diff mbox series

[3/5] iommu: Introduce the domain op enforce_cache_coherency()

Message ID 3-v1-ef02c60ddb76+12ca2-intel_no_snoop_jgg@nvidia.com (mailing list archive)
State New, archived
Headers show
Series Make the iommu driver no-snoop block feature consistent | expand

Commit Message

Jason Gunthorpe April 5, 2022, 4:16 p.m. UTC
This new mechanism will replace using IOMMU_CAP_CACHE_COHERENCY and
IOMMU_CACHE to control the no-snoop blocking behavior of the IOMMU.

Currently only Intel and AMD IOMMUs are known to support this
feature. They both implement it as an IOPTE bit, that when set, will cause
PCIe TLPs to that IOVA with the no-snoop bit set to be treated as though
the no-snoop bit was clear.

The new API is triggered by calling enforce_cache_coherency() before
mapping any IOVA to the domain which globally switches on no-snoop
blocking. This allows other implementations that might block no-snoop
globally and outside the IOPTE - AMD also documents such an HW capability.

Leave AMD out of sync with Intel and have it block no-snoop even for
in-kernel users. This can be trivially resolved in a follow up patch.

Only VFIO will call this new API.

Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/iommu/amd/iommu.c   |  7 +++++++
 drivers/iommu/intel/iommu.c | 14 +++++++++++++-
 include/linux/intel-iommu.h |  1 +
 include/linux/iommu.h       |  4 ++++
 4 files changed, 25 insertions(+), 1 deletion(-)

Comments

Alex Williamson April 5, 2022, 7:50 p.m. UTC | #1
On Tue,  5 Apr 2022 13:16:02 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> This new mechanism will replace using IOMMU_CAP_CACHE_COHERENCY and
> IOMMU_CACHE to control the no-snoop blocking behavior of the IOMMU.
> 
> Currently only Intel and AMD IOMMUs are known to support this
> feature. They both implement it as an IOPTE bit, that when set, will cause
> PCIe TLPs to that IOVA with the no-snoop bit set to be treated as though
> the no-snoop bit was clear.
> 
> The new API is triggered by calling enforce_cache_coherency() before
> mapping any IOVA to the domain which globally switches on no-snoop
> blocking. This allows other implementations that might block no-snoop
> globally and outside the IOPTE - AMD also documents such an HW capability.
> 
> Leave AMD out of sync with Intel and have it block no-snoop even for
> in-kernel users. This can be trivially resolved in a follow up patch.
> 
> Only VFIO will call this new API.
> 
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> ---
>  drivers/iommu/amd/iommu.c   |  7 +++++++
>  drivers/iommu/intel/iommu.c | 14 +++++++++++++-
>  include/linux/intel-iommu.h |  1 +
>  include/linux/iommu.h       |  4 ++++
>  4 files changed, 25 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index a1ada7bff44e61..e500b487eb3429 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -2271,6 +2271,12 @@ static int amd_iommu_def_domain_type(struct device *dev)
>  	return 0;
>  }
>  
> +static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
> +{
> +	/* IOMMU_PTE_FC is always set */
> +	return true;
> +}
> +
>  const struct iommu_ops amd_iommu_ops = {
>  	.capable = amd_iommu_capable,
>  	.domain_alloc = amd_iommu_domain_alloc,
> @@ -2293,6 +2299,7 @@ const struct iommu_ops amd_iommu_ops = {
>  		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
>  		.iotlb_sync	= amd_iommu_iotlb_sync,
>  		.free		= amd_iommu_domain_free,
> +		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
>  	}
>  };
>  
> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
> index df5c62ecf942b8..f08611a6cc4799 100644
> --- a/drivers/iommu/intel/iommu.c
> +++ b/drivers/iommu/intel/iommu.c
> @@ -4422,7 +4422,8 @@ static int intel_iommu_map(struct iommu_domain *domain,
>  		prot |= DMA_PTE_READ;
>  	if (iommu_prot & IOMMU_WRITE)
>  		prot |= DMA_PTE_WRITE;
> -	if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
> +	if (((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping) ||
> +	    dmar_domain->enforce_no_snoop)
>  		prot |= DMA_PTE_SNP;
>  
>  	max_addr = iova + size;
> @@ -4545,6 +4546,16 @@ static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
>  	return phys;
>  }
>  
> +static bool intel_iommu_enforce_cache_coherency(struct iommu_domain *domain)
> +{
> +	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
> +
> +	if (!dmar_domain->iommu_snooping)
> +		return false;
> +	dmar_domain->enforce_no_snoop = true;
> +	return true;
> +}

Don't we have issues if we try to set DMA_PTE_SNP on DMARs that don't
support it, ie. reserved register bit set in pte faults?  It seems
really inconsistent here that I could make a domain that supports
iommu_snooping, set enforce_no_snoop = true, then add another DMAR to
the domain that may not support iommu_snooping, I'd get false on the
subsequent enforcement test, but the dmar_domain is still trying to use
DMA_PTE_SNP.

There's also a disconnect, maybe just in the naming or documentation,
but if I call enforce_cache_coherency for a domain, that seems like the
domain should retain those semantics regardless of how it's modified,
ie. "enforced".  For example, if I tried to perform the above operation,
I should get a failure attaching the device that brings in the less
capable DMAR because the domain has been set to enforce this feature.

If the API is that I need to re-enforce_cache_coherency on every
modification of the domain, shouldn't dmar_domain->enforce_no_snoop
also return to a default value on domain changes?

Maybe this should be something like set_no_snoop_squashing with the
above semantics, it needs to be re-applied whenever the domain:device
composition changes?  Thanks,

Alex

> +
>  static bool intel_iommu_capable(enum iommu_cap cap)
>  {
>  	if (cap == IOMMU_CAP_CACHE_COHERENCY)
> @@ -4898,6 +4909,7 @@ const struct iommu_ops intel_iommu_ops = {
>  		.iotlb_sync		= intel_iommu_tlb_sync,
>  		.iova_to_phys		= intel_iommu_iova_to_phys,
>  		.free			= intel_iommu_domain_free,
> +		.enforce_cache_coherency = intel_iommu_enforce_cache_coherency,
>  	}
>  };
>  
> diff --git a/include/linux/intel-iommu.h b/include/linux/intel-iommu.h
> index 2f9891cb3d0014..1f930c0c225d94 100644
> --- a/include/linux/intel-iommu.h
> +++ b/include/linux/intel-iommu.h
> @@ -540,6 +540,7 @@ struct dmar_domain {
>  	u8 has_iotlb_device: 1;
>  	u8 iommu_coherency: 1;		/* indicate coherency of iommu access */
>  	u8 iommu_snooping: 1;		/* indicate snooping control feature */
> +	u8 enforce_no_snoop : 1;        /* Create IOPTEs with snoop control */
>  
>  	struct list_head devices;	/* all devices' list */
>  	struct iova_domain iovad;	/* iova's that belong to this domain */
> diff --git a/include/linux/iommu.h b/include/linux/iommu.h
> index 9208eca4b0d1ac..fe4f24c469c373 100644
> --- a/include/linux/iommu.h
> +++ b/include/linux/iommu.h
> @@ -272,6 +272,9 @@ struct iommu_ops {
>   * @iotlb_sync: Flush all queued ranges from the hardware TLBs and empty flush
>   *            queue
>   * @iova_to_phys: translate iova to physical address
> + * @enforce_cache_coherency: Prevent any kind of DMA from bypassing IOMMU_CACHE,
> + *                           including no-snoop TLPs on PCIe or other platform
> + *                           specific mechanisms.
>   * @enable_nesting: Enable nesting
>   * @set_pgtable_quirks: Set io page table quirks (IO_PGTABLE_QUIRK_*)
>   * @free: Release the domain after use.
> @@ -300,6 +303,7 @@ struct iommu_domain_ops {
>  	phys_addr_t (*iova_to_phys)(struct iommu_domain *domain,
>  				    dma_addr_t iova);
>  
> +	bool (*enforce_cache_coherency)(struct iommu_domain *domain);
>  	int (*enable_nesting)(struct iommu_domain *domain);
>  	int (*set_pgtable_quirks)(struct iommu_domain *domain,
>  				  unsigned long quirks);
Jason Gunthorpe April 5, 2022, 10:57 p.m. UTC | #2
On Tue, Apr 05, 2022 at 01:50:36PM -0600, Alex Williamson wrote:
> >  
> > +static bool intel_iommu_enforce_cache_coherency(struct iommu_domain *domain)
> > +{
> > +	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
> > +
> > +	if (!dmar_domain->iommu_snooping)
> > +		return false;
> > +	dmar_domain->enforce_no_snoop = true;
> > +	return true;
> > +}
> 
> Don't we have issues if we try to set DMA_PTE_SNP on DMARs that don't
> support it, ie. reserved register bit set in pte faults?  

The way the Intel driver is setup that is not possible. Currently it
does:

 static bool intel_iommu_capable(enum iommu_cap cap)
 {
	if (cap == IOMMU_CAP_CACHE_COHERENCY)
		return domain_update_iommu_snooping(NULL);

Which is a global property unrelated to any device.

Thus either all devices and all domains support iommu_snooping, or
none do.

It is unclear because for some reason the driver recalculates this
almost constant value on every device attach..

> There's also a disconnect, maybe just in the naming or documentation,
> but if I call enforce_cache_coherency for a domain, that seems like the
> domain should retain those semantics regardless of how it's
> modified,

Right, this is how I would expect it to work.

> ie. "enforced".  For example, if I tried to perform the above operation,
> I should get a failure attaching the device that brings in the less
> capable DMAR because the domain has been set to enforce this
> feature.

We don't have any code causing a failure like this because no driver
needs it.

> Maybe this should be something like set_no_snoop_squashing with the
> above semantics, it needs to be re-applied whenever the domain:device
> composition changes?  Thanks,

If we get a real driver that needs non-uniformity here we can revisit
what to do. There are a couple of good options depending on exactly
what the HW behavior is.

Is it more clear if I fold in the below? It helps show that the
decision to use DMA_PTE_SNP is a global choice based on
domain_update_iommu_snooping():

diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
index e5062461ab0640..fc789a9d955645 100644
--- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c
@@ -641,7 +641,6 @@ static unsigned long domain_super_pgsize_bitmap(struct dmar_domain *domain)
 static void domain_update_iommu_cap(struct dmar_domain *domain)
 {
 	domain_update_iommu_coherency(domain);
-	domain->iommu_snooping = domain_update_iommu_snooping(NULL);
 	domain->iommu_superpage = domain_update_iommu_superpage(domain, NULL);
 
 	/*
@@ -4283,7 +4282,6 @@ static int md_domain_init(struct dmar_domain *domain, int guest_width)
 	domain->agaw = width_to_agaw(adjust_width);
 
 	domain->iommu_coherency = false;
-	domain->iommu_snooping = false;
 	domain->iommu_superpage = 0;
 	domain->max_addr = 0;
 
@@ -4549,7 +4547,7 @@ static bool intel_iommu_enforce_cache_coherency(struct iommu_domain *domain)
 {
 	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
 
-	if (!dmar_domain->iommu_snooping)
+	if (!domain_update_iommu_snooping(NULL))
 		return false;
 	dmar_domain->enforce_no_snoop = true;
 	return true;
diff --git a/include/linux/intel-iommu.h b/include/linux/intel-iommu.h
index 1f930c0c225d94..bc39f633efdf03 100644
--- a/include/linux/intel-iommu.h
+++ b/include/linux/intel-iommu.h
@@ -539,7 +539,6 @@ struct dmar_domain {
 
 	u8 has_iotlb_device: 1;
 	u8 iommu_coherency: 1;		/* indicate coherency of iommu access */
-	u8 iommu_snooping: 1;		/* indicate snooping control feature */
 	u8 enforce_no_snoop : 1;        /* Create IOPTEs with snoop control */
 
 	struct list_head devices;	/* all devices' list */
Tian, Kevin April 5, 2022, 11:31 p.m. UTC | #3
> From: Jason Gunthorpe <jgg@nvidia.com>
> Sent: Wednesday, April 6, 2022 6:58 AM
> 
> On Tue, Apr 05, 2022 at 01:50:36PM -0600, Alex Williamson wrote:
> > >
> > > +static bool intel_iommu_enforce_cache_coherency(struct
> iommu_domain *domain)
> > > +{
> > > +	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
> > > +
> > > +	if (!dmar_domain->iommu_snooping)
> > > +		return false;
> > > +	dmar_domain->enforce_no_snoop = true;
> > > +	return true;
> > > +}
> >
> > Don't we have issues if we try to set DMA_PTE_SNP on DMARs that don't
> > support it, ie. reserved register bit set in pte faults?
> 
> The way the Intel driver is setup that is not possible. Currently it
> does:
> 
>  static bool intel_iommu_capable(enum iommu_cap cap)
>  {
> 	if (cap == IOMMU_CAP_CACHE_COHERENCY)
> 		return domain_update_iommu_snooping(NULL);
> 
> Which is a global property unrelated to any device.
> 
> Thus either all devices and all domains support iommu_snooping, or
> none do.
> 
> It is unclear because for some reason the driver recalculates this
> almost constant value on every device attach..

The reason is simply because iommu capability is a global flag 
Tian, Kevin April 6, 2022, 12:08 a.m. UTC | #4
> From: Tian, Kevin
> Sent: Wednesday, April 6, 2022 7:32 AM
> 
> > From: Jason Gunthorpe <jgg@nvidia.com>
> > Sent: Wednesday, April 6, 2022 6:58 AM
> >
> > On Tue, Apr 05, 2022 at 01:50:36PM -0600, Alex Williamson wrote:
> > > >
> > > > +static bool intel_iommu_enforce_cache_coherency(struct
> > iommu_domain *domain)
> > > > +{
> > > > +	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
> > > > +
> > > > +	if (!dmar_domain->iommu_snooping)
> > > > +		return false;
> > > > +	dmar_domain->enforce_no_snoop = true;
> > > > +	return true;
> > > > +}
> > >
> > > Don't we have issues if we try to set DMA_PTE_SNP on DMARs that don't
> > > support it, ie. reserved register bit set in pte faults?
> >
> > The way the Intel driver is setup that is not possible. Currently it
> > does:
> >
> >  static bool intel_iommu_capable(enum iommu_cap cap)
> >  {
> > 	if (cap == IOMMU_CAP_CACHE_COHERENCY)
> > 		return domain_update_iommu_snooping(NULL);
> >
> > Which is a global property unrelated to any device.
> >
> > Thus either all devices and all domains support iommu_snooping, or
> > none do.
> >
> > It is unclear because for some reason the driver recalculates this
> > almost constant value on every device attach..
> 
> The reason is simply because iommu capability is a global flag 
Tian, Kevin April 6, 2022, 7:09 a.m. UTC | #5
> From: Jason Gunthorpe <jgg@nvidia.com>
> Sent: Wednesday, April 6, 2022 12:16 AM
> 
> This new mechanism will replace using IOMMU_CAP_CACHE_COHERENCY
> and
> IOMMU_CACHE to control the no-snoop blocking behavior of the IOMMU.
> 
> Currently only Intel and AMD IOMMUs are known to support this
> feature. They both implement it as an IOPTE bit, that when set, will cause
> PCIe TLPs to that IOVA with the no-snoop bit set to be treated as though
> the no-snoop bit was clear.
> 
> The new API is triggered by calling enforce_cache_coherency() before
> mapping any IOVA to the domain which globally switches on no-snoop
> blocking. This allows other implementations that might block no-snoop
> globally and outside the IOPTE - AMD also documents such an HW capability.
> 
> Leave AMD out of sync with Intel and have it block no-snoop even for
> in-kernel users. This can be trivially resolved in a follow up patch.
> 
> Only VFIO will call this new API.

Is it too restrictive? In theory vdpa may also implement a contract with
KVM and then wants to call this new API too?
Jason Gunthorpe April 6, 2022, 12:27 p.m. UTC | #6
On Wed, Apr 06, 2022 at 07:09:49AM +0000, Tian, Kevin wrote:
> > From: Jason Gunthorpe <jgg@nvidia.com>
> > Sent: Wednesday, April 6, 2022 12:16 AM
> > 
> > This new mechanism will replace using IOMMU_CAP_CACHE_COHERENCY
> > and
> > IOMMU_CACHE to control the no-snoop blocking behavior of the IOMMU.
> > 
> > Currently only Intel and AMD IOMMUs are known to support this
> > feature. They both implement it as an IOPTE bit, that when set, will cause
> > PCIe TLPs to that IOVA with the no-snoop bit set to be treated as though
> > the no-snoop bit was clear.
> > 
> > The new API is triggered by calling enforce_cache_coherency() before
> > mapping any IOVA to the domain which globally switches on no-snoop
> > blocking. This allows other implementations that might block no-snoop
> > globally and outside the IOPTE - AMD also documents such an HW capability.
> > 
> > Leave AMD out of sync with Intel and have it block no-snoop even for
> > in-kernel users. This can be trivially resolved in a follow up patch.
> > 
> > Only VFIO will call this new API.
> 
> Is it too restrictive? In theory vdpa may also implement a contract with
> KVM and then wants to call this new API too?

I expect iommufd to handle this for all kernel users eventually.

Certainly vdpa should not be calling kvm functions.

Jason
diff mbox series

Patch

diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
index a1ada7bff44e61..e500b487eb3429 100644
--- a/drivers/iommu/amd/iommu.c
+++ b/drivers/iommu/amd/iommu.c
@@ -2271,6 +2271,12 @@  static int amd_iommu_def_domain_type(struct device *dev)
 	return 0;
 }
 
+static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
+{
+	/* IOMMU_PTE_FC is always set */
+	return true;
+}
+
 const struct iommu_ops amd_iommu_ops = {
 	.capable = amd_iommu_capable,
 	.domain_alloc = amd_iommu_domain_alloc,
@@ -2293,6 +2299,7 @@  const struct iommu_ops amd_iommu_ops = {
 		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
 		.iotlb_sync	= amd_iommu_iotlb_sync,
 		.free		= amd_iommu_domain_free,
+		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
 	}
 };
 
diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
index df5c62ecf942b8..f08611a6cc4799 100644
--- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c
@@ -4422,7 +4422,8 @@  static int intel_iommu_map(struct iommu_domain *domain,
 		prot |= DMA_PTE_READ;
 	if (iommu_prot & IOMMU_WRITE)
 		prot |= DMA_PTE_WRITE;
-	if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
+	if (((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping) ||
+	    dmar_domain->enforce_no_snoop)
 		prot |= DMA_PTE_SNP;
 
 	max_addr = iova + size;
@@ -4545,6 +4546,16 @@  static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
 	return phys;
 }
 
+static bool intel_iommu_enforce_cache_coherency(struct iommu_domain *domain)
+{
+	struct dmar_domain *dmar_domain = to_dmar_domain(domain);
+
+	if (!dmar_domain->iommu_snooping)
+		return false;
+	dmar_domain->enforce_no_snoop = true;
+	return true;
+}
+
 static bool intel_iommu_capable(enum iommu_cap cap)
 {
 	if (cap == IOMMU_CAP_CACHE_COHERENCY)
@@ -4898,6 +4909,7 @@  const struct iommu_ops intel_iommu_ops = {
 		.iotlb_sync		= intel_iommu_tlb_sync,
 		.iova_to_phys		= intel_iommu_iova_to_phys,
 		.free			= intel_iommu_domain_free,
+		.enforce_cache_coherency = intel_iommu_enforce_cache_coherency,
 	}
 };
 
diff --git a/include/linux/intel-iommu.h b/include/linux/intel-iommu.h
index 2f9891cb3d0014..1f930c0c225d94 100644
--- a/include/linux/intel-iommu.h
+++ b/include/linux/intel-iommu.h
@@ -540,6 +540,7 @@  struct dmar_domain {
 	u8 has_iotlb_device: 1;
 	u8 iommu_coherency: 1;		/* indicate coherency of iommu access */
 	u8 iommu_snooping: 1;		/* indicate snooping control feature */
+	u8 enforce_no_snoop : 1;        /* Create IOPTEs with snoop control */
 
 	struct list_head devices;	/* all devices' list */
 	struct iova_domain iovad;	/* iova's that belong to this domain */
diff --git a/include/linux/iommu.h b/include/linux/iommu.h
index 9208eca4b0d1ac..fe4f24c469c373 100644
--- a/include/linux/iommu.h
+++ b/include/linux/iommu.h
@@ -272,6 +272,9 @@  struct iommu_ops {
  * @iotlb_sync: Flush all queued ranges from the hardware TLBs and empty flush
  *            queue
  * @iova_to_phys: translate iova to physical address
+ * @enforce_cache_coherency: Prevent any kind of DMA from bypassing IOMMU_CACHE,
+ *                           including no-snoop TLPs on PCIe or other platform
+ *                           specific mechanisms.
  * @enable_nesting: Enable nesting
  * @set_pgtable_quirks: Set io page table quirks (IO_PGTABLE_QUIRK_*)
  * @free: Release the domain after use.
@@ -300,6 +303,7 @@  struct iommu_domain_ops {
 	phys_addr_t (*iova_to_phys)(struct iommu_domain *domain,
 				    dma_addr_t iova);
 
+	bool (*enforce_cache_coherency)(struct iommu_domain *domain);
 	int (*enable_nesting)(struct iommu_domain *domain);
 	int (*set_pgtable_quirks)(struct iommu_domain *domain,
 				  unsigned long quirks);