Re: [PATCH 09/12] KVM: MMU: introduce pte-list lockless walker

[Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>]

On 08/29/2013 05:31 PM, Gleb Natapov wrote:
> On Thu, Aug 29, 2013 at 02:50:51PM +0800, Xiao Guangrong wrote:
>> After more thinking, I still think rcu_assign_pointer() is unneeded when a entry
>> is removed. The remove-API does not care the order between unlink the entry and
>> the changes to its fields. It is the caller's responsibility:
>> - in the case of rcuhlist, the caller uses call_rcu()/synchronize_rcu(), etc to
>>   enforce all lookups exit and the later change on that entry is invisible to the
>>   lookups.
>>
>> - In the case of rculist_nulls, it seems refcounter is used to guarantee the order
>>   (see the example from Documentation/RCU/rculist_nulls.txt).
>>
>> - In our case, we allow the lookup to see the deleted desc even if it is in slab cache
>>   or its is initialized or it is re-added.
>>
> BTW is it a good idea? We can access deleted desc while it is allocated
> and initialized to zero by kmem_cache_zalloc(), are we sure we cannot
> see partially initialized desc->sptes[] entry? On related note what about
> 32 bit systems, they do not have atomic access to desc->sptes[].

Good eyes. This is a bug here.

It seems we do not have a good to fix this. How disable this optimization on
32 bit host, small changes:

 static inline void kvm_mmu_rcu_free_page_begin(struct kvm *kvm)
 {
+#ifdef CONFIG_X86_64
        rcu_read_lock();

        kvm->arch.rcu_free_shadow_page = true;
        /* Set the indicator before access shadow page. */
        smp_mb();
+#else
+       spin_lock(kvm->mmu_lock);
+#endif
 }

 static inline void kvm_mmu_rcu_free_page_end(struct kvm *kvm)
 {
+#ifdef CONFIG_X86_64
        /* Make sure that access shadow page has finished. */
        smp_mb();
        kvm->arch.rcu_free_shadow_page = false;

        rcu_read_unlock();
+#else
+       spin_unlock(kvm->mmu_lock);
+#endif
 }