From 19943b0e30b05d42e494ae6fef78156ebc8c637e Mon Sep 17 00:00:00 2001 From: David Woodhouse Date: Tue, 4 Aug 2009 16:19:20 +0100 Subject: [PATCH] intel-iommu: Unify hardware and software passthrough support 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 --- arch/x86/kernel/pci-swiotlb.c | 5 +- drivers/pci/intel-iommu.c | 174 +++++++++++++++------------------- 2 files changed, 76 insertions(+), 103 deletions(-) diff --git a/arch/x86/kernel/pci-swiotlb.c b/arch/x86/kernel/pci-swiotlb.c index 6af96ee44200..1e66b18f45cb 100644 --- a/arch/x86/kernel/pci-swiotlb.c +++ b/arch/x86/kernel/pci-swiotlb.c @@ -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; diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c index 097d5da2fae1..147b3b960b61 100644 --- a/drivers/pci/intel-iommu.c +++ b/drivers/pci/intel-iommu.c @@ -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(); -- 2.30.2