Transcript Kernel Modules - Northern Kentucky University

CSC 660: Advanced OS
Scheduling
CSC 660: Advanced Operating Systems
Slide #1
Topics
1. Basic Concepts
2. Scheduling Policy
3. The O(1) Scheduler
4. Runqueues
5. Priority Arrays
6. Calculating Priorities and Timeslices.
7. Scheduler Interrupts.
8. Sleeping and Waking.
9. The schedule() function
10. Multiprocessor Scheduling
11. Soft Realtime Scheduling
CSC 660: Advanced Operating Systems
Slide #2
Basic Concepts
Scheduler
Selects a process to run and allocates CPU to it.
Provides semblence of multitasking on single CPU.
Scheduler is invoked when:
Process blocks on an I/O operation.
A hardware interrupt occurs.
Process time slice expires.
Kernel thread yields to scheduler.
CSC 660: Advanced Operating Systems
Slide #3
Types of Processes
CPU Bound
Spend most time on computations.
Example: computer algebra systems.
I/O Bound
Spend most time on I/O.
Example: word processor.
Mixed
Alternate CPU and I/O activity.
Example: web browser.
CSC 660: Advanced Operating Systems
Slide #4
Alternating CPU and I/O Bursts
CSC 660: Advanced Operating Systems
Slide #5
Scheduling Policy
Scheduler executes policy, determining
1. When threads can execute.
2. How long threads can execute.
3. Where threads can execute.
CSC 660: Advanced Operating Systems
Slide #6
Scheduling Policy Goals
• Efficiency
– Maximize amount of work accomplished.
• Interactivity
– Respond as quickly as possible to user.
• Fairness
– Don’t allow any process to starve.
CSC 660: Advanced Operating Systems
Slide #7
Which goal is most important?
Depends on the target audience:
Desktop: interactivity
But kernel shouldn’t spend all its time in context
switch.
Server: efficiency
But should offer interactivity in order to serve
multiple users.
CSC 660: Advanced Operating Systems
Slide #8
Pre-2.6 Scheduler
O(n) algorithm at every process switch:
1. Scanned list of runnable processes.
2. Computed priority of each task.
3. Selected best task to run.
CSC 660: Advanced Operating Systems
Slide #9
The O(1) Scheduler
Replacement for O(n) 2.4 scheduler.
All algorithms run in constant time.
New data structures: runqueues and priority arrays.
Performs work in small pieces.
Additional new features
Improved SMP scalability, including NUMA.
Better processor affinity.
SMT scheduling.
CSC 660: Advanced Operating Systems
Slide #10
Runqueues
List of runnable processes on a processor.
Each runnable process is a member of
precisely one runqueue.
Runqueue data:
Lock to prevent concurrency problems.
Pointers to current and idle tasks.
Priority arrays which contain actual tasks.
Statistics
CSC 660: Advanced Operating Systems
Slide #11
Runqueues
struct runqueue {
spinlock_t lock;
unsigned long nr_running;
unsigned long long nr_switches;
unsigned long expired_timestamp,
nr_uninterruptible;
unsigned long long timestamp_last_tick;
task_t *curr, *idle;
struct mm_struct *prev_mm;
prio_array_t *active, *expired, arrays[2];
int best_expired_prio;
atomic_t nr_iowait;
}
CSC 660: Advanced Operating Systems
Slide #12
Priority Arrays
Each runqueue contains 2 priority arrays
Active array
Expired array
Basis for O(1) performance:
Scheduler always runs highest priority task.
Round robin for multiple equal priority tasks.
Priority array finds highest task O(1) operation.
Using two arrays allows transitions between
epochs by switching active and expired pointers.
CSC 660: Advanced Operating Systems
Slide #13
Priority Arrays
struct prio_array {
/* # of runnable tasks in array */
unsigned int nr_active;
/* bitmap: pri lvls contain tasks */
unsigned long bitmap[BITMAP_SIZE];
/* 1 list_head per priority (140) */
struct list_head queue[MAX_PRIO];
};
CSC 660: Advanced Operating Systems
Slide #14
Finding Highest Priority Task
1. Find first bit set in bitmap.
sched_find_first_bit()
2. Read corresponding queue[n]
If one process, give CPU to that one.
If multiple processes, round-robin schedule all
processes in queue for that priority.
idx = sched_find_first_bit(array->bitmap);
queue = array->queue + idx;
next = list_entry(queue->next, task_t, run_list);
CSC 660: Advanced Operating Systems
Slide #15
What if no runnable task exists?
System runs the swapper task (PID 0).
Each CPU has its own swapper process.
CSC 660: Advanced Operating Systems
Slide #16
Running out of Timeslice
1.
2.
3.
4.
Remove task from active priority array.
Calculate new priority and timeslice.
Add task to expired priority array.
Swap arrays when active array is empty.
array = rq->active;
if (unlikely(!array->nr_active)) {
rq->active = rq->expired;
rq->expired = array;
...
}
CSC 660: Advanced Operating Systems
Slide #17
Static and Dynamic Priorities
Initial priority value called the nice value.
Set via the nice() system call.
Static priority is nice value + 120.
Stored in current->static_prio.
Ranges from 100 (highest) to 139 (lowest).
Scheduling based on dynamic priority.
Bonuses and penalties according to interactivity.
Stored in current->prio.
Calculated by effective_prio() function.
CSC 660: Advanced Operating Systems
Slide #18
Dynamic Priority Policy
Increase priority of interactive processes.
Favor I/O-bound over CPU-bound.
Need heuristic for determining interactivity.
Use time spent sleeping vs. runnable time.
Sleep average
Stored in current->sleep_avg.
Incremented when task becomes runnable.
Decremented for each timer tick task runs.
Scaled to produce priority bonus ranging 0..10.
CSC 660: Advanced Operating Systems
Slide #19
Calculating Priority
/* Scale sleep_avg to range 0..MAX_BONUS */
#define CURRENT_BONUS(p) \
(NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
MAX_SLEEP_AVG)
static int effective_prio(task_t *p)
{
int bonus, prio;
bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
prio = p->static_prio - bonus;
return prio;
}
CSC 660: Advanced Operating Systems
Slide #20
Time Slices
Time slice duration critical to performance.
Too short: high overhead from context switches.
Too long: loss of apparent multitasking.
Interactive processes and time slices
Interactive processes have high priority.
Pre-empt CPU bound tasks on kbd/ptr interrupts.
Long time slices slow start of new tasks.
CSC 660: Advanced Operating Systems
Slide #21
Calculating Timeslice
Initial Timeslice
On fork(), parent + child divide remaining time evenly.
Stored in current->time_slice.
Recalculating Timeslices
Time Slice = (140 – static priority) x 20 if static < 120
= (140 – static priority) x 5 if static >= 120
Description
Highest
Nice
-20
Static Pri
100
Time Slice
800ms
Default
Lowest
0
+19
120
139
100ms
5ms
CSC 660: Advanced Operating Systems
Slide #22
Scheduler Interrupts
• Scheduler interrupt: scheduler_tick()
– Invoked every 1ms by a timer interrupt.
• Decrements task’s time slice.
• If a higher priority task exists,
– Higher priority task is given CPU.
– Current task remains in TASK_RUNNING state.
• If time slice expired,
– Moved to expired priority array.
– If highly interactive, may be re-inserted into active
priority array.
CSC 660: Advanced Operating Systems
Slide #23
Sleeping and Waking
Sleeping tasks are not in runqueues.
Require no CPU time until awakened.
Why sleep?
Waiting for I/O.
Waiting for other hardware events.
Waiting for a kernel semaphore.
CSC 660: Advanced Operating Systems
Slide #24
Sleeping
DECLARE_WAITQUEUE(wait, current);
/* q is a wait queue, wait is a q entry */
add_wait_queue(q, &wait);
while (!condition) {
set_current_state(TASK_INTERRUPTIBLE);
if (signal_pending(current))
/* Handle signal */
schedule()
}
set_current_state(TASK_RUNNING);
remove_wait_queue(q, &wait);
CSC 660: Advanced Operating Systems
Slide #25
Waking
wake_up() wakes up tasks on event
Exclusive: only wakes up one task on waitqueue
Non-exclusive: wakes all tasks on waitqueue
add_wait_queue
TASK_RUNNING
Signal
TASK_INTERRUPTIBLE
wake_up
CSC 660: Advanced Operating Systems
Slide #26
Multiprocessor Architectures
Classic
Memory shared by all CPUs.
Hyperthreading
Single CPU executing multiple on-chip threads.
NUMA
CPUs + RAM grouped in local nodes.
Reduces contention for accessing RAM.
Fast to access local RAM.
Slower to access remote RAM.
CSC 660: Advanced Operating Systems
Slide #27
Multiprocessor Scheduling
Each CPU has own runqueue.
Scheduler selects tasks from local runqueue.
CPU cache more likely to still be hot.
Periodic checks to balance load across CPUs.
Called by rebalance_tick().
Loops over all scheduling domains.
Calls load_balance() if balance interval
expired.
CSC 660: Advanced Operating Systems
Slide #28
load_balance()
1. Acquires this_rq->lock spin lock.
2.
3.
4.
5.
6.
Finds busiest CPU with > 1 process.
If no busiest or current CPU is busiest, terminates.
Obtains spin lock on busiest CPU.
Pull tasks from busiest CPU to local runqueue.
Releases locks.
CSC 660: Advanced Operating Systems
Slide #29
move_tasks()
Searches for runnable tasks in expired
runqueue.
Then scans active runqueue.
Call pull_task() to move task if all true:
Task not currently being executed.
Local CPU is in cpus_allowed bitmask.
At least one of the following is true:
Local CPU is idle.
Multiple attempts to move processes have failed.
Process is not cache hot.
CSC 660: Advanced Operating Systems
Slide #30
Realtime Scheduling
Hard Real-time
Guaranteed response within defined period.
Used for embedded systems: car engines.
Ex: RealTime Application Interface (RTAI)
Soft Real-time
Best effort to meet scheduling constraints.
Used for multimedia applications.
Currently provided by Linux.
Improved by Realtime Preemption Patch.
CSC 660: Advanced Operating Systems
Slide #31
Soft Realtime Scheduling
Scheduling Priorities
RT have higher priorities than any non-RT tasks.
RT priorities are static, ranging 1-99, not dynamic.
If RT tasks are runnable, no other tasks can run.
Scheduling Policies
SCHED_NORMAL (non-realtime)
SCHED_FIFO
SCHED_RR
CSC 660: Advanced Operating Systems
Slide #32
Realtime Policies
SCHED_FIFO
First-in First-out real-time Scheduling
Process uses CPU until:
It blocks or yields the CPU voluntarily.
A higher priority real-time process pre-empts it.
SCHED_RR
Round Robin real-time scheduling.
Process runs for time slice, then waits for other
equal priority real-time processes in runqueue.
CSC 660: Advanced Operating Systems
Slide #33
Realtime Process Replacement
Realtime processes replaced only when:
Pre-empted by a high-priority RT process.
Process performs a blocking operation.
Process is stopped or killed by a signal.
Process invokes sched_yield() system call.
SCHED_RR process has exhausted its time slice.
CSC 660: Advanced Operating Systems
Slide #34
Realtime System Calls
Scheduler Policy
sched_setscheduler()
sched_getscheduler()
Priority
sched_getparam()
sched_setparam()
CSC 660: Advanced Operating Systems
Slide #35
Yielding the Processor
sched_yield() system call
Moves regular task to expired priority array.
RT tasks moved to end of priority list.
Kernel tasks can yield the CPU too.
Call yield() function.
CSC 660: Advanced Operating Systems
Slide #36
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
Josh Aas, “Understanding the Linux 2.6.8.1 Scheduler,”
http://josh.trancesoftware.com/linux/, 2005.
Daniel P. Bovet and Marco Cesati, Understanding the Linux Kernel,
3rd edition, O’Reilly, 2005.
Corbet, “Realtime preemption and read-copy-update,” Linux Weekly
News, http://lwn.net/Articles/129511/, March 29, 2005.
Robert Love, Linux Kernel Development, 2nd edition, Prentice-Hall,
2005.
Claudia Rodriguez et al, The Linux Kernel Primer, Prentice-Hall,
2005.
RTAI, http://www.rtai.org/, 2006.
Peter Salzman et. al., Linux Kernel Module Programming Guide,
version 2.6.1, 2005.
Avi Silberchatz et. al., Operating System Concepts, 7th edition, 2004.
Andrew S. Tanenbaum, Modern Operating Systems, 3rd edition,
Prentice-Hall, 2005.
CSC 660: Advanced Operating Systems
Slide #37