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