[RFC 5/7] iommu/vt-d: Add support for KVA PASID mode

[Jacob Pan]

To support KVA fast mode, the VT-d driver must support domain allocation
of IOMMU_DOMAIN_KVA type. Since all devices in fast KVA mode share the
same kernel mapping, a single KVA domain is sufficient. This global KVA
domain contains the kernel mapping, i.e. init_mm.pgd.

The programming of the KVA domain follows the existing flow of auxiliary
domain attachment.

Signed-off-by: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx>
---
 drivers/iommu/intel/iommu.c | 59 ++++++++++++++++++++++++++++++++++---
 1 file changed, 55 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
index cbcfd178c16f..0dabd5f75acf 100644
--- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c
@@ -293,6 +293,9 @@ static inline void context_clear_entry(struct context_entry *context)
  *	3. Each iommu mapps to this domain if successful.
  */
 static struct dmar_domain *si_domain;
+
+/* This domain is used for shared virtual addressing with CPU kernel mapping */
+static struct dmar_domain *kva_domain;
 static int hw_pass_through = 1;
 
 #define for_each_domain_iommu(idx, domain)			\
@@ -1989,6 +1992,10 @@ static void domain_exit(struct dmar_domain *domain)
 	/* Remove associated devices and clear attached or cached domains */
 	domain_remove_dev_info(domain);
 
+	/* There is no IOMMU page table for KVA */
+	if (domain->pgd == (struct dma_pte *)init_mm.pgd)
+		return;
+
 	/* destroy iovas */
 	if (domain->domain.type == IOMMU_DOMAIN_DMA)
 		iommu_put_dma_cookie(&domain->domain);
@@ -2533,6 +2540,10 @@ static int domain_setup_first_level(struct intel_iommu *iommu,
 	int agaw, level;
 	int flags = 0;
 
+	if (domain->domain.type == IOMMU_DOMAIN_KVA) {
+		flags |= PASID_FLAG_SUPERVISOR_MODE;
+		goto do_setup;
+	}
 	/*
 	 * Skip top levels of page tables for iommu which has
 	 * less agaw than default. Unnecessary for PT mode.
@@ -2554,7 +2565,7 @@ static int domain_setup_first_level(struct intel_iommu *iommu,
 
 	if (domain->domain.type == IOMMU_DOMAIN_UNMANAGED)
 		flags |= PASID_FLAG_PAGE_SNOOP;
-
+do_setup:
 	return intel_pasid_setup_first_level(iommu, dev, (pgd_t *)pgd, pasid,
 					     domain->iommu_did[iommu->seq_id],
 					     flags);
@@ -2713,7 +2724,28 @@ static int iommu_domain_identity_map(struct dmar_domain *domain,
 }
 
 static int md_domain_init(struct dmar_domain *domain, int guest_width);
+#ifdef CONFIG_INTEL_IOMMU_SVM
+static int __init kva_domain_init(void)
+{
+	struct dmar_domain *dmar_domain;
+	struct iommu_domain *domain;
 
+	kva_domain = alloc_domain(0);
+	if (!kva_domain) {
+		pr_err("Can't allocate KVA domain\n");
+		return -EFAULT;
+	}
+	kva_domain->pgd = (struct dma_pte *)init_mm.pgd;
+	domain = &kva_domain->domain;
+	domain->type = IOMMU_DOMAIN_KVA;
+	/* REVISIT: may not need this other than sanity check */
+	domain->geometry.aperture_start = 0;
+	domain->geometry.aperture_end   =
+		__DOMAIN_MAX_ADDR(dmar_domain->gaw);
+	domain->geometry.force_aperture = true;
+	return 0;
+}
+#endif
 static int __init si_domain_init(int hw)
 {
 	struct dmar_rmrr_unit *rmrr;
@@ -3363,6 +3395,11 @@ static int __init init_dmars(void)
 			down_write(&dmar_global_lock);
 			if (ret)
 				goto free_iommu;
+			/* For in-kernel DMA with PASID in SVA */
+			ret = kva_domain_init();
+			if (ret)
+				goto free_iommu;
+
 		}
 #endif
 		ret = dmar_set_interrupt(iommu);
@@ -4558,6 +4595,9 @@ static struct iommu_domain *intel_iommu_domain_alloc(unsigned type)
 		domain->geometry.force_aperture = true;
 
 		return domain;
+	case IOMMU_DOMAIN_KVA:
+		/* Use a global domain for shared KVA mapping */
+		return &kva_domain->domain;
 	case IOMMU_DOMAIN_IDENTITY:
 		return &si_domain->domain;
 	default:
@@ -4583,7 +4623,8 @@ is_aux_domain(struct device *dev, struct iommu_domain *domain)
 	struct device_domain_info *info = get_domain_info(dev);
 
 	return info && info->auxd_enabled &&
-			domain->type == IOMMU_DOMAIN_UNMANAGED;
+		(domain->type == IOMMU_DOMAIN_UNMANAGED ||
+			domain->type == IOMMU_DOMAIN_KVA);
 }
 
 static inline struct subdev_domain_info *
@@ -4693,8 +4734,8 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 	if (ret)
 		goto attach_failed;
 
-	/* Setup the PASID entry for mediated devices: */
-	if (domain_use_first_level(domain))
+	/* Setup the PASID entry for devices do DMA with the default PASID */
+	if (domain_use_first_level(domain) || domain->domain.type == IOMMU_DOMAIN_KVA)
 		ret = domain_setup_first_level(iommu, domain, dev,
 					       domain->default_pasid);
 	else
@@ -4761,6 +4802,10 @@ static int prepare_domain_attach_device(struct iommu_domain *domain,
 	if (!iommu)
 		return -ENODEV;
 
+	if (domain->type == IOMMU_DOMAIN_KVA) {
+		pr_info("TODO: KVA dom check if device can do full 64bit DMA");
+		return 0;
+	}
 	/* check if this iommu agaw is sufficient for max mapped address */
 	addr_width = agaw_to_width(iommu->agaw);
 	if (addr_width > cap_mgaw(iommu->cap))
@@ -5588,6 +5633,12 @@ static int intel_enable_pasid_dma(struct device *dev, u32 pasid, int mode)
 		ret = domain_setup_first_level(info->iommu, info->domain, dev,
 						pasid);
 		break;
+	case IOMMU_DMA_PASID_KVA:
+		/*
+		 * KVA mode should be handled in the aux domain attach where the default
+		 * PASID of the aux domain is used for setting up PASID FL.
+		 */
+		fallthrough;
 	default:
 		dev_err(dev, "Invalid PASID DMA mode %d", mode);
 		ret = -EINVAL;
-- 
2.25.1