> On Apr 30, 2020, at 5:25 PM, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > It wasn't clear how "copy_to_mc()" could ever fault. Poisoning > after-the-fact? Why would that be preferable to just mapping a dummy > page? If the kernel gets an async memory error and maps a dummy page, then subsequent reads will subsequently succeed and return garbage when they should fail. If x86 had write-only pages, we could map a dummy write-only page. But we don’t, so I think we’re stuck. As for naming the kind of memory we’re taking about, ISTM there are two classes: DAX and monstrous memory-mapped non-persistent cache devices. Both could be Optane, I suppose. But I also think it’s legitimate to use these accessors to increase the chance of surviving a failure of normal memory. If a normal page happens to be page cache when it fails and if page cache access use these fancy accessors, then we might actually survive a failure. We could be ambitious: declare that all page cache and all get_user_page’d memory should use the new accessors. I doubt we’ll ever really succeed due to magical things like rseq and anything that thinks that users can set up their own memory as a kernel-accessed ring buffer, but I suppose we could try.
