[Intel-gfx] [PATCH 1/4] iommu/vt-d: Disable superpage for Geminilake igfx

[Lu Baolu]

On 7/14/21 4:30 AM, Ville Syrjälä wrote:
> On Tue, Jul 13, 2021 at 09:34:09AM +0800, Lu Baolu wrote:
>> On 7/12/21 11:47 PM, Ville Syrjälä wrote:
>>> On Mon, Jul 12, 2021 at 07:23:07AM +0800, Lu Baolu wrote:
>>>> On 7/10/21 12:47 AM, Ville Syrjala wrote:
>>>>> From: Ville Syrjälä<ville.syrjala at linux.intel.com>
>>>>>
>>>>> While running "gem_exec_big --r single" from igt-gpu-tools on
>>>>> Geminilake as soon as a 2M mapping is made I tend to get a DMAR
>>>>> write fault. Strangely the faulting address is always a 4K page
>>>>> and usually very far away from the 2M page that got mapped.
>>>>> But if no 2M mappings get used I can't reproduce the fault.
>>>>>
>>>>> I also tried to dump the PTE for the faulting address but it actually
>>>>> looks correct to me (ie. definitely seems to have the write bit set):
>>>>>     DMAR: DRHD: handling fault status reg 2
>>>>>     DMAR: [DMA Write] Request device [00:02.0] PASID ffffffff fault addr 7fa8a78000 [fault reason 05] PTE Write access is not set
>>>>>     DMAR: fault 7fa8a78000 (level=1) PTE = 149efc003
>>>>>
>>>>> So not really sure what's going on and this might just be full on duct
>>>>> tape, but it seems to work here. The machine has now survived a whole day
>>>>> running that test whereas with superpage enabled it fails in less than
>>>>> a minute usually.
>>>>>
>>>>> TODO: might be nice to disable superpage only for the igfx iommu
>>>>>          instead of both iommus
>>>> If all these quirks are about igfx dedicated iommu's, I would suggest to
>>>> disable superpage only for the igfx ones.
>>> Sure. Unfortunately there's no convenient mechanism to do that in
>>> the iommu driver that I can immediately see. So not something I
>>> can just whip up easily. Since you're actually familiar with the
>>> driver maybe you can come up with a decent solution for that?
>>>
>> How about something like below? [no compile, no test...]
>>
>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
>> index 1131b8efb050..2d51ef288a9e 100644
>> --- a/drivers/iommu/intel/iommu.c
>> +++ b/drivers/iommu/intel/iommu.c
>> @@ -338,6 +338,7 @@ static int intel_iommu_strict;
>>    static int intel_iommu_superpage = 1;
>>    static int iommu_identity_mapping;
>>    static int iommu_skip_te_disable;
>> +static int iommu_skip_igfx_superpage;
>>
>>    #define IDENTMAP_GFX		2
>>    #define IDENTMAP_AZALIA		4
>> @@ -652,6 +653,27 @@ static bool domain_update_iommu_snooping(struct
>> intel_iommu *skip)
>>    	return ret;
>>    }
>>
>> +static bool domain_use_super_page(struct dmar_domain *domain)
>> +{
>> +	struct dmar_drhd_unit *drhd;
>> +	struct intel_iommu *iommu;
>> +	bool ret = true;
>> +
>> +	if (!intel_iommu_superpage)
>> +		return false;
>> +
>> +	rcu_read_lock();
>> +	for_each_active_iommu(iommu, drhd) {
>> +		if (drhd->gfx_dedicated && iommu_skip_igfx_superpage) {
>> +			ret = false;
>> +			break
>                               ^
> Missing semicolon. Othwerwise seems to work great here. Thanks.
> 
> Are you going to turn this into a proper patch, or do you
> want me to just squash this into my patches and repost?
> 

Please go ahead with a new version.

Best regards,
baolu