xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision b455f0a1cc51f35a9e714554c6e3234fd806ce02)
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 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
56 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
57 
58 /**
59  * amdgpu_uvd_cs_ctx - Command submission parser context
60  *
61  * Used for emulating virtual memory support on UVD 4.2.
62  */
63 struct amdgpu_uvd_cs_ctx {
64 	struct amdgpu_cs_parser *parser;
65 	unsigned reg, count;
66 	unsigned data0, data1;
67 	unsigned idx;
68 	unsigned ib_idx;
69 
70 	/* does the IB has a msg command */
71 	bool has_msg_cmd;
72 
73 	/* minimum buffer sizes */
74 	unsigned *buf_sizes;
75 };
76 
77 #ifdef CONFIG_DRM_AMDGPU_CIK
78 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
79 MODULE_FIRMWARE(FIRMWARE_KABINI);
80 MODULE_FIRMWARE(FIRMWARE_KAVERI);
81 MODULE_FIRMWARE(FIRMWARE_HAWAII);
82 MODULE_FIRMWARE(FIRMWARE_MULLINS);
83 #endif
84 MODULE_FIRMWARE(FIRMWARE_TONGA);
85 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
86 MODULE_FIRMWARE(FIRMWARE_FIJI);
87 MODULE_FIRMWARE(FIRMWARE_STONEY);
88 
89 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
90 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
91 
92 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
93 {
94 	struct amdgpu_ring *ring;
95 	struct amd_sched_rq *rq;
96 	unsigned long bo_size;
97 	const char *fw_name;
98 	const struct common_firmware_header *hdr;
99 	unsigned version_major, version_minor, family_id;
100 	int i, r;
101 
102 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
103 
104 	switch (adev->asic_type) {
105 #ifdef CONFIG_DRM_AMDGPU_CIK
106 	case CHIP_BONAIRE:
107 		fw_name = FIRMWARE_BONAIRE;
108 		break;
109 	case CHIP_KABINI:
110 		fw_name = FIRMWARE_KABINI;
111 		break;
112 	case CHIP_KAVERI:
113 		fw_name = FIRMWARE_KAVERI;
114 		break;
115 	case CHIP_HAWAII:
116 		fw_name = FIRMWARE_HAWAII;
117 		break;
118 	case CHIP_MULLINS:
119 		fw_name = FIRMWARE_MULLINS;
120 		break;
121 #endif
122 	case CHIP_TONGA:
123 		fw_name = FIRMWARE_TONGA;
124 		break;
125 	case CHIP_FIJI:
126 		fw_name = FIRMWARE_FIJI;
127 		break;
128 	case CHIP_CARRIZO:
129 		fw_name = FIRMWARE_CARRIZO;
130 		break;
131 	case CHIP_STONEY:
132 		fw_name = FIRMWARE_STONEY;
133 		break;
134 	default:
135 		return -EINVAL;
136 	}
137 
138 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
139 	if (r) {
140 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
141 			fw_name);
142 		return r;
143 	}
144 
145 	r = amdgpu_ucode_validate(adev->uvd.fw);
146 	if (r) {
147 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
148 			fw_name);
149 		release_firmware(adev->uvd.fw);
150 		adev->uvd.fw = NULL;
151 		return r;
152 	}
153 
154 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
155 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
156 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
157 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
158 	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
159 		version_major, version_minor, family_id);
160 
161 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
162 		 +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
163 	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
164 			     AMDGPU_GEM_DOMAIN_VRAM,
165 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
166 			     NULL, NULL, &adev->uvd.vcpu_bo);
167 	if (r) {
168 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
169 		return r;
170 	}
171 
172 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
173 	if (r) {
174 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
175 		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
176 		return r;
177 	}
178 
179 	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
180 			  &adev->uvd.gpu_addr);
181 	if (r) {
182 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
183 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
184 		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
185 		return r;
186 	}
187 
188 	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
189 	if (r) {
190 		dev_err(adev->dev, "(%d) UVD map failed\n", r);
191 		return r;
192 	}
193 
194 	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
195 
196 	ring = &adev->uvd.ring;
197 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
198 	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
199 				  rq, amdgpu_sched_jobs);
200 	if (r != 0) {
201 		DRM_ERROR("Failed setting up UVD run queue.\n");
202 		return r;
203 	}
204 
205 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
206 		atomic_set(&adev->uvd.handles[i], 0);
207 		adev->uvd.filp[i] = NULL;
208 	}
209 
210 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
211 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
212 		adev->uvd.address_64_bit = true;
213 
214 	return 0;
215 }
216 
217 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
218 {
219 	int r;
220 
221 	if (adev->uvd.vcpu_bo == NULL)
222 		return 0;
223 
224 	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
225 
226 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
227 	if (!r) {
228 		amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
229 		amdgpu_bo_unpin(adev->uvd.vcpu_bo);
230 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
231 	}
232 
233 	amdgpu_bo_unref(&adev->uvd.vcpu_bo);
234 
235 	amdgpu_ring_fini(&adev->uvd.ring);
236 
237 	release_firmware(adev->uvd.fw);
238 
239 	return 0;
240 }
241 
242 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
243 {
244 	unsigned size;
245 	void *ptr;
246 	int i;
247 
248 	if (adev->uvd.vcpu_bo == NULL)
249 		return 0;
250 
251 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
252 		if (atomic_read(&adev->uvd.handles[i]))
253 			break;
254 
255 	if (i == AMDGPU_MAX_UVD_HANDLES)
256 		return 0;
257 
258 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
259 	ptr = adev->uvd.cpu_addr;
260 
261 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
262 	if (!adev->uvd.saved_bo)
263 		return -ENOMEM;
264 
265 	memcpy(adev->uvd.saved_bo, ptr, size);
266 
267 	return 0;
268 }
269 
270 int amdgpu_uvd_resume(struct amdgpu_device *adev)
271 {
272 	unsigned size;
273 	void *ptr;
274 
275 	if (adev->uvd.vcpu_bo == NULL)
276 		return -EINVAL;
277 
278 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
279 	ptr = adev->uvd.cpu_addr;
280 
281 	if (adev->uvd.saved_bo != NULL) {
282 		memcpy(ptr, adev->uvd.saved_bo, size);
283 		kfree(adev->uvd.saved_bo);
284 		adev->uvd.saved_bo = NULL;
285 	} else {
286 		const struct common_firmware_header *hdr;
287 		unsigned offset;
288 
289 		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
290 		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
291 		memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
292 			(adev->uvd.fw->size) - offset);
293 		size -= le32_to_cpu(hdr->ucode_size_bytes);
294 		ptr += le32_to_cpu(hdr->ucode_size_bytes);
295 		memset(ptr, 0, size);
296 	}
297 
298 	return 0;
299 }
300 
301 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
302 {
303 	struct amdgpu_ring *ring = &adev->uvd.ring;
304 	int i, r;
305 
306 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
307 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
308 		if (handle != 0 && adev->uvd.filp[i] == filp) {
309 			struct fence *fence;
310 
311 			amdgpu_uvd_note_usage(adev);
312 
313 			r = amdgpu_uvd_get_destroy_msg(ring, handle,
314 						       false, &fence);
315 			if (r) {
316 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
317 				continue;
318 			}
319 
320 			fence_wait(fence, false);
321 			fence_put(fence);
322 
323 			adev->uvd.filp[i] = NULL;
324 			atomic_set(&adev->uvd.handles[i], 0);
325 		}
326 	}
327 }
328 
329 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
330 {
331 	int i;
332 	for (i = 0; i < rbo->placement.num_placement; ++i) {
333 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
334 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
335 	}
336 }
337 
338 /**
339  * amdgpu_uvd_cs_pass1 - first parsing round
340  *
341  * @ctx: UVD parser context
342  *
343  * Make sure UVD message and feedback buffers are in VRAM and
344  * nobody is violating an 256MB boundary.
345  */
346 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
347 {
348 	struct amdgpu_bo_va_mapping *mapping;
349 	struct amdgpu_bo *bo;
350 	uint32_t cmd, lo, hi;
351 	uint64_t addr;
352 	int r = 0;
353 
354 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
355 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
356 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
357 
358 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
359 	if (mapping == NULL) {
360 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
361 		return -EINVAL;
362 	}
363 
364 	if (!ctx->parser->adev->uvd.address_64_bit) {
365 		/* check if it's a message or feedback command */
366 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
367 		if (cmd == 0x0 || cmd == 0x3) {
368 			/* yes, force it into VRAM */
369 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
370 			amdgpu_ttm_placement_from_domain(bo, domain);
371 		}
372 		amdgpu_uvd_force_into_uvd_segment(bo);
373 
374 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
375 	}
376 
377 	return r;
378 }
379 
380 /**
381  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
382  *
383  * @msg: pointer to message structure
384  * @buf_sizes: returned buffer sizes
385  *
386  * Peek into the decode message and calculate the necessary buffer sizes.
387  */
388 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
389 {
390 	unsigned stream_type = msg[4];
391 	unsigned width = msg[6];
392 	unsigned height = msg[7];
393 	unsigned dpb_size = msg[9];
394 	unsigned pitch = msg[28];
395 	unsigned level = msg[57];
396 
397 	unsigned width_in_mb = width / 16;
398 	unsigned height_in_mb = ALIGN(height / 16, 2);
399 	unsigned fs_in_mb = width_in_mb * height_in_mb;
400 
401 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
402 	unsigned min_ctx_size = 0;
403 
404 	image_size = width * height;
405 	image_size += image_size / 2;
406 	image_size = ALIGN(image_size, 1024);
407 
408 	switch (stream_type) {
409 	case 0: /* H264 */
410 	case 7: /* H264 Perf */
411 		switch(level) {
412 		case 30:
413 			num_dpb_buffer = 8100 / fs_in_mb;
414 			break;
415 		case 31:
416 			num_dpb_buffer = 18000 / fs_in_mb;
417 			break;
418 		case 32:
419 			num_dpb_buffer = 20480 / fs_in_mb;
420 			break;
421 		case 41:
422 			num_dpb_buffer = 32768 / fs_in_mb;
423 			break;
424 		case 42:
425 			num_dpb_buffer = 34816 / fs_in_mb;
426 			break;
427 		case 50:
428 			num_dpb_buffer = 110400 / fs_in_mb;
429 			break;
430 		case 51:
431 			num_dpb_buffer = 184320 / fs_in_mb;
432 			break;
433 		default:
434 			num_dpb_buffer = 184320 / fs_in_mb;
435 			break;
436 		}
437 		num_dpb_buffer++;
438 		if (num_dpb_buffer > 17)
439 			num_dpb_buffer = 17;
440 
441 		/* reference picture buffer */
442 		min_dpb_size = image_size * num_dpb_buffer;
443 
444 		/* macroblock context buffer */
445 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
446 
447 		/* IT surface buffer */
448 		min_dpb_size += width_in_mb * height_in_mb * 32;
449 		break;
450 
451 	case 1: /* VC1 */
452 
453 		/* reference picture buffer */
454 		min_dpb_size = image_size * 3;
455 
456 		/* CONTEXT_BUFFER */
457 		min_dpb_size += width_in_mb * height_in_mb * 128;
458 
459 		/* IT surface buffer */
460 		min_dpb_size += width_in_mb * 64;
461 
462 		/* DB surface buffer */
463 		min_dpb_size += width_in_mb * 128;
464 
465 		/* BP */
466 		tmp = max(width_in_mb, height_in_mb);
467 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
468 		break;
469 
470 	case 3: /* MPEG2 */
471 
472 		/* reference picture buffer */
473 		min_dpb_size = image_size * 3;
474 		break;
475 
476 	case 4: /* MPEG4 */
477 
478 		/* reference picture buffer */
479 		min_dpb_size = image_size * 3;
480 
481 		/* CM */
482 		min_dpb_size += width_in_mb * height_in_mb * 64;
483 
484 		/* IT surface buffer */
485 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
486 		break;
487 
488 	case 16: /* H265 */
489 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
490 		image_size = ALIGN(image_size, 256);
491 
492 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
493 		min_dpb_size = image_size * num_dpb_buffer;
494 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
495 					   * 16 * num_dpb_buffer + 52 * 1024;
496 		break;
497 
498 	default:
499 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
500 		return -EINVAL;
501 	}
502 
503 	if (width > pitch) {
504 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
505 		return -EINVAL;
506 	}
507 
508 	if (dpb_size < min_dpb_size) {
509 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
510 			  dpb_size, min_dpb_size);
511 		return -EINVAL;
512 	}
513 
514 	buf_sizes[0x1] = dpb_size;
515 	buf_sizes[0x2] = image_size;
516 	buf_sizes[0x4] = min_ctx_size;
517 	return 0;
518 }
519 
520 /**
521  * amdgpu_uvd_cs_msg - handle UVD message
522  *
523  * @ctx: UVD parser context
524  * @bo: buffer object containing the message
525  * @offset: offset into the buffer object
526  *
527  * Peek into the UVD message and extract the session id.
528  * Make sure that we don't open up to many sessions.
529  */
530 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
531 			     struct amdgpu_bo *bo, unsigned offset)
532 {
533 	struct amdgpu_device *adev = ctx->parser->adev;
534 	int32_t *msg, msg_type, handle;
535 	void *ptr;
536 	long r;
537 	int i;
538 
539 	if (offset & 0x3F) {
540 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
541 		return -EINVAL;
542 	}
543 
544 	r = amdgpu_bo_kmap(bo, &ptr);
545 	if (r) {
546 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
547 		return r;
548 	}
549 
550 	msg = ptr + offset;
551 
552 	msg_type = msg[1];
553 	handle = msg[2];
554 
555 	if (handle == 0) {
556 		DRM_ERROR("Invalid UVD handle!\n");
557 		return -EINVAL;
558 	}
559 
560 	switch (msg_type) {
561 	case 0:
562 		/* it's a create msg, calc image size (width * height) */
563 		amdgpu_bo_kunmap(bo);
564 
565 		/* try to alloc a new handle */
566 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
567 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
568 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
569 				return -EINVAL;
570 			}
571 
572 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
573 				adev->uvd.filp[i] = ctx->parser->filp;
574 				return 0;
575 			}
576 		}
577 
578 		DRM_ERROR("No more free UVD handles!\n");
579 		return -EINVAL;
580 
581 	case 1:
582 		/* it's a decode msg, calc buffer sizes */
583 		r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
584 		amdgpu_bo_kunmap(bo);
585 		if (r)
586 			return r;
587 
588 		/* validate the handle */
589 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
590 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
591 				if (adev->uvd.filp[i] != ctx->parser->filp) {
592 					DRM_ERROR("UVD handle collision detected!\n");
593 					return -EINVAL;
594 				}
595 				return 0;
596 			}
597 		}
598 
599 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
600 		return -ENOENT;
601 
602 	case 2:
603 		/* it's a destroy msg, free the handle */
604 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
605 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
606 		amdgpu_bo_kunmap(bo);
607 		return 0;
608 
609 	default:
610 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
611 		return -EINVAL;
612 	}
613 	BUG();
614 	return -EINVAL;
615 }
616 
617 /**
618  * amdgpu_uvd_cs_pass2 - second parsing round
619  *
620  * @ctx: UVD parser context
621  *
622  * Patch buffer addresses, make sure buffer sizes are correct.
623  */
624 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
625 {
626 	struct amdgpu_bo_va_mapping *mapping;
627 	struct amdgpu_bo *bo;
628 	uint32_t cmd, lo, hi;
629 	uint64_t start, end;
630 	uint64_t addr;
631 	int r;
632 
633 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
634 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
635 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
636 
637 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
638 	if (mapping == NULL)
639 		return -EINVAL;
640 
641 	start = amdgpu_bo_gpu_offset(bo);
642 
643 	end = (mapping->it.last + 1 - mapping->it.start);
644 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
645 
646 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
647 	start += addr;
648 
649 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
650 			    lower_32_bits(start));
651 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
652 			    upper_32_bits(start));
653 
654 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
655 	if (cmd < 0x4) {
656 		if ((end - start) < ctx->buf_sizes[cmd]) {
657 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
658 				  (unsigned)(end - start),
659 				  ctx->buf_sizes[cmd]);
660 			return -EINVAL;
661 		}
662 
663 	} else if (cmd == 0x206) {
664 		if ((end - start) < ctx->buf_sizes[4]) {
665 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
666 					  (unsigned)(end - start),
667 					  ctx->buf_sizes[4]);
668 			return -EINVAL;
669 		}
670 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
671 		DRM_ERROR("invalid UVD command %X!\n", cmd);
672 		return -EINVAL;
673 	}
674 
675 	if (!ctx->parser->adev->uvd.address_64_bit) {
676 		if ((start >> 28) != ((end - 1) >> 28)) {
677 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
678 				  start, end);
679 			return -EINVAL;
680 		}
681 
682 		if ((cmd == 0 || cmd == 0x3) &&
683 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
684 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
685 				  start, end);
686 			return -EINVAL;
687 		}
688 	}
689 
690 	if (cmd == 0) {
691 		ctx->has_msg_cmd = true;
692 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
693 		if (r)
694 			return r;
695 	} else if (!ctx->has_msg_cmd) {
696 		DRM_ERROR("Message needed before other commands are send!\n");
697 		return -EINVAL;
698 	}
699 
700 	return 0;
701 }
702 
703 /**
704  * amdgpu_uvd_cs_reg - parse register writes
705  *
706  * @ctx: UVD parser context
707  * @cb: callback function
708  *
709  * Parse the register writes, call cb on each complete command.
710  */
711 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
712 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
713 {
714 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
715 	int i, r;
716 
717 	ctx->idx++;
718 	for (i = 0; i <= ctx->count; ++i) {
719 		unsigned reg = ctx->reg + i;
720 
721 		if (ctx->idx >= ib->length_dw) {
722 			DRM_ERROR("Register command after end of CS!\n");
723 			return -EINVAL;
724 		}
725 
726 		switch (reg) {
727 		case mmUVD_GPCOM_VCPU_DATA0:
728 			ctx->data0 = ctx->idx;
729 			break;
730 		case mmUVD_GPCOM_VCPU_DATA1:
731 			ctx->data1 = ctx->idx;
732 			break;
733 		case mmUVD_GPCOM_VCPU_CMD:
734 			r = cb(ctx);
735 			if (r)
736 				return r;
737 			break;
738 		case mmUVD_ENGINE_CNTL:
739 			break;
740 		default:
741 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
742 			return -EINVAL;
743 		}
744 		ctx->idx++;
745 	}
746 	return 0;
747 }
748 
749 /**
750  * amdgpu_uvd_cs_packets - parse UVD packets
751  *
752  * @ctx: UVD parser context
753  * @cb: callback function
754  *
755  * Parse the command stream packets.
756  */
757 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
758 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
759 {
760 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
761 	int r;
762 
763 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
764 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
765 		unsigned type = CP_PACKET_GET_TYPE(cmd);
766 		switch (type) {
767 		case PACKET_TYPE0:
768 			ctx->reg = CP_PACKET0_GET_REG(cmd);
769 			ctx->count = CP_PACKET_GET_COUNT(cmd);
770 			r = amdgpu_uvd_cs_reg(ctx, cb);
771 			if (r)
772 				return r;
773 			break;
774 		case PACKET_TYPE2:
775 			++ctx->idx;
776 			break;
777 		default:
778 			DRM_ERROR("Unknown packet type %d !\n", type);
779 			return -EINVAL;
780 		}
781 	}
782 	return 0;
783 }
784 
785 /**
786  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
787  *
788  * @parser: Command submission parser context
789  *
790  * Parse the command stream, patch in addresses as necessary.
791  */
792 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
793 {
794 	struct amdgpu_uvd_cs_ctx ctx = {};
795 	unsigned buf_sizes[] = {
796 		[0x00000000]	=	2048,
797 		[0x00000001]	=	0xFFFFFFFF,
798 		[0x00000002]	=	0xFFFFFFFF,
799 		[0x00000003]	=	2048,
800 		[0x00000004]	=	0xFFFFFFFF,
801 	};
802 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
803 	int r;
804 
805 	if (ib->length_dw % 16) {
806 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
807 			  ib->length_dw);
808 		return -EINVAL;
809 	}
810 
811 	ctx.parser = parser;
812 	ctx.buf_sizes = buf_sizes;
813 	ctx.ib_idx = ib_idx;
814 
815 	/* first round, make sure the buffers are actually in the UVD segment */
816 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
817 	if (r)
818 		return r;
819 
820 	/* second round, patch buffer addresses into the command stream */
821 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
822 	if (r)
823 		return r;
824 
825 	if (!ctx.has_msg_cmd) {
826 		DRM_ERROR("UVD-IBs need a msg command!\n");
827 		return -EINVAL;
828 	}
829 
830 	amdgpu_uvd_note_usage(ctx.parser->adev);
831 
832 	return 0;
833 }
834 
835 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
836 			       bool direct, struct fence **fence)
837 {
838 	struct ttm_validate_buffer tv;
839 	struct ww_acquire_ctx ticket;
840 	struct list_head head;
841 	struct amdgpu_job *job;
842 	struct amdgpu_ib *ib;
843 	struct fence *f = NULL;
844 	struct amdgpu_device *adev = ring->adev;
845 	uint64_t addr;
846 	int i, r;
847 
848 	memset(&tv, 0, sizeof(tv));
849 	tv.bo = &bo->tbo;
850 
851 	INIT_LIST_HEAD(&head);
852 	list_add(&tv.head, &head);
853 
854 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
855 	if (r)
856 		return r;
857 
858 	if (!bo->adev->uvd.address_64_bit) {
859 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
860 		amdgpu_uvd_force_into_uvd_segment(bo);
861 	}
862 
863 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
864 	if (r)
865 		goto err;
866 
867 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
868 	if (r)
869 		goto err;
870 
871 	ib = &job->ibs[0];
872 	addr = amdgpu_bo_gpu_offset(bo);
873 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
874 	ib->ptr[1] = addr;
875 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
876 	ib->ptr[3] = addr >> 32;
877 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
878 	ib->ptr[5] = 0;
879 	for (i = 6; i < 16; ++i)
880 		ib->ptr[i] = PACKET2(0);
881 	ib->length_dw = 16;
882 
883 	if (direct) {
884 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
885 		job->fence = f;
886 		if (r)
887 			goto err_free;
888 
889 		amdgpu_job_free(job);
890 	} else {
891 		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
892 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
893 		if (r)
894 			goto err_free;
895 	}
896 
897 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
898 
899 	if (fence)
900 		*fence = fence_get(f);
901 	amdgpu_bo_unref(&bo);
902 	fence_put(f);
903 
904 	return 0;
905 
906 err_free:
907 	amdgpu_job_free(job);
908 
909 err:
910 	ttm_eu_backoff_reservation(&ticket, &head);
911 	return r;
912 }
913 
914 /* multiple fence commands without any stream commands in between can
915    crash the vcpu so just try to emmit a dummy create/destroy msg to
916    avoid this */
917 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
918 			      struct fence **fence)
919 {
920 	struct amdgpu_device *adev = ring->adev;
921 	struct amdgpu_bo *bo;
922 	uint32_t *msg;
923 	int r, i;
924 
925 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
926 			     AMDGPU_GEM_DOMAIN_VRAM,
927 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
928 			     NULL, NULL, &bo);
929 	if (r)
930 		return r;
931 
932 	r = amdgpu_bo_reserve(bo, false);
933 	if (r) {
934 		amdgpu_bo_unref(&bo);
935 		return r;
936 	}
937 
938 	r = amdgpu_bo_kmap(bo, (void **)&msg);
939 	if (r) {
940 		amdgpu_bo_unreserve(bo);
941 		amdgpu_bo_unref(&bo);
942 		return r;
943 	}
944 
945 	/* stitch together an UVD create msg */
946 	msg[0] = cpu_to_le32(0x00000de4);
947 	msg[1] = cpu_to_le32(0x00000000);
948 	msg[2] = cpu_to_le32(handle);
949 	msg[3] = cpu_to_le32(0x00000000);
950 	msg[4] = cpu_to_le32(0x00000000);
951 	msg[5] = cpu_to_le32(0x00000000);
952 	msg[6] = cpu_to_le32(0x00000000);
953 	msg[7] = cpu_to_le32(0x00000780);
954 	msg[8] = cpu_to_le32(0x00000440);
955 	msg[9] = cpu_to_le32(0x00000000);
956 	msg[10] = cpu_to_le32(0x01b37000);
957 	for (i = 11; i < 1024; ++i)
958 		msg[i] = cpu_to_le32(0x0);
959 
960 	amdgpu_bo_kunmap(bo);
961 	amdgpu_bo_unreserve(bo);
962 
963 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
964 }
965 
966 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
967 			       bool direct, struct fence **fence)
968 {
969 	struct amdgpu_device *adev = ring->adev;
970 	struct amdgpu_bo *bo;
971 	uint32_t *msg;
972 	int r, i;
973 
974 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
975 			     AMDGPU_GEM_DOMAIN_VRAM,
976 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
977 			     NULL, NULL, &bo);
978 	if (r)
979 		return r;
980 
981 	r = amdgpu_bo_reserve(bo, false);
982 	if (r) {
983 		amdgpu_bo_unref(&bo);
984 		return r;
985 	}
986 
987 	r = amdgpu_bo_kmap(bo, (void **)&msg);
988 	if (r) {
989 		amdgpu_bo_unreserve(bo);
990 		amdgpu_bo_unref(&bo);
991 		return r;
992 	}
993 
994 	/* stitch together an UVD destroy msg */
995 	msg[0] = cpu_to_le32(0x00000de4);
996 	msg[1] = cpu_to_le32(0x00000002);
997 	msg[2] = cpu_to_le32(handle);
998 	msg[3] = cpu_to_le32(0x00000000);
999 	for (i = 4; i < 1024; ++i)
1000 		msg[i] = cpu_to_le32(0x0);
1001 
1002 	amdgpu_bo_kunmap(bo);
1003 	amdgpu_bo_unreserve(bo);
1004 
1005 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1006 }
1007 
1008 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1009 {
1010 	struct amdgpu_device *adev =
1011 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1012 	unsigned i, fences, handles = 0;
1013 
1014 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1015 
1016 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1017 		if (atomic_read(&adev->uvd.handles[i]))
1018 			++handles;
1019 
1020 	if (fences == 0 && handles == 0) {
1021 		if (adev->pm.dpm_enabled) {
1022 			amdgpu_dpm_enable_uvd(adev, false);
1023 		} else {
1024 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1025 		}
1026 	} else {
1027 		schedule_delayed_work(&adev->uvd.idle_work,
1028 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1029 	}
1030 }
1031 
1032 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1033 {
1034 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1035 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1036 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1037 
1038 	if (set_clocks) {
1039 		if (adev->pm.dpm_enabled) {
1040 			amdgpu_dpm_enable_uvd(adev, true);
1041 		} else {
1042 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1043 		}
1044 	}
1045 }
1046