1 /* $NetBSD: i915_gem_object.c,v 1.9 2021/12/19 12:12:31 riastradh Exp $ */
2
3 /*
4 * Copyright © 2017 Intel Corporation
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
23 * IN THE SOFTWARE.
24 *
25 */
26
27 #include <sys/cdefs.h>
28 __KERNEL_RCSID(0, "$NetBSD: i915_gem_object.c,v 1.9 2021/12/19 12:12:31 riastradh Exp $");
29
30 #include <linux/bitmap.h>
31 #include <linux/sched/mm.h>
32
33 #include "display/intel_frontbuffer.h"
34 #include "gt/intel_gt.h"
35 #include "i915_drv.h"
36 #include "i915_gem_clflush.h"
37 #include "i915_gem_context.h"
38 #include "i915_gem_mman.h"
39 #include "i915_gem_object.h"
40 #include "i915_globals.h"
41 #include "i915_trace.h"
42
43 #include <linux/nbsd-namespace.h>
44
45 static struct i915_global_object {
46 struct i915_global base;
47 struct kmem_cache *slab_objects;
48 } global;
49
i915_gem_object_alloc(void)50 struct drm_i915_gem_object *i915_gem_object_alloc(void)
51 {
52 return kmem_cache_zalloc(global.slab_objects, GFP_KERNEL);
53 }
54
i915_gem_object_free(struct drm_i915_gem_object * obj)55 void i915_gem_object_free(struct drm_i915_gem_object *obj)
56 {
57 return kmem_cache_free(global.slab_objects, obj);
58 }
59
i915_gem_object_init(struct drm_i915_gem_object * obj,const struct drm_i915_gem_object_ops * ops,struct lock_class_key * key)60 void i915_gem_object_init(struct drm_i915_gem_object *obj,
61 const struct drm_i915_gem_object_ops *ops,
62 struct lock_class_key *key)
63 {
64 __mutex_init(&obj->mm.lock, "obj->mm.lock", key);
65
66 spin_lock_init(&obj->vma.lock);
67 INIT_LIST_HEAD(&obj->vma.list);
68 i915_vma_tree_init(obj);
69
70 INIT_LIST_HEAD(&obj->mm.link);
71
72 INIT_LIST_HEAD(&obj->lut_list);
73
74 spin_lock_init(&obj->mmo.lock);
75 #ifdef __NetBSD__
76 memset(obj->mmo.offsets, 0, sizeof(obj->mmo.offsets));
77 #else
78 obj->mmo.offsets = RB_ROOT;
79 #endif
80
81 init_rcu_head(&obj->rcu);
82
83 obj->ops = ops;
84
85 obj->mm.madv = I915_MADV_WILLNEED;
86 #ifndef __NetBSD__
87 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
88 mutex_init(&obj->mm.get_page.lock);
89 #endif
90 }
91
92 /**
93 * Mark up the object's coherency levels for a given cache_level
94 * @obj: #drm_i915_gem_object
95 * @cache_level: cache level
96 */
i915_gem_object_set_cache_coherency(struct drm_i915_gem_object * obj,unsigned int cache_level)97 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
98 unsigned int cache_level)
99 {
100 obj->cache_level = cache_level;
101
102 if (cache_level != I915_CACHE_NONE)
103 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
104 I915_BO_CACHE_COHERENT_FOR_WRITE);
105 else if (HAS_LLC(to_i915(obj->base.dev)))
106 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
107 else
108 obj->cache_coherent = 0;
109
110 obj->cache_dirty =
111 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE);
112 }
113
i915_gem_close_object(struct drm_gem_object * gem,struct drm_file * file)114 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
115 {
116 struct drm_i915_gem_object *obj = to_intel_bo(gem);
117 struct drm_i915_file_private *fpriv = file->driver_priv;
118 struct i915_mmap_offset *mmo, *mn;
119 struct i915_lut_handle *lut, *ln;
120 LIST_HEAD(close);
121
122 i915_gem_object_lock(obj);
123 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
124 struct i915_gem_context *ctx = lut->ctx;
125
126 if (ctx->file_priv != fpriv)
127 continue;
128
129 i915_gem_context_get(ctx);
130 list_move(&lut->obj_link, &close);
131 }
132 i915_gem_object_unlock(obj);
133
134 #ifdef __NetBSD__
135 __USE(mn);
136 for (enum i915_mmap_type t = 0; t < I915_MMAP_NTYPES; t++) {
137 if ((mmo = obj->mmo.offsets[t]) == NULL)
138 continue;
139 drm_vma_node_revoke(&mmo->vma_node, file);
140 }
141 #else
142 spin_lock(&obj->mmo.lock);
143 rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
144 drm_vma_node_revoke(&mmo->vma_node, file);
145 spin_unlock(&obj->mmo.lock);
146 #endif
147
148 list_for_each_entry_safe(lut, ln, &close, obj_link) {
149 struct i915_gem_context *ctx = lut->ctx;
150 struct i915_vma *vma;
151
152 /*
153 * We allow the process to have multiple handles to the same
154 * vma, in the same fd namespace, by virtue of flink/open.
155 */
156
157 mutex_lock(&ctx->mutex);
158 vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
159 if (vma) {
160 GEM_BUG_ON(vma->obj != obj);
161 GEM_BUG_ON(!atomic_read(&vma->open_count));
162 if (atomic_dec_and_test(&vma->open_count) &&
163 !i915_vma_is_ggtt(vma))
164 i915_vma_close(vma);
165 }
166 mutex_unlock(&ctx->mutex);
167
168 i915_gem_context_put(lut->ctx);
169 i915_lut_handle_free(lut);
170 i915_gem_object_put(obj);
171 }
172 }
173
__i915_gem_free_object_rcu(struct rcu_head * head)174 static void __i915_gem_free_object_rcu(struct rcu_head *head)
175 {
176 struct drm_i915_gem_object *obj =
177 container_of(head, typeof(*obj), rcu);
178 struct drm_i915_private *i915 = to_i915(obj->base.dev);
179
180 /* i915_gem_object_init */
181 spin_lock_destroy(&obj->mmo.lock);
182 spin_lock_destroy(&obj->vma.lock);
183 mutex_destroy(&obj->mm.lock);
184
185 dma_resv_fini(&obj->base._resv);
186 drm_vma_node_destroy(&obj->base.vma_node);
187 i915_gem_object_free(obj);
188
189 GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
190 atomic_dec(&i915->mm.free_count);
191 }
192
__i915_gem_free_objects(struct drm_i915_private * i915,struct llist_node * freed)193 static void __i915_gem_free_objects(struct drm_i915_private *i915,
194 struct llist_node *freed)
195 {
196 struct drm_i915_gem_object *obj, *on;
197 intel_wakeref_t wakeref;
198
199 wakeref = intel_runtime_pm_get(&i915->runtime_pm);
200 llist_for_each_entry_safe(obj, on, freed, freed) {
201 struct i915_mmap_offset *mmo, *mn;
202
203 trace_i915_gem_object_destroy(obj);
204
205 if (!list_empty(&obj->vma.list)) {
206 struct i915_vma *vma;
207
208 /*
209 * Note that the vma keeps an object reference while
210 * it is active, so it *should* not sleep while we
211 * destroy it. Our debug code errs insits it *might*.
212 * For the moment, play along.
213 */
214 spin_lock(&obj->vma.lock);
215 while ((vma = list_first_entry_or_null(&obj->vma.list,
216 struct i915_vma,
217 obj_link))) {
218 GEM_BUG_ON(vma->obj != obj);
219 spin_unlock(&obj->vma.lock);
220
221 __i915_vma_put(vma);
222
223 spin_lock(&obj->vma.lock);
224 }
225 spin_unlock(&obj->vma.lock);
226 }
227
228 i915_gem_object_release_mmap(obj);
229
230 #ifdef __NetBSD__
231 __USE(mn);
232 for (enum i915_mmap_type t = 0; t < I915_MMAP_NTYPES; t++) {
233 if ((mmo = obj->mmo.offsets[t]) == NULL)
234 continue;
235 drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
236 &mmo->vma_node);
237 drm_vma_node_destroy(&mmo->vma_node);
238 kfree(mmo);
239 }
240 memset(obj->mmo.offsets, 0, sizeof(obj->mmo.offsets));
241 #else
242 rbtree_postorder_for_each_entry_safe(mmo, mn,
243 &obj->mmo.offsets,
244 offset) {
245 drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
246 &mmo->vma_node);
247 kfree(mmo);
248 }
249 obj->mmo.offsets = RB_ROOT;
250 #endif
251
252 GEM_BUG_ON(atomic_read(&obj->bind_count));
253 GEM_BUG_ON(obj->userfault_count);
254 GEM_BUG_ON(!list_empty(&obj->lut_list));
255
256 atomic_set(&obj->mm.pages_pin_count, 0);
257 __i915_gem_object_put_pages(obj);
258 GEM_BUG_ON(i915_gem_object_has_pages(obj));
259 bitmap_free(obj->bit_17);
260
261 if (obj->base.import_attach)
262 drm_prime_gem_destroy(&obj->base, NULL);
263
264 drm_gem_free_mmap_offset(&obj->base);
265
266 if (obj->ops->release)
267 obj->ops->release(obj);
268
269 /* But keep the pointer alive for RCU-protected lookups */
270 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
271 }
272 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
273 }
274
i915_gem_flush_free_objects(struct drm_i915_private * i915)275 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
276 {
277 struct llist_node *freed = llist_del_all(&i915->mm.free_list);
278
279 if (unlikely(freed))
280 __i915_gem_free_objects(i915, freed);
281 }
282
__i915_gem_free_work(struct work_struct * work)283 static void __i915_gem_free_work(struct work_struct *work)
284 {
285 struct drm_i915_private *i915 =
286 container_of(work, struct drm_i915_private, mm.free_work);
287
288 i915_gem_flush_free_objects(i915);
289 }
290
i915_gem_free_object(struct drm_gem_object * gem_obj)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 struct drm_i915_private *i915 = to_i915(obj->base.dev);
295
296 GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
297
298 /*
299 * Before we free the object, make sure any pure RCU-only
300 * read-side critical sections are complete, e.g.
301 * i915_gem_busy_ioctl(). For the corresponding synchronized
302 * lookup see i915_gem_object_lookup_rcu().
303 */
304 atomic_inc(&i915->mm.free_count);
305
306 /*
307 * This serializes freeing with the shrinker. Since the free
308 * is delayed, first by RCU then by the workqueue, we want the
309 * shrinker to be able to free pages of unreferenced objects,
310 * or else we may oom whilst there are plenty of deferred
311 * freed objects.
312 */
313 i915_gem_object_make_unshrinkable(obj);
314
315 /*
316 * Since we require blocking on struct_mutex to unbind the freed
317 * object from the GPU before releasing resources back to the
318 * system, we can not do that directly from the RCU callback (which may
319 * be a softirq context), but must instead then defer that work onto a
320 * kthread. We use the RCU callback rather than move the freed object
321 * directly onto the work queue so that we can mix between using the
322 * worker and performing frees directly from subsequent allocations for
323 * crude but effective memory throttling.
324 */
325 if (llist_add(&obj->freed, &i915->mm.free_list))
326 queue_work(i915->wq, &i915->mm.free_work);
327 }
328
gpu_write_needs_clflush(struct drm_i915_gem_object * obj)329 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
330 {
331 return !(obj->cache_level == I915_CACHE_NONE ||
332 obj->cache_level == I915_CACHE_WT);
333 }
334
335 void
i915_gem_object_flush_write_domain(struct drm_i915_gem_object * obj,unsigned int flush_domains)336 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj,
337 unsigned int flush_domains)
338 {
339 struct i915_vma *vma;
340
341 assert_object_held(obj);
342
343 if (!(obj->write_domain & flush_domains))
344 return;
345
346 switch (obj->write_domain) {
347 case I915_GEM_DOMAIN_GTT:
348 spin_lock(&obj->vma.lock);
349 for_each_ggtt_vma(vma, obj) {
350 if (i915_vma_unset_ggtt_write(vma))
351 intel_gt_flush_ggtt_writes(vma->vm->gt);
352 }
353 spin_unlock(&obj->vma.lock);
354
355 i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
356 break;
357
358 case I915_GEM_DOMAIN_WC:
359 wmb();
360 break;
361
362 case I915_GEM_DOMAIN_CPU:
363 i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
364 break;
365
366 case I915_GEM_DOMAIN_RENDER:
367 if (gpu_write_needs_clflush(obj))
368 obj->cache_dirty = true;
369 break;
370 }
371
372 obj->write_domain = 0;
373 }
374
__i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)375 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
376 enum fb_op_origin origin)
377 {
378 struct intel_frontbuffer *front;
379
380 front = __intel_frontbuffer_get(obj);
381 if (front) {
382 intel_frontbuffer_flush(front, origin);
383 intel_frontbuffer_put(front);
384 }
385 }
386
__i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)387 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
388 enum fb_op_origin origin)
389 {
390 struct intel_frontbuffer *front;
391
392 front = __intel_frontbuffer_get(obj);
393 if (front) {
394 intel_frontbuffer_invalidate(front, origin);
395 intel_frontbuffer_put(front);
396 }
397 }
398
i915_gem_init__objects(struct drm_i915_private * i915)399 void i915_gem_init__objects(struct drm_i915_private *i915)
400 {
401 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
402 }
403
i915_global_objects_shrink(void)404 static void i915_global_objects_shrink(void)
405 {
406 kmem_cache_shrink(global.slab_objects);
407 }
408
i915_global_objects_exit(void)409 static void i915_global_objects_exit(void)
410 {
411 kmem_cache_destroy(global.slab_objects);
412 }
413
414 static struct i915_global_object global = { {
415 .shrink = i915_global_objects_shrink,
416 .exit = i915_global_objects_exit,
417 } };
418
i915_global_objects_init(void)419 int __init i915_global_objects_init(void)
420 {
421 global.slab_objects =
422 KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
423 if (!global.slab_objects)
424 return -ENOMEM;
425
426 i915_global_register(&global.base);
427 return 0;
428 }
429
430 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
431 #include "selftests/huge_gem_object.c"
432 #include "selftests/huge_pages.c"
433 #include "selftests/i915_gem_object.c"
434 #include "selftests/i915_gem_coherency.c"
435 #endif
436