Amer Filesys
Amer Filesys
Amer Filesys
File systems
Files
Directories & naming
File system implementation
Example file systems
Long-term information storage
Must store large amounts of data
Gigabytes > terabytes > petabytes
Stored information must survive the termination of
the process using it
Lifetime can be seconds to years
Must have some way of finding it!
Multiple processes must be able to access the
information concurrently
Naming files
Important to be able to find files after they’re created
Every file has at least one name
Name can be
Humanaccessible: “foo.c”, “my photo”, “Go Panthers!”, “Go Banana
Slugs!”
Machineusable: 4502, 33481
Case may or may not matter
Depends on the file system
Name may include information about the file’s contents
Certainly does for the user (the name should make it easy to figure out
what’s in it!)
Computer may use part of the name to determine the file type
Typical file extensions
File structures
1 record
1 byte
Sequence of bytes Sequence of records
S02 F01 W02
Tree
File types
Executable
file
Archive
Accessing a file
Sequential access
Read all bytes/records from the beginning
Cannot jump around
May rewind or back up, however
Convenient when medium was magnetic tape
Often useful when whole file is needed
Random access
Bytes (or records) read in any order
Essential for database systems
Read can be …
Move file marker (seek), then read or …
Read and then move file marker
File attributes
File operations
Create: make a new file Append: like write, but only
Delete: remove an existing at the end of the file
file Seek: move the “current”
Open: prepare a file to be pointer elsewhere in the file
accessed Get attributes: retrieve
Close: indicate that a file is attribute information
no longer being accessed Set attributes: modify
Read: get data from a file attribute information
Write: put data to a file Rename: change a file’s
name
Using file system calls
Using file system calls, continued
Memory-mapped files
Program Program
text text abc
Data Data xyz
Before mapping After mapping
Segmented process before mapping files into its address
space
Process after mapping
Existing file abc into one segment
Creating new segment for xyz
More on memory-mapped files
Memorymapped files are a convenient abstraction
Example: string search in a large file can be done just as
with memory!
Let the OS do the buffering (reads & writes) in the virtual
memory system
Some issues come up…
How long is the file?
Easy if readonly
Difficult if writes allowed: what if a write is past the end of file?
What happens if the file is shared: when do changes
appear to other processes?
When are writes flushed out to disk?
Clearly, easier to memory map readonly files…
Directories
Naming is nice, but limited
Humans like to group things together for
convenience
File systems allow this to be done with directories
(sometimes called folders)
Grouping makes it easier to
Find files in the first place: remember the enclosing
directories for the file
Locate related files (or just determine which files are
related)
Single-level directory systems
Root
directory
A A B C
foo bar baz blah
One directory in the file system
Example directory
Contains 4 files (foo, bar, baz, blah)
owned by 3 different people: A, B, and C (owners shown in red)
Problem: what if user B wants to create a file called foo?
Two-level directory system
Root
directory
A B C
A A B B C C C
foo bar foo baz bar foo blah
Solves naming problem: each user has her own directory
Multiple users can use the same file name
By default, users access files in their own directories
Extension: allow users to access files in others’ directories
Hierarchical directory system
Root
directory
A B C
A A A B B C C C
Papers foo Photos foo Papers bar foo blah
A A A B B
os.tex sunset Family foo.tex foo.ps
A A A
sunset kids Mom
Unix directory tree
Operations on directories
Create: make a new Readdir: read a directory
directory entry
Delete: remove a directory Rename: change the name
(usually must be empty) of a directory
Opendir: open a directory Similar to renaming a file
to allow searching it Link: create a new entry in
Closedir: close a directory a directory to link to an
(done searching) existing file
Unlink: remove an entry in
a directory
Remove the file if this is the
last link to this file
File system implementation issues
How are disks divided up into file systems?
How does the file system allocate blocks to files?
How does the file system manage free space?
How are directories handled?
How can the file system improve…
Performance?
Reliability?
Carving up the disk
Entire disk
Partition table
Master
Partition 1 Partition 2 Partition 3 Partition 4
boot record
A Free C Free E F
Contiguous allocation requires all blocks of a file to be
consecutive on disk
Problem: deleting files leaves “holes”
Similar to memory allocation issues
Compacting the disk can be a very slow procedure…
Contiguous allocation
Data in each file is stored in 0 1 2 3
consecutive blocks on disk
Simple & efficient indexing
Starting location (block #) on disk
(start) 4 5 6 7
Length of the file in blocks
(length)
Random access wellsupported
Difficult to grow files
Must preallocate all needed space
8 9 10 11
Wasteful of storage if file isn’t
using all of the space
Logical to physical mapping is easy
blocknum = (pos / 1024)
+ start; Start=5
offset_in_block = pos % Length=2902
1024;
Linked allocation
0 1 2 3
File is a linked list of disk
4 6
blocks
Blocks may be scattered
around the disk drive 4 5 6 7
Block contains both pointer
x x
to next block and data
Files may be as long as
needed 8 9 10 11
New blocks are allocated as 0
needed
Linked into list of blocks in
file Start=9 Start=3
Removed from list (bitmap) End=4 End=6
of free blocks Length=2902 Length=1500
Finding blocks with linked allocation
Directory structure is simple
Starting address looked up from directory
Directory only keeps track of first block (not others)
No wasted space all blocks can be used
Random access is difficult: must always start at first block!
Logical to physical mapping is done by
block = start;
offset_in_block = pos % 1020;
for (j = 0; j < pos / 1020; j++) {
block = block->next;
}
Assumes that next pointer is stored at end of block
May require a long time for seek to random location in file
Linked allocation using a RAM-based table
0 4
1 1
Links on disk are slow
2 1 Keep linked list in memory
3 2 Advantage: faster
4 2
5 1
Disadvantages
6 3 B
Have to copy it to disk at
7 1 some point
8 1 Have to keep inmemory and
9 0 A ondisk copy consistent
10 1
11 1
12 1
13 1
14 1
15 1
Using a block index for allocation
Store file block addresses in Name index size
an array grades 4 4802
Array itself is stored in a disk
block
0 1 2 3
Directory has a pointer to this
disk block
Nonexistent blocks indicated
by 1 6
4 5 6 7
Random access easy 9
Limit on file size? 7
0
8 8 9 10 11
Finding blocks with indexed allocation
Need location of index table: look up in directory
Random & sequential access both wellsupported:
look up block number in index table
Space utilization is good
No wasted disk blocks (allocate individually)
Files can grow and shrink easily
Overhead of a single disk block per file
Logical to physical mapping is done by
block = index[block % 1024];
offset_in_block = pos % 1024;
Limited file size: 256 pointers per index block, 1 KB
per file block > 256 KB per file limit
Larger files with indexed allocation
How can indexed allocation allow files larger than a single
index block?
Linked index blocks: similar to linked file blocks, but using
index blocks instead
Logical to physical mapping is done by
index = start;
blocknum = pos / 1024;
for (j = 0; j < blocknum /255); j++) {
index = index>next;
}
block = index[blocknum % 255];
offset_in_block = pos % 1024;
File size is now unlimited
Random access slow, but only for very large files
Two-level indexed allocation
Allow larger files by creating an index of index blocks
File size still limited, but much larger
Limit for 1 KB blocks = 1 KB * 256 * 256 = 226 bytes = 64 MB
Logical to physical mapping is done by
blocknum = pos / 1024;
index = start[blocknum / 256)];
block = index[blocknum % 256]
offset_in_block = pos % 1024;
Start is the only pointer kept in the directory
Overhead is now at least two blocks per file
This can be extended to more than two levels if larger files
are needed...
Block allocation with extents
Reduce space consumed by index pointers
Often, consecutive blocks in file are sequential on disk
Store <block,count> instead of just <block> in index
At each level, keep total count for the index for efficiency
Lookup procedure is:
Find correct index block by checking the starting file offset for each
index block
Find correct <block,count> entry by running through index block,
keeping track of how far into file the entry is
Find correct block in <block,count> pair
More efficient if file blocks tend to be consecutive on disk
Allocating blocks like this allows faster reads & writes
Lookup is somewhat more complex
Managing free space: bit vector
Keep a bit vector, with one entry per file block
Number bits from 0 through n1, where n is the number of file blocks
on the disk
If bit[j] == 0, block j is free
If bit[j] == 1, block j is in use by a file (for data or index)
If words are 32 bits long, calculate appropriate bit by:
wordnum = block / 32;
bitnum = block % 32;
Search for free blocks by looking for words with bits unset
(words != 0xffffffff)
Easy to find consecutive blocks for a single file
Bit map must be stored on disk, and consumes space
Assume 4 KB blocks, 8 GB disk => 2M blocks
2M bits = 221 bits = 218 bytes = 256KB overhead
Managing free space: linked list
Use a linked list to manage free blocks
Similar to linked list for file allocation
No wasted space for bitmap
No need for random access unless we want to find
consecutive blocks for a single file
Difficult to know how many blocks are free unless
it’s tracked elsewhere in the file system
Difficult to group nearby blocks together if they’re
freed at different times
Less efficient allocation of blocks to files
Files read & written more because consecutive blocks not
nearby
Issues with free space management
OS must protect data structures used for free space
management
OS must keep inmemory and ondisk structures consistent
Update free list when block is removed: change a pointer in the
previous block in the free list
Update bit map when block is allocated
Caution: ondisk map must never indicate that a block is free when it’s
part of a file
Solution: set bit[j] in free map to 1 on disk before using block[j] in a file
and setting bit[j] to 1 in memory
New problem: OS crash may leave bit[j] == 1 when block isn’t actually
used in a file
New solution: OS checks the file system when it boots up…
Managing free space is a big source of slowdown in file
systems
What’s in a directory?
Two types of information
File names
File metadata (size, timestamps, etc.)
Basic choices for directory information
Store all information in directory
Fixed size entries
Disk addresses and attributes in directory entry
Store names & pointers to index nodes (inodes)
attributes
games attributes games
mail attributes mail attributes
news attributes news
attributes
research attributes research
attributes
Storing all information Using pointers to
in the directory index nodes
Directory structure
Structure
Linear list of files (often itself stored in a file)
Simple to program
Slow to run
Increase speed by keeping it sorted (insertions are slower!)
Hash table: name hashed and looked up in file
Decreases search time: no linear searches!
May be difficult to expand
Can result in collisions (two files hash to same location)
Tree
Fast for searching
Easy to expand
Difficult to do in ondisk directory
Name length
Fixed: easy to program
Variable: more flexible, better for users
Handling long file names in a directory
Sharing files
Root
directory
A B C
A A A B B C C C
Papers foo Photos foo Photos bar foo blah
A A A B
os.tex sunset Family lake
A A ?
sunset kids ???
Solution: use links
A creates a file, and inserts into her directory
B shares the file by creating a link to it
A unlinks the file
B still links to the file
Owner is still A (unless B explicitly changes it)
A A B B
b.tex b.tex
a.tex a.tex
Block size
Dark line (left hand scale) gives data rate of a disk
Dotted line (right hand scale) gives disk space efficiency
All files 2KB
Disk quotas
Backing up a file system
A file system to be dumped
Squares are directories, circles are files
Shaded items, modified since last dump
Each directory & file labeled by inode number
Consistent Missing (“lost”) block
Duplicate block in free list Duplicate block in two files
File system cache
Many files are used repeatedly
Option: read it each time from disk
Better: keep a copy in memory
File system cache
Set of recently used file blocks
Keep blocks just referenced
Throw out old, unused blocks
Same kinds of algorithms as for virtual memory
More effort per reference is OK: file references are a lot less
frequent than memory references
Goal: eliminate as many disk accesses as possible!
Repeated reads & writes
Files deleted before they’re ever written to disk
File block cache data structures
Grouping data on disk
Log-structured file systems
Trends in disk & memory
Faster CPUs
Larger memories
Result
More memory > disk caches can also be larger
Increasing number of read requests can come from cache
Thus, most disk accesses will be writes
LFS structures entire disk as a log
All writes initially buffered in memory
Periodically write these to the end of the disk log
When file opened, locate inode, then find blocks
Issue: what happens when blocks are deleted?
Unix Fast File System indexing scheme
protection mode data
owner & group data
timestamps ...
size data
block count • data
link count
•
•
...
...
Direct pointers
•
•
•
data
data
•
• •
...
single indirect • • data
• •
double indirect • data
...
• •
•
triple indirect • • •
• • • data
inode
More on Unix FFS
First few block pointers kept in directory
Small files have no extra overhead for index blocks
Reading & writing small files is very fast!
Indirect structures only allocated if needed
For 4 KB file blocks (common in Unix), max file sizes are:
48 KB in directory (usually 12 direct blocks)
1024 * 4 KB = 4 MB of additional file data for single indirect
1024 * 1024 * 4 KB = 4 GB of additional file data for double indirect
1024 * 1024 * 1024 * 4 KB = 4 TB for triple indirect
Maximum of 5 accesses for any file block on disk
1 access to read inode & 1 to read file block
Maximum of 3 accesses to index blocks
Usually much fewer (12) because inode in memory
Directories in FFS
Directories in FFS are just Directory
special files inode number
Same basic mechanisms record length
Different internal structure name length
Directory entries contain
File name name
Inode number
Other Unix file systems inode number
have more complex record length
schemes name length
Not always simple files…
name
CD-ROM file system
Directory entry in MS-DOS
MS-DOS File Allocation Table
Bytes
Checksum
Storing a long name in Windows 98
Long name stored in Windows 98 so that it’s backwards
compatible with short names
Short name in “real” directory entry
Long name in “fake” directory entries: ignored by older systems
OS designers will go to great lengths to make new systems
work with older systems…