Skip to contentSkip to conversation

iommu/vt-d: Misc tweaks and fixes for vSVA

PATCH In review Jacob Pan 2 participants 9 messages Jul 22, 2020 – Jul 23, 2020

Subsystems (2)
lore
Share & tools
lore

Jacob Pan

author
#0 lore raw

v5 cover iommu/vt-d: Misc tweaks and fixes for vSVA

Hi Baolu and all,

This is a series to address some of the issues we found in vSVA support. Most of the patches deal with exception handling, we also removed some bits that are not currently supported.

Many thanks to Kevin Tian's review.

Jacob & Yi

Changelog:

v5 Fix off by one bug in 4/7

v4 Address reviews from Eric Auger
- Additional warning for unaligned mask and address in guest IOTLB invalidation
- Comments rewrite and code simplification
v3
- Use pr_err instead of WARN() for invalid user address range (6/7)
- Fix logic in PASID selective devTLB flush (3/7)

v2 Address reviews from Baolu
- Fixed addr field in devTLB flush (5/7)
- Assign address for single page devTLB invalidation (4/7)
- Coding style tweaks

Jacob Pan (4):
iommu/vt-d: Remove global page support in devTLB flush iommu/vt-d: Fix PASID devTLB invalidation iommu/vt-d: Warn on out-of-range invalidation address iommu/vt-d: Disable multiple GPASID-dev bind

Liu Yi L (3):
iommu/vt-d: Enforce PASID devTLB field mask
iommu/vt-d: Handle non-page aligned address
iommu/vt-d: Fix devTLB flush for vSVA

 drivers/iommu/intel/dmar.c  | 24 +++++++++++++++++++-----
 drivers/iommu/intel/iommu.c | 39 ++++++++++++++++++++++++---------------
 drivers/iommu/intel/pasid.c | 11 ++++++++++-
 drivers/iommu/intel/svm.c   | 22 +++++++++-------------
 include/linux/intel-iommu.h |  5 ++---
 5 files changed, 64 insertions(+), 37 deletions(-)
Signature · 1 line
2.7.4
author
#2 lore raw

PATCH v5 2/7 iommu/vt-d: Remove global page support in devTLB flush

Global pages support is removed from VT-d spec 3.0 for dev TLB invalidation. This patch is to remove the bits for vSVA. Similar change already made for the native SVA. See the link below.

---
 drivers/iommu/intel/dmar.c  | 4 +---
 drivers/iommu/intel/iommu.c | 4 ++--
 include/linux/intel-iommu.h | 3 +--
 3 files changed, 4 insertions(+), 7 deletions(-)
drivers/iommu/intel/dmar.c +1 -3
index 683b812c5c47..9be08b9400ee 100644
@@ -1438,8 +1438,7 @@ void qi_flush_piotlb(struct intel_iommu *iommu, u16 did, u32 pasid, u64 addr,
 
 /* PASID-based device IOTLB Invalidate */
 void qi_flush_dev_iotlb_pasid(struct intel_iommu *iommu, u16 sid, u16 pfsid,
-			      u32 pasid,  u16 qdep, u64 addr,
-			      unsigned int size_order, u64 granu)
+			      u32 pasid,  u16 qdep, u64 addr, unsigned int size_order)
 {
 	unsigned long mask = 1UL << (VTD_PAGE_SHIFT + size_order - 1);
 	struct qi_desc desc = {.qw1 = 0, .qw2 = 0, .qw3 = 0};
@@ -1447,7 +1446,6 @@ void qi_flush_dev_iotlb_pasid(struct intel_iommu *iommu, u16 sid, u16 pfsid,
 	desc.qw0 = QI_DEV_EIOTLB_PASID(pasid) | QI_DEV_EIOTLB_SID(sid) |
 		QI_DEV_EIOTLB_QDEP(qdep) | QI_DEIOTLB_TYPE |
 		QI_DEV_IOTLB_PFSID(pfsid);
-	desc.qw1 = QI_DEV_EIOTLB_GLOB(granu);
 
 	/*
 	 * If S bit is 0, we only flush a single page. If S bit is set,
drivers/iommu/intel/iommu.c +2 -2
index d759e7234e98..bdd1e7d81178 100644
@@ -5474,7 +5474,7 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 						info->pfsid, pasid,
 						info->ats_qdep,
 						inv_info->addr_info.addr,
-						size, granu);
+						size);
 			break;
 		case IOMMU_CACHE_INV_TYPE_DEV_IOTLB:
 			if (info->ats_enabled)
@@ -5482,7 +5482,7 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 						info->pfsid, pasid,
 						info->ats_qdep,
 						inv_info->addr_info.addr,
-						size, granu);
+						size);
 			else
 				pr_warn_ratelimited("Passdown device IOTLB flush w/o ATS!\n");
 			break;
include/linux/intel-iommu.h +1 -2
index 311117b50e93..c7a8aae36771 100644
@@ -381,7 +381,6 @@ enum {
 
 #define QI_DEV_EIOTLB_ADDR(a)	((u64)(a) & VTD_PAGE_MASK)
 #define QI_DEV_EIOTLB_SIZE	(((u64)1) << 11)
-#define QI_DEV_EIOTLB_GLOB(g)	((u64)(g) & 0x1)
 #define QI_DEV_EIOTLB_PASID(p)	((u64)((p) & 0xfffff) << 32)
 #define QI_DEV_EIOTLB_SID(sid)	((u64)((sid) & 0xffff) << 16)
 #define QI_DEV_EIOTLB_QDEP(qd)	((u64)((qd) & 0x1f) << 4)
@@ -705,7 +704,7 @@ void qi_flush_piotlb(struct intel_iommu *iommu, u16 did, u32 pasid, u64 addr,
 
 void qi_flush_dev_iotlb_pasid(struct intel_iommu *iommu, u16 sid, u16 pfsid,
 			      u32 pasid, u16 qdep, u64 addr,
-			      unsigned int size_order, u64 granu);
+			      unsigned int size_order);
 void qi_flush_pasid_cache(struct intel_iommu *iommu, u16 did, u64 granu,
 			  int pasid);
 
Signature · 1 line
2.7.4
author
#3 lore raw

PATCH v5 3/7 iommu/vt-d: Fix PASID devTLB invalidation

In reply to Jacob Pan

DevTLB flush can be used for both DMA request with and without PASIDs. The former uses PASID#0 (RID2PASID), latter uses non-zero PASID for SVA usage.

This patch adds a check for PASID value such that devTLB flush with PASID is used for SVA case. This is more efficient in that multiple PASIDs can be used by a single device, when tearing down a PASID entry we shall flush only the devTLB specific to a PASID.

Fixes 6f7db75e1c46 ("iommu/vt-d: Add second level page table")
---
 drivers/iommu/intel/pasid.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
drivers/iommu/intel/pasid.c +10 -1
index c81f0f17c6ba..fa0154cce537 100644
@@ -486,7 +486,16 @@ devtlb_invalidation_with_pasid(struct intel_iommu *iommu,
 	qdep = info->ats_qdep;
 	pfsid = info->pfsid;
 
-	qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
+	/*
+	 * When PASID 0 is used, it indicates RID2PASID(DMA request w/o PASID),
+	 * devTLB flush w/o PASID should be used. For non-zero PASID under
+	 * SVA usage, device could do DMA with multiple PASIDs. It is more
+	 * efficient to flush devTLB specific to the PASID.
+	 */
+	if (pasid == PASID_RID2PASID)
+		qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
+	else
+		qi_flush_dev_iotlb_pasid(iommu, sid, pfsid, pasid, qdep, 0, 64 - VTD_PAGE_SHIFT);
 }
 
 void intel_pasid_tear_down_entry(struct intel_iommu *iommu, struct device *dev,
Signature · 1 line
2.7.4
author
#5 lore raw

PATCH v5 4/7 iommu/vt-d: Handle non-page aligned address

In reply to Jacob Pan

From: Liu Yi L <[email protected]>

Address information for device TLB invalidation comes from userspace when device is directly assigned to a guest with vIOMMU support. VT-d requires page aligned address. This patch checks and enforce address to be page aligned, otherwise reserved bits can be set in the invalidation descriptor. Unrecoverable fault will be reported due to non-zero value in the reserved bits.

Fixes 61a06a16e36d8 ("iommu/vt-d: Support flushing more translation

cache types")

---
 drivers/iommu/intel/dmar.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)
drivers/iommu/intel/dmar.c +18 -2
index 9be08b9400ee..6038d02c5066 100644
@@ -1456,9 +1456,25 @@ void qi_flush_dev_iotlb_pasid(struct intel_iommu *iommu, u16 sid, u16 pfsid,
 	 * Max Invs Pending (MIP) is set to 0 for now until we have DIT in
 	 * ECAP.
 	 */
-	desc.qw1 |= addr & ~mask;
-	if (size_order)
+	if (addr & GENMASK_ULL(size_order + VTD_PAGE_SHIFT, 0))
+		pr_warn_ratelimited("Invalidate non-aligned address %llx, order %d\n", addr, size_order);
+
+	/* Take page address */
+	desc.qw1 = QI_DEV_EIOTLB_ADDR(addr);
+
+	if (size_order) {
+		/*
+		 * Existing 0s in address below size_order may be the least
+		 * significant bit, we must set them to 1s to avoid having
+		 * smaller size than desired.
+		 */
+		desc.qw1 |= GENMASK_ULL(size_order + VTD_PAGE_SHIFT - 1,
+					VTD_PAGE_SHIFT);
+		/* Clear size_order bit to indicate size */
+		desc.qw1 &= ~mask;
+		/* Set the S bit to indicate flushing more than 1 page */
 		desc.qw1 |= QI_DEV_EIOTLB_SIZE;
+	}
 
 	qi_submit_sync(iommu, &desc, 1, 0);
 }
Signature · 6 lines
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
author
#1 lore raw

PATCH v5 1/7 iommu/vt-d: Enforce PASID devTLB field mask

In reply to Jacob Pan

From: Liu Yi L <[email protected]>

Set proper masks to avoid invalid input spillover to reserved bits.

---
 include/linux/intel-iommu.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
include/linux/intel-iommu.h +2 -2
index 3e8fa1c7a1e6..311117b50e93 100644
@@ -381,8 +381,8 @@ enum {
 
 #define QI_DEV_EIOTLB_ADDR(a)	((u64)(a) & VTD_PAGE_MASK)
 #define QI_DEV_EIOTLB_SIZE	(((u64)1) << 11)
-#define QI_DEV_EIOTLB_GLOB(g)	((u64)g)
-#define QI_DEV_EIOTLB_PASID(p)	(((u64)p) << 32)
+#define QI_DEV_EIOTLB_GLOB(g)	((u64)(g) & 0x1)
+#define QI_DEV_EIOTLB_PASID(p)	((u64)((p) & 0xfffff) << 32)
 #define QI_DEV_EIOTLB_SID(sid)	((u64)((sid) & 0xffff) << 16)
 #define QI_DEV_EIOTLB_QDEP(qd)	((u64)((qd) & 0x1f) << 4)
 #define QI_DEV_EIOTLB_PFSID(pfsid) (((u64)(pfsid & 0xf) << 12) | \
Signature · 6 lines
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
author
#6 lore raw

PATCH v5 5/7 iommu/vt-d: Fix devTLB flush for vSVA

In reply to Jacob Pan

From: Liu Yi L <[email protected]>

For guest SVA usage, in order to optimize for less VMEXIT, guest request of IOTLB flush also includes device TLB.

On the host side, IOMMU driver performs IOTLB and implicit devTLB invalidation. When PASID-selective granularity is requested by the guest we need to derive the equivalent address range for devTLB instead of using the address information in the UAPI data. The reason for that is, unlike IOTLB flush, devTLB flush does not support PASID-selective granularity. This is to say, we need to set the following in the PASID based devTLB invalidation descriptor:
- entire 64 bit range in address ~(0x1 << 63)
- S bit = 1 (VT-d CH 6.5.2.6).

Without this fix, device TLB flush range is not set properly for PASID selective granularity. This patch also merged devTLB flush code for both implicit and explicit cases.

Fixes 6ee1b77ba3ac ("iommu/vt-d: Add svm/sva invalidate function")
---
 drivers/iommu/intel/iommu.c | 30 ++++++++++++++++++++----------
 1 file changed, 20 insertions(+), 10 deletions(-)
drivers/iommu/intel/iommu.c +20 -10
index bdd1e7d81178..c6999b9b6265 100644
@@ -5416,7 +5416,7 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 	sid = PCI_DEVID(bus, devfn);
 
 	/* Size is only valid in address selective invalidation */
-	if (inv_info->granularity != IOMMU_INV_GRANU_PASID)
+	if (inv_info->granularity == IOMMU_INV_GRANU_ADDR)
 		size = to_vtd_size(inv_info->addr_info.granule_size,
 				   inv_info->addr_info.nb_granules);
 
@@ -5425,6 +5425,7 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 			 IOMMU_CACHE_INV_TYPE_NR) {
 		int granu = 0;
 		u64 pasid = 0;
+		u64 addr = 0;
 
 		granu = to_vtd_granularity(cache_type, inv_info->granularity);
 		if (granu == -EINVAL) {
@@ -5464,24 +5465,33 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 					(granu == QI_GRAN_NONG_PASID) ? -1 : 1 << size,
 					inv_info->addr_info.flags & IOMMU_INV_ADDR_FLAGS_LEAF);
 
+			if (!info->ats_enabled)
+				break;
 			/*
 			 * Always flush device IOTLB if ATS is enabled. vIOMMU
 			 * in the guest may assume IOTLB flush is inclusive,
 			 * which is more efficient.
 			 */
-			if (info->ats_enabled)
-				qi_flush_dev_iotlb_pasid(iommu, sid,
-						info->pfsid, pasid,
-						info->ats_qdep,
-						inv_info->addr_info.addr,
-						size);
-			break;
+			fallthrough;
 		case IOMMU_CACHE_INV_TYPE_DEV_IOTLB:
+			/*
+			 * PASID based device TLB invalidation does not support
+			 * IOMMU_INV_GRANU_PASID granularity but only supports
+			 * IOMMU_INV_GRANU_ADDR.
+			 * The equivalent of that is we set the size to be the
+			 * entire range of 64 bit. User only provides PASID info
+			 * without address info. So we set addr to 0.
+			 */
+			if (inv_info->granularity == IOMMU_INV_GRANU_PASID) {
+				size = 64 - VTD_PAGE_SHIFT;
+				addr = 0;
+			} else if (inv_info->granularity == IOMMU_INV_GRANU_ADDR)
+				addr = inv_info->addr_info.addr;
+
 			if (info->ats_enabled)
 				qi_flush_dev_iotlb_pasid(iommu, sid,
 						info->pfsid, pasid,
-						info->ats_qdep,
-						inv_info->addr_info.addr,
+						info->ats_qdep, addr,
 						size);
 			else
 				pr_warn_ratelimited("Passdown device IOTLB flush w/o ATS!\n");
Signature · 6 lines
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
author
#7 lore raw

PATCH v5 6/7 iommu/vt-d: Warn on out-of-range invalidation address

In reply to Jacob Pan

For guest requested IOTLB invalidation, address and mask are provided as part of the invalidation data. VT-d HW silently ignores any address bits below the mask. SW shall also allow such case but give warning if address does not align with the mask. This patch relax the fault handling from error to warning and proceed with invalidation request with the given mask.

Fixes 6ee1b77ba3ac0 ("iommu/vt-d: Add svm/sva invalidate function")
---
 drivers/iommu/intel/iommu.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
drivers/iommu/intel/iommu.c +3 -4
index c6999b9b6265..92c7ad66e64c 100644
@@ -5447,13 +5447,12 @@ intel_iommu_sva_invalidate(struct iommu_domain *domain, struct device *dev,
 
 		switch (BIT(cache_type)) {
 		case IOMMU_CACHE_INV_TYPE_IOTLB:
+			/* HW will ignore LSB bits based on address mask */
 			if (inv_info->granularity == IOMMU_INV_GRANU_ADDR &&
 			    size &&
 			    (inv_info->addr_info.addr & ((BIT(VTD_PAGE_SHIFT + size)) - 1))) {
-				pr_err_ratelimited("Address out of range, 0x%llx, size order %llu\n",
-						   inv_info->addr_info.addr, size);
-				ret = -ERANGE;
-				goto out_unlock;
+				pr_err_ratelimited("User address not aligned, 0x%llx, size order %llu\n",
+					  inv_info->addr_info.addr, size);
 			}
 
 			/*
Signature · 6 lines
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
author
#8 lore raw

PATCH v5 7/7 iommu/vt-d: Disable multiple GPASID-dev bind

In reply to Jacob Pan

For the unlikely use case where multiple aux domains from the same pdev are attached to a single guest and then bound to a single process (thus same PASID) within that guest, we cannot easily support this case by refcounting the number of users. As there is only one SL page table per PASID while we have multiple aux domains thus multiple SL page tables for the same PASID.

Extra unbinding guest PASID can happen due to race between normal and exception cases. Termination of one aux domain may affect others unless we actively track and switch aux domains to ensure the validity of SL page tables and TLB states in the shared PASID entry.

Support for sharing second level PGDs across domains can reduce the complexity but this is not available due to the limitations on VFIO container architecture. We can revisit this decision once sharing PGDs are available.

Overall, the complexity and potential glitch do not warrant this unlikely use case thereby removed by this patch.

Fixes 56722a4398a30 ("iommu/vt-d: Add bind guest PASID support")
Cc Kevin Tian <[email protected]>
Cc Lu Baolu <[email protected]>
---
 drivers/iommu/intel/svm.c | 22 +++++++++-------------
 1 file changed, 9 insertions(+), 13 deletions(-)
drivers/iommu/intel/svm.c +9 -13
index 6c87c807a0ab..d386853121a2 100644
@@ -277,20 +277,16 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			goto out;
 		}
 
+		/*
+		 * Do not allow multiple bindings of the same device-PASID since
+		 * there is only one SL page tables per PASID. We may revisit
+		 * once sharing PGD across domains are supported.
+		 */
 		for_each_svm_dev(sdev, svm, dev) {
-			/*
-			 * For devices with aux domains, we should allow
-			 * multiple bind calls with the same PASID and pdev.
-			 */
-			if (iommu_dev_feature_enabled(dev,
-						      IOMMU_DEV_FEAT_AUX)) {
-				sdev->users++;
-			} else {
-				dev_warn_ratelimited(dev,
-						     "Already bound with PASID %u\n",
-						     svm->pasid);
-				ret = -EBUSY;
-			}
+			dev_warn_ratelimited(dev,
+					     "Already bound with PASID %u\n",
+					     svm->pasid);
+			ret = -EBUSY;
 			goto out;
 		}
 	} else {
Signature · 6 lines
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
maintainer Acked-by on v5 3/7
#4 lore raw

Re: [PATCH v5 3/7] iommu/vt-d: Fix PASID devTLB invalidation

In reply to Jacob Pan
Quoted 4 lines · Jacob Pan

DevTLB flush can be used for both DMA request with and without PASIDs. The former uses PASID#0 (RID2PASID), latter uses non-zero PASID for SVA usage.

Quoted 8 lines

This patch adds a check for PASID value such that devTLB flush with PASID is used for SVA case. This is more efficient in that multiple PASIDs can be used by a single device, when tearing down a PASID entry we shall flush only the devTLB specific to a PASID.

Fixes 6f7db75e1c46 ("iommu/vt-d: Add second level page table")

Best regards,
baolu

Quoted 27 lines
---
  drivers/iommu/intel/pasid.c | 11 ++++++++++-
  1 file changed, 10 insertions(+), 1 deletion(-)
drivers/iommu/intel/pasid.c +10 -1
index c81f0f17c6ba..fa0154cce537 100644
@@ -486,7 +486,16 @@ devtlb_invalidation_with_pasid(struct intel_iommu *iommu,
  	qdep = info->ats_qdep;
  	pfsid = info->pfsid;
 
-	qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
+	/*
+	 * When PASID 0 is used, it indicates RID2PASID(DMA request w/o PASID),
+	 * devTLB flush w/o PASID should be used. For non-zero PASID under
+	 * SVA usage, device could do DMA with multiple PASIDs. It is more
+	 * efficient to flush devTLB specific to the PASID.
+	 */
+	if (pasid == PASID_RID2PASID)
+		qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
+	else
+		qi_flush_dev_iotlb_pasid(iommu, sid, pfsid, pasid, qdep, 0, 64 - VTD_PAGE_SHIFT);
  }
 
  void intel_pasid_tear_down_entry(struct intel_iommu *iommu, struct device *dev,