From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sasha Levin Subject: [PATCH AUTOSEL 4.20 236/304] mlx5: update timecounter at least twice per counter overflow Date: Mon, 28 Jan 2019 10:42:33 -0500 Message-ID: <20190128154341.47195-236-sashal@kernel.org> References: <20190128154341.47195-1-sashal@kernel.org> Mime-Version: 1.0 Content-Transfer-Encoding: 8bit Return-path: In-Reply-To: <20190128154341.47195-1-sashal@kernel.org> Sender: netdev-owner@vger.kernel.org To: linux-kernel@vger.kernel.org, stable@vger.kernel.org Cc: Miroslav Lichvar , Richard Cochran , Ariel Levkovich , Saeed Mahameed , Sasha Levin , netdev@vger.kernel.org, linux-rdma@vger.kernel.org List-Id: linux-rdma@vger.kernel.org From: Miroslav Lichvar [ Upstream commit 5d8678365c90b9ce1fd2243ff5ea562609f6cec1 ] The timecounter needs to be updated at least once in half of the cyclecounter interval to prevent timecounter_cyc2time() interpreting a new timestamp as an old value and causing a backward jump. This would be an issue if the timecounter multiplier was so small that the update interval would not be limited by the 64-bit overflow in multiplication. Shorten the calculated interval to make sure the timecounter is updated in time even when the system clock is slowed down by up to 10%, the multiplier is increased by up to 10%, and the scheduled overflow check is late by 15%. Cc: Richard Cochran Cc: Ariel Levkovich Cc: Saeed Mahameed Signed-off-by: Miroslav Lichvar Signed-off-by: Saeed Mahameed Signed-off-by: Sasha Levin --- drivers/net/ethernet/mellanox/mlx5/core/lib/clock.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lib/clock.c b/drivers/net/ethernet/mellanox/mlx5/core/lib/clock.c index 0d90b1b4a3d3..2d6168ee99e8 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/lib/clock.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/lib/clock.c @@ -511,14 +511,14 @@ void mlx5_init_clock(struct mlx5_core_dev *mdev) ktime_to_ns(ktime_get_real())); /* Calculate period in seconds to call the overflow watchdog - to make - * sure counter is checked at least once every wrap around. + * sure counter is checked at least twice every wrap around. * The period is calculated as the minimum between max HW cycles count * (The clock source mask) and max amount of cycles that can be * multiplied by clock multiplier where the result doesn't exceed * 64bits. */ overflow_cycles = div64_u64(~0ULL >> 1, clock->cycles.mult); - overflow_cycles = min(overflow_cycles, clock->cycles.mask >> 1); + overflow_cycles = min(overflow_cycles, div_u64(clock->cycles.mask, 3)); ns = cyclecounter_cyc2ns(&clock->cycles, overflow_cycles, frac, &frac); -- 2.19.1