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 VkResult result
= anv_reloc_list_grow(list
, alloc
, 1);
155 if (result
!= VK_SUCCESS
)
158 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
159 index
= list
->num_relocs
++;
160 list
->reloc_bos
[index
] = target_bo
;
161 entry
= &list
->relocs
[index
];
162 entry
->target_handle
= target_bo
->gem_handle
;
163 entry
->delta
= delta
;
164 entry
->offset
= offset
;
165 entry
->presumed_offset
= target_bo
->offset
;
166 entry
->read_domains
= domain
;
167 entry
->write_domain
= domain
;
168 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
174 anv_reloc_list_append(struct anv_reloc_list
*list
,
175 const VkAllocationCallbacks
*alloc
,
176 struct anv_reloc_list
*other
, uint32_t offset
)
178 VkResult result
= anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
179 if (result
!= VK_SUCCESS
)
182 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
183 other
->num_relocs
* sizeof(other
->relocs
[0]));
184 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
185 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
187 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
188 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
190 list
->num_relocs
+= other
->num_relocs
;
194 /*-----------------------------------------------------------------------*
195 * Functions related to anv_batch
196 *-----------------------------------------------------------------------*/
199 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
201 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
202 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
203 if (result
!= VK_SUCCESS
) {
204 anv_batch_set_error(batch
, result
);
209 void *p
= batch
->next
;
211 batch
->next
+= num_dwords
* 4;
212 assert(batch
->next
<= batch
->end
);
218 anv_batch_emit_reloc(struct anv_batch
*batch
,
219 void *location
, struct anv_bo
*bo
, uint32_t delta
)
221 VkResult result
= anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
222 location
- batch
->start
, bo
, delta
);
223 if (result
!= VK_SUCCESS
) {
224 anv_batch_set_error(batch
, result
);
228 return bo
->offset
+ delta
;
232 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
234 uint32_t size
, offset
;
236 size
= other
->next
- other
->start
;
237 assert(size
% 4 == 0);
239 if (batch
->next
+ size
> batch
->end
) {
240 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
241 if (result
!= VK_SUCCESS
) {
242 anv_batch_set_error(batch
, result
);
247 assert(batch
->next
+ size
<= batch
->end
);
249 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
250 memcpy(batch
->next
, other
->start
, size
);
252 offset
= batch
->next
- batch
->start
;
253 VkResult result
= anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
254 other
->relocs
, offset
);
255 if (result
!= VK_SUCCESS
) {
256 anv_batch_set_error(batch
, result
);
263 /*-----------------------------------------------------------------------*
264 * Functions related to anv_batch_bo
265 *-----------------------------------------------------------------------*/
268 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
269 struct anv_batch_bo
**bbo_out
)
273 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
274 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
276 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
278 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
279 ANV_CMD_BUFFER_BATCH_SIZE
);
280 if (result
!= VK_SUCCESS
)
283 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
284 if (result
!= VK_SUCCESS
)
292 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
294 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
300 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
301 const struct anv_batch_bo
*other_bbo
,
302 struct anv_batch_bo
**bbo_out
)
306 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
307 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
309 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
311 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
313 if (result
!= VK_SUCCESS
)
316 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
318 if (result
!= VK_SUCCESS
)
321 bbo
->length
= other_bbo
->length
;
322 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
329 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
331 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
337 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
338 size_t batch_padding
)
340 batch
->next
= batch
->start
= bbo
->bo
.map
;
341 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
342 batch
->relocs
= &bbo
->relocs
;
343 bbo
->relocs
.num_relocs
= 0;
347 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
348 size_t batch_padding
)
350 batch
->start
= bbo
->bo
.map
;
351 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
352 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
353 batch
->relocs
= &bbo
->relocs
;
357 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
359 assert(batch
->start
== bbo
->bo
.map
);
360 bbo
->length
= batch
->next
- batch
->start
;
361 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
365 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
366 struct anv_batch
*batch
, size_t aditional
,
367 size_t batch_padding
)
369 assert(batch
->start
== bbo
->bo
.map
);
370 bbo
->length
= batch
->next
- batch
->start
;
372 size_t new_size
= bbo
->bo
.size
;
373 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
376 if (new_size
== bbo
->bo
.size
)
379 struct anv_bo new_bo
;
380 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
382 if (result
!= VK_SUCCESS
)
385 memcpy(new_bo
.map
, bbo
->bo
.map
, bbo
->length
);
387 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
390 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
396 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
397 struct anv_cmd_buffer
*cmd_buffer
)
399 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
400 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
401 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
405 anv_batch_bo_list_clone(const struct list_head
*list
,
406 struct anv_cmd_buffer
*cmd_buffer
,
407 struct list_head
*new_list
)
409 VkResult result
= VK_SUCCESS
;
411 list_inithead(new_list
);
413 struct anv_batch_bo
*prev_bbo
= NULL
;
414 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
415 struct anv_batch_bo
*new_bbo
= NULL
;
416 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
417 if (result
!= VK_SUCCESS
)
419 list_addtail(&new_bbo
->link
, new_list
);
422 /* As we clone this list of batch_bo's, they chain one to the
423 * other using MI_BATCH_BUFFER_START commands. We need to fix up
424 * those relocations as we go. Fortunately, this is pretty easy
425 * as it will always be the last relocation in the list.
427 uint32_t last_idx
= prev_bbo
->relocs
.num_relocs
- 1;
428 assert(prev_bbo
->relocs
.reloc_bos
[last_idx
] == &bbo
->bo
);
429 prev_bbo
->relocs
.reloc_bos
[last_idx
] = &new_bbo
->bo
;
435 if (result
!= VK_SUCCESS
) {
436 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
437 anv_batch_bo_destroy(bbo
, cmd_buffer
);
443 /*-----------------------------------------------------------------------*
444 * Functions related to anv_batch_bo
445 *-----------------------------------------------------------------------*/
447 static inline struct anv_batch_bo
*
448 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
450 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
454 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
456 return (struct anv_address
) {
457 .bo
= &cmd_buffer
->device
->surface_state_block_pool
.bo
,
458 .offset
= *(int32_t *)u_vector_head(&cmd_buffer
->bt_blocks
),
463 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
464 struct anv_bo
*bo
, uint32_t offset
)
466 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
467 * offsets. The high 16 bits are in the last dword, so we can use the gen8
468 * version in either case, as long as we set the instruction length in the
469 * header accordingly. This means that we always emit three dwords here
470 * and all the padding and adjustment we do in this file works for all
474 #define GEN7_MI_BATCH_BUFFER_START_length 2
475 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
477 const uint32_t gen7_length
=
478 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
479 const uint32_t gen8_length
=
480 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
482 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
483 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
484 gen7_length
: gen8_length
;
485 bbs
._2ndLevelBatchBuffer
= _1stlevelbatch
;
486 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
487 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
492 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
493 struct anv_batch_bo
*bbo
)
495 struct anv_batch
*batch
= &cmd_buffer
->batch
;
496 struct anv_batch_bo
*current_bbo
=
497 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
499 /* We set the end of the batch a little short so we would be sure we
500 * have room for the chaining command. Since we're about to emit the
501 * chaining command, let's set it back where it should go.
503 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
504 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
506 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
508 anv_batch_bo_finish(current_bbo
, batch
);
512 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
514 struct anv_cmd_buffer
*cmd_buffer
= _data
;
515 struct anv_batch_bo
*new_bbo
;
517 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
518 if (result
!= VK_SUCCESS
)
521 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
522 if (seen_bbo
== NULL
) {
523 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
524 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
528 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
530 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
532 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
538 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
540 struct anv_cmd_buffer
*cmd_buffer
= _data
;
541 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
543 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
544 GEN8_MI_BATCH_BUFFER_START_length
* 4);
549 /** Allocate a binding table
551 * This function allocates a binding table. This is a bit more complicated
552 * than one would think due to a combination of Vulkan driver design and some
553 * unfortunate hardware restrictions.
555 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
556 * the binding table pointer which means that all binding tables need to live
557 * in the bottom 64k of surface state base address. The way the GL driver has
558 * classically dealt with this restriction is to emit all surface states
559 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
560 * isn't really an option in Vulkan for a couple of reasons:
562 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
563 * to live in their own buffer and we have to be able to re-emit
564 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
565 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
566 * (it's not that hard to hit 64k of just binding tables), we allocate
567 * surface state objects up-front when VkImageView is created. In order
568 * for this to work, surface state objects need to be allocated from a
571 * 2) We tried to design the surface state system in such a way that it's
572 * already ready for bindless texturing. The way bindless texturing works
573 * on our hardware is that you have a big pool of surface state objects
574 * (with its own state base address) and the bindless handles are simply
575 * offsets into that pool. With the architecture we chose, we already
576 * have that pool and it's exactly the same pool that we use for regular
577 * surface states so we should already be ready for bindless.
579 * 3) For render targets, we need to be able to fill out the surface states
580 * later in vkBeginRenderPass so that we can assign clear colors
581 * correctly. One way to do this would be to just create the surface
582 * state data and then repeatedly copy it into the surface state BO every
583 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
584 * rather annoying and just being able to allocate them up-front and
585 * re-use them for the entire render pass.
587 * While none of these are technically blockers for emitting state on the fly
588 * like we do in GL, the ability to have a single surface state pool is
589 * simplifies things greatly. Unfortunately, it comes at a cost...
591 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
592 * place the binding tables just anywhere in surface state base address.
593 * Because 64k isn't a whole lot of space, we can't simply restrict the
594 * surface state buffer to 64k, we have to be more clever. The solution we've
595 * chosen is to have a block pool with a maximum size of 2G that starts at
596 * zero and grows in both directions. All surface states are allocated from
597 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
598 * binding tables from the bottom of the pool (negative offsets). Every time
599 * we allocate a new binding table block, we set surface state base address to
600 * point to the bottom of the binding table block. This way all of the
601 * binding tables in the block are in the bottom 64k of surface state base
602 * address. When we fill out the binding table, we add the distance between
603 * the bottom of our binding table block and zero of the block pool to the
604 * surface state offsets so that they are correct relative to out new surface
605 * state base address at the bottom of the binding table block.
607 * \see adjust_relocations_from_block_pool()
608 * \see adjust_relocations_too_block_pool()
610 * \param[in] entries The number of surface state entries the binding
611 * table should be able to hold.
613 * \param[out] state_offset The offset surface surface state base address
614 * where the surface states live. This must be
615 * added to the surface state offset when it is
616 * written into the binding table entry.
618 * \return An anv_state representing the binding table
621 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
622 uint32_t entries
, uint32_t *state_offset
)
624 struct anv_block_pool
*block_pool
=
625 &cmd_buffer
->device
->surface_state_block_pool
;
626 int32_t *bt_block
= u_vector_head(&cmd_buffer
->bt_blocks
);
627 struct anv_state state
;
629 state
.alloc_size
= align_u32(entries
* 4, 32);
631 if (cmd_buffer
->bt_next
+ state
.alloc_size
> block_pool
->block_size
)
632 return (struct anv_state
) { 0 };
634 state
.offset
= cmd_buffer
->bt_next
;
635 state
.map
= block_pool
->map
+ *bt_block
+ state
.offset
;
637 cmd_buffer
->bt_next
+= state
.alloc_size
;
639 assert(*bt_block
< 0);
640 *state_offset
= -(*bt_block
);
646 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
648 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
649 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
650 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
654 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
655 uint32_t size
, uint32_t alignment
)
657 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
662 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
664 struct anv_block_pool
*block_pool
=
665 &cmd_buffer
->device
->surface_state_block_pool
;
667 int32_t *offset
= u_vector_add(&cmd_buffer
->bt_blocks
);
668 if (offset
== NULL
) {
669 anv_batch_set_error(&cmd_buffer
->batch
, VK_ERROR_OUT_OF_HOST_MEMORY
);
670 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
673 *offset
= anv_block_pool_alloc_back(block_pool
);
674 cmd_buffer
->bt_next
= 0;
680 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
682 struct anv_batch_bo
*batch_bo
;
685 list_inithead(&cmd_buffer
->batch_bos
);
687 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
688 if (result
!= VK_SUCCESS
)
691 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
693 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
694 cmd_buffer
->batch
.user_data
= cmd_buffer
;
696 if (cmd_buffer
->device
->can_chain_batches
) {
697 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
699 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
702 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
703 GEN8_MI_BATCH_BUFFER_START_length
* 4);
705 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
706 sizeof(struct anv_bo
*),
707 8 * sizeof(struct anv_bo
*));
711 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
713 success
= u_vector_init(&cmd_buffer
->bt_blocks
, sizeof(int32_t),
714 8 * sizeof(int32_t));
718 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
719 &cmd_buffer
->pool
->alloc
);
720 if (result
!= VK_SUCCESS
)
722 cmd_buffer
->last_ss_pool_center
= 0;
724 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
725 if (result
!= VK_SUCCESS
)
731 u_vector_finish(&cmd_buffer
->bt_blocks
);
733 u_vector_finish(&cmd_buffer
->seen_bbos
);
735 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
741 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
744 u_vector_foreach(bt_block
, &cmd_buffer
->bt_blocks
) {
745 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
748 u_vector_finish(&cmd_buffer
->bt_blocks
);
750 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
752 u_vector_finish(&cmd_buffer
->seen_bbos
);
754 /* Destroy all of the batch buffers */
755 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
756 &cmd_buffer
->batch_bos
, link
) {
757 anv_batch_bo_destroy(bbo
, cmd_buffer
);
762 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
764 /* Delete all but the first batch bo */
765 assert(!list_empty(&cmd_buffer
->batch_bos
));
766 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
767 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
768 list_del(&bbo
->link
);
769 anv_batch_bo_destroy(bbo
, cmd_buffer
);
771 assert(!list_empty(&cmd_buffer
->batch_bos
));
773 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
775 GEN8_MI_BATCH_BUFFER_START_length
* 4);
777 while (u_vector_length(&cmd_buffer
->bt_blocks
) > 1) {
778 int32_t *bt_block
= u_vector_remove(&cmd_buffer
->bt_blocks
);
779 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
782 assert(u_vector_length(&cmd_buffer
->bt_blocks
) == 1);
783 cmd_buffer
->bt_next
= 0;
785 cmd_buffer
->surface_relocs
.num_relocs
= 0;
786 cmd_buffer
->last_ss_pool_center
= 0;
788 /* Reset the list of seen buffers */
789 cmd_buffer
->seen_bbos
.head
= 0;
790 cmd_buffer
->seen_bbos
.tail
= 0;
792 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
793 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
797 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
799 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
801 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
802 /* When we start a batch buffer, we subtract a certain amount of
803 * padding from the end to ensure that we always have room to emit a
804 * BATCH_BUFFER_START to chain to the next BO. We need to remove
805 * that padding before we end the batch; otherwise, we may end up
806 * with our BATCH_BUFFER_END in another BO.
808 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
809 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
811 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
813 /* Round batch up to an even number of dwords. */
814 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
815 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
817 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
820 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
822 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
) {
823 /* If this is a secondary command buffer, we need to determine the
824 * mode in which it will be executed with vkExecuteCommands. We
825 * determine this statically here so that this stays in sync with the
826 * actual ExecuteCommands implementation.
828 if (!cmd_buffer
->device
->can_chain_batches
) {
829 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
830 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
831 (batch_bo
->length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
832 /* If the secondary has exactly one batch buffer in its list *and*
833 * that batch buffer is less than half of the maximum size, we're
834 * probably better of simply copying it into our batch.
836 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
837 } else if (!(cmd_buffer
->usage_flags
&
838 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
839 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
841 /* When we chain, we need to add an MI_BATCH_BUFFER_START command
842 * with its relocation. In order to handle this we'll increment here
843 * so we can unconditionally decrement right before adding the
844 * MI_BATCH_BUFFER_START command.
846 batch_bo
->relocs
.num_relocs
++;
847 cmd_buffer
->batch
.next
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
849 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
854 static inline VkResult
855 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
856 struct list_head
*list
)
858 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
859 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
861 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
870 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
871 struct anv_cmd_buffer
*secondary
)
873 switch (secondary
->exec_mode
) {
874 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
875 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
877 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
878 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
879 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
880 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
881 GEN8_MI_BATCH_BUFFER_START_length
* 4);
882 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
885 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
886 struct anv_batch_bo
*first_bbo
=
887 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
888 struct anv_batch_bo
*last_bbo
=
889 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
891 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
893 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
894 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
895 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
896 const uint32_t inst_size
= GEN8_MI_BATCH_BUFFER_START_length
* 4;
898 /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we
899 * can emit a new command and relocation for the current splice. In
900 * order to handle the initial-use case, we incremented next and
901 * num_relocs in end_batch_buffer() so we can alyways just subtract
904 last_bbo
->relocs
.num_relocs
--;
905 secondary
->batch
.next
-= inst_size
;
906 emit_batch_buffer_start(secondary
, &this_bbo
->bo
, offset
);
907 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
909 /* After patching up the secondary buffer, we need to clflush the
910 * modified instruction in case we're on a !llc platform. We use a
911 * little loop to handle the case where the instruction crosses a cache
914 if (!primary
->device
->info
.has_llc
) {
915 void *inst
= secondary
->batch
.next
- inst_size
;
916 void *p
= (void *) (((uintptr_t) inst
) & ~CACHELINE_MASK
);
917 __builtin_ia32_mfence();
918 while (p
< secondary
->batch
.next
) {
919 __builtin_ia32_clflush(p
);
925 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
926 struct list_head copy_list
;
927 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
930 if (result
!= VK_SUCCESS
)
933 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
935 struct anv_batch_bo
*first_bbo
=
936 list_first_entry(©_list
, struct anv_batch_bo
, link
);
937 struct anv_batch_bo
*last_bbo
=
938 list_last_entry(©_list
, struct anv_batch_bo
, link
);
940 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
942 list_splicetail(©_list
, &primary
->batch_bos
);
944 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
945 GEN8_MI_BATCH_BUFFER_START_length
* 4);
949 assert(!"Invalid execution mode");
952 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
953 &secondary
->surface_relocs
, 0);
957 struct drm_i915_gem_execbuffer2 execbuf
;
959 struct drm_i915_gem_exec_object2
* objects
;
961 struct anv_bo
** bos
;
963 /* Allocated length of the 'objects' and 'bos' arrays */
964 uint32_t array_length
;
968 anv_execbuf_init(struct anv_execbuf
*exec
)
970 memset(exec
, 0, sizeof(*exec
));
974 anv_execbuf_finish(struct anv_execbuf
*exec
,
975 const VkAllocationCallbacks
*alloc
)
977 vk_free(alloc
, exec
->objects
);
978 vk_free(alloc
, exec
->bos
);
982 anv_execbuf_add_bo(struct anv_execbuf
*exec
,
984 struct anv_reloc_list
*relocs
,
985 const VkAllocationCallbacks
*alloc
)
987 struct drm_i915_gem_exec_object2
*obj
= NULL
;
989 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
990 obj
= &exec
->objects
[bo
->index
];
993 /* We've never seen this one before. Add it to the list and assign
994 * an id that we can use later.
996 if (exec
->bo_count
>= exec
->array_length
) {
997 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
999 struct drm_i915_gem_exec_object2
*new_objects
=
1000 vk_alloc(alloc
, new_len
* sizeof(*new_objects
),
1001 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1002 if (new_objects
== NULL
)
1003 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1005 struct anv_bo
**new_bos
=
1006 vk_alloc(alloc
, new_len
* sizeof(*new_bos
),
1007 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1008 if (new_bos
== NULL
) {
1009 vk_free(alloc
, new_objects
);
1010 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1013 if (exec
->objects
) {
1014 memcpy(new_objects
, exec
->objects
,
1015 exec
->bo_count
* sizeof(*new_objects
));
1016 memcpy(new_bos
, exec
->bos
,
1017 exec
->bo_count
* sizeof(*new_bos
));
1020 vk_free(alloc
, exec
->objects
);
1021 vk_free(alloc
, exec
->bos
);
1023 exec
->objects
= new_objects
;
1024 exec
->bos
= new_bos
;
1025 exec
->array_length
= new_len
;
1028 assert(exec
->bo_count
< exec
->array_length
);
1030 bo
->index
= exec
->bo_count
++;
1031 obj
= &exec
->objects
[bo
->index
];
1032 exec
->bos
[bo
->index
] = bo
;
1034 obj
->handle
= bo
->gem_handle
;
1035 obj
->relocation_count
= 0;
1036 obj
->relocs_ptr
= 0;
1038 obj
->offset
= bo
->offset
;
1039 obj
->flags
= bo
->is_winsys_bo
? EXEC_OBJECT_WRITE
: 0;
1044 if (relocs
!= NULL
&& obj
->relocation_count
== 0) {
1045 /* This is the first time we've ever seen a list of relocations for
1046 * this BO. Go ahead and set the relocations and then walk the list
1047 * of relocations and add them all.
1049 obj
->relocation_count
= relocs
->num_relocs
;
1050 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1052 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1053 /* A quick sanity check on relocations */
1054 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1055 anv_execbuf_add_bo(exec
, relocs
->reloc_bos
[i
], NULL
, alloc
);
1063 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1064 struct anv_reloc_list
*list
)
1066 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1067 list
->relocs
[i
].target_handle
= list
->reloc_bos
[i
]->index
;
1071 write_reloc(const struct anv_device
*device
, void *p
, uint64_t v
, bool flush
)
1073 unsigned reloc_size
= 0;
1074 if (device
->info
.gen
>= 8) {
1075 /* From the Broadwell PRM Vol. 2a, MI_LOAD_REGISTER_MEM::MemoryAddress:
1077 * "This field specifies the address of the memory location where the
1078 * register value specified in the DWord above will read from. The
1079 * address specifies the DWord location of the data. Range =
1080 * GraphicsVirtualAddress[63:2] for a DWord register GraphicsAddress
1081 * [63:48] are ignored by the HW and assumed to be in correct
1082 * canonical form [63:48] == [47]."
1084 const int shift
= 63 - 47;
1085 reloc_size
= sizeof(uint64_t);
1086 *(uint64_t *)p
= (((int64_t)v
) << shift
) >> shift
;
1088 reloc_size
= sizeof(uint32_t);
1092 if (flush
&& !device
->info
.has_llc
)
1093 anv_flush_range(p
, reloc_size
);
1097 adjust_relocations_from_state_pool(struct anv_block_pool
*pool
,
1098 struct anv_reloc_list
*relocs
,
1099 uint32_t last_pool_center_bo_offset
)
1101 assert(last_pool_center_bo_offset
<= pool
->center_bo_offset
);
1102 uint32_t delta
= pool
->center_bo_offset
- last_pool_center_bo_offset
;
1104 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1105 /* All of the relocations from this block pool to other BO's should
1106 * have been emitted relative to the surface block pool center. We
1107 * need to add the center offset to make them relative to the
1108 * beginning of the actual GEM bo.
1110 relocs
->relocs
[i
].offset
+= delta
;
1115 adjust_relocations_to_state_pool(struct anv_block_pool
*pool
,
1116 struct anv_bo
*from_bo
,
1117 struct anv_reloc_list
*relocs
,
1118 uint32_t last_pool_center_bo_offset
)
1120 assert(last_pool_center_bo_offset
<= pool
->center_bo_offset
);
1121 uint32_t delta
= pool
->center_bo_offset
- last_pool_center_bo_offset
;
1123 /* When we initially emit relocations into a block pool, we don't
1124 * actually know what the final center_bo_offset will be so we just emit
1125 * it as if center_bo_offset == 0. Now that we know what the center
1126 * offset is, we need to walk the list of relocations and adjust any
1127 * relocations that point to the pool bo with the correct offset.
1129 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1130 if (relocs
->reloc_bos
[i
] == &pool
->bo
) {
1131 /* Adjust the delta value in the relocation to correctly
1132 * correspond to the new delta. Initially, this value may have
1133 * been negative (if treated as unsigned), but we trust in
1134 * uint32_t roll-over to fix that for us at this point.
1136 relocs
->relocs
[i
].delta
+= delta
;
1138 /* Since the delta has changed, we need to update the actual
1139 * relocated value with the new presumed value. This function
1140 * should only be called on batch buffers, so we know it isn't in
1141 * use by the GPU at the moment.
1143 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1144 write_reloc(pool
->device
, from_bo
->map
+ relocs
->relocs
[i
].offset
,
1145 relocs
->relocs
[i
].presumed_offset
+
1146 relocs
->relocs
[i
].delta
, false);
1152 anv_reloc_list_apply(struct anv_device
*device
,
1153 struct anv_reloc_list
*list
,
1155 bool always_relocate
)
1157 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1158 struct anv_bo
*target_bo
= list
->reloc_bos
[i
];
1159 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1163 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1164 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1165 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1170 * This function applies the relocation for a command buffer and writes the
1171 * actual addresses into the buffers as per what we were told by the kernel on
1172 * the previous execbuf2 call. This should be safe to do because, for each
1173 * relocated address, we have two cases:
1175 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1176 * not in use by the GPU so updating the address is 100% ok. It won't be
1177 * in-use by the GPU (from our context) again until the next execbuf2
1178 * happens. If the kernel decides to move it in the next execbuf2, it
1179 * will have to do the relocations itself, but that's ok because it should
1180 * have all of the information needed to do so.
1182 * 2) The target BO is active (as seen by the kernel). In this case, it
1183 * hasn't moved since the last execbuffer2 call because GTT shuffling
1184 * *only* happens when the BO is idle. (From our perspective, it only
1185 * happens inside the execbuffer2 ioctl, but the shuffling may be
1186 * triggered by another ioctl, with full-ppgtt this is limited to only
1187 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1188 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1189 * address and the relocated value we are writing into the BO will be the
1190 * same as the value that is already there.
1192 * There is also a possibility that the target BO is active but the exact
1193 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1194 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1195 * may be stale but it's still safe to write the relocation because that
1196 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1197 * won't be until the next execbuf2 call.
1199 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1200 * need to bother. We want to do this because the surface state buffer is
1201 * used by every command buffer so, if the kernel does the relocations, it
1202 * will always be busy and the kernel will always stall. This is also
1203 * probably the fastest mechanism for doing relocations since the kernel would
1204 * have to make a full copy of all the relocations lists.
1207 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1208 struct anv_execbuf
*exec
)
1210 static int userspace_relocs
= -1;
1211 if (userspace_relocs
< 0)
1212 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1213 if (!userspace_relocs
)
1216 /* First, we have to check to see whether or not we can even do the
1217 * relocation. New buffers which have never been submitted to the kernel
1218 * don't have a valid offset so we need to let the kernel do relocations so
1219 * that we can get offsets for them. On future execbuf2 calls, those
1220 * buffers will have offsets and we will be able to skip relocating.
1221 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1223 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1224 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1228 /* Since surface states are shared between command buffers and we don't
1229 * know what order they will be submitted to the kernel, we don't know
1230 * what address is actually written in the surface state object at any
1231 * given time. The only option is to always relocate them.
1233 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1234 &cmd_buffer
->device
->surface_state_block_pool
.bo
,
1235 true /* always relocate surface states */);
1237 /* Since we own all of the batch buffers, we know what values are stored
1238 * in the relocated addresses and only have to update them if the offsets
1241 struct anv_batch_bo
**bbo
;
1242 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1243 anv_reloc_list_apply(cmd_buffer
->device
,
1244 &(*bbo
)->relocs
, &(*bbo
)->bo
, false);
1247 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1248 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1254 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1255 struct anv_cmd_buffer
*cmd_buffer
)
1257 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1258 struct anv_block_pool
*ss_pool
=
1259 &cmd_buffer
->device
->surface_state_block_pool
;
1261 struct anv_execbuf execbuf
;
1262 anv_execbuf_init(&execbuf
);
1264 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1265 cmd_buffer
->last_ss_pool_center
);
1267 anv_execbuf_add_bo(&execbuf
, &ss_pool
->bo
, &cmd_buffer
->surface_relocs
,
1268 &cmd_buffer
->pool
->alloc
);
1269 if (result
!= VK_SUCCESS
)
1272 /* First, we walk over all of the bos we've seen and add them and their
1273 * relocations to the validate list.
1275 struct anv_batch_bo
**bbo
;
1276 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1277 adjust_relocations_to_state_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1278 cmd_buffer
->last_ss_pool_center
);
1280 result
= anv_execbuf_add_bo(&execbuf
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1281 &cmd_buffer
->pool
->alloc
);
1282 if (result
!= VK_SUCCESS
)
1286 /* Now that we've adjusted all of the surface state relocations, we need to
1287 * record the surface state pool center so future executions of the command
1288 * buffer can adjust correctly.
1290 cmd_buffer
->last_ss_pool_center
= ss_pool
->center_bo_offset
;
1292 struct anv_batch_bo
*first_batch_bo
=
1293 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1295 /* The kernel requires that the last entry in the validation list be the
1296 * batch buffer to execute. We can simply swap the element
1297 * corresponding to the first batch_bo in the chain with the last
1298 * element in the list.
1300 if (first_batch_bo
->bo
.index
!= execbuf
.bo_count
- 1) {
1301 uint32_t idx
= first_batch_bo
->bo
.index
;
1302 uint32_t last_idx
= execbuf
.bo_count
- 1;
1304 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
.objects
[idx
];
1305 assert(execbuf
.bos
[idx
] == &first_batch_bo
->bo
);
1307 execbuf
.objects
[idx
] = execbuf
.objects
[last_idx
];
1308 execbuf
.bos
[idx
] = execbuf
.bos
[last_idx
];
1309 execbuf
.bos
[idx
]->index
= idx
;
1311 execbuf
.objects
[last_idx
] = tmp_obj
;
1312 execbuf
.bos
[last_idx
] = &first_batch_bo
->bo
;
1313 first_batch_bo
->bo
.index
= last_idx
;
1316 /* Now we go through and fixup all of the relocation lists to point to
1317 * the correct indices in the object array. We have to do this after we
1318 * reorder the list above as some of the indices may have changed.
1320 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1321 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1323 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1325 if (!cmd_buffer
->device
->info
.has_llc
) {
1326 __builtin_ia32_mfence();
1327 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1328 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1329 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1333 execbuf
.execbuf
= (struct drm_i915_gem_execbuffer2
) {
1334 .buffers_ptr
= (uintptr_t) execbuf
.objects
,
1335 .buffer_count
= execbuf
.bo_count
,
1336 .batch_start_offset
= 0,
1337 .batch_len
= batch
->next
- batch
->start
,
1342 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
|
1343 I915_EXEC_CONSTANTS_REL_GENERAL
,
1344 .rsvd1
= cmd_buffer
->device
->context_id
,
1348 if (relocate_cmd_buffer(cmd_buffer
, &execbuf
)) {
1349 /* If we were able to successfully relocate everything, tell the kernel
1350 * that it can skip doing relocations. The requirement for using
1353 * 1) The addresses written in the objects must match the corresponding
1354 * reloc.presumed_offset which in turn must match the corresponding
1355 * execobject.offset.
1357 * 2) To avoid stalling, execobject.offset should match the current
1358 * address of that object within the active context.
1360 * In order to satisfy all of the invariants that make userspace
1361 * relocations to be safe (see relocate_cmd_buffer()), we need to
1362 * further ensure that the addresses we use match those used by the
1363 * kernel for the most recent execbuf2.
1365 * The kernel may still choose to do relocations anyway if something has
1366 * moved in the GTT. In this case, the relocation list still needs to be
1367 * valid. All relocations on the batch buffers are already valid and
1368 * kept up-to-date. For surface state relocations, by applying the
1369 * relocations in relocate_cmd_buffer, we ensured that the address in
1370 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1371 * safe for the kernel to relocate them as needed.
1373 execbuf
.execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1375 /* In the case where we fall back to doing kernel relocations, we need
1376 * to ensure that the relocation list is valid. All relocations on the
1377 * batch buffers are already valid and kept up-to-date. Since surface
1378 * states are shared between command buffers and we don't know what
1379 * order they will be submitted to the kernel, we don't know what
1380 * address is actually written in the surface state object at any given
1381 * time. The only option is to set a bogus presumed offset and let the
1382 * kernel relocate them.
1384 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1385 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1388 result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1390 anv_execbuf_finish(&execbuf
, &cmd_buffer
->pool
->alloc
);