On Wed Aug 05 2020, Russell King - ARM Linux admin wrote:
> On Wed, Aug 05, 2020 at 01:04:31AM +0300, Grygorii Strashko wrote:
>> On 05/08/2020 00:44, Russell King - ARM Linux admin wrote:
>> > On Wed, Aug 05, 2020 at 12:34:47AM +0300, Grygorii Strashko wrote:
>> > > On 05/08/2020 00:07, Russell King - ARM Linux admin wrote:
>> > > > On Tue, Aug 04, 2020 at 11:56:12PM +0300, Grygorii Strashko wrote:
>> > > > > 
>> > > > > 
>> > > > > On 31/07/2020 13:06, Kurt Kanzenbach wrote:
>> > > > > > On Thu Jul 30 2020, Petr Machata wrote:
>> > > > > > > Kurt Kanzenbach <kurt@linutronix.de> writes:
>> > > > > > > 
>> > > > > > > > @@ -107,6 +107,37 @@ unsigned int ptp_classify_raw(const struct sk_buff *skb)
>> > > > > > > >     }
>> > > > > > > >     EXPORT_SYMBOL_GPL(ptp_classify_raw);
>> > > > > > > > +struct ptp_header *ptp_parse_header(struct sk_buff *skb, unsigned int type)
>> > > > > > > > +{
>> > > > > > > > +	u8 *data = skb_mac_header(skb);
>> > > > > > > > +	u8 *ptr = data;
>> > > > > > > 
>> > > > > > > One of the "data" and "ptr" variables is superfluous.
>> > > > > > 
>> > > > > > Yeah. Can be shortened to u8 *ptr = skb_mac_header(skb);
>> > > > > 
>> > > > > Actually usage of skb_mac_header(skb) breaks CPTS RX time-stamping on
>> > > > > am571x platform PATCH 6.
>> > > > > 
>> > > > > The CPSW RX timestamp requested after full packet put in SKB, but
>> > > > > before calling eth_type_trans().
>> > > > > 
>> > > > > So, skb->data pints on Eth header, but skb_mac_header() return garbage.
>> > > > > 
>> > > > > Below diff fixes it for me.
>> > > > 
>> > > > However, that's likely to break everyone else.
>> > > > 
>> > > > For example, anyone calling this from the mii_timestamper rxtstamp()
>> > > > method, the skb will have been classified with the MAC header pushed
>> > > > and restored, so skb->data points at the network header.
>> > > > 
>> > > > Your change means that ptp_parse_header() expects the MAC header to
>> > > > also be pushed.
>> > > > 
>> > > > Is it possible to adjust CPTS?
>> > > > 
>> > > > Looking at:
>> > > > drivers/net/ethernet/ti/cpsw.c... yes.
>> > > > drivers/net/ethernet/ti/cpsw_new.c... yes.
>> > > > drivers/net/ethernet/ti/netcp_core.c... unclear.
>> > > > 
>> > > > If not, maybe cpts should remain unconverted - I don't see any reason
>> > > > to provide a generic function for one user.
>> > > > 
>> > > 
>> > > Could it be an option to pass "u8 *ptr" instead of "const struct sk_buff *skb" as
>> > > input parameter to ptp_parse_header()?
>> > 
>> > It needs to read from the buffer, and in order to do that, it needs to
>> > validate that the buffer contains sufficient data.  So, at minimum it
>> > needs to be a pointer and size of valid data.
>> > 
>> > I was thinking about suggesting that as a core function, with a wrapper
>> > for the existing interface.
>> > 
>> 
>> Then length can be added.
>
> Actually, it needs more than that, because skb->data..skb->len already
> may contain the eth header or may not.
>
>> Otherwise not only CPTS can't benefit from this new API, but also
>> drivers like oki-semi/pch_gbe/pch_gbe_main.c -> pch_ptp_match()
>
> Again, this looks like it can be solved easily by swapping the position
> of these two calls:
>
>                         pch_rx_timestamp(adapter, skb);
>
>                         skb->protocol = eth_type_trans(skb, netdev);
>
>> or have to two have two APIs (name?).
>> 
>> ptp_parse_header1(struct sk_buff *skb, unsigned int type)
>> {
>> 	u8 *data = skb_mac_header(skb);
>> 
>> ptp_parse_header2(struct sk_buff *skb, unsigned int type)
>> {
>> 	u8 *data = skb->data;
>> 
>> everything else is the same.
>
> Actually, I really don't think we want 99% of users doing:
>
> 	hdr = ptp_parse_header(skb_mac_header(skb), skb->data, skb->len, type)
>
> or
>
> 	hdr = ptp_parse_header(skb_mac_header(skb), skb->data + skb->len, type);
>
> because that is what it will take, and this is starting to look
> really very horrid.

True.

>
> So, I repeat my question again: can netcp_core.c be adjusted to
> ensure that the skb mac header field is correctly set by calling
> eth_type_trans() prior to calling the rx hooks?  The other two
> cpts cases look easy to change, and the oki-semi also looks the
> same.

I think it's possible to adjust the netcp core. So, the time stamping is
done via

 gbe_rxhook() -> gbe_rxtstamp() -> cpts_rx_timestamp()

The hooks are called in netcp_process_one_rx_packet(). So, moving
eth_type_trans() before executing the hooks should work. Only one hook
is registered.

What do you think about it?

Thanks,
Kurt