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 #include <drm/drm_drv.h>
36 
37 #include "amdgpu.h"
38 #include "amdgpu_pm.h"
39 #include "amdgpu_uvd.h"
40 #include "cikd.h"
41 #include "uvd/uvd_4_2_d.h"
42 
43 #include "amdgpu_ras.h"
44 
45 /* 1 second timeout */
46 #define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
47 
48 /* Firmware versions for VI */
49 #define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
50 #define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
51 #define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
52 #define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
53 
54 /* Polaris10/11 firmware version */
55 #define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
56 
57 /* Firmware Names */
58 #ifdef CONFIG_DRM_AMDGPU_SI
59 #define FIRMWARE_TAHITI		"amdgpu/tahiti_uvd.bin"
60 #define FIRMWARE_VERDE		"amdgpu/verde_uvd.bin"
61 #define FIRMWARE_PITCAIRN	"amdgpu/pitcairn_uvd.bin"
62 #define FIRMWARE_OLAND		"amdgpu/oland_uvd.bin"
63 #endif
64 #ifdef CONFIG_DRM_AMDGPU_CIK
65 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_uvd.bin"
66 #define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
67 #define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
68 #define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
69 #define FIRMWARE_MULLINS	"amdgpu/mullins_uvd.bin"
70 #endif
71 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
72 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
73 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
74 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
75 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
76 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
77 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
78 #define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
79 
80 #define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
81 #define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
82 #define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
83 
84 /* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
85 #define UVD_GPCOM_VCPU_CMD		0x03c3
86 #define UVD_GPCOM_VCPU_DATA0	0x03c4
87 #define UVD_GPCOM_VCPU_DATA1	0x03c5
88 #define UVD_NO_OP				0x03ff
89 #define UVD_BASE_SI				0x3800
90 
91 /*
92  * amdgpu_uvd_cs_ctx - Command submission parser context
93  *
94  * Used for emulating virtual memory support on UVD 4.2.
95  */
96 struct amdgpu_uvd_cs_ctx {
97 	struct amdgpu_cs_parser *parser;
98 	unsigned reg, count;
99 	unsigned data0, data1;
100 	unsigned idx;
101 	unsigned ib_idx;
102 
103 	/* does the IB has a msg command */
104 	bool has_msg_cmd;
105 
106 	/* minimum buffer sizes */
107 	unsigned *buf_sizes;
108 };
109 
110 #ifdef CONFIG_DRM_AMDGPU_SI
111 MODULE_FIRMWARE(FIRMWARE_TAHITI);
112 MODULE_FIRMWARE(FIRMWARE_VERDE);
113 MODULE_FIRMWARE(FIRMWARE_PITCAIRN);
114 MODULE_FIRMWARE(FIRMWARE_OLAND);
115 #endif
116 #ifdef CONFIG_DRM_AMDGPU_CIK
117 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
118 MODULE_FIRMWARE(FIRMWARE_KABINI);
119 MODULE_FIRMWARE(FIRMWARE_KAVERI);
120 MODULE_FIRMWARE(FIRMWARE_HAWAII);
121 MODULE_FIRMWARE(FIRMWARE_MULLINS);
122 #endif
123 MODULE_FIRMWARE(FIRMWARE_TONGA);
124 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
125 MODULE_FIRMWARE(FIRMWARE_FIJI);
126 MODULE_FIRMWARE(FIRMWARE_STONEY);
127 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
128 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
129 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
130 MODULE_FIRMWARE(FIRMWARE_VEGAM);
131 
132 MODULE_FIRMWARE(FIRMWARE_VEGA10);
133 MODULE_FIRMWARE(FIRMWARE_VEGA12);
134 MODULE_FIRMWARE(FIRMWARE_VEGA20);
135 
136 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
137 
138 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
139 {
140 	unsigned long bo_size;
141 	const char *fw_name;
142 	const struct common_firmware_header *hdr;
143 	unsigned family_id;
144 	int i, j, r;
145 
146 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
147 
148 	switch (adev->asic_type) {
149 #ifdef CONFIG_DRM_AMDGPU_SI
150 	case CHIP_TAHITI:
151 		fw_name = FIRMWARE_TAHITI;
152 		break;
153 	case CHIP_VERDE:
154 		fw_name = FIRMWARE_VERDE;
155 		break;
156 	case CHIP_PITCAIRN:
157 		fw_name = FIRMWARE_PITCAIRN;
158 		break;
159 	case CHIP_OLAND:
160 		fw_name = FIRMWARE_OLAND;
161 		break;
162 #endif
163 #ifdef CONFIG_DRM_AMDGPU_CIK
164 	case CHIP_BONAIRE:
165 		fw_name = FIRMWARE_BONAIRE;
166 		break;
167 	case CHIP_KABINI:
168 		fw_name = FIRMWARE_KABINI;
169 		break;
170 	case CHIP_KAVERI:
171 		fw_name = FIRMWARE_KAVERI;
172 		break;
173 	case CHIP_HAWAII:
174 		fw_name = FIRMWARE_HAWAII;
175 		break;
176 	case CHIP_MULLINS:
177 		fw_name = FIRMWARE_MULLINS;
178 		break;
179 #endif
180 	case CHIP_TONGA:
181 		fw_name = FIRMWARE_TONGA;
182 		break;
183 	case CHIP_FIJI:
184 		fw_name = FIRMWARE_FIJI;
185 		break;
186 	case CHIP_CARRIZO:
187 		fw_name = FIRMWARE_CARRIZO;
188 		break;
189 	case CHIP_STONEY:
190 		fw_name = FIRMWARE_STONEY;
191 		break;
192 	case CHIP_POLARIS10:
193 		fw_name = FIRMWARE_POLARIS10;
194 		break;
195 	case CHIP_POLARIS11:
196 		fw_name = FIRMWARE_POLARIS11;
197 		break;
198 	case CHIP_POLARIS12:
199 		fw_name = FIRMWARE_POLARIS12;
200 		break;
201 	case CHIP_VEGA10:
202 		fw_name = FIRMWARE_VEGA10;
203 		break;
204 	case CHIP_VEGA12:
205 		fw_name = FIRMWARE_VEGA12;
206 		break;
207 	case CHIP_VEGAM:
208 		fw_name = FIRMWARE_VEGAM;
209 		break;
210 	case CHIP_VEGA20:
211 		fw_name = FIRMWARE_VEGA20;
212 		break;
213 	default:
214 		return -EINVAL;
215 	}
216 
217 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
218 	if (r) {
219 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
220 			fw_name);
221 		return r;
222 	}
223 
224 	r = amdgpu_ucode_validate(adev->uvd.fw);
225 	if (r) {
226 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
227 			fw_name);
228 		release_firmware(adev->uvd.fw);
229 		adev->uvd.fw = NULL;
230 		return r;
231 	}
232 
233 	/* Set the default UVD handles that the firmware can handle */
234 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
235 
236 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
237 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
238 
239 	if (adev->asic_type < CHIP_VEGA20) {
240 		unsigned version_major, version_minor;
241 
242 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
243 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
244 		DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
245 			version_major, version_minor, family_id);
246 
247 		/*
248 		 * Limit the number of UVD handles depending on microcode major
249 		 * and minor versions. The firmware version which has 40 UVD
250 		 * instances support is 1.80. So all subsequent versions should
251 		 * also have the same support.
252 		 */
253 		if ((version_major > 0x01) ||
254 		    ((version_major == 0x01) && (version_minor >= 0x50)))
255 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
256 
257 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
258 					(family_id << 8));
259 
260 		if ((adev->asic_type == CHIP_POLARIS10 ||
261 		     adev->asic_type == CHIP_POLARIS11) &&
262 		    (adev->uvd.fw_version < FW_1_66_16))
263 			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
264 				  version_major, version_minor);
265 	} else {
266 		unsigned int enc_major, enc_minor, dec_minor;
267 
268 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
269 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
270 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
271 		DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
272 			enc_major, enc_minor, dec_minor, family_id);
273 
274 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
275 
276 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
277 	}
278 
279 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
280 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
281 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
282 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
283 
284 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
285 		if (adev->uvd.harvest_config & (1 << j))
286 			continue;
287 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
288 					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
289 					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
290 		if (r) {
291 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
292 			return r;
293 		}
294 	}
295 
296 	for (i = 0; i < adev->uvd.max_handles; ++i) {
297 		atomic_set(&adev->uvd.handles[i], 0);
298 		adev->uvd.filp[i] = NULL;
299 	}
300 
301 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
302 	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
303 		adev->uvd.address_64_bit = true;
304 
305 	switch (adev->asic_type) {
306 	case CHIP_TONGA:
307 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
308 		break;
309 	case CHIP_CARRIZO:
310 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
311 		break;
312 	case CHIP_FIJI:
313 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
314 		break;
315 	case CHIP_STONEY:
316 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
317 		break;
318 	default:
319 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
320 	}
321 
322 	return 0;
323 }
324 
325 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
326 {
327 	int i, j;
328 
329 	drm_sched_entity_destroy(&adev->uvd.entity);
330 
331 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
332 		if (adev->uvd.harvest_config & (1 << j))
333 			continue;
334 		kvfree(adev->uvd.inst[j].saved_bo);
335 
336 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
337 				      &adev->uvd.inst[j].gpu_addr,
338 				      (void **)&adev->uvd.inst[j].cpu_addr);
339 
340 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
341 
342 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
343 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
344 	}
345 	release_firmware(adev->uvd.fw);
346 
347 	return 0;
348 }
349 
350 /**
351  * amdgpu_uvd_entity_init - init entity
352  *
353  * @adev: amdgpu_device pointer
354  *
355  */
356 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
357 {
358 	struct amdgpu_ring *ring;
359 	struct drm_gpu_scheduler *sched;
360 	int r;
361 
362 	ring = &adev->uvd.inst[0].ring;
363 	sched = &ring->sched;
364 	r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
365 				  &sched, 1, NULL);
366 	if (r) {
367 		DRM_ERROR("Failed setting up UVD kernel entity.\n");
368 		return r;
369 	}
370 
371 	return 0;
372 }
373 
374 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
375 {
376 	unsigned size;
377 	void *ptr;
378 	int i, j, idx;
379 	bool in_ras_intr = amdgpu_ras_intr_triggered();
380 
381 	cancel_delayed_work_sync(&adev->uvd.idle_work);
382 
383 	/* only valid for physical mode */
384 	if (adev->asic_type < CHIP_POLARIS10) {
385 		for (i = 0; i < adev->uvd.max_handles; ++i)
386 			if (atomic_read(&adev->uvd.handles[i]))
387 				break;
388 
389 		if (i == adev->uvd.max_handles)
390 			return 0;
391 	}
392 
393 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
394 		if (adev->uvd.harvest_config & (1 << j))
395 			continue;
396 		if (adev->uvd.inst[j].vcpu_bo == NULL)
397 			continue;
398 
399 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
400 		ptr = adev->uvd.inst[j].cpu_addr;
401 
402 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
403 		if (!adev->uvd.inst[j].saved_bo)
404 			return -ENOMEM;
405 
406 		if (drm_dev_enter(&adev->ddev, &idx)) {
407 			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
408 			if (in_ras_intr)
409 				memset(adev->uvd.inst[j].saved_bo, 0, size);
410 			else
411 				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
412 
413 			drm_dev_exit(idx);
414 		}
415 	}
416 
417 	if (in_ras_intr)
418 		DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
419 
420 	return 0;
421 }
422 
423 int amdgpu_uvd_resume(struct amdgpu_device *adev)
424 {
425 	unsigned size;
426 	void *ptr;
427 	int i, idx;
428 
429 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
430 		if (adev->uvd.harvest_config & (1 << i))
431 			continue;
432 		if (adev->uvd.inst[i].vcpu_bo == NULL)
433 			return -EINVAL;
434 
435 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
436 		ptr = adev->uvd.inst[i].cpu_addr;
437 
438 		if (adev->uvd.inst[i].saved_bo != NULL) {
439 			if (drm_dev_enter(&adev->ddev, &idx)) {
440 				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
441 				drm_dev_exit(idx);
442 			}
443 			kvfree(adev->uvd.inst[i].saved_bo);
444 			adev->uvd.inst[i].saved_bo = NULL;
445 		} else {
446 			const struct common_firmware_header *hdr;
447 			unsigned offset;
448 
449 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
450 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
451 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
452 				if (drm_dev_enter(&adev->ddev, &idx)) {
453 					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
454 						    le32_to_cpu(hdr->ucode_size_bytes));
455 					drm_dev_exit(idx);
456 				}
457 				size -= le32_to_cpu(hdr->ucode_size_bytes);
458 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
459 			}
460 			memset_io(ptr, 0, size);
461 			/* to restore uvd fence seq */
462 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
463 		}
464 	}
465 	return 0;
466 }
467 
468 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
469 {
470 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
471 	int i, r;
472 
473 	for (i = 0; i < adev->uvd.max_handles; ++i) {
474 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
475 
476 		if (handle != 0 && adev->uvd.filp[i] == filp) {
477 			struct dma_fence *fence;
478 
479 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
480 						       &fence);
481 			if (r) {
482 				DRM_ERROR("Error destroying UVD %d!\n", r);
483 				continue;
484 			}
485 
486 			dma_fence_wait(fence, false);
487 			dma_fence_put(fence);
488 
489 			adev->uvd.filp[i] = NULL;
490 			atomic_set(&adev->uvd.handles[i], 0);
491 		}
492 	}
493 }
494 
495 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
496 {
497 	int i;
498 	for (i = 0; i < abo->placement.num_placement; ++i) {
499 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
500 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
501 	}
502 }
503 
504 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
505 {
506 	uint32_t lo, hi;
507 	uint64_t addr;
508 
509 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
510 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
511 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
512 
513 	return addr;
514 }
515 
516 /**
517  * amdgpu_uvd_cs_pass1 - first parsing round
518  *
519  * @ctx: UVD parser context
520  *
521  * Make sure UVD message and feedback buffers are in VRAM and
522  * nobody is violating an 256MB boundary.
523  */
524 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
525 {
526 	struct ttm_operation_ctx tctx = { false, false };
527 	struct amdgpu_bo_va_mapping *mapping;
528 	struct amdgpu_bo *bo;
529 	uint32_t cmd;
530 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
531 	int r = 0;
532 
533 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
534 	if (r) {
535 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
536 		return r;
537 	}
538 
539 	if (!ctx->parser->adev->uvd.address_64_bit) {
540 		/* check if it's a message or feedback command */
541 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
542 		if (cmd == 0x0 || cmd == 0x3) {
543 			/* yes, force it into VRAM */
544 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
545 			amdgpu_bo_placement_from_domain(bo, domain);
546 		}
547 		amdgpu_uvd_force_into_uvd_segment(bo);
548 
549 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
550 	}
551 
552 	return r;
553 }
554 
555 /**
556  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
557  *
558  * @adev: amdgpu_device pointer
559  * @msg: pointer to message structure
560  * @buf_sizes: placeholder to put the different buffer lengths
561  *
562  * Peek into the decode message and calculate the necessary buffer sizes.
563  */
564 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
565 	unsigned buf_sizes[])
566 {
567 	unsigned stream_type = msg[4];
568 	unsigned width = msg[6];
569 	unsigned height = msg[7];
570 	unsigned dpb_size = msg[9];
571 	unsigned pitch = msg[28];
572 	unsigned level = msg[57];
573 
574 	unsigned width_in_mb = width / 16;
575 	unsigned height_in_mb = ALIGN(height / 16, 2);
576 	unsigned fs_in_mb = width_in_mb * height_in_mb;
577 
578 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
579 	unsigned min_ctx_size = ~0;
580 
581 	image_size = width * height;
582 	image_size += image_size / 2;
583 	image_size = ALIGN(image_size, 1024);
584 
585 	switch (stream_type) {
586 	case 0: /* H264 */
587 		switch(level) {
588 		case 30:
589 			num_dpb_buffer = 8100 / fs_in_mb;
590 			break;
591 		case 31:
592 			num_dpb_buffer = 18000 / fs_in_mb;
593 			break;
594 		case 32:
595 			num_dpb_buffer = 20480 / fs_in_mb;
596 			break;
597 		case 41:
598 			num_dpb_buffer = 32768 / fs_in_mb;
599 			break;
600 		case 42:
601 			num_dpb_buffer = 34816 / fs_in_mb;
602 			break;
603 		case 50:
604 			num_dpb_buffer = 110400 / fs_in_mb;
605 			break;
606 		case 51:
607 			num_dpb_buffer = 184320 / fs_in_mb;
608 			break;
609 		default:
610 			num_dpb_buffer = 184320 / fs_in_mb;
611 			break;
612 		}
613 		num_dpb_buffer++;
614 		if (num_dpb_buffer > 17)
615 			num_dpb_buffer = 17;
616 
617 		/* reference picture buffer */
618 		min_dpb_size = image_size * num_dpb_buffer;
619 
620 		/* macroblock context buffer */
621 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
622 
623 		/* IT surface buffer */
624 		min_dpb_size += width_in_mb * height_in_mb * 32;
625 		break;
626 
627 	case 1: /* VC1 */
628 
629 		/* reference picture buffer */
630 		min_dpb_size = image_size * 3;
631 
632 		/* CONTEXT_BUFFER */
633 		min_dpb_size += width_in_mb * height_in_mb * 128;
634 
635 		/* IT surface buffer */
636 		min_dpb_size += width_in_mb * 64;
637 
638 		/* DB surface buffer */
639 		min_dpb_size += width_in_mb * 128;
640 
641 		/* BP */
642 		tmp = max(width_in_mb, height_in_mb);
643 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
644 		break;
645 
646 	case 3: /* MPEG2 */
647 
648 		/* reference picture buffer */
649 		min_dpb_size = image_size * 3;
650 		break;
651 
652 	case 4: /* MPEG4 */
653 
654 		/* reference picture buffer */
655 		min_dpb_size = image_size * 3;
656 
657 		/* CM */
658 		min_dpb_size += width_in_mb * height_in_mb * 64;
659 
660 		/* IT surface buffer */
661 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
662 		break;
663 
664 	case 7: /* H264 Perf */
665 		switch(level) {
666 		case 30:
667 			num_dpb_buffer = 8100 / fs_in_mb;
668 			break;
669 		case 31:
670 			num_dpb_buffer = 18000 / fs_in_mb;
671 			break;
672 		case 32:
673 			num_dpb_buffer = 20480 / fs_in_mb;
674 			break;
675 		case 41:
676 			num_dpb_buffer = 32768 / fs_in_mb;
677 			break;
678 		case 42:
679 			num_dpb_buffer = 34816 / fs_in_mb;
680 			break;
681 		case 50:
682 			num_dpb_buffer = 110400 / fs_in_mb;
683 			break;
684 		case 51:
685 			num_dpb_buffer = 184320 / fs_in_mb;
686 			break;
687 		default:
688 			num_dpb_buffer = 184320 / fs_in_mb;
689 			break;
690 		}
691 		num_dpb_buffer++;
692 		if (num_dpb_buffer > 17)
693 			num_dpb_buffer = 17;
694 
695 		/* reference picture buffer */
696 		min_dpb_size = image_size * num_dpb_buffer;
697 
698 		if (!adev->uvd.use_ctx_buf){
699 			/* macroblock context buffer */
700 			min_dpb_size +=
701 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
702 
703 			/* IT surface buffer */
704 			min_dpb_size += width_in_mb * height_in_mb * 32;
705 		} else {
706 			/* macroblock context buffer */
707 			min_ctx_size =
708 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
709 		}
710 		break;
711 
712 	case 8: /* MJPEG */
713 		min_dpb_size = 0;
714 		break;
715 
716 	case 16: /* H265 */
717 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
718 		image_size = ALIGN(image_size, 256);
719 
720 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
721 		min_dpb_size = image_size * num_dpb_buffer;
722 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
723 					   * 16 * num_dpb_buffer + 52 * 1024;
724 		break;
725 
726 	default:
727 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
728 		return -EINVAL;
729 	}
730 
731 	if (width > pitch) {
732 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
733 		return -EINVAL;
734 	}
735 
736 	if (dpb_size < min_dpb_size) {
737 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
738 			  dpb_size, min_dpb_size);
739 		return -EINVAL;
740 	}
741 
742 	buf_sizes[0x1] = dpb_size;
743 	buf_sizes[0x2] = image_size;
744 	buf_sizes[0x4] = min_ctx_size;
745 	/* store image width to adjust nb memory pstate */
746 	adev->uvd.decode_image_width = width;
747 	return 0;
748 }
749 
750 /**
751  * amdgpu_uvd_cs_msg - handle UVD message
752  *
753  * @ctx: UVD parser context
754  * @bo: buffer object containing the message
755  * @offset: offset into the buffer object
756  *
757  * Peek into the UVD message and extract the session id.
758  * Make sure that we don't open up to many sessions.
759  */
760 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
761 			     struct amdgpu_bo *bo, unsigned offset)
762 {
763 	struct amdgpu_device *adev = ctx->parser->adev;
764 	int32_t *msg, msg_type, handle;
765 	void *ptr;
766 	long r;
767 	int i;
768 
769 	if (offset & 0x3F) {
770 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
771 		return -EINVAL;
772 	}
773 
774 	r = amdgpu_bo_kmap(bo, &ptr);
775 	if (r) {
776 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
777 		return r;
778 	}
779 
780 	msg = ptr + offset;
781 
782 	msg_type = msg[1];
783 	handle = msg[2];
784 
785 	if (handle == 0) {
786 		DRM_ERROR("Invalid UVD handle!\n");
787 		return -EINVAL;
788 	}
789 
790 	switch (msg_type) {
791 	case 0:
792 		/* it's a create msg, calc image size (width * height) */
793 		amdgpu_bo_kunmap(bo);
794 
795 		/* try to alloc a new handle */
796 		for (i = 0; i < adev->uvd.max_handles; ++i) {
797 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
798 				DRM_ERROR(")Handle 0x%x already in use!\n",
799 					  handle);
800 				return -EINVAL;
801 			}
802 
803 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
804 				adev->uvd.filp[i] = ctx->parser->filp;
805 				return 0;
806 			}
807 		}
808 
809 		DRM_ERROR("No more free UVD handles!\n");
810 		return -ENOSPC;
811 
812 	case 1:
813 		/* it's a decode msg, calc buffer sizes */
814 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
815 		amdgpu_bo_kunmap(bo);
816 		if (r)
817 			return r;
818 
819 		/* validate the handle */
820 		for (i = 0; i < adev->uvd.max_handles; ++i) {
821 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
822 				if (adev->uvd.filp[i] != ctx->parser->filp) {
823 					DRM_ERROR("UVD handle collision detected!\n");
824 					return -EINVAL;
825 				}
826 				return 0;
827 			}
828 		}
829 
830 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
831 		return -ENOENT;
832 
833 	case 2:
834 		/* it's a destroy msg, free the handle */
835 		for (i = 0; i < adev->uvd.max_handles; ++i)
836 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
837 		amdgpu_bo_kunmap(bo);
838 		return 0;
839 
840 	default:
841 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
842 	}
843 
844 	return -EINVAL;
845 }
846 
847 /**
848  * amdgpu_uvd_cs_pass2 - second parsing round
849  *
850  * @ctx: UVD parser context
851  *
852  * Patch buffer addresses, make sure buffer sizes are correct.
853  */
854 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
855 {
856 	struct amdgpu_bo_va_mapping *mapping;
857 	struct amdgpu_bo *bo;
858 	uint32_t cmd;
859 	uint64_t start, end;
860 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
861 	int r;
862 
863 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
864 	if (r) {
865 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
866 		return r;
867 	}
868 
869 	start = amdgpu_bo_gpu_offset(bo);
870 
871 	end = (mapping->last + 1 - mapping->start);
872 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
873 
874 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
875 	start += addr;
876 
877 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
878 			    lower_32_bits(start));
879 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
880 			    upper_32_bits(start));
881 
882 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
883 	if (cmd < 0x4) {
884 		if ((end - start) < ctx->buf_sizes[cmd]) {
885 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
886 				  (unsigned)(end - start),
887 				  ctx->buf_sizes[cmd]);
888 			return -EINVAL;
889 		}
890 
891 	} else if (cmd == 0x206) {
892 		if ((end - start) < ctx->buf_sizes[4]) {
893 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
894 					  (unsigned)(end - start),
895 					  ctx->buf_sizes[4]);
896 			return -EINVAL;
897 		}
898 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
899 		DRM_ERROR("invalid UVD command %X!\n", cmd);
900 		return -EINVAL;
901 	}
902 
903 	if (!ctx->parser->adev->uvd.address_64_bit) {
904 		if ((start >> 28) != ((end - 1) >> 28)) {
905 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
906 				  start, end);
907 			return -EINVAL;
908 		}
909 
910 		if ((cmd == 0 || cmd == 0x3) &&
911 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
912 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
913 				  start, end);
914 			return -EINVAL;
915 		}
916 	}
917 
918 	if (cmd == 0) {
919 		ctx->has_msg_cmd = true;
920 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
921 		if (r)
922 			return r;
923 	} else if (!ctx->has_msg_cmd) {
924 		DRM_ERROR("Message needed before other commands are send!\n");
925 		return -EINVAL;
926 	}
927 
928 	return 0;
929 }
930 
931 /**
932  * amdgpu_uvd_cs_reg - parse register writes
933  *
934  * @ctx: UVD parser context
935  * @cb: callback function
936  *
937  * Parse the register writes, call cb on each complete command.
938  */
939 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
940 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
941 {
942 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
943 	int i, r;
944 
945 	ctx->idx++;
946 	for (i = 0; i <= ctx->count; ++i) {
947 		unsigned reg = ctx->reg + i;
948 
949 		if (ctx->idx >= ib->length_dw) {
950 			DRM_ERROR("Register command after end of CS!\n");
951 			return -EINVAL;
952 		}
953 
954 		switch (reg) {
955 		case mmUVD_GPCOM_VCPU_DATA0:
956 			ctx->data0 = ctx->idx;
957 			break;
958 		case mmUVD_GPCOM_VCPU_DATA1:
959 			ctx->data1 = ctx->idx;
960 			break;
961 		case mmUVD_GPCOM_VCPU_CMD:
962 			r = cb(ctx);
963 			if (r)
964 				return r;
965 			break;
966 		case mmUVD_ENGINE_CNTL:
967 		case mmUVD_NO_OP:
968 			break;
969 		default:
970 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
971 			return -EINVAL;
972 		}
973 		ctx->idx++;
974 	}
975 	return 0;
976 }
977 
978 /**
979  * amdgpu_uvd_cs_packets - parse UVD packets
980  *
981  * @ctx: UVD parser context
982  * @cb: callback function
983  *
984  * Parse the command stream packets.
985  */
986 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
987 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
988 {
989 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
990 	int r;
991 
992 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
993 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
994 		unsigned type = CP_PACKET_GET_TYPE(cmd);
995 		switch (type) {
996 		case PACKET_TYPE0:
997 			ctx->reg = CP_PACKET0_GET_REG(cmd);
998 			ctx->count = CP_PACKET_GET_COUNT(cmd);
999 			r = amdgpu_uvd_cs_reg(ctx, cb);
1000 			if (r)
1001 				return r;
1002 			break;
1003 		case PACKET_TYPE2:
1004 			++ctx->idx;
1005 			break;
1006 		default:
1007 			DRM_ERROR("Unknown packet type %d !\n", type);
1008 			return -EINVAL;
1009 		}
1010 	}
1011 	return 0;
1012 }
1013 
1014 /**
1015  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1016  *
1017  * @parser: Command submission parser context
1018  * @ib_idx: Which indirect buffer to use
1019  *
1020  * Parse the command stream, patch in addresses as necessary.
1021  */
1022 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
1023 {
1024 	struct amdgpu_uvd_cs_ctx ctx = {};
1025 	unsigned buf_sizes[] = {
1026 		[0x00000000]	=	2048,
1027 		[0x00000001]	=	0xFFFFFFFF,
1028 		[0x00000002]	=	0xFFFFFFFF,
1029 		[0x00000003]	=	2048,
1030 		[0x00000004]	=	0xFFFFFFFF,
1031 	};
1032 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
1033 	int r;
1034 
1035 	parser->job->vm = NULL;
1036 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
1037 
1038 	if (ib->length_dw % 16) {
1039 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1040 			  ib->length_dw);
1041 		return -EINVAL;
1042 	}
1043 
1044 	ctx.parser = parser;
1045 	ctx.buf_sizes = buf_sizes;
1046 	ctx.ib_idx = ib_idx;
1047 
1048 	/* first round only required on chips without UVD 64 bit address support */
1049 	if (!parser->adev->uvd.address_64_bit) {
1050 		/* first round, make sure the buffers are actually in the UVD segment */
1051 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1052 		if (r)
1053 			return r;
1054 	}
1055 
1056 	/* second round, patch buffer addresses into the command stream */
1057 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1058 	if (r)
1059 		return r;
1060 
1061 	if (!ctx.has_msg_cmd) {
1062 		DRM_ERROR("UVD-IBs need a msg command!\n");
1063 		return -EINVAL;
1064 	}
1065 
1066 	return 0;
1067 }
1068 
1069 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1070 			       bool direct, struct dma_fence **fence)
1071 {
1072 	struct amdgpu_device *adev = ring->adev;
1073 	struct dma_fence *f = NULL;
1074 	struct amdgpu_job *job;
1075 	struct amdgpu_ib *ib;
1076 	uint32_t data[4];
1077 	uint64_t addr;
1078 	long r;
1079 	int i;
1080 	unsigned offset_idx = 0;
1081 	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1082 
1083 	amdgpu_bo_kunmap(bo);
1084 	amdgpu_bo_unpin(bo);
1085 
1086 	if (!ring->adev->uvd.address_64_bit) {
1087 		struct ttm_operation_ctx ctx = { true, false };
1088 
1089 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1090 		amdgpu_uvd_force_into_uvd_segment(bo);
1091 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1092 		if (r)
1093 			goto err;
1094 	}
1095 
1096 	r = amdgpu_job_alloc_with_ib(adev, 64, direct ? AMDGPU_IB_POOL_DIRECT :
1097 				     AMDGPU_IB_POOL_DELAYED, &job);
1098 	if (r)
1099 		goto err;
1100 
1101 	if (adev->asic_type >= CHIP_VEGA10) {
1102 		offset_idx = 1 + ring->me;
1103 		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1104 		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1105 	}
1106 
1107 	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1108 	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1109 	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1110 	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1111 
1112 	ib = &job->ibs[0];
1113 	addr = amdgpu_bo_gpu_offset(bo);
1114 	ib->ptr[0] = data[0];
1115 	ib->ptr[1] = addr;
1116 	ib->ptr[2] = data[1];
1117 	ib->ptr[3] = addr >> 32;
1118 	ib->ptr[4] = data[2];
1119 	ib->ptr[5] = 0;
1120 	for (i = 6; i < 16; i += 2) {
1121 		ib->ptr[i] = data[3];
1122 		ib->ptr[i+1] = 0;
1123 	}
1124 	ib->length_dw = 16;
1125 
1126 	if (direct) {
1127 		r = dma_resv_wait_timeout(bo->tbo.base.resv, true, false,
1128 					  msecs_to_jiffies(10));
1129 		if (r == 0)
1130 			r = -ETIMEDOUT;
1131 		if (r < 0)
1132 			goto err_free;
1133 
1134 		r = amdgpu_job_submit_direct(job, ring, &f);
1135 		if (r)
1136 			goto err_free;
1137 	} else {
1138 		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1139 				     AMDGPU_SYNC_ALWAYS,
1140 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1141 		if (r)
1142 			goto err_free;
1143 
1144 		r = amdgpu_job_submit(job, &adev->uvd.entity,
1145 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1146 		if (r)
1147 			goto err_free;
1148 	}
1149 
1150 	amdgpu_bo_fence(bo, f, false);
1151 	amdgpu_bo_unreserve(bo);
1152 	amdgpu_bo_unref(&bo);
1153 
1154 	if (fence)
1155 		*fence = dma_fence_get(f);
1156 	dma_fence_put(f);
1157 
1158 	return 0;
1159 
1160 err_free:
1161 	amdgpu_job_free(job);
1162 
1163 err:
1164 	amdgpu_bo_unreserve(bo);
1165 	amdgpu_bo_unref(&bo);
1166 	return r;
1167 }
1168 
1169 /* multiple fence commands without any stream commands in between can
1170    crash the vcpu so just try to emmit a dummy create/destroy msg to
1171    avoid this */
1172 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1173 			      struct dma_fence **fence)
1174 {
1175 	struct amdgpu_device *adev = ring->adev;
1176 	struct amdgpu_bo *bo = NULL;
1177 	uint32_t *msg;
1178 	int r, i;
1179 
1180 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1181 				      AMDGPU_GEM_DOMAIN_GTT,
1182 				      &bo, NULL, (void **)&msg);
1183 	if (r)
1184 		return r;
1185 
1186 	/* stitch together an UVD create msg */
1187 	msg[0] = cpu_to_le32(0x00000de4);
1188 	msg[1] = cpu_to_le32(0x00000000);
1189 	msg[2] = cpu_to_le32(handle);
1190 	msg[3] = cpu_to_le32(0x00000000);
1191 	msg[4] = cpu_to_le32(0x00000000);
1192 	msg[5] = cpu_to_le32(0x00000000);
1193 	msg[6] = cpu_to_le32(0x00000000);
1194 	msg[7] = cpu_to_le32(0x00000780);
1195 	msg[8] = cpu_to_le32(0x00000440);
1196 	msg[9] = cpu_to_le32(0x00000000);
1197 	msg[10] = cpu_to_le32(0x01b37000);
1198 	for (i = 11; i < 1024; ++i)
1199 		msg[i] = cpu_to_le32(0x0);
1200 
1201 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1202 }
1203 
1204 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1205 			       bool direct, struct dma_fence **fence)
1206 {
1207 	struct amdgpu_device *adev = ring->adev;
1208 	struct amdgpu_bo *bo = NULL;
1209 	uint32_t *msg;
1210 	int r, i;
1211 
1212 	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1213 				      AMDGPU_GEM_DOMAIN_GTT,
1214 				      &bo, NULL, (void **)&msg);
1215 	if (r)
1216 		return r;
1217 
1218 	/* stitch together an UVD destroy msg */
1219 	msg[0] = cpu_to_le32(0x00000de4);
1220 	msg[1] = cpu_to_le32(0x00000002);
1221 	msg[2] = cpu_to_le32(handle);
1222 	msg[3] = cpu_to_le32(0x00000000);
1223 	for (i = 4; i < 1024; ++i)
1224 		msg[i] = cpu_to_le32(0x0);
1225 
1226 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1227 }
1228 
1229 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1230 {
1231 	struct amdgpu_device *adev =
1232 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1233 	unsigned fences = 0, i, j;
1234 
1235 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1236 		if (adev->uvd.harvest_config & (1 << i))
1237 			continue;
1238 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1239 		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1240 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1241 		}
1242 	}
1243 
1244 	if (fences == 0) {
1245 		if (adev->pm.dpm_enabled) {
1246 			amdgpu_dpm_enable_uvd(adev, false);
1247 		} else {
1248 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1249 			/* shutdown the UVD block */
1250 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1251 							       AMD_PG_STATE_GATE);
1252 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1253 							       AMD_CG_STATE_GATE);
1254 		}
1255 	} else {
1256 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1257 	}
1258 }
1259 
1260 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1261 {
1262 	struct amdgpu_device *adev = ring->adev;
1263 	bool set_clocks;
1264 
1265 	if (amdgpu_sriov_vf(adev))
1266 		return;
1267 
1268 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1269 	if (set_clocks) {
1270 		if (adev->pm.dpm_enabled) {
1271 			amdgpu_dpm_enable_uvd(adev, true);
1272 		} else {
1273 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1274 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1275 							       AMD_CG_STATE_UNGATE);
1276 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1277 							       AMD_PG_STATE_UNGATE);
1278 		}
1279 	}
1280 }
1281 
1282 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1283 {
1284 	if (!amdgpu_sriov_vf(ring->adev))
1285 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1286 }
1287 
1288 /**
1289  * amdgpu_uvd_ring_test_ib - test ib execution
1290  *
1291  * @ring: amdgpu_ring pointer
1292  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1293  *
1294  * Test if we can successfully execute an IB
1295  */
1296 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1297 {
1298 	struct dma_fence *fence;
1299 	long r;
1300 
1301 	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1302 	if (r)
1303 		goto error;
1304 
1305 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1306 	if (r)
1307 		goto error;
1308 
1309 	r = dma_fence_wait_timeout(fence, false, timeout);
1310 	if (r == 0)
1311 		r = -ETIMEDOUT;
1312 	else if (r > 0)
1313 		r = 0;
1314 
1315 	dma_fence_put(fence);
1316 
1317 error:
1318 	return r;
1319 }
1320 
1321 /**
1322  * amdgpu_uvd_used_handles - returns used UVD handles
1323  *
1324  * @adev: amdgpu_device pointer
1325  *
1326  * Returns the number of UVD handles in use
1327  */
1328 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1329 {
1330 	unsigned i;
1331 	uint32_t used_handles = 0;
1332 
1333 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1334 		/*
1335 		 * Handles can be freed in any order, and not
1336 		 * necessarily linear. So we need to count
1337 		 * all non-zero handles.
1338 		 */
1339 		if (atomic_read(&adev->uvd.handles[i]))
1340 			used_handles++;
1341 	}
1342 
1343 	return used_handles;
1344 }
1345