xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_uvd.c (revision f3cfa6f6ce31685c6c4a758bc430e69eb99f50a4)
1 /*	$NetBSD: amdgpu_uvd.c,v 1.4 2019/01/01 08:07:47 maya Exp $	*/
2 
3 /*
4  * Copyright 2011 Advanced Micro Devices, Inc.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * The above copyright notice and this permission notice (including the
24  * next paragraph) shall be included in all copies or substantial portions
25  * of the Software.
26  *
27  */
28 /*
29  * Authors:
30  *    Christian König <deathsimple@vodafone.de>
31  */
32 
33 #include <sys/cdefs.h>
34 __KERNEL_RCSID(0, "$NetBSD: amdgpu_uvd.c,v 1.4 2019/01/01 08:07:47 maya Exp $");
35 
36 #include <linux/firmware.h>
37 #include <linux/module.h>
38 #include <asm/byteorder.h>
39 #include <drm/drmP.h>
40 #include <drm/drm.h>
41 
42 #include "amdgpu.h"
43 #include "amdgpu_pm.h"
44 #include "amdgpu_uvd.h"
45 #include "cikd.h"
46 #include "uvd/uvd_4_2_d.h"
47 
48 /* 1 second timeout */
49 #define UVD_IDLE_TIMEOUT_MS	1000
50 
51 /* Firmware Names */
52 #ifdef CONFIG_DRM_AMDGPU_CIK
53 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
54 #define FIRMWARE_KABINI 	"radeon/kabini_uvd.bin"
55 #define FIRMWARE_KAVERI 	"radeon/kaveri_uvd.bin"
56 #define FIRMWARE_HAWAII 	"radeon/hawaii_uvd.bin"
57 #define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
58 #endif
59 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
60 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
61 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
62 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
63 
64 /**
65  * amdgpu_uvd_cs_ctx - Command submission parser context
66  *
67  * Used for emulating virtual memory support on UVD 4.2.
68  */
69 struct amdgpu_uvd_cs_ctx {
70 	struct amdgpu_cs_parser *parser;
71 	unsigned reg, count;
72 	unsigned data0, data1;
73 	unsigned idx;
74 	unsigned ib_idx;
75 
76 	/* does the IB has a msg command */
77 	bool has_msg_cmd;
78 
79 	/* minimum buffer sizes */
80 	unsigned *buf_sizes;
81 };
82 
83 #ifdef CONFIG_DRM_AMDGPU_CIK
84 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
85 MODULE_FIRMWARE(FIRMWARE_KABINI);
86 MODULE_FIRMWARE(FIRMWARE_KAVERI);
87 MODULE_FIRMWARE(FIRMWARE_HAWAII);
88 MODULE_FIRMWARE(FIRMWARE_MULLINS);
89 #endif
90 MODULE_FIRMWARE(FIRMWARE_TONGA);
91 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
92 MODULE_FIRMWARE(FIRMWARE_FIJI);
93 MODULE_FIRMWARE(FIRMWARE_STONEY);
94 
95 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
96 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
97 
98 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
99 {
100 	unsigned long bo_size;
101 	const char *fw_name;
102 	const struct common_firmware_header *hdr;
103 	unsigned version_major, version_minor, family_id;
104 	int i, r;
105 
106 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
107 
108 	switch (adev->asic_type) {
109 #ifdef CONFIG_DRM_AMDGPU_CIK
110 	case CHIP_BONAIRE:
111 		fw_name = FIRMWARE_BONAIRE;
112 		break;
113 	case CHIP_KABINI:
114 		fw_name = FIRMWARE_KABINI;
115 		break;
116 	case CHIP_KAVERI:
117 		fw_name = FIRMWARE_KAVERI;
118 		break;
119 	case CHIP_HAWAII:
120 		fw_name = FIRMWARE_HAWAII;
121 		break;
122 	case CHIP_MULLINS:
123 		fw_name = FIRMWARE_MULLINS;
124 		break;
125 #endif
126 	case CHIP_TONGA:
127 		fw_name = FIRMWARE_TONGA;
128 		break;
129 	case CHIP_FIJI:
130 		fw_name = FIRMWARE_FIJI;
131 		break;
132 	case CHIP_CARRIZO:
133 		fw_name = FIRMWARE_CARRIZO;
134 		break;
135 	case CHIP_STONEY:
136 		fw_name = FIRMWARE_STONEY;
137 		break;
138 	default:
139 		return -EINVAL;
140 	}
141 
142 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
143 	if (r) {
144 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
145 			fw_name);
146 		return r;
147 	}
148 
149 	r = amdgpu_ucode_validate(adev->uvd.fw);
150 	if (r) {
151 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
152 			fw_name);
153 		release_firmware(adev->uvd.fw);
154 		adev->uvd.fw = NULL;
155 		return r;
156 	}
157 
158 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
159 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
160 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
161 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
162 	DRM_INFO("Found UVD firmware Version: %x.%x Family ID: %x\n",
163 		version_major, version_minor, family_id);
164 
165 	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
166 				(family_id << 8));
167 
168 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
169 		 +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
170 	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
171 			     AMDGPU_GEM_DOMAIN_VRAM,
172 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
173 			     NULL, NULL, &adev->uvd.vcpu_bo);
174 	if (r) {
175 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
176 		return r;
177 	}
178 
179 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
180 	if (r) {
181 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
182 		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
183 		return r;
184 	}
185 
186 	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
187 			  &adev->uvd.gpu_addr);
188 	if (r) {
189 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
190 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
191 		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
192 		return r;
193 	}
194 
195 	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
196 	if (r) {
197 		dev_err(adev->dev, "(%d) UVD map failed\n", r);
198 		return r;
199 	}
200 
201 	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
202 
203 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
204 		atomic_set(&adev->uvd.handles[i], 0);
205 		adev->uvd.filp[i] = NULL;
206 	}
207 
208 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
209 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
210 		adev->uvd.address_64_bit = true;
211 
212 	return 0;
213 }
214 
215 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
216 {
217 	int r;
218 
219 	if (adev->uvd.vcpu_bo == NULL)
220 		return 0;
221 
222 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
223 	if (!r) {
224 		amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
225 		amdgpu_bo_unpin(adev->uvd.vcpu_bo);
226 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
227 	}
228 
229 	amdgpu_bo_unref(&adev->uvd.vcpu_bo);
230 
231 	amdgpu_ring_fini(&adev->uvd.ring);
232 
233 	release_firmware(adev->uvd.fw);
234 
235 	return 0;
236 }
237 
238 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
239 {
240 	struct amdgpu_ring *ring = &adev->uvd.ring;
241 	int i, r;
242 
243 	if (adev->uvd.vcpu_bo == NULL)
244 		return 0;
245 
246 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
247 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
248 		if (handle != 0) {
249 			struct fence *fence;
250 
251 			amdgpu_uvd_note_usage(adev);
252 
253 			r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
254 			if (r) {
255 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
256 				continue;
257 			}
258 
259 			fence_wait(fence, false);
260 			fence_put(fence);
261 
262 			adev->uvd.filp[i] = NULL;
263 			atomic_set(&adev->uvd.handles[i], 0);
264 		}
265 	}
266 
267 	return 0;
268 }
269 
270 int amdgpu_uvd_resume(struct amdgpu_device *adev)
271 {
272 	unsigned size;
273 	void *ptr;
274 	const struct common_firmware_header *hdr;
275 	unsigned offset;
276 
277 	if (adev->uvd.vcpu_bo == NULL)
278 		return -EINVAL;
279 
280 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
281 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
282 	memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
283 		(adev->uvd.fw->size) - offset);
284 
285 	cancel_delayed_work_sync(&adev->uvd.idle_work);
286 
287 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
288 	size -= le32_to_cpu(hdr->ucode_size_bytes);
289 	ptr = adev->uvd.cpu_addr;
290 	ptr = (char *)ptr + le32_to_cpu(hdr->ucode_size_bytes);
291 
292 	memset(ptr, 0, size);
293 
294 	return 0;
295 }
296 
297 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
298 {
299 	struct amdgpu_ring *ring = &adev->uvd.ring;
300 	int i, r;
301 
302 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
303 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
304 		if (handle != 0 && adev->uvd.filp[i] == filp) {
305 			struct fence *fence;
306 
307 			amdgpu_uvd_note_usage(adev);
308 
309 			r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
310 			if (r) {
311 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
312 				continue;
313 			}
314 
315 			fence_wait(fence, false);
316 			fence_put(fence);
317 
318 			adev->uvd.filp[i] = NULL;
319 			atomic_set(&adev->uvd.handles[i], 0);
320 		}
321 	}
322 }
323 
324 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
325 {
326 	int i;
327 	for (i = 0; i < rbo->placement.num_placement; ++i) {
328 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
329 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
330 	}
331 }
332 
333 /**
334  * amdgpu_uvd_cs_pass1 - first parsing round
335  *
336  * @ctx: UVD parser context
337  *
338  * Make sure UVD message and feedback buffers are in VRAM and
339  * nobody is violating an 256MB boundary.
340  */
341 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
342 {
343 	struct amdgpu_bo_va_mapping *mapping;
344 	struct amdgpu_bo *bo;
345 	uint32_t cmd, lo, hi;
346 	uint64_t addr;
347 	int r = 0;
348 
349 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
350 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
351 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
352 
353 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
354 	if (mapping == NULL) {
355 		DRM_ERROR("Can't find BO for addr 0x%08"PRIx64"\n", addr);
356 		return -EINVAL;
357 	}
358 
359 	if (!ctx->parser->adev->uvd.address_64_bit) {
360 		/* check if it's a message or feedback command */
361 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
362 		if (cmd == 0x0 || cmd == 0x3) {
363 			/* yes, force it into VRAM */
364 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
365 			amdgpu_ttm_placement_from_domain(bo, domain);
366 		}
367 		amdgpu_uvd_force_into_uvd_segment(bo);
368 
369 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
370 	}
371 
372 	return r;
373 }
374 
375 /**
376  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
377  *
378  * @msg: pointer to message structure
379  * @buf_sizes: returned buffer sizes
380  *
381  * Peek into the decode message and calculate the necessary buffer sizes.
382  */
383 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
384 {
385 	unsigned stream_type = msg[4];
386 	unsigned width = msg[6];
387 	unsigned height = msg[7];
388 	unsigned dpb_size = msg[9];
389 	unsigned pitch = msg[28];
390 	unsigned level = msg[57];
391 
392 	unsigned width_in_mb = width / 16;
393 #ifdef __NetBSD__		/* XXX ALIGN means something else */
394 	unsigned height_in_mb = round_up(height / 16, 2);
395 #else
396 	unsigned height_in_mb = ALIGN(height / 16, 2);
397 #endif
398 	unsigned fs_in_mb = width_in_mb * height_in_mb;
399 
400 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
401 	unsigned min_ctx_size = 0;
402 
403 	image_size = width * height;
404 	image_size += image_size / 2;
405 #ifdef __NetBSD__		/* XXX ALIGN means something else */
406 	image_size = round_up(image_size, 1024);
407 #else
408 	image_size = ALIGN(image_size, 1024);
409 #endif
410 
411 	switch (stream_type) {
412 	case 0: /* H264 */
413 	case 7: /* H264 Perf */
414 		switch(level) {
415 		case 30:
416 			num_dpb_buffer = 8100 / fs_in_mb;
417 			break;
418 		case 31:
419 			num_dpb_buffer = 18000 / fs_in_mb;
420 			break;
421 		case 32:
422 			num_dpb_buffer = 20480 / fs_in_mb;
423 			break;
424 		case 41:
425 			num_dpb_buffer = 32768 / fs_in_mb;
426 			break;
427 		case 42:
428 			num_dpb_buffer = 34816 / fs_in_mb;
429 			break;
430 		case 50:
431 			num_dpb_buffer = 110400 / fs_in_mb;
432 			break;
433 		case 51:
434 			num_dpb_buffer = 184320 / fs_in_mb;
435 			break;
436 		default:
437 			num_dpb_buffer = 184320 / fs_in_mb;
438 			break;
439 		}
440 		num_dpb_buffer++;
441 		if (num_dpb_buffer > 17)
442 			num_dpb_buffer = 17;
443 
444 		/* reference picture buffer */
445 		min_dpb_size = image_size * num_dpb_buffer;
446 
447 		/* macroblock context buffer */
448 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
449 
450 		/* IT surface buffer */
451 		min_dpb_size += width_in_mb * height_in_mb * 32;
452 		break;
453 
454 	case 1: /* VC1 */
455 
456 		/* reference picture buffer */
457 		min_dpb_size = image_size * 3;
458 
459 		/* CONTEXT_BUFFER */
460 		min_dpb_size += width_in_mb * height_in_mb * 128;
461 
462 		/* IT surface buffer */
463 		min_dpb_size += width_in_mb * 64;
464 
465 		/* DB surface buffer */
466 		min_dpb_size += width_in_mb * 128;
467 
468 		/* BP */
469 		tmp = max(width_in_mb, height_in_mb);
470 #ifdef __NetBSD__		/* XXX ALIGN means something else */
471 		min_dpb_size += round_up(tmp * 7 * 16, 64);
472 #else
473 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
474 #endif
475 		break;
476 
477 	case 3: /* MPEG2 */
478 
479 		/* reference picture buffer */
480 		min_dpb_size = image_size * 3;
481 		break;
482 
483 	case 4: /* MPEG4 */
484 
485 		/* reference picture buffer */
486 		min_dpb_size = image_size * 3;
487 
488 		/* CM */
489 		min_dpb_size += width_in_mb * height_in_mb * 64;
490 
491 		/* IT surface buffer */
492 #ifdef __NetBSD__		/* XXX ALIGN means something else */
493 		min_dpb_size += round_up(width_in_mb * height_in_mb * 32, 64);
494 #else
495 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
496 #endif
497 		break;
498 
499 	case 16: /* H265 */
500 #ifdef __NetBSD__		/* XXX ALIGN means something else */
501 		image_size = (round_up(width, 16) * round_up(height, 16) * 3) / 2;
502 		image_size = round_up(image_size, 256);
503 #else
504 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
505 		image_size = ALIGN(image_size, 256);
506 #endif
507 
508 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
509 		min_dpb_size = image_size * num_dpb_buffer;
510 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
511 					   * 16 * num_dpb_buffer + 52 * 1024;
512 		break;
513 
514 	default:
515 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
516 		return -EINVAL;
517 	}
518 
519 	if (width > pitch) {
520 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
521 		return -EINVAL;
522 	}
523 
524 	if (dpb_size < min_dpb_size) {
525 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
526 			  dpb_size, min_dpb_size);
527 		return -EINVAL;
528 	}
529 
530 	buf_sizes[0x1] = dpb_size;
531 	buf_sizes[0x2] = image_size;
532 	buf_sizes[0x4] = min_ctx_size;
533 	return 0;
534 }
535 
536 /**
537  * amdgpu_uvd_cs_msg - handle UVD message
538  *
539  * @ctx: UVD parser context
540  * @bo: buffer object containing the message
541  * @offset: offset into the buffer object
542  *
543  * Peek into the UVD message and extract the session id.
544  * Make sure that we don't open up to many sessions.
545  */
546 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
547 			     struct amdgpu_bo *bo, unsigned offset)
548 {
549 	struct amdgpu_device *adev = ctx->parser->adev;
550 	int32_t *msg, msg_type, handle;
551 	void *ptr;
552 	long r;
553 	int i;
554 
555 	if (offset & 0x3F) {
556 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
557 		return -EINVAL;
558 	}
559 
560 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, true, false,
561 						MAX_SCHEDULE_TIMEOUT);
562 	if (r < 0) {
563 		DRM_ERROR("Failed waiting for UVD message (%ld)!\n", r);
564 		return r;
565 	}
566 
567 	r = amdgpu_bo_kmap(bo, &ptr);
568 	if (r) {
569 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
570 		return r;
571 	}
572 
573 	msg = (void *)((char *)ptr + offset);
574 
575 	msg_type = msg[1];
576 	handle = msg[2];
577 
578 	if (handle == 0) {
579 		DRM_ERROR("Invalid UVD handle!\n");
580 		return -EINVAL;
581 	}
582 
583 	switch (msg_type) {
584 	case 0:
585 		/* it's a create msg, calc image size (width * height) */
586 		amdgpu_bo_kunmap(bo);
587 
588 		/* try to alloc a new handle */
589 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
590 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
591 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
592 				return -EINVAL;
593 			}
594 
595 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
596 				adev->uvd.filp[i] = ctx->parser->filp;
597 				return 0;
598 			}
599 		}
600 
601 		DRM_ERROR("No more free UVD handles!\n");
602 		return -EINVAL;
603 
604 	case 1:
605 		/* it's a decode msg, calc buffer sizes */
606 		r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
607 		amdgpu_bo_kunmap(bo);
608 		if (r)
609 			return r;
610 
611 		/* validate the handle */
612 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
613 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
614 				if (adev->uvd.filp[i] != ctx->parser->filp) {
615 					DRM_ERROR("UVD handle collision detected!\n");
616 					return -EINVAL;
617 				}
618 				return 0;
619 			}
620 		}
621 
622 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
623 		return -ENOENT;
624 
625 	case 2:
626 		/* it's a destroy msg, free the handle */
627 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
628 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
629 		amdgpu_bo_kunmap(bo);
630 		return 0;
631 
632 	default:
633 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
634 		return -EINVAL;
635 	}
636 	BUG();
637 	return -EINVAL;
638 }
639 
640 /**
641  * amdgpu_uvd_cs_pass2 - second parsing round
642  *
643  * @ctx: UVD parser context
644  *
645  * Patch buffer addresses, make sure buffer sizes are correct.
646  */
647 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
648 {
649 	struct amdgpu_bo_va_mapping *mapping;
650 	struct amdgpu_bo *bo;
651 	struct amdgpu_ib *ib;
652 	uint32_t cmd, lo, hi;
653 	uint64_t start, end;
654 	uint64_t addr;
655 	int r;
656 
657 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
658 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
659 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
660 
661 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
662 	if (mapping == NULL)
663 		return -EINVAL;
664 
665 	start = amdgpu_bo_gpu_offset(bo);
666 
667 	end = (mapping->it.last + 1 - mapping->it.start);
668 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
669 
670 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
671 	start += addr;
672 
673 	ib = &ctx->parser->ibs[ctx->ib_idx];
674 	ib->ptr[ctx->data0] = start & 0xFFFFFFFF;
675 	ib->ptr[ctx->data1] = start >> 32;
676 
677 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
678 	if (cmd < 0x4) {
679 		if ((end - start) < ctx->buf_sizes[cmd]) {
680 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
681 				  (unsigned)(end - start),
682 				  ctx->buf_sizes[cmd]);
683 			return -EINVAL;
684 		}
685 
686 	} else if (cmd == 0x206) {
687 		if ((end - start) < ctx->buf_sizes[4]) {
688 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
689 					  (unsigned)(end - start),
690 					  ctx->buf_sizes[4]);
691 			return -EINVAL;
692 		}
693 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
694 		DRM_ERROR("invalid UVD command %X!\n", cmd);
695 		return -EINVAL;
696 	}
697 
698 	if (!ctx->parser->adev->uvd.address_64_bit) {
699 		if ((start >> 28) != ((end - 1) >> 28)) {
700 			DRM_ERROR("reloc %"PRIX64"-%"PRIX64" crossing 256MB boundary!\n",
701 				  start, end);
702 			return -EINVAL;
703 		}
704 
705 		if ((cmd == 0 || cmd == 0x3) &&
706 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
707 			DRM_ERROR("msg/fb buffer %"PRIX64"-%"PRIX64" out of 256MB segment!\n",
708 				  start, end);
709 			return -EINVAL;
710 		}
711 	}
712 
713 	if (cmd == 0) {
714 		ctx->has_msg_cmd = true;
715 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
716 		if (r)
717 			return r;
718 	} else if (!ctx->has_msg_cmd) {
719 		DRM_ERROR("Message needed before other commands are send!\n");
720 		return -EINVAL;
721 	}
722 
723 	return 0;
724 }
725 
726 /**
727  * amdgpu_uvd_cs_reg - parse register writes
728  *
729  * @ctx: UVD parser context
730  * @cb: callback function
731  *
732  * Parse the register writes, call cb on each complete command.
733  */
734 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
735 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
736 {
737 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
738 	int i, r;
739 
740 	ctx->idx++;
741 	for (i = 0; i <= ctx->count; ++i) {
742 		unsigned reg = ctx->reg + i;
743 
744 		if (ctx->idx >= ib->length_dw) {
745 			DRM_ERROR("Register command after end of CS!\n");
746 			return -EINVAL;
747 		}
748 
749 		switch (reg) {
750 		case mmUVD_GPCOM_VCPU_DATA0:
751 			ctx->data0 = ctx->idx;
752 			break;
753 		case mmUVD_GPCOM_VCPU_DATA1:
754 			ctx->data1 = ctx->idx;
755 			break;
756 		case mmUVD_GPCOM_VCPU_CMD:
757 			r = cb(ctx);
758 			if (r)
759 				return r;
760 			break;
761 		case mmUVD_ENGINE_CNTL:
762 			break;
763 		default:
764 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
765 			return -EINVAL;
766 		}
767 		ctx->idx++;
768 	}
769 	return 0;
770 }
771 
772 /**
773  * amdgpu_uvd_cs_packets - parse UVD packets
774  *
775  * @ctx: UVD parser context
776  * @cb: callback function
777  *
778  * Parse the command stream packets.
779  */
780 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
781 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
782 {
783 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
784 	int r;
785 
786 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
787 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
788 		unsigned type = CP_PACKET_GET_TYPE(cmd);
789 		switch (type) {
790 		case PACKET_TYPE0:
791 			ctx->reg = CP_PACKET0_GET_REG(cmd);
792 			ctx->count = CP_PACKET_GET_COUNT(cmd);
793 			r = amdgpu_uvd_cs_reg(ctx, cb);
794 			if (r)
795 				return r;
796 			break;
797 		case PACKET_TYPE2:
798 			++ctx->idx;
799 			break;
800 		default:
801 			DRM_ERROR("Unknown packet type %d !\n", type);
802 			return -EINVAL;
803 		}
804 	}
805 	return 0;
806 }
807 
808 /**
809  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
810  *
811  * @parser: Command submission parser context
812  *
813  * Parse the command stream, patch in addresses as necessary.
814  */
815 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
816 {
817 	struct amdgpu_uvd_cs_ctx ctx = {};
818 	unsigned buf_sizes[] = {
819 		[0x00000000]	=	2048,
820 		[0x00000001]	=	0xFFFFFFFF,
821 		[0x00000002]	=	0xFFFFFFFF,
822 		[0x00000003]	=	2048,
823 		[0x00000004]	=	0xFFFFFFFF,
824 	};
825 	struct amdgpu_ib *ib = &parser->ibs[ib_idx];
826 	int r;
827 
828 	if (ib->length_dw % 16) {
829 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
830 			  ib->length_dw);
831 		return -EINVAL;
832 	}
833 
834 	ctx.parser = parser;
835 	ctx.buf_sizes = buf_sizes;
836 	ctx.ib_idx = ib_idx;
837 
838 	/* first round, make sure the buffers are actually in the UVD segment */
839 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
840 	if (r)
841 		return r;
842 
843 	/* second round, patch buffer addresses into the command stream */
844 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
845 	if (r)
846 		return r;
847 
848 	if (!ctx.has_msg_cmd) {
849 		DRM_ERROR("UVD-IBs need a msg command!\n");
850 		return -EINVAL;
851 	}
852 
853 	amdgpu_uvd_note_usage(ctx.parser->adev);
854 
855 	return 0;
856 }
857 
858 static int amdgpu_uvd_free_job(
859 	struct amdgpu_job *job)
860 {
861 	amdgpu_ib_free(job->adev, job->ibs);
862 	kfree(job->ibs);
863 	return 0;
864 }
865 
866 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring,
867 			       struct amdgpu_bo *bo,
868 			       struct fence **fence)
869 {
870 	struct ttm_validate_buffer tv;
871 	struct ww_acquire_ctx ticket;
872 	struct list_head head;
873 	struct amdgpu_ib *ib = NULL;
874 	struct fence *f = NULL;
875 	struct amdgpu_device *adev = ring->adev;
876 	uint64_t addr;
877 	int i, r;
878 
879 	memset(&tv, 0, sizeof(tv));
880 	tv.bo = &bo->tbo;
881 
882 	INIT_LIST_HEAD(&head);
883 	list_add(&tv.head, &head);
884 
885 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
886 	if (r)
887 		return r;
888 
889 	if (!bo->adev->uvd.address_64_bit) {
890 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
891 		amdgpu_uvd_force_into_uvd_segment(bo);
892 	}
893 
894 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
895 	if (r)
896 		goto err;
897 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
898 	if (!ib) {
899 		r = -ENOMEM;
900 		goto err;
901 	}
902 	r = amdgpu_ib_get(ring, NULL, 64, ib);
903 	if (r)
904 		goto err1;
905 
906 	addr = amdgpu_bo_gpu_offset(bo);
907 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
908 	ib->ptr[1] = addr;
909 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
910 	ib->ptr[3] = addr >> 32;
911 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
912 	ib->ptr[5] = 0;
913 	for (i = 6; i < 16; ++i)
914 		ib->ptr[i] = PACKET2(0);
915 	ib->length_dw = 16;
916 
917 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
918 						 &amdgpu_uvd_free_job,
919 						 AMDGPU_FENCE_OWNER_UNDEFINED,
920 						 &f);
921 	if (r)
922 		goto err2;
923 
924 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
925 
926 	if (fence)
927 		*fence = fence_get(f);
928 	amdgpu_bo_unref(&bo);
929 	fence_put(f);
930 	if (amdgpu_enable_scheduler)
931 		return 0;
932 
933 	amdgpu_ib_free(ring->adev, ib);
934 	kfree(ib);
935 	return 0;
936 err2:
937 	amdgpu_ib_free(ring->adev, ib);
938 err1:
939 	kfree(ib);
940 err:
941 	ttm_eu_backoff_reservation(&ticket, &head);
942 	return r;
943 }
944 
945 /* multiple fence commands without any stream commands in between can
946    crash the vcpu so just try to emmit a dummy create/destroy msg to
947    avoid this */
948 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
949 			      struct fence **fence)
950 {
951 	struct amdgpu_device *adev = ring->adev;
952 	struct amdgpu_bo *bo;
953 	uint32_t *msg;
954 	int r, i;
955 
956 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
957 			     AMDGPU_GEM_DOMAIN_VRAM,
958 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
959 			     NULL, NULL, &bo);
960 	if (r)
961 		return r;
962 
963 	r = amdgpu_bo_reserve(bo, false);
964 	if (r) {
965 		amdgpu_bo_unref(&bo);
966 		return r;
967 	}
968 
969 	r = amdgpu_bo_kmap(bo, (void **)&msg);
970 	if (r) {
971 		amdgpu_bo_unreserve(bo);
972 		amdgpu_bo_unref(&bo);
973 		return r;
974 	}
975 
976 	/* stitch together an UVD create msg */
977 	msg[0] = cpu_to_le32(0x00000de4);
978 	msg[1] = cpu_to_le32(0x00000000);
979 	msg[2] = cpu_to_le32(handle);
980 	msg[3] = cpu_to_le32(0x00000000);
981 	msg[4] = cpu_to_le32(0x00000000);
982 	msg[5] = cpu_to_le32(0x00000000);
983 	msg[6] = cpu_to_le32(0x00000000);
984 	msg[7] = cpu_to_le32(0x00000780);
985 	msg[8] = cpu_to_le32(0x00000440);
986 	msg[9] = cpu_to_le32(0x00000000);
987 	msg[10] = cpu_to_le32(0x01b37000);
988 	for (i = 11; i < 1024; ++i)
989 		msg[i] = cpu_to_le32(0x0);
990 
991 	amdgpu_bo_kunmap(bo);
992 	amdgpu_bo_unreserve(bo);
993 
994 	return amdgpu_uvd_send_msg(ring, bo, fence);
995 }
996 
997 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
998 			       struct fence **fence)
999 {
1000 	struct amdgpu_device *adev = ring->adev;
1001 	struct amdgpu_bo *bo;
1002 	uint32_t *msg;
1003 	int r, i;
1004 
1005 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1006 			     AMDGPU_GEM_DOMAIN_VRAM,
1007 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1008 			     NULL, NULL, &bo);
1009 	if (r)
1010 		return r;
1011 
1012 	r = amdgpu_bo_reserve(bo, false);
1013 	if (r) {
1014 		amdgpu_bo_unref(&bo);
1015 		return r;
1016 	}
1017 
1018 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1019 	if (r) {
1020 		amdgpu_bo_unreserve(bo);
1021 		amdgpu_bo_unref(&bo);
1022 		return r;
1023 	}
1024 
1025 	/* stitch together an UVD destroy msg */
1026 	msg[0] = cpu_to_le32(0x00000de4);
1027 	msg[1] = cpu_to_le32(0x00000002);
1028 	msg[2] = cpu_to_le32(handle);
1029 	msg[3] = cpu_to_le32(0x00000000);
1030 	for (i = 4; i < 1024; ++i)
1031 		msg[i] = cpu_to_le32(0x0);
1032 
1033 	amdgpu_bo_kunmap(bo);
1034 	amdgpu_bo_unreserve(bo);
1035 
1036 	return amdgpu_uvd_send_msg(ring, bo, fence);
1037 }
1038 
1039 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1040 {
1041 	struct amdgpu_device *adev =
1042 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1043 	unsigned i, fences, handles = 0;
1044 
1045 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1046 
1047 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1048 		if (atomic_read(&adev->uvd.handles[i]))
1049 			++handles;
1050 
1051 	if (fences == 0 && handles == 0) {
1052 		if (adev->pm.dpm_enabled) {
1053 			amdgpu_dpm_enable_uvd(adev, false);
1054 		} else {
1055 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1056 		}
1057 	} else {
1058 		schedule_delayed_work(&adev->uvd.idle_work,
1059 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1060 	}
1061 }
1062 
1063 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1064 {
1065 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1066 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1067 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1068 
1069 	if (set_clocks) {
1070 		if (adev->pm.dpm_enabled) {
1071 			amdgpu_dpm_enable_uvd(adev, true);
1072 		} else {
1073 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1074 		}
1075 	}
1076 }
1077