blob: 697942a525e53567093f060b86c0a26bb13fd35a [file] [log] [blame]
// Copyright 2024 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/logging/counters.h"
#include "src/sandbox/compactible-external-entity-table.h"
#include "src/sandbox/external-entity-table-inl.h"
#include "src/sandbox/external-pointer.h"
namespace v8 {
namespace internal {
template <typename Entry, size_t size>
uint32_t CompactibleExternalEntityTable<Entry, size>::AllocateEntry(
Space* space) {
uint32_t index = Base::AllocateEntry(space);
// When we're compacting a space, we're trying to move all entries above a
// threshold index (the start of the evacuation area) into segments below
// that threshold. However, if the freelist becomes too short and we start
// allocating entries inside the area that is supposed to be evacuated, we
// need to abort compaction. This is not just an optimization but is also
// required for correctness: during sweeping we might otherwise assume that
// all entries inside the evacuation area have been moved and that these
// segments can therefore be deallocated. In particular, this check will also
// make sure that we abort compaction if we extend the space with a new
// segment and allocate at least one entry in it (if that segment is located
// after the threshold, otherwise it is unproblematic).
uint32_t start_of_evacuation_area =
if (V8_UNLIKELY(index >= start_of_evacuation_area)) {
return index;
template <typename Entry, size_t size>
typename CompactibleExternalEntityTable<Entry, size>::CompactionResult
CompactibleExternalEntityTable<Entry, size>::FinishCompaction(
Space* space, Histogram* counter) {
// When compacting, we can compute the number of unused segments at the end of
// the table and deallocate those after sweeping.
uint32_t start_of_evacuation_area =
bool evacuation_was_successful = false;
if (space->IsCompacting()) {
auto outcome = ExternalEntityTableCompactionOutcome::kAborted;
if (space->CompactingWasAborted()) {
// Compaction was aborted during marking because the freelist grew to
// short. In this case, it is not guaranteed that any segments will now be
// completely free. Extract the original start_of_evacuation_area value.
start_of_evacuation_area &= ~Space::kCompactionAbortedMarker;
} else {
// Entry evacuation was successful so all segments inside the evacuation
// area are now guaranteed to be free and so can be deallocated.
evacuation_was_successful = true;
outcome = ExternalEntityTableCompactionOutcome::kSuccess;
ExternalEntityTable<Entry, size>::kEntriesPerSegment));
return {start_of_evacuation_area, evacuation_was_successful};
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry, size>::MaybeCreateEvacuationEntry(
Space* space, uint32_t index, Address handle_location) {
// Check if the entry should be evacuated for table compaction.
// The current value of the start of the evacuation area is cached in a local
// variable here as it otherwise may be changed by another marking thread
// while this method runs, causing non-optimal behaviour (for example, the
// allocation of an evacuation entry _after_ the entry that is evacuated).
uint32_t start_of_evacuation_area =
if (index >= start_of_evacuation_area) {
uint32_t new_index =
Base::AllocateEntryBelow(space, start_of_evacuation_area);
if (new_index) {
DCHECK_LT(new_index, start_of_evacuation_area);
// Even though the new entry will only be accessed during sweeping, this
// still needs to be an atomic write as another thread may attempt (and
// fail) to allocate the same table entry, thereby causing a read from
// this memory location. Without an atomic store here, TSan would then
// complain about a data race.
} else {
// In this case, the application has allocated a sufficiently large
// number of entries from the freelist so that new entries would now be
// allocated inside the area that is being compacted. While it would be
// possible to shrink that area and continue compacting, we probably do
// not want to put more pressure on the freelist and so instead simply
// abort compaction here. Entries that have already been visited will
// still be compacted during Sweep, but there is no guarantee that any
// blocks at the end of the table will now be completely free.
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry, size>::Space::StartCompacting(
uint32_t start_of_evacuation_area) {
DCHECK_EQ(invalidated_fields_.size(), 0);,
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry, size>::Space::StopCompacting() {,
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry, size>::Space::AbortCompacting(
uint32_t start_of_evacuation_area) {
uint32_t compaction_aborted_marker =
start_of_evacuation_area | kCompactionAbortedMarker;
DCHECK_NE(compaction_aborted_marker, kNotCompactingMarker);,
template <typename Entry, size_t size>
bool CompactibleExternalEntityTable<Entry, size>::Space::IsCompacting() {
return start_of_evacuation_area_.load(std::memory_order_relaxed) !=
template <typename Entry, size_t size>
bool CompactibleExternalEntityTable<Entry,
size>::Space::CompactingWasAborted() {
auto value = start_of_evacuation_area_.load(std::memory_order_relaxed);
return (value & kCompactionAbortedMarker) == kCompactionAbortedMarker;
template <typename Entry, size_t size>
bool CompactibleExternalEntityTable<Entry, size>::Space::FieldWasInvalidated(
Address field_address) const {
return std::find(invalidated_fields_.begin(), invalidated_fields_.end(),
field_address) != invalidated_fields_.end();
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry,
size>::Space::ClearInvalidatedFields() {
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry, size>::Space::AddInvalidatedField(
Address field_address) {
if (IsCompacting()) {
base::MutexGuard guard(&invalidated_fields_mutex_);
template <typename Entry, size_t size>
void CompactibleExternalEntityTable<Entry,
size>::Space::StartCompactingIfNeeded() {
// Take the lock so that we can be sure that no other thread modifies the
// segments set concurrently.
base::MutexGuard guard(&this->mutex_);
// This method may be executed while other threads allocate entries from the
// freelist. In that case, this method may use incorrect data to determine if
// table compaction is necessary. That's fine however since in the worst
// case, compaction will simply be aborted right away if the freelist became
// too small.
uint32_t num_free_entries = this->freelist_length();
uint32_t num_total_entries = this->capacity();
// Current (somewhat arbitrary) heuristic: need compacting if the space is
// more than 1MB in size, is at least 10% empty, and if at least one segment
// can be freed after successful compaction.
double free_ratio = static_cast<double>(num_free_entries) /
uint32_t num_segments_to_evacuate =
(num_free_entries / 2) / Base::kEntriesPerSegment;
uint32_t space_size = num_total_entries * Base::kEntrySize;
bool should_compact = (space_size >= 1 * MB) && (free_ratio >= 0.10) &&
(num_segments_to_evacuate >= 1);
// However, if --stress-compaction is enabled, we compact whenever possible:
// whenever we have at least one segment worth of empty entries.
if (v8_flags.stress_compaction) {
should_compact = num_free_entries > Base::kEntriesPerSegment;
if (should_compact) {
// If we're compacting, attempt to free up the last N segments so that they
// can be decommitted afterwards.
auto first_segment_to_evacuate =
*std::prev(this->segments_.end(), num_segments_to_evacuate);
uint32_t start_of_evacuation_area = first_segment_to_evacuate.first_entry();
} // namespace internal
} // namespace v8