Automatic Checkpointing

1
Automatic Checkpointing

• Sean Stanek
• CprE 545

2
What is it?

• Saves complete checkpoint of program without
  requiring additional programming
• Advantages
• Easy to use
• Programmer does not need to worry about it
• Disadvantages
• May not work as desired
• May not work at all

3
How is it done?

• Process in memory has a state
• Data
• Code
• Stack
• Processor State
• OS/Process State

4
How is it done?

• Data
• Used to store a variety of static and variable
  data
• Data memory is composed of one or more segments
  of bytes
• Statically declared memory
• Variable memory
• Undeclared memory

5
How is it done?

• Code
• Binary data that defines a program
• Program code is actually the same as data memory
• Typically has different attributes on its memory
  segment to allow it to execute code
• Need to save these additional attributes

6
How is it done?

• Stack
• Stores local variables
• Stores procedure return addresses
• Stack is also just memory, but it is in a
  different location

7
How is it done?

• Processor State
• Registers
• General purpose registers
• Extra registers (FPU, etc.)
• Program counter (Instruction pointer)
• Special purpose registers
• Flags
• Potentially special attributes about the process

8
How is it done?

• OS/Process State
• Open files
• File positions
• File buffers
• File attributes
• Network connections
• Open libraries

9
How is it done?

• OS/Process State (continued)
• Process State
• Hooked signals
• Message queue
• Timers/Alarms
• OS/API hooks

10
When can it be used?

• Before starting a process
• Process Snooper / Spy
• Perform timed checkpoint
• After starting a process
• Attach to process to spy
• Timed checkpoint (but not always running)
• One-time checkpoint

11
How can it recover?

• Reload code, data, and stack segments
• Reload processor states
• Reload process states
• Reload OS states
• Continue execution of code from checkpoint
• Not always recoverable

12
What problems are there?

• An open file on disk has changed since last
  checkpoint
• Open file positions may be stale
• Buffered file data may be stale
• Might corrupt data by writing back to disk
• Just keep copy of file?
• Too much data
• Does not solve a bigger problem

13
What problems are there?

• Open network connections
• Impossible to recover remote state
• Both sides would have to checkpoint at same time
• Still probably would not fix the problem

14
What problems are there?

• Checkpoint after a fault but before detection
• Program may crash as a result of a fault from
  long before
• Keep multiple checkpoints?
• Would use very large amount of data
• If restricting number of checkpoints kept, still
  does not guarantee a working checkpoint
• May be difficult to find last correct checkpoint

15
Functions to Manage Checkpoints

• Unix/Linux
• ptrace() system call
• PTRACE_ATTACH
• PTRACE_PEEKDATA
• PTRACE_POKEDATA
• etc.
• Win32
• AttachProcess()
• ReadProcessMemory()
• WriteProcessMemory()
• etc.

16
CPU/OS specific data

• Would require additional programming for every
  new CPU and OS
• May not be able to retrieve all vital process
  states given user privilege level
• Might be easier to modify the OS to retrieve all
  this information, but then would not be an
  easy-to-use user mode process

17
Additional Uses

• If a program crashes, may be able to reload
  checkpoint in debug mode to track down bugs
• Word processing (already implemented, manually)
• Could recover other vital program data otherwise
  lost to program crash