Machine-Independent Virtual Memory Management

Noted by Lantian Zheng, Feb 1999

Goals for the VM manager of Mach

• Make it as portable as possible and without loss of efficiency
• Support for applications using large, possibly sparse address space
• Integration of memory management and communication
• Provide flexible memory-sharing mechanism

Strategies

• Separate the machine-independent codes from the machine-dependent codes
• External pagers (memory object manager) in user mode
• Copy-on-write mechanism

Components

• Pmap module
  • Run in the kernel
  • Concerned with managing the MMU (set up MMU registers and hardware page tables, and catch all page faults)
  • Depend on the MMU architecture and must be rewritten for each new machine.
• Machine-independent kernel code
  • concerned with processing page faults, managing address maps, replacing pages
• External pagers
  • run in the user mode
  • handle the backing store, page-in and page-out

Implementation Details

• User process has a large, linear virtual address space.
• Virtual memory can be allocated and deallocated by region, which is also the unit of memory sharing.
• Regions are mapped to memory objects by a data structure called address map, a memory object is a repository for data.
• Copy-on-write coping is implemented by shadow objects, while read/write sharing is implemented with a level of indirection--a sharing map.
• Comparison of different MMU architecture
  • The large size of a VAX page table has made pmap more complex
  • Inverted page table architecture simplified pmap, but it allows only one valid mapping for each physical page.
  • Segmentation with paging makes it possible to reasonably implement sparse addressing.
  • The main issue with multiprocessor is TLB consistency, since there's no hardware support for it.

Questions

• What do you think about the overhead caused by user-mode-pager? What's your opinion about microkernel? Is it outdated?
• Generally, high portability and high performance are contradict to each other. What do you think of the high performance showed by the data in Table 7-1?
• Does the existence of machine-independence codes make it more difficult to exploit the potential of MMU?
• What is the cost of copy-on-write mechanism? Is the cost small enough to justify copy-on-write?