Re: Subject: [PATCH] changed timespec64_to_ns to avoid underrun

[Thomas Gleixner]

Lukas,

On Wed, Aug 25 2021 at 10:12, OPENSOURCE Lukas Hannen wrote:

thanks for the patch. A few formal nitpicks:

The subject line lacks a subsystem prefix. You can figure that 
usually out by running:

   git log --format=oneline --abbrev-commit include/linux/time64.h

Look for the most used. So in this case it would be simply 'time:'

'changed ... underrun'

Please use imperative mood, i.e. 'Change'. But 'change' is redundant
here anyway because it would not be a patch if it would not change
anything.

So something like: 'Handle negative seconds correctly in timespec64_to_ns()'
would be more informative. Your subject line does not really match
the problem.

Also note the brackets which make it obvious that this talks about a
function.

You have a redundant 'Subject:' in the Subject: header

> This patch fixes a small oversight in timespec64_to_ns() that has

This patch is redundant as well. See Documentation/process/ and grep for
'This patch'

> resulted in negative seconds being erroneously clamped to KTIME_MAX
> due to a cast to unsigned long long (which expands to the 2's complement
> of a negative long long, even if the architecture does not implement
> negative numbers using 2's complement)
>
> This is especially relevant in the PTP context, since the ptp_clock_info
> struct (from include/linux/ptp_clock_kernel.h) specifies
>
>	int (*adjtime)(struct ptp_clock_info *ptp, s64 delta);
>	int (*gettime64)(struct ptp_clock_info *ptp, struct timespec64 *ts);
> 
> which is exactly the kind of timespec64 / nanoseconds mix in combination
> with negative values ( ns adjust times ) that can easily lead to calling
> timespec64_to_ns with a negative ts->tv_sec, which would in turn lead to
> instability of the ptp clock.

This is confusing at best.

The adjtime() callback has nothing to do with timespec64_to_ns(). The
conversion happens in the calling code.

gettime64() neither because that reads the time from the PTP clock which
cannot be negative and nothing uses timespec64_to_ns() there either.

The place where a negative seconds value must be handled correctly is
ptp_clock_adjtime().


So something like this:

timespec64_ns() prevents multiplication overflows by comparing the
seconds value of the timespec to KTIME_SEC_MAX. If the value is greater
or equal it returns KTIME_MAX.

But that check casts the signed seconds value to unsigned which makes
the comparision true for all negative values and therefore return
wrongly KTIME_MAX.

Negative second values are perfectly valid and required in some places,
e.g. ptp_clock_adjtime().

Remove the cast and add a check for the negative boundary which is
required to prevent undefined behaviour of the multiplication due to
multiplication underflow.

Hmm?

> Fixes: cb47755725da ("time: Prevent undefined behaviour in timespec64_to_ns()")'
> Signed-off-by: Lukas Hannen <lukas.hannen@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx>
>
> ---
> The Patch should apply cleanly to all the branches that the original

Emphasis on should. See below.

> commit cb47755725da ("time: Prevent undefined behaviour in timespec64_to_ns()")'
> was backported to.
>
> include/linux/time64.h | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/include/linux/time64.h b/include/linux/time64.h
> index 5117cb5b56561..81b9686a20799 100644
> --- a/include/linux/time64.h
> +++ b/include/linux/time64.h
> @@ -21,15 +21,17 @@ struct itimerspec64 {
>  };
>
>  /* Located here for timespec[64]_valid_strict */
>  #define TIME64_MAX                     ((s64)~((u64)1 << 63))
>  #define TIME64_MIN                     (-TIME64_MAX - 1)
>
>  #define KTIME_MAX                      ((s64)~((u64)1 << 63))
> +#define KTIME_MIN                      (-KTIME_MAX - 1)
>  #define KTIME_SEC_MAX                  (KTIME_MAX / NSEC_PER_SEC)
> +#define KTIME_SEC_MIN                  (KTIME_MIN / NSEC_PER_SEC)

Your patch is white space damaged, which makes it fail to apply.
Something replaced all tabs with spaces, most likely your mail client.

Please figure out a way to send patches either via git-email or by using
a email client which tries not to be smarter than the person using it.
Send the patch to yourself and validate that it applies cleanly.

I fixed it up for you this time.

Thanks,

        tglx