[PATCH v3 10/10] arm64: ptdump: Add support for guest stage-2 pagetables dumping

[Sebastian Ene]

On Wed, Nov 22, 2023 at 11:35:57PM +0000, Oliver Upton wrote:
> On Wed, Nov 15, 2023 at 05:16:40PM +0000, Sebastian Ene wrote:
> > +struct ptdump_registered_guest {
> > +	struct list_head		reg_list;
> > +	struct ptdump_info		info;
> > +	struct kvm_pgtable_snapshot	snapshot;
> > +	rwlock_t			*lock;
> > +};
> 
> Why can't we just store a pointer directly to struct kvm in ::private?

I don't think it will work unless we expect a struct kvm_pgtable
in stage2_ptdump_walk file_priv field. I think it is a good ideea and will
simplify things a little bit dropping kvm_pgtable_snapshot from here.

The current code has some fileds that are reduntant (the priv pointers)
because I also made this to work with protected guests where we can't
access their pagetables directly.

> Also, shouldn't you take a reference on struct kvm when the file is
> opened to protect against VM teardown?
>

I am not sure about the need could you please elaborate a bit ? On VM
teardown we expect ptdump_unregister_guest_stage2 to be invoked.

> > +static LIST_HEAD(ptdump_guest_list);
> > +static DEFINE_MUTEX(ptdump_list_lock);
> 
> What is the list for?
> 

I am keeping a list of registered guests with ptdump and the lock should
protect the list against concurent VM teardowns.

> >  static phys_addr_t ptdump_host_pa(void *addr)
> >  {
> >  	return __pa(addr);
> > @@ -757,6 +768,63 @@ static void stage2_ptdump_walk(struct seq_file *s, struct ptdump_info *info)
> >  	kvm_pgtable_walk(pgtable, start_ipa, end_ipa, &walker);
> >  }
> >  
> > +static void guest_stage2_ptdump_walk(struct seq_file *s,
> > +				     struct ptdump_info *info)
> > +{
> > +	struct ptdump_info_file_priv *f_priv =
> > +		container_of(info, struct ptdump_info_file_priv, info);
> > +	struct ptdump_registered_guest *guest = info->priv;
> > +
> > +	f_priv->file_priv = &guest->snapshot;
> > +
> > +	read_lock(guest->lock);
> > +	stage2_ptdump_walk(s, info);
> > +	read_unlock(guest->lock);
> 
> Taking the mmu lock for read allows other table walkers to add new
> mappings and adjust the granularity of existing ones. Should this
> instead take the mmu lock for write?
>

Thanks for pointing our, this is exactly what I was trying to avoid,
so yes I should use the write mmu lock in this case.

> > +}
> > +
> > +int ptdump_register_guest_stage2(struct kvm *kvm)
> > +{
> > +	struct ptdump_registered_guest *guest;
> > +	struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
> > +	struct kvm_pgtable *pgt = mmu->pgt;
> > +
> > +	guest = kzalloc(sizeof(struct ptdump_registered_guest), GFP_KERNEL);
> 
> You want GFP_KERNEL_ACCOUNT here.
> 

Right, thanks this is because it is an untrusted allocation triggered from
userspace.

> -- 
> Thanks,
> Oliver

Thank you,
Seb