Re: [RFC PATCH 11/20] x86/virt/seamldr: Abort updates if errors occurred midway

[Chao Gao]

>>   static void ack_state(void)
>>   {
>> -	if (atomic_dec_and_test(&tdp_data.thread_ack))
>> -		set_state(tdp_data.state + 1);
>> +	if (atomic_dec_and_test(&tdp_data.thread_ack)) {
>> +		/*
>> +		 * If an error occurred, abort the update by skipping to
>> +		 * the final state
>> +		 */
>> +		if (atomic_read(&tdp_data.failed))
>> +			set_state(TDP_DONE);
>> +		else
>> +			set_state(tdp_data.state + 1);
>> +	}
>>   }
>>   /*
>> @@ -285,6 +294,9 @@ static int do_seamldr_install_module(void *params)
>>   			default:
>>   				break;
>>   			}
>> +
>> +			if (ret)
>> +				atomic_inc(&tdp_data.failed);
>
>Should there be some explicit ordering requirement between setting an error
>and reading it in ack_state by a different CPU?

Only the last CPU that calls ack_state() will change the global state, either
advancing to the next state or setting it to TDP_DONE on error. so, we only
need to ensure that the last CPU can see the error. This is guaranteed because
the error is set before the call to ack_state().

+			if (ret)
+				atomic_inc(&tdp_data.failed);
			ack_state();