Re: [PATCH] iommu/intel: Fix __domain_mapping()'s usage of switch_to_super_page()

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

On 8/26/25 23:48, David Woodhouse wrote:
> On Tue, 2025-08-26 at 14:38 +0000, Eugene Koira wrote:
>> switch_to_super_page() assumes the memory range it's working on is aligned
>> to the target large page level. Unfortunately, __domain_mapping() doesn't
>> take this into account when using it, and will pass unaligned ranges
>> ultimately freeing a PTE range larger than expected.
>>
>> Take for example a mapping with the following iov_pfn range [0x3fe400,
>> 0x4c0600], which should be backed by the following mappings:
> 
> The range is [0x3fe400, 0x4c0600) ?
> 
>>     iov_pfn [0x3fe400, 0x3fffff] covered by 2MiB pages
>>     iov_pfn [0x400000, 0x4bffff] covered by 1GiB pages
>>     iov_pfn [0x4c0000, 0x4c05ff] covered by 2MiB pages
>>
>> Under this circumstance, __domain_mapping() will pass [0x400000, 0x4c05ff]
>> to switch_to_super_page() at a 1 GiB granularity, which will in turn
>> free PTEs all the way to iov_pfn 0x4fffff.
>>
>> Mitigate this by rounding down the iov_pfn range passed to
>> switch_to_super_page() in __domain_mapping()
>> to the target large page level.
>>
>> Additionally add range alignment checks to switch_to_super_page.
>>
>> Fixes: 9906b9352a35 ("iommu/vt-d: Avoid duplicate removing in __domain_mapping()")
>> Signed-off-by: Eugene Koira <eugkoira@amazon.com>
>> Cc: stable@vger.kernel.org
>> ---
>>   drivers/iommu/intel/iommu.c | 7 ++++++-
>>   1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
>> index 9c3ab9d9f69a..dff2d895b8ab 100644
>> --- a/drivers/iommu/intel/iommu.c
>> +++ b/drivers/iommu/intel/iommu.c
>> @@ -1575,6 +1575,10 @@ static void switch_to_super_page(struct dmar_domain *domain,
>>   	unsigned long lvl_pages = lvl_to_nr_pages(level);
>>   	struct dma_pte *pte = NULL;
>>   
>> +	if (WARN_ON(!IS_ALIGNED(start_pfn, lvl_pages) ||
>> +		    !IS_ALIGNED(end_pfn + 1, lvl_pages)))
>> +		return;
>> +
>>   	while (start_pfn <= end_pfn) {
>>   		if (!pte)
>>   			pte = pfn_to_dma_pte(domain, start_pfn, &level,
>> @@ -1650,7 +1654,8 @@ __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
>>   				unsigned long pages_to_remove;
>>   
>>   				pteval |= DMA_PTE_LARGE_PAGE;
>> -				pages_to_remove = min_t(unsigned long, nr_pages,
>> +				pages_to_remove = min_t(unsigned long,
>> +							round_down(nr_pages, lvl_pages),
>>   							nr_pte_to_next_page(pte) * lvl_pages);
>>   				end_pfn = iov_pfn + pages_to_remove - 1;
>>   				switch_to_super_page(domain, iov_pfn, end_pfn, largepage_lvl);
> 
> I'm mildly entertained by the fact that the *only* comment in this
> block of code is a lie. Would you care to address that while you're
> here? Maybe the comment could look something like...
> 
> 			/* If the new mapping is eligible for a large page, then
> 			 * remove all smaller pages that the existing pte at this
> 			 * level references.
> 			 * XX: do we even need to bother calling switch_to_super_page()
> 			 * if this PTE wasn't *present* before?
> 			 */
> 
> I bet it would benefit from one or two other one-line comments to make
> it clearer what's going on, too...
> 
> But in general, I think this looks sane even though this code makes my
> brain hurt. Could do with a test case, in an ideal world. Maybe we can
> work on that as part of the generic pagetable support which is coming?

Agreed. The generic pagetable work will replace this code, so this will
be removed. Therefore, we need a fix patch that can be backported before
the generic pagetable for vt-d lands.

Thanks,
baolu