Transcript osd06

CS 6560 Operating System
Design
Lecture 6:
Interrupts: Bottom Halves
Kernel Synchronization
References
• Our textbook: Robert Love, Linux Kernel
Development, 2nd edition, Novell Press,
2005.
• Understanding the LINUX Kernel, 3rd.
edition, O’Reilly, 2005. (covers 2.6)
• The kernel code and its own documentation.
Plan
• Chap 7: Bottom halves
• Chap 8: Concurrency Issues and Principles
• Chap 9: Concurrency methods in Linux 2.6
Recall: Top and Bottom Halves
• Conflicting needs for interrupts:
– Return quickly
– Get lots of work done
• Solution: Break up the interrupt processing into
two parts
– Top half: returns quickly, after setting up work to be
done
– Bottom half: scheduled by the kernel to get work done
at a later time
Bottom Halves
• Bottom Halves = General term for
mechanisms to defer work generated by
interrupt handlers (also an old name for an
obsolete mechanism for doing this)
When to run Bottom Half?
• When?
– In interrupt context - no sleeping allowed
– In process context - sleeping allowed
• Advantages to each, thus Linux 2.6 offers
both types of bottom halves.
Process and Interrupt Contexts
• Process Context: Execution of code in the
kernel that is scheduled by the process
scheduler.
• Interrupt Context: Execution of code in the
kernel that is responding to an interrupt and
is not scheduled by the process scheduler.
Bottom Half Mechanisms In
Linux 2.6
• The following are current mechanisms
– Softirq
– Tasklets
– Work queues
• Previous versions of the kernel had others
that are now obsolete:
– BH - removed in 2.5 kernel
– Task queues - removed in 2.5 kernel
Softirq
•
•
•
•
Runs in interrupt context - no blocking allowed
Statically defined at compile time - limited to 32
Available since 2.3 version of kernel
Handles a few special cases such as network and
scsi disk processing. Also handles tasklets
• Code is mainly in kernel/softirq.c (and other
places such as include/linux/irq_cpustat.h)
Properties
• A softirq is never preempted by another
softirq - only by an interrupt
• A softirq can run on more than one
processor at a time.
Implementation of Softirqs
• Data structure: softirq_action
– Has function reference and data reference
• Static array of these: softirq_vec[32]
• Handler function has simple prototype:
void softirq_handler(struct softirq_action *)
Registering Softirqs
• Softirqs are enumerated at compile time
• They are registered via the open_softirq
function.
• This just places them in the softirq_vec.
When do softirqs run?
• When needed, they are marked for running
with the raise_softirq function. This turns
on a bit in the softirq_pending word
• They are later run by the kernel with the
do_softirq function. This is called
– in the return from hardware interrupt code
– in the ksoftirq kernel thread
– elsewhere such as in the networking subsystem
Tasklets
• Tasklets is a system that is implemented on
top of soft_irq.
• While soft_irqs are reserved for certain
purposes, tasklets are the recommended
way to do low overhead bottom half
processing.
• They are dynamically generated and
managed.
Tasklet Operations
• Declaring: DECLARE_TASKLET and
DECLARE_TASKLET_DISABLED
• Scheduling (mark as pending):
tasklet_schedule
• Disabling: tasklet_disable and
tasklet_disable_nosync
• Enabling: tasklet_enable
• Removing: tasklet_kill
Workqueues
• Workqueues run in process context and contain a list of
work to be worked on.
• Each unit of work has a function and a pointer to some
data.
• Each workqueue has a name, a list of work, and a set of
worker threads, one for each CPU. The threads are named
after the name of the workqueue.
• Initially, there is one workqueue whose worker threads are
called “events”. Individual threads are called “events/0”,
“events/1”, and so on.
Workqueue Operations
• Creating Workqueues: create_workqueue makes new workqueue with
given name.
• Creating work: (give name, function, data)
– Statically: DECLARE_WORK
– Dynamically: INIT_WORK
• Schedule work:
– Generically: queue_work and queue_delayed_work
– Default workqueue: schedule_work and schedule_work_delayed
• Flush work:
– Generically: flush_workqueue
– Default workqueue: flush_scheduled_work
• Cancel work: cancel_delayed_work
Kernel Synchronization
• Chapters 8 and 9
• Issue: Sharing data in a multiprocessor,
preemptive environment
• Solutions:
– Limit sharing
– Use locks
Concurrency Issues
• True Concurrency: Multiple processors be
running code at the same time. They may
contend for the same data.
• Preemption: Running code may be
interrupted by the running of other code.
Running the new code may contend for the
same data as running the old.
Blocking and Sleeping
• Blocking: Code may stop running and be
waiting for an event to occur.
• Sleeping: This can only happen if there is a
process that is executing it.
Scalability
• Adding more processors adds more
overhead. So a quad core system does not
run four times faster.
Granularity
• Graduality: How much data should be
protected?
– Bits
– Words
– Structures
Danger of Data Contention
• Danger of data contention is corruption.
• Example: updates - read, modify, write
Solutions
• Don’t share:
• Don’t use global variables.
• Have separate copies of variables for each CPU
• Carefully design how data is shared and use a
variety of locking and synchronization methods.
Basic Principles
• Protect data not code.
• Critical region: Code that accesses shared
data
• Typical protection: bracket critical regions
Linux 2.6 uses
•
•
•
•
•
•
•
•
•
Per CPU copies of variables
Preemption disabling
Interrupt disabling
atomic operations
spinlocks
Semaphores
Completion variables
Seq locks
Barriers
Per CPU Variables
• Covered under memory management in
LKD, Chapter 11.
• Linux 2.4 used an explicit array, indexed by
CPU. The function get_cpu gives the index
number for the current CPU and disables
preemption. The function put_cpu enables
preemption.
• Linux 2.6 has a percpu interface.
Atomic Operations
• Linux 2.6 has atomic integer variables and
atomic bitwise operations
Atomic Integer Operations
• ATOMIC_INIT(int i)// assign the type atomic_t to i
• int atomic_read(atomic_t *v)// returns value of v atomically
• void atomic_set(atomic_t *v, int i)// atomically sets v equal to the
value of I
• void atomic_add(int i, atomic_t *v)// atomically add i to v
• void atomic_sub(int i, atomic_t *v)// atomically subtract i from v
• void atomic_inc(atomic_t *v)// atomically increment v
• void atomic_dec(atomic_t *v)// atomically decrement v
• And more - see page 133 for details on atomic_sub_and_test,
atomic_add_negative, atomic_dec_and_test, atomic_inc_and_test
Atomic Bitwise Operations
• Operations
–
–
–
–
Set, clear and flip individual bits
Test and set, clear, and flip individual bits
Return a bit value
Find first set bit and first cleared bit
• Atomic and non-atomic forms (prefix by double
underscore)
• Danger - bit position can overflow actual data
Spin Locks
• Why?
– Way of waiting without sleep
– Used in SMP systems
• Notes:
– sleep has high overhead and is not available in
interrupt context
– Spin locks must be only be held for a short time
- maybe five lines of code
Spin Lock Use
• Bracket critical region with spin_lock and
spin_unlock
• Example:
spin_lock(&my_lock)
/* critical section */
spin_unlock(&my_lock)
Spin Lock Implementation
• See include/asm-i386/spinlock.h
static inline void __raw_spin_lock(raw_spinlock_t *lock)
{
asm volatile("\n1:\t"
LOCK_PREFIX " ; decb %0\n\t"
"jns 3f\n"
"2:\t"
"rep;nop\n\t"
"cmpb $0,%0\n\t"
"jle 2b\n\t"
"jmp 1b\n"
"3:\n\t"
: "+m" (lock->slock) : : "memory");
}
Spin Lock Methods
•
•
•
•
•
•
•
•
•
spin_lock()// aquires spin lock
spin_unlock()// releases spin lock
spin_lock_irq() // disables local interrupts and acquires spin lock
spin_unlock_irq() // releases spin lock and enables interrupts
spin_lock_irqsave() // saves and disables local interrupts, acquires spin
lock
spin_unlock_irqsave() // restores and enables local interrupts and
releases spin lock
spin_lock_init()
spin_trylock()
spin_is_locked()
More Coming