]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/commitdiff
intel_agp: Use PCI DMA API correctly on chipsets new enough to have IOMMU
authorZhenyu Wang <zhenyu.z.wang@intel.com>
Mon, 27 Jul 2009 11:59:57 +0000 (12:59 +0100)
committerDavid Woodhouse <David.Woodhouse@intel.com>
Mon, 3 Aug 2009 08:04:50 +0000 (09:04 +0100)
When graphics dma remapping engine is active, we must fill
gart table with dma address from dmar engine, as now graphics
device access to graphics memory must go through dma remapping
table to get real physical address.

Add this support to all drivers which use intel_i915_insert_entries()

Signed-off-by: Zhenyu Wang <zhenyu.z.wang@intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
drivers/char/agp/intel-agp.c

index 21983456d672364d67a7ff458d03418c12c27f94..20fe82b99fdb85ac3f6f859761ecf146ebdf861b 100644 (file)
 #include <linux/agp_backend.h>
 #include "agp.h"
 
+/*
+ * If we have Intel graphics, we're not going to have anything other than
+ * an Intel IOMMU. So make the correct use of the PCI DMA API contingent
+ * on the Intel IOMMU support (CONFIG_DMAR).
+ * Only newer chipsets need to bother with this, of course.
+ */
+#ifdef CONFIG_DMAR
+#define USE_PCI_DMA_API 1
+#endif
+
 #define PCI_DEVICE_ID_INTEL_E7221_HB   0x2588
 #define PCI_DEVICE_ID_INTEL_E7221_IG   0x258a
 #define PCI_DEVICE_ID_INTEL_82946GZ_HB      0x2970
@@ -170,6 +180,131 @@ static struct _intel_private {
        int resource_valid;
 } intel_private;
 
+#ifdef USE_PCI_DMA_API
+static int intel_agp_map_page(void *addr, dma_addr_t *ret)
+{
+       *ret = pci_map_single(intel_private.pcidev, addr,
+                             PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+       if (pci_dma_mapping_error(intel_private.pcidev, *ret))
+               return -EINVAL;
+       return 0;
+}
+
+static void intel_agp_unmap_page(void *addr, dma_addr_t dma)
+{
+       pci_unmap_single(intel_private.pcidev, dma,
+                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+}
+
+static int intel_agp_map_memory(struct agp_memory *mem)
+{
+       struct scatterlist *sg;
+       int i;
+
+       DBG("try mapping %lu pages\n", (unsigned long)mem->page_count);
+
+       if ((mem->page_count * sizeof(*mem->sg_list)) < 2*PAGE_SIZE)
+               mem->sg_list = kcalloc(mem->page_count, sizeof(*mem->sg_list),
+                                      GFP_KERNEL);
+
+       if (mem->sg_list == NULL) {
+               mem->sg_list = vmalloc(mem->page_count * sizeof(*mem->sg_list));
+               mem->sg_vmalloc_flag = 1;
+       }
+
+       if (!mem->sg_list) {
+               mem->sg_vmalloc_flag = 0;
+               return -ENOMEM;
+       }
+       sg_init_table(mem->sg_list, mem->page_count);
+
+       sg = mem->sg_list;
+       for (i = 0 ; i < mem->page_count; i++, sg = sg_next(sg))
+               sg_set_page(sg, mem->pages[i], PAGE_SIZE, 0);
+
+       mem->num_sg = pci_map_sg(intel_private.pcidev, mem->sg_list,
+                                mem->page_count, PCI_DMA_BIDIRECTIONAL);
+       if (!mem->num_sg) {
+               if (mem->sg_vmalloc_flag)
+                       vfree(mem->sg_list);
+               else
+                       kfree(mem->sg_list);
+               mem->sg_list = NULL;
+               mem->sg_vmalloc_flag = 0;
+               return -ENOMEM;
+       }
+       return 0;
+}
+
+static void intel_agp_unmap_memory(struct agp_memory *mem)
+{
+       DBG("try unmapping %lu pages\n", (unsigned long)mem->page_count);
+
+       pci_unmap_sg(intel_private.pcidev, mem->sg_list,
+                    mem->page_count, PCI_DMA_BIDIRECTIONAL);
+       if (mem->sg_vmalloc_flag)
+               vfree(mem->sg_list);
+       else
+               kfree(mem->sg_list);
+       mem->sg_vmalloc_flag = 0;
+       mem->sg_list = NULL;
+       mem->num_sg = 0;
+}
+
+static void intel_agp_insert_sg_entries(struct agp_memory *mem,
+                                       off_t pg_start, int mask_type)
+{
+       struct scatterlist *sg;
+       int i, j;
+
+       j = pg_start;
+
+       WARN_ON(!mem->num_sg);
+
+       if (mem->num_sg == mem->page_count) {
+               for_each_sg(mem->sg_list, sg, mem->page_count, i) {
+                       writel(agp_bridge->driver->mask_memory(agp_bridge,
+                                       sg_dma_address(sg), mask_type),
+                                       intel_private.gtt+j);
+                       j++;
+               }
+       } else {
+               /* sg may merge pages, but we have to seperate
+                * per-page addr for GTT */
+               unsigned int len, m;
+
+               for_each_sg(mem->sg_list, sg, mem->num_sg, i) {
+                       len = sg_dma_len(sg) / PAGE_SIZE;
+                       for (m = 0; m < len; m++) {
+                               writel(agp_bridge->driver->mask_memory(agp_bridge,
+                                                                      sg_dma_address(sg) + m * PAGE_SIZE,
+                                                                      mask_type),
+                                      intel_private.gtt+j);
+                               j++;
+                       }
+               }
+       }
+       readl(intel_private.gtt+j-1);
+}
+
+#else
+
+static void intel_agp_insert_sg_entries(struct agp_memory *mem,
+                                       off_t pg_start, int mask_type)
+{
+       int i, j;
+
+       for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
+               writel(agp_bridge->driver->mask_memory(agp_bridge,
+                              phys_to_gart(page_to_phys(mem->pages[i])), mask_type),
+                      intel_private.gtt+j);
+       }
+
+       readl(intel_private.gtt+j-1);
+}
+
+#endif
+
 static int intel_i810_fetch_size(void)
 {
        u32 smram_miscc;
@@ -1003,9 +1138,13 @@ static int intel_i915_configure(void)
        writel(agp_bridge->gatt_bus_addr|I810_PGETBL_ENABLED, intel_private.registers+I810_PGETBL_CTL);
        readl(intel_private.registers+I810_PGETBL_CTL); /* PCI Posting. */
 
+#ifndef USE_PCI_DMA_API
+       agp_bridge->scratch_page_dma = agp_bridge->scratch_page;
+#endif
+
        if (agp_bridge->driver->needs_scratch_page) {
                for (i = intel_private.gtt_entries; i < current_size->num_entries; i++) {
-                       writel(agp_bridge->scratch_page, intel_private.gtt+i);
+                       writel(agp_bridge->scratch_page_dma, intel_private.gtt+i);
                }
                readl(intel_private.gtt+i-1);   /* PCI Posting. */
        }
@@ -1038,7 +1177,7 @@ static void intel_i915_chipset_flush(struct agp_bridge_data *bridge)
 static int intel_i915_insert_entries(struct agp_memory *mem, off_t pg_start,
                                     int type)
 {
-       int i, j, num_entries;
+       int num_entries;
        void *temp;
        int ret = -EINVAL;
        int mask_type;
@@ -1062,7 +1201,7 @@ static int intel_i915_insert_entries(struct agp_memory *mem, off_t pg_start,
        if ((pg_start + mem->page_count) > num_entries)
                goto out_err;
 
-       /* The i915 can't check the GTT for entries since its read only,
+       /* The i915 can't check the GTT for entries since it's read only;
         * depend on the caller to make the correct offset decisions.
         */
 
@@ -1078,14 +1217,7 @@ static int intel_i915_insert_entries(struct agp_memory *mem, off_t pg_start,
        if (!mem->is_flushed)
                global_cache_flush();
 
-       for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
-               writel(agp_bridge->driver->mask_memory(agp_bridge,
-                                      phys_to_gart(page_to_phys(mem->pages[i])),
-                                      mask_type),
-                      intel_private.gtt+j);
-       }
-
-       readl(intel_private.gtt+j-1);
+       intel_agp_insert_sg_entries(mem, pg_start, mask_type);
        agp_bridge->driver->tlb_flush(mem);
 
  out:
@@ -1110,7 +1242,7 @@ static int intel_i915_remove_entries(struct agp_memory *mem, off_t pg_start,
        }
 
        for (i = pg_start; i < (mem->page_count + pg_start); i++)
-               writel(agp_bridge->scratch_page, intel_private.gtt+i);
+               writel(agp_bridge->scratch_page_dma, intel_private.gtt+i);
 
        readl(intel_private.gtt+i-1);
 
@@ -2003,6 +2135,12 @@ static const struct agp_bridge_driver intel_915_driver = {
        .agp_destroy_pages      = agp_generic_destroy_pages,
        .agp_type_to_mask_type  = intel_i830_type_to_mask_type,
        .chipset_flush          = intel_i915_chipset_flush,
+#ifdef USE_PCI_DMA_API
+       .agp_map_page           = intel_agp_map_page,
+       .agp_unmap_page         = intel_agp_unmap_page,
+       .agp_map_memory         = intel_agp_map_memory,
+       .agp_unmap_memory       = intel_agp_unmap_memory,
+#endif
 };
 
 static const struct agp_bridge_driver intel_i965_driver = {
@@ -2031,6 +2169,12 @@ static const struct agp_bridge_driver intel_i965_driver = {
        .agp_destroy_pages      = agp_generic_destroy_pages,
        .agp_type_to_mask_type  = intel_i830_type_to_mask_type,
        .chipset_flush          = intel_i915_chipset_flush,
+#ifdef USE_PCI_DMA_API
+       .agp_map_page           = intel_agp_map_page,
+       .agp_unmap_page         = intel_agp_unmap_page,
+       .agp_map_memory         = intel_agp_map_memory,
+       .agp_unmap_memory       = intel_agp_unmap_memory,
+#endif
 };
 
 static const struct agp_bridge_driver intel_7505_driver = {
@@ -2085,6 +2229,12 @@ static const struct agp_bridge_driver intel_g33_driver = {
        .agp_destroy_pages      = agp_generic_destroy_pages,
        .agp_type_to_mask_type  = intel_i830_type_to_mask_type,
        .chipset_flush          = intel_i915_chipset_flush,
+#ifdef USE_PCI_DMA_API
+       .agp_map_page           = intel_agp_map_page,
+       .agp_unmap_page         = intel_agp_unmap_page,
+       .agp_map_memory         = intel_agp_map_memory,
+       .agp_unmap_memory       = intel_agp_unmap_memory,
+#endif
 };
 
 static int find_gmch(u16 device)