我發佈了關於我們實際完成的更新。我最終實現了自己的遠程內存分配器(下面的源代碼)。它在靈性上類似於answer Sam suggests,但使用boost侵入式RB樹來避免在釋放,加入等時進行一些日誌(N)查找。它是線程安全的並支持各種遠程指針/偏移量類型作爲模板參數。這在很多方面可能都不是很理想,但對於我們需要做的事情來說,它已經足夠好了。如果你發現錯誤,請告訴我。
/*
* Thread-safe remote memory allocator
*
* Author: Yuriy Romanenko
* Copyright (c) 2015 Lytro, Inc.
*
*/
#pragma once
#include <memory>
#include <mutex>
#include <cstdint>
#include <cstdio>
#include <functional>
#include <boost/intrusive/rbtree.hpp>
namespace bi = boost::intrusive;
template<typename remote_ptr_t = void*,
typename remote_size_t = size_t,
typename remote_uintptr_t = uintptr_t>
class RemoteAllocator
{
/* Internal structure used for keeping track of a contiguous block of
* remote memory. It can be on one or two of the following RB trees:
* Free Chunks (sorted by size)
* All Chunks (sorted by remote pointer)
*/
struct Chunk
{
bi::set_member_hook<> mRbFreeChunksHook;
bi::set_member_hook<> mRbAllChunksHook;
remote_uintptr_t mOffset;
remote_size_t mSize;
bool mFree;
Chunk(remote_uintptr_t off, remote_size_t sz, bool fr)
: mOffset(off), mSize(sz), mFree(fr)
{
}
bool contains(remote_uintptr_t off)
{
return (off >= mOffset) && (off < mOffset + mSize);
}
private:
Chunk(const Chunk&);
Chunk& operator=(const Chunk&);
};
struct ChunkCompareSize : public std::binary_function <Chunk,Chunk,bool>
{
bool operator() (const Chunk& x, const Chunk& y) const
{
return x.mSize < y.mSize;
}
};
struct ChunkCompareOffset : public std::binary_function <Chunk,Chunk,bool>
{
bool operator() (const Chunk& x, const Chunk& y) const
{
return x.mOffset < y.mOffset;
}
};
typedef bi::rbtree<Chunk,
bi::member_hook<Chunk,
bi::set_member_hook<>,
&Chunk::mRbFreeChunksHook>,
bi::compare<ChunkCompareSize> > FreeChunkTree;
typedef bi::rbtree<Chunk,
bi::member_hook<Chunk,
bi::set_member_hook<>,
&Chunk::mRbAllChunksHook>,
bi::compare<ChunkCompareOffset> > AllChunkTree;
// Thread safety lock
std::mutex mLock;
// Size of the entire pool
remote_size_t mSize;
// Start address of the pool
remote_ptr_t mStartAddr;
// Tree of free chunks
FreeChunkTree mFreeChunks;
// Tree of all chunks
AllChunkTree mAllChunks;
// This removes the chunk from both trees
Chunk *unlinkChunk(Chunk *c)
{
mAllChunks.erase(mAllChunks.iterator_to(*c));
if(c->mFree)
{
mFreeChunks.erase(mFreeChunks.iterator_to(*c));
}
return c;
}
// This reinserts the chunk into one or two trees, depending on mFree
Chunk *relinkChunk(Chunk *c)
{
mAllChunks.insert_equal(*c);
if(c->mFree)
{
mFreeChunks.insert_equal(*c);
}
return c;
}
/* This assumes c is 'free' and walks the mAllChunks tree to the left
* joining any contiguous free chunks into this one
*/
bool growFreeLeft(Chunk *c)
{
auto it = mAllChunks.iterator_to(*c);
if(it != mAllChunks.begin())
{
it--;
if(it->mFree)
{
Chunk *left = unlinkChunk(&(*it));
unlinkChunk(c);
c->mOffset = left->mOffset;
c->mSize = left->mSize + c->mSize;
delete left;
relinkChunk(c);
return true;
}
}
return false;
}
/* This assumes c is 'free' and walks the mAllChunks tree to the right
* joining any contiguous free chunks into this one
*/
bool growFreeRight(Chunk *c)
{
auto it = mAllChunks.iterator_to(*c);
it++;
if(it != mAllChunks.end())
{
if(it->mFree)
{
Chunk *right = unlinkChunk(&(*it));
unlinkChunk(c);
c->mSize = right->mSize + c->mSize;
delete right;
relinkChunk(c);
return true;
}
}
return false;
}
public:
RemoteAllocator(remote_size_t size, remote_ptr_t startAddr) :
mSize(size), mStartAddr(startAddr)
{
/* Initially we create one free chunk the size of the entire managed
* memory pool, and add it to both trees
*/
Chunk *all = new Chunk(reinterpret_cast<remote_uintptr_t>(mStartAddr),
mSize,
true);
mAllChunks.insert_equal(*all);
mFreeChunks.insert_equal(*all);
}
~RemoteAllocator()
{
auto it = mAllChunks.begin();
while(it != mAllChunks.end())
{
Chunk *pt = unlinkChunk(&(*it++));
delete pt;
}
}
remote_ptr_t malloc(remote_size_t bytes)
{
std::unique_lock<std::mutex> lock(mLock);
auto fit = mFreeChunks.lower_bound(
Chunk(reinterpret_cast<remote_uintptr_t>(mStartAddr),
bytes,
true));
/* Out of memory */
if(fit == mFreeChunks.end())
return remote_ptr_t{0};
Chunk *ret = &(*fit);
/* We need to split the chunk because it's not the exact size */
/* Let's remove the node */
mFreeChunks.erase(fit);
if(ret->mSize != bytes)
{
Chunk *right, *left = ret;
/* The following logic decides which way the heap grows
* based on allocation size. I am not 100% sure this actually
* helps with fragmentation with such a big threshold (50%)
*
* Check if we will occupy more than half of the chunk,
* in that case, use the left side. */
if(bytes > ret->mSize/2)
{
right = new Chunk(left->mOffset + bytes,
left->mSize - bytes,
true);
relinkChunk(right);
left->mSize = bytes;
left->mFree = false;
ret = left;
}
/* We'll be using less than half, let's use the right side. */
else
{
right = new Chunk(left->mOffset + left->mSize - bytes,
bytes,
false);
relinkChunk(right);
left->mSize = left->mSize - bytes;
mFreeChunks.insert_equal(*left);
ret = right;
}
}
else
{
ret->mFree = false;
}
return reinterpret_cast<remote_ptr_t>(ret->mOffset);
}
remote_ptr_t malloc_aligned(remote_size_t bytes, remote_size_t alignment)
{
remote_size_t bufSize = bytes + alignment;
remote_ptr_t mem = this->malloc(bufSize);
remote_ptr_t ret = mem;
if(mem)
{
remote_uintptr_t offset = reinterpret_cast<remote_uintptr_t>(mem);
if(offset % alignment)
{
offset = offset + (alignment - (offset % alignment));
}
ret = reinterpret_cast<remote_ptr_t>(offset);
}
return ret;
}
void free(remote_ptr_t ptr)
{
std::unique_lock<std::mutex> lock(mLock);
Chunk ref(reinterpret_cast<remote_uintptr_t>(ptr), 0, false);
auto it = mAllChunks.find(ref);
if(it == mAllChunks.end())
{
it = mAllChunks.upper_bound(ref);
it--;
}
if(!(it->contains(ref.mOffset)) || it->mFree)
throw std::runtime_error("Could not find chunk to free");
Chunk *chnk = &(*it);
chnk->mFree = true;
mFreeChunks.insert_equal(*chnk);
/* Maximize space */
while(growFreeLeft(chnk));
while(growFreeRight(chnk));
}
void debugDump()
{
std::unique_lock<std::mutex> lock(mLock);
int i = 0;
printf("----------- All chunks -----------\n");
for(auto it = mAllChunks.begin(); it != mAllChunks.end(); it++)
{
printf(" [%d] %lu -> %lu (%lu) %s\n",
i++,
it->mOffset,
it->mOffset + it->mSize,
it->mSize,
it->mFree ? "(FREE)" : "(NOT FREE)");
}
i = 0;
printf("----------- Free chunks -----------\n");
for(auto it = mFreeChunks.begin(); it != mFreeChunks.end(); it++)
{
printf(" [%d] %lu -> %lu (%lu) %s\n",
i++,
it->mOffset,
it->mOffset + it->mSize,
it->mSize,
it->mFree ? "(FREE)" : "(NOT FREE)");
}
}
};
緩衝區有多大(或很小)?它可以*非常*大? –
具體而言,我希望能夠管理一個16GB的外部緩衝區,分配大小從128字節一直到512MB。那*非常大嗎? –
在什麼電腦上?在超高速計算機上有一個TB的RAM,它不會那麼大!你應該編輯你的問題,而不是在評論中添加信息。 –