Deadlocks - II

1
Deadlocks - II
CS423UG Operating Systems

• Indranil Gupta
• Lecture 12
• Sep 21, 2005

2
Agenda

• Deadlock avoidance
• Deadlock detection

3
Review

• Resources can be of different types
• Necessary conditions for a Deadlock
• Mutual exclusion
• Hold and wait
• Non-preemption
• Circular wait
• Deadlock prevention
• Ensure at least one of conditions 1-4 is always
  false
• Use 2-phase locking

4
Deadlock Avoidance

• Each resource has multiple instances
• The system needs to know the maximum resource
  requirements of each process ahead of time
• Process p specifies, for each resource i,
  p.Maxi
• Bankers Algorithm (Dijkstra, 1965)
• Each customer tells banker the maximum number of
  resources it needs
• Customer borrows resources from banker
• Customer returns resources to banker
• Banker only lends resources if the system will be
  in a safe state after the loan
• Safe state - there is a lending sequence such
  that all customers will be able to take out a
  loan
• Unsafe state - there is a possibility of deadlock

5
Safe State and Unsafe State

• Safe State
• Can guarantee there is some scheduling order in
  which every process can run to completion even if
  all of them suddenly request their maximum number
  of resources immediately
• From safe state, the system can guarantee that
  all processes will finish
• Unsafe state no such guarantee
• A deadlock state is an unsafe state
• An unsafe state may not be a deadlock state
• Some process may be able to complete
• A conservative/pessimistic approach

6
How to Compute Safety

• Given
• n different types of resources
• Set P of processes, with each process p
  specifying p.Max1n
• For each p and i, the number of instances of
  resource i held by p currently p.Alloci
  (has to be lt p.Maxi)
• Current numbers of available resources of each
  type Available1n
• Let
• For each process p and resource i, let
  p.Needip.MAXi-p.Alloci
• while ( there exists a p in P such that
• for all i (p.Needi lt Availablei )
• )
• do for all i
• Availablei p.Alloci
• P P - p
• If P is empty then system is safe

7
Is Allocation (1 0 2) to P1 Safe?
If P1 requests max resources, can complete
8
Run Safety Test
If P1 requests max resources, can complete
9
Allocate to P1, Then

10
Release - P1 Finishes
Now P3 can acquire max resources and release
11
Release - P3 Finishes
Now P4 can acquire max resources and release
12
Release - P4 Finishes
Now P2 can acquire max resources and release
13
Release - P2 Finishes
Now P0 can acquire max resources and release
14
So P1 Allocation (1 0 2) Is Safe
15
Is allocation (0 2 0) to P0 Safe?
Try to Allocate 2 B to P0
16
Run Safety Test
No Processes may get max resources and release
17
Avoid Unsafe State- Do Not give (0,2,0) to P0
Return to Safe State and do not allocate resource
18
Suspend P0, Pending Request
When enough resources become available, P0 can be
woken
19
Results

• P0s request for 2 Bs cannot be granted because
  that would prevent any other process from
  completing if they need their maximum claim

20
Just Because Its Unsafe Doesnt mean

• P0 could have been allocated 2 Bs and a deadlock
  might not have occurred if
• P2 didnt use its maximum resources but finished
  using the resources it had

21
If no other process uses up Max
Then P0 would have finished
22
Tradeoff

• The bankers algorithm is conservative --- it
  reduces parallelism for safety sake

23
Banker Algorithm Issues

• Process may request more than claim
• A process may suffer indefinite postponement
  (starvation)
• Solution is to check for aged process
• Select an allocation sequence that includes aged
  process
• Only select requests that follow that sequence
  until aged process executes

24
Deadlock Detection

• Check to see if a deadlock has occurred!
• Single resource per type
• Can use wait-for graph
• Check for cycles
• O(n2)

25
Wait for Graphs
Resource Allocation Graph
Corresponding Wait For Graph
26
Multiple resources per type

• Run variant of bankers algorithm to see if
  processes can finish
• time O(mn2), m resources, n processes
• Optimistic version check only that at least one
  process can finish
• The deadlock detection is delayed
• Issues
• How often should the detection algorithm be run?
• How comprehensive should it be each time it is
  run?

27
Recovery From Deadlock

• Options
• Kill all deadlocked processes and release their
  resources
• Kill one deadlocked process at a time and release
  its resources, until no more deadlock
• Rollback (undo) all or one of the processes to a
  checkpoint that occurred before they requested
  any resources
• Note with rollback, starvation could occur
  (why?)

28
Deadlocks Summary

• In general, deadlock detection or avoidance is
  expensive
• Must evaluate cost of deadlock against detection
  or avoidance costs
• Deadlock avoidance and recovery may cause
  starvation
• Unix and Windows use Ostrich Algorithm

29
Reminders

• Reading for this lecture Chapter 3 (except
  Section 3.8)
• Reading for next lecture Sections 4.1-4.2
• Memory Management
• HW2 due this Friday.
• HW3 out this Friday, due in 1 week (short fuse!).
  Last HW before midterm.