Sweden-Number/documentation/address-space.sgml

<chapter id="address-space">
  <title> Address space management </title>

  <para>
    Every Win32 process in Wine has its own dedicated native process on the host system, and
    therefore its own address space. This section explores the layout of the Windows address space
    and how it is emulated.
  </para>

  <para>
    Firstly, a quick recap of how virtual memory works. Physical memory in RAM chips is split
    into <emphasis>frames</emphasis>, and the memory that each process sees is split
    into <emphasis>pages</emphasis>. Each process has its own 4 gigabytes of address space (4gig
    being the maximum space addressable with a 32 bit pointer). Pages can be mapped or unmapped:
    attempts to access an unmapped page cause an EXCEPTION_ACCESS_VIOLATION which has the
    easily recognizable code of 0xC0000005.  Any page can be mapped to any frame, therefore you can
    have multiple addresses which actually "contain" the same memory. Pages can also be mapped to
    things like files or swap space, in which case accessing that page will cause a disk access to
    read the contents into a free frame.
  </para>

  <sect1>
    <title>Initial layout</title>
  
    <para>
      When a Win32 process starts, it does not have a clear address space to use as it pleases. Many pages
      are already mapped by the operating system. In particular, the EXE file itself and any DLLs it
      needs are mapped into memory, and space has been reserved for the stack and a couple of heaps
      (zones used to allocate memory to the app from). Some of these things need to be at a fixed
      address, and others can be placed anywhere.
    </para>

    <para>
      The EXE file itself is usually mapped at address 0x400000 and up: indeed, most EXEs have
      their relocation records stripped which means they must be loaded at their base address and
      cannot be loaded at any other address.
    </para>

    <para>
      DLLs are internally much the same as EXE files but they have relocation records, which means
      that they can be mapped at any address in the address space. Remember we are not dealing with
      physical memory here, but rather virtual memory which is different for each
      process. Therefore OLEAUT32.DLL may be loaded at one address in one process, and a totally
      different one in another. Ensuring all the functions loaded into memory can find each other
      is the job of the Windows dynamic linker, which is a part of NTDLL.
    </para>

    <para>
      So, we have the EXE and its DLLs mapped into memory. Two other very important regions also
      exist: the stack and the process heap. The process heap is simply the equivalent of the libc
      malloc arena on UNIX: it's a region of memory managed by the OS which malloc/HeapAlloc
      partitions and hands out to the application. Windows applications can create several heaps but
      the process heap always exists. It's created as part of process initialization in
      dlls/ntdll/thread.c:thread_init().
    </para>

    <para>
      There is another heap created as part of process startup, the so-called shared or system
      heap. This is an undocumented service that exists only on Windows 9x: it is implemented in
      Wine so native win9x DLLs can be used. The shared heap is unusual in that anything allocated
      from it will be visible in every other process. This heap is always created at the
      SYSTEM_HEAP_BASE address or 0x80000000 and defaults to 16 megabytes in size.
    </para>

    <para>
      So far we've assumed the entire 4 gigs of address space is available for the application. In
      fact that's not so: only the lower 2 gigs are available, the upper 2 gigs are on Windows NT
      used by the operating system and hold the kernel (from 0x80000000). Why is the kernel mapped
      into every address space?  Mostly for performance: while it's possible to give the kernel its
      own address space too - this is what Ingo Molnars 4G/4G VM split patch does for Linux - it
      requires that every system call into the kernel switches address space. As that is a fairly
      expensive operation (requires flushing the translation lookaside buffers etc) and syscalls are
      made frequently it's best avoided by keeping the kernel mapped at a constant position in every
      processes address space.
    </para>

    <para>
      On Windows 9x, in fact only the upper gigabyte (0xC0000000 and up) is used by the kernel, the
      region from 2 to 3 gigs is a shared area used for loading system DLLs and for file
      mappings. The bottom 2 gigs on both NT and 9x are available for the programs memory allocation
      and stack.
    </para>

    <para>
      There are a few other magic locations. The bottom 64k of memory is deliberately left unmapped
      to catch null pointer dereferences. The region from 64k to 1mb+64k are reserved for DOS
      compatibility and contain various DOS data structures. Finally, the address space also
      contains mappings for the Wine binary itself, any native libaries Wine is using, the glibc
      malloc arena and so on.
    </para>
    
  </sect1>

  <sect1>
    <title> Laying out the address space </title>

    <para>
      Up until about the start of 2004, the Linux address space very much resembled the Windows 9x
      layout: the kernel sat in the top gigabyte, the bottom pages were unmapped to catch null
      pointer dereferences, and the rest was free. The kernels mmap algorithm was predictable: it
      would start by mapping files at low addresses and work up from there.
    </para>

    <para>
      The development of a series of new low level patches violated many of these assumptions, and
      resulted in Wine needing to force the Win32 address space layout upon the system. This
      section looks at why and how this is done.
    </para>

    <para>
      The exec-shield patch increases security by randomizing the kernels mmap algorithms. Rather
      than consistently choosing the same addresses given the same sequence of requests, the kernel
      will now choose randomized addresses. Because the Linux dynamic linker (ld-linux.so.2) loads
      DSOs into memory by using mmap, this means that DSOs are no longer loaded at predictable
      addresses, so making it harder to attack software by using buffer overflows. It also attempts
      to relocate certain binaries into a special low area of memory known as the ASCII armor so
      making it harder to jump into them when using string based attacks.
    </para>

    <para>
      Prelink is a technology that enhances startup times by precalculating ELF global offset
      tables then saving the results inside the native binaries themselves. By grid fitting each
      DSO into the address space, the dynamic linker does not have to perform as many relocations
      so allowing applications that heavily rely on dynamic linkage to be loaded into memory much
      quicker. Complex C++ applications such as Mozilla, OpenOffice and KDE can especially benefit
      from this technique.
    </para>

    <para>
      The 4G VM split patch was developed by Ingo Molnar. It gives the Linux kernel its own address
      space, thereby allowing processes to access the maximum addressable amount of memory on a
      32-bit machine: 4 gigabytes. It allows people with lots of RAM to fully utilise that in any
      given process at the cost of performance: as mentioned previously the reason behind giving
      the kernel a part of each processes address space was to avoid the overhead of switching on
      each syscall.
    </para>

    <para>
      Each of these changes alter the address space in a way incompatible with Windows. Prelink and
      exec-shield mean that the libraries Wine uses can be placed at any point in the address
      space: typically this meant that a library was sitting in the region that the EXE you wanted
      to run had to be loaded (remember that unlike DLLs, EXE files cannot be moved around in
      memory). The 4G VM split means that programs could receive pointers to the top gigabyte of
      address space which some are not prepared for (they may store extra information in the high
      bits of a pointer, for instance). In particular, in combination with exec-shield this one is
      especially deadly as it's possible the process heap could be allocated beyond
      ADDRESS_SPACE_LIMIT which causes Wine initialization to fail. 
    </para>

    <para>
      The solution to these problems is for Wine to reserve particular parts of the address space
      so that areas that we don't want the system to use will be avoided. We later on
      (re/de)allocate those areas as needed. One problem is that some of these mappings are put in
      place automatically by the dynamic linker: for instance any libraries that Wine
      is linked to (like libc, libwine, libpthread etc) will be mapped into memory before Wine even
      gets control. In order to solve that, Wine overrides the default ELF initialization sequence
      at a low level and reserves the needed areas by using direct syscalls into the kernel (ie
      without linking against any other code to do it) before restarting the standard
      initialization and letting the dynamic linker continue. This is referred to as the
      preloader and is found in loader/preloader.c.
    </para>

    <para>
      Once the usual ELF boot sequence has been completed, some native libraries may well have been
      mapped above the 3gig limit: however, this doesn't matter as 3G is a Windows limit, not a
      Linux limit. We still have to prevent the system from allocating anything else above there
      (like the heap or other DLLs) though so Wine performs a binary search over the upper gig of
      address space in order to iteratively fill in the holes with MAP_NORESERVE mappings so the
      address space is allocated but the memory to actually back it is not. This code can be found
      in libs/wine/mmap.c:reserve_area.
    </para>
    
  </sect1>
  
</chapter>