drm/ttm: Memory accounting rework.

Use inclusive zones to simplify accounting and its sysfs representation.
Use DMA32 accounting where applicable.

Add a sysfs interface to make the heuristically determined limits
readable and configurable.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>

drivers/gpu/drm/ttm/ttm_bo.c

@@ -70,7 +70,7 @@ static void ttm_bo_release_list(struct kref *list_kref)
 	if (bo->destroy)
 		bo->destroy(bo);
 	else {
-		ttm_mem_global_free(bdev->mem_glob, bo->acc_size, false);
+		ttm_mem_global_free(bdev->mem_glob, bo->acc_size);
 		kfree(bo);
 	}
 }
@@ -1065,14 +1065,14 @@ int ttm_buffer_object_create(struct ttm_bo_device *bdev,
 
 	size_t acc_size =
 	    ttm_bo_size(bdev, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
-	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false, false);
+	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
 	if (unlikely(ret != 0))
 		return ret;
 
 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
 
 	if (unlikely(bo == NULL)) {
-		ttm_mem_global_free(mem_glob, acc_size, false);
+		ttm_mem_global_free(mem_glob, acc_size);
 		return -ENOMEM;
 	}
 

drivers/gpu/drm/ttm/ttm_global.c

@@ -71,7 +71,7 @@ int ttm_global_item_ref(struct ttm_global_reference *ref)
 
 	mutex_lock(&item->mutex);
 	if (item->refcount == 0) {
-		item->object = kmalloc(ref->size, GFP_KERNEL);
+		item->object = kzalloc(ref->size, GFP_KERNEL);
 		if (unlikely(item->object == NULL)) {
 			ret = -ENOMEM;
 			goto out_err;
@@ -89,7 +89,6 @@ int ttm_global_item_ref(struct ttm_global_reference *ref)
 	mutex_unlock(&item->mutex);
 	return 0;
 out_err:
-	kfree(item->object);
 	mutex_unlock(&item->mutex);
 	item->object = NULL;
 	return ret;
@@ -105,7 +104,6 @@ void ttm_global_item_unref(struct ttm_global_reference *ref)
 	BUG_ON(ref->object != item->object);
 	if (--item->refcount == 0) {
 		ref->release(ref);
-		kfree(item->object);
 		item->object = NULL;
 	}
 	mutex_unlock(&item->mutex);

drivers/gpu/drm/ttm/ttm_tt.c

@@ -166,7 +166,7 @@ static void ttm_tt_free_user_pages(struct ttm_tt *ttm)
 			set_page_dirty_lock(page);
 
 		ttm->pages[i] = NULL;
-		ttm_mem_global_free(ttm->bdev->mem_glob, PAGE_SIZE, false);
+		ttm_mem_global_free(ttm->bdev->mem_glob, PAGE_SIZE);
 		put_page(page);
 	}
 	ttm->state = tt_unpopulated;
@@ -187,21 +187,14 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
 		if (!p)
 			return NULL;
 
-		if (PageHighMem(p)) {
-			ret =
-			    ttm_mem_global_alloc(mem_glob, PAGE_SIZE,
-						 false, false, true);
-			if (unlikely(ret != 0))
-				goto out_err;
+		ret = ttm_mem_global_alloc_page(mem_glob, p, false, false);
+		if (unlikely(ret != 0))
+			goto out_err;
+
+		if (PageHighMem(p))
 			ttm->pages[--ttm->first_himem_page] = p;
-		} else {
-			ret =
-			    ttm_mem_global_alloc(mem_glob, PAGE_SIZE,
-						 false, false, false);
-			if (unlikely(ret != 0))
-				goto out_err;
+		else
 			ttm->pages[++ttm->last_lomem_page] = p;
-		}
 	}
 	return p;
 out_err:
@@ -355,8 +348,8 @@ static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
 				printk(KERN_ERR TTM_PFX
 				       "Erroneous page count. "
 				       "Leaking pages.\n");
-			ttm_mem_global_free(ttm->bdev->mem_glob, PAGE_SIZE,
-					    PageHighMem(cur_page));
+			ttm_mem_global_free_page(ttm->bdev->mem_glob,
+						 cur_page);
 			__free_page(cur_page);
 		}
 	}
@@ -411,7 +404,7 @@ int ttm_tt_set_user(struct ttm_tt *ttm,
 	 */
 
 	ret = ttm_mem_global_alloc(mem_glob, num_pages * PAGE_SIZE,
-				   false, false, false);
+				   false, false);
 	if (unlikely(ret != 0))
 		return ret;
 
@@ -422,7 +415,7 @@ int ttm_tt_set_user(struct ttm_tt *ttm,
 
 	if (ret != num_pages && write) {
 		ttm_tt_free_user_pages(ttm);
-		ttm_mem_global_free(mem_glob, num_pages * PAGE_SIZE, false);
+		ttm_mem_global_free(mem_glob, num_pages * PAGE_SIZE);
 		return -ENOMEM;
 	}
 

include/drm/ttm/ttm_memory.h

@@ -32,6 +32,7 @@
 #include <linux/spinlock.h>
 #include <linux/wait.h>
 #include <linux/errno.h>
+#include <linux/kobject.h>
 
 /**
  * struct ttm_mem_shrink - callback to shrink TTM memory usage.
@@ -60,34 +61,33 @@ struct ttm_mem_shrink {
  * @queue: Wait queue for processes suspended waiting for memory.
  * @lock: Lock to protect the @shrink - and the memory accounting members,
  * that is, essentially the whole structure with some exceptions.
- * @emer_memory: Lowmem memory limit available for root.
- * @max_memory: Lowmem memory limit available for non-root.
- * @swap_limit: Lowmem memory limit where the shrink workqueue kicks in.
- * @used_memory: Currently used lowmem memory.
- * @used_total_memory: Currently used total (lowmem + highmem) memory.
- * @total_memory_swap_limit: Total memory limit where the shrink workqueue
- * kicks in.
- * @max_total_memory: Total memory available to non-root processes.
- * @emer_total_memory: Total memory available to root processes.
+ * @zones: Array of pointers to accounting zones.
+ * @num_zones: Number of populated entries in the @zones array.
+ * @zone_kernel: Pointer to the kernel zone.
+ * @zone_highmem: Pointer to the highmem zone if there is one.
+ * @zone_dma32: Pointer to the dma32 zone if there is one.
  *
  * Note that this structure is not per device. It should be global for all
  * graphics devices.
  */
 
+#define TTM_MEM_MAX_ZONES 2
+struct ttm_mem_zone;
 struct ttm_mem_global {
+	struct kobject kobj;
 	struct ttm_mem_shrink *shrink;
 	struct workqueue_struct *swap_queue;
 	struct work_struct work;
 	wait_queue_head_t queue;
 	spinlock_t lock;
-	uint64_t emer_memory;
-	uint64_t max_memory;
-	uint64_t swap_limit;
-	uint64_t used_memory;
-	uint64_t used_total_memory;
-	uint64_t total_memory_swap_limit;
-	uint64_t max_total_memory;
-	uint64_t emer_total_memory;
+	struct ttm_mem_zone *zones[TTM_MEM_MAX_ZONES];
+	unsigned int num_zones;
+	struct ttm_mem_zone *zone_kernel;
+#ifdef CONFIG_HIGHMEM
+	struct ttm_mem_zone *zone_highmem;
+#else
+	struct ttm_mem_zone *zone_dma32;
+#endif
 };
 
 /**
@@ -146,8 +146,13 @@ static inline void ttm_mem_unregister_shrink(struct ttm_mem_global *glob,
 extern int ttm_mem_global_init(struct ttm_mem_global *glob);
 extern void ttm_mem_global_release(struct ttm_mem_global *glob);
 extern int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
-				bool no_wait, bool interruptible, bool himem);
+				bool no_wait, bool interruptible);
 extern void ttm_mem_global_free(struct ttm_mem_global *glob,
-				uint64_t amount, bool himem);
+				uint64_t amount);
+extern int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
+				     struct page *page,
+				     bool no_wait, bool interruptible);
+extern void ttm_mem_global_free_page(struct ttm_mem_global *glob,
+				     struct page *page);
 extern size_t ttm_round_pot(size_t size);
 #endif

include/drm/ttm/ttm_module.h

@@ -32,6 +32,7 @@
 #define _TTM_MODULE_H_
 
 #include <linux/kernel.h>
+struct kobject;
 
 #define TTM_PFX "[TTM]"
 
@@ -54,5 +55,6 @@ extern void ttm_global_init(void);
 extern void ttm_global_release(void);
 extern int ttm_global_item_ref(struct ttm_global_reference *ref);
 extern void ttm_global_item_unref(struct ttm_global_reference *ref);
+extern struct kobject *ttm_get_kobj(void);
 
 #endif /* _TTM_MODULE_H_ */