drm/i915: Track unbound pages
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 20 Aug 2012 09:40:46 +0000 (11:40 +0200)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 21 Aug 2012 12:34:11 +0000 (14:34 +0200)
When dealing with a working set larger than the GATT, or even the
mappable aperture when touching through the GTT, we end up with evicting
objects only to rebind them at a new offset again later. Moving an
object into and out of the GTT requires clflushing the pages, thus
causing a double-clflush penalty for rebinding.

To avoid having to clflush on rebinding, we can track the pages as they
are evicted from the GTT and only relinquish those pages on memory
pressure.

As usual, if it were not for the handling of out-of-memory condition and
having to manually shrink our own bo caches, it would be a net reduction
of code. Alas.

Note: The patch also contains a few changes to the last-hope
evict_everything logic in i916_gem_execbuffer.c - we no longer try to
only evict the purgeable stuff in a first try (since that's superflous
and only helps in OOM corner-cases, not fragmented-gtt trashing
situations).

Also, the extraction of the get_pages retry loop from bind_to_gtt (and
other callsites) to get_pages should imo have been a separate patch.

v2: Ditch the newly added put_pages (for unbound objects only) in
i915_gem_reset. A quick irc discussion hasn't revealed any important
reason for this, so if we need this, I'd like to have a git blame'able
explanation for it.

v3: Undo the s/drm_malloc_ab/kmalloc/ in get_pages that Chris noticed.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Split out code movements and rant a bit in the commit message
with a few Notes. Done v2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_debugfs.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_dmabuf.c
drivers/gpu/drm/i915/i915_gem_evict.c
drivers/gpu/drm/i915/i915_gem_execbuffer.c
drivers/gpu/drm/i915/i915_gem_gtt.c
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/i915_trace.h

index a18e93687b8ba11dcec3087f0746de4f1a2f77c2..608d3ae2b5ba5dc50be4ee3615f398be67260959 100644 (file)
@@ -211,7 +211,7 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
                   dev_priv->mm.object_memory);
 
        size = count = mappable_size = mappable_count = 0;
-       count_objects(&dev_priv->mm.gtt_list, gtt_list);
+       count_objects(&dev_priv->mm.bound_list, gtt_list);
        seq_printf(m, "%u [%u] objects, %zu [%zu] bytes in gtt\n",
                   count, mappable_count, size, mappable_size);
 
@@ -225,8 +225,13 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
        seq_printf(m, "  %u [%u] inactive objects, %zu [%zu] bytes\n",
                   count, mappable_count, size, mappable_size);
 
+       size = count = 0;
+       list_for_each_entry(obj, &dev_priv->mm.unbound_list, gtt_list)
+               size += obj->base.size, ++count;
+       seq_printf(m, "%u unbound objects, %zu bytes\n", count, size);
+
        size = count = mappable_size = mappable_count = 0;
-       list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
                if (obj->fault_mappable) {
                        size += obj->gtt_space->size;
                        ++count;
@@ -264,7 +269,7 @@ static int i915_gem_gtt_info(struct seq_file *m, void* data)
                return ret;
 
        total_obj_size = total_gtt_size = count = 0;
-       list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
                if (list == PINNED_LIST && obj->pin_count == 0)
                        continue;
 
@@ -526,7 +531,8 @@ static int i915_gem_fence_regs_info(struct seq_file *m, void *data)
        for (i = 0; i < dev_priv->num_fence_regs; i++) {
                struct drm_i915_gem_object *obj = dev_priv->fence_regs[i].obj;
 
-               seq_printf(m, "Fenced object[%2d] = ", i);
+               seq_printf(m, "Fence %d, pin count = %d, object = ",
+                          i, dev_priv->fence_regs[i].pin_count);
                if (obj == NULL)
                        seq_printf(m, "unused");
                else
index ed3ba70923ac1e40a4df1bab8674363024d9766f..a2382a175d9268288217d160a16ecf357b8cf1c8 100644 (file)
@@ -685,7 +685,13 @@ typedef struct drm_i915_private {
                struct drm_mm gtt_space;
                /** List of all objects in gtt_space. Used to restore gtt
                 * mappings on resume */
-               struct list_head gtt_list;
+               struct list_head bound_list;
+               /**
+                * List of objects which are not bound to the GTT (thus
+                * are idle and not used by the GPU) but still have
+                * (presumably uncached) pages still attached.
+                */
+               struct list_head unbound_list;
 
                /** Usable portion of the GTT for GEM */
                unsigned long gtt_start;
@@ -1306,8 +1312,7 @@ int __must_check i915_gem_object_unbind(struct drm_i915_gem_object *obj);
 void i915_gem_release_mmap(struct drm_i915_gem_object *obj);
 void i915_gem_lastclose(struct drm_device *dev);
 
-int i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
-                                 gfp_t gfpmask);
+int __must_check i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj);
 int __must_check i915_mutex_lock_interruptible(struct drm_device *dev);
 int i915_gem_object_sync(struct drm_i915_gem_object *obj,
                         struct intel_ring_buffer *to);
@@ -1449,7 +1454,7 @@ int __must_check i915_gem_evict_something(struct drm_device *dev, int min_size,
                                          unsigned alignment,
                                          unsigned cache_level,
                                          bool mappable);
-int i915_gem_evict_everything(struct drm_device *dev, bool purgeable_only);
+int i915_gem_evict_everything(struct drm_device *dev);
 
 /* i915_gem_stolen.c */
 int i915_gem_init_stolen(struct drm_device *dev);
index 0f70c2acfefac48a740779df1518959a923cdb15..462a8f2ad3a200be2822662e7bece791f50102d3 100644 (file)
@@ -55,6 +55,8 @@ static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
 
 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
                                    struct shrink_control *sc);
+static long i915_gem_purge(struct drm_i915_private *dev_priv, long target);
+static void i915_gem_shrink_all(struct drm_i915_private *dev_priv);
 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
 
 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
@@ -140,7 +142,7 @@ int i915_mutex_lock_interruptible(struct drm_device *dev)
 static inline bool
 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
 {
-       return !obj->active;
+       return obj->gtt_space && !obj->active;
 }
 
 int
@@ -179,7 +181,7 @@ i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
 
        pinned = 0;
        mutex_lock(&dev->struct_mutex);
-       list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list)
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
                if (obj->pin_count)
                        pinned += obj->gtt_space->size;
        mutex_unlock(&dev->struct_mutex);
@@ -423,9 +425,11 @@ i915_gem_shmem_pread(struct drm_device *dev,
                 * anyway again before the next pread happens. */
                if (obj->cache_level == I915_CACHE_NONE)
                        needs_clflush = 1;
-               ret = i915_gem_object_set_to_gtt_domain(obj, false);
-               if (ret)
-                       return ret;
+               if (obj->gtt_space) {
+                       ret = i915_gem_object_set_to_gtt_domain(obj, false);
+                       if (ret)
+                               return ret;
+               }
        }
 
        offset = args->offset;
@@ -751,9 +755,11 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
                 * right away and we therefore have to clflush anyway. */
                if (obj->cache_level == I915_CACHE_NONE)
                        needs_clflush_after = 1;
-               ret = i915_gem_object_set_to_gtt_domain(obj, true);
-               if (ret)
-                       return ret;
+               if (obj->gtt_space) {
+                       ret = i915_gem_object_set_to_gtt_domain(obj, true);
+                       if (ret)
+                               return ret;
+               }
        }
        /* Same trick applies for invalidate partially written cachelines before
         * writing.  */
@@ -1366,17 +1372,28 @@ i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
        return obj->madv == I915_MADV_DONTNEED;
 }
 
-static void
+static int
 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
 {
        int page_count = obj->base.size / PAGE_SIZE;
-       int i;
+       int ret, i;
 
-       if (!obj->pages)
-               return;
+       if (obj->pages == NULL)
+               return 0;
 
+       BUG_ON(obj->gtt_space);
        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);
+               i915_gem_clflush_object(obj);
+               obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       }
+
        if (i915_gem_object_needs_bit17_swizzle(obj))
                i915_gem_object_save_bit_17_swizzle(obj);
 
@@ -1396,37 +1413,112 @@ i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
 
        drm_free_large(obj->pages);
        obj->pages = NULL;
+
+       list_del(&obj->gtt_list);
+
+       if (i915_gem_object_is_purgeable(obj))
+               i915_gem_object_truncate(obj);
+
+       return 0;
+}
+
+static long
+i915_gem_purge(struct drm_i915_private *dev_priv, long target)
+{
+       struct drm_i915_gem_object *obj, *next;
+       long count = 0;
+
+       list_for_each_entry_safe(obj, next,
+                                &dev_priv->mm.unbound_list,
+                                gtt_list) {
+               if (i915_gem_object_is_purgeable(obj) &&
+                   i915_gem_object_put_pages_gtt(obj) == 0) {
+                       count += obj->base.size >> PAGE_SHIFT;
+                       if (count >= target)
+                               return count;
+               }
+       }
+
+       list_for_each_entry_safe(obj, next,
+                                &dev_priv->mm.inactive_list,
+                                mm_list) {
+               if (i915_gem_object_is_purgeable(obj) &&
+                   i915_gem_object_unbind(obj) == 0 &&
+                   i915_gem_object_put_pages_gtt(obj) == 0) {
+                       count += obj->base.size >> PAGE_SHIFT;
+                       if (count >= target)
+                               return count;
+               }
+       }
+
+       return count;
+}
+
+static void
+i915_gem_shrink_all(struct drm_i915_private *dev_priv)
+{
+       struct drm_i915_gem_object *obj, *next;
+
+       i915_gem_evict_everything(dev_priv->dev);
+
+       list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, gtt_list)
+               i915_gem_object_put_pages_gtt(obj);
 }
 
 int
-i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
-                             gfp_t gfpmask)
+i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 {
+       struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
        int page_count, i;
        struct address_space *mapping;
-       struct inode *inode;
        struct page *page;
+       gfp_t gfp;
 
        if (obj->pages || obj->sg_table)
                return 0;
 
+       /* Assert that the object is not currently in any GPU domain. As it
+        * wasn't in the GTT, there shouldn't be any way it could have been in
+        * a GPU cache
+        */
+       BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
+       BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
+
        /* Get the list of pages out of our struct file.  They'll be pinned
         * at this point until we release them.
         */
        page_count = obj->base.size / PAGE_SIZE;
-       BUG_ON(obj->pages != NULL);
        obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
        if (obj->pages == NULL)
                return -ENOMEM;
 
-       inode = obj->base.filp->f_path.dentry->d_inode;
-       mapping = inode->i_mapping;
-       gfpmask |= mapping_gfp_mask(mapping);
-
+       /* Fail silently without starting the shrinker */
+       mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
+       gfp = mapping_gfp_mask(mapping);
+       gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
+       gfp &= ~(__GFP_IO | __GFP_WAIT);
        for (i = 0; i < page_count; i++) {
-               page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
-               if (IS_ERR(page))
-                       goto err_pages;
+               page = shmem_read_mapping_page_gfp(mapping, i, gfp);
+               if (IS_ERR(page)) {
+                       i915_gem_purge(dev_priv, page_count);
+                       page = shmem_read_mapping_page_gfp(mapping, i, gfp);
+               }
+               if (IS_ERR(page)) {
+                       /* We've tried hard to allocate the memory by reaping
+                        * our own buffer, now let the real VM do its job and
+                        * go down in flames if truly OOM.
+                        */
+                       gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD);
+                       gfp |= __GFP_IO | __GFP_WAIT;
+
+                       i915_gem_shrink_all(dev_priv);
+                       page = shmem_read_mapping_page_gfp(mapping, i, gfp);
+                       if (IS_ERR(page))
+                               goto err_pages;
+
+                       gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
+                       gfp &= ~(__GFP_IO | __GFP_WAIT);
+               }
 
                obj->pages[i] = page;
        }
@@ -1434,6 +1526,7 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
        if (i915_gem_object_needs_bit17_swizzle(obj))
                i915_gem_object_do_bit_17_swizzle(obj);
 
+       list_add_tail(&obj->gtt_list, &dev_priv->mm.unbound_list);
        return 0;
 
 err_pages:
@@ -1698,6 +1791,7 @@ void i915_gem_reset(struct drm_device *dev)
                obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
        }
 
+
        /* The fence registers are invalidated so clear them out */
        i915_gem_reset_fences(dev);
 }
@@ -2209,22 +2303,6 @@ i915_gem_object_unbind(struct drm_i915_gem_object *obj)
 
        i915_gem_object_finish_gtt(obj);
 
-       /* Move the object to the CPU domain to ensure that
-        * any possible CPU writes while it's not in the GTT
-        * are flushed when we go to remap it.
-        */
-       if (ret == 0)
-               ret = i915_gem_object_set_to_cpu_domain(obj, 1);
-       if (ret == -ERESTARTSYS)
-               return ret;
-       if (ret) {
-               /* In the event of a disaster, abandon all caches and
-                * hope for the best.
-                */
-               i915_gem_clflush_object(obj);
-               obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
-       }
-
        /* release the fence reg _after_ flushing */
        ret = i915_gem_object_put_fence(obj);
        if (ret)
@@ -2240,10 +2318,8 @@ i915_gem_object_unbind(struct drm_i915_gem_object *obj)
        }
        i915_gem_gtt_finish_object(obj);
 
-       i915_gem_object_put_pages_gtt(obj);
-
-       list_del_init(&obj->gtt_list);
-       list_del_init(&obj->mm_list);
+       list_del(&obj->mm_list);
+       list_move_tail(&obj->gtt_list, &dev_priv->mm.unbound_list);
        /* Avoid an unnecessary call to unbind on rebind. */
        obj->map_and_fenceable = true;
 
@@ -2251,10 +2327,7 @@ i915_gem_object_unbind(struct drm_i915_gem_object *obj)
        obj->gtt_space = NULL;
        obj->gtt_offset = 0;
 
-       if (i915_gem_object_is_purgeable(obj))
-               i915_gem_object_truncate(obj);
-
-       return ret;
+       return 0;
 }
 
 static int i915_ring_idle(struct intel_ring_buffer *ring)
@@ -2667,7 +2740,6 @@ i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
        struct drm_device *dev = obj->base.dev;
        drm_i915_private_t *dev_priv = dev->dev_private;
        struct drm_mm_node *free_space;
-       gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
        u32 size, fence_size, fence_alignment, unfenced_alignment;
        bool mappable, fenceable;
        int ret;
@@ -2707,6 +2779,10 @@ i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
                return -E2BIG;
        }
 
+       ret = i915_gem_object_get_pages_gtt(obj);
+       if (ret)
+               return ret;
+
  search_free:
        if (map_and_fenceable)
                free_space =
@@ -2733,9 +2809,6 @@ i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
                                                         false);
        }
        if (obj->gtt_space == NULL) {
-               /* If the gtt is empty and we're still having trouble
-                * fitting our object in, we're out of memory.
-                */
                ret = i915_gem_evict_something(dev, size, alignment,
                                               obj->cache_level,
                                               map_and_fenceable);
@@ -2752,55 +2825,20 @@ i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
                return -EINVAL;
        }
 
-       ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
-       if (ret) {
-               drm_mm_put_block(obj->gtt_space);
-               obj->gtt_space = NULL;
-
-               if (ret == -ENOMEM) {
-                       /* first try to reclaim some memory by clearing the GTT */
-                       ret = i915_gem_evict_everything(dev, false);
-                       if (ret) {
-                               /* now try to shrink everyone else */
-                               if (gfpmask) {
-                                       gfpmask = 0;
-                                       goto search_free;
-                               }
-
-                               return -ENOMEM;
-                       }
-
-                       goto search_free;
-               }
-
-               return ret;
-       }
 
        ret = i915_gem_gtt_prepare_object(obj);
        if (ret) {
-               i915_gem_object_put_pages_gtt(obj);
                drm_mm_put_block(obj->gtt_space);
                obj->gtt_space = NULL;
-
-               if (i915_gem_evict_everything(dev, false))
-                       return ret;
-
-               goto search_free;
+               return ret;
        }
 
        if (!dev_priv->mm.aliasing_ppgtt)
                i915_gem_gtt_bind_object(obj, obj->cache_level);
 
-       list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
+       list_move_tail(&obj->gtt_list, &dev_priv->mm.bound_list);
        list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
 
-       /* Assert that the object is not currently in any GPU domain. As it
-        * wasn't in the GTT, there shouldn't be any way it could have been in
-        * a GPU cache
-        */
-       BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
-       BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
-
        obj->gtt_offset = obj->gtt_space->start;
 
        fenceable =
@@ -3464,9 +3502,8 @@ i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
        if (obj->madv != __I915_MADV_PURGED)
                obj->madv = args->madv;
 
-       /* if the object is no longer bound, discard its backing storage */
-       if (i915_gem_object_is_purgeable(obj) &&
-           obj->gtt_space == NULL)
+       /* if the object is no longer attached, discard its backing storage */
+       if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
                i915_gem_object_truncate(obj);
 
        args->retained = obj->madv != __I915_MADV_PURGED;
@@ -3573,6 +3610,7 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj)
                dev_priv->mm.interruptible = was_interruptible;
        }
 
+       i915_gem_object_put_pages_gtt(obj);
        if (obj->base.map_list.map)
                drm_gem_free_mmap_offset(&obj->base);
 
@@ -3605,7 +3643,7 @@ i915_gem_idle(struct drm_device *dev)
 
        /* Under UMS, be paranoid and evict. */
        if (!drm_core_check_feature(dev, DRIVER_MODESET))
-               i915_gem_evict_everything(dev, false);
+               i915_gem_evict_everything(dev);
 
        i915_gem_reset_fences(dev);
 
@@ -3963,8 +4001,9 @@ i915_gem_load(struct drm_device *dev)
 
        INIT_LIST_HEAD(&dev_priv->mm.active_list);
        INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
+       INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
+       INIT_LIST_HEAD(&dev_priv->mm.bound_list);
        INIT_LIST_HEAD(&dev_priv->mm.fence_list);
-       INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
        for (i = 0; i < I915_NUM_RINGS; i++)
                init_ring_lists(&dev_priv->ring[i]);
        for (i = 0; i < I915_MAX_NUM_FENCES; i++)
@@ -4208,13 +4247,6 @@ void i915_gem_release(struct drm_device *dev, struct drm_file *file)
        spin_unlock(&file_priv->mm.lock);
 }
 
-static int
-i915_gpu_is_active(struct drm_device *dev)
-{
-       drm_i915_private_t *dev_priv = dev->dev_private;
-       return !list_empty(&dev_priv->mm.active_list);
-}
-
 static int
 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
 {
@@ -4223,60 +4255,26 @@ i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
                             struct drm_i915_private,
                             mm.inactive_shrinker);
        struct drm_device *dev = dev_priv->dev;
-       struct drm_i915_gem_object *obj, *next;
+       struct drm_i915_gem_object *obj;
        int nr_to_scan = sc->nr_to_scan;
        int cnt;
 
        if (!mutex_trylock(&dev->struct_mutex))
                return 0;
 
-       /* "fast-path" to count number of available objects */
-       if (nr_to_scan == 0) {
-               cnt = 0;
-               list_for_each_entry(obj,
-                                   &dev_priv->mm.inactive_list,
-                                   mm_list)
-                       cnt++;
-               mutex_unlock(&dev->struct_mutex);
-               return cnt / 100 * sysctl_vfs_cache_pressure;
+       if (nr_to_scan) {
+               nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan);
+               if (nr_to_scan > 0)
+                       i915_gem_shrink_all(dev_priv);
        }
 
-rescan:
-       /* first scan for clean buffers */
-       i915_gem_retire_requests(dev);
-
-       list_for_each_entry_safe(obj, next,
-                                &dev_priv->mm.inactive_list,
-                                mm_list) {
-               if (i915_gem_object_is_purgeable(obj)) {
-                       if (i915_gem_object_unbind(obj) == 0 &&
-                           --nr_to_scan == 0)
-                               break;
-               }
-       }
-
-       /* second pass, evict/count anything still on the inactive list */
        cnt = 0;
-       list_for_each_entry_safe(obj, next,
-                                &dev_priv->mm.inactive_list,
-                                mm_list) {
-               if (nr_to_scan &&
-                   i915_gem_object_unbind(obj) == 0)
-                       nr_to_scan--;
-               else
-                       cnt++;
-       }
+       list_for_each_entry(obj, &dev_priv->mm.unbound_list, gtt_list)
+               cnt += obj->base.size >> PAGE_SHIFT;
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
+               if (obj->pin_count == 0)
+                       cnt += obj->base.size >> PAGE_SHIFT;
 
-       if (nr_to_scan && i915_gpu_is_active(dev)) {
-               /*
-                * We are desperate for pages, so as a last resort, wait
-                * for the GPU to finish and discard whatever we can.
-                * This has a dramatic impact to reduce the number of
-                * OOM-killer events whilst running the GPU aggressively.
-                */
-               if (i915_gpu_idle(dev) == 0)
-                       goto rescan;
-       }
        mutex_unlock(&dev->struct_mutex);
-       return cnt / 100 * sysctl_vfs_cache_pressure;
+       return cnt;
 }
index ceaad5af01a468ef4c3655391e936ad7798eae6f..43c95307f99ed36980cfc69d7f23d9d635f0a08c 100644 (file)
@@ -33,7 +33,7 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme
        struct drm_i915_gem_object *obj = attachment->dmabuf->priv;
        struct drm_device *dev = obj->base.dev;
        int npages = obj->base.size / PAGE_SIZE;
-       struct sg_table *sg = NULL;
+       struct sg_table *sg;
        int ret;
        int nents;
 
@@ -41,10 +41,10 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme
        if (ret)
                return ERR_PTR(ret);
 
-       if (!obj->pages) {
-               ret = i915_gem_object_get_pages_gtt(obj, __GFP_NORETRY | __GFP_NOWARN);
-               if (ret)
-                       goto out;
+       ret = i915_gem_object_get_pages_gtt(obj);
+       if (ret) {
+               sg = ERR_PTR(ret);
+               goto out;
        }
 
        /* link the pages into an SG then map the sg */
@@ -89,12 +89,10 @@ static void *i915_gem_dmabuf_vmap(struct dma_buf *dma_buf)
                goto out_unlock;
        }
 
-       if (!obj->pages) {
-               ret = i915_gem_object_get_pages_gtt(obj, __GFP_NORETRY | __GFP_NOWARN);
-               if (ret) {
-                       mutex_unlock(&dev->struct_mutex);
-                       return ERR_PTR(ret);
-               }
+       ret = i915_gem_object_get_pages_gtt(obj);
+       if (ret) {
+               mutex_unlock(&dev->struct_mutex);
+               return ERR_PTR(ret);
        }
 
        obj->dma_buf_vmapping = vmap(obj->pages, obj->base.size / PAGE_SIZE, 0, PAGE_KERNEL);
index 7279c31d4a9ae53211671aa308fe9b443b0e9a2d..74635da7c4984bc2c22fd82cc0348431328e1df9 100644 (file)
@@ -148,7 +148,7 @@ found:
 }
 
 int
-i915_gem_evict_everything(struct drm_device *dev, bool purgeable_only)
+i915_gem_evict_everything(struct drm_device *dev)
 {
        drm_i915_private_t *dev_priv = dev->dev_private;
        struct drm_i915_gem_object *obj, *next;
@@ -160,7 +160,7 @@ i915_gem_evict_everything(struct drm_device *dev, bool purgeable_only)
        if (lists_empty)
                return -ENOSPC;
 
-       trace_i915_gem_evict_everything(dev, purgeable_only);
+       trace_i915_gem_evict_everything(dev);
 
        /* The gpu_idle will flush everything in the write domain to the
         * active list. Then we must move everything off the active list
@@ -174,12 +174,9 @@ i915_gem_evict_everything(struct drm_device *dev, bool purgeable_only)
 
        /* Having flushed everything, unbind() should never raise an error */
        list_for_each_entry_safe(obj, next,
-                                &dev_priv->mm.inactive_list, mm_list) {
-               if (!purgeable_only || obj->madv != I915_MADV_WILLNEED) {
-                       if (obj->pin_count == 0)
-                               WARN_ON(i915_gem_object_unbind(obj));
-               }
-       }
+                                &dev_priv->mm.inactive_list, mm_list)
+               if (obj->pin_count == 0)
+                       WARN_ON(i915_gem_object_unbind(obj));
 
        return 0;
 }
index afb312ee050ca972daf537061c68a1163df51e5a..834a636b44f078eb5c4d3016663f16adf7975c4c 100644 (file)
@@ -502,17 +502,12 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
                        }
                }
 
-               if (ret != -ENOSPC || retry > 1)
+               if (ret != -ENOSPC || retry++)
                        return ret;
 
-               /* First attempt, just clear anything that is purgeable.
-                * Second attempt, clear the entire GTT.
-                */
-               ret = i915_gem_evict_everything(ring->dev, retry == 0);
+               ret = i915_gem_evict_everything(ring->dev);
                if (ret)
                        return ret;
-
-               retry++;
        } while (1);
 
 err:
index 3b3b731a17c3418d132c4b0943ca7bbf4f305796..8329a14862ed23bbdebca3c6065f2471b14f4b21 100644 (file)
@@ -348,7 +348,7 @@ void i915_gem_restore_gtt_mappings(struct drm_device *dev)
        intel_gtt_clear_range(dev_priv->mm.gtt_start / PAGE_SIZE,
                              (dev_priv->mm.gtt_end - dev_priv->mm.gtt_start) / PAGE_SIZE);
 
-       list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
                i915_gem_clflush_object(obj);
                i915_gem_gtt_bind_object(obj, obj->cache_level);
        }
index a61b41a8c607878cb82d881d8645d12dba9805b6..002dceebbeabfa9e1f17dea7345f5fbf78486428 100644 (file)
@@ -1221,7 +1221,7 @@ static void i915_capture_error_state(struct drm_device *dev)
        list_for_each_entry(obj, &dev_priv->mm.active_list, mm_list)
                i++;
        error->active_bo_count = i;
-       list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list)
+       list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
                if (obj->pin_count)
                        i++;
        error->pinned_bo_count = i - error->active_bo_count;
@@ -1246,7 +1246,7 @@ static void i915_capture_error_state(struct drm_device *dev)
                error->pinned_bo_count =
                        capture_pinned_bo(error->pinned_bo,
                                          error->pinned_bo_count,
-                                         &dev_priv->mm.gtt_list);
+                                         &dev_priv->mm.bound_list);
 
        do_gettimeofday(&error->time);
 
index fe90b3a84a6d9db3b7cac327173749fe0fc0b0a9..3c4093d91f60d1e5b6e903d9a24f66285884ed82 100644 (file)
@@ -214,22 +214,18 @@ TRACE_EVENT(i915_gem_evict,
 );
 
 TRACE_EVENT(i915_gem_evict_everything,
-           TP_PROTO(struct drm_device *dev, bool purgeable),
-           TP_ARGS(dev, purgeable),
+           TP_PROTO(struct drm_device *dev),
+           TP_ARGS(dev),
 
            TP_STRUCT__entry(
                             __field(u32, dev)
-                            __field(bool, purgeable)
                            ),
 
            TP_fast_assign(
                           __entry->dev = dev->primary->index;
-                          __entry->purgeable = purgeable;
                          ),
 
-           TP_printk("dev=%d%s",
-                     __entry->dev,
-                     __entry->purgeable ? ", purgeable only" : "")
+           TP_printk("dev=%d", __entry->dev)
 );
 
 TRACE_EVENT(i915_gem_ring_dispatch,