The mvpp2 driver can't cope at all with the TX affinities being
changed from userspace, and spit an endless stream of
[ 91.779920] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
[ 91.779930] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
[ 91.780402] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
[ 91.780406] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
[ 91.780415] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
[ 91.780418] mvpp2 f4000000.ethernet eth2: wrong cpu on the end of Tx processing
rendering the box completely useless (I've measured around 600k
interrupts/s on a 8040 box) once irqbalance kicks in and start
doing its job.
Obviously, the driver was never designed with this in mind. So let's
work around the problem by preventing userspace from interacting
with these interrupts altogether.
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
---
drivers/net/ethernet/marvell/mvpp2.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/drivers/net/ethernet/marvell/mvpp2.c b/drivers/net/ethernet/marvell/mvpp2.c
index a37af5813f33..fcf9ba5eb8d1 100644
--- a/drivers/net/ethernet/marvell/mvpp2.c
+++ b/drivers/net/ethernet/marvell/mvpp2.c
@@ -6747,6 +6747,9 @@ static int mvpp2_irqs_init(struct mvpp2_port *port)
for (i = 0; i < port->nqvecs; i++) {
struct mvpp2_queue_vector *qv = port->qvecs + i;
+ if (qv->type == MVPP2_QUEUE_VECTOR_PRIVATE)
+ irq_set_status_flags(qv->irq, IRQ_NO_BALANCING);
+
err = request_irq(qv->irq, mvpp2_isr, 0, port->dev->name, qv);
if (err)
goto err;
@@ -6776,6 +6779,7 @@ static void mvpp2_irqs_deinit(struct mvpp2_port *port)
struct mvpp2_queue_vector *qv = port->qvecs + i;
irq_set_affinity_hint(qv->irq, NULL);
+ irq_clear_status_flags(qv->irq, IRQ_NO_BALANCING);
free_irq(qv->irq, qv);
}
}
--
2.11.0