]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0-only | |
2 | /* | |
3 | * Copyright © 2015 Broadcom | |
4 | */ | |
5 | ||
6 | /** | |
7 | * DOC: VC4 GEM BO management support | |
8 | * | |
9 | * The VC4 GPU architecture (both scanout and rendering) has direct | |
10 | * access to system memory with no MMU in between. To support it, we | |
11 | * use the GEM CMA helper functions to allocate contiguous ranges of | |
12 | * physical memory for our BOs. | |
13 | * | |
14 | * Since the CMA allocator is very slow, we keep a cache of recently | |
15 | * freed BOs around so that the kernel's allocation of objects for 3D | |
16 | * rendering can return quickly. | |
17 | */ | |
18 | ||
19 | #include <linux/dma-buf.h> | |
20 | ||
21 | #include "vc4_drv.h" | |
22 | #include "uapi/drm/vc4_drm.h" | |
23 | ||
24 | static const char * const bo_type_names[] = { | |
25 | "kernel", | |
26 | "V3D", | |
27 | "V3D shader", | |
28 | "dumb", | |
29 | "binner", | |
30 | "RCL", | |
31 | "BCL", | |
32 | "kernel BO cache", | |
33 | }; | |
34 | ||
35 | static bool is_user_label(int label) | |
36 | { | |
37 | return label >= VC4_BO_TYPE_COUNT; | |
38 | } | |
39 | ||
40 | static void vc4_bo_stats_print(struct drm_printer *p, struct vc4_dev *vc4) | |
41 | { | |
42 | int i; | |
43 | ||
44 | for (i = 0; i < vc4->num_labels; i++) { | |
45 | if (!vc4->bo_labels[i].num_allocated) | |
46 | continue; | |
47 | ||
48 | drm_printf(p, "%30s: %6dkb BOs (%d)\n", | |
49 | vc4->bo_labels[i].name, | |
50 | vc4->bo_labels[i].size_allocated / 1024, | |
51 | vc4->bo_labels[i].num_allocated); | |
52 | } | |
53 | ||
54 | mutex_lock(&vc4->purgeable.lock); | |
55 | if (vc4->purgeable.num) | |
56 | drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "userspace BO cache", | |
57 | vc4->purgeable.size / 1024, vc4->purgeable.num); | |
58 | ||
59 | if (vc4->purgeable.purged_num) | |
60 | drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "total purged BO", | |
61 | vc4->purgeable.purged_size / 1024, | |
62 | vc4->purgeable.purged_num); | |
63 | mutex_unlock(&vc4->purgeable.lock); | |
64 | } | |
65 | ||
66 | static int vc4_bo_stats_debugfs(struct seq_file *m, void *unused) | |
67 | { | |
68 | struct drm_info_node *node = (struct drm_info_node *)m->private; | |
69 | struct drm_device *dev = node->minor->dev; | |
70 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
71 | struct drm_printer p = drm_seq_file_printer(m); | |
72 | ||
73 | vc4_bo_stats_print(&p, vc4); | |
74 | ||
75 | return 0; | |
76 | } | |
77 | ||
78 | /* Takes ownership of *name and returns the appropriate slot for it in | |
79 | * the bo_labels[] array, extending it as necessary. | |
80 | * | |
81 | * This is inefficient and could use a hash table instead of walking | |
82 | * an array and strcmp()ing. However, the assumption is that user | |
83 | * labeling will be infrequent (scanout buffers and other long-lived | |
84 | * objects, or debug driver builds), so we can live with it for now. | |
85 | */ | |
86 | static int vc4_get_user_label(struct vc4_dev *vc4, const char *name) | |
87 | { | |
88 | int i; | |
89 | int free_slot = -1; | |
90 | ||
91 | for (i = 0; i < vc4->num_labels; i++) { | |
92 | if (!vc4->bo_labels[i].name) { | |
93 | free_slot = i; | |
94 | } else if (strcmp(vc4->bo_labels[i].name, name) == 0) { | |
95 | kfree(name); | |
96 | return i; | |
97 | } | |
98 | } | |
99 | ||
100 | if (free_slot != -1) { | |
101 | WARN_ON(vc4->bo_labels[free_slot].num_allocated != 0); | |
102 | vc4->bo_labels[free_slot].name = name; | |
103 | return free_slot; | |
104 | } else { | |
105 | u32 new_label_count = vc4->num_labels + 1; | |
106 | struct vc4_label *new_labels = | |
107 | krealloc(vc4->bo_labels, | |
108 | new_label_count * sizeof(*new_labels), | |
109 | GFP_KERNEL); | |
110 | ||
111 | if (!new_labels) { | |
112 | kfree(name); | |
113 | return -1; | |
114 | } | |
115 | ||
116 | free_slot = vc4->num_labels; | |
117 | vc4->bo_labels = new_labels; | |
118 | vc4->num_labels = new_label_count; | |
119 | ||
120 | vc4->bo_labels[free_slot].name = name; | |
121 | vc4->bo_labels[free_slot].num_allocated = 0; | |
122 | vc4->bo_labels[free_slot].size_allocated = 0; | |
123 | ||
124 | return free_slot; | |
125 | } | |
126 | } | |
127 | ||
128 | static void vc4_bo_set_label(struct drm_gem_object *gem_obj, int label) | |
129 | { | |
130 | struct vc4_bo *bo = to_vc4_bo(gem_obj); | |
131 | struct vc4_dev *vc4 = to_vc4_dev(gem_obj->dev); | |
132 | ||
133 | lockdep_assert_held(&vc4->bo_lock); | |
134 | ||
135 | if (label != -1) { | |
136 | vc4->bo_labels[label].num_allocated++; | |
137 | vc4->bo_labels[label].size_allocated += gem_obj->size; | |
138 | } | |
139 | ||
140 | vc4->bo_labels[bo->label].num_allocated--; | |
141 | vc4->bo_labels[bo->label].size_allocated -= gem_obj->size; | |
142 | ||
143 | if (vc4->bo_labels[bo->label].num_allocated == 0 && | |
144 | is_user_label(bo->label)) { | |
145 | /* Free user BO label slots on last unreference. | |
146 | * Slots are just where we track the stats for a given | |
147 | * name, and once a name is unused we can reuse that | |
148 | * slot. | |
149 | */ | |
150 | kfree(vc4->bo_labels[bo->label].name); | |
151 | vc4->bo_labels[bo->label].name = NULL; | |
152 | } | |
153 | ||
154 | bo->label = label; | |
155 | } | |
156 | ||
157 | static uint32_t bo_page_index(size_t size) | |
158 | { | |
159 | return (size / PAGE_SIZE) - 1; | |
160 | } | |
161 | ||
162 | static void vc4_bo_destroy(struct vc4_bo *bo) | |
163 | { | |
164 | struct drm_gem_object *obj = &bo->base.base; | |
165 | struct vc4_dev *vc4 = to_vc4_dev(obj->dev); | |
166 | ||
167 | lockdep_assert_held(&vc4->bo_lock); | |
168 | ||
169 | vc4_bo_set_label(obj, -1); | |
170 | ||
171 | if (bo->validated_shader) { | |
172 | kfree(bo->validated_shader->uniform_addr_offsets); | |
173 | kfree(bo->validated_shader->texture_samples); | |
174 | kfree(bo->validated_shader); | |
175 | bo->validated_shader = NULL; | |
176 | } | |
177 | ||
178 | drm_gem_cma_free_object(obj); | |
179 | } | |
180 | ||
181 | static void vc4_bo_remove_from_cache(struct vc4_bo *bo) | |
182 | { | |
183 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); | |
184 | ||
185 | lockdep_assert_held(&vc4->bo_lock); | |
186 | list_del(&bo->unref_head); | |
187 | list_del(&bo->size_head); | |
188 | } | |
189 | ||
190 | static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev, | |
191 | size_t size) | |
192 | { | |
193 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
194 | uint32_t page_index = bo_page_index(size); | |
195 | ||
196 | if (vc4->bo_cache.size_list_size <= page_index) { | |
197 | uint32_t new_size = max(vc4->bo_cache.size_list_size * 2, | |
198 | page_index + 1); | |
199 | struct list_head *new_list; | |
200 | uint32_t i; | |
201 | ||
202 | new_list = kmalloc_array(new_size, sizeof(struct list_head), | |
203 | GFP_KERNEL); | |
204 | if (!new_list) | |
205 | return NULL; | |
206 | ||
207 | /* Rebase the old cached BO lists to their new list | |
208 | * head locations. | |
209 | */ | |
210 | for (i = 0; i < vc4->bo_cache.size_list_size; i++) { | |
211 | struct list_head *old_list = | |
212 | &vc4->bo_cache.size_list[i]; | |
213 | ||
214 | if (list_empty(old_list)) | |
215 | INIT_LIST_HEAD(&new_list[i]); | |
216 | else | |
217 | list_replace(old_list, &new_list[i]); | |
218 | } | |
219 | /* And initialize the brand new BO list heads. */ | |
220 | for (i = vc4->bo_cache.size_list_size; i < new_size; i++) | |
221 | INIT_LIST_HEAD(&new_list[i]); | |
222 | ||
223 | kfree(vc4->bo_cache.size_list); | |
224 | vc4->bo_cache.size_list = new_list; | |
225 | vc4->bo_cache.size_list_size = new_size; | |
226 | } | |
227 | ||
228 | return &vc4->bo_cache.size_list[page_index]; | |
229 | } | |
230 | ||
231 | static void vc4_bo_cache_purge(struct drm_device *dev) | |
232 | { | |
233 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
234 | ||
235 | mutex_lock(&vc4->bo_lock); | |
236 | while (!list_empty(&vc4->bo_cache.time_list)) { | |
237 | struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list, | |
238 | struct vc4_bo, unref_head); | |
239 | vc4_bo_remove_from_cache(bo); | |
240 | vc4_bo_destroy(bo); | |
241 | } | |
242 | mutex_unlock(&vc4->bo_lock); | |
243 | } | |
244 | ||
245 | void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo) | |
246 | { | |
247 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); | |
248 | ||
249 | mutex_lock(&vc4->purgeable.lock); | |
250 | list_add_tail(&bo->size_head, &vc4->purgeable.list); | |
251 | vc4->purgeable.num++; | |
252 | vc4->purgeable.size += bo->base.base.size; | |
253 | mutex_unlock(&vc4->purgeable.lock); | |
254 | } | |
255 | ||
256 | static void vc4_bo_remove_from_purgeable_pool_locked(struct vc4_bo *bo) | |
257 | { | |
258 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); | |
259 | ||
260 | /* list_del_init() is used here because the caller might release | |
261 | * the purgeable lock in order to acquire the madv one and update the | |
262 | * madv status. | |
263 | * During this short period of time a user might decide to mark | |
264 | * the BO as unpurgeable, and if bo->madv is set to | |
265 | * VC4_MADV_DONTNEED it will try to remove the BO from the | |
266 | * purgeable list which will fail if the ->next/prev fields | |
267 | * are set to LIST_POISON1/LIST_POISON2 (which is what | |
268 | * list_del() does). | |
269 | * Re-initializing the list element guarantees that list_del() | |
270 | * will work correctly even if it's a NOP. | |
271 | */ | |
272 | list_del_init(&bo->size_head); | |
273 | vc4->purgeable.num--; | |
274 | vc4->purgeable.size -= bo->base.base.size; | |
275 | } | |
276 | ||
277 | void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo) | |
278 | { | |
279 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); | |
280 | ||
281 | mutex_lock(&vc4->purgeable.lock); | |
282 | vc4_bo_remove_from_purgeable_pool_locked(bo); | |
283 | mutex_unlock(&vc4->purgeable.lock); | |
284 | } | |
285 | ||
286 | static void vc4_bo_purge(struct drm_gem_object *obj) | |
287 | { | |
288 | struct vc4_bo *bo = to_vc4_bo(obj); | |
289 | struct drm_device *dev = obj->dev; | |
290 | ||
291 | WARN_ON(!mutex_is_locked(&bo->madv_lock)); | |
292 | WARN_ON(bo->madv != VC4_MADV_DONTNEED); | |
293 | ||
294 | drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping); | |
295 | ||
296 | dma_free_wc(dev->dev, obj->size, bo->base.vaddr, bo->base.paddr); | |
297 | bo->base.vaddr = NULL; | |
298 | bo->madv = __VC4_MADV_PURGED; | |
299 | } | |
300 | ||
301 | static void vc4_bo_userspace_cache_purge(struct drm_device *dev) | |
302 | { | |
303 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
304 | ||
305 | mutex_lock(&vc4->purgeable.lock); | |
306 | while (!list_empty(&vc4->purgeable.list)) { | |
307 | struct vc4_bo *bo = list_first_entry(&vc4->purgeable.list, | |
308 | struct vc4_bo, size_head); | |
309 | struct drm_gem_object *obj = &bo->base.base; | |
310 | size_t purged_size = 0; | |
311 | ||
312 | vc4_bo_remove_from_purgeable_pool_locked(bo); | |
313 | ||
314 | /* Release the purgeable lock while we're purging the BO so | |
315 | * that other people can continue inserting things in the | |
316 | * purgeable pool without having to wait for all BOs to be | |
317 | * purged. | |
318 | */ | |
319 | mutex_unlock(&vc4->purgeable.lock); | |
320 | mutex_lock(&bo->madv_lock); | |
321 | ||
322 | /* Since we released the purgeable pool lock before acquiring | |
323 | * the BO madv one, the user may have marked the BO as WILLNEED | |
324 | * and re-used it in the meantime. | |
325 | * Before purging the BO we need to make sure | |
326 | * - it is still marked as DONTNEED | |
327 | * - it has not been re-inserted in the purgeable list | |
328 | * - it is not used by HW blocks | |
329 | * If one of these conditions is not met, just skip the entry. | |
330 | */ | |
331 | if (bo->madv == VC4_MADV_DONTNEED && | |
332 | list_empty(&bo->size_head) && | |
333 | !refcount_read(&bo->usecnt)) { | |
334 | purged_size = bo->base.base.size; | |
335 | vc4_bo_purge(obj); | |
336 | } | |
337 | mutex_unlock(&bo->madv_lock); | |
338 | mutex_lock(&vc4->purgeable.lock); | |
339 | ||
340 | if (purged_size) { | |
341 | vc4->purgeable.purged_size += purged_size; | |
342 | vc4->purgeable.purged_num++; | |
343 | } | |
344 | } | |
345 | mutex_unlock(&vc4->purgeable.lock); | |
346 | } | |
347 | ||
348 | static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev, | |
349 | uint32_t size, | |
350 | enum vc4_kernel_bo_type type) | |
351 | { | |
352 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
353 | uint32_t page_index = bo_page_index(size); | |
354 | struct vc4_bo *bo = NULL; | |
355 | ||
356 | size = roundup(size, PAGE_SIZE); | |
357 | ||
358 | mutex_lock(&vc4->bo_lock); | |
359 | if (page_index >= vc4->bo_cache.size_list_size) | |
360 | goto out; | |
361 | ||
362 | if (list_empty(&vc4->bo_cache.size_list[page_index])) | |
363 | goto out; | |
364 | ||
365 | bo = list_first_entry(&vc4->bo_cache.size_list[page_index], | |
366 | struct vc4_bo, size_head); | |
367 | vc4_bo_remove_from_cache(bo); | |
368 | kref_init(&bo->base.base.refcount); | |
369 | ||
370 | out: | |
371 | if (bo) | |
372 | vc4_bo_set_label(&bo->base.base, type); | |
373 | mutex_unlock(&vc4->bo_lock); | |
374 | return bo; | |
375 | } | |
376 | ||
377 | /** | |
378 | * vc4_gem_create_object - Implementation of driver->gem_create_object. | |
379 | * @dev: DRM device | |
380 | * @size: Size in bytes of the memory the object will reference | |
381 | * | |
382 | * This lets the CMA helpers allocate object structs for us, and keep | |
383 | * our BO stats correct. | |
384 | */ | |
385 | struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size) | |
386 | { | |
387 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
388 | struct vc4_bo *bo; | |
389 | ||
390 | bo = kzalloc(sizeof(*bo), GFP_KERNEL); | |
391 | if (!bo) | |
392 | return ERR_PTR(-ENOMEM); | |
393 | ||
394 | bo->madv = VC4_MADV_WILLNEED; | |
395 | refcount_set(&bo->usecnt, 0); | |
396 | mutex_init(&bo->madv_lock); | |
397 | mutex_lock(&vc4->bo_lock); | |
398 | bo->label = VC4_BO_TYPE_KERNEL; | |
399 | vc4->bo_labels[VC4_BO_TYPE_KERNEL].num_allocated++; | |
400 | vc4->bo_labels[VC4_BO_TYPE_KERNEL].size_allocated += size; | |
401 | mutex_unlock(&vc4->bo_lock); | |
402 | ||
403 | return &bo->base.base; | |
404 | } | |
405 | ||
406 | struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size, | |
407 | bool allow_unzeroed, enum vc4_kernel_bo_type type) | |
408 | { | |
409 | size_t size = roundup(unaligned_size, PAGE_SIZE); | |
410 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
411 | struct drm_gem_cma_object *cma_obj; | |
412 | struct vc4_bo *bo; | |
413 | ||
414 | if (size == 0) | |
415 | return ERR_PTR(-EINVAL); | |
416 | ||
417 | /* First, try to get a vc4_bo from the kernel BO cache. */ | |
418 | bo = vc4_bo_get_from_cache(dev, size, type); | |
419 | if (bo) { | |
420 | if (!allow_unzeroed) | |
421 | memset(bo->base.vaddr, 0, bo->base.base.size); | |
422 | return bo; | |
423 | } | |
424 | ||
425 | cma_obj = drm_gem_cma_create(dev, size); | |
426 | if (IS_ERR(cma_obj)) { | |
427 | /* | |
428 | * If we've run out of CMA memory, kill the cache of | |
429 | * CMA allocations we've got laying around and try again. | |
430 | */ | |
431 | vc4_bo_cache_purge(dev); | |
432 | cma_obj = drm_gem_cma_create(dev, size); | |
433 | } | |
434 | ||
435 | if (IS_ERR(cma_obj)) { | |
436 | /* | |
437 | * Still not enough CMA memory, purge the userspace BO | |
438 | * cache and retry. | |
439 | * This is sub-optimal since we purge the whole userspace | |
440 | * BO cache which forces user that want to re-use the BO to | |
441 | * restore its initial content. | |
442 | * Ideally, we should purge entries one by one and retry | |
443 | * after each to see if CMA allocation succeeds. Or even | |
444 | * better, try to find an entry with at least the same | |
445 | * size. | |
446 | */ | |
447 | vc4_bo_userspace_cache_purge(dev); | |
448 | cma_obj = drm_gem_cma_create(dev, size); | |
449 | } | |
450 | ||
451 | if (IS_ERR(cma_obj)) { | |
452 | struct drm_printer p = drm_info_printer(vc4->dev->dev); | |
453 | DRM_ERROR("Failed to allocate from CMA:\n"); | |
454 | vc4_bo_stats_print(&p, vc4); | |
455 | return ERR_PTR(-ENOMEM); | |
456 | } | |
457 | bo = to_vc4_bo(&cma_obj->base); | |
458 | ||
459 | /* By default, BOs do not support the MADV ioctl. This will be enabled | |
460 | * only on BOs that are exposed to userspace (V3D, V3D_SHADER and DUMB | |
461 | * BOs). | |
462 | */ | |
463 | bo->madv = __VC4_MADV_NOTSUPP; | |
464 | ||
465 | mutex_lock(&vc4->bo_lock); | |
466 | vc4_bo_set_label(&cma_obj->base, type); | |
467 | mutex_unlock(&vc4->bo_lock); | |
468 | ||
469 | return bo; | |
470 | } | |
471 | ||
472 | int vc4_dumb_create(struct drm_file *file_priv, | |
473 | struct drm_device *dev, | |
474 | struct drm_mode_create_dumb *args) | |
475 | { | |
476 | int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8); | |
477 | struct vc4_bo *bo = NULL; | |
478 | int ret; | |
479 | ||
480 | if (args->pitch < min_pitch) | |
481 | args->pitch = min_pitch; | |
482 | ||
483 | if (args->size < args->pitch * args->height) | |
484 | args->size = args->pitch * args->height; | |
485 | ||
486 | bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_DUMB); | |
487 | if (IS_ERR(bo)) | |
488 | return PTR_ERR(bo); | |
489 | ||
490 | bo->madv = VC4_MADV_WILLNEED; | |
491 | ||
492 | ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle); | |
493 | drm_gem_object_put_unlocked(&bo->base.base); | |
494 | ||
495 | return ret; | |
496 | } | |
497 | ||
498 | static void vc4_bo_cache_free_old(struct drm_device *dev) | |
499 | { | |
500 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
501 | unsigned long expire_time = jiffies - msecs_to_jiffies(1000); | |
502 | ||
503 | lockdep_assert_held(&vc4->bo_lock); | |
504 | ||
505 | while (!list_empty(&vc4->bo_cache.time_list)) { | |
506 | struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list, | |
507 | struct vc4_bo, unref_head); | |
508 | if (time_before(expire_time, bo->free_time)) { | |
509 | mod_timer(&vc4->bo_cache.time_timer, | |
510 | round_jiffies_up(jiffies + | |
511 | msecs_to_jiffies(1000))); | |
512 | return; | |
513 | } | |
514 | ||
515 | vc4_bo_remove_from_cache(bo); | |
516 | vc4_bo_destroy(bo); | |
517 | } | |
518 | } | |
519 | ||
520 | /* Called on the last userspace/kernel unreference of the BO. Returns | |
521 | * it to the BO cache if possible, otherwise frees it. | |
522 | */ | |
523 | void vc4_free_object(struct drm_gem_object *gem_bo) | |
524 | { | |
525 | struct drm_device *dev = gem_bo->dev; | |
526 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
527 | struct vc4_bo *bo = to_vc4_bo(gem_bo); | |
528 | struct list_head *cache_list; | |
529 | ||
530 | /* Remove the BO from the purgeable list. */ | |
531 | mutex_lock(&bo->madv_lock); | |
532 | if (bo->madv == VC4_MADV_DONTNEED && !refcount_read(&bo->usecnt)) | |
533 | vc4_bo_remove_from_purgeable_pool(bo); | |
534 | mutex_unlock(&bo->madv_lock); | |
535 | ||
536 | mutex_lock(&vc4->bo_lock); | |
537 | /* If the object references someone else's memory, we can't cache it. | |
538 | */ | |
539 | if (gem_bo->import_attach) { | |
540 | vc4_bo_destroy(bo); | |
541 | goto out; | |
542 | } | |
543 | ||
544 | /* Don't cache if it was publicly named. */ | |
545 | if (gem_bo->name) { | |
546 | vc4_bo_destroy(bo); | |
547 | goto out; | |
548 | } | |
549 | ||
550 | /* If this object was partially constructed but CMA allocation | |
551 | * had failed, just free it. Can also happen when the BO has been | |
552 | * purged. | |
553 | */ | |
554 | if (!bo->base.vaddr) { | |
555 | vc4_bo_destroy(bo); | |
556 | goto out; | |
557 | } | |
558 | ||
559 | cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size); | |
560 | if (!cache_list) { | |
561 | vc4_bo_destroy(bo); | |
562 | goto out; | |
563 | } | |
564 | ||
565 | if (bo->validated_shader) { | |
566 | kfree(bo->validated_shader->uniform_addr_offsets); | |
567 | kfree(bo->validated_shader->texture_samples); | |
568 | kfree(bo->validated_shader); | |
569 | bo->validated_shader = NULL; | |
570 | } | |
571 | ||
572 | /* Reset madv and usecnt before adding the BO to the cache. */ | |
573 | bo->madv = __VC4_MADV_NOTSUPP; | |
574 | refcount_set(&bo->usecnt, 0); | |
575 | ||
576 | bo->t_format = false; | |
577 | bo->free_time = jiffies; | |
578 | list_add(&bo->size_head, cache_list); | |
579 | list_add(&bo->unref_head, &vc4->bo_cache.time_list); | |
580 | ||
581 | vc4_bo_set_label(&bo->base.base, VC4_BO_TYPE_KERNEL_CACHE); | |
582 | ||
583 | vc4_bo_cache_free_old(dev); | |
584 | ||
585 | out: | |
586 | mutex_unlock(&vc4->bo_lock); | |
587 | } | |
588 | ||
589 | static void vc4_bo_cache_time_work(struct work_struct *work) | |
590 | { | |
591 | struct vc4_dev *vc4 = | |
592 | container_of(work, struct vc4_dev, bo_cache.time_work); | |
593 | struct drm_device *dev = vc4->dev; | |
594 | ||
595 | mutex_lock(&vc4->bo_lock); | |
596 | vc4_bo_cache_free_old(dev); | |
597 | mutex_unlock(&vc4->bo_lock); | |
598 | } | |
599 | ||
600 | int vc4_bo_inc_usecnt(struct vc4_bo *bo) | |
601 | { | |
602 | int ret; | |
603 | ||
604 | /* Fast path: if the BO is already retained by someone, no need to | |
605 | * check the madv status. | |
606 | */ | |
607 | if (refcount_inc_not_zero(&bo->usecnt)) | |
608 | return 0; | |
609 | ||
610 | mutex_lock(&bo->madv_lock); | |
611 | switch (bo->madv) { | |
612 | case VC4_MADV_WILLNEED: | |
613 | if (!refcount_inc_not_zero(&bo->usecnt)) | |
614 | refcount_set(&bo->usecnt, 1); | |
615 | ret = 0; | |
616 | break; | |
617 | case VC4_MADV_DONTNEED: | |
618 | /* We shouldn't use a BO marked as purgeable if at least | |
619 | * someone else retained its content by incrementing usecnt. | |
620 | * Luckily the BO hasn't been purged yet, but something wrong | |
621 | * is happening here. Just throw an error instead of | |
622 | * authorizing this use case. | |
623 | */ | |
624 | case __VC4_MADV_PURGED: | |
625 | /* We can't use a purged BO. */ | |
626 | default: | |
627 | /* Invalid madv value. */ | |
628 | ret = -EINVAL; | |
629 | break; | |
630 | } | |
631 | mutex_unlock(&bo->madv_lock); | |
632 | ||
633 | return ret; | |
634 | } | |
635 | ||
636 | void vc4_bo_dec_usecnt(struct vc4_bo *bo) | |
637 | { | |
638 | /* Fast path: if the BO is still retained by someone, no need to test | |
639 | * the madv value. | |
640 | */ | |
641 | if (refcount_dec_not_one(&bo->usecnt)) | |
642 | return; | |
643 | ||
644 | mutex_lock(&bo->madv_lock); | |
645 | if (refcount_dec_and_test(&bo->usecnt) && | |
646 | bo->madv == VC4_MADV_DONTNEED) | |
647 | vc4_bo_add_to_purgeable_pool(bo); | |
648 | mutex_unlock(&bo->madv_lock); | |
649 | } | |
650 | ||
651 | static void vc4_bo_cache_time_timer(struct timer_list *t) | |
652 | { | |
653 | struct vc4_dev *vc4 = from_timer(vc4, t, bo_cache.time_timer); | |
654 | ||
655 | schedule_work(&vc4->bo_cache.time_work); | |
656 | } | |
657 | ||
658 | struct dma_buf * | |
659 | vc4_prime_export(struct drm_device *dev, struct drm_gem_object *obj, int flags) | |
660 | { | |
661 | struct vc4_bo *bo = to_vc4_bo(obj); | |
662 | struct dma_buf *dmabuf; | |
663 | int ret; | |
664 | ||
665 | if (bo->validated_shader) { | |
666 | DRM_DEBUG("Attempting to export shader BO\n"); | |
667 | return ERR_PTR(-EINVAL); | |
668 | } | |
669 | ||
670 | /* Note: as soon as the BO is exported it becomes unpurgeable, because | |
671 | * noone ever decrements the usecnt even if the reference held by the | |
672 | * exported BO is released. This shouldn't be a problem since we don't | |
673 | * expect exported BOs to be marked as purgeable. | |
674 | */ | |
675 | ret = vc4_bo_inc_usecnt(bo); | |
676 | if (ret) { | |
677 | DRM_ERROR("Failed to increment BO usecnt\n"); | |
678 | return ERR_PTR(ret); | |
679 | } | |
680 | ||
681 | dmabuf = drm_gem_prime_export(dev, obj, flags); | |
682 | if (IS_ERR(dmabuf)) | |
683 | vc4_bo_dec_usecnt(bo); | |
684 | ||
685 | return dmabuf; | |
686 | } | |
687 | ||
688 | vm_fault_t vc4_fault(struct vm_fault *vmf) | |
689 | { | |
690 | struct vm_area_struct *vma = vmf->vma; | |
691 | struct drm_gem_object *obj = vma->vm_private_data; | |
692 | struct vc4_bo *bo = to_vc4_bo(obj); | |
693 | ||
694 | /* The only reason we would end up here is when user-space accesses | |
695 | * BO's memory after it's been purged. | |
696 | */ | |
697 | mutex_lock(&bo->madv_lock); | |
698 | WARN_ON(bo->madv != __VC4_MADV_PURGED); | |
699 | mutex_unlock(&bo->madv_lock); | |
700 | ||
701 | return VM_FAULT_SIGBUS; | |
702 | } | |
703 | ||
704 | int vc4_mmap(struct file *filp, struct vm_area_struct *vma) | |
705 | { | |
706 | struct drm_gem_object *gem_obj; | |
707 | unsigned long vm_pgoff; | |
708 | struct vc4_bo *bo; | |
709 | int ret; | |
710 | ||
711 | ret = drm_gem_mmap(filp, vma); | |
712 | if (ret) | |
713 | return ret; | |
714 | ||
715 | gem_obj = vma->vm_private_data; | |
716 | bo = to_vc4_bo(gem_obj); | |
717 | ||
718 | if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) { | |
719 | DRM_DEBUG("mmaping of shader BOs for writing not allowed.\n"); | |
720 | return -EINVAL; | |
721 | } | |
722 | ||
723 | if (bo->madv != VC4_MADV_WILLNEED) { | |
724 | DRM_DEBUG("mmaping of %s BO not allowed\n", | |
725 | bo->madv == VC4_MADV_DONTNEED ? | |
726 | "purgeable" : "purged"); | |
727 | return -EINVAL; | |
728 | } | |
729 | ||
730 | /* | |
731 | * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the | |
732 | * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map | |
733 | * the whole buffer. | |
734 | */ | |
735 | vma->vm_flags &= ~VM_PFNMAP; | |
736 | ||
737 | /* This ->vm_pgoff dance is needed to make all parties happy: | |
738 | * - dma_mmap_wc() uses ->vm_pgoff as an offset within the allocated | |
739 | * mem-region, hence the need to set it to zero (the value set by | |
740 | * the DRM core is a virtual offset encoding the GEM object-id) | |
741 | * - the mmap() core logic needs ->vm_pgoff to be restored to its | |
742 | * initial value before returning from this function because it | |
743 | * encodes the offset of this GEM in the dev->anon_inode pseudo-file | |
744 | * and this information will be used when we invalidate userspace | |
745 | * mappings with drm_vma_node_unmap() (called from vc4_gem_purge()). | |
746 | */ | |
747 | vm_pgoff = vma->vm_pgoff; | |
748 | vma->vm_pgoff = 0; | |
749 | ret = dma_mmap_wc(bo->base.base.dev->dev, vma, bo->base.vaddr, | |
750 | bo->base.paddr, vma->vm_end - vma->vm_start); | |
751 | vma->vm_pgoff = vm_pgoff; | |
752 | ||
753 | if (ret) | |
754 | drm_gem_vm_close(vma); | |
755 | ||
756 | return ret; | |
757 | } | |
758 | ||
759 | int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) | |
760 | { | |
761 | struct vc4_bo *bo = to_vc4_bo(obj); | |
762 | ||
763 | if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) { | |
764 | DRM_DEBUG("mmaping of shader BOs for writing not allowed.\n"); | |
765 | return -EINVAL; | |
766 | } | |
767 | ||
768 | return drm_gem_cma_prime_mmap(obj, vma); | |
769 | } | |
770 | ||
771 | void *vc4_prime_vmap(struct drm_gem_object *obj) | |
772 | { | |
773 | struct vc4_bo *bo = to_vc4_bo(obj); | |
774 | ||
775 | if (bo->validated_shader) { | |
776 | DRM_DEBUG("mmaping of shader BOs not allowed.\n"); | |
777 | return ERR_PTR(-EINVAL); | |
778 | } | |
779 | ||
780 | return drm_gem_cma_prime_vmap(obj); | |
781 | } | |
782 | ||
783 | struct drm_gem_object * | |
784 | vc4_prime_import_sg_table(struct drm_device *dev, | |
785 | struct dma_buf_attachment *attach, | |
786 | struct sg_table *sgt) | |
787 | { | |
788 | struct drm_gem_object *obj; | |
789 | ||
790 | obj = drm_gem_cma_prime_import_sg_table(dev, attach, sgt); | |
791 | if (IS_ERR(obj)) | |
792 | return obj; | |
793 | ||
794 | obj->resv = attach->dmabuf->resv; | |
795 | ||
796 | return obj; | |
797 | } | |
798 | ||
799 | int vc4_create_bo_ioctl(struct drm_device *dev, void *data, | |
800 | struct drm_file *file_priv) | |
801 | { | |
802 | struct drm_vc4_create_bo *args = data; | |
803 | struct vc4_bo *bo = NULL; | |
804 | int ret; | |
805 | ||
806 | /* | |
807 | * We can't allocate from the BO cache, because the BOs don't | |
808 | * get zeroed, and that might leak data between users. | |
809 | */ | |
810 | bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_V3D); | |
811 | if (IS_ERR(bo)) | |
812 | return PTR_ERR(bo); | |
813 | ||
814 | bo->madv = VC4_MADV_WILLNEED; | |
815 | ||
816 | ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle); | |
817 | drm_gem_object_put_unlocked(&bo->base.base); | |
818 | ||
819 | return ret; | |
820 | } | |
821 | ||
822 | int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data, | |
823 | struct drm_file *file_priv) | |
824 | { | |
825 | struct drm_vc4_mmap_bo *args = data; | |
826 | struct drm_gem_object *gem_obj; | |
827 | ||
828 | gem_obj = drm_gem_object_lookup(file_priv, args->handle); | |
829 | if (!gem_obj) { | |
830 | DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle); | |
831 | return -EINVAL; | |
832 | } | |
833 | ||
834 | /* The mmap offset was set up at BO allocation time. */ | |
835 | args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node); | |
836 | ||
837 | drm_gem_object_put_unlocked(gem_obj); | |
838 | return 0; | |
839 | } | |
840 | ||
841 | int | |
842 | vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data, | |
843 | struct drm_file *file_priv) | |
844 | { | |
845 | struct drm_vc4_create_shader_bo *args = data; | |
846 | struct vc4_bo *bo = NULL; | |
847 | int ret; | |
848 | ||
849 | if (args->size == 0) | |
850 | return -EINVAL; | |
851 | ||
852 | if (args->size % sizeof(u64) != 0) | |
853 | return -EINVAL; | |
854 | ||
855 | if (args->flags != 0) { | |
856 | DRM_INFO("Unknown flags set: 0x%08x\n", args->flags); | |
857 | return -EINVAL; | |
858 | } | |
859 | ||
860 | if (args->pad != 0) { | |
861 | DRM_INFO("Pad set: 0x%08x\n", args->pad); | |
862 | return -EINVAL; | |
863 | } | |
864 | ||
865 | bo = vc4_bo_create(dev, args->size, true, VC4_BO_TYPE_V3D_SHADER); | |
866 | if (IS_ERR(bo)) | |
867 | return PTR_ERR(bo); | |
868 | ||
869 | bo->madv = VC4_MADV_WILLNEED; | |
870 | ||
871 | if (copy_from_user(bo->base.vaddr, | |
872 | (void __user *)(uintptr_t)args->data, | |
873 | args->size)) { | |
874 | ret = -EFAULT; | |
875 | goto fail; | |
876 | } | |
877 | /* Clear the rest of the memory from allocating from the BO | |
878 | * cache. | |
879 | */ | |
880 | memset(bo->base.vaddr + args->size, 0, | |
881 | bo->base.base.size - args->size); | |
882 | ||
883 | bo->validated_shader = vc4_validate_shader(&bo->base); | |
884 | if (!bo->validated_shader) { | |
885 | ret = -EINVAL; | |
886 | goto fail; | |
887 | } | |
888 | ||
889 | /* We have to create the handle after validation, to avoid | |
890 | * races for users to do doing things like mmap the shader BO. | |
891 | */ | |
892 | ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle); | |
893 | ||
894 | fail: | |
895 | drm_gem_object_put_unlocked(&bo->base.base); | |
896 | ||
897 | return ret; | |
898 | } | |
899 | ||
900 | /** | |
901 | * vc4_set_tiling_ioctl() - Sets the tiling modifier for a BO. | |
902 | * @dev: DRM device | |
903 | * @data: ioctl argument | |
904 | * @file_priv: DRM file for this fd | |
905 | * | |
906 | * The tiling state of the BO decides the default modifier of an fb if | |
907 | * no specific modifier was set by userspace, and the return value of | |
908 | * vc4_get_tiling_ioctl() (so that userspace can treat a BO it | |
909 | * received from dmabuf as the same tiling format as the producer | |
910 | * used). | |
911 | */ | |
912 | int vc4_set_tiling_ioctl(struct drm_device *dev, void *data, | |
913 | struct drm_file *file_priv) | |
914 | { | |
915 | struct drm_vc4_set_tiling *args = data; | |
916 | struct drm_gem_object *gem_obj; | |
917 | struct vc4_bo *bo; | |
918 | bool t_format; | |
919 | ||
920 | if (args->flags != 0) | |
921 | return -EINVAL; | |
922 | ||
923 | switch (args->modifier) { | |
924 | case DRM_FORMAT_MOD_NONE: | |
925 | t_format = false; | |
926 | break; | |
927 | case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: | |
928 | t_format = true; | |
929 | break; | |
930 | default: | |
931 | return -EINVAL; | |
932 | } | |
933 | ||
934 | gem_obj = drm_gem_object_lookup(file_priv, args->handle); | |
935 | if (!gem_obj) { | |
936 | DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle); | |
937 | return -ENOENT; | |
938 | } | |
939 | bo = to_vc4_bo(gem_obj); | |
940 | bo->t_format = t_format; | |
941 | ||
942 | drm_gem_object_put_unlocked(gem_obj); | |
943 | ||
944 | return 0; | |
945 | } | |
946 | ||
947 | /** | |
948 | * vc4_get_tiling_ioctl() - Gets the tiling modifier for a BO. | |
949 | * @dev: DRM device | |
950 | * @data: ioctl argument | |
951 | * @file_priv: DRM file for this fd | |
952 | * | |
953 | * Returns the tiling modifier for a BO as set by vc4_set_tiling_ioctl(). | |
954 | */ | |
955 | int vc4_get_tiling_ioctl(struct drm_device *dev, void *data, | |
956 | struct drm_file *file_priv) | |
957 | { | |
958 | struct drm_vc4_get_tiling *args = data; | |
959 | struct drm_gem_object *gem_obj; | |
960 | struct vc4_bo *bo; | |
961 | ||
962 | if (args->flags != 0 || args->modifier != 0) | |
963 | return -EINVAL; | |
964 | ||
965 | gem_obj = drm_gem_object_lookup(file_priv, args->handle); | |
966 | if (!gem_obj) { | |
967 | DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle); | |
968 | return -ENOENT; | |
969 | } | |
970 | bo = to_vc4_bo(gem_obj); | |
971 | ||
972 | if (bo->t_format) | |
973 | args->modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED; | |
974 | else | |
975 | args->modifier = DRM_FORMAT_MOD_NONE; | |
976 | ||
977 | drm_gem_object_put_unlocked(gem_obj); | |
978 | ||
979 | return 0; | |
980 | } | |
981 | ||
982 | int vc4_bo_cache_init(struct drm_device *dev) | |
983 | { | |
984 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
985 | int i; | |
986 | ||
987 | /* Create the initial set of BO labels that the kernel will | |
988 | * use. This lets us avoid a bunch of string reallocation in | |
989 | * the kernel's draw and BO allocation paths. | |
990 | */ | |
991 | vc4->bo_labels = kcalloc(VC4_BO_TYPE_COUNT, sizeof(*vc4->bo_labels), | |
992 | GFP_KERNEL); | |
993 | if (!vc4->bo_labels) | |
994 | return -ENOMEM; | |
995 | vc4->num_labels = VC4_BO_TYPE_COUNT; | |
996 | ||
997 | BUILD_BUG_ON(ARRAY_SIZE(bo_type_names) != VC4_BO_TYPE_COUNT); | |
998 | for (i = 0; i < VC4_BO_TYPE_COUNT; i++) | |
999 | vc4->bo_labels[i].name = bo_type_names[i]; | |
1000 | ||
1001 | mutex_init(&vc4->bo_lock); | |
1002 | ||
1003 | vc4_debugfs_add_file(dev, "bo_stats", vc4_bo_stats_debugfs, NULL); | |
1004 | ||
1005 | INIT_LIST_HEAD(&vc4->bo_cache.time_list); | |
1006 | ||
1007 | INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work); | |
1008 | timer_setup(&vc4->bo_cache.time_timer, vc4_bo_cache_time_timer, 0); | |
1009 | ||
1010 | return 0; | |
1011 | } | |
1012 | ||
1013 | void vc4_bo_cache_destroy(struct drm_device *dev) | |
1014 | { | |
1015 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
1016 | int i; | |
1017 | ||
1018 | del_timer(&vc4->bo_cache.time_timer); | |
1019 | cancel_work_sync(&vc4->bo_cache.time_work); | |
1020 | ||
1021 | vc4_bo_cache_purge(dev); | |
1022 | ||
1023 | for (i = 0; i < vc4->num_labels; i++) { | |
1024 | if (vc4->bo_labels[i].num_allocated) { | |
1025 | DRM_ERROR("Destroying BO cache with %d %s " | |
1026 | "BOs still allocated\n", | |
1027 | vc4->bo_labels[i].num_allocated, | |
1028 | vc4->bo_labels[i].name); | |
1029 | } | |
1030 | ||
1031 | if (is_user_label(i)) | |
1032 | kfree(vc4->bo_labels[i].name); | |
1033 | } | |
1034 | kfree(vc4->bo_labels); | |
1035 | } | |
1036 | ||
1037 | int vc4_label_bo_ioctl(struct drm_device *dev, void *data, | |
1038 | struct drm_file *file_priv) | |
1039 | { | |
1040 | struct vc4_dev *vc4 = to_vc4_dev(dev); | |
1041 | struct drm_vc4_label_bo *args = data; | |
1042 | char *name; | |
1043 | struct drm_gem_object *gem_obj; | |
1044 | int ret = 0, label; | |
1045 | ||
1046 | if (!args->len) | |
1047 | return -EINVAL; | |
1048 | ||
1049 | name = strndup_user(u64_to_user_ptr(args->name), args->len + 1); | |
1050 | if (IS_ERR(name)) | |
1051 | return PTR_ERR(name); | |
1052 | ||
1053 | gem_obj = drm_gem_object_lookup(file_priv, args->handle); | |
1054 | if (!gem_obj) { | |
1055 | DRM_ERROR("Failed to look up GEM BO %d\n", args->handle); | |
1056 | kfree(name); | |
1057 | return -ENOENT; | |
1058 | } | |
1059 | ||
1060 | mutex_lock(&vc4->bo_lock); | |
1061 | label = vc4_get_user_label(vc4, name); | |
1062 | if (label != -1) | |
1063 | vc4_bo_set_label(gem_obj, label); | |
1064 | else | |
1065 | ret = -ENOMEM; | |
1066 | mutex_unlock(&vc4->bo_lock); | |
1067 | ||
1068 | drm_gem_object_put_unlocked(gem_obj); | |
1069 | ||
1070 | return ret; | |
1071 | } |