Wrong use of MAX_JIFFY_OFFSET?

Wrong use of MAX_JIFFY_OFFSET?

[Liu Yu]

Hi guys,

I saw a couple of places in current kernel have this kind of code:

> static inline unsigned int elapsed_jiffies_msecs(unsigned long start)
> {
>         unsigned long end = jiffies;
>
>         if (end >= start)
>                return jiffies_to_msecs(end - start);
>
>         return jiffies_to_msecs(end + (MAX_JIFFY_OFFSET - start) + 1);
> }

As you know, jiffies has a type of unsigned long, so if we know which is the
end and
which is the start, then (end - start) can simply figure out how much
jiffies flies,
without worry about the overflow.

Look at the code above, assume that there is just an overflow happening on
jiffies: end=0 and start=~0UL.
Since end < start, then the return value of the function is
jiffies_to_msecs(MAX_JIFFY_OFFSET+2).
But shouldn't the correct value be jiffies_to_msecs(1)?

could someone tell me that am I missing anything?

Thanks,
Yu

Re: Wrong use of MAX_JIFFY_OFFSET?

[Andrew Morton]

On Wed, 28 Mar 2012 15:12:24 +0800
"Liu Yu" <liuyums@bwstor.com.cn> wrote:

> Hi guys,
> 
> I saw a couple of places in current kernel have this kind of code:
> 
> > static inline unsigned int elapsed_jiffies_msecs(unsigned long start)
> > {
> >         unsigned long end = jiffies;
> >
> >         if (end >= start)
> >                return jiffies_to_msecs(end - start);
> >
> >         return jiffies_to_msecs(end + (MAX_JIFFY_OFFSET - start) + 1);
> > }
> 
> As you know, jiffies has a type of unsigned long, so if we know which is the
> end and
> which is the start, then (end - start) can simply figure out how much
> jiffies flies,
> without worry about the overflow.
> 
> Look at the code above, assume that there is just an overflow happening on
> jiffies: end=0 and start=~0UL.
> Since end < start, then the return value of the function is
> jiffies_to_msecs(MAX_JIFFY_OFFSET+2).
> But shouldn't the correct value be jiffies_to_msecs(1)?
> 
> could someone tell me that am I missing anything?
> 

Seems right.  The code should be

static inline unsigned long elapsed_jiffies_msecs(unsigned long start)
{
	return jiffies_to_msecs(jiffies - start);
}

Note the return type.  jiffies_to_msecs() currently returns unsigned
int.  I think it should return unsigned long.  Even then, it can still
overflow with valid inputs on HZ=100 32-bit machines.

re: Wrong use of MAX_JIFFY_OFFSET?

[Liu Yu]

> -----邮件原件-----
> 发件人: Andrew Morton [mailto:akpm@linux-foundation.org]
> 发送时间: 2012年4月7日 3:05
> 收件人: Liu Yu
> 抄送: linux-kernel@vger.kernel.org
> 主题: Re: Wrong use of MAX_JIFFY_OFFSET?
> 
> On Wed, 28 Mar 2012 15:12:24 +0800
> "Liu Yu" <liuyums@bwstor.com.cn> wrote:
> 
> > Hi guys,
> >
> > I saw a couple of places in current kernel have this kind of code:
> >
> > > static inline unsigned int elapsed_jiffies_msecs(unsigned long start)
> > > {
> > >         unsigned long end = jiffies;
> > >
> > >         if (end >= start)
> > >                return jiffies_to_msecs(end - start);
> > >
> > >         return jiffies_to_msecs(end + (MAX_JIFFY_OFFSET - start) + 1);
> > > }
> >
> > As you know, jiffies has a type of unsigned long, so if we know which is
the
> > end and
> > which is the start, then (end - start) can simply figure out how much
> > jiffies flies,
> > without worry about the overflow.
> >
> > Look at the code above, assume that there is just an overflow happening
on
> > jiffies: end=0 and start=~0UL.
> > Since end < start, then the return value of the function is
> > jiffies_to_msecs(MAX_JIFFY_OFFSET+2).
> > But shouldn't the correct value be jiffies_to_msecs(1)?
> >
> > could someone tell me that am I missing anything?
> >
> 
> Seems right.  The code should be
> 
> static inline unsigned long elapsed_jiffies_msecs(unsigned long start)
> {
> 	return jiffies_to_msecs(jiffies - start);
> }
> 
> Note the return type.  jiffies_to_msecs() currently returns unsigned
> int.  I think it should return unsigned long.  Even then, it can still
> overflow with valid inputs on HZ=100 32-bit machines.

Thanks a lot for your reply :)

I saw similar code in
	drivers/net/wireless/ipw2x00/libipw_wx.c
	net/wireless/scan.c
	include/linux/jbd2.h

hope someone can fix this. :)

Thanks,
Yu

Wrong use of MAX_JIFFY_OFFSET?

[Liu Yu]

Hi guys,

I saw a couple of places in current kernel have this kind of code:

> static inline unsigned int elapsed_jiffies_msecs(unsigned long start)
> {
>         unsigned long end = jiffies;
>
>         if (end >= start)
>                return jiffies_to_msecs(end - start);
>
>         return jiffies_to_msecs(end + (MAX_JIFFY_OFFSET - start) + 1);
> }

As you know, jiffies has a type of unsigned long, so if we know which is the
end and
which is the start, then (end - start) can simply figure out how much
jiffies flies,
without worry about the overflow.

Look at the code above, assume that there is just an overflow happening on
jiffies: end=0 and start=~0UL.
Since end < start, then the return value of the function is
jiffies_to_msecs(MAX_JIFFY_OFFSET+2).
But shouldn't the correct value be jiffies_to_msecs(1)?

could someone tell me that am I missing anything?

Thanks,
Yu

RE: Wrong use of MAX_JIFFY_OFFSET?

[David Laight]

 

> -----Original Message-----
> From: netdev-owner@vger.kernel.org 
> [mailto:netdev-owner@vger.kernel.org] On Behalf Of Liu Yu
> Sent: 27 March 2012 02:25
> To: netdev@vger.kernel.org
> Subject: Wrong use of MAX_JIFFY_OFFSET?
> 
> Hi guys,
> 
> I saw a couple of places in current kernel have this kind of code:
> 
> > static inline unsigned int elapsed_jiffies_msecs(unsigned 
> long start)
> > {
> >         unsigned long end = jiffies;
> >
> >         if (end >= start)
> >                return jiffies_to_msecs(end - start);
> >
> >         return jiffies_to_msecs(end + (MAX_JIFFY_OFFSET - 
> start) + 1);
> > }
> 
> As you know, jiffies has a type of unsigned long, so if we 
> know which is the end and
> which is the start, then (end - start) can simply figure out how much
> jiffies flies, without worry about the overflow.
> 
> Look at the code above, assume that there is just an overflow 
> happening on
> jiffies: end=0 and start=~0UL.
> Since end < start, then the return value of the function is
> jiffies_to_msecs(MAX_JIFFY_OFFSET+2).
> But shouldn't the correct value be jiffies_to_msecs(1)?
> 
> could someone tell me that am I missing anything?

That code has confused me...
MAX_JIFFIES_OFFSET appears to have something to do with
timespec_to_jiffies() rather than being anything to so with
the way 'jiffies' itself wraps.

Since arithmetic wrap is well defined for unsigned values
the correct code should just be:
	return jiffies_to_ms(jiffies - start);
The code above is completly fubar when jiffies wraps.

	David