Loading...
--- libmalloc/libmalloc-374.100.5/src/nanov2_malloc.c
+++ libmalloc/libmalloc-521.100.59/src/nanov2_malloc.c
@@ -41,18 +41,38 @@
#pragma mark -
#pragma mark Externals for resolved functions
-MALLOC_NOEXPORT extern void *nanov2_allocate(nanozonev2_t *nanozone, size_t rounded_size,
- boolean_t clear);
-MALLOC_NOEXPORT extern void nanov2_free_to_block(nanozonev2_t *nanozone, void *ptr,
- nanov2_size_class_t size_class);
-MALLOC_NOEXPORT extern boolean_t nanov2_madvise_block(nanozonev2_t *nanozone,
- nanov2_block_meta_t *block_metap, nanov2_block_t *blockp,
- nanov2_size_class_t size_class);
MALLOC_NOEXPORT extern size_t nanov2_pointer_size(nanozonev2_t *nanozone, void *ptr,
boolean_t allow_inner);
MALLOC_NOEXPORT extern size_t nanov2_pressure_relief(nanozonev2_t *nanozone, size_t goal);
#if OS_VARIANT_RESOLVED
+MALLOC_ALWAYS_INLINE MALLOC_INLINE size_t
+nanov2_pointer_size_inline(nanozonev2_t *nanozone, void *ptr, boolean_t allow_inner,
+ nanov2_size_class_t *size_class_out, nanov2_block_meta_t **block_metap_out);
+
+MALLOC_ALWAYS_INLINE MALLOC_INLINE void *
+nanov2_allocate_from_block_inline(nanozonev2_t *nanozone,
+ nanov2_block_meta_t *block_metap, nanov2_size_class_t size_class,
+ nanov2_block_meta_t **madvise_block_metapp_out, bool *corruption);
+
+static void *
+nanov2_allocate_outlined(nanozonev2_t *nanozone,
+ nanov2_block_meta_t **block_metapp, size_t rounded_size,
+ nanov2_size_class_t size_class, int allocation_index,
+ nanov2_block_meta_t *madvise_block_metap, void *corrupt_slot,
+ bool clear);
+
+MALLOC_ALWAYS_INLINE MALLOC_INLINE nanov2_block_meta_t *
+nanov2_free_to_block_inline(nanozonev2_t *nanozone, void *ptr,
+ nanov2_size_class_t size_class, nanov2_block_meta_t *block_metap);
+
+static boolean_t nanov2_madvise_block_locked(
+ nanozonev2_t *nanozone, nanov2_block_meta_t *block_metap,
+ nanov2_block_t *blockp, nanov2_size_class_t size_class, uint32_t expected_state);
+static void nanov2_madvise_block(nanozonev2_t *nanozone,
+ nanov2_block_meta_t *block_metap, nanov2_size_class_t size_class,
+ uint32_t expected_state);
+
MALLOC_NOEXPORT extern nanov2_arena_t *nanov2_allocate_new_region(nanozonev2_t *nanozone);
#endif // OS_VARIANT_RESOLVED
@@ -637,13 +657,13 @@
// Default case is max magazines == physical number of CPUs, which
// must be > _malloc_cpu_number() >> hyper_shift, so the modulo
// operation is not required.
- return _malloc_cpu_number() >> hyper_shift;
+ return (_malloc_cpu_number() >> hyper_shift) & MAX_CURRENT_BLOCKS_MASK;
}
#else // CONFIG_NANO_USES_HYPER_SHIFT
if (os_likely(nano_common_max_magazines_is_ncpu)) {
// Default case is max magazines == logical number of CPUs, which
// must be > _malloc_cpu_number() so the modulo operation is not required.
- return _malloc_cpu_number();
+ return _malloc_cpu_number() & MAX_CURRENT_BLOCKS_MASK;
}
#endif // CONFIG_NANO_USES_HYPER_SHIFT
@@ -653,9 +673,11 @@
#endif // CONFIG_NANO_USES_HYPER_SHIFT
if (os_likely(_os_cpu_number_override == -1)) {
- return (_malloc_cpu_number() >> shift) % nano_common_max_magazines;
- }
- return (_os_cpu_number_override >> shift) % nano_common_max_magazines;
+ return ((_malloc_cpu_number() >> shift) % nano_common_max_magazines) &
+ MAX_CURRENT_BLOCKS_MASK;
+ }
+ return ((_os_cpu_number_override >> shift) % nano_common_max_magazines) &
+ MAX_CURRENT_BLOCKS_MASK;
}
#endif // OS_VARIANT_RESOLVED
@@ -1040,7 +1062,8 @@
MALLOC_NOEXPORT size_t
nanov2_size(nanozonev2_t *nanozone, const void *ptr)
{
- size_t size = nanov2_pointer_size(nanozone, (void *)ptr, FALSE);
+ size_t size = nanov2_pointer_size_inline(nanozone, (void *)ptr, FALSE,
+ NULL, NULL);
return size ? size : nanozone->helper_zone->size(nanozone->helper_zone, ptr);
}
@@ -1049,54 +1072,166 @@
{
size_t rounded_size = _nano_common_good_size(size);
if (rounded_size <= NANO_MAX_SIZE) {
- void *ptr = nanov2_allocate(nanozone, rounded_size, FALSE);
- if (ptr) {
- if (os_unlikely(size && (nanozone->debug_flags & MALLOC_DO_SCRIBBLE))) {
- memset(ptr, SCRIBBLE_BYTE, size);
+ nanov2_block_meta_t *madvise_block_metap = NULL;
+ nanov2_size_class_t size_class = nanov2_size_class_from_size(rounded_size);
+
+ // Get the index of the pointer to the block from which we are should be
+ // allocating. This currently depends on the physical CPU number.
+ int allocation_index = nanov2_get_allocation_block_index();
+
+ // Get the current allocation block meta data pointer. If this is NULL,
+ // we need to find a new allocation block.
+ nanov2_block_meta_t **block_metapp =
+ &nanozone->current_block[size_class][allocation_index];
+ nanov2_block_meta_t *block_metap = os_atomic_load(block_metapp, relaxed);
+ bool corruption = false;
+ void *ptr = NULL;
+ if (block_metap) {
+ // Fast path: we have a block -- try to allocate from it.
+ ptr = nanov2_allocate_from_block_inline(nanozone, block_metap,
+ size_class, &madvise_block_metap, &corruption);
+ if (ptr && !corruption) {
+ // Always clear the double-free guard so that we can recognize
+ // that this block is not on the free list.
+ nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
+ os_atomic_store(&slotp->double_free_guard, 0, relaxed);
+
+ // We know the body of the allocation is already clear, so we just
+ // need to clean up the next_slot word to get to all-zero. Do so in
+ // all cases, even if a cleared allocation is not requested, to
+ // prevent any leakage through the next_slot bits.
+ os_atomic_store(&slotp->next_slot, 0, relaxed);
+ return ptr;
+ }
+ }
+
+ return nanov2_allocate_outlined(nanozone, block_metapp, rounded_size,
+ size_class, allocation_index, madvise_block_metap, ptr, false);
+ }
+
+ // Too big for nano, so delegate to the helper zone.
+ return nanozone->helper_zone->malloc(nanozone->helper_zone, size);
+}
+
+MALLOC_ALWAYS_INLINE MALLOC_INLINE
+void
+nanov2_bzero(void *ptr, size_t size)
+{
+ // TODO: inline bzero from libplatform
+ bzero(ptr, size);
+}
+
+MALLOC_NOEXPORT void
+nanov2_free_definite_size(nanozonev2_t *nanozone, void *ptr, size_t size)
+{
+ if (ptr && nanov2_has_valid_signature(ptr)) {
+ nanov2_size_class_t size_class = nanov2_size_class_from_size(size);
+
+ if (malloc_zero_policy == MALLOC_ZERO_ON_FREE) {
+ if (size_class != 0) {
+ nanov2_bzero((char *)ptr + sizeof(nanov2_free_slot_t),
+ size - sizeof(nanov2_free_slot_t));
+ }
+ }
+
+ nanov2_block_meta_t *madvise_block_metap = nanov2_free_to_block_inline(
+ nanozone, ptr, size_class, NULL);
+ if (madvise_block_metap) {
+ nanov2_madvise_block(nanozone, madvise_block_metap, size_class,
+ SLOT_CAN_MADVISE);
+ }
+ return;
+ }
+ return nanozone->helper_zone->free_definite_size(nanozone->helper_zone, ptr,
+ size);
+}
+
+static void
+_nanov2_free(nanozonev2_t *nanozone, void *ptr, bool try)
+{
+ if (ptr) {
+ // Check whether it's a Nano pointer and get the size. If it's not
+ // Nano, pass it to the helper zone.
+ nanov2_size_class_t size_class;
+ nanov2_block_meta_t *block_metap;
+ size_t size = nanov2_pointer_size_inline(nanozone, ptr, FALSE,
+ &size_class, &block_metap);
+ if (size) {
+ if (malloc_zero_policy == MALLOC_ZERO_ON_FREE) {
+ if (size > sizeof(nanov2_free_slot_t)) {
+ nanov2_bzero((char *)ptr + sizeof(nanov2_free_slot_t),
+ size - sizeof(nanov2_free_slot_t));
+ }
+ }
+
+ nanov2_block_meta_t *madvise_block_metap = nanov2_free_to_block_inline(
+ nanozone, ptr, size_class, block_metap);
+ if (madvise_block_metap) {
+ nanov2_madvise_block(nanozone, madvise_block_metap, size_class,
+ SLOT_CAN_MADVISE);
+ }
+ return;
+ }
+ }
+ return try ? nanozone->helper_zone->try_free_default(nanozone->helper_zone, ptr) :
+ nanozone->helper_zone->free(nanozone->helper_zone, ptr);
+}
+
+MALLOC_NOEXPORT void
+nanov2_free(nanozonev2_t *nanozone, void *ptr)
+{
+ _nanov2_free(nanozone, ptr, false);
+}
+
+MALLOC_NOEXPORT void
+nanov2_try_free_default(nanozonev2_t *nanozone, void *ptr)
+{
+ _nanov2_free(nanozone, ptr, true);
+}
+
+MALLOC_ALWAYS_INLINE MALLOC_INLINE
+void *
+nanov2_malloc_zero(nanozonev2_t *nanozone, size_t rounded_size)
+{
+ nanov2_block_meta_t *madvise_block_metap = NULL;
+ nanov2_size_class_t size_class = nanov2_size_class_from_size(rounded_size);
+
+ // Get the index of the pointer to the block from which we are should be
+ // allocating. This currently depends on the physical CPU number.
+ int allocation_index = nanov2_get_allocation_block_index();
+
+ // Get the current allocation block meta data pointer. If this is NULL,
+ // we need to find a new allocation block.
+ nanov2_block_meta_t **block_metapp =
+ &nanozone->current_block[size_class][allocation_index];
+ nanov2_block_meta_t *block_metap = os_atomic_load(block_metapp, relaxed);
+ bool corruption = false;
+ void *ptr = NULL;
+ if (block_metap) {
+ // Fast path: we have a block -- try to allocate from it.
+ ptr = nanov2_allocate_from_block_inline(nanozone, block_metap,
+ size_class, &madvise_block_metap, &corruption);
+ if (ptr && !corruption) {
+ if (malloc_zero_policy == MALLOC_ZERO_ON_FREE) {
+ // Always clear the double-free guard so that we can recognize that
+ // this block is not on the free list.
+ nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
+ os_atomic_store(&slotp->double_free_guard, 0, relaxed);
+
+ // We know the body of the allocation is already clear, so we just
+ // need to clean up the next_slot word to get to all-zero. Do so in
+ // all cases, even if a cleared allocation is not requested, to
+ // prevent any leakage through the next_slot bits.
+ os_atomic_store(&slotp->next_slot, 0, relaxed);
+ } else {
+ nanov2_bzero(ptr, rounded_size);
}
return ptr;
}
}
- // If we reach this point, we couldn't allocate, so delegate to the
- // helper zone.
- return nanozone->helper_zone->malloc(nanozone->helper_zone, size);
-}
-
-MALLOC_NOEXPORT void
-nanov2_free_definite_size(nanozonev2_t *nanozone, void *ptr, size_t size)
-{
- // Check whether it's a Nano pointer and get the size. We should only get
- // here if it is and furthermore we already know that "size" is the actual
- // rounded size, so don't waste time rechecking that. This is just a
- // sanity check.
- if (ptr && nanov2_has_valid_signature(ptr)) {
- if (os_unlikely(nanozone->debug_flags & MALLOC_DO_SCRIBBLE)) {
- memset(ptr, SCRABBLE_BYTE, size);
- }
- nanov2_free_to_block(nanozone, ptr, nanov2_size_class_from_size(size));
- return;
- }
- return nanozone->helper_zone->free_definite_size(nanozone->helper_zone, ptr,
- size);
-}
-
-MALLOC_NOEXPORT void
-nanov2_free(nanozonev2_t *nanozone, void *ptr)
-{
- if (ptr && nanov2_has_valid_signature(ptr)) {
- // Check whether it's a Nano pointer and get the size. If it's not
- // Nano, pass it to the helper zone.
- size_t size = nanov2_pointer_size(nanozone, ptr, FALSE);
- if (size) {
- if (os_unlikely(nanozone->debug_flags & MALLOC_DO_SCRIBBLE)) {
- memset(ptr, SCRABBLE_BYTE, size);
- }
- nanov2_free_to_block(nanozone, ptr, nanov2_size_class_from_size(size));
- return;
- }
- }
- return nanozone->helper_zone->free(nanozone->helper_zone, ptr);
+ return nanov2_allocate_outlined(nanozone, block_metapp, rounded_size,
+ size_class, allocation_index, madvise_block_metap, ptr, true);
}
MALLOC_NOEXPORT void *
@@ -1108,15 +1243,23 @@
}
size_t rounded_size = _nano_common_good_size(total_bytes);
if (total_bytes <= NANO_MAX_SIZE) {
- void *ptr = nanov2_allocate(nanozone, rounded_size, TRUE);
- if (ptr) {
- return ptr;
- }
- }
-
- // If we reach this point, we couldn't allocate, so delegate to the
- // helper zone.
+ return nanov2_malloc_zero(nanozone, rounded_size);
+ }
+
+ // Too big for nano, so delegate to the helper zone.
return nanozone->helper_zone->calloc(nanozone->helper_zone, 1, total_bytes);
+}
+
+MALLOC_NOEXPORT void *
+nanov2_malloc_zero_on_alloc(nanozonev2_t *nanozone, size_t size)
+{
+ size_t rounded_size = _nano_common_good_size(size);
+ if (rounded_size <= NANO_MAX_SIZE) {
+ return nanov2_malloc_zero(nanozone, rounded_size);
+ }
+
+ // Too big for nano, so delegate to the helper zone.
+ return nanozone->helper_zone->malloc(nanozone->helper_zone, size);
}
#endif // OS_VARIANT_RESOLVED
@@ -1170,7 +1313,9 @@
}
} else {
// Same size or shrinking by less than half size. Keep the same
- // allocation and clear the area that's being released.
+ // allocation and scribble the area that's being released. Nothing
+ // to do for zero-on-free yet; that will be taken care of when the
+ // shrunk allocation is freed.
if (new_size != old_size) {
MALLOC_ASSERT(new_size < old_size);
if (os_unlikely(nanozone->debug_flags & MALLOC_DO_SCRIBBLE)) {
@@ -1217,7 +1362,8 @@
size_t rounded_size = _nano_common_good_size(size);
if (rounded_size <= NANO_MAX_SIZE) {
while (allocated < count) {
- void *ptr = nanov2_allocate(nanozone, rounded_size, FALSE);
+ // TODO: nanov2_malloc will redo _nano_common_good_size
+ void *ptr = nanov2_malloc(nanozone, rounded_size);
if (!ptr) {
break;
}
@@ -1248,19 +1394,20 @@
}
}
}
-#endif // OS_VARIANT_RESOLVED
-
-#if OS_VARIANT_NOTRESOLVED
-static void *
+
+MALLOC_NOEXPORT void *
nanov2_memalign(nanozonev2_t *nanozone, size_t alignment, size_t size)
{
- // Always delegate this to the helper zone.
+ // Serve directly if the requested alignment is trivially satisfied by our
+ // baseline alignment (16 bytes)
+ if (alignment <= NANO_REGIME_QUANTA_SIZE) {
+ return nanov2_malloc(nanozone, size);
+ }
+
+ // Otherwise delegate to the helper zone
return nanozone->helper_zone->memalign(nanozone->helper_zone, alignment,
size);
}
-#endif // OS_VARIANT_NOTRESOLVED
-
-#if OS_VARIANT_RESOLVED
size_t
nanov2_pressure_relief(nanozonev2_t *nanozone, size_t goal)
@@ -1306,8 +1453,9 @@
if (meta.next_slot == SLOT_CAN_MADVISE) {
nanov2_block_t *blockp = nanov2_block_address_from_meta_index(
nanozone, arena, i);
- if (nanov2_madvise_block(nanozone, block_metap,
- blockp, nanov2_size_class_for_ptr(nanozone, blockp))) {
+ if (nanov2_madvise_block_locked(nanozone, block_metap,
+ blockp, nanov2_size_class_for_ptr(nanozone, blockp),
+ SLOT_CAN_MADVISE)) {
total += NANOV2_BLOCK_SIZE;
}
}
@@ -1369,9 +1517,7 @@
kern_return_t kr;
bitarray_t slots;
- if (!reader) {
- reader = nano_common_default_reader;
- }
+ reader = reader_or_in_memory_fallback(reader, task);
kr = reader(task, zone_address, sizeof(nanozonev2_t), (void **)&nanozone);
if (kr) {
@@ -1476,7 +1622,7 @@
bitarray_zap(slots, log_size, next_slot);
void *ptr = nanov2_slot_in_block_ptr(blockp, size_class, next_slot);
nanov2_free_slot_t *slotp = NANOV2_ZONE_PTR_TO_MAPPED_PTR(nanov2_free_slot_t *, ptr, ptr_offset);
- next_slot = slotp->next_slot;
+ next_slot = (uint16_t)slotp->next_slot;
free_list_count++;
}
// Add a range for each slot that is not on the freelist,
@@ -1495,7 +1641,7 @@
ranges[range_count].size = slot_size;
range_count++;
}
- free(slots);
+ _free(slots);
}
if (range_count) {
// Notify the in-use pointers that we found.
@@ -1812,7 +1958,7 @@
malloc_statistics_t *stats)
{
printer = printer ? printer : nanov2_null_printer;
- reader = !reader && task == mach_task_self() ? _malloc_default_reader : reader;
+ reader = reader_or_in_memory_fallback(reader, task);
kern_return_t err;
@@ -1963,8 +2109,10 @@
// the allocation, or 0 if the pointer does not correspond to an active
// allocation. If allow_inner is true, the pointer need not point to the start
// of the allocation.
-size_t
-nanov2_pointer_size(nanozonev2_t *nanozone, void *ptr, boolean_t allow_inner)
+MALLOC_ALWAYS_INLINE MALLOC_INLINE size_t
+nanov2_pointer_size_inline(nanozonev2_t *nanozone, void *ptr,
+ boolean_t allow_inner, nanov2_size_class_t *size_class_out,
+ nanov2_block_meta_t **block_metap_out)
{
// First check the address signature.
if (!nanov2_has_valid_signature((void *)ptr)) {
@@ -2046,27 +2194,42 @@
return 0;
}
+ if (size_class_out) {
+ *size_class_out = size_class;
+ }
+ if (block_metap_out) {
+ *block_metap_out = block_metap;
+ }
return size;
+}
+
+size_t
+nanov2_pointer_size(nanozonev2_t *nanozone, void *ptr, boolean_t allow_inner)
+{
+ return nanov2_pointer_size_inline(nanozone, ptr, allow_inner, NULL, NULL);
}
#pragma mark -
#pragma mark Madvise Management
-// Given a pointer to a block and its metadata, calls madvise() on that block
-// if it is in state SLOT_CAN_MADVISE. Returns true on success, false if the
-// block is not in the correct state or if the state changed during the
-// operation.
+// Given a pointer to a block and its metadata, calls madvise() on that block if
+// it is still in the state we expect, either SLOT_CAN_MADVISE or SLOT_MADVISED
+// (the latter expected when we need to pessimistically re-madvise a block we
+// may have touched while racing to allocate against a transition to
+// SLOT_CAN_MADVISE). Returns true on success, false if the block is not in the
+// correct state or if the state changed during the operation.
//
// This function must be called with the zone's madvise_lock held
-boolean_t
-nanov2_madvise_block(nanozonev2_t *nanozone, nanov2_block_meta_t *block_metap,
- nanov2_block_t *blockp, nanov2_size_class_t size_class)
+static boolean_t
+nanov2_madvise_block_locked(nanozonev2_t *nanozone,
+ nanov2_block_meta_t *block_metap, nanov2_block_t *blockp,
+ nanov2_size_class_t size_class, uint32_t expected_state)
{
_malloc_lock_assert_owner(&nanozone->madvise_lock);
boolean_t madvised = FALSE;
nanov2_block_meta_t old_meta = os_atomic_load(block_metap, relaxed);
- if (old_meta.next_slot == SLOT_CAN_MADVISE) {
+ if (old_meta.next_slot == expected_state) {
// Nobody raced with us. We can safely madvise this block. First change
// the state to SLOT_MADVISING so that other threads don't try to
// grab the block for new allocations.
@@ -2108,6 +2271,18 @@
return madvised;
}
+static void
+nanov2_madvise_block(nanozonev2_t *nanozone, nanov2_block_meta_t *block_metap,
+ nanov2_size_class_t size_class, uint32_t expected_state)
+{
+ nanov2_block_t *blockp = nanov2_block_address_from_meta_ptr(nanozone,
+ block_metap);
+ _malloc_lock_lock(&nanozone->madvise_lock);
+ nanov2_madvise_block_locked(nanozone, block_metap, blockp, size_class,
+ expected_state);
+ _malloc_lock_unlock(&nanozone->madvise_lock);
+}
+
#endif // OS_VARIANT_RESOLVED
#pragma mark -
@@ -2115,26 +2290,49 @@
#if OS_VARIANT_NOTRESOLVED
-#if NANOV2_MULTIPLE_REGIONS
-static nanov2_addr_t nanov2_max_region_base = {
- .fields.nano_signature = NANOZONE_SIGNATURE,
- .fields.nano_region = NANOV2_MAX_REGION_NUMBER
-};
-#endif // NANOV2_MULTIPLE_REGIONS
-
+#if CONFIG_NANO_RESERVE_REGIONS
+// Update protection for region to DEFAULT
+static bool
+nanov2_unprotect_region(nanov2_region_t *region)
+{
+ MALLOC_TRACE(TRACE_nanov2_region_protection | DBG_FUNC_START,
+ (uint64_t)region, 0, 0, 0);
+ bool result = nano_common_unprotect_vm_space((mach_vm_address_t)region,
+ NANOV2_REGION_SIZE);
+ MALLOC_TRACE(TRACE_nanov2_region_protection | DBG_FUNC_END,
+ (uint64_t)region, result, 0, 0);
+ return result;
+}
+
+// Reserve VA at [base, base+num_regions*REGION_SIZE].
+// Note: permissions must still be granted on reserved region with `nanov2_unprotect_region`
+static bool
+nanov2_reserve_regions(nanov2_region_t *base, unsigned int num_regions)
+{
+ MALLOC_TRACE(TRACE_nanov2_region_reservation | DBG_FUNC_START,
+ (uint64_t)base, num_regions, 0, 0);
+ bool result = nano_common_reserve_vm_space((mach_vm_address_t)base,
+ (NANOV2_REGION_SIZE * (mach_vm_size_t)num_regions));
+ MALLOC_TRACE(TRACE_nanov2_region_reservation | DBG_FUNC_END,
+ (uint64_t)base, num_regions, result, 0);
+
+ return result;
+}
+#else
// Attempts to allocate VM space for a region at a given address and returns
// whether the allocation succeeded.
-static boolean_t
+static bool
nanov2_allocate_region(nanov2_region_t *region)
{
MALLOC_TRACE(TRACE_nanov2_region_allocation | DBG_FUNC_START,
(uint64_t)region, 0, 0, 0);
- boolean_t result = nano_common_allocate_vm_space((mach_vm_address_t)region,
+ bool result = nano_common_allocate_vm_space((mach_vm_address_t)region,
NANOV2_REGION_SIZE);
MALLOC_TRACE(TRACE_nanov2_region_allocation | DBG_FUNC_END,
(uint64_t)region, result, 0, 0);
return result;
}
+#endif // CONFIG_NANO_RESERVE_REGIONS
// Allocates a new region adjacent to the current one. If the allocation fails,
// keep sliding up by the size of a region until we either succeed or run out of
@@ -2146,11 +2344,26 @@
#if NANOV2_MULTIPLE_REGIONS
bool allocated = false;
+ nanov2_addr_t nanov2_max_region_base = {
+ .fields.nano_signature = NANOZONE_SIGNATURE,
+ .fields.nano_region = nano_max_region,
+ };
+
_malloc_lock_assert_owner(&nanozone->regions_lock);
nanov2_region_t *current_region = nanov2_current_region_base(
os_atomic_load(&nanozone->current_region_next_arena, relaxed));
nanov2_region_t *next_region = current_region + 1;
+
while ((void *)next_region <= nanov2_max_region_base.addr) {
+#if CONFIG_NANO_RESERVE_REGIONS
+ if (!nanov2_unprotect_region(next_region)) {
+ MALLOC_REPORT_FATAL_ERROR(next_region,
+ "Nano: Unable to raise protection on pre-allocated region");
+ }
+ nanozone->statistics.allocated_regions++;
+ allocated = true;
+ break;
+#else // CONFIG_NANO_RESERVE_REGIONS
if (nanov2_allocate_region(next_region)) {
nanozone->statistics.allocated_regions++;
allocated = true;
@@ -2163,6 +2376,7 @@
// atomically here. Published by the store-release of
// current_region_next_arena.
os_atomic_inc(&nanozone->statistics.region_address_clashes, relaxed);
+#endif // CONFIG_NANO_RESERVE_REGIONS
}
if (!allocated) {
@@ -2203,18 +2417,39 @@
#if OS_VARIANT_RESOLVED
+MALLOC_NOINLINE MALLOC_NORETURN
+static void
+nanov2_guard_corruption_detected(void *corrupt_slot)
+{
+ uint64_t guard = *(uint64_t *)corrupt_slot;
+ malloc_zone_error(MALLOC_ABORT_ON_CORRUPTION, true,
+ "Heap corruption detected, free list is damaged at %p\n"
+ "*** Incorrect guard value: %llu\n", corrupt_slot,
+ (unsigned long long)guard);
+ __builtin_unreachable();
+}
+
// Allocates memory from the block that corresponds to a given block meta data
// pointer. The memory is taken from the free list if possible, or from the
// unused region of the block if not. If the block is no longer in use or is
// full, NULL is returned and the caller is expected to find another block to
// allocate from.
-MALLOC_NOEXPORT
+MALLOC_ALWAYS_INLINE MALLOC_INLINE
void *
-nanov2_allocate_from_block(nanozonev2_t *nanozone,
- nanov2_block_meta_t *block_metap, nanov2_size_class_t size_class)
+nanov2_allocate_from_block_inline(nanozonev2_t *nanozone,
+ nanov2_block_meta_t *block_metap, nanov2_size_class_t size_class,
+ nanov2_block_meta_t **madvise_block_metap_out, bool *corruption)
{
nanov2_block_meta_view_t old_meta_view;
- old_meta_view.meta = os_atomic_load(block_metap, relaxed);
+
+ // Our loads of the block metadata use dependency ordering, which guarantees
+ // that any loads we do from a slot pointer derived from the metadata value
+ // as we do below will observe all of the stores preceding the store-release
+ // of that value we observed. This allows us to safely rely on the contents
+ // of the slot updated when it was last freed, including the double-free
+ // guard and zeroing done by zero-on-free (which is required for correctness
+ // in the case of calloc).
+ old_meta_view.meta = os_atomic_load(block_metap, dependency);
// Calculating blockp and ptr is relatively expensive. Do both lazily to
// minimize the time in the block starting with "again:" and ending with the
@@ -2255,7 +2490,7 @@
slot = old_meta_view.meta.next_slot - 1; // meta.next_slot is 1-based.
ptr = nanov2_slot_in_block_ptr(blockp, size_class, slot);
nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
- new_meta.next_slot = slot_full ? SLOT_FULL : slotp->next_slot;
+ new_meta.next_slot = slot_full ? SLOT_FULL : (uint16_t)slotp->next_slot;
}
// Write the updated meta data; try again if we raced with another thread.
@@ -2264,21 +2499,8 @@
if (old_meta_view.meta.next_slot == SLOT_CAN_MADVISE ||
old_meta_view.meta.next_slot == SLOT_MADVISING ||
old_meta_view.meta.next_slot == SLOT_MADVISED) {
- _malloc_lock_lock(&nanozone->madvise_lock);
- if (old_meta_view.meta.next_slot == SLOT_MADVISED) {
- // We raced against another thread madvising this block. We need
- // to redo the madvise because we may have touched it when
- // reading the next pointer in the freelist.
- if (!blockp) {
- blockp = nanov2_block_address_from_meta_ptr(nanozone, block_metap);
- }
- if (mvm_madvise_free(nanozone, nanov2_region_address_for_ptr(blockp),
- (uintptr_t)blockp, (uintptr_t)(blockp + 1), NULL, FALSE)) {
- malloc_zone_error(0, false,
- "Failed to remadvise block at blockp: %p, error: %d\n", blockp, errno);
- }
- }
- _malloc_lock_unlock(&nanozone->madvise_lock);
+ *madvise_block_metap_out = block_metap;
+ return NULL;
}
goto again;
}
@@ -2290,26 +2512,40 @@
ptr = nanov2_slot_in_block_ptr(blockp, size_class, slot);
}
- nanov2_free_slot_t *slotp = os_atomic_inject_dependency(ptr,
- (unsigned long)old_meta_view.bits);
if (from_free_list) {
// We grabbed the item from the free list. Check the free list canary
// and crash if it's not valid. We can't do this check before the
// cmpxchgv because another thread may race with us, claim the slot and
// write to it.
+ nanov2_free_slot_t *slotp = ptr;
uintptr_t guard = os_atomic_load(&slotp->double_free_guard, relaxed);
- if ((guard ^ nanozone->slot_freelist_cookie) != (uintptr_t)ptr) {
- malloc_zone_error(MALLOC_ABORT_ON_CORRUPTION, true,
- "Heap corruption detected, free list is damaged at %p\n"
- "*** Incorrect guard value: %lu\n", ptr, guard);
- __builtin_unreachable();
- }
- }
-
+ if (os_unlikely((guard ^ nanozone->slot_freelist_cookie) != (uintptr_t)ptr)) {
+ *corruption = true;
+ }
+ }
+
#if DEBUG_MALLOC
nanozone->statistics.size_class_statistics[size_class].total_allocations++;
#endif // DEBUG_MALLOC
+ return ptr;
+}
+
+static void *
+nanov2_allocate_from_block(nanozonev2_t *nanozone,
+ nanov2_block_meta_t *block_metap, nanov2_size_class_t size_class)
+{
+ nanov2_block_meta_t *madvise_block_metap = NULL;
+ bool corruption = false;
+ void *ptr = nanov2_allocate_from_block_inline(nanozone, block_metap,
+ size_class, &madvise_block_metap, &corruption);
+ if (os_unlikely(corruption)) {
+ nanov2_guard_corruption_detected(ptr);
+ }
+ if (madvise_block_metap) {
+ nanov2_madvise_block(nanozone, madvise_block_metap, size_class,
+ SLOT_MADVISED);
+ }
return ptr;
}
@@ -2657,39 +2893,36 @@
return NULL;
}
-// Allocates memory of a given size (which must be a multiple of the Nano
-// quantum size) and optionally clears it (for calloc).
+// This function is called when a fast-path allocation from a given (size_class,
+// allocation_index) has been tried and failed, and we need to act on
+// observations from that attempt and/or retry the allocation. Its rather
+// tortured calling contract is designed to allow the caller to avoid pushing a
+// frame and pass along as much of what it has already computed as possible.
//
-// Allocation is attempted first from the block last used for the caller's
-// context (which is initially the physical CPU by default). If there is no
-// last block, or the block is full or now out of use, find another one, if
-// possible. See the comments for nanov2_get_allocation_block() for the details.
+// If @corrupt_slot is non-NULL it means we detected corruption of the slot's
+// guard on the fast path, and we need to report that corruption.
+//
+// If @madvise_block_metap is non-NULL it means we raced with another thread
+// madvising the block we tried to allocate from and need to re-madvise it.
//
// If the allocation fails, NULL is returned.
-void *
-nanov2_allocate(nanozonev2_t *nanozone, size_t rounded_size, boolean_t clear)
+static void *
+nanov2_allocate_outlined(nanozonev2_t *nanozone, nanov2_block_meta_t **block_metapp,
+ size_t rounded_size, nanov2_size_class_t size_class,
+ int allocation_index, nanov2_block_meta_t *madvise_block_metap,
+ void *corrupt_slot, bool clear)
{
void *ptr = NULL;
- nanov2_size_class_t size_class = nanov2_size_class_from_size(rounded_size);
- MALLOC_ASSERT(size_class < NANO_SIZE_CLASSES);
- MALLOC_ASSERT(rounded_size != 0);
- nanov2_block_meta_t *block_metap;
- nanov2_block_meta_t **block_metapp;
-
- // Get the index of the pointer to the block from which we are should be
- // allocating. This currently depends on the physical CPU number.
- int allocation_index = nanov2_get_allocation_block_index() & MAX_CURRENT_BLOCKS_MASK;
-
- // Get the current allocation block meta data pointer. If this is NULL,
- // we need to find a new allocation block.
- block_metapp = &nanozone->current_block[size_class][allocation_index];
- block_metap = os_atomic_load(block_metapp, relaxed);
- if (block_metap) {
- // Fast path: we have a block -- try to allocate from it.
- ptr = nanov2_allocate_from_block(nanozone, block_metap, size_class);
- if (ptr) {
- goto done;
- }
+
+ if (os_unlikely(corrupt_slot)) {
+ nanov2_guard_corruption_detected(corrupt_slot);
+ }
+
+ // If we need to re-madvise the old block that we might have accidentally
+ // touched out of turn, do so now.
+ if (madvise_block_metap) {
+ nanov2_madvise_block(nanozone, madvise_block_metap, size_class,
+ SLOT_MADVISED);
}
// No current allocation block, or we were unable to allocate. We need to
@@ -2706,7 +2939,7 @@
_malloc_lock_s *lock = &nanozone->current_block_lock[size_class][allocation_index];
_malloc_lock_lock(lock);
- block_metap = os_atomic_load(block_metapp, relaxed);
+ nanov2_block_meta_t *block_metap = os_atomic_load(block_metapp, relaxed);
if (block_metap) {
ptr = nanov2_allocate_from_block(nanozone, block_metap, size_class);
if (ptr) {
@@ -2728,21 +2961,43 @@
// We could not find a block to allocate from -- make future
// allocations for this size class go to the helper zone until
// we have enough free space.
- _malloc_lock_lock(&nanozone->delegate_allocations_lock);
- nanozone->delegate_allocations |= 1 << size_class;
- _malloc_lock_unlock(&nanozone->delegate_allocations_lock);
+ os_atomic_or(&nanozone->delegate_allocations,
+ (uint16_t)(1 << size_class), relaxed);
+
+ ptr = nanozone->helper_zone->malloc(nanozone->helper_zone, rounded_size);
}
done:
- if (ptr) {
- if (clear) {
- memset(ptr, '\0', rounded_size);
- } else {
+ if (os_likely(ptr)) {
+ switch (malloc_zero_policy) {
+ case MALLOC_ZERO_ON_FREE: {
// Always clear the double-free guard so that we can recognize that
// this block is not on the free list.
nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
os_atomic_store(&slotp->double_free_guard, 0, relaxed);
- }
+
+ // We know the body of the allocation is already clear, so we just
+ // need to clean up the next_slot word to get to all-zero. Do so in
+ // all cases, even if a cleared allocation is not requested, to
+ // prevent any leakage through the next_slot bits.
+ os_atomic_store(&slotp->next_slot, 0, relaxed);
+ break;
+ }
+ case MALLOC_ZERO_NONE:
+ if (!clear) {
+ // Always clear the double-free guard so that we can recognize that
+ // this block is not on the free list.
+ nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
+ os_atomic_store(&slotp->double_free_guard, 0, relaxed);
+ break;
+ }
+ // fall through
+ case MALLOC_ZERO_ON_ALLOC:
+ memset(ptr, '\0', rounded_size);
+ break;
+ }
+ } else {
+ malloc_set_errno_fast(MZ_POSIX, ENOMEM);
}
return ptr;
}
@@ -2751,20 +3006,32 @@
#pragma mark Freeing
// Frees an allocation to its owning block and updates the block's state.
-// If the block becomes empty, it is marked as SLOT_CAN_MADVISE and is
-// madvised immediately if the policy is NANO_MADVISE_IMMEDIATE.
-void
-nanov2_free_to_block(nanozonev2_t *nanozone, void *ptr,
- nanov2_size_class_t size_class)
+//
+// If the block becomes empty, it is marked as SLOT_CAN_MADVISE and we return
+// the block to the caller to madvise if dictated by policy.
+MALLOC_ALWAYS_INLINE MALLOC_INLINE
+nanov2_block_meta_t *
+nanov2_free_to_block_inline(nanozonev2_t *nanozone, void *ptr,
+ nanov2_size_class_t size_class, nanov2_block_meta_t *block_metap)
{
nanov2_block_t *blockp = nanov2_block_address_for_ptr(ptr);
- nanov2_block_meta_t *block_metap = nanov2_meta_ptr_for_ptr(nanozone, ptr);
+ if (!block_metap) {
+ block_metap = nanov2_meta_ptr_for_ptr(nanozone, ptr);
+ }
// Release the slot memory onto the block's freelist.
nanov2_block_meta_t old_meta = os_atomic_load(block_metap, relaxed);
int slot_count = slots_by_size_class[size_class];
nanov2_block_meta_t new_meta;
boolean_t was_full;
+
+ nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
+ // All of the free slot content (double_free_guard, next_slot word and the
+ // zeroed remainder of the slot) must be visible when the os_atomic_cmpxchgv
+ // completes, so the metadata updates on either path below need a release
+ // barrier.
+ os_atomic_store(&slotp->double_free_guard,
+ nanozone->slot_freelist_cookie ^ (uintptr_t)ptr, relaxed);
again:
was_full = old_meta.next_slot == SLOT_FULL;
@@ -2774,12 +3041,14 @@
boolean_t freeing_last_active_slot = !was_full &&
new_meta.free_count == slots_by_size_class[size_class] - 1;
if (freeing_last_active_slot) {
+ os_atomic_store(&slotp->next_slot, SLOT_NULL, relaxed);
+
// Releasing the last active slot onto the free list. Mark the block as
// ready to be madvised if it's not in use, otherwise reset next_slot
// to SLOT_BUMP.
new_meta.next_slot = new_meta.in_use ? SLOT_BUMP : SLOT_CAN_MADVISE;
// Write the updated meta data; try again if we raced with another thread.
- if (!os_atomic_cmpxchgv(block_metap, old_meta, new_meta, &old_meta, relaxed)) {
+ if (!os_atomic_cmpxchgv(block_metap, old_meta, new_meta, &old_meta, release)) {
goto again;
}
@@ -2787,20 +3056,14 @@
// is to do so immediately.
if (new_meta.next_slot == SLOT_CAN_MADVISE &&
nanov2_madvise_policy == NANO_MADVISE_IMMEDIATE) {
- _malloc_lock_lock(&nanozone->madvise_lock);
- nanov2_madvise_block(nanozone, block_metap, blockp, size_class);
- _malloc_lock_unlock(&nanozone->madvise_lock);
+ return block_metap;
}
} else {
int slot_index = nanov2_slot_index_in_block(blockp, size_class, ptr);
new_meta.next_slot = slot_index + 1; // meta.next_slot is 1-based
- nanov2_free_slot_t *slotp = (nanov2_free_slot_t *)ptr;
- slotp->next_slot = was_full ? SLOT_BUMP : old_meta.next_slot;
- os_atomic_store(&slotp->double_free_guard,
- nanozone->slot_freelist_cookie ^ (uintptr_t)ptr, relaxed);
-
- // The double_free_guard change must be visible when the os_atomic_cmpxchgv
- // completes.
+ os_atomic_store(&slotp->next_slot,
+ was_full ? SLOT_BUMP : old_meta.next_slot, relaxed);
+
// Write the updated meta data; try again if we raced with another thread.
if (!os_atomic_cmpxchgv(block_metap, old_meta, new_meta, &old_meta, release)) {
goto again;
@@ -2814,14 +3077,14 @@
uint16_t class_mask = 1 << size_class;
if (!new_meta.in_use && (nanozone->delegate_allocations & class_mask) &&
(new_meta.free_count >= 0.75 * slot_count)) {
- _malloc_lock_lock(&nanozone->delegate_allocations_lock);
- nanozone->delegate_allocations &= ~class_mask;
- _malloc_lock_unlock(&nanozone->delegate_allocations_lock);
+ os_atomic_and(&nanozone->delegate_allocations, ~class_mask, relaxed);
}
#if DEBUG_MALLOC
nanozone->statistics.size_class_statistics[size_class].total_frees++;
#endif // DEBUG_MALLOC
+
+ return NULL;
}
#endif // OS_VARIANT_RESOLVED
@@ -2834,11 +3097,9 @@
malloc_zone_t *
nanov2_create_zone(malloc_zone_t *helper_zone, unsigned debug_flags)
{
- // Note: It is important that nanov2_create_zone resets _malloc_engaged_nano
- // if it is unable to enable the nanozone (and chooses not to abort). As
- // several functions rely on _malloc_engaged_nano to determine if they
- // should manipulate the nanozone, and these should not run if we failed
- // to create the zone.
+ // Note: It is not necessary that nanov2_create_zone resets _malloc_engaged_nano
+ // if it is unable to enable the nanozone - functions that need to determine
+ // whether the nanozone is preset should test initial_nano_zone.
MALLOC_ASSERT(_malloc_engaged_nano == NANO_V2);
// Get memory for the zone and disable Nano if we fail.
@@ -2850,9 +3111,14 @@
}
// Set up the basic_zone portion of the nanozonev2 structure
- nanozone->basic_zone.version = 12;
+ nanozone->basic_zone.version = 13;
nanozone->basic_zone.size = OS_RESOLVED_VARIANT_ADDR(nanov2_size);
- nanozone->basic_zone.malloc = OS_RESOLVED_VARIANT_ADDR(nanov2_malloc);
+ if (malloc_zero_policy == MALLOC_ZERO_ON_ALLOC) {
+ nanozone->basic_zone.malloc =
+ OS_RESOLVED_VARIANT_ADDR(nanov2_malloc_zero_on_alloc);
+ } else {
+ nanozone->basic_zone.malloc = OS_RESOLVED_VARIANT_ADDR(nanov2_malloc);
+ }
nanozone->basic_zone.calloc = OS_RESOLVED_VARIANT_ADDR(nanov2_calloc);
nanozone->basic_zone.valloc = (void *)nanov2_valloc;
nanozone->basic_zone.free = OS_RESOLVED_VARIANT_ADDR(nanov2_free);
@@ -2862,10 +3128,11 @@
nanozone->basic_zone.batch_free = OS_RESOLVED_VARIANT_ADDR(nanov2_batch_free);
nanozone->basic_zone.introspect =
(struct malloc_introspection_t *)&nanov2_introspect;
- nanozone->basic_zone.memalign = (void *)nanov2_memalign;
+ nanozone->basic_zone.memalign = OS_RESOLVED_VARIANT_ADDR(nanov2_memalign);
nanozone->basic_zone.free_definite_size = OS_RESOLVED_VARIANT_ADDR(nanov2_free_definite_size);
nanozone->basic_zone.pressure_relief = OS_RESOLVED_VARIANT_ADDR(nanov2_pressure_relief);
nanozone->basic_zone.claimed_address = OS_RESOLVED_VARIANT_ADDR(nanov2_claimed_address);
+ nanozone->basic_zone.try_free_default = OS_RESOLVED_VARIANT_ADDR(nanov2_try_free_default);
// Set these both to zero as required by CFAllocator.
nanozone->basic_zone.reserved1 = 0;
@@ -2904,14 +3171,30 @@
_malloc_lock_init(&nanozone->madvise_lock);
// Allocate the initial region. If this does not succeed, we disable Nano.
- nanov2_addr_t p = {.fields.nano_signature = NANOZONE_SIGNATURE};
- nanov2_region_t *region = (nanov2_region_t *)p.addr;
- boolean_t result = nanov2_allocate_region(region);
+ nanov2_region_t *region = (nanov2_region_t *)NANOZONE_BASE_REGION_ADDRESS;
+
+ bool result;
+#if CONFIG_NANO_RESERVE_REGIONS
+ unsigned int num_regions = (nano_max_region + 1);
+ result = nanov2_reserve_regions(region, num_regions);
+ if (result) {
+ result = nanov2_unprotect_region(region);
+ if (!result) {
+ malloc_report(ASL_LEVEL_ERR,
+ "unable to protect initial region\n");
+ nano_common_deallocate_pages((void *)region,
+ num_regions * (size_t)NANOV2_REGION_SIZE, 0);
+ }
+ }
+#else // CONFIG_NANO_RESERVE_REGIONS
+ result = nanov2_allocate_region(region);
+#endif // CONFIG_NANO_RESERVE_REGIONS
if (!result) {
- nano_common_deallocate_pages(nanozone, NANOZONEV2_ZONE_PAGED_SIZE, 0);
+ nano_common_deallocate_pages((void *)nanozone,
+ NANOZONEV2_ZONE_PAGED_SIZE, 0);
_malloc_engaged_nano = NANO_NONE;
malloc_report(ASL_LEVEL_NOTICE, "nano zone abandoned due to inability "
- "to preallocate reserved vm space.\n");
+ "to reserve vm space.\n");
return NULL;
}
nanov2_region_linkage_t *region_linkage =
@@ -2923,7 +3206,6 @@
os_atomic_store(&nanozone->current_region_next_arena,
((nanov2_arena_t *)region) + 1, release);
nanozone->statistics.allocated_regions = 1;
-
// Set up the guard blocks for the initial arena, if requested
nanov2_init_guard_blocks(nanozone, (nanov2_arena_t *)region);
@@ -2957,6 +3239,14 @@
{
// Just hand to the helper zone.
return nanozone->helper_zone->calloc(nanozone->helper_zone, num_items,
+ size);
+}
+
+static void *
+nanov2_forked_memalign(nanozonev2_t *nanozone, size_t alignment, size_t size)
+{
+ // Just hand to the helper zone.
+ return nanozone->helper_zone->memalign(nanozone->helper_zone, alignment,
size);
}
@@ -3098,10 +3388,11 @@
nanozone->basic_zone.batch_free = OS_RESOLVED_VARIANT_ADDR(nanov2_forked_batch_free);
nanozone->basic_zone.introspect =
(struct malloc_introspection_t *)&nanov2_introspect;// Unchanged
- nanozone->basic_zone.memalign = (void *)nanov2_memalign; // Unchanged
+ nanozone->basic_zone.memalign = (void *)nanov2_forked_memalign;
nanozone->basic_zone.free_definite_size =
OS_RESOLVED_VARIANT_ADDR(nanov2_forked_free_definite_size);
nanozone->basic_zone.claimed_address = nanov2_forked_claimed_address;
+ nanozone->basic_zone.try_free_default = NULL; // Fall back to old protocol
mprotect(nanozone, sizeof(nanozone->basic_zone), PROT_READ);
}