[PATCH v3 1/3] wifi: iwlwifi: pcie: optimize MSI-X interrupt affinity

[PATCH v3 1/3] wifi: iwlwifi: pcie: optimize MSI-X interrupt affinity

[Adrián García Casado]

Implement a balanced RSS queue distribution by skipping CPU0 for
high-rate MSI-X interrupts when multiple CPUs are online. This reduces
contention with system housekeeping tasks on the boot core and improves
overall throughput.

Signed-off-by: Adrián García Casado <adriangarciacasado42@gmail.com>
---
 drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c b/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
index 4560d92d7..87149f29e 100644
--- a/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
+++ b/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
@@ -1683,7 +1683,17 @@ static void iwl_pcie_irq_set_affinity(struct iwl_trans *trans,
 		 * Get the cpu prior to the place to search
 		 * (i.e. return will be > i - 1).
 		 */
+		/*
+		 * Balanced distribution: skip CPU0 for high-rate RSS queues
+		 * to avoid contention with system housekeeping.
+		 */
 		cpu = cpumask_next(i - offset, cpu_online_mask);
+		if (cpu >= nr_cpu_ids)
+			cpu = cpumask_first(cpu_online_mask);
+
+		if (cpu == 0 && num_online_cpus() > 1)
+			cpu = cpumask_next(0, cpu_online_mask);
+
 		cpumask_set_cpu(cpu, &trans_pcie->affinity_mask[i]);
 		ret = irq_set_affinity_hint(trans_pcie->msix_entries[i].vector,
 					    &trans_pcie->affinity_mask[i]);
-- 
2.47.3

[PATCH v3 1/3] wifi: iwlwifi: pcie: optimize MSI-X interrupt affinity

[Adrián García Casado]

Implement a balanced RSS queue distribution by skipping CPU0 for
high-rate MSI-X interrupts when multiple CPUs are online. This reduces
contention with system housekeeping tasks on the boot core and improves
overall throughput.

Signed-off-by: Adrián García Casado <adriangarciacasado42@gmail.com>
---
 drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c b/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
index 4560d92d7..87149f29e 100644
--- a/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
+++ b/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
@@ -1683,7 +1683,17 @@ static void iwl_pcie_irq_set_affinity(struct iwl_trans *trans,
 		 * Get the cpu prior to the place to search
 		 * (i.e. return will be > i - 1).
 		 */
+		/*
+		 * Balanced distribution: skip CPU0 for high-rate RSS queues
+		 * to avoid contention with system housekeeping.
+		 */
 		cpu = cpumask_next(i - offset, cpu_online_mask);
+		if (cpu >= nr_cpu_ids)
+			cpu = cpumask_first(cpu_online_mask);
+
+		if (cpu == 0 && num_online_cpus() > 1)
+			cpu = cpumask_next(0, cpu_online_mask);
+
 		cpumask_set_cpu(cpu, &trans_pcie->affinity_mask[i]);
 		ret = irq_set_affinity_hint(trans_pcie->msix_entries[i].vector,
 					    &trans_pcie->affinity_mask[i]);
-- 
2.47.3

Re: [PATCH v3 1/3] wifi: iwlwifi: pcie: optimize MSI-X interrupt affinity

[Johannes Berg]

Hi,

Somehow you messed up reposting this - the whole v2 series appears twice
and the v3 of this patch appears twice but without the rest of the
series?

> Implement a balanced RSS queue distribution by skipping CPU0 for
> high-rate MSI-X interrupts when multiple CPUs are online. This reduces
> contention with system housekeeping tasks on the boot core and improves
> overall throughput.

*Does* it improve throughput though? You should back up such claims with
numbers ... I doubt it would (unconditionally). It's not like the system
is going to be *really busy* all the time with housekeeping etc., and
the boot core is probably going to be a P-core, where you might have a
system with only 2 P-cores and 4-8 E (or even smaller) cores. This
likely also breaks some Intel-internal system requirements, but that
could be debated with the right folks.

Either way,

> --- a/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
> +++ b/drivers/net/wireless/intel/iwlwifi/pcie/gen1_2/trans.c
> @@ -1683,7 +1683,17 @@ static void iwl_pcie_irq_set_affinity(struct iwl_trans *trans,
>  		 * Get the cpu prior to the place to search
>  		 * (i.e. return will be > i - 1).
>  		 */
> +		/*
> +		 * Balanced distribution: skip CPU0 for high-rate RSS queues
> +		 * to avoid contention with system housekeeping.
> +		 */
>  		cpu = cpumask_next(i - offset, cpu_online_mask);
> +		if (cpu >= nr_cpu_ids)
> +			cpu = cpumask_first(cpu_online_mask);
> +
> +		if (cpu == 0 && num_online_cpus() > 1)
> +			cpu = cpumask_next(0, cpu_online_mask);
> +

this is wrong since you really then should allocate one queue less,
rather than mapping two queues to the same core.

johannes