On Fri, Apr 02, 2021 at 03:11:20PM +0000, Luck, Tony wrote: > >> Combined with my "mutex" patch (to get rid of races where 2nd process returns > >> early, but first process is still looking for mappings to unmap and tasks > >> to signal) this patch moves forward a bit. But I think it needs an > >> additional change here in kill_me_maybe() to just "return" if there is a > >> EHWPOISON return from memory_failure() > >> > > Got this, Thanks for your reply! > > I will dig into this! > > One problem with this approach is when the first task to find poison > fails to complete actions. Then the poison pages are not unmapped, > and just returning from kill_me_maybe() gets into a loop :-( Yes, that's the pain point. We need send SIGBUS to the current process in "already haredware poisoned" case of memory_failure(). SIGBUS should contain the error virtual address, but unfortunately walking the page table or using p->mce_vaddr is not always reliable now. So as a second-best approach, we can extend the "walking page table" approach such that we walk over the whole virtual address space to make sure that the number of entries pointing to the error page is exactly 1. If that's the case, then we can confidently send SIGBUS with it. If we find multiple entries pointing to the error page, then we give up guessing, then send a nomral SIGBUS to the current process. That's not worse than now, and I think we need wait in the hope that the virtual address will be available in MCE handler. Anyway I'll try to write a patch for this. Thanks, Naoya Horiguchi
