Re: [PATCH 2/5] hv: add helpers to handle hv_util device state

[Greg KH <gregkh@linuxfoundation.org>]

On Tue, Sep 15, 2015 at 05:37:51PM -0700, K. Y. Srinivasan wrote:
> From: Olaf Hering <olaf@aepfle.de>
> 
> The callbacks in kvp, vss and fcopy code are called both from the main thread
> as well as from interrupt context. If a state change is done by the main
> thread it is not immediately seen by the interrupt. As a result the
> state machine gets out of sync.
> 
> Force propagation of state changes via get/set helpers with a memory barrier.
> 
> Signed-off-by: Olaf Hering <olaf@aepfle.de>
> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
> ---
>  drivers/hv/hyperv_vmbus.h |   14 ++++++++++++++
>  1 files changed, 14 insertions(+), 0 deletions(-)
> 
> diff --git a/drivers/hv/hyperv_vmbus.h b/drivers/hv/hyperv_vmbus.h
> index 4b1eb6d..dee5798 100644
> --- a/drivers/hv/hyperv_vmbus.h
> +++ b/drivers/hv/hyperv_vmbus.h
> @@ -780,4 +780,18 @@ enum hvutil_device_state {
>  	HVUTIL_DEVICE_DYING,     /* driver unload is in progress */
>  };
>  
> +static inline void hvutil_device_set_state(enum hvutil_device_state *p,
> +					   enum hvutil_device_state s)
> +{
> +	*p = s;
> +	wmb();
> +}
> +
> +static inline enum hvutil_device_state
> +hvutil_device_get_state(enum hvutil_device_state *p)
> +{
> +	rmb();
> +	return *p;
> +}
> +
>  #endif /* _HYPERV_VMBUS_H */

This is crazy.  If you need to know the state of something (pun
intended) then you had better be using a lock, and not relying on a
random pointer to contain a random value and be able to do something
based on that.

This shows the code is broken, don't paper over things by throwing in
random read/write barriers, that is a HUGE flag that something bad is
happening here.

Just use a lock, that's what it is there for.

greg k-h