Re: [PATCH net v2] net: avoid registering new queue objects after device unregistration

Quoting Jakub Kicinski (2021-11-22 20:37:42)

On Mon, 22 Nov 2021 17:56:06 +0100 Antoine Tenart wrote:

quoted

Quoting Jakub Kicinski (2021-11-22 17:31:44)

quoted

On Mon, 22 Nov 2021 17:20:07 +0100 Antoine Tenart wrote:  

quoted

   veth_set_channels+0x1c3/0x550
   ethnl_set_channels+0x524/0x610  

The patch looks fine, but ethtool calls should not happen after
unregister IMHO. Too many drivers would be surprised. 

Maybe we should catch unregistered devices in ethnl_ops_begin()?  

That might be good to have indeed, I'll have a look. I'm not sure about
other paths that could do the same: if there are, we might want to keep
the check in this patch as well. Or WARN to catch future misuses?

My knee jerk reaction was to add a WARN() just because I can't think
why changing queue count after unregister would be needed. But it's not
really illegal and we do support it before register, so why not after
unregister..

It's a little tricky. The queue count can be changed before
registration (updating dev->real_num_rx/tx_queues only) but the queue
objects can't[1]. The symmetry would be to only allowing to update
dev->real_num_tx_queues after unregister (which seems useless).

This is what's actually done for Rx queues[2]. For Tx the issue we have
is a quirk that was added to allow qdiscs to remove their extra Tx
queues after unregister[3] as in unregister_netdevice_many the net
devices are unlisted and moved to NETREG_UNREGISTERING before the call
to dev_shutdown().

[1] https://elixir.bootlin.com/linux/v5.16-rc2/source/net/core/dev.c#L2918
[2] https://elixir.bootlin.com/linux/v5.16-rc2/source/net/core/dev.c#L2967
[3] 5c56580b74e5 ("net: Adjust TX queue kobjects if number of queues
    changes during unregister")

We can venture a warning with a comment saying that this is just for
catching bad uses and see if any driver hits it on a normal path?

With that a better fix seems to be what you suggested, in
ethnl_ops_begin. A WARN in netdev_queue_update_kobjects[4] detecting
misuses might be good as well as the logic is not that straightforward
here.

[4] Or in netif_set_real_num_tx_queues as it is already detecting when
    queues are being disabled. !disabling && NETREG_UNREGISTERING should
    be illegal or at least not update the kobjects.
    https://elixir.bootlin.com/linux/v5.16-rc2/source/net/core/dev.c#L2913

Thanks!
Antoine