Re: ethtool discrepancy between observed behaviour and comments

On Tue, Jun 15, 2021 at 07:30:39AM +0100, Kev Jackson wrote:

Hi,

I have been working with ethtool, ioctl and Mellanox multi-queue NICs and I think
I have discovered an issue with either the code or the comments that describe
the code or recent changes.

My focus here is simply the ethtool -L command to set the channels for a
multi-queue nic.  Running this command with the following params on a host with
a Mellanox NIC works fine:

ethtool -L eth0 combined 4

The code however seems to check that one of rx or tx must be set as part of the
command (channels.c):

        /* ensure there is at least one RX and one TX channel */
        if (!channels.combined_count && !channels.rx_count)
                err_attr = ETHTOOL_A_CHANNELS_RX_COUNT;
        else if (!channels.combined_count && !channels.tx_count)
                err_attr = ETHTOOL_A_CHANNELS_TX_COUNT;
        else
                err_attr = 0;

and (ioctl.c):

        /* ensure there is at least one RX and one TX channel */
        if (!channels.combined_count &&
            (!channels.rx_count || !channels.tx_count))
                return -EINVAL;

This check was added in commit 7be9251, with the comment:
"Having a channel config with no ability to RX or TX traffic is
clearly wrong. Check for this in the core so the drivers don't
have to."

However this comment and check is contradicted by a (much) older comment from
commit 8b5933c, "Most multiqueue drivers pair up RX and TX
queues so that most channels combine RX and TX work"

After working with ioctl and using a ETHTOOL_SCHANNELS command to try and set
the number of channels (ETHTOOL_GCHANNELS works perfectly fine), I noticed I was
always getting -EINVAL from ethtool_set_channels which led me to uncover these
differences between the behaviour of the ethtool binary and using the
ETHTOOL_SCHANNELS command via code.

After seeing this change I discovered the code in the latest ethtool is using
netlink.c commands and nl_schannels is the command used to set the channels (if
the kernel supports netlink).

nl_schannels was added in this commit: dd3ab09, which doesn't seem to have any
checks for setting rx_count/tx_count/combined_count (although I could have
missed them).

From putting all of this together, I have come to the conclusion that:
* ioctl / ETHTOOL_SCHANNELS is a legacy method of setting channels
* nl_schannels is the new / preferred method of setting channels
* ethtool has fallback code to run ioctl functions for commands which don't yet
* have a netlink equivalent

Our user experience is that ethtool -L currently does support (and
should continue to support) just setting combined_count rather than
having to set combined_count + one of rx_count/tx_count, which would
mean removing the check in the ioctl.c, ethtool_set_channels code to
make the netlink.c and ioctl.c commands consistent.

I don't understand what exactly are you reporting: do you have
a testcase where netlink and ioctl requests behave differently? Or do
you suggest that the sanity check should be also added to userspace
ethtool utility so that requests which do not make sense won't be passed
to kernel at all?

The logic behind the check is this: a combined channel can be used for
both receive and transmit, Rx and Tx channels only for one of them. Thus
the total number of receiving channels is combined_count + rx_count and
the total number of transmitting channels is combined_count + tx_count.

The conditions in ioctl and netlink paths are written in different ways
but they are equivalent: they prevent requests with no receiving (i.e.
combined or Rx) or no transmitting (i.e. combined or Tx) channel.
Examples:

  combined_count = 8, rx_count = 0, tx_count = 0 ... OK
  combined_count = 4, rx_count = 4, tx_count = 0 ... OK
  combined_count = 4, rx_count = 0, tx_count = 4 ... OK
  combined_count = 0, rx_count = 8, tx_count = 8 ... OK
  combined_count = 0, rx_count = 8, tx_count = 0 ... invalid
  combined_count = 0, rx_count = 0, tx_count = 8 ... invalid

Michal

Obviously the other approach is to add the check for setting one of rx_count /
tx_count into the nl_schannels function.

We're happy to provide a patch for either approach, but would like to
raise this as potentially a bug in the current code.

Thanks,
Kev