalloc pages
// ---------------------------------------------------------------
typedef struct mi_subproc_s mi_subproc_t;
// Segments are large allocated memory blocks (2MiB on 64 bit) from the OS.
// Inside segments we allocated fixed size _pages_ that contain blocks.
typedef struct mi_segment_s {
// constant fields
mi_memid_t memid; // memory id to track provenance
bool allow_decommit;
bool allow_purge;
size_t segment_size; // for huge pages this may be different from `MI_SEGMENT_SIZE`
mi_subproc_t* subproc; // segment belongs to sub process
// segment fields
struct mi_segment_s* next; // must be the first (non-constant) segment field -- see `segment.c:segment_init`
struct mi_segment_s* prev;
bool was_reclaimed; // true if it was reclaimed (used to limit reclaim-on-free reclamation)
bool dont_free; // can be temporarily true to ensure the segment is not freed
size_t abandoned; // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`)
size_t abandoned_visits; // count how often this segment is visited for reclaiming (to force reclaim if it is too long)
size_t used; // count of pages in use (`used <= capacity`)
size_t capacity; // count of available pages (`#free + used`)
size_t segment_info_size;// space we are using from the first page for segment meta-data and possible guard pages.
uintptr_t cookie; // verify addresses in secure mode: `_mi_ptr_cookie(segment) == segment->cookie`
struct mi_segment_s* abandoned_os_next; // only used for abandoned segments outside arena's, and only if `mi_option_visit_abandoned` is enabled
struct mi_segment_s* abandoned_os_prev;
// layout like this to optimize access in `mi_free`
_Atomic(mi_threadid_t) thread_id; // unique id of the thread owning this segment
size_t page_shift; // `1 << page_shift` == the page sizes == `page->block_size * page->reserved` (unless the first page, then `-segment_info_size`).
mi_page_kind_t page_kind; // kind of pages: small, medium, large, or huge
mi_page_t pages[1]; // up to `MI_SMALL_PAGES_PER_SEGMENT` pages
} mi_segment_t;
// ------------------------------------------------------
// Heaps
@ -522,21 +476,18 @@ typedef struct mi_stat_counter_s {
} mi_stat_counter_t;
typedef struct mi_stats_s {
mi_stat_count_t segments;
mi_stat_count_t pages;
mi_stat_count_t reserved;
mi_stat_count_t committed;
mi_stat_count_t reset;
mi_stat_count_t purged;
mi_stat_count_t page_committed;
mi_stat_count_t segments_abandoned;
mi_stat_count_t pages_abandoned;
mi_stat_count_t threads;
mi_stat_count_t normal;
mi_stat_count_t huge;
mi_stat_count_t giant;
mi_stat_count_t malloc;
mi_stat_count_t segments_cache;
mi_stat_counter_t pages_extended;
mi_stat_counter_t mmap_calls;
mi_stat_counter_t commit_calls;
@ -581,12 +532,12 @@ void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount);
// ------------------------------------------------------
struct mi_subproc_s {
_Atomic(size_t) abandoned_count; // count of abandoned segments for this sub-process
_Atomic(size_t) abandoned_os_list_count; // count of abandoned segments in the os-list
mi_lock_t abandoned_os_lock; // lock for the abandoned os segment list (outside of arena's) (this lock protect list operations)
_Atomic(size_t) abandoned_count; // count of abandoned pages for this sub-process
_Atomic(size_t) abandoned_os_list_count; // count of abandoned pages in the os-list
mi_lock_t abandoned_os_lock; // lock for the abandoned os pages list (outside of arena's) (this lock protect list operations)
mi_lock_t abandoned_os_visit_lock; // ensure only one thread per subproc visits the abandoned os list
mi_segment_t* abandoned_os_list; // doubly-linked list of abandoned segments outside of arena's (in OS allocated memory)
mi_segment_t* abandoned_os_list_tail; // the tail-end of the list
mi_page_t* abandoned_os_list; // doubly-linked list of abandoned pages outside of arena's (in OS allocated memory)
mi_page_t* abandoned_os_list_tail; // the tail-end of the list
mi_memid_t memid; // provenance of this memory block
};
@ -597,11 +548,6 @@ struct mi_subproc_s {
// Milliseconds as in `int64_t` to avoid overflows
typedef int64_t mi_msecs_t;
// Queue of segments
typedef struct mi_segment_queue_s {
mi_segment_t* first;
mi_segment_t* last;
} mi_segment_queue_t;
// OS thread local data
typedef struct mi_os_tld_s {
@ -609,28 +555,13 @@ typedef struct mi_os_tld_s {
mi_stats_t* stats; // points to tld stats
} mi_os_tld_t;
// Segments thread local data
typedef struct mi_segments_tld_s {
mi_segment_queue_t small_free; // queue of segments with free small pages
mi_segment_queue_t medium_free; // queue of segments with free medium pages
mi_page_queue_t pages_purge; // queue of freed pages that are delay purged
size_t count; // current number of segments;
size_t peak_count; // peak number of segments
size_t current_size; // current size of all segments
size_t peak_size; // peak size of all segments
size_t reclaim_count;// number of reclaimed (abandoned) segments
mi_subproc_t* subproc; // sub-process this thread belongs to.
mi_stats_t* stats; // points to tld stats
mi_os_tld_t* os; // points to os tld
} mi_segments_tld_t;
// Thread local data
struct mi_tld_s {
unsigned long long heartbeat; // monotonic heartbeat count
bool recurse; // true if deferred was called; used to prevent infinite recursion.
mi_heap_t* heap_backing; // backing heap of this thread (cannot be deleted)
mi_heap_t* heaps; // list of heaps in this thread (so we can abandon all when the thread terminates)
mi_segments_tld_t segments; // segment tld
mi_subproc_t* subproc; // sub-process this thread belongs to.
mi_os_tld_t os; // os tld
mi_stats_t stats; // statistics
};