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.
10 * DOC: VC4 GEM BO management support
12 * The VC4 GPU architecture (both scanout and rendering) has direct
13 * access to system memory with no MMU in between. To support it, we
14 * use the GEM CMA helper functions to allocate contiguous ranges of
15 * physical memory for our BOs.
17 * Since the CMA allocator is very slow, we keep a cache of recently
18 * freed BOs around so that the kernel's allocation of objects for 3D
19 * rendering can return quickly.
22 #include <linux/dma-buf.h>
25 #include "uapi/drm/vc4_drm.h"
27 static void vc4_bo_stats_dump(struct vc4_dev
*vc4
)
29 DRM_INFO("num bos allocated: %d\n",
30 vc4
->bo_stats
.num_allocated
);
31 DRM_INFO("size bos allocated: %dkb\n",
32 vc4
->bo_stats
.size_allocated
/ 1024);
33 DRM_INFO("num bos used: %d\n",
34 vc4
->bo_stats
.num_allocated
- vc4
->bo_stats
.num_cached
);
35 DRM_INFO("size bos used: %dkb\n",
36 (vc4
->bo_stats
.size_allocated
-
37 vc4
->bo_stats
.size_cached
) / 1024);
38 DRM_INFO("num bos cached: %d\n",
39 vc4
->bo_stats
.num_cached
);
40 DRM_INFO("size bos cached: %dkb\n",
41 vc4
->bo_stats
.size_cached
/ 1024);
44 #ifdef CONFIG_DEBUG_FS
45 int vc4_bo_stats_debugfs(struct seq_file
*m
, void *unused
)
47 struct drm_info_node
*node
= (struct drm_info_node
*)m
->private;
48 struct drm_device
*dev
= node
->minor
->dev
;
49 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
50 struct vc4_bo_stats stats
;
52 /* Take a snapshot of the current stats with the lock held. */
53 mutex_lock(&vc4
->bo_lock
);
54 stats
= vc4
->bo_stats
;
55 mutex_unlock(&vc4
->bo_lock
);
57 seq_printf(m
, "num bos allocated: %d\n",
59 seq_printf(m
, "size bos allocated: %dkb\n",
60 stats
.size_allocated
/ 1024);
61 seq_printf(m
, "num bos used: %d\n",
62 stats
.num_allocated
- stats
.num_cached
);
63 seq_printf(m
, "size bos used: %dkb\n",
64 (stats
.size_allocated
- stats
.size_cached
) / 1024);
65 seq_printf(m
, "num bos cached: %d\n",
67 seq_printf(m
, "size bos cached: %dkb\n",
68 stats
.size_cached
/ 1024);
74 static uint32_t bo_page_index(size_t size
)
76 return (size
/ PAGE_SIZE
) - 1;
79 /* Must be called with bo_lock held. */
80 static void vc4_bo_destroy(struct vc4_bo
*bo
)
82 struct drm_gem_object
*obj
= &bo
->base
.base
;
83 struct vc4_dev
*vc4
= to_vc4_dev(obj
->dev
);
85 if (bo
->validated_shader
) {
86 kfree(bo
->validated_shader
->texture_samples
);
87 kfree(bo
->validated_shader
);
88 bo
->validated_shader
= NULL
;
91 vc4
->bo_stats
.num_allocated
--;
92 vc4
->bo_stats
.size_allocated
-= obj
->size
;
94 reservation_object_fini(&bo
->_resv
);
96 drm_gem_cma_free_object(obj
);
99 /* Must be called with bo_lock held. */
100 static void vc4_bo_remove_from_cache(struct vc4_bo
*bo
)
102 struct drm_gem_object
*obj
= &bo
->base
.base
;
103 struct vc4_dev
*vc4
= to_vc4_dev(obj
->dev
);
105 vc4
->bo_stats
.num_cached
--;
106 vc4
->bo_stats
.size_cached
-= obj
->size
;
108 list_del(&bo
->unref_head
);
109 list_del(&bo
->size_head
);
112 static struct list_head
*vc4_get_cache_list_for_size(struct drm_device
*dev
,
115 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
116 uint32_t page_index
= bo_page_index(size
);
118 if (vc4
->bo_cache
.size_list_size
<= page_index
) {
119 uint32_t new_size
= max(vc4
->bo_cache
.size_list_size
* 2,
121 struct list_head
*new_list
;
124 new_list
= kmalloc_array(new_size
, sizeof(struct list_head
),
129 /* Rebase the old cached BO lists to their new list
132 for (i
= 0; i
< vc4
->bo_cache
.size_list_size
; i
++) {
133 struct list_head
*old_list
=
134 &vc4
->bo_cache
.size_list
[i
];
136 if (list_empty(old_list
))
137 INIT_LIST_HEAD(&new_list
[i
]);
139 list_replace(old_list
, &new_list
[i
]);
141 /* And initialize the brand new BO list heads. */
142 for (i
= vc4
->bo_cache
.size_list_size
; i
< new_size
; i
++)
143 INIT_LIST_HEAD(&new_list
[i
]);
145 kfree(vc4
->bo_cache
.size_list
);
146 vc4
->bo_cache
.size_list
= new_list
;
147 vc4
->bo_cache
.size_list_size
= new_size
;
150 return &vc4
->bo_cache
.size_list
[page_index
];
153 static void vc4_bo_cache_purge(struct drm_device
*dev
)
155 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
157 mutex_lock(&vc4
->bo_lock
);
158 while (!list_empty(&vc4
->bo_cache
.time_list
)) {
159 struct vc4_bo
*bo
= list_last_entry(&vc4
->bo_cache
.time_list
,
160 struct vc4_bo
, unref_head
);
161 vc4_bo_remove_from_cache(bo
);
164 mutex_unlock(&vc4
->bo_lock
);
167 static struct vc4_bo
*vc4_bo_get_from_cache(struct drm_device
*dev
,
170 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
171 uint32_t page_index
= bo_page_index(size
);
172 struct vc4_bo
*bo
= NULL
;
174 size
= roundup(size
, PAGE_SIZE
);
176 mutex_lock(&vc4
->bo_lock
);
177 if (page_index
>= vc4
->bo_cache
.size_list_size
)
180 if (list_empty(&vc4
->bo_cache
.size_list
[page_index
]))
183 bo
= list_first_entry(&vc4
->bo_cache
.size_list
[page_index
],
184 struct vc4_bo
, size_head
);
185 vc4_bo_remove_from_cache(bo
);
186 kref_init(&bo
->base
.base
.refcount
);
189 mutex_unlock(&vc4
->bo_lock
);
194 * vc4_gem_create_object - Implementation of driver->gem_create_object.
196 * @size: Size in bytes of the memory the object will reference
198 * This lets the CMA helpers allocate object structs for us, and keep
199 * our BO stats correct.
201 struct drm_gem_object
*vc4_create_object(struct drm_device
*dev
, size_t size
)
203 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
206 bo
= kzalloc(sizeof(*bo
), GFP_KERNEL
);
208 return ERR_PTR(-ENOMEM
);
210 mutex_lock(&vc4
->bo_lock
);
211 vc4
->bo_stats
.num_allocated
++;
212 vc4
->bo_stats
.size_allocated
+= size
;
213 mutex_unlock(&vc4
->bo_lock
);
214 bo
->resv
= &bo
->_resv
;
215 reservation_object_init(bo
->resv
);
217 return &bo
->base
.base
;
220 struct vc4_bo
*vc4_bo_create(struct drm_device
*dev
, size_t unaligned_size
,
223 size_t size
= roundup(unaligned_size
, PAGE_SIZE
);
224 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
225 struct drm_gem_cma_object
*cma_obj
;
229 return ERR_PTR(-EINVAL
);
231 /* First, try to get a vc4_bo from the kernel BO cache. */
232 bo
= vc4_bo_get_from_cache(dev
, size
);
235 memset(bo
->base
.vaddr
, 0, bo
->base
.base
.size
);
239 cma_obj
= drm_gem_cma_create(dev
, size
);
240 if (IS_ERR(cma_obj
)) {
242 * If we've run out of CMA memory, kill the cache of
243 * CMA allocations we've got laying around and try again.
245 vc4_bo_cache_purge(dev
);
247 cma_obj
= drm_gem_cma_create(dev
, size
);
248 if (IS_ERR(cma_obj
)) {
249 DRM_ERROR("Failed to allocate from CMA:\n");
250 vc4_bo_stats_dump(vc4
);
251 return ERR_PTR(-ENOMEM
);
254 return to_vc4_bo(&cma_obj
->base
);
257 int vc4_dumb_create(struct drm_file
*file_priv
,
258 struct drm_device
*dev
,
259 struct drm_mode_create_dumb
*args
)
261 int min_pitch
= DIV_ROUND_UP(args
->width
* args
->bpp
, 8);
262 struct vc4_bo
*bo
= NULL
;
265 if (args
->pitch
< min_pitch
)
266 args
->pitch
= min_pitch
;
268 if (args
->size
< args
->pitch
* args
->height
)
269 args
->size
= args
->pitch
* args
->height
;
271 bo
= vc4_bo_create(dev
, args
->size
, false);
275 ret
= drm_gem_handle_create(file_priv
, &bo
->base
.base
, &args
->handle
);
276 drm_gem_object_unreference_unlocked(&bo
->base
.base
);
281 /* Must be called with bo_lock held. */
282 static void vc4_bo_cache_free_old(struct drm_device
*dev
)
284 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
285 unsigned long expire_time
= jiffies
- msecs_to_jiffies(1000);
287 while (!list_empty(&vc4
->bo_cache
.time_list
)) {
288 struct vc4_bo
*bo
= list_last_entry(&vc4
->bo_cache
.time_list
,
289 struct vc4_bo
, unref_head
);
290 if (time_before(expire_time
, bo
->free_time
)) {
291 mod_timer(&vc4
->bo_cache
.time_timer
,
292 round_jiffies_up(jiffies
+
293 msecs_to_jiffies(1000)));
297 vc4_bo_remove_from_cache(bo
);
302 /* Called on the last userspace/kernel unreference of the BO. Returns
303 * it to the BO cache if possible, otherwise frees it.
305 void vc4_free_object(struct drm_gem_object
*gem_bo
)
307 struct drm_device
*dev
= gem_bo
->dev
;
308 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
309 struct vc4_bo
*bo
= to_vc4_bo(gem_bo
);
310 struct list_head
*cache_list
;
312 mutex_lock(&vc4
->bo_lock
);
313 /* If the object references someone else's memory, we can't cache it.
315 if (gem_bo
->import_attach
) {
320 /* Don't cache if it was publicly named. */
326 /* If this object was partially constructed but CMA allocation
327 * had failed, just free it.
329 if (!bo
->base
.vaddr
) {
334 cache_list
= vc4_get_cache_list_for_size(dev
, gem_bo
->size
);
340 if (bo
->validated_shader
) {
341 kfree(bo
->validated_shader
->texture_samples
);
342 kfree(bo
->validated_shader
);
343 bo
->validated_shader
= NULL
;
346 bo
->t_format
= false;
347 bo
->free_time
= jiffies
;
348 list_add(&bo
->size_head
, cache_list
);
349 list_add(&bo
->unref_head
, &vc4
->bo_cache
.time_list
);
351 vc4
->bo_stats
.num_cached
++;
352 vc4
->bo_stats
.size_cached
+= gem_bo
->size
;
354 vc4_bo_cache_free_old(dev
);
357 mutex_unlock(&vc4
->bo_lock
);
360 static void vc4_bo_cache_time_work(struct work_struct
*work
)
362 struct vc4_dev
*vc4
=
363 container_of(work
, struct vc4_dev
, bo_cache
.time_work
);
364 struct drm_device
*dev
= vc4
->dev
;
366 mutex_lock(&vc4
->bo_lock
);
367 vc4_bo_cache_free_old(dev
);
368 mutex_unlock(&vc4
->bo_lock
);
371 static void vc4_bo_cache_time_timer(unsigned long data
)
373 struct drm_device
*dev
= (struct drm_device
*)data
;
374 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
376 schedule_work(&vc4
->bo_cache
.time_work
);
379 struct reservation_object
*vc4_prime_res_obj(struct drm_gem_object
*obj
)
381 struct vc4_bo
*bo
= to_vc4_bo(obj
);
387 vc4_prime_export(struct drm_device
*dev
, struct drm_gem_object
*obj
, int flags
)
389 struct vc4_bo
*bo
= to_vc4_bo(obj
);
391 if (bo
->validated_shader
) {
392 DRM_ERROR("Attempting to export shader BO\n");
393 return ERR_PTR(-EINVAL
);
396 return drm_gem_prime_export(dev
, obj
, flags
);
399 int vc4_mmap(struct file
*filp
, struct vm_area_struct
*vma
)
401 struct drm_gem_object
*gem_obj
;
405 ret
= drm_gem_mmap(filp
, vma
);
409 gem_obj
= vma
->vm_private_data
;
410 bo
= to_vc4_bo(gem_obj
);
412 if (bo
->validated_shader
&& (vma
->vm_flags
& VM_WRITE
)) {
413 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
418 * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the
419 * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map
422 vma
->vm_flags
&= ~VM_PFNMAP
;
425 ret
= dma_mmap_wc(bo
->base
.base
.dev
->dev
, vma
, bo
->base
.vaddr
,
426 bo
->base
.paddr
, vma
->vm_end
- vma
->vm_start
);
428 drm_gem_vm_close(vma
);
433 int vc4_prime_mmap(struct drm_gem_object
*obj
, struct vm_area_struct
*vma
)
435 struct vc4_bo
*bo
= to_vc4_bo(obj
);
437 if (bo
->validated_shader
&& (vma
->vm_flags
& VM_WRITE
)) {
438 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
442 return drm_gem_cma_prime_mmap(obj
, vma
);
445 void *vc4_prime_vmap(struct drm_gem_object
*obj
)
447 struct vc4_bo
*bo
= to_vc4_bo(obj
);
449 if (bo
->validated_shader
) {
450 DRM_ERROR("mmaping of shader BOs not allowed.\n");
451 return ERR_PTR(-EINVAL
);
454 return drm_gem_cma_prime_vmap(obj
);
457 struct drm_gem_object
*
458 vc4_prime_import_sg_table(struct drm_device
*dev
,
459 struct dma_buf_attachment
*attach
,
460 struct sg_table
*sgt
)
462 struct drm_gem_object
*obj
;
465 obj
= drm_gem_cma_prime_import_sg_table(dev
, attach
, sgt
);
470 bo
->resv
= attach
->dmabuf
->resv
;
475 int vc4_create_bo_ioctl(struct drm_device
*dev
, void *data
,
476 struct drm_file
*file_priv
)
478 struct drm_vc4_create_bo
*args
= data
;
479 struct vc4_bo
*bo
= NULL
;
483 * We can't allocate from the BO cache, because the BOs don't
484 * get zeroed, and that might leak data between users.
486 bo
= vc4_bo_create(dev
, args
->size
, false);
490 ret
= drm_gem_handle_create(file_priv
, &bo
->base
.base
, &args
->handle
);
491 drm_gem_object_unreference_unlocked(&bo
->base
.base
);
496 int vc4_mmap_bo_ioctl(struct drm_device
*dev
, void *data
,
497 struct drm_file
*file_priv
)
499 struct drm_vc4_mmap_bo
*args
= data
;
500 struct drm_gem_object
*gem_obj
;
502 gem_obj
= drm_gem_object_lookup(file_priv
, args
->handle
);
504 DRM_ERROR("Failed to look up GEM BO %d\n", args
->handle
);
508 /* The mmap offset was set up at BO allocation time. */
509 args
->offset
= drm_vma_node_offset_addr(&gem_obj
->vma_node
);
511 drm_gem_object_unreference_unlocked(gem_obj
);
516 vc4_create_shader_bo_ioctl(struct drm_device
*dev
, void *data
,
517 struct drm_file
*file_priv
)
519 struct drm_vc4_create_shader_bo
*args
= data
;
520 struct vc4_bo
*bo
= NULL
;
526 if (args
->size
% sizeof(u64
) != 0)
529 if (args
->flags
!= 0) {
530 DRM_INFO("Unknown flags set: 0x%08x\n", args
->flags
);
534 if (args
->pad
!= 0) {
535 DRM_INFO("Pad set: 0x%08x\n", args
->pad
);
539 bo
= vc4_bo_create(dev
, args
->size
, true);
543 if (copy_from_user(bo
->base
.vaddr
,
544 (void __user
*)(uintptr_t)args
->data
,
549 /* Clear the rest of the memory from allocating from the BO
552 memset(bo
->base
.vaddr
+ args
->size
, 0,
553 bo
->base
.base
.size
- args
->size
);
555 bo
->validated_shader
= vc4_validate_shader(&bo
->base
);
556 if (!bo
->validated_shader
) {
561 /* We have to create the handle after validation, to avoid
562 * races for users to do doing things like mmap the shader BO.
564 ret
= drm_gem_handle_create(file_priv
, &bo
->base
.base
, &args
->handle
);
567 drm_gem_object_unreference_unlocked(&bo
->base
.base
);
573 * vc4_set_tiling_ioctl() - Sets the tiling modifier for a BO.
575 * @data: ioctl argument
576 * @file_priv: DRM file for this fd
578 * The tiling state of the BO decides the default modifier of an fb if
579 * no specific modifier was set by userspace, and the return value of
580 * vc4_get_tiling_ioctl() (so that userspace can treat a BO it
581 * received from dmabuf as the same tiling format as the producer
584 int vc4_set_tiling_ioctl(struct drm_device
*dev
, void *data
,
585 struct drm_file
*file_priv
)
587 struct drm_vc4_set_tiling
*args
= data
;
588 struct drm_gem_object
*gem_obj
;
592 if (args
->flags
!= 0)
595 switch (args
->modifier
) {
596 case DRM_FORMAT_MOD_NONE
:
599 case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED
:
606 gem_obj
= drm_gem_object_lookup(file_priv
, args
->handle
);
608 DRM_ERROR("Failed to look up GEM BO %d\n", args
->handle
);
611 bo
= to_vc4_bo(gem_obj
);
612 bo
->t_format
= t_format
;
614 drm_gem_object_unreference_unlocked(gem_obj
);
620 * vc4_get_tiling_ioctl() - Gets the tiling modifier for a BO.
622 * @data: ioctl argument
623 * @file_priv: DRM file for this fd
625 * Returns the tiling modifier for a BO as set by vc4_set_tiling_ioctl().
627 int vc4_get_tiling_ioctl(struct drm_device
*dev
, void *data
,
628 struct drm_file
*file_priv
)
630 struct drm_vc4_get_tiling
*args
= data
;
631 struct drm_gem_object
*gem_obj
;
634 if (args
->flags
!= 0 || args
->modifier
!= 0)
637 gem_obj
= drm_gem_object_lookup(file_priv
, args
->handle
);
639 DRM_ERROR("Failed to look up GEM BO %d\n", args
->handle
);
642 bo
= to_vc4_bo(gem_obj
);
645 args
->modifier
= DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED
;
647 args
->modifier
= DRM_FORMAT_MOD_NONE
;
649 drm_gem_object_unreference_unlocked(gem_obj
);
654 void vc4_bo_cache_init(struct drm_device
*dev
)
656 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
658 mutex_init(&vc4
->bo_lock
);
660 INIT_LIST_HEAD(&vc4
->bo_cache
.time_list
);
662 INIT_WORK(&vc4
->bo_cache
.time_work
, vc4_bo_cache_time_work
);
663 setup_timer(&vc4
->bo_cache
.time_timer
,
664 vc4_bo_cache_time_timer
,
668 void vc4_bo_cache_destroy(struct drm_device
*dev
)
670 struct vc4_dev
*vc4
= to_vc4_dev(dev
);
672 del_timer(&vc4
->bo_cache
.time_timer
);
673 cancel_work_sync(&vc4
->bo_cache
.time_work
);
675 vc4_bo_cache_purge(dev
);
677 if (vc4
->bo_stats
.num_allocated
) {
678 DRM_ERROR("Destroying BO cache while BOs still allocated:\n");
679 vc4_bo_stats_dump(vc4
);