Implementing software timers

软件方法实现定时器，这个是APUE习题10.5的参考答案。

Title: Implementing Software Timers
By: Don Libes

Originally appeared in the Nov. 1990 "C User's Journal" and is also
reprinted as Chapter 35 of "Obfuscated C and Other Mysteries",
John Wiley & Sons, 1993, ISBN 0-471-57805-3.

---------------------------------------------------------------------

This column will describe a set of functions to implement software
timers.  What are software timers and why do you need them?  Software
timers make up for inherent limitations in hardware timers.  For
example, while most computers have clock hardware, you can typically
only have the clock trigger an interrupt for one time in the future.

When running multiple tasks, you will want to have the clock keep
track of multiple timers concurrently so that interrupts can be
generated correctly even if the time periods overlap.  Operating
systems do this all the time.

Robert Ward discussed the related problem of building a general
purpose scheduler ("Practical Schedulers for Real-Time Applications")
in the April 1990 CUJ.  In the "Additional Ideas" section, Robert
described the usefulness of a timer scheduling queue.  "Events can
specify the timing of other events by putting a timer programming
request in a special queue."  That is exactly what the software in
this column will do.  (Thanks for the lead in, Robert.)  You may want
to reread at least the beginning of his column right now, although it
isn't really necessary.

The code in this column has other uses as well.  For example, you can
use it to simulate multiple timers in environments such as a UNIX
process which only allows the user one software timer.  Even if you
aren't interested in software timers, I think you will find this an
intriguing column.  Using simple techniques and data structures, this
C code produces very powerful results.  The code was very tricky to
get right, and my commentary should be interesting if only as some
more practice in reading and writing C code.

Timers

By implementing the timers as a separate piece of software, we can
reduce the complexity of the scheduler.  Some people like this kind of
modularization, and some don't.  Similarly some operating systems do
this, and some don't.  I like it.  It makes the code easier to write,
to read, and to correct (oops).

The basic idea of a timer is that they allow tasks to be run at some
time in the future.  When their time arrives, they are scheduled to be
run.  The responsibility of actually running them is then turned over
to someone else, such as the scheduler.  In order to communicate with
the scheduler, we'll set up a common data structure called a timer
(listing 1).  I've also included a few other miscellaneous definitions
that will be needed later on.  For instance, the TIME typedef is used
to declare all relative time variables.  You can complete this
definition based on what your needs are.

#include 

#define TRUE	1
#define FALSE	0

#define MAX_TIMERS	...	/* number of timers */
typedef ... TIME;		/* how time is actually stored */
#define VERY_LONG_TIME	...	/* longest time possible */

struct timer {
	int inuse;		/* TRUE if in use */
	TIME time;		/* relative time to wait */
	char *event;		/* set to TRUE at timeout */
} timers[MAX_TIMERS];		/* set of timers */

listing 1

Each timer will be represented by a timer struct.  The set of timers
will be maintained in an array, timers.  The first element of each
timer declares whether the timer is in use.  The second element of a
timer is the amount of time being waited for.  As time passes, this
will be periodically updated.  event is a pointer to a value that is
initially set to 0.  When it is time to run the task, *event is set to
1.  We can imagine that the scheduler also keeps an event pointer.
Every so often, it reexamines it.  When it finds it has been set to 1,
it knows that the timer has expired and the associated task can be
run.

[Notice how simple this is.  Other schedulers or other scheduler data
structures could enable runnability, without worrying or even knowing
about timers.]

The code in listing 2 initializes the timers.  It runs through the
array setting each inuse flag to FALSE.  This for loop will become
idiomatic to you by the end of this column.

void
timers_init() {
	struct timer *t;

	for (t=timers;t<&timers[MAX_TIMERS];t++)
		t->inuse = FALSE;
}

listing 2

Now we can write the routines to schedule the timers.  First, I'll
show timer_undeclare, which is a little simpler than its counterpart,
timer_declare.

There are a variety of ways to keep track of the timers.  Machines
which don't have sophisticated clock hardware usually call an
interrupt handler at every clock tick.  The software then maintains
the system time in a register, as well as checking for timer entries
that have expired.

More intelligent machines can maintain the clock in hardware, only
interrupting the CPU after a given time period has expired.  By having
the clock interrupt for when an event is waiting, you can get a
tremendous speedup.  This technique is also common in software
simulations and thread implementation.

Reading the clock may require an operating system call, but for our
purposes we will assume the variable time_now to be automatically
updated by the hardware for just this purpose.  volatile indicates
that the variable should not be cached in a register but read from
storage each time.

	volatile TIME time_now;

We will define several variables for shorthands.  timer_next will
point to the timer entry that we next expect to expire.
time_timer_set will contain the system time when the hardware timer
was last set.

struct timer *timer_next = NULL;/* timer we expect to run down next */
TIME time_timer_set;		/* time when physical timer was set */

void timers_update();		/* see discussion below */

void
timer_undeclare(t)
struct timer *t;
{
	disable_interrupts();
	if (!t->inuse) {
		enable_interrupts();
		return;
	}

	t->inuse = FALSE;

	/* check if we were waiting on this one */
	if (t == timer_next) {
		timers_update(time_now - time_timer_set);
		if (timer_next) {
			start_physical_timer(timer_next->time);
			time_timer_set = time_now;
		}
	}
	enable_interrupts();
}

Listing 3

Undeclaring Timers - Why and How?

timer_undeclare does just what its name implies, it undeclares a
timer.  Undeclaring timers is actually an important operation in some
applications.  For example, network code sets timers like crazy.  In
some protocols, each packet sent generates a timer.  If the sender
doesn't receive an acknowledgement after a given interval, the timer
forces it to resend a packet.  If the sender does receive an
acknowledgement, it undeclares the timer.  If things are going well,
every single timer declared is later undeclared.

timer_undeclare (listing 3) is performed with interrupts disabled.
This is necessary because we are going to have an interrupt handler
that can access the same data.  Because this data is shared, access
must be strictly controlled.  I've shown the interrupt manipulation as
a function call, but you must use whatever is appropriate to your
system.  This is very system dependent.

timer_undeclare starts by checking the validity of the argument as a
timer entry.  We will see later that the system clock can implicitly
undeclare timer entries.  Thus we must make a reasonable attempt to
assure ourselves that a timer to be undeclared is still declared.

Once assured the timer is valid, timer_undeclare marks the entry
invalid.  If the timer happens to be the very one next expected to
expire, the physical timer must be restarted for the next shorter
timer.  Before doing that, all the timer entries have to be updated by
the amount of time that has elapsed since the timer was last set.
This is done by timers_update which also calculates the next shortest
timer.  Looking for the shortest timer in that function is a little
obscure but happens to be very convenient since timers_update has to
look at every timer anyway.

timers_update (listing 4) goes through the timers, subtracting the
given time from each.  If any reach 0 this way, they are triggered by
setting the event flag.  Any lag in the difference between when a
timer was requested and timers_update is called, is accounted for by
basing the latency against time_now and also collecting timers that
have "gone negative" in timers_update.  (Why might a timer go
negative?)  Lastly, we also remember the lowest nonzero timer to wait
for as timer_next.

timer_last is just a temporary.  It is a permanently non-schedulable
timer that will only show up when all the other timers have been
scheduled.

/* subtract time from all timers, enabling any that run out along the way */
void
timers_update(time)
TIME time;
{
	static struct timer timer_last = {
		FALSE			/* in use */,
		VERY_LONG_TIME		/* time */,
		NULL			/* event pointer */
	};

	struct timer *t;

	timer_next = &timer_last;

	for (t=timers;t<&timers[MAX_TIMERS];t++) {
		if (t->inuse) {
			if (time < t->time) { /* unexpired */
				t->time -= time;
				if (t->time < timer_next->time)
					timer_next = t;
			} else { /* expired */
				/* tell scheduler */
				*t->event = TRUE;
				t->inuse = 0;	/* remove timer */
			}
		}
	}

	/* reset timer_next if no timers found */
	if (!timer_next->inuse) timer_next = 0;
}

listing 4

Declaring Timers

timer_declare (listing 5) takes a time and an event address as
arguments.  When the time expires, the value that event points to will
be set.  (This occurs in timers_update under the comment /* tell
scheduler */.)  timer_declare returns a pointer to a timer.  This
pointer is the same one that timer_undeclare takes as an argument.

struct timer *
timer_declare(time,event)
unsigned int time;		/* time to wait in 10msec ticks */
char *event;
{
	struct timer *t;

	disable_interrupts();

	for (t=timers;t<&timers[MAX_TIMERS];t++) {
		if (!t->inuse) break;
	}

	/* out of timers? */
	if (t == &timers[MAX_TIMERS]) {
		enable_interrupts();
		return(0);
	}

	/* install new timer */
	t->event = event;
	t->time = time;
	if (!timer_next) {
		/* no timers set at all, so this is shortest */
		time_timer_set = time_now;
		start_physical_timer((timer_next = t)->time);
	} else if ((time + time_now) < (timer_next->time + time_timer_set)) {
		/* new timer is shorter than current one, so */
		timers_update(time_now - time_timer_set);
		time_timer_set = time_now;
		start_physical_timer((timer_next = t)->time);
	} else {
		/* new timer is longer, than current one */
	}
	t->inuse = TRUE;
	enable_interrupts();
	return(t);
}

listing 5

As with its counterpart, interrupts are disabled in timer_declare to
prevent concurrent access to the shared data structure.

The first thing timer_declare does is to allocate a timer.  If none
are available, a NULL is returned so that the caller can fail or retry
later.

Once a timer is allocated and initialized, we must check if the
physical timer must be changed.  There are three cases:

1) There are no other timers;

   In this case, we go ahead and start the physical timer with the
   time of this timer.

2) There are other timers, but this new one is the shortest of all the others;

   In this case, we must restart the physical timer to the new time.
   But before we do that, we must update all the other timers by the
   amount of time that has elapsed since the physical timer was last
   set.

3) There are other timers, and this new one is not the shortest.

   There is nothing to do in this case.  However, for legibility it is
   broken into its own case which contains only a comment.  That way
   it is clear what is going on when the previous else-if test fails.

Before enabling interrupts and returning, the timer's inuse flag is
set.  The reason it is done afterwards rather than with the earlier
timer settings is that this prevents timers_update from updating it
with a time period that occurred before it was even declared.

Handling Timer Interrupts

The only remaining routine is the interrupt handler (listing 6)
actually called when the physical clock expires.  When the interrupt
handler is called, we are guaranteed that the time described by
timer_next has elapsed.

void
timer_interrupt_handler() {
	timers_update(time_now - time_timer_set);
	
	/* start physical timer for next shortest time if one exists */
	if (timer_next) {
		time_timer_set = time_now;
		start_physical_timer(timer_next->time);
	}
}

listing 6

Each time the interrupt handler is called, a timer has expired.  By
calling timers_update, all the timers will be decremented and any
timers that have expired will have their event flags enabled.  This
will also set up timer_next so that the physical timer can be
restarted for the next timer we expect to occur.

Let's examine one special case.  Suppose we have only one timer set
up.  Now imagine that we have called timer_undeclare and just as
interrupts are disabled, the physical clock ticks down all the way.
Since interrupts are disabled, the interrupt will be delivered
immediately after interrupts are enabled.  But they will be enabled
after the timer has been deleted.  So we see a situation where an
interrupt will be delivered for a timer that no longer exists.  What
occurs in the interrupt handler?

timers_update is called.  It finds nothing to update.  As a
consequence of this, timer_next is set to 0.  The remainder of the
interrupt handler already handles the case of no remaining timers, and
the handler returns normally.

This is an example of the kind of special casing you have to keep in
mind when writing the code.  (In fact, my first implementation didn't
handle this right, and it was painful to debug.  Debuggers don't work
very well when fooling around with interrupts!)

Conclusion

I have presented an implementation of timers.  The code is carefully
designed so that it is relatively free of special demands it places on
a scheduler.  For example, it doesn't close off the scheduler from
using a different kind of timer at the same time.

One thought that may have occurred to you while reading this, is why
the timers are maintained as an array rather than say, a linked list.
Using a linked list would avoid the overhead of stepping through
arrays (which can be almost entirely empty).  Keeping the list sorted
by time would make the timers_update function much simpler.

On the other hand, it would complicate the other functions.  For
example, timer_undeclare, would either require you to use a
doubly-linked list, or to search the entire list from the beginning
each time.  Another point is that real-time systems typically avoid
dynamic structures to begin with.  For example, using malloc/free from
a process-wide heap can cause an indeterminate amount of time that is
difficult to estimate.  If I was to recode this using linked lists, I
would use a malloc implementation from a small pool of timer-only
buffers, which in effect is very similar to what I've done here with
arrays.  There would be a tradeoff in space and time, which you might
prefer or not depending upon your application.

If you decide to recode or just modify my implementation, be very
careful.  Always imagine the worst thing that can happen when two
processes attempt to access the same data structure at the same time.
Happy interruptions!

Thanks

Debugging timing routines is very different than other code, since
unrelated events in the computer can make your programs behave
differently.  Even putting in printf statements can change critical
execution paths.  It is extremely aggravating when problems disappear
only when you are debugging.  Furthermore, most debuggers do not work
well when interrupts are disabled.  Ed Barkmeyer was of great help
debugging the timer code and teaching me to persevere when I saw code
behaving in ways that had to be impossible.  Thanks to Sarah Wallace
and Randy Miller who debugged this column and also forced me to make
all the explanations much clearer.