netdev: question on ndo_set_rx_mode() API

netdev: question on ndo_set_rx_mode() API

[Murali Karicheri]

Hello Netdev experts,

I am seeing an issue with netcp driver that has a mutex lock/unlock() 
call. When kernel hack debug options are enabled, I see a warning that 
this function is taking a mutex that can sleep and is not allowed. I am 
working to fix this. Looking at other drivers, I see many drivers such 
as e1000_main.c are not holding any driver specific lock as part of the 
API implementation. So my first attempt is to remove the mutex. But 
wondering what kind of synchronization is required in this API to run it 
properly on an SMP kernel. Based on my search following files are 
calling dev_change_flags() which in turn calls ndo_set_rx_mode()

Set-1
======
net/ipv4/devinet.c
net/ipv4/ipconfig.c
net/core/dev_ioctl.c

Set-2
=====
net/8021q/vlan.c
net/core/rtnetlink.c

Set-1 seems to call this with rtnl_lock (mutex) held. So there is 
already protection between processes that calls this function and driver 
doesn't need to provide any explicit synchronization. Is this correct?

For Set-2, I can't figure out in what context this is calling this API 
Can someone help me understand this?

However I see below.

Documentation/networking/netdevices.txt explains,

ndo_set_rx_mode:
Synchronization: netif_addr_lock spinlock.
Context: BHs disabled

So is there any synchronization required from the driver perspective? If 
yes, what kind of synchronization is needed? Thanks in advance for your 
response.

-- 
Murali Karicheri
Linux Kernel, Keystone

Re: netdev: question on ndo_set_rx_mode() API

[Murali Karicheri]

On 09/18/2015 01:07 PM, Murali Karicheri wrote:
> Hello Netdev experts,
>
> I am seeing an issue with netcp driver that has a mutex lock/unlock()
> call. When kernel hack debug options are enabled, I see a warning that
> this function is taking a mutex that can sleep and is not allowed. I am
> working to fix this. Looking at other drivers, I see many drivers such
> as e1000_main.c are not holding any driver specific lock as part of the
> API implementation. So my first attempt is to remove the mutex. But
> wondering what kind of synchronization is required in this API to run it
> properly on an SMP kernel. Based on my search following files are
> calling dev_change_flags() which in turn calls ndo_set_rx_mode()
>
> Set-1
> ======
> net/ipv4/devinet.c
> net/ipv4/ipconfig.c
> net/core/dev_ioctl.c
>
> Set-2
> =====
> net/8021q/vlan.c
> net/core/rtnetlink.c
>
> Set-1 seems to call this with rtnl_lock (mutex) held. So there is
> already protection between processes that calls this function and driver
> doesn't need to provide any explicit synchronization. Is this correct?
>
> For Set-2, I can't figure out in what context this is calling this API
> Can someone help me understand this?
>
> However I see below.
>
> Documentation/networking/netdevices.txt explains,
>
> ndo_set_rx_mode:
> Synchronization: netif_addr_lock spinlock.
> Context: BHs disabled
>
> So is there any synchronization required from the driver perspective? If
> yes, what kind of synchronization is needed? Thanks in advance for your
> response.
>
Any help??

Thanks.

-- 
Murali Karicheri
Linux Kernel, Keystone