Virtual Memory 1

1
Chapter 13. Virtual Memory

• Introduction
• Demand Paging
• Hardware Requirements
• 4.3 BSD Virtual Memory
• 4.3 BSD Memory Management Operations

2
Introduction

• Memory management unit (MMU)
• Responsible for getting data to and from main
  memory
• Virtual memory
• The notion of an address space as distinct from
  memory locations
• Translation tables
• Page-based

3
Introduction (cont)

• Memory management in the Stone Age
• Software overlay
• Swapping
• Demand paging
• Segmentation

4
Demand Paging

• Primary goal
• To allow a process to run in a virtual address
  space
• To perform translations from virtual to physical
  addresses transparently to the process
• Functional requirements
• Address space management
• fork( ), exec( )
• Address translation
• Virtual address ? Virtual page number offset

5
Demand Paging (cont)

• Physical memory management
• Memory protection
• Memory sharing
• Monitoring system load
• Other facilities
• Memory-mapped files
• Dynamically linked shared libraries
• etc.

6
Demand Paging (cont)

• Page swapping

uninitialized data pages zero-filled on first
access
dirty pages saved initialized data
main memory
swap area on disk
text and initialized data
executable file
subsequent faults on outswapped pages
stack and heap pages allocated on first access
7
Demand Paging (cont)

• Translation maps

address space map
backup store map
page fault
hardware address translations
Data
process virtual page number
physical page number
physical memory map
8
Demand Paging (cont)

• Page replacement policies
• Local v.s. global replacement policy
• Locality of reference
• Concept of working set
• Least recently used (LRU) policy

9
Hardware Requirements

• MMU
• Translation of virtual addresses
• Using page tables or TLBs
• Page table entry
• Physical page frame number
• Protection information
• a valid bit
• a modified bit
• a referenced bit

10
Hardware Requirements (cont)

• Page table
• Hardware-prescribed
• Located in main memory
• MMU uses only the active tables, whose locations
  are loaded in h/w page table registers
• Typically, on a uniprocessor, there are two
  active page tables - one for kernel and one for
  the currently running process
• Paging the page table to reduce the required
  memory

11
Hardware Requirements (cont)

• Address translation may fail for three reasons
• Bounds error
• Validation error
• Protection error
• MMU caches
• A high-speed cache that is searched before each
  memory access
• Translation lookaside buffer (TLB)

12
Hardware Requirements (cont)

• Intel x86
• Unix implementations on the x86 hides the notion
  of segmentation from user processes, who see a
  flat address space
• Two-level page table
• 4k byte page size
• Control register CR3 stores the physical page
  number of the current page directory
• CR3 register needs to be reset on each context
  switch

13
Hardware Requirements (cont)

• Page table entry
• Physical page number, protection field
• Valid, referenced, and modified bit
• x86 supports 4 privilege levels of which UNIX
  uses only two
• The kernel runs in the innermost ring, which is
  the most privileged
• User code runs in the outermost, least privileged
  ring

14
Address Translation on Intel x86
page directory of current process
CR3
31
11
0
one of the page tables of current process
PFN
31
11
0
PFN
31
21
11
0
DIR
PAGE
OFFSET
virtual address
31
11
0
PFN
OFFSET
physical address
15
Intel x86 Page Table Entry
31
12
6
5
2
1
0
PFN
D
A
U
W
P
PFN Page Frame Number D Dirty A Accessed
(Referenced) U User (0) / Supervisor (1) W Read
(0) / Write (1) P Present (valid)
16
4.3BSD

• Target platform VAX-11
• Several BSD-based implementations emulate the VAX
  memory architecture in software, including its
  address space layout and page table entry format
• Core map
• Describes physical memory
• Page tables
• Describe virtual memory

17
4.3BSD (cont)

• Disk maps
• Describe the swap areas
• Resource maps
• Manage allocation of resources such as page
  tables and swap space

18
4.3BSD Physical Memory
error buffer
nonpaged pool
paged pool
cmap (in nonpaged pool)
19
4.3BSD Physical Memory (cont)

• Core map struct cmap
• Is a kernel data structure, allocated at boot
  time and resident in the nonpaged pool
• One entry for each frame in the paged pool
• Core map entry
• Name
• lttype, owner, virtual page numbergt
• Text page cache
• Synchronization

20
4.3BSD Physical to Virtual address Translation
proc
cmap
type data owner VPN
page table
type data owner VPN
page table
text
21
4.3BSD Address Space

• Uses VAX-11 address space model
• 32-bit machine with 512-byte page size
• 4 Giga byte address space is divided into four
  regions of equal size
• P0
• Program region
• Text and data section of the process

22
4.3BSD Address Space (cont)

• P1
• Control region
• User stack, u area, kernel stack
• S0
• System region
• Kernel text and data
• 4th region
• Reserved and not supported by current VAX h/w

23
4.3BSD Page Tables

• Single system page table
• Map the kernel text and data
• Each process has two page tables to map its P0
  and P1
• Mapped by a set of contiguous PTEs in the
  Userptmap section of the system page table
• State of a particular page of a process
• Resident
• The page is in physical memory, and the page
  table entry contains its physical page frame
  number

24
4.3BSD Page Tables (cont)

• Fill-on-demand
• Fill-from-text Text and initialized data pages
  are read in from the executable file upon first
  access
• Zero-fill Uninitialized data, heap, and stack
  pages are created and filled with zero when
  required
• Outswapped
• These pages may be recovered from their swap area
  locations

25
4.3BSD Page Tables (cont)

• Kernel maintains information about all
  nonresident pages
• For swapped out pages
• Kernel must stores their location on the swap
  device
• For zero-fill pages
• Kernel only needs to recognize them as such
• For fill-from-text pages
• Kernel must determine their location in the
  filesystem

26
4.3BSD Page Tables (cont)

• Page table entry
• Since all nonresident pages have the valid bit
  clear, other fields can be replaced by other
  information that tracks these pages
• Kernel maintains separate swap maps to locate
  those pages on the swap device

27
4.3BSD Page Tables Entry
(a) VAX-11 page table entry format
31
26
20
0
Page Frame Number (PFN)
PROT
V
M
unused
valid
modified
(b) Ordinary page table entry
31
26
20
0
25
Page Frame Number (PFN)
PROT
V
M
0
valid
modified
fill-on-demand
fill-from-text (1) or zero-fill (0)
File System Block Number
PROT
V
1
F
31
26
23
0
(c) Fill-on-demand page table entry
28
4.3BSD Swap Space

• Per region dmap structure

data region
swap space
0
dmap
dmmin
2dmmin
4dmmin
29
4.3BSD Swap Space (cont)

• Text pages
• Once the page is brought into memory, the
  fill-on-demand PTE is overwritten by the page
  frame number
• As a result, retrieving the page from the file
  involving recomputing its location and perhaps,
  accessing one or more indirect blocks
• To avoid that, such pages are saved in swap as
  well

30
4.3BSD Swap Space (cont)

• U area holds the maps for the data and stack
  region
• The text region swap map is part of the text
  structure

31
4.3BSD Memory Management Operations

• Process creation
• Swap space, u area
• Page tables
• Kernel must allocate contiguous PTEs in Userptmap
  to map page tables for their process
• Text region
• The child is added to the list of processes
  sharing the text structure used by the parent
• Data and Stack
• Data and stack must be copied one page at a time
• Copy-on-write, vfork( )

32
4.3BSD Memory Management Operations (cont)

• Page table handling
• Two types of page faults
• validation, protection
• Validation
• No PTE for that page (bounds error)
• PTE is marked invalid
• Pagein( ) is called to handle the fault
• The faulting virtual address ? PTE

33
Pagein( ) (1/2)
start
Fill on demand?
No
Yes
No
PFN 0?
(next page)
Yes
in transit text page?
Yes
text page onhash queue?
Yes
set wanted flag
No
(next page)
No
on free list?
No
allocate new page
sleep on text struct
Yes
take page off free list
read page from swap
start over when woken up
set valid bit
34
Pagein( ) (2/2)
Fill on demand?
Yes
(previous page)
Yes
zero-fill?
(previous page)
No
allocate new page
Yes
text page onhash queue?
fill it with zeros
No
take page off free list
Yes
page in buffer cache?
flush cache copy to disk
No
Read page from file
mark page modified
35
4.3BSD Memory Management Operations (cont)

• Free page list
• Ideally, to keep all garbage pages at the head of
  the free list, followed by some useful pages in
  LRU order
• 4.3BSD replaces the least recently used policy by
  a not recently used policy
• Uses referenced bit and two passes over each page
• VAX-11 does not support a referenced bit in the
  h/w, so BSD simulates the referenced bit in
  software
• Pagedaemon process is responsible for page
  replacement

36
4.3BSD Memory Management Operations (cont)

• Swapping
• Problem of thrashing
• Swapper process monitors the system load, and
  swap processes in and out when needed
• Swapper will swap out a process in the following
  cases
• Userptmap fragmentation
• Memory shortfall
• Inactive processes

37
4.3BSD Memory Management Operations (cont)

• Swapper performs the following task when swapping
  out a process
• Allocates swap space for the u area, kernel
  stack, and page tables
• Detach the process from its text region
• Save the resident data and stack on swap
• Release the system PTEs in Userptmap
• Record the swap location of the u area in the
  proc structure

38
Analysis of 4.3BSD

• No support for execution of remote programs
• No support for sharing of memory
• vfork( ) is not a true substitute for fork( )
• The lack of copy-on-write hurts the performance
  of Ap that rely extensively on fork
• No support for memory mapped files

39
Analysis of 4.3BSD (cont)

• Reserves enough swap space in advance to page out
  very single page in the process address space
• No support for using swap space on remote nodes