Re: [Xen-devel] [PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

Re: [Xen-devel] [PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

[Juergen Gross]

On 25/04/2019 11:01, Jan Beulich wrote:
>>>> On 23.04.19 at 20:36, <jgross@suse.com> wrote:
>> On 23/04/2019 19:05, Stefano Stabellini wrote:
>>> On Tue, 23 Apr 2019, Juergen Gross wrote:
>>>> Instead of always calling xen_destroy_contiguous_region() in case the
>>>> memory is DMA-able for the used device, do so only in case it has been
>>>> made DMA-able via xen_create_contiguous_region() before.
>>>>
>>>> This will avoid a lot of xen_destroy_contiguous_region() calls for
>>>> 64-bit capable devices.
>>>>
>>>> As the memory in question is owned by swiotlb-xen the PG_owner_priv_1
>>>> flag of the first allocated page can be used for remembering.
>>>
>>> Although the patch looks OK, this sentence puzzles me. Why do you say
>>> that the memory in question is owned by swiotlb-xen? Because it was
>>> returned by xen_alloc_coherent_pages? Both the x86 and the Arm
>>> implementation return fresh new memory, hence, it should be safe to set
>>> the PageOwnerPriv1 flag?
>>>
>>> My concern with this approach is with the semantics of PG_owner_priv_1.
>>> Is a page marked with PG_owner_priv_1 only supposed to be used by the
>>> owner?
>>
>> The owner of the page is free to use the flag.
>>
>> Like Grant pages are marked by the grant driver using this flag. And
>> Xen page tables are using it in PV-guests for indicating a "Pinned"
>> page table.
> 
> Considering the background of the series, isn't such multi-purpose use
> of the flag a possible problem? You're already suspecting a wrong call
> into here. The function finding the flag set (but for another reason)
> might add to the confusion. But I realize there are only so many page
> flags available.

Right.

> Perhaps the freeing function should, first thing, check the handed
> space actually matches the criteria (within dma_mask and contiguous)?

Yes, I think this is a good idea.


Juergen

[PATCH 0/3] xen/swiotlb: fix an issue and improve swiotlb-xen

[Juergen Gross]

While hunting an issue in swiotlb-xen I stumbled over a wrong test
and found some areas for improvement.

Juergen Gross (3):
  xen/swiotlb: fix condition for calling xen_destroy_contiguous_region()
  xen/swiotlb: simplify range_straddles_page_boundary()
  xen/swiotlb: remember having called xen_create_contiguous_region()

 drivers/xen/swiotlb-xen.c | 37 ++++++++++++-------------------------
 1 file changed, 12 insertions(+), 25 deletions(-)

-- 
2.16.4

[PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

[Juergen Gross]

Instead of always calling xen_destroy_contiguous_region() in case the
memory is DMA-able for the used device, do so only in case it has been
made DMA-able via xen_create_contiguous_region() before.

This will avoid a lot of xen_destroy_contiguous_region() calls for
64-bit capable devices.

As the memory in question is owned by swiotlb-xen the PG_owner_priv_1
flag of the first allocated page can be used for remembering.

Signed-off-by: Juergen Gross <jgross@suse.com>
---
 drivers/xen/swiotlb-xen.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
index 43b6e65ae256..a72f181d8e20 100644
--- a/drivers/xen/swiotlb-xen.c
+++ b/drivers/xen/swiotlb-xen.c
@@ -321,6 +321,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
 			return NULL;
 		}
+		SetPageOwnerPriv1(virt_to_page(ret));
 	}
 	memset(ret, 0, size);
 	return ret;
@@ -344,9 +345,11 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
 	/* Convert the size to actually allocated. */
 	size = 1UL << (order + XEN_PAGE_SHIFT);
 
-	if ((dev_addr + size - 1 <= dma_mask) &&
-	    !WARN_ON(range_straddles_page_boundary(phys, size)))
-		xen_destroy_contiguous_region(phys, order);
+	if (PageOwnerPriv1(virt_to_page(vaddr))) {
+		if (!WARN_ON(range_straddles_page_boundary(phys, size)))
+			xen_destroy_contiguous_region(phys, order);
+		ClearPageOwnerPriv1(virt_to_page(vaddr));
+	}
 
 	xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
 }
-- 
2.16.4

Re: [PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

[Stefano Stabellini]

On Tue, 23 Apr 2019, Juergen Gross wrote:
> Instead of always calling xen_destroy_contiguous_region() in case the
> memory is DMA-able for the used device, do so only in case it has been
> made DMA-able via xen_create_contiguous_region() before.
> 
> This will avoid a lot of xen_destroy_contiguous_region() calls for
> 64-bit capable devices.
> 
> As the memory in question is owned by swiotlb-xen the PG_owner_priv_1
> flag of the first allocated page can be used for remembering.

Although the patch looks OK, this sentence puzzles me. Why do you say
that the memory in question is owned by swiotlb-xen? Because it was
returned by xen_alloc_coherent_pages? Both the x86 and the Arm
implementation return fresh new memory, hence, it should be safe to set
the PageOwnerPriv1 flag?

My concern with this approach is with the semantics of PG_owner_priv_1.
Is a page marked with PG_owner_priv_1 only supposed to be used by the
owner?


> Signed-off-by: Juergen Gross <jgross@suse.com>
> ---
>  drivers/xen/swiotlb-xen.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> index 43b6e65ae256..a72f181d8e20 100644
> --- a/drivers/xen/swiotlb-xen.c
> +++ b/drivers/xen/swiotlb-xen.c
> @@ -321,6 +321,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
>  			xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
>  			return NULL;
>  		}
> +		SetPageOwnerPriv1(virt_to_page(ret));
>  	}
>  	memset(ret, 0, size);
>  	return ret;
> @@ -344,9 +345,11 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
>  	/* Convert the size to actually allocated. */
>  	size = 1UL << (order + XEN_PAGE_SHIFT);
>  
> -	if ((dev_addr + size - 1 <= dma_mask) &&
> -	    !WARN_ON(range_straddles_page_boundary(phys, size)))
> -		xen_destroy_contiguous_region(phys, order);
> +	if (PageOwnerPriv1(virt_to_page(vaddr))) {
> +		if (!WARN_ON(range_straddles_page_boundary(phys, size)))
> +			xen_destroy_contiguous_region(phys, order);
> +		ClearPageOwnerPriv1(virt_to_page(vaddr));
> +	}
>  
>  	xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
>  }

Re: [PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

[Juergen Gross]

On 23/04/2019 19:05, Stefano Stabellini wrote:
> On Tue, 23 Apr 2019, Juergen Gross wrote:
>> Instead of always calling xen_destroy_contiguous_region() in case the
>> memory is DMA-able for the used device, do so only in case it has been
>> made DMA-able via xen_create_contiguous_region() before.
>>
>> This will avoid a lot of xen_destroy_contiguous_region() calls for
>> 64-bit capable devices.
>>
>> As the memory in question is owned by swiotlb-xen the PG_owner_priv_1
>> flag of the first allocated page can be used for remembering.
> 
> Although the patch looks OK, this sentence puzzles me. Why do you say
> that the memory in question is owned by swiotlb-xen? Because it was
> returned by xen_alloc_coherent_pages? Both the x86 and the Arm
> implementation return fresh new memory, hence, it should be safe to set
> the PageOwnerPriv1 flag?
> 
> My concern with this approach is with the semantics of PG_owner_priv_1.
> Is a page marked with PG_owner_priv_1 only supposed to be used by the
> owner?

The owner of the page is free to use the flag.

Like Grant pages are marked by the grant driver using this flag. And
Xen page tables are using it in PV-guests for indicating a "Pinned"
page table.


Juergen

Re: [Xen-devel] [PATCH 3/3] xen/swiotlb: remember having called xen_create_contiguous_region()

[Jan Beulich]

>>> On 23.04.19 at 20:36, <jgross@suse.com> wrote:
> On 23/04/2019 19:05, Stefano Stabellini wrote:
>> On Tue, 23 Apr 2019, Juergen Gross wrote:
>>> Instead of always calling xen_destroy_contiguous_region() in case the
>>> memory is DMA-able for the used device, do so only in case it has been
>>> made DMA-able via xen_create_contiguous_region() before.
>>>
>>> This will avoid a lot of xen_destroy_contiguous_region() calls for
>>> 64-bit capable devices.
>>>
>>> As the memory in question is owned by swiotlb-xen the PG_owner_priv_1
>>> flag of the first allocated page can be used for remembering.
>> 
>> Although the patch looks OK, this sentence puzzles me. Why do you say
>> that the memory in question is owned by swiotlb-xen? Because it was
>> returned by xen_alloc_coherent_pages? Both the x86 and the Arm
>> implementation return fresh new memory, hence, it should be safe to set
>> the PageOwnerPriv1 flag?
>> 
>> My concern with this approach is with the semantics of PG_owner_priv_1.
>> Is a page marked with PG_owner_priv_1 only supposed to be used by the
>> owner?
> 
> The owner of the page is free to use the flag.
> 
> Like Grant pages are marked by the grant driver using this flag. And
> Xen page tables are using it in PV-guests for indicating a "Pinned"
> page table.

Considering the background of the series, isn't such multi-purpose use
of the flag a possible problem? You're already suspecting a wrong call
into here. The function finding the flag set (but for another reason)
might add to the confusion. But I realize there are only so many page
flags available.

Perhaps the freeing function should, first thing, check the handed
space actually matches the criteria (within dma_mask and contiguous)?

Jan