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 #include "amdgpu_ras.h"
44 /* 1 second timeout */
45 #define UVD_IDLE_TIMEOUT msecs_to_jiffies(1000)
47 /* Firmware versions for VI */
48 #define FW_1_65_10 ((1 << 24) | (65 << 16) | (10 << 8))
49 #define FW_1_87_11 ((1 << 24) | (87 << 16) | (11 << 8))
50 #define FW_1_87_12 ((1 << 24) | (87 << 16) | (12 << 8))
51 #define FW_1_37_15 ((1 << 24) | (37 << 16) | (15 << 8))
53 /* Polaris10/11 firmware version */
54 #define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
57 #ifdef CONFIG_DRM_AMDGPU_SI
58 #define FIRMWARE_TAHITI "amdgpu/tahiti_uvd.bin"
59 #define FIRMWARE_VERDE "amdgpu/verde_uvd.bin"
60 #define FIRMWARE_PITCAIRN "amdgpu/pitcairn_uvd.bin"
61 #define FIRMWARE_OLAND "amdgpu/oland_uvd.bin"
63 #ifdef CONFIG_DRM_AMDGPU_CIK
64 #define FIRMWARE_BONAIRE "amdgpu/bonaire_uvd.bin"
65 #define FIRMWARE_KABINI "amdgpu/kabini_uvd.bin"
66 #define FIRMWARE_KAVERI "amdgpu/kaveri_uvd.bin"
67 #define FIRMWARE_HAWAII "amdgpu/hawaii_uvd.bin"
68 #define FIRMWARE_MULLINS "amdgpu/mullins_uvd.bin"
70 #define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
71 #define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
72 #define FIRMWARE_FIJI "amdgpu/fiji_uvd.bin"
73 #define FIRMWARE_STONEY "amdgpu/stoney_uvd.bin"
74 #define FIRMWARE_POLARIS10 "amdgpu/polaris10_uvd.bin"
75 #define FIRMWARE_POLARIS11 "amdgpu/polaris11_uvd.bin"
76 #define FIRMWARE_POLARIS12 "amdgpu/polaris12_uvd.bin"
77 #define FIRMWARE_VEGAM "amdgpu/vegam_uvd.bin"
79 #define FIRMWARE_VEGA10 "amdgpu/vega10_uvd.bin"
80 #define FIRMWARE_VEGA12 "amdgpu/vega12_uvd.bin"
81 #define FIRMWARE_VEGA20 "amdgpu/vega20_uvd.bin"
83 /* These are common relative offsets for all asics, from uvd_7_0_offset.h, */
84 #define UVD_GPCOM_VCPU_CMD 0x03c3
85 #define UVD_GPCOM_VCPU_DATA0 0x03c4
86 #define UVD_GPCOM_VCPU_DATA1 0x03c5
87 #define UVD_NO_OP 0x03ff
88 #define UVD_BASE_SI 0x3800
91 * amdgpu_uvd_cs_ctx - Command submission parser context
93 * Used for emulating virtual memory support on UVD 4.2.
95 struct amdgpu_uvd_cs_ctx
{
96 struct amdgpu_cs_parser
*parser
;
98 unsigned data0
, data1
;
102 /* does the IB has a msg command */
105 /* minimum buffer sizes */
109 #ifdef CONFIG_DRM_AMDGPU_SI
110 MODULE_FIRMWARE(FIRMWARE_TAHITI
);
111 MODULE_FIRMWARE(FIRMWARE_VERDE
);
112 MODULE_FIRMWARE(FIRMWARE_PITCAIRN
);
113 MODULE_FIRMWARE(FIRMWARE_OLAND
);
115 #ifdef CONFIG_DRM_AMDGPU_CIK
116 MODULE_FIRMWARE(FIRMWARE_BONAIRE
);
117 MODULE_FIRMWARE(FIRMWARE_KABINI
);
118 MODULE_FIRMWARE(FIRMWARE_KAVERI
);
119 MODULE_FIRMWARE(FIRMWARE_HAWAII
);
120 MODULE_FIRMWARE(FIRMWARE_MULLINS
);
122 MODULE_FIRMWARE(FIRMWARE_TONGA
);
123 MODULE_FIRMWARE(FIRMWARE_CARRIZO
);
124 MODULE_FIRMWARE(FIRMWARE_FIJI
);
125 MODULE_FIRMWARE(FIRMWARE_STONEY
);
126 MODULE_FIRMWARE(FIRMWARE_POLARIS10
);
127 MODULE_FIRMWARE(FIRMWARE_POLARIS11
);
128 MODULE_FIRMWARE(FIRMWARE_POLARIS12
);
129 MODULE_FIRMWARE(FIRMWARE_VEGAM
);
131 MODULE_FIRMWARE(FIRMWARE_VEGA10
);
132 MODULE_FIRMWARE(FIRMWARE_VEGA12
);
133 MODULE_FIRMWARE(FIRMWARE_VEGA20
);
135 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
);
137 int amdgpu_uvd_sw_init(struct amdgpu_device
*adev
)
139 unsigned long bo_size
;
141 const struct common_firmware_header
*hdr
;
145 INIT_DELAYED_WORK(&adev
->uvd
.idle_work
, amdgpu_uvd_idle_work_handler
);
147 switch (adev
->asic_type
) {
148 #ifdef CONFIG_DRM_AMDGPU_SI
150 fw_name
= FIRMWARE_TAHITI
;
153 fw_name
= FIRMWARE_VERDE
;
156 fw_name
= FIRMWARE_PITCAIRN
;
159 fw_name
= FIRMWARE_OLAND
;
162 #ifdef CONFIG_DRM_AMDGPU_CIK
164 fw_name
= FIRMWARE_BONAIRE
;
167 fw_name
= FIRMWARE_KABINI
;
170 fw_name
= FIRMWARE_KAVERI
;
173 fw_name
= FIRMWARE_HAWAII
;
176 fw_name
= FIRMWARE_MULLINS
;
180 fw_name
= FIRMWARE_TONGA
;
183 fw_name
= FIRMWARE_FIJI
;
186 fw_name
= FIRMWARE_CARRIZO
;
189 fw_name
= FIRMWARE_STONEY
;
192 fw_name
= FIRMWARE_POLARIS10
;
195 fw_name
= FIRMWARE_POLARIS11
;
198 fw_name
= FIRMWARE_POLARIS12
;
201 fw_name
= FIRMWARE_VEGA10
;
204 fw_name
= FIRMWARE_VEGA12
;
207 fw_name
= FIRMWARE_VEGAM
;
210 fw_name
= FIRMWARE_VEGA20
;
216 r
= request_firmware(&adev
->uvd
.fw
, fw_name
, adev
->dev
);
218 dev_err(adev
->dev
, "amdgpu_uvd: Can't load firmware \"%s\"\n",
223 r
= amdgpu_ucode_validate(adev
->uvd
.fw
);
225 dev_err(adev
->dev
, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
227 release_firmware(adev
->uvd
.fw
);
232 /* Set the default UVD handles that the firmware can handle */
233 adev
->uvd
.max_handles
= AMDGPU_DEFAULT_UVD_HANDLES
;
235 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
236 family_id
= le32_to_cpu(hdr
->ucode_version
) & 0xff;
238 if (adev
->asic_type
< CHIP_VEGA20
) {
239 unsigned version_major
, version_minor
;
241 version_major
= (le32_to_cpu(hdr
->ucode_version
) >> 24) & 0xff;
242 version_minor
= (le32_to_cpu(hdr
->ucode_version
) >> 8) & 0xff;
243 DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
244 version_major
, version_minor
, family_id
);
247 * Limit the number of UVD handles depending on microcode major
248 * and minor versions. The firmware version which has 40 UVD
249 * instances support is 1.80. So all subsequent versions should
250 * also have the same support.
252 if ((version_major
> 0x01) ||
253 ((version_major
== 0x01) && (version_minor
>= 0x50)))
254 adev
->uvd
.max_handles
= AMDGPU_MAX_UVD_HANDLES
;
256 adev
->uvd
.fw_version
= ((version_major
<< 24) | (version_minor
<< 16) |
259 if ((adev
->asic_type
== CHIP_POLARIS10
||
260 adev
->asic_type
== CHIP_POLARIS11
) &&
261 (adev
->uvd
.fw_version
< FW_1_66_16
))
262 DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
263 version_major
, version_minor
);
265 unsigned int enc_major
, enc_minor
, dec_minor
;
267 dec_minor
= (le32_to_cpu(hdr
->ucode_version
) >> 8) & 0xff;
268 enc_minor
= (le32_to_cpu(hdr
->ucode_version
) >> 24) & 0x3f;
269 enc_major
= (le32_to_cpu(hdr
->ucode_version
) >> 30) & 0x3;
270 DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
271 enc_major
, enc_minor
, dec_minor
, family_id
);
273 adev
->uvd
.max_handles
= AMDGPU_MAX_UVD_HANDLES
;
275 adev
->uvd
.fw_version
= le32_to_cpu(hdr
->ucode_version
);
278 bo_size
= AMDGPU_UVD_STACK_SIZE
+ AMDGPU_UVD_HEAP_SIZE
279 + AMDGPU_UVD_SESSION_SIZE
* adev
->uvd
.max_handles
;
280 if (adev
->firmware
.load_type
!= AMDGPU_FW_LOAD_PSP
)
281 bo_size
+= AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr
->ucode_size_bytes
) + 8);
283 for (j
= 0; j
< adev
->uvd
.num_uvd_inst
; j
++) {
284 if (adev
->uvd
.harvest_config
& (1 << j
))
286 r
= amdgpu_bo_create_kernel(adev
, bo_size
, PAGE_SIZE
,
287 AMDGPU_GEM_DOMAIN_VRAM
, &adev
->uvd
.inst
[j
].vcpu_bo
,
288 &adev
->uvd
.inst
[j
].gpu_addr
, &adev
->uvd
.inst
[j
].cpu_addr
);
290 dev_err(adev
->dev
, "(%d) failed to allocate UVD bo\n", r
);
295 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
296 atomic_set(&adev
->uvd
.handles
[i
], 0);
297 adev
->uvd
.filp
[i
] = NULL
;
300 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
301 if (!amdgpu_device_ip_block_version_cmp(adev
, AMD_IP_BLOCK_TYPE_UVD
, 5, 0))
302 adev
->uvd
.address_64_bit
= true;
304 switch (adev
->asic_type
) {
306 adev
->uvd
.use_ctx_buf
= adev
->uvd
.fw_version
>= FW_1_65_10
;
309 adev
->uvd
.use_ctx_buf
= adev
->uvd
.fw_version
>= FW_1_87_11
;
312 adev
->uvd
.use_ctx_buf
= adev
->uvd
.fw_version
>= FW_1_87_12
;
315 adev
->uvd
.use_ctx_buf
= adev
->uvd
.fw_version
>= FW_1_37_15
;
318 adev
->uvd
.use_ctx_buf
= adev
->asic_type
>= CHIP_POLARIS10
;
324 int amdgpu_uvd_sw_fini(struct amdgpu_device
*adev
)
328 cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
329 drm_sched_entity_destroy(&adev
->uvd
.entity
);
331 for (j
= 0; j
< adev
->uvd
.num_uvd_inst
; ++j
) {
332 if (adev
->uvd
.harvest_config
& (1 << j
))
334 kvfree(adev
->uvd
.inst
[j
].saved_bo
);
336 amdgpu_bo_free_kernel(&adev
->uvd
.inst
[j
].vcpu_bo
,
337 &adev
->uvd
.inst
[j
].gpu_addr
,
338 (void **)&adev
->uvd
.inst
[j
].cpu_addr
);
340 amdgpu_ring_fini(&adev
->uvd
.inst
[j
].ring
);
342 for (i
= 0; i
< AMDGPU_MAX_UVD_ENC_RINGS
; ++i
)
343 amdgpu_ring_fini(&adev
->uvd
.inst
[j
].ring_enc
[i
]);
345 release_firmware(adev
->uvd
.fw
);
351 * amdgpu_uvd_entity_init - init entity
353 * @adev: amdgpu_device pointer
356 int amdgpu_uvd_entity_init(struct amdgpu_device
*adev
)
358 struct amdgpu_ring
*ring
;
359 struct drm_gpu_scheduler
*sched
;
362 ring
= &adev
->uvd
.inst
[0].ring
;
363 sched
= &ring
->sched
;
364 r
= drm_sched_entity_init(&adev
->uvd
.entity
, DRM_SCHED_PRIORITY_NORMAL
,
367 DRM_ERROR("Failed setting up UVD kernel entity.\n");
374 int amdgpu_uvd_suspend(struct amdgpu_device
*adev
)
379 bool in_ras_intr
= amdgpu_ras_intr_triggered();
381 cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
383 /* only valid for physical mode */
384 if (adev
->asic_type
< CHIP_POLARIS10
) {
385 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
)
386 if (atomic_read(&adev
->uvd
.handles
[i
]))
389 if (i
== adev
->uvd
.max_handles
)
393 for (j
= 0; j
< adev
->uvd
.num_uvd_inst
; ++j
) {
394 if (adev
->uvd
.harvest_config
& (1 << j
))
396 if (adev
->uvd
.inst
[j
].vcpu_bo
== NULL
)
399 size
= amdgpu_bo_size(adev
->uvd
.inst
[j
].vcpu_bo
);
400 ptr
= adev
->uvd
.inst
[j
].cpu_addr
;
402 adev
->uvd
.inst
[j
].saved_bo
= kvmalloc(size
, GFP_KERNEL
);
403 if (!adev
->uvd
.inst
[j
].saved_bo
)
406 /* re-write 0 since err_event_athub will corrupt VCPU buffer */
408 memset(adev
->uvd
.inst
[j
].saved_bo
, 0, size
);
410 memcpy_fromio(adev
->uvd
.inst
[j
].saved_bo
, ptr
, size
);
414 DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
419 int amdgpu_uvd_resume(struct amdgpu_device
*adev
)
425 for (i
= 0; i
< adev
->uvd
.num_uvd_inst
; i
++) {
426 if (adev
->uvd
.harvest_config
& (1 << i
))
428 if (adev
->uvd
.inst
[i
].vcpu_bo
== NULL
)
431 size
= amdgpu_bo_size(adev
->uvd
.inst
[i
].vcpu_bo
);
432 ptr
= adev
->uvd
.inst
[i
].cpu_addr
;
434 if (adev
->uvd
.inst
[i
].saved_bo
!= NULL
) {
435 memcpy_toio(ptr
, adev
->uvd
.inst
[i
].saved_bo
, size
);
436 kvfree(adev
->uvd
.inst
[i
].saved_bo
);
437 adev
->uvd
.inst
[i
].saved_bo
= NULL
;
439 const struct common_firmware_header
*hdr
;
442 hdr
= (const struct common_firmware_header
*)adev
->uvd
.fw
->data
;
443 if (adev
->firmware
.load_type
!= AMDGPU_FW_LOAD_PSP
) {
444 offset
= le32_to_cpu(hdr
->ucode_array_offset_bytes
);
445 memcpy_toio(adev
->uvd
.inst
[i
].cpu_addr
, adev
->uvd
.fw
->data
+ offset
,
446 le32_to_cpu(hdr
->ucode_size_bytes
));
447 size
-= le32_to_cpu(hdr
->ucode_size_bytes
);
448 ptr
+= le32_to_cpu(hdr
->ucode_size_bytes
);
450 memset_io(ptr
, 0, size
);
451 /* to restore uvd fence seq */
452 amdgpu_fence_driver_force_completion(&adev
->uvd
.inst
[i
].ring
);
458 void amdgpu_uvd_free_handles(struct amdgpu_device
*adev
, struct drm_file
*filp
)
460 struct amdgpu_ring
*ring
= &adev
->uvd
.inst
[0].ring
;
463 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
464 uint32_t handle
= atomic_read(&adev
->uvd
.handles
[i
]);
466 if (handle
!= 0 && adev
->uvd
.filp
[i
] == filp
) {
467 struct dma_fence
*fence
;
469 r
= amdgpu_uvd_get_destroy_msg(ring
, handle
, false,
472 DRM_ERROR("Error destroying UVD %d!\n", r
);
476 dma_fence_wait(fence
, false);
477 dma_fence_put(fence
);
479 adev
->uvd
.filp
[i
] = NULL
;
480 atomic_set(&adev
->uvd
.handles
[i
], 0);
485 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo
*abo
)
488 for (i
= 0; i
< abo
->placement
.num_placement
; ++i
) {
489 abo
->placements
[i
].fpfn
= 0 >> PAGE_SHIFT
;
490 abo
->placements
[i
].lpfn
= (256 * 1024 * 1024) >> PAGE_SHIFT
;
494 static u64
amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx
*ctx
)
499 lo
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
);
500 hi
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
);
501 addr
= ((uint64_t)lo
) | (((uint64_t)hi
) << 32);
507 * amdgpu_uvd_cs_pass1 - first parsing round
509 * @ctx: UVD parser context
511 * Make sure UVD message and feedback buffers are in VRAM and
512 * nobody is violating an 256MB boundary.
514 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx
*ctx
)
516 struct ttm_operation_ctx tctx
= { false, false };
517 struct amdgpu_bo_va_mapping
*mapping
;
518 struct amdgpu_bo
*bo
;
520 uint64_t addr
= amdgpu_uvd_get_addr_from_ctx(ctx
);
523 r
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
, &mapping
);
525 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr
);
529 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
530 /* check if it's a message or feedback command */
531 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
532 if (cmd
== 0x0 || cmd
== 0x3) {
533 /* yes, force it into VRAM */
534 uint32_t domain
= AMDGPU_GEM_DOMAIN_VRAM
;
535 amdgpu_bo_placement_from_domain(bo
, domain
);
537 amdgpu_uvd_force_into_uvd_segment(bo
);
539 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &tctx
);
546 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
548 * @adev: amdgpu_device pointer
549 * @msg: pointer to message structure
550 * @buf_sizes: placeholder to put the different buffer lengths
552 * Peek into the decode message and calculate the necessary buffer sizes.
554 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device
*adev
, uint32_t *msg
,
555 unsigned buf_sizes
[])
557 unsigned stream_type
= msg
[4];
558 unsigned width
= msg
[6];
559 unsigned height
= msg
[7];
560 unsigned dpb_size
= msg
[9];
561 unsigned pitch
= msg
[28];
562 unsigned level
= msg
[57];
564 unsigned width_in_mb
= width
/ 16;
565 unsigned height_in_mb
= ALIGN(height
/ 16, 2);
566 unsigned fs_in_mb
= width_in_mb
* height_in_mb
;
568 unsigned image_size
, tmp
, min_dpb_size
, num_dpb_buffer
;
569 unsigned min_ctx_size
= ~0;
571 image_size
= width
* height
;
572 image_size
+= image_size
/ 2;
573 image_size
= ALIGN(image_size
, 1024);
575 switch (stream_type
) {
579 num_dpb_buffer
= 8100 / fs_in_mb
;
582 num_dpb_buffer
= 18000 / fs_in_mb
;
585 num_dpb_buffer
= 20480 / fs_in_mb
;
588 num_dpb_buffer
= 32768 / fs_in_mb
;
591 num_dpb_buffer
= 34816 / fs_in_mb
;
594 num_dpb_buffer
= 110400 / fs_in_mb
;
597 num_dpb_buffer
= 184320 / fs_in_mb
;
600 num_dpb_buffer
= 184320 / fs_in_mb
;
604 if (num_dpb_buffer
> 17)
607 /* reference picture buffer */
608 min_dpb_size
= image_size
* num_dpb_buffer
;
610 /* macroblock context buffer */
611 min_dpb_size
+= width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
613 /* IT surface buffer */
614 min_dpb_size
+= width_in_mb
* height_in_mb
* 32;
619 /* reference picture buffer */
620 min_dpb_size
= image_size
* 3;
623 min_dpb_size
+= width_in_mb
* height_in_mb
* 128;
625 /* IT surface buffer */
626 min_dpb_size
+= width_in_mb
* 64;
628 /* DB surface buffer */
629 min_dpb_size
+= width_in_mb
* 128;
632 tmp
= max(width_in_mb
, height_in_mb
);
633 min_dpb_size
+= ALIGN(tmp
* 7 * 16, 64);
638 /* reference picture buffer */
639 min_dpb_size
= image_size
* 3;
644 /* reference picture buffer */
645 min_dpb_size
= image_size
* 3;
648 min_dpb_size
+= width_in_mb
* height_in_mb
* 64;
650 /* IT surface buffer */
651 min_dpb_size
+= ALIGN(width_in_mb
* height_in_mb
* 32, 64);
654 case 7: /* H264 Perf */
657 num_dpb_buffer
= 8100 / fs_in_mb
;
660 num_dpb_buffer
= 18000 / fs_in_mb
;
663 num_dpb_buffer
= 20480 / fs_in_mb
;
666 num_dpb_buffer
= 32768 / fs_in_mb
;
669 num_dpb_buffer
= 34816 / fs_in_mb
;
672 num_dpb_buffer
= 110400 / fs_in_mb
;
675 num_dpb_buffer
= 184320 / fs_in_mb
;
678 num_dpb_buffer
= 184320 / fs_in_mb
;
682 if (num_dpb_buffer
> 17)
685 /* reference picture buffer */
686 min_dpb_size
= image_size
* num_dpb_buffer
;
688 if (!adev
->uvd
.use_ctx_buf
){
689 /* macroblock context buffer */
691 width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
693 /* IT surface buffer */
694 min_dpb_size
+= width_in_mb
* height_in_mb
* 32;
696 /* macroblock context buffer */
698 width_in_mb
* height_in_mb
* num_dpb_buffer
* 192;
707 image_size
= (ALIGN(width
, 16) * ALIGN(height
, 16) * 3) / 2;
708 image_size
= ALIGN(image_size
, 256);
710 num_dpb_buffer
= (le32_to_cpu(msg
[59]) & 0xff) + 2;
711 min_dpb_size
= image_size
* num_dpb_buffer
;
712 min_ctx_size
= ((width
+ 255) / 16) * ((height
+ 255) / 16)
713 * 16 * num_dpb_buffer
+ 52 * 1024;
717 DRM_ERROR("UVD codec not handled %d!\n", stream_type
);
722 DRM_ERROR("Invalid UVD decoding target pitch!\n");
726 if (dpb_size
< min_dpb_size
) {
727 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
728 dpb_size
, min_dpb_size
);
732 buf_sizes
[0x1] = dpb_size
;
733 buf_sizes
[0x2] = image_size
;
734 buf_sizes
[0x4] = min_ctx_size
;
735 /* store image width to adjust nb memory pstate */
736 adev
->uvd
.decode_image_width
= width
;
741 * amdgpu_uvd_cs_msg - handle UVD message
743 * @ctx: UVD parser context
744 * @bo: buffer object containing the message
745 * @offset: offset into the buffer object
747 * Peek into the UVD message and extract the session id.
748 * Make sure that we don't open up to many sessions.
750 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx
*ctx
,
751 struct amdgpu_bo
*bo
, unsigned offset
)
753 struct amdgpu_device
*adev
= ctx
->parser
->adev
;
754 int32_t *msg
, msg_type
, handle
;
760 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
764 r
= amdgpu_bo_kmap(bo
, &ptr
);
766 DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r
);
776 DRM_ERROR("Invalid UVD handle!\n");
782 /* it's a create msg, calc image size (width * height) */
783 amdgpu_bo_kunmap(bo
);
785 /* try to alloc a new handle */
786 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
787 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
788 DRM_ERROR(")Handle 0x%x already in use!\n",
793 if (!atomic_cmpxchg(&adev
->uvd
.handles
[i
], 0, handle
)) {
794 adev
->uvd
.filp
[i
] = ctx
->parser
->filp
;
799 DRM_ERROR("No more free UVD handles!\n");
803 /* it's a decode msg, calc buffer sizes */
804 r
= amdgpu_uvd_cs_msg_decode(adev
, msg
, ctx
->buf_sizes
);
805 amdgpu_bo_kunmap(bo
);
809 /* validate the handle */
810 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
811 if (atomic_read(&adev
->uvd
.handles
[i
]) == handle
) {
812 if (adev
->uvd
.filp
[i
] != ctx
->parser
->filp
) {
813 DRM_ERROR("UVD handle collision detected!\n");
820 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle
);
824 /* it's a destroy msg, free the handle */
825 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
)
826 atomic_cmpxchg(&adev
->uvd
.handles
[i
], handle
, 0);
827 amdgpu_bo_kunmap(bo
);
831 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type
);
839 * amdgpu_uvd_cs_pass2 - second parsing round
841 * @ctx: UVD parser context
843 * Patch buffer addresses, make sure buffer sizes are correct.
845 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx
*ctx
)
847 struct amdgpu_bo_va_mapping
*mapping
;
848 struct amdgpu_bo
*bo
;
851 uint64_t addr
= amdgpu_uvd_get_addr_from_ctx(ctx
);
854 r
= amdgpu_cs_find_mapping(ctx
->parser
, addr
, &bo
, &mapping
);
856 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr
);
860 start
= amdgpu_bo_gpu_offset(bo
);
862 end
= (mapping
->last
+ 1 - mapping
->start
);
863 end
= end
* AMDGPU_GPU_PAGE_SIZE
+ start
;
865 addr
-= mapping
->start
* AMDGPU_GPU_PAGE_SIZE
;
868 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data0
,
869 lower_32_bits(start
));
870 amdgpu_set_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->data1
,
871 upper_32_bits(start
));
873 cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
) >> 1;
875 if ((end
- start
) < ctx
->buf_sizes
[cmd
]) {
876 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
877 (unsigned)(end
- start
),
878 ctx
->buf_sizes
[cmd
]);
882 } else if (cmd
== 0x206) {
883 if ((end
- start
) < ctx
->buf_sizes
[4]) {
884 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd
,
885 (unsigned)(end
- start
),
889 } else if ((cmd
!= 0x100) && (cmd
!= 0x204)) {
890 DRM_ERROR("invalid UVD command %X!\n", cmd
);
894 if (!ctx
->parser
->adev
->uvd
.address_64_bit
) {
895 if ((start
>> 28) != ((end
- 1) >> 28)) {
896 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
901 if ((cmd
== 0 || cmd
== 0x3) &&
902 (start
>> 28) != (ctx
->parser
->adev
->uvd
.inst
->gpu_addr
>> 28)) {
903 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
910 ctx
->has_msg_cmd
= true;
911 r
= amdgpu_uvd_cs_msg(ctx
, bo
, addr
);
914 } else if (!ctx
->has_msg_cmd
) {
915 DRM_ERROR("Message needed before other commands are send!\n");
923 * amdgpu_uvd_cs_reg - parse register writes
925 * @ctx: UVD parser context
926 * @cb: callback function
928 * Parse the register writes, call cb on each complete command.
930 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx
*ctx
,
931 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
933 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
937 for (i
= 0; i
<= ctx
->count
; ++i
) {
938 unsigned reg
= ctx
->reg
+ i
;
940 if (ctx
->idx
>= ib
->length_dw
) {
941 DRM_ERROR("Register command after end of CS!\n");
946 case mmUVD_GPCOM_VCPU_DATA0
:
947 ctx
->data0
= ctx
->idx
;
949 case mmUVD_GPCOM_VCPU_DATA1
:
950 ctx
->data1
= ctx
->idx
;
952 case mmUVD_GPCOM_VCPU_CMD
:
957 case mmUVD_ENGINE_CNTL
:
961 DRM_ERROR("Invalid reg 0x%X!\n", reg
);
970 * amdgpu_uvd_cs_packets - parse UVD packets
972 * @ctx: UVD parser context
973 * @cb: callback function
975 * Parse the command stream packets.
977 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx
*ctx
,
978 int (*cb
)(struct amdgpu_uvd_cs_ctx
*ctx
))
980 struct amdgpu_ib
*ib
= &ctx
->parser
->job
->ibs
[ctx
->ib_idx
];
983 for (ctx
->idx
= 0 ; ctx
->idx
< ib
->length_dw
; ) {
984 uint32_t cmd
= amdgpu_get_ib_value(ctx
->parser
, ctx
->ib_idx
, ctx
->idx
);
985 unsigned type
= CP_PACKET_GET_TYPE(cmd
);
988 ctx
->reg
= CP_PACKET0_GET_REG(cmd
);
989 ctx
->count
= CP_PACKET_GET_COUNT(cmd
);
990 r
= amdgpu_uvd_cs_reg(ctx
, cb
);
998 DRM_ERROR("Unknown packet type %d !\n", type
);
1006 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1008 * @parser: Command submission parser context
1009 * @ib_idx: Which indirect buffer to use
1011 * Parse the command stream, patch in addresses as necessary.
1013 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser
*parser
, uint32_t ib_idx
)
1015 struct amdgpu_uvd_cs_ctx ctx
= {};
1016 unsigned buf_sizes
[] = {
1017 [0x00000000] = 2048,
1018 [0x00000001] = 0xFFFFFFFF,
1019 [0x00000002] = 0xFFFFFFFF,
1020 [0x00000003] = 2048,
1021 [0x00000004] = 0xFFFFFFFF,
1023 struct amdgpu_ib
*ib
= &parser
->job
->ibs
[ib_idx
];
1026 parser
->job
->vm
= NULL
;
1027 ib
->gpu_addr
= amdgpu_sa_bo_gpu_addr(ib
->sa_bo
);
1029 if (ib
->length_dw
% 16) {
1030 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1035 ctx
.parser
= parser
;
1036 ctx
.buf_sizes
= buf_sizes
;
1037 ctx
.ib_idx
= ib_idx
;
1039 /* first round only required on chips without UVD 64 bit address support */
1040 if (!parser
->adev
->uvd
.address_64_bit
) {
1041 /* first round, make sure the buffers are actually in the UVD segment */
1042 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass1
);
1047 /* second round, patch buffer addresses into the command stream */
1048 r
= amdgpu_uvd_cs_packets(&ctx
, amdgpu_uvd_cs_pass2
);
1052 if (!ctx
.has_msg_cmd
) {
1053 DRM_ERROR("UVD-IBs need a msg command!\n");
1060 static int amdgpu_uvd_send_msg(struct amdgpu_ring
*ring
, struct amdgpu_bo
*bo
,
1061 bool direct
, struct dma_fence
**fence
)
1063 struct amdgpu_device
*adev
= ring
->adev
;
1064 struct dma_fence
*f
= NULL
;
1065 struct amdgpu_job
*job
;
1066 struct amdgpu_ib
*ib
;
1071 unsigned offset_idx
= 0;
1072 unsigned offset
[3] = { UVD_BASE_SI
, 0, 0 };
1074 amdgpu_bo_kunmap(bo
);
1075 amdgpu_bo_unpin(bo
);
1077 if (!ring
->adev
->uvd
.address_64_bit
) {
1078 struct ttm_operation_ctx ctx
= { true, false };
1080 amdgpu_bo_placement_from_domain(bo
, AMDGPU_GEM_DOMAIN_VRAM
);
1081 amdgpu_uvd_force_into_uvd_segment(bo
);
1082 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &ctx
);
1087 r
= amdgpu_job_alloc_with_ib(adev
, 64, direct
? AMDGPU_IB_POOL_DIRECT
:
1088 AMDGPU_IB_POOL_DELAYED
, &job
);
1092 if (adev
->asic_type
>= CHIP_VEGA10
) {
1093 offset_idx
= 1 + ring
->me
;
1094 offset
[1] = adev
->reg_offset
[UVD_HWIP
][0][1];
1095 offset
[2] = adev
->reg_offset
[UVD_HWIP
][1][1];
1098 data
[0] = PACKET0(offset
[offset_idx
] + UVD_GPCOM_VCPU_DATA0
, 0);
1099 data
[1] = PACKET0(offset
[offset_idx
] + UVD_GPCOM_VCPU_DATA1
, 0);
1100 data
[2] = PACKET0(offset
[offset_idx
] + UVD_GPCOM_VCPU_CMD
, 0);
1101 data
[3] = PACKET0(offset
[offset_idx
] + UVD_NO_OP
, 0);
1104 addr
= amdgpu_bo_gpu_offset(bo
);
1105 ib
->ptr
[0] = data
[0];
1107 ib
->ptr
[2] = data
[1];
1108 ib
->ptr
[3] = addr
>> 32;
1109 ib
->ptr
[4] = data
[2];
1111 for (i
= 6; i
< 16; i
+= 2) {
1112 ib
->ptr
[i
] = data
[3];
1118 r
= dma_resv_wait_timeout_rcu(bo
->tbo
.base
.resv
,
1120 msecs_to_jiffies(10));
1126 r
= amdgpu_job_submit_direct(job
, ring
, &f
);
1130 r
= amdgpu_sync_resv(adev
, &job
->sync
, bo
->tbo
.base
.resv
,
1132 AMDGPU_FENCE_OWNER_UNDEFINED
);
1136 r
= amdgpu_job_submit(job
, &adev
->uvd
.entity
,
1137 AMDGPU_FENCE_OWNER_UNDEFINED
, &f
);
1142 amdgpu_bo_fence(bo
, f
, false);
1143 amdgpu_bo_unreserve(bo
);
1144 amdgpu_bo_unref(&bo
);
1147 *fence
= dma_fence_get(f
);
1153 amdgpu_job_free(job
);
1156 amdgpu_bo_unreserve(bo
);
1157 amdgpu_bo_unref(&bo
);
1161 /* multiple fence commands without any stream commands in between can
1162 crash the vcpu so just try to emmit a dummy create/destroy msg to
1164 int amdgpu_uvd_get_create_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
1165 struct dma_fence
**fence
)
1167 struct amdgpu_device
*adev
= ring
->adev
;
1168 struct amdgpu_bo
*bo
= NULL
;
1172 r
= amdgpu_bo_create_reserved(adev
, 1024, PAGE_SIZE
,
1173 AMDGPU_GEM_DOMAIN_VRAM
,
1174 &bo
, NULL
, (void **)&msg
);
1178 /* stitch together an UVD create msg */
1179 msg
[0] = cpu_to_le32(0x00000de4);
1180 msg
[1] = cpu_to_le32(0x00000000);
1181 msg
[2] = cpu_to_le32(handle
);
1182 msg
[3] = cpu_to_le32(0x00000000);
1183 msg
[4] = cpu_to_le32(0x00000000);
1184 msg
[5] = cpu_to_le32(0x00000000);
1185 msg
[6] = cpu_to_le32(0x00000000);
1186 msg
[7] = cpu_to_le32(0x00000780);
1187 msg
[8] = cpu_to_le32(0x00000440);
1188 msg
[9] = cpu_to_le32(0x00000000);
1189 msg
[10] = cpu_to_le32(0x01b37000);
1190 for (i
= 11; i
< 1024; ++i
)
1191 msg
[i
] = cpu_to_le32(0x0);
1193 return amdgpu_uvd_send_msg(ring
, bo
, true, fence
);
1196 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring
*ring
, uint32_t handle
,
1197 bool direct
, struct dma_fence
**fence
)
1199 struct amdgpu_device
*adev
= ring
->adev
;
1200 struct amdgpu_bo
*bo
= NULL
;
1204 r
= amdgpu_bo_create_reserved(adev
, 1024, PAGE_SIZE
,
1205 AMDGPU_GEM_DOMAIN_VRAM
,
1206 &bo
, NULL
, (void **)&msg
);
1210 /* stitch together an UVD destroy msg */
1211 msg
[0] = cpu_to_le32(0x00000de4);
1212 msg
[1] = cpu_to_le32(0x00000002);
1213 msg
[2] = cpu_to_le32(handle
);
1214 msg
[3] = cpu_to_le32(0x00000000);
1215 for (i
= 4; i
< 1024; ++i
)
1216 msg
[i
] = cpu_to_le32(0x0);
1218 return amdgpu_uvd_send_msg(ring
, bo
, direct
, fence
);
1221 static void amdgpu_uvd_idle_work_handler(struct work_struct
*work
)
1223 struct amdgpu_device
*adev
=
1224 container_of(work
, struct amdgpu_device
, uvd
.idle_work
.work
);
1225 unsigned fences
= 0, i
, j
;
1227 for (i
= 0; i
< adev
->uvd
.num_uvd_inst
; ++i
) {
1228 if (adev
->uvd
.harvest_config
& (1 << i
))
1230 fences
+= amdgpu_fence_count_emitted(&adev
->uvd
.inst
[i
].ring
);
1231 for (j
= 0; j
< adev
->uvd
.num_enc_rings
; ++j
) {
1232 fences
+= amdgpu_fence_count_emitted(&adev
->uvd
.inst
[i
].ring_enc
[j
]);
1237 if (adev
->pm
.dpm_enabled
) {
1238 amdgpu_dpm_enable_uvd(adev
, false);
1240 amdgpu_asic_set_uvd_clocks(adev
, 0, 0);
1241 /* shutdown the UVD block */
1242 amdgpu_device_ip_set_powergating_state(adev
, AMD_IP_BLOCK_TYPE_UVD
,
1244 amdgpu_device_ip_set_clockgating_state(adev
, AMD_IP_BLOCK_TYPE_UVD
,
1248 schedule_delayed_work(&adev
->uvd
.idle_work
, UVD_IDLE_TIMEOUT
);
1252 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring
*ring
)
1254 struct amdgpu_device
*adev
= ring
->adev
;
1257 if (amdgpu_sriov_vf(adev
))
1260 set_clocks
= !cancel_delayed_work_sync(&adev
->uvd
.idle_work
);
1262 if (adev
->pm
.dpm_enabled
) {
1263 amdgpu_dpm_enable_uvd(adev
, true);
1265 amdgpu_asic_set_uvd_clocks(adev
, 53300, 40000);
1266 amdgpu_device_ip_set_clockgating_state(adev
, AMD_IP_BLOCK_TYPE_UVD
,
1267 AMD_CG_STATE_UNGATE
);
1268 amdgpu_device_ip_set_powergating_state(adev
, AMD_IP_BLOCK_TYPE_UVD
,
1269 AMD_PG_STATE_UNGATE
);
1274 void amdgpu_uvd_ring_end_use(struct amdgpu_ring
*ring
)
1276 if (!amdgpu_sriov_vf(ring
->adev
))
1277 schedule_delayed_work(&ring
->adev
->uvd
.idle_work
, UVD_IDLE_TIMEOUT
);
1281 * amdgpu_uvd_ring_test_ib - test ib execution
1283 * @ring: amdgpu_ring pointer
1284 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1286 * Test if we can successfully execute an IB
1288 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring
*ring
, long timeout
)
1290 struct dma_fence
*fence
;
1293 r
= amdgpu_uvd_get_create_msg(ring
, 1, NULL
);
1297 r
= amdgpu_uvd_get_destroy_msg(ring
, 1, true, &fence
);
1301 r
= dma_fence_wait_timeout(fence
, false, timeout
);
1307 dma_fence_put(fence
);
1314 * amdgpu_uvd_used_handles - returns used UVD handles
1316 * @adev: amdgpu_device pointer
1318 * Returns the number of UVD handles in use
1320 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device
*adev
)
1323 uint32_t used_handles
= 0;
1325 for (i
= 0; i
< adev
->uvd
.max_handles
; ++i
) {
1327 * Handles can be freed in any order, and not
1328 * necessarily linear. So we need to count
1329 * all non-zero handles.
1331 if (atomic_read(&adev
->uvd
.handles
[i
]))
1335 return used_handles
;