Message ID | 1428683240-22879-1-git-send-email-bpoirier@suse.de |
---|---|
State | Changes Requested, archived |
Delegated to: | David Miller |
Headers | show |
Hi Benjamin, On 4/10/2015 7:27 PM, Benjamin Poirier wrote: > By default, the number of tx queues is limited by the number of online cpus in > mlx4_en_get_profile(). However, this limit no longer holds after the ethtool > .set_channels method has been called. In that situation, the driver may access > invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. I must say I don't see the above issue with the current code. Whatever is the modified value of priv->num_tx_rings_p_up, it will set XPS only on queues which have been set with CPU affinity mask (no access to invalid bits). It's true that when priv->num_tx_rings_p_up > nr_cpus. not all queues will be set with XPS. This is because the code tries to preserve 1:1 mapping of queues to cores, to avoid a double mapping of queues to cores. I guess it's ok to break the 1:1 mapping in this condition, but the commit message should say that instead of invalid bits. Please fix me if I'm wrong. > Signed-off-by: Benjamin Poirier <bpoirier@suse.de> > --- > drivers/net/ethernet/mellanox/mlx4/en_tx.c | 8 +++++--- > 1 file changed, 5 insertions(+), 3 deletions(-) > > diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > index 55f9f5c..8c234ec 100644 > --- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c > +++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > @@ -143,8 +143,10 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv, > ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type; > ring->queue_index = queue_index; > > - if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index)) > - cpumask_set_cpu(queue_index, &ring->affinity_mask); > + if (queue_index < priv->num_tx_rings_p_up) > + cpumask_set_cpu_local_first(queue_index, > + priv->mdev->dev->numa_node, > + &ring->affinity_mask); Moving from cpumask_set_cpu to cpumask_set_cpu_local_first is great, but should come in a different commit, since the behavior of the XPS is changed here (xps_cpus[tx_ring[queue_index]] != queue_index from now). Commit should state of this behavior change. Thanks a lot Benjamin. > > *pring = ring; > return 0; > @@ -213,7 +215,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, > > err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context, > &ring->qp, &ring->qp_state); > - if (!user_prio && cpu_online(ring->queue_index)) > + if (!cpumask_empty(&ring->affinity_mask)) > netif_set_xps_queue(priv->dev, &ring->affinity_mask, > ring->queue_index); > -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 2015/04/12 10:03, Ido Shamay wrote: > Hi Benjamin, > > On 4/10/2015 7:27 PM, Benjamin Poirier wrote: > >By default, the number of tx queues is limited by the number of online cpus in > >mlx4_en_get_profile(). However, this limit no longer holds after the ethtool > >.set_channels method has been called. In that situation, the driver may access > >invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. > > I must say I don't see the above issue with the current code. > Whatever is the modified value of priv->num_tx_rings_p_up, it will set XPS > only on queues which have > been set with CPU affinity mask (no access to invalid bits). The problem is not with the call to netif_set_xps_queue() it is with the calls to cpu_online() and cpumask_set_cpu(). For example, if the user calls `ethtool -L ethX tx 32`, queue_index in mlx4_en_create_tx_ring() can be up to 255. Depending on CONFIG_NR_CPUS and CONFIG_CPUMASK_OFFSTACK this may result in calls to cpu_online() and cpumask_set_cpu() with cpu >= nr_cpumask_bits which is an invalid usage of the cpumask api. The driver will potentially read or write beyond the end of the bitmap. With CONFIG_CPUMASK_OFFSTACK=y and CONFIG_DEBUG_PER_CPU_MAPS=y, the aforementioned ethtool call on a system with <32 cpus triggers the warning in cpumask_check(). > > It's true that when priv->num_tx_rings_p_up > nr_cpus. not all queues will > be set with XPS. > This is because the code tries to preserve 1:1 mapping of queues to cores, > to avoid a double mapping > of queues to cores. > I guess it's ok to break the 1:1 mapping in this condition, but the commit > message should say that instead > of invalid bits. Please fix me if I'm wrong. > > >Signed-off-by: Benjamin Poirier <bpoirier@suse.de> > >--- > > drivers/net/ethernet/mellanox/mlx4/en_tx.c | 8 +++++--- > > 1 file changed, 5 insertions(+), 3 deletions(-) > > > >diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > >index 55f9f5c..8c234ec 100644 > >--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c > >+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > >@@ -143,8 +143,10 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv, > > ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type; > > ring->queue_index = queue_index; > >- if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index)) > >- cpumask_set_cpu(queue_index, &ring->affinity_mask); > >+ if (queue_index < priv->num_tx_rings_p_up) > >+ cpumask_set_cpu_local_first(queue_index, > >+ priv->mdev->dev->numa_node, > >+ &ring->affinity_mask); > Moving from cpumask_set_cpu to cpumask_set_cpu_local_first is great, but > should come in a different commit, since > the behavior of the XPS is changed here (xps_cpus[tx_ring[queue_index]] != > queue_index from now). > Commit should state of this behavior change. > Thanks a lot Benjamin. > > *pring = ring; > > return 0; > >@@ -213,7 +215,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, > > err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context, > > &ring->qp, &ring->qp_state); > >- if (!user_prio && cpu_online(ring->queue_index)) > >+ if (!cpumask_empty(&ring->affinity_mask)) > > netif_set_xps_queue(priv->dev, &ring->affinity_mask, > > ring->queue_index); > -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 2015/04/13 17:22, Benjamin Poirier wrote: > On 2015/04/12 10:03, Ido Shamay wrote: > > Hi Benjamin, > > > > On 4/10/2015 7:27 PM, Benjamin Poirier wrote: > > >By default, the number of tx queues is limited by the number of online cpus in > > >mlx4_en_get_profile(). However, this limit no longer holds after the ethtool > > >.set_channels method has been called. In that situation, the driver may access > > >invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. > > > > I must say I don't see the above issue with the current code. > > Whatever is the modified value of priv->num_tx_rings_p_up, it will set XPS > > only on queues which have > > been set with CPU affinity mask (no access to invalid bits). > > The problem is not with the call to netif_set_xps_queue() it is with the > calls to cpu_online() and cpumask_set_cpu(). > > For example, if the user calls `ethtool -L ethX tx 32`, queue_index in > mlx4_en_create_tx_ring() can be up to 255. Depending on CONFIG_NR_CPUS > and CONFIG_CPUMASK_OFFSTACK this may result in calls to cpu_online() and > cpumask_set_cpu() with cpu >= nr_cpumask_bits which is an invalid usage > of the cpumask api. The driver will potentially read or write beyond the > end of the bitmap. With CONFIG_CPUMASK_OFFSTACK=y and > CONFIG_DEBUG_PER_CPU_MAPS=y, the aforementioned ethtool call on a system > with <32 cpus triggers the warning in cpumask_check(). > Mellanox, can you please ack the patch as submitted, or clarify what changes you'd like to see given my reply above, or submit a fix of your own for this problem Thanks, -Benjamin > > > > It's true that when priv->num_tx_rings_p_up > nr_cpus. not all queues will > > be set with XPS. > > This is because the code tries to preserve 1:1 mapping of queues to cores, > > to avoid a double mapping > > of queues to cores. > > I guess it's ok to break the 1:1 mapping in this condition, but the commit > > message should say that instead > > of invalid bits. Please fix me if I'm wrong. > > > > >Signed-off-by: Benjamin Poirier <bpoirier@suse.de> > > >--- > > > drivers/net/ethernet/mellanox/mlx4/en_tx.c | 8 +++++--- > > > 1 file changed, 5 insertions(+), 3 deletions(-) > > > > > >diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > > >index 55f9f5c..8c234ec 100644 > > >--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c > > >+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c > > >@@ -143,8 +143,10 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv, > > > ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type; > > > ring->queue_index = queue_index; > > >- if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index)) > > >- cpumask_set_cpu(queue_index, &ring->affinity_mask); > > >+ if (queue_index < priv->num_tx_rings_p_up) > > >+ cpumask_set_cpu_local_first(queue_index, > > >+ priv->mdev->dev->numa_node, > > >+ &ring->affinity_mask); > > Moving from cpumask_set_cpu to cpumask_set_cpu_local_first is great, but > > should come in a different commit, since > > the behavior of the XPS is changed here (xps_cpus[tx_ring[queue_index]] != > > queue_index from now). > > Commit should state of this behavior change. > > Thanks a lot Benjamin. > > > *pring = ring; > > > return 0; > > >@@ -213,7 +215,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, > > > err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context, > > > &ring->qp, &ring->qp_state); > > >- if (!user_prio && cpu_online(ring->queue_index)) > > >+ if (!cpumask_empty(&ring->affinity_mask)) > > > netif_set_xps_queue(priv->dev, &ring->affinity_mask, > > > ring->queue_index); > > > -- > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > Please read the FAQ at http://www.tux.org/lkml/ -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 4/28/2015 6:26 AM, Benjamin Poirier wrote: > On 2015/04/13 17:22, Benjamin Poirier wrote: >> On 2015/04/12 10:03, Ido Shamay wrote: >>> Hi Benjamin, >>> >>> On 4/10/2015 7:27 PM, Benjamin Poirier wrote: >>>> By default, the number of tx queues is limited by the number of online cpus in >>>> mlx4_en_get_profile(). However, this limit no longer holds after the ethtool >>>> .set_channels method has been called. In that situation, the driver may access >>>> invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. >>> I must say I don't see the above issue with the current code. >>> Whatever is the modified value of priv->num_tx_rings_p_up, it will set XPS >>> only on queues which have >>> been set with CPU affinity mask (no access to invalid bits). >> The problem is not with the call to netif_set_xps_queue() it is with the >> calls to cpu_online() and cpumask_set_cpu(). >> >> For example, if the user calls `ethtool -L ethX tx 32`, queue_index in >> mlx4_en_create_tx_ring() can be up to 255. Depending on CONFIG_NR_CPUS >> and CONFIG_CPUMASK_OFFSTACK this may result in calls to cpu_online() and >> cpumask_set_cpu() with cpu >= nr_cpumask_bits which is an invalid usage >> of the cpumask api. The driver will potentially read or write beyond the >> end of the bitmap. With CONFIG_CPUMASK_OFFSTACK=y and >> CONFIG_DEBUG_PER_CPU_MAPS=y, the aforementioned ethtool call on a system >> with <32 cpus triggers the warning in cpumask_check(). >> > Mellanox, can you please > ack the patch as submitted, or > clarify what changes you'd like to see given my reply above, or > submit a fix of your own for this problem > > Thanks, > -Benjamin Hi Benjamin, After further review and better understanding of the issue, we are okay with your patch as is. Thanks for the good work. Acked-by: Ido Shamay <idos@mellanox.com> -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Apr 10, 2015 at 7:27 PM, Benjamin Poirier <bpoirier@suse.de> wrote: > By default, the number of tx queues is limited by the number of online cpus in > mlx4_en_get_profile(). However, this limit no longer holds after the ethtool > .set_channels method has been called. In that situation, the driver may access > invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. > Hi Benjamin, Can this fix be related to a specific commit? if yes, would be good if you can add here a Fixes: line so it would be easier to spot down to which stable kernels the fix should go. Or. > Signed-off-by: Benjamin Poirier <bpoirier@suse.de> -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c index 55f9f5c..8c234ec 100644 --- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c +++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c @@ -143,8 +143,10 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv, ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type; ring->queue_index = queue_index; - if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index)) - cpumask_set_cpu(queue_index, &ring->affinity_mask); + if (queue_index < priv->num_tx_rings_p_up) + cpumask_set_cpu_local_first(queue_index, + priv->mdev->dev->numa_node, + &ring->affinity_mask); *pring = ring; return 0; @@ -213,7 +215,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context, &ring->qp, &ring->qp_state); - if (!user_prio && cpu_online(ring->queue_index)) + if (!cpumask_empty(&ring->affinity_mask)) netif_set_xps_queue(priv->dev, &ring->affinity_mask, ring->queue_index);
By default, the number of tx queues is limited by the number of online cpus in mlx4_en_get_profile(). However, this limit no longer holds after the ethtool .set_channels method has been called. In that situation, the driver may access invalid bits of certain cpumask variables when queue_index > nr_cpu_ids. Signed-off-by: Benjamin Poirier <bpoirier@suse.de> --- drivers/net/ethernet/mellanox/mlx4/en_tx.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)