On 6/3/21 9:04 PM, Dexuan Cui wrote:
From: Leonid Bloch <leonidb@xxxxxxxxxxxxxx>
Sent: Thursday, June 3, 2021 5:35 AM
To: KY Srinivasan <kys@xxxxxxxxxxxxx>; Haiyang Zhang
<haiyangz@xxxxxxxxxxxxx>; Stephen Hemminger
<sthemmin@xxxxxxxxxxxxx>; Wei Liu <wei.liu@xxxxxxxxxx>; Dexuan Cui
<decui@xxxxxxxxxxxxx>
Cc: linux-hyperv@xxxxxxxxxxxxxxx; netdev@xxxxxxxxxxxxxxx
Subject: [BUG] hv_netvsc: Unbind exits before the VFs bound to it are
unregistered
Hi,
When I try to unbind a network interface from hv_netvsc and bind it to
uio_hv_generic, once in a while I get the following kernel panic (please
note the first two lines: it seems as uio_hv_generic is registered
before the VF bound to hv_netvsc is unregistered):
[Jun 3 09:04] hv_vmbus: registering driver uio_hv_generic
[ +0.002215] hv_netvsc 5e089342-8a78-4b76-9729-25c81bd338fc eth2: VF
unregistering: eth5
[ +1.088078] BUG: scheduling while atomic: swapper/8/0/0x00010003
[ +0.000001] BUG: scheduling while atomic: swapper/3/0/0x00010003
[ +0.000001] BUG: scheduling while atomic: swapper/6/0/0x00010003
[ +0.000000] BUG: scheduling while atomic: swapper/7/0/0x00010003
[ +0.000005] Modules linked in:
[ +0.000001] Modules linked in:
[ +0.000001] uio_hv_generic
[ +0.000000] Modules linked in:
[ +0.000000] Modules linked in:
[ +0.000001] uio_hv_generic uio
[ +0.000001] uio
[ +0.000000] uio_hv_generic
[ +0.000000] uio_hv_generic
...
I run kernel 5.10.27, unmodified, besides RT patch v36, on Azure Stack
Edge platform, software version 2105 (2.2.1606.3320).
I perform the bind-unbind using the following script (please note the
comment inline):
net_uuid="f8615163-df3e-46c5-913f-f2d2f965ed0e"
dev_uuid="$(basename "$(readlink "/sys/class/net/eth1/device")")"
modprobe uio_hv_generic
echo "${net_uuid}" > /sys/bus/vmbus/drivers/uio_hv_generic/new_id
printf "%s" "${dev_uuid}" > /sys/bus/vmbus/drivers/hv_netvsc/unbind
### If I insert 'sleep 1' here - all works correctly
printf "%s" "${dev_uuid}" > /sys/bus/vmbus/drivers/uio_hv_generic/bind
Thanks,
Leonid.
It would be great if you can test the mainline kernel, which I suspect also
has the bug.
It looks like netvsc_remove() -> netvsc_unregister_vf() does the unbinding work
in a synchronous mannter. I don't know why the bug happens.
Right now I don't have a DPDK setup to test this, but I think the bug can
be worked around by unbinding the PCI VF device from the pci-hyperv driver
before unbinding the netvsc device, and re-binding the VF device after binding
the netvsc device to uio_hv_generic.
Thanks,
-- Dexuan
Hi Dexuan,
Thanks for your reply. I can check for myself only next week, as I am
out of office now, but do you think that the reason might be using
cancel_delayed_work_sync(), instead of cancel_delayed_work() in
netvsc_unregister_vf()?
And if the above is not correct, can you please advise on a way of
finding the corresponding VF device from userspace, given the kernel
name of the parent device? I did not find it in sysfs so far.
Thanks,
Leonid.
