2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include "anv_private.h"
32 #include "genxml/gen8_pack.h"
34 #include "util/debug.h"
36 /** \file anv_batch_chain.c
38 * This file contains functions related to anv_cmd_buffer as a data
39 * structure. This involves everything required to create and destroy
40 * the actual batch buffers as well as link them together and handle
41 * relocations and surface state. It specifically does *not* contain any
42 * handling of actual vkCmd calls beyond vkCmdExecuteCommands.
45 /*-----------------------------------------------------------------------*
46 * Functions related to anv_reloc_list
47 *-----------------------------------------------------------------------*/
50 anv_reloc_list_init_clone(struct anv_reloc_list
*list
,
51 const VkAllocationCallbacks
*alloc
,
52 const struct anv_reloc_list
*other_list
)
55 list
->num_relocs
= other_list
->num_relocs
;
56 list
->array_length
= other_list
->array_length
;
59 list
->array_length
= 256;
63 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->relocs
), 8,
64 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
66 if (list
->relocs
== NULL
)
67 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
70 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->reloc_bos
), 8,
71 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
73 if (list
->reloc_bos
== NULL
) {
74 vk_free(alloc
, list
->relocs
);
75 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
79 memcpy(list
->relocs
, other_list
->relocs
,
80 list
->array_length
* sizeof(*list
->relocs
));
81 memcpy(list
->reloc_bos
, other_list
->reloc_bos
,
82 list
->array_length
* sizeof(*list
->reloc_bos
));
89 anv_reloc_list_init(struct anv_reloc_list
*list
,
90 const VkAllocationCallbacks
*alloc
)
92 return anv_reloc_list_init_clone(list
, alloc
, NULL
);
96 anv_reloc_list_finish(struct anv_reloc_list
*list
,
97 const VkAllocationCallbacks
*alloc
)
99 vk_free(alloc
, list
->relocs
);
100 vk_free(alloc
, list
->reloc_bos
);
104 anv_reloc_list_grow(struct anv_reloc_list
*list
,
105 const VkAllocationCallbacks
*alloc
,
106 size_t num_additional_relocs
)
108 if (list
->num_relocs
+ num_additional_relocs
<= list
->array_length
)
111 size_t new_length
= list
->array_length
* 2;
112 while (new_length
< list
->num_relocs
+ num_additional_relocs
)
115 struct drm_i915_gem_relocation_entry
*new_relocs
=
116 vk_alloc(alloc
, new_length
* sizeof(*list
->relocs
), 8,
117 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
118 if (new_relocs
== NULL
)
119 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
121 struct anv_bo
**new_reloc_bos
=
122 vk_alloc(alloc
, new_length
* sizeof(*list
->reloc_bos
), 8,
123 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
124 if (new_reloc_bos
== NULL
) {
125 vk_free(alloc
, new_relocs
);
126 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
129 memcpy(new_relocs
, list
->relocs
, list
->num_relocs
* sizeof(*list
->relocs
));
130 memcpy(new_reloc_bos
, list
->reloc_bos
,
131 list
->num_relocs
* sizeof(*list
->reloc_bos
));
133 vk_free(alloc
, list
->relocs
);
134 vk_free(alloc
, list
->reloc_bos
);
136 list
->array_length
= new_length
;
137 list
->relocs
= new_relocs
;
138 list
->reloc_bos
= new_reloc_bos
;
144 anv_reloc_list_add(struct anv_reloc_list
*list
,
145 const VkAllocationCallbacks
*alloc
,
146 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
)
148 struct drm_i915_gem_relocation_entry
*entry
;
151 const uint32_t domain
=
152 target_bo
->is_winsys_bo
? I915_GEM_DOMAIN_RENDER
: 0;
154 anv_reloc_list_grow(list
, alloc
, 1);
155 /* TODO: Handle failure */
157 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
158 index
= list
->num_relocs
++;
159 list
->reloc_bos
[index
] = target_bo
;
160 entry
= &list
->relocs
[index
];
161 entry
->target_handle
= target_bo
->gem_handle
;
162 entry
->delta
= delta
;
163 entry
->offset
= offset
;
164 entry
->presumed_offset
= target_bo
->offset
;
165 entry
->read_domains
= domain
;
166 entry
->write_domain
= domain
;
167 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
169 return target_bo
->offset
+ delta
;
173 anv_reloc_list_append(struct anv_reloc_list
*list
,
174 const VkAllocationCallbacks
*alloc
,
175 struct anv_reloc_list
*other
, uint32_t offset
)
177 anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
178 /* TODO: Handle failure */
180 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
181 other
->num_relocs
* sizeof(other
->relocs
[0]));
182 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
183 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
185 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
186 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
188 list
->num_relocs
+= other
->num_relocs
;
191 /*-----------------------------------------------------------------------*
192 * Functions related to anv_batch
193 *-----------------------------------------------------------------------*/
196 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
198 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
199 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
200 if (result
!= VK_SUCCESS
) {
201 anv_batch_set_error(batch
, result
);
206 void *p
= batch
->next
;
208 batch
->next
+= num_dwords
* 4;
209 assert(batch
->next
<= batch
->end
);
215 anv_batch_emit_reloc(struct anv_batch
*batch
,
216 void *location
, struct anv_bo
*bo
, uint32_t delta
)
218 return anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
219 location
- batch
->start
, bo
, delta
);
223 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
225 uint32_t size
, offset
;
227 size
= other
->next
- other
->start
;
228 assert(size
% 4 == 0);
230 if (batch
->next
+ size
> batch
->end
) {
231 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
232 if (result
!= VK_SUCCESS
) {
233 anv_batch_set_error(batch
, result
);
238 assert(batch
->next
+ size
<= batch
->end
);
240 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
241 memcpy(batch
->next
, other
->start
, size
);
243 offset
= batch
->next
- batch
->start
;
244 anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
245 other
->relocs
, offset
);
250 /*-----------------------------------------------------------------------*
251 * Functions related to anv_batch_bo
252 *-----------------------------------------------------------------------*/
255 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
256 struct anv_batch_bo
**bbo_out
)
260 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
261 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
263 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
265 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
266 ANV_CMD_BUFFER_BATCH_SIZE
);
267 if (result
!= VK_SUCCESS
)
270 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
271 if (result
!= VK_SUCCESS
)
279 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
281 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
287 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
288 const struct anv_batch_bo
*other_bbo
,
289 struct anv_batch_bo
**bbo_out
)
293 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
294 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
296 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
298 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
300 if (result
!= VK_SUCCESS
)
303 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
305 if (result
!= VK_SUCCESS
)
308 bbo
->length
= other_bbo
->length
;
309 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
316 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
318 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
324 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
325 size_t batch_padding
)
327 batch
->next
= batch
->start
= bbo
->bo
.map
;
328 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
329 batch
->relocs
= &bbo
->relocs
;
330 bbo
->relocs
.num_relocs
= 0;
334 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
335 size_t batch_padding
)
337 batch
->start
= bbo
->bo
.map
;
338 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
339 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
340 batch
->relocs
= &bbo
->relocs
;
344 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
346 assert(batch
->start
== bbo
->bo
.map
);
347 bbo
->length
= batch
->next
- batch
->start
;
348 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
352 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
353 struct anv_batch
*batch
, size_t aditional
,
354 size_t batch_padding
)
356 assert(batch
->start
== bbo
->bo
.map
);
357 bbo
->length
= batch
->next
- batch
->start
;
359 size_t new_size
= bbo
->bo
.size
;
360 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
363 if (new_size
== bbo
->bo
.size
)
366 struct anv_bo new_bo
;
367 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
369 if (result
!= VK_SUCCESS
)
372 memcpy(new_bo
.map
, bbo
->bo
.map
, bbo
->length
);
374 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
377 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
383 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
384 struct anv_cmd_buffer
*cmd_buffer
)
386 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
387 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
388 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
392 anv_batch_bo_list_clone(const struct list_head
*list
,
393 struct anv_cmd_buffer
*cmd_buffer
,
394 struct list_head
*new_list
)
396 VkResult result
= VK_SUCCESS
;
398 list_inithead(new_list
);
400 struct anv_batch_bo
*prev_bbo
= NULL
;
401 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
402 struct anv_batch_bo
*new_bbo
= NULL
;
403 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
404 if (result
!= VK_SUCCESS
)
406 list_addtail(&new_bbo
->link
, new_list
);
409 /* As we clone this list of batch_bo's, they chain one to the
410 * other using MI_BATCH_BUFFER_START commands. We need to fix up
411 * those relocations as we go. Fortunately, this is pretty easy
412 * as it will always be the last relocation in the list.
414 uint32_t last_idx
= prev_bbo
->relocs
.num_relocs
- 1;
415 assert(prev_bbo
->relocs
.reloc_bos
[last_idx
] == &bbo
->bo
);
416 prev_bbo
->relocs
.reloc_bos
[last_idx
] = &new_bbo
->bo
;
422 if (result
!= VK_SUCCESS
) {
423 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
424 anv_batch_bo_destroy(bbo
, cmd_buffer
);
430 /*-----------------------------------------------------------------------*
431 * Functions related to anv_batch_bo
432 *-----------------------------------------------------------------------*/
434 static inline struct anv_batch_bo
*
435 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
437 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
441 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
443 return (struct anv_address
) {
444 .bo
= &cmd_buffer
->device
->surface_state_block_pool
.bo
,
445 .offset
= *(int32_t *)u_vector_head(&cmd_buffer
->bt_blocks
),
450 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
451 struct anv_bo
*bo
, uint32_t offset
)
453 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
454 * offsets. The high 16 bits are in the last dword, so we can use the gen8
455 * version in either case, as long as we set the instruction length in the
456 * header accordingly. This means that we always emit three dwords here
457 * and all the padding and adjustment we do in this file works for all
461 #define GEN7_MI_BATCH_BUFFER_START_length 2
462 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
464 const uint32_t gen7_length
=
465 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
466 const uint32_t gen8_length
=
467 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
469 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
470 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
471 gen7_length
: gen8_length
;
472 bbs
._2ndLevelBatchBuffer
= _1stlevelbatch
;
473 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
474 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
479 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
480 struct anv_batch_bo
*bbo
)
482 struct anv_batch
*batch
= &cmd_buffer
->batch
;
483 struct anv_batch_bo
*current_bbo
=
484 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
486 /* We set the end of the batch a little short so we would be sure we
487 * have room for the chaining command. Since we're about to emit the
488 * chaining command, let's set it back where it should go.
490 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
491 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
493 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
495 anv_batch_bo_finish(current_bbo
, batch
);
499 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
501 struct anv_cmd_buffer
*cmd_buffer
= _data
;
502 struct anv_batch_bo
*new_bbo
;
504 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
505 if (result
!= VK_SUCCESS
)
508 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
509 if (seen_bbo
== NULL
) {
510 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
511 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
515 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
517 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
519 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
525 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
527 struct anv_cmd_buffer
*cmd_buffer
= _data
;
528 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
530 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
531 GEN8_MI_BATCH_BUFFER_START_length
* 4);
536 /** Allocate a binding table
538 * This function allocates a binding table. This is a bit more complicated
539 * than one would think due to a combination of Vulkan driver design and some
540 * unfortunate hardware restrictions.
542 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
543 * the binding table pointer which means that all binding tables need to live
544 * in the bottom 64k of surface state base address. The way the GL driver has
545 * classically dealt with this restriction is to emit all surface states
546 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
547 * isn't really an option in Vulkan for a couple of reasons:
549 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
550 * to live in their own buffer and we have to be able to re-emit
551 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
552 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
553 * (it's not that hard to hit 64k of just binding tables), we allocate
554 * surface state objects up-front when VkImageView is created. In order
555 * for this to work, surface state objects need to be allocated from a
558 * 2) We tried to design the surface state system in such a way that it's
559 * already ready for bindless texturing. The way bindless texturing works
560 * on our hardware is that you have a big pool of surface state objects
561 * (with its own state base address) and the bindless handles are simply
562 * offsets into that pool. With the architecture we chose, we already
563 * have that pool and it's exactly the same pool that we use for regular
564 * surface states so we should already be ready for bindless.
566 * 3) For render targets, we need to be able to fill out the surface states
567 * later in vkBeginRenderPass so that we can assign clear colors
568 * correctly. One way to do this would be to just create the surface
569 * state data and then repeatedly copy it into the surface state BO every
570 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
571 * rather annoying and just being able to allocate them up-front and
572 * re-use them for the entire render pass.
574 * While none of these are technically blockers for emitting state on the fly
575 * like we do in GL, the ability to have a single surface state pool is
576 * simplifies things greatly. Unfortunately, it comes at a cost...
578 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
579 * place the binding tables just anywhere in surface state base address.
580 * Because 64k isn't a whole lot of space, we can't simply restrict the
581 * surface state buffer to 64k, we have to be more clever. The solution we've
582 * chosen is to have a block pool with a maximum size of 2G that starts at
583 * zero and grows in both directions. All surface states are allocated from
584 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
585 * binding tables from the bottom of the pool (negative offsets). Every time
586 * we allocate a new binding table block, we set surface state base address to
587 * point to the bottom of the binding table block. This way all of the
588 * binding tables in the block are in the bottom 64k of surface state base
589 * address. When we fill out the binding table, we add the distance between
590 * the bottom of our binding table block and zero of the block pool to the
591 * surface state offsets so that they are correct relative to out new surface
592 * state base address at the bottom of the binding table block.
594 * \see adjust_relocations_from_block_pool()
595 * \see adjust_relocations_too_block_pool()
597 * \param[in] entries The number of surface state entries the binding
598 * table should be able to hold.
600 * \param[out] state_offset The offset surface surface state base address
601 * where the surface states live. This must be
602 * added to the surface state offset when it is
603 * written into the binding table entry.
605 * \return An anv_state representing the binding table
608 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
609 uint32_t entries
, uint32_t *state_offset
)
611 struct anv_block_pool
*block_pool
=
612 &cmd_buffer
->device
->surface_state_block_pool
;
613 int32_t *bt_block
= u_vector_head(&cmd_buffer
->bt_blocks
);
614 struct anv_state state
;
616 state
.alloc_size
= align_u32(entries
* 4, 32);
618 if (cmd_buffer
->bt_next
+ state
.alloc_size
> block_pool
->block_size
)
619 return (struct anv_state
) { 0 };
621 state
.offset
= cmd_buffer
->bt_next
;
622 state
.map
= block_pool
->map
+ *bt_block
+ state
.offset
;
624 cmd_buffer
->bt_next
+= state
.alloc_size
;
626 assert(*bt_block
< 0);
627 *state_offset
= -(*bt_block
);
633 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
635 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
636 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
637 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
641 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
642 uint32_t size
, uint32_t alignment
)
644 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
649 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
651 struct anv_block_pool
*block_pool
=
652 &cmd_buffer
->device
->surface_state_block_pool
;
654 int32_t *offset
= u_vector_add(&cmd_buffer
->bt_blocks
);
656 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
658 *offset
= anv_block_pool_alloc_back(block_pool
);
659 cmd_buffer
->bt_next
= 0;
665 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
667 struct anv_batch_bo
*batch_bo
;
670 list_inithead(&cmd_buffer
->batch_bos
);
672 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
673 if (result
!= VK_SUCCESS
)
676 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
678 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
679 cmd_buffer
->batch
.user_data
= cmd_buffer
;
681 if (cmd_buffer
->device
->can_chain_batches
) {
682 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
684 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
687 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
688 GEN8_MI_BATCH_BUFFER_START_length
* 4);
690 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
691 sizeof(struct anv_bo
*),
692 8 * sizeof(struct anv_bo
*));
696 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
698 success
= u_vector_init(&cmd_buffer
->bt_blocks
, sizeof(int32_t),
699 8 * sizeof(int32_t));
703 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
704 &cmd_buffer
->pool
->alloc
);
705 if (result
!= VK_SUCCESS
)
707 cmd_buffer
->last_ss_pool_center
= 0;
709 anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
714 u_vector_finish(&cmd_buffer
->bt_blocks
);
716 u_vector_finish(&cmd_buffer
->seen_bbos
);
718 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
724 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
727 u_vector_foreach(bt_block
, &cmd_buffer
->bt_blocks
) {
728 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
731 u_vector_finish(&cmd_buffer
->bt_blocks
);
733 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
735 u_vector_finish(&cmd_buffer
->seen_bbos
);
737 /* Destroy all of the batch buffers */
738 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
739 &cmd_buffer
->batch_bos
, link
) {
740 anv_batch_bo_destroy(bbo
, cmd_buffer
);
745 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
747 /* Delete all but the first batch bo */
748 assert(!list_empty(&cmd_buffer
->batch_bos
));
749 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
750 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
751 list_del(&bbo
->link
);
752 anv_batch_bo_destroy(bbo
, cmd_buffer
);
754 assert(!list_empty(&cmd_buffer
->batch_bos
));
756 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
758 GEN8_MI_BATCH_BUFFER_START_length
* 4);
760 while (u_vector_length(&cmd_buffer
->bt_blocks
) > 1) {
761 int32_t *bt_block
= u_vector_remove(&cmd_buffer
->bt_blocks
);
762 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
765 assert(u_vector_length(&cmd_buffer
->bt_blocks
) == 1);
766 cmd_buffer
->bt_next
= 0;
768 cmd_buffer
->surface_relocs
.num_relocs
= 0;
769 cmd_buffer
->last_ss_pool_center
= 0;
771 /* Reset the list of seen buffers */
772 cmd_buffer
->seen_bbos
.head
= 0;
773 cmd_buffer
->seen_bbos
.tail
= 0;
775 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
776 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
780 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
782 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
784 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
785 /* When we start a batch buffer, we subtract a certain amount of
786 * padding from the end to ensure that we always have room to emit a
787 * BATCH_BUFFER_START to chain to the next BO. We need to remove
788 * that padding before we end the batch; otherwise, we may end up
789 * with our BATCH_BUFFER_END in another BO.
791 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
792 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
794 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
796 /* Round batch up to an even number of dwords. */
797 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
798 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
800 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
803 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
805 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
) {
806 /* If this is a secondary command buffer, we need to determine the
807 * mode in which it will be executed with vkExecuteCommands. We
808 * determine this statically here so that this stays in sync with the
809 * actual ExecuteCommands implementation.
811 if (!cmd_buffer
->device
->can_chain_batches
) {
812 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
813 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
814 (batch_bo
->length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
815 /* If the secondary has exactly one batch buffer in its list *and*
816 * that batch buffer is less than half of the maximum size, we're
817 * probably better of simply copying it into our batch.
819 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
820 } else if (!(cmd_buffer
->usage_flags
&
821 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
822 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
824 /* When we chain, we need to add an MI_BATCH_BUFFER_START command
825 * with its relocation. In order to handle this we'll increment here
826 * so we can unconditionally decrement right before adding the
827 * MI_BATCH_BUFFER_START command.
829 batch_bo
->relocs
.num_relocs
++;
830 cmd_buffer
->batch
.next
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
832 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
837 static inline VkResult
838 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
839 struct list_head
*list
)
841 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
842 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
844 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
853 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
854 struct anv_cmd_buffer
*secondary
)
856 switch (secondary
->exec_mode
) {
857 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
858 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
860 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
861 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
862 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
863 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
864 GEN8_MI_BATCH_BUFFER_START_length
* 4);
865 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
868 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
869 struct anv_batch_bo
*first_bbo
=
870 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
871 struct anv_batch_bo
*last_bbo
=
872 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
874 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
876 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
877 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
878 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
879 const uint32_t inst_size
= GEN8_MI_BATCH_BUFFER_START_length
* 4;
881 /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we
882 * can emit a new command and relocation for the current splice. In
883 * order to handle the initial-use case, we incremented next and
884 * num_relocs in end_batch_buffer() so we can alyways just subtract
887 last_bbo
->relocs
.num_relocs
--;
888 secondary
->batch
.next
-= inst_size
;
889 emit_batch_buffer_start(secondary
, &this_bbo
->bo
, offset
);
890 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
892 /* After patching up the secondary buffer, we need to clflush the
893 * modified instruction in case we're on a !llc platform. We use a
894 * little loop to handle the case where the instruction crosses a cache
897 if (!primary
->device
->info
.has_llc
) {
898 void *inst
= secondary
->batch
.next
- inst_size
;
899 void *p
= (void *) (((uintptr_t) inst
) & ~CACHELINE_MASK
);
900 __builtin_ia32_mfence();
901 while (p
< secondary
->batch
.next
) {
902 __builtin_ia32_clflush(p
);
908 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
909 struct list_head copy_list
;
910 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
913 if (result
!= VK_SUCCESS
)
916 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
918 struct anv_batch_bo
*first_bbo
=
919 list_first_entry(©_list
, struct anv_batch_bo
, link
);
920 struct anv_batch_bo
*last_bbo
=
921 list_last_entry(©_list
, struct anv_batch_bo
, link
);
923 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
925 list_splicetail(©_list
, &primary
->batch_bos
);
927 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
928 GEN8_MI_BATCH_BUFFER_START_length
* 4);
932 assert(!"Invalid execution mode");
935 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
936 &secondary
->surface_relocs
, 0);
940 struct drm_i915_gem_execbuffer2 execbuf
;
942 struct drm_i915_gem_exec_object2
* objects
;
944 struct anv_bo
** bos
;
946 /* Allocated length of the 'objects' and 'bos' arrays */
947 uint32_t array_length
;
951 anv_execbuf_init(struct anv_execbuf
*exec
)
953 memset(exec
, 0, sizeof(*exec
));
957 anv_execbuf_finish(struct anv_execbuf
*exec
,
958 const VkAllocationCallbacks
*alloc
)
960 vk_free(alloc
, exec
->objects
);
961 vk_free(alloc
, exec
->bos
);
965 anv_execbuf_add_bo(struct anv_execbuf
*exec
,
967 struct anv_reloc_list
*relocs
,
968 const VkAllocationCallbacks
*alloc
)
970 struct drm_i915_gem_exec_object2
*obj
= NULL
;
972 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
973 obj
= &exec
->objects
[bo
->index
];
976 /* We've never seen this one before. Add it to the list and assign
977 * an id that we can use later.
979 if (exec
->bo_count
>= exec
->array_length
) {
980 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
982 struct drm_i915_gem_exec_object2
*new_objects
=
983 vk_alloc(alloc
, new_len
* sizeof(*new_objects
),
984 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
985 if (new_objects
== NULL
)
986 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
988 struct anv_bo
**new_bos
=
989 vk_alloc(alloc
, new_len
* sizeof(*new_bos
),
990 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
991 if (new_bos
== NULL
) {
992 vk_free(alloc
, new_objects
);
993 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
997 memcpy(new_objects
, exec
->objects
,
998 exec
->bo_count
* sizeof(*new_objects
));
999 memcpy(new_bos
, exec
->bos
,
1000 exec
->bo_count
* sizeof(*new_bos
));
1003 vk_free(alloc
, exec
->objects
);
1004 vk_free(alloc
, exec
->bos
);
1006 exec
->objects
= new_objects
;
1007 exec
->bos
= new_bos
;
1008 exec
->array_length
= new_len
;
1011 assert(exec
->bo_count
< exec
->array_length
);
1013 bo
->index
= exec
->bo_count
++;
1014 obj
= &exec
->objects
[bo
->index
];
1015 exec
->bos
[bo
->index
] = bo
;
1017 obj
->handle
= bo
->gem_handle
;
1018 obj
->relocation_count
= 0;
1019 obj
->relocs_ptr
= 0;
1021 obj
->offset
= bo
->offset
;
1022 obj
->flags
= bo
->is_winsys_bo
? EXEC_OBJECT_WRITE
: 0;
1027 if (relocs
!= NULL
&& obj
->relocation_count
== 0) {
1028 /* This is the first time we've ever seen a list of relocations for
1029 * this BO. Go ahead and set the relocations and then walk the list
1030 * of relocations and add them all.
1032 obj
->relocation_count
= relocs
->num_relocs
;
1033 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1035 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1036 /* A quick sanity check on relocations */
1037 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1038 anv_execbuf_add_bo(exec
, relocs
->reloc_bos
[i
], NULL
, alloc
);
1046 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1047 struct anv_reloc_list
*list
)
1049 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1050 list
->relocs
[i
].target_handle
= list
->reloc_bos
[i
]->index
;
1054 write_reloc(const struct anv_device
*device
, void *p
, uint64_t v
, bool flush
)
1056 unsigned reloc_size
= 0;
1057 if (device
->info
.gen
>= 8) {
1058 /* From the Broadwell PRM Vol. 2a, MI_LOAD_REGISTER_MEM::MemoryAddress:
1060 * "This field specifies the address of the memory location where the
1061 * register value specified in the DWord above will read from. The
1062 * address specifies the DWord location of the data. Range =
1063 * GraphicsVirtualAddress[63:2] for a DWord register GraphicsAddress
1064 * [63:48] are ignored by the HW and assumed to be in correct
1065 * canonical form [63:48] == [47]."
1067 const int shift
= 63 - 47;
1068 reloc_size
= sizeof(uint64_t);
1069 *(uint64_t *)p
= (((int64_t)v
) << shift
) >> shift
;
1071 reloc_size
= sizeof(uint32_t);
1075 if (flush
&& !device
->info
.has_llc
)
1076 anv_flush_range(p
, reloc_size
);
1080 adjust_relocations_from_state_pool(struct anv_block_pool
*pool
,
1081 struct anv_reloc_list
*relocs
,
1082 uint32_t last_pool_center_bo_offset
)
1084 assert(last_pool_center_bo_offset
<= pool
->center_bo_offset
);
1085 uint32_t delta
= pool
->center_bo_offset
- last_pool_center_bo_offset
;
1087 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1088 /* All of the relocations from this block pool to other BO's should
1089 * have been emitted relative to the surface block pool center. We
1090 * need to add the center offset to make them relative to the
1091 * beginning of the actual GEM bo.
1093 relocs
->relocs
[i
].offset
+= delta
;
1098 adjust_relocations_to_state_pool(struct anv_block_pool
*pool
,
1099 struct anv_bo
*from_bo
,
1100 struct anv_reloc_list
*relocs
,
1101 uint32_t last_pool_center_bo_offset
)
1103 assert(last_pool_center_bo_offset
<= pool
->center_bo_offset
);
1104 uint32_t delta
= pool
->center_bo_offset
- last_pool_center_bo_offset
;
1106 /* When we initially emit relocations into a block pool, we don't
1107 * actually know what the final center_bo_offset will be so we just emit
1108 * it as if center_bo_offset == 0. Now that we know what the center
1109 * offset is, we need to walk the list of relocations and adjust any
1110 * relocations that point to the pool bo with the correct offset.
1112 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1113 if (relocs
->reloc_bos
[i
] == &pool
->bo
) {
1114 /* Adjust the delta value in the relocation to correctly
1115 * correspond to the new delta. Initially, this value may have
1116 * been negative (if treated as unsigned), but we trust in
1117 * uint32_t roll-over to fix that for us at this point.
1119 relocs
->relocs
[i
].delta
+= delta
;
1121 /* Since the delta has changed, we need to update the actual
1122 * relocated value with the new presumed value. This function
1123 * should only be called on batch buffers, so we know it isn't in
1124 * use by the GPU at the moment.
1126 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1127 write_reloc(pool
->device
, from_bo
->map
+ relocs
->relocs
[i
].offset
,
1128 relocs
->relocs
[i
].presumed_offset
+
1129 relocs
->relocs
[i
].delta
, false);
1135 anv_reloc_list_apply(struct anv_device
*device
,
1136 struct anv_reloc_list
*list
,
1138 bool always_relocate
)
1140 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1141 struct anv_bo
*target_bo
= list
->reloc_bos
[i
];
1142 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1146 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1147 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1148 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1153 * This function applies the relocation for a command buffer and writes the
1154 * actual addresses into the buffers as per what we were told by the kernel on
1155 * the previous execbuf2 call. This should be safe to do because, for each
1156 * relocated address, we have two cases:
1158 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1159 * not in use by the GPU so updating the address is 100% ok. It won't be
1160 * in-use by the GPU (from our context) again until the next execbuf2
1161 * happens. If the kernel decides to move it in the next execbuf2, it
1162 * will have to do the relocations itself, but that's ok because it should
1163 * have all of the information needed to do so.
1165 * 2) The target BO is active (as seen by the kernel). In this case, it
1166 * hasn't moved since the last execbuffer2 call because GTT shuffling
1167 * *only* happens when the BO is idle. (From our perspective, it only
1168 * happens inside the execbuffer2 ioctl, but the shuffling may be
1169 * triggered by another ioctl, with full-ppgtt this is limited to only
1170 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1171 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1172 * address and the relocated value we are writing into the BO will be the
1173 * same as the value that is already there.
1175 * There is also a possibility that the target BO is active but the exact
1176 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1177 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1178 * may be stale but it's still safe to write the relocation because that
1179 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1180 * won't be until the next execbuf2 call.
1182 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1183 * need to bother. We want to do this because the surface state buffer is
1184 * used by every command buffer so, if the kernel does the relocations, it
1185 * will always be busy and the kernel will always stall. This is also
1186 * probably the fastest mechanism for doing relocations since the kernel would
1187 * have to make a full copy of all the relocations lists.
1190 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1191 struct anv_execbuf
*exec
)
1193 static int userspace_relocs
= -1;
1194 if (userspace_relocs
< 0)
1195 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1196 if (!userspace_relocs
)
1199 /* First, we have to check to see whether or not we can even do the
1200 * relocation. New buffers which have never been submitted to the kernel
1201 * don't have a valid offset so we need to let the kernel do relocations so
1202 * that we can get offsets for them. On future execbuf2 calls, those
1203 * buffers will have offsets and we will be able to skip relocating.
1204 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1206 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1207 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1211 /* Since surface states are shared between command buffers and we don't
1212 * know what order they will be submitted to the kernel, we don't know
1213 * what address is actually written in the surface state object at any
1214 * given time. The only option is to always relocate them.
1216 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1217 &cmd_buffer
->device
->surface_state_block_pool
.bo
,
1218 true /* always relocate surface states */);
1220 /* Since we own all of the batch buffers, we know what values are stored
1221 * in the relocated addresses and only have to update them if the offsets
1224 struct anv_batch_bo
**bbo
;
1225 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1226 anv_reloc_list_apply(cmd_buffer
->device
,
1227 &(*bbo
)->relocs
, &(*bbo
)->bo
, false);
1230 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1231 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1237 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1238 struct anv_cmd_buffer
*cmd_buffer
)
1240 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1241 struct anv_block_pool
*ss_pool
=
1242 &cmd_buffer
->device
->surface_state_block_pool
;
1244 struct anv_execbuf execbuf
;
1245 anv_execbuf_init(&execbuf
);
1247 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1248 cmd_buffer
->last_ss_pool_center
);
1249 anv_execbuf_add_bo(&execbuf
, &ss_pool
->bo
, &cmd_buffer
->surface_relocs
,
1250 &cmd_buffer
->pool
->alloc
);
1252 /* First, we walk over all of the bos we've seen and add them and their
1253 * relocations to the validate list.
1255 struct anv_batch_bo
**bbo
;
1256 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1257 adjust_relocations_to_state_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1258 cmd_buffer
->last_ss_pool_center
);
1260 anv_execbuf_add_bo(&execbuf
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1261 &cmd_buffer
->pool
->alloc
);
1264 /* Now that we've adjusted all of the surface state relocations, we need to
1265 * record the surface state pool center so future executions of the command
1266 * buffer can adjust correctly.
1268 cmd_buffer
->last_ss_pool_center
= ss_pool
->center_bo_offset
;
1270 struct anv_batch_bo
*first_batch_bo
=
1271 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1273 /* The kernel requires that the last entry in the validation list be the
1274 * batch buffer to execute. We can simply swap the element
1275 * corresponding to the first batch_bo in the chain with the last
1276 * element in the list.
1278 if (first_batch_bo
->bo
.index
!= execbuf
.bo_count
- 1) {
1279 uint32_t idx
= first_batch_bo
->bo
.index
;
1280 uint32_t last_idx
= execbuf
.bo_count
- 1;
1282 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
.objects
[idx
];
1283 assert(execbuf
.bos
[idx
] == &first_batch_bo
->bo
);
1285 execbuf
.objects
[idx
] = execbuf
.objects
[last_idx
];
1286 execbuf
.bos
[idx
] = execbuf
.bos
[last_idx
];
1287 execbuf
.bos
[idx
]->index
= idx
;
1289 execbuf
.objects
[last_idx
] = tmp_obj
;
1290 execbuf
.bos
[last_idx
] = &first_batch_bo
->bo
;
1291 first_batch_bo
->bo
.index
= last_idx
;
1294 /* Now we go through and fixup all of the relocation lists to point to
1295 * the correct indices in the object array. We have to do this after we
1296 * reorder the list above as some of the indices may have changed.
1298 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1299 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1301 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1303 if (!cmd_buffer
->device
->info
.has_llc
) {
1304 __builtin_ia32_mfence();
1305 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1306 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1307 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1311 execbuf
.execbuf
= (struct drm_i915_gem_execbuffer2
) {
1312 .buffers_ptr
= (uintptr_t) execbuf
.objects
,
1313 .buffer_count
= execbuf
.bo_count
,
1314 .batch_start_offset
= 0,
1315 .batch_len
= batch
->next
- batch
->start
,
1320 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
|
1321 I915_EXEC_CONSTANTS_REL_GENERAL
,
1322 .rsvd1
= cmd_buffer
->device
->context_id
,
1326 if (relocate_cmd_buffer(cmd_buffer
, &execbuf
)) {
1327 /* If we were able to successfully relocate everything, tell the kernel
1328 * that it can skip doing relocations. The requirement for using
1331 * 1) The addresses written in the objects must match the corresponding
1332 * reloc.presumed_offset which in turn must match the corresponding
1333 * execobject.offset.
1335 * 2) To avoid stalling, execobject.offset should match the current
1336 * address of that object within the active context.
1338 * In order to satisfy all of the invariants that make userspace
1339 * relocations to be safe (see relocate_cmd_buffer()), we need to
1340 * further ensure that the addresses we use match those used by the
1341 * kernel for the most recent execbuf2.
1343 * The kernel may still choose to do relocations anyway if something has
1344 * moved in the GTT. In this case, the relocation list still needs to be
1345 * valid. All relocations on the batch buffers are already valid and
1346 * kept up-to-date. For surface state relocations, by applying the
1347 * relocations in relocate_cmd_buffer, we ensured that the address in
1348 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1349 * safe for the kernel to relocate them as needed.
1351 execbuf
.execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1353 /* In the case where we fall back to doing kernel relocations, we need
1354 * to ensure that the relocation list is valid. All relocations on the
1355 * batch buffers are already valid and kept up-to-date. Since surface
1356 * states are shared between command buffers and we don't know what
1357 * order they will be submitted to the kernel, we don't know what
1358 * address is actually written in the surface state object at any given
1359 * time. The only option is to set a bogus presumed offset and let the
1360 * kernel relocate them.
1362 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1363 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1366 VkResult result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1368 anv_execbuf_finish(&execbuf
, &cmd_buffer
->pool
->alloc
);