On Tue, Apr 26, 2022 at 6:30 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > On Tue, Apr 26, 2022, Mingwei Zhang wrote: > > On Tue, Apr 26, 2022 at 6:16 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > > > > On Tue, Apr 26, 2022, Mingwei Zhang wrote: > > > > > I completely agree that lookup_address() and friends are unnecessarily fragile, > > > > > but I think that attempting to harden them to fix this KVM bug will open a can > > > > > of worms and end up delaying getting KVM fixed. > > > > > > > > So basically, we need to: > > > > - choose perf_get_page_size() instead of using any of the > > > > lookup_address*() in mm. > > > > - add a wrapper layer to adapt: 1) irq disabling/enabling and 2) size > > > > -> level translation. > > > > > > > > Agree? > > > > > > Drat, I didn't see that it returns the page size, not the level. That's a bit > > > unfortunate. It definitely makes me less averse to fixing lookup_address_in_pgd() > > > > > > Hrm. I guess since we know there's at least one broken user, and in theory > > > fixing lookup_address_in_pgd() should do no harm to users that don't need protection, > > > it makes sense to just fix lookup_address_in_pgd() and see if the x86 maintainers > > > push back. > > > > Yeah, fixing lookup_address_in_pgd() should be cleaner(), since the > > page fault usage case does not need irq save/restore. But the other > > one needs it. So, we can easily fix the function with READ_ONCE and > > lockless staff. But wrapping the function with irq save/restore from > > the KVM side. > > I think it makes sense to do the save/restore in lookup_address_in_pgd(). The > Those helpers are exported, so odds are good there are broken users that will > benefit from fixing all paths. no, lookup_address_in_pgd() is probably just broken for KVM. In other call sites, some may already disable IRQ, so doing that again inside lookup_address_in_pgd() will be bad. I am looking at here: https://elixir.bootlin.com/linux/latest/source/arch/arm/kernel/traps.c#L304 so, the save/restore are done in oops_begin() and oops_end(), which is wrapping show_fault_oops() that calls lookup_address_in_pgd(). So, I think we need to ensure the READ_ONCE. hmm, regarding the lockless macros, Paolo is right, for x86 it makes no difference. s390 seems to have a different implementation, but kvm_mmu_max_mapping_level() as well as host_pfn_mapping_level are both functions in x86 mmu. Thanks. -Mingwei
