2 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
14 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
15 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
16 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
17 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
27 #include "radeon_drm_cs.h"
29 #include "util/u_hash_table.h"
30 #include "util/u_memory.h"
31 #include "util/simple_list.h"
32 #include "os/os_thread.h"
33 #include "os/os_mman.h"
34 #include "util/os_time.h"
36 #include "state_tracker/drm_driver.h"
38 #include <sys/ioctl.h>
45 static struct pb_buffer
*
46 radeon_winsys_bo_create(struct radeon_winsys
*rws
,
49 enum radeon_bo_domain domain
,
50 enum radeon_bo_flag flags
);
52 static inline struct radeon_bo
*radeon_bo(struct pb_buffer
*bo
)
54 return (struct radeon_bo
*)bo
;
57 struct radeon_bo_va_hole
{
58 struct list_head list
;
63 static bool radeon_real_bo_is_busy(struct radeon_bo
*bo
)
65 struct drm_radeon_gem_busy args
= {0};
67 args
.handle
= bo
->handle
;
68 return drmCommandWriteRead(bo
->rws
->fd
, DRM_RADEON_GEM_BUSY
,
69 &args
, sizeof(args
)) != 0;
72 static bool radeon_bo_is_busy(struct radeon_bo
*bo
)
78 return radeon_real_bo_is_busy(bo
);
80 mtx_lock(&bo
->rws
->bo_fence_lock
);
81 for (num_idle
= 0; num_idle
< bo
->u
.slab
.num_fences
; ++num_idle
) {
82 if (radeon_real_bo_is_busy(bo
->u
.slab
.fences
[num_idle
])) {
86 radeon_bo_reference(&bo
->u
.slab
.fences
[num_idle
], NULL
);
88 memmove(&bo
->u
.slab
.fences
[0], &bo
->u
.slab
.fences
[num_idle
],
89 (bo
->u
.slab
.num_fences
- num_idle
) * sizeof(bo
->u
.slab
.fences
[0]));
90 bo
->u
.slab
.num_fences
-= num_idle
;
91 mtx_unlock(&bo
->rws
->bo_fence_lock
);
96 static void radeon_real_bo_wait_idle(struct radeon_bo
*bo
)
98 struct drm_radeon_gem_wait_idle args
= {0};
100 args
.handle
= bo
->handle
;
101 while (drmCommandWrite(bo
->rws
->fd
, DRM_RADEON_GEM_WAIT_IDLE
,
102 &args
, sizeof(args
)) == -EBUSY
);
105 static void radeon_bo_wait_idle(struct radeon_bo
*bo
)
108 radeon_real_bo_wait_idle(bo
);
110 mtx_lock(&bo
->rws
->bo_fence_lock
);
111 while (bo
->u
.slab
.num_fences
) {
112 struct radeon_bo
*fence
= NULL
;
113 radeon_bo_reference(&fence
, bo
->u
.slab
.fences
[0]);
114 mtx_unlock(&bo
->rws
->bo_fence_lock
);
116 /* Wait without holding the fence lock. */
117 radeon_real_bo_wait_idle(fence
);
119 mtx_lock(&bo
->rws
->bo_fence_lock
);
120 if (bo
->u
.slab
.num_fences
&& fence
== bo
->u
.slab
.fences
[0]) {
121 radeon_bo_reference(&bo
->u
.slab
.fences
[0], NULL
);
122 memmove(&bo
->u
.slab
.fences
[0], &bo
->u
.slab
.fences
[1],
123 (bo
->u
.slab
.num_fences
- 1) * sizeof(bo
->u
.slab
.fences
[0]));
124 bo
->u
.slab
.num_fences
--;
126 radeon_bo_reference(&fence
, NULL
);
128 mtx_unlock(&bo
->rws
->bo_fence_lock
);
132 static bool radeon_bo_wait(struct pb_buffer
*_buf
, uint64_t timeout
,
133 enum radeon_bo_usage usage
)
135 struct radeon_bo
*bo
= radeon_bo(_buf
);
138 /* No timeout. Just query. */
140 return !bo
->num_active_ioctls
&& !radeon_bo_is_busy(bo
);
142 abs_timeout
= os_time_get_absolute_timeout(timeout
);
144 /* Wait if any ioctl is being submitted with this buffer. */
145 if (!os_wait_until_zero_abs_timeout(&bo
->num_active_ioctls
, abs_timeout
))
148 /* Infinite timeout. */
149 if (abs_timeout
== PIPE_TIMEOUT_INFINITE
) {
150 radeon_bo_wait_idle(bo
);
154 /* Other timeouts need to be emulated with a loop. */
155 while (radeon_bo_is_busy(bo
)) {
156 if (os_time_get_nano() >= abs_timeout
)
164 static enum radeon_bo_domain
get_valid_domain(enum radeon_bo_domain domain
)
166 /* Zero domains the driver doesn't understand. */
167 domain
&= RADEON_DOMAIN_VRAM_GTT
;
169 /* If no domain is set, we must set something... */
171 domain
= RADEON_DOMAIN_VRAM_GTT
;
176 static enum radeon_bo_domain
radeon_bo_get_initial_domain(
177 struct pb_buffer
*buf
)
179 struct radeon_bo
*bo
= (struct radeon_bo
*)buf
;
180 struct drm_radeon_gem_op args
;
182 if (bo
->rws
->info
.drm_minor
< 38)
183 return RADEON_DOMAIN_VRAM_GTT
;
185 memset(&args
, 0, sizeof(args
));
186 args
.handle
= bo
->handle
;
187 args
.op
= RADEON_GEM_OP_GET_INITIAL_DOMAIN
;
189 if (drmCommandWriteRead(bo
->rws
->fd
, DRM_RADEON_GEM_OP
,
190 &args
, sizeof(args
))) {
191 fprintf(stderr
, "radeon: failed to get initial domain: %p 0x%08X\n",
193 /* Default domain as returned by get_valid_domain. */
194 return RADEON_DOMAIN_VRAM_GTT
;
197 /* GEM domains and winsys domains are defined the same. */
198 return get_valid_domain(args
.value
);
201 static uint64_t radeon_bomgr_find_va(const struct radeon_info
*info
,
202 struct radeon_vm_heap
*heap
,
203 uint64_t size
, uint64_t alignment
)
205 struct radeon_bo_va_hole
*hole
, *n
;
206 uint64_t offset
= 0, waste
= 0;
208 /* All VM address space holes will implicitly start aligned to the
209 * size alignment, so we don't need to sanitize the alignment here
211 size
= align(size
, info
->gart_page_size
);
213 mtx_lock(&heap
->mutex
);
214 /* first look for a hole */
215 LIST_FOR_EACH_ENTRY_SAFE(hole
, n
, &heap
->holes
, list
) {
216 offset
= hole
->offset
;
217 waste
= offset
% alignment
;
218 waste
= waste
? alignment
- waste
: 0;
220 if (offset
>= (hole
->offset
+ hole
->size
)) {
223 if (!waste
&& hole
->size
== size
) {
224 offset
= hole
->offset
;
225 list_del(&hole
->list
);
227 mtx_unlock(&heap
->mutex
);
230 if ((hole
->size
- waste
) > size
) {
232 n
= CALLOC_STRUCT(radeon_bo_va_hole
);
234 n
->offset
= hole
->offset
;
235 list_add(&n
->list
, &hole
->list
);
237 hole
->size
-= (size
+ waste
);
238 hole
->offset
+= size
+ waste
;
239 mtx_unlock(&heap
->mutex
);
242 if ((hole
->size
- waste
) == size
) {
244 mtx_unlock(&heap
->mutex
);
249 offset
= heap
->start
;
250 waste
= offset
% alignment
;
251 waste
= waste
? alignment
- waste
: 0;
253 if (offset
+ waste
+ size
> heap
->end
) {
254 mtx_unlock(&heap
->mutex
);
259 n
= CALLOC_STRUCT(radeon_bo_va_hole
);
262 list_add(&n
->list
, &heap
->holes
);
265 heap
->start
+= size
+ waste
;
266 mtx_unlock(&heap
->mutex
);
270 static uint64_t radeon_bomgr_find_va64(struct radeon_drm_winsys
*ws
,
271 uint64_t size
, uint64_t alignment
)
275 /* Try to allocate from the 64-bit address space first.
276 * If it doesn't exist (start = 0) or if it doesn't have enough space,
277 * fall back to the 32-bit address space.
280 va
= radeon_bomgr_find_va(&ws
->info
, &ws
->vm64
, size
, alignment
);
282 va
= radeon_bomgr_find_va(&ws
->info
, &ws
->vm32
, size
, alignment
);
286 static void radeon_bomgr_free_va(const struct radeon_info
*info
,
287 struct radeon_vm_heap
*heap
,
288 uint64_t va
, uint64_t size
)
290 struct radeon_bo_va_hole
*hole
= NULL
;
292 size
= align(size
, info
->gart_page_size
);
294 mtx_lock(&heap
->mutex
);
295 if ((va
+ size
) == heap
->start
) {
297 /* Delete uppermost hole if it reaches the new top */
298 if (!LIST_IS_EMPTY(&heap
->holes
)) {
299 hole
= container_of(heap
->holes
.next
, hole
, list
);
300 if ((hole
->offset
+ hole
->size
) == va
) {
301 heap
->start
= hole
->offset
;
302 list_del(&hole
->list
);
307 struct radeon_bo_va_hole
*next
;
309 hole
= container_of(&heap
->holes
, hole
, list
);
310 LIST_FOR_EACH_ENTRY(next
, &heap
->holes
, list
) {
311 if (next
->offset
< va
)
316 if (&hole
->list
!= &heap
->holes
) {
317 /* Grow upper hole if it's adjacent */
318 if (hole
->offset
== (va
+ size
)) {
321 /* Merge lower hole if it's adjacent */
322 if (next
!= hole
&& &next
->list
!= &heap
->holes
&&
323 (next
->offset
+ next
->size
) == va
) {
324 next
->size
+= hole
->size
;
325 list_del(&hole
->list
);
332 /* Grow lower hole if it's adjacent */
333 if (next
!= hole
&& &next
->list
!= &heap
->holes
&&
334 (next
->offset
+ next
->size
) == va
) {
339 /* FIXME on allocation failure we just lose virtual address space
340 * maybe print a warning
342 next
= CALLOC_STRUCT(radeon_bo_va_hole
);
346 list_add(&next
->list
, &hole
->list
);
350 mtx_unlock(&heap
->mutex
);
353 void radeon_bo_destroy(struct pb_buffer
*_buf
)
355 struct radeon_bo
*bo
= radeon_bo(_buf
);
356 struct radeon_drm_winsys
*rws
= bo
->rws
;
357 struct drm_gem_close args
;
359 assert(bo
->handle
&& "must not be called for slab entries");
361 memset(&args
, 0, sizeof(args
));
363 mtx_lock(&rws
->bo_handles_mutex
);
364 util_hash_table_remove(rws
->bo_handles
, (void*)(uintptr_t)bo
->handle
);
365 if (bo
->flink_name
) {
366 util_hash_table_remove(rws
->bo_names
,
367 (void*)(uintptr_t)bo
->flink_name
);
369 mtx_unlock(&rws
->bo_handles_mutex
);
372 os_munmap(bo
->u
.real
.ptr
, bo
->base
.size
);
374 if (rws
->info
.has_virtual_memory
) {
375 if (rws
->va_unmap_working
) {
376 struct drm_radeon_gem_va va
;
378 va
.handle
= bo
->handle
;
380 va
.operation
= RADEON_VA_UNMAP
;
381 va
.flags
= RADEON_VM_PAGE_READABLE
|
382 RADEON_VM_PAGE_WRITEABLE
|
383 RADEON_VM_PAGE_SNOOPED
;
386 if (drmCommandWriteRead(rws
->fd
, DRM_RADEON_GEM_VA
, &va
,
388 va
.operation
== RADEON_VA_RESULT_ERROR
) {
389 fprintf(stderr
, "radeon: Failed to deallocate virtual address for buffer:\n");
390 fprintf(stderr
, "radeon: size : %"PRIu64
" bytes\n", bo
->base
.size
);
391 fprintf(stderr
, "radeon: va : 0x%"PRIx64
"\n", bo
->va
);
395 radeon_bomgr_free_va(&rws
->info
,
396 bo
->va
< rws
->vm32
.end
? &rws
->vm32
: &rws
->vm64
,
397 bo
->va
, bo
->base
.size
);
401 args
.handle
= bo
->handle
;
402 drmIoctl(rws
->fd
, DRM_IOCTL_GEM_CLOSE
, &args
);
404 mtx_destroy(&bo
->u
.real
.map_mutex
);
406 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
407 rws
->allocated_vram
-= align(bo
->base
.size
, rws
->info
.gart_page_size
);
408 else if (bo
->initial_domain
& RADEON_DOMAIN_GTT
)
409 rws
->allocated_gtt
-= align(bo
->base
.size
, rws
->info
.gart_page_size
);
411 if (bo
->u
.real
.map_count
>= 1) {
412 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
413 bo
->rws
->mapped_vram
-= bo
->base
.size
;
415 bo
->rws
->mapped_gtt
-= bo
->base
.size
;
416 bo
->rws
->num_mapped_buffers
--;
422 static void radeon_bo_destroy_or_cache(struct pb_buffer
*_buf
)
424 struct radeon_bo
*bo
= radeon_bo(_buf
);
426 assert(bo
->handle
&& "must not be called for slab entries");
428 if (bo
->u
.real
.use_reusable_pool
)
429 pb_cache_add_buffer(&bo
->u
.real
.cache_entry
);
431 radeon_bo_destroy(_buf
);
434 void *radeon_bo_do_map(struct radeon_bo
*bo
)
436 struct drm_radeon_gem_mmap args
= {0};
440 /* If the buffer is created from user memory, return the user pointer. */
447 offset
= bo
->va
- bo
->u
.slab
.real
->va
;
448 bo
= bo
->u
.slab
.real
;
451 /* Map the buffer. */
452 mtx_lock(&bo
->u
.real
.map_mutex
);
453 /* Return the pointer if it's already mapped. */
454 if (bo
->u
.real
.ptr
) {
455 bo
->u
.real
.map_count
++;
456 mtx_unlock(&bo
->u
.real
.map_mutex
);
457 return (uint8_t*)bo
->u
.real
.ptr
+ offset
;
459 args
.handle
= bo
->handle
;
461 args
.size
= (uint64_t)bo
->base
.size
;
462 if (drmCommandWriteRead(bo
->rws
->fd
,
466 mtx_unlock(&bo
->u
.real
.map_mutex
);
467 fprintf(stderr
, "radeon: gem_mmap failed: %p 0x%08X\n",
472 ptr
= os_mmap(0, args
.size
, PROT_READ
|PROT_WRITE
, MAP_SHARED
,
473 bo
->rws
->fd
, args
.addr_ptr
);
474 if (ptr
== MAP_FAILED
) {
475 /* Clear the cache and try again. */
476 pb_cache_release_all_buffers(&bo
->rws
->bo_cache
);
478 ptr
= os_mmap(0, args
.size
, PROT_READ
|PROT_WRITE
, MAP_SHARED
,
479 bo
->rws
->fd
, args
.addr_ptr
);
480 if (ptr
== MAP_FAILED
) {
481 mtx_unlock(&bo
->u
.real
.map_mutex
);
482 fprintf(stderr
, "radeon: mmap failed, errno: %i\n", errno
);
486 bo
->u
.real
.ptr
= ptr
;
487 bo
->u
.real
.map_count
= 1;
489 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
490 bo
->rws
->mapped_vram
+= bo
->base
.size
;
492 bo
->rws
->mapped_gtt
+= bo
->base
.size
;
493 bo
->rws
->num_mapped_buffers
++;
495 mtx_unlock(&bo
->u
.real
.map_mutex
);
496 return (uint8_t*)bo
->u
.real
.ptr
+ offset
;
499 static void *radeon_bo_map(struct pb_buffer
*buf
,
500 struct radeon_winsys_cs
*rcs
,
501 enum pipe_transfer_usage usage
)
503 struct radeon_bo
*bo
= (struct radeon_bo
*)buf
;
504 struct radeon_drm_cs
*cs
= (struct radeon_drm_cs
*)rcs
;
506 /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */
507 if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
508 /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */
509 if (usage
& PIPE_TRANSFER_DONTBLOCK
) {
510 if (!(usage
& PIPE_TRANSFER_WRITE
)) {
513 * Since we are mapping for read, we don't need to wait
514 * if the GPU is using the buffer for read too
515 * (neither one is changing it).
517 * Only check whether the buffer is being used for write. */
518 if (cs
&& radeon_bo_is_referenced_by_cs_for_write(cs
, bo
)) {
519 cs
->flush_cs(cs
->flush_data
, PIPE_FLUSH_ASYNC
, NULL
);
523 if (!radeon_bo_wait((struct pb_buffer
*)bo
, 0,
524 RADEON_USAGE_WRITE
)) {
528 if (cs
&& radeon_bo_is_referenced_by_cs(cs
, bo
)) {
529 cs
->flush_cs(cs
->flush_data
, PIPE_FLUSH_ASYNC
, NULL
);
533 if (!radeon_bo_wait((struct pb_buffer
*)bo
, 0,
534 RADEON_USAGE_READWRITE
)) {
539 uint64_t time
= os_time_get_nano();
541 if (!(usage
& PIPE_TRANSFER_WRITE
)) {
544 * Since we are mapping for read, we don't need to wait
545 * if the GPU is using the buffer for read too
546 * (neither one is changing it).
548 * Only check whether the buffer is being used for write. */
549 if (cs
&& radeon_bo_is_referenced_by_cs_for_write(cs
, bo
)) {
550 cs
->flush_cs(cs
->flush_data
, 0, NULL
);
552 radeon_bo_wait((struct pb_buffer
*)bo
, PIPE_TIMEOUT_INFINITE
,
555 /* Mapping for write. */
557 if (radeon_bo_is_referenced_by_cs(cs
, bo
)) {
558 cs
->flush_cs(cs
->flush_data
, 0, NULL
);
560 /* Try to avoid busy-waiting in radeon_bo_wait. */
561 if (p_atomic_read(&bo
->num_active_ioctls
))
562 radeon_drm_cs_sync_flush(rcs
);
566 radeon_bo_wait((struct pb_buffer
*)bo
, PIPE_TIMEOUT_INFINITE
,
567 RADEON_USAGE_READWRITE
);
570 bo
->rws
->buffer_wait_time
+= os_time_get_nano() - time
;
574 return radeon_bo_do_map(bo
);
577 static void radeon_bo_unmap(struct pb_buffer
*_buf
)
579 struct radeon_bo
*bo
= (struct radeon_bo
*)_buf
;
585 bo
= bo
->u
.slab
.real
;
587 mtx_lock(&bo
->u
.real
.map_mutex
);
588 if (!bo
->u
.real
.ptr
) {
589 mtx_unlock(&bo
->u
.real
.map_mutex
);
590 return; /* it's not been mapped */
593 assert(bo
->u
.real
.map_count
);
594 if (--bo
->u
.real
.map_count
) {
595 mtx_unlock(&bo
->u
.real
.map_mutex
);
596 return; /* it's been mapped multiple times */
599 os_munmap(bo
->u
.real
.ptr
, bo
->base
.size
);
600 bo
->u
.real
.ptr
= NULL
;
602 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
603 bo
->rws
->mapped_vram
-= bo
->base
.size
;
605 bo
->rws
->mapped_gtt
-= bo
->base
.size
;
606 bo
->rws
->num_mapped_buffers
--;
608 mtx_unlock(&bo
->u
.real
.map_mutex
);
611 static const struct pb_vtbl radeon_bo_vtbl
= {
612 radeon_bo_destroy_or_cache
613 /* other functions are never called */
616 static struct radeon_bo
*radeon_create_bo(struct radeon_drm_winsys
*rws
,
617 unsigned size
, unsigned alignment
,
618 unsigned initial_domains
,
622 struct radeon_bo
*bo
;
623 struct drm_radeon_gem_create args
;
626 memset(&args
, 0, sizeof(args
));
628 assert(initial_domains
);
629 assert((initial_domains
&
630 ~(RADEON_GEM_DOMAIN_GTT
| RADEON_GEM_DOMAIN_VRAM
)) == 0);
633 args
.alignment
= alignment
;
634 args
.initial_domain
= initial_domains
;
637 /* If VRAM is just stolen system memory, allow both VRAM and
638 * GTT, whichever has free space. If a buffer is evicted from
639 * VRAM to GTT, it will stay there.
641 if (!rws
->info
.has_dedicated_vram
)
642 args
.initial_domain
|= RADEON_DOMAIN_GTT
;
644 if (flags
& RADEON_FLAG_GTT_WC
)
645 args
.flags
|= RADEON_GEM_GTT_WC
;
646 if (flags
& RADEON_FLAG_NO_CPU_ACCESS
)
647 args
.flags
|= RADEON_GEM_NO_CPU_ACCESS
;
649 if (drmCommandWriteRead(rws
->fd
, DRM_RADEON_GEM_CREATE
,
650 &args
, sizeof(args
))) {
651 fprintf(stderr
, "radeon: Failed to allocate a buffer:\n");
652 fprintf(stderr
, "radeon: size : %u bytes\n", size
);
653 fprintf(stderr
, "radeon: alignment : %u bytes\n", alignment
);
654 fprintf(stderr
, "radeon: domains : %u\n", args
.initial_domain
);
655 fprintf(stderr
, "radeon: flags : %u\n", args
.flags
);
659 assert(args
.handle
!= 0);
661 bo
= CALLOC_STRUCT(radeon_bo
);
665 pipe_reference_init(&bo
->base
.reference
, 1);
666 bo
->base
.alignment
= alignment
;
668 bo
->base
.size
= size
;
669 bo
->base
.vtbl
= &radeon_bo_vtbl
;
671 bo
->handle
= args
.handle
;
673 bo
->initial_domain
= initial_domains
;
674 bo
->hash
= __sync_fetch_and_add(&rws
->next_bo_hash
, 1);
675 (void) mtx_init(&bo
->u
.real
.map_mutex
, mtx_plain
);
678 pb_cache_init_entry(&rws
->bo_cache
, &bo
->u
.real
.cache_entry
, &bo
->base
,
682 if (rws
->info
.has_virtual_memory
) {
683 struct drm_radeon_gem_va va
;
684 unsigned va_gap_size
;
686 va_gap_size
= rws
->check_vm
? MAX2(4 * alignment
, 64 * 1024) : 0;
688 if (flags
& RADEON_FLAG_32BIT
) {
689 bo
->va
= radeon_bomgr_find_va(&rws
->info
, &rws
->vm32
,
690 size
+ va_gap_size
, alignment
);
691 assert(bo
->va
+ size
< rws
->vm32
.end
);
693 bo
->va
= radeon_bomgr_find_va64(rws
, size
+ va_gap_size
, alignment
);
696 va
.handle
= bo
->handle
;
698 va
.operation
= RADEON_VA_MAP
;
699 va
.flags
= RADEON_VM_PAGE_READABLE
|
700 RADEON_VM_PAGE_WRITEABLE
|
701 RADEON_VM_PAGE_SNOOPED
;
703 r
= drmCommandWriteRead(rws
->fd
, DRM_RADEON_GEM_VA
, &va
, sizeof(va
));
704 if (r
&& va
.operation
== RADEON_VA_RESULT_ERROR
) {
705 fprintf(stderr
, "radeon: Failed to allocate virtual address for buffer:\n");
706 fprintf(stderr
, "radeon: size : %d bytes\n", size
);
707 fprintf(stderr
, "radeon: alignment : %d bytes\n", alignment
);
708 fprintf(stderr
, "radeon: domains : %d\n", args
.initial_domain
);
709 fprintf(stderr
, "radeon: va : 0x%016llx\n", (unsigned long long)bo
->va
);
710 radeon_bo_destroy(&bo
->base
);
713 mtx_lock(&rws
->bo_handles_mutex
);
714 if (va
.operation
== RADEON_VA_RESULT_VA_EXIST
) {
715 struct pb_buffer
*b
= &bo
->base
;
716 struct radeon_bo
*old_bo
=
717 util_hash_table_get(rws
->bo_vas
, (void*)(uintptr_t)va
.offset
);
719 mtx_unlock(&rws
->bo_handles_mutex
);
720 pb_reference(&b
, &old_bo
->base
);
724 util_hash_table_set(rws
->bo_vas
, (void*)(uintptr_t)bo
->va
, bo
);
725 mtx_unlock(&rws
->bo_handles_mutex
);
728 if (initial_domains
& RADEON_DOMAIN_VRAM
)
729 rws
->allocated_vram
+= align(size
, rws
->info
.gart_page_size
);
730 else if (initial_domains
& RADEON_DOMAIN_GTT
)
731 rws
->allocated_gtt
+= align(size
, rws
->info
.gart_page_size
);
736 bool radeon_bo_can_reclaim(struct pb_buffer
*_buf
)
738 struct radeon_bo
*bo
= radeon_bo(_buf
);
740 if (radeon_bo_is_referenced_by_any_cs(bo
))
743 return radeon_bo_wait(_buf
, 0, RADEON_USAGE_READWRITE
);
746 bool radeon_bo_can_reclaim_slab(void *priv
, struct pb_slab_entry
*entry
)
748 struct radeon_bo
*bo
= NULL
; /* fix container_of */
749 bo
= container_of(entry
, bo
, u
.slab
.entry
);
751 return radeon_bo_can_reclaim(&bo
->base
);
754 static void radeon_bo_slab_destroy(struct pb_buffer
*_buf
)
756 struct radeon_bo
*bo
= radeon_bo(_buf
);
760 pb_slab_free(&bo
->rws
->bo_slabs
, &bo
->u
.slab
.entry
);
763 static const struct pb_vtbl radeon_winsys_bo_slab_vtbl
= {
764 radeon_bo_slab_destroy
765 /* other functions are never called */
768 struct pb_slab
*radeon_bo_slab_alloc(void *priv
, unsigned heap
,
770 unsigned group_index
)
772 struct radeon_drm_winsys
*ws
= priv
;
773 struct radeon_slab
*slab
= CALLOC_STRUCT(radeon_slab
);
774 enum radeon_bo_domain domains
= radeon_domain_from_heap(heap
);
775 enum radeon_bo_flag flags
= radeon_flags_from_heap(heap
);
781 slab
->buffer
= radeon_bo(radeon_winsys_bo_create(&ws
->base
,
782 64 * 1024, 64 * 1024,
787 assert(slab
->buffer
->handle
);
789 slab
->base
.num_entries
= slab
->buffer
->base
.size
/ entry_size
;
790 slab
->base
.num_free
= slab
->base
.num_entries
;
791 slab
->entries
= CALLOC(slab
->base
.num_entries
, sizeof(*slab
->entries
));
795 LIST_INITHEAD(&slab
->base
.free
);
797 base_hash
= __sync_fetch_and_add(&ws
->next_bo_hash
, slab
->base
.num_entries
);
799 for (unsigned i
= 0; i
< slab
->base
.num_entries
; ++i
) {
800 struct radeon_bo
*bo
= &slab
->entries
[i
];
802 bo
->base
.alignment
= entry_size
;
803 bo
->base
.usage
= slab
->buffer
->base
.usage
;
804 bo
->base
.size
= entry_size
;
805 bo
->base
.vtbl
= &radeon_winsys_bo_slab_vtbl
;
807 bo
->va
= slab
->buffer
->va
+ i
* entry_size
;
808 bo
->initial_domain
= domains
;
809 bo
->hash
= base_hash
+ i
;
810 bo
->u
.slab
.entry
.slab
= &slab
->base
;
811 bo
->u
.slab
.entry
.group_index
= group_index
;
812 bo
->u
.slab
.real
= slab
->buffer
;
814 LIST_ADDTAIL(&bo
->u
.slab
.entry
.head
, &slab
->base
.free
);
820 radeon_bo_reference(&slab
->buffer
, NULL
);
826 void radeon_bo_slab_free(void *priv
, struct pb_slab
*pslab
)
828 struct radeon_slab
*slab
= (struct radeon_slab
*)pslab
;
830 for (unsigned i
= 0; i
< slab
->base
.num_entries
; ++i
) {
831 struct radeon_bo
*bo
= &slab
->entries
[i
];
832 for (unsigned j
= 0; j
< bo
->u
.slab
.num_fences
; ++j
)
833 radeon_bo_reference(&bo
->u
.slab
.fences
[j
], NULL
);
834 FREE(bo
->u
.slab
.fences
);
838 radeon_bo_reference(&slab
->buffer
, NULL
);
842 static unsigned eg_tile_split(unsigned tile_split
)
844 switch (tile_split
) {
845 case 0: tile_split
= 64; break;
846 case 1: tile_split
= 128; break;
847 case 2: tile_split
= 256; break;
848 case 3: tile_split
= 512; break;
850 case 4: tile_split
= 1024; break;
851 case 5: tile_split
= 2048; break;
852 case 6: tile_split
= 4096; break;
857 static unsigned eg_tile_split_rev(unsigned eg_tile_split
)
859 switch (eg_tile_split
) {
871 static void radeon_bo_get_metadata(struct pb_buffer
*_buf
,
872 struct radeon_bo_metadata
*md
)
874 struct radeon_bo
*bo
= radeon_bo(_buf
);
875 struct drm_radeon_gem_set_tiling args
;
877 assert(bo
->handle
&& "must not be called for slab entries");
879 memset(&args
, 0, sizeof(args
));
881 args
.handle
= bo
->handle
;
883 drmCommandWriteRead(bo
->rws
->fd
,
884 DRM_RADEON_GEM_GET_TILING
,
888 md
->u
.legacy
.microtile
= RADEON_LAYOUT_LINEAR
;
889 md
->u
.legacy
.macrotile
= RADEON_LAYOUT_LINEAR
;
890 if (args
.tiling_flags
& RADEON_TILING_MICRO
)
891 md
->u
.legacy
.microtile
= RADEON_LAYOUT_TILED
;
892 else if (args
.tiling_flags
& RADEON_TILING_MICRO_SQUARE
)
893 md
->u
.legacy
.microtile
= RADEON_LAYOUT_SQUARETILED
;
895 if (args
.tiling_flags
& RADEON_TILING_MACRO
)
896 md
->u
.legacy
.macrotile
= RADEON_LAYOUT_TILED
;
898 md
->u
.legacy
.bankw
= (args
.tiling_flags
>> RADEON_TILING_EG_BANKW_SHIFT
) & RADEON_TILING_EG_BANKW_MASK
;
899 md
->u
.legacy
.bankh
= (args
.tiling_flags
>> RADEON_TILING_EG_BANKH_SHIFT
) & RADEON_TILING_EG_BANKH_MASK
;
900 md
->u
.legacy
.tile_split
= (args
.tiling_flags
>> RADEON_TILING_EG_TILE_SPLIT_SHIFT
) & RADEON_TILING_EG_TILE_SPLIT_MASK
;
901 md
->u
.legacy
.mtilea
= (args
.tiling_flags
>> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT
) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK
;
902 md
->u
.legacy
.tile_split
= eg_tile_split(md
->u
.legacy
.tile_split
);
903 md
->u
.legacy
.scanout
= bo
->rws
->gen
>= DRV_SI
&& !(args
.tiling_flags
& RADEON_TILING_R600_NO_SCANOUT
);
906 static void radeon_bo_set_metadata(struct pb_buffer
*_buf
,
907 struct radeon_bo_metadata
*md
)
909 struct radeon_bo
*bo
= radeon_bo(_buf
);
910 struct drm_radeon_gem_set_tiling args
;
912 assert(bo
->handle
&& "must not be called for slab entries");
914 memset(&args
, 0, sizeof(args
));
916 os_wait_until_zero(&bo
->num_active_ioctls
, PIPE_TIMEOUT_INFINITE
);
918 if (md
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
919 args
.tiling_flags
|= RADEON_TILING_MICRO
;
920 else if (md
->u
.legacy
.microtile
== RADEON_LAYOUT_SQUARETILED
)
921 args
.tiling_flags
|= RADEON_TILING_MICRO_SQUARE
;
923 if (md
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
924 args
.tiling_flags
|= RADEON_TILING_MACRO
;
926 args
.tiling_flags
|= (md
->u
.legacy
.bankw
& RADEON_TILING_EG_BANKW_MASK
) <<
927 RADEON_TILING_EG_BANKW_SHIFT
;
928 args
.tiling_flags
|= (md
->u
.legacy
.bankh
& RADEON_TILING_EG_BANKH_MASK
) <<
929 RADEON_TILING_EG_BANKH_SHIFT
;
930 if (md
->u
.legacy
.tile_split
) {
931 args
.tiling_flags
|= (eg_tile_split_rev(md
->u
.legacy
.tile_split
) &
932 RADEON_TILING_EG_TILE_SPLIT_MASK
) <<
933 RADEON_TILING_EG_TILE_SPLIT_SHIFT
;
935 args
.tiling_flags
|= (md
->u
.legacy
.mtilea
& RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK
) <<
936 RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT
;
938 if (bo
->rws
->gen
>= DRV_SI
&& !md
->u
.legacy
.scanout
)
939 args
.tiling_flags
|= RADEON_TILING_R600_NO_SCANOUT
;
941 args
.handle
= bo
->handle
;
942 args
.pitch
= md
->u
.legacy
.stride
;
944 drmCommandWriteRead(bo
->rws
->fd
,
945 DRM_RADEON_GEM_SET_TILING
,
950 static struct pb_buffer
*
951 radeon_winsys_bo_create(struct radeon_winsys
*rws
,
954 enum radeon_bo_domain domain
,
955 enum radeon_bo_flag flags
)
957 struct radeon_drm_winsys
*ws
= radeon_drm_winsys(rws
);
958 struct radeon_bo
*bo
;
961 assert(!(flags
& RADEON_FLAG_SPARSE
)); /* not supported */
963 /* Only 32-bit sizes are supported. */
967 /* VRAM implies WC. This is not optional. */
968 if (domain
& RADEON_DOMAIN_VRAM
)
969 flags
|= RADEON_FLAG_GTT_WC
;
970 /* NO_CPU_ACCESS is valid with VRAM only. */
971 if (domain
!= RADEON_DOMAIN_VRAM
)
972 flags
&= ~RADEON_FLAG_NO_CPU_ACCESS
;
974 /* Sub-allocate small buffers from slabs. */
975 if (!(flags
& RADEON_FLAG_NO_SUBALLOC
) &&
976 size
<= (1 << RADEON_SLAB_MAX_SIZE_LOG2
) &&
977 ws
->info
.has_virtual_memory
&&
978 alignment
<= MAX2(1 << RADEON_SLAB_MIN_SIZE_LOG2
, util_next_power_of_two(size
))) {
979 struct pb_slab_entry
*entry
;
980 int heap
= radeon_get_heap_index(domain
, flags
);
982 if (heap
< 0 || heap
>= RADEON_MAX_SLAB_HEAPS
)
985 entry
= pb_slab_alloc(&ws
->bo_slabs
, size
, heap
);
987 /* Clear the cache and try again. */
988 pb_cache_release_all_buffers(&ws
->bo_cache
);
990 entry
= pb_slab_alloc(&ws
->bo_slabs
, size
, heap
);
996 bo
= container_of(entry
, bo
, u
.slab
.entry
);
998 pipe_reference_init(&bo
->base
.reference
, 1);
1004 /* This flag is irrelevant for the cache. */
1005 flags
&= ~RADEON_FLAG_NO_SUBALLOC
;
1007 /* Align size to page size. This is the minimum alignment for normal
1008 * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
1009 * like constant/uniform buffers, can benefit from better and more reuse.
1011 size
= align(size
, ws
->info
.gart_page_size
);
1012 alignment
= align(alignment
, ws
->info
.gart_page_size
);
1014 bool use_reusable_pool
= flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
;
1016 /* Shared resources don't use cached heaps. */
1017 if (use_reusable_pool
) {
1018 heap
= radeon_get_heap_index(domain
, flags
);
1019 assert(heap
>= 0 && heap
< RADEON_MAX_CACHED_HEAPS
);
1021 bo
= radeon_bo(pb_cache_reclaim_buffer(&ws
->bo_cache
, size
, alignment
,
1027 bo
= radeon_create_bo(ws
, size
, alignment
, domain
, flags
, heap
);
1029 /* Clear the cache and try again. */
1030 if (ws
->info
.has_virtual_memory
)
1031 pb_slabs_reclaim(&ws
->bo_slabs
);
1032 pb_cache_release_all_buffers(&ws
->bo_cache
);
1033 bo
= radeon_create_bo(ws
, size
, alignment
, domain
, flags
, heap
);
1038 bo
->u
.real
.use_reusable_pool
= use_reusable_pool
;
1040 mtx_lock(&ws
->bo_handles_mutex
);
1041 util_hash_table_set(ws
->bo_handles
, (void*)(uintptr_t)bo
->handle
, bo
);
1042 mtx_unlock(&ws
->bo_handles_mutex
);
1047 static struct pb_buffer
*radeon_winsys_bo_from_ptr(struct radeon_winsys
*rws
,
1048 void *pointer
, uint64_t size
)
1050 struct radeon_drm_winsys
*ws
= radeon_drm_winsys(rws
);
1051 struct drm_radeon_gem_userptr args
;
1052 struct radeon_bo
*bo
;
1055 bo
= CALLOC_STRUCT(radeon_bo
);
1059 memset(&args
, 0, sizeof(args
));
1060 args
.addr
= (uintptr_t)pointer
;
1061 args
.size
= align(size
, ws
->info
.gart_page_size
);
1062 args
.flags
= RADEON_GEM_USERPTR_ANONONLY
|
1063 RADEON_GEM_USERPTR_VALIDATE
|
1064 RADEON_GEM_USERPTR_REGISTER
;
1065 if (drmCommandWriteRead(ws
->fd
, DRM_RADEON_GEM_USERPTR
,
1066 &args
, sizeof(args
))) {
1071 assert(args
.handle
!= 0);
1073 mtx_lock(&ws
->bo_handles_mutex
);
1075 /* Initialize it. */
1076 pipe_reference_init(&bo
->base
.reference
, 1);
1077 bo
->handle
= args
.handle
;
1078 bo
->base
.alignment
= 0;
1079 bo
->base
.size
= size
;
1080 bo
->base
.vtbl
= &radeon_bo_vtbl
;
1082 bo
->user_ptr
= pointer
;
1084 bo
->initial_domain
= RADEON_DOMAIN_GTT
;
1085 bo
->hash
= __sync_fetch_and_add(&ws
->next_bo_hash
, 1);
1086 (void) mtx_init(&bo
->u
.real
.map_mutex
, mtx_plain
);
1088 util_hash_table_set(ws
->bo_handles
, (void*)(uintptr_t)bo
->handle
, bo
);
1090 mtx_unlock(&ws
->bo_handles_mutex
);
1092 if (ws
->info
.has_virtual_memory
) {
1093 struct drm_radeon_gem_va va
;
1095 bo
->va
= radeon_bomgr_find_va64(ws
, bo
->base
.size
, 1 << 20);
1097 va
.handle
= bo
->handle
;
1098 va
.operation
= RADEON_VA_MAP
;
1101 va
.flags
= RADEON_VM_PAGE_READABLE
|
1102 RADEON_VM_PAGE_WRITEABLE
|
1103 RADEON_VM_PAGE_SNOOPED
;
1105 r
= drmCommandWriteRead(ws
->fd
, DRM_RADEON_GEM_VA
, &va
, sizeof(va
));
1106 if (r
&& va
.operation
== RADEON_VA_RESULT_ERROR
) {
1107 fprintf(stderr
, "radeon: Failed to assign virtual address space\n");
1108 radeon_bo_destroy(&bo
->base
);
1111 mtx_lock(&ws
->bo_handles_mutex
);
1112 if (va
.operation
== RADEON_VA_RESULT_VA_EXIST
) {
1113 struct pb_buffer
*b
= &bo
->base
;
1114 struct radeon_bo
*old_bo
=
1115 util_hash_table_get(ws
->bo_vas
, (void*)(uintptr_t)va
.offset
);
1117 mtx_unlock(&ws
->bo_handles_mutex
);
1118 pb_reference(&b
, &old_bo
->base
);
1122 util_hash_table_set(ws
->bo_vas
, (void*)(uintptr_t)bo
->va
, bo
);
1123 mtx_unlock(&ws
->bo_handles_mutex
);
1126 ws
->allocated_gtt
+= align(bo
->base
.size
, ws
->info
.gart_page_size
);
1128 return (struct pb_buffer
*)bo
;
1131 static struct pb_buffer
*radeon_winsys_bo_from_handle(struct radeon_winsys
*rws
,
1132 struct winsys_handle
*whandle
,
1136 struct radeon_drm_winsys
*ws
= radeon_drm_winsys(rws
);
1137 struct radeon_bo
*bo
;
1142 if (!offset
&& whandle
->offset
!= 0) {
1143 fprintf(stderr
, "attempt to import unsupported winsys offset %u\n",
1148 /* We must maintain a list of pairs <handle, bo>, so that we always return
1149 * the same BO for one particular handle. If we didn't do that and created
1150 * more than one BO for the same handle and then relocated them in a CS,
1151 * we would hit a deadlock in the kernel.
1153 * The list of pairs is guarded by a mutex, of course. */
1154 mtx_lock(&ws
->bo_handles_mutex
);
1156 if (whandle
->type
== DRM_API_HANDLE_TYPE_SHARED
) {
1157 /* First check if there already is an existing bo for the handle. */
1158 bo
= util_hash_table_get(ws
->bo_names
, (void*)(uintptr_t)whandle
->handle
);
1159 } else if (whandle
->type
== DRM_API_HANDLE_TYPE_FD
) {
1160 /* We must first get the GEM handle, as fds are unreliable keys */
1161 r
= drmPrimeFDToHandle(ws
->fd
, whandle
->handle
, &handle
);
1164 bo
= util_hash_table_get(ws
->bo_handles
, (void*)(uintptr_t)handle
);
1166 /* Unknown handle type */
1171 /* Increase the refcount. */
1172 struct pb_buffer
*b
= NULL
;
1173 pb_reference(&b
, &bo
->base
);
1177 /* There isn't, create a new one. */
1178 bo
= CALLOC_STRUCT(radeon_bo
);
1183 if (whandle
->type
== DRM_API_HANDLE_TYPE_SHARED
) {
1184 struct drm_gem_open open_arg
= {};
1185 memset(&open_arg
, 0, sizeof(open_arg
));
1187 open_arg
.name
= whandle
->handle
;
1188 if (drmIoctl(ws
->fd
, DRM_IOCTL_GEM_OPEN
, &open_arg
)) {
1192 handle
= open_arg
.handle
;
1193 size
= open_arg
.size
;
1194 bo
->flink_name
= whandle
->handle
;
1195 } else if (whandle
->type
== DRM_API_HANDLE_TYPE_FD
) {
1196 size
= lseek(whandle
->handle
, 0, SEEK_END
);
1198 * Could check errno to determine whether the kernel is new enough, but
1199 * it doesn't really matter why this failed, just that it failed.
1201 if (size
== (off_t
)-1) {
1205 lseek(whandle
->handle
, 0, SEEK_SET
);
1208 assert(handle
!= 0);
1210 bo
->handle
= handle
;
1212 /* Initialize it. */
1213 pipe_reference_init(&bo
->base
.reference
, 1);
1214 bo
->base
.alignment
= 0;
1215 bo
->base
.size
= (unsigned) size
;
1216 bo
->base
.vtbl
= &radeon_bo_vtbl
;
1219 bo
->hash
= __sync_fetch_and_add(&ws
->next_bo_hash
, 1);
1220 (void) mtx_init(&bo
->u
.real
.map_mutex
, mtx_plain
);
1223 util_hash_table_set(ws
->bo_names
, (void*)(uintptr_t)bo
->flink_name
, bo
);
1225 util_hash_table_set(ws
->bo_handles
, (void*)(uintptr_t)bo
->handle
, bo
);
1228 mtx_unlock(&ws
->bo_handles_mutex
);
1231 *stride
= whandle
->stride
;
1233 *offset
= whandle
->offset
;
1235 if (ws
->info
.has_virtual_memory
&& !bo
->va
) {
1236 struct drm_radeon_gem_va va
;
1238 bo
->va
= radeon_bomgr_find_va64(ws
, bo
->base
.size
, 1 << 20);
1240 va
.handle
= bo
->handle
;
1241 va
.operation
= RADEON_VA_MAP
;
1244 va
.flags
= RADEON_VM_PAGE_READABLE
|
1245 RADEON_VM_PAGE_WRITEABLE
|
1246 RADEON_VM_PAGE_SNOOPED
;
1248 r
= drmCommandWriteRead(ws
->fd
, DRM_RADEON_GEM_VA
, &va
, sizeof(va
));
1249 if (r
&& va
.operation
== RADEON_VA_RESULT_ERROR
) {
1250 fprintf(stderr
, "radeon: Failed to assign virtual address space\n");
1251 radeon_bo_destroy(&bo
->base
);
1254 mtx_lock(&ws
->bo_handles_mutex
);
1255 if (va
.operation
== RADEON_VA_RESULT_VA_EXIST
) {
1256 struct pb_buffer
*b
= &bo
->base
;
1257 struct radeon_bo
*old_bo
=
1258 util_hash_table_get(ws
->bo_vas
, (void*)(uintptr_t)va
.offset
);
1260 mtx_unlock(&ws
->bo_handles_mutex
);
1261 pb_reference(&b
, &old_bo
->base
);
1265 util_hash_table_set(ws
->bo_vas
, (void*)(uintptr_t)bo
->va
, bo
);
1266 mtx_unlock(&ws
->bo_handles_mutex
);
1269 bo
->initial_domain
= radeon_bo_get_initial_domain((void*)bo
);
1271 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
1272 ws
->allocated_vram
+= align(bo
->base
.size
, ws
->info
.gart_page_size
);
1273 else if (bo
->initial_domain
& RADEON_DOMAIN_GTT
)
1274 ws
->allocated_gtt
+= align(bo
->base
.size
, ws
->info
.gart_page_size
);
1276 return (struct pb_buffer
*)bo
;
1279 mtx_unlock(&ws
->bo_handles_mutex
);
1283 static bool radeon_winsys_bo_get_handle(struct pb_buffer
*buffer
,
1284 unsigned stride
, unsigned offset
,
1285 unsigned slice_size
,
1286 struct winsys_handle
*whandle
)
1288 struct drm_gem_flink flink
;
1289 struct radeon_bo
*bo
= radeon_bo(buffer
);
1290 struct radeon_drm_winsys
*ws
= bo
->rws
;
1292 /* Don't allow exports of slab entries. */
1296 memset(&flink
, 0, sizeof(flink
));
1298 bo
->u
.real
.use_reusable_pool
= false;
1300 if (whandle
->type
== DRM_API_HANDLE_TYPE_SHARED
) {
1301 if (!bo
->flink_name
) {
1302 flink
.handle
= bo
->handle
;
1304 if (ioctl(ws
->fd
, DRM_IOCTL_GEM_FLINK
, &flink
)) {
1308 bo
->flink_name
= flink
.name
;
1310 mtx_lock(&ws
->bo_handles_mutex
);
1311 util_hash_table_set(ws
->bo_names
, (void*)(uintptr_t)bo
->flink_name
, bo
);
1312 mtx_unlock(&ws
->bo_handles_mutex
);
1314 whandle
->handle
= bo
->flink_name
;
1315 } else if (whandle
->type
== DRM_API_HANDLE_TYPE_KMS
) {
1316 whandle
->handle
= bo
->handle
;
1317 } else if (whandle
->type
== DRM_API_HANDLE_TYPE_FD
) {
1318 if (drmPrimeHandleToFD(ws
->fd
, bo
->handle
, DRM_CLOEXEC
, (int*)&whandle
->handle
))
1322 whandle
->stride
= stride
;
1323 whandle
->offset
= offset
;
1324 whandle
->offset
+= slice_size
* whandle
->layer
;
1329 static bool radeon_winsys_bo_is_user_ptr(struct pb_buffer
*buf
)
1331 return ((struct radeon_bo
*)buf
)->user_ptr
!= NULL
;
1334 static bool radeon_winsys_bo_is_suballocated(struct pb_buffer
*buf
)
1336 return !((struct radeon_bo
*)buf
)->handle
;
1339 static uint64_t radeon_winsys_bo_va(struct pb_buffer
*buf
)
1341 return ((struct radeon_bo
*)buf
)->va
;
1344 static unsigned radeon_winsys_bo_get_reloc_offset(struct pb_buffer
*buf
)
1346 struct radeon_bo
*bo
= radeon_bo(buf
);
1351 return bo
->va
- bo
->u
.slab
.real
->va
;
1354 void radeon_drm_bo_init_functions(struct radeon_drm_winsys
*ws
)
1356 ws
->base
.buffer_set_metadata
= radeon_bo_set_metadata
;
1357 ws
->base
.buffer_get_metadata
= radeon_bo_get_metadata
;
1358 ws
->base
.buffer_map
= radeon_bo_map
;
1359 ws
->base
.buffer_unmap
= radeon_bo_unmap
;
1360 ws
->base
.buffer_wait
= radeon_bo_wait
;
1361 ws
->base
.buffer_create
= radeon_winsys_bo_create
;
1362 ws
->base
.buffer_from_handle
= radeon_winsys_bo_from_handle
;
1363 ws
->base
.buffer_from_ptr
= radeon_winsys_bo_from_ptr
;
1364 ws
->base
.buffer_is_user_ptr
= radeon_winsys_bo_is_user_ptr
;
1365 ws
->base
.buffer_is_suballocated
= radeon_winsys_bo_is_suballocated
;
1366 ws
->base
.buffer_get_handle
= radeon_winsys_bo_get_handle
;
1367 ws
->base
.buffer_get_virtual_address
= radeon_winsys_bo_va
;
1368 ws
->base
.buffer_get_reloc_offset
= radeon_winsys_bo_get_reloc_offset
;
1369 ws
->base
.buffer_get_initial_domain
= radeon_bo_get_initial_domain
;