kiper220/mimalloc
1
0
Fork 0
mirror of https://github.com/microsoft/mimalloc.git synced 2025-05-06 23:39:31 +03:00
Code Issues Projects Releases Packages Wiki Activity Actions

initial no more pairmap

Browse source
This commit is contained in:
daanx 2024-12-06 14:53:24 -08:00
parent 7443ee317e
commit ec9c61c066
5 changed files with 465 additions and 711 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
include/mimalloc/internal.h
View file

@ -700,7 +700,9 @@ static inline bool mi_page_try_claim_ownership(mi_page_t* page) {
return ((old&1)==0);
}
static inline void _mi_page_unown(mi_page_t* page) {
// release ownership of a page. This may free the page if all blocks were concurrently
// freed in the meantime. Returns true if the page was freed.
static inline bool _mi_page_unown(mi_page_t* page) {
mi_assert_internal(mi_page_is_owned(page));
mi_assert_internal(mi_page_is_abandoned(page));
mi_thread_free_t tf_new;
@ -712,13 +714,14 @@ static inline void _mi_page_unown(mi_page_t* page) {
if (mi_page_all_free(page)) { // it may become free just before unowning it
_mi_arena_page_unabandon(page);
_mi_arena_page_free(page);
return;
return true;
}
tf_old = mi_atomic_load_relaxed(&page->xthread_free);
}
mi_assert_internal(mi_tf_block(tf_old)==NULL);
tf_new = mi_tf_create(NULL, false);
} while (!mi_atomic_cas_weak_release(&page->xthread_free, &tf_old, tf_new));
return false;
}
//-----------------------------------------------------------

8
include/mimalloc/types.h
View file

@ -117,16 +117,16 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_ARENA_SLICE_SHIFT (13 + MI_SIZE_SHIFT) // 64 KiB (32 KiB on 32-bit)
#endif
#endif
#ifndef MI_BITMAP_CHUNK_BITS_SHIFT
#define MI_BITMAP_CHUNK_BITS_SHIFT (6 + MI_SIZE_SHIFT) // optimized for 512 bits per chunk (avx512)
#ifndef MI_BCHUNK_BITS_SHIFT
#define MI_BCHUNK_BITS_SHIFT (6 + MI_SIZE_SHIFT) // optimized for 512 bits per chunk (avx512)
#endif
#define MI_BITMAP_CHUNK_BITS (1 << MI_BITMAP_CHUNK_BITS_SHIFT)
#define MI_BCHUNK_BITS (1 << MI_BCHUNK_BITS_SHIFT)
#define MI_ARENA_SLICE_SIZE (MI_ZU(1) << MI_ARENA_SLICE_SHIFT)
#define MI_ARENA_SLICE_ALIGN (MI_ARENA_SLICE_SIZE)
#define MI_ARENA_MIN_OBJ_SLICES (1)
#define MI_ARENA_MAX_OBJ_SLICES (MI_BITMAP_CHUNK_BITS) // 32 MiB (for now, cannot cross chunk boundaries)
#define MI_ARENA_MAX_OBJ_SLICES (MI_BCHUNK_BITS) // 32 MiB (for now, cannot cross chunk boundaries)
#define MI_ARENA_MIN_OBJ_SIZE (MI_ARENA_MIN_OBJ_SLICES * MI_ARENA_SLICE_SIZE)
#define MI_ARENA_MAX_OBJ_SIZE (MI_ARENA_MAX_OBJ_SLICES * MI_ARENA_SLICE_SIZE)

62
src/arena.c
View file

@ -48,7 +48,7 @@ typedef struct mi_arena_s {
mi_bitmap_t* slices_committed; // is the slice committed? (i.e. accessible)
mi_bitmap_t* slices_purge; // can the slice be purged? (slice in purge => slice in free)
mi_bitmap_t* slices_dirty; // is the slice potentially non-zero?
mi_pairmap_t pages_abandoned[MI_BIN_COUNT]; // abandoned pages per size bin (a set bit means the start of the page)
mi_bitmap_t* pages_abandoned[MI_BIN_COUNT]; // abandoned pages per size bin (a set bit means the start of the page)
// the full queue contains abandoned full pages
// followed by the bitmaps (whose size depends on the arena size)
} mi_arena_t;
@ -476,16 +476,24 @@ void* _mi_arena_alloc(size_t size, bool commit, bool allow_large, mi_arena_id_t
Arena page allocation
----------------------------------------------------------- */
static bool mi_arena_claim_abandoned(size_t slice_index, void* arg1, void* arg2) {
mi_arena_t* arena = (mi_arena_t*)arg1;
mi_subproc_t* subproc = (mi_subproc_t*)arg2;
static bool mi_arena_claim_abandoned(size_t slice_index, void* arg1, void* arg2, bool* keep_abandoned) {
// found an abandoned page of the right size
// it is set busy for now so we can read safely even with concurrent mi_free reclaiming
// try to claim ownership atomically
mi_page_t* page = (mi_page_t*)mi_arena_slice_start(arena, slice_index);
if (subproc != page->subproc) return false;
if (!mi_page_try_claim_ownership(page)) return false;
mi_arena_t* const arena = (mi_arena_t*)arg1;
mi_subproc_t* const subproc = (mi_subproc_t*)arg2;
mi_page_t* const page = (mi_page_t*)mi_arena_slice_start(arena, slice_index);
// can we claim ownership?
if (!mi_page_try_claim_ownership(page)) {
*keep_abandoned = true;
return false;
}
if (subproc != page->subproc) {
// wrong sub-process.. we need to unown again, and perhaps not keep it abandoned
const bool freed = _mi_page_unown(page);
*keep_abandoned = !freed;
return false;
}
// yes, we can reclaim it
*keep_abandoned = false;
return true;
}
@ -505,9 +513,9 @@ static mi_page_t* mi_arena_page_try_find_abandoned(size_t slice_count, size_t bl
mi_forall_arenas(req_arena_id, allow_large, tseq, arena_id, arena)
{
size_t slice_index;
mi_pairmap_t* const pairmap = &arena->pages_abandoned[bin];
mi_bitmap_t* const bitmap = arena->pages_abandoned[bin];
if (mi_pairmap_try_find_and_set_busy(pairmap, tseq, &slice_index, &mi_arena_claim_abandoned, arena, subproc)) {
if (mi_bitmap_try_find_and_claim(bitmap, tseq, &slice_index, &mi_arena_claim_abandoned, arena, subproc)) {
// found an abandoned page of the right size
// and claimed ownership.
mi_page_t* page = (mi_page_t*)mi_arena_slice_start(arena, slice_index);
@ -694,7 +702,7 @@ void _mi_arena_page_free(mi_page_t* page) {
mi_assert_internal(mi_bitmap_is_clearN(arena->slices_free, slice_index, slice_count));
mi_assert_internal(mi_bitmap_is_setN(arena->slices_committed, slice_index, slice_count));
mi_assert_internal(mi_bitmap_is_clearN(arena->slices_purge, slice_index, slice_count));
mi_assert_internal(mi_pairmap_is_clear(&arena->pages_abandoned[bin], slice_index));
mi_assert_internal(mi_bitmap_is_clearN(arena->pages_abandoned[bin], slice_index, 1));
}
#endif
@ -728,8 +736,8 @@ static void mi_arena_page_abandon_no_stat(mi_page_t* page) {
mi_assert_internal(mi_bitmap_is_setN(arena->slices_dirty, slice_index, slice_count));
mi_page_set_abandoned_mapped(page);
bool were_zero = mi_pairmap_set(&arena->pages_abandoned[bin], slice_index);
MI_UNUSED(were_zero); mi_assert_internal(were_zero);
const bool wasclear = mi_bitmap_set(arena->pages_abandoned[bin], slice_index);
MI_UNUSED(wasclear); mi_assert_internal(wasclear);
mi_atomic_increment_relaxed(&subproc->abandoned_count[bin]);
}
else {
@ -783,7 +791,7 @@ void _mi_arena_page_unabandon(mi_page_t* page) {
mi_assert_internal(mi_bitmap_is_clearN(arena->slices_purge, slice_index, slice_count));
// this busy waits until a concurrent reader (from alloc_abandoned) is done
mi_pairmap_clear_once_not_busy(&arena->pages_abandoned[bin], slice_index);
mi_bitmap_clear_once_set(arena->pages_abandoned[bin], slice_index);
mi_page_clear_abandoned_mapped(page);
mi_atomic_decrement_relaxed(&page->subproc->abandoned_count[bin]);
}
@ -956,12 +964,12 @@ static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id, mi_stats_t*
}
static size_t mi_arena_info_slices_needed(size_t slice_count, size_t* bitmap_base) {
if (slice_count == 0) slice_count = MI_BITMAP_CHUNK_BITS;
mi_assert_internal((slice_count % MI_BITMAP_CHUNK_BITS) == 0);
const size_t base_size = _mi_align_up(sizeof(mi_arena_t), MI_BITMAP_CHUNK_SIZE);
const size_t bitmaps_size = 4 * mi_bitmap_size(slice_count,NULL);
const size_t pairmaps_size = MI_BIN_COUNT * 2 * mi_bitmap_size(slice_count,NULL);
const size_t size = base_size + bitmaps_size + pairmaps_size;
if (slice_count == 0) slice_count = MI_BCHUNK_BITS;
mi_assert_internal((slice_count % MI_BCHUNK_BITS) == 0);
const size_t base_size = _mi_align_up(sizeof(mi_arena_t), MI_BCHUNK_SIZE);
const size_t bitmaps_count = 4 + MI_BIN_COUNT; // free, commit, dirty, purge, and abandonded
const size_t bitmaps_size = bitmaps_count * mi_bitmap_size(slice_count,NULL);
const size_t size = base_size + bitmaps_size;
const size_t os_page_size = _mi_os_page_size();
const size_t info_size = _mi_align_up(size, os_page_size) + os_page_size; // + guard page
@ -992,7 +1000,7 @@ static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int
if (arena_id != NULL) { *arena_id = _mi_arena_id_none(); }
const size_t slice_count = _mi_align_down(size / MI_ARENA_SLICE_SIZE, MI_BITMAP_CHUNK_BITS);
const size_t slice_count = _mi_align_down(size / MI_ARENA_SLICE_SIZE, MI_BCHUNK_BITS);
if (slice_count > MI_BITMAP_MAX_BIT_COUNT) { // 16 GiB for now
// todo: allow larger areas (either by splitting it up in arena's or having larger arena's)
_mi_warning_message("cannot use OS memory since it is too large (size %zu MiB, maximum is %zu MiB)", size/MI_MiB, mi_size_of_slices(MI_BITMAP_MAX_BIT_COUNT)/MI_MiB);
@ -1034,7 +1042,7 @@ static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int
arena->slices_dirty = mi_arena_bitmap_init(slice_count,&base);
arena->slices_purge = mi_arena_bitmap_init(slice_count,&base);
for( size_t i = 0; i < MI_ARENA_BIN_COUNT; i++) {
mi_pairmap_init(&arena->pages_abandoned[i], mi_arena_bitmap_init(slice_count, &base), mi_arena_bitmap_init(slice_count, &base));
arena->pages_abandoned[i] = mi_arena_bitmap_init(slice_count,&base);
}
mi_assert_internal(mi_size_of_slices(info_slices) >= (size_t)(base - mi_arena_start(arena)));
@ -1112,9 +1120,9 @@ static size_t mi_debug_show_bitmap(const char* prefix, const char* header, size_
size_t bit_count = 0;
size_t bit_set_count = 0;
for (size_t i = 0; i < mi_bitmap_chunk_count(bitmap) && bit_count < slice_count; i++) {
char buf[MI_BITMAP_CHUNK_BITS + 64]; _mi_memzero(buf, sizeof(buf));
mi_bitmap_chunk_t* chunk = &bitmap->chunks[i];
for (size_t j = 0, k = 0; j < MI_BITMAP_CHUNK_FIELDS; j++) {
char buf[MI_BCHUNK_BITS + 64]; _mi_memzero(buf, sizeof(buf));
mi_bchunk_t* chunk = &bitmap->chunks[i];
for (size_t j = 0, k = 0; j < MI_BCHUNK_FIELDS; j++) {
if (j > 0 && (j % 4) == 0) {
buf[k++] = '\n';
_mi_memcpy(buf+k, prefix, strlen(prefix)); k += strlen(prefix);

929
src/bitmap.c
View file

File diff suppressed because it is too large Load diff

148
src/bitmap.h
View file

@ -19,35 +19,34 @@ Concurrent bitmap that can set/reset sequences of bits atomically
each bit usually represents a single MI_ARENA_SLICE_SIZE in an arena (64 KiB).
We need 16K bits to represent a 1GiB arena.
`mi_bitmap_chunk_t`: a chunk of bfield's of a total of MI_BITMAP_CHUNK_BITS (= 512)
`mi_bchunk_t`: a chunk of bfield's of a total of MI_BCHUNK_BITS (= 512 on 64-bit, 256 on 32-bit)
allocations never span across chunks -- so MI_ARENA_MAX_OBJ_SIZE is the number
of bits in a chunk times the MI_ARENA_SLICE_SIZE (512 * 64KiB = 32 MiB).
These chunks are cache-aligned and we can use AVX2/AVX512/SVE/SVE2/etc. instructions
These chunks are cache-aligned and we can use AVX2/AVX512/NEON/SVE/SVE2/etc. instructions
to scan for bits (perhaps) more efficiently.
`mi_chunkmap_t`: for each chunk we track if it has (potentially) any bit set.
`mi_bchunkmap_t` == `mi_bchunk_t`: for each chunk we track if it has (potentially) any bit set.
The chunkmap has 1 bit per chunk that is set if the chunk potentially has a bit set.
This is used to avoid scanning every chunk. (and thus strictly an optimization)
It is conservative: it is fine to a bit in the chunk map even if the chunk turns out
to have no bits set.
to have no bits set. It is also allowed to briefly have a clear bit even if the
chunk has bits set, as long as we guarantee that we set the bit later on -- this
allows us to set the chunkmap bit after we set a bit in the corresponding chunk.
When we (potentially) set a bit in a chunk, we first update the chunkmap.
However, when we clear a bit in a chunk, and the chunk is indeed all clear, we
cannot safely clear the bit corresponding to the chunk in the chunkmap since it
may race with another thread setting a bit in the same chunk (and we may clear the
bit even though a bit is set in the chunk which is not allowed).
may race with another thread setting a bit in the same chunk. Therefore, when
clearing, we first test if a chunk is clear, then clear the chunkmap bit, and
then test again to catch any set bits that we missed.
To fix this, the chunkmap contains 32-bits of bits for chunks, and a 32-bit "epoch"
counter that is increased everytime a bit is set. We only clear a bit if the epoch
stayed the same over our clear operation (so we know no other thread in the mean
time set a bit in any of the chunks corresponding to the chunkmap).
Since increasing the epoch and setting a bit must be atomic, we use only half-word
bits (32) (we could use 128-bit atomics if needed since modern hardware supports this)
Since the chunkmap may thus be briefly out-of-sync, this means that we may sometimes
not find a free page even though it's there (but we accept this as we avoid taking
full locks). (Another way to do this is to use an epoch but we like to avoid that complexity
for now).
`mi_bitmap_t`: a bitmap with N chunks. A bitmap always has MI_BITMAP_MAX_CHUNK_FIELDS (=16)
and can support arena's from few chunks up to 16 chunkmap's = 16 * 32 chunks = 16 GiB
The `chunk_count` can be anything from 1 to the max supported by the chunkmap's but
each chunk is always complete (512 bits, so 512 * 64KiB = 32MiB memory area's).
`mi_bitmap_t`: a bitmap with N chunks. A bitmap has a chunkmap of MI_BCHUNK_BITS (512)
and thus has at most 512 chunks (=2^18 bits x 64 KiB slices = 16 GiB max arena size).
The minimum is 1 chunk which is a 32 MiB arena.
For now, the implementation assumes MI_HAS_FAST_BITSCAN and uses trailing-zero-count
and pop-count (but we think it can be adapted work reasonably well on older hardware too)
@ -59,57 +58,46 @@ typedef size_t mi_bfield_t;
#define MI_BFIELD_BITS_SHIFT (MI_SIZE_SHIFT+3)
#define MI_BFIELD_BITS (1 << MI_BFIELD_BITS_SHIFT)
#define MI_BFIELD_SIZE (MI_BFIELD_BITS/8)
#define MI_BFIELD_BITS_MOD_MASK (MI_BFIELD_BITS - 1)
#define MI_BFIELD_LO_BIT8 (((~(mi_bfield_t)0))/0xFF) // 0x01010101 ..
#define MI_BFIELD_HI_BIT8 (MI_BFIELD_LO_BIT8 << 7) // 0x80808080 ..
#define MI_BITMAP_CHUNK_SIZE (MI_BITMAP_CHUNK_BITS / 8)
#define MI_BITMAP_CHUNK_FIELDS (MI_BITMAP_CHUNK_BITS / MI_BFIELD_BITS)
#define MI_BITMAP_CHUNK_BITS_MOD_MASK (MI_BITMAP_CHUNK_BITS - 1)
// A bitmap chunk contains 512 bits of bfields on 64_bit (256 on 32-bit)
typedef mi_decl_align(MI_BITMAP_CHUNK_SIZE) struct mi_bitmap_chunk_s {
_Atomic(mi_bfield_t) bfields[MI_BITMAP_CHUNK_FIELDS];
} mi_bitmap_chunk_t;
#define MI_BCHUNK_SIZE (MI_BCHUNK_BITS / 8)
#define MI_BCHUNK_FIELDS (MI_BCHUNK_BITS / MI_BFIELD_BITS) // 8 on both 64- and 32-bit
// for now 32-bit epoch + 32-bit bit-set (note: with ABA instructions we can double this)
typedef uint64_t mi_chunkmap_t;
typedef uint32_t mi_epoch_t;
typedef uint32_t mi_cmap_t;
// A bitmap chunk contains 512 bits on 64-bit (256 on 32-bit)
typedef mi_decl_align(MI_BCHUNK_SIZE) struct mi_bchunk_s {
_Atomic(mi_bfield_t) bfields[MI_BCHUNK_FIELDS];
} mi_bchunk_t;
#define MI_CHUNKMAP_BITS (32) // 1 chunkmap tracks 32 chunks
// The chunkmap has one bit per corresponding chunk that is set if the chunk potentially has bits set.
// The chunkmap is itself a chunk.
typedef mi_bchunk_t mi_bchunkmap_t;
#define MI_BITMAP_MAX_CHUNKMAPS (16)
#define MI_BITMAP_MAX_CHUNK_COUNT (MI_BITMAP_MAX_CHUNKMAPS * MI_CHUNKMAP_BITS)
#define MI_BITMAP_MIN_CHUNK_COUNT (1 * MI_CHUNKMAP_BITS) // 1 GiB arena
#define MI_BCHUNKMAP_BITS MI_BCHUNK_BITS
#define MI_BITMAP_MAX_BIT_COUNT (MI_BITMAP_MAX_CHUNK_COUNT * MI_BITMAP_CHUNK_BITS) // 16 GiB arena
#define MI_BITMAP_MIN_BIT_COUNT (MI_BITMAP_MIN_CHUNK_COUNT * MI_BITMAP_CHUNK_BITS) // 1 GiB arena
#define MI_BITMAP_MAX_CHUNK_COUNT (MI_BCHUNKMAP_BITS)
#define MI_BITMAP_MIN_CHUNK_COUNT (1)
#define MI_BITMAP_MAX_BIT_COUNT (MI_BITMAP_MAX_CHUNK_COUNT * MI_BCHUNK_BITS) // 16 GiB arena
#define MI_BITMAP_MIN_BIT_COUNT (MI_BITMAP_MIN_CHUNK_COUNT * MI_BCHUNK_BITS) // 32 MiB arena
// An atomic bitmap
typedef mi_decl_align(MI_BITMAP_CHUNK_SIZE) struct mi_bitmap_s {
_Atomic(size_t) chunk_map_count; // valid chunk_maps entries
_Atomic(size_t) chunk_count; // total count of chunks
size_t padding[MI_BITMAP_CHUNK_SIZE/MI_SIZE_SIZE - 2]; // suppress warning on msvc
_Atomic(mi_chunkmap_t) chunk_maps[MI_BITMAP_MAX_CHUNKMAPS];
mi_bitmap_chunk_t chunks[MI_BITMAP_MIN_BIT_COUNT]; // or more, up to MI_BITMAP_MAX_CHUNK_COUNT
typedef mi_decl_align(MI_BCHUNK_SIZE) struct mi_bitmap_s {
_Atomic(size_t) chunk_count; // total count of chunks (0 < N <= MI_BCHUNKMAP_BITS)
size_t _padding[MI_BCHUNK_SIZE/MI_SIZE_SIZE - 1]; // suppress warning on msvc
mi_bchunkmap_t chunkmap;
mi_bchunk_t chunks[1]; // or more, up to MI_BITMAP_MAX_CHUNK_COUNT
} mi_bitmap_t;
static inline size_t mi_bitmap_chunk_map_count(const mi_bitmap_t* bitmap) {
return mi_atomic_load_relaxed(&bitmap->chunk_map_count);
}
static inline size_t mi_bitmap_chunk_count(const mi_bitmap_t* bitmap) {
return mi_atomic_load_relaxed(&bitmap->chunk_count);
}
static inline size_t mi_bitmap_max_bits(const mi_bitmap_t* bitmap) {
return (mi_bitmap_chunk_count(bitmap) * MI_BITMAP_CHUNK_BITS);
return (mi_bitmap_chunk_count(bitmap) * MI_BCHUNK_BITS);
}
@ -134,9 +122,22 @@ size_t mi_bitmap_init(mi_bitmap_t* bitmap, size_t bit_count, bool already_zero);
// Set/clear a sequence of `n` bits in the bitmap (and can cross chunks). Not atomic so only use if local to a thread.
void mi_bitmap_unsafe_setN(mi_bitmap_t* bitmap, size_t idx, size_t n);
// Set/clear a bit in the bitmap; returns `true` if atomically transitioned from 0 to 1 (or 1 to 0)
bool mi_bitmap_xset(mi_xset_t set, mi_bitmap_t* bitmap, size_t idx);
static inline bool mi_bitmap_set(mi_bitmap_t* bitmap, size_t idx) {
return mi_bitmap_xset(MI_BIT_SET, bitmap, idx);
}
static inline bool mi_bitmap_clear(mi_bitmap_t* bitmap, size_t idx) {
return mi_bitmap_xset(MI_BIT_CLEAR, bitmap, idx);
}
// Set/clear a sequence of `n` bits in the bitmap; returns `true` if atomically transitioned from all 0's to 1's (or all 1's to 0's).
// `n` cannot cross chunk boundaries (and `n <= MI_BITMAP_CHUNK_BITS`)!
// If `already_xset` is not NULL, it is set to true if all the bits were already all set/cleared.
// `n` cannot cross chunk boundaries (and `n <= MI_BCHUNK_BITS`)!
// If `already_xset` is not NULL, it is to all the bits were already all set/cleared.
bool mi_bitmap_xsetN(mi_xset_t set, mi_bitmap_t* bitmap, size_t idx, size_t n, size_t* already_xset);
static inline bool mi_bitmap_setN(mi_bitmap_t* bitmap, size_t idx, size_t n, size_t* already_set) {
@ -162,7 +163,7 @@ static inline bool mi_bitmap_is_clearN(mi_bitmap_t* bitmap, size_t idx, size_t n
// Try to set/clear a sequence of `n` bits in the bitmap; returns `true` if atomically transitioned from 0's to 1's (or 1's to 0's)
// and false otherwise leaving the bitmask as is.
// `n` cannot cross chunk boundaries (and `n <= MI_BITMAP_CHUNK_BITS`)!
// `n` cannot cross chunk boundaries (and `n <= MI_BCHUNK_BITS`)!
mi_decl_nodiscard bool mi_bitmap_try_xsetN(mi_xset_t set, mi_bitmap_t* bitmap, size_t idx, size_t n);
static inline bool mi_bitmap_try_setN(mi_bitmap_t* bitmap, size_t idx, size_t n) {
@ -177,48 +178,11 @@ static inline bool mi_bitmap_try_clearN(mi_bitmap_t* bitmap, size_t idx, size_t
// Returns true on success, and in that case sets the index: `0 <= *pidx <= MI_BITMAP_MAX_BITS-n`.
mi_decl_nodiscard bool mi_bitmap_try_find_and_clearN(mi_bitmap_t* bitmap, size_t n, size_t tseq, size_t* pidx);
typedef bool (mi_claim_fun_t)(size_t slice_index, void* arg1, void* arg2, bool* keep_set);
mi_decl_nodiscard bool mi_bitmap_try_find_and_claim(mi_bitmap_t* bitmap, size_t tseq, size_t* pidx,
mi_claim_fun_t* claim, void* arg1, void* arg2);
/* --------------------------------------------------------------------------------
Atomic bitmap for a pair of bits.
The valid pairs are CLEAR (0), SET (3), or BUSY (2).
These bit pairs are used in the abandoned pages maps: when set, the entry has
an available page. When we scan for an available abandoned page and find an entry SET,
we first set it to BUSY, and try to claim the page atomically (since it can race
with a concurrent `mi_free` which also tries to claim the page). However, unlike `mi_free`,
we cannot be sure that a concurrent `mi_free` also didn't free (and decommit) the page
just when we got the entry. Therefore, a page can only be freed after `mi_arena_unabandon`
which (busy) waits until the BUSY flag is cleared to ensure all readers are done.
(and pair-bit operations must therefore be release_acquire).
-------------------------------------------------------------------------------- */
#define MI_PAIR_CLEAR (0)
#define MI_PAIR_UNUSED (1) // should never occur
#define MI_PAIR_BUSY (2)
#define MI_PAIR_SET (3)
// 0b....0101010101010101
#define MI_BFIELD_LO_BIT2 ((MI_BFIELD_LO_BIT8 << 6)|(MI_BFIELD_LO_BIT8 << 4)|(MI_BFIELD_LO_BIT8 << 2)|MI_BFIELD_LO_BIT8)
// A pairmap manipulates pairs of bits (and consists of 2 bitmaps)
typedef struct mi_pairmap_s {
mi_bitmap_t* bitmap1;
mi_bitmap_t* bitmap2;
} mi_pairmap_t;
// initialize a pairmap to all clear; avoid a mem_zero if `already_zero` is true
void mi_pairmap_init(mi_pairmap_t* pairmap, mi_bitmap_t* bm1, mi_bitmap_t* bm2);
bool mi_pairmap_set(mi_pairmap_t* pairmap, size_t pair_idx);
bool mi_pairmap_clear(mi_pairmap_t* pairmap, size_t pair_idx);
bool mi_pairmap_is_clear(mi_pairmap_t* pairmap, size_t pair_idx);
void mi_pairmap_clear_once_not_busy(mi_pairmap_t* pairmap, size_t pair_idx);
typedef bool (mi_bitmap_claim_while_busy_fun_t)(size_t pair_index, void* arg1, void* arg2);
mi_decl_nodiscard bool mi_pairmap_try_find_and_set_busy(mi_pairmap_t* pairmap, size_t tseq, size_t* pidx,
mi_bitmap_claim_while_busy_fun_t* claim, void* arg1 ,void* arg2
);
void mi_bitmap_clear_once_set(mi_bitmap_t* bitmap, size_t idx);
#endif // MI_BITMAP_H