[PATCH 04/13] iommu coherency

[PATCH 04/13] iommu coherency

[Han, Weidong]

[-- Attachment #1: Type: text/plain, Size: 1489 bytes --]

in dmar_domain, more than one iommus may be included in iommu_bmp. Due to "Coherency" capability may be different across iommus, set this variable to indicate iommu access is coherent or not. Only when all related iommus in a dmar_domain are all coherent, iommu access of this domain is coherent.

Signed-off-by: Weidong Han <weidong.han@intel.com>
---
 drivers/pci/intel-iommu.c     |    6 ++++++
 include/linux/dma_remapping.h |    2 ++
 2 files changed, 8 insertions(+), 0 deletions(-)

diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
index a18e0b4..fa1507b 100644
--- a/drivers/pci/intel-iommu.c
+++ b/drivers/pci/intel-iommu.c
@@ -982,6 +982,12 @@ static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
 	domain->id = num;
 	memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
 	set_bit(iommu->seq_id, &domain->iommu_bmp);
+
+	if (ecap_coherent(iommu->ecap))
+		domain->iommu_coherency = 1;
+	else
+		domain->iommu_coherency = 0;
+
 	iommu->domains[num] = domain;
 	spin_unlock_irqrestore(&iommu->lock, flags);
 
diff --git a/include/linux/dma_remapping.h b/include/linux/dma_remapping.h
index c9d99c9..add2111 100644
--- a/include/linux/dma_remapping.h
+++ b/include/linux/dma_remapping.h
@@ -175,6 +175,8 @@ struct dmar_domain {
 
 #define DOMAIN_FLAG_MULTIPLE_DEVICES 1
 	int		flags;
+
+	int		iommu_coherency;/* iommu access is coherent or not */
 };
 
 /* PCI domain-device relationship */
-- 
1.5.1

[-- Attachment #2: 0004-iommu-coherency.patch --]
[-- Type: application/octet-stream, Size: 1634 bytes --]

From 306ef75fd396827592b823822da81568fe7dfc0b Mon Sep 17 00:00:00 2001
From: Weidong Han <weidong.han@intel.com>
Date: Tue, 2 Dec 2008 14:20:42 +0800
Subject: [PATCH] iommu coherency

in dmar_domain, more than one iommus may be included in iommu_bmp. Due to "Coherency" capability may be different across iommus, set this variable to indicate iommu access is coherent or not. Only when all related iommus in a dmar_domain are all coherent, iommu access of this domain is coherent.

Signed-off-by: Weidong Han <weidong.han@intel.com>
---
 drivers/pci/intel-iommu.c     |    6 ++++++
 include/linux/dma_remapping.h |    2 ++
 2 files changed, 8 insertions(+), 0 deletions(-)

diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
index a18e0b4..fa1507b 100644
--- a/drivers/pci/intel-iommu.c
+++ b/drivers/pci/intel-iommu.c
@@ -982,6 +982,12 @@ static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
 	domain->id = num;
 	memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
 	set_bit(iommu->seq_id, &domain->iommu_bmp);
+
+	if (ecap_coherent(iommu->ecap))
+		domain->iommu_coherency = 1;
+	else
+		domain->iommu_coherency = 0;
+
 	iommu->domains[num] = domain;
 	spin_unlock_irqrestore(&iommu->lock, flags);
 
diff --git a/include/linux/dma_remapping.h b/include/linux/dma_remapping.h
index c9d99c9..add2111 100644
--- a/include/linux/dma_remapping.h
+++ b/include/linux/dma_remapping.h
@@ -175,6 +175,8 @@ struct dmar_domain {
 
 #define DOMAIN_FLAG_MULTIPLE_DEVICES 1
 	int		flags;
+
+	int		iommu_coherency;/* iommu access is coherent or not */
 };
 
 /* PCI domain-device relationship */
-- 
1.5.1

Re: [PATCH 04/13] iommu coherency

[Mark McLoughlin]

On Tue, 2008-12-02 at 22:22 +0800, Han, Weidong wrote:

> in dmar_domain, more than one iommus may be included in iommu_bmp. Due
> to "Coherency" capability may be different across iommus, set this
> variable to indicate iommu access is coherent or not. Only when all
> related iommus in a dmar_domain are all coherent, iommu access of this
> domain is coherent.
> 
> Signed-off-by: Weidong Han <weidong.han@intel.com>
> ---
>  drivers/pci/intel-iommu.c     |    6 ++++++
>  include/linux/dma_remapping.h |    2 ++
>  2 files changed, 8 insertions(+), 0 deletions(-)
> 
> diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
> index a18e0b4..fa1507b 100644
> --- a/drivers/pci/intel-iommu.c
> +++ b/drivers/pci/intel-iommu.c
> @@ -982,6 +982,12 @@ static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
>  	domain->id = num;
>  	memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
>  	set_bit(iommu->seq_id, &domain->iommu_bmp);
> +
> +	if (ecap_coherent(iommu->ecap))
> +		domain->iommu_coherency = 1;
> +	else
> +		domain->iommu_coherency = 0;

If you allocate a non-coherent iommu, followed by a coherent iommu, then
iommu_coherency ends up as 1

In patch 6/13 you add domain_update_iommu_coherency(). It would make
more sense to add that function in this patch and use it here.

Cheers,
Mark.

RE: [PATCH 04/13] iommu coherency

[Han, Weidong]

Mark McLoughlin wrote:
> On Tue, 2008-12-02 at 22:22 +0800, Han, Weidong wrote:
> 
>> in dmar_domain, more than one iommus may be included in iommu_bmp.
>> Due 
>> to "Coherency" capability may be different across iommus, set this
>> variable to indicate iommu access is coherent or not. Only when all
>> related iommus in a dmar_domain are all coherent, iommu access of
>> this 
>> domain is coherent.
>> 
>> Signed-off-by: Weidong Han <weidong.han@intel.com>
>> ---
>>  drivers/pci/intel-iommu.c     |    6 ++++++
>>  include/linux/dma_remapping.h |    2 ++
>>  2 files changed, 8 insertions(+), 0 deletions(-)
>> 
>> diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
>> index a18e0b4..fa1507b 100644
>> --- a/drivers/pci/intel-iommu.c
>> +++ b/drivers/pci/intel-iommu.c
>> @@ -982,6 +982,12 @@ static struct dmar_domain *
>>  	iommu_alloc_domain(struct intel_iommu *iommu)  	domain->id = num;
>>  	memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
>> set_bit(iommu->seq_id, &domain->iommu_bmp); +
>> +	if (ecap_coherent(iommu->ecap))
>> +		domain->iommu_coherency = 1;
>> +	else
>> +		domain->iommu_coherency = 0;
> 
> If you allocate a non-coherent iommu, followed by a coherent iommu,
> then iommu_coherency ends up as 1
> 
> In patch 6/13 you add domain_update_iommu_coherency(). It would make
> more sense to add that function in this patch and use it here.

there is also an assumption that iommu_alloc_domain() is only used by "native VT-d", so the domain is related to only one iommu. I will add ASSERT() here.

Regards,
Weidong

> 
> Cheers,
> Mark.