2 * Copyright 2011 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Christian König <deathsimple@vodafone.de>
31 #include <linux/firmware.h>
32 #include <linux/module.h>
37 #include "amdgpu_pm.h"
38 #include "amdgpu_uvd.h"
40 #include "uvd/uvd_4_2_d.h"
42 /* 1 second timeout */
43 #define UVD_IDLE_TIMEOUT_MS 1000
44 /* Polaris10/11 firmware version */
45 #define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 #define FIRMWARE_BONAIRE "radeon/bonaire_uvd.bin"
50 #define FIRMWARE_KABINI "radeon/kabini_uvd.bin"
51 #define FIRMWARE_KAVERI "radeon/kaveri_uvd.bin"
52 #define FIRMWARE_HAWAII "radeon/hawaii_uvd.bin"
53 #define FIRMWARE_MULLINS "radeon/mullins_uvd.bin"
55 #define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
56 #define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
57 #define FIRMWARE_FIJI "amdgpu/fiji_uvd.bin"
58 #define FIRMWARE_STONEY "amdgpu/stoney_uvd.bin"
59 #define FIRMWARE_POLARIS10 "amdgpu/polaris10_uvd.bin"
60 #define FIRMWARE_POLARIS11 "amdgpu/polaris11_uvd.bin"
63 * amdgpu_uvd_cs_ctx - Command submission parser context
65 * Used for emulating virtual memory support on UVD 4.2.
67 struct amdgpu_uvd_cs_ctx
{
68 struct amdgpu_cs_parser
*parser
;
70 unsigned data0
, data1
;
74 /* does the IB has a msg command */
77 /* minimum buffer sizes */
81 #ifdef CONFIG_DRM_AMDGPU_CIK
82 MODULE_FIRMWARE(FIRMWARE_BONAIRE
);
83 MODULE_FIRMWARE(FIRMWARE_KABINI
);
84 MODULE_FIRMWARE(FIRMWARE_KAVERI
);
85 MODULE_FIRMWARE(FIRMWARE_HAWAII
);
86 MODULE_FIRMWARE(FIRMWARE_MULLINS
);
88 MODULE_FIRMWARE(FIRMWARE_TONGA
);
89 MODULE_FIRMWARE(FIRMWARE_CARRIZO
);
90 MODULE_FIRMWARE(FIRMWARE_FIJI
);
91 MODULE_FIRMWARE(FIRMWARE_STONEY
);
92 MODULE_FIRMWARE(FIRMWARE_POLARIS10
);
93 MODULE_FIRMWARE(FIRMWARE_POLARIS11
);
95 static void amdgpu_uvd_note_usage(struct amdgpu_device
*adev
);
96 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
);
98 int amdgpu_uvd_sw_init(struct amdgpu_device
*adev
)
100 struct amdgpu_ring
*ring
;
101 struct amd_sched_rq
*rq
;
102 unsigned long bo_size
;
104 const struct common_firmware_header
*hdr
;
105 unsigned version_major
, version_minor
, family_id
;
108 INIT_DELAYED_WORK(&adev
->uvd
.idle_work
, amdgpu_uvd_idle_work_handler
);
110 switch (adev
->asic_type
) {
111 #ifdef CONFIG_DRM_AMDGPU_CIK
113 fw_name
= FIRMWARE_BONAIRE
;
116 fw_name
= FIRMWARE_KABINI
;
119 fw_name
= FIRMWARE_KAVERI
;
122 fw_name
= FIRMWARE_HAWAII
;
125 fw_name
= FIRMWARE_MULLINS
;
129 fw_name
= FIRMWARE_TONGA
;
132 fw_name
= FIRMWARE_FIJI
;
135 fw_name
= FIRMWARE_CARRIZO
;
138 fw_name
= FIRMWARE_STONEY
;
141 fw_name
= FIRMWARE_POLARIS10
;
144 fw_name
= FIRMWARE_POLARIS11
;
150 r
= request_firmware(&adev
->uvd
.fw
, fw_name
, adev
->dev
);
152 dev_err(adev
->dev
, "amdgpu_uvd: Can't load firmware \"%s\"\n",
157 r
= amdgpu_ucode_validate(adev
->uvd
.fw
);
159 dev_err(adev
->dev
, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
161 release_firmware(adev
->uvd
.fw
);
166 /* Set the default UVD handles that the firmware can handle */
167 adev
->uvd
.max_handles
= AMDGPU_DEFAULT_UVD_HANDLES
;
169 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
170 family_id
= le32_to_cpu(hdr
->ucode_version
) & 0xff;
171 version_major
= (le32_to_cpu(hdr
->ucode_version
) >> 24) & 0xff;
172 version_minor
= (le32_to_cpu(hdr
->ucode_version
) >> 8) & 0xff;
173 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
174 version_major
, version_minor
, family_id
);
177 * Limit the number of UVD handles depending on microcode major
178 * and minor versions. The firmware version which has 40 UVD
179 * instances support is 1.80. So all subsequent versions should
180 * also have the same support.
182 if ((version_major
> 0x01) ||
183 ((version_major
== 0x01) && (version_minor
>= 0x50)))
184 adev
->uvd
.max_handles
= AMDGPU_MAX_UVD_HANDLES
;
186 adev
->uvd
.fw_version
= ((version_major
<< 24) | (version_minor
<< 16) |
189 if ((adev
->asic_type
== CHIP_POLARIS10
||
190 adev
->asic_type
== CHIP_POLARIS11
) &&
191 (adev
->uvd
.fw_version
< FW_1_66_16
))
192 DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
193 version_major
, version_minor
);
195 bo_size
= AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr
->ucode_size_bytes
) + 8)
196 + AMDGPU_UVD_STACK_SIZE
+ AMDGPU_UVD_HEAP_SIZE
197 + AMDGPU_UVD_SESSION_SIZE
* adev
->uvd
.max_handles
;
198 r
= amdgpu_bo_create(adev
, bo_size
, PAGE_SIZE
, true,
199 AMDGPU_GEM_DOMAIN_VRAM
,
200 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
201 NULL
, NULL
, &adev
->uvd
.vcpu_bo
);
203 dev_err(adev
->dev
, "(%d) failed to allocate UVD bo\n", r
);
207 r
= amdgpu_bo_reserve(adev
->uvd
.vcpu_bo
, false);
209 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
210 dev_err(adev
->dev
, "(%d) failed to reserve UVD bo\n", r
);
214 r
= amdgpu_bo_pin(adev
->uvd
.vcpu_bo
, AMDGPU_GEM_DOMAIN_VRAM
,
215 &adev
->uvd
.gpu_addr
);
217 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
218 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
219 dev_err(adev
->dev
, "(%d) UVD bo pin failed\n", r
);
223 r
= amdgpu_bo_kmap(adev
->uvd
.vcpu_bo
, &adev
->uvd
.cpu_addr
);
225 dev_err(adev
->dev
, "(%d) UVD map failed\n", r
);
229 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
231 ring
= &adev
->uvd
.ring
;
232 rq
= &ring
->sched
.sched_rq
[AMD_SCHED_PRIORITY_NORMAL
];
233 r
= amd_sched_entity_init(&ring
->sched
, &adev
->uvd
.entity
,
234 rq
, amdgpu_sched_jobs
);
236 DRM_ERROR("Failed setting up UVD run queue.\n");
240 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
241 atomic_set(&adev
->uvd
.handles
[i
], 0);
242 adev
->uvd
.filp
[i
] = NULL
;
245 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
246 if (!amdgpu_ip_block_version_cmp(adev
, AMD_IP_BLOCK_TYPE_UVD
, 5, 0))
247 adev
->uvd
.address_64_bit
= true;
252 int amdgpu_uvd_sw_fini(struct amdgpu_device
*adev
)
256 kfree(adev
->uvd
.saved_bo
);
258 amd_sched_entity_fini(&adev
->uvd
.ring
.sched
, &adev
->uvd
.entity
);
260 if (adev
->uvd
.vcpu_bo
) {
261 r
= amdgpu_bo_reserve(adev
->uvd
.vcpu_bo
, false);
263 amdgpu_bo_kunmap(adev
->uvd
.vcpu_bo
);
264 amdgpu_bo_unpin(adev
->uvd
.vcpu_bo
);
265 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
268 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
271 amdgpu_ring_fini(&adev
->uvd
.ring
);
273 release_firmware(adev
->uvd
.fw
);
278 int amdgpu_uvd_suspend(struct amdgpu_device
*adev
)
284 if (adev
->uvd
.vcpu_bo
== NULL
)
287 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
)
288 if (atomic_read(&adev
->uvd
.handles
[i
]))
291 if (i
== AMDGPU_MAX_UVD_HANDLES
)
294 cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
296 size
= amdgpu_bo_size(adev
->uvd
.vcpu_bo
);
297 ptr
= adev
->uvd
.cpu_addr
;
299 adev
->uvd
.saved_bo
= kmalloc(size
, GFP_KERNEL
);
300 if (!adev
->uvd
.saved_bo
)
303 memcpy(adev
->uvd
.saved_bo
, ptr
, size
);
308 int amdgpu_uvd_resume(struct amdgpu_device
*adev
)
313 if (adev
->uvd
.vcpu_bo
== NULL
)
316 size
= amdgpu_bo_size(adev
->uvd
.vcpu_bo
);
317 ptr
= adev
->uvd
.cpu_addr
;
319 if (adev
->uvd
.saved_bo
!= NULL
) {
320 memcpy(ptr
, adev
->uvd
.saved_bo
, size
);
321 kfree(adev
->uvd
.saved_bo
);
322 adev
->uvd
.saved_bo
= NULL
;
324 const struct common_firmware_header
*hdr
;
327 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
328 offset
= le32_to_cpu(hdr
->ucode_array_offset_bytes
);
329 memcpy(adev
->uvd
.cpu_addr
, (adev
->uvd
.fw
->data
) + offset
,
330 (adev
->uvd
.fw
->size
) - offset
);
331 size
-= le32_to_cpu(hdr
->ucode_size_bytes
);
332 ptr
+= le32_to_cpu(hdr
->ucode_size_bytes
);
333 memset(ptr
, 0, size
);
339 void amdgpu_uvd_free_handles(struct amdgpu_device
*adev
, struct drm_file
*filp
)
341 struct amdgpu_ring
*ring
= &adev
->uvd
.ring
;
344 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
345 uint32_t handle
= atomic_read(&adev
->uvd
.handles
[i
]);
346 if (handle
!= 0 && adev
->uvd
.filp
[i
] == filp
) {
349 amdgpu_uvd_note_usage(adev
);
351 r
= amdgpu_uvd_get_destroy_msg(ring
, handle
,
354 DRM_ERROR("Error destroying UVD (%d)!\n", r
);
358 fence_wait(fence
, false);
361 adev
->uvd
.filp
[i
] = NULL
;
362 atomic_set(&adev
->uvd
.handles
[i
], 0);
367 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo
*rbo
)
370 for (i
= 0; i
< rbo
->placement
.num_placement
; ++i
) {
371 rbo
->placements
[i
].fpfn
= 0 >> PAGE_SHIFT
;
372 rbo
->placements
[i
].lpfn
= (256 * 1024 * 1024) >> PAGE_SHIFT
;
377 * amdgpu_uvd_cs_pass1 - first parsing round
379 * @ctx: UVD parser context
381 * Make sure UVD message and feedback buffers are in VRAM and
382 * nobody is violating an 256MB boundary.
384 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx
*ctx
)
386 struct amdgpu_bo_va_mapping
*mapping
;
387 struct amdgpu_bo
*bo
;
388 uint32_t cmd
, lo
, hi
;
392 lo
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
);
393 hi
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
);
394 addr
= ((uint64_t)lo
) | (((uint64_t)hi
) << 32);
396 mapping
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
);
397 if (mapping
== NULL
) {
398 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr
);
402 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
403 /* check if it's a message or feedback command */
404 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
405 if (cmd
== 0x0 || cmd
== 0x3) {
406 /* yes, force it into VRAM */
407 uint32_t domain
= AMDGPU_GEM_DOMAIN_VRAM
;
408 amdgpu_ttm_placement_from_domain(bo
, domain
);
410 amdgpu_uvd_force_into_uvd_segment(bo
);
412 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
419 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
421 * @msg: pointer to message structure
422 * @buf_sizes: returned buffer sizes
424 * Peek into the decode message and calculate the necessary buffer sizes.
426 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device
*adev
, uint32_t *msg
,
427 unsigned buf_sizes
[])
429 unsigned stream_type
= msg
[4];
430 unsigned width
= msg
[6];
431 unsigned height
= msg
[7];
432 unsigned dpb_size
= msg
[9];
433 unsigned pitch
= msg
[28];
434 unsigned level
= msg
[57];
436 unsigned width_in_mb
= width
/ 16;
437 unsigned height_in_mb
= ALIGN(height
/ 16, 2);
438 unsigned fs_in_mb
= width_in_mb
* height_in_mb
;
440 unsigned image_size
, tmp
, min_dpb_size
, num_dpb_buffer
;
441 unsigned min_ctx_size
= 0;
443 image_size
= width
* height
;
444 image_size
+= image_size
/ 2;
445 image_size
= ALIGN(image_size
, 1024);
447 switch (stream_type
) {
451 num_dpb_buffer
= 8100 / fs_in_mb
;
454 num_dpb_buffer
= 18000 / fs_in_mb
;
457 num_dpb_buffer
= 20480 / fs_in_mb
;
460 num_dpb_buffer
= 32768 / fs_in_mb
;
463 num_dpb_buffer
= 34816 / fs_in_mb
;
466 num_dpb_buffer
= 110400 / fs_in_mb
;
469 num_dpb_buffer
= 184320 / fs_in_mb
;
472 num_dpb_buffer
= 184320 / fs_in_mb
;
476 if (num_dpb_buffer
> 17)
479 /* reference picture buffer */
480 min_dpb_size
= image_size
* num_dpb_buffer
;
482 /* macroblock context buffer */
483 min_dpb_size
+= width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
485 /* IT surface buffer */
486 min_dpb_size
+= width_in_mb
* height_in_mb
* 32;
491 /* reference picture buffer */
492 min_dpb_size
= image_size
* 3;
495 min_dpb_size
+= width_in_mb
* height_in_mb
* 128;
497 /* IT surface buffer */
498 min_dpb_size
+= width_in_mb
* 64;
500 /* DB surface buffer */
501 min_dpb_size
+= width_in_mb
* 128;
504 tmp
= max(width_in_mb
, height_in_mb
);
505 min_dpb_size
+= ALIGN(tmp
* 7 * 16, 64);
510 /* reference picture buffer */
511 min_dpb_size
= image_size
* 3;
516 /* reference picture buffer */
517 min_dpb_size
= image_size
* 3;
520 min_dpb_size
+= width_in_mb
* height_in_mb
* 64;
522 /* IT surface buffer */
523 min_dpb_size
+= ALIGN(width_in_mb
* height_in_mb
* 32, 64);
526 case 7: /* H264 Perf */
529 num_dpb_buffer
= 8100 / fs_in_mb
;
532 num_dpb_buffer
= 18000 / fs_in_mb
;
535 num_dpb_buffer
= 20480 / fs_in_mb
;
538 num_dpb_buffer
= 32768 / fs_in_mb
;
541 num_dpb_buffer
= 34816 / fs_in_mb
;
544 num_dpb_buffer
= 110400 / fs_in_mb
;
547 num_dpb_buffer
= 184320 / fs_in_mb
;
550 num_dpb_buffer
= 184320 / fs_in_mb
;
554 if (num_dpb_buffer
> 17)
557 /* reference picture buffer */
558 min_dpb_size
= image_size
* num_dpb_buffer
;
560 if (adev
->asic_type
< CHIP_POLARIS10
){
561 /* macroblock context buffer */
563 width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
565 /* IT surface buffer */
566 min_dpb_size
+= width_in_mb
* height_in_mb
* 32;
568 /* macroblock context buffer */
570 width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
575 image_size
= (ALIGN(width
, 16) * ALIGN(height
, 16) * 3) / 2;
576 image_size
= ALIGN(image_size
, 256);
578 num_dpb_buffer
= (le32_to_cpu(msg
[59]) & 0xff) + 2;
579 min_dpb_size
= image_size
* num_dpb_buffer
;
580 min_ctx_size
= ((width
+ 255) / 16) * ((height
+ 255) / 16)
581 * 16 * num_dpb_buffer
+ 52 * 1024;
585 DRM_ERROR("UVD codec not handled %d!\n", stream_type
);
590 DRM_ERROR("Invalid UVD decoding target pitch!\n");
594 if (dpb_size
< min_dpb_size
) {
595 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
596 dpb_size
, min_dpb_size
);
600 buf_sizes
[0x1] = dpb_size
;
601 buf_sizes
[0x2] = image_size
;
602 buf_sizes
[0x4] = min_ctx_size
;
607 * amdgpu_uvd_cs_msg - handle UVD message
609 * @ctx: UVD parser context
610 * @bo: buffer object containing the message
611 * @offset: offset into the buffer object
613 * Peek into the UVD message and extract the session id.
614 * Make sure that we don't open up to many sessions.
616 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx
*ctx
,
617 struct amdgpu_bo
*bo
, unsigned offset
)
619 struct amdgpu_device
*adev
= ctx
->parser
->adev
;
620 int32_t *msg
, msg_type
, handle
;
626 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
630 r
= amdgpu_bo_kmap(bo
, &ptr
);
632 DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r
);
642 DRM_ERROR("Invalid UVD handle!\n");
648 /* it's a create msg, calc image size (width * height) */
649 amdgpu_bo_kunmap(bo
);
651 /* try to alloc a new handle */
652 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
653 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
654 DRM_ERROR("Handle 0x%x already in use!\n", handle
);
658 if (!atomic_cmpxchg(&adev
->uvd
.handles
[i
], 0, handle
)) {
659 adev
->uvd
.filp
[i
] = ctx
->parser
->filp
;
664 DRM_ERROR("No more free UVD handles!\n");
668 /* it's a decode msg, calc buffer sizes */
669 r
= amdgpu_uvd_cs_msg_decode(adev
, msg
, ctx
->buf_sizes
);
670 amdgpu_bo_kunmap(bo
);
674 /* validate the handle */
675 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
676 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
677 if (adev
->uvd
.filp
[i
] != ctx
->parser
->filp
) {
678 DRM_ERROR("UVD handle collision detected!\n");
685 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle
);
689 /* it's a destroy msg, free the handle */
690 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
)
691 atomic_cmpxchg(&adev
->uvd
.handles
[i
], handle
, 0);
692 amdgpu_bo_kunmap(bo
);
696 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type
);
704 * amdgpu_uvd_cs_pass2 - second parsing round
706 * @ctx: UVD parser context
708 * Patch buffer addresses, make sure buffer sizes are correct.
710 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx
*ctx
)
712 struct amdgpu_bo_va_mapping
*mapping
;
713 struct amdgpu_bo
*bo
;
714 uint32_t cmd
, lo
, hi
;
719 lo
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
);
720 hi
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
);
721 addr
= ((uint64_t)lo
) | (((uint64_t)hi
) << 32);
723 mapping
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
);
727 start
= amdgpu_bo_gpu_offset(bo
);
729 end
= (mapping
->it
.last
+ 1 - mapping
->it
.start
);
730 end
= end
* AMDGPU_GPU_PAGE_SIZE
+ start
;
732 addr
-= ((uint64_t)mapping
->it
.start
) * AMDGPU_GPU_PAGE_SIZE
;
735 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
,
736 lower_32_bits(start
));
737 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
,
738 upper_32_bits(start
));
740 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
742 if ((end
- start
) < ctx
->buf_sizes
[cmd
]) {
743 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
744 (unsigned)(end
- start
),
745 ctx
->buf_sizes
[cmd
]);
749 } else if (cmd
== 0x206) {
750 if ((end
- start
) < ctx
->buf_sizes
[4]) {
751 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
752 (unsigned)(end
- start
),
756 } else if ((cmd
!= 0x100) && (cmd
!= 0x204)) {
757 DRM_ERROR("invalid UVD command %X!\n", cmd
);
761 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
762 if ((start
>> 28) != ((end
- 1) >> 28)) {
763 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
768 if ((cmd
== 0 || cmd
== 0x3) &&
769 (start
>> 28) != (ctx
->parser
->adev
->uvd
.gpu_addr
>> 28)) {
770 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
777 ctx
->has_msg_cmd
= true;
778 r
= amdgpu_uvd_cs_msg(ctx
, bo
, addr
);
781 } else if (!ctx
->has_msg_cmd
) {
782 DRM_ERROR("Message needed before other commands are send!\n");
790 * amdgpu_uvd_cs_reg - parse register writes
792 * @ctx: UVD parser context
793 * @cb: callback function
795 * Parse the register writes, call cb on each complete command.
797 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx
*ctx
,
798 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
800 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
804 for (i
= 0; i
<= ctx
->count
; ++i
) {
805 unsigned reg
= ctx
->reg
+ i
;
807 if (ctx
->idx
>= ib
->length_dw
) {
808 DRM_ERROR("Register command after end of CS!\n");
813 case mmUVD_GPCOM_VCPU_DATA0
:
814 ctx
->data0
= ctx
->idx
;
816 case mmUVD_GPCOM_VCPU_DATA1
:
817 ctx
->data1
= ctx
->idx
;
819 case mmUVD_GPCOM_VCPU_CMD
:
824 case mmUVD_ENGINE_CNTL
:
827 DRM_ERROR("Invalid reg 0x%X!\n", reg
);
836 * amdgpu_uvd_cs_packets - parse UVD packets
838 * @ctx: UVD parser context
839 * @cb: callback function
841 * Parse the command stream packets.
843 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx
*ctx
,
844 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
846 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
849 for (ctx
->idx
= 0 ; ctx
->idx
< ib
->length_dw
; ) {
850 uint32_t cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
);
851 unsigned type
= CP_PACKET_GET_TYPE(cmd
);
854 ctx
->reg
= CP_PACKET0_GET_REG(cmd
);
855 ctx
->count
= CP_PACKET_GET_COUNT(cmd
);
856 r
= amdgpu_uvd_cs_reg(ctx
, cb
);
864 DRM_ERROR("Unknown packet type %d !\n", type
);
872 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
874 * @parser: Command submission parser context
876 * Parse the command stream, patch in addresses as necessary.
878 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser
*parser
, uint32_t ib_idx
)
880 struct amdgpu_uvd_cs_ctx ctx
= {};
881 unsigned buf_sizes
[] = {
883 [0x00000001] = 0xFFFFFFFF,
884 [0x00000002] = 0xFFFFFFFF,
886 [0x00000004] = 0xFFFFFFFF,
888 struct amdgpu_ib
*ib
= &parser
->job
->ibs
[ib_idx
];
891 if (ib
->length_dw
% 16) {
892 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
898 ctx
.buf_sizes
= buf_sizes
;
901 /* first round, make sure the buffers are actually in the UVD segment */
902 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass1
);
906 /* second round, patch buffer addresses into the command stream */
907 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass2
);
911 if (!ctx
.has_msg_cmd
) {
912 DRM_ERROR("UVD-IBs need a msg command!\n");
916 amdgpu_uvd_note_usage(ctx
.parser
->adev
);
921 static int amdgpu_uvd_send_msg(struct amdgpu_ring
*ring
, struct amdgpu_bo
*bo
,
922 bool direct
, struct fence
**fence
)
924 struct ttm_validate_buffer tv
;
925 struct ww_acquire_ctx ticket
;
926 struct list_head head
;
927 struct amdgpu_job
*job
;
928 struct amdgpu_ib
*ib
;
929 struct fence
*f
= NULL
;
930 struct amdgpu_device
*adev
= ring
->adev
;
934 memset(&tv
, 0, sizeof(tv
));
937 INIT_LIST_HEAD(&head
);
938 list_add(&tv
.head
, &head
);
940 r
= ttm_eu_reserve_buffers(&ticket
, &head
, true, NULL
);
944 if (!bo
->adev
->uvd
.address_64_bit
) {
945 amdgpu_ttm_placement_from_domain(bo
, AMDGPU_GEM_DOMAIN_VRAM
);
946 amdgpu_uvd_force_into_uvd_segment(bo
);
949 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, true, false);
953 r
= amdgpu_job_alloc_with_ib(adev
, 64, &job
);
958 addr
= amdgpu_bo_gpu_offset(bo
);
959 ib
->ptr
[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0
, 0);
961 ib
->ptr
[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1
, 0);
962 ib
->ptr
[3] = addr
>> 32;
963 ib
->ptr
[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD
, 0);
965 for (i
= 6; i
< 16; ++i
)
966 ib
->ptr
[i
] = PACKET2(0);
970 r
= amdgpu_ib_schedule(ring
, 1, ib
, NULL
, NULL
, &f
);
975 amdgpu_job_free(job
);
977 r
= amdgpu_job_submit(job
, ring
, &adev
->uvd
.entity
,
978 AMDGPU_FENCE_OWNER_UNDEFINED
, &f
);
983 ttm_eu_fence_buffer_objects(&ticket
, &head
, f
);
986 *fence
= fence_get(f
);
987 amdgpu_bo_unref(&bo
);
993 amdgpu_job_free(job
);
996 ttm_eu_backoff_reservation(&ticket
, &head
);
1000 /* multiple fence commands without any stream commands in between can
1001 crash the vcpu so just try to emmit a dummy create/destroy msg to
1003 int amdgpu_uvd_get_create_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
1004 struct fence
**fence
)
1006 struct amdgpu_device
*adev
= ring
->adev
;
1007 struct amdgpu_bo
*bo
;
1011 r
= amdgpu_bo_create(adev
, 1024, PAGE_SIZE
, true,
1012 AMDGPU_GEM_DOMAIN_VRAM
,
1013 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
1018 r
= amdgpu_bo_reserve(bo
, false);
1020 amdgpu_bo_unref(&bo
);
1024 r
= amdgpu_bo_kmap(bo
, (void **)&msg
);
1026 amdgpu_bo_unreserve(bo
);
1027 amdgpu_bo_unref(&bo
);
1031 /* stitch together an UVD create msg */
1032 msg
[0] = cpu_to_le32(0x00000de4);
1033 msg
[1] = cpu_to_le32(0x00000000);
1034 msg
[2] = cpu_to_le32(handle
);
1035 msg
[3] = cpu_to_le32(0x00000000);
1036 msg
[4] = cpu_to_le32(0x00000000);
1037 msg
[5] = cpu_to_le32(0x00000000);
1038 msg
[6] = cpu_to_le32(0x00000000);
1039 msg
[7] = cpu_to_le32(0x00000780);
1040 msg
[8] = cpu_to_le32(0x00000440);
1041 msg
[9] = cpu_to_le32(0x00000000);
1042 msg
[10] = cpu_to_le32(0x01b37000);
1043 for (i
= 11; i
< 1024; ++i
)
1044 msg
[i
] = cpu_to_le32(0x0);
1046 amdgpu_bo_kunmap(bo
);
1047 amdgpu_bo_unreserve(bo
);
1049 return amdgpu_uvd_send_msg(ring
, bo
, true, fence
);
1052 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
1053 bool direct
, struct fence
**fence
)
1055 struct amdgpu_device
*adev
= ring
->adev
;
1056 struct amdgpu_bo
*bo
;
1060 r
= amdgpu_bo_create(adev
, 1024, PAGE_SIZE
, true,
1061 AMDGPU_GEM_DOMAIN_VRAM
,
1062 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
1067 r
= amdgpu_bo_reserve(bo
, false);
1069 amdgpu_bo_unref(&bo
);
1073 r
= amdgpu_bo_kmap(bo
, (void **)&msg
);
1075 amdgpu_bo_unreserve(bo
);
1076 amdgpu_bo_unref(&bo
);
1080 /* stitch together an UVD destroy msg */
1081 msg
[0] = cpu_to_le32(0x00000de4);
1082 msg
[1] = cpu_to_le32(0x00000002);
1083 msg
[2] = cpu_to_le32(handle
);
1084 msg
[3] = cpu_to_le32(0x00000000);
1085 for (i
= 4; i
< 1024; ++i
)
1086 msg
[i
] = cpu_to_le32(0x0);
1088 amdgpu_bo_kunmap(bo
);
1089 amdgpu_bo_unreserve(bo
);
1091 return amdgpu_uvd_send_msg(ring
, bo
, direct
, fence
);
1094 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
)
1096 struct amdgpu_device
*adev
=
1097 container_of(work
, struct amdgpu_device
, uvd
.idle_work
.work
);
1098 unsigned i
, fences
, handles
= 0;
1100 fences
= amdgpu_fence_count_emitted(&adev
->uvd
.ring
);
1102 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
)
1103 if (atomic_read(&adev
->uvd
.handles
[i
]))
1106 if (fences
== 0 && handles
== 0) {
1107 if (adev
->pm
.dpm_enabled
) {
1108 amdgpu_dpm_enable_uvd(adev
, false);
1109 /* just work around for uvd clock remain high even
1110 * when uvd dpm disabled on Polaris10 */
1111 if (adev
->asic_type
== CHIP_POLARIS10
)
1112 amdgpu_asic_set_uvd_clocks(adev
, 0, 0);
1114 amdgpu_asic_set_uvd_clocks(adev
, 0, 0);
1117 schedule_delayed_work(&adev
->uvd
.idle_work
,
1118 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS
));
1122 static void amdgpu_uvd_note_usage(struct amdgpu_device
*adev
)
1124 bool set_clocks
= !cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
1125 set_clocks
&= schedule_delayed_work(&adev
->uvd
.idle_work
,
1126 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS
));
1129 if (adev
->pm
.dpm_enabled
) {
1130 amdgpu_dpm_enable_uvd(adev
, true);
1132 amdgpu_asic_set_uvd_clocks(adev
, 53300, 40000);