[PATCH 1/1 v3] inside-secure irq balance

[PATCH 1/1 v3] inside-secure irq balance

[Sven Auhagen]

Balance the irqs of the inside secure driver over all
available cpus.
Currently all interrupts are handled by the first CPU.

From my testing with IPSec AES-GCM 256
on my MCbin with 4 Cores I get a 50% speed increase:

Before the patch: 99.73 Kpps
With the patch: 151.25 Kpps

Signed-off-by: Sven Auhagen <sven.auhagen@voleatech.de>
---
v3:
* use NUMA_NO_NODE constant

v2:
* use cpumask_local_spread and remove affinity on
  module remove

 drivers/crypto/inside-secure/safexcel.c | 13 +++++++++++--
 drivers/crypto/inside-secure/safexcel.h |  3 +++
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/drivers/crypto/inside-secure/safexcel.c b/drivers/crypto/inside-secure/safexcel.c
index 2cb53fbae841..fb8e0d8732f8 100644
--- a/drivers/crypto/inside-secure/safexcel.c
+++ b/drivers/crypto/inside-secure/safexcel.c
@@ -1135,11 +1135,12 @@ static irqreturn_t safexcel_irq_ring_thread(int irq, void *data)
 
 static int safexcel_request_ring_irq(void *pdev, int irqid,
 				     int is_pci_dev,
+				     int ring_id,
 				     irq_handler_t handler,
 				     irq_handler_t threaded_handler,
 				     struct safexcel_ring_irq_data *ring_irq_priv)
 {
-	int ret, irq;
+	int ret, irq, cpu;
 	struct device *dev;
 
 	if (IS_ENABLED(CONFIG_PCI) && is_pci_dev) {
@@ -1177,6 +1178,10 @@ static int safexcel_request_ring_irq(void *pdev, int irqid,
 		return ret;
 	}
 
+	/* Set affinity */
+	cpu = cpumask_local_spread(ring_id, NUMA_NO_NODE);
+	irq_set_affinity_hint(irq, get_cpu_mask(cpu));
+
 	return irq;
 }
 
@@ -1611,6 +1616,7 @@ static int safexcel_probe_generic(void *pdev,
 		irq = safexcel_request_ring_irq(pdev,
 						EIP197_IRQ_NUMBER(i, is_pci_dev),
 						is_pci_dev,
+						i,
 						safexcel_irq_ring,
 						safexcel_irq_ring_thread,
 						ring_irq);
@@ -1619,6 +1625,7 @@ static int safexcel_probe_generic(void *pdev,
 			return irq;
 		}
 
+		priv->ring[i].irq = irq;
 		priv->ring[i].work_data.priv = priv;
 		priv->ring[i].work_data.ring = i;
 		INIT_WORK(&priv->ring[i].work_data.work,
@@ -1756,8 +1763,10 @@ static int safexcel_remove(struct platform_device *pdev)
 	clk_disable_unprepare(priv->reg_clk);
 	clk_disable_unprepare(priv->clk);
 
-	for (i = 0; i < priv->config.rings; i++)
+	for (i = 0; i < priv->config.rings; i++) {
+		irq_set_affinity_hint(priv->ring[i].irq, NULL);
 		destroy_workqueue(priv->ring[i].workqueue);
+	}
 
 	return 0;
 }
diff --git a/drivers/crypto/inside-secure/safexcel.h b/drivers/crypto/inside-secure/safexcel.h
index 94016c505abb..7c5fe382d272 100644
--- a/drivers/crypto/inside-secure/safexcel.h
+++ b/drivers/crypto/inside-secure/safexcel.h
@@ -707,6 +707,9 @@ struct safexcel_ring {
 	 */
 	struct crypto_async_request *req;
 	struct crypto_async_request *backlog;
+
+	/* irq of this ring */
+	int irq;
 };
 
 /* EIP integration context flags */
-- 
2.20.1

Re: [PATCH 1/1 v3] inside-secure irq balance

[Herbert Xu]

On Tue, Jul 21, 2020 at 06:37:59AM +0200, Sven Auhagen wrote:
> Balance the irqs of the inside secure driver over all
> available cpus.
> Currently all interrupts are handled by the first CPU.
> 
> >From my testing with IPSec AES-GCM 256
> on my MCbin with 4 Cores I get a 50% speed increase:
> 
> Before the patch: 99.73 Kpps
> With the patch: 151.25 Kpps
> 
> Signed-off-by: Sven Auhagen <sven.auhagen@voleatech.de>
> ---
> v3:
> * use NUMA_NO_NODE constant
> 
> v2:
> * use cpumask_local_spread and remove affinity on
>   module remove
> 
>  drivers/crypto/inside-secure/safexcel.c | 13 +++++++++++--
>  drivers/crypto/inside-secure/safexcel.h |  3 +++
>  2 files changed, 14 insertions(+), 2 deletions(-)

Patch applied.  Thanks.
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt