On Tue, Oct 25, 2022 at 5:06 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > On Tue, Oct 25, 2022 at 04:18:20PM +0200, Jann Horn wrote: > > On Tue, Oct 25, 2022 at 4:02 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > > On Mon, Oct 24, 2022 at 09:58:07PM +0200, Jann Horn wrote: > > > > > > > Unless I'm completely misunderstanding what's going on here, the whole > > > > "remove_table" thing only happens when you "remove a table", meaning > > > > you free an entire *pagetable*. Just zapping PTEs doesn't trigger that > > > > logic. > > > > > > Aah; yes true. OTOH even it that were not so, I think it would still be > > > broken because the current code relies on the TLB flush to have > > > completed, whereas the RCU scheme is effectively async and can be > > > considered pending until the callback runs. > > > > > > Hurmph... easiest fix is probably to dis-allow kvm_flush_tlb_multi() > > > for i386-pae builds. > > > > > > Something like so... nobody in his right mind should care about i386-pae > > > virt performance much. > > > > I think Xen and HyperV have similar codepaths. > > hyperv_flush_tlb_multi() looks like it uses remote flush hypercalls, > > xen_flush_tlb_multi() too. > > Sure (not updated). > > > On top of that, I think that theoretically, Linux doesn't even ensure > > that you have a TLB flush in between tearing down one PTE and > > installing another PTE (see > > https://lore.kernel.org/all/CAG48ez1Oz4tT-N2Y=Zs6jumu=zOp7SQRZ=V2c+b5bT9P4retJA@xxxxxxxxxxxxxx/), > > but I haven't tested that, and if it is true, I'm also not entirely > > sure if it's correct (in the sense that it only creates incoherent-TLB > > states when userspace is doing something stupid like racing > > MADV_DONTNEED and page faults on the same region). > > > > I think the more clearly correct fix would be to get rid of the split > > loads and use CMPXCHG16B instead (probably destroying the performance > > of GUP-fast completely), but that's complicated because some of the > > architectures that use the split loads path don't have cmpxchg_double > > (or at least don't have it wired up). > > cmpxchg8b; but no, I think we want to fix MADV_DONTNEED, incoherent TLB > states are a pain nobody needs. > > Something like so should force TLB flushes before dropping pte_lock (not > looked at the various pmd level things yet). [...] > #endif /* _LINUX_MM_H */ > diff --git a/mm/memory.c b/mm/memory.c > index f88c351aecd4..9bb63b3fbee1 100644 > --- a/mm/memory.c > +++ b/mm/memory.c > @@ -1440,6 +1440,11 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb, > tlb_remove_tlb_entry(tlb, pte, addr); > zap_install_uffd_wp_if_needed(vma, addr, pte, details, > ptent); > + > + if (!force_flush && !tlb->fullmm && details && > + details->zap_flags & ZAP_FLAG_FORCE_FLUSH) > + force_flush = 1; > + Hmm... I guess that might work, assuming that there is no other codepath we might race with that first turns the present PTE into a non-present PTE but keeps the flush queued for later. At least codepaths that use the tlb_batched infrastructure are unproblematic...
