Re: [PATCH] One possible bugfix for CLOCK_REALTIME timer.

From: George Anzinger
Date: Thu Jun 03 2004 - 19:38:47 EST


Hu, Boris wrote:
Thanks for your detailed comments. :)


Hu, Boris wrote:

<<posix-abs_timer-bugfix.diff>> George,

There is one bug of CLOCK_REALTIME timer reported by Adam at
http://sources.redhat.com/ml/libc-alpha/2004-05/msg00177.html.

Here is one possible bugfix and it is against linux-2.6.6. All
TIMER_ABSTIME cloks will be collected in k_clock struct and updated

in

sys_clock_settime. Could you review it? Thanks.

Thanks for the poke :). Could you make the following changes:

First, put the list in the posix timer structure (k_itimer), not in
timer_list.
This means one more dereference when doing things, but it does not

push

into
the timer_list structure which is mostly used for other things.


Done.


Second, I don't see the timer being removed from the list (should

happen

when
ever it is inactive). Timers that repeat should be out of the list

while

waiting for the signal to be picked up and put back in when add_timer

is

again
called.


I tried to add the removed codes to set_timer_inactive() but it would
trigger a strange oops. I am still investigating on it. Is there any
recommending places except set_timer_inactive()?

Hm, set_timer_inactive() is called from the timer create routine. Should not need to remove it here... Also, this function is used for SMP issues and, in some cases (do_timer_settime is one) it is not called if not on an SMP system. Also, it seems not to be called from sys_timer_delete(). This last would be a real problem as we are about to return the memory the list runs through it.

So, I think you will just have to find the places were we delete a timer, that and the timer call back function should do it.



Also, you should test to see if the clock is one that can be set prior

to

putting the timer in the abs timer list. We must not correct timers

on

CLOCK_MONOTONIC.



Done.


Now, for the correction. Sys_clock_settime() is the wrong place for

this

as the
clock can also be set a number of other ways. The right place is in
clock_was_set(), which is called if time is set in any of the several

ways.

The
next thing is to determine how far the clock was moved. I think the

best

way to
do this is to keep a copy of the wall_to_monotonic var in a private
location.
This should be set to be exactly wall_to_monotonic when the system is
booted (in
the same function you are setting up the clock lists) and at the end

of

clock_was_set(). When clock_was_set() is called the difference

between

this
value and wall_to_monotonic is exactly how far the clock was moved.

(Be

careful
on the sign of this movement.)



Done.


Finally, be careful about races. Timers can expire while

clock_was_set()

is
running. The removal code should take the timer lock as well as the
abs_time
list lock (at least I think this would be wise, but I could be wrong

here).


IMHO, we need not take the timer locks. We del_timer() first and if the
timer has expired, we simply removed it from the abs_timer_list which is
protected by abs_timer_lock.

Me thinks you will need to do a bit more to convince me that locks are not needed here. Lets see if I can explain my concerns.

First, I think the clock_was_set() function needs to serialize it self so only one cpu/ task can be in it at a time. This, I think, can/is done with the abs list lock.

Second there is the possibility that a timer in the list will already be set via the correct time. To avoid this possibility I suggest (this is a change from my suggestion of yesterday) putting the current value of wall_to_monotonic in the k_timer structure when the timer is calculated. This value must be obtained under the xtime read lock (which we already take to calculate the timer). In this way of doing things, clock_was_set() would take the xtime write lock, possibly for each entry in the abs list. It would use the wall_to_monotonic time in the structure rather than keeping a local copy, and it would update that time once the correction was made.

We still haven't covered the case where time is set while a timer is being set. I.e. where the expire time is calculated but the result has not yet been put in the timer structure and add_timer has not yet been called. It is here that the timer lock would seem to be the right thing to do. We would require that the timer be put in the abs list while the xtime read lock is held (careful here as this is now a sequence lock and we only want to add the timer to the list once). This is complicated. We must take the locks in the same order so we can not take the timer lock while holding the abs list lock. The, IMHO, simple thing to do is to have clock_was_set() copy the whole abs list (this is just a simple pointer manipulation). Then it can scan this list and move each entry to the abs list as it updates the timers. The timer lock would be taken for each timer during this update. This is to allow timers that are on the way to add_timer to get there. This code should not remove timers from the abs list, or rather, each timer it finds in the moved list should be put back in the abs list even if it is not in the system timer list (it just means that someone else is removing it). Both the removal from the moved list and the insert into the abs list should be done under the same abs list lock but it must be dropped while taking the timer lock.

There is a possible race here with the timer delete code. Here is how I would solve this. First, with the list being moved, you need only be concerned with the first entry in the list (as you will remove it as part of processing and then do the next first entry). So, first lock the abs list. Then find the timer ID for the first entry. Unlock the abs list, and lock the timer using the ID. The existing lock code will take care of possible races WRT existence. Once the timer is locked, re lock the abs list and if the given timer is still the first entry, a.) remove it b.) if the system delete timer fails, just take the abs list lock and reinsert the timer, else under the xtime readlock compare wall_to_monotonic with the timers value c.) and put it in the abs list and the add_timer list.

Note that the case of an expired timer that is still in the abs list is handled by just reinserting the timer. This means that the expire call back code needs to check for a clock reset that may have made the expire invalid.

I am sure there are other ways of doing this, but the main locking issues are:

1.) Getting the timer's value pegged to a given clock set value (done by copying wall_to_monotonic to the k_timer struct)., This must be done under the xtime read lock.

2.) Covering the race between the timer adjustment code and possible POSIX timer deletion (done by NOT assuming the timer is there just because we found it in the abs timer list, although we do pin it down long enough to get it's ID by locking this list). This also requires us to take the timer lock which means we have to drop the abs list lock.

3.) Covering the race between the timer expiring during the system clock setting processing. Done by having the timer call back code verify that the same value of wall_to_monotonic is still in play.

We note that the time used for NOW in the repeating timer update needs to also satisfy 1. above.

And in passing, note that the overrun count will show something going on when the clock is moved forward and not when it is move backward.



And a minor issue, the community seems to prefer the C comment style

to

the C++
style of comments...



Done.


Thanks for your effort in this matter.

George


diff -urN -X rt.ia32/base/dontdiff
linux-2.6.6/include/linux/posix-timers.h
linux-2.6.6.dev/include/linux/posix-timers.h
--- linux-2.6.6/include/linux/posix-timers.h 2004-05-10
10:32:29.000000000 +0800
+++ linux-2.6.6.dev/include/linux/posix-timers.h 2004-06-02
10:30:57.000000000 +0800
@@ -1,9 +1,14 @@
#ifndef _linux_POSIX_TIMERS_H
#define _linux_POSIX_TIMERS_H

+#include <linux/list.h>
+#include <linux/spinlock.h>
+
struct k_clock {
int res; /* in nano seconds */
- int (*clock_set) (struct timespec * tp);
+ struct list_head abs_timer_list;
+ spinlock_t abs_timer_lock;
+ int (*clock_set) (struct timespec * tp);
int (*clock_get) (struct timespec * tp);
int (*nsleep) (int flags,
struct timespec * new_setting,
diff -urN -X rt.ia32/base/dontdiff linux-2.6.6/include/linux/timer.h
linux-2.6.6.dev/include/linux/timer.h
--- linux-2.6.6/include/linux/timer.h 2004-05-10

10:32:54.000000000

+0800
+++ linux-2.6.6.dev/include/linux/timer.h 2004-06-02
19:16:08.000000000 +0800
@@ -9,6 +9,7 @@

struct timer_list {
struct list_head entry;
+ struct list_head abs_timer_entry;
unsigned long expires;

spinlock_t lock;
diff -urN -X rt.ia32/base/dontdiff linux-2.6.6/kernel/posix-timers.c
linux-2.6.6.dev/kernel/posix-timers.c
--- linux-2.6.6/kernel/posix-timers.c 2004-05-10

10:32:37.000000000

+0800
+++ linux-2.6.6.dev/kernel/posix-timers.c 2004-06-02
19:12:31.000000000 +0800
@@ -7,6 +7,9 @@
*
* Copyright (C) 2002 2003 by MontaVista
Software.
*
+ * 2004-06-01 Fix CLOCK_REALTIME clock/timer TIMER_ABSTIME bug.
+ * Copyright (C) 2004 Boris Hu
+ *
* This program is free software; you can redistribute it and/or

modify

* it under the terms of the GNU General Public License as

published by

* the Free Software Foundation; either version 2 of the License,

or

(at
@@ -200,7 +203,9 @@
*/
static __init int init_posix_timers(void)
{
- struct k_clock clock_realtime = {.res = CLOCK_REALTIME_RES };
+ struct k_clock clock_realtime = {.res = CLOCK_REALTIME_RES,
+ .abs_timer_lock = SPIN_LOCK_UNLOCKED

+ };
struct k_clock clock_monotonic = {.res = CLOCK_REALTIME_RES,
.clock_get = do_posix_clock_monotonic_gettime,
.clock_set = do_posix_clock_monotonic_settime
@@ -388,6 +393,7 @@
return;
}
posix_clocks[clock_id] = *new_clock;
+ INIT_LIST_HEAD(&posix_clocks[clock_id].abs_timer_list);
}

static struct k_itimer * alloc_posix_timer(void)
@@ -843,8 +849,15 @@
if (((timr->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE))
{
if (timr->it_timer.expires == jiffies)
timer_notify_task(timr);
- else
+ else {
add_timer(&timr->it_timer);
+ if (flags & TIMER_ABSTIME) {
+ spin_lock(&clock->abs_timer_lock);
+
list_add_tail(&(timr->it_timer.abs_timer_entry),
+
&(clock->abs_timer_list));
+

spin_unlock(&clock->abs_timer_lock);

+ }
+ }
}
return 0;
}
@@ -1085,16 +1098,61 @@
sys_clock_settime(clockid_t which_clock, const struct timespec

__user

*tp)
{
struct timespec new_tp;
+ struct timespec before, now;
+ struct k_clock *clock;
+ long ret;
+ struct timer_list *timer, *tmp;
+ struct timespec delta;
+ u64 exp = 0;

- if ((unsigned) which_clock >= MAX_CLOCKS ||
+ if ((unsigned) which_clock >= MAX_CLOCKS ||
!posix_clocks[which_clock].res)
return -EINVAL;
- if (copy_from_user(&new_tp, tp, sizeof (*tp)))
+
+ clock = &posix_clocks[which_clock];
+
+ if (copy_from_user(&new_tp, tp, sizeof (*tp)))
return -EFAULT;
if (posix_clocks[which_clock].clock_set)
return posix_clocks[which_clock].clock_set(&new_tp);

- return do_sys_settimeofday(&new_tp, NULL);
+ do_posix_gettime(clock, &before);
+
+ ret = do_sys_settimeofday(&new_tp, NULL);
+
+ /*
+ * Check if there exist TIMER_ABSTIME timers to update.
+ */
+ if (which_clock != CLOCK_REALTIME ||
+ list_empty(&clock->abs_timer_list))
+ return ret;
+
+ do_posix_gettime(clock, &now);
+ delta.tv_sec = before.tv_sec - now.tv_sec;
+ delta.tv_nsec = before.tv_nsec - now.tv_nsec;
+ while (delta.tv_nsec >= NSEC_PER_SEC) {
+ delta.tv_sec ++;
+ delta.tv_nsec -= NSEC_PER_SEC;
+ }
+ while (delta.tv_nsec < 0) {
+ delta.tv_sec --;
+ delta.tv_nsec += NSEC_PER_SEC;
+ }
+
+ tstojiffie(&delta, clock->res, &exp);
+
+ spin_lock(&(clock->abs_timer_lock));
+ list_for_each_entry_safe(timer, tmp,
+ &clock->abs_timer_list,
+ abs_timer_entry) {
+ if (timer && del_timer(timer)) { //the timer has

not

run
+ timer->expires += exp;
+ add_timer(timer);
+ } else
+ list_del(&timer->abs_timer_entry);
+ }
+ spin_unlock(&(clock->abs_timer_lock));
+ return ret;
}

asmlinkage long

Good Luck !
Boris Hu (Hu Jiangtao)
Intel China Software Center
86-021-5257-4545#1277
iNET: 8-752-1277
************************************
There are my thoughts, not my employer's.
************************************
"gpg --recv-keys --keyserver wwwkeys.pgp.net 0FD7685F"
{0FD7685F:CFD6 6F5C A2CB 7881 725B CEA0 956F 9F14 0FD7 685F}



--
George Anzinger george@xxxxxxxxxx
High-res-timers: http://sourceforge.net/projects/high-res-timers/
Preemption patch: http://www.kernel.org/pub/linux/kernel/people/rml





--
George Anzinger george@xxxxxxxxxx
High-res-timers: http://sourceforge.net/projects/high-res-timers/
Preemption patch: http://www.kernel.org/pub/linux/kernel/people/rml

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/