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 #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 	cancel_delayed_work_sync(&adev->uvd.idle_work);
303 	drm_sched_entity_destroy(&adev->uvd.entity);
304 
305 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
306 		if (adev->uvd.harvest_config & (1 << j))
307 			continue;
308 		kvfree(adev->uvd.inst[j].saved_bo);
309 
310 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
311 				      &adev->uvd.inst[j].gpu_addr,
312 				      (void **)&adev->uvd.inst[j].cpu_addr);
313 
314 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
315 
316 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
317 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
318 	}
319 	release_firmware(adev->uvd.fw);
320 
321 	return 0;
322 }
323 
324 /**
325  * amdgpu_uvd_entity_init - init entity
326  *
327  * @adev: amdgpu_device pointer
328  *
329  */
330 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
331 {
332 	struct amdgpu_ring *ring;
333 	struct drm_sched_rq *rq;
334 	int r;
335 
336 	ring = &adev->uvd.inst[0].ring;
337 	rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
338 	r = drm_sched_entity_init(&adev->uvd.entity, &rq, 1, NULL);
339 	if (r) {
340 		DRM_ERROR("Failed setting up UVD kernel entity.\n");
341 		return r;
342 	}
343 
344 	return 0;
345 }
346 
347 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
348 {
349 	unsigned size;
350 	void *ptr;
351 	int i, j;
352 	bool in_ras_intr = amdgpu_ras_intr_triggered();
353 
354 	cancel_delayed_work_sync(&adev->uvd.idle_work);
355 
356 	/* only valid for physical mode */
357 	if (adev->asic_type < CHIP_POLARIS10) {
358 		for (i = 0; i < adev->uvd.max_handles; ++i)
359 			if (atomic_read(&adev->uvd.handles[i]))
360 				break;
361 
362 		if (i == adev->uvd.max_handles)
363 			return 0;
364 	}
365 
366 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
367 		if (adev->uvd.harvest_config & (1 << j))
368 			continue;
369 		if (adev->uvd.inst[j].vcpu_bo == NULL)
370 			continue;
371 
372 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
373 		ptr = adev->uvd.inst[j].cpu_addr;
374 
375 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
376 		if (!adev->uvd.inst[j].saved_bo)
377 			return -ENOMEM;
378 
379 		/* re-write 0 since err_event_athub will corrupt VCPU buffer */
380 		if (in_ras_intr)
381 			memset(adev->uvd.inst[j].saved_bo, 0, size);
382 		else
383 			memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
384 	}
385 
386 	if (in_ras_intr)
387 		DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
388 
389 	return 0;
390 }
391 
392 int amdgpu_uvd_resume(struct amdgpu_device *adev)
393 {
394 	unsigned size;
395 	void *ptr;
396 	int i;
397 
398 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
399 		if (adev->uvd.harvest_config & (1 << i))
400 			continue;
401 		if (adev->uvd.inst[i].vcpu_bo == NULL)
402 			return -EINVAL;
403 
404 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
405 		ptr = adev->uvd.inst[i].cpu_addr;
406 
407 		if (adev->uvd.inst[i].saved_bo != NULL) {
408 			memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
409 			kvfree(adev->uvd.inst[i].saved_bo);
410 			adev->uvd.inst[i].saved_bo = NULL;
411 		} else {
412 			const struct common_firmware_header *hdr;
413 			unsigned offset;
414 
415 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
416 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
417 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
418 				memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
419 					    le32_to_cpu(hdr->ucode_size_bytes));
420 				size -= le32_to_cpu(hdr->ucode_size_bytes);
421 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
422 			}
423 			memset_io(ptr, 0, size);
424 			/* to restore uvd fence seq */
425 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
426 		}
427 	}
428 	return 0;
429 }
430 
431 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
432 {
433 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
434 	int i, r;
435 
436 	for (i = 0; i < adev->uvd.max_handles; ++i) {
437 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
438 
439 		if (handle != 0 && adev->uvd.filp[i] == filp) {
440 			struct dma_fence *fence;
441 
442 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
443 						       &fence);
444 			if (r) {
445 				DRM_ERROR("Error destroying UVD %d!\n", r);
446 				continue;
447 			}
448 
449 			dma_fence_wait(fence, false);
450 			dma_fence_put(fence);
451 
452 			adev->uvd.filp[i] = NULL;
453 			atomic_set(&adev->uvd.handles[i], 0);
454 		}
455 	}
456 }
457 
458 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
459 {
460 	int i;
461 	for (i = 0; i < abo->placement.num_placement; ++i) {
462 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
463 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
464 	}
465 }
466 
467 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
468 {
469 	uint32_t lo, hi;
470 	uint64_t addr;
471 
472 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
473 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
474 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
475 
476 	return addr;
477 }
478 
479 /**
480  * amdgpu_uvd_cs_pass1 - first parsing round
481  *
482  * @ctx: UVD parser context
483  *
484  * Make sure UVD message and feedback buffers are in VRAM and
485  * nobody is violating an 256MB boundary.
486  */
487 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
488 {
489 	struct ttm_operation_ctx tctx = { false, false };
490 	struct amdgpu_bo_va_mapping *mapping;
491 	struct amdgpu_bo *bo;
492 	uint32_t cmd;
493 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
494 	int r = 0;
495 
496 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
497 	if (r) {
498 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
499 		return r;
500 	}
501 
502 	if (!ctx->parser->adev->uvd.address_64_bit) {
503 		/* check if it's a message or feedback command */
504 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
505 		if (cmd == 0x0 || cmd == 0x3) {
506 			/* yes, force it into VRAM */
507 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
508 			amdgpu_bo_placement_from_domain(bo, domain);
509 		}
510 		amdgpu_uvd_force_into_uvd_segment(bo);
511 
512 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
513 	}
514 
515 	return r;
516 }
517 
518 /**
519  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
520  *
521  * @msg: pointer to message structure
522  * @buf_sizes: returned buffer sizes
523  *
524  * Peek into the decode message and calculate the necessary buffer sizes.
525  */
526 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
527 	unsigned buf_sizes[])
528 {
529 	unsigned stream_type = msg[4];
530 	unsigned width = msg[6];
531 	unsigned height = msg[7];
532 	unsigned dpb_size = msg[9];
533 	unsigned pitch = msg[28];
534 	unsigned level = msg[57];
535 
536 	unsigned width_in_mb = width / 16;
537 	unsigned height_in_mb = ALIGN(height / 16, 2);
538 	unsigned fs_in_mb = width_in_mb * height_in_mb;
539 
540 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
541 	unsigned min_ctx_size = ~0;
542 
543 	image_size = width * height;
544 	image_size += image_size / 2;
545 	image_size = ALIGN(image_size, 1024);
546 
547 	switch (stream_type) {
548 	case 0: /* H264 */
549 		switch(level) {
550 		case 30:
551 			num_dpb_buffer = 8100 / fs_in_mb;
552 			break;
553 		case 31:
554 			num_dpb_buffer = 18000 / fs_in_mb;
555 			break;
556 		case 32:
557 			num_dpb_buffer = 20480 / fs_in_mb;
558 			break;
559 		case 41:
560 			num_dpb_buffer = 32768 / fs_in_mb;
561 			break;
562 		case 42:
563 			num_dpb_buffer = 34816 / fs_in_mb;
564 			break;
565 		case 50:
566 			num_dpb_buffer = 110400 / fs_in_mb;
567 			break;
568 		case 51:
569 			num_dpb_buffer = 184320 / fs_in_mb;
570 			break;
571 		default:
572 			num_dpb_buffer = 184320 / fs_in_mb;
573 			break;
574 		}
575 		num_dpb_buffer++;
576 		if (num_dpb_buffer > 17)
577 			num_dpb_buffer = 17;
578 
579 		/* reference picture buffer */
580 		min_dpb_size = image_size * num_dpb_buffer;
581 
582 		/* macroblock context buffer */
583 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
584 
585 		/* IT surface buffer */
586 		min_dpb_size += width_in_mb * height_in_mb * 32;
587 		break;
588 
589 	case 1: /* VC1 */
590 
591 		/* reference picture buffer */
592 		min_dpb_size = image_size * 3;
593 
594 		/* CONTEXT_BUFFER */
595 		min_dpb_size += width_in_mb * height_in_mb * 128;
596 
597 		/* IT surface buffer */
598 		min_dpb_size += width_in_mb * 64;
599 
600 		/* DB surface buffer */
601 		min_dpb_size += width_in_mb * 128;
602 
603 		/* BP */
604 		tmp = max(width_in_mb, height_in_mb);
605 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
606 		break;
607 
608 	case 3: /* MPEG2 */
609 
610 		/* reference picture buffer */
611 		min_dpb_size = image_size * 3;
612 		break;
613 
614 	case 4: /* MPEG4 */
615 
616 		/* reference picture buffer */
617 		min_dpb_size = image_size * 3;
618 
619 		/* CM */
620 		min_dpb_size += width_in_mb * height_in_mb * 64;
621 
622 		/* IT surface buffer */
623 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
624 		break;
625 
626 	case 7: /* H264 Perf */
627 		switch(level) {
628 		case 30:
629 			num_dpb_buffer = 8100 / fs_in_mb;
630 			break;
631 		case 31:
632 			num_dpb_buffer = 18000 / fs_in_mb;
633 			break;
634 		case 32:
635 			num_dpb_buffer = 20480 / fs_in_mb;
636 			break;
637 		case 41:
638 			num_dpb_buffer = 32768 / fs_in_mb;
639 			break;
640 		case 42:
641 			num_dpb_buffer = 34816 / fs_in_mb;
642 			break;
643 		case 50:
644 			num_dpb_buffer = 110400 / fs_in_mb;
645 			break;
646 		case 51:
647 			num_dpb_buffer = 184320 / fs_in_mb;
648 			break;
649 		default:
650 			num_dpb_buffer = 184320 / fs_in_mb;
651 			break;
652 		}
653 		num_dpb_buffer++;
654 		if (num_dpb_buffer > 17)
655 			num_dpb_buffer = 17;
656 
657 		/* reference picture buffer */
658 		min_dpb_size = image_size * num_dpb_buffer;
659 
660 		if (!adev->uvd.use_ctx_buf){
661 			/* macroblock context buffer */
662 			min_dpb_size +=
663 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
664 
665 			/* IT surface buffer */
666 			min_dpb_size += width_in_mb * height_in_mb * 32;
667 		} else {
668 			/* macroblock context buffer */
669 			min_ctx_size =
670 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
671 		}
672 		break;
673 
674 	case 8: /* MJPEG */
675 		min_dpb_size = 0;
676 		break;
677 
678 	case 16: /* H265 */
679 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
680 		image_size = ALIGN(image_size, 256);
681 
682 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
683 		min_dpb_size = image_size * num_dpb_buffer;
684 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
685 					   * 16 * num_dpb_buffer + 52 * 1024;
686 		break;
687 
688 	default:
689 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
690 		return -EINVAL;
691 	}
692 
693 	if (width > pitch) {
694 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
695 		return -EINVAL;
696 	}
697 
698 	if (dpb_size < min_dpb_size) {
699 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
700 			  dpb_size, min_dpb_size);
701 		return -EINVAL;
702 	}
703 
704 	buf_sizes[0x1] = dpb_size;
705 	buf_sizes[0x2] = image_size;
706 	buf_sizes[0x4] = min_ctx_size;
707 	/* store image width to adjust nb memory pstate */
708 	adev->uvd.decode_image_width = width;
709 	return 0;
710 }
711 
712 /**
713  * amdgpu_uvd_cs_msg - handle UVD message
714  *
715  * @ctx: UVD parser context
716  * @bo: buffer object containing the message
717  * @offset: offset into the buffer object
718  *
719  * Peek into the UVD message and extract the session id.
720  * Make sure that we don't open up to many sessions.
721  */
722 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
723 			     struct amdgpu_bo *bo, unsigned offset)
724 {
725 	struct amdgpu_device *adev = ctx->parser->adev;
726 	int32_t *msg, msg_type, handle;
727 	void *ptr;
728 	long r;
729 	int i;
730 
731 	if (offset & 0x3F) {
732 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
733 		return -EINVAL;
734 	}
735 
736 	r = amdgpu_bo_kmap(bo, &ptr);
737 	if (r) {
738 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
739 		return r;
740 	}
741 
742 	msg = ptr + offset;
743 
744 	msg_type = msg[1];
745 	handle = msg[2];
746 
747 	if (handle == 0) {
748 		DRM_ERROR("Invalid UVD handle!\n");
749 		return -EINVAL;
750 	}
751 
752 	switch (msg_type) {
753 	case 0:
754 		/* it's a create msg, calc image size (width * height) */
755 		amdgpu_bo_kunmap(bo);
756 
757 		/* try to alloc a new handle */
758 		for (i = 0; i < adev->uvd.max_handles; ++i) {
759 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
760 				DRM_ERROR(")Handle 0x%x already in use!\n",
761 					  handle);
762 				return -EINVAL;
763 			}
764 
765 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
766 				adev->uvd.filp[i] = ctx->parser->filp;
767 				return 0;
768 			}
769 		}
770 
771 		DRM_ERROR("No more free UVD handles!\n");
772 		return -ENOSPC;
773 
774 	case 1:
775 		/* it's a decode msg, calc buffer sizes */
776 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
777 		amdgpu_bo_kunmap(bo);
778 		if (r)
779 			return r;
780 
781 		/* validate the handle */
782 		for (i = 0; i < adev->uvd.max_handles; ++i) {
783 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
784 				if (adev->uvd.filp[i] != ctx->parser->filp) {
785 					DRM_ERROR("UVD handle collision detected!\n");
786 					return -EINVAL;
787 				}
788 				return 0;
789 			}
790 		}
791 
792 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
793 		return -ENOENT;
794 
795 	case 2:
796 		/* it's a destroy msg, free the handle */
797 		for (i = 0; i < adev->uvd.max_handles; ++i)
798 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
799 		amdgpu_bo_kunmap(bo);
800 		return 0;
801 
802 	default:
803 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
804 		return -EINVAL;
805 	}
806 	BUG();
807 	return -EINVAL;
808 }
809 
810 /**
811  * amdgpu_uvd_cs_pass2 - second parsing round
812  *
813  * @ctx: UVD parser context
814  *
815  * Patch buffer addresses, make sure buffer sizes are correct.
816  */
817 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
818 {
819 	struct amdgpu_bo_va_mapping *mapping;
820 	struct amdgpu_bo *bo;
821 	uint32_t cmd;
822 	uint64_t start, end;
823 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
824 	int r;
825 
826 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
827 	if (r) {
828 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
829 		return r;
830 	}
831 
832 	start = amdgpu_bo_gpu_offset(bo);
833 
834 	end = (mapping->last + 1 - mapping->start);
835 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
836 
837 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
838 	start += addr;
839 
840 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
841 			    lower_32_bits(start));
842 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
843 			    upper_32_bits(start));
844 
845 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
846 	if (cmd < 0x4) {
847 		if ((end - start) < ctx->buf_sizes[cmd]) {
848 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
849 				  (unsigned)(end - start),
850 				  ctx->buf_sizes[cmd]);
851 			return -EINVAL;
852 		}
853 
854 	} else if (cmd == 0x206) {
855 		if ((end - start) < ctx->buf_sizes[4]) {
856 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
857 					  (unsigned)(end - start),
858 					  ctx->buf_sizes[4]);
859 			return -EINVAL;
860 		}
861 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
862 		DRM_ERROR("invalid UVD command %X!\n", cmd);
863 		return -EINVAL;
864 	}
865 
866 	if (!ctx->parser->adev->uvd.address_64_bit) {
867 		if ((start >> 28) != ((end - 1) >> 28)) {
868 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
869 				  start, end);
870 			return -EINVAL;
871 		}
872 
873 		if ((cmd == 0 || cmd == 0x3) &&
874 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
875 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
876 				  start, end);
877 			return -EINVAL;
878 		}
879 	}
880 
881 	if (cmd == 0) {
882 		ctx->has_msg_cmd = true;
883 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
884 		if (r)
885 			return r;
886 	} else if (!ctx->has_msg_cmd) {
887 		DRM_ERROR("Message needed before other commands are send!\n");
888 		return -EINVAL;
889 	}
890 
891 	return 0;
892 }
893 
894 /**
895  * amdgpu_uvd_cs_reg - parse register writes
896  *
897  * @ctx: UVD parser context
898  * @cb: callback function
899  *
900  * Parse the register writes, call cb on each complete command.
901  */
902 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
903 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
904 {
905 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
906 	int i, r;
907 
908 	ctx->idx++;
909 	for (i = 0; i <= ctx->count; ++i) {
910 		unsigned reg = ctx->reg + i;
911 
912 		if (ctx->idx >= ib->length_dw) {
913 			DRM_ERROR("Register command after end of CS!\n");
914 			return -EINVAL;
915 		}
916 
917 		switch (reg) {
918 		case mmUVD_GPCOM_VCPU_DATA0:
919 			ctx->data0 = ctx->idx;
920 			break;
921 		case mmUVD_GPCOM_VCPU_DATA1:
922 			ctx->data1 = ctx->idx;
923 			break;
924 		case mmUVD_GPCOM_VCPU_CMD:
925 			r = cb(ctx);
926 			if (r)
927 				return r;
928 			break;
929 		case mmUVD_ENGINE_CNTL:
930 		case mmUVD_NO_OP:
931 			break;
932 		default:
933 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
934 			return -EINVAL;
935 		}
936 		ctx->idx++;
937 	}
938 	return 0;
939 }
940 
941 /**
942  * amdgpu_uvd_cs_packets - parse UVD packets
943  *
944  * @ctx: UVD parser context
945  * @cb: callback function
946  *
947  * Parse the command stream packets.
948  */
949 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
950 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
951 {
952 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
953 	int r;
954 
955 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
956 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
957 		unsigned type = CP_PACKET_GET_TYPE(cmd);
958 		switch (type) {
959 		case PACKET_TYPE0:
960 			ctx->reg = CP_PACKET0_GET_REG(cmd);
961 			ctx->count = CP_PACKET_GET_COUNT(cmd);
962 			r = amdgpu_uvd_cs_reg(ctx, cb);
963 			if (r)
964 				return r;
965 			break;
966 		case PACKET_TYPE2:
967 			++ctx->idx;
968 			break;
969 		default:
970 			DRM_ERROR("Unknown packet type %d !\n", type);
971 			return -EINVAL;
972 		}
973 	}
974 	return 0;
975 }
976 
977 /**
978  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
979  *
980  * @parser: Command submission parser context
981  *
982  * Parse the command stream, patch in addresses as necessary.
983  */
984 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
985 {
986 	struct amdgpu_uvd_cs_ctx ctx = {};
987 	unsigned buf_sizes[] = {
988 		[0x00000000]	=	2048,
989 		[0x00000001]	=	0xFFFFFFFF,
990 		[0x00000002]	=	0xFFFFFFFF,
991 		[0x00000003]	=	2048,
992 		[0x00000004]	=	0xFFFFFFFF,
993 	};
994 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
995 	int r;
996 
997 	parser->job->vm = NULL;
998 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
999 
1000 	if (ib->length_dw % 16) {
1001 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1002 			  ib->length_dw);
1003 		return -EINVAL;
1004 	}
1005 
1006 	ctx.parser = parser;
1007 	ctx.buf_sizes = buf_sizes;
1008 	ctx.ib_idx = ib_idx;
1009 
1010 	/* first round only required on chips without UVD 64 bit address support */
1011 	if (!parser->adev->uvd.address_64_bit) {
1012 		/* first round, make sure the buffers are actually in the UVD segment */
1013 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1014 		if (r)
1015 			return r;
1016 	}
1017 
1018 	/* second round, patch buffer addresses into the command stream */
1019 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1020 	if (r)
1021 		return r;
1022 
1023 	if (!ctx.has_msg_cmd) {
1024 		DRM_ERROR("UVD-IBs need a msg command!\n");
1025 		return -EINVAL;
1026 	}
1027 
1028 	return 0;
1029 }
1030 
1031 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1032 			       bool direct, struct dma_fence **fence)
1033 {
1034 	struct amdgpu_device *adev = ring->adev;
1035 	struct dma_fence *f = NULL;
1036 	struct amdgpu_job *job;
1037 	struct amdgpu_ib *ib;
1038 	uint32_t data[4];
1039 	uint64_t addr;
1040 	long r;
1041 	int i;
1042 	unsigned offset_idx = 0;
1043 	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1044 
1045 	amdgpu_bo_kunmap(bo);
1046 	amdgpu_bo_unpin(bo);
1047 
1048 	if (!ring->adev->uvd.address_64_bit) {
1049 		struct ttm_operation_ctx ctx = { true, false };
1050 
1051 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1052 		amdgpu_uvd_force_into_uvd_segment(bo);
1053 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1054 		if (r)
1055 			goto err;
1056 	}
1057 
1058 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1059 	if (r)
1060 		goto err;
1061 
1062 	if (adev->asic_type >= CHIP_VEGA10) {
1063 		offset_idx = 1 + ring->me;
1064 		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1065 		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1066 	}
1067 
1068 	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1069 	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1070 	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1071 	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1072 
1073 	ib = &job->ibs[0];
1074 	addr = amdgpu_bo_gpu_offset(bo);
1075 	ib->ptr[0] = data[0];
1076 	ib->ptr[1] = addr;
1077 	ib->ptr[2] = data[1];
1078 	ib->ptr[3] = addr >> 32;
1079 	ib->ptr[4] = data[2];
1080 	ib->ptr[5] = 0;
1081 	for (i = 6; i < 16; i += 2) {
1082 		ib->ptr[i] = data[3];
1083 		ib->ptr[i+1] = 0;
1084 	}
1085 	ib->length_dw = 16;
1086 
1087 	if (direct) {
1088 		r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv,
1089 							true, false,
1090 							msecs_to_jiffies(10));
1091 		if (r == 0)
1092 			r = -ETIMEDOUT;
1093 		if (r < 0)
1094 			goto err_free;
1095 
1096 		r = amdgpu_job_submit_direct(job, ring, &f);
1097 		if (r)
1098 			goto err_free;
1099 	} else {
1100 		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1101 				     AMDGPU_FENCE_OWNER_UNDEFINED, false);
1102 		if (r)
1103 			goto err_free;
1104 
1105 		r = amdgpu_job_submit(job, &adev->uvd.entity,
1106 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1107 		if (r)
1108 			goto err_free;
1109 	}
1110 
1111 	amdgpu_bo_fence(bo, f, false);
1112 	amdgpu_bo_unreserve(bo);
1113 	amdgpu_bo_unref(&bo);
1114 
1115 	if (fence)
1116 		*fence = dma_fence_get(f);
1117 	dma_fence_put(f);
1118 
1119 	return 0;
1120 
1121 err_free:
1122 	amdgpu_job_free(job);
1123 
1124 err:
1125 	amdgpu_bo_unreserve(bo);
1126 	amdgpu_bo_unref(&bo);
1127 	return r;
1128 }
1129 
1130 /* multiple fence commands without any stream commands in between can
1131    crash the vcpu so just try to emmit a dummy create/destroy msg to
1132    avoid this */
1133 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1134 			      struct dma_fence **fence)
1135 {
1136 	struct amdgpu_device *adev = ring->adev;
1137 	struct amdgpu_bo *bo = NULL;
1138 	uint32_t *msg;
1139 	int r, i;
1140 
1141 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1142 				      AMDGPU_GEM_DOMAIN_VRAM,
1143 				      &bo, NULL, (void **)&msg);
1144 	if (r)
1145 		return r;
1146 
1147 	/* stitch together an UVD create msg */
1148 	msg[0] = cpu_to_le32(0x00000de4);
1149 	msg[1] = cpu_to_le32(0x00000000);
1150 	msg[2] = cpu_to_le32(handle);
1151 	msg[3] = cpu_to_le32(0x00000000);
1152 	msg[4] = cpu_to_le32(0x00000000);
1153 	msg[5] = cpu_to_le32(0x00000000);
1154 	msg[6] = cpu_to_le32(0x00000000);
1155 	msg[7] = cpu_to_le32(0x00000780);
1156 	msg[8] = cpu_to_le32(0x00000440);
1157 	msg[9] = cpu_to_le32(0x00000000);
1158 	msg[10] = cpu_to_le32(0x01b37000);
1159 	for (i = 11; i < 1024; ++i)
1160 		msg[i] = cpu_to_le32(0x0);
1161 
1162 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1163 }
1164 
1165 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1166 			       bool direct, struct dma_fence **fence)
1167 {
1168 	struct amdgpu_device *adev = ring->adev;
1169 	struct amdgpu_bo *bo = NULL;
1170 	uint32_t *msg;
1171 	int r, i;
1172 
1173 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1174 				      AMDGPU_GEM_DOMAIN_VRAM,
1175 				      &bo, NULL, (void **)&msg);
1176 	if (r)
1177 		return r;
1178 
1179 	/* stitch together an UVD destroy msg */
1180 	msg[0] = cpu_to_le32(0x00000de4);
1181 	msg[1] = cpu_to_le32(0x00000002);
1182 	msg[2] = cpu_to_le32(handle);
1183 	msg[3] = cpu_to_le32(0x00000000);
1184 	for (i = 4; i < 1024; ++i)
1185 		msg[i] = cpu_to_le32(0x0);
1186 
1187 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1188 }
1189 
1190 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1191 {
1192 	struct amdgpu_device *adev =
1193 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1194 	unsigned fences = 0, i, j;
1195 
1196 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1197 		if (adev->uvd.harvest_config & (1 << i))
1198 			continue;
1199 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1200 		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1201 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1202 		}
1203 	}
1204 
1205 	if (fences == 0) {
1206 		if (adev->pm.dpm_enabled) {
1207 			amdgpu_dpm_enable_uvd(adev, false);
1208 		} else {
1209 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1210 			/* shutdown the UVD block */
1211 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1212 							       AMD_PG_STATE_GATE);
1213 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1214 							       AMD_CG_STATE_GATE);
1215 		}
1216 	} else {
1217 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1218 	}
1219 }
1220 
1221 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1222 {
1223 	struct amdgpu_device *adev = ring->adev;
1224 	bool set_clocks;
1225 
1226 	if (amdgpu_sriov_vf(adev))
1227 		return;
1228 
1229 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1230 	if (set_clocks) {
1231 		if (adev->pm.dpm_enabled) {
1232 			amdgpu_dpm_enable_uvd(adev, true);
1233 		} else {
1234 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1235 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1236 							       AMD_CG_STATE_UNGATE);
1237 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1238 							       AMD_PG_STATE_UNGATE);
1239 		}
1240 	}
1241 }
1242 
1243 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1244 {
1245 	if (!amdgpu_sriov_vf(ring->adev))
1246 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1247 }
1248 
1249 /**
1250  * amdgpu_uvd_ring_test_ib - test ib execution
1251  *
1252  * @ring: amdgpu_ring pointer
1253  *
1254  * Test if we can successfully execute an IB
1255  */
1256 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1257 {
1258 	struct dma_fence *fence;
1259 	long r;
1260 
1261 	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1262 	if (r)
1263 		goto error;
1264 
1265 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1266 	if (r)
1267 		goto error;
1268 
1269 	r = dma_fence_wait_timeout(fence, false, timeout);
1270 	if (r == 0)
1271 		r = -ETIMEDOUT;
1272 	else if (r > 0)
1273 		r = 0;
1274 
1275 	dma_fence_put(fence);
1276 
1277 error:
1278 	return r;
1279 }
1280 
1281 /**
1282  * amdgpu_uvd_used_handles - returns used UVD handles
1283  *
1284  * @adev: amdgpu_device pointer
1285  *
1286  * Returns the number of UVD handles in use
1287  */
1288 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1289 {
1290 	unsigned i;
1291 	uint32_t used_handles = 0;
1292 
1293 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1294 		/*
1295 		 * Handles can be freed in any order, and not
1296 		 * necessarily linear. So we need to count
1297 		 * all non-zero handles.
1298 		 */
1299 		if (atomic_read(&adev->uvd.handles[i]))
1300 			used_handles++;
1301 	}
1302 
1303 	return used_handles;
1304 }
1305