Re: [PATCH] xen/netfront: destroy queues before real_num_tx_queues is zeroed

[Jakub Kicinski <kuba@xxxxxxxxxx>]

On Mon, 21 Feb 2022 07:27:32 +0100 Juergen Gross wrote:
> On 20.02.22 14:42, Marek Marczykowski-Górecki wrote:
> > xennet_destroy_queues() relies on info->netdev->real_num_tx_queues to
> > delete queues. Since d7dac083414eb5bb99a6d2ed53dc2c1b405224e5
> > ("net-sysfs: update the queue counts in the unregistration path"),
> > unregister_netdev() indirectly sets real_num_tx_queues to 0. Those two
> > facts together means, that xennet_destroy_queues() called from
> > xennet_remove() cannot do its job, because it's called after
> > unregister_netdev(). This results in kfree-ing queues that are still
> > linked in napi, which ultimately crashes:
> > 
> >      BUG: kernel NULL pointer dereference, address: 0000000000000000
> >      #PF: supervisor read access in kernel mode
> >      #PF: error_code(0x0000) - not-present page
> >      PGD 0 P4D 0
> >      Oops: 0000 [#1] PREEMPT SMP PTI
> >      CPU: 1 PID: 52 Comm: xenwatch Tainted: G        W         
> > 5.16.10-1.32.fc32.qubes.x86_64+ #226
> >      RIP: 0010:free_netdev+0xa3/0x1a0
> >      Code: ff 48 89 df e8 2e e9 00 00 48 8b 43 50 48 8b 08 48 8d b8 a0 fe 
> > ff ff 48 8d a9 a0 fe ff ff 49 39 c4 75 26 eb 47 e8 ed c1 66 ff <48> 8b 85 
> > 60 01 00 00 48 8d 95 60 01 00 00 48 89 ef 48 2d 60 01 00
> >      RSP: 0000:ffffc90000bcfd00 EFLAGS: 00010286
> >      RAX: 0000000000000000 RBX: ffff88800edad000 RCX: 0000000000000000
> >      RDX: 0000000000000001 RSI: ffffc90000bcfc30 RDI: 00000000ffffffff
> >      RBP: fffffffffffffea0 R08: 0000000000000000 R09: 0000000000000000
> >      R10: 0000000000000000 R11: 0000000000000001 R12: ffff88800edad050
> >      R13: ffff8880065f8f88 R14: 0000000000000000 R15: ffff8880066c6680
> >      FS:  0000000000000000(0000) GS:ffff8880f3300000(0000) 
> > knlGS:0000000000000000
> >      CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >      CR2: 0000000000000000 CR3: 00000000e998c006 CR4: 00000000003706e0
> >      Call Trace:
> >       <TASK>
> >       xennet_remove+0x13d/0x300 [xen_netfront]
> >       xenbus_dev_remove+0x6d/0xf0
> >       __device_release_driver+0x17a/0x240
> >       device_release_driver+0x24/0x30
> >       bus_remove_device+0xd8/0x140
> >       device_del+0x18b/0x410
> >       ? _raw_spin_unlock+0x16/0x30
> >       ? klist_iter_exit+0x14/0x20
> >       ? xenbus_dev_request_and_reply+0x80/0x80
> >       device_unregister+0x13/0x60
> >       xenbus_dev_changed+0x18e/0x1f0
> >       xenwatch_thread+0xc0/0x1a0
> >       ? do_wait_intr_irq+0xa0/0xa0
> >       kthread+0x16b/0x190
> >       ? set_kthread_struct+0x40/0x40
> >       ret_from_fork+0x22/0x30
> >       </TASK>
> > 
> > Fix this by calling xennet_destroy_queues() from xennet_close() too,
> > when real_num_tx_queues is still available. This ensures that queues are
> > destroyed when real_num_tx_queues is set to 0, regardless of how
> > unregister_netdev() was called.
> > 
> > Originally reported at
> > https://github.com/QubesOS/qubes-issues/issues/7257
> > 
> > Fixes: d7dac083414eb5bb9 ("net-sysfs: update the queue counts in the 
> > unregistration path")
> > Cc: stable@xxxxxxxxxxxxxxx # 5.16+
> > Signed-off-by: Marek Marczykowski-Górecki <marmarek@xxxxxxxxxxxxxxxxxxxxxx>
> > 
> > ---
> > While this fixes the issue, I'm not sure if that is the correct thing
> > to do. xennet_remove() calls xennet_destroy_queues() under rtnl_lock,
> > which may be important here? Just moving xennet_destroy_queues() before  
> 
> I checked some of the call paths leading to xennet_close(), and all of
> those contained an ASSERT_RTNL(), so it seems the rtnl_lock is already
> taken here. Could you test with adding an ASSERT_RTNL() in
> xennet_destroy_queues()?
> 
> > unregister_netdev() in xennet_remove() did not helped - it crashed in
> > another way (use-after-free in xennet_close()).  
> 
> Yes, this would need to basically do the xennet_close() handling in
> xennet_destroy() instead, which I believe is not really an option.

I think the patch makes open/close asymmetric, tho. After ifup ; ifdown;
the next ifup will fail because queues are already destroyed, no?
IOW xennet_open() expects the queues were created at an earlier stage.

Maybe we can move the destroy to ndo_uninit? (and create to ndo_init?)