=============================================================================
Layout of linear memory:

   From       To         Size  Usage
   ---------------------------------------------------------------
   00000000 - 000fffff   1M    Dos memory (low 1M).
   00100000 - 003fffff   3M    VCPI server controlled.
   00400000 - 007fefff   4M-   DPMI server code and data.
   007ff000 - 007fffff   4K    Temporary page for `physical_poke'.
   00800000 - ffffffff   4G-   Dynamically allocated for clients.

=============================================================================
Layout of GDT:

See "Defines.Inc".  A selector of 0x40 is valid for clients for
compatibility with broken clients wanting access to bios data area.

=============================================================================
Layout of LDT:

Completely up to the client.

=============================================================================
Mode switching:

Routines _go32 and [_goreal] perform the switches.  These preserve as
many registers as possible and communicates segment register values on
the stack.  For details see "Dpmi.Asm".

=============================================================================
Client memory allocation:

Linear address space is allocated in 4M chunks.  For each chunk a page
table is also allocated.  Allocated page tables will be mapped into the
address space just before the client's memory.  In particular, for each
4M requested an extra 4K will be allocated.  Committed memory will always
be allocated for the page tables.  Note that since the first page table
maps itself into linear space some magic must set it up in the first
place.  An example with between 4M and 8M allocated:

        Page Dir             Page Tables           User Pages
                                  v------+
        +-------+             +-------+  |         +-------+
CR3---->|  1M   |-...  +----->|       |--+  +----->|       |
        +-------+      |      +-------+     |      +-------+
        | Server|-...  |      |       |--+  |  +-->|       |
        +-------+      |      +-------+  |  |  |   +-------+
        | 1st 4M|------+      |       |--|--+  |   |       |
        +-------+             +-------+  |     |   +-------+
        | 2nd 4M|------+      |       |--|-----+   |       |
        +-------+      |      +-------+  |         +-------+
        .       .      |      .       .  |         .       .
        .       .      |      .       .  |         .       .
        +-------+      |      +-------+  |         +-------+
        |       |      |      |       |--|--...    |       |
        +-------+      |      +-------+  |         +-------+
                       |                 |
                       |          +---- -+
                       |          v
                       |      +-------+
                       +----->|       |-...
                              +-------+
                              |       |-...
                              +-------+
                              |       |-...
                              +-------+
                              |   0   |
                              +-------+
                              .       .
                              .       .
                              +-------+
                              |   0   |
                              +-------+

Client memory handles are the linear address of the memory block's
descriptor.  All handles are checked for validity by scanning the
current client's list of memory blocks.  While this at first might
seem inefficient it is also very safe and speed during memory
allocation is hardly important.  Apart from the limit due to linear
address space getting exhausted (i.e., 1022 blocks) there is no
limit on the number of memory blocks allowed for one client.