From 4e1b20246e9cace556cd8095520e92b287a70748 Mon Sep 17 00:00:00 2001 From: Laurentiu Tudor Date: Tue, 26 Jul 2016 16:38:18 +0300 Subject: [PATCH] arm/arm64: KVM: drop qman mmio cacheable mapping hack Instead of hardcoding checks for qman cacheable mmio region physical addresses extract mapping information from the user-space mapping. The involves several steps; - get access to a pte part of the user-space mapping by using get_locked_pte() / pte_unmap_unlock() apis - extract memtype (normal / device), shareability from the pte - convert to S2 translation bits in newly added function stage1_to_stage2_pgprot() - finish making the s2 translation with the obtained bits Another explored option was using vm_area_struct::vm_page_prot which is set in vfio-mc mmap code to the correct page bits. However, experiments show that these bits are later altered in the generic mmap code (e.g. the shareability bit is always set on arm64). The only place where the original bits can still be found is the user-space mapping, using the method described above. Signed-off-by: Laurentiu Tudor [Bharat - Fixed mem_type check issue] [changed "ifdef ARM64" to CONFIG_ARM64] Signed-off-by: Bharat Bhushan [Ioana - added a sanity check for hugepages] Signed-off-by: Ioana Ciornei [Fixed format issues] Signed-off-by: Diana Craciun --- virt/kvm/arm/mmu.c | 55 ++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 53 insertions(+), 2 deletions(-) diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c index b1fdb926bc6f..f2c50685bd3d 100644 --- a/virt/kvm/arm/mmu.c +++ b/virt/kvm/arm/mmu.c @@ -1370,6 +1370,30 @@ out: return ret; } +#ifdef CONFIG_ARM64 +static pgprot_t stage1_to_stage2_pgprot(pgprot_t prot) +{ + switch (pgprot_val(prot) & PTE_ATTRINDX_MASK) { + case PTE_ATTRINDX(MT_DEVICE_nGnRE): + case PTE_ATTRINDX(MT_DEVICE_nGnRnE): + case PTE_ATTRINDX(MT_DEVICE_GRE): + return PAGE_S2_DEVICE; + case PTE_ATTRINDX(MT_NORMAL_NC): + case PTE_ATTRINDX(MT_NORMAL): + return (pgprot_val(prot) & PTE_SHARED) + ? PAGE_S2 + : PAGE_S2_NS; + } + + return PAGE_S2_DEVICE; +} +#else +static pgprot_t stage1_to_stage2_pgprot(pgprot_t prot) +{ + return PAGE_S2_DEVICE; +} +#endif + static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap) { kvm_pfn_t pfn = *pfnp; @@ -1713,8 +1737,23 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, * 3 levels, i.e, PMD is not folded. */ if (vma_pagesize == PMD_SIZE || - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) { gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; + } else { + if (!is_vm_hugetlb_page(vma)) { + pte_t *pte; + spinlock_t *ptl; + pgprot_t prot; + + pte = get_locked_pte(current->mm, memslot->userspace_addr, &ptl); + prot = stage1_to_stage2_pgprot(__pgprot(pte_val(*pte))); + pte_unmap_unlock(pte, ptl); +#ifdef CONFIG_ARM64 + if (pgprot_val(prot) == pgprot_val(PAGE_S2_NS)) + mem_type = PAGE_S2_NS; +#endif + } + } up_read(¤t->mm->mmap_sem); /* We need minimum second+third level pages */ @@ -1743,6 +1782,11 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, if (is_error_noslot_pfn(pfn)) return -EFAULT; +#ifdef CONFIG_ARM64 + if (pgprot_val(mem_type) == pgprot_val(PAGE_S2_NS)) { + flags |= KVM_S2PTE_FLAG_IS_IOMAP; + } else +#endif if (kvm_is_device_pfn(pfn)) { mem_type = PAGE_S2_DEVICE; flags |= KVM_S2PTE_FLAG_IS_IOMAP; @@ -2322,6 +2366,9 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm, gpa_t gpa = mem->guest_phys_addr + (vm_start - mem->userspace_addr); phys_addr_t pa; + pgprot_t prot; + pte_t *pte; + spinlock_t *ptl; pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT; pa += vm_start - vma->vm_start; @@ -2332,9 +2379,13 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm, goto out; } + pte = get_locked_pte(current->mm, mem->userspace_addr, &ptl); + prot = stage1_to_stage2_pgprot(__pgprot(pte_val(*pte))); + pte_unmap_unlock(pte, ptl); + ret = kvm_phys_addr_ioremap(kvm, gpa, pa, vm_end - vm_start, - writable, PAGE_S2_DEVICE); + writable, prot); if (ret) break; }