On 7/19/19 11:17 PM, Michael S. Tsirkin wrote:
On Fri, Jul 19, 2019 at 09:31:34PM +0800, Fei Li wrote:
From: Fam Zheng <zhengfei...@bytedance.com>
This allows the backend to _not_ trap mmio read of the status register
after injecting IRQ in the data path, which can improve the performance
significantly by avoiding a vmexit for each interrupt.
More importantly it also makes it possible for Firecracker to hook up
virtio-mmio with vhost-net, in which case there isn't a way to implement
proper status register handling.
For a complete backend that does set either INT_CONFIG bit or INT_VRING
bit upon generating irq, what happens hasn't changed.
Signed-off-by: Fam Zheng <zhengfei...@bytedance.com>
This has a side effect of skipping vring callbacks
if they trigger at the same time with a config
interrupt.
I don't see why this is safe.
Good point! I think the block can be moved out from the else block and
run unconditionally then.
Fam
---
drivers/virtio/virtio_mmio.c | 4 +---
1 file changed, 1 insertion(+), 3 deletions(-)
diff --git a/drivers/virtio/virtio_mmio.c b/drivers/virtio/virtio_mmio.c
index e09edb5c5e06..9b42502b2204 100644
--- a/drivers/virtio/virtio_mmio.c
+++ b/drivers/virtio/virtio_mmio.c
@@ -295,9 +295,7 @@ static irqreturn_t vm_interrupt(int irq, void *opaque)
if (unlikely(status & VIRTIO_MMIO_INT_CONFIG)) {
virtio_config_changed(&vm_dev->vdev);
ret = IRQ_HANDLED;
- }
-
- if (likely(status & VIRTIO_MMIO_INT_VRING)) {
+ } else {
spin_lock_irqsave(&vm_dev->lock, flags);
list_for_each_entry(info, &vm_dev->virtqueues, node)
ret |= vring_interrupt(irq, info->vq);
--
2.11.0
