On 06/30/2016 07:55 PM, Linus Torvalds wrote: > On Thu, Jun 30, 2016 at 5:12 PM, Dave Hansen <dave@xxxxxxxx> wrote: >> From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx> >> The Intel(R) Xeon Phi(TM) Processor x200 Family (codename: Knights >> Landing) has an erratum where a processor thread setting the Accessed >> or Dirty bits may not do so atomically against its checks for the >> Present bit. This may cause a thread (which is about to page fault) >> to set A and/or D, even though the Present bit had already been >> atomically cleared. > > So I don't think your approach is wrong, but I suspect this is > overkill, and what we should instead just do is to not use the A/D > bits at all in the swap representation. We actually don't even use Dirty today. It's (implicitly) used to determine pte_none(), but it ends up being masked out for the swp_offset/type() calculations entirely, much to my surprise. I think what you suggest will work if we don't consider A/D in pte_none(). I think there are a bunch of code path where assume that !pte_present() && !pte_none() means swap. > The swap-entry representation was a bit tight on 32-bit page table > entries, but in 64-bit ones, I think we have tons of bits, don't we? > So we could decide just to not use those two bits on x86. Yeah, we've definitely got space. I'll go poke around and make sure that this works everywhere. I agree that throwing 32-bit non-PAE under the bus is definitely worth it here. Nobody will care about that in a million years. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>