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 <linux/sched/mm.h>
26 
27 #include "display/intel_frontbuffer.h"
28 #include "gt/intel_gt.h"
29 #include "i915_drv.h"
30 #include "i915_gem_clflush.h"
31 #include "i915_gem_context.h"
32 #include "i915_gem_mman.h"
33 #include "i915_gem_object.h"
34 #include "i915_globals.h"
35 #include "i915_trace.h"
36 
37 static struct i915_global_object {
38 	struct i915_global base;
39 	struct kmem_cache *slab_objects;
40 } global;
41 
42 struct drm_i915_gem_object *i915_gem_object_alloc(void)
43 {
44 	return kmem_cache_zalloc(global.slab_objects, GFP_KERNEL);
45 }
46 
47 void i915_gem_object_free(struct drm_i915_gem_object *obj)
48 {
49 	return kmem_cache_free(global.slab_objects, obj);
50 }
51 
52 void i915_gem_object_init(struct drm_i915_gem_object *obj,
53 			  const struct drm_i915_gem_object_ops *ops,
54 			  struct lock_class_key *key)
55 {
56 	__mutex_init(&obj->mm.lock, "obj->mm.lock", key);
57 
58 	spin_lock_init(&obj->vma.lock);
59 	INIT_LIST_HEAD(&obj->vma.list);
60 
61 	INIT_LIST_HEAD(&obj->mm.link);
62 
63 	INIT_LIST_HEAD(&obj->lut_list);
64 
65 	spin_lock_init(&obj->mmo.lock);
66 	obj->mmo.offsets = RB_ROOT;
67 
68 	init_rcu_head(&obj->rcu);
69 
70 	obj->ops = ops;
71 
72 	obj->mm.madv = I915_MADV_WILLNEED;
73 	INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
74 	mutex_init(&obj->mm.get_page.lock);
75 }
76 
77 /**
78  * Mark up the object's coherency levels for a given cache_level
79  * @obj: #drm_i915_gem_object
80  * @cache_level: cache level
81  */
82 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
83 					 unsigned int cache_level)
84 {
85 	obj->cache_level = cache_level;
86 
87 	if (cache_level != I915_CACHE_NONE)
88 		obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
89 				       I915_BO_CACHE_COHERENT_FOR_WRITE);
90 	else if (HAS_LLC(to_i915(obj->base.dev)))
91 		obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
92 	else
93 		obj->cache_coherent = 0;
94 
95 	obj->cache_dirty =
96 		!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE);
97 }
98 
99 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
100 {
101 	struct drm_i915_gem_object *obj = to_intel_bo(gem);
102 	struct drm_i915_file_private *fpriv = file->driver_priv;
103 	struct i915_mmap_offset *mmo, *mn;
104 	struct i915_lut_handle *lut, *ln;
105 	LIST_HEAD(close);
106 
107 	i915_gem_object_lock(obj);
108 	list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
109 		struct i915_gem_context *ctx = lut->ctx;
110 
111 		if (ctx->file_priv != fpriv)
112 			continue;
113 
114 		i915_gem_context_get(ctx);
115 		list_move(&lut->obj_link, &close);
116 	}
117 	i915_gem_object_unlock(obj);
118 
119 	spin_lock(&obj->mmo.lock);
120 	rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
121 		drm_vma_node_revoke(&mmo->vma_node, file);
122 	spin_unlock(&obj->mmo.lock);
123 
124 	list_for_each_entry_safe(lut, ln, &close, obj_link) {
125 		struct i915_gem_context *ctx = lut->ctx;
126 		struct i915_vma *vma;
127 
128 		/*
129 		 * We allow the process to have multiple handles to the same
130 		 * vma, in the same fd namespace, by virtue of flink/open.
131 		 */
132 
133 		mutex_lock(&ctx->mutex);
134 		vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
135 		if (vma) {
136 			GEM_BUG_ON(vma->obj != obj);
137 			GEM_BUG_ON(!atomic_read(&vma->open_count));
138 			if (atomic_dec_and_test(&vma->open_count) &&
139 			    !i915_vma_is_ggtt(vma))
140 				i915_vma_close(vma);
141 		}
142 		mutex_unlock(&ctx->mutex);
143 
144 		i915_gem_context_put(lut->ctx);
145 		i915_lut_handle_free(lut);
146 		i915_gem_object_put(obj);
147 	}
148 }
149 
150 static void __i915_gem_free_object_rcu(struct rcu_head *head)
151 {
152 	struct drm_i915_gem_object *obj =
153 		container_of(head, typeof(*obj), rcu);
154 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
155 
156 	dma_resv_fini(&obj->base._resv);
157 	i915_gem_object_free(obj);
158 
159 	GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
160 	atomic_dec(&i915->mm.free_count);
161 }
162 
163 static void __i915_gem_free_objects(struct drm_i915_private *i915,
164 				    struct llist_node *freed)
165 {
166 	struct drm_i915_gem_object *obj, *on;
167 	intel_wakeref_t wakeref;
168 
169 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
170 	llist_for_each_entry_safe(obj, on, freed, freed) {
171 		struct i915_mmap_offset *mmo, *mn;
172 
173 		trace_i915_gem_object_destroy(obj);
174 
175 		if (!list_empty(&obj->vma.list)) {
176 			struct i915_vma *vma;
177 
178 			/*
179 			 * Note that the vma keeps an object reference while
180 			 * it is active, so it *should* not sleep while we
181 			 * destroy it. Our debug code errs insits it *might*.
182 			 * For the moment, play along.
183 			 */
184 			spin_lock(&obj->vma.lock);
185 			while ((vma = list_first_entry_or_null(&obj->vma.list,
186 							       struct i915_vma,
187 							       obj_link))) {
188 				GEM_BUG_ON(vma->obj != obj);
189 				spin_unlock(&obj->vma.lock);
190 
191 				__i915_vma_put(vma);
192 
193 				spin_lock(&obj->vma.lock);
194 			}
195 			spin_unlock(&obj->vma.lock);
196 		}
197 
198 		i915_gem_object_release_mmap(obj);
199 
200 		rbtree_postorder_for_each_entry_safe(mmo, mn,
201 						     &obj->mmo.offsets,
202 						     offset) {
203 			drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
204 					      &mmo->vma_node);
205 			kfree(mmo);
206 		}
207 		obj->mmo.offsets = RB_ROOT;
208 
209 		GEM_BUG_ON(atomic_read(&obj->bind_count));
210 		GEM_BUG_ON(obj->userfault_count);
211 		GEM_BUG_ON(!list_empty(&obj->lut_list));
212 
213 		atomic_set(&obj->mm.pages_pin_count, 0);
214 		__i915_gem_object_put_pages(obj);
215 		GEM_BUG_ON(i915_gem_object_has_pages(obj));
216 		bitmap_free(obj->bit_17);
217 
218 		if (obj->base.import_attach)
219 			drm_prime_gem_destroy(&obj->base, NULL);
220 
221 		drm_gem_free_mmap_offset(&obj->base);
222 
223 		if (obj->ops->release)
224 			obj->ops->release(obj);
225 
226 		/* But keep the pointer alive for RCU-protected lookups */
227 		call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
228 		cond_resched();
229 	}
230 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
231 }
232 
233 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
234 {
235 	struct llist_node *freed = llist_del_all(&i915->mm.free_list);
236 
237 	if (unlikely(freed))
238 		__i915_gem_free_objects(i915, freed);
239 }
240 
241 static void __i915_gem_free_work(struct work_struct *work)
242 {
243 	struct drm_i915_private *i915 =
244 		container_of(work, struct drm_i915_private, mm.free_work);
245 
246 	i915_gem_flush_free_objects(i915);
247 }
248 
249 void i915_gem_free_object(struct drm_gem_object *gem_obj)
250 {
251 	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
252 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
253 
254 	GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
255 
256 	/*
257 	 * Before we free the object, make sure any pure RCU-only
258 	 * read-side critical sections are complete, e.g.
259 	 * i915_gem_busy_ioctl(). For the corresponding synchronized
260 	 * lookup see i915_gem_object_lookup_rcu().
261 	 */
262 	atomic_inc(&i915->mm.free_count);
263 
264 	/*
265 	 * This serializes freeing with the shrinker. Since the free
266 	 * is delayed, first by RCU then by the workqueue, we want the
267 	 * shrinker to be able to free pages of unreferenced objects,
268 	 * or else we may oom whilst there are plenty of deferred
269 	 * freed objects.
270 	 */
271 	i915_gem_object_make_unshrinkable(obj);
272 
273 	/*
274 	 * Since we require blocking on struct_mutex to unbind the freed
275 	 * object from the GPU before releasing resources back to the
276 	 * system, we can not do that directly from the RCU callback (which may
277 	 * be a softirq context), but must instead then defer that work onto a
278 	 * kthread. We use the RCU callback rather than move the freed object
279 	 * directly onto the work queue so that we can mix between using the
280 	 * worker and performing frees directly from subsequent allocations for
281 	 * crude but effective memory throttling.
282 	 */
283 	if (llist_add(&obj->freed, &i915->mm.free_list))
284 		queue_work(i915->wq, &i915->mm.free_work);
285 }
286 
287 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
288 {
289 	return !(obj->cache_level == I915_CACHE_NONE ||
290 		 obj->cache_level == I915_CACHE_WT);
291 }
292 
293 void
294 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj,
295 				   unsigned int flush_domains)
296 {
297 	struct i915_vma *vma;
298 
299 	assert_object_held(obj);
300 
301 	if (!(obj->write_domain & flush_domains))
302 		return;
303 
304 	switch (obj->write_domain) {
305 	case I915_GEM_DOMAIN_GTT:
306 		spin_lock(&obj->vma.lock);
307 		for_each_ggtt_vma(vma, obj) {
308 			if (i915_vma_unset_ggtt_write(vma))
309 				intel_gt_flush_ggtt_writes(vma->vm->gt);
310 		}
311 		spin_unlock(&obj->vma.lock);
312 
313 		i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
314 		break;
315 
316 	case I915_GEM_DOMAIN_WC:
317 		wmb();
318 		break;
319 
320 	case I915_GEM_DOMAIN_CPU:
321 		i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
322 		break;
323 
324 	case I915_GEM_DOMAIN_RENDER:
325 		if (gpu_write_needs_clflush(obj))
326 			obj->cache_dirty = true;
327 		break;
328 	}
329 
330 	obj->write_domain = 0;
331 }
332 
333 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
334 					 enum fb_op_origin origin)
335 {
336 	struct intel_frontbuffer *front;
337 
338 	front = __intel_frontbuffer_get(obj);
339 	if (front) {
340 		intel_frontbuffer_flush(front, origin);
341 		intel_frontbuffer_put(front);
342 	}
343 }
344 
345 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
346 					      enum fb_op_origin origin)
347 {
348 	struct intel_frontbuffer *front;
349 
350 	front = __intel_frontbuffer_get(obj);
351 	if (front) {
352 		intel_frontbuffer_invalidate(front, origin);
353 		intel_frontbuffer_put(front);
354 	}
355 }
356 
357 void i915_gem_init__objects(struct drm_i915_private *i915)
358 {
359 	INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
360 }
361 
362 static void i915_global_objects_shrink(void)
363 {
364 	kmem_cache_shrink(global.slab_objects);
365 }
366 
367 static void i915_global_objects_exit(void)
368 {
369 	kmem_cache_destroy(global.slab_objects);
370 }
371 
372 static struct i915_global_object global = { {
373 	.shrink = i915_global_objects_shrink,
374 	.exit = i915_global_objects_exit,
375 } };
376 
377 int __init i915_global_objects_init(void)
378 {
379 	global.slab_objects =
380 		KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
381 	if (!global.slab_objects)
382 		return -ENOMEM;
383 
384 	i915_global_register(&global.base);
385 	return 0;
386 }
387 
388 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
389 #include "selftests/huge_gem_object.c"
390 #include "selftests/huge_pages.c"
391 #include "selftests/i915_gem_object.c"
392 #include "selftests/i915_gem_coherency.c"
393 #endif
394