----- On Nov 9, 2018, at 10:37 AM, Jonathan Rajotte
jonathan.rajotte-jul...@efficios.com wrote:
> Problem:
>
> The previous algorithm was not *wrong* but was prone to C type conversion
> error.
>
> Previous algorithm:
>
> ((A.sec - B.sec) * MSEC_PER_SEC) + ((A.nsec - B.nsec) / NSEC_PER_SEC)
>
> ((A.sec - B.sec) * 1000000ULL) + ((A.nsec - B.nsec) / 1000000ULL)
>
> Note that "(A.nsec - B.nsec)" can be negative.
>
> A *quick* fix here could be to cast the NSEC_PER_SEC to long allowing a
> valid division operation.
>
> The problem was introduced by 395d6b02dda3db1acd08936f49c1dc8efc48e613.
>
> Solution:
>
> Perform the time diff using a temporary timespec struct and perform the
> carry operation manually. This prevent negative value for nsec and
> helps prevent C type promotion problems.
>
> Signed-off-by: Jonathan Rajotte <jonathan.rajotte-jul...@efficios.com>
> ---
>
> I initially misplaced the issue.
> As mentioned in the commit message, we can also cast NSEC_PER_SEC to long to
> fix
> the issue. I prefer the use of a timespec struct for expressiveness.
>
> Also the case were we pass a time_b > time_a is handled by returning a
> difference of zero. We might want to assert here since it should never happen.
>
> src/common/sessiond-comm/inet.c | 28 ++++++++++++++++++++--------
> src/common/sessiond-comm/inet6.c | 28 ++++++++++++++++++++--------
> 2 files changed, 40 insertions(+), 16 deletions(-)
>
> diff --git a/src/common/sessiond-comm/inet.c b/src/common/sessiond-comm/inet.c
> index e0b3e7a96..94a3c3a67 100644
> --- a/src/common/sessiond-comm/inet.c
> +++ b/src/common/sessiond-comm/inet.c
> @@ -119,16 +119,28 @@ static
> unsigned long time_diff_ms(struct timespec *time_a,
> struct timespec *time_b)
> {
> - time_t sec_diff;
> - long nsec_diff;
> - unsigned long result_ms;
> + struct timespec diff;
>
> - sec_diff = time_a->tv_sec - time_b->tv_sec;
> - nsec_diff = time_a->tv_nsec - time_b->tv_nsec;
> + diff.tv_sec = time_a->tv_sec - time_b->tv_sec;
> + diff.tv_nsec = time_a->tv_nsec - time_b->tv_nsec;
>
> - result_ms = sec_diff * MSEC_PER_SEC;
> - result_ms += nsec_diff / NSEC_PER_MSEC;
> - return result_ms;
> + if (diff.tv_nsec < 0) {
> + /* Perform borrow operation */
> + diff.tv_sec -= 1;
> + diff.tv_nsec += NSEC_PER_SEC;
> + }
> +
> + if (diff.tv_sec < 0) {
> + /* We went back in time. Put all to zero */
> + diff.tv_sec = 0;
> + diff.tv_nsec = 0;
> + }
Why would a diff operation implicitly floor everything to 0 when
time_a < time_b ?
> +
> + /*
> + * Transform into MSEC, diff.tv_nsec is always positive due to borrow
> + * operation. diff.tb_sec is always positive.
tb_sec -> tv_nsec
> + */
> + return (diff.tv_sec * MSEC_PER_SEC) + (diff.tv_nsec / NSEC_PER_MSEC);
How do we handle time delta difference that overflows unsigned long ?
> }
>
> static
> diff --git a/src/common/sessiond-comm/inet6.c
> b/src/common/sessiond-comm/inet6.c
> index dfb5fc5d1..07043e93f 100644
> --- a/src/common/sessiond-comm/inet6.c
> +++ b/src/common/sessiond-comm/inet6.c
> @@ -117,16 +117,28 @@ static
> unsigned long time_diff_ms(struct timespec *time_a,
> struct timespec *time_b)
This should be moved to a common implementation file. It's a cut and
paste from inet.c.
Thanks,
Mathieu
> {
> - time_t sec_diff;
> - long nsec_diff;
> - unsigned long result_ms;
> + struct timespec diff;
>
> - sec_diff = time_a->tv_sec - time_b->tv_sec;
> - nsec_diff = time_a->tv_nsec - time_b->tv_nsec;
> + diff.tv_sec = time_a->tv_sec - time_b->tv_sec;
> + diff.tv_nsec = time_a->tv_nsec - time_b->tv_nsec;
>
> - result_ms = sec_diff * MSEC_PER_SEC;
> - result_ms += nsec_diff / NSEC_PER_MSEC;
> - return result_ms;
> + if (diff.tv_nsec < 0) {
> + /* Perform borrow operation */
> + diff.tv_sec -= 1;
> + diff.tv_nsec += NSEC_PER_SEC;
> + }
> +
> + if (diff.tv_sec < 0) {
> + /* We went back in time. Put all to zero */
> + diff.tv_sec = 0;
> + diff.tv_nsec = 0;
> + }
> +
> + /*
> + * Transform into MSEC, diff.tv_nsec is always positive due to borrow
> + * operation. diff.tb_sec is always positive.
> + */
> + return (diff.tv_sec * MSEC_PER_SEC) + (diff.tv_nsec / NSEC_PER_MSEC);
> }
>
> static
> --
> 2.17.1
>
> _______________________________________________
> lttng-dev mailing list
> lttng-dev@lists.lttng.org
> https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
_______________________________________________
lttng-dev mailing list
lttng-dev@lists.lttng.org
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
