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
46 #ifdef CONFIG_DRM_AMDGPU_CIK
47 #define FIRMWARE_BONAIRE "radeon/bonaire_uvd.bin"
48 #define FIRMWARE_KABINI "radeon/kabini_uvd.bin"
49 #define FIRMWARE_KAVERI "radeon/kaveri_uvd.bin"
50 #define FIRMWARE_HAWAII "radeon/hawaii_uvd.bin"
51 #define FIRMWARE_MULLINS "radeon/mullins_uvd.bin"
53 #define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
54 #define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
55 #define FIRMWARE_FIJI "amdgpu/fiji_uvd.bin"
56 #define FIRMWARE_STONEY "amdgpu/stoney_uvd.bin"
59 * amdgpu_uvd_cs_ctx - Command submission parser context
61 * Used for emulating virtual memory support on UVD 4.2.
63 struct amdgpu_uvd_cs_ctx
{
64 struct amdgpu_cs_parser
*parser
;
66 unsigned data0
, data1
;
70 /* does the IB has a msg command */
73 /* minimum buffer sizes */
77 #ifdef CONFIG_DRM_AMDGPU_CIK
78 MODULE_FIRMWARE(FIRMWARE_BONAIRE
);
79 MODULE_FIRMWARE(FIRMWARE_KABINI
);
80 MODULE_FIRMWARE(FIRMWARE_KAVERI
);
81 MODULE_FIRMWARE(FIRMWARE_HAWAII
);
82 MODULE_FIRMWARE(FIRMWARE_MULLINS
);
84 MODULE_FIRMWARE(FIRMWARE_TONGA
);
85 MODULE_FIRMWARE(FIRMWARE_CARRIZO
);
86 MODULE_FIRMWARE(FIRMWARE_FIJI
);
87 MODULE_FIRMWARE(FIRMWARE_STONEY
);
89 static void amdgpu_uvd_note_usage(struct amdgpu_device
*adev
);
90 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
);
92 int amdgpu_uvd_sw_init(struct amdgpu_device
*adev
)
94 struct amdgpu_ring
*ring
;
95 struct amd_sched_rq
*rq
;
96 unsigned long bo_size
;
98 const struct common_firmware_header
*hdr
;
99 unsigned version_major
, version_minor
, family_id
;
102 INIT_DELAYED_WORK(&adev
->uvd
.idle_work
, amdgpu_uvd_idle_work_handler
);
104 switch (adev
->asic_type
) {
105 #ifdef CONFIG_DRM_AMDGPU_CIK
107 fw_name
= FIRMWARE_BONAIRE
;
110 fw_name
= FIRMWARE_KABINI
;
113 fw_name
= FIRMWARE_KAVERI
;
116 fw_name
= FIRMWARE_HAWAII
;
119 fw_name
= FIRMWARE_MULLINS
;
123 fw_name
= FIRMWARE_TONGA
;
126 fw_name
= FIRMWARE_FIJI
;
129 fw_name
= FIRMWARE_CARRIZO
;
132 fw_name
= FIRMWARE_STONEY
;
138 r
= request_firmware(&adev
->uvd
.fw
, fw_name
, adev
->dev
);
140 dev_err(adev
->dev
, "amdgpu_uvd: Can't load firmware \"%s\"\n",
145 r
= amdgpu_ucode_validate(adev
->uvd
.fw
);
147 dev_err(adev
->dev
, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
149 release_firmware(adev
->uvd
.fw
);
154 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
155 family_id
= le32_to_cpu(hdr
->ucode_version
) & 0xff;
156 version_major
= (le32_to_cpu(hdr
->ucode_version
) >> 24) & 0xff;
157 version_minor
= (le32_to_cpu(hdr
->ucode_version
) >> 8) & 0xff;
158 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
159 version_major
, version_minor
, family_id
);
161 bo_size
= AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr
->ucode_size_bytes
) + 8)
162 + AMDGPU_UVD_STACK_SIZE
+ AMDGPU_UVD_HEAP_SIZE
;
163 r
= amdgpu_bo_create(adev
, bo_size
, PAGE_SIZE
, true,
164 AMDGPU_GEM_DOMAIN_VRAM
,
165 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
166 NULL
, NULL
, &adev
->uvd
.vcpu_bo
);
168 dev_err(adev
->dev
, "(%d) failed to allocate UVD bo\n", r
);
172 r
= amdgpu_bo_reserve(adev
->uvd
.vcpu_bo
, false);
174 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
175 dev_err(adev
->dev
, "(%d) failed to reserve UVD bo\n", r
);
179 r
= amdgpu_bo_pin(adev
->uvd
.vcpu_bo
, AMDGPU_GEM_DOMAIN_VRAM
,
180 &adev
->uvd
.gpu_addr
);
182 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
183 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
184 dev_err(adev
->dev
, "(%d) UVD bo pin failed\n", r
);
188 r
= amdgpu_bo_kmap(adev
->uvd
.vcpu_bo
, &adev
->uvd
.cpu_addr
);
190 dev_err(adev
->dev
, "(%d) UVD map failed\n", r
);
194 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
196 ring
= &adev
->uvd
.ring
;
197 rq
= &ring
->sched
.sched_rq
[AMD_SCHED_PRIORITY_NORMAL
];
198 r
= amd_sched_entity_init(&ring
->sched
, &adev
->uvd
.entity
,
199 rq
, amdgpu_sched_jobs
);
201 DRM_ERROR("Failed setting up UVD run queue.\n");
205 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
) {
206 atomic_set(&adev
->uvd
.handles
[i
], 0);
207 adev
->uvd
.filp
[i
] = NULL
;
210 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
211 if (!amdgpu_ip_block_version_cmp(adev
, AMD_IP_BLOCK_TYPE_UVD
, 5, 0))
212 adev
->uvd
.address_64_bit
= true;
217 int amdgpu_uvd_sw_fini(struct amdgpu_device
*adev
)
221 if (adev
->uvd
.vcpu_bo
== NULL
)
224 amd_sched_entity_fini(&adev
->uvd
.ring
.sched
, &adev
->uvd
.entity
);
226 r
= amdgpu_bo_reserve(adev
->uvd
.vcpu_bo
, false);
228 amdgpu_bo_kunmap(adev
->uvd
.vcpu_bo
);
229 amdgpu_bo_unpin(adev
->uvd
.vcpu_bo
);
230 amdgpu_bo_unreserve(adev
->uvd
.vcpu_bo
);
233 amdgpu_bo_unref(&adev
->uvd
.vcpu_bo
);
235 amdgpu_ring_fini(&adev
->uvd
.ring
);
237 release_firmware(adev
->uvd
.fw
);
242 int amdgpu_uvd_suspend(struct amdgpu_device
*adev
)
248 if (adev
->uvd
.vcpu_bo
== NULL
)
251 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
)
252 if (atomic_read(&adev
->uvd
.handles
[i
]))
255 if (i
== AMDGPU_MAX_UVD_HANDLES
)
258 cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
260 size
= amdgpu_bo_size(adev
->uvd
.vcpu_bo
);
261 ptr
= adev
->uvd
.cpu_addr
;
263 adev
->uvd
.saved_bo
= kmalloc(size
, GFP_KERNEL
);
264 if (!adev
->uvd
.saved_bo
)
267 memcpy(adev
->uvd
.saved_bo
, ptr
, size
);
272 int amdgpu_uvd_resume(struct amdgpu_device
*adev
)
277 if (adev
->uvd
.vcpu_bo
== NULL
)
280 size
= amdgpu_bo_size(adev
->uvd
.vcpu_bo
);
281 ptr
= adev
->uvd
.cpu_addr
;
283 if (adev
->uvd
.saved_bo
!= NULL
) {
284 memcpy(ptr
, adev
->uvd
.saved_bo
, size
);
285 kfree(adev
->uvd
.saved_bo
);
286 adev
->uvd
.saved_bo
= NULL
;
288 const struct common_firmware_header
*hdr
;
291 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
292 offset
= le32_to_cpu(hdr
->ucode_array_offset_bytes
);
293 memcpy(adev
->uvd
.cpu_addr
, (adev
->uvd
.fw
->data
) + offset
,
294 (adev
->uvd
.fw
->size
) - offset
);
295 size
-= le32_to_cpu(hdr
->ucode_size_bytes
);
296 ptr
+= le32_to_cpu(hdr
->ucode_size_bytes
);
297 memset(ptr
, 0, size
);
303 void amdgpu_uvd_free_handles(struct amdgpu_device
*adev
, struct drm_file
*filp
)
305 struct amdgpu_ring
*ring
= &adev
->uvd
.ring
;
308 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
) {
309 uint32_t handle
= atomic_read(&adev
->uvd
.handles
[i
]);
310 if (handle
!= 0 && adev
->uvd
.filp
[i
] == filp
) {
313 amdgpu_uvd_note_usage(adev
);
315 r
= amdgpu_uvd_get_destroy_msg(ring
, handle
,
318 DRM_ERROR("Error destroying UVD (%d)!\n", r
);
322 fence_wait(fence
, false);
325 adev
->uvd
.filp
[i
] = NULL
;
326 atomic_set(&adev
->uvd
.handles
[i
], 0);
331 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo
*rbo
)
334 for (i
= 0; i
< rbo
->placement
.num_placement
; ++i
) {
335 rbo
->placements
[i
].fpfn
= 0 >> PAGE_SHIFT
;
336 rbo
->placements
[i
].lpfn
= (256 * 1024 * 1024) >> PAGE_SHIFT
;
341 * amdgpu_uvd_cs_pass1 - first parsing round
343 * @ctx: UVD parser context
345 * Make sure UVD message and feedback buffers are in VRAM and
346 * nobody is violating an 256MB boundary.
348 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx
*ctx
)
350 struct amdgpu_bo_va_mapping
*mapping
;
351 struct amdgpu_bo
*bo
;
352 uint32_t cmd
, lo
, hi
;
356 lo
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
);
357 hi
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
);
358 addr
= ((uint64_t)lo
) | (((uint64_t)hi
) << 32);
360 mapping
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
);
361 if (mapping
== NULL
) {
362 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr
);
366 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
367 /* check if it's a message or feedback command */
368 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
369 if (cmd
== 0x0 || cmd
== 0x3) {
370 /* yes, force it into VRAM */
371 uint32_t domain
= AMDGPU_GEM_DOMAIN_VRAM
;
372 amdgpu_ttm_placement_from_domain(bo
, domain
);
374 amdgpu_uvd_force_into_uvd_segment(bo
);
376 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
383 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
385 * @msg: pointer to message structure
386 * @buf_sizes: returned buffer sizes
388 * Peek into the decode message and calculate the necessary buffer sizes.
390 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg
, unsigned buf_sizes
[])
392 unsigned stream_type
= msg
[4];
393 unsigned width
= msg
[6];
394 unsigned height
= msg
[7];
395 unsigned dpb_size
= msg
[9];
396 unsigned pitch
= msg
[28];
397 unsigned level
= msg
[57];
399 unsigned width_in_mb
= width
/ 16;
400 unsigned height_in_mb
= ALIGN(height
/ 16, 2);
401 unsigned fs_in_mb
= width_in_mb
* height_in_mb
;
403 unsigned image_size
, tmp
, min_dpb_size
, num_dpb_buffer
;
404 unsigned min_ctx_size
= 0;
406 image_size
= width
* height
;
407 image_size
+= image_size
/ 2;
408 image_size
= ALIGN(image_size
, 1024);
410 switch (stream_type
) {
412 case 7: /* H264 Perf */
415 num_dpb_buffer
= 8100 / fs_in_mb
;
418 num_dpb_buffer
= 18000 / fs_in_mb
;
421 num_dpb_buffer
= 20480 / fs_in_mb
;
424 num_dpb_buffer
= 32768 / fs_in_mb
;
427 num_dpb_buffer
= 34816 / fs_in_mb
;
430 num_dpb_buffer
= 110400 / fs_in_mb
;
433 num_dpb_buffer
= 184320 / fs_in_mb
;
436 num_dpb_buffer
= 184320 / fs_in_mb
;
440 if (num_dpb_buffer
> 17)
443 /* reference picture buffer */
444 min_dpb_size
= image_size
* num_dpb_buffer
;
446 /* macroblock context buffer */
447 min_dpb_size
+= width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
449 /* IT surface buffer */
450 min_dpb_size
+= width_in_mb
* height_in_mb
* 32;
455 /* reference picture buffer */
456 min_dpb_size
= image_size
* 3;
459 min_dpb_size
+= width_in_mb
* height_in_mb
* 128;
461 /* IT surface buffer */
462 min_dpb_size
+= width_in_mb
* 64;
464 /* DB surface buffer */
465 min_dpb_size
+= width_in_mb
* 128;
468 tmp
= max(width_in_mb
, height_in_mb
);
469 min_dpb_size
+= ALIGN(tmp
* 7 * 16, 64);
474 /* reference picture buffer */
475 min_dpb_size
= image_size
* 3;
480 /* reference picture buffer */
481 min_dpb_size
= image_size
* 3;
484 min_dpb_size
+= width_in_mb
* height_in_mb
* 64;
486 /* IT surface buffer */
487 min_dpb_size
+= ALIGN(width_in_mb
* height_in_mb
* 32, 64);
491 image_size
= (ALIGN(width
, 16) * ALIGN(height
, 16) * 3) / 2;
492 image_size
= ALIGN(image_size
, 256);
494 num_dpb_buffer
= (le32_to_cpu(msg
[59]) & 0xff) + 2;
495 min_dpb_size
= image_size
* num_dpb_buffer
;
496 min_ctx_size
= ((width
+ 255) / 16) * ((height
+ 255) / 16)
497 * 16 * num_dpb_buffer
+ 52 * 1024;
501 DRM_ERROR("UVD codec not handled %d!\n", stream_type
);
506 DRM_ERROR("Invalid UVD decoding target pitch!\n");
510 if (dpb_size
< min_dpb_size
) {
511 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
512 dpb_size
, min_dpb_size
);
516 buf_sizes
[0x1] = dpb_size
;
517 buf_sizes
[0x2] = image_size
;
518 buf_sizes
[0x4] = min_ctx_size
;
523 * amdgpu_uvd_cs_msg - handle UVD message
525 * @ctx: UVD parser context
526 * @bo: buffer object containing the message
527 * @offset: offset into the buffer object
529 * Peek into the UVD message and extract the session id.
530 * Make sure that we don't open up to many sessions.
532 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx
*ctx
,
533 struct amdgpu_bo
*bo
, unsigned offset
)
535 struct amdgpu_device
*adev
= ctx
->parser
->adev
;
536 int32_t *msg
, msg_type
, handle
;
542 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
546 r
= amdgpu_bo_kmap(bo
, &ptr
);
548 DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r
);
558 DRM_ERROR("Invalid UVD handle!\n");
564 /* it's a create msg, calc image size (width * height) */
565 amdgpu_bo_kunmap(bo
);
567 /* try to alloc a new handle */
568 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
) {
569 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
570 DRM_ERROR("Handle 0x%x already in use!\n", handle
);
574 if (!atomic_cmpxchg(&adev
->uvd
.handles
[i
], 0, handle
)) {
575 adev
->uvd
.filp
[i
] = ctx
->parser
->filp
;
580 DRM_ERROR("No more free UVD handles!\n");
584 /* it's a decode msg, calc buffer sizes */
585 r
= amdgpu_uvd_cs_msg_decode(msg
, ctx
->buf_sizes
);
586 amdgpu_bo_kunmap(bo
);
590 /* validate the handle */
591 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
) {
592 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
593 if (adev
->uvd
.filp
[i
] != ctx
->parser
->filp
) {
594 DRM_ERROR("UVD handle collision detected!\n");
601 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle
);
605 /* it's a destroy msg, free the handle */
606 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
)
607 atomic_cmpxchg(&adev
->uvd
.handles
[i
], handle
, 0);
608 amdgpu_bo_kunmap(bo
);
612 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type
);
620 * amdgpu_uvd_cs_pass2 - second parsing round
622 * @ctx: UVD parser context
624 * Patch buffer addresses, make sure buffer sizes are correct.
626 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx
*ctx
)
628 struct amdgpu_bo_va_mapping
*mapping
;
629 struct amdgpu_bo
*bo
;
630 uint32_t cmd
, lo
, hi
;
635 lo
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
);
636 hi
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
);
637 addr
= ((uint64_t)lo
) | (((uint64_t)hi
) << 32);
639 mapping
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
);
643 start
= amdgpu_bo_gpu_offset(bo
);
645 end
= (mapping
->it
.last
+ 1 - mapping
->it
.start
);
646 end
= end
* AMDGPU_GPU_PAGE_SIZE
+ start
;
648 addr
-= ((uint64_t)mapping
->it
.start
) * AMDGPU_GPU_PAGE_SIZE
;
651 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
,
652 lower_32_bits(start
));
653 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
,
654 upper_32_bits(start
));
656 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
658 if ((end
- start
) < ctx
->buf_sizes
[cmd
]) {
659 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
660 (unsigned)(end
- start
),
661 ctx
->buf_sizes
[cmd
]);
665 } else if (cmd
== 0x206) {
666 if ((end
- start
) < ctx
->buf_sizes
[4]) {
667 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
668 (unsigned)(end
- start
),
672 } else if ((cmd
!= 0x100) && (cmd
!= 0x204)) {
673 DRM_ERROR("invalid UVD command %X!\n", cmd
);
677 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
678 if ((start
>> 28) != ((end
- 1) >> 28)) {
679 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
684 if ((cmd
== 0 || cmd
== 0x3) &&
685 (start
>> 28) != (ctx
->parser
->adev
->uvd
.gpu_addr
>> 28)) {
686 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
693 ctx
->has_msg_cmd
= true;
694 r
= amdgpu_uvd_cs_msg(ctx
, bo
, addr
);
697 } else if (!ctx
->has_msg_cmd
) {
698 DRM_ERROR("Message needed before other commands are send!\n");
706 * amdgpu_uvd_cs_reg - parse register writes
708 * @ctx: UVD parser context
709 * @cb: callback function
711 * Parse the register writes, call cb on each complete command.
713 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx
*ctx
,
714 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
716 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
720 for (i
= 0; i
<= ctx
->count
; ++i
) {
721 unsigned reg
= ctx
->reg
+ i
;
723 if (ctx
->idx
>= ib
->length_dw
) {
724 DRM_ERROR("Register command after end of CS!\n");
729 case mmUVD_GPCOM_VCPU_DATA0
:
730 ctx
->data0
= ctx
->idx
;
732 case mmUVD_GPCOM_VCPU_DATA1
:
733 ctx
->data1
= ctx
->idx
;
735 case mmUVD_GPCOM_VCPU_CMD
:
740 case mmUVD_ENGINE_CNTL
:
743 DRM_ERROR("Invalid reg 0x%X!\n", reg
);
752 * amdgpu_uvd_cs_packets - parse UVD packets
754 * @ctx: UVD parser context
755 * @cb: callback function
757 * Parse the command stream packets.
759 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx
*ctx
,
760 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
762 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
765 for (ctx
->idx
= 0 ; ctx
->idx
< ib
->length_dw
; ) {
766 uint32_t cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
);
767 unsigned type
= CP_PACKET_GET_TYPE(cmd
);
770 ctx
->reg
= CP_PACKET0_GET_REG(cmd
);
771 ctx
->count
= CP_PACKET_GET_COUNT(cmd
);
772 r
= amdgpu_uvd_cs_reg(ctx
, cb
);
780 DRM_ERROR("Unknown packet type %d !\n", type
);
788 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
790 * @parser: Command submission parser context
792 * Parse the command stream, patch in addresses as necessary.
794 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser
*parser
, uint32_t ib_idx
)
796 struct amdgpu_uvd_cs_ctx ctx
= {};
797 unsigned buf_sizes
[] = {
799 [0x00000001] = 0xFFFFFFFF,
800 [0x00000002] = 0xFFFFFFFF,
802 [0x00000004] = 0xFFFFFFFF,
804 struct amdgpu_ib
*ib
= &parser
->job
->ibs
[ib_idx
];
807 if (ib
->length_dw
% 16) {
808 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
814 ctx
.buf_sizes
= buf_sizes
;
817 /* first round, make sure the buffers are actually in the UVD segment */
818 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass1
);
822 /* second round, patch buffer addresses into the command stream */
823 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass2
);
827 if (!ctx
.has_msg_cmd
) {
828 DRM_ERROR("UVD-IBs need a msg command!\n");
832 amdgpu_uvd_note_usage(ctx
.parser
->adev
);
837 static int amdgpu_uvd_send_msg(struct amdgpu_ring
*ring
, struct amdgpu_bo
*bo
,
838 bool direct
, struct fence
**fence
)
840 struct ttm_validate_buffer tv
;
841 struct ww_acquire_ctx ticket
;
842 struct list_head head
;
843 struct amdgpu_job
*job
;
844 struct amdgpu_ib
*ib
;
845 struct fence
*f
= NULL
;
846 struct amdgpu_device
*adev
= ring
->adev
;
850 memset(&tv
, 0, sizeof(tv
));
853 INIT_LIST_HEAD(&head
);
854 list_add(&tv
.head
, &head
);
856 r
= ttm_eu_reserve_buffers(&ticket
, &head
, true, NULL
);
860 if (!bo
->adev
->uvd
.address_64_bit
) {
861 amdgpu_ttm_placement_from_domain(bo
, AMDGPU_GEM_DOMAIN_VRAM
);
862 amdgpu_uvd_force_into_uvd_segment(bo
);
865 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, true, false);
869 r
= amdgpu_job_alloc_with_ib(adev
, 64, &job
);
874 addr
= amdgpu_bo_gpu_offset(bo
);
875 ib
->ptr
[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0
, 0);
877 ib
->ptr
[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1
, 0);
878 ib
->ptr
[3] = addr
>> 32;
879 ib
->ptr
[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD
, 0);
881 for (i
= 6; i
< 16; ++i
)
882 ib
->ptr
[i
] = PACKET2(0);
886 r
= amdgpu_ib_schedule(ring
, 1, ib
, NULL
, &f
);
891 amdgpu_job_free(job
);
893 r
= amdgpu_job_submit(job
, ring
, &adev
->uvd
.entity
,
894 AMDGPU_FENCE_OWNER_UNDEFINED
, &f
);
899 ttm_eu_fence_buffer_objects(&ticket
, &head
, f
);
902 *fence
= fence_get(f
);
903 amdgpu_bo_unref(&bo
);
909 amdgpu_job_free(job
);
912 ttm_eu_backoff_reservation(&ticket
, &head
);
916 /* multiple fence commands without any stream commands in between can
917 crash the vcpu so just try to emmit a dummy create/destroy msg to
919 int amdgpu_uvd_get_create_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
920 struct fence
**fence
)
922 struct amdgpu_device
*adev
= ring
->adev
;
923 struct amdgpu_bo
*bo
;
927 r
= amdgpu_bo_create(adev
, 1024, PAGE_SIZE
, true,
928 AMDGPU_GEM_DOMAIN_VRAM
,
929 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
934 r
= amdgpu_bo_reserve(bo
, false);
936 amdgpu_bo_unref(&bo
);
940 r
= amdgpu_bo_kmap(bo
, (void **)&msg
);
942 amdgpu_bo_unreserve(bo
);
943 amdgpu_bo_unref(&bo
);
947 /* stitch together an UVD create msg */
948 msg
[0] = cpu_to_le32(0x00000de4);
949 msg
[1] = cpu_to_le32(0x00000000);
950 msg
[2] = cpu_to_le32(handle
);
951 msg
[3] = cpu_to_le32(0x00000000);
952 msg
[4] = cpu_to_le32(0x00000000);
953 msg
[5] = cpu_to_le32(0x00000000);
954 msg
[6] = cpu_to_le32(0x00000000);
955 msg
[7] = cpu_to_le32(0x00000780);
956 msg
[8] = cpu_to_le32(0x00000440);
957 msg
[9] = cpu_to_le32(0x00000000);
958 msg
[10] = cpu_to_le32(0x01b37000);
959 for (i
= 11; i
< 1024; ++i
)
960 msg
[i
] = cpu_to_le32(0x0);
962 amdgpu_bo_kunmap(bo
);
963 amdgpu_bo_unreserve(bo
);
965 return amdgpu_uvd_send_msg(ring
, bo
, true, fence
);
968 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
969 bool direct
, struct fence
**fence
)
971 struct amdgpu_device
*adev
= ring
->adev
;
972 struct amdgpu_bo
*bo
;
976 r
= amdgpu_bo_create(adev
, 1024, PAGE_SIZE
, true,
977 AMDGPU_GEM_DOMAIN_VRAM
,
978 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
983 r
= amdgpu_bo_reserve(bo
, false);
985 amdgpu_bo_unref(&bo
);
989 r
= amdgpu_bo_kmap(bo
, (void **)&msg
);
991 amdgpu_bo_unreserve(bo
);
992 amdgpu_bo_unref(&bo
);
996 /* stitch together an UVD destroy msg */
997 msg
[0] = cpu_to_le32(0x00000de4);
998 msg
[1] = cpu_to_le32(0x00000002);
999 msg
[2] = cpu_to_le32(handle
);
1000 msg
[3] = cpu_to_le32(0x00000000);
1001 for (i
= 4; i
< 1024; ++i
)
1002 msg
[i
] = cpu_to_le32(0x0);
1004 amdgpu_bo_kunmap(bo
);
1005 amdgpu_bo_unreserve(bo
);
1007 return amdgpu_uvd_send_msg(ring
, bo
, direct
, fence
);
1010 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
)
1012 struct amdgpu_device
*adev
=
1013 container_of(work
, struct amdgpu_device
, uvd
.idle_work
.work
);
1014 unsigned i
, fences
, handles
= 0;
1016 fences
= amdgpu_fence_count_emitted(&adev
->uvd
.ring
);
1018 for (i
= 0; i
< AMDGPU_MAX_UVD_HANDLES
; ++i
)
1019 if (atomic_read(&adev
->uvd
.handles
[i
]))
1022 if (fences
== 0 && handles
== 0) {
1023 if (adev
->pm
.dpm_enabled
) {
1024 amdgpu_dpm_enable_uvd(adev
, false);
1026 amdgpu_asic_set_uvd_clocks(adev
, 0, 0);
1029 schedule_delayed_work(&adev
->uvd
.idle_work
,
1030 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS
));
1034 static void amdgpu_uvd_note_usage(struct amdgpu_device
*adev
)
1036 bool set_clocks
= !cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
1037 set_clocks
&= schedule_delayed_work(&adev
->uvd
.idle_work
,
1038 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS
));
1041 if (adev
->pm
.dpm_enabled
) {
1042 amdgpu_dpm_enable_uvd(adev
, true);
1044 amdgpu_asic_set_uvd_clocks(adev
, 53300, 40000);