Re: [PATCH v3] xen: avoid crash in disable_hotplug_cpu

[Juergen Gross]

On 07/09/18 08:30, Olaf Hering wrote:
> The command 'xl vcpu-set 0 0', issued in dom0, will crash dom0:
> 
> BUG: unable to handle kernel NULL pointer dereference at 00000000000002d8
> PGD 0 P4D 0
> Oops: 0000 [#1] PREEMPT SMP NOPTI
> CPU: 7 PID: 65 Comm: xenwatch Not tainted 4.19.0-rc2-1.ga9462db-default #1 openSUSE Tumbleweed (unreleased)
> Hardware name: Intel Corporation S5520UR/S5520UR, BIOS S5500.86B.01.00.0050.050620101605 05/06/2010
> RIP: e030:device_offline+0x9/0xb0
> Code: 77 24 00 e9 ce fe ff ff 48 8b 13 e9 68 ff ff ff 48 8b 13 e9 29 ff ff ff 48 8b 13 e9 ea fe ff ff 90 66 66 66 66 90 41 54 55 53 <f6> 87 d8 02 00 00 01 0f 85 88 00 00 00 48 c7 c2 20 09 60 81 31 f6
> RSP: e02b:ffffc90040f27e80 EFLAGS: 00010203
> RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
> RDX: ffff8801f3800000 RSI: ffffc90040f27e70 RDI: 0000000000000000
> RBP: 0000000000000000 R08: ffffffff820e47b3 R09: 0000000000000000
> R10: 0000000000007ff0 R11: 0000000000000000 R12: ffffffff822e6d30
> R13: dead000000000200 R14: dead000000000100 R15: ffffffff8158b4e0
> FS:  00007ffa595158c0(0000) GS:ffff8801f39c0000(0000) knlGS:0000000000000000
> CS:  e033 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00000000000002d8 CR3: 00000001d9602000 CR4: 0000000000002660
> Call Trace:
>  handle_vcpu_hotplug_event+0xb5/0xc0
>  xenwatch_thread+0x80/0x140
>  ? wait_woken+0x80/0x80
>  kthread+0x112/0x130
>  ? kthread_create_worker_on_cpu+0x40/0x40
>  ret_from_fork+0x3a/0x50
> 
> This happens because handle_vcpu_hotplug_event is called twice. In the
> first iteration cpu_present is still true, in the second iteration
> cpu_present is false which causes get_cpu_device to return NULL.
> In case of cpu#0, cpu_online is apparently always true.
> 
> Fix this crash by checking if the cpu can be hotplugged, which is false
> for a cpu that was just removed.
> 
> Also check if the cpu was actually offlined by device_remove, otherwise
> leave the cpu_present state as it is.
> 
> Signed-off-by: Olaf Hering <olaf@xxxxxxxxx>
> ---
>  drivers/xen/cpu_hotplug.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/drivers/xen/cpu_hotplug.c b/drivers/xen/cpu_hotplug.c
> index d4265c8ebb22..68f9f663da08 100644
> --- a/drivers/xen/cpu_hotplug.c
> +++ b/drivers/xen/cpu_hotplug.c
> @@ -19,11 +19,15 @@ static void enable_hotplug_cpu(int cpu)
>  
>  static void disable_hotplug_cpu(int cpu)
>  {
> +	if (!cpu_is_hotpluggable(cpu))
> +		return;
>  	if (cpu_online(cpu)) {
>  		lock_device_hotplug();
>  		device_offline(get_cpu_device(cpu));
>  		unlock_device_hotplug();
>  	}
> +	if (cpu_online(cpu))
> +		return;
>  	if (cpu_present(cpu))
>  		xen_arch_unregister_cpu(cpu);

Could you merge the two if conditions?

if (!cpu_online(cpu) && cpu_present(cpu))
	xen_arch_unregister_cpu(cpu);

And while not really important, as we are called in the xenstore watch
thread only, it might be a good idea to move the first cpu_online()
test into the lock, i.e.:

lock_device_hotplug();
if (cpu_online(cpu))
	device_offline(get_cpu_device(cpu));
unlock_device_hotplug();

This will make the code robust against reentry.


Juergen