Leveldb source code analysis 2.1-the core design of the memory pool implementation

introduction

In order to use memory more reasonably, a custom memory pool is implemented in leveldb.

1 data structure

Leveldb's memory pool implementation class is Arena, and its core data fields are as follows:

First, use a vector type of blocks_ to store the allocated blocks of memory; use alloc_ptr_ to point to the first address of the unused memory in the newly applied memory; use alloc_bytes_remaining_ Indicates the byte size of unused memory in the newly applied memory; in particular, memory_usage_ counts the sum of memory occupied by all bock and blocks_ variables.

2 Source code interpretation

2.1 Source code interpretation

There are Allocate and AllocateAligned methods to apply for memory, the latter has more start address alignment function.

1. AllocateAligned function

char* Arena::AllocateAligned(size_t bytes) {
    
    
  //进行最低8字节的对齐
  const int align = (sizeof(void*) > 8) ? sizeof(void*) : 8;
  static_assert((align & (align - 1)) == 0,
                "Pointer size should be a power of 2");
  size_t current_mod = reinterpret_cast<uintptr_t>(alloc_ptr_) & (align - 1);
  size_t slop = (current_mod == 0 ? 0 : align - current_mod);
  size_t needed = bytes + slop;
  char* result;
  if (needed <= alloc_bytes_remaining_) {
    
      //如果最新内存块有空余空间
    result = alloc_ptr_ + slop; //移动对齐补全字节
    alloc_ptr_ += needed; //重置未使用内存指针
    alloc_bytes_remaining_ -= needed; //重置未使用内存大小
  } else {
    
      //如果最新内存块没有空余空间
    result = AllocateFallback(bytes);
  }
  assert((reinterpret_cast<uintptr_t>(result) & (align - 1)) == 0);
  return result;
}

2. AllocateFallback function

char* Arena::AllocateFallback(size_t bytes) {
    
    
  if (bytes > kBlockSize / 4) {
    
    
    // 如果申请内存大小大于1K，那么进入下面，申请bytes大小然后直接使用
    char* result = AllocateNewBlock(bytes);
    return result;
  }

  // 如果申请内存大小大于1K,直接申请4K大内存后，还需要重置alloc_ptr_和alloc_bytes_remaining_
  alloc_ptr_ = AllocateNewBlock(kBlockSize);
  alloc_bytes_remaining_ = kBlockSize;

  char* result = alloc_ptr_;
  alloc_ptr_ += bytes;
  alloc_bytes_remaining_ -= bytes;
  return result;
}

3. AllocateNewBlock function

char* Arena::AllocateNewBlock(size_t block_bytes) {
    
    
  char* result = new char[block_bytes];
  blocks_.push_back(result);
  // 加sizeof(char*)是因为blocks_还占用一个指针大小的存储空间
  memory_usage_.fetch_add(block_bytes + sizeof(char*),
                          std::memory_order_relaxed);
  return result;
}

2.2 Found the problem

By reading the Arena source code, memory will be wasted in the following scenarios:

1. First, call AllocateAligned to apply for 512Byte memory, then a large 4K memory will be applied, and the remaining 3.5K memory can be allocated;
2. Then, call AllocateAligned to apply for 512Byte memory 6 times, and finally the remaining 512Byte memory can be allocated;
3. Continue to apply for 600Byte of memory. At this time, a new piece of 4K of large memory will be new. At this time, alloc_ptr_ and alloc_bytes_remaining_ record the free space of this 4K new large memory. The remaining 512Bytes in the previous 4K will be wasted .