> It looks like memory_failure_queue() is only invoked when source user > page has a poison, not if the poison > > is in source kernel page. Any idea why? If it is a user page, Linux can unmap it from that process. Send signal(SIGBUS) to the process if it was actively consuming the poison (as opposed to poison found by patrol scrubber or by speculative access). Linux doesn't (yet) have a way to workaround poison in a source kernel page. I think the only place it will ever get that is for pages in the page cache. Action to somehow unmap them from the file? -Tony
