On Mon, Jan 27, 2025 at 7:01 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > On Mon, Jan 27, 2025, Paolo Bonzini wrote: > > On Sat, Jan 25, 2025 at 1:44 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > I like the special casing, it makes the oddballs stand out, which in turn (hopefully) > > > makes developers pause and take note. I.e. the SRCU walkers are all normal readers, > > > the set_nx_huge_pages() "never" path is a write in disguise, and > > > kvm_hyperv_tsc_notifier() is a very special snowflake. > > > > set_nx_huge_pages() is not a writer in disguise. Rather, it's > > a *real* writer for nx_hugepage_mitigation_hard_disabled which is > > also protected by kvm_lock; > > Heh, agreed, I was trying to say that it's a write that is disguised as a reader. > > > and there's a (mostly theoretical) bug in set_nx_huge_pages_recovery_param() > > which reads it without taking the lock. > > It's arguably not a bug. Userspace has no visibility into the order in which > param writes are processed. If there are racing writes to the period/ratio and > "never", both outcomes are legal (rejected with -EPERM or period/ratio changes). > If nx_hugepage_mitigation_hard_disabled becomes set after the params are changed, > then vm_list is guaranteed to be empty, so the wakeup walk is still a nop. I think we have enough arguably necessary lockless cases to think about linearizability of sysfs writes and reads of vm_list. :) If you agree, set_nx_huge_pages() and set_nx_huge_pages_recovery_param() can be changed to simply have a guard(mutex)(&kvm_lock) around them, instead of protecting just the vm_list walk. Paolo
