On 2022-08-10 09:21, Ido Schimmel wrote:
On Tue, Aug 09, 2022 at 10:00:49PM +0200, netdev@xxxxxxxxxxxxxxxxxxxx
wrote:
On 2022-08-09 11:20, Ido Schimmel wrote:
> On Mon, Aug 01, 2022 at 05:33:49PM +0200, netdev@xxxxxxxxxxxxxxxxxxxx
> wrote:
> > On 2022-07-13 14:39, Ido Schimmel wrote:
> >
> > >
> > > What are "Storm Prevention" and "zero-DPV" FDB entries?
> > >
> >
> > For the zero-DPV entries, I can summarize:
> >
> > Since a CPU can become saturated from constant SA Miss Violations
> > from a
> > denied source, source MAC address are masked by loading a zero-DPV
> > (Destination Port Vector) entry in the ATU. As the address now
> > appears in
> > the database it will not cause more Miss Violations. ANY port trying
> > to send
> > a frame to this unauthorized address is discarded. Any locked port
> > trying to
> > use this unauthorized address has its frames discarded too (as the
> > ports SA
> > bit is not set in the ATU entry).
>
> What happens to unlocked ports that have learning enabled and are trying
> to use this address as SMAC? AFAICT, at least in the bridge driver, the
> locked entry will roam, but will keep the "locked" flag, which is
> probably not what we want. Let's see if we can agree on these semantics
> for a "locked" entry:
The next version of this will block forwarding to locked entries in
the
bridge, so they will behave like the zero-DPV entries.
I'm talking about roaming, not forwarding. Let's say you have a locked
entry with MAC X pointing to port Y. Now you get a packet with SMAC X
from port Z which is unlocked. Will the FDB entry roam to port Z? I
think it should, but at least in current implementation it seems that
the "locked" flag will not be reset and having locked entries pointing
to an unlocked port looks like a bug.
Remember that zero-DPV entries blackhole (mask) the MAC, so whenever a
packet appears with the same MAC on another port it is just dropped in
the HW, so there is no possibility of doing any CPU processing in this
case. Only after the timeout (5 min) can the MAC get a normal ATU on an
open port.
For the bridge to do what you suggest, a FDB search would be needed
afaics, and this might be in a process sensitive part of the code, thus
leading to too heavy a cost.
>
> 1. It discards packets with matching DMAC, regardless of ingress port. I
> read the document [1] you linked to in a different reply and could not
> find anything against this approach, so this might be fine or at least
> not very significant.
>
> Note that this means that "locked" entries need to be notified to device
> drivers so that they will install a matching entry in the HW FDB.
Okay, so as V4 does (just without the error noted).
>
> 2. It is not refreshed and has ageing enabled. That is, after initial
> installation it will be removed by the bridge driver after configured
> ageing time unless converted to a regular (unlocked) entry.
>
> I assume this allows you to remove the timer implementation from your
> driver and let the bridge driver notify you about the removal of this
> entry.
Okay, but only if the scheme is not so that the driver creates the
locked
entries itself, unless you indicate that the driver notifies the
bridge,
which then notifies back to the driver and installs the zero-DPV
entry? If
not I think the current implementation for the mv88e6xxx is fine.
I don't see a problem in having the driver notifying the bridge about
the installation of this entry and the bridge notifying the driver that
the entry needs to be removed. It removes complexity from device
drivers
like mv88e6xxx and doesn't add extra complexity to the bridge driver.
Actually, there is one complication, 'SWITCHDEV_FDB_ADD_TO_BRIDGE' will
add the locked entry as externally learned, which means the bridge will
not age it. Might need something like this:
diff --git a/net/bridge/br_fdb.c b/net/bridge/br_fdb.c
index e7f4fccb6adb..5f73d0b44ed9 100644
--- a/net/bridge/br_fdb.c
+++ b/net/bridge/br_fdb.c
@@ -530,7 +530,8 @@ void br_fdb_cleanup(struct work_struct *work)
unsigned long this_timer = f->updated + delay;
if (test_bit(BR_FDB_STATIC, &f->flags) ||
- test_bit(BR_FDB_ADDED_BY_EXT_LEARN, &f->flags)) {
+ (test_bit(BR_FDB_ADDED_BY_EXT_LEARN, &f->flags) &&
+ !test_bit(BR_FDB_ENTRY_LOCKED, &f->flags))) {
if (test_bit(BR_FDB_NOTIFY, &f->flags)) {
if (time_after(this_timer, now))
work_delay = min(work_delay,
There is a case of ownership of the FDB/ATU entry, which if I remember
correctly, will point to the current implementation being the right way
to do it, thus having the driver keeping ownership of the entry and
thereby also ageing it, but I think Vladimir should have his say here.
>
> 3. With regards to roaming, the entry cannot roam between locked ports
> (they need to have learning disabled anyway), but can roam to an
> unlocked port, in which case it becomes a regular entry that can roam
> and age.
>
> If we agree on these semantics, then I can try to verify that at least
> Spectrum can support them (it seems mv88e6xxx can).
The consensus here is that at least for the mv88e6xxx, learning should
be on
and link local learning should be blocked by the userspace setting you
pointed to earlier.
Why learning needs to be on in the bridge (not mv88e6xxx) driver?
I think it is seen as 'cheating' to enable learning only in the driver
behind the scenes, so kind of hackish. E.g. 'bridge -d link show' will
then report 'learning off', while learning is on in the driver.
And learning needs to be on for the driver as discussed earlier, which
only gives rise to the link local learning problem, which is then solved
by the user space setting.
