drm/i915/gvt: Map shadow page before using it in shadow page table
authorChuanxiao Dong <chuanxiao.dong@intel.com>
Thu, 9 Feb 2017 03:37:11 +0000 (11:37 +0800)
committerZhenyu Wang <zhenyuw@linux.intel.com>
Thu, 9 Feb 2017 06:17:14 +0000 (14:17 +0800)
MFN usually refers to "Machine Frame Number" in virtulization world.
Currently GVT-g populates the shadow PPGTT/GGTT page table with MFN
according to the translation between MFN and Guest PFN.

When host IOMMU is enabled, GPU DMA transactions go through the IOMMU,
GPU needs an IOVA<->MFN mapping to walk the shadow page table in host
main memory.

This patch will map a page in IOMMU page table before using it in shadow
page table and release the map when a shadow page is going to be freed.

Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Chuanxiao Dong <chuanxiao.dong@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
drivers/gpu/drm/i915/gvt/gtt.c

index 47dec4acf7ff12951eb592e2b115953e961f6bdf..28c92346db0e4e3615c2b5c484b4421eeb097e6e 100644 (file)
@@ -606,21 +606,33 @@ struct intel_vgpu_guest_page *intel_vgpu_find_guest_page(
 static inline int init_shadow_page(struct intel_vgpu *vgpu,
                struct intel_vgpu_shadow_page *p, int type)
 {
+       struct device *kdev = &vgpu->gvt->dev_priv->drm.pdev->dev;
+       dma_addr_t daddr;
+
+       daddr = dma_map_page(kdev, p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
+       if (dma_mapping_error(kdev, daddr)) {
+               gvt_err("fail to map dma addr\n");
+               return -EINVAL;
+       }
+
        p->vaddr = page_address(p->page);
        p->type = type;
 
        INIT_HLIST_NODE(&p->node);
 
-       p->mfn = intel_gvt_hypervisor_virt_to_mfn(p->vaddr);
-       if (p->mfn == INTEL_GVT_INVALID_ADDR)
-               return -EFAULT;
-
+       p->mfn = daddr >> GTT_PAGE_SHIFT;
        hash_add(vgpu->gtt.shadow_page_hash_table, &p->node, p->mfn);
        return 0;
 }
 
-static inline void clean_shadow_page(struct intel_vgpu_shadow_page *p)
+static inline void clean_shadow_page(struct intel_vgpu *vgpu,
+               struct intel_vgpu_shadow_page *p)
 {
+       struct device *kdev = &vgpu->gvt->dev_priv->drm.pdev->dev;
+
+       dma_unmap_page(kdev, p->mfn << GTT_PAGE_SHIFT, 4096,
+                       PCI_DMA_BIDIRECTIONAL);
+
        if (!hlist_unhashed(&p->node))
                hash_del(&p->node);
 }
@@ -670,7 +682,7 @@ static void ppgtt_free_shadow_page(struct intel_vgpu_ppgtt_spt *spt)
 {
        trace_spt_free(spt->vgpu->id, spt, spt->shadow_page.type);
 
-       clean_shadow_page(&spt->shadow_page);
+       clean_shadow_page(spt->vgpu, &spt->shadow_page);
        intel_vgpu_clean_guest_page(spt->vgpu, &spt->guest_page);
        list_del_init(&spt->post_shadow_list);
 
@@ -1875,8 +1887,9 @@ static int alloc_scratch_pages(struct intel_vgpu *vgpu,
        int page_entry_num = GTT_PAGE_SIZE >>
                                vgpu->gvt->device_info.gtt_entry_size_shift;
        void *scratch_pt;
-       unsigned long mfn;
        int i;
+       struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
+       dma_addr_t daddr;
 
        if (WARN_ON(type < GTT_TYPE_PPGTT_PTE_PT || type >= GTT_TYPE_MAX))
                return -EINVAL;
@@ -1887,16 +1900,18 @@ static int alloc_scratch_pages(struct intel_vgpu *vgpu,
                return -ENOMEM;
        }
 
-       mfn = intel_gvt_hypervisor_virt_to_mfn(scratch_pt);
-       if (mfn == INTEL_GVT_INVALID_ADDR) {
-               gvt_err("fail to translate vaddr:0x%lx\n", (unsigned long)scratch_pt);
-               free_page((unsigned long)scratch_pt);
-               return -EFAULT;
+       daddr = dma_map_page(dev, virt_to_page(scratch_pt), 0,
+                       4096, PCI_DMA_BIDIRECTIONAL);
+       if (dma_mapping_error(dev, daddr)) {
+               gvt_err("fail to dmamap scratch_pt\n");
+               __free_page(virt_to_page(scratch_pt));
+               return -ENOMEM;
        }
-       gtt->scratch_pt[type].page_mfn = mfn;
+       gtt->scratch_pt[type].page_mfn =
+               (unsigned long)(daddr >> GTT_PAGE_SHIFT);
        gtt->scratch_pt[type].page = virt_to_page(scratch_pt);
        gvt_dbg_mm("vgpu%d create scratch_pt: type %d mfn=0x%lx\n",
-                       vgpu->id, type, mfn);
+                       vgpu->id, type, gtt->scratch_pt[type].page_mfn);
 
        /* Build the tree by full filled the scratch pt with the entries which
         * point to the next level scratch pt or scratch page. The
@@ -1930,9 +1945,14 @@ static int alloc_scratch_pages(struct intel_vgpu *vgpu,
 static int release_scratch_page_tree(struct intel_vgpu *vgpu)
 {
        int i;
+       struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
+       dma_addr_t daddr;
 
        for (i = GTT_TYPE_PPGTT_PTE_PT; i < GTT_TYPE_MAX; i++) {
                if (vgpu->gtt.scratch_pt[i].page != NULL) {
+                       daddr = (dma_addr_t)(vgpu->gtt.scratch_pt[i].page_mfn <<
+                                       GTT_PAGE_SHIFT);
+                       dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
                        __free_page(vgpu->gtt.scratch_pt[i].page);
                        vgpu->gtt.scratch_pt[i].page = NULL;
                        vgpu->gtt.scratch_pt[i].page_mfn = 0;
@@ -2192,6 +2212,8 @@ int intel_gvt_init_gtt(struct intel_gvt *gvt)
 {
        int ret;
        void *page;
+       struct device *dev = &gvt->dev_priv->drm.pdev->dev;
+       dma_addr_t daddr;
 
        gvt_dbg_core("init gtt\n");
 
@@ -2209,14 +2231,16 @@ int intel_gvt_init_gtt(struct intel_gvt *gvt)
                gvt_err("fail to allocate scratch ggtt page\n");
                return -ENOMEM;
        }
-       gvt->gtt.scratch_ggtt_page = virt_to_page(page);
 
-       gvt->gtt.scratch_ggtt_mfn = intel_gvt_hypervisor_virt_to_mfn(page);
-       if (gvt->gtt.scratch_ggtt_mfn == INTEL_GVT_INVALID_ADDR) {
-               gvt_err("fail to translate scratch ggtt page\n");
-               __free_page(gvt->gtt.scratch_ggtt_page);
-               return -EFAULT;
+       daddr = dma_map_page(dev, virt_to_page(page), 0,
+                       4096, PCI_DMA_BIDIRECTIONAL);
+       if (dma_mapping_error(dev, daddr)) {
+               gvt_err("fail to dmamap scratch ggtt page\n");
+               __free_page(virt_to_page(page));
+               return -ENOMEM;
        }
+       gvt->gtt.scratch_ggtt_page = virt_to_page(page);
+       gvt->gtt.scratch_ggtt_mfn = (unsigned long)(daddr >> GTT_PAGE_SHIFT);
 
        if (enable_out_of_sync) {
                ret = setup_spt_oos(gvt);
@@ -2239,6 +2263,12 @@ int intel_gvt_init_gtt(struct intel_gvt *gvt)
  */
 void intel_gvt_clean_gtt(struct intel_gvt *gvt)
 {
+       struct device *dev = &gvt->dev_priv->drm.pdev->dev;
+       dma_addr_t daddr = (dma_addr_t)(gvt->gtt.scratch_ggtt_mfn <<
+                                       GTT_PAGE_SHIFT);
+
+       dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
+
        __free_page(gvt->gtt.scratch_ggtt_page);
 
        if (enable_out_of_sync)