[Xen-devel] [PATCH v2] drm/xen-front: Make shmem backed display buffer coherent

[Oleksandr Andrushchenko]

On 1/24/19 5:02 PM, Julien Grall wrote:
>
>
> On 24/01/2019 14:34, Oleksandr Andrushchenko wrote:
>> Hello, Julien!
>
> Hi,
>
>> On 1/22/19 1:44 PM, Julien Grall wrote:
>>>
>>>
>>> On 1/22/19 10:28 AM, Oleksandr Andrushchenko wrote:
>>>> Hello, Julien!
>>>
>>> Hi,
>>>
>>>> On 1/21/19 7:09 PM, Julien Grall wrote:
>>>> Well, I didn't get the attributes of pages at the backend side, but 
>>>> IMO
>>>> those
>>>> do not matter in my use-case (for simplicity I am not using
>>>> zero-copying at
>>>> backend side):
>>>
>>> They are actually important no matter what is your use case. If you
>>> access the same physical page with different attributes, then you are
>>> asking for trouble.
>> So, we have:
>>
>> DomU: frontend side
>> ====================
>> !PTE_RDONLY + PTE_PXN + PTE_SHARED + PTE_AF + PTE_UXN +
>> PTE_ATTRINDX(MT_NORMAL)
>
> I still don't understand how you came up with MT_NORMAL when you seem 
> to confirm...
>
>>
>> DomD: backend side
>> ====================
>> PTE_USER + !PTE_RDONLY + PTE_PXN + PTE_NG + PTE_CONT + PTE_TABLE_BIT +
>> PTE_UXN + PTE_ATTRINDX(MT_NORMAL)
>>
>>   From the above it seems that I don't violate cached/non-cached
>> agreement here
>>>
>>> This is why Xen imposes all the pages shared to have their memory
>>> attributes following some rules. Actually, speaking with Mark R., we
>>> may want to tight a bit more the attributes.
>>>
>>>>
>>>> 1. Frontend device allocates display buffer pages which come from 
>>>> shmem
>>>> and have these attributes:
>>>> !PTE_RDONLY + PTE_PXN + PTE_SHARED + PTE_AF + PTE_UXN +
>>>> PTE_ATTRINDX(MT_NORMAL)
>>>
>>> My knowledge of Xen DRM is inexistent. However, looking at the code in
>>> 5.0-rc2, I don't seem to find the same attributes. For instance
>>> xen_drm_front_gem_prime_vmap and gem_mmap_obj are using
>>> pgprot_writecombine. So it looks like, the mapping will be
>>> non-cacheable on Arm64.
>>>
>>> Can you explain how you came up to these attributes?
>> pgprot_writecombine is PTE_ATTRINDX(MT_NORMAL_NC), so it seems to be
>> applicable here? [1]
>
> ... that MT_NORMAL_NC is used for the frontend pages.
>
> MT_NORMAL_NC is different from MT_NORMAL. The use of the former will 
> result to non-cacheable memory while the latter will result to 
> cacheable memory.
>
> To me, this looks like the exact reason why you see artifact on the 
> display buffer. As the author of this code, can you explain why you 
> decided to use pgprot_writecombine here instead of relying on the 
> default VMA prot?
>
> [...]
Sorry for late reply, it took me quite some time to re-check all the 
use-cases
that we have with PV DRM.
First of all I found a bug in my debug code which reported page 
attributes and
now I can confirm that all the use-cases sue MT_NORMAL, both backend and 
frontend.
You are right: there is no reason to have pgprot_writecombine and indeed 
I have
to rely on almost default VMA prot. I came up with the following (used 
by omap drm,
for example):

diff --git a/drivers/gpu/drm/xen/xen_drm_front_gem.c 
b/drivers/gpu/drm/xen/xen_drm_front_gem.c
index ae28ad4b4254..867800a2ed42 100644
--- a/drivers/gpu/drm/xen/xen_drm_front_gem.c
+++ b/drivers/gpu/drm/xen/xen_drm_front_gem.c
@@ -238,8 +238,7 @@ static int gem_mmap_obj(struct xen_gem_object *xen_obj,
         vma->vm_flags &= ~VM_PFNMAP;
         vma->vm_flags |= VM_MIXEDMAP;
         vma->vm_pgoff = 0;
-       vma->vm_page_prot =
- pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
+       vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);

         {
                 int cnt = xen_obj->num_pages > 5 ? 5 : xen_obj->num_pages;
@@ -295,7 +294,7 @@ void *xen_drm_front_gem_prime_vmap(struct 
drm_gem_object *gem_obj)
                 return NULL;

         return vmap(xen_obj->pages, xen_obj->num_pages,
-                   VM_MAP, pgprot_writecombine(PAGE_KERNEL));
+                   VM_MAP, PAGE_KERNEL);
  }

With the above all the artifacts are gone now as page attributes are the 
same across
all domains. So, I consider this as a fix and will send it as v3 or better
drop this patch and have a new one.

THANK YOU for helping figuring this out!
>
>>> We actually never required to use cache flush in other PV protocol, so
>>> I still don't understand why the PV DRM should be different here.
>> Well, you are right. But at the same time not flushing the buffer makes
>> troubles,
>> so this is why I am trying to figure out what is wrong here.
>
> The cache flush is likely hiding the real problem rather than solving it.
>
It does hide the real issue. And I have confirmed this
>>>
>>> To me, it looks like that you are either missing some barriers
>> Barriers for the buffer? Not sure what you mean here.
>
> If you share information between two entities, you may need some 
> ordering so the information are seen consistently by the consumer 
> side. This can be achieved by using barriers.
>
>> Even more, we have
>> a use case
>> when the buffer is not touched by CPU in DomD and is solely owned by 
>> the HW.
>
> Memory ordering issues are subtle. The fact that one of your use-case 
> works does not imply that barriers are not necessary. I am not saying 
> there are a missing barriers, I am only pointed out potential reasons.
>
> Anyway, I don't think your problem is a missing barriers here. It is 
> more likely because of mismatch memory attributes (see above).
>
Yes, please see above
> Cheers,
>
Thank you all for helping with the correct fix,
Oleksandr