Date:   Mon, 27 Dec 2021 00:12:15 +0200
From:   Petko Manolov <petkan@...leusys.com>
To:     Matthias-Christian Ott <ott@...ix.org>
Cc:     linux-usb@...r.kernel.org, netdev@...r.kernel.org
Subject: Re: [PATCH] net: usb: pegasus: Request Ethernet FCS from hardware

On 21-12-26 17:12:24, Matthias-Christian Ott wrote:
> On 26/12/2021 17:02, Petko Manolov wrote:
> > On 21-12-26 14:25:02, Matthias-Christian Ott wrote:
> >> Commit 1a8deec09d12 ("pegasus: fixes reported packet length") tried to
> >> configure the hardware to not include the FCS/CRC of Ethernet frames.
> >> Unfortunately, this does not work with the D-Link DSB-650TX (USB IDs
> >> 2001:4002 and 2001:400b): the transferred "packets" (in the terminology
> >> of the hardware) still contain 4 additional octets. For IP packets in
> >> Ethernet this is not a problem as IP packets contain their own lengths
> >> fields but other protocols also see Ethernet frames that include the FCS
> >> in the payload which might be a problem for some protocols.
> >>
> >> I was not able to open the D-Link DSB-650TX as the case is a very tight
> >> press fit and opening it would likely destroy it. However, according to
> >> the source code the earlier revision of the D-Link DSB-650TX (USB ID
> >> 2001:4002) is a Pegasus (possibly AN986) and not Pegasus II (AN8511)
> >> device. I also tried it with the later revision of the D-Link DSB-650TX
> >> (USB ID 2001:400b) which is a Pegasus II device according to the source
> >> code and had the same results. Therefore, I'm not sure whether the RXCS
> >> (rx_crc_sent) field of the EC0 (Ethernet control_0) register has any
> >> effect or in which revision of the hardware it is usable and has an
> >> effect. As a result, it seems best to me to revert commit
> >> 1a8deec09d12 ("pegasus: fixes reported packet length") and to set the
> >> RXCS (rx_crc_sent) field of the EC0 (Ethernet control_0) register so
> >> that the FCS/CRC is always included.
> >>
> >> Fixes: 1a8deec09d12 ("pegasus: fixes reported packet length")
> >> Signed-off-by: Matthias-Christian Ott <ott@...ix.org>
> >> ---
> >>  drivers/net/usb/pegasus.c | 15 ++++++++++++++-
> >>  1 file changed, 14 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/net/usb/pegasus.c b/drivers/net/usb/pegasus.c
> >> index c4cd40b090fd..140d11ae6688 100644
> >> --- a/drivers/net/usb/pegasus.c
> >> +++ b/drivers/net/usb/pegasus.c
> >> @@ -422,7 +422,13 @@ static int enable_net_traffic(struct net_device *dev, struct usb_device *usb)
> >>  	ret = read_mii_word(pegasus, pegasus->phy, MII_LPA, &linkpart);
> >>  	if (ret < 0)
> >>  		goto fail;
> >> -	data[0] = 0xc8; /* TX & RX enable, append status, no CRC */
> >> +	/* At least two hardware revisions of the D-Link DSB-650TX (USB IDs
> >> +	 * 2001:4002 and 2001:400b) include the Ethernet FCS in the packets,
> >> +	 * even if RXCS is set to 0 in the EC0 register and the hardware is
> >> +	 * instructed to not include the Ethernet FCS in the packet.Therefore,
> >> +	 * it seems best to set RXCS to 1 and later ignore the Ethernet FCS.
> >> +	 */
> >> +	data[0] = 0xc9; /* TX & RX enable, append status, CRC */
> >>  	data[1] = 0;
> >>  	if (linkpart & (ADVERTISE_100FULL | ADVERTISE_10FULL))
> >>  		data[1] |= 0x20;	/* set full duplex */
> >> @@ -513,6 +519,13 @@ static void read_bulk_callback(struct urb *urb)
> >>  		pkt_len = buf[count - 3] << 8;
> >>  		pkt_len += buf[count - 4];
> >>  		pkt_len &= 0xfff;
> >> +		/* The FCS at the end of the packet is ignored. So subtract
> >> +		 * its length to ignore it.
> >> +		 */
> >> +		pkt_len -= ETH_FCS_LEN;
> >> +		/* Subtract the length of the received status at the end of the
> >> +		 * packet as it is not part of the Ethernet frame.
> >> +		 */
> >>  		pkt_len -= 4;
> >>  	}
> > 
> > Nice catch.  However, changing these constants for all devices isn't such a
> > good idea.  I'd rather use vendor and device IDs to distinguish these two
> > cases in the above code.
> 
> I don't think that it would hurt to include the FCS for all devices. I only
> have the datasheets for the ADM8511/X and the ADM8513 but it seems that all
> devices that are supported by the driver also include the RXCS field in EC0.
> This was also the previous behaviour before commit 1a8deec09d12 and seemed to
> have worked. It also only adds four octet that have to be transferred and it
> seems to avoid exceptions for different devices which seems to be a good idea,
> in particular, because it is not easy to acquire all of the supported devices
> as they are no longer sold or manufactured.

The fix that commit 1a8deec09d12 introduces is real (the commit message makes
sense) and i don't feel confident to revert it so lightly.  I think i have all
relevant datasheets somewhere, along with a couple of old "pegasus I" devices,
which i could use for testing. Not at home right now, the aforementioned testing
will have to wait a couple of days.

> That being said, if you are going to veto this change otherwise, I can of
> course just add the FCS back for the two USB IDs, even though it likely
> affects other devices as well.

Like i said, i don't want to hurry up and revert something that looks like a
valid fix.  Especially after five years worth of kernel releases and no
complaints related to 1a8deec09d12.  This should mean two things: a) the driver
isn't used anymore, or b) this commit fixes a real problem.

However, if it turn out that your fix is the right one, it goes in without fuss.
So lets see what it is...


cheers,
Petko

Powered by blists - more mailing lists