Vlan EVENT patch

[Jack Vogel]

On Wed, Jun 11, 2008 at 7:16 PM, Sam Leffler <sam at errno.com> wrote:
> [trimming cc list to reduce spamage]
>
> Pyun YongHyeon wrote:
>>
>> On Wed, Jun 11, 2008 at 09:52:23AM -0700, Jack Vogel wrote:
>>  > On Wed, Jun 11, 2008 at 12:33 AM, Pyun YongHyeon <pyunyh at gmail.com>
>> wrote:
>>  > > On Tue, Jun 10, 2008 at 09:51:53AM -0700, Jack Vogel wrote:
>>  > >  > This is a small patch that Sam came up with for me, it will allow
>>  > >  > drivers to know
>>  > >  > when a vlan attaches.
>>  > >  >
>>  > >  > It is transparent to any code that doesn't want to change, but
>> this
>>  > >  > will allow my
>>  > >  > drivers to finally utilize the vlan hardware filter (something
>> Linux has had for
>>  > >  > ever but we lacked).
>>  > >  >
>>  > >
>>  > > Just curious, is there any rule how to use that new capability?
>>  > > Because drivers will receive events whenever VLAN tags are
>>  > > added/removed they would know how to act for these events. If
>>  > > promiscuous mode is on for interface, driver should not filter any
>>  > > VLAN tagged frames, right?
>>  > > If users want to disable VLAN hardware filtering feature what is
>>  > > best way to perform this? Introducing a new flag like
>>  > > IFCAP_VLAN_HWFILT or add a new sysctl that control this feature?
>>  > > I guess VIA Rhine III also have VLAN hardware filtering capability
>>  > > so it would be even better if we have a way to share common part.
>>  >  > All the patch does is have the vlan driver generate events when it
>> attaches
>>  > or detaches from a NIC, there are no rules, however I can tell you what
>>  > I'm coding into this in the Intel drivers.
>>  >  > The way it works is the driver registers a callback for each event,
>> I will
>>  > call that [igb,em,ixgbe]_register_vlan(), and unregister obviously.
>>  >  > Right now, the drivers just generically enable VLAN capability
>> because
>>  > there is never a trigger to know IF and WHEN you need to do so, but
>>  > with this change the VLAN capability will only get turned on by the
>>  > registration routine.
>>  >  > Most significantly, now when the pseudo device it gives the driver
>>  > the VLAN tag, this will mean the driver will be able from the start
>>  > to use the VFTA, the hardware filter, for each vlan attach the driver
>>  > will add the ID into this table.
>>  >  > The unregister event will turn the table entry off, and if this is
>> the
>>  > last VLAN being detached it will then disable the features.
>>  >  > Oh yes, these routines will also take care of the size change of
>>  > the frame due to the tag. I already have the changes in place in
>>  > the igb drive, and they are working great.
>>  >  > I do not understand why you think you need a flag to disable this,
>>  > yes it could be done, but why? If you need to do some sort of
>>  > debugging won't a system not using vlans and in promiscuous
>>  > mode do just fine?
>>  >
>> I guess this would be the same reason why FreeBSD have a way to
>> disable checksum offload for buggy hardware. Diabling all hardware
>> VLAN assistance due to broken VLAN filtering doesn't look right.
>>
>>  > It just seems to violate the whole reason for doing vlans in the
>>  > first place, however perhaps I am missing something? I do not
>>  > believe the Linux driver has some way to disable use of the table
>>  > but I'll double check on that.
>>  >  > Remember, this change requires NO driver changes unless they
>>  > wish to take advantage of the ability.
>>
>> Yes.
>>
>>  >  > Cheers,
>>  >  > Jack
>>
>> Thanks.
>
> Sounds like there needs to be a h/w vlan assist capability bit that controls
> this in the driver.  Then you'd have a way to disable via ifconfig w/ a
> trivial mod.

I don't want to create stuff in ifconfig when I'm not convinced
of the need. If there were, as he says, 'buggy hardware', specifically
buggy Intel hardware, then either our drivers would have had special
errata or workarounds in it for that, but none of the OS drivers have
any special code for issues involving VFTA (the filter) or other VLAN
controlling components that I am aware of.

If there are other network drivers that are buggy in this regard then why
encumber the generic interface due to that, let the drivers deal with it,
via sysctl or something of the sort.

There are enough real cases of hardware problems we need to address in
code that I don't just want to modify code on the mere theoretical possibility
of such.

How bout this, we put the change into HEAD, I add support as I've planned into
the em and igb drivers, and then we let them get tested, if a real problem comes
up, THEN we worry about adding special case code, how's that?

Regards,

Jack