remove threadid from pages and keep page flags separate

This commit is contained in:
daan 2019-08-23 14:08:00 -07:00
parent b511309709
commit acde83543f
7 changed files with 37 additions and 58 deletions

View File

@ -317,39 +317,24 @@ static inline mi_page_queue_t* mi_page_queue(const mi_heap_t* heap, size_t size)
} }
//----------------------------------------------------------- //-----------------------------------------------------------
// Page flags // Page flags
//----------------------------------------------------------- //-----------------------------------------------------------
static inline uintptr_t mi_page_thread_id(const mi_page_t* page) {
return (page->flags & ~MI_PAGE_FLAGS_MASK);
}
static inline void mi_page_init_flags(mi_page_t* page, uintptr_t thread_id) {
mi_assert_internal((thread_id & MI_PAGE_FLAGS_MASK) == 0);
page->flags = thread_id;
}
static inline void mi_page_set_thread_id(mi_page_t* page, uintptr_t thread_id) {
mi_assert_internal((thread_id & MI_PAGE_FLAGS_MASK) == 0);
page->flags = thread_id | (page->flags & MI_PAGE_FLAGS_MASK);
}
static inline bool mi_page_is_in_full(const mi_page_t* page) { static inline bool mi_page_is_in_full(const mi_page_t* page) {
return ((page->flags & 0x01) != 0); return page->flags.in_full;
} }
static inline void mi_page_set_in_full(mi_page_t* page, bool in_full) { static inline void mi_page_set_in_full(mi_page_t* page, bool in_full) {
if (in_full) page->flags |= 0x01; page->flags.in_full = in_full;
else page->flags &= ~0x01;
} }
static inline bool mi_page_has_aligned(const mi_page_t* page) { static inline bool mi_page_has_aligned(const mi_page_t* page) {
return ((page->flags & 0x02) != 0); return page->flags.has_aligned;
} }
static inline void mi_page_set_has_aligned(mi_page_t* page, bool has_aligned) { static inline void mi_page_set_has_aligned(mi_page_t* page, bool has_aligned) {
if (has_aligned) page->flags |= 0x02; page->flags.has_aligned = has_aligned;
else page->flags &= ~0x02;
} }

View File

@ -124,12 +124,15 @@ typedef enum mi_delayed_e {
} mi_delayed_t; } mi_delayed_t;
// Use the bottom 2 bits for the `in_full` and `has_aligned` flags // The `in_full` and `has_aligned` page flags are put in a union to efficiently
// and the rest for the threadid (we assume tid's never use those lower 2 bits). // test if both are false (`value == 0`) in the `mi_free` routine.
// This allows a single test in `mi_free` to check for unlikely cases typedef union mi_page_flags_u {
// (namely, non-local free, aligned free, or freeing in a full page) uint16_t value;
#define MI_PAGE_FLAGS_MASK ((uintptr_t)0x03) struct {
typedef uintptr_t mi_page_flags_t; bool in_full;
bool has_aligned;
};
} mi_page_flags_t;
// Thread free list. // Thread free list.
// We use the bottom 2 bits of the pointer for mi_delayed_t flags // We use the bottom 2 bits of the pointer for mi_delayed_t flags
@ -163,12 +166,12 @@ typedef struct mi_page_s {
// layout like this to optimize access in `mi_malloc` and `mi_free` // layout like this to optimize access in `mi_malloc` and `mi_free`
uint16_t capacity; // number of blocks committed uint16_t capacity; // number of blocks committed
uint16_t reserved; // number of blocks reserved in memory uint16_t reserved; // number of blocks reserved in memory
// 16 bits padding mi_page_flags_t flags; // `in_full` and `has_aligned` flags (16 bits)
mi_block_t* free; // list of available free blocks (`malloc` allocates from this list) mi_block_t* free; // list of available free blocks (`malloc` allocates from this list)
#if MI_SECURE #if MI_SECURE
uintptr_t cookie; // random cookie to encode the free lists uintptr_t cookie; // random cookie to encode the free lists
#endif #endif
mi_page_flags_t flags; // threadid:62 | has_aligned:1 | in_full:1
size_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`) size_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`)
mi_block_t* local_free; // list of deferred free blocks by this thread (migrates to `free`) mi_block_t* local_free; // list of deferred free blocks by this thread (migrates to `free`)
@ -181,12 +184,11 @@ typedef struct mi_page_s {
struct mi_page_s* next; // next page owned by this thread with the same `block_size` struct mi_page_s* next; // next page owned by this thread with the same `block_size`
struct mi_page_s* prev; // previous page owned by this thread with the same `block_size` struct mi_page_s* prev; // previous page owned by this thread with the same `block_size`
// improve page index calculation // improve page index calculation
#if (MI_INTPTR_SIZE==8 && MI_SECURE==0) // without padding: 10 words on 64-bit, 11 on 32-bit. Secure adds one word
void* padding[1]; // 12 words on 64-bit #if (MI_INTPTR_SIZE==8 && MI_SECURE>0) || (MI_INTPTR_SIZE==4 && MI_SECURE==0)
#elif MI_INTPTR_SIZE==4 void* padding[1]; // 12 words on 64-bit in secure mode, 12 words on 32-bit plain
// void* padding[1]; // 12 words on 32-bit #endif
#endif
} mi_page_t; } mi_page_t;

View File

@ -225,19 +225,19 @@ void mi_free(void* p) mi_attr_noexcept
} }
#endif #endif
const uintptr_t tid = _mi_thread_id();
mi_page_t* const page = _mi_segment_page_of(segment, p); mi_page_t* const page = _mi_segment_page_of(segment, p);
#if (MI_STAT>1) #if (MI_STAT>1)
mi_heap_t* heap = mi_heap_get_default(); mi_heap_t* heap = mi_heap_get_default();
mi_heap_stat_decrease( heap, malloc, mi_usable_size(p)); mi_heap_stat_decrease(heap, malloc, mi_usable_size(p));
if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) {
mi_heap_stat_decrease( heap, normal[_mi_bin(page->block_size)], 1); mi_heap_stat_decrease(heap, normal[_mi_bin(page->block_size)], 1);
} }
// huge page stat is accounted for in `_mi_page_retire` // huge page stat is accounted for in `_mi_page_retire`
#endif #endif
const uintptr_t tid = _mi_thread_id(); if (mi_likely(tid == segment->thread_id && page->flags.value == 0)) { // the thread id matches and it is not a full page, nor has aligned blocks
if (mi_likely(tid == page->flags)) { // if equal, the thread id matches and it is not a full page, nor has aligned blocks
// local, and not full or aligned // local, and not full or aligned
mi_block_t* block = (mi_block_t*)p; mi_block_t* block = (mi_block_t*)p;
mi_block_set_next(page, block, page->local_free); mi_block_set_next(page, block, page->local_free);
@ -247,7 +247,7 @@ void mi_free(void* p) mi_attr_noexcept
} }
else { else {
// non-local, aligned blocks, or a full page; use the more generic path // non-local, aligned blocks, or a full page; use the more generic path
mi_free_generic(segment, page, tid == mi_page_thread_id(page), p); mi_free_generic(segment, page, tid == segment->thread_id, p);
} }
} }

View File

@ -13,15 +13,16 @@ terms of the MIT license. A copy of the license can be found in the file
// Empty page used to initialize the small free pages array // Empty page used to initialize the small free pages array
const mi_page_t _mi_page_empty = { const mi_page_t _mi_page_empty = {
0, false, false, false, 0, 0, 0, false, false, false, 0, 0,
{ 0 },
NULL, // free NULL, // free
#if MI_SECURE #if MI_SECURE
0, 0,
#endif #endif
0, 0, // flags, used 0, // used
NULL, 0, 0, NULL, 0, 0,
0, NULL, NULL, NULL 0, NULL, NULL, NULL
#if (MI_INTPTR_SIZE==8 && MI_SECURE==0) #if (MI_INTPTR_SIZE==8 && MI_SECURE>0) || (MI_INTPTR_SIZE==4 && MI_SECURE==0)
, { NULL } , { NULL } // padding
#endif #endif
}; };
@ -350,7 +351,7 @@ void mi_thread_init(void) mi_attr_noexcept
pthread_setspecific(mi_pthread_key, (void*)(_mi_thread_id()|1)); // set to a dummy value so that `mi_pthread_done` is called pthread_setspecific(mi_pthread_key, (void*)(_mi_thread_id()|1)); // set to a dummy value so that `mi_pthread_done` is called
#endif #endif
#if (MI_DEBUG>0) // not in release mode as that leads to crashes on Windows dynamic override #if (MI_DEBUG>0) && !defined(NDEBUG) // not in release mode as that leads to crashes on Windows dynamic override
_mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id()); _mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id());
#endif #endif
} }

View File

@ -217,7 +217,7 @@ static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment
} }
else { else {
// else fall back to regular large OS pages // else fall back to regular large OS pages
_mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) page instead (error %lx)\n", err); _mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) pages instead (error %lx)\n", err);
} }
} }
#endif #endif

View File

@ -75,7 +75,6 @@ static bool mi_page_is_valid_init(mi_page_t* page) {
mi_segment_t* segment = _mi_page_segment(page); mi_segment_t* segment = _mi_page_segment(page);
uint8_t* start = _mi_page_start(segment,page,NULL); uint8_t* start = _mi_page_start(segment,page,NULL);
mi_assert_internal(start == _mi_segment_page_start(segment,page,page->block_size,NULL)); mi_assert_internal(start == _mi_segment_page_start(segment,page,page->block_size,NULL));
mi_assert_internal(segment->thread_id==0 || segment->thread_id == mi_page_thread_id(page));
//mi_assert_internal(start + page->capacity*page->block_size == page->top); //mi_assert_internal(start + page->capacity*page->block_size == page->top);
mi_assert_internal(mi_page_list_is_valid(page,page->free)); mi_assert_internal(mi_page_list_is_valid(page,page->free));

View File

@ -549,14 +549,11 @@ static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) {
mi_assert_internal(segment->used > 0); mi_assert_internal(segment->used > 0);
mi_assert_internal(segment->abandoned_next == NULL); mi_assert_internal(segment->abandoned_next == NULL);
mi_assert_expensive(mi_segment_is_valid(segment)); mi_assert_expensive(mi_segment_is_valid(segment));
#if MI_DEBUG>1
for (size_t i = 0; i < segment->capacity; i++) {
mi_assert_internal(!segment->pages[i].segment_in_use || mi_page_thread_id(&segment->pages[i]) == 0);
}
#endif
// remove the segment from the free page queue if needed // remove the segment from the free page queue if needed
mi_segment_remove_from_free_queue(segment,tld); mi_segment_remove_from_free_queue(segment,tld);
mi_assert_internal(segment->next == NULL && segment->prev == NULL); mi_assert_internal(segment->next == NULL && segment->prev == NULL);
// all pages in the segment are abandoned; add it to the abandoned list // all pages in the segment are abandoned; add it to the abandoned list
_mi_stat_increase(&tld->stats->segments_abandoned, 1); _mi_stat_increase(&tld->stats->segments_abandoned, 1);
mi_segments_track_size(-((long)segment->segment_size), tld); mi_segments_track_size(-((long)segment->segment_size), tld);
@ -570,11 +567,10 @@ static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) {
} }
void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld) { void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld) {
mi_assert(page != NULL && mi_page_thread_id(page) != 0); mi_assert(page != NULL);
mi_segment_t* segment = _mi_page_segment(page); mi_segment_t* segment = _mi_page_segment(page);
mi_assert_expensive(mi_segment_is_valid(segment)); mi_assert_expensive(mi_segment_is_valid(segment));
segment->abandoned++; segment->abandoned++;
mi_page_set_thread_id(page, 0);
_mi_stat_increase(&tld->stats->pages_abandoned, 1); _mi_stat_increase(&tld->stats->pages_abandoned, 1);
mi_assert_internal(segment->abandoned <= segment->used); mi_assert_internal(segment->abandoned <= segment->used);
if (segment->used == segment->abandoned) { if (segment->used == segment->abandoned) {
@ -626,7 +622,6 @@ bool _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segmen
} }
else { else {
// otherwise reclaim it // otherwise reclaim it
mi_page_set_thread_id(page,segment->thread_id);
_mi_page_reclaim(heap,page); _mi_page_reclaim(heap,page);
} }
} }
@ -656,8 +651,7 @@ bool _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segmen
static mi_page_t* mi_segment_page_alloc_in(mi_segment_t* segment, mi_segments_tld_t* tld) { static mi_page_t* mi_segment_page_alloc_in(mi_segment_t* segment, mi_segments_tld_t* tld) {
mi_assert_internal(mi_segment_has_free(segment)); mi_assert_internal(mi_segment_has_free(segment));
mi_page_t* page = mi_segment_find_free(segment, tld->stats); mi_page_t* page = mi_segment_find_free(segment, tld->stats);
page->segment_in_use = true; page->segment_in_use = true;
mi_page_init_flags(page,segment->thread_id);
segment->used++; segment->used++;
mi_assert_internal(segment->used <= segment->capacity); mi_assert_internal(segment->used <= segment->capacity);
if (segment->used == segment->capacity) { if (segment->used == segment->capacity) {
@ -697,7 +691,6 @@ static mi_page_t* mi_segment_large_page_alloc(mi_segments_tld_t* tld, mi_os_tld_
segment->used = 1; segment->used = 1;
mi_page_t* page = &segment->pages[0]; mi_page_t* page = &segment->pages[0];
page->segment_in_use = true; page->segment_in_use = true;
mi_page_init_flags(page,segment->thread_id);
return page; return page;
} }
@ -709,7 +702,6 @@ static mi_page_t* mi_segment_huge_page_alloc(size_t size, mi_segments_tld_t* tld
segment->used = 1; segment->used = 1;
mi_page_t* page = &segment->pages[0]; mi_page_t* page = &segment->pages[0];
page->segment_in_use = true; page->segment_in_use = true;
mi_page_init_flags(page,segment->thread_id);
return page; return page;
} }