Machine-Independent Virtual Memory Management for
Paged Uniprocessor and Multiprocessor Architectures
Richard Rashid, et al. - 1988
---------------------------------------------------

Summary: This paper describes the Mach virtual memory system, which
was designed with an emphasis on portability without sacrificing
functionality or performance.

Functional Features: 

   1. Read-write sharing w/ copy-on-write 
   2. User-level pagers 
   3. Memory-mapped files

Design Decisions, Motivations, and Issues:

   D1. To address data, the kernel uses byte offsets into memory
       objects (analogous to UNIX files) rather than VPN/offset.

   M1. Avoids linking the implementation to a particular notion of
       physical page size.

   I1. VA translation:

       - VA region found in Address Map (sorted linked list that maps
	 contiguous regions of VA-space to contiguous regions of
	 memory objects) to obtain memory_object/offset.

       - PPN found in hash table keyed on memory_object/offset.

       Sharing:
       
       - Copy-on-write achieved through a shadow object, which has a
         pointer back to the memory object it is shadowing.  Changed
         pages are written to the shadow objects, but unchanged pages
         are accessed via the pointer to the original object.

       - Read/write sharing is achieved with an extra level of
         indirection.  Tasks sharing a region of a memory object have
         entries in their address maps point to a sharing map rather
         than to the memory object itself.  This sharing map is an
         address map that points to the memory object.

   D2. Message-based IPC

   M2. Extends nicely to multiprocessor and distributed systems
       because there is a consistent interface to all resources
       independent of the machine's hardware and OS.

   I2. Most message-based systems suffer in performance, so Mach's VM
       and IPC are integrated to make sending large messages efficient
       using copy-on-write (performance same as memory remapping).

   D3. pmap (physical address map; e.g. for a VAX a pmap is a page
       table) management is the only machine-dependent component.  For
       performance, the pmap is not required to contain all currently
       valid mappings; all but a few essential kernel
       virtual-to-physical mappings can be thrown away to improve
       either space or speed efficiency.

Discussion:

   - Why can't read/write sharing be done by having address maps point
     to a shared memory object?

   - Why are there current and maximum values for protection
     (described on page 3, last paragraph)?

   - Is memory copying the main reason for bad performance in
     message-based IPC systems?  Are there other factors?