1 /*
2  * Copyright © 2017 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24 
25 #include "display/intel_frontbuffer.h"
26 
27 #include "i915_drv.h"
28 #include "i915_gem_clflush.h"
29 #include "i915_gem_context.h"
30 #include "i915_gem_object.h"
31 #include "i915_globals.h"
32 
33 static struct i915_global_object {
34 	struct i915_global base;
35 	struct kmem_cache *slab_objects;
36 } global;
37 
38 struct drm_i915_gem_object *i915_gem_object_alloc(void)
39 {
40 	return kmem_cache_zalloc(global.slab_objects, GFP_KERNEL);
41 }
42 
43 void i915_gem_object_free(struct drm_i915_gem_object *obj)
44 {
45 	return kmem_cache_free(global.slab_objects, obj);
46 }
47 
48 static void
49 frontbuffer_retire(struct i915_active_request *active,
50 		   struct i915_request *request)
51 {
52 	struct drm_i915_gem_object *obj =
53 		container_of(active, typeof(*obj), frontbuffer_write);
54 
55 	intel_fb_obj_flush(obj, ORIGIN_CS);
56 }
57 
58 void i915_gem_object_init(struct drm_i915_gem_object *obj,
59 			  const struct drm_i915_gem_object_ops *ops)
60 {
61 	mutex_init(&obj->mm.lock);
62 
63 	spin_lock_init(&obj->vma.lock);
64 	INIT_LIST_HEAD(&obj->vma.list);
65 
66 	INIT_LIST_HEAD(&obj->lut_list);
67 	INIT_LIST_HEAD(&obj->batch_pool_link);
68 
69 	init_rcu_head(&obj->rcu);
70 
71 	obj->ops = ops;
72 
73 	obj->frontbuffer_ggtt_origin = ORIGIN_GTT;
74 	i915_active_request_init(&obj->frontbuffer_write,
75 				 NULL, frontbuffer_retire);
76 
77 	obj->mm.madv = I915_MADV_WILLNEED;
78 	INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
79 	mutex_init(&obj->mm.get_page.lock);
80 }
81 
82 /**
83  * Mark up the object's coherency levels for a given cache_level
84  * @obj: #drm_i915_gem_object
85  * @cache_level: cache level
86  */
87 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
88 					 unsigned int cache_level)
89 {
90 	obj->cache_level = cache_level;
91 
92 	if (cache_level != I915_CACHE_NONE)
93 		obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
94 				       I915_BO_CACHE_COHERENT_FOR_WRITE);
95 	else if (HAS_LLC(to_i915(obj->base.dev)))
96 		obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
97 	else
98 		obj->cache_coherent = 0;
99 
100 	obj->cache_dirty =
101 		!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE);
102 }
103 
104 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
105 {
106 	struct drm_i915_gem_object *obj = to_intel_bo(gem);
107 	struct drm_i915_file_private *fpriv = file->driver_priv;
108 	struct i915_lut_handle *lut, *ln;
109 	LIST_HEAD(close);
110 
111 	i915_gem_object_lock(obj);
112 	list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
113 		struct i915_gem_context *ctx = lut->ctx;
114 
115 		if (ctx->file_priv != fpriv)
116 			continue;
117 
118 		i915_gem_context_get(ctx);
119 		list_move(&lut->obj_link, &close);
120 	}
121 	i915_gem_object_unlock(obj);
122 
123 	list_for_each_entry_safe(lut, ln, &close, obj_link) {
124 		struct i915_gem_context *ctx = lut->ctx;
125 		struct i915_vma *vma;
126 
127 		/*
128 		 * We allow the process to have multiple handles to the same
129 		 * vma, in the same fd namespace, by virtue of flink/open.
130 		 */
131 
132 		mutex_lock(&ctx->mutex);
133 		vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
134 		if (vma) {
135 			GEM_BUG_ON(vma->obj != obj);
136 			GEM_BUG_ON(!atomic_read(&vma->open_count));
137 			if (atomic_dec_and_test(&vma->open_count) &&
138 			    !i915_vma_is_ggtt(vma))
139 				i915_vma_close(vma);
140 		}
141 		mutex_unlock(&ctx->mutex);
142 
143 		i915_gem_context_put(lut->ctx);
144 		i915_lut_handle_free(lut);
145 		i915_gem_object_put(obj);
146 	}
147 }
148 
149 static void __i915_gem_free_objects(struct drm_i915_private *i915,
150 				    struct llist_node *freed)
151 {
152 	struct drm_i915_gem_object *obj, *on;
153 	intel_wakeref_t wakeref;
154 
155 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
156 	llist_for_each_entry_safe(obj, on, freed, freed) {
157 		struct i915_vma *vma, *vn;
158 
159 		trace_i915_gem_object_destroy(obj);
160 
161 		mutex_lock(&i915->drm.struct_mutex);
162 
163 		GEM_BUG_ON(i915_gem_object_is_active(obj));
164 		list_for_each_entry_safe(vma, vn, &obj->vma.list, obj_link) {
165 			GEM_BUG_ON(i915_vma_is_active(vma));
166 			vma->flags &= ~I915_VMA_PIN_MASK;
167 			i915_vma_destroy(vma);
168 		}
169 		GEM_BUG_ON(!list_empty(&obj->vma.list));
170 		GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma.tree));
171 
172 		/*
173 		 * This serializes freeing with the shrinker. Since the free
174 		 * is delayed, first by RCU then by the workqueue, we want the
175 		 * shrinker to be able to free pages of unreferenced objects,
176 		 * or else we may oom whilst there are plenty of deferred
177 		 * freed objects.
178 		 */
179 		if (i915_gem_object_has_pages(obj) &&
180 		    i915_gem_object_is_shrinkable(obj)) {
181 			unsigned long flags;
182 
183 			spin_lock_irqsave(&i915->mm.obj_lock, flags);
184 			list_del_init(&obj->mm.link);
185 			spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
186 		}
187 
188 		mutex_unlock(&i915->drm.struct_mutex);
189 
190 		GEM_BUG_ON(atomic_read(&obj->bind_count));
191 		GEM_BUG_ON(obj->userfault_count);
192 		GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits));
193 		GEM_BUG_ON(!list_empty(&obj->lut_list));
194 
195 		if (obj->ops->release)
196 			obj->ops->release(obj);
197 
198 		atomic_set(&obj->mm.pages_pin_count, 0);
199 		__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
200 		GEM_BUG_ON(i915_gem_object_has_pages(obj));
201 
202 		if (obj->base.import_attach)
203 			drm_prime_gem_destroy(&obj->base, NULL);
204 
205 		drm_gem_object_release(&obj->base);
206 
207 		bitmap_free(obj->bit_17);
208 		i915_gem_object_free(obj);
209 
210 		GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
211 		atomic_dec(&i915->mm.free_count);
212 
213 		cond_resched();
214 	}
215 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
216 }
217 
218 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
219 {
220 	struct llist_node *freed;
221 
222 	/* Free the oldest, most stale object to keep the free_list short */
223 	freed = NULL;
224 	if (!llist_empty(&i915->mm.free_list)) { /* quick test for hotpath */
225 		/* Only one consumer of llist_del_first() allowed */
226 		spin_lock(&i915->mm.free_lock);
227 		freed = llist_del_first(&i915->mm.free_list);
228 		spin_unlock(&i915->mm.free_lock);
229 	}
230 	if (unlikely(freed)) {
231 		freed->next = NULL;
232 		__i915_gem_free_objects(i915, freed);
233 	}
234 }
235 
236 static void __i915_gem_free_work(struct work_struct *work)
237 {
238 	struct drm_i915_private *i915 =
239 		container_of(work, struct drm_i915_private, mm.free_work);
240 	struct llist_node *freed;
241 
242 	/*
243 	 * All file-owned VMA should have been released by this point through
244 	 * i915_gem_close_object(), or earlier by i915_gem_context_close().
245 	 * However, the object may also be bound into the global GTT (e.g.
246 	 * older GPUs without per-process support, or for direct access through
247 	 * the GTT either for the user or for scanout). Those VMA still need to
248 	 * unbound now.
249 	 */
250 
251 	spin_lock(&i915->mm.free_lock);
252 	while ((freed = llist_del_all(&i915->mm.free_list))) {
253 		spin_unlock(&i915->mm.free_lock);
254 
255 		__i915_gem_free_objects(i915, freed);
256 		if (need_resched())
257 			return;
258 
259 		spin_lock(&i915->mm.free_lock);
260 	}
261 	spin_unlock(&i915->mm.free_lock);
262 }
263 
264 static void __i915_gem_free_object_rcu(struct rcu_head *head)
265 {
266 	struct drm_i915_gem_object *obj =
267 		container_of(head, typeof(*obj), rcu);
268 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
269 
270 	/*
271 	 * We reuse obj->rcu for the freed list, so we had better not treat
272 	 * it like a rcu_head from this point forwards. And we expect all
273 	 * objects to be freed via this path.
274 	 */
275 	destroy_rcu_head(&obj->rcu);
276 
277 	/*
278 	 * Since we require blocking on struct_mutex to unbind the freed
279 	 * object from the GPU before releasing resources back to the
280 	 * system, we can not do that directly from the RCU callback (which may
281 	 * be a softirq context), but must instead then defer that work onto a
282 	 * kthread. We use the RCU callback rather than move the freed object
283 	 * directly onto the work queue so that we can mix between using the
284 	 * worker and performing frees directly from subsequent allocations for
285 	 * crude but effective memory throttling.
286 	 */
287 	if (llist_add(&obj->freed, &i915->mm.free_list))
288 		queue_work(i915->wq, &i915->mm.free_work);
289 }
290 
291 void i915_gem_free_object(struct drm_gem_object *gem_obj)
292 {
293 	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
294 
295 	/*
296 	 * Before we free the object, make sure any pure RCU-only
297 	 * read-side critical sections are complete, e.g.
298 	 * i915_gem_busy_ioctl(). For the corresponding synchronized
299 	 * lookup see i915_gem_object_lookup_rcu().
300 	 */
301 	atomic_inc(&to_i915(obj->base.dev)->mm.free_count);
302 	call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
303 }
304 
305 static inline enum fb_op_origin
306 fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain)
307 {
308 	return (domain == I915_GEM_DOMAIN_GTT ?
309 		obj->frontbuffer_ggtt_origin : ORIGIN_CPU);
310 }
311 
312 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
313 {
314 	return !(obj->cache_level == I915_CACHE_NONE ||
315 		 obj->cache_level == I915_CACHE_WT);
316 }
317 
318 void
319 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj,
320 				   unsigned int flush_domains)
321 {
322 	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
323 	struct i915_vma *vma;
324 
325 	assert_object_held(obj);
326 
327 	if (!(obj->write_domain & flush_domains))
328 		return;
329 
330 	switch (obj->write_domain) {
331 	case I915_GEM_DOMAIN_GTT:
332 		i915_gem_flush_ggtt_writes(dev_priv);
333 
334 		intel_fb_obj_flush(obj,
335 				   fb_write_origin(obj, I915_GEM_DOMAIN_GTT));
336 
337 		for_each_ggtt_vma(vma, obj) {
338 			if (vma->iomap)
339 				continue;
340 
341 			i915_vma_unset_ggtt_write(vma);
342 		}
343 		break;
344 
345 	case I915_GEM_DOMAIN_WC:
346 		wmb();
347 		break;
348 
349 	case I915_GEM_DOMAIN_CPU:
350 		i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
351 		break;
352 
353 	case I915_GEM_DOMAIN_RENDER:
354 		if (gpu_write_needs_clflush(obj))
355 			obj->cache_dirty = true;
356 		break;
357 	}
358 
359 	obj->write_domain = 0;
360 }
361 
362 void i915_gem_init__objects(struct drm_i915_private *i915)
363 {
364 	INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
365 }
366 
367 static void i915_global_objects_shrink(void)
368 {
369 	kmem_cache_shrink(global.slab_objects);
370 }
371 
372 static void i915_global_objects_exit(void)
373 {
374 	kmem_cache_destroy(global.slab_objects);
375 }
376 
377 static struct i915_global_object global = { {
378 	.shrink = i915_global_objects_shrink,
379 	.exit = i915_global_objects_exit,
380 } };
381 
382 int __init i915_global_objects_init(void)
383 {
384 	global.slab_objects =
385 		KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
386 	if (!global.slab_objects)
387 		return -ENOMEM;
388 
389 	i915_global_register(&global.base);
390 	return 0;
391 }
392 
393 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
394 #include "selftests/huge_gem_object.c"
395 #include "selftests/huge_pages.c"
396 #include "selftests/i915_gem_object.c"
397 #include "selftests/i915_gem_coherency.c"
398 #endif
399