* Re: Bug in timekeeping.c ktime_get_ts() Kernel 2.2.29.6-rt23
[not found] <4EA20108F62E0346A80AFB03445370F210F31385D4@de01ex08.GLOBAL.JHCN.NET>
@ 2009-10-12 15:48 ` Thomas Gleixner
0 siblings, 0 replies; only message in thread
From: Thomas Gleixner @ 2009-10-12 15:48 UTC (permalink / raw)
To: Mühlbauer Gerhard; +Cc: linux-rt-users@vger.kernel.org, mingo@elte.hu
[-- Attachment #1: Type: TEXT/PLAIN, Size: 3411 bytes --]
Gerhard,
On Mon, 12 Oct 2009, Mühlbauer Gerhard wrote:
> Hello!
> We have found a bug in the function ktime_get_ts() in the rt23
> realtime patches. The function set_normalized_timespec() gets
> called with a sum of 3 different nsec values. Although each value
> should be within a range of 0..100000000, the sum can be within a
> range of 0..3000000000 and by this an overflow happens when
> assigning this to a long variable. As a result, the function
> clock_gettime(CLOCK_MONOTONIC) can create results that jump back and
> forth in time each second by a step of 2^32nsec (=4sec). The
> following patch solves this problem when attached after
> patch-2.6.29.6-rt23.
correct. We fixed that problem in mainline a few weeks ago, but I did
not come around to do an update to 2.6.29-rt.
While your patch solves the problem I recommend using the mainline fix
below as it covers all eventual use cases of set_normalized_timespec().
I also uploaded a 2.6.29-rt24 patch set to the usual place which has
the fix applied.
Thanks,
tglx
---
commit 12e09337fe238981cb0c87543306e23775d1a143
Author: Thomas Gleixner <tglx@linutronix.de>
Date: Mon Sep 14 23:37:40 2009 +0200
time: Prevent 32 bit overflow with set_normalized_timespec()
set_normalized_timespec() nsec argument is of type long. The recent
timekeeping changes of ktime_get_ts() feed
ts->tv_nsec + tomono.tv_nsec + nsecs
to set_normalized_timespec(). On 32 bit machines that sum can be
larger than (1 << 31) and therefor result in a negative value which
screws up the result completely.
Make the nsec argument of set_normalized_timespec() s64 to fix the
problem at hand. This also prevents similar problems for future users
of set_normalized_timespec().
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Carsten Emde <carsten.emde@osadl.org>
LKML-Reference: <new-submission>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: John Stultz <johnstul@us.ibm.com>
diff --git a/include/linux/time.h b/include/linux/time.h
index 256232f..56787c0 100644
--- a/include/linux/time.h
+++ b/include/linux/time.h
@@ -75,7 +75,7 @@ extern unsigned long mktime(const unsigned int year, const unsigned int mon,
const unsigned int day, const unsigned int hour,
const unsigned int min, const unsigned int sec);
-extern void set_normalized_timespec(struct timespec *ts, time_t sec, long nsec);
+extern void set_normalized_timespec(struct timespec *ts, time_t sec, s64 nsec);
extern struct timespec timespec_add_safe(const struct timespec lhs,
const struct timespec rhs);
diff --git a/kernel/time.c b/kernel/time.c
index 2951194..2e2e469 100644
--- a/kernel/time.c
+++ b/kernel/time.c
@@ -370,13 +370,20 @@ EXPORT_SYMBOL(mktime);
* 0 <= tv_nsec < NSEC_PER_SEC
* For negative values only the tv_sec field is negative !
*/
-void set_normalized_timespec(struct timespec *ts, time_t sec, long nsec)
+void set_normalized_timespec(struct timespec *ts, time_t sec, s64 nsec)
{
while (nsec >= NSEC_PER_SEC) {
+ /*
+ * The following asm() prevents the compiler from
+ * optimising this loop into a modulo operation. See
+ * also __iter_div_u64_rem() in include/linux/time.h
+ */
+ asm("" : "+rm"(nsec));
nsec -= NSEC_PER_SEC;
++sec;
}
while (nsec < 0) {
+ asm("" : "+rm"(nsec));
nsec += NSEC_PER_SEC;
--sec;
}
^ permalink raw reply related [flat|nested] only message in thread