gh-112532: Tag mimalloc heaps and pages (#113742)

* gh-112532: Tag mimalloc heaps and pages

Mimalloc pages are data structures that contain contiguous allocations
of the same block size. Note that they are distinct from operating
system pages. Mimalloc pages are contained in segments.

When a thread exits, it abandons any segments and contained pages that
have live allocations. These segments and pages may be later reclaimed
by another thread. To support GC and certain thread-safety guarantees in
free-threaded builds, we want pages to only be reclaimed by the
corresponding heap in the claimant thread. For example, we want pages
containing GC objects to only be claimed by GC heaps.

This allows heaps and pages to be tagged with an integer tag that is
used to ensure that abandoned pages are only claimed by heaps with the
same tag. Heaps can be initialized with a tag (0-15); any page allocated
by that heap copies the corresponding tag.

* Fix conversion warning
This commit is contained in:
Sam Gross 2024-01-05 15:08:50 -05:00 committed by GitHub
parent eb53750757
commit 0b7476080b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
7 changed files with 35 additions and 16 deletions

View File

@ -155,7 +155,7 @@ size_t _mi_bin_size(uint8_t bin); // for stats
uint8_t _mi_bin(size_t size); // for stats
// "heap.c"
void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id);
void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id, bool no_reclaim, uint8_t tag);
void _mi_heap_destroy_pages(mi_heap_t* heap);
void _mi_heap_collect_abandon(mi_heap_t* heap);
void _mi_heap_set_default_direct(mi_heap_t* heap);

View File

@ -311,6 +311,7 @@ typedef struct mi_page_s {
uint32_t slice_offset; // distance from the actual page data slice (0 if a 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 tag : 4; // tag from the owning heap
// 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`
@ -551,6 +552,7 @@ struct mi_heap_s {
size_t page_retired_max; // largest retired index into the `pages` array.
mi_heap_t* next; // list of heaps per thread
bool no_reclaim; // `true` if this heap should not reclaim abandoned pages
uint8_t tag; // custom identifier for this heap
};

View File

@ -209,7 +209,7 @@ mi_heap_t* mi_heap_get_backing(void) {
return bheap;
}
void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id)
void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id, bool no_reclaim, uint8_t tag)
{
_mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(mi_heap_t));
heap->tld = tld;
@ -224,17 +224,19 @@ void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id)
heap->cookie = _mi_heap_random_next(heap) | 1;
heap->keys[0] = _mi_heap_random_next(heap);
heap->keys[1] = _mi_heap_random_next(heap);
heap->no_reclaim = no_reclaim;
heap->tag = tag;
// push on the thread local heaps list
heap->next = heap->tld->heaps;
heap->tld->heaps = heap;
}
mi_decl_nodiscard mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id) {
mi_heap_t* bheap = mi_heap_get_backing();
mi_heap_t* heap = mi_heap_malloc_tp(bheap, mi_heap_t); // todo: OS allocate in secure mode?
if (heap == NULL) return NULL;
_mi_heap_init_ex(heap, bheap->tld, arena_id);
heap->no_reclaim = true; // don't reclaim abandoned pages or otherwise destroy is unsafe
// push on the thread local heaps list
heap->next = heap->tld->heaps;
heap->tld->heaps = heap;
// don't reclaim abandoned pages or otherwise destroy is unsafe
_mi_heap_init_ex(heap, bheap->tld, arena_id, true, 0);
return heap;
}

View File

@ -14,7 +14,7 @@ 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
const mi_page_t _mi_page_empty = {
0, false, false, false,
0, false, false, false, 0,
0, // capacity
0, // reserved capacity
{ 0 }, // flags
@ -121,7 +121,8 @@ mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
0, // page count
MI_BIN_FULL, 0, // page retired min/max
NULL, // next
false
false,
0
};
#define tld_empty_stats ((mi_stats_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,stats)))
@ -298,7 +299,7 @@ static bool _mi_heap_init(void) {
if (td == NULL) return false;
_mi_tld_init(&td->tld, &td->heap);
_mi_heap_init_ex(&td->heap, &td->tld, _mi_arena_id_none());
_mi_heap_init_ex(&td->heap, &td->tld, _mi_arena_id_none(), false, 0);
_mi_heap_set_default_direct(&td->heap);
}
return false;
@ -311,7 +312,6 @@ void _mi_tld_init(mi_tld_t* tld, mi_heap_t* bheap) {
tld->segments.abandoned = &_mi_abandoned_default;
tld->os.stats = &tld->stats;
tld->heap_backing = bheap;
tld->heaps = bheap;
}
// Free the thread local default heap (called from `mi_thread_done`)

View File

@ -660,6 +660,7 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
mi_assert_internal(block_size > 0);
// set fields
mi_page_set_heap(page, heap);
page->tag = heap->tag;
page->xblock_size = (block_size < MI_HUGE_BLOCK_SIZE ? (uint32_t)block_size : MI_HUGE_BLOCK_SIZE); // initialize before _mi_segment_page_start
size_t page_size;
const void* page_start = _mi_segment_page_start(segment, page, &page_size);

View File

@ -1299,6 +1299,18 @@ static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, s
return has_page;
}
static mi_heap_t* mi_heap_by_tag(mi_heap_t* heap, uint8_t tag) {
if (heap->tag == tag) {
return heap;
}
for (mi_heap_t *curr = heap->tld->heaps; curr != NULL; curr = curr->next) {
if (curr->tag == tag) {
return curr;
}
}
return NULL;
}
// Reclaim an abandoned segment; returns NULL if the segment was freed
// set `right_page_reclaimed` to `true` if it reclaimed a page of the right `block_size` that was not full.
static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap, size_t requested_block_size, bool* right_page_reclaimed, mi_segments_tld_t* tld) {
@ -1321,6 +1333,7 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
if (mi_slice_is_used(slice)) {
// in use: reclaim the page in our heap
mi_page_t* page = mi_slice_to_page(slice);
mi_heap_t* target_heap = mi_heap_by_tag(heap, page->tag);
mi_assert_internal(page->is_committed);
mi_assert_internal(mi_page_thread_free_flag(page)==MI_NEVER_DELAYED_FREE);
mi_assert_internal(mi_page_heap(page) == NULL);
@ -1328,7 +1341,7 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
_mi_stat_decrease(&tld->stats->pages_abandoned, 1);
segment->abandoned--;
// set the heap again and allow delayed free again
mi_page_set_heap(page, heap);
mi_page_set_heap(page, target_heap);
_mi_page_use_delayed_free(page, MI_USE_DELAYED_FREE, true); // override never (after heap is set)
_mi_page_free_collect(page, false); // ensure used count is up to date
if (mi_page_all_free(page)) {
@ -1337,8 +1350,9 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
}
else {
// otherwise reclaim it into the heap
_mi_page_reclaim(heap, page);
if (requested_block_size == page->xblock_size && mi_page_has_any_available(page)) {
_mi_page_reclaim(target_heap, page);
if (requested_block_size == page->xblock_size && mi_page_has_any_available(page) &&
heap == target_heap) {
if (right_page_reclaimed != NULL) { *right_page_reclaimed = true; }
}
}

View File

@ -2539,8 +2539,8 @@ tstate_mimalloc_bind(PyThreadState *tstate)
tld->segments.abandoned = &tstate->interp->mimalloc.abandoned_pool;
// Initialize each heap
for (Py_ssize_t i = 0; i < _Py_MIMALLOC_HEAP_COUNT; i++) {
_mi_heap_init_ex(&mts->heaps[i], tld, _mi_arena_id_none());
for (uint8_t i = 0; i < _Py_MIMALLOC_HEAP_COUNT; i++) {
_mi_heap_init_ex(&mts->heaps[i], tld, _mi_arena_id_none(), false, i);
}
// By default, object allocations use _Py_MIMALLOC_HEAP_OBJECT.