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 <[email protected]>
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 #include "amdgpu_ras.h"
43 
44 /* 1 second timeout */
45 #define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
46 
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))
52 
53 /* Polaris10/11 firmware version */
54 #define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
55 
56 /* Firmware Names */
57 #ifdef CONFIG_DRM_AMDGPU_CIK
58 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_uvd.bin"
59 #define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
60 #define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
61 #define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
62 #define FIRMWARE_MULLINS	"amdgpu/mullins_uvd.bin"
63 #endif
64 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
65 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
66 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
67 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
68 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
69 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
70 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
71 #define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
72 
73 #define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
74 #define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
75 #define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
76 
77 /* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
78 #define UVD_GPCOM_VCPU_CMD		0x03c3
79 #define UVD_GPCOM_VCPU_DATA0	0x03c4
80 #define UVD_GPCOM_VCPU_DATA1	0x03c5
81 #define UVD_NO_OP				0x03ff
82 #define UVD_BASE_SI				0x3800
83 
84 /**
85  * amdgpu_uvd_cs_ctx - Command submission parser context
86  *
87  * Used for emulating virtual memory support on UVD 4.2.
88  */
89 struct amdgpu_uvd_cs_ctx {
90 	struct amdgpu_cs_parser *parser;
91 	unsigned reg, count;
92 	unsigned data0, data1;
93 	unsigned idx;
94 	unsigned ib_idx;
95 
96 	/* does the IB has a msg command */
97 	bool has_msg_cmd;
98 
99 	/* minimum buffer sizes */
100 	unsigned *buf_sizes;
101 };
102 
103 #ifdef CONFIG_DRM_AMDGPU_CIK
104 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
105 MODULE_FIRMWARE(FIRMWARE_KABINI);
106 MODULE_FIRMWARE(FIRMWARE_KAVERI);
107 MODULE_FIRMWARE(FIRMWARE_HAWAII);
108 MODULE_FIRMWARE(FIRMWARE_MULLINS);
109 #endif
110 MODULE_FIRMWARE(FIRMWARE_TONGA);
111 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
112 MODULE_FIRMWARE(FIRMWARE_FIJI);
113 MODULE_FIRMWARE(FIRMWARE_STONEY);
114 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
115 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
116 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
117 MODULE_FIRMWARE(FIRMWARE_VEGAM);
118 
119 MODULE_FIRMWARE(FIRMWARE_VEGA10);
120 MODULE_FIRMWARE(FIRMWARE_VEGA12);
121 MODULE_FIRMWARE(FIRMWARE_VEGA20);
122 
123 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
124 
125 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
126 {
127 	unsigned long bo_size;
128 	const char *fw_name;
129 	const struct common_firmware_header *hdr;
130 	unsigned family_id;
131 	int i, j, r;
132 
133 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
134 
135 	switch (adev->asic_type) {
136 #ifdef CONFIG_DRM_AMDGPU_CIK
137 	case CHIP_BONAIRE:
138 		fw_name = FIRMWARE_BONAIRE;
139 		break;
140 	case CHIP_KABINI:
141 		fw_name = FIRMWARE_KABINI;
142 		break;
143 	case CHIP_KAVERI:
144 		fw_name = FIRMWARE_KAVERI;
145 		break;
146 	case CHIP_HAWAII:
147 		fw_name = FIRMWARE_HAWAII;
148 		break;
149 	case CHIP_MULLINS:
150 		fw_name = FIRMWARE_MULLINS;
151 		break;
152 #endif
153 	case CHIP_TONGA:
154 		fw_name = FIRMWARE_TONGA;
155 		break;
156 	case CHIP_FIJI:
157 		fw_name = FIRMWARE_FIJI;
158 		break;
159 	case CHIP_CARRIZO:
160 		fw_name = FIRMWARE_CARRIZO;
161 		break;
162 	case CHIP_STONEY:
163 		fw_name = FIRMWARE_STONEY;
164 		break;
165 	case CHIP_POLARIS10:
166 		fw_name = FIRMWARE_POLARIS10;
167 		break;
168 	case CHIP_POLARIS11:
169 		fw_name = FIRMWARE_POLARIS11;
170 		break;
171 	case CHIP_POLARIS12:
172 		fw_name = FIRMWARE_POLARIS12;
173 		break;
174 	case CHIP_VEGA10:
175 		fw_name = FIRMWARE_VEGA10;
176 		break;
177 	case CHIP_VEGA12:
178 		fw_name = FIRMWARE_VEGA12;
179 		break;
180 	case CHIP_VEGAM:
181 		fw_name = FIRMWARE_VEGAM;
182 		break;
183 	case CHIP_VEGA20:
184 		fw_name = FIRMWARE_VEGA20;
185 		break;
186 	default:
187 		return -EINVAL;
188 	}
189 
190 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
191 	if (r) {
192 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
193 			fw_name);
194 		return r;
195 	}
196 
197 	r = amdgpu_ucode_validate(adev->uvd.fw);
198 	if (r) {
199 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
200 			fw_name);
201 		release_firmware(adev->uvd.fw);
202 		adev->uvd.fw = NULL;
203 		return r;
204 	}
205 
206 	/* Set the default UVD handles that the firmware can handle */
207 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
208 
209 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
210 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
211 
212 	if (adev->asic_type < CHIP_VEGA20) {
213 		unsigned version_major, version_minor;
214 
215 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
216 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
217 		DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
218 			version_major, version_minor, family_id);
219 
220 		/*
221 		 * Limit the number of UVD handles depending on microcode major
222 		 * and minor versions. The firmware version which has 40 UVD
223 		 * instances support is 1.80. So all subsequent versions should
224 		 * also have the same support.
225 		 */
226 		if ((version_major > 0x01) ||
227 		    ((version_major == 0x01) && (version_minor >= 0x50)))
228 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
229 
230 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
231 					(family_id << 8));
232 
233 		if ((adev->asic_type == CHIP_POLARIS10 ||
234 		     adev->asic_type == CHIP_POLARIS11) &&
235 		    (adev->uvd.fw_version < FW_1_66_16))
236 			DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
237 				  version_major, version_minor);
238 	} else {
239 		unsigned int enc_major, enc_minor, dec_minor;
240 
241 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
242 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
243 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
244 		DRM_INFO("Found UVD firmware ENC: %hu.%hu DEC: .%hu Family ID: %hu\n",
245 			enc_major, enc_minor, dec_minor, family_id);
246 
247 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
248 
249 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
250 	}
251 
252 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
253 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
254 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
255 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
256 
257 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
258 		if (adev->uvd.harvest_config & (1 << j))
259 			continue;
260 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
261 					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
262 					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
263 		if (r) {
264 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
265 			return r;
266 		}
267 	}
268 
269 	for (i = 0; i < adev->uvd.max_handles; ++i) {
270 		atomic_set(&adev->uvd.handles[i], 0);
271 		adev->uvd.filp[i] = NULL;
272 	}
273 
274 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
275 	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
276 		adev->uvd.address_64_bit = true;
277 
278 	switch (adev->asic_type) {
279 	case CHIP_TONGA:
280 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
281 		break;
282 	case CHIP_CARRIZO:
283 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
284 		break;
285 	case CHIP_FIJI:
286 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
287 		break;
288 	case CHIP_STONEY:
289 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
290 		break;
291 	default:
292 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
293 	}
294 
295 	return 0;
296 }
297 
298 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
299 {
300 	int i, j;
301 
302 	drm_sched_entity_destroy(&adev->uvd.entity);
303 
304 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
305 		if (adev->uvd.harvest_config & (1 << j))
306 			continue;
307 		kvfree(adev->uvd.inst[j].saved_bo);
308 
309 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
310 				      &adev->uvd.inst[j].gpu_addr,
311 				      (void **)&adev->uvd.inst[j].cpu_addr);
312 
313 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
314 
315 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
316 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
317 	}
318 	release_firmware(adev->uvd.fw);
319 
320 	return 0;
321 }
322 
323 /**
324  * amdgpu_uvd_entity_init - init entity
325  *
326  * @adev: amdgpu_device pointer
327  *
328  */
329 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
330 {
331 	struct amdgpu_ring *ring;
332 	struct drm_sched_rq *rq;
333 	int r;
334 
335 	ring = &adev->uvd.inst[0].ring;
336 	rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
337 	r = drm_sched_entity_init(&adev->uvd.entity, &rq, 1, NULL);
338 	if (r) {
339 		DRM_ERROR("Failed setting up UVD kernel entity.\n");
340 		return r;
341 	}
342 
343 	return 0;
344 }
345 
346 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
347 {
348 	unsigned size;
349 	void *ptr;
350 	int i, j;
351 
352 	cancel_delayed_work_sync(&adev->uvd.idle_work);
353 
354 	/* only valid for physical mode */
355 	if (adev->asic_type < CHIP_POLARIS10) {
356 		for (i = 0; i < adev->uvd.max_handles; ++i)
357 			if (atomic_read(&adev->uvd.handles[i]))
358 				break;
359 
360 		if (i == adev->uvd.max_handles)
361 			return 0;
362 	}
363 
364 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
365 		if (adev->uvd.harvest_config & (1 << j))
366 			continue;
367 		if (adev->uvd.inst[j].vcpu_bo == NULL)
368 			continue;
369 
370 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
371 		ptr = adev->uvd.inst[j].cpu_addr;
372 
373 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
374 		if (!adev->uvd.inst[j].saved_bo)
375 			return -ENOMEM;
376 
377 		/* re-write 0 since err_event_athub will corrupt VCPU buffer */
378 		if (amdgpu_ras_intr_triggered()) {
379 			DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
380 			memset(adev->uvd.inst[j].saved_bo, 0, size);
381 		} else {
382 			memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
383 		}
384 	}
385 	return 0;
386 }
387 
388 int amdgpu_uvd_resume(struct amdgpu_device *adev)
389 {
390 	unsigned size;
391 	void *ptr;
392 	int i;
393 
394 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
395 		if (adev->uvd.harvest_config & (1 << i))
396 			continue;
397 		if (adev->uvd.inst[i].vcpu_bo == NULL)
398 			return -EINVAL;
399 
400 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
401 		ptr = adev->uvd.inst[i].cpu_addr;
402 
403 		if (adev->uvd.inst[i].saved_bo != NULL) {
404 			memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
405 			kvfree(adev->uvd.inst[i].saved_bo);
406 			adev->uvd.inst[i].saved_bo = NULL;
407 		} else {
408 			const struct common_firmware_header *hdr;
409 			unsigned offset;
410 
411 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
412 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
413 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
414 				memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
415 					    le32_to_cpu(hdr->ucode_size_bytes));
416 				size -= le32_to_cpu(hdr->ucode_size_bytes);
417 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
418 			}
419 			memset_io(ptr, 0, size);
420 			/* to restore uvd fence seq */
421 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
422 		}
423 	}
424 	return 0;
425 }
426 
427 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
428 {
429 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
430 	int i, r;
431 
432 	for (i = 0; i < adev->uvd.max_handles; ++i) {
433 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
434 
435 		if (handle != 0 && adev->uvd.filp[i] == filp) {
436 			struct dma_fence *fence;
437 
438 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
439 						       &fence);
440 			if (r) {
441 				DRM_ERROR("Error destroying UVD %d!\n", r);
442 				continue;
443 			}
444 
445 			dma_fence_wait(fence, false);
446 			dma_fence_put(fence);
447 
448 			adev->uvd.filp[i] = NULL;
449 			atomic_set(&adev->uvd.handles[i], 0);
450 		}
451 	}
452 }
453 
454 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
455 {
456 	int i;
457 	for (i = 0; i < abo->placement.num_placement; ++i) {
458 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
459 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
460 	}
461 }
462 
463 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
464 {
465 	uint32_t lo, hi;
466 	uint64_t addr;
467 
468 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
469 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
470 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
471 
472 	return addr;
473 }
474 
475 /**
476  * amdgpu_uvd_cs_pass1 - first parsing round
477  *
478  * @ctx: UVD parser context
479  *
480  * Make sure UVD message and feedback buffers are in VRAM and
481  * nobody is violating an 256MB boundary.
482  */
483 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
484 {
485 	struct ttm_operation_ctx tctx = { false, false };
486 	struct amdgpu_bo_va_mapping *mapping;
487 	struct amdgpu_bo *bo;
488 	uint32_t cmd;
489 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
490 	int r = 0;
491 
492 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
493 	if (r) {
494 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
495 		return r;
496 	}
497 
498 	if (!ctx->parser->adev->uvd.address_64_bit) {
499 		/* check if it's a message or feedback command */
500 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
501 		if (cmd == 0x0 || cmd == 0x3) {
502 			/* yes, force it into VRAM */
503 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
504 			amdgpu_bo_placement_from_domain(bo, domain);
505 		}
506 		amdgpu_uvd_force_into_uvd_segment(bo);
507 
508 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
509 	}
510 
511 	return r;
512 }
513 
514 /**
515  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
516  *
517  * @msg: pointer to message structure
518  * @buf_sizes: returned buffer sizes
519  *
520  * Peek into the decode message and calculate the necessary buffer sizes.
521  */
522 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
523 	unsigned buf_sizes[])
524 {
525 	unsigned stream_type = msg[4];
526 	unsigned width = msg[6];
527 	unsigned height = msg[7];
528 	unsigned dpb_size = msg[9];
529 	unsigned pitch = msg[28];
530 	unsigned level = msg[57];
531 
532 	unsigned width_in_mb = width / 16;
533 	unsigned height_in_mb = ALIGN(height / 16, 2);
534 	unsigned fs_in_mb = width_in_mb * height_in_mb;
535 
536 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
537 	unsigned min_ctx_size = ~0;
538 
539 	image_size = width * height;
540 	image_size += image_size / 2;
541 	image_size = ALIGN(image_size, 1024);
542 
543 	switch (stream_type) {
544 	case 0: /* H264 */
545 		switch(level) {
546 		case 30:
547 			num_dpb_buffer = 8100 / fs_in_mb;
548 			break;
549 		case 31:
550 			num_dpb_buffer = 18000 / fs_in_mb;
551 			break;
552 		case 32:
553 			num_dpb_buffer = 20480 / fs_in_mb;
554 			break;
555 		case 41:
556 			num_dpb_buffer = 32768 / fs_in_mb;
557 			break;
558 		case 42:
559 			num_dpb_buffer = 34816 / fs_in_mb;
560 			break;
561 		case 50:
562 			num_dpb_buffer = 110400 / fs_in_mb;
563 			break;
564 		case 51:
565 			num_dpb_buffer = 184320 / fs_in_mb;
566 			break;
567 		default:
568 			num_dpb_buffer = 184320 / fs_in_mb;
569 			break;
570 		}
571 		num_dpb_buffer++;
572 		if (num_dpb_buffer > 17)
573 			num_dpb_buffer = 17;
574 
575 		/* reference picture buffer */
576 		min_dpb_size = image_size * num_dpb_buffer;
577 
578 		/* macroblock context buffer */
579 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
580 
581 		/* IT surface buffer */
582 		min_dpb_size += width_in_mb * height_in_mb * 32;
583 		break;
584 
585 	case 1: /* VC1 */
586 
587 		/* reference picture buffer */
588 		min_dpb_size = image_size * 3;
589 
590 		/* CONTEXT_BUFFER */
591 		min_dpb_size += width_in_mb * height_in_mb * 128;
592 
593 		/* IT surface buffer */
594 		min_dpb_size += width_in_mb * 64;
595 
596 		/* DB surface buffer */
597 		min_dpb_size += width_in_mb * 128;
598 
599 		/* BP */
600 		tmp = max(width_in_mb, height_in_mb);
601 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
602 		break;
603 
604 	case 3: /* MPEG2 */
605 
606 		/* reference picture buffer */
607 		min_dpb_size = image_size * 3;
608 		break;
609 
610 	case 4: /* MPEG4 */
611 
612 		/* reference picture buffer */
613 		min_dpb_size = image_size * 3;
614 
615 		/* CM */
616 		min_dpb_size += width_in_mb * height_in_mb * 64;
617 
618 		/* IT surface buffer */
619 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
620 		break;
621 
622 	case 7: /* H264 Perf */
623 		switch(level) {
624 		case 30:
625 			num_dpb_buffer = 8100 / fs_in_mb;
626 			break;
627 		case 31:
628 			num_dpb_buffer = 18000 / fs_in_mb;
629 			break;
630 		case 32:
631 			num_dpb_buffer = 20480 / fs_in_mb;
632 			break;
633 		case 41:
634 			num_dpb_buffer = 32768 / fs_in_mb;
635 			break;
636 		case 42:
637 			num_dpb_buffer = 34816 / fs_in_mb;
638 			break;
639 		case 50:
640 			num_dpb_buffer = 110400 / fs_in_mb;
641 			break;
642 		case 51:
643 			num_dpb_buffer = 184320 / fs_in_mb;
644 			break;
645 		default:
646 			num_dpb_buffer = 184320 / fs_in_mb;
647 			break;
648 		}
649 		num_dpb_buffer++;
650 		if (num_dpb_buffer > 17)
651 			num_dpb_buffer = 17;
652 
653 		/* reference picture buffer */
654 		min_dpb_size = image_size * num_dpb_buffer;
655 
656 		if (!adev->uvd.use_ctx_buf){
657 			/* macroblock context buffer */
658 			min_dpb_size +=
659 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
660 
661 			/* IT surface buffer */
662 			min_dpb_size += width_in_mb * height_in_mb * 32;
663 		} else {
664 			/* macroblock context buffer */
665 			min_ctx_size =
666 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
667 		}
668 		break;
669 
670 	case 8: /* MJPEG */
671 		min_dpb_size = 0;
672 		break;
673 
674 	case 16: /* H265 */
675 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
676 		image_size = ALIGN(image_size, 256);
677 
678 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
679 		min_dpb_size = image_size * num_dpb_buffer;
680 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
681 					   * 16 * num_dpb_buffer + 52 * 1024;
682 		break;
683 
684 	default:
685 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
686 		return -EINVAL;
687 	}
688 
689 	if (width > pitch) {
690 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
691 		return -EINVAL;
692 	}
693 
694 	if (dpb_size < min_dpb_size) {
695 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
696 			  dpb_size, min_dpb_size);
697 		return -EINVAL;
698 	}
699 
700 	buf_sizes[0x1] = dpb_size;
701 	buf_sizes[0x2] = image_size;
702 	buf_sizes[0x4] = min_ctx_size;
703 	/* store image width to adjust nb memory pstate */
704 	adev->uvd.decode_image_width = width;
705 	return 0;
706 }
707 
708 /**
709  * amdgpu_uvd_cs_msg - handle UVD message
710  *
711  * @ctx: UVD parser context
712  * @bo: buffer object containing the message
713  * @offset: offset into the buffer object
714  *
715  * Peek into the UVD message and extract the session id.
716  * Make sure that we don't open up to many sessions.
717  */
718 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
719 			     struct amdgpu_bo *bo, unsigned offset)
720 {
721 	struct amdgpu_device *adev = ctx->parser->adev;
722 	int32_t *msg, msg_type, handle;
723 	void *ptr;
724 	long r;
725 	int i;
726 
727 	if (offset & 0x3F) {
728 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
729 		return -EINVAL;
730 	}
731 
732 	r = amdgpu_bo_kmap(bo, &ptr);
733 	if (r) {
734 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
735 		return r;
736 	}
737 
738 	msg = ptr + offset;
739 
740 	msg_type = msg[1];
741 	handle = msg[2];
742 
743 	if (handle == 0) {
744 		DRM_ERROR("Invalid UVD handle!\n");
745 		return -EINVAL;
746 	}
747 
748 	switch (msg_type) {
749 	case 0:
750 		/* it's a create msg, calc image size (width * height) */
751 		amdgpu_bo_kunmap(bo);
752 
753 		/* try to alloc a new handle */
754 		for (i = 0; i < adev->uvd.max_handles; ++i) {
755 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
756 				DRM_ERROR(")Handle 0x%x already in use!\n",
757 					  handle);
758 				return -EINVAL;
759 			}
760 
761 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
762 				adev->uvd.filp[i] = ctx->parser->filp;
763 				return 0;
764 			}
765 		}
766 
767 		DRM_ERROR("No more free UVD handles!\n");
768 		return -ENOSPC;
769 
770 	case 1:
771 		/* it's a decode msg, calc buffer sizes */
772 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
773 		amdgpu_bo_kunmap(bo);
774 		if (r)
775 			return r;
776 
777 		/* validate the handle */
778 		for (i = 0; i < adev->uvd.max_handles; ++i) {
779 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
780 				if (adev->uvd.filp[i] != ctx->parser->filp) {
781 					DRM_ERROR("UVD handle collision detected!\n");
782 					return -EINVAL;
783 				}
784 				return 0;
785 			}
786 		}
787 
788 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
789 		return -ENOENT;
790 
791 	case 2:
792 		/* it's a destroy msg, free the handle */
793 		for (i = 0; i < adev->uvd.max_handles; ++i)
794 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
795 		amdgpu_bo_kunmap(bo);
796 		return 0;
797 
798 	default:
799 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
800 		return -EINVAL;
801 	}
802 	BUG();
803 	return -EINVAL;
804 }
805 
806 /**
807  * amdgpu_uvd_cs_pass2 - second parsing round
808  *
809  * @ctx: UVD parser context
810  *
811  * Patch buffer addresses, make sure buffer sizes are correct.
812  */
813 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
814 {
815 	struct amdgpu_bo_va_mapping *mapping;
816 	struct amdgpu_bo *bo;
817 	uint32_t cmd;
818 	uint64_t start, end;
819 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
820 	int r;
821 
822 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
823 	if (r) {
824 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
825 		return r;
826 	}
827 
828 	start = amdgpu_bo_gpu_offset(bo);
829 
830 	end = (mapping->last + 1 - mapping->start);
831 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
832 
833 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
834 	start += addr;
835 
836 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
837 			    lower_32_bits(start));
838 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
839 			    upper_32_bits(start));
840 
841 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
842 	if (cmd < 0x4) {
843 		if ((end - start) < ctx->buf_sizes[cmd]) {
844 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
845 				  (unsigned)(end - start),
846 				  ctx->buf_sizes[cmd]);
847 			return -EINVAL;
848 		}
849 
850 	} else if (cmd == 0x206) {
851 		if ((end - start) < ctx->buf_sizes[4]) {
852 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
853 					  (unsigned)(end - start),
854 					  ctx->buf_sizes[4]);
855 			return -EINVAL;
856 		}
857 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
858 		DRM_ERROR("invalid UVD command %X!\n", cmd);
859 		return -EINVAL;
860 	}
861 
862 	if (!ctx->parser->adev->uvd.address_64_bit) {
863 		if ((start >> 28) != ((end - 1) >> 28)) {
864 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
865 				  start, end);
866 			return -EINVAL;
867 		}
868 
869 		if ((cmd == 0 || cmd == 0x3) &&
870 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
871 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
872 				  start, end);
873 			return -EINVAL;
874 		}
875 	}
876 
877 	if (cmd == 0) {
878 		ctx->has_msg_cmd = true;
879 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
880 		if (r)
881 			return r;
882 	} else if (!ctx->has_msg_cmd) {
883 		DRM_ERROR("Message needed before other commands are send!\n");
884 		return -EINVAL;
885 	}
886 
887 	return 0;
888 }
889 
890 /**
891  * amdgpu_uvd_cs_reg - parse register writes
892  *
893  * @ctx: UVD parser context
894  * @cb: callback function
895  *
896  * Parse the register writes, call cb on each complete command.
897  */
898 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
899 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
900 {
901 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
902 	int i, r;
903 
904 	ctx->idx++;
905 	for (i = 0; i <= ctx->count; ++i) {
906 		unsigned reg = ctx->reg + i;
907 
908 		if (ctx->idx >= ib->length_dw) {
909 			DRM_ERROR("Register command after end of CS!\n");
910 			return -EINVAL;
911 		}
912 
913 		switch (reg) {
914 		case mmUVD_GPCOM_VCPU_DATA0:
915 			ctx->data0 = ctx->idx;
916 			break;
917 		case mmUVD_GPCOM_VCPU_DATA1:
918 			ctx->data1 = ctx->idx;
919 			break;
920 		case mmUVD_GPCOM_VCPU_CMD:
921 			r = cb(ctx);
922 			if (r)
923 				return r;
924 			break;
925 		case mmUVD_ENGINE_CNTL:
926 		case mmUVD_NO_OP:
927 			break;
928 		default:
929 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
930 			return -EINVAL;
931 		}
932 		ctx->idx++;
933 	}
934 	return 0;
935 }
936 
937 /**
938  * amdgpu_uvd_cs_packets - parse UVD packets
939  *
940  * @ctx: UVD parser context
941  * @cb: callback function
942  *
943  * Parse the command stream packets.
944  */
945 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
946 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
947 {
948 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
949 	int r;
950 
951 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
952 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
953 		unsigned type = CP_PACKET_GET_TYPE(cmd);
954 		switch (type) {
955 		case PACKET_TYPE0:
956 			ctx->reg = CP_PACKET0_GET_REG(cmd);
957 			ctx->count = CP_PACKET_GET_COUNT(cmd);
958 			r = amdgpu_uvd_cs_reg(ctx, cb);
959 			if (r)
960 				return r;
961 			break;
962 		case PACKET_TYPE2:
963 			++ctx->idx;
964 			break;
965 		default:
966 			DRM_ERROR("Unknown packet type %d !\n", type);
967 			return -EINVAL;
968 		}
969 	}
970 	return 0;
971 }
972 
973 /**
974  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
975  *
976  * @parser: Command submission parser context
977  *
978  * Parse the command stream, patch in addresses as necessary.
979  */
980 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
981 {
982 	struct amdgpu_uvd_cs_ctx ctx = {};
983 	unsigned buf_sizes[] = {
984 		[0x00000000]	=	2048,
985 		[0x00000001]	=	0xFFFFFFFF,
986 		[0x00000002]	=	0xFFFFFFFF,
987 		[0x00000003]	=	2048,
988 		[0x00000004]	=	0xFFFFFFFF,
989 	};
990 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
991 	int r;
992 
993 	parser->job->vm = NULL;
994 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
995 
996 	if (ib->length_dw % 16) {
997 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
998 			  ib->length_dw);
999 		return -EINVAL;
1000 	}
1001 
1002 	ctx.parser = parser;
1003 	ctx.buf_sizes = buf_sizes;
1004 	ctx.ib_idx = ib_idx;
1005 
1006 	/* first round only required on chips without UVD 64 bit address support */
1007 	if (!parser->adev->uvd.address_64_bit) {
1008 		/* first round, make sure the buffers are actually in the UVD segment */
1009 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1010 		if (r)
1011 			return r;
1012 	}
1013 
1014 	/* second round, patch buffer addresses into the command stream */
1015 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1016 	if (r)
1017 		return r;
1018 
1019 	if (!ctx.has_msg_cmd) {
1020 		DRM_ERROR("UVD-IBs need a msg command!\n");
1021 		return -EINVAL;
1022 	}
1023 
1024 	return 0;
1025 }
1026 
1027 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1028 			       bool direct, struct dma_fence **fence)
1029 {
1030 	struct amdgpu_device *adev = ring->adev;
1031 	struct dma_fence *f = NULL;
1032 	struct amdgpu_job *job;
1033 	struct amdgpu_ib *ib;
1034 	uint32_t data[4];
1035 	uint64_t addr;
1036 	long r;
1037 	int i;
1038 	unsigned offset_idx = 0;
1039 	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1040 
1041 	amdgpu_bo_kunmap(bo);
1042 	amdgpu_bo_unpin(bo);
1043 
1044 	if (!ring->adev->uvd.address_64_bit) {
1045 		struct ttm_operation_ctx ctx = { true, false };
1046 
1047 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1048 		amdgpu_uvd_force_into_uvd_segment(bo);
1049 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1050 		if (r)
1051 			goto err;
1052 	}
1053 
1054 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1055 	if (r)
1056 		goto err;
1057 
1058 	if (adev->asic_type >= CHIP_VEGA10) {
1059 		offset_idx = 1 + ring->me;
1060 		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1061 		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1062 	}
1063 
1064 	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1065 	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1066 	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1067 	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1068 
1069 	ib = &job->ibs[0];
1070 	addr = amdgpu_bo_gpu_offset(bo);
1071 	ib->ptr[0] = data[0];
1072 	ib->ptr[1] = addr;
1073 	ib->ptr[2] = data[1];
1074 	ib->ptr[3] = addr >> 32;
1075 	ib->ptr[4] = data[2];
1076 	ib->ptr[5] = 0;
1077 	for (i = 6; i < 16; i += 2) {
1078 		ib->ptr[i] = data[3];
1079 		ib->ptr[i+1] = 0;
1080 	}
1081 	ib->length_dw = 16;
1082 
1083 	if (direct) {
1084 		r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv,
1085 							true, false,
1086 							msecs_to_jiffies(10));
1087 		if (r == 0)
1088 			r = -ETIMEDOUT;
1089 		if (r < 0)
1090 			goto err_free;
1091 
1092 		r = amdgpu_job_submit_direct(job, ring, &f);
1093 		if (r)
1094 			goto err_free;
1095 	} else {
1096 		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1097 				     AMDGPU_FENCE_OWNER_UNDEFINED, false);
1098 		if (r)
1099 			goto err_free;
1100 
1101 		r = amdgpu_job_submit(job, &adev->uvd.entity,
1102 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1103 		if (r)
1104 			goto err_free;
1105 	}
1106 
1107 	amdgpu_bo_fence(bo, f, false);
1108 	amdgpu_bo_unreserve(bo);
1109 	amdgpu_bo_unref(&bo);
1110 
1111 	if (fence)
1112 		*fence = dma_fence_get(f);
1113 	dma_fence_put(f);
1114 
1115 	return 0;
1116 
1117 err_free:
1118 	amdgpu_job_free(job);
1119 
1120 err:
1121 	amdgpu_bo_unreserve(bo);
1122 	amdgpu_bo_unref(&bo);
1123 	return r;
1124 }
1125 
1126 /* multiple fence commands without any stream commands in between can
1127    crash the vcpu so just try to emmit a dummy create/destroy msg to
1128    avoid this */
1129 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1130 			      struct dma_fence **fence)
1131 {
1132 	struct amdgpu_device *adev = ring->adev;
1133 	struct amdgpu_bo *bo = NULL;
1134 	uint32_t *msg;
1135 	int r, i;
1136 
1137 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1138 				      AMDGPU_GEM_DOMAIN_VRAM,
1139 				      &bo, NULL, (void **)&msg);
1140 	if (r)
1141 		return r;
1142 
1143 	/* stitch together an UVD create msg */
1144 	msg[0] = cpu_to_le32(0x00000de4);
1145 	msg[1] = cpu_to_le32(0x00000000);
1146 	msg[2] = cpu_to_le32(handle);
1147 	msg[3] = cpu_to_le32(0x00000000);
1148 	msg[4] = cpu_to_le32(0x00000000);
1149 	msg[5] = cpu_to_le32(0x00000000);
1150 	msg[6] = cpu_to_le32(0x00000000);
1151 	msg[7] = cpu_to_le32(0x00000780);
1152 	msg[8] = cpu_to_le32(0x00000440);
1153 	msg[9] = cpu_to_le32(0x00000000);
1154 	msg[10] = cpu_to_le32(0x01b37000);
1155 	for (i = 11; i < 1024; ++i)
1156 		msg[i] = cpu_to_le32(0x0);
1157 
1158 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1159 }
1160 
1161 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1162 			       bool direct, struct dma_fence **fence)
1163 {
1164 	struct amdgpu_device *adev = ring->adev;
1165 	struct amdgpu_bo *bo = NULL;
1166 	uint32_t *msg;
1167 	int r, i;
1168 
1169 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1170 				      AMDGPU_GEM_DOMAIN_VRAM,
1171 				      &bo, NULL, (void **)&msg);
1172 	if (r)
1173 		return r;
1174 
1175 	/* stitch together an UVD destroy msg */
1176 	msg[0] = cpu_to_le32(0x00000de4);
1177 	msg[1] = cpu_to_le32(0x00000002);
1178 	msg[2] = cpu_to_le32(handle);
1179 	msg[3] = cpu_to_le32(0x00000000);
1180 	for (i = 4; i < 1024; ++i)
1181 		msg[i] = cpu_to_le32(0x0);
1182 
1183 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1184 }
1185 
1186 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1187 {
1188 	struct amdgpu_device *adev =
1189 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1190 	unsigned fences = 0, i, j;
1191 
1192 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1193 		if (adev->uvd.harvest_config & (1 << i))
1194 			continue;
1195 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1196 		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1197 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1198 		}
1199 	}
1200 
1201 	if (fences == 0) {
1202 		if (adev->pm.dpm_enabled) {
1203 			amdgpu_dpm_enable_uvd(adev, false);
1204 		} else {
1205 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1206 			/* shutdown the UVD block */
1207 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1208 							       AMD_PG_STATE_GATE);
1209 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1210 							       AMD_CG_STATE_GATE);
1211 		}
1212 	} else {
1213 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1214 	}
1215 }
1216 
1217 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1218 {
1219 	struct amdgpu_device *adev = ring->adev;
1220 	bool set_clocks;
1221 
1222 	if (amdgpu_sriov_vf(adev))
1223 		return;
1224 
1225 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1226 	if (set_clocks) {
1227 		if (adev->pm.dpm_enabled) {
1228 			amdgpu_dpm_enable_uvd(adev, true);
1229 		} else {
1230 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1231 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1232 							       AMD_CG_STATE_UNGATE);
1233 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1234 							       AMD_PG_STATE_UNGATE);
1235 		}
1236 	}
1237 }
1238 
1239 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1240 {
1241 	if (!amdgpu_sriov_vf(ring->adev))
1242 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1243 }
1244 
1245 /**
1246  * amdgpu_uvd_ring_test_ib - test ib execution
1247  *
1248  * @ring: amdgpu_ring pointer
1249  *
1250  * Test if we can successfully execute an IB
1251  */
1252 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1253 {
1254 	struct dma_fence *fence;
1255 	long r;
1256 
1257 	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1258 	if (r)
1259 		goto error;
1260 
1261 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1262 	if (r)
1263 		goto error;
1264 
1265 	r = dma_fence_wait_timeout(fence, false, timeout);
1266 	if (r == 0)
1267 		r = -ETIMEDOUT;
1268 	else if (r > 0)
1269 		r = 0;
1270 
1271 	dma_fence_put(fence);
1272 
1273 error:
1274 	return r;
1275 }
1276 
1277 /**
1278  * amdgpu_uvd_used_handles - returns used UVD handles
1279  *
1280  * @adev: amdgpu_device pointer
1281  *
1282  * Returns the number of UVD handles in use
1283  */
1284 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1285 {
1286 	unsigned i;
1287 	uint32_t used_handles = 0;
1288 
1289 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1290 		/*
1291 		 * Handles can be freed in any order, and not
1292 		 * necessarily linear. So we need to count
1293 		 * all non-zero handles.
1294 		 */
1295 		if (atomic_read(&adev->uvd.handles[i]))
1296 			used_handles++;
1297 	}
1298 
1299 	return used_handles;
1300 }
1301