chromium/v8/src/heap/free-list.h

// Copyright 2020 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.

#ifndef V8_HEAP_FREE_LIST_H_
#define V8_HEAP_FREE_LIST_H_

#include <atomic>

#include "src/base/macros.h"
#include "src/common/globals.h"
#include "src/heap/allocation-result.h"
#include "src/heap/mutable-page-metadata.h"
#include "src/objects/free-space.h"
#include "src/objects/map.h"
#include "src/utils/utils.h"
#include "testing/gtest/include/gtest/gtest_prod.h"  // nogncheck

namespace v8 {
namespace internal {

namespace heap {
class HeapTester;
class TestCodePageAllocatorScope;
}  // namespace heap

class AllocationObserver;
class FreeList;
class Isolate;
class LargeObjectSpace;
class LargePageMetadata;
class LinearAllocationArea;
class PageMetadata;
class PagedSpace;
class SemiSpace;

FreeListCategoryType;

static constexpr FreeListCategoryType kFirstCategory = …;
static constexpr FreeListCategoryType kInvalidCategory = …;

enum FreeMode { … };

// A free list category maintains a linked list of free memory blocks.
class FreeListCategory { … };

// A free list maintains free blocks of memory. The free list is organized in
// a way to encourage objects allocated around the same time to be near each
// other. The normal way to allocate is intended to be by bumping a 'top'
// pointer until it hits a 'limit' pointer.  When the limit is hit we need to
// find a new space to allocate from. This is done with the free list, which is
// divided up into rough categories to cut down on waste. Having finer
// categories would scatter allocation more.
class FreeList { … };

// Use 24 Freelists: on per 16 bytes between 24 and 256, and then a few ones for
// larger sizes. See the variable |categories_min| for the size of each
// Freelist.  Allocation is done using a best-fit strategy (considering only the
// first element of each category though).
// Performances are expected to be worst than FreeListLegacy, but memory
// consumption should be lower (since fragmentation should be lower).
class V8_EXPORT_PRIVATE FreeListMany : public FreeList { … };

// Same as FreeListMany but uses a cache to know which categories are empty.
// The cache (|next_nonempty_category|) is maintained in a way such that for
// each category c, next_nonempty_category[c] contains the first non-empty
// category greater or equal to c, that may hold an object of size c.
// Allocation is done using the same strategy as FreeListMany (ie, best fit).
class V8_EXPORT_PRIVATE FreeListManyCached : public FreeListMany { … };

// Same as FreeListManyCached but uses a fast path.
// The fast path overallocates by at least 1.85k bytes. The idea of this 1.85k
// is: we want the fast path to always overallocate, even for larger
// categories. Therefore, we have two choices: either overallocate by
// "size_in_bytes * something" or overallocate by "size_in_bytes +
// something". We choose the later, as the former will tend to overallocate too
// much for larger objects. The 1.85k (= 2048 - 128) has been chosen such that
// for tiny objects (size <= 128 bytes), the first category considered is the
// 36th (which holds objects of 2k to 3k), while for larger objects, the first
// category considered will be one that guarantees a 1.85k+ bytes
// overallocation. Using 2k rather than 1.85k would have resulted in either a
// more complex logic for SelectFastAllocationFreeListCategoryType, or the 36th
// category (2k to 3k) not being used; both of which are undesirable.
// A secondary fast path is used for tiny objects (size <= 128), in order to
// consider categories from 256 to 2048 bytes for them.
// Note that this class uses a precise GetPageForSize (inherited from
// FreeListMany), which makes its fast path less fast in the Scavenger. This is
// done on purpose, since this class's only purpose is to be used by
// FreeListManyCachedOrigin, which is precise for the scavenger.
class V8_EXPORT_PRIVATE FreeListManyCachedFastPathBase
    : public FreeListManyCached { … };

class FreeListManyCachedFastPath : public FreeListManyCachedFastPathBase { … };

class FreeListManyCachedFastPathForNewSpace
    : public FreeListManyCachedFastPathBase { … };

// Uses FreeListManyCached if in the GC; FreeListManyCachedFastPath otherwise.
// The reasoning behind this FreeList is the following: the GC runs in
// parallel, and therefore, more expensive allocations there are less
// noticeable. On the other hand, the generated code and runtime need to be very
// fast. Therefore, the strategy for the former is one that is not very
// efficient, but reduces fragmentation (FreeListManyCached), while the strategy
// for the later is one that is very efficient, but introduces some
// fragmentation (FreeListManyCachedFastPath).
class V8_EXPORT_PRIVATE FreeListManyCachedOrigin
    : public FreeListManyCachedFastPath { … };

}  // namespace internal
}  // namespace v8

#endif  // V8_HEAP_FREE_LIST_H_