Loading...
1/*
2 * Copyright 2011 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
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:
12 *
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.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Christian König <deathsimple@vodafone.de>
29 */
30
31#include <linux/firmware.h>
32#include <linux/module.h>
33
34#include <drm/drm.h>
35
36#include "amdgpu.h"
37#include "amdgpu_pm.h"
38#include "amdgpu_uvd.h"
39#include "cikd.h"
40#include "uvd/uvd_4_2_d.h"
41
42/* 1 second timeout */
43#define UVD_IDLE_TIMEOUT msecs_to_jiffies(1000)
44
45/* Firmware versions for VI */
46#define FW_1_65_10 ((1 << 24) | (65 << 16) | (10 << 8))
47#define FW_1_87_11 ((1 << 24) | (87 << 16) | (11 << 8))
48#define FW_1_87_12 ((1 << 24) | (87 << 16) | (12 << 8))
49#define FW_1_37_15 ((1 << 24) | (37 << 16) | (15 << 8))
50
51/* Polaris10/11 firmware version */
52#define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
53
54/* Firmware Names */
55#ifdef CONFIG_DRM_AMDGPU_CIK
56#define FIRMWARE_BONAIRE "amdgpu/bonaire_uvd.bin"
57#define FIRMWARE_KABINI "amdgpu/kabini_uvd.bin"
58#define FIRMWARE_KAVERI "amdgpu/kaveri_uvd.bin"
59#define FIRMWARE_HAWAII "amdgpu/hawaii_uvd.bin"
60#define FIRMWARE_MULLINS "amdgpu/mullins_uvd.bin"
61#endif
62#define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
63#define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
64#define FIRMWARE_FIJI "amdgpu/fiji_uvd.bin"
65#define FIRMWARE_STONEY "amdgpu/stoney_uvd.bin"
66#define FIRMWARE_POLARIS10 "amdgpu/polaris10_uvd.bin"
67#define FIRMWARE_POLARIS11 "amdgpu/polaris11_uvd.bin"
68#define FIRMWARE_POLARIS12 "amdgpu/polaris12_uvd.bin"
69#define FIRMWARE_VEGAM "amdgpu/vegam_uvd.bin"
70
71#define FIRMWARE_VEGA10 "amdgpu/vega10_uvd.bin"
72#define FIRMWARE_VEGA12 "amdgpu/vega12_uvd.bin"
73#define FIRMWARE_VEGA20 "amdgpu/vega20_uvd.bin"
74
75/* These are common relative offsets for all asics, from uvd_7_0_offset.h, */
76#define UVD_GPCOM_VCPU_CMD 0x03c3
77#define UVD_GPCOM_VCPU_DATA0 0x03c4
78#define UVD_GPCOM_VCPU_DATA1 0x03c5
79#define UVD_NO_OP 0x03ff
80#define UVD_BASE_SI 0x3800
81
82/**
83 * amdgpu_uvd_cs_ctx - Command submission parser context
84 *
85 * Used for emulating virtual memory support on UVD 4.2.
86 */
87struct amdgpu_uvd_cs_ctx {
88 struct amdgpu_cs_parser *parser;
89 unsigned reg, count;
90 unsigned data0, data1;
91 unsigned idx;
92 unsigned ib_idx;
93
94 /* does the IB has a msg command */
95 bool has_msg_cmd;
96
97 /* minimum buffer sizes */
98 unsigned *buf_sizes;
99};
100
101#ifdef CONFIG_DRM_AMDGPU_CIK
102MODULE_FIRMWARE(FIRMWARE_BONAIRE);
103MODULE_FIRMWARE(FIRMWARE_KABINI);
104MODULE_FIRMWARE(FIRMWARE_KAVERI);
105MODULE_FIRMWARE(FIRMWARE_HAWAII);
106MODULE_FIRMWARE(FIRMWARE_MULLINS);
107#endif
108MODULE_FIRMWARE(FIRMWARE_TONGA);
109MODULE_FIRMWARE(FIRMWARE_CARRIZO);
110MODULE_FIRMWARE(FIRMWARE_FIJI);
111MODULE_FIRMWARE(FIRMWARE_STONEY);
112MODULE_FIRMWARE(FIRMWARE_POLARIS10);
113MODULE_FIRMWARE(FIRMWARE_POLARIS11);
114MODULE_FIRMWARE(FIRMWARE_POLARIS12);
115MODULE_FIRMWARE(FIRMWARE_VEGAM);
116
117MODULE_FIRMWARE(FIRMWARE_VEGA10);
118MODULE_FIRMWARE(FIRMWARE_VEGA12);
119MODULE_FIRMWARE(FIRMWARE_VEGA20);
120
121static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
122
123int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
124{
125 unsigned long bo_size;
126 const char *fw_name;
127 const struct common_firmware_header *hdr;
128 unsigned family_id;
129 int i, j, r;
130
131 INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
132
133 switch (adev->asic_type) {
134#ifdef CONFIG_DRM_AMDGPU_CIK
135 case CHIP_BONAIRE:
136 fw_name = FIRMWARE_BONAIRE;
137 break;
138 case CHIP_KABINI:
139 fw_name = FIRMWARE_KABINI;
140 break;
141 case CHIP_KAVERI:
142 fw_name = FIRMWARE_KAVERI;
143 break;
144 case CHIP_HAWAII:
145 fw_name = FIRMWARE_HAWAII;
146 break;
147 case CHIP_MULLINS:
148 fw_name = FIRMWARE_MULLINS;
149 break;
150#endif
151 case CHIP_TONGA:
152 fw_name = FIRMWARE_TONGA;
153 break;
154 case CHIP_FIJI:
155 fw_name = FIRMWARE_FIJI;
156 break;
157 case CHIP_CARRIZO:
158 fw_name = FIRMWARE_CARRIZO;
159 break;
160 case CHIP_STONEY:
161 fw_name = FIRMWARE_STONEY;
162 break;
163 case CHIP_POLARIS10:
164 fw_name = FIRMWARE_POLARIS10;
165 break;
166 case CHIP_POLARIS11:
167 fw_name = FIRMWARE_POLARIS11;
168 break;
169 case CHIP_POLARIS12:
170 fw_name = FIRMWARE_POLARIS12;
171 break;
172 case CHIP_VEGA10:
173 fw_name = FIRMWARE_VEGA10;
174 break;
175 case CHIP_VEGA12:
176 fw_name = FIRMWARE_VEGA12;
177 break;
178 case CHIP_VEGAM:
179 fw_name = FIRMWARE_VEGAM;
180 break;
181 case CHIP_VEGA20:
182 fw_name = FIRMWARE_VEGA20;
183 break;
184 default:
185 return -EINVAL;
186 }
187
188 r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
189 if (r) {
190 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
191 fw_name);
192 return r;
193 }
194
195 r = amdgpu_ucode_validate(adev->uvd.fw);
196 if (r) {
197 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
198 fw_name);
199 release_firmware(adev->uvd.fw);
200 adev->uvd.fw = NULL;
201 return r;
202 }
203
204 /* Set the default UVD handles that the firmware can handle */
205 adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
206
207 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
208 family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
209
210 if (adev->asic_type < CHIP_VEGA20) {
211 unsigned version_major, version_minor;
212
213 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
214 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
215 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
216 version_major, version_minor, family_id);
217
218 /*
219 * Limit the number of UVD handles depending on microcode major
220 * and minor versions. The firmware version which has 40 UVD
221 * instances support is 1.80. So all subsequent versions should
222 * also have the same support.
223 */
224 if ((version_major > 0x01) ||
225 ((version_major == 0x01) && (version_minor >= 0x50)))
226 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
227
228 adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
229 (family_id << 8));
230
231 if ((adev->asic_type == CHIP_POLARIS10 ||
232 adev->asic_type == CHIP_POLARIS11) &&
233 (adev->uvd.fw_version < FW_1_66_16))
234 DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
235 version_major, version_minor);
236 } else {
237 unsigned int enc_major, enc_minor, dec_minor;
238
239 dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
240 enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
241 enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
242 DRM_INFO("Found UVD firmware ENC: %hu.%hu DEC: .%hu Family ID: %hu\n",
243 enc_major, enc_minor, dec_minor, family_id);
244
245 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
246
247 adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
248 }
249
250 bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
251 + AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
252 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
253 bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
254
255 for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
256 if (adev->uvd.harvest_config & (1 << j))
257 continue;
258 r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
259 AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
260 &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
261 if (r) {
262 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
263 return r;
264 }
265 }
266
267 for (i = 0; i < adev->uvd.max_handles; ++i) {
268 atomic_set(&adev->uvd.handles[i], 0);
269 adev->uvd.filp[i] = NULL;
270 }
271
272 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
273 if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
274 adev->uvd.address_64_bit = true;
275
276 switch (adev->asic_type) {
277 case CHIP_TONGA:
278 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
279 break;
280 case CHIP_CARRIZO:
281 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
282 break;
283 case CHIP_FIJI:
284 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
285 break;
286 case CHIP_STONEY:
287 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
288 break;
289 default:
290 adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
291 }
292
293 return 0;
294}
295
296int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
297{
298 int i, j;
299
300 drm_sched_entity_destroy(&adev->uvd.entity);
301
302 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
303 if (adev->uvd.harvest_config & (1 << j))
304 continue;
305 kvfree(adev->uvd.inst[j].saved_bo);
306
307 amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
308 &adev->uvd.inst[j].gpu_addr,
309 (void **)&adev->uvd.inst[j].cpu_addr);
310
311 amdgpu_ring_fini(&adev->uvd.inst[j].ring);
312
313 for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
314 amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
315 }
316 release_firmware(adev->uvd.fw);
317
318 return 0;
319}
320
321/**
322 * amdgpu_uvd_entity_init - init entity
323 *
324 * @adev: amdgpu_device pointer
325 *
326 */
327int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
328{
329 struct amdgpu_ring *ring;
330 struct drm_sched_rq *rq;
331 int r;
332
333 ring = &adev->uvd.inst[0].ring;
334 rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
335 r = drm_sched_entity_init(&adev->uvd.entity, &rq, 1, NULL);
336 if (r) {
337 DRM_ERROR("Failed setting up UVD kernel entity.\n");
338 return r;
339 }
340
341 return 0;
342}
343
344int amdgpu_uvd_suspend(struct amdgpu_device *adev)
345{
346 unsigned size;
347 void *ptr;
348 int i, j;
349
350 cancel_delayed_work_sync(&adev->uvd.idle_work);
351
352 /* only valid for physical mode */
353 if (adev->asic_type < CHIP_POLARIS10) {
354 for (i = 0; i < adev->uvd.max_handles; ++i)
355 if (atomic_read(&adev->uvd.handles[i]))
356 break;
357
358 if (i == adev->uvd.max_handles)
359 return 0;
360 }
361
362 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
363 if (adev->uvd.harvest_config & (1 << j))
364 continue;
365 if (adev->uvd.inst[j].vcpu_bo == NULL)
366 continue;
367
368 size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
369 ptr = adev->uvd.inst[j].cpu_addr;
370
371 adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
372 if (!adev->uvd.inst[j].saved_bo)
373 return -ENOMEM;
374
375 memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
376 }
377 return 0;
378}
379
380int amdgpu_uvd_resume(struct amdgpu_device *adev)
381{
382 unsigned size;
383 void *ptr;
384 int i;
385
386 for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
387 if (adev->uvd.harvest_config & (1 << i))
388 continue;
389 if (adev->uvd.inst[i].vcpu_bo == NULL)
390 return -EINVAL;
391
392 size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
393 ptr = adev->uvd.inst[i].cpu_addr;
394
395 if (adev->uvd.inst[i].saved_bo != NULL) {
396 memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
397 kvfree(adev->uvd.inst[i].saved_bo);
398 adev->uvd.inst[i].saved_bo = NULL;
399 } else {
400 const struct common_firmware_header *hdr;
401 unsigned offset;
402
403 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
404 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
405 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
406 memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
407 le32_to_cpu(hdr->ucode_size_bytes));
408 size -= le32_to_cpu(hdr->ucode_size_bytes);
409 ptr += le32_to_cpu(hdr->ucode_size_bytes);
410 }
411 memset_io(ptr, 0, size);
412 /* to restore uvd fence seq */
413 amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
414 }
415 }
416 return 0;
417}
418
419void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
420{
421 struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
422 int i, r;
423
424 for (i = 0; i < adev->uvd.max_handles; ++i) {
425 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
426
427 if (handle != 0 && adev->uvd.filp[i] == filp) {
428 struct dma_fence *fence;
429
430 r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
431 &fence);
432 if (r) {
433 DRM_ERROR("Error destroying UVD %d!\n", r);
434 continue;
435 }
436
437 dma_fence_wait(fence, false);
438 dma_fence_put(fence);
439
440 adev->uvd.filp[i] = NULL;
441 atomic_set(&adev->uvd.handles[i], 0);
442 }
443 }
444}
445
446static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
447{
448 int i;
449 for (i = 0; i < abo->placement.num_placement; ++i) {
450 abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
451 abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
452 }
453}
454
455static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
456{
457 uint32_t lo, hi;
458 uint64_t addr;
459
460 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
461 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
462 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
463
464 return addr;
465}
466
467/**
468 * amdgpu_uvd_cs_pass1 - first parsing round
469 *
470 * @ctx: UVD parser context
471 *
472 * Make sure UVD message and feedback buffers are in VRAM and
473 * nobody is violating an 256MB boundary.
474 */
475static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
476{
477 struct ttm_operation_ctx tctx = { false, false };
478 struct amdgpu_bo_va_mapping *mapping;
479 struct amdgpu_bo *bo;
480 uint32_t cmd;
481 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
482 int r = 0;
483
484 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
485 if (r) {
486 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
487 return r;
488 }
489
490 if (!ctx->parser->adev->uvd.address_64_bit) {
491 /* check if it's a message or feedback command */
492 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
493 if (cmd == 0x0 || cmd == 0x3) {
494 /* yes, force it into VRAM */
495 uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
496 amdgpu_bo_placement_from_domain(bo, domain);
497 }
498 amdgpu_uvd_force_into_uvd_segment(bo);
499
500 r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
501 }
502
503 return r;
504}
505
506/**
507 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
508 *
509 * @msg: pointer to message structure
510 * @buf_sizes: returned buffer sizes
511 *
512 * Peek into the decode message and calculate the necessary buffer sizes.
513 */
514static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
515 unsigned buf_sizes[])
516{
517 unsigned stream_type = msg[4];
518 unsigned width = msg[6];
519 unsigned height = msg[7];
520 unsigned dpb_size = msg[9];
521 unsigned pitch = msg[28];
522 unsigned level = msg[57];
523
524 unsigned width_in_mb = width / 16;
525 unsigned height_in_mb = ALIGN(height / 16, 2);
526 unsigned fs_in_mb = width_in_mb * height_in_mb;
527
528 unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
529 unsigned min_ctx_size = ~0;
530
531 image_size = width * height;
532 image_size += image_size / 2;
533 image_size = ALIGN(image_size, 1024);
534
535 switch (stream_type) {
536 case 0: /* H264 */
537 switch(level) {
538 case 30:
539 num_dpb_buffer = 8100 / fs_in_mb;
540 break;
541 case 31:
542 num_dpb_buffer = 18000 / fs_in_mb;
543 break;
544 case 32:
545 num_dpb_buffer = 20480 / fs_in_mb;
546 break;
547 case 41:
548 num_dpb_buffer = 32768 / fs_in_mb;
549 break;
550 case 42:
551 num_dpb_buffer = 34816 / fs_in_mb;
552 break;
553 case 50:
554 num_dpb_buffer = 110400 / fs_in_mb;
555 break;
556 case 51:
557 num_dpb_buffer = 184320 / fs_in_mb;
558 break;
559 default:
560 num_dpb_buffer = 184320 / fs_in_mb;
561 break;
562 }
563 num_dpb_buffer++;
564 if (num_dpb_buffer > 17)
565 num_dpb_buffer = 17;
566
567 /* reference picture buffer */
568 min_dpb_size = image_size * num_dpb_buffer;
569
570 /* macroblock context buffer */
571 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
572
573 /* IT surface buffer */
574 min_dpb_size += width_in_mb * height_in_mb * 32;
575 break;
576
577 case 1: /* VC1 */
578
579 /* reference picture buffer */
580 min_dpb_size = image_size * 3;
581
582 /* CONTEXT_BUFFER */
583 min_dpb_size += width_in_mb * height_in_mb * 128;
584
585 /* IT surface buffer */
586 min_dpb_size += width_in_mb * 64;
587
588 /* DB surface buffer */
589 min_dpb_size += width_in_mb * 128;
590
591 /* BP */
592 tmp = max(width_in_mb, height_in_mb);
593 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
594 break;
595
596 case 3: /* MPEG2 */
597
598 /* reference picture buffer */
599 min_dpb_size = image_size * 3;
600 break;
601
602 case 4: /* MPEG4 */
603
604 /* reference picture buffer */
605 min_dpb_size = image_size * 3;
606
607 /* CM */
608 min_dpb_size += width_in_mb * height_in_mb * 64;
609
610 /* IT surface buffer */
611 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
612 break;
613
614 case 7: /* H264 Perf */
615 switch(level) {
616 case 30:
617 num_dpb_buffer = 8100 / fs_in_mb;
618 break;
619 case 31:
620 num_dpb_buffer = 18000 / fs_in_mb;
621 break;
622 case 32:
623 num_dpb_buffer = 20480 / fs_in_mb;
624 break;
625 case 41:
626 num_dpb_buffer = 32768 / fs_in_mb;
627 break;
628 case 42:
629 num_dpb_buffer = 34816 / fs_in_mb;
630 break;
631 case 50:
632 num_dpb_buffer = 110400 / fs_in_mb;
633 break;
634 case 51:
635 num_dpb_buffer = 184320 / fs_in_mb;
636 break;
637 default:
638 num_dpb_buffer = 184320 / fs_in_mb;
639 break;
640 }
641 num_dpb_buffer++;
642 if (num_dpb_buffer > 17)
643 num_dpb_buffer = 17;
644
645 /* reference picture buffer */
646 min_dpb_size = image_size * num_dpb_buffer;
647
648 if (!adev->uvd.use_ctx_buf){
649 /* macroblock context buffer */
650 min_dpb_size +=
651 width_in_mb * height_in_mb * num_dpb_buffer * 192;
652
653 /* IT surface buffer */
654 min_dpb_size += width_in_mb * height_in_mb * 32;
655 } else {
656 /* macroblock context buffer */
657 min_ctx_size =
658 width_in_mb * height_in_mb * num_dpb_buffer * 192;
659 }
660 break;
661
662 case 8: /* MJPEG */
663 min_dpb_size = 0;
664 break;
665
666 case 16: /* H265 */
667 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
668 image_size = ALIGN(image_size, 256);
669
670 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
671 min_dpb_size = image_size * num_dpb_buffer;
672 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
673 * 16 * num_dpb_buffer + 52 * 1024;
674 break;
675
676 default:
677 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
678 return -EINVAL;
679 }
680
681 if (width > pitch) {
682 DRM_ERROR("Invalid UVD decoding target pitch!\n");
683 return -EINVAL;
684 }
685
686 if (dpb_size < min_dpb_size) {
687 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
688 dpb_size, min_dpb_size);
689 return -EINVAL;
690 }
691
692 buf_sizes[0x1] = dpb_size;
693 buf_sizes[0x2] = image_size;
694 buf_sizes[0x4] = min_ctx_size;
695 /* store image width to adjust nb memory pstate */
696 adev->uvd.decode_image_width = width;
697 return 0;
698}
699
700/**
701 * amdgpu_uvd_cs_msg - handle UVD message
702 *
703 * @ctx: UVD parser context
704 * @bo: buffer object containing the message
705 * @offset: offset into the buffer object
706 *
707 * Peek into the UVD message and extract the session id.
708 * Make sure that we don't open up to many sessions.
709 */
710static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
711 struct amdgpu_bo *bo, unsigned offset)
712{
713 struct amdgpu_device *adev = ctx->parser->adev;
714 int32_t *msg, msg_type, handle;
715 void *ptr;
716 long r;
717 int i;
718
719 if (offset & 0x3F) {
720 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
721 return -EINVAL;
722 }
723
724 r = amdgpu_bo_kmap(bo, &ptr);
725 if (r) {
726 DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
727 return r;
728 }
729
730 msg = ptr + offset;
731
732 msg_type = msg[1];
733 handle = msg[2];
734
735 if (handle == 0) {
736 DRM_ERROR("Invalid UVD handle!\n");
737 return -EINVAL;
738 }
739
740 switch (msg_type) {
741 case 0:
742 /* it's a create msg, calc image size (width * height) */
743 amdgpu_bo_kunmap(bo);
744
745 /* try to alloc a new handle */
746 for (i = 0; i < adev->uvd.max_handles; ++i) {
747 if (atomic_read(&adev->uvd.handles[i]) == handle) {
748 DRM_ERROR(")Handle 0x%x already in use!\n",
749 handle);
750 return -EINVAL;
751 }
752
753 if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
754 adev->uvd.filp[i] = ctx->parser->filp;
755 return 0;
756 }
757 }
758
759 DRM_ERROR("No more free UVD handles!\n");
760 return -ENOSPC;
761
762 case 1:
763 /* it's a decode msg, calc buffer sizes */
764 r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
765 amdgpu_bo_kunmap(bo);
766 if (r)
767 return r;
768
769 /* validate the handle */
770 for (i = 0; i < adev->uvd.max_handles; ++i) {
771 if (atomic_read(&adev->uvd.handles[i]) == handle) {
772 if (adev->uvd.filp[i] != ctx->parser->filp) {
773 DRM_ERROR("UVD handle collision detected!\n");
774 return -EINVAL;
775 }
776 return 0;
777 }
778 }
779
780 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
781 return -ENOENT;
782
783 case 2:
784 /* it's a destroy msg, free the handle */
785 for (i = 0; i < adev->uvd.max_handles; ++i)
786 atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
787 amdgpu_bo_kunmap(bo);
788 return 0;
789
790 default:
791 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
792 return -EINVAL;
793 }
794 BUG();
795 return -EINVAL;
796}
797
798/**
799 * amdgpu_uvd_cs_pass2 - second parsing round
800 *
801 * @ctx: UVD parser context
802 *
803 * Patch buffer addresses, make sure buffer sizes are correct.
804 */
805static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
806{
807 struct amdgpu_bo_va_mapping *mapping;
808 struct amdgpu_bo *bo;
809 uint32_t cmd;
810 uint64_t start, end;
811 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
812 int r;
813
814 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
815 if (r) {
816 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
817 return r;
818 }
819
820 start = amdgpu_bo_gpu_offset(bo);
821
822 end = (mapping->last + 1 - mapping->start);
823 end = end * AMDGPU_GPU_PAGE_SIZE + start;
824
825 addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
826 start += addr;
827
828 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
829 lower_32_bits(start));
830 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
831 upper_32_bits(start));
832
833 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
834 if (cmd < 0x4) {
835 if ((end - start) < ctx->buf_sizes[cmd]) {
836 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
837 (unsigned)(end - start),
838 ctx->buf_sizes[cmd]);
839 return -EINVAL;
840 }
841
842 } else if (cmd == 0x206) {
843 if ((end - start) < ctx->buf_sizes[4]) {
844 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
845 (unsigned)(end - start),
846 ctx->buf_sizes[4]);
847 return -EINVAL;
848 }
849 } else if ((cmd != 0x100) && (cmd != 0x204)) {
850 DRM_ERROR("invalid UVD command %X!\n", cmd);
851 return -EINVAL;
852 }
853
854 if (!ctx->parser->adev->uvd.address_64_bit) {
855 if ((start >> 28) != ((end - 1) >> 28)) {
856 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
857 start, end);
858 return -EINVAL;
859 }
860
861 if ((cmd == 0 || cmd == 0x3) &&
862 (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
863 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
864 start, end);
865 return -EINVAL;
866 }
867 }
868
869 if (cmd == 0) {
870 ctx->has_msg_cmd = true;
871 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
872 if (r)
873 return r;
874 } else if (!ctx->has_msg_cmd) {
875 DRM_ERROR("Message needed before other commands are send!\n");
876 return -EINVAL;
877 }
878
879 return 0;
880}
881
882/**
883 * amdgpu_uvd_cs_reg - parse register writes
884 *
885 * @ctx: UVD parser context
886 * @cb: callback function
887 *
888 * Parse the register writes, call cb on each complete command.
889 */
890static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
891 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
892{
893 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
894 int i, r;
895
896 ctx->idx++;
897 for (i = 0; i <= ctx->count; ++i) {
898 unsigned reg = ctx->reg + i;
899
900 if (ctx->idx >= ib->length_dw) {
901 DRM_ERROR("Register command after end of CS!\n");
902 return -EINVAL;
903 }
904
905 switch (reg) {
906 case mmUVD_GPCOM_VCPU_DATA0:
907 ctx->data0 = ctx->idx;
908 break;
909 case mmUVD_GPCOM_VCPU_DATA1:
910 ctx->data1 = ctx->idx;
911 break;
912 case mmUVD_GPCOM_VCPU_CMD:
913 r = cb(ctx);
914 if (r)
915 return r;
916 break;
917 case mmUVD_ENGINE_CNTL:
918 case mmUVD_NO_OP:
919 break;
920 default:
921 DRM_ERROR("Invalid reg 0x%X!\n", reg);
922 return -EINVAL;
923 }
924 ctx->idx++;
925 }
926 return 0;
927}
928
929/**
930 * amdgpu_uvd_cs_packets - parse UVD packets
931 *
932 * @ctx: UVD parser context
933 * @cb: callback function
934 *
935 * Parse the command stream packets.
936 */
937static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
938 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
939{
940 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
941 int r;
942
943 for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
944 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
945 unsigned type = CP_PACKET_GET_TYPE(cmd);
946 switch (type) {
947 case PACKET_TYPE0:
948 ctx->reg = CP_PACKET0_GET_REG(cmd);
949 ctx->count = CP_PACKET_GET_COUNT(cmd);
950 r = amdgpu_uvd_cs_reg(ctx, cb);
951 if (r)
952 return r;
953 break;
954 case PACKET_TYPE2:
955 ++ctx->idx;
956 break;
957 default:
958 DRM_ERROR("Unknown packet type %d !\n", type);
959 return -EINVAL;
960 }
961 }
962 return 0;
963}
964
965/**
966 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
967 *
968 * @parser: Command submission parser context
969 *
970 * Parse the command stream, patch in addresses as necessary.
971 */
972int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
973{
974 struct amdgpu_uvd_cs_ctx ctx = {};
975 unsigned buf_sizes[] = {
976 [0x00000000] = 2048,
977 [0x00000001] = 0xFFFFFFFF,
978 [0x00000002] = 0xFFFFFFFF,
979 [0x00000003] = 2048,
980 [0x00000004] = 0xFFFFFFFF,
981 };
982 struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
983 int r;
984
985 parser->job->vm = NULL;
986 ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
987
988 if (ib->length_dw % 16) {
989 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
990 ib->length_dw);
991 return -EINVAL;
992 }
993
994 ctx.parser = parser;
995 ctx.buf_sizes = buf_sizes;
996 ctx.ib_idx = ib_idx;
997
998 /* first round only required on chips without UVD 64 bit address support */
999 if (!parser->adev->uvd.address_64_bit) {
1000 /* first round, make sure the buffers are actually in the UVD segment */
1001 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1002 if (r)
1003 return r;
1004 }
1005
1006 /* second round, patch buffer addresses into the command stream */
1007 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1008 if (r)
1009 return r;
1010
1011 if (!ctx.has_msg_cmd) {
1012 DRM_ERROR("UVD-IBs need a msg command!\n");
1013 return -EINVAL;
1014 }
1015
1016 return 0;
1017}
1018
1019static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1020 bool direct, struct dma_fence **fence)
1021{
1022 struct amdgpu_device *adev = ring->adev;
1023 struct dma_fence *f = NULL;
1024 struct amdgpu_job *job;
1025 struct amdgpu_ib *ib;
1026 uint32_t data[4];
1027 uint64_t addr;
1028 long r;
1029 int i;
1030 unsigned offset_idx = 0;
1031 unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1032
1033 amdgpu_bo_kunmap(bo);
1034 amdgpu_bo_unpin(bo);
1035
1036 if (!ring->adev->uvd.address_64_bit) {
1037 struct ttm_operation_ctx ctx = { true, false };
1038
1039 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1040 amdgpu_uvd_force_into_uvd_segment(bo);
1041 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1042 if (r)
1043 goto err;
1044 }
1045
1046 r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1047 if (r)
1048 goto err;
1049
1050 if (adev->asic_type >= CHIP_VEGA10) {
1051 offset_idx = 1 + ring->me;
1052 offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1053 offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1054 }
1055
1056 data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1057 data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1058 data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1059 data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1060
1061 ib = &job->ibs[0];
1062 addr = amdgpu_bo_gpu_offset(bo);
1063 ib->ptr[0] = data[0];
1064 ib->ptr[1] = addr;
1065 ib->ptr[2] = data[1];
1066 ib->ptr[3] = addr >> 32;
1067 ib->ptr[4] = data[2];
1068 ib->ptr[5] = 0;
1069 for (i = 6; i < 16; i += 2) {
1070 ib->ptr[i] = data[3];
1071 ib->ptr[i+1] = 0;
1072 }
1073 ib->length_dw = 16;
1074
1075 if (direct) {
1076 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv,
1077 true, false,
1078 msecs_to_jiffies(10));
1079 if (r == 0)
1080 r = -ETIMEDOUT;
1081 if (r < 0)
1082 goto err_free;
1083
1084 r = amdgpu_job_submit_direct(job, ring, &f);
1085 if (r)
1086 goto err_free;
1087 } else {
1088 r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1089 AMDGPU_FENCE_OWNER_UNDEFINED, false);
1090 if (r)
1091 goto err_free;
1092
1093 r = amdgpu_job_submit(job, &adev->uvd.entity,
1094 AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1095 if (r)
1096 goto err_free;
1097 }
1098
1099 amdgpu_bo_fence(bo, f, false);
1100 amdgpu_bo_unreserve(bo);
1101 amdgpu_bo_unref(&bo);
1102
1103 if (fence)
1104 *fence = dma_fence_get(f);
1105 dma_fence_put(f);
1106
1107 return 0;
1108
1109err_free:
1110 amdgpu_job_free(job);
1111
1112err:
1113 amdgpu_bo_unreserve(bo);
1114 amdgpu_bo_unref(&bo);
1115 return r;
1116}
1117
1118/* multiple fence commands without any stream commands in between can
1119 crash the vcpu so just try to emmit a dummy create/destroy msg to
1120 avoid this */
1121int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1122 struct dma_fence **fence)
1123{
1124 struct amdgpu_device *adev = ring->adev;
1125 struct amdgpu_bo *bo = NULL;
1126 uint32_t *msg;
1127 int r, i;
1128
1129 r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1130 AMDGPU_GEM_DOMAIN_VRAM,
1131 &bo, NULL, (void **)&msg);
1132 if (r)
1133 return r;
1134
1135 /* stitch together an UVD create msg */
1136 msg[0] = cpu_to_le32(0x00000de4);
1137 msg[1] = cpu_to_le32(0x00000000);
1138 msg[2] = cpu_to_le32(handle);
1139 msg[3] = cpu_to_le32(0x00000000);
1140 msg[4] = cpu_to_le32(0x00000000);
1141 msg[5] = cpu_to_le32(0x00000000);
1142 msg[6] = cpu_to_le32(0x00000000);
1143 msg[7] = cpu_to_le32(0x00000780);
1144 msg[8] = cpu_to_le32(0x00000440);
1145 msg[9] = cpu_to_le32(0x00000000);
1146 msg[10] = cpu_to_le32(0x01b37000);
1147 for (i = 11; i < 1024; ++i)
1148 msg[i] = cpu_to_le32(0x0);
1149
1150 return amdgpu_uvd_send_msg(ring, bo, true, fence);
1151}
1152
1153int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1154 bool direct, struct dma_fence **fence)
1155{
1156 struct amdgpu_device *adev = ring->adev;
1157 struct amdgpu_bo *bo = NULL;
1158 uint32_t *msg;
1159 int r, i;
1160
1161 r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1162 AMDGPU_GEM_DOMAIN_VRAM,
1163 &bo, NULL, (void **)&msg);
1164 if (r)
1165 return r;
1166
1167 /* stitch together an UVD destroy msg */
1168 msg[0] = cpu_to_le32(0x00000de4);
1169 msg[1] = cpu_to_le32(0x00000002);
1170 msg[2] = cpu_to_le32(handle);
1171 msg[3] = cpu_to_le32(0x00000000);
1172 for (i = 4; i < 1024; ++i)
1173 msg[i] = cpu_to_le32(0x0);
1174
1175 return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1176}
1177
1178static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1179{
1180 struct amdgpu_device *adev =
1181 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1182 unsigned fences = 0, i, j;
1183
1184 for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1185 if (adev->uvd.harvest_config & (1 << i))
1186 continue;
1187 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1188 for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1189 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1190 }
1191 }
1192
1193 if (fences == 0) {
1194 if (adev->pm.dpm_enabled) {
1195 amdgpu_dpm_enable_uvd(adev, false);
1196 } else {
1197 amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1198 /* shutdown the UVD block */
1199 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1200 AMD_PG_STATE_GATE);
1201 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1202 AMD_CG_STATE_GATE);
1203 }
1204 } else {
1205 schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1206 }
1207}
1208
1209void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1210{
1211 struct amdgpu_device *adev = ring->adev;
1212 bool set_clocks;
1213
1214 if (amdgpu_sriov_vf(adev))
1215 return;
1216
1217 set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1218 if (set_clocks) {
1219 if (adev->pm.dpm_enabled) {
1220 amdgpu_dpm_enable_uvd(adev, true);
1221 } else {
1222 amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1223 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1224 AMD_CG_STATE_UNGATE);
1225 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1226 AMD_PG_STATE_UNGATE);
1227 }
1228 }
1229}
1230
1231void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1232{
1233 if (!amdgpu_sriov_vf(ring->adev))
1234 schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1235}
1236
1237/**
1238 * amdgpu_uvd_ring_test_ib - test ib execution
1239 *
1240 * @ring: amdgpu_ring pointer
1241 *
1242 * Test if we can successfully execute an IB
1243 */
1244int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1245{
1246 struct dma_fence *fence;
1247 long r;
1248
1249 r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1250 if (r)
1251 goto error;
1252
1253 r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1254 if (r)
1255 goto error;
1256
1257 r = dma_fence_wait_timeout(fence, false, timeout);
1258 if (r == 0)
1259 r = -ETIMEDOUT;
1260 else if (r > 0)
1261 r = 0;
1262
1263 dma_fence_put(fence);
1264
1265error:
1266 return r;
1267}
1268
1269/**
1270 * amdgpu_uvd_used_handles - returns used UVD handles
1271 *
1272 * @adev: amdgpu_device pointer
1273 *
1274 * Returns the number of UVD handles in use
1275 */
1276uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1277{
1278 unsigned i;
1279 uint32_t used_handles = 0;
1280
1281 for (i = 0; i < adev->uvd.max_handles; ++i) {
1282 /*
1283 * Handles can be freed in any order, and not
1284 * necessarily linear. So we need to count
1285 * all non-zero handles.
1286 */
1287 if (atomic_read(&adev->uvd.handles[i]))
1288 used_handles++;
1289 }
1290
1291 return used_handles;
1292}
1/*
2 * Copyright 2011 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
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:
12 *
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.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Christian König <deathsimple@vodafone.de>
29 */
30
31#include <linux/firmware.h>
32#include <linux/module.h>
33#include <drm/drmP.h>
34#include <drm/drm.h>
35
36#include "amdgpu.h"
37#include "amdgpu_pm.h"
38#include "amdgpu_uvd.h"
39#include "cikd.h"
40#include "uvd/uvd_4_2_d.h"
41
42/* 1 second timeout */
43#define UVD_IDLE_TIMEOUT_MS 1000
44
45/* Firmware Names */
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"
52#endif
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"
57
58/**
59 * amdgpu_uvd_cs_ctx - Command submission parser context
60 *
61 * Used for emulating virtual memory support on UVD 4.2.
62 */
63struct amdgpu_uvd_cs_ctx {
64 struct amdgpu_cs_parser *parser;
65 unsigned reg, count;
66 unsigned data0, data1;
67 unsigned idx;
68 unsigned ib_idx;
69
70 /* does the IB has a msg command */
71 bool has_msg_cmd;
72
73 /* minimum buffer sizes */
74 unsigned *buf_sizes;
75};
76
77#ifdef CONFIG_DRM_AMDGPU_CIK
78MODULE_FIRMWARE(FIRMWARE_BONAIRE);
79MODULE_FIRMWARE(FIRMWARE_KABINI);
80MODULE_FIRMWARE(FIRMWARE_KAVERI);
81MODULE_FIRMWARE(FIRMWARE_HAWAII);
82MODULE_FIRMWARE(FIRMWARE_MULLINS);
83#endif
84MODULE_FIRMWARE(FIRMWARE_TONGA);
85MODULE_FIRMWARE(FIRMWARE_CARRIZO);
86MODULE_FIRMWARE(FIRMWARE_FIJI);
87MODULE_FIRMWARE(FIRMWARE_STONEY);
88
89static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
90static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
91
92int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
93{
94 struct amdgpu_ring *ring;
95 struct amd_sched_rq *rq;
96 unsigned long bo_size;
97 const char *fw_name;
98 const struct common_firmware_header *hdr;
99 unsigned version_major, version_minor, family_id;
100 int i, r;
101
102 INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
103
104 switch (adev->asic_type) {
105#ifdef CONFIG_DRM_AMDGPU_CIK
106 case CHIP_BONAIRE:
107 fw_name = FIRMWARE_BONAIRE;
108 break;
109 case CHIP_KABINI:
110 fw_name = FIRMWARE_KABINI;
111 break;
112 case CHIP_KAVERI:
113 fw_name = FIRMWARE_KAVERI;
114 break;
115 case CHIP_HAWAII:
116 fw_name = FIRMWARE_HAWAII;
117 break;
118 case CHIP_MULLINS:
119 fw_name = FIRMWARE_MULLINS;
120 break;
121#endif
122 case CHIP_TONGA:
123 fw_name = FIRMWARE_TONGA;
124 break;
125 case CHIP_FIJI:
126 fw_name = FIRMWARE_FIJI;
127 break;
128 case CHIP_CARRIZO:
129 fw_name = FIRMWARE_CARRIZO;
130 break;
131 case CHIP_STONEY:
132 fw_name = FIRMWARE_STONEY;
133 break;
134 default:
135 return -EINVAL;
136 }
137
138 r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
139 if (r) {
140 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
141 fw_name);
142 return r;
143 }
144
145 r = amdgpu_ucode_validate(adev->uvd.fw);
146 if (r) {
147 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
148 fw_name);
149 release_firmware(adev->uvd.fw);
150 adev->uvd.fw = NULL;
151 return r;
152 }
153
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);
160
161 adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
162 (family_id << 8));
163
164 bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
165 + AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
166 r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
167 AMDGPU_GEM_DOMAIN_VRAM,
168 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
169 NULL, NULL, &adev->uvd.vcpu_bo);
170 if (r) {
171 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
172 return r;
173 }
174
175 r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
176 if (r) {
177 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
178 dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
179 return r;
180 }
181
182 r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
183 &adev->uvd.gpu_addr);
184 if (r) {
185 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
186 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
187 dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
188 return r;
189 }
190
191 r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
192 if (r) {
193 dev_err(adev->dev, "(%d) UVD map failed\n", r);
194 return r;
195 }
196
197 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
198
199 ring = &adev->uvd.ring;
200 rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
201 r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
202 rq, amdgpu_sched_jobs);
203 if (r != 0) {
204 DRM_ERROR("Failed setting up UVD run queue.\n");
205 return r;
206 }
207
208 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
209 atomic_set(&adev->uvd.handles[i], 0);
210 adev->uvd.filp[i] = NULL;
211 }
212
213 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
214 if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
215 adev->uvd.address_64_bit = true;
216
217 return 0;
218}
219
220int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
221{
222 int r;
223
224 if (adev->uvd.vcpu_bo == NULL)
225 return 0;
226
227 amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
228
229 r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
230 if (!r) {
231 amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
232 amdgpu_bo_unpin(adev->uvd.vcpu_bo);
233 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
234 }
235
236 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
237
238 amdgpu_ring_fini(&adev->uvd.ring);
239
240 release_firmware(adev->uvd.fw);
241
242 return 0;
243}
244
245int amdgpu_uvd_suspend(struct amdgpu_device *adev)
246{
247 unsigned size;
248 void *ptr;
249 int i;
250
251 if (adev->uvd.vcpu_bo == NULL)
252 return 0;
253
254 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
255 if (atomic_read(&adev->uvd.handles[i]))
256 break;
257
258 if (i == AMDGPU_MAX_UVD_HANDLES)
259 return 0;
260
261 cancel_delayed_work_sync(&adev->uvd.idle_work);
262
263 size = amdgpu_bo_size(adev->uvd.vcpu_bo);
264 ptr = adev->uvd.cpu_addr;
265
266 adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
267 if (!adev->uvd.saved_bo)
268 return -ENOMEM;
269
270 memcpy(adev->uvd.saved_bo, ptr, size);
271
272 return 0;
273}
274
275int amdgpu_uvd_resume(struct amdgpu_device *adev)
276{
277 unsigned size;
278 void *ptr;
279
280 if (adev->uvd.vcpu_bo == NULL)
281 return -EINVAL;
282
283 size = amdgpu_bo_size(adev->uvd.vcpu_bo);
284 ptr = adev->uvd.cpu_addr;
285
286 if (adev->uvd.saved_bo != NULL) {
287 memcpy(ptr, adev->uvd.saved_bo, size);
288 kfree(adev->uvd.saved_bo);
289 adev->uvd.saved_bo = NULL;
290 } else {
291 const struct common_firmware_header *hdr;
292 unsigned offset;
293
294 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
295 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
296 memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
297 (adev->uvd.fw->size) - offset);
298 size -= le32_to_cpu(hdr->ucode_size_bytes);
299 ptr += le32_to_cpu(hdr->ucode_size_bytes);
300 memset(ptr, 0, size);
301 }
302
303 return 0;
304}
305
306void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
307{
308 struct amdgpu_ring *ring = &adev->uvd.ring;
309 int i, r;
310
311 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
312 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
313 if (handle != 0 && adev->uvd.filp[i] == filp) {
314 struct fence *fence;
315
316 amdgpu_uvd_note_usage(adev);
317
318 r = amdgpu_uvd_get_destroy_msg(ring, handle,
319 false, &fence);
320 if (r) {
321 DRM_ERROR("Error destroying UVD (%d)!\n", r);
322 continue;
323 }
324
325 fence_wait(fence, false);
326 fence_put(fence);
327
328 adev->uvd.filp[i] = NULL;
329 atomic_set(&adev->uvd.handles[i], 0);
330 }
331 }
332}
333
334static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
335{
336 int i;
337 for (i = 0; i < rbo->placement.num_placement; ++i) {
338 rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
339 rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
340 }
341}
342
343/**
344 * amdgpu_uvd_cs_pass1 - first parsing round
345 *
346 * @ctx: UVD parser context
347 *
348 * Make sure UVD message and feedback buffers are in VRAM and
349 * nobody is violating an 256MB boundary.
350 */
351static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
352{
353 struct amdgpu_bo_va_mapping *mapping;
354 struct amdgpu_bo *bo;
355 uint32_t cmd, lo, hi;
356 uint64_t addr;
357 int r = 0;
358
359 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
360 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
361 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
362
363 mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
364 if (mapping == NULL) {
365 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
366 return -EINVAL;
367 }
368
369 if (!ctx->parser->adev->uvd.address_64_bit) {
370 /* check if it's a message or feedback command */
371 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
372 if (cmd == 0x0 || cmd == 0x3) {
373 /* yes, force it into VRAM */
374 uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
375 amdgpu_ttm_placement_from_domain(bo, domain);
376 }
377 amdgpu_uvd_force_into_uvd_segment(bo);
378
379 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
380 }
381
382 return r;
383}
384
385/**
386 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
387 *
388 * @msg: pointer to message structure
389 * @buf_sizes: returned buffer sizes
390 *
391 * Peek into the decode message and calculate the necessary buffer sizes.
392 */
393static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
394{
395 unsigned stream_type = msg[4];
396 unsigned width = msg[6];
397 unsigned height = msg[7];
398 unsigned dpb_size = msg[9];
399 unsigned pitch = msg[28];
400 unsigned level = msg[57];
401
402 unsigned width_in_mb = width / 16;
403 unsigned height_in_mb = ALIGN(height / 16, 2);
404 unsigned fs_in_mb = width_in_mb * height_in_mb;
405
406 unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
407 unsigned min_ctx_size = 0;
408
409 image_size = width * height;
410 image_size += image_size / 2;
411 image_size = ALIGN(image_size, 1024);
412
413 switch (stream_type) {
414 case 0: /* H264 */
415 case 7: /* H264 Perf */
416 switch(level) {
417 case 30:
418 num_dpb_buffer = 8100 / fs_in_mb;
419 break;
420 case 31:
421 num_dpb_buffer = 18000 / fs_in_mb;
422 break;
423 case 32:
424 num_dpb_buffer = 20480 / fs_in_mb;
425 break;
426 case 41:
427 num_dpb_buffer = 32768 / fs_in_mb;
428 break;
429 case 42:
430 num_dpb_buffer = 34816 / fs_in_mb;
431 break;
432 case 50:
433 num_dpb_buffer = 110400 / fs_in_mb;
434 break;
435 case 51:
436 num_dpb_buffer = 184320 / fs_in_mb;
437 break;
438 default:
439 num_dpb_buffer = 184320 / fs_in_mb;
440 break;
441 }
442 num_dpb_buffer++;
443 if (num_dpb_buffer > 17)
444 num_dpb_buffer = 17;
445
446 /* reference picture buffer */
447 min_dpb_size = image_size * num_dpb_buffer;
448
449 /* macroblock context buffer */
450 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
451
452 /* IT surface buffer */
453 min_dpb_size += width_in_mb * height_in_mb * 32;
454 break;
455
456 case 1: /* VC1 */
457
458 /* reference picture buffer */
459 min_dpb_size = image_size * 3;
460
461 /* CONTEXT_BUFFER */
462 min_dpb_size += width_in_mb * height_in_mb * 128;
463
464 /* IT surface buffer */
465 min_dpb_size += width_in_mb * 64;
466
467 /* DB surface buffer */
468 min_dpb_size += width_in_mb * 128;
469
470 /* BP */
471 tmp = max(width_in_mb, height_in_mb);
472 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
473 break;
474
475 case 3: /* MPEG2 */
476
477 /* reference picture buffer */
478 min_dpb_size = image_size * 3;
479 break;
480
481 case 4: /* MPEG4 */
482
483 /* reference picture buffer */
484 min_dpb_size = image_size * 3;
485
486 /* CM */
487 min_dpb_size += width_in_mb * height_in_mb * 64;
488
489 /* IT surface buffer */
490 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
491 break;
492
493 case 16: /* H265 */
494 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
495 image_size = ALIGN(image_size, 256);
496
497 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
498 min_dpb_size = image_size * num_dpb_buffer;
499 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
500 * 16 * num_dpb_buffer + 52 * 1024;
501 break;
502
503 default:
504 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
505 return -EINVAL;
506 }
507
508 if (width > pitch) {
509 DRM_ERROR("Invalid UVD decoding target pitch!\n");
510 return -EINVAL;
511 }
512
513 if (dpb_size < min_dpb_size) {
514 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
515 dpb_size, min_dpb_size);
516 return -EINVAL;
517 }
518
519 buf_sizes[0x1] = dpb_size;
520 buf_sizes[0x2] = image_size;
521 buf_sizes[0x4] = min_ctx_size;
522 return 0;
523}
524
525/**
526 * amdgpu_uvd_cs_msg - handle UVD message
527 *
528 * @ctx: UVD parser context
529 * @bo: buffer object containing the message
530 * @offset: offset into the buffer object
531 *
532 * Peek into the UVD message and extract the session id.
533 * Make sure that we don't open up to many sessions.
534 */
535static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
536 struct amdgpu_bo *bo, unsigned offset)
537{
538 struct amdgpu_device *adev = ctx->parser->adev;
539 int32_t *msg, msg_type, handle;
540 void *ptr;
541 long r;
542 int i;
543
544 if (offset & 0x3F) {
545 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
546 return -EINVAL;
547 }
548
549 r = amdgpu_bo_kmap(bo, &ptr);
550 if (r) {
551 DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
552 return r;
553 }
554
555 msg = ptr + offset;
556
557 msg_type = msg[1];
558 handle = msg[2];
559
560 if (handle == 0) {
561 DRM_ERROR("Invalid UVD handle!\n");
562 return -EINVAL;
563 }
564
565 switch (msg_type) {
566 case 0:
567 /* it's a create msg, calc image size (width * height) */
568 amdgpu_bo_kunmap(bo);
569
570 /* try to alloc a new handle */
571 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
572 if (atomic_read(&adev->uvd.handles[i]) == handle) {
573 DRM_ERROR("Handle 0x%x already in use!\n", handle);
574 return -EINVAL;
575 }
576
577 if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
578 adev->uvd.filp[i] = ctx->parser->filp;
579 return 0;
580 }
581 }
582
583 DRM_ERROR("No more free UVD handles!\n");
584 return -EINVAL;
585
586 case 1:
587 /* it's a decode msg, calc buffer sizes */
588 r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
589 amdgpu_bo_kunmap(bo);
590 if (r)
591 return r;
592
593 /* validate the handle */
594 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
595 if (atomic_read(&adev->uvd.handles[i]) == handle) {
596 if (adev->uvd.filp[i] != ctx->parser->filp) {
597 DRM_ERROR("UVD handle collision detected!\n");
598 return -EINVAL;
599 }
600 return 0;
601 }
602 }
603
604 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
605 return -ENOENT;
606
607 case 2:
608 /* it's a destroy msg, free the handle */
609 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
610 atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
611 amdgpu_bo_kunmap(bo);
612 return 0;
613
614 default:
615 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
616 return -EINVAL;
617 }
618 BUG();
619 return -EINVAL;
620}
621
622/**
623 * amdgpu_uvd_cs_pass2 - second parsing round
624 *
625 * @ctx: UVD parser context
626 *
627 * Patch buffer addresses, make sure buffer sizes are correct.
628 */
629static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
630{
631 struct amdgpu_bo_va_mapping *mapping;
632 struct amdgpu_bo *bo;
633 uint32_t cmd, lo, hi;
634 uint64_t start, end;
635 uint64_t addr;
636 int r;
637
638 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
639 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
640 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
641
642 mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
643 if (mapping == NULL)
644 return -EINVAL;
645
646 start = amdgpu_bo_gpu_offset(bo);
647
648 end = (mapping->it.last + 1 - mapping->it.start);
649 end = end * AMDGPU_GPU_PAGE_SIZE + start;
650
651 addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
652 start += addr;
653
654 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
655 lower_32_bits(start));
656 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
657 upper_32_bits(start));
658
659 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
660 if (cmd < 0x4) {
661 if ((end - start) < ctx->buf_sizes[cmd]) {
662 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
663 (unsigned)(end - start),
664 ctx->buf_sizes[cmd]);
665 return -EINVAL;
666 }
667
668 } else if (cmd == 0x206) {
669 if ((end - start) < ctx->buf_sizes[4]) {
670 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
671 (unsigned)(end - start),
672 ctx->buf_sizes[4]);
673 return -EINVAL;
674 }
675 } else if ((cmd != 0x100) && (cmd != 0x204)) {
676 DRM_ERROR("invalid UVD command %X!\n", cmd);
677 return -EINVAL;
678 }
679
680 if (!ctx->parser->adev->uvd.address_64_bit) {
681 if ((start >> 28) != ((end - 1) >> 28)) {
682 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
683 start, end);
684 return -EINVAL;
685 }
686
687 if ((cmd == 0 || cmd == 0x3) &&
688 (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
689 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
690 start, end);
691 return -EINVAL;
692 }
693 }
694
695 if (cmd == 0) {
696 ctx->has_msg_cmd = true;
697 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
698 if (r)
699 return r;
700 } else if (!ctx->has_msg_cmd) {
701 DRM_ERROR("Message needed before other commands are send!\n");
702 return -EINVAL;
703 }
704
705 return 0;
706}
707
708/**
709 * amdgpu_uvd_cs_reg - parse register writes
710 *
711 * @ctx: UVD parser context
712 * @cb: callback function
713 *
714 * Parse the register writes, call cb on each complete command.
715 */
716static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
717 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
718{
719 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
720 int i, r;
721
722 ctx->idx++;
723 for (i = 0; i <= ctx->count; ++i) {
724 unsigned reg = ctx->reg + i;
725
726 if (ctx->idx >= ib->length_dw) {
727 DRM_ERROR("Register command after end of CS!\n");
728 return -EINVAL;
729 }
730
731 switch (reg) {
732 case mmUVD_GPCOM_VCPU_DATA0:
733 ctx->data0 = ctx->idx;
734 break;
735 case mmUVD_GPCOM_VCPU_DATA1:
736 ctx->data1 = ctx->idx;
737 break;
738 case mmUVD_GPCOM_VCPU_CMD:
739 r = cb(ctx);
740 if (r)
741 return r;
742 break;
743 case mmUVD_ENGINE_CNTL:
744 break;
745 default:
746 DRM_ERROR("Invalid reg 0x%X!\n", reg);
747 return -EINVAL;
748 }
749 ctx->idx++;
750 }
751 return 0;
752}
753
754/**
755 * amdgpu_uvd_cs_packets - parse UVD packets
756 *
757 * @ctx: UVD parser context
758 * @cb: callback function
759 *
760 * Parse the command stream packets.
761 */
762static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
763 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
764{
765 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
766 int r;
767
768 for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
769 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
770 unsigned type = CP_PACKET_GET_TYPE(cmd);
771 switch (type) {
772 case PACKET_TYPE0:
773 ctx->reg = CP_PACKET0_GET_REG(cmd);
774 ctx->count = CP_PACKET_GET_COUNT(cmd);
775 r = amdgpu_uvd_cs_reg(ctx, cb);
776 if (r)
777 return r;
778 break;
779 case PACKET_TYPE2:
780 ++ctx->idx;
781 break;
782 default:
783 DRM_ERROR("Unknown packet type %d !\n", type);
784 return -EINVAL;
785 }
786 }
787 return 0;
788}
789
790/**
791 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
792 *
793 * @parser: Command submission parser context
794 *
795 * Parse the command stream, patch in addresses as necessary.
796 */
797int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
798{
799 struct amdgpu_uvd_cs_ctx ctx = {};
800 unsigned buf_sizes[] = {
801 [0x00000000] = 2048,
802 [0x00000001] = 0xFFFFFFFF,
803 [0x00000002] = 0xFFFFFFFF,
804 [0x00000003] = 2048,
805 [0x00000004] = 0xFFFFFFFF,
806 };
807 struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
808 int r;
809
810 if (ib->length_dw % 16) {
811 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
812 ib->length_dw);
813 return -EINVAL;
814 }
815
816 ctx.parser = parser;
817 ctx.buf_sizes = buf_sizes;
818 ctx.ib_idx = ib_idx;
819
820 /* first round, make sure the buffers are actually in the UVD segment */
821 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
822 if (r)
823 return r;
824
825 /* second round, patch buffer addresses into the command stream */
826 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
827 if (r)
828 return r;
829
830 if (!ctx.has_msg_cmd) {
831 DRM_ERROR("UVD-IBs need a msg command!\n");
832 return -EINVAL;
833 }
834
835 amdgpu_uvd_note_usage(ctx.parser->adev);
836
837 return 0;
838}
839
840static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
841 bool direct, struct fence **fence)
842{
843 struct ttm_validate_buffer tv;
844 struct ww_acquire_ctx ticket;
845 struct list_head head;
846 struct amdgpu_job *job;
847 struct amdgpu_ib *ib;
848 struct fence *f = NULL;
849 struct amdgpu_device *adev = ring->adev;
850 uint64_t addr;
851 int i, r;
852
853 memset(&tv, 0, sizeof(tv));
854 tv.bo = &bo->tbo;
855
856 INIT_LIST_HEAD(&head);
857 list_add(&tv.head, &head);
858
859 r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
860 if (r)
861 return r;
862
863 if (!bo->adev->uvd.address_64_bit) {
864 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
865 amdgpu_uvd_force_into_uvd_segment(bo);
866 }
867
868 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
869 if (r)
870 goto err;
871
872 r = amdgpu_job_alloc_with_ib(adev, 64, &job);
873 if (r)
874 goto err;
875
876 ib = &job->ibs[0];
877 addr = amdgpu_bo_gpu_offset(bo);
878 ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
879 ib->ptr[1] = addr;
880 ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
881 ib->ptr[3] = addr >> 32;
882 ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
883 ib->ptr[5] = 0;
884 for (i = 6; i < 16; ++i)
885 ib->ptr[i] = PACKET2(0);
886 ib->length_dw = 16;
887
888 if (direct) {
889 r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
890 job->fence = f;
891 if (r)
892 goto err_free;
893
894 amdgpu_job_free(job);
895 } else {
896 r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
897 AMDGPU_FENCE_OWNER_UNDEFINED, &f);
898 if (r)
899 goto err_free;
900 }
901
902 ttm_eu_fence_buffer_objects(&ticket, &head, f);
903
904 if (fence)
905 *fence = fence_get(f);
906 amdgpu_bo_unref(&bo);
907 fence_put(f);
908
909 return 0;
910
911err_free:
912 amdgpu_job_free(job);
913
914err:
915 ttm_eu_backoff_reservation(&ticket, &head);
916 return r;
917}
918
919/* multiple fence commands without any stream commands in between can
920 crash the vcpu so just try to emmit a dummy create/destroy msg to
921 avoid this */
922int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
923 struct fence **fence)
924{
925 struct amdgpu_device *adev = ring->adev;
926 struct amdgpu_bo *bo;
927 uint32_t *msg;
928 int r, i;
929
930 r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
931 AMDGPU_GEM_DOMAIN_VRAM,
932 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
933 NULL, NULL, &bo);
934 if (r)
935 return r;
936
937 r = amdgpu_bo_reserve(bo, false);
938 if (r) {
939 amdgpu_bo_unref(&bo);
940 return r;
941 }
942
943 r = amdgpu_bo_kmap(bo, (void **)&msg);
944 if (r) {
945 amdgpu_bo_unreserve(bo);
946 amdgpu_bo_unref(&bo);
947 return r;
948 }
949
950 /* stitch together an UVD create msg */
951 msg[0] = cpu_to_le32(0x00000de4);
952 msg[1] = cpu_to_le32(0x00000000);
953 msg[2] = cpu_to_le32(handle);
954 msg[3] = cpu_to_le32(0x00000000);
955 msg[4] = cpu_to_le32(0x00000000);
956 msg[5] = cpu_to_le32(0x00000000);
957 msg[6] = cpu_to_le32(0x00000000);
958 msg[7] = cpu_to_le32(0x00000780);
959 msg[8] = cpu_to_le32(0x00000440);
960 msg[9] = cpu_to_le32(0x00000000);
961 msg[10] = cpu_to_le32(0x01b37000);
962 for (i = 11; i < 1024; ++i)
963 msg[i] = cpu_to_le32(0x0);
964
965 amdgpu_bo_kunmap(bo);
966 amdgpu_bo_unreserve(bo);
967
968 return amdgpu_uvd_send_msg(ring, bo, true, fence);
969}
970
971int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
972 bool direct, struct fence **fence)
973{
974 struct amdgpu_device *adev = ring->adev;
975 struct amdgpu_bo *bo;
976 uint32_t *msg;
977 int r, i;
978
979 r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
980 AMDGPU_GEM_DOMAIN_VRAM,
981 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
982 NULL, NULL, &bo);
983 if (r)
984 return r;
985
986 r = amdgpu_bo_reserve(bo, false);
987 if (r) {
988 amdgpu_bo_unref(&bo);
989 return r;
990 }
991
992 r = amdgpu_bo_kmap(bo, (void **)&msg);
993 if (r) {
994 amdgpu_bo_unreserve(bo);
995 amdgpu_bo_unref(&bo);
996 return r;
997 }
998
999 /* stitch together an UVD destroy msg */
1000 msg[0] = cpu_to_le32(0x00000de4);
1001 msg[1] = cpu_to_le32(0x00000002);
1002 msg[2] = cpu_to_le32(handle);
1003 msg[3] = cpu_to_le32(0x00000000);
1004 for (i = 4; i < 1024; ++i)
1005 msg[i] = cpu_to_le32(0x0);
1006
1007 amdgpu_bo_kunmap(bo);
1008 amdgpu_bo_unreserve(bo);
1009
1010 return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1011}
1012
1013static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1014{
1015 struct amdgpu_device *adev =
1016 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1017 unsigned i, fences, handles = 0;
1018
1019 fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1020
1021 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1022 if (atomic_read(&adev->uvd.handles[i]))
1023 ++handles;
1024
1025 if (fences == 0 && handles == 0) {
1026 if (adev->pm.dpm_enabled) {
1027 amdgpu_dpm_enable_uvd(adev, false);
1028 } else {
1029 amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1030 }
1031 } else {
1032 schedule_delayed_work(&adev->uvd.idle_work,
1033 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1034 }
1035}
1036
1037static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1038{
1039 bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1040 set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1041 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1042
1043 if (set_clocks) {
1044 if (adev->pm.dpm_enabled) {
1045 amdgpu_dpm_enable_uvd(adev, true);
1046 } else {
1047 amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1048 }
1049 }
1050}