Re: [PATCH v8 01/17] KVM: s390: pv: leak the topmost page table when destroy fails

[Janis Schoetterl-Glausch]

On 3/3/22 16:05, Claudio Imbrenda wrote:
> On Thu, 3 Mar 2022 15:40:42 +0100
> Janis Schoetterl-Glausch <scgl@xxxxxxxxxxxxx> wrote:
> 
>> On 3/2/22 19:11, Claudio Imbrenda wrote:
>>> Each secure guest must have a unique ASCE (address space control
>>> element); we must avoid that new guests use the same page for their
>>> ASCE, to avoid errors.
>>>
>>> Since the ASCE mostly consists of the address of the topmost page table
>>> (plus some flags), we must not return that memory to the pool unless
>>> the ASCE is no longer in use.
>>>
>>> Only a successful Destroy Secure Configuration UVC will make the ASCE
>>> reusable again.
>>>
>>> If the Destroy Configuration UVC fails, the ASCE cannot be reused for a
>>> secure guest (either for the ASCE or for other memory areas). To avoid
>>> a collision, it must not be used again. This is a permanent error and
>>> the page becomes in practice unusable, so we set it aside and leak it.
>>> On failure we already leak other memory that belongs to the ultravisor
>>> (i.e. the variable and base storage for a guest) and not leaking the
>>> topmost page table was an oversight.
>>>
>>> This error (and thus the leakage) should not happen unless the hardware
>>> is broken or KVM has some unknown serious bug.
>>>
>>> Signed-off-by: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx>
>>> Fixes: 29b40f105ec8d55 ("KVM: s390: protvirt: Add initial vm and cpu lifecycle handling")
>>> ---
>>>  arch/s390/include/asm/gmap.h |  2 +
>>>  arch/s390/kvm/pv.c           |  9 +++--
>>>  arch/s390/mm/gmap.c          | 71 ++++++++++++++++++++++++++++++++++++
>>>  3 files changed, 79 insertions(+), 3 deletions(-)
>>>   
>> [...]
>>
>>> +/**
>>> + * s390_replace_asce - Try to replace the current ASCE of a gmap with
>>> + * another equivalent one.
>>> + * @gmap the gmap
>>> + *
>>> + * If the allocation of the new top level page table fails, the ASCE is not
>>> + * replaced.
>>> + * In any case, the old ASCE is always removed from the list. Therefore the
>>> + * caller has to make sure to save a pointer to it beforehands, unless an
>>> + * intentional leak is intended.
>>> + */
>>> +int s390_replace_asce(struct gmap *gmap)
>>> +{
>>> +	unsigned long asce;
>>> +	struct page *page;
>>> +	void *table;
>>> +
>>> +	s390_remove_old_asce(gmap);
>>> +
>>> +	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
>>> +	if (!page)
>>> +		return -ENOMEM;
>>> +	table = page_to_virt(page);
>>> +	memcpy(table, gmap->table, 1UL << (CRST_ALLOC_ORDER + PAGE_SHIFT));  
>>
>> Is concurrent modification of *gmap->table possible during the copy?
> 
> that would only be possible if the guest touches memory in such way
> that the table needs to be changed.
> 
> this function is only called when the guest is not running (e.g. during
> reboot), so nobody should touch the table

Is that asserted?
I guess if modifications to the table are block concurrent and entries are only
freed when the vm is destroyed you cannot intentionally do any funny business.
> 
>>
>>> +
>>> +	/*
>>> +	 * The caller has to deal with the old ASCE, but here we make sure
>>> +	 * the new one is properly added to the list of page tables, so that
>>> +	 * it will be freed when the VM is torn down.
>>> +	 */
>>> +	spin_lock(&gmap->guest_table_lock);
>>> +	list_add(&page->lru, &gmap->crst_list);
>>> +	spin_unlock(&gmap->guest_table_lock);
>>> +
>>> +	/* Set new table origin while preserving existing ASCE control bits */
>>> +	asce = (gmap->asce & _ASCE_ORIGIN) | __pa(table);
>>> +	WRITE_ONCE(gmap->asce, asce);
>>> +	WRITE_ONCE(gmap->mm->context.gmap_asce, asce);
>>> +	WRITE_ONCE(gmap->table, table);
>>> +
>>> +	return 0;
>>> +}
>>> +EXPORT_SYMBOL_GPL(s390_replace_asce);  
>>
>