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 DMA helper functions to allocate contiguous ranges of |
12 | * physical memory for our BOs. |
13 | * |
14 | * Since the DMA 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 <drm/drm_fourcc.h> |
22 | |
23 | #include "vc4_drv.h" |
24 | #include "uapi/drm/vc4_drm.h" |
25 | |
26 | static const struct drm_gem_object_funcs vc4_gem_object_funcs; |
27 | |
28 | static const char * const bo_type_names[] = { |
29 | "kernel" , |
30 | "V3D" , |
31 | "V3D shader" , |
32 | "dumb" , |
33 | "binner" , |
34 | "RCL" , |
35 | "BCL" , |
36 | "kernel BO cache" , |
37 | }; |
38 | |
39 | static bool is_user_label(int label) |
40 | { |
41 | return label >= VC4_BO_TYPE_COUNT; |
42 | } |
43 | |
44 | static void vc4_bo_stats_print(struct drm_printer *p, struct vc4_dev *vc4) |
45 | { |
46 | int i; |
47 | |
48 | for (i = 0; i < vc4->num_labels; i++) { |
49 | if (!vc4->bo_labels[i].num_allocated) |
50 | continue; |
51 | |
52 | drm_printf(p, f: "%30s: %6dkb BOs (%d)\n" , |
53 | vc4->bo_labels[i].name, |
54 | vc4->bo_labels[i].size_allocated / 1024, |
55 | vc4->bo_labels[i].num_allocated); |
56 | } |
57 | |
58 | mutex_lock(&vc4->purgeable.lock); |
59 | if (vc4->purgeable.num) |
60 | drm_printf(p, f: "%30s: %6zdkb BOs (%d)\n" , "userspace BO cache" , |
61 | vc4->purgeable.size / 1024, vc4->purgeable.num); |
62 | |
63 | if (vc4->purgeable.purged_num) |
64 | drm_printf(p, f: "%30s: %6zdkb BOs (%d)\n" , "total purged BO" , |
65 | vc4->purgeable.purged_size / 1024, |
66 | vc4->purgeable.purged_num); |
67 | mutex_unlock(lock: &vc4->purgeable.lock); |
68 | } |
69 | |
70 | static int vc4_bo_stats_debugfs(struct seq_file *m, void *unused) |
71 | { |
72 | struct drm_debugfs_entry *entry = m->private; |
73 | struct drm_device *dev = entry->dev; |
74 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
75 | struct drm_printer p = drm_seq_file_printer(f: m); |
76 | |
77 | vc4_bo_stats_print(p: &p, vc4); |
78 | |
79 | return 0; |
80 | } |
81 | |
82 | /* Takes ownership of *name and returns the appropriate slot for it in |
83 | * the bo_labels[] array, extending it as necessary. |
84 | * |
85 | * This is inefficient and could use a hash table instead of walking |
86 | * an array and strcmp()ing. However, the assumption is that user |
87 | * labeling will be infrequent (scanout buffers and other long-lived |
88 | * objects, or debug driver builds), so we can live with it for now. |
89 | */ |
90 | static int vc4_get_user_label(struct vc4_dev *vc4, const char *name) |
91 | { |
92 | int i; |
93 | int free_slot = -1; |
94 | |
95 | for (i = 0; i < vc4->num_labels; i++) { |
96 | if (!vc4->bo_labels[i].name) { |
97 | free_slot = i; |
98 | } else if (strcmp(vc4->bo_labels[i].name, name) == 0) { |
99 | kfree(objp: name); |
100 | return i; |
101 | } |
102 | } |
103 | |
104 | if (free_slot != -1) { |
105 | WARN_ON(vc4->bo_labels[free_slot].num_allocated != 0); |
106 | vc4->bo_labels[free_slot].name = name; |
107 | return free_slot; |
108 | } else { |
109 | u32 new_label_count = vc4->num_labels + 1; |
110 | struct vc4_label *new_labels = |
111 | krealloc(objp: vc4->bo_labels, |
112 | new_size: new_label_count * sizeof(*new_labels), |
113 | GFP_KERNEL); |
114 | |
115 | if (!new_labels) { |
116 | kfree(objp: name); |
117 | return -1; |
118 | } |
119 | |
120 | free_slot = vc4->num_labels; |
121 | vc4->bo_labels = new_labels; |
122 | vc4->num_labels = new_label_count; |
123 | |
124 | vc4->bo_labels[free_slot].name = name; |
125 | vc4->bo_labels[free_slot].num_allocated = 0; |
126 | vc4->bo_labels[free_slot].size_allocated = 0; |
127 | |
128 | return free_slot; |
129 | } |
130 | } |
131 | |
132 | static void vc4_bo_set_label(struct drm_gem_object *gem_obj, int label) |
133 | { |
134 | struct vc4_bo *bo = to_vc4_bo(gem_obj); |
135 | struct vc4_dev *vc4 = to_vc4_dev(gem_obj->dev); |
136 | |
137 | lockdep_assert_held(&vc4->bo_lock); |
138 | |
139 | if (label != -1) { |
140 | vc4->bo_labels[label].num_allocated++; |
141 | vc4->bo_labels[label].size_allocated += gem_obj->size; |
142 | } |
143 | |
144 | vc4->bo_labels[bo->label].num_allocated--; |
145 | vc4->bo_labels[bo->label].size_allocated -= gem_obj->size; |
146 | |
147 | if (vc4->bo_labels[bo->label].num_allocated == 0 && |
148 | is_user_label(label: bo->label)) { |
149 | /* Free user BO label slots on last unreference. |
150 | * Slots are just where we track the stats for a given |
151 | * name, and once a name is unused we can reuse that |
152 | * slot. |
153 | */ |
154 | kfree(objp: vc4->bo_labels[bo->label].name); |
155 | vc4->bo_labels[bo->label].name = NULL; |
156 | } |
157 | |
158 | bo->label = label; |
159 | } |
160 | |
161 | static uint32_t bo_page_index(size_t size) |
162 | { |
163 | return (size / PAGE_SIZE) - 1; |
164 | } |
165 | |
166 | static void vc4_bo_destroy(struct vc4_bo *bo) |
167 | { |
168 | struct drm_gem_object *obj = &bo->base.base; |
169 | struct vc4_dev *vc4 = to_vc4_dev(obj->dev); |
170 | |
171 | lockdep_assert_held(&vc4->bo_lock); |
172 | |
173 | vc4_bo_set_label(gem_obj: obj, label: -1); |
174 | |
175 | if (bo->validated_shader) { |
176 | kfree(objp: bo->validated_shader->uniform_addr_offsets); |
177 | kfree(objp: bo->validated_shader->texture_samples); |
178 | kfree(objp: bo->validated_shader); |
179 | bo->validated_shader = NULL; |
180 | } |
181 | |
182 | mutex_destroy(lock: &bo->madv_lock); |
183 | drm_gem_dma_free(dma_obj: &bo->base); |
184 | } |
185 | |
186 | static void vc4_bo_remove_from_cache(struct vc4_bo *bo) |
187 | { |
188 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
189 | |
190 | lockdep_assert_held(&vc4->bo_lock); |
191 | list_del(entry: &bo->unref_head); |
192 | list_del(entry: &bo->size_head); |
193 | } |
194 | |
195 | static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev, |
196 | size_t size) |
197 | { |
198 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
199 | uint32_t page_index = bo_page_index(size); |
200 | |
201 | if (vc4->bo_cache.size_list_size <= page_index) { |
202 | uint32_t new_size = max(vc4->bo_cache.size_list_size * 2, |
203 | page_index + 1); |
204 | struct list_head *new_list; |
205 | uint32_t i; |
206 | |
207 | new_list = kmalloc_array(n: new_size, size: sizeof(struct list_head), |
208 | GFP_KERNEL); |
209 | if (!new_list) |
210 | return NULL; |
211 | |
212 | /* Rebase the old cached BO lists to their new list |
213 | * head locations. |
214 | */ |
215 | for (i = 0; i < vc4->bo_cache.size_list_size; i++) { |
216 | struct list_head *old_list = |
217 | &vc4->bo_cache.size_list[i]; |
218 | |
219 | if (list_empty(head: old_list)) |
220 | INIT_LIST_HEAD(list: &new_list[i]); |
221 | else |
222 | list_replace(old: old_list, new: &new_list[i]); |
223 | } |
224 | /* And initialize the brand new BO list heads. */ |
225 | for (i = vc4->bo_cache.size_list_size; i < new_size; i++) |
226 | INIT_LIST_HEAD(list: &new_list[i]); |
227 | |
228 | kfree(objp: vc4->bo_cache.size_list); |
229 | vc4->bo_cache.size_list = new_list; |
230 | vc4->bo_cache.size_list_size = new_size; |
231 | } |
232 | |
233 | return &vc4->bo_cache.size_list[page_index]; |
234 | } |
235 | |
236 | static void vc4_bo_cache_purge(struct drm_device *dev) |
237 | { |
238 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
239 | |
240 | mutex_lock(&vc4->bo_lock); |
241 | while (!list_empty(head: &vc4->bo_cache.time_list)) { |
242 | struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list, |
243 | struct vc4_bo, unref_head); |
244 | vc4_bo_remove_from_cache(bo); |
245 | vc4_bo_destroy(bo); |
246 | } |
247 | mutex_unlock(lock: &vc4->bo_lock); |
248 | } |
249 | |
250 | void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo) |
251 | { |
252 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
253 | |
254 | if (WARN_ON_ONCE(vc4->is_vc5)) |
255 | return; |
256 | |
257 | mutex_lock(&vc4->purgeable.lock); |
258 | list_add_tail(new: &bo->size_head, head: &vc4->purgeable.list); |
259 | vc4->purgeable.num++; |
260 | vc4->purgeable.size += bo->base.base.size; |
261 | mutex_unlock(lock: &vc4->purgeable.lock); |
262 | } |
263 | |
264 | static void vc4_bo_remove_from_purgeable_pool_locked(struct vc4_bo *bo) |
265 | { |
266 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
267 | |
268 | if (WARN_ON_ONCE(vc4->is_vc5)) |
269 | return; |
270 | |
271 | /* list_del_init() is used here because the caller might release |
272 | * the purgeable lock in order to acquire the madv one and update the |
273 | * madv status. |
274 | * During this short period of time a user might decide to mark |
275 | * the BO as unpurgeable, and if bo->madv is set to |
276 | * VC4_MADV_DONTNEED it will try to remove the BO from the |
277 | * purgeable list which will fail if the ->next/prev fields |
278 | * are set to LIST_POISON1/LIST_POISON2 (which is what |
279 | * list_del() does). |
280 | * Re-initializing the list element guarantees that list_del() |
281 | * will work correctly even if it's a NOP. |
282 | */ |
283 | list_del_init(entry: &bo->size_head); |
284 | vc4->purgeable.num--; |
285 | vc4->purgeable.size -= bo->base.base.size; |
286 | } |
287 | |
288 | void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo) |
289 | { |
290 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
291 | |
292 | mutex_lock(&vc4->purgeable.lock); |
293 | vc4_bo_remove_from_purgeable_pool_locked(bo); |
294 | mutex_unlock(lock: &vc4->purgeable.lock); |
295 | } |
296 | |
297 | static void vc4_bo_purge(struct drm_gem_object *obj) |
298 | { |
299 | struct vc4_bo *bo = to_vc4_bo(obj); |
300 | struct drm_device *dev = obj->dev; |
301 | |
302 | WARN_ON(!mutex_is_locked(&bo->madv_lock)); |
303 | WARN_ON(bo->madv != VC4_MADV_DONTNEED); |
304 | |
305 | drm_vma_node_unmap(node: &obj->vma_node, file_mapping: dev->anon_inode->i_mapping); |
306 | |
307 | dma_free_wc(dev: dev->dev, size: obj->size, cpu_addr: bo->base.vaddr, dma_addr: bo->base.dma_addr); |
308 | bo->base.vaddr = NULL; |
309 | bo->madv = __VC4_MADV_PURGED; |
310 | } |
311 | |
312 | static void vc4_bo_userspace_cache_purge(struct drm_device *dev) |
313 | { |
314 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
315 | |
316 | mutex_lock(&vc4->purgeable.lock); |
317 | while (!list_empty(head: &vc4->purgeable.list)) { |
318 | struct vc4_bo *bo = list_first_entry(&vc4->purgeable.list, |
319 | struct vc4_bo, size_head); |
320 | struct drm_gem_object *obj = &bo->base.base; |
321 | size_t purged_size = 0; |
322 | |
323 | vc4_bo_remove_from_purgeable_pool_locked(bo); |
324 | |
325 | /* Release the purgeable lock while we're purging the BO so |
326 | * that other people can continue inserting things in the |
327 | * purgeable pool without having to wait for all BOs to be |
328 | * purged. |
329 | */ |
330 | mutex_unlock(lock: &vc4->purgeable.lock); |
331 | mutex_lock(&bo->madv_lock); |
332 | |
333 | /* Since we released the purgeable pool lock before acquiring |
334 | * the BO madv one, the user may have marked the BO as WILLNEED |
335 | * and re-used it in the meantime. |
336 | * Before purging the BO we need to make sure |
337 | * - it is still marked as DONTNEED |
338 | * - it has not been re-inserted in the purgeable list |
339 | * - it is not used by HW blocks |
340 | * If one of these conditions is not met, just skip the entry. |
341 | */ |
342 | if (bo->madv == VC4_MADV_DONTNEED && |
343 | list_empty(head: &bo->size_head) && |
344 | !refcount_read(r: &bo->usecnt)) { |
345 | purged_size = bo->base.base.size; |
346 | vc4_bo_purge(obj); |
347 | } |
348 | mutex_unlock(lock: &bo->madv_lock); |
349 | mutex_lock(&vc4->purgeable.lock); |
350 | |
351 | if (purged_size) { |
352 | vc4->purgeable.purged_size += purged_size; |
353 | vc4->purgeable.purged_num++; |
354 | } |
355 | } |
356 | mutex_unlock(lock: &vc4->purgeable.lock); |
357 | } |
358 | |
359 | static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev, |
360 | uint32_t size, |
361 | enum vc4_kernel_bo_type type) |
362 | { |
363 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
364 | uint32_t page_index = bo_page_index(size); |
365 | struct vc4_bo *bo = NULL; |
366 | |
367 | mutex_lock(&vc4->bo_lock); |
368 | if (page_index >= vc4->bo_cache.size_list_size) |
369 | goto out; |
370 | |
371 | if (list_empty(head: &vc4->bo_cache.size_list[page_index])) |
372 | goto out; |
373 | |
374 | bo = list_first_entry(&vc4->bo_cache.size_list[page_index], |
375 | struct vc4_bo, size_head); |
376 | vc4_bo_remove_from_cache(bo); |
377 | kref_init(kref: &bo->base.base.refcount); |
378 | |
379 | out: |
380 | if (bo) |
381 | vc4_bo_set_label(gem_obj: &bo->base.base, label: type); |
382 | mutex_unlock(lock: &vc4->bo_lock); |
383 | return bo; |
384 | } |
385 | |
386 | /** |
387 | * vc4_create_object - Implementation of driver->gem_create_object. |
388 | * @dev: DRM device |
389 | * @size: Size in bytes of the memory the object will reference |
390 | * |
391 | * This lets the DMA helpers allocate object structs for us, and keep |
392 | * our BO stats correct. |
393 | */ |
394 | struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size) |
395 | { |
396 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
397 | struct vc4_bo *bo; |
398 | |
399 | if (WARN_ON_ONCE(vc4->is_vc5)) |
400 | return ERR_PTR(error: -ENODEV); |
401 | |
402 | bo = kzalloc(size: sizeof(*bo), GFP_KERNEL); |
403 | if (!bo) |
404 | return ERR_PTR(error: -ENOMEM); |
405 | |
406 | bo->madv = VC4_MADV_WILLNEED; |
407 | refcount_set(r: &bo->usecnt, n: 0); |
408 | |
409 | mutex_init(&bo->madv_lock); |
410 | |
411 | mutex_lock(&vc4->bo_lock); |
412 | bo->label = VC4_BO_TYPE_KERNEL; |
413 | vc4->bo_labels[VC4_BO_TYPE_KERNEL].num_allocated++; |
414 | vc4->bo_labels[VC4_BO_TYPE_KERNEL].size_allocated += size; |
415 | mutex_unlock(lock: &vc4->bo_lock); |
416 | |
417 | bo->base.base.funcs = &vc4_gem_object_funcs; |
418 | |
419 | return &bo->base.base; |
420 | } |
421 | |
422 | struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size, |
423 | bool allow_unzeroed, enum vc4_kernel_bo_type type) |
424 | { |
425 | size_t size = roundup(unaligned_size, PAGE_SIZE); |
426 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
427 | struct drm_gem_dma_object *dma_obj; |
428 | struct vc4_bo *bo; |
429 | |
430 | if (WARN_ON_ONCE(vc4->is_vc5)) |
431 | return ERR_PTR(error: -ENODEV); |
432 | |
433 | if (size == 0) |
434 | return ERR_PTR(error: -EINVAL); |
435 | |
436 | /* First, try to get a vc4_bo from the kernel BO cache. */ |
437 | bo = vc4_bo_get_from_cache(dev, size, type); |
438 | if (bo) { |
439 | if (!allow_unzeroed) |
440 | memset(bo->base.vaddr, 0, bo->base.base.size); |
441 | return bo; |
442 | } |
443 | |
444 | dma_obj = drm_gem_dma_create(drm: dev, size); |
445 | if (IS_ERR(ptr: dma_obj)) { |
446 | /* |
447 | * If we've run out of DMA memory, kill the cache of |
448 | * DMA allocations we've got laying around and try again. |
449 | */ |
450 | vc4_bo_cache_purge(dev); |
451 | dma_obj = drm_gem_dma_create(drm: dev, size); |
452 | } |
453 | |
454 | if (IS_ERR(ptr: dma_obj)) { |
455 | /* |
456 | * Still not enough DMA memory, purge the userspace BO |
457 | * cache and retry. |
458 | * This is sub-optimal since we purge the whole userspace |
459 | * BO cache which forces user that want to re-use the BO to |
460 | * restore its initial content. |
461 | * Ideally, we should purge entries one by one and retry |
462 | * after each to see if DMA allocation succeeds. Or even |
463 | * better, try to find an entry with at least the same |
464 | * size. |
465 | */ |
466 | vc4_bo_userspace_cache_purge(dev); |
467 | dma_obj = drm_gem_dma_create(drm: dev, size); |
468 | } |
469 | |
470 | if (IS_ERR(ptr: dma_obj)) { |
471 | struct drm_printer p = drm_info_printer(dev: vc4->base.dev); |
472 | DRM_ERROR("Failed to allocate from GEM DMA helper:\n" ); |
473 | vc4_bo_stats_print(p: &p, vc4); |
474 | return ERR_PTR(error: -ENOMEM); |
475 | } |
476 | bo = to_vc4_bo(&dma_obj->base); |
477 | |
478 | /* By default, BOs do not support the MADV ioctl. This will be enabled |
479 | * only on BOs that are exposed to userspace (V3D, V3D_SHADER and DUMB |
480 | * BOs). |
481 | */ |
482 | bo->madv = __VC4_MADV_NOTSUPP; |
483 | |
484 | mutex_lock(&vc4->bo_lock); |
485 | vc4_bo_set_label(gem_obj: &dma_obj->base, label: type); |
486 | mutex_unlock(lock: &vc4->bo_lock); |
487 | |
488 | return bo; |
489 | } |
490 | |
491 | int vc4_bo_dumb_create(struct drm_file *file_priv, |
492 | struct drm_device *dev, |
493 | struct drm_mode_create_dumb *args) |
494 | { |
495 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
496 | struct vc4_bo *bo = NULL; |
497 | int ret; |
498 | |
499 | if (WARN_ON_ONCE(vc4->is_vc5)) |
500 | return -ENODEV; |
501 | |
502 | ret = vc4_dumb_fixup_args(args); |
503 | if (ret) |
504 | return ret; |
505 | |
506 | bo = vc4_bo_create(dev, unaligned_size: args->size, allow_unzeroed: false, type: VC4_BO_TYPE_DUMB); |
507 | if (IS_ERR(ptr: bo)) |
508 | return PTR_ERR(ptr: bo); |
509 | |
510 | bo->madv = VC4_MADV_WILLNEED; |
511 | |
512 | ret = drm_gem_handle_create(file_priv, obj: &bo->base.base, handlep: &args->handle); |
513 | drm_gem_object_put(obj: &bo->base.base); |
514 | |
515 | return ret; |
516 | } |
517 | |
518 | static void vc4_bo_cache_free_old(struct drm_device *dev) |
519 | { |
520 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
521 | unsigned long expire_time = jiffies - msecs_to_jiffies(m: 1000); |
522 | |
523 | lockdep_assert_held(&vc4->bo_lock); |
524 | |
525 | while (!list_empty(head: &vc4->bo_cache.time_list)) { |
526 | struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list, |
527 | struct vc4_bo, unref_head); |
528 | if (time_before(expire_time, bo->free_time)) { |
529 | mod_timer(timer: &vc4->bo_cache.time_timer, |
530 | expires: round_jiffies_up(j: jiffies + |
531 | msecs_to_jiffies(m: 1000))); |
532 | return; |
533 | } |
534 | |
535 | vc4_bo_remove_from_cache(bo); |
536 | vc4_bo_destroy(bo); |
537 | } |
538 | } |
539 | |
540 | /* Called on the last userspace/kernel unreference of the BO. Returns |
541 | * it to the BO cache if possible, otherwise frees it. |
542 | */ |
543 | static void vc4_free_object(struct drm_gem_object *gem_bo) |
544 | { |
545 | struct drm_device *dev = gem_bo->dev; |
546 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
547 | struct vc4_bo *bo = to_vc4_bo(gem_bo); |
548 | struct list_head *cache_list; |
549 | |
550 | /* Remove the BO from the purgeable list. */ |
551 | mutex_lock(&bo->madv_lock); |
552 | if (bo->madv == VC4_MADV_DONTNEED && !refcount_read(r: &bo->usecnt)) |
553 | vc4_bo_remove_from_purgeable_pool(bo); |
554 | mutex_unlock(lock: &bo->madv_lock); |
555 | |
556 | mutex_lock(&vc4->bo_lock); |
557 | /* If the object references someone else's memory, we can't cache it. |
558 | */ |
559 | if (gem_bo->import_attach) { |
560 | vc4_bo_destroy(bo); |
561 | goto out; |
562 | } |
563 | |
564 | /* Don't cache if it was publicly named. */ |
565 | if (gem_bo->name) { |
566 | vc4_bo_destroy(bo); |
567 | goto out; |
568 | } |
569 | |
570 | /* If this object was partially constructed but DMA allocation |
571 | * had failed, just free it. Can also happen when the BO has been |
572 | * purged. |
573 | */ |
574 | if (!bo->base.vaddr) { |
575 | vc4_bo_destroy(bo); |
576 | goto out; |
577 | } |
578 | |
579 | cache_list = vc4_get_cache_list_for_size(dev, size: gem_bo->size); |
580 | if (!cache_list) { |
581 | vc4_bo_destroy(bo); |
582 | goto out; |
583 | } |
584 | |
585 | if (bo->validated_shader) { |
586 | kfree(objp: bo->validated_shader->uniform_addr_offsets); |
587 | kfree(objp: bo->validated_shader->texture_samples); |
588 | kfree(objp: bo->validated_shader); |
589 | bo->validated_shader = NULL; |
590 | } |
591 | |
592 | /* Reset madv and usecnt before adding the BO to the cache. */ |
593 | bo->madv = __VC4_MADV_NOTSUPP; |
594 | refcount_set(r: &bo->usecnt, n: 0); |
595 | |
596 | bo->t_format = false; |
597 | bo->free_time = jiffies; |
598 | list_add(new: &bo->size_head, head: cache_list); |
599 | list_add(new: &bo->unref_head, head: &vc4->bo_cache.time_list); |
600 | |
601 | vc4_bo_set_label(gem_obj: &bo->base.base, label: VC4_BO_TYPE_KERNEL_CACHE); |
602 | |
603 | vc4_bo_cache_free_old(dev); |
604 | |
605 | out: |
606 | mutex_unlock(lock: &vc4->bo_lock); |
607 | } |
608 | |
609 | static void vc4_bo_cache_time_work(struct work_struct *work) |
610 | { |
611 | struct vc4_dev *vc4 = |
612 | container_of(work, struct vc4_dev, bo_cache.time_work); |
613 | struct drm_device *dev = &vc4->base; |
614 | |
615 | mutex_lock(&vc4->bo_lock); |
616 | vc4_bo_cache_free_old(dev); |
617 | mutex_unlock(lock: &vc4->bo_lock); |
618 | } |
619 | |
620 | int vc4_bo_inc_usecnt(struct vc4_bo *bo) |
621 | { |
622 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
623 | int ret; |
624 | |
625 | if (WARN_ON_ONCE(vc4->is_vc5)) |
626 | return -ENODEV; |
627 | |
628 | /* Fast path: if the BO is already retained by someone, no need to |
629 | * check the madv status. |
630 | */ |
631 | if (refcount_inc_not_zero(r: &bo->usecnt)) |
632 | return 0; |
633 | |
634 | mutex_lock(&bo->madv_lock); |
635 | switch (bo->madv) { |
636 | case VC4_MADV_WILLNEED: |
637 | if (!refcount_inc_not_zero(r: &bo->usecnt)) |
638 | refcount_set(r: &bo->usecnt, n: 1); |
639 | ret = 0; |
640 | break; |
641 | case VC4_MADV_DONTNEED: |
642 | /* We shouldn't use a BO marked as purgeable if at least |
643 | * someone else retained its content by incrementing usecnt. |
644 | * Luckily the BO hasn't been purged yet, but something wrong |
645 | * is happening here. Just throw an error instead of |
646 | * authorizing this use case. |
647 | */ |
648 | case __VC4_MADV_PURGED: |
649 | /* We can't use a purged BO. */ |
650 | default: |
651 | /* Invalid madv value. */ |
652 | ret = -EINVAL; |
653 | break; |
654 | } |
655 | mutex_unlock(lock: &bo->madv_lock); |
656 | |
657 | return ret; |
658 | } |
659 | |
660 | void vc4_bo_dec_usecnt(struct vc4_bo *bo) |
661 | { |
662 | struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev); |
663 | |
664 | if (WARN_ON_ONCE(vc4->is_vc5)) |
665 | return; |
666 | |
667 | /* Fast path: if the BO is still retained by someone, no need to test |
668 | * the madv value. |
669 | */ |
670 | if (refcount_dec_not_one(r: &bo->usecnt)) |
671 | return; |
672 | |
673 | mutex_lock(&bo->madv_lock); |
674 | if (refcount_dec_and_test(r: &bo->usecnt) && |
675 | bo->madv == VC4_MADV_DONTNEED) |
676 | vc4_bo_add_to_purgeable_pool(bo); |
677 | mutex_unlock(lock: &bo->madv_lock); |
678 | } |
679 | |
680 | static void vc4_bo_cache_time_timer(struct timer_list *t) |
681 | { |
682 | struct vc4_dev *vc4 = from_timer(vc4, t, bo_cache.time_timer); |
683 | |
684 | schedule_work(work: &vc4->bo_cache.time_work); |
685 | } |
686 | |
687 | static struct dma_buf *vc4_prime_export(struct drm_gem_object *obj, int flags) |
688 | { |
689 | struct vc4_bo *bo = to_vc4_bo(obj); |
690 | struct dma_buf *dmabuf; |
691 | int ret; |
692 | |
693 | if (bo->validated_shader) { |
694 | DRM_DEBUG("Attempting to export shader BO\n" ); |
695 | return ERR_PTR(error: -EINVAL); |
696 | } |
697 | |
698 | /* Note: as soon as the BO is exported it becomes unpurgeable, because |
699 | * noone ever decrements the usecnt even if the reference held by the |
700 | * exported BO is released. This shouldn't be a problem since we don't |
701 | * expect exported BOs to be marked as purgeable. |
702 | */ |
703 | ret = vc4_bo_inc_usecnt(bo); |
704 | if (ret) { |
705 | DRM_ERROR("Failed to increment BO usecnt\n" ); |
706 | return ERR_PTR(error: ret); |
707 | } |
708 | |
709 | dmabuf = drm_gem_prime_export(obj, flags); |
710 | if (IS_ERR(ptr: dmabuf)) |
711 | vc4_bo_dec_usecnt(bo); |
712 | |
713 | return dmabuf; |
714 | } |
715 | |
716 | static vm_fault_t vc4_fault(struct vm_fault *vmf) |
717 | { |
718 | struct vm_area_struct *vma = vmf->vma; |
719 | struct drm_gem_object *obj = vma->vm_private_data; |
720 | struct vc4_bo *bo = to_vc4_bo(obj); |
721 | |
722 | /* The only reason we would end up here is when user-space accesses |
723 | * BO's memory after it's been purged. |
724 | */ |
725 | mutex_lock(&bo->madv_lock); |
726 | WARN_ON(bo->madv != __VC4_MADV_PURGED); |
727 | mutex_unlock(lock: &bo->madv_lock); |
728 | |
729 | return VM_FAULT_SIGBUS; |
730 | } |
731 | |
732 | static int vc4_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) |
733 | { |
734 | struct vc4_bo *bo = to_vc4_bo(obj); |
735 | |
736 | if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) { |
737 | DRM_DEBUG("mmapping of shader BOs for writing not allowed.\n" ); |
738 | return -EINVAL; |
739 | } |
740 | |
741 | if (bo->madv != VC4_MADV_WILLNEED) { |
742 | DRM_DEBUG("mmapping of %s BO not allowed\n" , |
743 | bo->madv == VC4_MADV_DONTNEED ? |
744 | "purgeable" : "purged" ); |
745 | return -EINVAL; |
746 | } |
747 | |
748 | return drm_gem_dma_mmap(dma_obj: &bo->base, vma); |
749 | } |
750 | |
751 | static const struct vm_operations_struct vc4_vm_ops = { |
752 | .fault = vc4_fault, |
753 | .open = drm_gem_vm_open, |
754 | .close = drm_gem_vm_close, |
755 | }; |
756 | |
757 | static const struct drm_gem_object_funcs vc4_gem_object_funcs = { |
758 | .free = vc4_free_object, |
759 | .export = vc4_prime_export, |
760 | .get_sg_table = drm_gem_dma_object_get_sg_table, |
761 | .vmap = drm_gem_dma_object_vmap, |
762 | .mmap = vc4_gem_object_mmap, |
763 | .vm_ops = &vc4_vm_ops, |
764 | }; |
765 | |
766 | static int vc4_grab_bin_bo(struct vc4_dev *vc4, struct vc4_file *vc4file) |
767 | { |
768 | if (!vc4->v3d) |
769 | return -ENODEV; |
770 | |
771 | if (vc4file->bin_bo_used) |
772 | return 0; |
773 | |
774 | return vc4_v3d_bin_bo_get(vc4, used: &vc4file->bin_bo_used); |
775 | } |
776 | |
777 | int vc4_create_bo_ioctl(struct drm_device *dev, void *data, |
778 | struct drm_file *file_priv) |
779 | { |
780 | struct drm_vc4_create_bo *args = data; |
781 | struct vc4_file *vc4file = file_priv->driver_priv; |
782 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
783 | struct vc4_bo *bo = NULL; |
784 | int ret; |
785 | |
786 | if (WARN_ON_ONCE(vc4->is_vc5)) |
787 | return -ENODEV; |
788 | |
789 | ret = vc4_grab_bin_bo(vc4, vc4file); |
790 | if (ret) |
791 | return ret; |
792 | |
793 | /* |
794 | * We can't allocate from the BO cache, because the BOs don't |
795 | * get zeroed, and that might leak data between users. |
796 | */ |
797 | bo = vc4_bo_create(dev, unaligned_size: args->size, allow_unzeroed: false, type: VC4_BO_TYPE_V3D); |
798 | if (IS_ERR(ptr: bo)) |
799 | return PTR_ERR(ptr: bo); |
800 | |
801 | bo->madv = VC4_MADV_WILLNEED; |
802 | |
803 | ret = drm_gem_handle_create(file_priv, obj: &bo->base.base, handlep: &args->handle); |
804 | drm_gem_object_put(obj: &bo->base.base); |
805 | |
806 | return ret; |
807 | } |
808 | |
809 | int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data, |
810 | struct drm_file *file_priv) |
811 | { |
812 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
813 | struct drm_vc4_mmap_bo *args = data; |
814 | struct drm_gem_object *gem_obj; |
815 | |
816 | if (WARN_ON_ONCE(vc4->is_vc5)) |
817 | return -ENODEV; |
818 | |
819 | gem_obj = drm_gem_object_lookup(filp: file_priv, handle: args->handle); |
820 | if (!gem_obj) { |
821 | DRM_DEBUG("Failed to look up GEM BO %d\n" , args->handle); |
822 | return -EINVAL; |
823 | } |
824 | |
825 | /* The mmap offset was set up at BO allocation time. */ |
826 | args->offset = drm_vma_node_offset_addr(node: &gem_obj->vma_node); |
827 | |
828 | drm_gem_object_put(obj: gem_obj); |
829 | return 0; |
830 | } |
831 | |
832 | int |
833 | vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data, |
834 | struct drm_file *file_priv) |
835 | { |
836 | struct drm_vc4_create_shader_bo *args = data; |
837 | struct vc4_file *vc4file = file_priv->driver_priv; |
838 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
839 | struct vc4_bo *bo = NULL; |
840 | int ret; |
841 | |
842 | if (WARN_ON_ONCE(vc4->is_vc5)) |
843 | return -ENODEV; |
844 | |
845 | if (args->size == 0) |
846 | return -EINVAL; |
847 | |
848 | if (args->size % sizeof(u64) != 0) |
849 | return -EINVAL; |
850 | |
851 | if (args->flags != 0) { |
852 | DRM_INFO("Unknown flags set: 0x%08x\n" , args->flags); |
853 | return -EINVAL; |
854 | } |
855 | |
856 | if (args->pad != 0) { |
857 | DRM_INFO("Pad set: 0x%08x\n" , args->pad); |
858 | return -EINVAL; |
859 | } |
860 | |
861 | ret = vc4_grab_bin_bo(vc4, vc4file); |
862 | if (ret) |
863 | return ret; |
864 | |
865 | bo = vc4_bo_create(dev, unaligned_size: args->size, allow_unzeroed: true, type: VC4_BO_TYPE_V3D_SHADER); |
866 | if (IS_ERR(ptr: bo)) |
867 | return PTR_ERR(ptr: bo); |
868 | |
869 | bo->madv = VC4_MADV_WILLNEED; |
870 | |
871 | if (copy_from_user(to: bo->base.vaddr, |
872 | from: (void __user *)(uintptr_t)args->data, |
873 | n: 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(shader_obj: &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, obj: &bo->base.base, handlep: &args->handle); |
893 | |
894 | fail: |
895 | drm_gem_object_put(obj: &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 vc4_dev *vc4 = to_vc4_dev(dev); |
916 | struct drm_vc4_set_tiling *args = data; |
917 | struct drm_gem_object *gem_obj; |
918 | struct vc4_bo *bo; |
919 | bool t_format; |
920 | |
921 | if (WARN_ON_ONCE(vc4->is_vc5)) |
922 | return -ENODEV; |
923 | |
924 | if (args->flags != 0) |
925 | return -EINVAL; |
926 | |
927 | switch (args->modifier) { |
928 | case DRM_FORMAT_MOD_NONE: |
929 | t_format = false; |
930 | break; |
931 | case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: |
932 | t_format = true; |
933 | break; |
934 | default: |
935 | return -EINVAL; |
936 | } |
937 | |
938 | gem_obj = drm_gem_object_lookup(filp: file_priv, handle: args->handle); |
939 | if (!gem_obj) { |
940 | DRM_DEBUG("Failed to look up GEM BO %d\n" , args->handle); |
941 | return -ENOENT; |
942 | } |
943 | bo = to_vc4_bo(gem_obj); |
944 | bo->t_format = t_format; |
945 | |
946 | drm_gem_object_put(obj: gem_obj); |
947 | |
948 | return 0; |
949 | } |
950 | |
951 | /** |
952 | * vc4_get_tiling_ioctl() - Gets the tiling modifier for a BO. |
953 | * @dev: DRM device |
954 | * @data: ioctl argument |
955 | * @file_priv: DRM file for this fd |
956 | * |
957 | * Returns the tiling modifier for a BO as set by vc4_set_tiling_ioctl(). |
958 | */ |
959 | int vc4_get_tiling_ioctl(struct drm_device *dev, void *data, |
960 | struct drm_file *file_priv) |
961 | { |
962 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
963 | struct drm_vc4_get_tiling *args = data; |
964 | struct drm_gem_object *gem_obj; |
965 | struct vc4_bo *bo; |
966 | |
967 | if (WARN_ON_ONCE(vc4->is_vc5)) |
968 | return -ENODEV; |
969 | |
970 | if (args->flags != 0 || args->modifier != 0) |
971 | return -EINVAL; |
972 | |
973 | gem_obj = drm_gem_object_lookup(filp: file_priv, handle: args->handle); |
974 | if (!gem_obj) { |
975 | DRM_DEBUG("Failed to look up GEM BO %d\n" , args->handle); |
976 | return -ENOENT; |
977 | } |
978 | bo = to_vc4_bo(gem_obj); |
979 | |
980 | if (bo->t_format) |
981 | args->modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED; |
982 | else |
983 | args->modifier = DRM_FORMAT_MOD_NONE; |
984 | |
985 | drm_gem_object_put(obj: gem_obj); |
986 | |
987 | return 0; |
988 | } |
989 | |
990 | int vc4_bo_debugfs_init(struct drm_minor *minor) |
991 | { |
992 | struct drm_device *drm = minor->dev; |
993 | struct vc4_dev *vc4 = to_vc4_dev(drm); |
994 | |
995 | if (!vc4->v3d) |
996 | return -ENODEV; |
997 | |
998 | drm_debugfs_add_file(dev: drm, name: "bo_stats" , show: vc4_bo_stats_debugfs, NULL); |
999 | |
1000 | return 0; |
1001 | } |
1002 | |
1003 | static void vc4_bo_cache_destroy(struct drm_device *dev, void *unused); |
1004 | int vc4_bo_cache_init(struct drm_device *dev) |
1005 | { |
1006 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
1007 | int ret; |
1008 | int i; |
1009 | |
1010 | if (WARN_ON_ONCE(vc4->is_vc5)) |
1011 | return -ENODEV; |
1012 | |
1013 | /* Create the initial set of BO labels that the kernel will |
1014 | * use. This lets us avoid a bunch of string reallocation in |
1015 | * the kernel's draw and BO allocation paths. |
1016 | */ |
1017 | vc4->bo_labels = kcalloc(n: VC4_BO_TYPE_COUNT, size: sizeof(*vc4->bo_labels), |
1018 | GFP_KERNEL); |
1019 | if (!vc4->bo_labels) |
1020 | return -ENOMEM; |
1021 | vc4->num_labels = VC4_BO_TYPE_COUNT; |
1022 | |
1023 | BUILD_BUG_ON(ARRAY_SIZE(bo_type_names) != VC4_BO_TYPE_COUNT); |
1024 | for (i = 0; i < VC4_BO_TYPE_COUNT; i++) |
1025 | vc4->bo_labels[i].name = bo_type_names[i]; |
1026 | |
1027 | ret = drmm_mutex_init(dev, &vc4->bo_lock); |
1028 | if (ret) { |
1029 | kfree(objp: vc4->bo_labels); |
1030 | return ret; |
1031 | } |
1032 | |
1033 | INIT_LIST_HEAD(list: &vc4->bo_cache.time_list); |
1034 | |
1035 | INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work); |
1036 | timer_setup(&vc4->bo_cache.time_timer, vc4_bo_cache_time_timer, 0); |
1037 | |
1038 | return drmm_add_action_or_reset(dev, vc4_bo_cache_destroy, NULL); |
1039 | } |
1040 | |
1041 | static void vc4_bo_cache_destroy(struct drm_device *dev, void *unused) |
1042 | { |
1043 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
1044 | int i; |
1045 | |
1046 | del_timer(timer: &vc4->bo_cache.time_timer); |
1047 | cancel_work_sync(work: &vc4->bo_cache.time_work); |
1048 | |
1049 | vc4_bo_cache_purge(dev); |
1050 | |
1051 | for (i = 0; i < vc4->num_labels; i++) { |
1052 | if (vc4->bo_labels[i].num_allocated) { |
1053 | DRM_ERROR("Destroying BO cache with %d %s " |
1054 | "BOs still allocated\n" , |
1055 | vc4->bo_labels[i].num_allocated, |
1056 | vc4->bo_labels[i].name); |
1057 | } |
1058 | |
1059 | if (is_user_label(label: i)) |
1060 | kfree(objp: vc4->bo_labels[i].name); |
1061 | } |
1062 | kfree(objp: vc4->bo_labels); |
1063 | } |
1064 | |
1065 | int vc4_label_bo_ioctl(struct drm_device *dev, void *data, |
1066 | struct drm_file *file_priv) |
1067 | { |
1068 | struct vc4_dev *vc4 = to_vc4_dev(dev); |
1069 | struct drm_vc4_label_bo *args = data; |
1070 | char *name; |
1071 | struct drm_gem_object *gem_obj; |
1072 | int ret = 0, label; |
1073 | |
1074 | if (WARN_ON_ONCE(vc4->is_vc5)) |
1075 | return -ENODEV; |
1076 | |
1077 | if (!args->len) |
1078 | return -EINVAL; |
1079 | |
1080 | name = strndup_user(u64_to_user_ptr(args->name), args->len + 1); |
1081 | if (IS_ERR(ptr: name)) |
1082 | return PTR_ERR(ptr: name); |
1083 | |
1084 | gem_obj = drm_gem_object_lookup(filp: file_priv, handle: args->handle); |
1085 | if (!gem_obj) { |
1086 | DRM_ERROR("Failed to look up GEM BO %d\n" , args->handle); |
1087 | kfree(objp: name); |
1088 | return -ENOENT; |
1089 | } |
1090 | |
1091 | mutex_lock(&vc4->bo_lock); |
1092 | label = vc4_get_user_label(vc4, name); |
1093 | if (label != -1) |
1094 | vc4_bo_set_label(gem_obj, label); |
1095 | else |
1096 | ret = -ENOMEM; |
1097 | mutex_unlock(lock: &vc4->bo_lock); |
1098 | |
1099 | drm_gem_object_put(obj: gem_obj); |
1100 | |
1101 | return ret; |
1102 | } |
1103 | |