1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2015-2018 Broadcom */
5 * DOC: V3D GEM BO management support
7 * Compared to VC4 (V3D 2.x), V3D 3.3 introduces an MMU between the
8 * GPU and the bus, allowing us to use shmem objects for our storage
11 * Physically contiguous objects may still be imported to V3D, but the
12 * driver doesn't allocate physically contiguous objects on its own.
13 * Display engines requiring physically contiguous allocations should
14 * look into Mesa's "renderonly" support (as used by the Mesa pl111
15 * driver) for an example of how to integrate with V3D.
17 * Long term, we should support evicting pages from the MMU when under
18 * memory pressure (thus the v3d_bo_get_pages() refcounting), but
19 * that's not a high priority since our systems tend to not have swap.
22 #include <linux/dma-buf.h>
23 #include <linux/pfn_t.h>
24 #include <linux/vmalloc.h>
27 #include "uapi/drm/v3d_drm.h"
29 /* Called DRM core on the last userspace/kernel unreference of the
32 void v3d_free_object(struct drm_gem_object *obj)
34 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
35 struct v3d_bo *bo = to_v3d_bo(obj);
40 v3d_mmu_remove_ptes(bo);
42 mutex_lock(&v3d->bo_lock);
43 v3d->bo_stats.num_allocated--;
44 v3d->bo_stats.pages_allocated -= obj->size >> V3D_MMU_PAGE_SHIFT;
45 mutex_unlock(&v3d->bo_lock);
47 spin_lock(&v3d->mm_lock);
48 drm_mm_remove_node(&bo->node);
49 spin_unlock(&v3d->mm_lock);
51 /* GPU execution may have dirtied any pages in the BO. */
52 bo->base.pages_mark_dirty_on_put = true;
54 drm_gem_shmem_free(&bo->base);
57 static const struct drm_gem_object_funcs v3d_gem_funcs = {
58 .free = v3d_free_object,
59 .print_info = drm_gem_shmem_object_print_info,
60 .pin = drm_gem_shmem_object_pin,
61 .unpin = drm_gem_shmem_object_unpin,
62 .get_sg_table = drm_gem_shmem_object_get_sg_table,
63 .vmap = drm_gem_shmem_object_vmap,
64 .vunmap = drm_gem_shmem_object_vunmap,
65 .mmap = drm_gem_shmem_object_mmap,
66 .vm_ops = &drm_gem_shmem_vm_ops,
69 /* gem_create_object function for allocating a BO struct and doing
72 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size)
75 struct drm_gem_object *obj;
78 return ERR_PTR(-EINVAL);
80 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
82 return ERR_PTR(-ENOMEM);
85 obj->funcs = &v3d_gem_funcs;
86 bo->base.map_wc = true;
87 INIT_LIST_HEAD(&bo->unref_head);
89 return &bo->base.base;
93 v3d_bo_create_finish(struct drm_gem_object *obj)
95 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
96 struct v3d_bo *bo = to_v3d_bo(obj);
100 /* So far we pin the BO in the MMU for its lifetime, so use
101 * shmem's helper for getting a lifetime sgt.
103 sgt = drm_gem_shmem_get_pages_sgt(&bo->base);
107 spin_lock(&v3d->mm_lock);
108 /* Allocate the object's space in the GPU's page tables.
109 * Inserting PTEs will happen later, but the offset is for the
110 * lifetime of the BO.
112 ret = drm_mm_insert_node_generic(&v3d->mm, &bo->node,
113 obj->size >> V3D_MMU_PAGE_SHIFT,
114 GMP_GRANULARITY >> V3D_MMU_PAGE_SHIFT, 0, 0);
115 spin_unlock(&v3d->mm_lock);
119 /* Track stats for /debug/dri/n/bo_stats. */
120 mutex_lock(&v3d->bo_lock);
121 v3d->bo_stats.num_allocated++;
122 v3d->bo_stats.pages_allocated += obj->size >> V3D_MMU_PAGE_SHIFT;
123 mutex_unlock(&v3d->bo_lock);
125 v3d_mmu_insert_ptes(bo);
130 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
131 size_t unaligned_size)
133 struct drm_gem_shmem_object *shmem_obj;
137 shmem_obj = drm_gem_shmem_create(dev, unaligned_size);
138 if (IS_ERR(shmem_obj))
139 return ERR_CAST(shmem_obj);
140 bo = to_v3d_bo(&shmem_obj->base);
143 ret = v3d_bo_create_finish(&shmem_obj->base);
150 drm_gem_shmem_free(shmem_obj);
154 struct drm_gem_object *
155 v3d_prime_import_sg_table(struct drm_device *dev,
156 struct dma_buf_attachment *attach,
157 struct sg_table *sgt)
159 struct drm_gem_object *obj;
162 obj = drm_gem_shmem_prime_import_sg_table(dev, attach, sgt);
166 ret = v3d_bo_create_finish(obj);
168 drm_gem_shmem_free(&to_v3d_bo(obj)->base);
175 void v3d_get_bo_vaddr(struct v3d_bo *bo)
177 struct drm_gem_shmem_object *obj = &bo->base;
179 bo->vaddr = vmap(obj->pages, obj->base.size >> PAGE_SHIFT, VM_MAP,
180 pgprot_writecombine(PAGE_KERNEL));
183 void v3d_put_bo_vaddr(struct v3d_bo *bo)
189 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
190 struct drm_file *file_priv)
192 struct drm_v3d_create_bo *args = data;
193 struct v3d_bo *bo = NULL;
196 if (args->flags != 0) {
197 DRM_INFO("unknown create_bo flags: %d\n", args->flags);
201 bo = v3d_bo_create(dev, file_priv, PAGE_ALIGN(args->size));
205 args->offset = bo->node.start << V3D_MMU_PAGE_SHIFT;
207 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
208 drm_gem_object_put(&bo->base.base);
213 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
214 struct drm_file *file_priv)
216 struct drm_v3d_mmap_bo *args = data;
217 struct drm_gem_object *gem_obj;
219 if (args->flags != 0) {
220 DRM_INFO("unknown mmap_bo flags: %d\n", args->flags);
224 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
226 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
230 args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
231 drm_gem_object_put(gem_obj);
236 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
237 struct drm_file *file_priv)
239 struct drm_v3d_get_bo_offset *args = data;
240 struct drm_gem_object *gem_obj;
243 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
245 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
248 bo = to_v3d_bo(gem_obj);
250 args->offset = bo->node.start << V3D_MMU_PAGE_SHIFT;
252 drm_gem_object_put(gem_obj);
257 v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
258 struct drm_file *file_priv)
261 struct drm_v3d_wait_bo *args = data;
262 ktime_t start = ktime_get();
264 unsigned long timeout_jiffies =
265 nsecs_to_jiffies_timeout(args->timeout_ns);
270 ret = drm_gem_dma_resv_wait(file_priv, args->handle,
271 true, timeout_jiffies);
273 /* Decrement the user's timeout, in case we got interrupted
274 * such that the ioctl will be restarted.
276 delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
277 if (delta_ns < args->timeout_ns)
278 args->timeout_ns -= delta_ns;
280 args->timeout_ns = 0;
282 /* Asked to wait beyond the jiffie/scheduler precision? */
283 if (ret == -ETIME && args->timeout_ns)