From: longli@xxxxxxxxxxxxxxxxx <longli@xxxxxxxxxxxxxxxxx> Sent: Wednesday, February 26, 2025 4:46 PM
>
> Hyper-V may offer a non latency sensitive device with subchannels without
> monitor bit enabled. The decision is entirely on the Hyper-V host not
> configurable within guest.
>
> When a device has subchannels, also signal events for the subchannel
> if its monitor bit is disabled.
>
> Signed-off-by: Long Li <longli@xxxxxxxxxxxxx>
> ---
> drivers/uio/uio_hv_generic.c | 30 ++++++++++++++++++++++++------
> 1 file changed, 24 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/uio/uio_hv_generic.c b/drivers/uio/uio_hv_generic.c
> index 3976360d0096..8b6df598a728 100644
> --- a/drivers/uio/uio_hv_generic.c
> +++ b/drivers/uio/uio_hv_generic.c
> @@ -65,6 +65,16 @@ struct hv_uio_private_data {
> char send_name[32];
> };
>
> +static void set_event(struct vmbus_channel *channel, s32 irq_state)
> +{
> + channel->inbound.ring_buffer->interrupt_mask = !irq_state;
> + if (!channel->offermsg.monitor_allocated && irq_state) {
> + /* MB is needed for host to see the interrupt mask first */
> + virt_mb();
> + vmbus_setevent(channel);
A minor point, but vmbus_setevent() checks the "monitor_allocated"
flag, and if not set, then calls vmbus_set_event(). Since
monitor_allocated() is already checked here, couldn't vmbus_set_event()
be called directly? The existing code calls vmbus_setevent() so keeping
vmbus_setevent() is less of a change, but it is still doing a redundant
check.
> + }
> +}
> +
> /*
> * This is the irqcontrol callback to be registered to uio_info.
> * It can be used to disable/enable interrupt from user space processes.
> @@ -79,12 +89,13 @@ hv_uio_irqcontrol(struct uio_info *info, s32 irq_state)
> {
> struct hv_uio_private_data *pdata = info->priv;
> struct hv_device *dev = pdata->device;
> + struct vmbus_channel *primary, *sc;
>
> - dev->channel->inbound.ring_buffer->interrupt_mask = !irq_state;
> - virt_mb();
> + primary = dev->channel;
> + set_event(primary, irq_state);
>
> - if (!dev->channel->offermsg.monitor_allocated && irq_state)
> - vmbus_setevent(dev->channel);
> + list_for_each_entry(sc, &primary->sc_list, sc_list)
> + set_event(sc, irq_state);
Walking the sc_list usually requires holding vmbus_connection.channel_mutex.
Is there a reason it's safe to walk the list here without the mutex?
>
> return 0;
> }
> @@ -95,12 +106,19 @@ hv_uio_irqcontrol(struct uio_info *info, s32 irq_state)
> static void hv_uio_channel_cb(void *context)
> {
> struct vmbus_channel *chan = context;
> - struct hv_device *hv_dev = chan->device_obj;
> - struct hv_uio_private_data *pdata = hv_get_drvdata(hv_dev);
> + struct hv_device *hv_dev;
> + struct hv_uio_private_data *pdata;
>
> chan->inbound.ring_buffer->interrupt_mask = 1;
> virt_mb();
>
> + /*
> + * The callback may come from a subchannel, in which case look
> + * for the hv device in the primary channel
> + */
> + hv_dev = chan->primary_channel ?
> + chan->primary_channel->device_obj : chan->device_obj;
This certainly looks correct and necessary. But how did this work in the
past? Wouldn't DPDK running on a synthetic NIC have gotten callbacks on
a subchannel? I'm just trying to understand whether this a bug fix, or if
not, what new scenario requires this change. I'm not understanding how
this change is related to the commit message.
Michael
> + pdata = hv_get_drvdata(hv_dev);
> uio_event_notify(&pdata->info);
> }
>
> --
> 2.34.1
>
