2 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
3 * Copyright © 2015 Advanced Micro Devices, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
16 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
18 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 * The above copyright notice and this permission notice (including the
24 * next paragraph) shall be included in all copies or substantial portions
28 #include "amdgpu_cs.h"
30 #include "util/os_time.h"
31 #include "util/u_hash_table.h"
32 #include "state_tracker/drm_driver.h"
33 #include <amdgpu_drm.h>
38 #ifndef AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
39 #define AMDGPU_GEM_CREATE_VM_ALWAYS_VALID (1 << 6)
42 #ifndef AMDGPU_VA_RANGE_HIGH
43 #define AMDGPU_VA_RANGE_HIGH 0x2
46 /* Set to 1 for verbose output showing committed sparse buffer ranges. */
47 #define DEBUG_SPARSE_COMMITS 0
49 struct amdgpu_sparse_backing_chunk
{
53 static struct pb_buffer
*
54 amdgpu_bo_create(struct radeon_winsys
*rws
,
57 enum radeon_bo_domain domain
,
58 enum radeon_bo_flag flags
);
59 static void amdgpu_bo_unmap(struct pb_buffer
*buf
);
61 static bool amdgpu_bo_wait(struct pb_buffer
*_buf
, uint64_t timeout
,
62 enum radeon_bo_usage usage
)
64 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
65 struct amdgpu_winsys
*ws
= bo
->ws
;
69 if (p_atomic_read(&bo
->num_active_ioctls
))
73 abs_timeout
= os_time_get_absolute_timeout(timeout
);
75 /* Wait if any ioctl is being submitted with this buffer. */
76 if (!os_wait_until_zero_abs_timeout(&bo
->num_active_ioctls
, abs_timeout
))
81 /* We can't use user fences for shared buffers, because user fences
82 * are local to this process only. If we want to wait for all buffer
83 * uses in all processes, we have to use amdgpu_bo_wait_for_idle.
85 bool buffer_busy
= true;
88 r
= amdgpu_bo_wait_for_idle(bo
->bo
, timeout
, &buffer_busy
);
90 fprintf(stderr
, "%s: amdgpu_bo_wait_for_idle failed %i\n", __func__
,
99 simple_mtx_lock(&ws
->bo_fence_lock
);
101 for (idle_fences
= 0; idle_fences
< bo
->num_fences
; ++idle_fences
) {
102 if (!amdgpu_fence_wait(bo
->fences
[idle_fences
], 0, false))
106 /* Release the idle fences to avoid checking them again later. */
107 for (unsigned i
= 0; i
< idle_fences
; ++i
)
108 amdgpu_fence_reference(&bo
->fences
[i
], NULL
);
110 memmove(&bo
->fences
[0], &bo
->fences
[idle_fences
],
111 (bo
->num_fences
- idle_fences
) * sizeof(*bo
->fences
));
112 bo
->num_fences
-= idle_fences
;
114 buffer_idle
= !bo
->num_fences
;
115 simple_mtx_unlock(&ws
->bo_fence_lock
);
119 bool buffer_idle
= true;
121 simple_mtx_lock(&ws
->bo_fence_lock
);
122 while (bo
->num_fences
&& buffer_idle
) {
123 struct pipe_fence_handle
*fence
= NULL
;
124 bool fence_idle
= false;
126 amdgpu_fence_reference(&fence
, bo
->fences
[0]);
128 /* Wait for the fence. */
129 simple_mtx_unlock(&ws
->bo_fence_lock
);
130 if (amdgpu_fence_wait(fence
, abs_timeout
, true))
134 simple_mtx_lock(&ws
->bo_fence_lock
);
136 /* Release an idle fence to avoid checking it again later, keeping in
137 * mind that the fence array may have been modified by other threads.
139 if (fence_idle
&& bo
->num_fences
&& bo
->fences
[0] == fence
) {
140 amdgpu_fence_reference(&bo
->fences
[0], NULL
);
141 memmove(&bo
->fences
[0], &bo
->fences
[1],
142 (bo
->num_fences
- 1) * sizeof(*bo
->fences
));
146 amdgpu_fence_reference(&fence
, NULL
);
148 simple_mtx_unlock(&ws
->bo_fence_lock
);
154 static enum radeon_bo_domain
amdgpu_bo_get_initial_domain(
155 struct pb_buffer
*buf
)
157 return ((struct amdgpu_winsys_bo
*)buf
)->initial_domain
;
160 static void amdgpu_bo_remove_fences(struct amdgpu_winsys_bo
*bo
)
162 for (unsigned i
= 0; i
< bo
->num_fences
; ++i
)
163 amdgpu_fence_reference(&bo
->fences
[i
], NULL
);
170 void amdgpu_bo_destroy(struct pb_buffer
*_buf
)
172 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
173 struct amdgpu_winsys
*ws
= bo
->ws
;
175 assert(bo
->bo
&& "must not be called for slab entries");
177 if (!bo
->is_user_ptr
&& bo
->cpu_ptr
) {
179 amdgpu_bo_unmap(&bo
->base
);
181 assert(bo
->is_user_ptr
|| bo
->u
.real
.map_count
== 0);
183 if (ws
->debug_all_bos
) {
184 simple_mtx_lock(&ws
->global_bo_list_lock
);
185 LIST_DEL(&bo
->u
.real
.global_list_item
);
187 simple_mtx_unlock(&ws
->global_bo_list_lock
);
190 simple_mtx_lock(&ws
->bo_export_table_lock
);
191 util_hash_table_remove(ws
->bo_export_table
, bo
->bo
);
192 simple_mtx_unlock(&ws
->bo_export_table_lock
);
194 amdgpu_bo_va_op(bo
->bo
, 0, bo
->base
.size
, bo
->va
, 0, AMDGPU_VA_OP_UNMAP
);
195 amdgpu_va_range_free(bo
->u
.real
.va_handle
);
196 amdgpu_bo_free(bo
->bo
);
198 amdgpu_bo_remove_fences(bo
);
200 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
201 ws
->allocated_vram
-= align64(bo
->base
.size
, ws
->info
.gart_page_size
);
202 else if (bo
->initial_domain
& RADEON_DOMAIN_GTT
)
203 ws
->allocated_gtt
-= align64(bo
->base
.size
, ws
->info
.gart_page_size
);
205 simple_mtx_destroy(&bo
->lock
);
209 static void amdgpu_bo_destroy_or_cache(struct pb_buffer
*_buf
)
211 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
213 assert(bo
->bo
); /* slab buffers have a separate vtbl */
215 if (bo
->u
.real
.use_reusable_pool
)
216 pb_cache_add_buffer(&bo
->u
.real
.cache_entry
);
218 amdgpu_bo_destroy(_buf
);
221 static void amdgpu_clean_up_buffer_managers(struct amdgpu_winsys
*ws
)
223 for (unsigned i
= 0; i
< NUM_SLAB_ALLOCATORS
; i
++)
224 pb_slabs_reclaim(&ws
->bo_slabs
[i
]);
226 pb_cache_release_all_buffers(&ws
->bo_cache
);
229 static bool amdgpu_bo_do_map(struct amdgpu_winsys_bo
*bo
, void **cpu
)
231 assert(!bo
->sparse
&& bo
->bo
&& !bo
->is_user_ptr
);
232 int r
= amdgpu_bo_cpu_map(bo
->bo
, cpu
);
234 /* Clean up buffer managers and try again. */
235 amdgpu_clean_up_buffer_managers(bo
->ws
);
236 r
= amdgpu_bo_cpu_map(bo
->bo
, cpu
);
241 if (p_atomic_inc_return(&bo
->u
.real
.map_count
) == 1) {
242 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
243 bo
->ws
->mapped_vram
+= bo
->base
.size
;
244 else if (bo
->initial_domain
& RADEON_DOMAIN_GTT
)
245 bo
->ws
->mapped_gtt
+= bo
->base
.size
;
246 bo
->ws
->num_mapped_buffers
++;
252 static void *amdgpu_bo_map(struct pb_buffer
*buf
,
253 struct radeon_cmdbuf
*rcs
,
254 enum pipe_transfer_usage usage
)
256 struct amdgpu_winsys_bo
*bo
= (struct amdgpu_winsys_bo
*)buf
;
257 struct amdgpu_winsys_bo
*real
;
258 struct amdgpu_cs
*cs
= (struct amdgpu_cs
*)rcs
;
262 /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */
263 if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
264 /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */
265 if (usage
& PIPE_TRANSFER_DONTBLOCK
) {
266 if (!(usage
& PIPE_TRANSFER_WRITE
)) {
269 * Since we are mapping for read, we don't need to wait
270 * if the GPU is using the buffer for read too
271 * (neither one is changing it).
273 * Only check whether the buffer is being used for write. */
274 if (cs
&& amdgpu_bo_is_referenced_by_cs_with_usage(cs
, bo
,
275 RADEON_USAGE_WRITE
)) {
276 cs
->flush_cs(cs
->flush_data
,
277 RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
281 if (!amdgpu_bo_wait((struct pb_buffer
*)bo
, 0,
282 RADEON_USAGE_WRITE
)) {
286 if (cs
&& amdgpu_bo_is_referenced_by_cs(cs
, bo
)) {
287 cs
->flush_cs(cs
->flush_data
,
288 RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
292 if (!amdgpu_bo_wait((struct pb_buffer
*)bo
, 0,
293 RADEON_USAGE_READWRITE
)) {
298 uint64_t time
= os_time_get_nano();
300 if (!(usage
& PIPE_TRANSFER_WRITE
)) {
303 * Since we are mapping for read, we don't need to wait
304 * if the GPU is using the buffer for read too
305 * (neither one is changing it).
307 * Only check whether the buffer is being used for write. */
309 if (amdgpu_bo_is_referenced_by_cs_with_usage(cs
, bo
,
310 RADEON_USAGE_WRITE
)) {
311 cs
->flush_cs(cs
->flush_data
,
312 RADEON_FLUSH_START_NEXT_GFX_IB_NOW
, NULL
);
314 /* Try to avoid busy-waiting in amdgpu_bo_wait. */
315 if (p_atomic_read(&bo
->num_active_ioctls
))
316 amdgpu_cs_sync_flush(rcs
);
320 amdgpu_bo_wait((struct pb_buffer
*)bo
, PIPE_TIMEOUT_INFINITE
,
323 /* Mapping for write. */
325 if (amdgpu_bo_is_referenced_by_cs(cs
, bo
)) {
326 cs
->flush_cs(cs
->flush_data
,
327 RADEON_FLUSH_START_NEXT_GFX_IB_NOW
, NULL
);
329 /* Try to avoid busy-waiting in amdgpu_bo_wait. */
330 if (p_atomic_read(&bo
->num_active_ioctls
))
331 amdgpu_cs_sync_flush(rcs
);
335 amdgpu_bo_wait((struct pb_buffer
*)bo
, PIPE_TIMEOUT_INFINITE
,
336 RADEON_USAGE_READWRITE
);
339 bo
->ws
->buffer_wait_time
+= os_time_get_nano() - time
;
343 /* Buffer synchronization has been checked, now actually map the buffer. */
350 real
= bo
->u
.slab
.real
;
351 offset
= bo
->va
- real
->va
;
354 if (usage
& RADEON_TRANSFER_TEMPORARY
) {
355 if (real
->is_user_ptr
) {
358 if (!amdgpu_bo_do_map(real
, &cpu
))
362 cpu
= p_atomic_read(&real
->cpu_ptr
);
364 simple_mtx_lock(&real
->lock
);
365 /* Must re-check due to the possibility of a race. Re-check need not
366 * be atomic thanks to the lock. */
369 if (!amdgpu_bo_do_map(real
, &cpu
)) {
370 simple_mtx_unlock(&real
->lock
);
373 p_atomic_set(&real
->cpu_ptr
, cpu
);
375 simple_mtx_unlock(&real
->lock
);
379 return (uint8_t*)cpu
+ offset
;
382 static void amdgpu_bo_unmap(struct pb_buffer
*buf
)
384 struct amdgpu_winsys_bo
*bo
= (struct amdgpu_winsys_bo
*)buf
;
385 struct amdgpu_winsys_bo
*real
;
392 real
= bo
->bo
? bo
: bo
->u
.slab
.real
;
393 assert(real
->u
.real
.map_count
!= 0 && "too many unmaps");
394 if (p_atomic_dec_zero(&real
->u
.real
.map_count
)) {
395 assert(!real
->cpu_ptr
&&
396 "too many unmaps or forgot RADEON_TRANSFER_TEMPORARY flag");
398 if (real
->initial_domain
& RADEON_DOMAIN_VRAM
)
399 real
->ws
->mapped_vram
-= real
->base
.size
;
400 else if (real
->initial_domain
& RADEON_DOMAIN_GTT
)
401 real
->ws
->mapped_gtt
-= real
->base
.size
;
402 real
->ws
->num_mapped_buffers
--;
405 amdgpu_bo_cpu_unmap(real
->bo
);
408 static const struct pb_vtbl amdgpu_winsys_bo_vtbl
= {
409 amdgpu_bo_destroy_or_cache
410 /* other functions are never called */
413 static void amdgpu_add_buffer_to_global_list(struct amdgpu_winsys_bo
*bo
)
415 struct amdgpu_winsys
*ws
= bo
->ws
;
419 if (ws
->debug_all_bos
) {
420 simple_mtx_lock(&ws
->global_bo_list_lock
);
421 LIST_ADDTAIL(&bo
->u
.real
.global_list_item
, &ws
->global_bo_list
);
423 simple_mtx_unlock(&ws
->global_bo_list_lock
);
427 static struct amdgpu_winsys_bo
*amdgpu_create_bo(struct amdgpu_winsys
*ws
,
430 enum radeon_bo_domain initial_domain
,
434 struct amdgpu_bo_alloc_request request
= {0};
435 amdgpu_bo_handle buf_handle
;
437 struct amdgpu_winsys_bo
*bo
;
438 amdgpu_va_handle va_handle
;
439 unsigned va_gap_size
;
442 /* VRAM or GTT must be specified, but not both at the same time. */
443 assert(util_bitcount(initial_domain
& RADEON_DOMAIN_VRAM_GTT
) == 1);
445 /* Gfx9: Overallocate the size to the next power of two for faster address
446 * translation if we don't waste too much memory.
448 if (ws
->info
.chip_class
>= GFX9
) {
449 uint64_t next_pot_size
= util_next_power_of_two64(size
);
451 /* For slightly lower than 4 GB allocations, at most 32 MB are wasted.
452 * For slightly lower than 256 MB allocations, at most 2 MB are wasted.
453 * For slightly lower than 64 MB allocations, at most 512 KB are wasted.
455 * Waste at most 0.79% (1/127) of the size if we decide to overallocate.
457 uint64_t max_overalloc
= next_pot_size
>> 7;
459 /* If the next power-of-two size is <= the page size, waste up to
460 * 6.25% (1/16) of the size if we decide to overallocate.
462 if (next_pot_size
<= ws
->info
.pte_fragment_size
)
463 max_overalloc
= next_pot_size
>> 4;
465 if (size
+ max_overalloc
>= next_pot_size
)
466 size
= next_pot_size
;
469 bo
= CALLOC_STRUCT(amdgpu_winsys_bo
);
475 pb_cache_init_entry(&ws
->bo_cache
, &bo
->u
.real
.cache_entry
, &bo
->base
,
478 request
.alloc_size
= size
;
479 request
.phys_alignment
= alignment
;
481 if (initial_domain
& RADEON_DOMAIN_VRAM
)
482 request
.preferred_heap
|= AMDGPU_GEM_DOMAIN_VRAM
;
483 if (initial_domain
& RADEON_DOMAIN_GTT
)
484 request
.preferred_heap
|= AMDGPU_GEM_DOMAIN_GTT
;
486 /* Since VRAM and GTT have almost the same performance on APUs, we could
487 * just set GTT. However, in order to decrease GTT(RAM) usage, which is
488 * shared with the OS, allow VRAM placements too. The idea is not to use
489 * VRAM usefully, but to use it so that it's not unused and wasted.
491 if (!ws
->info
.has_dedicated_vram
)
492 request
.preferred_heap
|= AMDGPU_GEM_DOMAIN_GTT
;
494 if (flags
& RADEON_FLAG_NO_CPU_ACCESS
)
495 request
.flags
|= AMDGPU_GEM_CREATE_NO_CPU_ACCESS
;
496 if (flags
& RADEON_FLAG_GTT_WC
)
497 request
.flags
|= AMDGPU_GEM_CREATE_CPU_GTT_USWC
;
498 if (flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
&&
499 ws
->info
.has_local_buffers
)
500 request
.flags
|= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
;
501 if (ws
->zero_all_vram_allocs
&&
502 (request
.preferred_heap
& AMDGPU_GEM_DOMAIN_VRAM
))
503 request
.flags
|= AMDGPU_GEM_CREATE_VRAM_CLEARED
;
505 r
= amdgpu_bo_alloc(ws
->dev
, &request
, &buf_handle
);
507 fprintf(stderr
, "amdgpu: Failed to allocate a buffer:\n");
508 fprintf(stderr
, "amdgpu: size : %"PRIu64
" bytes\n", size
);
509 fprintf(stderr
, "amdgpu: alignment : %u bytes\n", alignment
);
510 fprintf(stderr
, "amdgpu: domains : %u\n", initial_domain
);
514 va_gap_size
= ws
->check_vm
? MAX2(4 * alignment
, 64 * 1024) : 0;
516 uint64_t vm_alignment
= alignment
;
518 /* Increase the VM alignment for faster address translation. */
519 if (size
>= ws
->info
.pte_fragment_size
)
520 vm_alignment
= MAX2(vm_alignment
, ws
->info
.pte_fragment_size
);
522 /* Gfx9: Increase the VM alignment to the most significant bit set
523 * in the size for faster address translation.
525 if (ws
->info
.chip_class
>= GFX9
) {
526 unsigned msb
= util_last_bit64(size
); /* 0 = no bit is set */
527 uint64_t msb_alignment
= msb
? 1ull << (msb
- 1) : 0;
529 vm_alignment
= MAX2(vm_alignment
, msb_alignment
);
532 r
= amdgpu_va_range_alloc(ws
->dev
, amdgpu_gpu_va_range_general
,
533 size
+ va_gap_size
, vm_alignment
, 0, &va
, &va_handle
,
534 (flags
& RADEON_FLAG_32BIT
? AMDGPU_VA_RANGE_32_BIT
: 0) |
535 AMDGPU_VA_RANGE_HIGH
);
539 unsigned vm_flags
= AMDGPU_VM_PAGE_READABLE
|
540 AMDGPU_VM_PAGE_EXECUTABLE
;
542 if (!(flags
& RADEON_FLAG_READ_ONLY
))
543 vm_flags
|= AMDGPU_VM_PAGE_WRITEABLE
;
545 r
= amdgpu_bo_va_op_raw(ws
->dev
, buf_handle
, 0, size
, va
, vm_flags
,
550 simple_mtx_init(&bo
->lock
, mtx_plain
);
551 pipe_reference_init(&bo
->base
.reference
, 1);
552 bo
->base
.alignment
= alignment
;
554 bo
->base
.size
= size
;
555 bo
->base
.vtbl
= &amdgpu_winsys_bo_vtbl
;
559 bo
->u
.real
.va_handle
= va_handle
;
560 bo
->initial_domain
= initial_domain
;
561 bo
->unique_id
= __sync_fetch_and_add(&ws
->next_bo_unique_id
, 1);
562 bo
->is_local
= !!(request
.flags
& AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
);
564 if (initial_domain
& RADEON_DOMAIN_VRAM
)
565 ws
->allocated_vram
+= align64(size
, ws
->info
.gart_page_size
);
566 else if (initial_domain
& RADEON_DOMAIN_GTT
)
567 ws
->allocated_gtt
+= align64(size
, ws
->info
.gart_page_size
);
569 amdgpu_bo_export(bo
->bo
, amdgpu_bo_handle_type_kms
, &bo
->u
.real
.kms_handle
);
571 amdgpu_add_buffer_to_global_list(bo
);
576 amdgpu_va_range_free(va_handle
);
579 amdgpu_bo_free(buf_handle
);
586 bool amdgpu_bo_can_reclaim(struct pb_buffer
*_buf
)
588 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
590 if (amdgpu_bo_is_referenced_by_any_cs(bo
)) {
594 return amdgpu_bo_wait(_buf
, 0, RADEON_USAGE_READWRITE
);
597 bool amdgpu_bo_can_reclaim_slab(void *priv
, struct pb_slab_entry
*entry
)
599 struct amdgpu_winsys_bo
*bo
= NULL
; /* fix container_of */
600 bo
= container_of(entry
, bo
, u
.slab
.entry
);
602 return amdgpu_bo_can_reclaim(&bo
->base
);
605 static struct pb_slabs
*get_slabs(struct amdgpu_winsys
*ws
, uint64_t size
)
607 /* Find the correct slab allocator for the given size. */
608 for (unsigned i
= 0; i
< NUM_SLAB_ALLOCATORS
; i
++) {
609 struct pb_slabs
*slabs
= &ws
->bo_slabs
[i
];
611 if (size
<= 1 << (slabs
->min_order
+ slabs
->num_orders
- 1))
619 static void amdgpu_bo_slab_destroy(struct pb_buffer
*_buf
)
621 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
625 pb_slab_free(get_slabs(bo
->ws
, bo
->base
.size
), &bo
->u
.slab
.entry
);
628 static const struct pb_vtbl amdgpu_winsys_bo_slab_vtbl
= {
629 amdgpu_bo_slab_destroy
630 /* other functions are never called */
633 struct pb_slab
*amdgpu_bo_slab_alloc(void *priv
, unsigned heap
,
635 unsigned group_index
)
637 struct amdgpu_winsys
*ws
= priv
;
638 struct amdgpu_slab
*slab
= CALLOC_STRUCT(amdgpu_slab
);
639 enum radeon_bo_domain domains
= radeon_domain_from_heap(heap
);
640 enum radeon_bo_flag flags
= radeon_flags_from_heap(heap
);
642 unsigned slab_size
= 0;
647 /* Determine the slab buffer size. */
648 for (unsigned i
= 0; i
< NUM_SLAB_ALLOCATORS
; i
++) {
649 struct pb_slabs
*slabs
= &ws
->bo_slabs
[i
];
650 unsigned max_entry_size
= 1 << (slabs
->min_order
+ slabs
->num_orders
- 1);
652 if (entry_size
<= max_entry_size
) {
653 /* The slab size is twice the size of the largest possible entry. */
654 slab_size
= max_entry_size
* 2;
656 /* The largest slab should have the same size as the PTE fragment
657 * size to get faster address translation.
659 if (i
== NUM_SLAB_ALLOCATORS
- 1 &&
660 slab_size
< ws
->info
.pte_fragment_size
)
661 slab_size
= ws
->info
.pte_fragment_size
;
665 assert(slab_size
!= 0);
667 slab
->buffer
= amdgpu_winsys_bo(amdgpu_bo_create(&ws
->base
,
668 slab_size
, slab_size
,
673 slab
->base
.num_entries
= slab
->buffer
->base
.size
/ entry_size
;
674 slab
->base
.num_free
= slab
->base
.num_entries
;
675 slab
->entries
= CALLOC(slab
->base
.num_entries
, sizeof(*slab
->entries
));
679 LIST_INITHEAD(&slab
->base
.free
);
681 base_id
= __sync_fetch_and_add(&ws
->next_bo_unique_id
, slab
->base
.num_entries
);
683 for (unsigned i
= 0; i
< slab
->base
.num_entries
; ++i
) {
684 struct amdgpu_winsys_bo
*bo
= &slab
->entries
[i
];
686 simple_mtx_init(&bo
->lock
, mtx_plain
);
687 bo
->base
.alignment
= entry_size
;
688 bo
->base
.usage
= slab
->buffer
->base
.usage
;
689 bo
->base
.size
= entry_size
;
690 bo
->base
.vtbl
= &amdgpu_winsys_bo_slab_vtbl
;
692 bo
->va
= slab
->buffer
->va
+ i
* entry_size
;
693 bo
->initial_domain
= domains
;
694 bo
->unique_id
= base_id
+ i
;
695 bo
->u
.slab
.entry
.slab
= &slab
->base
;
696 bo
->u
.slab
.entry
.group_index
= group_index
;
698 if (slab
->buffer
->bo
) {
699 /* The slab is not suballocated. */
700 bo
->u
.slab
.real
= slab
->buffer
;
702 /* The slab is allocated out of a bigger slab. */
703 bo
->u
.slab
.real
= slab
->buffer
->u
.slab
.real
;
704 assert(bo
->u
.slab
.real
->bo
);
707 LIST_ADDTAIL(&bo
->u
.slab
.entry
.head
, &slab
->base
.free
);
713 amdgpu_winsys_bo_reference(&slab
->buffer
, NULL
);
719 void amdgpu_bo_slab_free(void *priv
, struct pb_slab
*pslab
)
721 struct amdgpu_slab
*slab
= amdgpu_slab(pslab
);
723 for (unsigned i
= 0; i
< slab
->base
.num_entries
; ++i
) {
724 amdgpu_bo_remove_fences(&slab
->entries
[i
]);
725 simple_mtx_destroy(&slab
->entries
[i
].lock
);
729 amdgpu_winsys_bo_reference(&slab
->buffer
, NULL
);
733 #if DEBUG_SPARSE_COMMITS
735 sparse_dump(struct amdgpu_winsys_bo
*bo
, const char *func
)
737 fprintf(stderr
, "%s: %p (size=%"PRIu64
", num_va_pages=%u) @ %s\n"
739 __func__
, bo
, bo
->base
.size
, bo
->u
.sparse
.num_va_pages
, func
);
741 struct amdgpu_sparse_backing
*span_backing
= NULL
;
742 uint32_t span_first_backing_page
= 0;
743 uint32_t span_first_va_page
= 0;
744 uint32_t va_page
= 0;
747 struct amdgpu_sparse_backing
*backing
= 0;
748 uint32_t backing_page
= 0;
750 if (va_page
< bo
->u
.sparse
.num_va_pages
) {
751 backing
= bo
->u
.sparse
.commitments
[va_page
].backing
;
752 backing_page
= bo
->u
.sparse
.commitments
[va_page
].page
;
756 (backing
!= span_backing
||
757 backing_page
!= span_first_backing_page
+ (va_page
- span_first_va_page
))) {
758 fprintf(stderr
, " %u..%u: backing=%p:%u..%u\n",
759 span_first_va_page
, va_page
- 1, span_backing
,
760 span_first_backing_page
,
761 span_first_backing_page
+ (va_page
- span_first_va_page
) - 1);
766 if (va_page
>= bo
->u
.sparse
.num_va_pages
)
769 if (backing
&& !span_backing
) {
770 span_backing
= backing
;
771 span_first_backing_page
= backing_page
;
772 span_first_va_page
= va_page
;
778 fprintf(stderr
, "Backing:\n");
780 list_for_each_entry(struct amdgpu_sparse_backing
, backing
, &bo
->u
.sparse
.backing
, list
) {
781 fprintf(stderr
, " %p (size=%"PRIu64
")\n", backing
, backing
->bo
->base
.size
);
782 for (unsigned i
= 0; i
< backing
->num_chunks
; ++i
)
783 fprintf(stderr
, " %u..%u\n", backing
->chunks
[i
].begin
, backing
->chunks
[i
].end
);
789 * Attempt to allocate the given number of backing pages. Fewer pages may be
790 * allocated (depending on the fragmentation of existing backing buffers),
791 * which will be reflected by a change to *pnum_pages.
793 static struct amdgpu_sparse_backing
*
794 sparse_backing_alloc(struct amdgpu_winsys_bo
*bo
, uint32_t *pstart_page
, uint32_t *pnum_pages
)
796 struct amdgpu_sparse_backing
*best_backing
;
798 uint32_t best_num_pages
;
804 /* This is a very simple and inefficient best-fit algorithm. */
805 list_for_each_entry(struct amdgpu_sparse_backing
, backing
, &bo
->u
.sparse
.backing
, list
) {
806 for (unsigned idx
= 0; idx
< backing
->num_chunks
; ++idx
) {
807 uint32_t cur_num_pages
= backing
->chunks
[idx
].end
- backing
->chunks
[idx
].begin
;
808 if ((best_num_pages
< *pnum_pages
&& cur_num_pages
> best_num_pages
) ||
809 (best_num_pages
> *pnum_pages
&& cur_num_pages
< best_num_pages
)) {
810 best_backing
= backing
;
812 best_num_pages
= cur_num_pages
;
817 /* Allocate a new backing buffer if necessary. */
819 struct pb_buffer
*buf
;
823 best_backing
= CALLOC_STRUCT(amdgpu_sparse_backing
);
827 best_backing
->max_chunks
= 4;
828 best_backing
->chunks
= CALLOC(best_backing
->max_chunks
,
829 sizeof(*best_backing
->chunks
));
830 if (!best_backing
->chunks
) {
835 assert(bo
->u
.sparse
.num_backing_pages
< DIV_ROUND_UP(bo
->base
.size
, RADEON_SPARSE_PAGE_SIZE
));
837 size
= MIN3(bo
->base
.size
/ 16,
839 bo
->base
.size
- (uint64_t)bo
->u
.sparse
.num_backing_pages
* RADEON_SPARSE_PAGE_SIZE
);
840 size
= MAX2(size
, RADEON_SPARSE_PAGE_SIZE
);
842 buf
= amdgpu_bo_create(&bo
->ws
->base
, size
, RADEON_SPARSE_PAGE_SIZE
,
844 bo
->u
.sparse
.flags
| RADEON_FLAG_NO_SUBALLOC
);
846 FREE(best_backing
->chunks
);
851 /* We might have gotten a bigger buffer than requested via caching. */
852 pages
= buf
->size
/ RADEON_SPARSE_PAGE_SIZE
;
854 best_backing
->bo
= amdgpu_winsys_bo(buf
);
855 best_backing
->num_chunks
= 1;
856 best_backing
->chunks
[0].begin
= 0;
857 best_backing
->chunks
[0].end
= pages
;
859 list_add(&best_backing
->list
, &bo
->u
.sparse
.backing
);
860 bo
->u
.sparse
.num_backing_pages
+= pages
;
863 best_num_pages
= pages
;
866 *pnum_pages
= MIN2(*pnum_pages
, best_num_pages
);
867 *pstart_page
= best_backing
->chunks
[best_idx
].begin
;
868 best_backing
->chunks
[best_idx
].begin
+= *pnum_pages
;
870 if (best_backing
->chunks
[best_idx
].begin
>= best_backing
->chunks
[best_idx
].end
) {
871 memmove(&best_backing
->chunks
[best_idx
], &best_backing
->chunks
[best_idx
+ 1],
872 sizeof(*best_backing
->chunks
) * (best_backing
->num_chunks
- best_idx
- 1));
873 best_backing
->num_chunks
--;
880 sparse_free_backing_buffer(struct amdgpu_winsys_bo
*bo
,
881 struct amdgpu_sparse_backing
*backing
)
883 struct amdgpu_winsys
*ws
= backing
->bo
->ws
;
885 bo
->u
.sparse
.num_backing_pages
-= backing
->bo
->base
.size
/ RADEON_SPARSE_PAGE_SIZE
;
887 simple_mtx_lock(&ws
->bo_fence_lock
);
888 amdgpu_add_fences(backing
->bo
, bo
->num_fences
, bo
->fences
);
889 simple_mtx_unlock(&ws
->bo_fence_lock
);
891 list_del(&backing
->list
);
892 amdgpu_winsys_bo_reference(&backing
->bo
, NULL
);
893 FREE(backing
->chunks
);
898 * Return a range of pages from the given backing buffer back into the
902 sparse_backing_free(struct amdgpu_winsys_bo
*bo
,
903 struct amdgpu_sparse_backing
*backing
,
904 uint32_t start_page
, uint32_t num_pages
)
906 uint32_t end_page
= start_page
+ num_pages
;
908 unsigned high
= backing
->num_chunks
;
910 /* Find the first chunk with begin >= start_page. */
912 unsigned mid
= low
+ (high
- low
) / 2;
914 if (backing
->chunks
[mid
].begin
>= start_page
)
920 assert(low
>= backing
->num_chunks
|| end_page
<= backing
->chunks
[low
].begin
);
921 assert(low
== 0 || backing
->chunks
[low
- 1].end
<= start_page
);
923 if (low
> 0 && backing
->chunks
[low
- 1].end
== start_page
) {
924 backing
->chunks
[low
- 1].end
= end_page
;
926 if (low
< backing
->num_chunks
&& end_page
== backing
->chunks
[low
].begin
) {
927 backing
->chunks
[low
- 1].end
= backing
->chunks
[low
].end
;
928 memmove(&backing
->chunks
[low
], &backing
->chunks
[low
+ 1],
929 sizeof(*backing
->chunks
) * (backing
->num_chunks
- low
- 1));
930 backing
->num_chunks
--;
932 } else if (low
< backing
->num_chunks
&& end_page
== backing
->chunks
[low
].begin
) {
933 backing
->chunks
[low
].begin
= start_page
;
935 if (backing
->num_chunks
>= backing
->max_chunks
) {
936 unsigned new_max_chunks
= 2 * backing
->max_chunks
;
937 struct amdgpu_sparse_backing_chunk
*new_chunks
=
938 REALLOC(backing
->chunks
,
939 sizeof(*backing
->chunks
) * backing
->max_chunks
,
940 sizeof(*backing
->chunks
) * new_max_chunks
);
944 backing
->max_chunks
= new_max_chunks
;
945 backing
->chunks
= new_chunks
;
948 memmove(&backing
->chunks
[low
+ 1], &backing
->chunks
[low
],
949 sizeof(*backing
->chunks
) * (backing
->num_chunks
- low
));
950 backing
->chunks
[low
].begin
= start_page
;
951 backing
->chunks
[low
].end
= end_page
;
952 backing
->num_chunks
++;
955 if (backing
->num_chunks
== 1 && backing
->chunks
[0].begin
== 0 &&
956 backing
->chunks
[0].end
== backing
->bo
->base
.size
/ RADEON_SPARSE_PAGE_SIZE
)
957 sparse_free_backing_buffer(bo
, backing
);
962 static void amdgpu_bo_sparse_destroy(struct pb_buffer
*_buf
)
964 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
967 assert(!bo
->bo
&& bo
->sparse
);
969 r
= amdgpu_bo_va_op_raw(bo
->ws
->dev
, NULL
, 0,
970 (uint64_t)bo
->u
.sparse
.num_va_pages
* RADEON_SPARSE_PAGE_SIZE
,
971 bo
->va
, 0, AMDGPU_VA_OP_CLEAR
);
973 fprintf(stderr
, "amdgpu: clearing PRT VA region on destroy failed (%d)\n", r
);
976 while (!list_empty(&bo
->u
.sparse
.backing
)) {
977 struct amdgpu_sparse_backing
*dummy
= NULL
;
978 sparse_free_backing_buffer(bo
,
979 container_of(bo
->u
.sparse
.backing
.next
,
983 amdgpu_va_range_free(bo
->u
.sparse
.va_handle
);
984 FREE(bo
->u
.sparse
.commitments
);
985 simple_mtx_destroy(&bo
->lock
);
989 static const struct pb_vtbl amdgpu_winsys_bo_sparse_vtbl
= {
990 amdgpu_bo_sparse_destroy
991 /* other functions are never called */
994 static struct pb_buffer
*
995 amdgpu_bo_sparse_create(struct amdgpu_winsys
*ws
, uint64_t size
,
996 enum radeon_bo_domain domain
,
997 enum radeon_bo_flag flags
)
999 struct amdgpu_winsys_bo
*bo
;
1001 uint64_t va_gap_size
;
1004 /* We use 32-bit page numbers; refuse to attempt allocating sparse buffers
1005 * that exceed this limit. This is not really a restriction: we don't have
1006 * that much virtual address space anyway.
1008 if (size
> (uint64_t)INT32_MAX
* RADEON_SPARSE_PAGE_SIZE
)
1011 bo
= CALLOC_STRUCT(amdgpu_winsys_bo
);
1015 simple_mtx_init(&bo
->lock
, mtx_plain
);
1016 pipe_reference_init(&bo
->base
.reference
, 1);
1017 bo
->base
.alignment
= RADEON_SPARSE_PAGE_SIZE
;
1018 bo
->base
.size
= size
;
1019 bo
->base
.vtbl
= &amdgpu_winsys_bo_sparse_vtbl
;
1021 bo
->initial_domain
= domain
;
1022 bo
->unique_id
= __sync_fetch_and_add(&ws
->next_bo_unique_id
, 1);
1024 bo
->u
.sparse
.flags
= flags
& ~RADEON_FLAG_SPARSE
;
1026 bo
->u
.sparse
.num_va_pages
= DIV_ROUND_UP(size
, RADEON_SPARSE_PAGE_SIZE
);
1027 bo
->u
.sparse
.commitments
= CALLOC(bo
->u
.sparse
.num_va_pages
,
1028 sizeof(*bo
->u
.sparse
.commitments
));
1029 if (!bo
->u
.sparse
.commitments
)
1030 goto error_alloc_commitments
;
1032 LIST_INITHEAD(&bo
->u
.sparse
.backing
);
1034 /* For simplicity, we always map a multiple of the page size. */
1035 map_size
= align64(size
, RADEON_SPARSE_PAGE_SIZE
);
1036 va_gap_size
= ws
->check_vm
? 4 * RADEON_SPARSE_PAGE_SIZE
: 0;
1037 r
= amdgpu_va_range_alloc(ws
->dev
, amdgpu_gpu_va_range_general
,
1038 map_size
+ va_gap_size
, RADEON_SPARSE_PAGE_SIZE
,
1039 0, &bo
->va
, &bo
->u
.sparse
.va_handle
,
1040 AMDGPU_VA_RANGE_HIGH
);
1042 goto error_va_alloc
;
1044 r
= amdgpu_bo_va_op_raw(bo
->ws
->dev
, NULL
, 0, size
, bo
->va
,
1045 AMDGPU_VM_PAGE_PRT
, AMDGPU_VA_OP_MAP
);
1052 amdgpu_va_range_free(bo
->u
.sparse
.va_handle
);
1054 FREE(bo
->u
.sparse
.commitments
);
1055 error_alloc_commitments
:
1056 simple_mtx_destroy(&bo
->lock
);
1062 amdgpu_bo_sparse_commit(struct pb_buffer
*buf
, uint64_t offset
, uint64_t size
,
1065 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(buf
);
1066 struct amdgpu_sparse_commitment
*comm
;
1067 uint32_t va_page
, end_va_page
;
1072 assert(offset
% RADEON_SPARSE_PAGE_SIZE
== 0);
1073 assert(offset
<= bo
->base
.size
);
1074 assert(size
<= bo
->base
.size
- offset
);
1075 assert(size
% RADEON_SPARSE_PAGE_SIZE
== 0 || offset
+ size
== bo
->base
.size
);
1077 comm
= bo
->u
.sparse
.commitments
;
1078 va_page
= offset
/ RADEON_SPARSE_PAGE_SIZE
;
1079 end_va_page
= va_page
+ DIV_ROUND_UP(size
, RADEON_SPARSE_PAGE_SIZE
);
1081 simple_mtx_lock(&bo
->lock
);
1083 #if DEBUG_SPARSE_COMMITS
1084 sparse_dump(bo
, __func__
);
1088 while (va_page
< end_va_page
) {
1089 uint32_t span_va_page
;
1091 /* Skip pages that are already committed. */
1092 if (comm
[va_page
].backing
) {
1097 /* Determine length of uncommitted span. */
1098 span_va_page
= va_page
;
1099 while (va_page
< end_va_page
&& !comm
[va_page
].backing
)
1102 /* Fill the uncommitted span with chunks of backing memory. */
1103 while (span_va_page
< va_page
) {
1104 struct amdgpu_sparse_backing
*backing
;
1105 uint32_t backing_start
, backing_size
;
1107 backing_size
= va_page
- span_va_page
;
1108 backing
= sparse_backing_alloc(bo
, &backing_start
, &backing_size
);
1114 r
= amdgpu_bo_va_op_raw(bo
->ws
->dev
, backing
->bo
->bo
,
1115 (uint64_t)backing_start
* RADEON_SPARSE_PAGE_SIZE
,
1116 (uint64_t)backing_size
* RADEON_SPARSE_PAGE_SIZE
,
1117 bo
->va
+ (uint64_t)span_va_page
* RADEON_SPARSE_PAGE_SIZE
,
1118 AMDGPU_VM_PAGE_READABLE
|
1119 AMDGPU_VM_PAGE_WRITEABLE
|
1120 AMDGPU_VM_PAGE_EXECUTABLE
,
1121 AMDGPU_VA_OP_REPLACE
);
1123 ok
= sparse_backing_free(bo
, backing
, backing_start
, backing_size
);
1124 assert(ok
&& "sufficient memory should already be allocated");
1130 while (backing_size
) {
1131 comm
[span_va_page
].backing
= backing
;
1132 comm
[span_va_page
].page
= backing_start
;
1140 r
= amdgpu_bo_va_op_raw(bo
->ws
->dev
, NULL
, 0,
1141 (uint64_t)(end_va_page
- va_page
) * RADEON_SPARSE_PAGE_SIZE
,
1142 bo
->va
+ (uint64_t)va_page
* RADEON_SPARSE_PAGE_SIZE
,
1143 AMDGPU_VM_PAGE_PRT
, AMDGPU_VA_OP_REPLACE
);
1149 while (va_page
< end_va_page
) {
1150 struct amdgpu_sparse_backing
*backing
;
1151 uint32_t backing_start
;
1152 uint32_t span_pages
;
1154 /* Skip pages that are already uncommitted. */
1155 if (!comm
[va_page
].backing
) {
1160 /* Group contiguous spans of pages. */
1161 backing
= comm
[va_page
].backing
;
1162 backing_start
= comm
[va_page
].page
;
1163 comm
[va_page
].backing
= NULL
;
1168 while (va_page
< end_va_page
&&
1169 comm
[va_page
].backing
== backing
&&
1170 comm
[va_page
].page
== backing_start
+ span_pages
) {
1171 comm
[va_page
].backing
= NULL
;
1176 if (!sparse_backing_free(bo
, backing
, backing_start
, span_pages
)) {
1177 /* Couldn't allocate tracking data structures, so we have to leak */
1178 fprintf(stderr
, "amdgpu: leaking PRT backing memory\n");
1185 simple_mtx_unlock(&bo
->lock
);
1190 static unsigned eg_tile_split(unsigned tile_split
)
1192 switch (tile_split
) {
1193 case 0: tile_split
= 64; break;
1194 case 1: tile_split
= 128; break;
1195 case 2: tile_split
= 256; break;
1196 case 3: tile_split
= 512; break;
1198 case 4: tile_split
= 1024; break;
1199 case 5: tile_split
= 2048; break;
1200 case 6: tile_split
= 4096; break;
1205 static unsigned eg_tile_split_rev(unsigned eg_tile_split
)
1207 switch (eg_tile_split
) {
1213 case 1024: return 4;
1214 case 2048: return 5;
1215 case 4096: return 6;
1219 static void amdgpu_buffer_get_metadata(struct pb_buffer
*_buf
,
1220 struct radeon_bo_metadata
*md
)
1222 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
1223 struct amdgpu_bo_info info
= {0};
1224 uint64_t tiling_flags
;
1227 assert(bo
->bo
&& "must not be called for slab entries");
1229 r
= amdgpu_bo_query_info(bo
->bo
, &info
);
1233 tiling_flags
= info
.metadata
.tiling_info
;
1235 if (bo
->ws
->info
.chip_class
>= GFX9
) {
1236 md
->u
.gfx9
.swizzle_mode
= AMDGPU_TILING_GET(tiling_flags
, SWIZZLE_MODE
);
1238 md
->u
.legacy
.microtile
= RADEON_LAYOUT_LINEAR
;
1239 md
->u
.legacy
.macrotile
= RADEON_LAYOUT_LINEAR
;
1241 if (AMDGPU_TILING_GET(tiling_flags
, ARRAY_MODE
) == 4) /* 2D_TILED_THIN1 */
1242 md
->u
.legacy
.macrotile
= RADEON_LAYOUT_TILED
;
1243 else if (AMDGPU_TILING_GET(tiling_flags
, ARRAY_MODE
) == 2) /* 1D_TILED_THIN1 */
1244 md
->u
.legacy
.microtile
= RADEON_LAYOUT_TILED
;
1246 md
->u
.legacy
.pipe_config
= AMDGPU_TILING_GET(tiling_flags
, PIPE_CONFIG
);
1247 md
->u
.legacy
.bankw
= 1 << AMDGPU_TILING_GET(tiling_flags
, BANK_WIDTH
);
1248 md
->u
.legacy
.bankh
= 1 << AMDGPU_TILING_GET(tiling_flags
, BANK_HEIGHT
);
1249 md
->u
.legacy
.tile_split
= eg_tile_split(AMDGPU_TILING_GET(tiling_flags
, TILE_SPLIT
));
1250 md
->u
.legacy
.mtilea
= 1 << AMDGPU_TILING_GET(tiling_flags
, MACRO_TILE_ASPECT
);
1251 md
->u
.legacy
.num_banks
= 2 << AMDGPU_TILING_GET(tiling_flags
, NUM_BANKS
);
1252 md
->u
.legacy
.scanout
= AMDGPU_TILING_GET(tiling_flags
, MICRO_TILE_MODE
) == 0; /* DISPLAY */
1255 md
->size_metadata
= info
.metadata
.size_metadata
;
1256 memcpy(md
->metadata
, info
.metadata
.umd_metadata
, sizeof(md
->metadata
));
1259 static void amdgpu_buffer_set_metadata(struct pb_buffer
*_buf
,
1260 struct radeon_bo_metadata
*md
)
1262 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(_buf
);
1263 struct amdgpu_bo_metadata metadata
= {0};
1264 uint64_t tiling_flags
= 0;
1266 assert(bo
->bo
&& "must not be called for slab entries");
1268 if (bo
->ws
->info
.chip_class
>= GFX9
) {
1269 tiling_flags
|= AMDGPU_TILING_SET(SWIZZLE_MODE
, md
->u
.gfx9
.swizzle_mode
);
1271 if (md
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
1272 tiling_flags
|= AMDGPU_TILING_SET(ARRAY_MODE
, 4); /* 2D_TILED_THIN1 */
1273 else if (md
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
1274 tiling_flags
|= AMDGPU_TILING_SET(ARRAY_MODE
, 2); /* 1D_TILED_THIN1 */
1276 tiling_flags
|= AMDGPU_TILING_SET(ARRAY_MODE
, 1); /* LINEAR_ALIGNED */
1278 tiling_flags
|= AMDGPU_TILING_SET(PIPE_CONFIG
, md
->u
.legacy
.pipe_config
);
1279 tiling_flags
|= AMDGPU_TILING_SET(BANK_WIDTH
, util_logbase2(md
->u
.legacy
.bankw
));
1280 tiling_flags
|= AMDGPU_TILING_SET(BANK_HEIGHT
, util_logbase2(md
->u
.legacy
.bankh
));
1281 if (md
->u
.legacy
.tile_split
)
1282 tiling_flags
|= AMDGPU_TILING_SET(TILE_SPLIT
, eg_tile_split_rev(md
->u
.legacy
.tile_split
));
1283 tiling_flags
|= AMDGPU_TILING_SET(MACRO_TILE_ASPECT
, util_logbase2(md
->u
.legacy
.mtilea
));
1284 tiling_flags
|= AMDGPU_TILING_SET(NUM_BANKS
, util_logbase2(md
->u
.legacy
.num_banks
)-1);
1286 if (md
->u
.legacy
.scanout
)
1287 tiling_flags
|= AMDGPU_TILING_SET(MICRO_TILE_MODE
, 0); /* DISPLAY_MICRO_TILING */
1289 tiling_flags
|= AMDGPU_TILING_SET(MICRO_TILE_MODE
, 1); /* THIN_MICRO_TILING */
1292 metadata
.tiling_info
= tiling_flags
;
1293 metadata
.size_metadata
= md
->size_metadata
;
1294 memcpy(metadata
.umd_metadata
, md
->metadata
, sizeof(md
->metadata
));
1296 amdgpu_bo_set_metadata(bo
->bo
, &metadata
);
1299 static struct pb_buffer
*
1300 amdgpu_bo_create(struct radeon_winsys
*rws
,
1303 enum radeon_bo_domain domain
,
1304 enum radeon_bo_flag flags
)
1306 struct amdgpu_winsys
*ws
= amdgpu_winsys(rws
);
1307 struct amdgpu_winsys_bo
*bo
;
1310 /* VRAM implies WC. This is not optional. */
1311 assert(!(domain
& RADEON_DOMAIN_VRAM
) || flags
& RADEON_FLAG_GTT_WC
);
1313 /* NO_CPU_ACCESS is valid with VRAM only. */
1314 assert(domain
== RADEON_DOMAIN_VRAM
|| !(flags
& RADEON_FLAG_NO_CPU_ACCESS
));
1316 /* Sparse buffers must have NO_CPU_ACCESS set. */
1317 assert(!(flags
& RADEON_FLAG_SPARSE
) || flags
& RADEON_FLAG_NO_CPU_ACCESS
);
1319 struct pb_slabs
*last_slab
= &ws
->bo_slabs
[NUM_SLAB_ALLOCATORS
- 1];
1320 unsigned max_slab_entry_size
= 1 << (last_slab
->min_order
+ last_slab
->num_orders
- 1);
1322 /* Sub-allocate small buffers from slabs. */
1323 if (!(flags
& (RADEON_FLAG_NO_SUBALLOC
| RADEON_FLAG_SPARSE
)) &&
1324 size
<= max_slab_entry_size
&&
1325 /* The alignment must be at most the size of the smallest slab entry or
1326 * the next power of two. */
1327 alignment
<= MAX2(1 << ws
->bo_slabs
[0].min_order
, util_next_power_of_two(size
))) {
1328 struct pb_slab_entry
*entry
;
1329 int heap
= radeon_get_heap_index(domain
, flags
);
1331 if (heap
< 0 || heap
>= RADEON_MAX_SLAB_HEAPS
)
1334 struct pb_slabs
*slabs
= get_slabs(ws
, size
);
1335 entry
= pb_slab_alloc(slabs
, size
, heap
);
1337 /* Clean up buffer managers and try again. */
1338 amdgpu_clean_up_buffer_managers(ws
);
1340 entry
= pb_slab_alloc(slabs
, size
, heap
);
1346 bo
= container_of(entry
, bo
, u
.slab
.entry
);
1348 pipe_reference_init(&bo
->base
.reference
, 1);
1354 if (flags
& RADEON_FLAG_SPARSE
) {
1355 assert(RADEON_SPARSE_PAGE_SIZE
% alignment
== 0);
1357 return amdgpu_bo_sparse_create(ws
, size
, domain
, flags
);
1360 /* This flag is irrelevant for the cache. */
1361 flags
&= ~RADEON_FLAG_NO_SUBALLOC
;
1363 /* Align size to page size. This is the minimum alignment for normal
1364 * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
1365 * like constant/uniform buffers, can benefit from better and more reuse.
1367 size
= align64(size
, ws
->info
.gart_page_size
);
1368 alignment
= align(alignment
, ws
->info
.gart_page_size
);
1370 bool use_reusable_pool
= flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
;
1372 if (use_reusable_pool
) {
1373 heap
= radeon_get_heap_index(domain
, flags
);
1374 assert(heap
>= 0 && heap
< RADEON_MAX_CACHED_HEAPS
);
1376 /* Get a buffer from the cache. */
1377 bo
= (struct amdgpu_winsys_bo
*)
1378 pb_cache_reclaim_buffer(&ws
->bo_cache
, size
, alignment
, 0, heap
);
1383 /* Create a new one. */
1384 bo
= amdgpu_create_bo(ws
, size
, alignment
, domain
, flags
, heap
);
1386 /* Clean up buffer managers and try again. */
1387 amdgpu_clean_up_buffer_managers(ws
);
1389 bo
= amdgpu_create_bo(ws
, size
, alignment
, domain
, flags
, heap
);
1394 bo
->u
.real
.use_reusable_pool
= use_reusable_pool
;
1398 static struct pb_buffer
*amdgpu_bo_from_handle(struct radeon_winsys
*rws
,
1399 struct winsys_handle
*whandle
,
1403 struct amdgpu_winsys
*ws
= amdgpu_winsys(rws
);
1404 struct amdgpu_winsys_bo
*bo
= NULL
;
1405 enum amdgpu_bo_handle_type type
;
1406 struct amdgpu_bo_import_result result
= {0};
1408 amdgpu_va_handle va_handle
= NULL
;
1409 struct amdgpu_bo_info info
= {0};
1410 enum radeon_bo_domain initial
= 0;
1413 switch (whandle
->type
) {
1414 case WINSYS_HANDLE_TYPE_SHARED
:
1415 type
= amdgpu_bo_handle_type_gem_flink_name
;
1417 case WINSYS_HANDLE_TYPE_FD
:
1418 type
= amdgpu_bo_handle_type_dma_buf_fd
;
1425 *stride
= whandle
->stride
;
1427 *offset
= whandle
->offset
;
1429 r
= amdgpu_bo_import(ws
->dev
, type
, whandle
->handle
, &result
);
1433 simple_mtx_lock(&ws
->bo_export_table_lock
);
1434 bo
= util_hash_table_get(ws
->bo_export_table
, result
.buf_handle
);
1436 /* If the amdgpu_winsys_bo instance already exists, bump the reference
1437 * counter and return it.
1440 p_atomic_inc(&bo
->base
.reference
.count
);
1441 simple_mtx_unlock(&ws
->bo_export_table_lock
);
1443 /* Release the buffer handle, because we don't need it anymore.
1444 * This function is returning an existing buffer, which has its own
1447 amdgpu_bo_free(result
.buf_handle
);
1451 /* Get initial domains. */
1452 r
= amdgpu_bo_query_info(result
.buf_handle
, &info
);
1456 r
= amdgpu_va_range_alloc(ws
->dev
, amdgpu_gpu_va_range_general
,
1457 result
.alloc_size
, 1 << 20, 0, &va
, &va_handle
,
1458 AMDGPU_VA_RANGE_HIGH
);
1462 bo
= CALLOC_STRUCT(amdgpu_winsys_bo
);
1466 r
= amdgpu_bo_va_op(result
.buf_handle
, 0, result
.alloc_size
, va
, 0, AMDGPU_VA_OP_MAP
);
1470 if (info
.preferred_heap
& AMDGPU_GEM_DOMAIN_VRAM
)
1471 initial
|= RADEON_DOMAIN_VRAM
;
1472 if (info
.preferred_heap
& AMDGPU_GEM_DOMAIN_GTT
)
1473 initial
|= RADEON_DOMAIN_GTT
;
1475 /* Initialize the structure. */
1476 simple_mtx_init(&bo
->lock
, mtx_plain
);
1477 pipe_reference_init(&bo
->base
.reference
, 1);
1478 bo
->base
.alignment
= info
.phys_alignment
;
1479 bo
->bo
= result
.buf_handle
;
1480 bo
->base
.size
= result
.alloc_size
;
1481 bo
->base
.vtbl
= &amdgpu_winsys_bo_vtbl
;
1484 bo
->u
.real
.va_handle
= va_handle
;
1485 bo
->initial_domain
= initial
;
1486 bo
->unique_id
= __sync_fetch_and_add(&ws
->next_bo_unique_id
, 1);
1487 bo
->is_shared
= true;
1489 if (bo
->initial_domain
& RADEON_DOMAIN_VRAM
)
1490 ws
->allocated_vram
+= align64(bo
->base
.size
, ws
->info
.gart_page_size
);
1491 else if (bo
->initial_domain
& RADEON_DOMAIN_GTT
)
1492 ws
->allocated_gtt
+= align64(bo
->base
.size
, ws
->info
.gart_page_size
);
1494 amdgpu_bo_export(bo
->bo
, amdgpu_bo_handle_type_kms
, &bo
->u
.real
.kms_handle
);
1496 amdgpu_add_buffer_to_global_list(bo
);
1498 util_hash_table_set(ws
->bo_export_table
, bo
->bo
, bo
);
1499 simple_mtx_unlock(&ws
->bo_export_table_lock
);
1504 simple_mtx_unlock(&ws
->bo_export_table_lock
);
1508 amdgpu_va_range_free(va_handle
);
1509 amdgpu_bo_free(result
.buf_handle
);
1513 static bool amdgpu_bo_get_handle(struct pb_buffer
*buffer
,
1514 unsigned stride
, unsigned offset
,
1515 unsigned slice_size
,
1516 struct winsys_handle
*whandle
)
1518 struct amdgpu_winsys_bo
*bo
= amdgpu_winsys_bo(buffer
);
1519 struct amdgpu_winsys
*ws
= bo
->ws
;
1520 enum amdgpu_bo_handle_type type
;
1523 /* Don't allow exports of slab entries and sparse buffers. */
1527 bo
->u
.real
.use_reusable_pool
= false;
1529 switch (whandle
->type
) {
1530 case WINSYS_HANDLE_TYPE_SHARED
:
1531 type
= amdgpu_bo_handle_type_gem_flink_name
;
1533 case WINSYS_HANDLE_TYPE_FD
:
1534 type
= amdgpu_bo_handle_type_dma_buf_fd
;
1536 case WINSYS_HANDLE_TYPE_KMS
:
1537 type
= amdgpu_bo_handle_type_kms
;
1543 r
= amdgpu_bo_export(bo
->bo
, type
, &whandle
->handle
);
1547 simple_mtx_lock(&ws
->bo_export_table_lock
);
1548 util_hash_table_set(ws
->bo_export_table
, bo
->bo
, bo
);
1549 simple_mtx_unlock(&ws
->bo_export_table_lock
);
1551 whandle
->stride
= stride
;
1552 whandle
->offset
= offset
;
1553 whandle
->offset
+= slice_size
* whandle
->layer
;
1554 bo
->is_shared
= true;
1558 static struct pb_buffer
*amdgpu_bo_from_ptr(struct radeon_winsys
*rws
,
1559 void *pointer
, uint64_t size
)
1561 struct amdgpu_winsys
*ws
= amdgpu_winsys(rws
);
1562 amdgpu_bo_handle buf_handle
;
1563 struct amdgpu_winsys_bo
*bo
;
1565 amdgpu_va_handle va_handle
;
1566 /* Avoid failure when the size is not page aligned */
1567 uint64_t aligned_size
= align64(size
, ws
->info
.gart_page_size
);
1569 bo
= CALLOC_STRUCT(amdgpu_winsys_bo
);
1573 if (amdgpu_create_bo_from_user_mem(ws
->dev
, pointer
,
1574 aligned_size
, &buf_handle
))
1577 if (amdgpu_va_range_alloc(ws
->dev
, amdgpu_gpu_va_range_general
,
1578 aligned_size
, 1 << 12, 0, &va
, &va_handle
,
1579 AMDGPU_VA_RANGE_HIGH
))
1580 goto error_va_alloc
;
1582 if (amdgpu_bo_va_op(buf_handle
, 0, aligned_size
, va
, 0, AMDGPU_VA_OP_MAP
))
1585 /* Initialize it. */
1586 bo
->is_user_ptr
= true;
1587 pipe_reference_init(&bo
->base
.reference
, 1);
1588 simple_mtx_init(&bo
->lock
, mtx_plain
);
1589 bo
->bo
= buf_handle
;
1590 bo
->base
.alignment
= 0;
1591 bo
->base
.size
= size
;
1592 bo
->base
.vtbl
= &amdgpu_winsys_bo_vtbl
;
1594 bo
->cpu_ptr
= pointer
;
1596 bo
->u
.real
.va_handle
= va_handle
;
1597 bo
->initial_domain
= RADEON_DOMAIN_GTT
;
1598 bo
->unique_id
= __sync_fetch_and_add(&ws
->next_bo_unique_id
, 1);
1600 ws
->allocated_gtt
+= aligned_size
;
1602 amdgpu_add_buffer_to_global_list(bo
);
1604 amdgpu_bo_export(bo
->bo
, amdgpu_bo_handle_type_kms
, &bo
->u
.real
.kms_handle
);
1606 return (struct pb_buffer
*)bo
;
1609 amdgpu_va_range_free(va_handle
);
1612 amdgpu_bo_free(buf_handle
);
1619 static bool amdgpu_bo_is_user_ptr(struct pb_buffer
*buf
)
1621 return ((struct amdgpu_winsys_bo
*)buf
)->is_user_ptr
;
1624 static bool amdgpu_bo_is_suballocated(struct pb_buffer
*buf
)
1626 struct amdgpu_winsys_bo
*bo
= (struct amdgpu_winsys_bo
*)buf
;
1628 return !bo
->bo
&& !bo
->sparse
;
1631 static uint64_t amdgpu_bo_get_va(struct pb_buffer
*buf
)
1633 return ((struct amdgpu_winsys_bo
*)buf
)->va
;
1636 void amdgpu_bo_init_functions(struct amdgpu_winsys
*ws
)
1638 ws
->base
.buffer_set_metadata
= amdgpu_buffer_set_metadata
;
1639 ws
->base
.buffer_get_metadata
= amdgpu_buffer_get_metadata
;
1640 ws
->base
.buffer_map
= amdgpu_bo_map
;
1641 ws
->base
.buffer_unmap
= amdgpu_bo_unmap
;
1642 ws
->base
.buffer_wait
= amdgpu_bo_wait
;
1643 ws
->base
.buffer_create
= amdgpu_bo_create
;
1644 ws
->base
.buffer_from_handle
= amdgpu_bo_from_handle
;
1645 ws
->base
.buffer_from_ptr
= amdgpu_bo_from_ptr
;
1646 ws
->base
.buffer_is_user_ptr
= amdgpu_bo_is_user_ptr
;
1647 ws
->base
.buffer_is_suballocated
= amdgpu_bo_is_suballocated
;
1648 ws
->base
.buffer_get_handle
= amdgpu_bo_get_handle
;
1649 ws
->base
.buffer_commit
= amdgpu_bo_sparse_commit
;
1650 ws
->base
.buffer_get_virtual_address
= amdgpu_bo_get_va
;
1651 ws
->base
.buffer_get_initial_domain
= amdgpu_bo_get_initial_domain
;