On Fri, 18 Oct 2019 at 16:54, Russell King - ARM Linux admin <li...@armlinux.org.uk> wrote: > > On Fri, Oct 18, 2019 at 04:37:55PM +0300, Vladimir Oltean wrote: > > On Fri, 18 Oct 2019 at 16:23, Russell King - ARM Linux admin > > <li...@armlinux.org.uk> wrote: > > > > > > On Fri, Oct 18, 2019 at 04:09:30PM +0300, Vladimir Oltean wrote: > > > > Hi Andrew, > > > > > > > > On Fri, 18 Oct 2019 at 16:01, Andrew Lunn <and...@lunn.ch> wrote: > > > > > > > > > > > Well, that's the tricky part. You're sending a frame out, with no > > > > > > guarantee you'll get the same frame back in. So I'm not sure that > > > > > > any > > > > > > identifiers put inside the frame will survive. > > > > > > How do the tests pan out for you? Do you actually get to trigger > > > > > > this > > > > > > check? As I mentioned, my NIC drops the frames with bad FCS. > > > > > > > > > > My experience is, the NIC drops the frame and increments some the > > > > > counter about bad FCS. I do very occasionally see a frame delivered, > > > > > but i guess that is 1/65536 where the FCS just happens to be good by > > > > > accident. So i think some other algorithm should be used which is > > > > > unlikely to be good when the FCS is accidentally good, or just check > > > > > the contents of the packet, you know what is should contain. > > > > > > > > > > Are there any NICs which don't do hardware FCS? Is that something we > > > > > realistically need to consider? > > > > > > > > > > > Yes, but remember, nobody guarantees that a frame with DMAC > > > > > > ff:ff:ff:ff:ff:ff on egress will still have it on its way back. > > > > > > Again, > > > > > > this all depends on how you plan to manage the rx-all ethtool > > > > > > feature. > > > > > > > > > > Humm. Never heard that before. Are you saying some NICs rewrite the > > > > > DMAN? > > > > > > > > > > > > > I'm just trying to understand the circumstances under which this > > > > kernel thread makes sense. > > > > Checking for FCS validity means that the intention was to enable the > > > > reception of frames with bad FCS. > > > > Bad FCS after bad RGMII setup/hold times doesn't mean there's a small > > > > guy in there who rewrites the checksum. It means that frame octets get > > > > garbled. All octets are just as likely to get garbled, including the > > > > SFD, preamble, DMAC, etc. > > > > All I'm saying is that, if the intention of the patch is to actually > > > > process the FCS of frames before and after, then it should actually > > > > put the interface in promiscuous mode, so that frames with a > > > > non-garbled SFD and preamble can still be received, even though their > > > > DMAC was the one that got garbled. > > > > > > Isn't the point of this to see which RGMII setting results in a working > > > setup? > > > > > > So, is it not true that what we're after is receiving a _correct_ frame > > > that corresponds to the frame that was sent out? > > > > > > > Only true if the MAC does not drop bad frames by itself. Then the FCS > > check in the kernel thread is superfluous. > > If a MAC driver doesn't drop bad frames, then surely it's buggy, since > there isn't (afaik) a way of marking a received skb with a FCS error. > Therefore, forwarding frames with bad FCS into the Linux networking > stack will allow the reception of bad frames as if they were good. > > All the network drivers I've looked at (and written), when encountering > a packet with an error, update the statistic counters and drop the > errored packet. > > Do you know of any that don't? >
I don't think you are following the big picture of what I am saying. I was trying to follow Florian's intention (first make sure I understand it) and suggest that the FCS checking code in the patch he submitted is not doing what it was intended to. I am getting apparent FCS mismatches reported by the program, when I know full well that the MAC I am testing on would have dropped those frames were they really invalid. We aren't saying anything in contradiction. > -- > RMK's Patch system: https://www.armlinux.org.uk/developer/patches/ > FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up > According to speedtest.net: 11.9Mbps down 500kbps up Thanks, -Vladimir