Running

1
Running FSCK in the Background

• Marshall Kirk McKusick
• USENIX BSDCon 2002, Feb

2
Abstract

• After uncontrolled system crash
• Traditional FFS
• Requires fsck
• Very time consuming
• Not suitable for servers when high availability
  is critical

3
Abstract

• After uncontrolled system crash
• FFS with soft update
• Only possible file system inconsistencies are
  some lost resources
• System can be brought up without fsck
• fsck can be performed whenever convenient
• But, also not suitable for 24x7 environment

4
Abstract

• Background fsck
• fsck can be performed anytime while system is
  fully functional
• HOW?
• First, take a snapshot of file system
• Run fsck on snapshot as background process

5
Snapshot
6
Snapshot

• Snapshot
• Author added snapshot capability in BSD
• As lost resources will not be used until they
  have been found and marked as free in bitmap,
• There are no limits on how long a background
  fsck can take to find and recover them
• Suspension time caused by snapshot creation
• Less than 1sec
• Independent of the size of file system

7
Snapshot

• Snapshot
• Other useful functionalities
• Ability to do reliable dumps of active file
  system
• Ability to provide back-ups of the file system
• These functionalities was first popularized by
  NetApp WAFL (Hitz el al, 1994)

8
Snapshot Performance

• Environment
• Computer 2 x Celeron 350MHz, 256M RAM
• OS FreeBSD 5.0
• Disk
• Small file system
• 0.5 Gb, 8K block, 1K fragment, 90 full, 70874
  files
• Large file system
• 7.7 Gb, 16K block, 2K fragment, 90 full, 520715
  files
• Load four continuously running simultaneous
  Andrew benchmarks

9
Snapshot Performance

• Results
• Snapshot creation times
• On idle file system

• On active file system

10
Snapshot Performance

• Results
• Snapshot removal times
• On idle file system

• On active file system

11
Background fsck
12
Background fsck

• Implementation
• Take a snapshot
• Run its traditional algorithms over snapshot
• 5 phases

13
Background fsck

• Problems when using background fsck
• Background fsck cannot update the on-disk image
  of file system
• Because, the file system state will be different
  from that of snapshot
• User-level program cannot obtain kernel-level
  locks needed to provide consistent updates of
  file system data structures

14
Background fsck

• Solution
• Five operations were implemented in the kernel
  (added system call)
• Set/clear superblock flags
• Adjust an inode block count
• Adjust an inode ref count
• Free a range of inodes
• Free a range of blocks/fragments

15
Background fsck Performance

• Results
• Traditional fsck times
• On idle file system

16
Background fsck Performance

• Results
• Background fsck times
• On idle file system

• On active file system