Re: [PATCH v2] iommu/amd: Disable IOMMU on Stoney Ridge systems

[Kai-Heng Feng]

> On Dec 20, 2019, at 03:15, Deucher, Alexander <Alexander.Deucher@xxxxxxx> wrote:
> 
>> -----Original Message-----
>> From: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
>> Sent: Wednesday, December 18, 2019 12:45 PM
>> To: Joerg Roedel <joro@xxxxxxxxxx>
>> Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>; Deucher, Alexander
>> <Alexander.Deucher@xxxxxxx>; iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx; Kernel
>> development list <linux-kernel@xxxxxxxxxxxxxxx>
>> Subject: Re: [PATCH v2] iommu/amd: Disable IOMMU on Stoney Ridge
>> systems
>> 
>> 
>> 
>>> On Dec 17, 2019, at 17:53, Joerg Roedel <joro@xxxxxxxxxx> wrote:
>>> 
>>> On Fri, Dec 06, 2019 at 01:57:41PM +0800, Kai-Heng Feng wrote:
>>>> Hi Joerg,
>>>> 
>>>>> On Dec 3, 2019, at 01:00, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
>>>>> 
>>>>> On Fri, Nov 29, 2019 at 10:21:54PM +0800, Kai-Heng Feng wrote:
>>>>>> Serious screen flickering when Stoney Ridge outputs to a 4K monitor.
>>>>>> 
>>>>>> According to Alex Deucher, IOMMU isn't enabled on Windows, so let's
>>>>>> do the same here to avoid screen flickering on 4K monitor.
>>>>> 
>>>>> Disabling the IOMMU entirely seem pretty severe.  Isn't it enough to
>>>>> identity map the GPU device?
>>>> 
>>>> Ok, there's set_device_exclusion_range() to exclude the device from
>> IOMMU.
>>>> However I don't know how to generate range_start and range_length,
>> which are read from ACPI.
>>> 
>>> set_device_exclusion_range() is not the solution here. The best is if
>>> the GPU device is put into a passthrough domain at boot, in which it
>>> will be identity mapped. DMA still goes through the IOMMU in this
>>> case, but it only needs to lookup the device-table, page-table walks
>>> will not be done anymore.
>>> 
>>> The best way to implement this is to put it into the
>>> amd_iommu_add_device() in drivers/iommu/amd_iommu.c. There is this
>>> check:
>>> 
>>>       if (dev_data->iommu_v2)
>>> 		iommu_request_dm_for_dev(dev);
>>> 
>>> The iommu_request_dm_for_dev() function causes the device to be
>>> identity mapped. The check can be extended to also check for a device
>>> white-list for devices that need identity mapping.
>> 
>> My patch looks like this but the original behavior (4K screen flickering) is still
>> the same:
> 
> Does reverting the patch to disable ATS along with this patch help?

Unfortunately it doesn't help.

Kai-Heng

> 
> Alex
> 
>> 
>> diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
>> index bd25674ee4db..f913a25c9e92 100644
>> --- a/drivers/iommu/amd_iommu.c
>> +++ b/drivers/iommu/amd_iommu.c
>> @@ -42,6 +42,7 @@
>> #include <asm/iommu.h>
>> #include <asm/gart.h>
>> #include <asm/dma.h>
>> +#include <asm/pci-direct.h>
>> 
>> #include "amd_iommu_proto.h"
>> #include "amd_iommu_types.h"
>> @@ -2159,6 +2160,8 @@ static int amd_iommu_add_device(struct device
>> *dev)
>>        struct iommu_domain *domain;
>>        struct amd_iommu *iommu;
>>        int ret, devid;
>> +       bool need_identity_mapping = false;
>> +       u32 header;
>> 
>>        if (!check_device(dev) || get_dev_data(dev))
>>                return 0;
>> @@ -2184,7 +2187,11 @@ static int amd_iommu_add_device(struct device
>> *dev)
>> 
>>        BUG_ON(!dev_data);
>> 
>> -       if (dev_data->iommu_v2)
>> +       header = read_pci_config(0, PCI_BUS_NUM(devid), PCI_SLOT(devid),
>> PCI_FUNC(devid));
>> +       if ((header & 0xffff) == 0x1002 && (header >> 16) == 0x98e4)
>> +               need_identity_mapping = true;
>> +
>> +       if (dev_data->iommu_v2 || need_identity_mapping)
>>                iommu_request_dm_for_dev(dev);
>> 
>>        /* Domains are initialized for this device - have a look what we ended up
>> with */
>> 
>> 
>> $ dmesg | grep -i direct
>> [    0.011446] Using GB pages for direct mapping
>> [    0.703369] pci 0000:00:01.0: Using iommu direct mapping
>> [    0.703830] pci 0000:00:08.0: Using iommu direct mapping
>> 
>> So the graphics device (pci 0000:00:01.0:) is using direct mapping after the
>> change.
>> 
>> Kai-Heng
>> 
>>> 
>>> HTH,
>>> 
>>> 	Joerg