xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision bbde9fc1824aab58bc78c084163007dd6c03fe5b)
1 /*
2  * Copyright 2011 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <deathsimple@vodafone.de>
29  */
30 
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 #include <drm/drmP.h>
34 #include <drm/drm.h>
35 
36 #include "amdgpu.h"
37 #include "amdgpu_pm.h"
38 #include "amdgpu_uvd.h"
39 #include "cikd.h"
40 #include "uvd/uvd_4_2_d.h"
41 
42 /* 1 second timeout */
43 #define UVD_IDLE_TIMEOUT_MS	1000
44 
45 /* Firmware Names */
46 #ifdef CONFIG_DRM_AMDGPU_CIK
47 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
48 #define FIRMWARE_KABINI 	"radeon/kabini_uvd.bin"
49 #define FIRMWARE_KAVERI 	"radeon/kaveri_uvd.bin"
50 #define FIRMWARE_HAWAII 	"radeon/hawaii_uvd.bin"
51 #define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
52 #endif
53 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
54 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
55 
56 /**
57  * amdgpu_uvd_cs_ctx - Command submission parser context
58  *
59  * Used for emulating virtual memory support on UVD 4.2.
60  */
61 struct amdgpu_uvd_cs_ctx {
62 	struct amdgpu_cs_parser *parser;
63 	unsigned reg, count;
64 	unsigned data0, data1;
65 	unsigned idx;
66 	unsigned ib_idx;
67 
68 	/* does the IB has a msg command */
69 	bool has_msg_cmd;
70 
71 	/* minimum buffer sizes */
72 	unsigned *buf_sizes;
73 };
74 
75 #ifdef CONFIG_DRM_AMDGPU_CIK
76 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
77 MODULE_FIRMWARE(FIRMWARE_KABINI);
78 MODULE_FIRMWARE(FIRMWARE_KAVERI);
79 MODULE_FIRMWARE(FIRMWARE_HAWAII);
80 MODULE_FIRMWARE(FIRMWARE_MULLINS);
81 #endif
82 MODULE_FIRMWARE(FIRMWARE_TONGA);
83 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
84 
85 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
86 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
87 
88 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
89 {
90 	unsigned long bo_size;
91 	const char *fw_name;
92 	const struct common_firmware_header *hdr;
93 	unsigned version_major, version_minor, family_id;
94 	int i, r;
95 
96 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
97 
98 	switch (adev->asic_type) {
99 #ifdef CONFIG_DRM_AMDGPU_CIK
100 	case CHIP_BONAIRE:
101 		fw_name = FIRMWARE_BONAIRE;
102 		break;
103 	case CHIP_KABINI:
104 		fw_name = FIRMWARE_KABINI;
105 		break;
106 	case CHIP_KAVERI:
107 		fw_name = FIRMWARE_KAVERI;
108 		break;
109 	case CHIP_HAWAII:
110 		fw_name = FIRMWARE_HAWAII;
111 		break;
112 	case CHIP_MULLINS:
113 		fw_name = FIRMWARE_MULLINS;
114 		break;
115 #endif
116 	case CHIP_TONGA:
117 		fw_name = FIRMWARE_TONGA;
118 		break;
119 	case CHIP_CARRIZO:
120 		fw_name = FIRMWARE_CARRIZO;
121 		break;
122 	default:
123 		return -EINVAL;
124 	}
125 
126 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
127 	if (r) {
128 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
129 			fw_name);
130 		return r;
131 	}
132 
133 	r = amdgpu_ucode_validate(adev->uvd.fw);
134 	if (r) {
135 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
136 			fw_name);
137 		release_firmware(adev->uvd.fw);
138 		adev->uvd.fw = NULL;
139 		return r;
140 	}
141 
142 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
143 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
144 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
145 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
146 	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
147 		version_major, version_minor, family_id);
148 
149 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
150 		 +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
151 	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
152 			     AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->uvd.vcpu_bo);
153 	if (r) {
154 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
155 		return r;
156 	}
157 
158 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
159 	if (r) {
160 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
161 		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
162 		return r;
163 	}
164 
165 	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
166 			  &adev->uvd.gpu_addr);
167 	if (r) {
168 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
169 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
170 		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
171 		return r;
172 	}
173 
174 	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
175 	if (r) {
176 		dev_err(adev->dev, "(%d) UVD map failed\n", r);
177 		return r;
178 	}
179 
180 	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
181 
182 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
183 		atomic_set(&adev->uvd.handles[i], 0);
184 		adev->uvd.filp[i] = NULL;
185 	}
186 
187 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
188 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
189 		adev->uvd.address_64_bit = true;
190 
191 	return 0;
192 }
193 
194 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
195 {
196 	int r;
197 
198 	if (adev->uvd.vcpu_bo == NULL)
199 		return 0;
200 
201 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
202 	if (!r) {
203 		amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
204 		amdgpu_bo_unpin(adev->uvd.vcpu_bo);
205 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
206 	}
207 
208 	amdgpu_bo_unref(&adev->uvd.vcpu_bo);
209 
210 	amdgpu_ring_fini(&adev->uvd.ring);
211 
212 	release_firmware(adev->uvd.fw);
213 
214 	return 0;
215 }
216 
217 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
218 {
219 	unsigned size;
220 	void *ptr;
221 	const struct common_firmware_header *hdr;
222 	int i;
223 
224 	if (adev->uvd.vcpu_bo == NULL)
225 		return 0;
226 
227 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
228 		if (atomic_read(&adev->uvd.handles[i]))
229 			break;
230 
231 	if (i == AMDGPU_MAX_UVD_HANDLES)
232 		return 0;
233 
234 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
235 
236 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
237 	size -= le32_to_cpu(hdr->ucode_size_bytes);
238 
239 	ptr = adev->uvd.cpu_addr;
240 	ptr += le32_to_cpu(hdr->ucode_size_bytes);
241 
242 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
243 	memcpy(adev->uvd.saved_bo, ptr, size);
244 
245 	return 0;
246 }
247 
248 int amdgpu_uvd_resume(struct amdgpu_device *adev)
249 {
250 	unsigned size;
251 	void *ptr;
252 	const struct common_firmware_header *hdr;
253 	unsigned offset;
254 
255 	if (adev->uvd.vcpu_bo == NULL)
256 		return -EINVAL;
257 
258 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
259 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
260 	memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
261 		(adev->uvd.fw->size) - offset);
262 
263 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
264 	size -= le32_to_cpu(hdr->ucode_size_bytes);
265 	ptr = adev->uvd.cpu_addr;
266 	ptr += le32_to_cpu(hdr->ucode_size_bytes);
267 
268 	if (adev->uvd.saved_bo != NULL) {
269 		memcpy(ptr, adev->uvd.saved_bo, size);
270 		kfree(adev->uvd.saved_bo);
271 		adev->uvd.saved_bo = NULL;
272 	} else
273 		memset(ptr, 0, size);
274 
275 	return 0;
276 }
277 
278 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
279 {
280 	struct amdgpu_ring *ring = &adev->uvd.ring;
281 	int i, r;
282 
283 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
284 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
285 		if (handle != 0 && adev->uvd.filp[i] == filp) {
286 			struct amdgpu_fence *fence;
287 
288 			amdgpu_uvd_note_usage(adev);
289 
290 			r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
291 			if (r) {
292 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
293 				continue;
294 			}
295 
296 			amdgpu_fence_wait(fence, false);
297 			amdgpu_fence_unref(&fence);
298 
299 			adev->uvd.filp[i] = NULL;
300 			atomic_set(&adev->uvd.handles[i], 0);
301 		}
302 	}
303 }
304 
305 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
306 {
307 	int i;
308 	for (i = 0; i < rbo->placement.num_placement; ++i) {
309 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
310 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
311 	}
312 }
313 
314 /**
315  * amdgpu_uvd_cs_pass1 - first parsing round
316  *
317  * @ctx: UVD parser context
318  *
319  * Make sure UVD message and feedback buffers are in VRAM and
320  * nobody is violating an 256MB boundary.
321  */
322 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
323 {
324 	struct amdgpu_bo_va_mapping *mapping;
325 	struct amdgpu_bo *bo;
326 	uint32_t cmd, lo, hi;
327 	uint64_t addr;
328 	int r = 0;
329 
330 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
331 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
332 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
333 
334 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
335 	if (mapping == NULL) {
336 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
337 		return -EINVAL;
338 	}
339 
340 	if (!ctx->parser->adev->uvd.address_64_bit) {
341 		/* check if it's a message or feedback command */
342 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
343 		if (cmd == 0x0 || cmd == 0x3) {
344 			/* yes, force it into VRAM */
345 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
346 			amdgpu_ttm_placement_from_domain(bo, domain);
347 		}
348 		amdgpu_uvd_force_into_uvd_segment(bo);
349 
350 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
351 	}
352 
353 	return r;
354 }
355 
356 /**
357  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
358  *
359  * @msg: pointer to message structure
360  * @buf_sizes: returned buffer sizes
361  *
362  * Peek into the decode message and calculate the necessary buffer sizes.
363  */
364 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
365 {
366 	unsigned stream_type = msg[4];
367 	unsigned width = msg[6];
368 	unsigned height = msg[7];
369 	unsigned dpb_size = msg[9];
370 	unsigned pitch = msg[28];
371 	unsigned level = msg[57];
372 
373 	unsigned width_in_mb = width / 16;
374 	unsigned height_in_mb = ALIGN(height / 16, 2);
375 	unsigned fs_in_mb = width_in_mb * height_in_mb;
376 
377 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
378 
379 	image_size = width * height;
380 	image_size += image_size / 2;
381 	image_size = ALIGN(image_size, 1024);
382 
383 	switch (stream_type) {
384 	case 0: /* H264 */
385 	case 7: /* H264 Perf */
386 		switch(level) {
387 		case 30:
388 			num_dpb_buffer = 8100 / fs_in_mb;
389 			break;
390 		case 31:
391 			num_dpb_buffer = 18000 / fs_in_mb;
392 			break;
393 		case 32:
394 			num_dpb_buffer = 20480 / fs_in_mb;
395 			break;
396 		case 41:
397 			num_dpb_buffer = 32768 / fs_in_mb;
398 			break;
399 		case 42:
400 			num_dpb_buffer = 34816 / fs_in_mb;
401 			break;
402 		case 50:
403 			num_dpb_buffer = 110400 / fs_in_mb;
404 			break;
405 		case 51:
406 			num_dpb_buffer = 184320 / fs_in_mb;
407 			break;
408 		default:
409 			num_dpb_buffer = 184320 / fs_in_mb;
410 			break;
411 		}
412 		num_dpb_buffer++;
413 		if (num_dpb_buffer > 17)
414 			num_dpb_buffer = 17;
415 
416 		/* reference picture buffer */
417 		min_dpb_size = image_size * num_dpb_buffer;
418 
419 		/* macroblock context buffer */
420 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
421 
422 		/* IT surface buffer */
423 		min_dpb_size += width_in_mb * height_in_mb * 32;
424 		break;
425 
426 	case 1: /* VC1 */
427 
428 		/* reference picture buffer */
429 		min_dpb_size = image_size * 3;
430 
431 		/* CONTEXT_BUFFER */
432 		min_dpb_size += width_in_mb * height_in_mb * 128;
433 
434 		/* IT surface buffer */
435 		min_dpb_size += width_in_mb * 64;
436 
437 		/* DB surface buffer */
438 		min_dpb_size += width_in_mb * 128;
439 
440 		/* BP */
441 		tmp = max(width_in_mb, height_in_mb);
442 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
443 		break;
444 
445 	case 3: /* MPEG2 */
446 
447 		/* reference picture buffer */
448 		min_dpb_size = image_size * 3;
449 		break;
450 
451 	case 4: /* MPEG4 */
452 
453 		/* reference picture buffer */
454 		min_dpb_size = image_size * 3;
455 
456 		/* CM */
457 		min_dpb_size += width_in_mb * height_in_mb * 64;
458 
459 		/* IT surface buffer */
460 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
461 		break;
462 
463 	case 16: /* H265 */
464 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
465 		image_size = ALIGN(image_size, 256);
466 
467 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
468 		min_dpb_size = image_size * num_dpb_buffer;
469 		break;
470 
471 	default:
472 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
473 		return -EINVAL;
474 	}
475 
476 	if (width > pitch) {
477 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
478 		return -EINVAL;
479 	}
480 
481 	if (dpb_size < min_dpb_size) {
482 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
483 			  dpb_size, min_dpb_size);
484 		return -EINVAL;
485 	}
486 
487 	buf_sizes[0x1] = dpb_size;
488 	buf_sizes[0x2] = image_size;
489 	return 0;
490 }
491 
492 /**
493  * amdgpu_uvd_cs_msg - handle UVD message
494  *
495  * @ctx: UVD parser context
496  * @bo: buffer object containing the message
497  * @offset: offset into the buffer object
498  *
499  * Peek into the UVD message and extract the session id.
500  * Make sure that we don't open up to many sessions.
501  */
502 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
503 			     struct amdgpu_bo *bo, unsigned offset)
504 {
505 	struct amdgpu_device *adev = ctx->parser->adev;
506 	int32_t *msg, msg_type, handle;
507 	struct fence *f;
508 	void *ptr;
509 
510 	int i, r;
511 
512 	if (offset & 0x3F) {
513 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
514 		return -EINVAL;
515 	}
516 
517 	f = reservation_object_get_excl(bo->tbo.resv);
518 	if (f) {
519 		r = amdgpu_fence_wait((struct amdgpu_fence *)f, false);
520 		if (r) {
521 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
522 			return r;
523 		}
524 	}
525 
526 	r = amdgpu_bo_kmap(bo, &ptr);
527 	if (r) {
528 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
529 		return r;
530 	}
531 
532 	msg = ptr + offset;
533 
534 	msg_type = msg[1];
535 	handle = msg[2];
536 
537 	if (handle == 0) {
538 		DRM_ERROR("Invalid UVD handle!\n");
539 		return -EINVAL;
540 	}
541 
542 	if (msg_type == 1) {
543 		/* it's a decode msg, calc buffer sizes */
544 		r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
545 		amdgpu_bo_kunmap(bo);
546 		if (r)
547 			return r;
548 
549 	} else if (msg_type == 2) {
550 		/* it's a destroy msg, free the handle */
551 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
552 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
553 		amdgpu_bo_kunmap(bo);
554 		return 0;
555 	} else {
556 		/* it's a create msg */
557 		amdgpu_bo_kunmap(bo);
558 
559 		if (msg_type != 0) {
560 			DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
561 			return -EINVAL;
562 		}
563 
564 		/* it's a create msg, no special handling needed */
565 	}
566 
567 	/* create or decode, validate the handle */
568 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
569 		if (atomic_read(&adev->uvd.handles[i]) == handle)
570 			return 0;
571 	}
572 
573 	/* handle not found try to alloc a new one */
574 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
575 		if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
576 			adev->uvd.filp[i] = ctx->parser->filp;
577 			return 0;
578 		}
579 	}
580 
581 	DRM_ERROR("No more free UVD handles!\n");
582 	return -EINVAL;
583 }
584 
585 /**
586  * amdgpu_uvd_cs_pass2 - second parsing round
587  *
588  * @ctx: UVD parser context
589  *
590  * Patch buffer addresses, make sure buffer sizes are correct.
591  */
592 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
593 {
594 	struct amdgpu_bo_va_mapping *mapping;
595 	struct amdgpu_bo *bo;
596 	struct amdgpu_ib *ib;
597 	uint32_t cmd, lo, hi;
598 	uint64_t start, end;
599 	uint64_t addr;
600 	int r;
601 
602 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
603 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
604 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
605 
606 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
607 	if (mapping == NULL)
608 		return -EINVAL;
609 
610 	start = amdgpu_bo_gpu_offset(bo);
611 
612 	end = (mapping->it.last + 1 - mapping->it.start);
613 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
614 
615 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
616 	start += addr;
617 
618 	ib = &ctx->parser->ibs[ctx->ib_idx];
619 	ib->ptr[ctx->data0] = start & 0xFFFFFFFF;
620 	ib->ptr[ctx->data1] = start >> 32;
621 
622 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
623 	if (cmd < 0x4) {
624 		if ((end - start) < ctx->buf_sizes[cmd]) {
625 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
626 				  (unsigned)(end - start),
627 				  ctx->buf_sizes[cmd]);
628 			return -EINVAL;
629 		}
630 
631 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
632 		DRM_ERROR("invalid UVD command %X!\n", cmd);
633 		return -EINVAL;
634 	}
635 
636 	if (!ctx->parser->adev->uvd.address_64_bit) {
637 		if ((start >> 28) != ((end - 1) >> 28)) {
638 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
639 				  start, end);
640 			return -EINVAL;
641 		}
642 
643 		if ((cmd == 0 || cmd == 0x3) &&
644 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
645 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
646 				  start, end);
647 			return -EINVAL;
648 		}
649 	}
650 
651 	if (cmd == 0) {
652 		ctx->has_msg_cmd = true;
653 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
654 		if (r)
655 			return r;
656 	} else if (!ctx->has_msg_cmd) {
657 		DRM_ERROR("Message needed before other commands are send!\n");
658 		return -EINVAL;
659 	}
660 
661 	return 0;
662 }
663 
664 /**
665  * amdgpu_uvd_cs_reg - parse register writes
666  *
667  * @ctx: UVD parser context
668  * @cb: callback function
669  *
670  * Parse the register writes, call cb on each complete command.
671  */
672 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
673 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
674 {
675 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
676 	int i, r;
677 
678 	ctx->idx++;
679 	for (i = 0; i <= ctx->count; ++i) {
680 		unsigned reg = ctx->reg + i;
681 
682 		if (ctx->idx >= ib->length_dw) {
683 			DRM_ERROR("Register command after end of CS!\n");
684 			return -EINVAL;
685 		}
686 
687 		switch (reg) {
688 		case mmUVD_GPCOM_VCPU_DATA0:
689 			ctx->data0 = ctx->idx;
690 			break;
691 		case mmUVD_GPCOM_VCPU_DATA1:
692 			ctx->data1 = ctx->idx;
693 			break;
694 		case mmUVD_GPCOM_VCPU_CMD:
695 			r = cb(ctx);
696 			if (r)
697 				return r;
698 			break;
699 		case mmUVD_ENGINE_CNTL:
700 			break;
701 		default:
702 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
703 			return -EINVAL;
704 		}
705 		ctx->idx++;
706 	}
707 	return 0;
708 }
709 
710 /**
711  * amdgpu_uvd_cs_packets - parse UVD packets
712  *
713  * @ctx: UVD parser context
714  * @cb: callback function
715  *
716  * Parse the command stream packets.
717  */
718 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
719 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
720 {
721 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
722 	int r;
723 
724 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
725 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
726 		unsigned type = CP_PACKET_GET_TYPE(cmd);
727 		switch (type) {
728 		case PACKET_TYPE0:
729 			ctx->reg = CP_PACKET0_GET_REG(cmd);
730 			ctx->count = CP_PACKET_GET_COUNT(cmd);
731 			r = amdgpu_uvd_cs_reg(ctx, cb);
732 			if (r)
733 				return r;
734 			break;
735 		case PACKET_TYPE2:
736 			++ctx->idx;
737 			break;
738 		default:
739 			DRM_ERROR("Unknown packet type %d !\n", type);
740 			return -EINVAL;
741 		}
742 	}
743 	return 0;
744 }
745 
746 /**
747  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
748  *
749  * @parser: Command submission parser context
750  *
751  * Parse the command stream, patch in addresses as necessary.
752  */
753 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
754 {
755 	struct amdgpu_uvd_cs_ctx ctx = {};
756 	unsigned buf_sizes[] = {
757 		[0x00000000]	=	2048,
758 		[0x00000001]	=	32 * 1024 * 1024,
759 		[0x00000002]	=	2048 * 1152 * 3,
760 		[0x00000003]	=	2048,
761 	};
762 	struct amdgpu_ib *ib = &parser->ibs[ib_idx];
763 	int r;
764 
765 	if (ib->length_dw % 16) {
766 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
767 			  ib->length_dw);
768 		return -EINVAL;
769 	}
770 
771 	ctx.parser = parser;
772 	ctx.buf_sizes = buf_sizes;
773 	ctx.ib_idx = ib_idx;
774 
775 	/* first round, make sure the buffers are actually in the UVD segment */
776 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
777 	if (r)
778 		return r;
779 
780 	/* second round, patch buffer addresses into the command stream */
781 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
782 	if (r)
783 		return r;
784 
785 	if (!ctx.has_msg_cmd) {
786 		DRM_ERROR("UVD-IBs need a msg command!\n");
787 		return -EINVAL;
788 	}
789 
790 	amdgpu_uvd_note_usage(ctx.parser->adev);
791 
792 	return 0;
793 }
794 
795 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring,
796 			       struct amdgpu_bo *bo,
797 			       struct amdgpu_fence **fence)
798 {
799 	struct ttm_validate_buffer tv;
800 	struct ww_acquire_ctx ticket;
801 	struct list_head head;
802 	struct amdgpu_ib ib;
803 	uint64_t addr;
804 	int i, r;
805 
806 	memset(&tv, 0, sizeof(tv));
807 	tv.bo = &bo->tbo;
808 
809 	INIT_LIST_HEAD(&head);
810 	list_add(&tv.head, &head);
811 
812 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
813 	if (r)
814 		return r;
815 
816 	if (!bo->adev->uvd.address_64_bit) {
817 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
818 		amdgpu_uvd_force_into_uvd_segment(bo);
819 	}
820 
821 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
822 	if (r)
823 		goto err;
824 
825 	r = amdgpu_ib_get(ring, NULL, 64, &ib);
826 	if (r)
827 		goto err;
828 
829 	addr = amdgpu_bo_gpu_offset(bo);
830 	ib.ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
831 	ib.ptr[1] = addr;
832 	ib.ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
833 	ib.ptr[3] = addr >> 32;
834 	ib.ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
835 	ib.ptr[5] = 0;
836 	for (i = 6; i < 16; ++i)
837 		ib.ptr[i] = PACKET2(0);
838 	ib.length_dw = 16;
839 
840 	r = amdgpu_ib_schedule(ring->adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
841 	if (r)
842 		goto err;
843 	ttm_eu_fence_buffer_objects(&ticket, &head, &ib.fence->base);
844 
845 	if (fence)
846 		*fence = amdgpu_fence_ref(ib.fence);
847 
848 	amdgpu_ib_free(ring->adev, &ib);
849 	amdgpu_bo_unref(&bo);
850 	return 0;
851 
852 err:
853 	ttm_eu_backoff_reservation(&ticket, &head);
854 	return r;
855 }
856 
857 /* multiple fence commands without any stream commands in between can
858    crash the vcpu so just try to emmit a dummy create/destroy msg to
859    avoid this */
860 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
861 			      struct amdgpu_fence **fence)
862 {
863 	struct amdgpu_device *adev = ring->adev;
864 	struct amdgpu_bo *bo;
865 	uint32_t *msg;
866 	int r, i;
867 
868 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
869 			     AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo);
870 	if (r)
871 		return r;
872 
873 	r = amdgpu_bo_reserve(bo, false);
874 	if (r) {
875 		amdgpu_bo_unref(&bo);
876 		return r;
877 	}
878 
879 	r = amdgpu_bo_kmap(bo, (void **)&msg);
880 	if (r) {
881 		amdgpu_bo_unreserve(bo);
882 		amdgpu_bo_unref(&bo);
883 		return r;
884 	}
885 
886 	/* stitch together an UVD create msg */
887 	msg[0] = cpu_to_le32(0x00000de4);
888 	msg[1] = cpu_to_le32(0x00000000);
889 	msg[2] = cpu_to_le32(handle);
890 	msg[3] = cpu_to_le32(0x00000000);
891 	msg[4] = cpu_to_le32(0x00000000);
892 	msg[5] = cpu_to_le32(0x00000000);
893 	msg[6] = cpu_to_le32(0x00000000);
894 	msg[7] = cpu_to_le32(0x00000780);
895 	msg[8] = cpu_to_le32(0x00000440);
896 	msg[9] = cpu_to_le32(0x00000000);
897 	msg[10] = cpu_to_le32(0x01b37000);
898 	for (i = 11; i < 1024; ++i)
899 		msg[i] = cpu_to_le32(0x0);
900 
901 	amdgpu_bo_kunmap(bo);
902 	amdgpu_bo_unreserve(bo);
903 
904 	return amdgpu_uvd_send_msg(ring, bo, fence);
905 }
906 
907 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
908 			       struct amdgpu_fence **fence)
909 {
910 	struct amdgpu_device *adev = ring->adev;
911 	struct amdgpu_bo *bo;
912 	uint32_t *msg;
913 	int r, i;
914 
915 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
916 			     AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo);
917 	if (r)
918 		return r;
919 
920 	r = amdgpu_bo_reserve(bo, false);
921 	if (r) {
922 		amdgpu_bo_unref(&bo);
923 		return r;
924 	}
925 
926 	r = amdgpu_bo_kmap(bo, (void **)&msg);
927 	if (r) {
928 		amdgpu_bo_unreserve(bo);
929 		amdgpu_bo_unref(&bo);
930 		return r;
931 	}
932 
933 	/* stitch together an UVD destroy msg */
934 	msg[0] = cpu_to_le32(0x00000de4);
935 	msg[1] = cpu_to_le32(0x00000002);
936 	msg[2] = cpu_to_le32(handle);
937 	msg[3] = cpu_to_le32(0x00000000);
938 	for (i = 4; i < 1024; ++i)
939 		msg[i] = cpu_to_le32(0x0);
940 
941 	amdgpu_bo_kunmap(bo);
942 	amdgpu_bo_unreserve(bo);
943 
944 	return amdgpu_uvd_send_msg(ring, bo, fence);
945 }
946 
947 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
948 {
949 	struct amdgpu_device *adev =
950 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
951 	unsigned i, fences, handles = 0;
952 
953 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
954 
955 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
956 		if (atomic_read(&adev->uvd.handles[i]))
957 			++handles;
958 
959 	if (fences == 0 && handles == 0) {
960 		if (adev->pm.dpm_enabled) {
961 			amdgpu_dpm_enable_uvd(adev, false);
962 		} else {
963 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
964 		}
965 	} else {
966 		schedule_delayed_work(&adev->uvd.idle_work,
967 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
968 	}
969 }
970 
971 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
972 {
973 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
974 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
975 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
976 
977 	if (set_clocks) {
978 		if (adev->pm.dpm_enabled) {
979 			amdgpu_dpm_enable_uvd(adev, true);
980 		} else {
981 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
982 		}
983 	}
984 }
985