| Message ID | 1484587805-12666-4-git-send-email-tariqt@mellanox.com |
|---|---|
| State | Changes Requested, archived |
| Delegated to: | David Miller |
| Headers | show |
On Mon, Jan 16, 2017 at 7:29 PM, Tariq Toukan <tariqt@mellanox.com> wrote: > From: Daniel Jurgens <danielj@mellanox.com> > > Use CPUs on the close NUMA when setting the EQ affinity hints. Dan, are we sure there are no down-sides for always doing this? this code is probably there for many years and we're introducing here new behaviour to potentially to many Ms installs when they get distro update that includes this patch.
On 1/16/2017 3:44 PM, Or Gerlitz wrote: > On Mon, Jan 16, 2017 at 7:29 PM, Tariq Toukan <tariqt@mellanox.com> wrote: >> From: Daniel Jurgens <danielj@mellanox.com> >> >> Use CPUs on the close NUMA when setting the EQ affinity hints. > Dan, are we sure there are no down-sides for always doing this? this > code is probably there for many years and we're introducing here new > behaviour to potentially to many Ms installs when they get distro > update that includes this patch. > I don't see a downside, this just favors using the node local CPUs before others. I don't understand your 2nd sentence there. "Ms installs"?
From: Daniel Jurgens <danielj@mellanox.com> Date: Mon, 16 Jan 2017 21:54:36 +0000 > I don't understand your 2nd sentence there. "Ms installs"? I think he meant "Mellanox's installs", meaning installations of Linux using Mellanox hardware. At least that's how I read it.
On Mon, Jan 16, 2017 at 11:54 PM, Daniel Jurgens <danielj@mellanox.com> wrote: > On 1/16/2017 3:44 PM, Or Gerlitz wrote: >> On Mon, Jan 16, 2017 at 7:29 PM, Tariq Toukan <tariqt@mellanox.com> wrote: >>> From: Daniel Jurgens <danielj@mellanox.com> >>> >>> Use CPUs on the close NUMA when setting the EQ affinity hints. >> Dan, are we sure there are no down-sides for always doing this? this >> code is probably there for many years and we're introducing here new >> behaviour to potentially to many Ms installs when they get distro >> update that includes this patch. > I don't see a downside, this just favors using the node local CPUs before others. OK, so this just favors before others and not limits (not in front of the code now)? would be good to improve the change log and make this clear. > I don't understand your 2nd sentence there. "Ms installs"? Millions of installs that run Linux driver.
On 1/16/2017 3:59 PM, Or Gerlitz wrote: > On Mon, Jan 16, 2017 at 11:54 PM, Daniel Jurgens <danielj@mellanox.com> wrote: >> On 1/16/2017 3:44 PM, Or Gerlitz wrote: >>> On Mon, Jan 16, 2017 at 7:29 PM, Tariq Toukan <tariqt@mellanox.com> wrote: >>>> From: Daniel Jurgens <danielj@mellanox.com> >>>> >>>> Use CPUs on the close NUMA when setting the EQ affinity hints. >>> Dan, are we sure there are no down-sides for always doing this? this >>> code is probably there for many years and we're introducing here new >>> behaviour to potentially to many Ms installs when they get distro >>> update that includes this patch. > >> I don't see a downside, this just favors using the node local CPUs before others. > OK, so this just favors before others and not limits (not in front of > the code now)? would be good to improve the change log and make this > clear. Right, doesn't limit. Just favors the local node. I can work with Tariq to update the commit message if you think it necessary.
diff --git a/drivers/net/ethernet/mellanox/mlx4/main.c b/drivers/net/ethernet/mellanox/mlx4/main.c index 6efd66b0b822..058531bf7947 100644 --- a/drivers/net/ethernet/mellanox/mlx4/main.c +++ b/drivers/net/ethernet/mellanox/mlx4/main.c @@ -2843,7 +2843,8 @@ static int mlx4_init_affinity_hint(struct mlx4_dev *dev, int port, int eqn) if (!zalloc_cpumask_var(&eq->affinity_mask, GFP_KERNEL)) return -ENOMEM; - cpumask_set_cpu(requested_cpu, eq->affinity_mask); + cpumask_set_cpu(cpumask_local_spread(requested_cpu, dev->numa_node), + eq->affinity_mask); return 0; }