Add Enhanced Shared Interrupt (ESI) CPU affinity support to improve
UFS performance on Qualcomm platforms.

By Default, the IRQ core route interrupts to a limited number of
cores while other cores remain idle. This patch enables dynamic
interrupt affinity adjustment for better performance tuning by
distributing ESI interrupts across all online CPUs in round-robin
fashion.

This reduces CPU contention and enables better performance optimization
on Qualcomm UFS controllers by utilizing all available online CPUs.

Signed-off-by: Nitin Rawat <nitin.rawat@oss.qualcomm.com>
---
 drivers/ufs/host/ufs-qcom.c | 50 +++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)

diff --git a/drivers/ufs/host/ufs-qcom.c b/drivers/ufs/host/ufs-qcom.c
index 8ebee0cc5313..c43bb75d208c 100644
--- a/drivers/ufs/host/ufs-qcom.c
+++ b/drivers/ufs/host/ufs-qcom.c
@@ -2070,6 +2070,55 @@ static irqreturn_t ufs_qcom_mcq_esi_handler(int irq, void *data)
 	return IRQ_HANDLED;
 }

+/**
+ * ufs_qcom_set_esi_affinity - Set CPU affinity hints for ESI interrupts
+ * @hba: UFS host controller instance
+ *
+ * Sets CPU affinity hints for ESI interrupts to distribute them across
+ * online CPUs for better performance in round-robin fashion.
+ */
+static void ufs_qcom_set_esi_affinity(struct ufs_hba *hba)
+{
+	struct msi_desc *desc;
+	int ret, i = 0, nr_irqs = 0;
+	const cpumask_t *mask;
+	int cpu;
+
+	__msi_lock_descs(hba->dev);
+	/* Count the number of MSI descriptors */
+	msi_for_each_desc(desc, hba->dev, MSI_DESC_ALL) {
+		nr_irqs++;
+	}
+	__msi_unlock_descs(hba->dev);
+
+	if (nr_irqs == 0)
+		return;
+
+	__msi_lock_descs(hba->dev);
+	/* Set affinity hints for each interrupt in round-robin fashion */
+	msi_for_each_desc(desc, hba->dev, MSI_DESC_ALL) {
+		if (i >= nr_irqs)
+			break;
+
+		/* Distribute interrupts across online CPUs in round-robin */
+		cpu = cpumask_nth(i % num_online_cpus(), cpu_online_mask);
+		mask = get_cpu_mask(cpu);
+		if (!cpumask_subset(mask, cpu_online_mask)) {
+			dev_err(hba->dev, "Invalid CPU %d in map, using online CPUs\n",
+				cpu);
+			mask = cpu_online_mask;
+		}
+
+		ret = irq_set_affinity_hint(desc->irq, mask);
+		if (ret < 0)
+			dev_err(hba->dev, "Failed to set affinity hint to CPU %d for ESI IRQ %d, err = %d\n",
+				cpu, desc->irq, ret);
+
+		i++;
+	}
+	__msi_unlock_descs(hba->dev);
+}
+
 static int ufs_qcom_config_esi(struct ufs_hba *hba)
 {
 	struct ufs_qcom_host *host = ufshcd_get_variant(hba);
@@ -2122,6 +2171,7 @@ static int ufs_qcom_config_esi(struct ufs_hba *hba)
 			    REG_UFS_CFG3);
 	}
 	ufshcd_mcq_enable_esi(hba);
+	ufs_qcom_set_esi_affinity(hba);
 	host->esi_enabled = true;
 	return 0;
 }
--
2.34.1

On 1/22/26 6:13 AM, Nitin Rawat wrote:
> +static void ufs_qcom_set_esi_affinity(struct ufs_hba *hba)
> +{
> +	struct msi_desc *desc;
> +	int ret, i = 0, nr_irqs = 0;
> +	const cpumask_t *mask;
> +	int cpu;
> +
> +	__msi_lock_descs(hba->dev);
> +	/* Count the number of MSI descriptors */
> +	msi_for_each_desc(desc, hba->dev, MSI_DESC_ALL) {
> +		nr_irqs++;
> +	}
> +	__msi_unlock_descs(hba->dev);
> +
> +	if (nr_irqs == 0)
> +		return;
> +
> +	__msi_lock_descs(hba->dev);
> +	/* Set affinity hints for each interrupt in round-robin fashion */
> +	msi_for_each_desc(desc, hba->dev, MSI_DESC_ALL) {
> +		if (i >= nr_irqs)
> +			break;
> +
> +		/* Distribute interrupts across online CPUs in round-robin */
> +		cpu = cpumask_nth(i % num_online_cpus(), cpu_online_mask);
> +		mask = get_cpu_mask(cpu);
> +		if (!cpumask_subset(mask, cpu_online_mask)) {
> +			dev_err(hba->dev, "Invalid CPU %d in map, using online CPUs\n",
> +				cpu);
> +			mask = cpu_online_mask;
> +		}
> +
> +		ret = irq_set_affinity_hint(desc->irq, mask);
> +		if (ret < 0)
> +			dev_err(hba->dev, "Failed to set affinity hint to CPU %d for ESI IRQ %d, err = %d\n",
> +				cpu, desc->irq, ret);
> +
> +		i++;
> +	}
> +	__msi_unlock_descs(hba->dev);
> +}

Why an entirely new function for setting interrupt affinity? Why isn't
irq_create_affinity_masks() good enough? Are you aware that
devm_platform_get_irqs_affinity() calls irq_create_affinity_masks()?

Thanks,

Bart.