How to dynamically expand a Memory Mapped File
Well, you can!!.
Here is my implementation of a growable memory mapped file:
using System;
using System.Collections.Generic;
using System.ComponentModel;
using System.IO;
using System.IO.MemoryMappedFiles;
namespace MmbpTree
{
public unsafe sealed class GrowableMemoryMappedFile : IDisposable
{
private const int AllocationGranularity = 64 * 1024;
private class MemoryMappedArea
{
public MemoryMappedFile Mmf;
public byte* Address;
public long Size;
}
private FileStream fs;
private List<MemoryMappedArea> areas = new List<MemoryMappedArea>();
private long[] offsets;
private byte*[] addresses;
public long Length
{
get {
CheckDisposed();
return fs.Length;
}
}
public GrowableMemoryMappedFile(string filePath, long initialFileSize)
{
if (initialFileSize <= 0 || initialFileSize % AllocationGranularity != 0)
{
throw new ArgumentException("The initial file size must be a multiple of 64Kb and grater than zero");
}
bool existingFile = File.Exists(filePath);
fs = new FileStream(filePath, FileMode.OpenOrCreate, FileAccess.ReadWrite, FileShare.None);
if (existingFile)
{
if (fs.Length <= 0 || fs.Length % AllocationGranularity != 0)
{
throw new ArgumentException("Invalid file. Its lenght must be a multiple of 64Kb and greater than zero");
}
}
else
{
fs.SetLength(initialFileSize);
}
CreateFirstArea();
}
private void CreateFirstArea()
{
var mmf = MemoryMappedFile.CreateFromFile(fs, null, fs.Length, MemoryMappedFileAccess.ReadWrite, null, HandleInheritability.None, true);
var address = Win32FileMapping.MapViewOfFileEx(mmf.SafeMemoryMappedFileHandle.DangerousGetHandle(),
Win32FileMapping.FileMapAccess.Read | Win32FileMapping.FileMapAccess.Write,
0, 0, new UIntPtr((ulong) fs.Length), null);
if (address == null) throw new Win32Exception();
var area = new MemoryMappedArea
{
Address = address,
Mmf = mmf,
Size = fs.Length
};
areas.Add(area);
addresses = new byte*[] { address };
offsets = new long[] { 0 };
}
public void Grow(long bytesToGrow)
{
CheckDisposed();
if (bytesToGrow <= 0 || bytesToGrow % AllocationGranularity != 0) {
throw new ArgumentException("The growth must be a multiple of 64Kb and greater than zero");
}
long offset = fs.Length;
fs.SetLength(fs.Length + bytesToGrow);
var mmf = MemoryMappedFile.CreateFromFile(fs, null, fs.Length, MemoryMappedFileAccess.ReadWrite, null, HandleInheritability.None, true);
uint* offsetPointer = (uint*)&offset;
var lastArea = areas[areas.Count - 1];
byte* desiredAddress = lastArea.Address + lastArea.Size;
var address = Win32FileMapping.MapViewOfFileEx(mmf.SafeMemoryMappedFileHandle.DangerousGetHandle(),
Win32FileMapping.FileMapAccess.Read | Win32FileMapping.FileMapAccess.Write,
offsetPointer[1], offsetPointer[0], new UIntPtr((ulong)bytesToGrow), desiredAddress);
if (address == null) {
address = Win32FileMapping.MapViewOfFileEx(mmf.SafeMemoryMappedFileHandle.DangerousGetHandle(),
Win32FileMapping.FileMapAccess.Read | Win32FileMapping.FileMapAccess.Write,
offsetPointer[1], offsetPointer[0], new UIntPtr((ulong)bytesToGrow), null);
}
if (address == null) throw new Win32Exception();
var area = new MemoryMappedArea {
Address = address,
Mmf = mmf,
Size = bytesToGrow
};
areas.Add(area);
if (desiredAddress != address) {
offsets = offsets.Add(offset);
addresses = addresses.Add(address);
}
}
public byte* GetPointer(long offset)
{
CheckDisposed();
int i = offsets.Length;
if (i <= 128) // linear search is more efficient for small arrays. Experiments show 140 as the cutpoint on x64 and 100 on x86.
{
while (--i > 0 && offsets[i] > offset);
}
else // binary search is more efficient for large arrays
{
i = Array.BinarySearch<long>(offsets, offset);
if (i < 0) i = ~i - 1;
}
return addresses[i] + offset - offsets[i];
}
private bool isDisposed;
public void Dispose()
{
if (isDisposed) return;
isDisposed = true;
foreach (var a in this.areas)
{
Win32FileMapping.UnmapViewOfFile(a.Address);
a.Mmf.Dispose();
}
fs.Dispose();
areas.Clear();
}
private void CheckDisposed()
{
if (isDisposed) throw new ObjectDisposedException(this.GetType().Name);
}
public void Flush()
{
CheckDisposed();
foreach (var area in areas)
{
if (!Win32FileMapping.FlushViewOfFile(area.Address, new IntPtr(area.Size))) {
throw new Win32Exception();
}
}
fs.Flush(true);
}
}
}
Here is the Win32FileMapping
class:
using System;
using System.Runtime.InteropServices;
namespace MmbpTree
{
public static unsafe class Win32FileMapping
{
[Flags]
public enum FileMapAccess : uint
{
Copy = 0x01,
Write = 0x02,
Read = 0x04,
AllAccess = 0x08,
Execute = 0x20,
}
[DllImport("kernel32.dll", SetLastError = true)]
public static extern byte* MapViewOfFileEx(IntPtr mappingHandle,
FileMapAccess access,
uint offsetHigh,
uint offsetLow,
UIntPtr bytesToMap,
byte* desiredAddress);
[DllImport("kernel32.dll", SetLastError = true)]
public static extern bool UnmapViewOfFile(byte* address);
[DllImport("kernel32.dll", SetLastError = true)]
[return: MarshalAs(UnmanagedType.Bool)]
public static extern bool FlushViewOfFile(byte* address, IntPtr bytesToFlush);
}
}
And here you have the Extensions
class:
using System;
namespace MmbpTree
{
public static class Extensions
{
public static T[] Add<T>(this T[] array, T element)
{
var result = new T[array.Length + 1];
Array.Copy(array, result, array.Length);
result[array.Length] = element;
return result;
}
public static unsafe byte*[] Add(this byte*[] array, byte* element)
{
var result = new byte*[array.Length + 1];
Array.Copy(array, result, array.Length);
result[array.Length] = element;
return result;
}
}
}
As you can see I take the unsafe approach. It's the only way to get the performance benefits of memory mapped files.
To work with this you need to consider the following concepts:
- The block or page. This is your minimal region of continuous memory address and storage space you work with. The size of a block or page must be a multiple of underlying system page size (4Kb).
- The initial file size. It must be a multiple of the block or page size and it must be a multiple of the system allocation granularity (64Kb).
- The file growth. It must be a multiple of the block or page size and it must be a multiple of the system allocation granularity (64Kb).
For example you may want to work with a page size of 1Mb, a file growth of 64Mb and an initial size of 1Gb. You can get a pointer to a page by calling GetPointer
, grow the file using Grow
and flush the file using Flush
:
const int InitialSize = 1024 * 1024 * 1024;
const int FileGrowth = 64 * 1024 * 1024;
const int PageSize = 1024 * 1024;
using (var gmmf = new GrowableMemoryMappedFile("mmf.bin", InitialSize))
{
var pageNumber = 32;
var pointer = gmmf.GetPointer(pageNumber * PageSize);
// you can read the page content:
byte firstPageByte = pointer[0];
byte lastPageByte = pointer[PageSize - 1];
// or write it
pointer[0] = 3;
pointer[PageSize -1] = 43;
/* allocate more pages when needed */
gmmf.Grow(FileGrowth);
/* use new allocated pages */
/* flushing the file writes to the underlying file */
gmmf.Flush();
}
The reason that the code does not compile is because it uses a non-existing overload. Either create a filestream yourself and pass it to the correct overload (assuming 2000 will be your new size):
FileStream fs = new FileStream("C:\MyFile.dat", FileMode.Open);
MemoryMappedFile mmf = MemoryMappedFile.CreateFromFile(fs, "someName", 2000,
MemoryMappedFileAccess.ReadWriteExecute, null, HandleInheritablity.None, false);
Or use this overload to skip the filstream creation:
MemoryMappedFile mmf = MemoryMappedFile.CreateFromFile("C:\MyFile.dat",
FileMode.Open, "someName", 2000);
Once you map a file in memory, you cannot increase its size. This is a known limitation of memory mapped files.
...you must calculate or estimate the size of the finished file because file mapping objects are static in size; once created, their size cannot be increased or decreased.
One strategy would be to use chunks stored in non-persisted memory mapped files of a given size, say 1GB or 2GB. You would manage these through a top level ViewAccessor
of your own design (probably doing basic passthru of the methods you need from the MemoryMappedViewAccessor
).
Edit: or you could just create a non-persisted memory mapped file of a maximal size you expect to use (say 8GB to start, with a parameter to tune it on start-up of your application) and retrieve MemoryMappedViewAccessor
's per logical chunk. The non-persisted file will not use physical resources until each view is requested.