Merge tag 'drm-intel-next-2016-10-24' of git://anongit.freedesktop.org/drm-intel...
[linux-2.6-block.git] / drivers / gpu / drm / i915 / i915_gem_execbuffer.c
index 7adb4c77cc7f449698f0207d5f51cbf037cff179..e52affdcc125d5585cc3fa45bd25c60aed540084 100644 (file)
@@ -370,8 +370,7 @@ static void reloc_cache_fini(struct reloc_cache *cache)
 
                        ggtt->base.clear_range(&ggtt->base,
                                               cache->node.start,
-                                              cache->node.size,
-                                              true);
+                                              cache->node.size);
                        drm_mm_remove_node(&cache->node);
                } else {
                        i915_vma_unpin((struct i915_vma *)cache->node.mm);
@@ -429,7 +428,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
        }
 
        if (cache->vaddr) {
-               io_mapping_unmap_atomic(unmask_page(cache->vaddr));
+               io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
        } else {
                struct i915_vma *vma;
                int ret;
@@ -474,7 +473,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
                offset += page << PAGE_SHIFT;
        }
 
-       vaddr = io_mapping_map_atomic_wc(&cache->i915->ggtt.mappable, offset);
+       vaddr = (void __force *) io_mapping_map_atomic_wc(&cache->i915->ggtt.mappable, offset);
        cache->page = page;
        cache->vaddr = (unsigned long)vaddr;
 
@@ -552,27 +551,13 @@ repeat:
        return 0;
 }
 
-static bool object_is_idle(struct drm_i915_gem_object *obj)
-{
-       unsigned long active = i915_gem_object_get_active(obj);
-       int idx;
-
-       for_each_active(active, idx) {
-               if (!i915_gem_active_is_idle(&obj->last_read[idx],
-                                            &obj->base.dev->struct_mutex))
-                       return false;
-       }
-
-       return true;
-}
-
 static int
 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
                                   struct eb_vmas *eb,
                                   struct drm_i915_gem_relocation_entry *reloc,
                                   struct reloc_cache *cache)
 {
-       struct drm_device *dev = obj->base.dev;
+       struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
        struct drm_gem_object *target_obj;
        struct drm_i915_gem_object *target_i915_obj;
        struct i915_vma *target_vma;
@@ -591,7 +576,7 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
        /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
         * pipe_control writes because the gpu doesn't properly redirect them
         * through the ppgtt for non_secure batchbuffers. */
-       if (unlikely(IS_GEN6(dev) &&
+       if (unlikely(IS_GEN6(dev_priv) &&
            reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
                ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
                                    PIN_GLOBAL);
@@ -649,10 +634,6 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
                return -EINVAL;
        }
 
-       /* We can't wait for rendering with pagefaults disabled */
-       if (pagefault_disabled() && !object_is_idle(obj))
-               return -EFAULT;
-
        ret = relocate_entry(obj, reloc, cache, target_offset);
        if (ret)
                return ret;
@@ -679,12 +660,23 @@ i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
        remain = entry->relocation_count;
        while (remain) {
                struct drm_i915_gem_relocation_entry *r = stack_reloc;
-               int count = remain;
-               if (count > ARRAY_SIZE(stack_reloc))
-                       count = ARRAY_SIZE(stack_reloc);
+               unsigned long unwritten;
+               unsigned int count;
+
+               count = min_t(unsigned int, remain, ARRAY_SIZE(stack_reloc));
                remain -= count;
 
-               if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]))) {
+               /* This is the fast path and we cannot handle a pagefault
+                * whilst holding the struct mutex lest the user pass in the
+                * relocations contained within a mmaped bo. For in such a case
+                * we, the page fault handler would call i915_gem_fault() and
+                * we would try to acquire the struct mutex again. Obviously
+                * this is bad and so lockdep complains vehemently.
+                */
+               pagefault_disable();
+               unwritten = __copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]));
+               pagefault_enable();
+               if (unlikely(unwritten)) {
                        ret = -EFAULT;
                        goto out;
                }
@@ -696,11 +688,26 @@ i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
                        if (ret)
                                goto out;
 
-                       if (r->presumed_offset != offset &&
-                           __put_user(r->presumed_offset,
-                                      &user_relocs->presumed_offset)) {
-                               ret = -EFAULT;
-                               goto out;
+                       if (r->presumed_offset != offset) {
+                               pagefault_disable();
+                               unwritten = __put_user(r->presumed_offset,
+                                                      &user_relocs->presumed_offset);
+                               pagefault_enable();
+                               if (unlikely(unwritten)) {
+                                       /* Note that reporting an error now
+                                        * leaves everything in an inconsistent
+                                        * state as we have *already* changed
+                                        * the relocation value inside the
+                                        * object. As we have not changed the
+                                        * reloc.presumed_offset or will not
+                                        * change the execobject.offset, on the
+                                        * call we may not rewrite the value
+                                        * inside the object, leaving it
+                                        * dangling and causing a GPU hang.
+                                        */
+                                       ret = -EFAULT;
+                                       goto out;
+                               }
                        }
 
                        user_relocs++;
@@ -740,20 +747,11 @@ i915_gem_execbuffer_relocate(struct eb_vmas *eb)
        struct i915_vma *vma;
        int ret = 0;
 
-       /* This is the fast path and we cannot handle a pagefault whilst
-        * holding the struct mutex lest the user pass in the relocations
-        * contained within a mmaped bo. For in such a case we, the page
-        * fault handler would call i915_gem_fault() and we would try to
-        * acquire the struct mutex again. Obviously this is bad and so
-        * lockdep complains vehemently.
-        */
-       pagefault_disable();
        list_for_each_entry(vma, &eb->vmas, exec_list) {
                ret = i915_gem_execbuffer_relocate_vma(vma, eb);
                if (ret)
                        break;
        }
-       pagefault_enable();
 
        return ret;
 }
@@ -1599,12 +1597,12 @@ eb_select_engine(struct drm_i915_private *dev_priv,
                        return NULL;
                }
 
-               engine = &dev_priv->engine[_VCS(bsd_idx)];
+               engine = dev_priv->engine[_VCS(bsd_idx)];
        } else {
-               engine = &dev_priv->engine[user_ring_map[user_ring_id]];
+               engine = dev_priv->engine[user_ring_map[user_ring_id]];
        }
 
-       if (!intel_engine_initialized(engine)) {
+       if (!engine) {
                DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
                return NULL;
        }