2 * Copyright © 2008 Jérôme Glisse
3 * Copyright © 2010 Marek Olšák <maraeo@gmail.com>
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
29 * Marek Olšák <maraeo@gmail.com>
31 * Based on work from libdrm_radeon by:
32 * Aapo Tahkola <aet@rasterburn.org>
33 * Nicolai Haehnle <prefect_@gmx.net>
34 * Jérôme Glisse <glisse@freedesktop.org>
38 This file replaces libdrm's radeon_cs_gem with our own implemention.
39 It's optimized specifically for Radeon DRM.
40 Reloc writes and space checking are faster and simpler than their
41 counterparts in libdrm (the time complexity of all the functions
42 is O(1) in nearly all scenarios, thanks to hashing).
46 cs_add_reloc(cs, buf, read_domain, write_domain) adds a new relocation and
47 also adds the size of 'buf' to the used_gart and used_vram winsys variables
48 based on the domains, which are simply or'd for the accounting purposes.
49 The adding is skipped if the reloc is already present in the list, but it
50 accounts any newly-referenced domains.
52 cs_validate is then called, which just checks:
53 used_vram/gart < vram/gart_size * 0.8
54 The 0.8 number allows for some memory fragmentation. If the validation
55 fails, the pipe driver flushes CS and tries do the validation again,
56 i.e. it validates only that one operation. If it fails again, it drops
57 the operation on the floor and prints some nasty message to stderr.
58 (done in the pipe driver)
60 cs_write_reloc(cs, buf) just writes a reloc that has been added using
61 cs_add_reloc. The read_domain and write_domain parameters have been removed,
62 because we already specify them in cs_add_reloc.
65 #include "radeon_drm_cs.h"
67 #include "util/u_memory.h"
68 #include "os/os_time.h"
76 #define RELOC_DWORDS (sizeof(struct drm_radeon_cs_reloc) / sizeof(uint32_t))
78 static struct pipe_fence_handle
*
79 radeon_cs_create_fence(struct radeon_winsys_cs
*rcs
);
80 static void radeon_fence_reference(struct pipe_fence_handle
**dst
,
81 struct pipe_fence_handle
*src
);
83 static struct radeon_winsys_ctx
*radeon_drm_ctx_create(struct radeon_winsys
*ws
)
85 /* No context support here. Just return the winsys pointer
86 * as the "context". */
87 return (struct radeon_winsys_ctx
*)ws
;
90 static void radeon_drm_ctx_destroy(struct radeon_winsys_ctx
*ctx
)
92 /* No context support here. */
95 static boolean
radeon_init_cs_context(struct radeon_cs_context
*csc
,
96 struct radeon_drm_winsys
*ws
)
102 csc
->relocs_bo
= (struct radeon_bo
**)
103 CALLOC(1, csc
->nrelocs
* sizeof(struct radeon_bo
*));
104 if (!csc
->relocs_bo
) {
108 csc
->relocs
= (struct drm_radeon_cs_reloc
*)
109 CALLOC(1, csc
->nrelocs
* sizeof(struct drm_radeon_cs_reloc
));
111 FREE(csc
->relocs_bo
);
115 csc
->chunks
[0].chunk_id
= RADEON_CHUNK_ID_IB
;
116 csc
->chunks
[0].length_dw
= 0;
117 csc
->chunks
[0].chunk_data
= (uint64_t)(uintptr_t)csc
->buf
;
118 csc
->chunks
[1].chunk_id
= RADEON_CHUNK_ID_RELOCS
;
119 csc
->chunks
[1].length_dw
= 0;
120 csc
->chunks
[1].chunk_data
= (uint64_t)(uintptr_t)csc
->relocs
;
121 csc
->chunks
[2].chunk_id
= RADEON_CHUNK_ID_FLAGS
;
122 csc
->chunks
[2].length_dw
= 2;
123 csc
->chunks
[2].chunk_data
= (uint64_t)(uintptr_t)&csc
->flags
;
125 csc
->chunk_array
[0] = (uint64_t)(uintptr_t)&csc
->chunks
[0];
126 csc
->chunk_array
[1] = (uint64_t)(uintptr_t)&csc
->chunks
[1];
127 csc
->chunk_array
[2] = (uint64_t)(uintptr_t)&csc
->chunks
[2];
129 csc
->cs
.chunks
= (uint64_t)(uintptr_t)csc
->chunk_array
;
131 for (i
= 0; i
< Elements(csc
->reloc_indices_hashlist
); i
++) {
132 csc
->reloc_indices_hashlist
[i
] = -1;
137 static void radeon_cs_context_cleanup(struct radeon_cs_context
*csc
)
141 for (i
= 0; i
< csc
->crelocs
; i
++) {
142 p_atomic_dec(&csc
->relocs_bo
[i
]->num_cs_references
);
143 radeon_bo_reference(&csc
->relocs_bo
[i
], NULL
);
147 csc
->validated_crelocs
= 0;
148 csc
->chunks
[0].length_dw
= 0;
149 csc
->chunks
[1].length_dw
= 0;
153 for (i
= 0; i
< Elements(csc
->reloc_indices_hashlist
); i
++) {
154 csc
->reloc_indices_hashlist
[i
] = -1;
158 static void radeon_destroy_cs_context(struct radeon_cs_context
*csc
)
160 radeon_cs_context_cleanup(csc
);
161 FREE(csc
->relocs_bo
);
166 static struct radeon_winsys_cs
*
167 radeon_drm_cs_create(struct radeon_winsys_ctx
*ctx
,
168 enum ring_type ring_type
,
169 void (*flush
)(void *ctx
, unsigned flags
,
170 struct pipe_fence_handle
**fence
),
172 struct radeon_winsys_cs_handle
*trace_buf
)
174 struct radeon_drm_winsys
*ws
= (struct radeon_drm_winsys
*)ctx
;
175 struct radeon_drm_cs
*cs
;
177 cs
= CALLOC_STRUCT(radeon_drm_cs
);
181 pipe_semaphore_init(&cs
->flush_completed
, 1);
184 cs
->flush_cs
= flush
;
185 cs
->flush_data
= flush_ctx
;
186 cs
->trace_buf
= (struct radeon_bo
*)trace_buf
;
188 if (!radeon_init_cs_context(&cs
->csc1
, cs
->ws
)) {
192 if (!radeon_init_cs_context(&cs
->csc2
, cs
->ws
)) {
193 radeon_destroy_cs_context(&cs
->csc1
);
198 /* Set the first command buffer as current. */
201 cs
->base
.buf
= cs
->csc
->buf
;
202 cs
->base
.ring_type
= ring_type
;
203 cs
->base
.max_dw
= ARRAY_SIZE(cs
->csc
->buf
);
205 p_atomic_inc(&ws
->num_cs
);
209 #define OUT_CS(cs, value) (cs)->buf[(cs)->cdw++] = (value)
211 static inline void update_reloc(struct drm_radeon_cs_reloc
*reloc
,
212 enum radeon_bo_domain rd
,
213 enum radeon_bo_domain wd
,
215 enum radeon_bo_domain
*added_domains
)
217 *added_domains
= (rd
| wd
) & ~(reloc
->read_domains
| reloc
->write_domain
);
219 reloc
->read_domains
|= rd
;
220 reloc
->write_domain
|= wd
;
221 reloc
->flags
= MAX2(reloc
->flags
, priority
);
224 int radeon_get_reloc(struct radeon_cs_context
*csc
, struct radeon_bo
*bo
)
226 unsigned hash
= bo
->handle
& (Elements(csc
->reloc_indices_hashlist
)-1);
227 int i
= csc
->reloc_indices_hashlist
[hash
];
229 /* not found or found */
230 if (i
== -1 || csc
->relocs_bo
[i
] == bo
)
233 /* Hash collision, look for the BO in the list of relocs linearly. */
234 for (i
= csc
->crelocs
- 1; i
>= 0; i
--) {
235 if (csc
->relocs_bo
[i
] == bo
) {
236 /* Put this reloc in the hash list.
237 * This will prevent additional hash collisions if there are
238 * several consecutive get_reloc calls for the same buffer.
240 * Example: Assuming buffers A,B,C collide in the hash list,
241 * the following sequence of relocs:
242 * AAAAAAAAAAABBBBBBBBBBBBBBCCCCCCCC
243 * will collide here: ^ and here: ^,
244 * meaning that we should get very few collisions in the end. */
245 csc
->reloc_indices_hashlist
[hash
] = i
;
252 static unsigned radeon_add_reloc(struct radeon_drm_cs
*cs
,
253 struct radeon_bo
*bo
,
254 enum radeon_bo_usage usage
,
255 enum radeon_bo_domain domains
,
257 enum radeon_bo_domain
*added_domains
)
259 struct radeon_cs_context
*csc
= cs
->csc
;
260 struct drm_radeon_cs_reloc
*reloc
;
261 unsigned hash
= bo
->handle
& (Elements(csc
->reloc_indices_hashlist
)-1);
262 enum radeon_bo_domain rd
= usage
& RADEON_USAGE_READ
? domains
: 0;
263 enum radeon_bo_domain wd
= usage
& RADEON_USAGE_WRITE
? domains
: 0;
266 priority
= MIN2(priority
, 15);
269 i
= radeon_get_reloc(csc
, bo
);
272 reloc
= &csc
->relocs
[i
];
273 update_reloc(reloc
, rd
, wd
, priority
, added_domains
);
275 /* For async DMA, every add_reloc call must add a buffer to the list
276 * no matter how many duplicates there are. This is due to the fact
277 * the DMA CS checker doesn't use NOP packets for offset patching,
278 * but always uses the i-th buffer from the list to patch the i-th
279 * offset. If there are N offsets in a DMA CS, there must also be N
280 * buffers in the relocation list.
282 * This doesn't have to be done if virtual memory is enabled,
283 * because there is no offset patching with virtual memory.
285 if (cs
->base
.ring_type
!= RING_DMA
|| cs
->ws
->info
.r600_virtual_address
) {
290 /* New relocation, check if the backing array is large enough. */
291 if (csc
->crelocs
>= csc
->nrelocs
) {
295 size
= csc
->nrelocs
* sizeof(struct radeon_bo
*);
296 csc
->relocs_bo
= realloc(csc
->relocs_bo
, size
);
298 size
= csc
->nrelocs
* sizeof(struct drm_radeon_cs_reloc
);
299 csc
->relocs
= realloc(csc
->relocs
, size
);
301 csc
->chunks
[1].chunk_data
= (uint64_t)(uintptr_t)csc
->relocs
;
304 /* Initialize the new relocation. */
305 csc
->relocs_bo
[csc
->crelocs
] = NULL
;
306 radeon_bo_reference(&csc
->relocs_bo
[csc
->crelocs
], bo
);
307 p_atomic_inc(&bo
->num_cs_references
);
308 reloc
= &csc
->relocs
[csc
->crelocs
];
309 reloc
->handle
= bo
->handle
;
310 reloc
->read_domains
= rd
;
311 reloc
->write_domain
= wd
;
312 reloc
->flags
= priority
;
314 csc
->reloc_indices_hashlist
[hash
] = csc
->crelocs
;
316 csc
->chunks
[1].length_dw
+= RELOC_DWORDS
;
318 *added_domains
= rd
| wd
;
319 return csc
->crelocs
++;
322 static unsigned radeon_drm_cs_add_reloc(struct radeon_winsys_cs
*rcs
,
323 struct radeon_winsys_cs_handle
*buf
,
324 enum radeon_bo_usage usage
,
325 enum radeon_bo_domain domains
,
326 enum radeon_bo_priority priority
)
328 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
329 struct radeon_bo
*bo
= (struct radeon_bo
*)buf
;
330 enum radeon_bo_domain added_domains
;
331 unsigned index
= radeon_add_reloc(cs
, bo
, usage
, domains
, priority
, &added_domains
);
333 if (added_domains
& RADEON_DOMAIN_GTT
)
334 cs
->csc
->used_gart
+= bo
->base
.size
;
335 if (added_domains
& RADEON_DOMAIN_VRAM
)
336 cs
->csc
->used_vram
+= bo
->base
.size
;
341 static int radeon_drm_cs_get_reloc(struct radeon_winsys_cs
*rcs
,
342 struct radeon_winsys_cs_handle
*buf
)
344 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
346 return radeon_get_reloc(cs
->csc
, (struct radeon_bo
*)buf
);
349 static boolean
radeon_drm_cs_validate(struct radeon_winsys_cs
*rcs
)
351 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
353 cs
->csc
->used_gart
< cs
->ws
->info
.gart_size
* 0.8 &&
354 cs
->csc
->used_vram
< cs
->ws
->info
.vram_size
* 0.8;
357 cs
->csc
->validated_crelocs
= cs
->csc
->crelocs
;
359 /* Remove lately-added relocations. The validation failed with them
360 * and the CS is about to be flushed because of that. Keep only
361 * the already-validated relocations. */
364 for (i
= cs
->csc
->validated_crelocs
; i
< cs
->csc
->crelocs
; i
++) {
365 p_atomic_dec(&cs
->csc
->relocs_bo
[i
]->num_cs_references
);
366 radeon_bo_reference(&cs
->csc
->relocs_bo
[i
], NULL
);
368 cs
->csc
->crelocs
= cs
->csc
->validated_crelocs
;
370 /* Flush if there are any relocs. Clean up otherwise. */
371 if (cs
->csc
->crelocs
) {
372 cs
->flush_cs(cs
->flush_data
, RADEON_FLUSH_ASYNC
, NULL
);
374 radeon_cs_context_cleanup(cs
->csc
);
376 assert(cs
->base
.cdw
== 0);
377 if (cs
->base
.cdw
!= 0) {
378 fprintf(stderr
, "radeon: Unexpected error in %s.\n", __func__
);
385 static boolean
radeon_drm_cs_memory_below_limit(struct radeon_winsys_cs
*rcs
, uint64_t vram
, uint64_t gtt
)
387 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
389 vram
+= cs
->csc
->used_vram
;
390 gtt
+= cs
->csc
->used_gart
;
392 /* Anything that goes above the VRAM size should go to GTT. */
393 if (vram
> cs
->ws
->info
.vram_size
)
394 gtt
+= vram
- cs
->ws
->info
.vram_size
;
396 /* Now we just need to check if we have enough GTT. */
397 return gtt
< cs
->ws
->info
.gart_size
* 0.7;
400 void radeon_drm_cs_emit_ioctl_oneshot(struct radeon_drm_cs
*cs
, struct radeon_cs_context
*csc
)
405 r
= drmCommandWriteRead(csc
->fd
, DRM_RADEON_CS
,
406 &csc
->cs
, sizeof(struct drm_radeon_cs
));
409 fprintf(stderr
, "radeon: Not enough memory for command submission.\n");
410 else if (debug_get_bool_option("RADEON_DUMP_CS", FALSE
)) {
413 fprintf(stderr
, "radeon: The kernel rejected CS, dumping...\n");
414 for (i
= 0; i
< csc
->chunks
[0].length_dw
; i
++) {
415 fprintf(stderr
, "0x%08X\n", csc
->buf
[i
]);
418 fprintf(stderr
, "radeon: The kernel rejected CS, "
419 "see dmesg for more information.\n");
424 radeon_dump_cs_on_lockup(cs
, csc
);
427 for (i
= 0; i
< csc
->crelocs
; i
++)
428 p_atomic_dec(&csc
->relocs_bo
[i
]->num_active_ioctls
);
430 radeon_cs_context_cleanup(csc
);
434 * Make sure previous submission of this cs are completed
436 void radeon_drm_cs_sync_flush(struct radeon_winsys_cs
*rcs
)
438 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
440 /* Wait for any pending ioctl to complete. */
441 if (cs
->ws
->thread
) {
442 pipe_semaphore_wait(&cs
->flush_completed
);
443 pipe_semaphore_signal(&cs
->flush_completed
);
447 DEBUG_GET_ONCE_BOOL_OPTION(noop
, "RADEON_NOOP", FALSE
)
449 static void radeon_drm_cs_flush(struct radeon_winsys_cs
*rcs
,
451 struct pipe_fence_handle
**fence
,
452 uint32_t cs_trace_id
)
454 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
455 struct radeon_cs_context
*tmp
;
457 switch (cs
->base
.ring_type
) {
459 /* pad DMA ring to 8 DWs */
460 if (cs
->ws
->info
.chip_class
<= SI
) {
462 OUT_CS(&cs
->base
, 0xf0000000); /* NOP packet */
465 OUT_CS(&cs
->base
, 0x00000000); /* NOP packet */
469 /* pad DMA ring to 8 DWs to meet CP fetch alignment requirements
470 * r6xx, requires at least 4 dw alignment to avoid a hw bug.
471 * hawaii with old firmware needs type2 nop packet.
472 * accel_working2 with value 3 indicates the new firmware.
474 if (cs
->ws
->info
.chip_class
<= SI
||
475 (cs
->ws
->info
.family
== CHIP_HAWAII
&&
476 cs
->ws
->accel_working2
< 3)) {
478 OUT_CS(&cs
->base
, 0x80000000); /* type2 nop packet */
481 OUT_CS(&cs
->base
, 0xffff1000); /* type3 nop packet */
485 while (rcs
->cdw
& 15)
486 OUT_CS(&cs
->base
, 0x80000000); /* type2 nop packet */
492 if (rcs
->cdw
> rcs
->max_dw
) {
493 fprintf(stderr
, "radeon: command stream overflowed\n");
497 radeon_fence_reference(fence
, NULL
);
498 *fence
= radeon_cs_create_fence(rcs
);
501 radeon_drm_cs_sync_flush(rcs
);
503 /* Swap command streams. */
508 cs
->cst
->cs_trace_id
= cs_trace_id
;
510 /* If the CS is not empty or overflowed, emit it in a separate thread. */
511 if (cs
->base
.cdw
&& cs
->base
.cdw
<= cs
->base
.max_dw
&& !debug_get_option_noop()) {
514 crelocs
= cs
->cst
->crelocs
;
516 cs
->cst
->chunks
[0].length_dw
= cs
->base
.cdw
;
518 for (i
= 0; i
< crelocs
; i
++) {
519 /* Update the number of active asynchronous CS ioctls for the buffer. */
520 p_atomic_inc(&cs
->cst
->relocs_bo
[i
]->num_active_ioctls
);
523 switch (cs
->base
.ring_type
) {
525 cs
->cst
->flags
[0] = 0;
526 cs
->cst
->flags
[1] = RADEON_CS_RING_DMA
;
527 cs
->cst
->cs
.num_chunks
= 3;
528 if (cs
->ws
->info
.r600_virtual_address
) {
529 cs
->cst
->flags
[0] |= RADEON_CS_USE_VM
;
534 cs
->cst
->flags
[0] = 0;
535 cs
->cst
->flags
[1] = RADEON_CS_RING_UVD
;
536 cs
->cst
->cs
.num_chunks
= 3;
540 cs
->cst
->flags
[0] = 0;
541 cs
->cst
->flags
[1] = RADEON_CS_RING_VCE
;
542 cs
->cst
->cs
.num_chunks
= 3;
548 cs
->cst
->flags
[0] = 0;
549 cs
->cst
->flags
[1] = RADEON_CS_RING_GFX
;
550 cs
->cst
->cs
.num_chunks
= 2;
551 if (flags
& RADEON_FLUSH_KEEP_TILING_FLAGS
) {
552 cs
->cst
->flags
[0] |= RADEON_CS_KEEP_TILING_FLAGS
;
553 cs
->cst
->cs
.num_chunks
= 3;
555 if (cs
->ws
->info
.r600_virtual_address
) {
556 cs
->cst
->flags
[0] |= RADEON_CS_USE_VM
;
557 cs
->cst
->cs
.num_chunks
= 3;
559 if (flags
& RADEON_FLUSH_END_OF_FRAME
) {
560 cs
->cst
->flags
[0] |= RADEON_CS_END_OF_FRAME
;
561 cs
->cst
->cs
.num_chunks
= 3;
563 if (cs
->base
.ring_type
== RING_COMPUTE
) {
564 cs
->cst
->flags
[1] = RADEON_CS_RING_COMPUTE
;
565 cs
->cst
->cs
.num_chunks
= 3;
570 if (cs
->ws
->thread
) {
571 pipe_semaphore_wait(&cs
->flush_completed
);
572 radeon_drm_ws_queue_cs(cs
->ws
, cs
);
573 if (!(flags
& RADEON_FLUSH_ASYNC
))
574 radeon_drm_cs_sync_flush(rcs
);
576 radeon_drm_cs_emit_ioctl_oneshot(cs
, cs
->cst
);
579 radeon_cs_context_cleanup(cs
->cst
);
582 /* Prepare a new CS. */
583 cs
->base
.buf
= cs
->csc
->buf
;
586 cs
->ws
->num_cs_flushes
++;
589 static void radeon_drm_cs_destroy(struct radeon_winsys_cs
*rcs
)
591 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
593 radeon_drm_cs_sync_flush(rcs
);
594 pipe_semaphore_destroy(&cs
->flush_completed
);
595 radeon_cs_context_cleanup(&cs
->csc1
);
596 radeon_cs_context_cleanup(&cs
->csc2
);
597 p_atomic_dec(&cs
->ws
->num_cs
);
598 radeon_destroy_cs_context(&cs
->csc1
);
599 radeon_destroy_cs_context(&cs
->csc2
);
603 static boolean
radeon_bo_is_referenced(struct radeon_winsys_cs
*rcs
,
604 struct radeon_winsys_cs_handle
*_buf
,
605 enum radeon_bo_usage usage
)
607 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
608 struct radeon_bo
*bo
= (struct radeon_bo
*)_buf
;
611 if (!bo
->num_cs_references
)
614 index
= radeon_get_reloc(cs
->csc
, bo
);
618 if ((usage
& RADEON_USAGE_WRITE
) && cs
->csc
->relocs
[index
].write_domain
)
620 if ((usage
& RADEON_USAGE_READ
) && cs
->csc
->relocs
[index
].read_domains
)
628 static struct pipe_fence_handle
*
629 radeon_cs_create_fence(struct radeon_winsys_cs
*rcs
)
631 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
632 struct pb_buffer
*fence
;
634 /* Create a fence, which is a dummy BO. */
635 fence
= cs
->ws
->base
.buffer_create(&cs
->ws
->base
, 1, 1, TRUE
,
636 RADEON_DOMAIN_GTT
, 0);
637 /* Add the fence as a dummy relocation. */
638 cs
->ws
->base
.cs_add_reloc(rcs
, cs
->ws
->base
.buffer_get_cs_handle(fence
),
639 RADEON_USAGE_READWRITE
, RADEON_DOMAIN_GTT
,
641 return (struct pipe_fence_handle
*)fence
;
644 static bool radeon_fence_wait(struct radeon_winsys
*ws
,
645 struct pipe_fence_handle
*fence
,
648 struct pb_buffer
*rfence
= (struct pb_buffer
*)fence
;
651 return ws
->buffer_wait(rfence
, 0, RADEON_USAGE_READWRITE
);
653 if (timeout
!= PIPE_TIMEOUT_INFINITE
) {
654 int64_t start_time
= os_time_get();
656 /* Convert to microseconds. */
659 /* Wait in a loop. */
660 while (!ws
->buffer_wait(rfence
, 0, RADEON_USAGE_READWRITE
)) {
661 if (os_time_get() - start_time
>= timeout
) {
669 ws
->buffer_wait(rfence
, PIPE_TIMEOUT_INFINITE
, RADEON_USAGE_READWRITE
);
673 static void radeon_fence_reference(struct pipe_fence_handle
**dst
,
674 struct pipe_fence_handle
*src
)
676 pb_reference((struct pb_buffer
**)dst
, (struct pb_buffer
*)src
);
679 void radeon_drm_cs_init_functions(struct radeon_drm_winsys
*ws
)
681 ws
->base
.ctx_create
= radeon_drm_ctx_create
;
682 ws
->base
.ctx_destroy
= radeon_drm_ctx_destroy
;
683 ws
->base
.cs_create
= radeon_drm_cs_create
;
684 ws
->base
.cs_destroy
= radeon_drm_cs_destroy
;
685 ws
->base
.cs_add_reloc
= radeon_drm_cs_add_reloc
;
686 ws
->base
.cs_get_reloc
= radeon_drm_cs_get_reloc
;
687 ws
->base
.cs_validate
= radeon_drm_cs_validate
;
688 ws
->base
.cs_memory_below_limit
= radeon_drm_cs_memory_below_limit
;
689 ws
->base
.cs_flush
= radeon_drm_cs_flush
;
690 ws
->base
.cs_is_buffer_referenced
= radeon_bo_is_referenced
;
691 ws
->base
.cs_sync_flush
= radeon_drm_cs_sync_flush
;
692 ws
->base
.fence_wait
= radeon_fence_wait
;
693 ws
->base
.fence_reference
= radeon_fence_reference
;