Re: [PATCH AUTOSEL 5.11 51/67] Drivers: hv: vmbus: Resolve race condition in vmbus_onoffer_rescind()

[Andrea Parri]

On Wed, Feb 24, 2021 at 07:50:09AM -0500, Sasha Levin wrote:
> From: "Andrea Parri (Microsoft)" <parri.andrea@xxxxxxxxx>
> 
> [ Upstream commit e4d221b42354b2e2ddb9187a806afb651eee2cda ]
> 
> An erroneous or malicious host could send multiple rescind messages for
> a same channel.  In vmbus_onoffer_rescind(), the guest maps the channel
> ID to obtain a pointer to the channel object and it eventually releases
> such object and associated data.  The host could time rescind messages
> and lead to an use-after-free.  Add a new flag to the channel structure
> to make sure that only one instance of vmbus_onoffer_rescind() can get
> the reference to the channel object.
> 
> Reported-by: Juan Vazquez <juvazq@xxxxxxxxxxxxx>
> Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@xxxxxxxxx>
> Reviewed-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>
> Link: https://lore.kernel.org/r/20201209070827.29335-6-parri.andrea@xxxxxxxxx
> Signed-off-by: Wei Liu <wei.liu@xxxxxxxxxx>
> Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>

Sasha - This patch is one of a group of patches where a Linux guest running on
Hyper-V will start assuming that hypervisor behavior might be malicious, and
guards against such behavior.  Because this is a new assumption, these patches
are more properly treated as new functionality rather than as bug fixes.  So I
would propose that we *not* bring such patches back to stable branches.

Thanks,
  Andrea


> ---
>  drivers/hv/channel_mgmt.c | 12 ++++++++++++
>  include/linux/hyperv.h    |  1 +
>  2 files changed, 13 insertions(+)
> 
> diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
> index 1d44bb635bb84..a9f58840f85dc 100644
> --- a/drivers/hv/channel_mgmt.c
> +++ b/drivers/hv/channel_mgmt.c
> @@ -1049,6 +1049,18 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
>  
>  	mutex_lock(&vmbus_connection.channel_mutex);
>  	channel = relid2channel(rescind->child_relid);
> +	if (channel != NULL) {
> +		/*
> +		 * Guarantee that no other instance of vmbus_onoffer_rescind()
> +		 * has got a reference to the channel object.  Synchronize on
> +		 * &vmbus_connection.channel_mutex.
> +		 */
> +		if (channel->rescind_ref) {
> +			mutex_unlock(&vmbus_connection.channel_mutex);
> +			return;
> +		}
> +		channel->rescind_ref = true;
> +	}
>  	mutex_unlock(&vmbus_connection.channel_mutex);
>  
>  	if (channel == NULL) {
> diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
> index 5ddb479c4d4cb..ef3573e99d989 100644
> --- a/include/linux/hyperv.h
> +++ b/include/linux/hyperv.h
> @@ -803,6 +803,7 @@ struct vmbus_channel {
>  	u8 monitor_bit;
>  
>  	bool rescind; /* got rescind msg */
> +	bool rescind_ref; /* got rescind msg, got channel reference */
>  	struct completion rescind_event;
>  
>  	u32 ringbuffer_gpadlhandle;
> -- 
> 2.27.0
>