2 * Copyright © 2015 Broadcom
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License version 2 as
6 * published by the Free Software Foundation.
9 /* DOC: VC4 GEM BO management support.
11 * The VC4 GPU architecture (both scanout and rendering) has direct
12 * access to system memory with no MMU in between. To support it, we
13 * use the GEM CMA helper functions to allocate contiguous ranges of
14 * physical memory for our BOs.
16 * Since the CMA allocator is very slow, we keep a cache of recently
17 * freed BOs around so that the kernel's allocation of objects for 3D
18 * rendering can return quickly.
22 #include "uapi/drm/vc4_drm.h"
24 static void vc4_bo_stats_dump(struct vc4_dev *vc4)
26 DRM_INFO("num bos allocated: %d\n",
27 vc4->bo_stats.num_allocated);
28 DRM_INFO("size bos allocated: %dkb\n",
29 vc4->bo_stats.size_allocated / 1024);
30 DRM_INFO("num bos used: %d\n",
31 vc4->bo_stats.num_allocated - vc4->bo_stats.num_cached);
32 DRM_INFO("size bos used: %dkb\n",
33 (vc4->bo_stats.size_allocated -
34 vc4->bo_stats.size_cached) / 1024);
35 DRM_INFO("num bos cached: %d\n",
36 vc4->bo_stats.num_cached);
37 DRM_INFO("size bos cached: %dkb\n",
38 vc4->bo_stats.size_cached / 1024);
41 #ifdef CONFIG_DEBUG_FS
42 int vc4_bo_stats_debugfs(struct seq_file *m, void *unused)
44 struct drm_info_node *node = (struct drm_info_node *)m->private;
45 struct drm_device *dev = node->minor->dev;
46 struct vc4_dev *vc4 = to_vc4_dev(dev);
47 struct vc4_bo_stats stats;
49 /* Take a snapshot of the current stats with the lock held. */
50 mutex_lock(&vc4->bo_lock);
51 stats = vc4->bo_stats;
52 mutex_unlock(&vc4->bo_lock);
54 seq_printf(m, "num bos allocated: %d\n",
56 seq_printf(m, "size bos allocated: %dkb\n",
57 stats.size_allocated / 1024);
58 seq_printf(m, "num bos used: %d\n",
59 stats.num_allocated - stats.num_cached);
60 seq_printf(m, "size bos used: %dkb\n",
61 (stats.size_allocated - stats.size_cached) / 1024);
62 seq_printf(m, "num bos cached: %d\n",
64 seq_printf(m, "size bos cached: %dkb\n",
65 stats.size_cached / 1024);
71 static uint32_t bo_page_index(size_t size)
73 return (size / PAGE_SIZE) - 1;
76 /* Must be called with bo_lock held. */
77 static void vc4_bo_destroy(struct vc4_bo *bo)
79 struct drm_gem_object *obj = &bo->base.base;
80 struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
82 if (bo->validated_shader) {
83 kfree(bo->validated_shader->uniform_addr_offsets);
84 kfree(bo->validated_shader->texture_samples);
85 kfree(bo->validated_shader);
86 bo->validated_shader = NULL;
89 vc4->bo_stats.num_allocated--;
90 vc4->bo_stats.size_allocated -= obj->size;
91 drm_gem_cma_free_object(obj);
94 /* Must be called with bo_lock held. */
95 static void vc4_bo_remove_from_cache(struct vc4_bo *bo)
97 struct drm_gem_object *obj = &bo->base.base;
98 struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
100 vc4->bo_stats.num_cached--;
101 vc4->bo_stats.size_cached -= obj->size;
103 list_del(&bo->unref_head);
104 list_del(&bo->size_head);
107 static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev,
110 struct vc4_dev *vc4 = to_vc4_dev(dev);
111 uint32_t page_index = bo_page_index(size);
113 if (vc4->bo_cache.size_list_size <= page_index) {
114 uint32_t new_size = max(vc4->bo_cache.size_list_size * 2,
116 struct list_head *new_list;
119 new_list = kmalloc_array(new_size, sizeof(struct list_head),
124 /* Rebase the old cached BO lists to their new list
127 for (i = 0; i < vc4->bo_cache.size_list_size; i++) {
128 struct list_head *old_list =
129 &vc4->bo_cache.size_list[i];
131 if (list_empty(old_list))
132 INIT_LIST_HEAD(&new_list[i]);
134 list_replace(old_list, &new_list[i]);
136 /* And initialize the brand new BO list heads. */
137 for (i = vc4->bo_cache.size_list_size; i < new_size; i++)
138 INIT_LIST_HEAD(&new_list[i]);
140 kfree(vc4->bo_cache.size_list);
141 vc4->bo_cache.size_list = new_list;
142 vc4->bo_cache.size_list_size = new_size;
145 return &vc4->bo_cache.size_list[page_index];
148 static void vc4_bo_cache_purge(struct drm_device *dev)
150 struct vc4_dev *vc4 = to_vc4_dev(dev);
152 mutex_lock(&vc4->bo_lock);
153 while (!list_empty(&vc4->bo_cache.time_list)) {
154 struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
155 struct vc4_bo, unref_head);
156 vc4_bo_remove_from_cache(bo);
159 mutex_unlock(&vc4->bo_lock);
162 static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
165 struct vc4_dev *vc4 = to_vc4_dev(dev);
166 uint32_t page_index = bo_page_index(size);
167 struct vc4_bo *bo = NULL;
169 size = roundup(size, PAGE_SIZE);
171 mutex_lock(&vc4->bo_lock);
172 if (page_index >= vc4->bo_cache.size_list_size)
175 if (list_empty(&vc4->bo_cache.size_list[page_index]))
178 bo = list_first_entry(&vc4->bo_cache.size_list[page_index],
179 struct vc4_bo, size_head);
180 vc4_bo_remove_from_cache(bo);
181 kref_init(&bo->base.base.refcount);
184 mutex_unlock(&vc4->bo_lock);
189 * vc4_gem_create_object - Implementation of driver->gem_create_object.
191 * This lets the CMA helpers allocate object structs for us, and keep
192 * our BO stats correct.
194 struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
196 struct vc4_dev *vc4 = to_vc4_dev(dev);
199 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
203 mutex_lock(&vc4->bo_lock);
204 vc4->bo_stats.num_allocated++;
205 vc4->bo_stats.size_allocated += size;
206 mutex_unlock(&vc4->bo_lock);
208 return &bo->base.base;
211 struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
214 size_t size = roundup(unaligned_size, PAGE_SIZE);
215 struct vc4_dev *vc4 = to_vc4_dev(dev);
216 struct drm_gem_cma_object *cma_obj;
219 return ERR_PTR(-EINVAL);
221 /* First, try to get a vc4_bo from the kernel BO cache. */
223 struct vc4_bo *bo = vc4_bo_get_from_cache(dev, size);
229 cma_obj = drm_gem_cma_create(dev, size);
230 if (IS_ERR(cma_obj)) {
232 * If we've run out of CMA memory, kill the cache of
233 * CMA allocations we've got laying around and try again.
235 vc4_bo_cache_purge(dev);
237 cma_obj = drm_gem_cma_create(dev, size);
238 if (IS_ERR(cma_obj)) {
239 DRM_ERROR("Failed to allocate from CMA:\n");
240 vc4_bo_stats_dump(vc4);
241 return ERR_PTR(-ENOMEM);
245 return to_vc4_bo(&cma_obj->base);
248 int vc4_dumb_create(struct drm_file *file_priv,
249 struct drm_device *dev,
250 struct drm_mode_create_dumb *args)
252 int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
253 struct vc4_bo *bo = NULL;
256 if (args->pitch < min_pitch)
257 args->pitch = min_pitch;
259 if (args->size < args->pitch * args->height)
260 args->size = args->pitch * args->height;
262 bo = vc4_bo_create(dev, args->size, false);
266 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
267 drm_gem_object_unreference_unlocked(&bo->base.base);
272 /* Must be called with bo_lock held. */
273 static void vc4_bo_cache_free_old(struct drm_device *dev)
275 struct vc4_dev *vc4 = to_vc4_dev(dev);
276 unsigned long expire_time = jiffies - msecs_to_jiffies(1000);
278 while (!list_empty(&vc4->bo_cache.time_list)) {
279 struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
280 struct vc4_bo, unref_head);
281 if (time_before(expire_time, bo->free_time)) {
282 mod_timer(&vc4->bo_cache.time_timer,
283 round_jiffies_up(jiffies +
284 msecs_to_jiffies(1000)));
288 vc4_bo_remove_from_cache(bo);
293 /* Called on the last userspace/kernel unreference of the BO. Returns
294 * it to the BO cache if possible, otherwise frees it.
296 void vc4_free_object(struct drm_gem_object *gem_bo)
298 struct drm_device *dev = gem_bo->dev;
299 struct vc4_dev *vc4 = to_vc4_dev(dev);
300 struct vc4_bo *bo = to_vc4_bo(gem_bo);
301 struct list_head *cache_list;
303 mutex_lock(&vc4->bo_lock);
304 /* If the object references someone else's memory, we can't cache it.
306 if (gem_bo->import_attach) {
311 /* Don't cache if it was publicly named. */
317 /* If this object was partially constructed but CMA allocation
318 * had failed, just free it.
320 if (!bo->base.vaddr) {
325 cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size);
331 if (bo->validated_shader) {
332 kfree(bo->validated_shader->uniform_addr_offsets);
333 kfree(bo->validated_shader->texture_samples);
334 kfree(bo->validated_shader);
335 bo->validated_shader = NULL;
338 bo->free_time = jiffies;
339 list_add(&bo->size_head, cache_list);
340 list_add(&bo->unref_head, &vc4->bo_cache.time_list);
342 vc4->bo_stats.num_cached++;
343 vc4->bo_stats.size_cached += gem_bo->size;
345 vc4_bo_cache_free_old(dev);
348 mutex_unlock(&vc4->bo_lock);
351 static void vc4_bo_cache_time_work(struct work_struct *work)
353 struct vc4_dev *vc4 =
354 container_of(work, struct vc4_dev, bo_cache.time_work);
355 struct drm_device *dev = vc4->dev;
357 mutex_lock(&vc4->bo_lock);
358 vc4_bo_cache_free_old(dev);
359 mutex_unlock(&vc4->bo_lock);
362 static void vc4_bo_cache_time_timer(unsigned long data)
364 struct drm_device *dev = (struct drm_device *)data;
365 struct vc4_dev *vc4 = to_vc4_dev(dev);
367 schedule_work(&vc4->bo_cache.time_work);
371 vc4_prime_export(struct drm_device *dev, struct drm_gem_object *obj, int flags)
373 struct vc4_bo *bo = to_vc4_bo(obj);
375 if (bo->validated_shader) {
376 DRM_ERROR("Attempting to export shader BO\n");
377 return ERR_PTR(-EINVAL);
380 return drm_gem_prime_export(dev, obj, flags);
383 int vc4_mmap(struct file *filp, struct vm_area_struct *vma)
385 struct drm_gem_object *gem_obj;
389 ret = drm_gem_mmap(filp, vma);
393 gem_obj = vma->vm_private_data;
394 bo = to_vc4_bo(gem_obj);
396 if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
397 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
402 * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the
403 * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map
406 vma->vm_flags &= ~VM_PFNMAP;
409 ret = dma_mmap_wc(bo->base.base.dev->dev, vma, bo->base.vaddr,
410 bo->base.paddr, vma->vm_end - vma->vm_start);
412 drm_gem_vm_close(vma);
417 int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
419 struct vc4_bo *bo = to_vc4_bo(obj);
421 if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
422 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
426 return drm_gem_cma_prime_mmap(obj, vma);
429 void *vc4_prime_vmap(struct drm_gem_object *obj)
431 struct vc4_bo *bo = to_vc4_bo(obj);
433 if (bo->validated_shader) {
434 DRM_ERROR("mmaping of shader BOs not allowed.\n");
435 return ERR_PTR(-EINVAL);
438 return drm_gem_cma_prime_vmap(obj);
441 int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
442 struct drm_file *file_priv)
444 struct drm_vc4_create_bo *args = data;
445 struct vc4_bo *bo = NULL;
449 * We can't allocate from the BO cache, because the BOs don't
450 * get zeroed, and that might leak data between users.
452 bo = vc4_bo_create(dev, args->size, false);
456 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
457 drm_gem_object_unreference_unlocked(&bo->base.base);
462 int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
463 struct drm_file *file_priv)
465 struct drm_vc4_mmap_bo *args = data;
466 struct drm_gem_object *gem_obj;
468 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
470 DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
474 /* The mmap offset was set up at BO allocation time. */
475 args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
477 drm_gem_object_unreference_unlocked(gem_obj);
482 vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
483 struct drm_file *file_priv)
485 struct drm_vc4_create_shader_bo *args = data;
486 struct vc4_bo *bo = NULL;
492 if (args->size % sizeof(u64) != 0)
495 if (args->flags != 0) {
496 DRM_INFO("Unknown flags set: 0x%08x\n", args->flags);
500 if (args->pad != 0) {
501 DRM_INFO("Pad set: 0x%08x\n", args->pad);
505 bo = vc4_bo_create(dev, args->size, true);
509 if (copy_from_user(bo->base.vaddr,
510 (void __user *)(uintptr_t)args->data,
515 /* Clear the rest of the memory from allocating from the BO
518 memset(bo->base.vaddr + args->size, 0,
519 bo->base.base.size - args->size);
521 bo->validated_shader = vc4_validate_shader(&bo->base);
522 if (!bo->validated_shader) {
527 /* We have to create the handle after validation, to avoid
528 * races for users to do doing things like mmap the shader BO.
530 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
533 drm_gem_object_unreference_unlocked(&bo->base.base);
538 void vc4_bo_cache_init(struct drm_device *dev)
540 struct vc4_dev *vc4 = to_vc4_dev(dev);
542 mutex_init(&vc4->bo_lock);
544 INIT_LIST_HEAD(&vc4->bo_cache.time_list);
546 INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work);
547 setup_timer(&vc4->bo_cache.time_timer,
548 vc4_bo_cache_time_timer,
552 void vc4_bo_cache_destroy(struct drm_device *dev)
554 struct vc4_dev *vc4 = to_vc4_dev(dev);
556 del_timer(&vc4->bo_cache.time_timer);
557 cancel_work_sync(&vc4->bo_cache.time_work);
559 vc4_bo_cache_purge(dev);
561 if (vc4->bo_stats.num_allocated) {
562 DRM_ERROR("Destroying BO cache while BOs still allocated:\n");
563 vc4_bo_stats_dump(vc4);