Re: [net] net-sysfs: avoid registering new queue objects after device unregistration

Quoting David Miller (2021-10-26 16:30:57)

From: Antoine Tenart <atenart@kernel.org>
Date: Tue, 26 Oct 2021 15:38:22 +0200

quoted

netdev_queue_update_kobjects can be called after device unregistration
started (and device_del was called) resulting in two issues: possible
registration of new queue kobjects (leading to the following trace) and
providing a wrong 'old_num' number (because real_num_tx_queues is not
updated in the unregistration path).

  BUG: KASAN: use-after-free in kobject_get+0x14/0x90
  Read of size 1 at addr ffff88801961248c by task ethtool/755

  CPU: 0 PID: 755 Comm: ethtool Not tainted 5.15.0-rc6+ #778
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-4.fc34 04/014
  Call Trace:
   dump_stack_lvl+0x57/0x72
   print_address_description.constprop.0+0x1f/0x140
   kasan_report.cold+0x7f/0x11b
   kobject_get+0x14/0x90
   kobject_add_internal+0x3d1/0x450
   kobject_init_and_add+0xba/0xf0
   netdev_queue_update_kobjects+0xcf/0x200
   netif_set_real_num_tx_queues+0xb4/0x310
   veth_set_channels+0x1c3/0x550
   ethnl_set_channels+0x524/0x610

The fix for both is to only allow unregistering queue kobjects after a
net device started its unregistration and to ensure we know the current
Tx queue number (we update dev->real_num_tx_queues before returning).
This relies on the fact that dev->real_num_tx_queues is used for
'old_num' expect when firstly allocating queues.

(Rx queues are not affected as net_rx_queue_update_kobjects can't be
called after a net device started its unregistration).

Fixes: 5c56580b74e5 ("net: Adjust TX queue kobjects if number of queues changes during unregister")
Signed-off-by: Antoine Tenart <atenart@kernel.org>

netdev_queue_update_kobjects is a confusing function name, it sounds
like it handles both rx and tx.  It only handles tx so
net_tx_queue_update_kobjects is more appropriate.

Agreed.

Could you rename the function in this patch please?

As this is targeting stable kernels, shouldn't the rename be a separate
patch sent to net-next instead? (And it's not the only function that
should be renamed if we take this path, such as netdev_queue_add_kobject
and the functions in struct kobj_type netdev_queue_ktype).

Thanks!
Antoine