CS 5600

Computer Systems

Lecture 10: File Systems

 What are We Doing Today?
• Last week we talked extensively about hard
drives and SSDs
– How they work
– Performance characteristics
• This week is all about managing storage
– Disks/SSDs offer a blank slate of empty blocks
– How do we store files on these devices, and keep
track of them?
– How do we maintain high performance?
– How do we maintain consistency in the face of
random crashes?
2
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
3
 Building the Root File System
• One of the first tasks of an OS during bootup is to
build the root file system
1. Locate all bootable media
– Internal and external hard disks
– SSDs
– Floppy disks, CDs, DVDs, USB sticks
2. Locate all the partitions on each media
– Read MBR(s), extended partition tables, etc.
3. Mount one or more partitions
– Makes the file system(s) available for access
4
 The Master Boot Record
Address Size
Description
Hex Dec. (Bytes) Includes the starting
0x000 0 Bootstrap code area 446 LBA and length of
0x1BE 446 Partition Entry #1 16 the partition
0x1CE 462 Partition Entry #2 16
0x1DE 478 Partition Entry #3 16
0x1EE 494 Partition Entry #4 16
0x1FE 510 Magic Number 2
Total: 512
Disk 1

Partition 1 Partition 2 Partition 3 Partition 4

MBR
(ext3) (swap) (NTFS) (FAT32)
Disk 2

Partition 1
MBR
(NTFS) 5
 Extended Partitions
• In some cases, you may want >4 partitions
• Modern OSes support extended partitions

Logical Logical
Disk 1

Partition 1 Partition 2 Ext. Partition 3 Partition 4

MBR Partition 1 Partition 2
(ext3) (swap) Part. (Extended Partition) (FAT32)
(NTFS) (NTFS)

• Extended partitions may use OS-specific partition

table formats (meta-data)
– Thus, other OSes may not be able to read the logical
partitions 6
 Types of Root File Systems
•[cbw@ativ9
Windows ~] df -hexposes a multi-rooted
Filesystem Size Used Avail Use% Mounted on
system39G 14G 23G 38% /
/dev/sda7
1 drive, 4
/dev/sda2 296M 48M 249M 16% /boot/efi partitions
– Each device and partition is assigned
/dev/sda5 127G 86G 42G 68% /media/cbw/Data
a letter
/dev/sda4 61G 34G 27G 57% /media/cbw/Windows 1drive, 1
/dev/sdb1 1.9G 352K 1.9G 1% /media/cbw/NDSS-2013
– Internally, a single root is maintained partition

• Linux has a single root

– One partition is mounted as /
– All other partitions are mounted
somewhere under /
• Typically, the partition containing
the kernel is mounted as / or C: 7
 Mounting a File System
1. Read the super block for the target file system
– Contains meta-data about the file system
– Version, size, locations of key structures on disk, etc.
2. Determine the mount point
– On Windows: pick a drive letter
– On Linux: mount the new file system under a
specific directory
Filesystem Size Used Avail Use% Mounted on
/dev/sda5 127G 86G 42G 68% /media/cbw/Data
/dev/sda4 61G 34G 27G 57% /media/cbw/Windows
/dev/sdb1 1.9G 352K 1.9G 1% /media/cbw/NDSS-2013
8
 Virtual File System Interface
• Problem: the OS may mount several partitions
containing different underlying file systems
– It would be bad if processes had to use different APIs
for different file systems
• Linux uses a Virtual File System interface (VFS)
– Exposes POSIX APIs to processes
– Forwards requests to lower-level file system specific
drivers
• Windows uses a similar system

9
 VFS Flowchart
Processes (usually) don’t Relatively simple to
need to know about low- add additional file
level file system details system drivers

Process 1 Process 2 Process 3

Virtual File System Interface

Kernel

ext3 Driver NTFS Driver FAT32 Driver

ext3 Partition NTFS Partition FAT32 Partition

10
 Mount isn’t Just for Bootup
• When you plug storage
devices into your running
system, mount is executed
in the background
• Example: plugging in a USB
stick
• What does it mean to
“safely eject” a device?
– Flush cached writes to that
device
– Cleanly unmount the file
system on that device 11
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
12
 Status Check
• At this point, the OS can locate and mount
partitions
• Next step: what is the on-disk layout of the file
system?
– We expect certain features from a file system
• Named files
• Nested hierarchy of directories
• Meta-data like creation time, file permissions, etc.
– How do we design on-disk structures that support
these features?
13
 The Directory Tree
cs5600
home cbw

bin

/ (root) python

tmp amislove

• Navigated using a path

– E.g. /home/amislove/music.mp3 14
 Absolute and Relative Paths
• Two types of file system paths
– Absolute
• Full path from the root to the object
• Example: /home/cbw/cs5600/hw4.pdf
• Example: C:\Users\cbw\Documents\
– Relative
• OS keeps track of the working directory for each process
• Path relative to the current working directory
• Examples [working directory = /home/cbw]:
– syllabus.docx [  /home/cbw/syllabus.docx]
– cs5600/hw4.pdf [  /home/cbw/cs5600/hw4.pdf]
– ./cs5600/hw4.pdf [  /home/cbw/cs5600/hw4.pdf]
– ../amislove/music.mp3 [  /home/amislove/music.mp3] 15
 Files
• A file is a composed of two components
– The file data itself
• One or more blocks (sectors) of binary data
• A file can contain anything
– Meta-data about the file
• Name, total size
• What directory is it in?
• Created time, modified time, access time
• Hidden or system file?
• Owner and owner’s group
• Permissions: read/write/execute
16
 File Extensions
• File name are often written in dotted notation
– E.g. program.exe, image.jpg, music.mp3
• A file’s extension does not mean anything
– Any file (regardless of its contents) can be given any
name or extension
Has the data in the
Rename file changed from
music to an image?

• Graphical shells (like Windows explorer) use

extensions to try and match files  programs
– This mapping may fail for a variety of reasons 17
 More File Meta-Data
• Files have additional meta-data that is not
typically shown to users
– Unique identifier (file names may not be unique)
– Structure that maps the file to blocks on the disk
• Managing the mapping from files to blocks is one
of the key jobs of the file system

Disk 18
 Mapping Files to Blocks
• Every file is composed of >=1 blocks
• Key question: how do we map a file to its blocks?
List of blocks As (start, length) pairs

[1] [4, 5, 7, 8] [6] (1, 1) (4, 4) (9, 1)

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9

• Problem? • Problem?
– Really large files – Fragmentation
– E.g. try to add a new
file with 3 blocks 19
 Directories
• Traditionally, file systems have used a hierarchical,
tree-structured namespace
– Directories are objects that contain other objects
• i.e. a directory may (or may not) have children
– Files are leaves in the tree
• By default, directories contain at least two entries
“..” points the the “.” self pointer
parents directory
.
..
/ (root) bin
20
python
 More on Directories
• Directories have associated meta-data
– Name, number of entries
– Created time, modified time, access time
– Permissions (read/write), owner, and group
• The file system must encode directories and store
them on the disk
– Typically, directories are stored as a special type of file
– File contains a list of entries inside the directory, plus
some meta-data for each entry

21
 Example Directory File

Windows

C:\ Name Index Dir? Perms

. 2 Y rwx
Users
Windows 3 Y rwx
Users 4 Y rwx
pagefile.sys 5 N r
pagefile.sys

0 1 2 3 4 5 6 7 8 9

Disk C:\

22
 Directory File Implementation
• Each directory file stores many entries
• Key Question: how do you encode the entries?
Unordered List of Entries Sorted List of Entries
• Other alternatives: hash tables, B-trees
Name Index Dir? Perms Name Index Dir? Perms
• More on B-trees later…
. 2 Y rwx .
• In practice, implementing directory files is2complicated
Y rwx
Windows 3 Y rwx
• Example: do filenames havepagefile.sys 5
a fixed, maximum N r
length
Users
or 4variable
Y rwx
length? Users 4 Y rwx
pagefile.sys 5 N r Windows 3 Y rwx

• Good: O(1) to add new entries • Good: O(log n) to search for

– Just append to the file an entry
• Bad: O(n) to search for an • Bad: O(n) to add new entries
entry – Entire file has the be rewritten
 File Allocation Tables (FAT)
• Simple file system popularized by MS-DOS
– First introduced in 1977
– Most devices today use the FAT32 spec from 1996
– FAT12, FAT16, VFAT, FAT32, etc.
• Still quite popular today
– Default format for USB sticks and memory cards
– Used for EFI boot partitions
• Name comes from the index table used to track
directories and files
24
• Stores basic info about the file system
• FAT version, location of boot files
• Total number of blocks
• Index of the root directory in the FAT

• File allocation table (FAT)

• Marks which blocks are free or in-use
• Linked-list structure to manage large files

• Store file and directory data

• Each block is a fixed size (4KB – 64KB)
• Files may span multiple blocks

Super
Disk Block

25
 • Directories are special files
– File contains a list of entries inside
Windows the directory
• Possible values for FAT entries:
C:\ – 0 – entry is empty
Users – 1 – reserved by the OS
– 1 < N < 0xFFFF – next block in a chain
– 0xFFFF – end of a chain

2 3 4 5 6 7 8 9 2 3 4 5 6 7 8 9
Super
Disk Block C:\

Name Index Dir? Perms

Root directory . 2 Y rwx
index = 2 Windows 3 Y rwx
Users 4 Y rwx
pagefile.sys 5 N r 26
 Fat Table Entries
• len(FAT) == Number of clusters on the disk
– Max number of files/directories is bounded
– Decided when you format the partition
• The FAT version roughly corresponds to the size in
bits of each FAT entry
– E.g. FAT16  each FAT entry is 16 bits
– More bits  larger disks are supported

27
 Fragmentation
• Blocks for a file need not be contiguous

56 57 58 59 60 61 62 63 64 65 67 68

FAT 0 0 65 0 0 0xFF 0 58 0 67 61 0
FF

56 57 58 59 60 61 62 63 64 65 67 68

Blocks

Possible values for FAT entries:

• 0 – entry is empty
• 1 < N < 0xFFFF – next block in a chain
• 0xFFFF – end of a chain 28
 FAT: The Good and the Bad
• The Good – FAT supports:
– Hierarchical tree of directories and files
– Variable length files
– Basic file and directory meta-data
• The Bad
– At most, FAT32 supports 2TB disks
– Locating free chunks requires scanning the entire FAT
– Prone to internal and external fragmentation
• Large blocks  internal fragmentation
– Reads require a lot of random seeking
29
 Lots of Seeking
FAT may have very low
• Consider the following code: spatial locality, thus a
lot of random seeking
int fd = open(“my_file.txt”, “r”);
int r = read(fd, buffer, 1024 * 4 * 4); // 4 4KB blocks

56 57 58 59 60 61 62 63 64 65 67 68

FAT 67 0xFF 0 0xFF 63 0 56 57 0 0 59 60

FF FF

56 57 58 59 60 61 62 63 64 65 67 68

Blocks
30
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
31
 Status Check
• At this point, we have on-disk structures for:
– Building a directory tree
– Storing variable length files
• But, the efficiency of FAT is very low
– Lots of seeking over file chains in FAT
– Only way to identify free space is to scan over the
entire FAT
• Linux file system uses more efficient structures
– Extended File System (ext) uses index nodes (inodes)
to track files and directories
32
 Size Distribution of Files
• FAT uses a linked list for all files
– Simple and uniform mechanism
– … but, it is not optimized for short or long files
• Question: are short or long files more common?
– Studies over the last 30 years show that short files
are much more common
– 2KB is the most common file size
– Average file size is 200KB (biased upward by a few
very large files)
• Key idea: optimize the file system for many small
files 33
• Super block, storing:
• Size and location of bitmaps
• Number and location of inodes
• Number and location of data blocks
• Index of root inodes
• Table of inodes
Bitmap of free & • Each inode is a file/directory
used data blocks • Includes meta-data and lists
of associated data blocks
Bitmap of free &
used inodes
Data blocks (4KB each)

34
 • Directories are files
• Contains the list of
bin
entries in the directory
/ Name inode
• . Each inode 0
can directly point to 12
home cbw
blocks
bin 1
• Can also indirectly point to blocks
home 2
at 1, 2, and 3 levels of depth
initrd.img 3

Inode Data
Inodes Data Blocks
Bitmap Bitmap

SB

Root inode = 0 35
 ext2 inodes
Size (bytes) Name What is this field for?
2 mode Read/write/execute?
2 uid User ID of the file owner
4 size Size of the file in bytes
4 time Last access time
4 ctime Creation time
4 mtime Last modification time
4 dtime Deletion time
2 gid Group ID of the file
2 links_count How many hard links point to this file?
4 blocks How many data blocks are allocated to this file?
4 flags File or directory? Plus, other simple flags
60 block 15 direct and indirect pointers to data blocks
36
 inode Block Pointers
• Each inode is the root of an unbalanced tree of
data blocks 15 total pointers
inode

Triple
Single Double
Indirect
Indirect Indirect

12 blocks *
4KB = 48KB

1024 blocks * 230 blocks * 4KB = 4TB

4KB = 4MB
1024 * 1024 blocks * 4KB = 4GB
37
 Advantages of inodes
• Optimized for file systems with many small files
– Each inode can directly point to 48KB of data
– Only one layer of indirection needed for 4MB files
• Faster file access
– Greater meta-data locality  less random seeking
– No need to traverse long, chained FAT entries
• Easier free space management
– Bitmaps can be cached in memory for fast access
– inode and data space handled independently

38
 File Reading Example
Bitmaps inodes Data Blocks
data inode root tmp file root tmp file[0] file[1] file[3]
read
open(“/tmp/file”)

read
read
read
read
Update the last read
Time

read() accessed time read

of the file write
read
read() read
write
read
read() read
write
 Bitmaps inodes Data Blocks
File data inode root tmp file root tmp file[0]
read
Create read
and read

open(“/tmp/file”)
read
Write read
Example write
Update the write
modified time write
of the directory write
Time

read
read
write()
write
write
write
 ext2 inodes, Again
Size (bytes) Name What is this field for?
2 mode Read/write/execute?
2 uid User ID of the file owner
4 size Size of the file in bytes
4 time Last access time
4 ctime Creation time
4 mtime Last modification time
4 dtime Deletion time
2 gid Group ID of the file
2 links_count How many hard links point to this file?
4 blocks How many data blocks are allocated to this file?
4 flags File or directory? Plus, other simple flags
60 block 15 direct and indirect pointers to data blocks
41
 Hard Link Example
• Multiple directory entries may point to the same
inode
[amislove@ativ9 ~] ln –T ../cbw/my_file cbw_file

cbw my_file 1. Add an entry to the “amislove”

directory
home
2. Increase the link_count of the
amislove cbw_file “my_file” inode

Inode Data
Inodes Data Blocks
Bitmap Bitmap

SB

42
 Hard Link Details
• Hard links give you the ability to create many
aliases of the same underlying file
– Can be in different directories
• Target file will not be marked invalid (deleted)
until link_count == 0
– This is why POSIX “delete” is called unlink()
• Disadvantage of hard links
– Inodes are only unique within a single file system
– Thus, can only point to files in the same partition

43
 Soft Links
• Soft links are special files that include the path
to another file
– Also known as symbolic links
– On Windows, known as shortcuts
– File may be on another partition or device

44
 Soft Link Example
[amislove@ativ9 ~] ln –s ../cbw/my_file cbw_file

cbw my_file
1. Create a soft link file
home 2. Add it to the current
amislove cbw_file directory

Inode Data
Inodes Data Blocks
Bitmap Bitmap

SB

45
 ext: The Good and the Bad
• The Good – ext file system (inodes) support:
– All the typical file/directory features
– Hard and soft links
– More performant (less seeking) than FAT
• The Bad: poor locality
– ext is optimized for a particular file size distribution
– However, it is not optimized for spinning disks
– inodes and associated data are far apart on the disk!
Inode Data
Bitmap Bitmap Inodes Data Blocks

SB
46
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
47
 Status Check
• At this point, we’ve moved from FAT to ext
– inodes are imbalanced trees of data blocks
– Optimized for the common case: small files
• Problem: ext has poor locality
– inodes are far from their corresponding data
– This is going to result in long seeks across the disk
• Problem: ext is prone to fragmentation
– ext chooses the first available blocks for new data
– No attempt is made to keep the blocks of a file
contiguous
48
 Fast File System (FFS)
• FFS developed at Berkeley in 1984
– First attempt at a disk aware file system
– i.e. optimized for performance on spinning disks
• Observation: processes tend to access files that
are in the same (or close) directories
– Spatial locality
• Key idea: place groups of directories and their
files into cylinder groups
– Introduced into ext2, called block groups

49
 Block Groups
• In ext, there is a single set of key data structures
– One data bitmap, one inode bitmap
– One inode table, one array of data blocks
• In ext2, each block group contains its own key
data structures
Inode Data
Bitmap Bitmap Inodes Data Blocks

Block Block Block Block Block Block

SB
Group 1 Group 2 Group 3 Group 4 Group 5 Group 6 50
 Allocation Policy
• ext2 attempts to keep related files and
directories within the same block group

amislove home cbw

Block Block Block Block Block Block

SB
Group 1 Group 2 Group 3 Group 4 Group 5 Group 6
 ext2: The Good and the Bad
• The good – ext2 supports:
– All the features of ext…
– … with even better performance (because of
increased spatial locality)
• The bad
– Large files must cross block groups
– As the file system becomes more complex, the
chance of file system corruption grows
• E.g. invalid inodes, incorrect directory entries, etc.

52
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
53
 Status Check

• At this point, we have a full featured file system

– Directories
– Fine-grained data allocation
– Hard/soft links
• File system is optimized for spinning disks
– inodes are optimized for small files
– Block groups improve locality
• What’s next?
– Consistency and reliability
54
 Maintaining Consistency

• Many operations results in multiple,

independent writes to the file system
– Example: append a block to an existing file
1. Update the free data bitmap
2. Update the inode
3. Write the user data
• What happens if the computer crashes in the
middle of this process?

55
 File Append Example
owner: christo • These three operations can
permissions: rw
size: 1
2 potentially be done in any order
pointer: 4
pointer: null
5
• … but the system can crash at
pointer: null any time
pointer: null

Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

Update
Update Write the
the data
the inode data
bitmap 56
 Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
D1 D2

Result: file system is consistent, but the data is lost Write the data

v2
v1
D1

Update the inode Result: inode points to garbage data, and file
system is inconsistent (data bitmap vs. inode)

v1
D1

Update the Result: space leakage, and file system is

data bitmap inconsistent (data bitmap vs. inode)
 Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

Result: inode points to data, but file system is inconsistent

v1
D1 D2

Result: file system is inconsistent, and the data is

useless since it’s not associated with an inode

v1
v2
D1

Result: file system is consistent, but the inode

points to garbage data
 The Crash Consistency Problem
• The disk guarantees that sector writes are atomic
– No way to make multi-sector writes atomic
• How to ensure consistency after a crash?
1. Don’t bother to ensure consistency
• Accept that the file system may be inconsistent after a crash
• Run a program that fixes the file system during bootup
• File system checker (fsck)
2. Use a transaction log to make multi-writes atomic
• Log stores a history of all writes to the disk
• After a crash the log can be “replayed” to finish updates
• Journaling file system
59
 Approach 1: File System Checker
• Key idea: fix inconsistent file systems during
bootup
– Unix utility called fsck (chkdsk on Windows)
– Scans the entire file system multiple times, identifying
and correcting inconsistencies
• Why during bootup?
– No other file system activity can be going on
– After fsck runs, bootup/mounting can continue

60
 fsck Tasks
• Superblock: validate the superblock, replace it
with a backup if it is corrupted
• Free blocks and inodes: rebuild the bitmaps by
scanning all inodes
• Reachability: make sure all inodes are reachable
from the root of the file system
• inodes: delete all corrupted inodes, and rebuild
their link counts by walking the directory tree
• directories: verify the integrity of all directories
• … and many other minor consistency checks 61
 fsck: the Good and the Bad
• Advantages of fsck
– Doesn’t require the file system to do any work to
ensure consistency
– Makes the file system implementation simpler
• Disadvantages of fsck
– Very complicated to implement the fsck program
• Many possible inconsistencies that must be identified
• Many difficult corner cases to consider and handle
– fsck is super slow
• Scans the entire file system multiple times
• Imagine how long it would take to fsck a 40 TB RAID array
62
63
 Approach 2: Journaling
• Problem: fsck is slow because it checks the entire
file system after a crash
– What if we knew where the last writes were before
the crash, and just checked those?
• Key idea: make writes transactional by using a
write-ahead log
– Commonly referred to as a journal
• Ext3 and NTFS use journaling
Block Block Block
Superblock Journal …
Group 0 Group 1 Group N
64
 Write-Ahead Log
• Key idea: writes to disk are first written into a log
– After the log is written, the writes execute normally
– In essence, the log records transactions
• What happens after a crash…
– If the writes to the log are interrupted?
• The transaction is incomplete
• The user’s data is lost, but the file system is consistent
– If the writes to the log succeed, but the normal
writes are interrupted?
• The file system may be inconsistent, but…
• The log has exactly the right information to fix the problem
65
 Data Journaling Example
• Assume we are appending to a file
– Three writes: inode v2, data bitmap v2, data D2
• Before executing these writes, first log them
Journal

TxB TxE
I v2 B v2 D2
ID=1 ID=1

1. Begin a new transaction with a unique ID=k

2. Write the updated meta-data block(s)
3. Write the file data block(s)
4. Write an end-of-transaction with ID=k 66
 Commits and Checkpoints
• We say a transaction is committed after all writes
to the log are complete
• After a transaction is committed, the OS
checkpoints the update Committed!

Journal TxB I v2 B v2 D2 TxE

Inode Data Checkpointed!
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

• Final step: free the checkpointed transaction 67

 Journal Implementation
• Journals are typically implemented as a circular
buffer
– Journal is append-only
• OS maintains pointers to the front and back of
the transactions in the buffer
– As transactions are freed, the back is moved up
• Thus, the contents of the journal are never
deleted, they are just overwritten over time

68
 Data Journaling Timeline

Journal File System

TxB Meta-data Data TxE Meta-data Data
Issue Issue Issue Issue
Complete
Complete
Complete
Time
Complete
Issue Issue
Complete
Complete

69
 Crash Recovery (1)
• What if the system crashes during logging?
– If the transaction is not committed, data is lost
– But, the file system remains consistent

Journal TxB I v2 B v2 D2
Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
D1

70
 Crash Recovery (2)
• What if the system crashes during the checkpoint?
– File system may be inconsistent
– During reboot, transactions that are committed but not
free are replayed in order
– Thus, no data is lost and consistency is restored

Journal TxB I v2 B v2 D2 TxE

Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

71
 Corrupted Transactions
• Problem: the disk scheduler may not execute
writes in-order
– Transactions in the log may appear committed, when
in fact they are invalid
Journal TxB I v2 B v2 D2 TxE

• Transaction looks valid,

• Solution: add a checksum to TxB
but the data is missing!
• During recovery, reject transactions
• During replay, garbage
with invalid checksums
data is written to the
• Implemented on Linux in ext4
file system
72
 Journaling: The Good and the Bad
• Advantages of journaling
– Robust, fast file system recovery
• No need to scan the entire journal or file system
– Relatively straight forward to implement
• Disadvantages of journaling
– Write traffic to the disk is doubled
• Especially the file data, which is probably large
– Deletes are very hard to correctly log
• Example in a few slides…

73
 Making Journaling Faster
• Journaling adds a lot of write overhead
• OSes typically batch updates to the journal
– Buffer sequential writes in memory, then issue one
large write to the log
– Example: ext3 batches updates for 5 seconds
• Tradeoff between performance and persistence
– Long batch interval = fewer, larger writes to the log
• Improved performance due to large sequential writes
– But, if there is a crash, everything in the buffer will be
lost
74
 Meta-Data Journaling
• The most expensive part of data journaling is
writing the file data twice
– Meta-data is small (~1 sector), file data is large
• ext3 implements meta-data journaling
Journal TxB I v2 B v2 TxE
Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

75
 Meta-Journaling Timeline

Journal File System

TxB Meta-data TxE Meta-data Data
Issue Issue Issue
Complete Transaction
Complete can only be
Complete
committed
after the meta-
Time

Issue
Complete
data and data
Issue
are written
Complete

76
 Crash Recovery Redux (1)
• What if the system crashes during logging?
– If the transaction is not committed, data is lost
– D2 will eventually be overwritten
– The file system remains consistent

Journal TxB I v2 B v2
Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
D1 D2

77
 Crash Recovery Redux (2)
• What if the system crashes during the checkpoint?
– File system may be inconsistent
– During reboot, transactions that are committed but not
free are replayed in order
– Thus, no data is lost and consistency is restored

Journal TxB I v2 B v2 TxE

Inode Data
Inodes Data Blocks
Bitmap Bitmap
v1
v2
D1 D2

78
 Delete and Block Reuse
Journal TxB dir dir TxE TxB dir dir TxE TxB f1 f1 TxE

Inode Data
Inodes Data Blocks
Bitmap Bitmap
dir
f1
dir
f1

The block that previously held directory

info is reused to hold file data

1. Create a directory: inode and data are written

2. Delete the directory: inode is removed
3. Create a file: inode and data are written
79
 The Trouble With Delete
• What happens when the log is replayed?

Journal TxB dir dir TxE TxB dir dir TxE TxB f1 f1 TxE

Data Blocks

f1
dir

file data is overwritten file data is not in the

by directory meta-data log, thus it is lost! :(

80
 Handling Delete
• Strategy 1: don’t reuse blocks until the delete is
checkpointed and freed
• Strategy 2: add a revoke record to the log
– ext3 used revoke records

TxB Rx TxB TxB

Journal ID=1
dir TxE
ID=1 ID=2
dir TxE
ID=3
f1 TxE

If the log is replayed,

ignore transaction ID=1

81
 Journaling Wrap-Up
• Today, most OSes use journaling file systems
– ext3/ext4 on Linux
– NTFS on Windows
• Provides excellent crash recovery with relatively
low space and performance overhead
• Next-gen OSes will likely move to file systems
with copy-on-write semantics
– btrfs and zfs on Linux

82
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
83
 Status Check
• At this point:
– We not only have a fast file system
– But it is also resilient against corruption
• What’s next?
– More efficiency improvements!

84
 Revisiting inodes
• Recall: inodes use indirection to acquire
additional blocks of pointers
• Problem: inodes are not efficient for large files
– Example: for a 100MB file, you need 25600 block
pointers (assuming 4KB blocks)
• This is unavoidable if the file is 100% fragmented
– However, what if large groups of blocks are
contiguous?

85
 From Pointers to Extents
• Modern file systems try hard to minimize
fragmentation
– Since it results in many seeks, thus low performance
• Extents are better suited for contiguous files
Each extent
inode inode
includes a block
block 1 block 1
pointer and a
block 2 length 1
length
block 3 block 2
block 4 length 2
block 5 block 3
block 6 length 3
86
 Implementing Extents
• ext4 and NTFS use extents
• ext4 inodes include 4 extents instead of block
pointers
– Each extent can address at most 128MB of
contiguous space (assuming 4KB blocks)
– If more extents are needed, a data block is allocated
– Similar to a block of indirect pointers

87
 Revisiting Directories
• In ext, ext2, and ext3, each directory is a file with
a list of entries
– Entries are not stored in sorted order
– Some entries may be blank, if they have been deleted
• Problem: searching for files in large directories
takes O(n) time
– Practically, you can’t store >10K files in a directory
– It takes way too long to locate and open files

88
 From Lists to B-Trees
• ext4 and NTFS encode directories as B-Trees to
improve lookup time to O(log N)
• A B-Tree is a type of balanced tree that is
optimized for storage on disk
– Items are stored in sorted order in blocks
– Each block stores between m and 2m items
• Suppose items i and j are in the root of the tree
– The root must have 3 children, since it has 2 items
– The three child groups contain items a < i, i < a < j,
and a > j
89
 Example B-Tree
• ext4 uses a B-Tree variant known as a H-Tree
– The H stands for hash (sometime called B+Tree)
• Suppose you try to open(“my_file”, “r”)
hash(“my_file”) = 0x0000C194
H-Tree Root
0x00AD1102 0xCFF1A412

H-Tree Node H-Tree Node H-Tree Node

0x0000C195 0x00018201

H-Tree Leaf
H-Tree Leaf H-Tree Leaf
0x0000A0D1 0x0000C194

my_file  inode 90
 ext4: The Good and the Bad
• The good – ext4 (and NTFS) supports:
– All of the basic file system functionality we require
– Improved performance from ext3’s block groups
– Additional performance gains from extents and B-
Tree directory files
• The bad:
– ext4 is an incremental improvement over ext3
– Next-gen file systems have even nicer features
• Copy-on-write semantics (btrfs and ZFS)

91
• Partitions and Mounting
• Basics (FAT)
• inodes and Blocks (ext)
• Block Groups (ext2)
• Journaling (ext3)
• Extents and B-Trees (ext4)
• Log-based File Systems
92
 Status Check
• At this point:
– We have arrived at a modern file system like ext4
• What’s next?
– Go back to the drawing board and reevaluate from
first-principals

93
 Reevaluating Disk Performance
• How has computer hardware been evolving?
– RAM has become cheaper and grown larger :)
– Random access seek times have remained very slow :(
• This changing dynamic alters how disks are used
– More data can be cached in RAM = less disk reads
– Thus, writes will dominate disk I/O
• Can we create a file system that is optimized for
sequential writes?

94
 Log-structured File System
• Key idea: buffer all writes (including meta-data)
in memory
– Write these long segments to disk sequentially
– Treat the disk as a circular buffer, i.e. don’t overwrite
• Advantages:
– All writes are large and sequential
• Big question:
– How do you manage meta-data and maintain
structure in this kind of design?

95
 Treating the Disk as a Log
• Same concept as data journaling
– Data and meta-data get appended to a log
– Stale data isn’t overwritten, its replaced

Data Data inode Data inode Data inode

Disk

Block 1 Block 2 1 Block 5 2 Block 1 1

96
 Buffering Writes
• LFS buffers writes in-memory into chunks
Memory

Data Data Data Data inode Data inode

Block 1 Block 2 Block 3 Block 4 1 Block 5 2
Disk

Giant Log

• Chunks get appended to the log once they are

sufficiently large 97
 How to Find inodes
• In a typical file system, the inodes are stored at
fixed locations (relatively easy to find)
• How do you find inodes in the log?
– Remember, there may be multiple copies of a given
inode
• Solution: add a level of indirection
– The traditional inode map can be broken into pieces
– When a portion of the inode map is updated, write it
to the log!

98
Memory inode Maps

inode
Data Data Data Data inode Data inode
map
Block 1 Block 2 Block 3 Block 4 1 Block 5 2
N
Disk

Giant Log

• New problem: the inode map is scattered

throughout the log
– How do we find the most up-to-date pieces?
99
 The Checkpoint Region
• The superblock in LFS contains pointers to all of
the up-to-date inode maps
– The checkpoint region is always cached in memory
– Written periodically to disk, say ~30 seconds
– Only part of LFS that isn’t maintained in the log

Data Data Data Data Data inode

Disk

inode inode
CR Block Block Block Block
1
Block
2
map
1 2 3 4 5 N

100
 How to Read a File in LFS
• Suppose you want to read inode 1
1. Look up inode 1 in the checkpoint region
• inode map containing inode 1 is in sector X
2. Read the inode map at sector X
• inode 1 is in sector Y
3. Read inode 1
• File data is in sectors A, B, C, etc.

Data Data Data Data Data inode

Disk

inode inode
CR Block Block Block Block
1
Block
2
map
1 2 3 4 5 N
101
 Directories in LFS
• Directories are stored just like in typical file systems
– Directory data stored in a file
– inode points to the directory file
– Directory file contains name  inode mappings

Data Data Data Data Dir inode

Disk

inode inode
CR Block Block Block Block
1
Data
2
map
1 2 3 4 1 N
102
 Garbage
• Over time, the log is going to fill up with stale
data
– Highly fragmented: live data mixed with stale data
• Periodically, the log must be garbage collected

Data Data inode Data inode Data inode

Disk

Block 1 Block 2 1 Block 5 2 Block 1 1

103
 Garbage Collection in LFS
• Each cluster has a summary block
– Contains the block  inode mapping for each block in the cluster
• check
• To Which liveness,
blocks arethestale?
GC reads each file with blocks in the
• Pointers from other
cluster
– Ifclusters are invisible
the current info doesn’t match the summary, blocks are stale
Memory

S D1 D1 i1 D2 i2

Summary block
Disk

S D1 D1 i1 D2 i2 S D1 i1 D3 D3 D3 S D1 D2 i2 i1

Cluster 1 Cluster 2 104

 An Idea Whose Time Has Come
• LFS seems like a very strange design
– Totally unlike traditional file system structures
– Doesn’t map well to our ideas about directory
heirarchies
• Initially, people did not like LFS
• However, today it’s features are widely used

105
 File Systems for SSDs
• SSD hardware constraints
– To implement wear leveling, writes must be spread
across the blocks of flash
– Periodically, old blocks need to be garbage collected
to prevent write-amplification
• Does this sounds familiar?
• LFS is the ideal file system for SSDs!
• Internally, SSDs manage all files in a LFS
– This is transparent to the OS and end-users
– Ideal for wear-leveling and avoiding write-
amplification
106
 Copy-on-write
• Modern file systems incorporate ideas from LFS
• Copy-on-write sematics
– Updated data is written to empty space on disk,
rather than overwriting the original data
– Helps prevent data corruption, improves sequential
write performance
• Pioneered by LFS, now used in ZFS and btrfs
– btrfs will probably be the next default file system in
Linux

107
 Versioning File Systems
• LFS keeps old copies of data by default
• Old versions of files may be useful!
– Example: accidental file deletion
– Example: accidentally doing open(file, ‘w’) on a file
full of data
• Turn LFS flaw into a virtue
• Many modern file systems are versioned
– Old copies of data are exposed to the user
– The user may roll-back a file to recover old versions

108