P j t3 Project 3: Virtual Memory

Jin-Soo Jin Soo Kim ( [email protected]) Computer Systems Laboratory Sungkyunkwan University htt // l kk d http://csl.skku.edu

Introduction (1)
Paging g g in the x86 architecture

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Introduction (2)
Current Pintos VM implementation p
Use paging Page g size: 4KB Each process has its own page tables
The page directory is allocated when the process is created (pagedir_create()@userprog/pagedir.c) (struct thread*)t>pagedir points to the page directory (load()@userprog/process.c userprog/process c) The (secondary) page tables are dynamically created if necessary (lookup_page()@userprog/pagedir.c) For kernel region, processes have the same mapping (PHYS_BASE ~ 0xffffffff)

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Introduction (3)
Current Pintos VM implementation p ( (contd) )
No demand paging
When a process is created, all the contents of code and data segments are read into the physical memory (load_segment()@userprog/process.c)

Fixed stack size

Only one stack page is allocated to each process (setup_stack()@userprog/process.c)

No page faults in the user mode

Everything needed by each process is in the physical memory

Page faults may occur only in the kernel mode

If you use the optimistic approach to accessing arguments When Wh i invalid lid pointers i are passed d via i system calls ll
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Project 3 Overview
Requirements q
Lazy loading (or demand paging) Swapping pp g in/out / p pages g from/to / swap p disk Dynamic stack growth Memory e o y mapped apped files es

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Lazy Loading (1)

Why? y
An executable file holds code and data images Ap process will not need all the p pages g immediately y

How to?
Use the executable file as the backing store
Only when a page is needed at run time, load the corresponding code/data page into the physical memory Loaded pages will have valid PTEs

Handling page faults

Accesses to not-yet-loaded pages will cause page faults Find the corresponding location in the executable file Read in the page from the executable file Setup the corresponding PTE
6

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Lazy Loading (2)

Loading g code/data from the executable file
In load_segment()@userprog/process.c Each p page g is filled with data using g p page g _zero_bytes y and page_read_bytes
page_zero_bytes + page_read_bytes = PGSIZE

All zeroed page (page_zero_bytes == PGSIZE)

Allocate a new page and initialize it with zeroes

Full ll code/data d d page (page_read_bytes == PGSIZE)

Allocate a new page and read its contents from the executable file

Partial page (0<page_read_bytes <PGSIZE)

Read page_read_bytes page read bytes from the executable file and fill the rest of the page with zeroes
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Lazy Loading (3)

The supplemental pp p page g table
The page table with additional data about each page Main p purposes p
On a page fault, find out what data should be there for the faulted virtual page On O a process t termination, i ti d decide id what h t resources t to f free

Possible organizations
Per-segment Per segment Per-page

Implementation
You can use any data structure for the supplemental page table. <hash.h>will be useful (lib/kernel/hash.[ch])
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Lazy Loading (4)

Strategy gy
Pagefault handler
1 Gettheinformation 1. onthefaultedpage VA x86 page table Pagefault 2.Allocateanewframe The supplemental pagetable

4 Updateinformation 4. 3.Readfromtheexecutablefile 5.Installthenewpage

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Swapping (1)
Why? y
You may run out of the physical memory Your programs p g memory y footprint p can be larger g than the physical memory size

How to?
Find a victim page in the physical memory Swap out the victim page to the swap disk Extend your supplemental page table to indicate the victim p page g has been swapped pp out When the page is accessed later, swap in the page from the swap disk to the physical memory
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

10

Swapping (2)
Swap p disk
Use the following command to create an 4 MB swap disk in the vm/build directory
$pintosmkdisk swap.dsk4

Alternatively, you can tell Pintos to use a temporary 4 MB swap di 4-MB disk kf for a single i l run with i h swapdisk=4 di k
Used during makecheck

A swap di disk k consists i t of f swap slots l t

A swap slot is a continuous, page-size region of disk space on the swap p disk
swap slot 0 swap slot 1 swap slot 2

...
11

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

Swapping (3)
Accessing g swap p disk
The swap disk is automatically attached as hd1:1 when you run Pintos. Use the disk interface in devices/disk.h
A size of a disk sector is 512 bytes You can read or write one sector at a time
struct disk*disk_get disk get (int chan_no, chan no int dev_no); dev no); disk_sector_t disk_size (struct disk*); voiddisk_read (struct disk*,disk_sector_t,void*); voiddisk_write disk write (struct disk*, ,disk_sector_t, disk sector t, constvoid*);

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

12

Swapping (4)
Managing g g swap p slots
Pick an unused swap slot for evicting a page from its from to the swap disk Free a swap slot when its page is read back or the process is terminated Allocate lazily, i.e., only when they are actually required by eviction

The swap table

The swap table tracks in-use and free swap slots <bitmap.h>will be useful (lib/kernel/bitmap.[ch])
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

13

Swapping (5)
Page g replacement p p policy y
You should implement a global page replacement algorithm that approximates LRU
Do not use FIFO or RANDOM The second chance or clock algorithm is OK Bonus B if you i implement l t your own page replacement l t policy li better than the second chance algorithm

Get/Clear Accessed and Dirty bits in the PTE

pagedir_is_dirty(), pagedir_set_dirty() pagedir_is_accessed(), pagedir_set_accessed()

Other processes should be able to run while you are performing I/O due to page faults
Some synchronization effort will be required
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

14

Swapping (6)
The frame table
Allows efficient implementation of eviction policy One entry y for each frame that contains a user p page g
Each entry contains a pointer to the page, if any, that currently occupies it, and other data of your choice

Use the h f frame table bl while h l you choose h a victim page to evict when no frames are free Code pages can be shared among those processes created from the same executable file (optional)

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

15

Swapping (7)
User p pool vs. kernel p pool
The physical memory is divided into the user pool and the kernel pool
Running out of pages in the user pool just causes user programs to page Running out of pages in the kernel pool means a disaster The size of the user pool can be limited (ul option)

The frames used for user pages should be obtained from the user pool
By calling palloc_get_page (PAL_USER)

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

16

Swapping (8)
Frame allocation
On top of the current page allocator (threads/palloc.c)
palloc_get_page(),palloc_free_page()

If there are free frames in the user pool, allocate one by calling palloc_get_page() If none is free
Choose a victim page using your page replacement policy Remove R references f to the h f frame f from any page table bl that h refers to it If the frame is modified, , write the p page g to the file system y or to the swap disk Return the frame

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

17

Stack Growth (1)

Growing g the stack segment g
Allocate additional pages as necessary Devise a heuristic that attempts p to distinguish g stack accesses from other accesses
Bug if a program writes to the stack below the stack pointer However, in x86, it is possible to fault 4 ~ 32 bytes below the stack pointer

You may impose some absolute limit on stack size The first stack page need not be allocated lazily
The page is initialized with the command line arguments

All stack pages should be candidates for eviction

An evicted stack page should be written to swap
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

18

Stack Growth (2)

How to obtain the user stack pointer? p
You need the current value of the user programs stack pointer on page fault
Compare it with the faulted address

When the page fault occurred in the user mode

Use (struct intr_frame *)f>esp

When the page fault occurred in the kernel mode

struct intr_frame i f i not is t saved db by th the processor (struct intr_frame *)f>esp yields an undefined value Save esp into struct threadon the initial transition from user to kernel mode

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

19

Memory Mapped Files (1)

Example p
Writes the contents of a file to the console
#include<stdio.h> <stdio h> #include<syscall.h> int main(int argc,char*argv[]) { void*data=(void*)0x10000000; int fd =open(argv[1]); mapid_t map=mmap (fd,data); write(1,data,filesize(fd)); munmap (map); return0; }

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

20

Memory Mapped Files (2)

System y calls to implement p
mapid_t mmap (int fd,void*addr); voidmunmap (mapid_t (mapid t mapping);

mmap() fails if
fd is i 0 or 1 The file has a length of zero bytes addr is 0 addr is not page-aligned The range of pages mapped overlaps any exisitng set of mapped pages

All mappings are implicitly unmapped when a process exits

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

21

Memory Mapped Files (3)

Managing g g mapped pp files
Lazily load pages in mmap regions
For the final mapped page, set the bytes beyond the end of the file to zero

Use the mmapd file itself as backing store for mapping

All pages written to by the process are written back to the file

Closing or removing a file does not unmap any of its mappings

Once created, a mapping is valid until munmap() is called or the process exits

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

22

Summary (1)
Pages g
Code page (clean) Data p page g ( (clean/dirty) / y) Stack page (dirty) mmaped aped page (c (clean/dirty) ea /d ty)

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

23

Summary (2)
When y you attach a new frame,
It may be just initialized to zero It may y be read from a file It may be read from a swap slot

When you evict a frame,

It may b be j just td dropped d It may be swapped out to a swap slot It may b be written itt t to a fil file

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

24

Tips (1)
Suggested gg order of implementation p
Lazy loading
Modify load_segment()and page_fault() Construct the supplemental page table You should be able to run all user programs of Project 2

Frame allocation/deallocation layer

Add a new interface that can allocate or free a frame Construct the frame table as you allocate a new frame Assume there is enough g p physical y memory y No eviction is necessary You should be able to run all user programs of Project 2

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

25

Tips (2)
Suggested gg order of implementation p ( (contd) )
Page replacement policy
Develop your own page replacement policy Need to interact with the supplemental page table and the frame table First, First try to evict read read-only only pages and make sure it has no problem And then, implement the swap table and test your code to access the swap disk Finally, implement the full-fledged page replacement policy

Stack growth
Extend your page fault handler

Memory mapped files

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

26

Tips (3)
No files in the vm directory y
You should add your files in the directory The Pintos documentation says y

Adding your own source files (src/Makefile.build)

CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

27

Submission
Due
December 15, 11:59PM Fill out the design g document (vm.txt) and p put it in your source tree (pintos/src/vm) Tar and gzip your Pintos source codes
$cd pintos $(cd src/vm;makeclean) $tarcvzf f TeamName.tar.gz ./src /

Send it to the instructor via e-mail (NOT to the GoogleGroups!!) Hand in the printed version of your design document during the demo session (after final exam)
CSE3008: Operating Systems | Fall 2009 | Jin-Soo Kim ([email protected])

28