drm/i915: Make the physical object coherent with GTT
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 4 Nov 2014 12:51:40 +0000 (04:51 -0800)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 14 Nov 2014 09:29:18 +0000 (10:29 +0100)
Currently objects for which the hardware needs a contiguous physical
address are allocated a shadow backing storage to satisfy the contraint.
This shadow buffer is not wired into the normal obj->pages and so the
physical object is incoherent with accesses via the GPU, GTT and CPU. By
setting up the appropriate scatter-gather table, we can allow userspace
to access the physical object via either a GTT mmaping of or by rendering
into the GEM bo. However, keeping the CPU mmap of the shmemfs backing
storage coherent with the contiguous shadow is not yet possible.
Fortuituously, CPU mmaps of objects requiring physical addresses are not
expected to be coherent anyway.

This allows the physical constraint of the GEM object to be transparent
to userspace and allow it to efficiently render into or update them via
the GTT and GPU.

v2: Fix leak of pci handle spotted by Ville
v3: Remove the now duplicate call to detach_phys_object during free.
v4: Wait for rendering before pwrite. As this patch makes it possible to
render into the phys object, we should make it correct as well!

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_dma.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
include/uapi/drm/i915_drm.h

index 9a7353302b3f9a09e44a63c65dfbb2c68a9f240d..5dc37f0233b2a54710a37b8945ddff31040cef7a 100644 (file)
@@ -1027,6 +1027,9 @@ static int i915_getparam(struct drm_device *dev, void *data,
        case I915_PARAM_CMD_PARSER_VERSION:
                value = i915_cmd_parser_get_version();
                break;
+       case I915_PARAM_HAS_COHERENT_PHYS_GTT:
+               value = 1;
+               break;
        default:
                DRM_DEBUG("Unknown parameter %d\n", param->param);
                return -EINVAL;
index f830596faa9eb0f522aaf7ee3bef57328aa212d9..3df9ef32d0117780d0eb080910946483ddb0d7aa 100644 (file)
@@ -1957,10 +1957,10 @@ struct drm_i915_gem_object {
        unsigned long user_pin_count;
        struct drm_file *pin_filp;
 
-       /** for phy allocated objects */
-       struct drm_dma_handle *phys_handle;
-
        union {
+               /** for phy allocated objects */
+               struct drm_dma_handle *phys_handle;
+
                struct i915_gem_userptr {
                        uintptr_t ptr;
                        unsigned read_only :1;
index 3e0cabe9b54498caa60428652ce6f9245f7d9535..86cf428b6c4e251bcdc7c4351cc1e95753789d84 100644 (file)
@@ -208,40 +208,137 @@ i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
        return 0;
 }
 
-static void i915_gem_object_detach_phys(struct drm_i915_gem_object *obj)
+static int
+i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
 {
-       drm_dma_handle_t *phys = obj->phys_handle;
+       struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
+       char *vaddr = obj->phys_handle->vaddr;
+       struct sg_table *st;
+       struct scatterlist *sg;
+       int i;
 
-       if (!phys)
-               return;
+       if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
+               return -EINVAL;
+
+       for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
+               struct page *page;
+               char *src;
+
+               page = shmem_read_mapping_page(mapping, i);
+               if (IS_ERR(page))
+                       return PTR_ERR(page);
+
+               src = kmap_atomic(page);
+               memcpy(vaddr, src, PAGE_SIZE);
+               drm_clflush_virt_range(vaddr, PAGE_SIZE);
+               kunmap_atomic(src);
+
+               page_cache_release(page);
+               vaddr += PAGE_SIZE;
+       }
+
+       i915_gem_chipset_flush(obj->base.dev);
+
+       st = kmalloc(sizeof(*st), GFP_KERNEL);
+       if (st == NULL)
+               return -ENOMEM;
+
+       if (sg_alloc_table(st, 1, GFP_KERNEL)) {
+               kfree(st);
+               return -ENOMEM;
+       }
+
+       sg = st->sgl;
+       sg->offset = 0;
+       sg->length = obj->base.size;
 
-       if (obj->madv == I915_MADV_WILLNEED) {
+       sg_dma_address(sg) = obj->phys_handle->busaddr;
+       sg_dma_len(sg) = obj->base.size;
+
+       obj->pages = st;
+       obj->has_dma_mapping = true;
+       return 0;
+}
+
+static void
+i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
+{
+       int ret;
+
+       BUG_ON(obj->madv == __I915_MADV_PURGED);
+
+       ret = i915_gem_object_set_to_cpu_domain(obj, true);
+       if (ret) {
+               /* In the event of a disaster, abandon all caches and
+                * hope for the best.
+                */
+               WARN_ON(ret != -EIO);
+               obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       }
+
+       if (obj->madv == I915_MADV_DONTNEED)
+               obj->dirty = 0;
+
+       if (obj->dirty) {
                struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
-               char *vaddr = phys->vaddr;
+               char *vaddr = obj->phys_handle->vaddr;
                int i;
 
                for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
-                       struct page *page = shmem_read_mapping_page(mapping, i);
-                       if (!IS_ERR(page)) {
-                               char *dst = kmap_atomic(page);
-                               memcpy(dst, vaddr, PAGE_SIZE);
-                               drm_clflush_virt_range(dst, PAGE_SIZE);
-                               kunmap_atomic(dst);
-
-                               set_page_dirty(page);
+                       struct page *page;
+                       char *dst;
+
+                       page = shmem_read_mapping_page(mapping, i);
+                       if (IS_ERR(page))
+                               continue;
+
+                       dst = kmap_atomic(page);
+                       drm_clflush_virt_range(vaddr, PAGE_SIZE);
+                       memcpy(dst, vaddr, PAGE_SIZE);
+                       kunmap_atomic(dst);
+
+                       set_page_dirty(page);
+                       if (obj->madv == I915_MADV_WILLNEED)
                                mark_page_accessed(page);
-                               page_cache_release(page);
-                       }
+                       page_cache_release(page);
                        vaddr += PAGE_SIZE;
                }
-               i915_gem_chipset_flush(obj->base.dev);
+               obj->dirty = 0;
        }
 
-#ifdef CONFIG_X86
-       set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE);
-#endif
-       drm_pci_free(obj->base.dev, phys);
-       obj->phys_handle = NULL;
+       sg_free_table(obj->pages);
+       kfree(obj->pages);
+
+       obj->has_dma_mapping = false;
+}
+
+static void
+i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
+{
+       drm_pci_free(obj->base.dev, obj->phys_handle);
+}
+
+static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
+       .get_pages = i915_gem_object_get_pages_phys,
+       .put_pages = i915_gem_object_put_pages_phys,
+       .release = i915_gem_object_release_phys,
+};
+
+static int
+drop_pages(struct drm_i915_gem_object *obj)
+{
+       struct i915_vma *vma, *next;
+       int ret;
+
+       drm_gem_object_reference(&obj->base);
+       list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link)
+               if (i915_vma_unbind(vma))
+                       break;
+
+       ret = i915_gem_object_put_pages(obj);
+       drm_gem_object_unreference(&obj->base);
+
+       return ret;
 }
 
 int
@@ -249,9 +346,7 @@ i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
                            int align)
 {
        drm_dma_handle_t *phys;
-       struct address_space *mapping;
-       char *vaddr;
-       int i;
+       int ret;
 
        if (obj->phys_handle) {
                if ((unsigned long)obj->phys_handle->vaddr & (align -1))
@@ -266,41 +361,19 @@ i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
        if (obj->base.filp == NULL)
                return -EINVAL;
 
+       ret = drop_pages(obj);
+       if (ret)
+               return ret;
+
        /* create a new object */
        phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
        if (!phys)
                return -ENOMEM;
 
-       vaddr = phys->vaddr;
-#ifdef CONFIG_X86
-       set_memory_wc((unsigned long)vaddr, phys->size / PAGE_SIZE);
-#endif
-       mapping = file_inode(obj->base.filp)->i_mapping;
-       for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
-               struct page *page;
-               char *src;
-
-               page = shmem_read_mapping_page(mapping, i);
-               if (IS_ERR(page)) {
-#ifdef CONFIG_X86
-                       set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE);
-#endif
-                       drm_pci_free(obj->base.dev, phys);
-                       return PTR_ERR(page);
-               }
-
-               src = kmap_atomic(page);
-               memcpy(vaddr, src, PAGE_SIZE);
-               kunmap_atomic(src);
-
-               mark_page_accessed(page);
-               page_cache_release(page);
-
-               vaddr += PAGE_SIZE;
-       }
-
        obj->phys_handle = phys;
-       return 0;
+       obj->ops = &i915_gem_phys_ops;
+
+       return i915_gem_object_get_pages(obj);
 }
 
 static int
@@ -311,6 +384,14 @@ i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
        struct drm_device *dev = obj->base.dev;
        void *vaddr = obj->phys_handle->vaddr + args->offset;
        char __user *user_data = to_user_ptr(args->data_ptr);
+       int ret;
+
+       /* We manually control the domain here and pretend that it
+        * remains coherent i.e. in the GTT domain, like shmem_pwrite.
+        */
+       ret = i915_gem_object_wait_rendering(obj, false);
+       if (ret)
+               return ret;
 
        if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
                unsigned long unwritten;
@@ -326,6 +407,7 @@ i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
                        return -EFAULT;
        }
 
+       drm_clflush_virt_range(vaddr, args->size);
        i915_gem_chipset_flush(dev);
        return 0;
 }
@@ -1046,11 +1128,6 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
         * pread/pwrite currently are reading and writing from the CPU
         * perspective, requiring manual detiling by the client.
         */
-       if (obj->phys_handle) {
-               ret = i915_gem_phys_pwrite(obj, args, file);
-               goto out;
-       }
-
        if (obj->tiling_mode == I915_TILING_NONE &&
            obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
            cpu_write_needs_clflush(obj)) {
@@ -1060,8 +1137,12 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
                 * textures). Fallback to the shmem path in that case. */
        }
 
-       if (ret == -EFAULT || ret == -ENOSPC)
-               ret = i915_gem_shmem_pwrite(dev, obj, args, file);
+       if (ret == -EFAULT || ret == -ENOSPC) {
+               if (obj->phys_handle)
+                       ret = i915_gem_phys_pwrite(obj, args, file);
+               else
+                       ret = i915_gem_shmem_pwrite(dev, obj, args, file);
+       }
 
 out:
        drm_gem_object_unreference(&obj->base);
@@ -3509,7 +3590,7 @@ i915_gem_clflush_object(struct drm_i915_gem_object *obj,
         * Stolen memory is always coherent with the GPU as it is explicitly
         * marked as wc by the system, or the system is cache-coherent.
         */
-       if (obj->stolen)
+       if (obj->stolen || obj->phys_handle)
                return false;
 
        /* If the GPU is snooping the contents of the CPU cache,
@@ -4471,8 +4552,6 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj)
                }
        }
 
-       i915_gem_object_detach_phys(obj);
-
        /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
         * before progressing. */
        if (obj->stolen)
index 2ec0efcaa7197e9fe10de889acc6e54bb0abf9b6..250262265ee31c3099b00850a29c6729889dcee9 100644 (file)
@@ -340,6 +340,7 @@ typedef struct drm_i915_irq_wait {
 #define I915_PARAM_HAS_EXEC_HANDLE_LUT   26
 #define I915_PARAM_HAS_WT               27
 #define I915_PARAM_CMD_PARSER_VERSION   28
+#define I915_PARAM_HAS_COHERENT_PHYS_GTT 29
 
 typedef struct drm_i915_getparam {
        int param;