intel-iommu: Unify hardware and software passthrough support
authorDavid Woodhouse <David.Woodhouse@intel.com>
Tue, 4 Aug 2009 15:19:20 +0000 (16:19 +0100)
committerDavid Woodhouse <David.Woodhouse@intel.com>
Tue, 4 Aug 2009 15:19:23 +0000 (16:19 +0100)
This makes the hardware passthrough mode work a lot more like the
software version, so that the behaviour of a kernel with 'iommu=pt'
is the same whether the hardware supports passthrough or not.

In particular:
 - We use a single si_domain for the pass-through devices.
 - 32-bit devices can be taken out of the pass-through domain so that
   they don't have to use swiotlb.
 - Devices will work again after being removed from a KVM guest.
 - A potential oops on OOM (in init_context_pass_through()) is fixed.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
arch/x86/kernel/pci-swiotlb.c
drivers/pci/intel-iommu.c

index 6af96ee4420080775f5afa466f2bad335434d935..1e66b18f45cb6da71321a0cc5f66cf9688e980f7 100644 (file)
@@ -71,9 +71,8 @@ void __init pci_swiotlb_init(void)
 {
        /* don't initialize swiotlb if iommu=off (no_iommu=1) */
 #ifdef CONFIG_X86_64
-       if ((!iommu_detected && !no_iommu && max_pfn > MAX_DMA32_PFN) ||
-               iommu_pass_through)
-              swiotlb = 1;
+       if ((!iommu_detected && !no_iommu && max_pfn > MAX_DMA32_PFN))
+               swiotlb = 1;
 #endif
        if (swiotlb_force)
                swiotlb = 1;
index 097d5da2fae17e4d8ac493ed7b96d3dfc136f161..147b3b960b619c6e6c9c21f155729f5c473f08b3 100644 (file)
@@ -251,7 +251,8 @@ static inline int first_pte_in_page(struct dma_pte *pte)
  *     2. It maps to each iommu if successful.
  *     3. Each iommu mapps to this domain if successful.
  */
-struct dmar_domain *si_domain;
+static struct dmar_domain *si_domain;
+static int hw_pass_through = 1;
 
 /* devices under the same p2p bridge are owned in one domain */
 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
@@ -1948,14 +1949,24 @@ static int iommu_prepare_identity_map(struct pci_dev *pdev,
        struct dmar_domain *domain;
        int ret;
 
-       printk(KERN_INFO
-              "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
-              pci_name(pdev), start, end);
-
        domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
        if (!domain)
                return -ENOMEM;
 
+       /* For _hardware_ passthrough, don't bother. But for software
+          passthrough, we do it anyway -- it may indicate a memory
+          range which is reserved in E820, so which didn't get set
+          up to start with in si_domain */
+       if (domain == si_domain && hw_pass_through) {
+               printk("Ignoring identity map for HW passthrough device %s [0x%Lx - 0x%Lx]\n",
+                      pci_name(pdev), start, end);
+               return 0;
+       }
+
+       printk(KERN_INFO
+              "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
+              pci_name(pdev), start, end);
+
        ret = iommu_domain_identity_map(domain, start, end);
        if (ret)
                goto error;
@@ -2006,23 +2017,6 @@ static inline void iommu_prepare_isa(void)
 }
 #endif /* !CONFIG_DMAR_FLPY_WA */
 
-/* Initialize each context entry as pass through.*/
-static int __init init_context_pass_through(void)
-{
-       struct pci_dev *pdev = NULL;
-       struct dmar_domain *domain;
-       int ret;
-
-       for_each_pci_dev(pdev) {
-               domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
-               ret = domain_context_mapping(domain, pdev,
-                                            CONTEXT_TT_PASS_THROUGH);
-               if (ret)
-                       return ret;
-       }
-       return 0;
-}
-
 static int md_domain_init(struct dmar_domain *domain, int guest_width);
 
 static int __init si_domain_work_fn(unsigned long start_pfn,
@@ -2037,7 +2031,7 @@ static int __init si_domain_work_fn(unsigned long start_pfn,
 
 }
 
-static int si_domain_init(void)
+static int si_domain_init(int hw)
 {
        struct dmar_drhd_unit *drhd;
        struct intel_iommu *iommu;
@@ -2064,6 +2058,9 @@ static int si_domain_init(void)
 
        si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
 
+       if (hw)
+               return 0;
+
        for_each_online_node(nid) {
                work_with_active_regions(nid, si_domain_work_fn, &ret);
                if (ret)
@@ -2155,24 +2152,26 @@ static int iommu_should_identity_map(struct pci_dev *pdev, int startup)
        return 1;
 }
 
-static int iommu_prepare_static_identity_mapping(void)
+static int iommu_prepare_static_identity_mapping(int hw)
 {
        struct pci_dev *pdev = NULL;
        int ret;
 
-       ret = si_domain_init();
+       ret = si_domain_init(hw);
        if (ret)
                return -EFAULT;
 
        for_each_pci_dev(pdev) {
                if (iommu_should_identity_map(pdev, 1)) {
-                       printk(KERN_INFO "IOMMU: identity mapping for device %s\n",
-                              pci_name(pdev));
+                       printk(KERN_INFO "IOMMU: %s identity mapping for device %s\n",
+                              hw ? "hardware" : "software", pci_name(pdev));
 
                        ret = domain_context_mapping(si_domain, pdev,
+                                                    hw ? CONTEXT_TT_PASS_THROUGH :
                                                     CONTEXT_TT_MULTI_LEVEL);
                        if (ret)
                                return ret;
+
                        ret = domain_add_dev_info(si_domain, pdev);
                        if (ret)
                                return ret;
@@ -2189,14 +2188,6 @@ int __init init_dmars(void)
        struct pci_dev *pdev;
        struct intel_iommu *iommu;
        int i, ret;
-       int pass_through = 1;
-
-       /*
-        * In case pass through can not be enabled, iommu tries to use identity
-        * mapping.
-        */
-       if (iommu_pass_through)
-               iommu_identity_mapping = 1;
 
        /*
         * for each drhd
@@ -2250,14 +2241,8 @@ int __init init_dmars(void)
                        goto error;
                }
                if (!ecap_pass_through(iommu->ecap))
-                       pass_through = 0;
+                       hw_pass_through = 0;
        }
-       if (iommu_pass_through)
-               if (!pass_through) {
-                       printk(KERN_INFO
-                              "Pass Through is not supported by hardware.\n");
-                       iommu_pass_through = 0;
-               }
 
        /*
         * Start from the sane iommu hardware state.
@@ -2312,64 +2297,57 @@ int __init init_dmars(void)
                }
        }
 
+       if (iommu_pass_through)
+               iommu_identity_mapping = 1;
+#ifdef CONFIG_DMAR_BROKEN_GFX_WA
+       else
+               iommu_identity_mapping = 2;
+#endif
        /*
-        * If pass through is set and enabled, context entries of all pci
-        * devices are intialized by pass through translation type.
+        * If pass through is not set or not enabled, setup context entries for
+        * identity mappings for rmrr, gfx, and isa and may fall back to static
+        * identity mapping if iommu_identity_mapping is set.
         */
-       if (iommu_pass_through) {
-               ret = init_context_pass_through();
+       if (iommu_identity_mapping) {
+               ret = iommu_prepare_static_identity_mapping(hw_pass_through);
                if (ret) {
-                       printk(KERN_ERR "IOMMU: Pass through init failed.\n");
-                       iommu_pass_through = 0;
+                       printk(KERN_CRIT "Failed to setup IOMMU pass-through\n");
+                       goto error;
                }
        }
-
        /*
-        * If pass through is not set or not enabled, setup context entries for
-        * identity mappings for rmrr, gfx, and isa and may fall back to static
-        * identity mapping if iommu_identity_mapping is set.
+        * For each rmrr
+        *   for each dev attached to rmrr
+        *   do
+        *     locate drhd for dev, alloc domain for dev
+        *     allocate free domain
+        *     allocate page table entries for rmrr
+        *     if context not allocated for bus
+        *           allocate and init context
+        *           set present in root table for this bus
+        *     init context with domain, translation etc
+        *    endfor
+        * endfor
         */
-       if (!iommu_pass_through) {
-#ifdef CONFIG_DMAR_BROKEN_GFX_WA
-               if (!iommu_identity_mapping)
-                       iommu_identity_mapping = 2;
-#endif
-               if (iommu_identity_mapping)
-                       iommu_prepare_static_identity_mapping();
-               /*
-                * For each rmrr
-                *   for each dev attached to rmrr
-                *   do
-                *     locate drhd for dev, alloc domain for dev
-                *     allocate free domain
-                *     allocate page table entries for rmrr
-                *     if context not allocated for bus
-                *           allocate and init context
-                *           set present in root table for this bus
-                *     init context with domain, translation etc
-                *    endfor
-                * endfor
-                */
-               printk(KERN_INFO "IOMMU: Setting RMRR:\n");
-               for_each_rmrr_units(rmrr) {
-                       for (i = 0; i < rmrr->devices_cnt; i++) {
-                               pdev = rmrr->devices[i];
-                               /*
-                                * some BIOS lists non-exist devices in DMAR
-                                * table.
-                                */
-                               if (!pdev)
-                                       continue;
-                               ret = iommu_prepare_rmrr_dev(rmrr, pdev);
-                               if (ret)
-                                       printk(KERN_ERR
-                                "IOMMU: mapping reserved region failed\n");
-                       }
+       printk(KERN_INFO "IOMMU: Setting RMRR:\n");
+       for_each_rmrr_units(rmrr) {
+               for (i = 0; i < rmrr->devices_cnt; i++) {
+                       pdev = rmrr->devices[i];
+                       /*
+                        * some BIOS lists non-exist devices in DMAR
+                        * table.
+                        */
+                       if (!pdev)
+                               continue;
+                       ret = iommu_prepare_rmrr_dev(rmrr, pdev);
+                       if (ret)
+                               printk(KERN_ERR
+                                      "IOMMU: mapping reserved region failed\n");
                }
-
-               iommu_prepare_isa();
        }
 
+       iommu_prepare_isa();
+
        /*
         * for each drhd
         *   enable fault log
@@ -2536,7 +2514,10 @@ static int iommu_no_mapping(struct device *dev)
                        ret = domain_add_dev_info(si_domain, pdev);
                        if (ret)
                                return 0;
-                       ret = domain_context_mapping(si_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
+                       ret = domain_context_mapping(si_domain, pdev,
+                                                    hw_pass_through ?
+                                                    CONTEXT_TT_PASS_THROUGH :
+                                                    CONTEXT_TT_MULTI_LEVEL);
                        if (!ret) {
                                printk(KERN_INFO "64bit %s uses identity mapping\n",
                                       pci_name(pdev));
@@ -3202,7 +3183,7 @@ int __init intel_iommu_init(void)
         * Check the need for DMA-remapping initialization now.
         * Above initialization will also be used by Interrupt-remapping.
         */
-       if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
+       if (no_iommu || swiotlb || dmar_disabled)
                return -ENODEV;
 
        iommu_init_mempool();
@@ -3222,14 +3203,7 @@ int __init intel_iommu_init(void)
 
        init_timer(&unmap_timer);
        force_iommu = 1;
-
-       if (!iommu_pass_through) {
-               printk(KERN_INFO
-                      "Multi-level page-table translation for DMAR.\n");
-               dma_ops = &intel_dma_ops;
-       } else
-               printk(KERN_INFO
-                      "DMAR: Pass through translation for DMAR.\n");
+       dma_ops = &intel_dma_ops;
 
        init_iommu_sysfs();