RE: [PATCH 1/1] Drivers: hv: vmbus: Limit channel interrupt scan to relid high water mark

[mhklkml]

From: vdso@xxxxxxxxxxx <vdso@xxxxxxxxxxx> Sent: Friday, February 20, 2026 5:48 PM
>
> Hi Michael,
> 
> Boots for me on an x86_64 machine. Got a typo fix and a question for you.
> Tagging as reviewed and tested regardless :)
> 
> > On 02/20/2026 8:40 AM  Michael Kelley <mhklkml@xxxxxxxxxxxx> wrote:
> >
> >
> > From: Michael Kelley <mhklinux@xxxxxxxxxxx>
> >
> > When checking for VMBus channel interrutps, current code always scans the
> 
> /s/interrutps/interrupts
> 
> > full SynIC receive interrupt bit array to get the relid of the
> > interrupting channels. The array has HV_EVENT_FLAGS_COUNT (2048) bits.
> > But VMs rarely have more than 100 channels, and the relid is typically
> > a small integer that is densely assigned by the Hyper-V host. It's
> > wasteful to scan 2048 bits when it is highly unlikely that anything will
> > be found past bit 100. The waste is double with Confidential VMBus because
> > there are two receive interrupt arrays that must be scanned: one for the
> > hypervisor SynIC and one for the paravisor SynIC.
> >
> > Improve the scanning by tracking the largest relid that has been offered
> > by the Hyper-V host. Then when checking for VMBus channel interrupts, only
> > scan up to this high water mark.
> >
> > When channels are rescinded, it's not worth the complexity to recalculate
> > the high water mark. Hyper-V tends to reuse the rescinded relids for any
> > new channels that are subsequently added, and the performance benefit of
> > exactly tracking the high water mark would be minimal.
> >
> > Signed-off-by: Michael Kelley <mhklinux@xxxxxxxxxxx>
> 
> Tested-by: Roman Kisel <vdso@xxxxxxxxxxx>
> Reviewed-by: Roman Kisel <vdso@xxxxxxxxxxx>

Thanks!

> 
> > ---
> >  drivers/hv/channel_mgmt.c | 16 ++++++++++++----
> >  drivers/hv/hyperv_vmbus.h |  3 ++-
> >  drivers/hv/vmbus_drv.c    |  7 +------
> >  3 files changed, 15 insertions(+), 11 deletions(-)
> >
> > diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
> > index 74fed2c073d4..61f7dffd0f50 100644
> > --- a/drivers/hv/channel_mgmt.c
> > +++ b/drivers/hv/channel_mgmt.c
> > @@ -384,8 +384,18 @@ static void free_channel(struct vmbus_channel *channel)
> >
> >  void vmbus_channel_map_relid(struct vmbus_channel *channel)
> >  {
> > -	if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
> > +	u32 new_relid = channel->offermsg.child_relid;
> > +
> > +	if (WARN_ON(new_relid >= MAX_CHANNEL_RELIDS))
> >  		return;
> > +
> > +	/*
> > +	 * This function is always called in the tasklet for the connect CPU.
> > +	 * So updating the relid hiwater mark does not need to be atomic.
> > +	 */
> > +	if (new_relid > READ_ONCE(vmbus_connection.relid_hiwater))
> > +		WRITE_ONCE(vmbus_connection.relid_hiwater, new_relid);
> > +
> >  	/*
> >  	 * The mapping of the channel's relid is visible from the CPUs that
> >  	 * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
> > @@ -411,9 +421,7 @@ void vmbus_channel_map_relid(struct vmbus_channel *channel)
> >  	 *      of the VMBus driver and vmbus_chan_sched() can not run before
> >  	 *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
> >  	 */
> > -	virt_store_mb(
> > -		vmbus_connection.channels[channel->offermsg.child_relid],
> > -		channel);
> > +	virt_store_mb(vmbus_connection.channels[new_relid], channel);
> >  }
> >
> >  void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
> > diff --git a/drivers/hv/hyperv_vmbus.h b/drivers/hv/hyperv_vmbus.h
> > index 7bd8f8486e85..2c90c81a3b0f 100644
> > --- a/drivers/hv/hyperv_vmbus.h
> > +++ b/drivers/hv/hyperv_vmbus.h
> > @@ -276,8 +276,9 @@ struct vmbus_connection {
> >  	struct list_head chn_list;
> >  	struct mutex channel_mutex;
> >
> > -	/* Array of channels */
> > +	/* Array of channel pointers, indexed by relid */
> >  	struct vmbus_channel **channels;
> > +	u32 relid_hiwater;
> >
> >  	/*
> >  	 * An offer message is handled first on the work_queue, and then
> > diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c
> > index 3e7a52918ce0..a96da105b593 100644
> > --- a/drivers/hv/vmbus_drv.c
> > +++ b/drivers/hv/vmbus_drv.c
> > @@ -1258,17 +1258,12 @@ static void vmbus_chan_sched(void *event_page_addr)
> >  		return;
> >  	event = (union hv_synic_event_flags *)event_page_addr + VMBUS_MESSAGE_SINT;
> >
> > -	maxbits = HV_EVENT_FLAGS_COUNT;
> > +	maxbits = READ_ONCE(vmbus_connection.relid_hiwater) + 1;
> 
> Worth checking that "maxbits <= HV_EVENT_FLAGS_COUNT" to protect from
> corruptions, etc. or would be too paranoidal?

We definitely want to validate what Hyper-V returns to the guest as a relid,
and drop any values that are "too big", so we don't go indexing off into
bogus memory. But that validation is done in vmbus_channel_map_relid()
with a WARN_ON() before setting relid_hiwater.  So there's no way for
relid_hiwater to be bogus, and additional validation here in
vmbus_chan_sched() really isn't necessary.

Michael

> 
> >  	recv_int_page = event->flags;
> >
> >  	if (unlikely(!recv_int_page))
> >  		return;
> >
> > -	/*
> > -	 * Suggested-by: Michael Kelley <mhklinux@xxxxxxxxxxx>
> > -	 * One possible optimization would be to keep track of the largest relID that's in use,
> > -	 * and only scan up to that relID.
> > -	 */
> >  	for_each_set_bit(relid, recv_int_page, maxbits) {
> >  		void (*callback_fn)(void *context);
> >  		struct vmbus_channel *channel;
> > --
> > 2.25.1