Re: [PATCH v3 1/1] iommu/sva: Invalidate KVA range on kernel TLB flush

[Baolu Lu <baolu.lu@linux.intel.com>]

On 8/8/2025 10:57 AM, Tian, Kevin wrote:
>> From: Jason Gunthorpe <jgg@nvidia.com>
>> Sent: Friday, August 8, 2025 3:52 AM
>>
>> On Thu, Aug 07, 2025 at 10:40:39PM +0800, Baolu Lu wrote:
>>> +static void kernel_pte_work_func(struct work_struct *work)
>>> +{
>>> +	struct page *page, *next;
>>> +
>>> +	iommu_sva_invalidate_kva_range(0, TLB_FLUSH_ALL);
>>> +
>>> +	guard(spinlock)(&kernel_pte_work.lock);
>>> +	list_for_each_entry_safe(page, next, &kernel_pte_work.list, lru) {
>>> +		list_del_init(&page->lru);
>>
>> Please don't add new usages of lru, we are trying to get rid of this. :(
>>
>> I think the memory should be struct ptdesc, use that..
>>
> 
> btw with this change we should also defer free of the pmd page:
> 
> pud_free_pmd_page()
> 	...
> 	for (i = 0; i < PTRS_PER_PMD; i++) {
> 		if (!pmd_none(pmd_sv[i])) {
> 			pte = (pte_t *)pmd_page_vaddr(pmd_sv[i]);
> 			pte_free_kernel(&init_mm, pte);
> 		}
> 	}
> 
> 	free_page((unsigned long)pmd_sv);
> 
> Otherwise the risk still exists if the pmd page is repurposed before the
> pte work is scheduled.

You're right that freeing high-level page table pages also requires an
IOTLB flush before the pages are freed. But I question the practical
risk of the race given the extremely small time window. If this is a
real concern, a potential mitigation would be to clear the U/S bits in
all page table entries for kernel address space? But I am not confident
in making that change at this time as I am unsure of the side effects it
might cause.

> 
> another observation - pte_free_kernel is not used in remove_pagetable ()
> and __change_page_attr(). Is it straightforward to put it in those paths
> or do we need duplicate some deferring logic there?

The remove_pagetable() function is called in the path where memory is
hot-removed from the system, right? If so, there should be no issue, as
the threat model here is a page table page being freed and repurposed
while it's still cached in the IOTLB. In the hot-remove case, the memory
is removed and will not be reused, so that's fine as far as I can see.

The same to __change_page_attr(), which only changes the attributes of a
page table entry while the underlying page remains in use.

Thanks,
baolu