Re: [PATCH net] net: dsa: microchip: fix initial port flush problem

[Simon Horman <horms@kernel.org>]

On Sun, Jun 02, 2024 at 05:01:18PM +0300, Vladimir Oltean wrote:
> On Sat, Jun 01, 2024 at 01:05:45PM +0100, Simon Horman wrote:
> > On Fri, May 31, 2024 at 07:19:54PM +0000, Tristram.Ha@microchip.com wrote:
> > > > Subject: Re: [PATCH net] net: dsa: microchip: fix initial port flush problem
> > > > 
> > > > EXTERNAL EMAIL: Do not click links or open attachments unless you know the content
> > > > is safe
> > > > 
> > > > On Tue, May 28, 2024 at 02:35:45PM -0700, Tristram.Ha@microchip.com wrote:
> > > > > From: Tristram Ha <tristram.ha@microchip.com>
> > > > >
> > > > > The very first flush in any port will flush all learned addresses in all
> > > > > ports.  This can be observed by unplugging a cable from one port while
> > > > > additional ports are connected and dumping the fdb entries.
> > > > >
> > > > > This problem is caused by the initially wrong value programmed to the
> > > > > register.  After the first flush the value is reset back to the normal so
> > > > > the next port flush will not cause such problem again.
> > > > 
> > > > Hi Tristram,
> > > > 
> > > > I think it would be worth spelling out why it is correct to:
> > > > 1. Not set SW_FLUSH_STP_TABLE or SW_FLUSH_MSTP_TABLE; and
> > > > 2. Preserve the value of the other bits of REG_SW_LUE_CTRL_1
> > > 
> > > Setting SW_FLUSH_STP_TABLE and SW_FLUSH_MSTP_TABLE bits are wrong as they
> > > are action bits.  The bit should be set only when doing an action like
> > > flushing.
> > 
> > Understood, thanks. And I guess that only bits that are being configured
> > should be changed, thus the values other bits are preserved with this
> > change.
> > 
> > FWIIW, I do think it would be worth adding something about this to the
> > patch description.
> 
> I agree the description is confusing and I had to look it up in the
> datasheet to understand.
> 
> I would suggest something along the lines of:
> 
> Setting the SW_FLUSH_STP_TABLE | SW_FLUSH_MSTP_TABLE bits of
> REG_SW_LUE_CTRL_1 does not do anything right away. They are
> just one-shot modifiers of the upcoming flush action executed by
> ksz9477_flush_dyn_mac_table().
> 
> It is wrong to set these bits at ksz9477_reset_switch() time, because
> it makes ksz9477_flush_dyn_mac_table() have an unexpected and incorrect
> behavior during its first run. When DSA calls ksz_port_fast_age() on a
> single port for the first time, due to this modifier being set, the
> entire FDB will be flushed of dynamically learned entries, across all
> ports.
> 
> Additionally, there is another mistake in the original code, which is
> that the value read from the REG_SW_LUE_CTRL_1 is immediately discarded,
> rather than preserved. The relevant bit which is set by default in this
> register (but we are mistakenly clearing) is:
> 
> Bit 3: Multicast Source Address Filtering
> 1 = Forward packets with a multicast source address
> 0 = Drop packets with a multicast source address

Thanks, that makes things a lot clearer to me.

> Tristram, now a question to you: why would we want to forward packets
> with a multicast source address? It looks like clearing that field is
> one of those things which were accidentally correct.
> 
> The cleanest way to not make a functional change where none is intended
> is to simply delete the read.

FWIIW, I thought about that too. But I was concerned that perhaps the read
has a side effect, because I don't know the hw well enough to say otherwise.