Loading...
--- libmalloc/libmalloc-317.140.5/src/magazine_medium.c
+++ libmalloc/libmalloc-792.41.1/src/magazine_medium.c
@@ -32,15 +32,29 @@
static MALLOC_INLINE uint64_t
medium_sliding_madvise_granularity(magazine_t *magazine)
{
+ uint64_t granularity = MEDIUM_MADVISE_MIN;
// Use a sliding madvise granularity based on how many bytes the region
// currently has allocated. This way we will advise at a finer granularity
// as the region becomes more and more empty.
// region_trailer_t *t = REGION_TRAILER_FOR_MEDIUM_REGION(region);
- if (magazine->mag_num_bytes_in_objects == 0) {
- return MEDIUM_MADVISE_MIN;
- }
- return MAX(MEDIUM_MADVISE_MIN, 1 << (64 -
- __builtin_clzl(magazine->mag_num_bytes_in_objects >> MEDIUM_MADVISE_SHIFT)));
+ if (magazine->mag_num_bytes_in_objects > 0) {
+ if (magazine_medium_madvise_window_scale_factor == 1) {
+ // NB: This code has a bug causing undefined behavior whenever the result of clzl is > 32
+ // because it's shifting a signed integer. It should be 1ULL << (64 - ...) but
+ // fixing this bug will introduce memory regressions, so for now we've only fixed it on very
+ // large memory configs that are scaling up their window anyways.
+ // See rdar://82128925 for details
+ granularity = MAX(granularity, 1 << (64 -
+ __builtin_clzl(magazine->mag_num_bytes_in_objects >> MEDIUM_MADVISE_SHIFT)));
+ } else {
+ granularity = MAX(granularity, 1ULL << (64 -
+ __builtin_clzl(magazine->mag_num_bytes_in_objects >> MEDIUM_MADVISE_SHIFT)));
+ if (os_mul_overflow(granularity, magazine_medium_madvise_window_scale_factor, &granularity)) {
+ return UINT64_MAX;
+ }
+ }
+ }
+ return granularity;
}
static MALLOC_INLINE void
@@ -116,25 +130,6 @@
medium_meta_header_set_middle(msize_t *meta_headers, msize_t index)
{
meta_headers[index] = 0;
-}
-
-static MALLOC_INLINE MALLOC_ALWAYS_INLINE
-mag_index_t
-medium_mag_get_thread_index(void)
-{
-#if CONFIG_MEDIUM_USES_HYPER_SHIFT
- if (os_likely(_os_cpu_number_override == -1)) {
- return _malloc_cpu_number() >> hyper_shift;
- } else {
- return _os_cpu_number_override >> hyper_shift;
- }
-#else // CONFIG_MEDIUM_USES_HYPER_SHIFT
- if (os_likely(_os_cpu_number_override == -1)) {
- return _malloc_cpu_number();
- } else {
- return _os_cpu_number_override;
- }
-#endif // CONFIG_MEDIUM_USES_HYPER_SHIFT
}
#pragma mark in-place free list
@@ -995,11 +990,6 @@
{
magazine_t *depot_ptr = &(rack->magazines[DEPOT_MAGAZINE_INDEX]);
- /* FIXME: Would Uniprocessor benefit from recirc and MADV_FREE? */
- if (rack->num_magazines == 1) { // Uniprocessor, single magazine, so no recirculation necessary
- return 0;
- }
-
#if DEBUG_MALLOC
if (DEPOT_MAGAZINE_INDEX == mag_index) {
malloc_zone_error(rack->debug_flags, true, "medium_get_region_from_depot called for magazine index -1\n", NULL, NULL);
@@ -1012,22 +1002,32 @@
// Appropriate a Depot'd region that can satisfy requested msize.
region_trailer_t *node;
region_t sparse_region;
+ msize_t try_msize = msize;
while (1) {
- sparse_region = medium_find_msize_region(rack, depot_ptr, DEPOT_MAGAZINE_INDEX, msize);
+ sparse_region = medium_find_msize_region(rack, depot_ptr, DEPOT_MAGAZINE_INDEX, try_msize);
if (NULL == sparse_region) { // Depot empty?
SZONE_MAGAZINE_PTR_UNLOCK(depot_ptr);
return 0;
}
node = REGION_TRAILER_FOR_MEDIUM_REGION(sparse_region);
- if (0 >= node->pinned_to_depot) {
+ if (0 == node->pinned_to_depot) {
+ // Found one!
break;
}
- SZONE_MAGAZINE_PTR_UNLOCK(depot_ptr);
- yield();
- SZONE_MAGAZINE_PTR_LOCK(depot_ptr);
+ // Try the next msize up - maybe the head of its free list will be in
+ // a region we can use. Once we get the region we'll still allocate the
+ // original msize.
+ try_msize++;
+
+ if (try_msize > NUM_MEDIUM_SLOTS) {
+ // Tried all the msizes but couldn't get a usable region. Let's
+ // give up for now and we'll allocate a new region from the kernel.
+ SZONE_MAGAZINE_PTR_UNLOCK(depot_ptr);
+ return 0;
+ }
}
// disconnect node from Depot
@@ -1038,7 +1038,7 @@
// Transfer ownership of the region
MAGAZINE_INDEX_FOR_MEDIUM_REGION(sparse_region) = mag_index;
- node->pinned_to_depot = 0;
+ MALLOC_ASSERT(node->pinned_to_depot == 0);
// Iterate the region putting its free entries on its new (locked) magazine's free list
size_t bytes_inplay = medium_free_reattach_region(rack, medium_mag_ptr, sparse_region);
@@ -1129,7 +1129,7 @@
SZONE_MAGAZINE_PTR_LOCK(medium_depot_ptr);
MAGAZINE_INDEX_FOR_MEDIUM_REGION(medium) = DEPOT_MAGAZINE_INDEX;
- REGION_TRAILER_FOR_MEDIUM_REGION(medium)->pinned_to_depot = 0;
+ MALLOC_ASSERT(REGION_TRAILER_FOR_MEDIUM_REGION(medium)->pinned_to_depot == 0);
size_t bytes_inplay = medium_free_reattach_region(rack, medium_depot_ptr, medium);
@@ -1241,7 +1241,17 @@
// If the target region is madvisable, then madvise whatever we can but
// bound it by the safe_start/end pointers to make sure we don't clobber
// the free-list.
- if ((vote_force == 2) || (dirty_msz >= trigger_msize)) {
+ bool should_madvise = (vote_force == 2) || (dirty_msz >= trigger_msize);
+ if (magazine_medium_madvise_window_scale_factor > 1) {
+ // trigger_msize is an unsigned short, but it's possible that trigger level was larger than UINT16_MAX
+ // even after we shifted it. In this case the window rolls around to 0. This is even more likely
+ // if we're scaling the window up. We should fix the truncation bug above, but doing so
+ // will cause a memory regression. For now, we avoid using trigger_msize iff we're scaling
+ // the window up since truncating to 0 would be self-defeating in this case.
+ // See rdar://82128639 for details
+ should_madvise = (vote_force == 2) || (MEDIUM_BYTES_FOR_MSIZE(dirty_msz) >= trigger_level);
+ }
+ if (should_madvise) {
uintptr_t lo = MAX((uintptr_t)MEDIUM_PTR_FOR_META_INDEX(region, range_idx),
safe_start_ptr);
uintptr_t hi = MIN((uintptr_t)MEDIUM_PTR_FOR_META_INDEX(region, range_idx) +
@@ -1480,11 +1490,7 @@
region_trailer_t *node = REGION_TRAILER_FOR_MEDIUM_REGION(region);
size_t bytes_used = node->bytes_used;
- /* FIXME: Would Uniprocessor benefit from recirc and MADV_FREE? */
- if (rack->num_magazines == 1) { // Uniprocessor, single magazine, so no recirculation necessary
- /* NOTHING */
- return TRUE; // Caller must do SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr)
- } else if (DEPOT_MAGAZINE_INDEX != mag_index) {
+ if (DEPOT_MAGAZINE_INDEX != mag_index) {
// Emptiness discriminant
if (bytes_used < DENSITY_THRESHOLD(MEDIUM_REGION_PAYLOAD_BYTES)) {
/* Region has crossed threshold from density to sparsity. Mark it "suitable" on the
@@ -1870,7 +1876,8 @@
// use is adjusted by the medium_meta_header_set_middle() call below.
medium_meta_header_set_in_use(meta_headers, index + new_msize, leftover_msize);
- if (madv_headers[index] & MEDIUM_IS_ADVISED) {
+ /* Propagate the madvise information from the block we're using to the leftover block. */
+ if (madv_headers[next_index] & MEDIUM_IS_ADVISED) {
medium_madvise_header_mark_clean(madv_headers, index + new_msize, leftover_msize);
} else {
medium_madvise_header_mark_dirty(madv_headers, index + new_msize, leftover_msize);
@@ -2271,7 +2278,7 @@
if ((ptr = medium_free_list_get_ptr(rack, *free_list))) {
this_msize = MEDIUM_PTR_SIZE(ptr);
was_madvised = (medium_madvise_header_dirty_len(
- MEDIUM_MADVISE_HEADER_FOR_PTR(ptr), this_msize) == 0);
+ MEDIUM_MADVISE_HEADER_FOR_PTR(ptr), MEDIUM_META_INDEX_FOR_PTR(ptr)) == 0);
medium_free_list_remove_ptr(rack, medium_mag_ptr, *free_list, this_msize);
goto add_leftover_and_proceed;
}
@@ -2377,7 +2384,7 @@
medium_malloc_should_clear(rack_t *rack, msize_t msize, boolean_t cleared_requested)
{
void *ptr;
- mag_index_t mag_index = medium_mag_get_thread_index() % rack->num_magazines;
+ mag_index_t mag_index = rack_get_thread_index(rack) % rack->num_magazines;
magazine_t *medium_mag_ptr = &(rack->magazines[mag_index]);
MALLOC_TRACE(TRACE_medium_malloc, (uintptr_t)rack, MEDIUM_BYTES_FOR_MSIZE(msize), (uintptr_t)medium_mag_ptr, cleared_requested);
@@ -2426,16 +2433,17 @@
// The magazine is exhausted. A new region (heap) must be allocated to satisfy this call to malloc().
// The allocation, an mmap() system call, will be performed outside the magazine spin locks by the first
- // thread that suffers the exhaustion. That thread sets "alloc_underway" and enters a critical section.
- // Threads arriving here later are excluded from the critical section, yield the CPU, and then retry the
- // allocation. After some time the magazine is resupplied, the original thread leaves with its allocation,
- // and retry-ing threads succeed in the code just above.
- if (!medium_mag_ptr->alloc_underway) {
+ // thread that suffers the exhaustion. That thread accquires the magazine_alloc_lock, then drops the
+ // magazine lock to allow freeing threads to proceed. Allocating thrads that arrive later are excluded
+ // from the critial section by the alloc lock. When those are unblocked, they succeed in the code above.
+ //
+ // Note that we need to trylock the alloc lock to avoid a deadlock, since we can't block on the alloc
+ // lock while holding the magazine lock
+ if (os_likely(_malloc_lock_trylock(&medium_mag_ptr->magazine_alloc_lock))) {
+ // We got the alloc lock, so we are the thread that should allocate a new region
void *fresh_region;
// time to create a new region (do this outside the magazine lock)
- medium_mag_ptr->alloc_underway = TRUE;
- OSMemoryBarrier();
SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
fresh_region = mvm_allocate_pages(MEDIUM_REGION_SIZE,
MEDIUM_BLOCKS_ALIGN,
@@ -2448,9 +2456,8 @@
fresh_region, MEDIUM_REGION_SIZE);
if (!fresh_region) { // out of memory!
- medium_mag_ptr->alloc_underway = FALSE;
- OSMemoryBarrier();
SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
+ _malloc_lock_unlock(&medium_mag_ptr->magazine_alloc_lock);
return NULL;
}
@@ -2459,14 +2466,20 @@
mag_index, msize, fresh_region);
// we don't clear because this freshly allocated space is pristine
- medium_mag_ptr->alloc_underway = FALSE;
- OSMemoryBarrier();
SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
+ _malloc_lock_unlock(&medium_mag_ptr->magazine_alloc_lock);
CHECK(szone, __PRETTY_FUNCTION__);
return ptr;
} else {
+ // We failed to get the alloc lock, so someone else is allocating.
+ // Drop the magazine lock...
SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
- yield();
+
+ // Wait for the other thread on the alloc lock
+ _malloc_lock_lock(&medium_mag_ptr->magazine_alloc_lock);
+ _malloc_lock_unlock(&medium_mag_ptr->magazine_alloc_lock);
+
+ // Reacquire the magazine lock to go around the loop again
SZONE_MAGAZINE_PTR_LOCK(medium_mag_ptr);
}
}