CPSC313: Computer Hardware

and Operating Systems

Unit 4: File Systems
Representing files on disk

CPSC 313 1
 Admin

Quiz 4:
 Starts in a bit under two weeks.
 Don't forget to make your reservation.

Labs:
 Lab 8 is due Sunday November 17th.
 Lab 9 will be released tomorrow.

Next week:
 No lectures or office hours from Monday to Wednesday.
 No tutorials.
CPSC 313 2
Where we are

Unit Map:
 ...
 P20: File Systems implementation overview
 4.3. How we represent files
 P21: Why fixed-size block file systems?
 4.4. Building a file index
 P22: Getting File System metadata
 4.5. Naming
 ...
CPSC 313 3
Today

Learning Outcomes
 Define:

Buffer cache
 Compare/contrast different implementations of file indexing

Topics:
 Review how data moves between persistent storage and main memory
 Review the layers of abstraction in the file system
 Index structures for fixed-size block allocation

CPSC 313 4
Recall: From persistent media to memory
2. We will use main memory as a
cache for persistent data.

Execution
Cores
Execution
Cores
buffer cache
f f
a a

Memory
L1 Cache L1 Cache
s
s
t L2
Cache
t
really slow Disk
reasonably fast
(Persistent)

Mother board or system board 1. “Large” unit of transfer

(called an IO for
Input/Output):
>= sector (4KB)
The buffer cache:
• An in-memory cache of persistent data pages
• Managed by the file system (part of the operating system)
• A software cache (i.e., managed by SW not HW)
CPSC 313 5
 Posix API: hierarchical name space, byte-streams, open, close, read, write
open close read write

Map: Name to file Allocate and manage

metadata file descriptors

We are still here

Find location of in-memory file
file metadata object: VNODE

Map: file offsets to

disk blocks

Persistent storage: Numbered disk blocks, checksums and ECC, bad block handling

CPSC 313 6
 Our Current Part of the Layers-of-Abstraction:
File system API (open/close/read/write/seek) -- application programmers see this
byte offset o
a stream of bytes
byte 0 byte size - 1

The file system’s view: LBN = floor(o / file_system_block_size)

A sequence of logical blocks: the size of a block is whatever the file system wants
LBN 0 (logical block numbers) LBN max
= ceil(filesize /file system block size) - 1

The disk’s view: PBN = something … based on file’s block map or index
A sequence of (maybe not consecutive) disk sectors
PBN 0 (physical block numbers or disk addresses) PBN
max
CPSC 313 7
Recall: Constraints on Indexes

The index itself must be in persistent storage, or we couldn’t find
the blocks of our files on reboot.

Thus, indexes "live" in disk blocks, too. They take space!

As usual, we want to efficiently support:
 Sequential access
 Random access
 Sparse files
 Both small and large files

CPSC 313 8
 Recall: Impractical Strict Single-Extent-based
Data Structure:
Inode

The inode has the disk address of the first
nblocks
block in the extent and the length of the extent.


Pros:
…

 Very fast sequential access

 Minimal meta-data

Cons:
 Cannot (efficiently) support sparse files
 Leads to too much external fragmentation
 Does not match POSIX API

CPSC 313 9
 File Representation: Flat Index (practical?)
Data Structure:
Inode
Index

The index is a fixed sized array (else we have the memory
management problem we had with extents). Thus:
 The index consumes some number of disk blocks
 Growing the index is not possible
 There is some maximum file size

That's true for most indexes; however it's worse here.
Index

This inode has a reference to the (first block of) the index.
Or the design could have the index “live” inside the inode.
CPSC 313 10
Example (flat index)
File system parameters
 4096-byte blocks This means
maximum file size is 4096*16384=64MB
 4-byte block numbers 4096 byte blocks @ 4 bytes/entry = 1024 entries/block
16384 entries @ 1024 entries/block = 16 index blocks
 inode: 16384 index entries
Example of sparse file with 3 blocks
 LBN 2 is at PBN 123
 LBN 6 is at PBN 456
 LBN 3074 is at PBN 789
 Let’s draw this file’s meta data. (Use 0 to indicate a missing block.)
11
 Example (Flat index)
Inode
0 0
Example of sparse file with 3 blocks
1 0
2 123
LBN 2 is at PBN 123
3 0 123
4 0
LBN 6 is at PBN 456
5 0
LBN 3074 is at PBN 789
6 456
7 0 456
⁞

3073 0

3074 789
789
3075 0

16383 0
12
 Example (Flat index)
Inode
0 0
Example of sparse file with 3 blocks
1 0
2 123
LBN 2 is at PBN 123
3 0 123
4 0
LBN 6 is at PBN 456
5 0 LBN 3074 is at PBN 789
6 456
7 0 456
⁞

3073 0

3074 789 789

3075 0 Note: We draw these as pointers, but they are
⁞ not pointers-in-memory; they are disk addresses!
16383 0
13
 File Representation: Flat Index (practical?)
Data Structure:
Inode
Index

The index is a fixed sized array (else we have the memory
management problem we had with extents). Thus:
 The index consumes some number of disk blocks
 Growing the index is not possible
 There is some maximum file size

Index

Pros:
 Can represent sparse files (set block pointer to 0, which
is an invalid disk address).
 Sequential and random access are efficient in terms of
metadata blocks that need to be fetched.

Cons:
4-KB block, 4-byte block #, 1-GB max file size.
How big is the index?
How many blocks required for smallest non-
empty file?
CPSC 313 14
 File Representation: Flat Index (practical?)
Data Structure:
Inode
Index

The index is a fixed sized array (else we have the memory
management problem we had with extents). Thus:

al  The index consumes some number of disk blocks

ti c  Growing the index is not possible

 There is some maximum file size

ra c
Index 
Pros:

p  Can represent sparse files (set block pointer to 0, which

Im 
is an invalid disk address).
Sequential and random access are efficient in terms of
metadata blocks that need to be fetched.

Cons:
4-KB block, 4-byte block #, 1-GB max file size.
 Either we have to allocate really big indices or we
How big is the index? 256 blocks
impose unreasonable constraints on file size.
How many blocks required for smallest non-
empty file? 257 blocks  Small and large files consume exactly the same amount
CPSC 313
of index space. 15
 File Representation: Multi-level Index (Tree)
Data Structure:
Inode
• The inode stores a disk address.
• It may refer to a data block (for a file with one block).

In some versions, the inode always references a separate
index root, even for 1 (or perhaps 0) block files.

CPSC 313 16
 File Representation: Multi-level Index (Tree)
Data Structure:
Inode
• The inode stores a disk address.
• It may refer to a data block (for a file with one block).
• If there is more than one data block, it refers to a metadata
block packed with data addresses of blocks: an indirect block

We call these indirect blocks:

blocks that point to data blocks.

CPSC 313 17
 File Representation: Multi-level Index (Tree)
Inode
Data Structure:
• The inode stores a disk address.
• It may refer to a data block (for a file with one block).
• If there is more than one data block, it refers to a metadata block
packed with data addresses of blocks: an indirect block
• When that fills up, switch to storing the address of a block packed
with addresses of indirect blocks: a double-indirect block. Etc.

We call these double

indirect blocks: they
point to indirect blocks

indirect
blocks
CPSC 313 18
Example (Multi-level index)
Indirect blocks are also 4096 bytes.
File system parameters How many disk addresses can we store in a block?
4096/4 = 1024
 4096-byte blocks Files with more than 1024 blocks require a double-
 indirect block.
4-byte block numbers
 inode: stores one disk address (root of the tree)
Example of sparse file with 3 blocks
 LBN 2 is at PBN 123, indirect block is at 124
 LBN 6 is at PBN 456, indirect block is at 124
 LBN 3074 is at PBN 789, indirect block is at 345
 draw this file’s meta data : use 0 to indicate a block is missing.
19
 Indirect blocks
0 0 • LBN 2 is at PBN 123, indirect block is at 124

Example (Tree index) • LBN 6 is at PBN 456, indirect block is at 124

1 0
2 123
• LBN 3074 is at PBN 789, indirect block is at 345
3 0
4 0
Inode 5 0
129
Double-indirect block 6 456
0-1023 124 7 0
123

1024-2047 0 ⁞

2048-3071 0 1023 0
124
124 456
3072-4095 345

4096-5119 0

1047552-1048575 0

129

20
 Indirect blocks
0 0 • LBN 2 is at PBN 123, indirect block is at 124

Example (Tree index) • LBN 6 is at PBN 456, indirect block is at 124

1 0
2 123
3 0
• LBN 3074 is at PBN 789, indirect block is at 345
4 0
Inode 5 0
129
Double-indirect block 6 456
0-1023 124 7 0 123
⁞
1024-2047 0
1023 0
2048-3071 0
124
456
3072-4095 345
3072 0

4096-5119 0 3073 0
3074 789
⁞
3075 0
789
1047552-1048575 0 3076 0
3077 0
129 3078 0
3079 0

4095 0
21
345
 File Representation: Multi-level Index (Tree)
Inode

Pros:
 Can represent sparse files.
 If block size >> disk address size, an access
requires few intermediate blocks.
 Can grow easily.

Cons:
We call these double  Even for 2-block files, we perform two IOs
indirect blocks: they
point to indirect blocks
(one to get the root of the index and one to
get the data blocks, ignoring the inode!).
 The file size determines how deep the index
indirect
blocks
is, so it’s a bit more complicated to navigate
CPSC 313
the data structure. 22
 File Representation: Hybrid Index
Data Structure:
Inode • The index is a fixed-size array, small enough to fit in the inode.
• Most entries are direct pointers (point to data blocks)
• A few entries are single-, double-, or triple-indirect (or more)
Direct block
pointers

Double indirect
block pointer
…
…

Indirect block pointer

CPSC 313 23
 Widely used!
File Representation: Hybrid Index
Data Structure:
Inode • The index is a fixed-size array, small enough to fit in the inode.
• Most entries are direct pointers (point to data blocks)
• A few entries are single-, double-, or triple-indirect (or more)
Direct block
pointers

Double indirect
block pointer
…
…

Indirect block pointer

CPSC 313 24
Example (Hybrid Index)
File system parameters
• 4096-byte blocks
• 4-byte block numbers
• inode: 4 direct, 1 indirect, and 1 double-indirect blocks
Example of sparse file with 3 blocks
• LBN 2 is at PBN 123
• LBN 6 is at PBN 456 and indirect block is at 125
• LBN 3074 is at PBN 789, double indirect block is at 129, indirect at 345
• draw this file’s meta data : use 0 to indicate a block is missing.

25
 • LBN 2 is at PBN 123
• LBN 6 is at PBN 456, indirect at 125
Example (Hybrid index) • LBN 3074 is at PBN 789,
• double indirect at 129
Inode • indirect at 345
123
Direct block

0
pointers

0
123
0

indirect 125
129

double
indirect

26
 • LBN 2 is at PBN 123
• LBN 6 is at PBN 456, indirect at 125
Example (Hybrid index) 4 0
• LBN 3074 is at PBN 789,
5 0 • double indirect at 129
6 456
Inode • indirect at 345
123 7 0
Direct block

0 8 0
pointers

0 9 0
123 ⁞
0
1027 0
indirect 125
125 456
129

double
indirect Computing the index into the direct block:
 lookup = 6 // Initialize lookup to the LBN
 lookup -= 4 // Subtract the number of direct pointers

This is the index into our indirect block

27
 • LBN 2 is at PBN 123
• LBN 6 is at PBN 456, indirect at 125
Example (Hybrid index) 4 0
• LBN 3074 is at PBN 789,
5 0 • double indirect at 129
6 456
Inode • indirect at 345
123 7 0
Direct block

0 8 0
pointers

0 9 0
123 ⁞
0
1027 0
indirect 125
125
1028-2051 0
456
129

double 2052-3075 345

indirect 3076-4099 0

4100-5123 0 Why is the indirect at index 1?


lookup = LBN - nDirect (4) - nIndirect (1024) = 3074-1028 = 2046
5124-6047 0

dIndex = lookup / daddrs_per_block = 2046/1024 = 1
6048-7071 0

1047555-1048579 0

129 28
 • LBN 2 is at PBN 123
• LBN 6 is at PBN 456, indirect at 125
Example (Hybrid index) 4 0
• LBN 3074 is at PBN 789,
5 0 • double indirect at 129
6 456
Inode • indirect at 345
123 7 0
Direct block

0 8 0
pointers

0 9 0
123 ⁞
0 1027 0
indirect 125 125
1028-2051 0
456
129

double 2052-3075 345 Indirect block

indirect 3076-4099 0 2052 0
2053 0
4100-5123 0 789
2054 0
5124-6047 0 2055 0

6048-7071 0 ⁞

⁞ 3073 0
Why is the direct block located at index 1022? 3074 789
1047555-1048579 0

lookup = 2046 (last slide) 3075 0

iIndex = lookup % 1024129 = 1022 345 29
 Widely used!
File Representation: Hybrid Index
Data Structure:
Inode • The index is a fixed-size array, small enough to fit in the inode.
• Most entries are direct pointers (point to data blocks)
• A few entries are single-, double-, or triple-indirect (or more)
Direct block
pointers


Pros:
 Small index
Double indirect  Efficient for small files and sparse files
block pointer
…  Files can grow large
 Good for random/sequential if block size >> PBN
size
…


Cons:
…

Indirect block pointer  Slightly more complicated to map from LBN to

CPSC 313 30
PBN
File Structure Summary
Hybrid Index
Inode
Extent-based
Inode

Flat Index Inode Multi-level Index

Inode
Index

…
…
…

…
…
Impractical
CPSC 313 31
Exercise
• Compute lots of different things about the different representations
to:
• Develop a good understanding of how each structure works
• Get a sense of how they compare to each other.

CPSC 313 32
Wrapping Up

File index structures must be persistent
 Index structure (mostly) in block-sized units (exception is what is in the
inode)

Good design is:
 Space efficient for small files
 Performant for large files
 Flexible: allows for both sequential and random access

In our case studies, we’ll see examples and variants of the
designs we discussed today.
CPSC 313 33