Linux Virtual File System

Linux Virtual File System

Robert LedfordLeif Wickland

CS518 Fall 2004

I/O, I/O, It's off to disk I go-o-o, A bit or byte to read or write, I/O, I/O, I/O...

Overview

• What is a file system?

• Historical view of file systems

• Another layer of indirection

• Do I have to?

• File systems have layers

• How is it done?

• Sign me up

What is a file system?

• Speaking broadly, a file system is the logical means for an operating system to store and retrieve data on the computers hard disks, be they local drives, network-available volumes, or exported shares in a storage area network (SAN)


• There is some ambiguity in the term “file system”. The term can be used to mean any of the following:- The type of a file system refers to a specific implementation such as ext2, reiserfs or nfs, each implementation contains the methods and data structures that an operating system uses to keep track of files on a disk or partition

http://e2fsprogs.sourceforge.net/ext2.html

http://www.namesys.com/

http://nfs.sourceforge.net/


- An instance of a file system refers to a file system type residing at a location such as /dev/hda4

- Additionally a file system can refer to the methods and data structures that an operating system uses to keep track of files on a disk or partition


• Linux keeps regular files and directories on block devices such as disks

• A Linux installation may have several physical disk units, each containing one or more file system types

• Partitioning a disk into several file system instances makes it easier for administrators to manage the data stored there


Overheadview

sector track

cylinder

Disk blocks are composed of one or more contiguous sectors

The same track on each platter in a disk makes a cylinder; partitions are groups of contiguous cylinders


• Why have multiple partitions?

• Encapsulate your data:

- Since file system corruption is local to a partition, you stand to lose only some of your data if an accident occurs


• Increase disk space efficiency:- You can format partitions with varying block sizes, depending on your usage- If your data is in a large number of small files (less than 1k) and your partition uses 4k sized blocks, you are wasting 3k for every file- In general, you waste on average one half of a block for every file, so matching block size to the average size of your files is important if you have many files


• Limit data growth:

- Runaway processes or maniacal users can consume so much disk space that the operating system no longer has room on the hard drive for its bookkeeping operations

- This can lead to disaster. By segregating space, you ensure that things other than the operating system die when allocated disk space is exhausted


• Partitioning tools and utilities:- fdiskrledford@leonard > sudo fdisk -l /dev/had

Disk /dev/hda: 255 heads, 63 sectors, 4863 cylindersUnits = cylinders of 16065 * 512 bytesDevice Boot Start End Blocks Id System/dev/hda1 * 1 26 208813+ 83 Linux/dev/hda2 27 1070 8385930 83 Linux/dev/hda3 1071 1853 6289447+ 83 Linux/dev/hda4 1854 4863 24177825 f Win95 Ext'd (LBA)/dev/hda5 1854 2375 4192933+ 83 Linux/dev/hda6 2376 2897 4192933+ 83 Linux/dev/hda7 2898 3028 1052226 82 Linux swap/dev/hda8 3029 4863 14739606 83 Linux

- parted: GNU partition editor

http://www.gnu.org/software/parted/parted.html


• Review Questions - Where does Linux keep regular file types?

• On block devices such as disks.

- On average, how much of a block is wasted for every file?

• On average ½ of a block is wasted for every file.


• Review:• File system instances reside on partitions• Partitioning is a means to divide a single

hard drive into many logical drives• A partition is a contiguous set of blocks on

a drive that are treated as an independent disk

• A partition table is an index that relates sections of the hard drive to partitions


Entire Disk

Partition Table Disk PartitionsMaster BootRecord

Boot Block Super Block Inode List Data Blocks

A Possible File System Instance Layout


• The central structural concepts of a file system type are:- Boot Block

- Super Block

- Inode List

- Data Block




• Boot Block: - Occupies the beginning of a file system

- Typically residing at the first sector, it may also contain the bootstrap code that is read into the machine at boot time

- Although only one boot block is required to boot the system, every file system may contain a boot block

Boot Block


• Super Block: - Describes the state of a file system

- How large it is

- How many files it can store

- Where to find free space in the file system

- Additional data that assists the file management system with operating on the file system

Boot Block Super Block


• Super Block:- Duplicate copies of the super block may reside through out the file system in case the super block is corrupted

Boot Block Super Block


• Inode List: - An inode is the internal representation of a file

contains the description of the disk layout of the file data

- file owner

- permissions

- The inode list contains all of the inodes present in an instance of a file system

Boot Block Super Block Inode List


• Data Blocks: - Contain the file data in the file system

- Additional administrative data

- An allocated data block can belong to one and only one file in the file system



• On a Linux system, a user or user program sees a single file hierarchy, rooted at /

• Every file and directory can trace its origins on a tree back to the root directory


/

bin boot dev etc home lib mnt proc root sbin tmp usr var


• Review Questions - Where is the boot strap code located

• In the Boot Block.

- What contains information about a file’s owner and the file’s permissions?

• The inode.

- What is the index that relates sections of the hard drive to partitions?

• The Partition Table.


• A file system implements the basic operations to manipulate files and directories

• These basic operations include:- Opening of files

- Closing of files

- Creation of directories

- Listing of the contents of directories

- Removal of files from a directory


• The kernel deals on a logical level with file systems rather than with disks

• The separate file systems that the system may use are not accessed by device identifiers

• Instead they are combined into a single hierarchical tree structure that represents the file systems as one whole single entity


• So what is a file system?

- A file system is a set of abstract data types that are implemented for the storage, hierarchical organization, manipulation, navigation, access, and retrieval of data

Overview




• Do I have to?


• How is it done?

• Sign me up

Historical view of file systems

• Managing storage has long been a prominent role for operating systems.– This role was so important to the MS-DOS OS

that it was named after that function. – DOS stands for Disk Operating System.– Created in 1980.


• The file system hasn’t changed much since the 1960s.– A research paper was presented in 1965

describing “A General-Purpose File System For Secondary Storage.”

• Laid out the notion of a hierarchal file system much as is used today.


• File system features in 1965 paper– Files– Directories– Links– Access permissions– Create, access and modify times– Path nomenclature: Directory:directory:file – All devices mount into a unified hierarchy– Backing up

• Everybody knew it was a good idea back then, too.


• This type of file system was implemented in Multics.

• Unix was created as an “emasculated” version of Multics.– Project started as gaming system in 1969.

• The designers of Unix had worked on Multics and brought to Unix a Multics-style file system.


• chdir: change directory• chmod: change permission• chown: change owner• chroot: change root• close: close a file• creat: create a file• dup: copy file descriptor• link: add a file reference• lseek: set open file cursor

• mknod: make a special file• mount: graft in a file system• open: open a file system• pipe: create a pipe• read: read from a file• stat: get file status• umount: opposite of mount• unlink: delete file• write: write to a file

File System API in Unix System V, c. 1983


• Tannenbaum wrote Minix, a pedantic version of Unix, in 1987.– Of course, it included a Unix-style file API.


• Linus Torvalds introduced Linux in 1991.– He developed Linux on Minix.– Consequently it was convenient for the OSes

to share a file system and file API.– Thus, Linux inherited the same style file

system as presented in the 1965 paper.– Today Linux supports a superset of the file

system features available in Unix System V.


• Review Questions– In what year was the paper released that

described the file system design that is the ancestor of Linux’ file system?

• 1965– Yes, that’s about 40 years ago

Overview




• Do I have to?


• How is it done?

• Sign me up

Another layer of indirection

• Multics, Unix, Minux, and Linux originally supported only one file system.– They only understood one type of layout on

disk for directories and files.– Because of its origins, Linux initially supported

just the Minix file system.• Limited to small partitions and short filenames

– However, it wasn’t long before people wanted more from their file systems


• The problem:– Linux was

implemented like this.

User Program

Minix FS Interface

Hard Drive

User Program

Minix FSOther FS

Hard Drive BHard Drive A

– To add support for another file system in a similar manner was unsavory and didn’t scale.

• User program must call a separate API for each type of file system.


• “Any problem in computer science can be solved with another layer of indirection.”– David Wheeler (chief programmer for the

EDSAC project in the early 1950s)

Virtual File System


• The solution was to add a layer of indirection to the file system stack.

• In Linux this layer is called the virtual file system (VFS).

• User programs access any file system through a consistent API.

• All File Systems implement an API which is called by the VFS.

User Program

Minix FSOther FS

Hard Drive BHard Drive A


• The VFS is– Another layer of indirection– A file system- and device-agnostic layer of the

operating system– A consistent API for user applications to

access storage independent of the underlying device or type of file system

Task 1 Task 2 Task n…user space

kernel space

VIRTUAL FILE SYSTEM

minix ext2 msdos proc

device driverfor hard disk

device driver for floppy disk

Buffer Cache

softwarehardware

Hard Disk Floppy Disk

Linux Kernel

Robbed from http://www.cs.usfca.edu/~cruse/cs326/lesson22.ppt

Overview




• Do I have to?


• How is it done?

• Sign me up

Do I have to?

• Is a VFS worth doing?– What do you think?

Do I have to?

• Cons– Harms system

performance• Another layer of

indirection

– Adds to the size of the system because more code must be written

– A conceptually simpler system

• Pros– Enables using multiple

file systems • Facilitates research

– Makes the computer more useful

• My Linux box has ext2, ext3, FAT32, and NTFS partitions mounted

– Facilitates code reuse– Simplifies

implementation

Do I have to?

• Is a VFS worth doing?– Ultimately, the answer is yes for general

purpose operating systems.– All modern commercial operating systems do.– What would you do if you had to design an

OS’ file system? Would you use a VFS?

Overview




• Do I have to?


• How is it done?

• Sign me up

File systems have layers

• Like onions and ogres…

• In this context “file system” means the software stack that extends from the user application to the hardware.


• The file system of Unix System V has three layers– File/Directory API– Inodes– Buffers

File/Directory API

Buffers

Inodes

Storage Device

User Program

File system


• Unix System V File/Directory API– Functions like open()

and read()• Called by user

programs

– Directories are implemented as files

• Contain children’s– Name

– Inode Number

File/Directory API

Buffers

Inodes

Storage Device

User Program

File system


• Unix System V Inodes– Allocate disk blocks for

files.– Record file attributes

• Owner• Access Permissions• File Size• Type

– File– Directory– Special

• Not File Names

File/Directory API

Buffers

Inodes

Storage Device

User Program

File system

If file names aren’t stored in inodes,where are they stored?File names are stored in the parentdirectory entry.


• Unix System V Inodes, continued– Stored on disk– Cached in memory

• With added details– Device the inode is

from– The inode number– If the file is a mount

point– Much more…

– Many pathnames may point to a single inode

File/Directory API

Buffers

Inodes

Storage Device

User Program

File system


• Implementation of an inodeFile Attributes

Direct 0Direct 1Direct 2Direct 3Direct 4Direct 5Direct 6

Direct 8Direct 7

Direct 9Single IndirectDouble IndirectTriple Indirect

Direct Block

Direct Block

Direct Block

.

.

.

Indirect Block

Inode

Inode

Inode

Indirect BlockIndirect Blocks

Inodes Indirect2 Blocks

Inodes Inodes Indirect3 Blocks


• Notes on previous diagram– All the direct pointers of an inode are used

before using an indirect pointer– All of the slots in a single indirect inode are

consumed before starting to use double indirect inodes

• Likewise for double indirect


• Unix System V Buffers– In memory copy of

contents of a disk block

• Same size

– Mechanism through which caching is achieved

• Read ahead• Delayed write

File/Directory API

Buffers

Inodes

Storage Device

User Program

File system


• Review Questions– Where is the type of a directory entry stored?

• In the inode the directory entry points to.

– Where is the name of an inode stored?• In the directory entry which points to it.

– Can more than one file point to a given inode?• Yes. Many files may point to the same inode.

– Can more than inode point to a disk block?• No. Inodes point to zero or more blocks, but a block may be

referenced by zero or one inode.

– How is a directory different that a file?• A directory is a type of file, as indicated by the inode, which

contains a listing of the directory’s children


• How do you get from a file name to a file’s contents?

– Recursive procedure1. Start with inode of current or root directory

• If path begins with ‘/’ use root; otherwise use current• Inodes for both are cached for the process

2. Get disk block(s) pointed to by inode for directory3. Search directory listing for next part of path. 4. If found, get the inode pointed to by entry5. If inode says child is a directory, go to 2. If child

is a file, get disk blocks(s) inode points to.

• Resolving a file path to a file’s contents– Consider how we’d

resolve the path /foo/foobarred/found to that file’s contents in this example directory tree.

– See following slides for an example


/

foo bar.bashrc

foobarred.ssh

lost found


• Resolving a file path to a file’s contentsInodes

01234567

foo: 1; .bashrc:5; bar: 4

lost: 6; found: 7

I’m found

All my secrets would go in here…

foobarred: 2; .ssh: 3

I’m lost

-empty-

I get into bars with my aliases

0

1

2

3

4

5

6

7

Disk Blocks

# Type Disk BlockDir 0Dir 7Dir 2File 3Dir 5File 4File 6File 1

Contents#

Step through the following slides to see the process

of resolving the file name /foo/foobarred/found

to its contents



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Inode for root directory is known to be in slot 0.

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Root directory inode points to block 0

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Root directory listing says that inode 1 is for child

foo.

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Foo’s inode says that it’s a directory and points to

block 7

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Foo’s directory listing says that inode 2 is for child foobarred

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Foobarred’s inode says that it is a directory and its

listing is in block 2

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Directory listing of foobarred says that

child found has inode 7.

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


Inode says that found is a file and 1

is its first block

Contents#



01234567


lost: 6; found: 7

I’m found



I’m lost

-empty-


0

1

2

3

4

5

6

7

Disk Data Blocks


The content of /foo/foobarred/found

is “I’m found”

Contents#


• Review Questions– What are the layers of the System V file

system?• File/Directory API• Inodes• Buffers


• Problems in System V file system implementation– Searching a directory listing for a child is time

consuming• Directory listing is unsorted

– Allows entries to be inserted and removed cheaply– Makes searching expensive– Requires that a linear be performed; can’t use binary– Implies that time to find entry increases linearly with the

number of directory entries


• Problems in System V file system implementation– The listing for the directory must be read from

disk at each step in a path• Can cause the disk head to jump around

– For example, if you want to read the file /foo/foobarred/found, you have to read and search the three directory listing along the way

• Detrimental to performance


• The Linux solution to these shortcomings– Add another layer of indirection– Layer is called the dcache

• Short for directory cache• Caches the contents of directory listings

– The dcache is composed of dentries• Short for directory entry

– A dentry is a cached association from a path name to an inode

• Also caches relationships to other dentries


• With the addition of the dcache, what does the Linux file system software stack look like, compared to the System V file system software stack?– Glad you asked– Next slide, please

Dcache

File/Directory API

Inodes

Real FS Implementation

User Program

Storage Device

Buffers


• Since version 2.1, the Linux virtual file system has had three layers– File/Directory API– Dcache– Inodes

• Buffers are not part of VFS

• Dcache doesn’t exist in System V

Virtual file system

Layers of the Linux kernel FS


• Modern Linux File/Directory API– Implements a superset

of the System V API– User programs interact

with the File Directory API through path names or integer file descriptors

• A file descriptors is an index into an array of pointers to file structures

– More on that later

Dcache

File/Directory API

Inodes


User Program

Storage Device

Buffers

Virtual file system



• Modern Linux Dcache– Improves performance– Composed of dentries

• Caches the inode that a path name points to

• Caches relationships to other dentries

– Parent– Siblings– Children

• Has a hash value for the name of the file it represents

– Speeds up string comparisons

Dcache

File/Directory API

Inodes


User Program

Storage Device

Buffers

Virtual file system



• Modern Linux Inodes– Much like System V

inodes– Contain pointers to file

system implementation specific operations

Dcache

File/Directory API

Inodes


User Program

Storage Device

Buffers

Virtual file system



• Modern Linux Real File System Implementation– Can be compiled into

the kernel or can be loaded as a kernel module

Dcache

File/Directory API

Inodes


User Program

Storage Device

Buffers

Virtual file system



• Modern Linux File Buffers– Integrated with the

virtual memory cache since 1999

– See virtual memory presentation for more information

Dcache

File/Directory API

Inodes


User Program

Storage Device

Buffers

Virtual file system



• Review Questions– What’s the purpose of a dentry?

• A dentry exists to cache a directory entry in order to improve performance

– What are some of the things a dentry links to?• A dentry links to

– Its parent– A list of its children– A list of its siblings– Its file’s in memory inode

Overview




• Do I have to?


• How is it done?

• Sign me up

How is it done?

• Down to the nitty-gritty code details– There’s nothing to fear here

How is it done?

• There are quite a few C structures that the kernel employs to keep track of open files– struct task_struct (View sched.h code)

• Details about a user process• Has a pointer to a files_struct• During a syscall, the kernel has a pointer the

current process’ task_struct

– struct files_struct (View file.h code)• Tracks the open files for a process• Has an array of pointers to file struct instances.

– When a user program opens a file, it’s given an integer index into this array

How is it done?

• File related kernel structures, continued– struct file (View fs.h code)

• The kernel maintains an array of these with an element for each open file system-wide

• Has a pointer to dentry; plus file owner, read/write cursor position, and more

– struct dentry (View dcache.h code)• Points to the inode for the file

– struct inode (View fs.h code)• Stores its block number and the device it’s from• Points to operations in the FS implementation which can

read the disk

How is it done?

• The following diagram shows the relationship of the aforementioned structures

task_struct struct

files_struct pointer

Kernel’s array of all file structs

other stuff

Other stuffDentry pointer

Dcache

User Program

File handle (integer value)

In memory inode cache

A Dentry

A Dentry

Other stuffInode pointer

A file struct

A file struct

inode struct

inode struct

inode struct

Storage Device

Disk Inode

Disk Inode

Data Blocks

Inside the Linux Kernel

FS instance

files_struct struct

other stuff

file struct ptr array

file struct ptr

file struct ptrfile struct ptr

file struct ptr

How is it done?

• Review Questions– When a user program calls open() to open a file, a

non-negative return value indicates success. What does the function of that non-negative number?

• The return value is the index into the array of file structure pointers in the files_struct structure.

– What is the cardinality between user processes and files_struct instances?

• There is a one-to-one relationship between files_struct instances and user processes.

– Why?• Because there are one-to-one relationships between user

processes and task_struct instances and between task_struct instances and files_struct instances.

How is it done?

• Review Questions– If two user processes open the same file, are

two or one file struct instances created?• Trick question

– Normally, two instances are created– However, if a process opened a file and then forked to

create another process, parent and child have the file open and both share one file struct instance

How is it done?

• Great. Now we see how the data relates• But we’d like to see some action• Consider the following C program#include <unistd.h> void main() { const int seekFromStart = 0; const int rdWrCreate = 00102; char ch = 'A'; int fd = open("fsSample.txt", rdWrCreate); write(fd, &ch, 1); lseek(fd, 0, seekFromStart); read(fd, &ch, 1); close(fd); printf("The character read was: %c\n", ch); }

• What actually happens when those file system API functions are called?

• Links in diagrams go to Linux 2.6 source

getname()sys_open()

get_unused_fd()

How is it done?

System call layerMake a copy of the path name string in kernel space

Reserve an unused element in the array of file struct pointers in the process’ files_struct instance (diagram)

filp_open()The workhorse: see the next slide for more detail.Will return a file struct instance for the opened file.

fd_install()putname()

open()

User program Takes the path name of the file to open as an

argument

http://www.cs.montana.edu/~leifw/src/namei.c.html#getname

http://www.cs.montana.edu/~leifw/src/open.c.html#sys_open

http://www.cs.montana.edu/~leifw/src/open.c.html#get_unused_fd

http://www.cs.montana.edu/~leifw/src/open.c.html#filp_open

How is it done?

Get the dentry for that directory.Call link_path_walk() which performs file name resolution. See next slide for more.

Get a dentry for the file we’re opening. Call real FS implementation to populate structure if not cached.

vfs_create()

If the file doesn’t exist, call the real FS implementation to create it.

may_open()Check that the user has permission to open the file.

path_lookup()

Determine if the path is relative to the current directory, the root directory, or a process specific root.

__lookup_hash()sys_open()

filp_open()

System call layer

open_namei()dentry_open()

open()

User program

http://www.cs.montana.edu/~leifw/src/namei.c.html#link_path_walk

http://www.cs.montana.edu/~leifw/src/namei.c.html#vfs_create

http://www.cs.montana.edu/~leifw/src/namei.c.html#may_open

http://www.cs.montana.edu/~leifw/src/namei.c.html#path_lookup

http://www.cs.montana.edu/~leifw/src/namei.c.html#__lookup_hash


http://www.cs.montana.edu/~leifw/src/namei.c.html#open_namei

sys_open()

How is it done?

open()

System call layer

filp_open()

open_namei()

While there are segments left in the path name

If the segment is “..”

Fail if user does not have permission to inode

path_lookup()

link_path_walk()

Get a dentry for child of inode named by segment

Set inode to the parent of inode, allowing for crossing mount points or inode being root dir

Set inode to the inode for directory that path_lookup() determined the path name was relative to

Parse the next “/” separated segment in path name

If the dentry for the segment isn’t cachedAsk FS implementation for dentry and its inode

If the dentry points to a symbolic linkGet the dentry and inode for the item pointed to

If the dentry is for a mount pointGet the dentry and inode for mounted root dir

Set inode to dentry’s inode

Purpose: Resolve the path name to a dentryUser program

Return the dentry that owns inode

http://www.cs.montana.edu/~leifw/src/namei.c.html#link_path_walk

How is it done?

Allocate a file struct instance

Returned a dentry for the path namesys_open()

filp_open()

System call layer

open_namei()dentry_open()

open()

User program

Populate that file struct from the dentry

http://www.cs.montana.edu/~leifw/src/namei.c.html#open_namei

http://www.cs.montana.edu/~leifw/src/open.c.html#dentry_open

getname()sys_open()

get_unused_fd()

How is it done?

System call layer

filp_open()

Returned a file struct instance for the opened file

fd_install()putname()

Set the previously reserved file struct pointer to the file struct instance returned by filp_open()

Deallocate the kernel copy of the filename

open()

User program


http://www.cs.montana.edu/~leifw/src/open.c.html#fd_install

http://www.cs.montana.edu/~leifw/src/fs.h.html#putname

sys_open()

How is it done?

System call layer

Returns the index where fd_install() put the new file struct pointer. Value is called the file descriptor, or FD.

open()

User program

fget_light()sys_read()

file_pos_read()

How is it done?

System call layerGet the file struct instance for the FD

vfs_read()file_pos_write()fput_light()

read()

User program Takes the open file’s FD as an argument

Copy the read/write cursor location from file struct

Call the real FS implementation to read a specified number of bytes from the file

Increment the read/write cursor by the number of bytes read

Release the file struct instance for the FD

http://www.cs.montana.edu/~leifw/src/file_table.c.html#fget_light

http://www.cs.montana.edu/~leifw/src/read_write.c.html#sys_read

http://www.cs.montana.edu/~leifw/src/read_write.c.html#file_pos_read

http://www.cs.montana.edu/~leifw/src/read_write.c.html#vfs_read

http://www.cs.montana.edu/~leifw/src/read_write.c.html#file_pos_write

fget_light()sys_lseek()

How is it done?

System call layer Get the FD’s file struct instance

vfs_llseek()fput_light()

lseek()


default_llseek()

Release the FD’s file struct instance

Get the read/write cursor location from file structUpdate the cursor according to the argumentsSave the cursor location back into the file struct


http://www.cs.montana.edu/~leifw/src/read_write.c.html#sys_lseek

http://www.cs.montana.edu/~leifw/src/read_write.c.html#vfs_llseek

fget_light()sys_write()

file_pos_read()

How is it done?

System call layerGet the file struct instance for the FD

vfs_write()file_pos_write()fput_light()

write()


Copy the read/write cursor location from file struct

Call the real FS implementation to write the specified bytes at the write cursor location

Increment the read/write cursor by the number of bytes written

Release the file struct instance for the FD


http://www.cs.montana.edu/~leifw/src/read_write.c.html#sys_write

http://www.cs.montana.edu/~leifw/src/read_write.c.html#file_pos_read

http://www.cs.montana.edu/~leifw/src/read_write.c.html#vfs_write

http://www.cs.montana.edu/~leifw/src/read_write.c.html#file_pos_write

filp_close()

sys_close()

How is it done?

System call layer

Saves the value of FD so that an unused slot can be found quickly when next one is needed

Calls the real FS implementation to flush any unwritten data to storage

close()

User programTakes the FD of the file to close as an argument

Frees FD’s slot in the array of file struct pointers

If this is the last reference to the file struct instance

fput()

Allow the real FS implementation to free resources Deallocate the dentry for the file

http://www.cs.montana.edu/~leifw/src/open.c.html#filp_close

http://www.cs.montana.edu/~leifw/src/open.c.html#sys_close

Overview




• Do I have to?


• How is it done?

• Sign me up

Sign me up.

• So how do the Linux kernel and the Virtual File System find out about a particular file system implementation?

• (e.g., ext2, nfs, reiserfs)

Sign me up.

• When you perform a Linux installation you are prompted to see if you want each of the supported file systems in your build

• When the kernel is actually built, the file system startup code contains calls to the initialization routines of all of the built in file systems

Sign me up.

• Linux file systems may also be built as modules and, in this case, they may be demand loaded as they are needed or loaded by hand using insmod

• Whenever a file system module is loaded it registers itself with the kernel and unregisters itself when it is unloaded

Sign me up.

• Each specific file system's initialization routine registers itself with the Virtual File System and is represented by a:

file_system_type

data structure which contains the name of the file system and a pointer to its VFS Super Block read routine

Sign me up.

• When a file system is registered the file_system_type data structure is populated with data specific to the file system

• And the file_system_type structure is put into a list, pointed at by the file_systems pointer

Sign me up.

• The kernel utilizes the file_systems list to check if a specific file system has been registered

• And to assist with the mapping of the Virtual File System’s operations to a specific implementation of a file system’s operations by using the VFS Super Block read routine

Sign me up.

file_systems

name (ext2)

fs_flags

*read_super()

*next

name (reiserfs)

fs_flags

*read_super()

*next

name (nfs)

fs_flags

*read_super()

*next

file_system_type file_system_type file_system_type

Registered File Systems Example

file_systems pointer

list of file_system_type

structures

Sign me up.

• Where is the file_system_type structure declared?

• What does it look like?

linux/fs.h

struct file_system_type {const char *name;int fs_flags;struct super_block *(*read_super) (struct super_block *, void *,int);struct file_system_type * next;

};

Sign me up.

• Some details of the file_system_type structure:- name:

-The name of the file system type, such as ext2, nfs, reiserfs etc.- This field is used as a key and it is not possible to register a file system that is already in use- Function find_filesystem() utilizes the

name field to check if a file system has already been registered

Sign me up.

- fs_flags:

- A number of flags which record features of the file system

- If the fs_flag is set to FS_REQUIRES_DEV then a block device must be given when mounting the file system

- Not all file systems need a device to hold them. The /proc file system, for example, does not require a block device

Sign me up.

- fs_flags: (continued)- Aside: The command:$cat /proc/filesystemsdisplays file system information from the file_systems list of file_system_structures- In particular, it is possible to see if a file system requires a device or not by noting

the presence or lack of the “nodev” in front of a file system listing

Sign me up.

- fs_flags: (continued)

ledford@esus ~ $ cat /proc/filesystemsnodev rootfsnodev bdevnodev procnodev sockfsnodev tmpfsnodev shmnodev pipefsnodev binfmt_miscext3ext2nodev ramfsiso9660nodev nfsnodev smbfsnodev autofsreiserfsnodev devptsxfs

Sign me up.

- next:

- A pointer used for chaining all the file_system_type structures together

file_systems

name (ext2)

fs_flags

*read_super()

*next

name (reiserfs)

fs_flags

*read_super()

*next

name (nfs)

fs_flags

*read_super()

*next


Sign me up.

- read_super

- This routine is called by the VFS when an instance of the file system is mounted

struct super_block *(*read_super) (struct super_block *, void *, int);

- What is going on here?

- A VFS super_block structure is being populated by the function call read_super with data specific to a particular file system

Sign me up.struct super_block *(*read_super) (struct super_block *, void *, int);

- The void* pointer points to data that has been passed down from the mount system call- The trailing int signifies whether or not

read_super should be silent about errors- This is only set when mounting the root file system- Several file systems may be tried when attempting to mount the root

file system so avoiding unsightly errors is desired

Sign me up.

• So what gets called when we register a file system?

• Linux finds out about new file system types by calls to:

register_filesystem()

• And forgets about them by the calls to its counterpart:

unregister_filesystem()

Sign me up.

The formal declarations are:

#include <linux/fs.h>

…

int register_filesystem(struct file_system_type * fs);

…

int unregister_filesystem(struct file_system_type * fs);

Sign me up.

• register_filesystem()linux/fs/super.c…int register_filesystem(struct file_system_type * fs){ struct file_system_type ** tmp;

if (!fs) return -EINVAL; if (fs->next) return -EBUSY; tmp = &file_systems; while (*tmp) { if (strcmp((*tmp)->name, fs->name) == 0) return -EBUSY; tmp = &(*tmp)->next; } *tmp = fs; return 0;}

• So what’s happening here?- Essentially register_filesystem() takes a, file system specific, populated file_system_type structure as a parameter performs some checks and, if successful, adds the file_system_type structure to the file_systems list

Sign me up.

• unregister_filesystem()linux/fs/super.c…int unregister_filesystem(struct file_system_type * fs){#ifdef CONFIG_MODULES struct file_system_type ** tmp;

tmp = &file_systems; while (*tmp) { if (fs == *tmp) { *tmp = fs->next; fs->next = NULL; return 0; } tmp = &(*tmp)->next; }#endif return -EINVAL;}

• So what’s happening here?- unregister_filesystem() takes a, file system specific, populated file_system_type structure as a parameter and removes the file_system_type structure from the file_systems list if present

Sign me up.

• How is the file_system_type structure populated with file system specific data?

- Each file system implementation defines a file_system_type structure with data specific to that file system

Sign me up.

• Example:

- The ext2 file_system_type structurelinux/fs/ext2/super.c

…

static struct file_system_type ext2_fs_type = {

"ext2",

FS_REQUIRES_DEV /* | FS_IBASKET */, /* ibaskets have unresolved bugs */

ext2_read_super,

NULL

};

- Here we can see the fields: name, fs_flags, read_super and next being populated

Sign me up.

• How is the register_filesystem() called from a specific file system?

- Each file system implementation defines an init function that calls register_filesystem()

Sign me up.

• Example:

- The ext2 init_ext2_fs() functionlinux/fs/ext2/super.c

…

static int __init init_ext2_fs(void)

{

return register_filesystem(&ext2_fs_type);

}

- This is pretty self-explanatory

Sign me up.

• How is the unregister_filesystem() called from a specific file system?

- Each file system implementation defines an exit function that calls unregister_filesystem()

Sign me up.

• Example:

- The ext2 exit_ext2_fs() functionlinux/fs/ext2/super.c

…

static void __exit exit_ext2_fs(void)

{

unregister_filesystem(&ext2_fs_type);

}

- Again, This is self-explanatory

Sign me up.

• So how and where are the calls to the init and exit functions made?

linux/fs/ext2/super.c…module_init(init_ext2_fs)module_exit(exit_ext2_fs)

- Calls to module_init() and module_exit() begin the, file system specific, registration and unregistration processes

Sign me up.

• But how are module_init() and module_exit() called?- module_init() is called when the module is loaded, if built as a module, with a call to insmod

- Or it is called at the same time as all of the init calls are made during the kernel boot

process- module_exit() is called when the module is unloaded, if built as a module, with a call to rmmod

Sign me up.

• So how does this all fit together to register a file system?

Sign me up.module_init(init_ext2_fs)

init_ext2_fs(void)

register_filesystem(&ext2_fs_type)

file_systems

name (ext2)

fs_flags

*read_super()

*next

file_system_type

Called at boot time or with

insmod

name (ext2)

fs_flags

*read_super()

*next

file_system_typeTakes a file_system_type

structure specific to ext2 as a parameter

Registering the ext2 file system

Populates the file_systems list with the

file_systems_type structure for ext2

Sign me up.

• So how does this all fit together to unregister a file system?

Sign me up.Sign me up.module_exit(exit_foo_fs)

exit_foo_fs(void)

unregister_filesystem(&foo_fs_type)

file_systems

name (foo)

fs_flags

*read_super()

*next

file_system_type

Called with rmmod

name (foo)

fs_flags

*read_super()

*next

file_system_typeTakes a file_system_type

structure specific to foo as a parameter

Unregistering the foo file system

Remove from the file_systems list the file_systems_type structure for foo

module_exit(exit_foo_fs)

exit_foo_fs(void)


file_systems

name (foo)

fs_flags

*read_super()

*next

file_system_type

Called with rmmod



Sign me up.

Sign me up.Sign me up.Sign me up.module_exit(exit_foo_fs)

exit_foo_fs(void)


file_systems

name (ext2)

fs_flags

*read_super()

*next

file_system_type

Called with rmmod



Update the file_systems pointer to point at what

next was pointing at in the file_system_type structure for foo

In this case it is ext2

Sign me up.

• Whoptie doo, Basil.. what does it all mean?- For the Virtual File System layer to work with a specific file system implementation it must have some knowledge of the file system- This knowledge is acquired in part by registering the file system- Once this is done the Virtual File System can map calls to files, on a specific file system, through the correct structures and perform the requested operations

Sign me up.

• What’s next?

- After the file system has been registered we must mount it in order to use it

- A file system is mounted at boot or with the use of the mount command

- To unmount a file system the umount command is used

Sign me up.

• So what does mounting do?

- When a file system is mounted the file_systems list is searched to see if the file system has been registered

Sign me up.

file_systems

name (nfs)

fs_flags

*read_super()

*next

name (reiserfs)

fs_flags

*read_super()

*next

name (ext2)

fs_flags

*read_super()

*next


Looking for ext2 in the file_systems list

Is name ext2? Is name ext2? Is name ext2?

NO NO YES

Sign me up.

- If the file system is found in the list then the read_super() function in the file_system_type structure for the particular file system is called

- The call to read super occurs in fs/super.c do_mount()

name (ext2)

fs_flags

*read_super()

*next

file_system_type

Sign me up.

• Review read_super()- Like we saw earlier the read_super() function

populates a VFS super_block with data from a particular file system’s super_block

Sign me up.

• So what the heck is a super_block?- A super_block is a structure that maintains information about a particular file system- The Virtual File System has a super_block structure that is populated with data from a specific file system’s implementation of the super_block- Is this confusing?

Sign me up.

data 1

data 2

data 3

data 4

super_block

data 1

data 2

data 3

data 4

ext2_super_block

… …

data n data n

VFS ext2

Sign me up.

• A little more in depth– The call to the VFS read_super() makes a call

to a specific file systems version of read_super()

– The file systems version of read_super() populates a VFS super_block structure with data from it’s version of the super_block and returns the populated VFS super_block back to the VFS read_super()

• Confused again?

Sign me up.read_super()

ext2_read_super()

data 1

data 2

super_block

data 1

data 2

ext2_super_block

… …data n data 2

VFS version

ext2 version

Returns populated suber_blockback to read_super

Populates

super_block

Sign me up.

• So why go through all this trouble?– In order for the VFS layer to operate on the

data or inodes residing on a specific file system it needs to know what the file system specific data/inode operations are

– Part of the VFS super_block structure is a pointer to a VFS super_operations structure

– This super_operations structure is populated during read_super with the operations specific to a file system

Sign me up.

• Links to the source for the above operations:– VFS read_super()– ext2_read_super()– VFS super_block– ext2_super_block– VFS super_operations– ext2_sops // The ext2 version of the VFS

super_operations stucture

Sign me up.

• What’s next?– Now that we’ve populated a VFS super_block

with file system specific data we need to maintain some lists in the VFS layer so we know what file systems have been mounted

– In read_super there is a call to insert_super()– insert_super places the VFS super_block

generated by read_super into the super_blocks list

Sign me up.

fs/super.c

static void insert_super(struct super_block *s, structfile_system_type *type)

{

s->s_type = type;

list_add(&s->s_list, super_blocks.prev);

list_add(&s->s_instances, &type->fs_supers);

spin_unlock(&sb_lock);

get_filesystem(type);

}

Sign me up.read_super()

ext2_read_super()

VFS version

ext2 version

Populates VFS super_block

Returns to read_super()

read_super()

insert_super()

Places super_block in list

super_blocks

data 1

data 2

super_block

…data n

super_block is added to the super_blocks list

Sign me up.

• In addition to the super_blocks list another list, vfsmntlist, is also maintained with the currently mounted file systems- vfsmntlist is a list of vfsmount structures- The vfsmount list is populated by calls toadd_vfsmnt()- The function do_mount() calls add_vfsmount() after the call to read_super()- Within the vfsmount structure is a pointer to the VFS super_block for the mounted file system

Sign me up.Sign me up.do_mount()

add_vfsmnt()

Adds a vfsmount structureto the vfsmntlist

Returns to do_mount()

read_super()

vfsmntlist

data 1

data 2

vfsmount

…

data n

vfsmount structure is added to vfsmntlist

do_mount()

Sign me up.

• In addition to the pointer to the VFS super_block of the file system, the vfsmount structure contains: – The device number of the block device

holding the file system– And the directory where this file system is

mounted

Sign me up.

• Review• In order to use a file system:

– It must be registered with VFS• Registration builds a VFS super_block structure from the

specific file systems super_block• And places a file_system_type structure entry in the

file_systems list– It must be mounted with VFS

• Mounting adds an entry into the super_blocks and the vfsmntlist lists

• By utilizing the information in these data structures VFS is able to resolve operations to specific file system implementations

Overview




• Do I have to?


• How is it done?

• Sign me up

Resources

• http://www.tldp.org/LDP/lki/lki-3.html

• http://mm.iit.uni-miskolc.hu/Data/texts/Linux/SAG/node74.html

• http://www.enterprisestorageforum.com/technology/features/article.php/2026611

• http://www.multicians.org/fjcc4.html

Resources

• http://www.computerhope.com/history/unix.htm

• http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dnpag/html/intpatt-ch03.asp

• http://e2fsprogs.sourceforge.net/ext2intro.html

• http://www.cs.usfca.edu/~cruse/cs326/lesson22.ppt

Resources

• http://bama.ua.edu/~dunna001/journeyman/html/x323.htm#AEN390

• http://ldp.rtin.bz/LDP/lki/lki-3.html

• http://people.netfilter.org/~rusty/unreliable-guides/kernel-hacking/routines-init-again.html

• http://www.cse.unsw.edu.au/~neilb/oss/linux-commentary/vfs-3.html

Resources

• http://www.cs.wits.ac.za/~adi/courses/linuxadmin/content/module2doc.html

• http://www.tldp.org/LDP/tlk/fs/filesystem.html

• http://www.faqs.org/docs/kernel_2_4/lki-3.html

• http://www.tldp.org/HOWTO/Partition/partition-4.html#AEN487

Resources

• http://www.science.unitn.it/~fiorella/guidelinux/tlk/node94.html

• The design of the UNIX Operating System, Maurice J. Bach

• Linux File Systems, Moshe Bar

Linux Virtual File System

Documents

Transcript of Linux Virtual File System