On Mon, Apr 22, 2019 at 07:15:32PM +0200, Borislav Petkov wrote: > On Mon, Apr 22, 2019 at 03:59:16PM +0000, Luck, Tony wrote: > > > Err, this all sounds to me like the storm detection code should > > > *automatically* disable the CEC in such cases, I'd say. > > > > Sounds good. But we should distinguish storms that have many different > > addresses from storms that just ping a few addresses. CEC will see counts > > hit the threshold in the latter case, but it might not be able to take the > > pages > > offline (because they are locked, or in-use by kernel). > > > > So I think the change might be to the return value from NOTIFY_STOP to > > NOTIFY_DONE > > ... but only if we are in the middle of a storm AND the CEC array is full. > > Well, regardless of this specific use case, isn't that a generic enough > action that we should do always? I mean, the aspect of falling back to > logging to external agent.
Yes. Automating this would be a very good idea. > However, currently we don't signal that the CEC is full - we simply > remove the LRU element in cec_add_elem() before we insert the new one. > > We can either return a specific retval to say, CEC is full and we had to > delete an elem or we can add a cec_is_full() accessor... A lot depends on why the CEC is full, and which entry is being deleted to make room. In the case of many errors at different addresses we are deleting the entry with the lowest count. But all of the entries have low counts because we are just thrashing the array with many different addresses. In this situation a warning would be helpful. But in the case where the system has been up for months and we very slowly accumlated logs of bit flips. The periodic spring cleaning means they all have generation "00", but we never actually drop an old entry because of age. In this case dropping one entry to make space for a new one is fine and doesn't need any action. Perhaps we can distinguish the cases by the generation? If we are dropping an entry that was recently added, then it will still have generation "11" (or at least not "00"). Use that to trigger an action? -Tony
