CS 3733 Operating Systems

• Each page table entry has a bit indicating whether the page actually occupies a page frame.
• Initially when the process is allocated resources, the page table is set up with all (pure demand paging) or some of the entries marked invalid.
• The entire program image is available on disk. If a page is not in main memory when it is referenced, the page is brought in by the pager and validated.
• A page fault occurrs when a memory reference is to a page that is not in main memory.

• Page table entry is retrieved. (If we're lucky this is from the TLB otherwise from main memory.)
• The hardware checks to see if the page operation is permitted (e.g. writable) and if not, the program aborts.
• The hardware checks to see if the page table entry is valid. If it is, everything is as before.
• If an invalid page is referenced, a page fault trap will occur and the following must be done:
  • Remove the process from the CPU. When restarted, the instruction which caused the page fault will be reexecuted.
    (Be careful to undo operations such as MOV -(R2), (R3)+.)
  • Find a free frame from the free frame list and update this list. (The method used to find this is the page replacement algorithm.)
  • Schedule a disk operation to read page into frame.
  • Update the page table.

• If the process never attempts to access a particular page of memory, it is never brought in.
• PSECTS in VAX assembly language are defined on page boundaries. You can put restrictions such as noexc or nowrt on pages and keep the protection information in the page table.

• Service the page fault interrupt:
  • trap to OS
  • save registers and state of faulting process
  • determine that it was a page fault interrupt
  • find location of requested page on disk
  • select a page to be replaced
  • deal with a dirty replaced page
  • lock selected frame so that it won't be paged out
  • issue I/O request to disk controller
  • move faulting process to waiting queue
  • select another process to be run from the ready queue
  • restore state of selected process so that it can be run

• Swap or page in the page:
  • disk controller queues request to read page
  • wait in queue for device until scheduled
  • wait for disk read
  • transfer the page from disk to main memory using DMA
  • generate I/O complete interrupt

• Restart the process:
  • trap to OS
  • save registers and state of interrupted process (this process is not related to the faulted process)
  • determine it was an I/O complete interrupt
  • update page table and other tables
  • move faulted process from wait to ready queue
  • restore registers and state of interrupted program
  • resume interrupted instruction

The first and third tasks take a few hundred instructions each (perhaps 100 microseconds). Page switching takes perhaps 25 milliseconds. Thus effective memory access time is dominated by the page fault rate.

Skill: Understand the mechanics of demand paging.