xref: /linux/drivers/gpu/drm/i915/gem/i915_gem_object.c (revision 5a7eeb8ba143d860050ecea924a8f074f02d8023)
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, ops->name ?: "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 	if (IS_ENABLED(CONFIG_LOCKDEP) && i915_gem_object_is_shrinkable(obj))
77 		i915_gem_shrinker_taints_mutex(to_i915(obj->base.dev),
78 					       &obj->mm.lock);
79 }
80 
81 /**
82  * Mark up the object's coherency levels for a given cache_level
83  * @obj: #drm_i915_gem_object
84  * @cache_level: cache level
85  */
86 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
87 					 unsigned int cache_level)
88 {
89 	obj->cache_level = cache_level;
90 
91 	if (cache_level != I915_CACHE_NONE)
92 		obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
93 				       I915_BO_CACHE_COHERENT_FOR_WRITE);
94 	else if (HAS_LLC(to_i915(obj->base.dev)))
95 		obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
96 	else
97 		obj->cache_coherent = 0;
98 
99 	obj->cache_dirty =
100 		!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE);
101 }
102 
103 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
104 {
105 	struct drm_i915_gem_object *obj = to_intel_bo(gem);
106 	struct drm_i915_file_private *fpriv = file->driver_priv;
107 	struct i915_mmap_offset *mmo, *mn;
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 	spin_lock(&obj->mmo.lock);
124 	rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
125 		drm_vma_node_revoke(&mmo->vma_node, file);
126 	spin_unlock(&obj->mmo.lock);
127 
128 	list_for_each_entry_safe(lut, ln, &close, obj_link) {
129 		struct i915_gem_context *ctx = lut->ctx;
130 		struct i915_vma *vma;
131 
132 		/*
133 		 * We allow the process to have multiple handles to the same
134 		 * vma, in the same fd namespace, by virtue of flink/open.
135 		 */
136 
137 		mutex_lock(&ctx->mutex);
138 		vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
139 		if (vma) {
140 			GEM_BUG_ON(vma->obj != obj);
141 			GEM_BUG_ON(!atomic_read(&vma->open_count));
142 			i915_vma_close(vma);
143 		}
144 		mutex_unlock(&ctx->mutex);
145 
146 		i915_gem_context_put(lut->ctx);
147 		i915_lut_handle_free(lut);
148 		i915_gem_object_put(obj);
149 	}
150 }
151 
152 static void __i915_gem_free_object_rcu(struct rcu_head *head)
153 {
154 	struct drm_i915_gem_object *obj =
155 		container_of(head, typeof(*obj), rcu);
156 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
157 
158 	dma_resv_fini(&obj->base._resv);
159 	i915_gem_object_free(obj);
160 
161 	GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
162 	atomic_dec(&i915->mm.free_count);
163 }
164 
165 static void __i915_gem_free_objects(struct drm_i915_private *i915,
166 				    struct llist_node *freed)
167 {
168 	struct drm_i915_gem_object *obj, *on;
169 
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(obj->userfault_count);
210 		GEM_BUG_ON(!list_empty(&obj->lut_list));
211 
212 		atomic_set(&obj->mm.pages_pin_count, 0);
213 		__i915_gem_object_put_pages(obj);
214 		GEM_BUG_ON(i915_gem_object_has_pages(obj));
215 		bitmap_free(obj->bit_17);
216 
217 		if (obj->base.import_attach)
218 			drm_prime_gem_destroy(&obj->base, NULL);
219 
220 		drm_gem_free_mmap_offset(&obj->base);
221 
222 		if (obj->ops->release)
223 			obj->ops->release(obj);
224 
225 		/* But keep the pointer alive for RCU-protected lookups */
226 		call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
227 		cond_resched();
228 	}
229 }
230 
231 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
232 {
233 	struct llist_node *freed = llist_del_all(&i915->mm.free_list);
234 
235 	if (unlikely(freed))
236 		__i915_gem_free_objects(i915, freed);
237 }
238 
239 static void __i915_gem_free_work(struct work_struct *work)
240 {
241 	struct drm_i915_private *i915 =
242 		container_of(work, struct drm_i915_private, mm.free_work);
243 
244 	i915_gem_flush_free_objects(i915);
245 }
246 
247 void i915_gem_free_object(struct drm_gem_object *gem_obj)
248 {
249 	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
250 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
251 
252 	GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
253 
254 	/*
255 	 * Before we free the object, make sure any pure RCU-only
256 	 * read-side critical sections are complete, e.g.
257 	 * i915_gem_busy_ioctl(). For the corresponding synchronized
258 	 * lookup see i915_gem_object_lookup_rcu().
259 	 */
260 	atomic_inc(&i915->mm.free_count);
261 
262 	/*
263 	 * This serializes freeing with the shrinker. Since the free
264 	 * is delayed, first by RCU then by the workqueue, we want the
265 	 * shrinker to be able to free pages of unreferenced objects,
266 	 * or else we may oom whilst there are plenty of deferred
267 	 * freed objects.
268 	 */
269 	i915_gem_object_make_unshrinkable(obj);
270 
271 	/*
272 	 * Since we require blocking on struct_mutex to unbind the freed
273 	 * object from the GPU before releasing resources back to the
274 	 * system, we can not do that directly from the RCU callback (which may
275 	 * be a softirq context), but must instead then defer that work onto a
276 	 * kthread. We use the RCU callback rather than move the freed object
277 	 * directly onto the work queue so that we can mix between using the
278 	 * worker and performing frees directly from subsequent allocations for
279 	 * crude but effective memory throttling.
280 	 */
281 	if (llist_add(&obj->freed, &i915->mm.free_list))
282 		queue_work(i915->wq, &i915->mm.free_work);
283 }
284 
285 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
286 {
287 	return !(obj->cache_level == I915_CACHE_NONE ||
288 		 obj->cache_level == I915_CACHE_WT);
289 }
290 
291 void
292 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj,
293 				   unsigned int flush_domains)
294 {
295 	struct i915_vma *vma;
296 
297 	assert_object_held(obj);
298 
299 	if (!(obj->write_domain & flush_domains))
300 		return;
301 
302 	switch (obj->write_domain) {
303 	case I915_GEM_DOMAIN_GTT:
304 		spin_lock(&obj->vma.lock);
305 		for_each_ggtt_vma(vma, obj) {
306 			if (i915_vma_unset_ggtt_write(vma))
307 				intel_gt_flush_ggtt_writes(vma->vm->gt);
308 		}
309 		spin_unlock(&obj->vma.lock);
310 
311 		i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
312 		break;
313 
314 	case I915_GEM_DOMAIN_WC:
315 		wmb();
316 		break;
317 
318 	case I915_GEM_DOMAIN_CPU:
319 		i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
320 		break;
321 
322 	case I915_GEM_DOMAIN_RENDER:
323 		if (gpu_write_needs_clflush(obj))
324 			obj->cache_dirty = true;
325 		break;
326 	}
327 
328 	obj->write_domain = 0;
329 }
330 
331 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
332 					 enum fb_op_origin origin)
333 {
334 	struct intel_frontbuffer *front;
335 
336 	front = __intel_frontbuffer_get(obj);
337 	if (front) {
338 		intel_frontbuffer_flush(front, origin);
339 		intel_frontbuffer_put(front);
340 	}
341 }
342 
343 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
344 					      enum fb_op_origin origin)
345 {
346 	struct intel_frontbuffer *front;
347 
348 	front = __intel_frontbuffer_get(obj);
349 	if (front) {
350 		intel_frontbuffer_invalidate(front, origin);
351 		intel_frontbuffer_put(front);
352 	}
353 }
354 
355 void i915_gem_init__objects(struct drm_i915_private *i915)
356 {
357 	INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
358 }
359 
360 static void i915_global_objects_shrink(void)
361 {
362 	kmem_cache_shrink(global.slab_objects);
363 }
364 
365 static void i915_global_objects_exit(void)
366 {
367 	kmem_cache_destroy(global.slab_objects);
368 }
369 
370 static struct i915_global_object global = { {
371 	.shrink = i915_global_objects_shrink,
372 	.exit = i915_global_objects_exit,
373 } };
374 
375 int __init i915_global_objects_init(void)
376 {
377 	global.slab_objects =
378 		KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
379 	if (!global.slab_objects)
380 		return -ENOMEM;
381 
382 	i915_global_register(&global.base);
383 	return 0;
384 }
385 
386 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
387 #include "selftests/huge_gem_object.c"
388 #include "selftests/huge_pages.c"
389 #include "selftests/i915_gem_object.c"
390 #include "selftests/i915_gem_coherency.c"
391 #endif
392