Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/* Copyright (C) 2015-2018 Broadcom */
3
4/**
5 * DOC: V3D GEM BO management support
6 *
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
9 * instead of CMA.
10 *
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.
16 *
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.
20 */
21
22#include <linux/dma-buf.h>
23#include <linux/pfn_t.h>
24
25#include "v3d_drv.h"
26#include "uapi/drm/v3d_drm.h"
27
28/* Called DRM core on the last userspace/kernel unreference of the
29 * BO.
30 */
31void v3d_free_object(struct drm_gem_object *obj)
32{
33 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
34 struct v3d_bo *bo = to_v3d_bo(obj);
35
36 v3d_mmu_remove_ptes(bo);
37
38 mutex_lock(&v3d->bo_lock);
39 v3d->bo_stats.num_allocated--;
40 v3d->bo_stats.pages_allocated -= obj->size >> PAGE_SHIFT;
41 mutex_unlock(&v3d->bo_lock);
42
43 spin_lock(&v3d->mm_lock);
44 drm_mm_remove_node(&bo->node);
45 spin_unlock(&v3d->mm_lock);
46
47 /* GPU execution may have dirtied any pages in the BO. */
48 bo->base.pages_mark_dirty_on_put = true;
49
50 drm_gem_shmem_free_object(obj);
51}
52
53static const struct drm_gem_object_funcs v3d_gem_funcs = {
54 .free = v3d_free_object,
55 .print_info = drm_gem_shmem_print_info,
56 .pin = drm_gem_shmem_pin,
57 .unpin = drm_gem_shmem_unpin,
58 .get_sg_table = drm_gem_shmem_get_sg_table,
59 .vmap = drm_gem_shmem_vmap,
60 .vunmap = drm_gem_shmem_vunmap,
61 .mmap = drm_gem_shmem_mmap,
62};
63
64/* gem_create_object function for allocating a BO struct and doing
65 * early setup.
66 */
67struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size)
68{
69 struct v3d_bo *bo;
70 struct drm_gem_object *obj;
71
72 if (size == 0)
73 return NULL;
74
75 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
76 if (!bo)
77 return NULL;
78 obj = &bo->base.base;
79
80 obj->funcs = &v3d_gem_funcs;
81
82 INIT_LIST_HEAD(&bo->unref_head);
83
84 return &bo->base.base;
85}
86
87static int
88v3d_bo_create_finish(struct drm_gem_object *obj)
89{
90 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
91 struct v3d_bo *bo = to_v3d_bo(obj);
92 struct sg_table *sgt;
93 int ret;
94
95 /* So far we pin the BO in the MMU for its lifetime, so use
96 * shmem's helper for getting a lifetime sgt.
97 */
98 sgt = drm_gem_shmem_get_pages_sgt(&bo->base.base);
99 if (IS_ERR(sgt))
100 return PTR_ERR(sgt);
101
102 spin_lock(&v3d->mm_lock);
103 /* Allocate the object's space in the GPU's page tables.
104 * Inserting PTEs will happen later, but the offset is for the
105 * lifetime of the BO.
106 */
107 ret = drm_mm_insert_node_generic(&v3d->mm, &bo->node,
108 obj->size >> PAGE_SHIFT,
109 GMP_GRANULARITY >> PAGE_SHIFT, 0, 0);
110 spin_unlock(&v3d->mm_lock);
111 if (ret)
112 return ret;
113
114 /* Track stats for /debug/dri/n/bo_stats. */
115 mutex_lock(&v3d->bo_lock);
116 v3d->bo_stats.num_allocated++;
117 v3d->bo_stats.pages_allocated += obj->size >> PAGE_SHIFT;
118 mutex_unlock(&v3d->bo_lock);
119
120 v3d_mmu_insert_ptes(bo);
121
122 return 0;
123}
124
125struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
126 size_t unaligned_size)
127{
128 struct drm_gem_shmem_object *shmem_obj;
129 struct v3d_bo *bo;
130 int ret;
131
132 shmem_obj = drm_gem_shmem_create(dev, unaligned_size);
133 if (IS_ERR(shmem_obj))
134 return ERR_CAST(shmem_obj);
135 bo = to_v3d_bo(&shmem_obj->base);
136
137 ret = v3d_bo_create_finish(&shmem_obj->base);
138 if (ret)
139 goto free_obj;
140
141 return bo;
142
143free_obj:
144 drm_gem_shmem_free_object(&shmem_obj->base);
145 return ERR_PTR(ret);
146}
147
148struct drm_gem_object *
149v3d_prime_import_sg_table(struct drm_device *dev,
150 struct dma_buf_attachment *attach,
151 struct sg_table *sgt)
152{
153 struct drm_gem_object *obj;
154 int ret;
155
156 obj = drm_gem_shmem_prime_import_sg_table(dev, attach, sgt);
157 if (IS_ERR(obj))
158 return obj;
159
160 ret = v3d_bo_create_finish(obj);
161 if (ret) {
162 drm_gem_shmem_free_object(obj);
163 return ERR_PTR(ret);
164 }
165
166 return obj;
167}
168
169int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
170 struct drm_file *file_priv)
171{
172 struct drm_v3d_create_bo *args = data;
173 struct v3d_bo *bo = NULL;
174 int ret;
175
176 if (args->flags != 0) {
177 DRM_INFO("unknown create_bo flags: %d\n", args->flags);
178 return -EINVAL;
179 }
180
181 bo = v3d_bo_create(dev, file_priv, PAGE_ALIGN(args->size));
182 if (IS_ERR(bo))
183 return PTR_ERR(bo);
184
185 args->offset = bo->node.start << PAGE_SHIFT;
186
187 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
188 drm_gem_object_put(&bo->base.base);
189
190 return ret;
191}
192
193int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
194 struct drm_file *file_priv)
195{
196 struct drm_v3d_mmap_bo *args = data;
197 struct drm_gem_object *gem_obj;
198
199 if (args->flags != 0) {
200 DRM_INFO("unknown mmap_bo flags: %d\n", args->flags);
201 return -EINVAL;
202 }
203
204 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
205 if (!gem_obj) {
206 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
207 return -ENOENT;
208 }
209
210 args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
211 drm_gem_object_put(gem_obj);
212
213 return 0;
214}
215
216int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
217 struct drm_file *file_priv)
218{
219 struct drm_v3d_get_bo_offset *args = data;
220 struct drm_gem_object *gem_obj;
221 struct v3d_bo *bo;
222
223 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
224 if (!gem_obj) {
225 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
226 return -ENOENT;
227 }
228 bo = to_v3d_bo(gem_obj);
229
230 args->offset = bo->node.start << PAGE_SHIFT;
231
232 drm_gem_object_put(gem_obj);
233 return 0;
234}
1// SPDX-License-Identifier: GPL-2.0+
2/* Copyright (C) 2015-2018 Broadcom */
3
4/**
5 * DOC: V3D GEM BO management support
6 *
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
9 * instead of CMA.
10 *
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.
16 *
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.
20 */
21
22#include <linux/dma-buf.h>
23#include <linux/pfn_t.h>
24#include <linux/vmalloc.h>
25
26#include "v3d_drv.h"
27#include "uapi/drm/v3d_drm.h"
28
29static enum drm_gem_object_status v3d_gem_status(struct drm_gem_object *obj)
30{
31 struct v3d_bo *bo = to_v3d_bo(obj);
32 enum drm_gem_object_status res = 0;
33
34 if (bo->base.pages)
35 res |= DRM_GEM_OBJECT_RESIDENT;
36
37 return res;
38}
39
40/* Called DRM core on the last userspace/kernel unreference of the
41 * BO.
42 */
43void v3d_free_object(struct drm_gem_object *obj)
44{
45 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
46 struct v3d_bo *bo = to_v3d_bo(obj);
47
48 if (bo->vaddr)
49 v3d_put_bo_vaddr(bo);
50
51 v3d_mmu_remove_ptes(bo);
52
53 mutex_lock(&v3d->bo_lock);
54 v3d->bo_stats.num_allocated--;
55 v3d->bo_stats.pages_allocated -= obj->size >> V3D_MMU_PAGE_SHIFT;
56 mutex_unlock(&v3d->bo_lock);
57
58 spin_lock(&v3d->mm_lock);
59 drm_mm_remove_node(&bo->node);
60 spin_unlock(&v3d->mm_lock);
61
62 /* GPU execution may have dirtied any pages in the BO. */
63 bo->base.pages_mark_dirty_on_put = true;
64
65 drm_gem_shmem_free(&bo->base);
66}
67
68static const struct drm_gem_object_funcs v3d_gem_funcs = {
69 .free = v3d_free_object,
70 .print_info = drm_gem_shmem_object_print_info,
71 .pin = drm_gem_shmem_object_pin,
72 .unpin = drm_gem_shmem_object_unpin,
73 .get_sg_table = drm_gem_shmem_object_get_sg_table,
74 .vmap = drm_gem_shmem_object_vmap,
75 .vunmap = drm_gem_shmem_object_vunmap,
76 .mmap = drm_gem_shmem_object_mmap,
77 .status = v3d_gem_status,
78 .vm_ops = &drm_gem_shmem_vm_ops,
79};
80
81/* gem_create_object function for allocating a BO struct and doing
82 * early setup.
83 */
84struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size)
85{
86 struct v3d_bo *bo;
87 struct drm_gem_object *obj;
88
89 if (size == 0)
90 return ERR_PTR(-EINVAL);
91
92 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
93 if (!bo)
94 return ERR_PTR(-ENOMEM);
95 obj = &bo->base.base;
96
97 obj->funcs = &v3d_gem_funcs;
98 bo->base.map_wc = true;
99 INIT_LIST_HEAD(&bo->unref_head);
100
101 return &bo->base.base;
102}
103
104static int
105v3d_bo_create_finish(struct drm_gem_object *obj)
106{
107 struct v3d_dev *v3d = to_v3d_dev(obj->dev);
108 struct v3d_bo *bo = to_v3d_bo(obj);
109 struct sg_table *sgt;
110 u64 align;
111 int ret;
112
113 /* So far we pin the BO in the MMU for its lifetime, so use
114 * shmem's helper for getting a lifetime sgt.
115 */
116 sgt = drm_gem_shmem_get_pages_sgt(&bo->base);
117 if (IS_ERR(sgt))
118 return PTR_ERR(sgt);
119
120 if (!v3d->gemfs)
121 align = SZ_4K;
122 else if (obj->size >= SZ_1M)
123 align = SZ_1M;
124 else if (obj->size >= SZ_64K)
125 align = SZ_64K;
126 else
127 align = SZ_4K;
128
129 spin_lock(&v3d->mm_lock);
130 /* Allocate the object's space in the GPU's page tables.
131 * Inserting PTEs will happen later, but the offset is for the
132 * lifetime of the BO.
133 */
134 ret = drm_mm_insert_node_generic(&v3d->mm, &bo->node,
135 obj->size >> V3D_MMU_PAGE_SHIFT,
136 align >> V3D_MMU_PAGE_SHIFT, 0, 0);
137 spin_unlock(&v3d->mm_lock);
138 if (ret)
139 return ret;
140
141 /* Track stats for /debug/dri/n/bo_stats. */
142 mutex_lock(&v3d->bo_lock);
143 v3d->bo_stats.num_allocated++;
144 v3d->bo_stats.pages_allocated += obj->size >> V3D_MMU_PAGE_SHIFT;
145 mutex_unlock(&v3d->bo_lock);
146
147 v3d_mmu_insert_ptes(bo);
148
149 return 0;
150}
151
152struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
153 size_t unaligned_size)
154{
155 struct drm_gem_shmem_object *shmem_obj;
156 struct v3d_dev *v3d = to_v3d_dev(dev);
157 struct v3d_bo *bo;
158 int ret;
159
160 shmem_obj = drm_gem_shmem_create_with_mnt(dev, unaligned_size,
161 v3d->gemfs);
162 if (IS_ERR(shmem_obj))
163 return ERR_CAST(shmem_obj);
164 bo = to_v3d_bo(&shmem_obj->base);
165 bo->vaddr = NULL;
166
167 ret = v3d_bo_create_finish(&shmem_obj->base);
168 if (ret)
169 goto free_obj;
170
171 return bo;
172
173free_obj:
174 drm_gem_shmem_free(shmem_obj);
175 return ERR_PTR(ret);
176}
177
178struct drm_gem_object *
179v3d_prime_import_sg_table(struct drm_device *dev,
180 struct dma_buf_attachment *attach,
181 struct sg_table *sgt)
182{
183 struct drm_gem_object *obj;
184 int ret;
185
186 obj = drm_gem_shmem_prime_import_sg_table(dev, attach, sgt);
187 if (IS_ERR(obj))
188 return obj;
189
190 ret = v3d_bo_create_finish(obj);
191 if (ret) {
192 drm_gem_shmem_free(&to_v3d_bo(obj)->base);
193 return ERR_PTR(ret);
194 }
195
196 return obj;
197}
198
199void v3d_get_bo_vaddr(struct v3d_bo *bo)
200{
201 struct drm_gem_shmem_object *obj = &bo->base;
202
203 bo->vaddr = vmap(obj->pages, obj->base.size >> PAGE_SHIFT, VM_MAP,
204 pgprot_writecombine(PAGE_KERNEL));
205}
206
207void v3d_put_bo_vaddr(struct v3d_bo *bo)
208{
209 vunmap(bo->vaddr);
210 bo->vaddr = NULL;
211}
212
213int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
214 struct drm_file *file_priv)
215{
216 struct drm_v3d_create_bo *args = data;
217 struct v3d_bo *bo = NULL;
218 int ret;
219
220 if (args->flags != 0) {
221 DRM_INFO("unknown create_bo flags: %d\n", args->flags);
222 return -EINVAL;
223 }
224
225 bo = v3d_bo_create(dev, file_priv, PAGE_ALIGN(args->size));
226 if (IS_ERR(bo))
227 return PTR_ERR(bo);
228
229 args->offset = bo->node.start << V3D_MMU_PAGE_SHIFT;
230
231 ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
232 drm_gem_object_put(&bo->base.base);
233
234 return ret;
235}
236
237int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
238 struct drm_file *file_priv)
239{
240 struct drm_v3d_mmap_bo *args = data;
241 struct drm_gem_object *gem_obj;
242
243 if (args->flags != 0) {
244 DRM_INFO("unknown mmap_bo flags: %d\n", args->flags);
245 return -EINVAL;
246 }
247
248 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
249 if (!gem_obj) {
250 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
251 return -ENOENT;
252 }
253
254 args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
255 drm_gem_object_put(gem_obj);
256
257 return 0;
258}
259
260int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
261 struct drm_file *file_priv)
262{
263 struct drm_v3d_get_bo_offset *args = data;
264 struct drm_gem_object *gem_obj;
265 struct v3d_bo *bo;
266
267 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
268 if (!gem_obj) {
269 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
270 return -ENOENT;
271 }
272 bo = to_v3d_bo(gem_obj);
273
274 args->offset = bo->node.start << V3D_MMU_PAGE_SHIFT;
275
276 drm_gem_object_put(gem_obj);
277 return 0;
278}
279
280int
281v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
282 struct drm_file *file_priv)
283{
284 int ret;
285 struct drm_v3d_wait_bo *args = data;
286 ktime_t start = ktime_get();
287 u64 delta_ns;
288 unsigned long timeout_jiffies =
289 nsecs_to_jiffies_timeout(args->timeout_ns);
290
291 if (args->pad != 0)
292 return -EINVAL;
293
294 ret = drm_gem_dma_resv_wait(file_priv, args->handle,
295 true, timeout_jiffies);
296
297 /* Decrement the user's timeout, in case we got interrupted
298 * such that the ioctl will be restarted.
299 */
300 delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
301 if (delta_ns < args->timeout_ns)
302 args->timeout_ns -= delta_ns;
303 else
304 args->timeout_ns = 0;
305
306 /* Asked to wait beyond the jiffy/scheduler precision? */
307 if (ret == -ETIME && args->timeout_ns)
308 ret = -EAGAIN;
309
310 return ret;
311}