Loading...
--- libmalloc/libmalloc-283/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
@@ -118,25 +132,6 @@
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 _os_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 _os_cpu_number();
- } else {
- return _os_cpu_number_override;
- }
-#endif // CONFIG_MEDIUM_USES_HYPER_SHIFT
-}
-
#pragma mark in-place free list
static MALLOC_INLINE void
@@ -312,7 +307,7 @@
static MALLOC_INLINE void
medium_oob_free_entry_set_ptr(oob_free_entry_t oobe, void *ptr)
{
- oobe->ptr = MEDIUM_IS_OOB | (MEDIUM_OFFSET_FOR_PTR(ptr) >> SHIFT_MEDIUM_QUANTUM);
+ oobe->ptr = MEDIUM_IS_OOB | (MEDIUM_REGION_OFFSET_FOR_PTR(ptr) >> SHIFT_MEDIUM_QUANTUM);
}
static MALLOC_INLINE void
@@ -712,7 +707,7 @@
// than performing this workaround.
//
if (medium_mag_ptr->mag_bytes_free_at_end) {
- last_block = MEDIUM_REGION_END(medium_mag_ptr->mag_last_region) - medium_mag_ptr->mag_bytes_free_at_end;
+ last_block = MEDIUM_REGION_HEAP_END(medium_mag_ptr->mag_last_region) - medium_mag_ptr->mag_bytes_free_at_end;
last_msize = MEDIUM_MSIZE_FOR_BYTES(medium_mag_ptr->mag_bytes_free_at_end);
last_index = MEDIUM_META_INDEX_FOR_PTR(last_block);
@@ -762,12 +757,11 @@
int
medium_free_detach_region(rack_t *rack, magazine_t *medium_mag_ptr, region_t r)
{
- unsigned char *ptr = MEDIUM_REGION_ADDRESS(r);
- msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(ptr);
- uintptr_t start = (uintptr_t)MEDIUM_REGION_ADDRESS(r);
+ uintptr_t start = (uintptr_t)MEDIUM_REGION_HEAP_BASE(r);
uintptr_t current = start;
- uintptr_t limit = (uintptr_t)MEDIUM_REGION_END(r);
+ uintptr_t limit = (uintptr_t)MEDIUM_REGION_HEAP_END(r);
int total_alloc = 0;
+ msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(start);
while (current < limit) {
unsigned index = MEDIUM_META_INDEX_FOR_PTR(current);
@@ -797,12 +791,11 @@
size_t
medium_free_reattach_region(rack_t *rack, magazine_t *medium_mag_ptr, region_t r)
{
- unsigned char *ptr = MEDIUM_REGION_ADDRESS(r);
- msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(ptr);
- uintptr_t start = (uintptr_t)MEDIUM_REGION_ADDRESS(r);
+ uintptr_t start = (uintptr_t)MEDIUM_REGION_HEAP_BASE(r);
uintptr_t current = start;
- uintptr_t limit = (uintptr_t)MEDIUM_REGION_END(r);
+ uintptr_t limit = (uintptr_t)MEDIUM_REGION_HEAP_END(r);
size_t total_alloc = 0;
+ msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(start);
while (current < limit) {
unsigned index = MEDIUM_META_INDEX_FOR_PTR(current);
@@ -834,9 +827,9 @@
void
medium_free_scan_madvise_free(rack_t *rack, magazine_t *depot_ptr, region_t r)
{
- uintptr_t start = (uintptr_t)MEDIUM_REGION_ADDRESS(r);
+ uintptr_t start = (uintptr_t)MEDIUM_REGION_HEAP_BASE(r);
uintptr_t current = start;
- uintptr_t limit = (uintptr_t)MEDIUM_REGION_END(r);
+ uintptr_t limit = (uintptr_t)MEDIUM_REGION_HEAP_END(r);
msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(start);
msize_t *madv_headers = MEDIUM_MADVISE_HEADER_FOR_PTR(start);
@@ -868,6 +861,7 @@
medium_advisory_t mat = (medium_advisory_t)pgLo;
mat->next = advisories;
mat->size = pgHi - pgLo;
+ advisories = mat;
}
break;
}
@@ -895,6 +889,7 @@
medium_advisory_t mat = (medium_advisory_t)pgLo;
mat->next = advisories;
mat->size = pgHi - pgLo;
+ advisories = mat;
}
memset(&madv_headers[index], 0, sizeof(uint16_t) * alloc_msize);
@@ -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);
@@ -1072,19 +1072,31 @@
for (mag_index = 0; mag_index < rack->num_magazines; mag_index++) {
size_t index;
for (index = 0; index < rack->region_generation->num_regions_allocated; ++index) {
- SZONE_LOCK(MEDIUM_SZONE_FROM_RACK(rack));
+ rack_region_lock(rack);
region_t medium = rack->region_generation->hashed_regions[index];
if (!medium || medium == HASHRING_REGION_DEALLOCATED) {
- SZONE_UNLOCK(MEDIUM_SZONE_FROM_RACK(rack));
+ rack_region_unlock(rack);
continue;
}
+ region_trailer_t *trailer =
+ REGION_TRAILER_FOR_MEDIUM_REGION(medium);
+ // Make sure that the owning magazine doesn't try and take this out
+ // from under our feet.
+ trailer->dispose_flags |= RACK_DISPOSE_DELAY;
+ rack_region_unlock(rack);
+
magazine_t *mag_ptr = mag_lock_zine_for_region_trailer(rack->magazines,
- REGION_TRAILER_FOR_MEDIUM_REGION(medium),
- MAGAZINE_INDEX_FOR_MEDIUM_REGION(medium));
-
- SZONE_UNLOCK(MEDIUM_SZONE_FROM_RACK(rack));
+ trailer, MAGAZINE_INDEX_FOR_MEDIUM_REGION(medium));
+
+ // If acquiring the region lock was enough to prevent the owning
+ // magazine from deallocating the region, free it now so we don't
+ // do wasted work.
+ if (rack_region_maybe_dispose(rack, medium, MEDIUM_REGION_SIZE, trailer)) {
+ SZONE_MAGAZINE_PTR_UNLOCK(mag_ptr);
+ continue;
+ }
/* Ordering is important here, the magazine of a region may potentially change
* during mag_lock_zine_for_region_trailer, so src_mag_index must be taken
@@ -1117,7 +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);
@@ -1158,7 +1170,6 @@
{
region_trailer_t *node = REGION_TRAILER_FOR_MEDIUM_REGION(region);
msize_t *madvh = MEDIUM_MADVISE_HEADER_FOR_PTR(ptr);
-
msize_t trigger_msize = trigger_level >> SHIFT_MEDIUM_QUANTUM;
size_t free_header_size = sizeof(medium_inplace_free_entry_s) + sizeof(msize_t);
@@ -1207,7 +1218,7 @@
}
msize_t right_dirty_msz = 0;
- if (right_end_idx < src_end_idx) {
+ if (right_end_idx > src_end_idx) {
// Same as above, if we had trailing data coalesced with this entry
// and that was not madvised, consider it, too.
right_dirty_msz = medium_madvise_header_dirty_len(madvh, right_start_idx);
@@ -1220,7 +1231,7 @@
medium_madvise_header_mark_middle(madvh, right_end_idx);
}
- // We absolutely can't madvise lower the the free-list entry pointer plus
+ // We absolutely can't madvise lower than the free-list entry pointer plus
// the header size. When the entry is OOB, there's no header or footer to
// store in memory.
uintptr_t safe_start_ptr = round_page_kernel(rangep + free_header_size);
@@ -1230,21 +1241,33 @@
// 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)) {
- uintptr_t lo = MAX(MEDIUM_PTR_FOR_META_INDEX(region, range_idx),
+ 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(MEDIUM_PTR_FOR_META_INDEX(region, range_idx) +
+ uintptr_t hi = MIN((uintptr_t)MEDIUM_PTR_FOR_META_INDEX(region, range_idx) +
MEDIUM_BYTES_FOR_MSIZE(range_msz), safe_end_ptr);
// The page that contains the freelist entry needs to be marked as not
- // having been madvised.
+ // having been madvised. Note that the quantum is larger than the kernel page size
+ // so if safe_start_ptr and rangep are on different pages, we just mark
+ // the whole block as clean.
if (range_idx < MEDIUM_META_INDEX_FOR_PTR(safe_start_ptr)) {
medium_madvise_header_mark_dirty(madvh, range_idx,
MEDIUM_META_INDEX_FOR_PTR(safe_start_ptr) - range_idx);
}
if (range_idx + range_msz > MEDIUM_META_INDEX_FOR_PTR(safe_end_ptr)) {
medium_madvise_header_mark_dirty(madvh,
- MEDIUM_META_INDEX_FOR_PTR(safe_end_ptr) + 1, range_idx +
+ MEDIUM_META_INDEX_FOR_PTR(safe_end_ptr), range_idx +
range_msz - MEDIUM_META_INDEX_FOR_PTR(safe_end_ptr));
}
@@ -1269,10 +1292,12 @@
// We chose not to madvise, we need to re-mark the region as dirty
// for when we come back to it later.
if (left_dirty_msz < left_msz) {
+ /* The preceding block was clean. */
medium_madvise_header_mark_clean(madvh, range_idx,
left_msz - left_dirty_msz);
}
if (right_dirty_msz < right_msz) {
+ /* The trailing block was clean. */
medium_madvise_header_mark_clean(madvh, right_start_idx +
right_dirty_msz, right_msz - right_dirty_msz);
}
@@ -1359,24 +1384,10 @@
int objects_in_use = medium_free_detach_region(rack, depot_ptr, sparse_region);
if (0 == objects_in_use) {
- // Invalidate the hash table entry for this region with HASHRING_REGION_DEALLOCATED.
- // Using HASHRING_REGION_DEALLOCATED preserves the collision chain, using HASHRING_OPEN_ENTRY (0) would not.
- rgnhdl_t pSlot = hash_lookup_region_no_lock(rack->region_generation->hashed_regions,
- rack->region_generation->num_regions_allocated,
- rack->region_generation->num_regions_allocated_shift,
- sparse_region);
- if (NULL == pSlot) {
- malloc_zone_error(rack->debug_flags, true, "medium_free_try_depot_unmap_no_lock hash lookup failed: %p\n", sparse_region);
+ if (!rack_region_remove(rack, sparse_region, node)) {
return NULL;
}
- *pSlot = HASHRING_REGION_DEALLOCATED;
depot_ptr->num_bytes_in_magazine -= MEDIUM_REGION_PAYLOAD_BYTES;
- // Atomically increment num_regions_dealloc
-#ifdef __LP64___
- OSAtomicIncrement64(&rack->num_regions_dealloc);
-#else
- OSAtomicIncrement32((int32_t *)&rack->num_regions_dealloc);
-#endif
// Caller will transfer ownership of the region back to the OS with no locks held
MAGMALLOC_DEALLOCREGION(MEDIUM_SZONE_FROM_RACK(rack), (void *)sparse_region, (int)MEDIUM_REGION_SIZE); // DTrace USDT Probe
@@ -1448,16 +1459,20 @@
MAGMALLOC_RECIRCREGION(MEDIUM_SZONE_FROM_RACK(rack), (int)mag_index, (void *)sparse_region, MEDIUM_REGION_SIZE,
(int)BYTES_USED_FOR_MEDIUM_REGION(sparse_region)); // DTrace USDT Probe
-#if !CONFIG_AGGRESSIVE_MADVISE
- // Mark free'd dirty pages with MADV_FREE to reduce memory pressure
- medium_free_scan_madvise_free(rack, depot_ptr, sparse_region);
+#if CONFIG_AGGRESSIVE_MADVISE
+ if (!aggressive_madvise_enabled)
#endif
+ {
+ // Mark free'd dirty pages with MADV_FREE to reduce memory pressure
+ medium_free_scan_madvise_free(rack, depot_ptr, sparse_region);
+ }
// If the region is entirely empty vm_deallocate() it outside the depot lock
region_t r_dealloc = medium_free_try_depot_unmap_no_lock(rack, depot_ptr, node);
SZONE_MAGAZINE_PTR_UNLOCK(depot_ptr);
if (r_dealloc) {
- mvm_deallocate_pages(r_dealloc, MEDIUM_REGION_SIZE, 0);
+ mvm_deallocate_pages(r_dealloc, MEDIUM_REGION_SIZE,
+ MALLOC_FIX_GUARD_PAGE_FLAGS(rack->debug_flags));
}
return FALSE; // Caller need not unlock the originating magazine
}
@@ -1475,11 +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
@@ -1502,12 +1513,15 @@
}
} else {
-#if !CONFIG_AGGRESSIVE_MADVISE
- // We are free'ing into the depot, so madvise as we do so unless we were madvising every incoming
- // allocation anyway.
- medium_madvise_free_range_no_lock(rack, medium_mag_ptr,
- vm_kernel_page_size, region, freee, msize, headptr, headsize);
+#if CONFIG_AGGRESSIVE_MADVISE
+ if (!aggressive_madvise_enabled)
#endif
+ {
+ // We are free'ing into the depot, so madvise as we do so unless we were madvising every incoming
+ // allocation anyway.
+ medium_madvise_free_range_no_lock(rack, medium_mag_ptr,
+ vm_kernel_page_size, region, freee, msize, headptr, headsize);
+ }
if (0 < bytes_used || 0 < node->pinned_to_depot) {
/* Depot'd region is still live. Leave it in place on the Depot's recirculation list
@@ -1518,7 +1532,7 @@
region_t r_dealloc = medium_free_try_depot_unmap_no_lock(rack, medium_mag_ptr, node);
SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
if (r_dealloc) {
- mvm_deallocate_pages(r_dealloc, MEDIUM_REGION_SIZE, 0);
+ mvm_deallocate_pages(r_dealloc, MEDIUM_REGION_SIZE, MALLOC_FIX_GUARD_PAGE_FLAGS(rack->debug_flags));
}
return FALSE; // Caller need not unlock
}
@@ -1533,14 +1547,12 @@
msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(ptr);
unsigned index = MEDIUM_META_INDEX_FOR_PTR(ptr);
size_t original_size = MEDIUM_BYTES_FOR_MSIZE(msize);
- unsigned char *next_block = ((unsigned char *)ptr + original_size);
+ void *next_block = ptr + original_size;
msize_t next_index = index + msize;
MALLOC_TRACE(TRACE_medium_free, (uintptr_t)rack, (uintptr_t)medium_mag_ptr, (uintptr_t)ptr, MEDIUM_BYTES_FOR_MSIZE(msize));
-#if CONFIG_AGGRESSIVE_MADVISE || CONFIG_RECIRC_DEPOT
void *original_ptr = ptr;
-#endif
#if DEBUG_MALLOC
if (!msize) {
@@ -1554,6 +1566,9 @@
ptr, msize);
}
#endif
+
+ // Check that the region cookie is intact.
+ region_check_cookie(region, ®ION_COOKIE_FOR_MEDIUM_REGION(region));
// We try to coalesce this block with the preceeding one
if (index > 0 && (meta_headers[index - 1] & MEDIUM_IS_FREE)) {
@@ -1576,7 +1591,7 @@
}
// Try to coalesce with this block with the next block
- if ((next_block < MEDIUM_REGION_END(region)) && (meta_headers[next_index] & MEDIUM_IS_FREE)) {
+ if ((next_block < MEDIUM_REGION_HEAP_END(region)) && (meta_headers[next_index] & MEDIUM_IS_FREE)) {
msize_t next_msize = meta_headers[next_index] & ~MEDIUM_IS_FREE;
free_list_t next = medium_free_list_find_by_ptr(rack, medium_mag_ptr, next_block, next_msize);
medium_free_list_remove_ptr(rack, medium_mag_ptr, next, next_msize);
@@ -1598,15 +1613,24 @@
medium_mag_ptr->mag_num_objects--;
// Update this region's bytes in use count
- region_trailer_t *node = REGION_TRAILER_FOR_MEDIUM_REGION(region);
- size_t bytes_used = node->bytes_used - original_size;
- node->bytes_used = (unsigned int)bytes_used;
-
- // Always attempt to madvise free regions that exceed the conditional
- // madvise limit size.
- medium_madvise_free_range_conditional_no_lock(rack, medium_mag_ptr,
- medium_sliding_madvise_granularity(medium_mag_ptr), region, &freee,
- msize, original_ptr, original_size);
+ region_trailer_t *trailer = REGION_TRAILER_FOR_MEDIUM_REGION(region);
+ size_t bytes_used = trailer->bytes_used - original_size;
+ trailer->bytes_used = (unsigned int)bytes_used;
+
+#if CONFIG_AGGRESSIVE_MADVISE
+ if (aggressive_madvise_enabled) {
+ medium_madvise_free_range_no_lock(rack, medium_mag_ptr,
+ vm_kernel_page_size, region, freee, msize, original_ptr, original_size);
+ }
+ else
+#endif
+ {
+ // Always attempt to madvise free regions that exceed the conditional
+ // madvise limit size.
+ medium_madvise_free_range_conditional_no_lock(rack, medium_mag_ptr,
+ medium_sliding_madvise_granularity(medium_mag_ptr), region, &freee,
+ msize, original_ptr, original_size);
+ }
// Caller must do SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr) if this function
// returns TRUE.
@@ -1634,16 +1658,18 @@
medium_finalize_region(rack, medium_mag_ptr);
}
+ medium_region_t region = (medium_region_t)aligned_address;
+
// Tag the region at "aligned_address" as belonging to us,
// and so put it under the protection of the magazine lock we are holding.
// Do this before advertising "aligned_address" on the hash ring(!)
- MAGAZINE_INDEX_FOR_MEDIUM_REGION(aligned_address) = mag_index;
+ MAGAZINE_INDEX_FOR_MEDIUM_REGION(region) = mag_index;
// Insert the new region into the hash ring
- rack_region_insert(rack, (region_t)aligned_address);
-
- medium_mag_ptr->mag_last_region = aligned_address;
- BYTES_USED_FOR_MEDIUM_REGION(aligned_address) = MEDIUM_BYTES_FOR_MSIZE(msize);
+ rack_region_insert(rack, region);
+
+ medium_mag_ptr->mag_last_region = region;
+ BYTES_USED_FOR_MEDIUM_REGION(region) = MEDIUM_BYTES_FOR_MSIZE(msize);
#if CONFIG_ASLR_INTERNAL
int offset_msize = malloc_entropy[1] & MEDIUM_ENTROPY_MASK;
@@ -1652,13 +1678,13 @@
offset_msize = strtol(getenv("MallocASLRForce"), NULL, 0) & MEDIUM_ENTROPY_MASK;
}
if (getenv("MallocASLRPrint")) {
- malloc_report(ASL_LEVEL_INFO, "Region: %p offset: %d\n", aligned_address, offset_msize);
+ malloc_report(ASL_LEVEL_INFO, "Region: %p offset: %d\n", region, offset_msize);
}
#endif
#else
int offset_msize = 0;
#endif
- ptr = (void *)((uintptr_t)aligned_address +
+ ptr = (void *)(MEDIUM_REGION_HEAP_BASE(region) +
MEDIUM_BYTES_FOR_MSIZE(offset_msize));
medium_meta_header_set_in_use(MEDIUM_META_HEADER_FOR_PTR(ptr),
offset_msize, msize);
@@ -1694,7 +1720,7 @@
// connect to magazine as last node
recirc_list_splice_last(rack, medium_mag_ptr,
- REGION_TRAILER_FOR_MEDIUM_REGION(aligned_address));
+ REGION_TRAILER_FOR_MEDIUM_REGION(region));
return ptr;
}
@@ -1760,7 +1786,8 @@
medium_claimed_address(rack_t *rack, void *ptr)
{
region_t r = medium_region_for_ptr_no_lock(rack, ptr);
- return r && ptr < (void *)MEDIUM_REGION_END(r);
+ return r && ptr >= MEDIUM_REGION_HEAP_BASE(r)
+ && ptr < MEDIUM_REGION_HEAP_END(r);
}
void *
@@ -1849,7 +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);
@@ -1870,7 +1898,7 @@
* Try to expand into unused space immediately after this block.
*/
msize_t unused_msize = MEDIUM_MSIZE_FOR_BYTES(medium_mag_ptr->mag_bytes_free_at_end);
- void *unused_start = MEDIUM_REGION_END(MEDIUM_REGION_FOR_PTR(ptr)) - medium_mag_ptr->mag_bytes_free_at_end;
+ void *unused_start = MEDIUM_REGION_HEAP_END(MEDIUM_REGION_FOR_PTR(ptr)) - medium_mag_ptr->mag_bytes_free_at_end;
if (medium_mag_ptr->mag_last_region == MEDIUM_REGION_FOR_PTR(ptr)
&& coalesced_msize < unused_msize && unused_start == ptr + old_size) {
// Extend the in-use for this block to the new size
@@ -1916,6 +1944,12 @@
/* there's some left, so put the remainder back */
leftover = (unsigned char *)ptr + MEDIUM_BYTES_FOR_MSIZE(new_msize);
medium_free_list_add_ptr(rack, medium_mag_ptr, leftover, leftover_msize);
+ msize_t leftover_index = MEDIUM_META_INDEX_FOR_PTR(leftover);
+ if (madv_headers[leftover_index] & MEDIUM_IS_ADVISED) {
+ medium_madvise_header_mark_clean(madv_headers, leftover_index, leftover_msize);
+ } else {
+ medium_madvise_header_mark_dirty(madv_headers, leftover_index, leftover_msize);
+ }
}
medium_meta_header_set_in_use(meta_headers, index, new_msize);
medium_madvise_header_mark_dirty(madv_headers, index, new_msize);
@@ -1964,9 +1998,9 @@
medium_check_region(rack_t *rack, region_t region, size_t region_index,
unsigned counter)
{
- unsigned char *ptr = MEDIUM_REGION_ADDRESS(region);
+ void *ptr = MEDIUM_REGION_HEAP_BASE(region);
msize_t *meta_headers = MEDIUM_META_HEADER_FOR_PTR(ptr);
- unsigned char *region_end = MEDIUM_REGION_END(region);
+ void *region_end = MEDIUM_REGION_HEAP_END(region);
msize_t prev_free = 0;
unsigned index;
msize_t msize_and_free;
@@ -2038,9 +2072,10 @@
return 0;
}
if (MEDIUM_PREVIOUS_MSIZE(follower) != msize) {
- MEDIUM_CHECK_FAIL("*** invariant broken for medium free %p followed by %p in region [%p-%p] "
+ MEDIUM_CHECK_FAIL("*** invariant broken for medium free %p followed by %p in region %p [%p-%p] "
"(end marker incorrect) should be %d; in fact %d\n",
- ptr, follower, MEDIUM_REGION_ADDRESS(region), region_end, msize, MEDIUM_PREVIOUS_MSIZE(follower));
+ ptr, follower, region, MEDIUM_REGION_HEAP_BASE(region),
+ region_end, msize, MEDIUM_PREVIOUS_MSIZE(follower));
return 0;
}
ptr = (unsigned char *)follower;
@@ -2100,20 +2135,21 @@
for (index = 0; index < num_regions; ++index) {
region = regions[index];
if (HASHRING_OPEN_ENTRY != region && HASHRING_REGION_DEALLOCATED != region) {
- range.address = (vm_address_t)MEDIUM_REGION_ADDRESS(region);
- range.size = MEDIUM_REGION_SIZE;
+ range.address = (vm_address_t)MEDIUM_REGION_HEAP_BASE(region);
+ range.size = MEDIUM_HEAP_SIZE;
if (type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) {
- admin_range.address = range.address + MEDIUM_METADATA_START;
+ admin_range.address = MEDIUM_REGION_METADATA(region);
admin_range.size = MEDIUM_METADATA_SIZE;
recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &admin_range, 1);
}
if (type_mask & (MALLOC_PTR_REGION_RANGE_TYPE | MALLOC_ADMIN_REGION_RANGE_TYPE)) {
ptr_range.address = range.address;
- ptr_range.size = NUM_MEDIUM_BLOCKS * MEDIUM_QUANTUM;
+ ptr_range.size = MEDIUM_HEAP_SIZE;
recorder(task, context, MALLOC_PTR_REGION_RANGE_TYPE, &ptr_range, 1);
}
if (type_mask & MALLOC_PTR_IN_USE_RANGE_TYPE) {
- err = reader(task, range.address, range.size, (void **)&mapped_region);
+ err = reader(task, (vm_address_t)region,
+ (vm_size_t)MEDIUM_REGION_SIZE, (void **)&mapped_region);
if (err) {
return err;
}
@@ -2126,13 +2162,13 @@
// Each magazine could have a pointer to a cached free block from
// this region. Count the regions that have such a pointer.
for (mag_index = 0; mag_index < szone->medium_rack.num_magazines; mag_index++) {
- if ((void *)range.address == (medium_mag_base + mag_index)->mag_last_free_rgn) {
+ if (region == (medium_mag_base + mag_index)->mag_last_free_rgn) {
cached_free_blocks++;
}
}
#endif // CONFIG_MEDIUM_CACHE
- block_header = (msize_t *)(mapped_region + MEDIUM_METADATA_START + sizeof(region_trailer_t));
+ block_header = MEDIUM_META_HEADER_FOR_REGION(mapped_region);
block_index = 0;
block_limit = NUM_MEDIUM_BLOCKS;
if (region == medium_mag_ptr->mag_last_region) {
@@ -2146,7 +2182,7 @@
return KERN_FAILURE; // Somethings amiss. Avoid looping at this block_index.
}
if (!(msize_and_free & MEDIUM_IS_FREE)) {
- vm_address_t ptr = range.address + MEDIUM_BYTES_FOR_MSIZE(block_index);
+ void *ptr = MEDIUM_REGION_HEAP_BASE(region) + MEDIUM_BYTES_FOR_MSIZE(block_index);
#if CONFIG_MEDIUM_CACHE
// If there are still magazines that have cached free
// blocks in this region, check whether this is one of
@@ -2154,7 +2190,7 @@
boolean_t block_cached = false;
if (cached_free_blocks) {
for (mag_index = 0; mag_index < szone->medium_rack.num_magazines; mag_index++) {
- if ((void *)ptr == (medium_mag_base + mag_index)->mag_last_free) {
+ if (ptr == (medium_mag_base + mag_index)->mag_last_free) {
block_cached = true;
cached_free_blocks--;
break;
@@ -2166,7 +2202,7 @@
}
#endif // CONFIG_MEDIUM_CACHE
// Block in use
- buffer[count].address = ptr;
+ buffer[count].address = (vm_address_t)ptr;
buffer[count].size = MEDIUM_BYTES_FOR_MSIZE(msize);
count++;
if (count >= MAX_RECORDER_BUFFER) {
@@ -2242,7 +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;
}
@@ -2254,7 +2290,7 @@
try_medium_from_end:
// Let's see if we can use medium_mag_ptr->mag_bytes_free_at_end
if (medium_mag_ptr->mag_bytes_free_at_end >= MEDIUM_BYTES_FOR_MSIZE(msize)) {
- ptr = MEDIUM_REGION_END(medium_mag_ptr->mag_last_region) -
+ ptr = MEDIUM_REGION_HEAP_END(medium_mag_ptr->mag_last_region) -
medium_mag_ptr->mag_bytes_free_at_end;
medium_mag_ptr->mag_bytes_free_at_end -= MEDIUM_BYTES_FOR_MSIZE(msize);
if (medium_mag_ptr->mag_bytes_free_at_end) {
@@ -2315,10 +2351,13 @@
medium_mag_ptr->mag_num_objects++;
medium_mag_ptr->mag_num_bytes_in_objects += MEDIUM_BYTES_FOR_MSIZE(this_msize);
- // Update this region's bytes in use count
- region_trailer_t *node = REGION_TRAILER_FOR_MEDIUM_REGION(MEDIUM_REGION_FOR_PTR(ptr));
- size_t bytes_used = node->bytes_used + MEDIUM_BYTES_FOR_MSIZE(this_msize);
- node->bytes_used = (unsigned int)bytes_used;
+ // Check that the region cookie is intact and update the region's bytes in use count
+ medium_region_t region = MEDIUM_REGION_FOR_PTR(ptr);
+ region_check_cookie(region, ®ION_COOKIE_FOR_MEDIUM_REGION(region));
+
+ region_trailer_t *trailer = REGION_TRAILER_FOR_MEDIUM_REGION(region);
+ size_t bytes_used = trailer->bytes_used + MEDIUM_BYTES_FOR_MSIZE(this_msize);
+ trailer->bytes_used = (unsigned int)bytes_used;
// Emptiness discriminant
if (bytes_used < DENSITY_THRESHOLD(MEDIUM_REGION_PAYLOAD_BYTES)) {
@@ -2327,7 +2366,7 @@
} else {
/* Region has crossed threshold from sparsity to density. Mark in not "suitable" on the
* recirculation candidates list. */
- node->recirc_suitable = FALSE;
+ trailer->recirc_suitable = FALSE;
}
#if DEBUG_MALLOC
if (LOG(szone, ptr)) {
@@ -2345,7 +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);
@@ -2394,20 +2433,22 @@
// 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_securely(MEDIUM_REGION_SIZE,
- MEDIUM_BLOCKS_ALIGN, VM_MEMORY_MALLOC_MEDIUM,
- rack->debug_flags);
+ fresh_region = mvm_allocate_pages(MEDIUM_REGION_SIZE,
+ MEDIUM_BLOCKS_ALIGN,
+ MALLOC_FIX_GUARD_PAGE_FLAGS(rack->debug_flags),
+ VM_MEMORY_MALLOC_MEDIUM);
SZONE_MAGAZINE_PTR_LOCK(medium_mag_ptr);
// DTrace USDT Probe
@@ -2415,24 +2456,30 @@
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;
}
+ region_set_cookie(®ION_COOKIE_FOR_MEDIUM_REGION(fresh_region));
ptr = medium_malloc_from_region_no_lock(rack, medium_mag_ptr,
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);
}
}
@@ -2479,9 +2526,6 @@
static MALLOC_NOINLINE void
free_medium_botch(rack_t *rack, void *ptr)
{
- mag_index_t mag_index = MAGAZINE_INDEX_FOR_MEDIUM_REGION(MEDIUM_REGION_FOR_PTR(ptr));
- magazine_t *medium_mag_ptr = &(rack->magazines[mag_index]);
- SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
malloc_zone_error(rack->debug_flags, true, "double free for ptr %p\n", ptr);
}
@@ -2514,6 +2558,7 @@
/* check that we don't already have this pointer in the cache */
if (ptr == ptr2) {
+ SZONE_MAGAZINE_PTR_UNLOCK(medium_mag_ptr);
free_medium_botch(rack, ptr);
return;
}
@@ -2635,9 +2680,9 @@
{
unsigned counts[1024];
unsigned in_use = 0;
- uintptr_t start = (uintptr_t)MEDIUM_REGION_ADDRESS(region);
+ uintptr_t start = (uintptr_t)MEDIUM_REGION_HEAP_BASE(region);
uintptr_t current = start + bytes_at_start;
- uintptr_t limit = (uintptr_t)MEDIUM_REGION_END(region) - bytes_at_end;
+ uintptr_t limit = (uintptr_t)MEDIUM_REGION_HEAP_END(region) - bytes_at_end;
uintptr_t mapped_start;
msize_t msize_and_free;
msize_t msize;
@@ -2705,8 +2750,9 @@
}
if ((b = _simple_salloc()) != NULL) {
mag_index_t mag_index = MAGAZINE_INDEX_FOR_MEDIUM_REGION(mapped_region);
- _simple_sprintf(b, "Medium region [%p-%p, %y] \t", (void *)start,
- MEDIUM_REGION_END(region), (int)MEDIUM_REGION_SIZE);
+ _simple_sprintf(b, "Medium region %p [%p-%p, %y] \t", region,
+ (void *)start, MEDIUM_REGION_HEAP_END(region),
+ (int)MEDIUM_REGION_SIZE);
if (mag_index == DEPOT_MAGAZINE_INDEX) {
_simple_sprintf(b, "Recirc depot \t");
} else {
@@ -2773,8 +2819,9 @@
}
malloc_report(MALLOC_REPORT_NOLOG | MALLOC_REPORT_NOPREFIX,
- "Medium region [%p-%p, %y, %y]\n", (void *)region,
- MEDIUM_REGION_END(region), (int)MEDIUM_REGION_SIZE,
+ "Medium region %p [%p-%p, %y, %y]\n", (void *)region,
+ MEDIUM_REGION_HEAP_BASE(region),
+ MEDIUM_REGION_HEAP_END(region), (int)MEDIUM_REGION_SIZE,
((medium_region_t)region)->trailer.bytes_used);
for (size_t x = 0; x < NUM_MEDIUM_BLOCKS; x++) {